Datasets:

kilian-group
/

arxiv-classifier

Dataset card Data Studio Files Files and versions Community

Subset (4)

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

Split (2)

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

paper_id stringlengths 9 16	version stringclasses 26 values	yymm stringclasses 311 values	created timestamp[s]	title stringlengths 6 335	secondary_subfield sequencelengths 1 8	abstract stringlengths 25 3.93k	primary_subfield stringclasses 124 values	field stringclasses 20 values	fulltext stringlengths 0 2.84M
1312.4646	3	1312	2015-12-15T10:26:56	K-homological finiteness and hyperbolic groups	[ "math.OA", "math.GR", "math.KT" ]	Motivated by classical facts concerning closed manifolds, we introduce a strong finiteness property in K-homology. We say that a C-algebra has uniformly summable K-homology if all its K-homology classes can be represented by Fredholm modules which are finitely summable over the same dense subalgebra, and with the same degree of summability. We show that two types of C-algebras associated to hyperbolic groups - the C-crossed product for the boundary action, and the reduced group C-algebra - have uniformly summable K-homology. We provide explicit summability degrees, as well as explicit finitely summable representatives for the K-homology classes.	math.OA	math	K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS HEATH EMERSON AND BOGDAN NICA Abstract. Motivated by classical facts concerning closed manifolds, we introduce a strong finiteness property in K-homology. We say that a C-algebra has uniformly summable K- homology if all its K-homology classes can be represented by Fredholm modules which are finitely summable over the same dense subalgebra, and with the same degree of summability. We show that two types of C-algebras associated to hyperbolic groups – the C-crossed product for the boundary action, and the reduced group C-algebra – have uniformly sum- mable K-homology. We provide explicit summability degrees, as well as explicit finitely summable representatives for the K-homology classes. Introduction A Fredholm module over a unital C-algebra A consists of a representation π : A → B(H) of A on a Hilbert space, and an operator T ∈ B(H), which in the even case is an essential unitary and in the odd case an essential projection, and which in both cases is required to essentially commute with the representation: [π(a), T ] ∈ K(H) for all a ∈ A. The definition of Fredholm module was motivated by elliptic operator theory: if M is a smooth, compact manifold, then any zero-order elliptic pseudodifferential operator T on sections L2(E) of a bundle over M , determines a Fredholm module with the obvious representation of C(M ) on L2(E) by multiplication. Since elliptic operators induce maps on K-theory by an index theoretic construction, this led Atiyah, and subsequently Kasparov, to describe the K-homology of a C- algebra A as equivalence classes of Fredholm modules over A. Connes' early work on cyclic cohomology, the noncommutative analogue of de Rham the- ory, and on the noncommutative Chern character, a map from K-homology to cyclic theory, suggested the importance of the finite summability condition on a Fredholm module that [π(a), T ] ∈ Lp(H) for dense a ∈ A where Lp(H) is the Schatten ideal of p-summable compact operators. Connes showed how to associate a canonical cyclic cocycle, the so-called character, to a finitely summable Fredholm module, and how to use the character for computing the index pairing between the K-theory of A and the K-homology class of the Fredholm module. This is just one aspect of a larger landscape, that of quantized calculus [7, Ch.IV], depending on finite summability. Examples of finitely summable Fredholm modules over C-algebras are thus of considerable interest in noncommutative geometry, and by this stage there have been numerous construc- tions of them, but one can state a theorem about their existence in the classical situation, using elliptic operator theory and Poincar´e duality. If M is a closed manifold, then M has Poincar´e duality with its tangent bundle. This implies that every K-homology class for M is represented by a pseudodifferential operator of order zero. Classical spectral estimates for pseudodifferential operators imply that the singular values of commutators [f, T ], where f is Date: July 19, 2018. 2000 Mathematics Subject Classification. 19K33, 20F67, 58B34. Key words and phrases. K-homology, finitely summable Fredholm modules, boundaries of hyperbolic groups. 1 2 HEATH EMERSON AND BOGDAN NICA a smooth function and T is pseudodifferential, satisfy the asymptotic law sn = O(n−1/ dim M ). It follows from this that every K-homology class for M is represented by a Fredholm module which is p-summable over C∞(M ) for each p > dim M . The moral is that the K-homology of a closed manifold exhibits finiteness, in the sense that all K-homology classes can be represented by finitely summable Fredholm modules, in a strong form: both the smooth subalgebra and the degree of summability may be taken uniform across all K-homology classes. The main focus of the present article is on noncommutative manifestations of this strong finiteness phenomenon. We say that a C-algebra has uniformly summable K-homology if all its K-homology classes can be represented by Fredholm modules which are finitely summable over the same smooth subalgebra, and with the same degree of summability. Theorem A. Let Γ be a regular, torsion-free hyperbolic group, and let ∂Γ denote its boundary. Then the crossed-product C-algebra C(∂Γ) ⋊ Γ has uniformly summable K-homology. The summability degree is the Hausdorff dimension of the boundary, more precisely a suit- able interpretation thereof, and it is obtained by analytic means from the Γ-invariant Holder structure of the boundary. The same structure is responsible for the smooth subalgebra. Fredholm modules for the crossed-product C-algebra C(∂Γ) ⋊ Γ restrict to Fredholm mod- ules for the reduced group C-algebra C∗ r Γ, preserving summability. However, additional effort and new ingredients are needed in order to represent all the K-homology classes of C∗ r Γ. Our main result in this direction, Theorem 1.5 below, is somewhat technical due to certain issues connected with the Baum - Connes conjecture. At this point, we merely quote the following concrete consequence. Theorem B. Let Γ be one of the following: • a finitely generated free group; • a torsion-free cocompact lattice in SO(n, 1) or SU(n, 1); • a torsion-free C′(1/6) small-cancellation group with one more generator than relators. r Γ has uniformly summable K-homology over the group Then the reduced group C-algebra C∗ algebra CΓ. Hyperbolicity, in the sense of Gromov, is a coarse notion of negative curvature. A hyperbolic space is a geodesic metric space all of whose geodesic triangles are uniformly thin. A group is said to be hyperbolic if it admits a geometric - that is, isometric, proper, and cocompact - action on a hyperbolic space. Fundamental examples of hyperbolic groups are finitely generated free groups, cocompact lattices in SO(n, 1), SU(n, 1), or Sp(n, 1) ('classical hyperbolic groups'), as well as C′(1/6) small-cancellation groups. Throughout this paper, hyperbolic groups are assumed to be non-elementary, meaning that we discard the virtually cyclic groups. The boundary ∂Γ of a hyperbolic group Γ is a compact Hausdorff space carrying a natural action of Γ by homeomorphisms. Our standing assumption that Γ is non-elementary translates into ∂Γ having uncountably many points. For instance, the boundary of a free group is a Cantor set, and the boundary of a classical hyperbolic group is a sphere. The Γ-action on ∂Γ is topologically amenable [1]. This means that the C-algebra C(∂Γ) ⋊ Γ is nuclear, and that the full and the reduced crossed products coincide [3]. The boundary compactification Γ = Γ ∪ ∂Γ carries a Γ-action as well, and it is a coarse compactification – in the sense that a ball of uniform size in Γ becomes small in the topology of the compact space Γ when translated out to the boundary. The action of Γ on the boundary ∂Γ is minimal and exhibits a north- south dynamics, making the C-algebra crossed-product C(∂Γ) ⋊ Γ simple and purely infinite [2, 30]. If Γ is a free group, then C(∂Γ) ⋊ Γ is canonically a Cuntz - Krieger algebra with transition matrix simply coding that a generator cannot be followed by its inverse. If Γ is a K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 3 surface group, Γ = π1(M ) where M is a closed surface of genus at least 2, then ∂Γ is the 1-sphere S1 and the groupoid Γ ⋉ ∂Γ is strongly Morita equivalent to the holonomy groupoid of the canonical foliation of M ×Γ S1 by the projections to M ×Γ S1 of copies of M . The K-theory of the boundary crossed-product C(∂Γ) ⋊ Γ has been investigated in [11]. The main result of this reference is a Gysin sequence for boundary actions which computes the map on K-theory induced by the inclusion i : C∗ r Γ → C(∂Γ) ⋊ Γ. Its K-homology version is described in this article, and it plays an important role in our results. Another K-theoretic feature of C(∂Γ) ⋊ Γ is that it exhibits Poincar´e self-duality: the K-theory and the K-homology of C(∂Γ) ⋊ Γ are isomorphic with a parity shift [10]. The isomorphism is implemented by a cup-cap product map ∆∩ : K∗(C(∂Γ) ⋊ Γ) → K∗+1(C(∂Γ) ⋊ Γ) with a suitable class ∆ ∈ K1((C(∂Γ) ⋊ Γ) ⊗ (C(∂Γ) ⋊ Γ)). For the purposes of this paper, the important feature of Poincar´e self-duality is the surjectivity of ∆∩, which leads to an explicit description of the K-homology of C(∂Γ) ⋊ Γ as a function of its K-theory. The proof of Poincar´e duality given in [10] requires Γ to be torsion-free, and its boundary ∂Γ to admit a continuous self-map without any fixed points. This technical condition on the boundary is the regularity assumption in Theorem A. We do not know of any hyperbolic group which fails to be regular. Our rank-1 results in Theorem B are in stark contrast to the higher rank situation. As shown by Puschnigg [36], no non-trivial K-homology class for the reduced C-algebra of a higher rank lattice can be represented by a Fredholm module which is finitely summable over the group algebra. This very opposite behaviour with respect to K-homological finiteness is reminiscent of the following sharp distinction between rank-1 lattices and higher rank lattices, also involving a notion of finite summability. Every hyperbolic group admits a proper isometric action on an Lp-space for large enough p > 1 [42], [35], but every isometric action of a higher rank lattice on an Lp-space, p > 1, fixes a point [4]. It should be noted that, despite our strong finiteness results at the level of Fredholm modules (i.e., bounded K-cycles), neither the boundary crossed-product C(∂Γ) ⋊ Γ nor the reduced C- algebra C∗ r Γ support any finitely summable spectral triples (i.e., unbounded K-cycles). For C(∂Γ) ⋊ Γ this is due to the lack of a trace [6, Thm.8], whereas for C∗ r Γ the reason is the non-amenability of Γ [6, Thm.19], [7, Thm.1 in IV.9.α]. The present work expands, and supersedes, our preprint [14]. Other results. Lott [31] employed elliptic operator methods to construct K-homology cycles, some of which are finitely summable, for the crossed-product C-algebra arising from the action of a subgroup of SO(n, 1) on its limit set. Lott's constructions are not obviously related to ours. Rave [37] proved that AF C-algebras have, in the terminology of this paper, uniformly summable K-homology. Incidentally, [37] also contains a good account of the fact that the commutative C-algebra C(M ) of a closed manifold M has uniformly summable K-homology. Very recently, Goffeng and Mesland [16] have addressed the issue of uniform summability for K-homology in the case of Cuntz - Krieger C-algebras. This is a family which bears some analogies to the boundary C-crossed products considered herein. In [16] it is shown, among other things, that the odd K-homology of a Cuntz - Krieger C-algebra is uniformly summable. Acknowledgements. We thank Nigel Higson, Vadim Kaimanovich, Misha Kapovich, Bruce Kleiner, Georges Skandalis, and Bob Yuncken for correspondence or discussions. The first author acknowledges support from an NSERC Discovery grant. The second author thanks the Pacific Institute for the Mathematical Sciences, and the Alexander von Humboldt Foundation for their support. We also thank the last referee for a careful reading of the paper, and for constructive comments. 4 HEATH EMERSON AND BOGDAN NICA 1. Outline We now describe our results and our approach in more detail. Let us start with a con- ceptual clarification of what exactly is the boundary of a hyperbolic group. Having fixed a (non-elementary) hyperbolic group Γ, by a geometric model for Γ we mean a hyperbolic space on which Γ acts geometrically. Some groups come with ready-made geometric models, e.g., for a cocompact lattice in SO(n, 1) the n-dimensional real hyperbolic space Hn is such a space. Otherwise, a geometric model can be manufactured as the Cayley graph with respect to a finite generating set - for instance, free groups admit regular trees as geometric models. The important point is that boundaries of geometric models for Γ are Γ-equivariantly homeomor- phic. Thus, the boundary of Γ should be understood as the boundary of any geometric model for Γ. Let X be a geometric model for Γ, so ∂X is a topological avatar of the boundary of Γ. There is a natural collection of visual metrics on ∂X, all assigning a finite Hausdorff dimension to ∂X. The visual dimension of the boundary, denoted visdim ∂X, is the infimal Hausdorff dimension over the family of visual metrics. In particular, the visual dimension is at least as large as the topological dimension. The Hausdorff measures defined by visual metrics are comparable, in the sense that they are within constant multiples of each other. This prompts us to define a visual probability measure on ∂X as a Borel probability measure which is comparable with some (equivalently, each) Hausdorff measure defined by a visual metric. Consider now the crossed product C(∂X)⋊Γ, a C-avatar of C(∂Γ)⋊Γ. Endowing ∂X with a visual probability measure µ, we obtain a faithful representation of C(∂X) on L2(∂X, µ) by multiplication. This induces, in turn, a faithful representation λµ, the left regular representation with respect to µ, of C(∂X) ⋊ Γ on ℓ2(Γ, L2(∂X, µ)). We also let Pℓ2Γ be the projection onto ℓ2Γ, regarded as constant functions on ∂X. The basic idea for our construction of Fredholm modules, and the relationship to Poincar´e self-duality for C(∂Γ) ⋊ Γ begins with the following result. Theorem 1.1 (Basic K-cycle). With the above notations, the pair (cid:0)λµ, Pℓ2Γ(cid:1) is an odd Fredholm module over C(∂Γ) ⋊ Γ. Moreover, (λµ, Pℓ2Γ) is p-summable for every p > max{2, visdim ∂X}, and it represents the Poincar´e dual ∆ ∩ [1] ∈ K1(C(∂Γ) ⋊ Γ) of the unit class [1] ∈ K0(C(∂Γ) ⋊ Γ). A certain compatibility between the constructions going into Theorem 1.1 and the Poincar´e duality of [10] implies that one can 'twist' the basic K-cycle above with projections or unitaries in C(∂Γ) ⋊ Γ in a certain way, generalizing Theorem 1.1 to cover arbitrary K-homology classes – leading to the following essential fact about K-homology classes for C(∂Γ) ⋊ Γ. Theorem 1.2 (Twisted K-cycles). Let Γ be regular and torsion-free. Then the following hold. • Every class in K1(C(∂Γ) ⋊ Γ) is represented by an odd Fredholm module of the form e projection in C(∂Γ) ⋊ Γ. Moreover, the projection e ∈ C(∂Γ) ⋊ Γ can be chosen so that the Fredholm module is p-summable for every p > max{2, visdim ∂X}. • Every class in K0(C(∂Γ) ⋊ Γ) is represented by a balanced even Fredholm module of the form (cid:0)λµ, Pℓ2Γλop µ (e)Pℓ2Γ(cid:1), (cid:0)λµ, Pℓ2Γλop µ (u)Pℓ2Γ + (1 − Pℓ2Γ)(cid:1), u unitary in C(∂Γ) ⋊ Γ. Moreover, the unitary u ∈ C(∂Γ) ⋊ Γ can be chosen so that the Fredholm module is p- summable for every p > max{2, visdim ∂X}. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 5 Here λop µ is the right regular representation of C(∂X) ⋊ Γ on ℓ2(Γ, L2(∂X, µ)), the conjugate of λµ by an appropriate self-adjoint unitary J : C(∂X) ⋊ Γ → C(∂X) ⋊ Γ. While all projections and all unitaries in C(∂Γ) ⋊ Γ yield Fredholm modules as above, one needs to restrict to a suitable smooth subalgebra in order to get finite summability. For each p, the twisted K-cycles are p-summable over one and the same dense ∗-subalgebra, the algebraic crossed-product Lip(∂X, d) ⋊alg Γ where d is a visual metric on ∂X of Hausdorff dimension at most p. We thus obtain Theorem A, in the following more precise form. Theorem 1.3 (Uniform summability). Let Γ be regular and torsion-free. Then the K-homology of C(∂Γ) ⋊ Γ is uniformly p-summable for every p > max{2, visdim ∂X}. We now turn our attention to the K-homology of the reduced C∗-algebra C∗ r Γ. Here issues related to the Baum - Connes conjecture mean that, in general, our methods only yield results about the 'γ-part' of the K-homology of C∗ r Γ. The key tool is the following Gysin sequence, which computes the restriction map i∗ : K∗(C(∂Γ) ⋊ Γ) → γK∗(C∗ r Γ) on K-homology induced by the inclusion i : C∗ r Γ → C(∂Γ) ⋊ Γ. This sequence is the K-homology version of the one in [11], which computes the map i∗ : K∗(C∗ r Γ) → K∗(C(∂Γ) ⋊ Γ) induced by i on K-theory. Theorem 1.4 (Gysin sequence for K-homology). Let Γ be torsion-free. Let γ ∈ KKΓ be the γ-element for Γ, and γK∗(C∗ Then there is an exact sequence r Γ) the corresponding summand of the K-homology of C∗ 0 (C, C) r Γ. 0 −→ K1(BΓ) −→ K0(C(∂Γ) ⋊ Γ) −→ γK0(C∗ r Γ) i∗ i∗ r Γ) −→ 0 where Eul is the map Eul(a) = χ(Γ) index(a) [pnt] ∈ K0(BΓ), and where index is the ordinary Fredholm index map KKΓ(C, C) → Z, [pnt] the K-homology class of a point. K0(BΓ) −→ K1(C(∂Γ) ⋊ Γ) −→ γK1(C∗ We note that the torsion assumption could be dropped, at the expense of elaborating the sequence in the way that was done in [11]. The Gysin sequence, combined with Theorem 1.2, yields the following. Theorem 1.5 (Twisted K-cycles over the reduced C-algebra). Let Γ be regular and torsion- free. Then the following hold. • Every class in the γ-part γK1(C∗ odd Fredholm module of the form r Γ) of the odd K-homology of C∗ r Γ is represented by an yEul (cid:0)λ, Pℓ2Γλop µ (e)Pℓ2Γ(cid:1), e projection in C(∂Γ) ⋊ Γ. Moreover, the projection e ∈ C(∂Γ) ⋊ Γ can be chosen so that the Fredholm module is p-summable over CΓ for every p > max{2, visdim ∂X}. • If χ(Γ) = 0, then, similarly, every class in the γ-part γK0(C∗ balanced even Fredholm module of the form r Γ) is represented by a (cid:0)λ, Pℓ2Γλop µ (u)Pℓ2Γ(cid:1), u unitary in C(∂Γ) ⋊ Γ. Moreover, the unitary u ∈ C(∂Γ) ⋊ Γ can be chosen so that the Fredholm module is p- summable over CΓ for every p > max{2, visdim ∂X}. γK0(C∗ module as above. r Γ) is a reduced γ-element, then every class in the γ-part r Γ) is, up to an integral multiple of γr, represented by a balanced even Fredholm • If χ(Γ) 6= 0 and if γr ∈ γK0(C∗ 6 HEATH EMERSON AND BOGDAN NICA Here λ denotes, as usual, the regular representation of Γ. A 'reduced' γ-element is roughly the same as a γ-element (a class in K0(C∗Γ) which factors in a certain way), but one which is defined over C∗ r Γ rather than C∗Γ. We do not know whether, in general, there exists a reduced γ-element with a finitely sum- mable representative, for general hyperbolic groups. We also do not know whether, in general, the γ-element acts as the identity on K∗(C∗ r Γ). But for the class of a-T-menable groups we do know, thanks to Higson - Kasparov [18], that γ = 1. Specializing Theorem 1.5 to this class, we obtain: Theorem 1.6 (Uniform summability for a-T-menable groups). Assume that Γ is regular, torsion-free, and a-T-menable. Then the odd K-homology K1(C∗ r Γ) is uniformly p-summable over CΓ for every p > max{2, visdim ∂X}. If χ(Γ) = 0, then the even K-homology K0(C∗ r Γ) is uniformly p-summable over CΓ for every p > max{2, visdim ∂X}. If the γ-element γr ∈ K0(C∗ r Γ) can be represented by a p(γr)-summable Fredholm module over CΓ, then the even K- homology K0(C∗ r Γ) is uniformly p-summable over CΓ for every p > max{2, visdim ∂X, p(γr)}. A-T-menable hyperbolic groups include finitely generated free groups, cocompact lattices in SO(n, 1) and SU(n, 1), and C′(1/6) small-cancellation groups – the latter by [41]. Applying the previous theorem to each one of these classes, we obtain the following consequences. Corollary 1.7. Let Γ be a finitely generated free group. Then the K-homology of C∗ uniformly p-summable over CΓ for every p > 2. r Γ is Corollary 1.8. If Γ is a torsion-free cocompact lattice in SO(n, 1), then the K-homology of C∗ r Γ is uniformly n+-summable over CΓ when n ≥ 3, respectively p-summable over CΓ for every p > 2, when n = 2. If Γ is a torsion-free cocompact lattice in SU(n, 1), then the K-homology of C∗ r Γ is uniformly (2n)+-summable over CΓ. These two corollaries rely on the existence of finitely summable representatives for the γ- element, due to Julg - Valette [21] in the free group case, Kasparov [25] in the SO(n, 1) case, respectively Julg - Kasparov [20] in the SU(n, 1) case. For small-cancellation groups, the outcome is less satisfactory. We have to apply the vanishing Euler characteristic criterion of Theorem 1.6, as we are lacking information on the finite summability of the γ-element, and we also do not have an explicit formula for the visual dimension. Corollary 1.9. Let Γ be a torsion-free group given by a C′(1/6) presentation hS Ri. Then the odd K-homology K1(C∗ r Γ) is uniformly summable over CΓ, and the same is true for the even K-homology K0(C∗ r Γ) provided that S − R = 1. Further problems. Our work calls attention to the following questions. The first two have already been mentioned. r Γ)? reduced γ-element γr ∈ K0(C∗ • Let Γ be a hyperbolic group. Does there exist a finitely summable representative for a • Let Γ be a hyperbolic group. Does γ act as the identity on the K-homology of C∗ • Let Γ be a hyperbolic group. Is the K-homology of C∗ We conjecture the answer to the first problem to be positive for fundamental groups of compact, negatively curved manifolds. It is quite plausible for the answer to be positive in general. Note that our results cover the case when χ(Γ) = 0. r Γ uniformly summable over CΓ? r Γ? The second problem may well have a negative answer. Positive solutions to the first two problems also settle the third. But it is entirely possible that the third problem can be attacked from a different perspective. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 7 2. Preliminaries on K-homology 2.1. Fredholm modules and K-homology. We recall some definitions, while fixing nota- tions along the way. For further details, we refer to Connes [7, Ch.IV] and Higson - Roe [19, Ch.8]. Let A be a unital C-algebra. As usual, B(H) and K(H) denote the bounded operators, respectively the ideal of compact operators on a (separable) Hilbert space H. Definition 2.1 (Atiyah, Kasparov). An odd Fredholm module for A is a pair (π, P ), where π : A → B(H) is a representation, P : H → H is an essential projection in the sense that P ∗ − P, P 2 − P ∈ K(H), and such that [P, π(a)] = P π(a) − π(a)P ∈ K(H) for all a ∈ A. An even Fredholm module for A is a pair (π±, U ), where π± : A → B(H±) are representa- tions, U : H+ → H− is an essential unitary in the sense that U ∗U − 1 ∈ K(H+), U U ∗ − 1 ∈ K(H−), and such that π+(a) − U ∗π−(a)U ∈ K(H+) for all a ∈ A. If H+ = H− =: H and π+ = π− =: π, then we say that the even Fredholm module is balanced, and we simply write it (π, U ). Fredholm modules are the cycles in Kasparov's K-homology groups, and for that reason they are also called K-cycles. Here is an outline of the odd case, leading to the odd K-homology group K1(A). The equivalence relation defined by Kasparov on odd Fredholm modules is generated by unitary equivalence, operator homotopy, and addition of degenerates. Unitary equivalence has the obvious meaning. Two Fredholm modules (π, P0) and (π, P1) are operator homotopic if there is a norm-continuous path of essential projections (Pt)t∈[0,1] such that (π, Pt) is a Fredholm module at all times t ∈ [0, 1]. Thus, the representation π is fixed throughout an operator homotopy. A Fredholm module (π, P ) is degenerate if P is a projection which commutes with the representation π. Under direct summation of K-homology classes, K1(A) is an abelian group. Modulo essentially the same equivalence relation as in the odd case, even Fredholm modules up to equivalence are the classes in the even K-homology group K0(A). Every class in K0(A) can be represented by a balanced even Fredholm module. 2.2. Finitely summable Fredholm modules. The singular values {sn(T )}n≥1 of a compact operator T ∈ K(H) are the eigenvalues of T, arranged in non-increasing order and repeated according to their multiplicity. The compactness of T means that sn(T ) → 0. For p ≥ 1, the Schatten ideals Lp(H) and Lp+(H) are defined as follows: Lp(H) =nT ∈ K(H) : X sn(T )p < ∞o, Lp+(H) =nT ∈ K(H) : sn(T ) = O(n−1/p)o. (Actually, the definition of L1+(H) is slightly different, and it will not be used in this paper.) We have Lp(H) ⊂ Lp+(H) ⊂ Lq(H) for all q > p. The summable Fredholm modules are those which satisfy a restricted version of Defini- tion 2.1, in which the ideals Lp(H) or Lp+(H) replace the ideal of compact operators K(H). Definition 2.2 (Connes). An odd Fredholm module (π, P ) is p-summable (over A) if P ∗ − P, P 2 − P ∈ Lp(H) and [P, π(a)] ∈ Lp(H) for all a in a dense subalgebra A of A. A balanced even Fredholm module (π, U ) is p-summable (over A) if U U ∗ − 1, U ∗U − 1 ∈ Lp(H), and [U, π(a)] ∈ Lp(H) for all a in a dense subalgebra A of A. The notion of p+-summable Fredholm module is defined analogously. We note that p- summability implies p+-summability, which in turn implies q-summability for all q > p. The property that every K-homology class is representable by a finitely summable Fredholm module could be deemed as K-homological finiteness. Even sharper would be to require that finite summability can be achieved in a uniform way throughout the K-homology classes. Such 8 HEATH EMERSON AND BOGDAN NICA a uniformity could be imposed on the degree of summability, or on the summability subalgebra, or both. We propose the following definition. Definition 2.3. The K-homology of a C-algebra A is uniformly p-summable (over A) if there is a dense subalgebra A of A such that every K-homology class of A can be represented by a Fredholm module which is p-summable over A. There is an obvious variation for p+-summability. The motivating example for this strong notion of K-homological finiteness is the following: the K-homology of the commutative C- algebra C(M ), where M is a smooth closed manifold, is uniformly (dim M )+-summable over the smooth subalgebra C∞(M ). 3. The basic K-cycle Throughout this section, G is a discrete countable group acting by homeomorphisms on a compact metrizable space X. To avoid trivialities, we assume that X is not a singleton. We consider the reduced crossed-product C(X) ⋊r G associated to the topological dynamics G y X. 3.1. Left regular representation, G-expectation and G-deviation. Let µ be a Borel probability measure on X which has full support, meaning that non-empty open subsets have positive measure, and it is G-quasi-invariant, i.e., the action of G preserves the measure class of µ. We do not assume µ to be G-invariant, in fact a highly non-invariant µ will turn out to be the most interesting case for our purposes. The faithful representation of C(X) on L2(X, µ) by multiplication induces a faithful rep- resentation λµ of C(X) ⋊r G on ℓ2(G, L2(X, µ)), the left regular representation with respect to µ. In fact, the C-algebra C(X) ⋊r G can be defined as the norm completion of the alge- braic crossed-product C(X) ⋊alg G in the regular representation λµ. Concretely, λµ is given as follows: λµ(φ)(cid:16)X ψhδh(cid:17) =X(h−1.φ)ψhδh, λµ(g)(cid:16)X ψhδh(cid:17) =X ψhδgh λµ(g)λµ(φ)λµ(g−1) holds. where φ ∈ C(X), g ∈ G, and P ψhδh ∈ ℓ2(G, L2(X, µ)). The covariance relation λµ(g.φ) = On the probability space (X, µ), momentarily devoid of the G-action, there are two impor- tant numerical characteristics attached to a continuous functions on X: the expectation and the standard deviation. Namely, for φ ∈ C(X) we put (3.1) σφ(g) =s 1 2ZZ φ(gx) − φ(gy)2 dµ(x) dµ(y). As an illustration of the dynamical expectation for a non-trivial group action, consider the case of a group G ⊆ SU(1, 1) acting by linear fractional transformations on the unit circle Eφ =Z φ dµ, σφ =qE(cid:0)φ2(cid:1) −(cid:12)(cid:12)Eφ(cid:12)(cid:12)2 . When we bring in the G-action, we are led to consider the following dynamical counterparts. Definition 3.1. The G-expectation and the G-deviation of φ ∈ C(X) with respect to µ are the functions Eφ : G → C and σφ : G → [0,∞) given as follows: Eφ(g) =Z g−1.φ dµ =Z φ dg∗µ, σφ =qE(cid:0)φ2(cid:1) −(cid:12)(cid:12)Eφ(cid:12)(cid:12)2 . An explicit, and useful, formula for the G-deviation is K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 9 S1 = {ζ ∈ C : ζ = 1}. With respect to the normalized Lebesgue measure, Eφ(g) is the value of the Poisson transform of φ on the unit disk at the point g(0). Definition 3.2. We say that µ has C0-deviation if σφ ∈ C0(G) for all φ ∈ C(X), respectively ℓp-deviation if σφ ∈ ℓpG for all φ in a dense subalgebra of C(X). 3.2. The basic K-cycle. We view ℓ2G as the constant-coefficient subspace of ℓ2(G, L2(X, µ)). The corresponding projection Pℓ2G is given by coefficient-wise integration: Pℓ2G(cid:16)X ψhδh(cid:17) =X(cid:18)Z ψh dµ(cid:19)δh. We are interested in the event that (λµ, Pℓ2G) is a Fredholm module – or, even better, a summable one – for C(X) ⋊r G. When this happens, we refer to (λµ, Pℓ2G) as the basic K-cycle associated to µ. The Fredholmness and the summability of (λµ, Pℓ2G) can be conveniently expressed in terms of the decay of the G-deviation. First, we record a general observation regarding Fredholmness and summability in the odd case. Lemma 3.3. Let A be a unital C-algebra, let π : A → B(H) be a representation, and let P be a projection in B(H). Denote by s(a) := P π(a)P the corresponding compression. Then (π, P ) is a Fredholm module for A if and only ifps(a2) − s(a)2 ∈ K(H) for all a ∈ A. Furthermore, (π, P ) is p-summable over a dense ∗-subalgebra A ⊆ A if and only ifps(a2) − s(a)2 ∈ Lp(H) for all a ∈ A. Proof. Let Π(a) = (1 − P )π(a)P ; this is the lower left corner of the 2-by-2 matrix defined by the decomposition of π with respect to P . Using the relations [π(a), P ] = Π(a) − Π(a∗)∗, Π(a) = (1 − P )[π(a), P ] we see that (π, P ) is a Fredholm module if and only if Π(a) ∈ K(H) for all a ∈ A, and that (π, P ) is p-summable over A if and only if Π(a) ∈ Lp(H) for all a ∈ A. Now Π(a)∗Π(a) = P π(a∗)(1 − P )π(a)P = P π(a∗a)P −(cid:0)P π(a)∗P(cid:1)(cid:0)P π(a)P(cid:1) = s(a∗a) − s(a)∗s(a) shows that Π(a) =ps(a2) − s(a)2. Proposition 3.4. The pair (λµ, Pℓ2G) is a Fredholm module if and only if µ has C0-deviation. If µ has ℓp-deviation, then (λµ, Pℓ2G) is a p-summable Fredholm module; for p ≥ 2, the converse holds. (cid:3) Proof. The projection Pℓ2G compresses the space restriction λµC(X) to multiplication by the G-expectation on ℓ2G: (3.2) Pℓ2Gλµ(φ)Pℓ2G = M(Eφ) for all φ ∈ C(X). Hence for sµ(φ) := Pℓ2Gλµ(φ)Pℓ2G we have qsµ(φ2) − sµ(φ)2 = M(σφ). By the proof of Lemma 3.3, µ has C0-deviation if and only if [λµ(φ), Pℓ2G] is compact for all φ ∈ C(X). As Pℓ2G commutes with the group restriction λµG, the latter condition is equivalent to having [λµ(a), Pℓ2G] compact for all a ∈ C(X) ⋊alg G, which is equivalent to (λµ, Pℓ2G) being Fredholm. The summable analogue is argued in a similar way. For sufficiency, assume that µ has ℓp- deviation. Then there is a G-invariant, dense ∗-subalgebra A(X) ⊆ C(X) such that σφ ∈ ℓpG 10 HEATH EMERSON AND BOGDAN NICA for all φ ∈ A(X). As above, we deduce that [λµ(a), Pℓ2G] is a p-summable operator for all a ∈ A(X) ⋊alg G. Thus (λµ, Pℓ2G) is p-summable. For the converse, we bring in another expectation, namely the bounded linear map E : C(X) ⋊r G ։ C(X) defined by E(P φg g) = φ1 over C(X) ⋊alg Γ. We claim that (3.3) for all h ∈ G and a ∈ C(X) ⋊r G. Indeed, using along the way the fact that Π(φg2 )∗Π(φg1 ) is a multiplication operator on ℓ2G, we have: kΠ(a)δhk2 ≥ σ(E(a))(h) (cid:10)Π(a)δh, Π(a)δh(cid:11) = Xg1,g2(cid:10)Π(φg1 )δg1h, Π(φg2 )δg2h(cid:11) = Xg1,g2(cid:10)Π(φg2 )∗Π(φg1 )δg1h, δg2h(cid:11) =Xg (cid:10)Π(φg)∗Π(φg)δgh, δgh(cid:11) =Xg (cid:10)M(σφg)2δgh, δgh(cid:11) =Xg (σφg)2(gh) ≥ (σφ1)2(h) =(cid:0)σ(E(a))(h)(cid:1)2 Now assume that (λµ, Pℓ2G) is a p-summable Fredholm module for C(X)⋊rG. Then Π(a) is a p- summable operator for all a in a dense subalgebra A of C(X)⋊rG. For p ≥ 2, the p-summability of Π(a) implies the p-summability of {kΠ(a)ξιk2}ι∈I for any orthonormal system (ξι)ι∈I ([38, we have shown that σφ ∈ ℓpG for all φ ∈ E(A). It follows that {φ ∈ C(X) : σφ ∈ ℓpG}, which is always a subalgebra of C(X), is dense. We conclude that µ has ℓp-deviation. Thm.1.18]). In particular {kΠ(a)δhk2}h∈G is p-summable, so σ(cid:0)E(a)(cid:1) ∈ ℓpG by (3.3). Thus, (cid:3) At the current level of generality, we cannot address the question whether (λµ, Pℓ2G), when a Fredholm module, is homologically non-trivial or not. It is, however, clear that it is non- degenerate, given our assumptions that X is not a singleton and µ is fully-supported. 3.3. C0-deviation and the convergence property. Let Prob(X) denote the space of Borel probability measures on X, and equip Prob(X) with the weak∗ convergence induced by C(X): by definition, νι → ν if R φ dνι →R φ dν for all φ ∈ C(X). The space Prob(X) is compact. In particular, push-forwards of µ by elements of G must accumulate. We make the following definition. Definition 3.5. The probability measure µ is said to have the convergence property if the accumulation points of the G-orbit Gµ ⊆ Prob(X) are all point measures. We think of the convergence property for a probability measure as a measurable analogue of an established notion in topological dynamics, that of a convergence group action. Let us recall the definition, originally due to Gehring and Martin for the case of group actions on spheres or closed balls, and then subsequently extended by Tukia, Freden, Bowditch to the general case of group actions on compact metrizable spaces. The action of G on X is said to be a convergence action if the following holds: for each sequence (gn) ⊆ G with gn → ∞, there is a subsequence (gni ) with attracting and repelling points x+, x− ∈ X in the sense that gniz → x+ uniformly outside neighbourhoods of x−. See, for instance, [23, Sec.5] for further details and references. We note the following simple fact. Proposition 3.6. Let G y X be a convergence action. If points in X are µ-negligible, then µ has the convergence property. Proof. Let (gn)∗µ converge in Prob(X), where gn → ∞ in G. Without loss of generality (gn) has attracting and repelling points x+, x− ∈ X. We claim that (gn)∗µ → δx+. Indeed, let n .φ converges pointwise to the constant function φ(x+) on X − {x−}. φ ∈ C(X). Then g−1 K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 11 As x− is µ-negligible, we get R φ d(gn)∗µ = R g−1 convergence theorem. n .φ dµ → φ(x+) by Lebesgue's dominated (cid:3) The convergence property is relevant for our discussion, in light of the following characteri- zation. Proposition 3.7. The probability measure µ has C0-deviation if and only if it has the con- vergence property. Proof. Assume that µ has C0-deviation, and let ν ∈ Prob(X) be the limit of a sequence (gn)∗µ with gn → ∞ in G. For each φ ∈ C(X) we have, on the one hand, that σφ(gn) converges to 0, and on the other hand that σφ(gn) converges to the standard deviation of φ with respect to ν. Therefore,R φ2 dν = R φ dν2 for all φ ∈ C(X). This continues to hold throughout L2(X, ν), by the density of C(X) in L2(X, ν) – Borel probability measures on compact metrizable spaces are automatically Radon. Taking characteristic functions of measurable sets, we see that ν is {0, 1}-valued. But the only {0, 1}-valued Borel probability measures on X are the point measures: choosing a compatible metric on X, there exists a sequence of full-measure balls with radius converging to 0, hence a point having full measure. The converse implication is left to the reader. (cid:3) In §6 we will show that suitable measures on the boundary of a Gromov hyperbolic group have the convergence property with respect to the boundary action of the group (which is a convergence action.) 3.4. Double ergodicity and ℓp-deviation. We address the condition p ≥ 2, encountered in Proposition 3.4. Namely, we show that double ergodicity of µ is an obstruction to having ℓp-deviation with p ≤ 2. Proposition 3.8. If µ × µ is ergodic for the diagonal action of G on X × X, and X has no isolated points, then a function φ ∈ C(X) with σGφ ∈ ℓ2G must be constant. In particular, if µ has ℓp-deviation then p > 2. Proof. Arguing by contradiction, we assume that φ ∈ C(X) is a non-constant function with the property that σφ ∈ ℓ2G. By (3.1), we have kσφk2 ℓ2G = 1 φ(gx) − φ(gy)2 dµ(x)dµ(y) 2ZZ Xg∈G Therefore S(x, y) = Pg∈G φ(gx) − φ(gy)2 defines a G-invariant L2 map on X × X. By ergodicity, S is a.e. constant, say S(x, y) = C for almost all (x, y) ∈ X × X. There exists c > 0 such that the open subset V = {(x, y) : φ(x) − φ(y) > c} ⊆ X × X is non-empty. As X × X has no isolated points, for each positive integer N there exist disjoint, non-empty open subsets U1, . . . , UN ⊆ V . Using again the ergodicity assumption, we have that each G · Ui = ∪g∈G gUi is either negligible or of full measure. Since non-empty open subsets of X × X have positive measure, the latter alternative must occur. It follows that ∩N i=1 G · Ui has full measure. Let (x, y) in ∩N i=1 G · Ui with S(x, y) = C. Thus, for each i we have some gi ∈ G such that (gix, giy) ∈ Ui. Now the gi's are distinct since the Ui's are disjoint, and φ(gix) − φ(giy) > c since Ui ⊆ V , so C = S(x, y) ≥ NXi=1 φ(gix) − φ(giy)2 > N c2. (cid:3) As N is arbitrary, this is a contradiction. 12 HEATH EMERSON AND BOGDAN NICA 4. Further properties of the basic K-cycle We now investigate the behaviour of the basic K-cycle under two operations: changing the measure µ, respectively passing to a finite-index subgroup of G. We keep the notations of the previous section. 4.1. Comparable measures. The pair (λµ, Pℓ2G) is constructed in reference to the measure µ, which is not part of the given topological setting. Nevertheless, its relevant features – Fredholmness, degree of summability, K-homology class – are canonical over the measure class of µ. The suitable equivalence here is the following: a Borel probability measure µ′ on X is said to be comparable to µ if µ′ ≍ µ, in the sense that C1µ ≤ µ′ ≤ C2µ for some positive constants C1, C2. Clearly, comparability is finer than the usual equivalence of measures which, we recall, means that each measure is absolutely continuous with respect to the other. Formula (3.1) shows that comparable measures have comparable G-deviations, hence the following: Proposition 4.1. Let µ and µ′ be comparable probability measures. Then (λµ, Pℓ2G) is a Fredholm module if and only if (λµ′ , Pℓ2G) is a Fredholm module. For p ≥ 2, (λµ, Pℓ2G) is p-summable if and only if (λµ′ , Pℓ2G) is p-summable. Most importantly, basic K-cycles associated to comparable measures define one and the same homology class: Proposition 4.2. Let µ and µ′ be comparable probability measures having C0-deviation. Then the Fredholm modules (λµ, Pℓ2G) and (λµ′ , Pℓ2G) are K-homologous. Proof. Let ρ = dµ′/dµ be the Radon-Nikodym derivative, so ρ is essentially bounded from above and from below by the comparability constants of µ and µ′. First, we have a unitary U : ℓ2(G, L2(X, µ′)) → ℓ2(G, L2(X, µ)), X ψhδh 7→X√ρ ψhδh which intertwines the corresponding regular representations of C(X) ⋊r G, that is, U λµ′ U ∗ = λµ. We may therefore exchange (λµ′ , P ′ ℓ2G is used in order to emphasize the dependence on µ′. We now claim that the Fredholm modules (λµ, U P ′ ℓ2GU ∗) and (λµ, Pℓ2G) are operator homotopic. Note that ℓ2GU ∗), where the notation P ′ ℓ2G) for (λµ, U P ′ U P ′ ℓ2GU ∗(cid:0)X ψhδh(cid:1) =X√ρ(cid:18)Z √ρ ψh dµ(cid:19)δh, and that √ρ ∈ L∞(X, µ) with k√ρkL2(X,µ) = 1. For η ∈ L∞(X, µ) satisfying kηkL2(X,µ) = 1, let M (η) be the corresponding multiplication operator on ℓ2(G, L2(X, µ)). Then P (η) = M (¯η)Pℓ2GM (η), X ψhδh 7→X ¯η(cid:18)Z η ψh dµ(cid:19)δh. is a projection, namely the projection of ℓ2(G, L2(X, µ)) onto M (¯η)ℓ2G. We have [P (η), λµ] = M (¯η)[Pℓ2G, λµ]M (η) since M (η) and M (¯η) commute with λµ, so (λµ, P (η)) is a Fredholm module. On the other hand, we have kP (η1) − P (η2)k ≤ 2kη1 − η2kL2(X,µ); this follows from the fact that (cid:13)(cid:13)(cid:13)(cid:13)¯η1Z η1ψ dµ − ¯η2Z η2ψ dµ(cid:13)(cid:13)(cid:13)(cid:13)2 ≤ 2kη1 − η2k2kψk2 for all ψ ∈ L2(X, µ). Now let η(t) = (cos t) 1 + (i sin t) √ρ, where 0 ≤ t ≤ π/2. Then η(t) ∈ L∞(X, µ), and η(t) describes a continuous path in the unit sphere of L2(X, µ) between the constant function 1 and i√ρ. Consequently, P (η(t)) describes a norm-continuous path between P (1) = Pℓ2G and P (i√ρ) = P (√ρ) = U P ′ (cid:3) ℓ2GU ∗. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 13 4.2. Finite-index subgroups. Let H ≤ G be a subgroup. Restriction of representations from C(X) ⋊r G to C(X) ⋊r H takes Fredholm modules for C(X) ⋊r G to Fredholm modules for C(X) ⋊r H, and it defines a natural homomorphism of abelian groups res : K1(C(X) ⋊r G) → K1(C(X) ⋊r H). Assume that µ has C0-deviation. On the one hand, restricting (λG µ , Pℓ2G) yields a Fredholm module for C(X) ⋊r H. On the other hand, we can form the Fredholm module (λH µ , Pℓ2H ) for C(X) ⋊r H. The homological relation between these two Fredholm modules for C(X) ⋊r H is particularly simple in the case when H has finite index in G. Proposition 4.3. Assume that µ has C0-deviation, and that {g∗µ}g∈G forms a family of mutually comparable measures. If H is a finite-index subgroup of G, then in K1(C(X) ⋊r H). res(cid:2)(λG µ , Pℓ2G)(cid:3) = [G : H](cid:2)(λH µ , Pℓ2H )(cid:3) Proof. Put n = [G : H], and pick a transversal t1, . . . , tn for the right H-cosets. The coset decomposition ℓ2(G, L2(X, µ)) = ⊕n 1 ℓ2(Hti, L2(X, µ)) yields res(cid:2)(λG µ , Pℓ2G)(cid:3) = ⊕n 1(cid:2)(λti , Pℓ2(Hti))(cid:3) Vt : ℓ2(H, L2(X, µ)) → ℓ2(H, L2(X, t∗µ)), X ψhδh 7→X(t.ψh)δh t λtRt, Pℓ2H ) and (λH makes (R∗ t∗µ, Pℓ2H ) equivalent. On the other hand, the assumption that {g∗µ}g∈G consists of mutually comparable measures implies, in light of Proposition 4.2, that (λH µ , Pℓ2H ) are homologous. Summarizing, we have t∗µ, Pℓ2H ) and (λH in K1(C(X) ⋊r H), as desired. (cid:2)(res(λG µ ), Pℓ2G)(cid:3) = ⊕n 1 (cid:2)(λH µ , Pℓ2H )(cid:3) (cid:3) 5. Preliminaries on boundaries of hyperbolic spaces This section is devoted to the metric-measure structure on the boundary of a hyperbolic space in the sense of Gromov [17]. In §5.1 we recall some basic facts on hyperbolic spaces and their boundaries. In §5.2 we focus on the family of visual metrics, and their induced Hausdorff measures, on the boundary of a hyperbolic space. The content of these two subsections is, to a large extent, standard [15], [40, Sec.5]. In §5.3 we discuss geometric group actions, and their boundary effect. We describe results of Coornaert [8] providing remarkable finiteness properties for the visual metric-measure structure. In §5.4 we introduce a suitable notion of 'Hausdorff dimension' for the boundary. in K1(C(X) ⋊r H), where λti denotes the representation of C(X) ⋊r H on ℓ2(Hti, L2(X, µ)). Now consider (λt, Pℓ2(Ht)) for t ∈ {t1, . . . , tn}. The unitary Rt : ℓ2(H, L2(X, µ)) → ℓ2(Ht, L2(X, µ)), X ψhδh 7→X ψhδht implements an equivalence between (λt, Pℓ2(Ht)) and (R∗ R∗ t λtRt on ℓ2(H, L2(X, µ)) is given by t λtRt, Pℓ2H ). The representation R∗ t λtRt(h′)(cid:16)X ψhδh(cid:17) =X ψhδh′h R∗ t λtRt(φ)(cid:16)X ψhδh(cid:17) =X t−1.(h−1.φ)ψhδh, for φ ∈ C(X) and h′ ∈ H. Next, the unitary 14 HEATH EMERSON AND BOGDAN NICA 5.1. The boundary of a hyperbolic space. Let (X, d) be a proper geodesic space. The Gromov product of x, y ∈ X with respect to o ∈ X is defined by the formula Definition 5.1 (Gromov). The space X is hyperbolic if there exists a constant δ ≥ 0 such that, for all x, y, z, o ∈ X, we have (x, y)o := 1 2(cid:0)d(o, x) + d(o, y) − d(x, y)(cid:1). (x, y)o ≥ min(cid:8)(x, z)o, (y, z)o(cid:9) − δ. Let X be a hyperbolic space, and fix a basepoint o ∈ X. A sequence (xi) ⊆ X converges to infinity if (xi, xj )o → ∞ as i, j → ∞. Two sequences (xi), (yi) converging to infinity are asymptotic if (xi, yi)o → ∞ as i → ∞. The asymptotic relation is an equivalence on sequences converging to infinity. Both convergence to infinity, and the asymptotic relation, are independent of the chosen basepoint o ∈ X. The boundary of X, denoted ∂X, is the set of asymptotic classes of sequences converging to infinity. A sequence (xi) ⊆ X converges to ξ ∈ ∂X if (xi) converges to infinity, and the asymptotic class of (xi) is ξ. i)o : xi → ξ, x′ The Gromov product on ∂X × ∂X is defined as follows: If ξ = ξ′, then (ξ, ξ′)o = ∞. If ξ 6= ξ′, then the sequence (xi, x′ and x′ (ξ, ξ′)o := inf(cid:8) lim inf (xi, x′ i → ξ′, hence (ξ, ξ′)o < ∞. It turns out that i)o ≤ lim sup (xi, x′ i)o ≤ (ξ, ξ′)o + 2δ (x, ξ)o := inf(cid:8) lim inf (x, xi)o : xi → ξ(cid:9) (5.1) Similarly, the Gromov product on X × ∂X is defined by setting (ξ, ξ′)o ≤ lim inf (xi, x′ and we have i → ξ′(cid:9) i)o is bounded whenever xi → ξ (xi → ξ, x′ i → ξ′). (5.2) (x, ξ)o ≤ lim inf (x, xi)o ≤ lim sup (x, xi)o ≤ (x, ξ)o + δ (xi → ξ). 5.2. Visual metrics. Equipped with a canonical topology defined in terms of the Gromov product, the boundary ∂X is compact and metrizable (see [15, Ch.7, §2]). But the metric structure on ∂X, which is of great importance in this paper, is a more subtle issue. Definition 5.2. A visual metric on ∂X is a metric dǫ satisfying dǫ ≍ exp(−ǫ(·,·)o) for some ǫ > 0, called the visual parameter of dǫ. This definition is independent of the chosen basepoint o ∈ X, and every visual metric for any determines the canonical topology on ∂X. α ∈ (0, 1]; consequently, if ǫ is a visual parameter then so is any ǫ′ ∈ (0, ǫ]. Fact 5.3 (Scaling). Let dǫ and dǫ′ be two visual metrics. Then: • dǫ and dǫ′ are Holder equivalent: d1/ǫ • the corresponding Hausdorff dimensions are inversely proportional to the visual parameter: ǫ hdim(∂X, dǫ) = ǫ′ hdim(∂X, dǫ′); If dǫ is a visual metric, then so is dα ǫ ǫ ≍ d1/ǫ′ ǫ′ ; • the corresponding Hausdorff measures are comparable: µǫ ≍ µǫ′. Visual metrics do exist, provided that the visual parameter is small with respect to 1/δ where δ is a constant of hyperbolicity. Furthermore, there is a companion metric-like map on X × ∂X, which is visual in the corresponding way [40, Prop.5.16]: Fact 5.4 (Small visual range). Let ǫ > 0 be such that ǫδ < 1/5. Then: • there exists a visual metric dǫ on ∂X, having visual parameter ǫ; K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 15 • there exists dǫ : X × ∂X → [0,∞) satisfying 1 2 exp(−ǫ(x, ξ)o) ≤ dǫ(x, ξ) ≤ exp(−ǫ(x, ξ)o) and dǫ(x, ξ) − dǫ(x, ξ′) ≤ dǫ(ξ, ξ′) ≤ dǫ(x, ξ) + dǫ(x, ξ′) for all x ∈ X and ξ, ξ′ ∈ ∂X. The small range for visual parameters is neither optimal, nor particularly natural, and we have to consider the entire 'cone' of visual metrics on ∂X. Statements about visual metrics on ∂X are sometimes proved by first dealing with visual parameters in the small range, and then extended by scaling (Fact 5.3). 5.3. Geometric group actions. Underlying the following definition is the fundamental fact that hyperbolicity is invariant under quasi-isometries. Definition 5.5 (Gromov). A group Γ is hyperbolic if it acts geometrically, that is to say isometrically, properly and cocompactly, on a hyperbolic space. We refer to a space carrying a geometric action of a hyperbolic group Γ as a geometric model for Γ. For example, Cayley graphs of Γ with respect to various finite generating sets are geometric models for Γ. There could be, however, more natural geometric models for a given hyperbolic group, e.g., for a surface group of genus at least 2 the natural geometric model is the standard hyperbolic plane. Geometric models for Γ have Γ-equivariantly homeomorphic boundaries, and each one of them is a topological realization of ∂Γ. Now let X be a hyperbolic space admitting a geometric action of a (hyperbolic) group Γ. In what follows, we assume that Γ is non-elementary, that is, Γ is neither finite, nor virtually infinite cyclic. In terms of the space X, the non-elementary hypothesis on Γ means that ∂X is infinite as a set. The action of Γ on X extends to the boundary ∂X. The boundary action is a convergence action, in the sense of §3.3. We also have (gx, gx′)o ≥ (x, x′)o − d(o, go) for all g ∈ Γ and x, x′ ∈ X, which implies that (5.3) for all g ∈ Γ and ξ, ξ′ ∈ ∂X. Therefore Γ acts by Lipschitz maps on (∂X, dǫ) for any choice of visual metric dǫ on ∂X. (gξ, gξ′)o ≥ (ξ, ξ′)o − d(o, go) Definition 5.6. The exponent, or the volume entropy of X is the finite positive number given by eX = infns > 0 : Xg∈Γ exp(−sd(o, go)) < ∞o = lim sup R→∞ 1 R ln(cid:12)(cid:12){g ∈ Γ : d(o, go) ≤ R}(cid:12)(cid:12). The two formulas give two interpretations of the exponent, namely critical exponent as well as growth exponent. As the notation eX already suggests, the definition is independent of the basepoint o ∈ X and of the group Γ acting geometrically on X. The Patterson - Sullivan theory developed by Coornaert in [8] plays a crucial role in under- standing the growth of Γ-orbits in X, and the Hausdorff dimensions and measures associated to visual metrics on ∂X. The following hold. Fact 5.7 (Orbit growth). Let o be a basepoint in X. Then (cid:12)(cid:12){g ∈ Γ : d(o, go) ≤ R}(cid:12)(cid:12) ≍ Fact 5.8 (Hausdorff measure and dimension). Let dǫ be a visual metric on ∂X. Then: exp(eX R). 16 HEATH EMERSON AND BOGDAN NICA • the Hausdorff dimension hdim(∂X, dǫ) equals eX /ǫ; • the Hausdorff measure µǫ is Ahlfors regular, that is µǫ(Br) ≍ rhdim(∂X,dǫ) uniformly over all closed balls Br of radius 0 ≤ r ≤ diam(∂X, dǫ). Fact 5.7 is [8, Thm.7.2]. For sufficiently small visual parameters, Fact 5.8 follows from [8, Prop.7.4, Cor.7.5, Cor.7.6]; using Fact 5.3, it extends to arbitrary visual parameters. In particular, ∂X has finite mass under the Hausdorff measure defined by a visual metric. This allows for the following definition. Definition 5.9. A visual probability measure on ∂X is a Borel probability measure µ satisfying µ ≍ µǫ for some (equivalently, each) Hausdorff measure µǫ defined by a visual metric dǫ. 5.4. Visual dimension. In general, there is no canonical choice of metric on the boundary ∂X of a hyperbolic space X. So the notion of Hausdorff dimension for the boundary has to be understood relative to a family of admissible metrics. For visual metrics, one obtains the following notion of dimension. Definition 5.10. The visual dimension of ∂X, denoted visdim ∂X, is the infimal Hausdorff dimension of (∂X, d) as d runs over the visual metrics on ∂X. Clearly visdim ∂X ≥ topdim ∂X, as the topological dimension of a metric space is at most the Hausdorff dimension with respect to a compatible metric. A more appropriate comparison term for the visual dimension is that of conformal dimension. The following chain of inequalities holds: visdim ∂X ≥ A-confdim ∂X ≥ confdim ∂X ≥ topdim ∂X. The conformal dimension of ∂X is a notion of metric dimension which only depends on the quasi-isometry type of X. It resolves the metric ambiguity at the boundary by taking all possible metrics which are equivalent to a visual metric in a suitable sense. The original definition, due to Pansu, uses quasi-conformal equivalence; more recently, the closely related quasi-Mobius equivalence seems to be favored. Then confdim ∂X is defined as the infimal Hausdorff dimension of (∂X, d) as d runs over all metrics which are equivalent to a visual metric. See [23, Section 14]. From a measure-theoretic point of view, the equivalence relation used for defining the conformal dimension is too loose. For hyperbolic spaces admitting geometric group actions, the notion of Ahlfors conformal dimension strikes a compromise by restricting the equivalence relation to Ahlfors regular metrics. Namely, A-confdim ∂X is defined as the infimal Hausdorff dimension of (∂X, d) as d runs over all Ahlfors regular metrics on ∂X which are quasi-Mobius equivalent to a visual metric. The Ahlfors conformal dimension is a key concept for much of the current work on boundaries of hyperbolic spaces from the perspective of analysis on metric spaces [28]. We illustrate Definition 5.10 on the following important examples. Example 5.11. Let T be a regular tree of degree at least 3. Topologically, the boundary ∂T is a Cantor set. The Gromov product (·,·)o on T extends canonically and continuously to ∂T , and exp(−(·,·)o) is an ultrametric on ∂T . Each ǫ > 0 is a visual parameter, so visdim ∂T = 0. Example 5.12. The boundary of the standard hyperbolic space Hn, where n ≥ 2, is the sphere Sn−1. The usual spherical metric is a visual metric with visual parameter ǫ = 1. In fact ǫ = 1 is the largest possible visual parameter. Indeed, the Lipschitz functions with respect to a visual metric are dense in C(Sn−1). On the other hand, they are the ǫ-Holder functions with respect to the spherical metric, ǫ being the visual parameter. Now observe that, on a geodesic metric space, only the constant functions are ǫ-Holder for ǫ > 1. Thus visdim ∂Hn = n − 1, the Hausdorff dimension with respect to the spherical metric. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 17 More generally, let us consider the (non-compact) rank-1 symmetric spaces. These are the real, complex, quaternionic, or octonionic hyperbolic spaces Hn K, where n ≥ 2 respectively n = 2 in the exceptional octonionic case. Put k = dimR K ∈ {1, 2, 4, 8}. Topologically, the boundary ∂Hn K, the so-called Carnot metric, is a visual metric with visual parameter ǫ = 1. As the Carnot metric is geodesic, no parameter greater than 1 is a visual parameter. Therefore visdim ∂Hn K, the Hausdorff dimension of ∂Hn K equipped with the Carnot metric. The latter is explicitly given by the Mitchell - Pansu formula: hdim ∂Hn K is a sphere of dimension nk − 1. The standard metric on ∂Hn K = hdim ∂Hn K = topdim ∂Hn K + k − 1 (= nk + k − 2). The notion of visual dimension for a hyperbolic space can be taken a step further. Namely, one could define a visual dimension for a hyperbolic group Γ as the infimal visual dimension of ∂X, where X runs over the (isometry classes of) geometric models for Γ. Our results are, in fact, most conveniently expressed in terms of such a visual dimension for the group. 6. The basic K-cycle for boundary actions of hyperbolic groups In this section, we realize the paradigm of Section 3 in the case of a hyperbolic group acting on its boundary. First, some standing notations for the rest of the paper. The main characters are Γ : a non-elementary hyperbolic group; X : a hyperbolic space on which Γ acts geometrically, i.e., a geometric model for Γ; µ : a visual probability measure on ∂X. In order to be able to do geometric analysis on the boundary, we also fix dǫ : a visual metric on ∂X, and we denote Dǫ : the Hausdorff dimension of (∂X, dǫ). Lemma 6.1. Let o ∈ X be a basepoint. Then there exists C′ > 0 such that, for all g ∈ Γ, we have If Dǫ > 2, then there exists C > 0 such that, for all g ∈ Γ, we have (cid:18)ZZ dǫ(gξ, gξ′)2 dµ(ξ) dµ(ξ′)(cid:19)1/2 (cid:18)ZZ dǫ(gξ, gξ′)2 dµ(ξ) dµ(ξ′)(cid:19)1/2 ≥ C′ exp(−ǫ d(o, go)). ≤ C exp(−ǫ d(o, go)). The proof is deferred to the end of the section. The important part is the second estimate; the purpose of the first part is to show that we are getting the correct asymptotics. Theorem 6.2. (λµ, Pℓ2Γ) is a Fredholm module for C(∂X) ⋊ Γ which is D+ ǫ -summable when Dǫ > 2, respectively p-summable for every p > 2 when Dǫ ≤ 2. The summability holds over the dense ∗-subalgebra Lip(∂X, dǫ) ⋊alg Γ, where Lip(∂X, dǫ) is the Γ-invariant algebra of Lipschitz functions on ∂X. Proof. We first prove the claim in the case when Dǫ > 2. Using formula (3.1), we have σφ(g) ≤ kφkLips 1 2ZZ dǫ(gξ, gξ′)2 dµ(ξ) dµ(ξ′) for all φ ∈ Lip(∂X, dǫ). It follows from Lemma 6.1 that there exists C > 0 such that σφ(g) ≤ CkφkLip exp(−ǫ d(o, go)) 18 HEATH EMERSON AND BOGDAN NICA for all φ ∈ Lip(∂X, dǫ) and g ∈ Γ. Let T denote the multiplication by g 7→ exp(−ǫ d(o, go)), viewed as an operator on ℓ2Γ. We claim that T ∈ LDǫ+(ℓ2Γ). Once we know this, it follows that multiplication by σφ is in LDǫ+(ℓ2Γ) for all φ ∈ Lip(∂X, dǫ), and the proof of Proposition 3.4 shows that (λµǫ , Pℓ2Γ) is a D+ ǫ -summable Fredholm module. In order to prove our claim that sn(T ) = O(n−1/Dǫ ), we first control a subsequence of singular values for T . Let mk denote the size of the "ball" {g ∈ Γ : d(o, go) ≤ k}; thus mk ≍ exp(eX k) by Fact 5.7. We have For an arbitrary positive integer n, let k be such that mk + 1 ≤ n ≤ mk+1 + 1. Then ≤ C2 n−1/Dǫ smk+1(T ) < exp(−ǫk) = exp(eX k)−1/Dǫ ≤ C1 m−1/Dǫ (cid:17)1/Dǫ sn(T ) ≤ smk+1(T ) ≤ C1 m−1/Dǫ ≤ C1 n−1/Dǫ(cid:16) mk+1 + 1 mk for some constant C2 independent of n and k. The claim is proved for Dǫ > 2. k . k Now assume that Dǫ ≤ 2, and let p > 2. Let also α ∈ (0, 1] so that Dǫ/α, which is the ǫ ), satisfies p > Dǫ/α > 2. By the previous part of the proof, we ǫ )⋊algΓ. As Lip(∂X, dǫ) is contained (cid:3) Hausdorff dimension of (∂X, dα know that (λµ, Pℓ2Γ) is (Dǫ/α)+-summable over Lip(∂X, dα in Lip(∂X, dα ǫ ), we conclude that (λµ, Pℓ2Γ) is p-summable over Lip(∂X, dǫ) ⋊alg Γ. In Theorem 6.2, the summability of the basic K-cycle (λµ, Pℓ2Γ) always occurs above 2. We do not know whether this phenomenon is due to some structural obstruction. There is, however, an obstruction to our method of controlling summability by the decay of the Γ-deviation, and this is the fact that visual probability measures are doubly ergodic [22]. Indeed, by Proposition 3.8 the Γ-deviation has to decay faster than ℓ2 in the presence of double ergodicity. We may optimize Theorem 6.2 by varying the visual metric. Our notion of visual dimension is in fact tailored for this very purpose. Corollary 6.3. (λµ, Pℓ2Γ) is p-summable for every p > max{2, visdim ∂X}. Furthermore, (λµ, Pℓ2Γ) is (visdim ∂X)+-summable provided visdim ∂X is attained and greater than 2. Now let us return to Lemma 6.1, whose proof we have postponed. Proof of Lemma 6.1. The first estimate is straightforward. As in (5.3), we have (gξ, gξ′)o ≤ d(o, go) + (ξ, ξ′)o for all ξ, ξ′ ∈ ∂X. Therefore dǫ(gξ, gξ′) ≥ c1 exp(−ǫ d(o, go)) dǫ(ξ, ξ′) for some c1 > 0, which implies that (cid:18)ZZ dǫ(gξ, gξ′)2 dµ(ξ) dµ(ξ′)(cid:19)1/2 ≥ c2 exp(−ǫ d(o, go)). The second, conditional estimate is more involved. First, we assume that the visual parameter ǫ is in the small visual range and that dǫ is a visual metric enjoying the properties listed in Fact 5.4. We let α > 0, and we show the following: if Dǫ > 2α, then there exists C > 0 such that for all g ∈ Γ we have (cid:18)ZZ dǫ(gξ, gξ′)2α dµ(ξ) dµ(ξ′)(cid:19)1/2 ≤ C exp(−αǫ d(o, go)). Let ξ, ξ′ ∈ ∂X. Observe that (gx, gx′)o +(g−1o, x)o +(g−1o, x′)o ≥ d(o, go) whenever x, x′ ∈ X; indeed, this is just a complicated rewriting of the triangle inequality d(o, x)+d(o, x′) ≥ d(x, x′). Letting x → ξ, x → ξ′ and using (5.1) and (5.2), we obtain that (gξ, gξ′)o + (g−1o, ξ)o + (g−1o, ξ′)o ≥ d(o, go) − 4δ. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 19 Thus there is C1 > 0 such that dǫ(gξ, gξ′) ≤ C1 exp(−ǫ d(o, go)) dǫ(g−1o, ξ)−1 dǫ(g−1o, ξ′)−1 for all g ∈ Γ and ξ, ξ′ ∈ ∂X. It follows that (cid:18)ZZ dǫ(gξ, gξ′)2α dµ(ξ) dµ(ξ′)(cid:19)1/2 1 exp(−αǫ d(o, go)) Z dǫ(g−1o, ξ)−2α dµ(ξ) ≤ Cα and our next goal is to show that the integral on the right hand side is bounded above inde- pendently of g ∈ Γ. At this point, the technical advantage of using ǫ in the small visual range becomes apparent: the function dǫ(g−1o,·) is Lipschitz, in particular measurable, on ∂X. For each positive integer k we put ∆k =(cid:8)ξ ∈ ∂X : exp(−ǫk) ≤ dǫ(g−1o, ξ) ≤ exp(−ǫ(k − 1))(cid:9). (Although we will not need this fact, we remark that dǫ(g−1o,·) is bounded below by a constant multiple of exp(−ǫ d(o, go)), so ∆k is in fact empty for k ≫ d(o, go).) From the hyperbolic is C2 ≥ 0 such that inequality (ξ, ξ′)o ≥ min(cid:8)(g−1o, ξ)o, (g−1o, ξ′)o(cid:9) − δ, where ξ, ξ′ ∈ ∂X, we deduce that there for all ξ, ξ′ ∈ ∂X. This inequality implies that diam(∆k) ≤ eǫC2 exp(−ǫk). It now follows from Fact 5.8 that there exists C3 > 0, independent of k, such that dǫ(ξ, ξ′) ≤ C2 max(cid:8)dǫ(g−1o, ξ), dǫ(g−1o, ξ′)(cid:9) µ(∆k) ≤ C3(cid:0) exp(−ǫk)(cid:1)eX /ǫ = C3 exp(−eX k). Using this diameter bound, we immediately get the desired integral estimate: Z dǫ(g−1o, ξ)−2α dµ(ξ) ≤Xk≥1Z∆k ≤ C3Xk≥1 dǫ(g−1o, ξ)−2α dµ(ξ) ≤Xk≥1 exp(cid:0)(2αǫ − eX )k(cid:1) exp(2αǫk)µ(∆k) and the latter series converges since Dǫ = eX /ǫ > 2α by assumption. Now let ǫ be an arbitrary visual parameter. Pick ǫ0 in the small visual range, and let dǫ0 be a corresponding visual metric. By Fact 5.3, we have (cid:18)ZZ dǫ(gξ, gξ′)2 dµ(ξ) dµ(ξ′)(cid:19)1/2 ≍(cid:18)ZZ dǫ0(gξ, gξ′)2ǫ/ǫ0 dµ(ξ) dµ(ξ′)(cid:19)1/2 . According to the lemma's hypothesis, hdim(∂X, dǫ0) = (ǫ/ǫ0) hdim(∂X, dǫ) is greater than 2ǫ/ǫ0. Hence the previous part of the proof shows that the right-hand side is bounded above by a constant multiple of exp(−(ǫ/ǫ0)ǫ0 d(o, go)) = exp(−ǫ d(o, go)). (cid:3) For the sake of conciseness, we adopt the following for the rest of the paper. Notation. We write D> ǫ -summable to mean (D+ ǫ -summable when Dǫ > 2, p-summable for every p > 2 when Dǫ ≤ 2. 20 HEATH EMERSON AND BOGDAN NICA 7. The boundary extension All the basic K-cycles coming from visual probability measures on ∂X are K-homologous, by Proposition 4.2. The purpose of this section is to describe the K1-class of a basic K-cycle as the class of a certain canonical extension of C(∂X) ⋊ Γ by the compact operators on ℓ2Γ. This extension encodes the compactification of Γ obtained by attaching the boundary ∂X. Let Γ = Γ ∪ ∂X be the compactification of Γ by the boundary of the geometric model X. By definition, g → ω ∈ ∂X in Γ if go → ω in X for some (equivalently, each) base point o ∈ X. From the exact sequence of Γ-C-algebras 0 → C0(Γ) → C(Γ) → C(∂X) → 0 we obtain an exact sequence of C-crossed products by Γ: (7.1) 0 −→ C0(Γ) ⋊ Γ −→ C(Γ) ⋊ Γ −→ C(∂X) ⋊ Γ −→ 0 Each C-algebra in (7.1) is nuclear; in particular, the full and the reduced crossed products agree. The faithful representation of C(Γ) on ℓ2Γ by multiplication induces a faithful repre- sentation π : C(Γ) ⋊ Γ → B(ℓ2Γ), which restricts to the standard isomorphism between the ideal term C0(Γ) ⋊ Γ and the compact operators K(ℓ2Γ). Thus (7.1) defines a class in the Brown - Douglas - Fillmore group Ext(C(∂Γ) ⋊ Γ). The nuclearity of C(∂Γ) ⋊ Γ implies that Ext(C(∂Γ) ⋊ Γ) and K1(C(∂Γ) ⋊ Γ) are isomorphic. The map Ext → K1 is given by the Stinespring construction, which dilates a completely positive map to an odd Fredholm module. The compactification of Γ, hence the exact sequence (7.1) as well, is defined in reference to the chosen geometric model X. However, and this is an important point, the boundaries of two geometric models for Γ are Γ-equivariantly homeomorphic. It follows that, up to isomorphism of extensions, the exact sequence (7.1) is independent of the chosen geometric model X. Definition 7.1. The boundary extension class [∂Γ] ∈ Ext(C(∂Γ) ⋊ Γ) is the class defined by the extension (7.1). We will show that the Fredholm module (λµ, Pℓ2Γ) represents [∂Γ]. The initial ingredient is the following lemma, which should be compared with Proposition 3.6 and its proof. Lemma 7.2. If g → ω ∈ ∂X in Γ, then g∗µ → δω in Prob(∂X). Proof. Fix φ ∈ C(∂X). We have (cid:12)(cid:12)(cid:12)(cid:12)Z φ dg∗µ − φ(ω)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)Z φ(gξ) − φ(ω) dµ(ξ)(cid:12)(cid:12)(cid:12)(cid:12) ≤Z (cid:12)(cid:12)φ(gξ) − φ(ω)(cid:12)(cid:12) dµ(ξ). and we show that the right-hand integral converges to 0 as g → ω in Γ. Let t > 0, and let dǫ be a visual metric on ∂X with parameter ǫ in the small visual range so that Fact 5.4 applies. The uniform continuity of φ provides us with some R > 0 such that φ(ξ)− φ(ξ′) < t whenever dǫ(ξ, ξ′) < R. The set Z(g) = {ξ ∈ ∂X : dǫ(gξ, ω) ≥ R} is measurable, since ξ 7→ dǫ(gξ, ω) is continuous. We write: Z (cid:12)(cid:12)φ(gξ) − φ(ω)(cid:12)(cid:12) dµ(ξ) =Z∂X\Z(g)(cid:12)(cid:12)φ(gξ) − φ(ω)(cid:12)(cid:12) dµ(ξ) +ZZ(g)(cid:12)(cid:12)φ(gξ) − φ(ω)(cid:12)(cid:12) dµ(ξ) ≤ t + 2kφk∞ µ(Z(g)) It suffices to show that µ(Z(g)) → 0 as g → ω. Let o ∈ X be a basepoint. One easily checks that (gx, w)o + (g−1o, x)o ≥ (go, w)o for x, w ∈ X. Letting x → ξ and w → ω, and using (5.1) and (5.2), we get (gξ, ω)o + (g−1o, ξ)o ≥ (go, ω)o − 3δ. It follows that there is C1 > 0 such that dǫ(gξ, ω) dǫ(g−1o, ξ) ≤ C1 dǫ(go, ω) K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 21 C2 max(cid:8)dǫ(g−1o, ξ), dǫ(g−1o, ξ′)(cid:9) for all ξ, ξ′ ∈ Z(g). In turn, both dǫ(g−1o, ξ) and dǫ(g−1o, ξ′) for all g ∈ Γ and ξ, ω ∈ ∂X. Now by hyperbolicity there exists C2 > 0 such that dǫ(ξ, ξ′) ≤ are at most C1R−1 dǫ(go, ω) by the inequality above. Thus diam(Z(g)) ≤ C3dǫ(go, ω). By Ahlfors regularity (Fact 5.8) µ(Z(g)) ≤ C4 dǫ(go, ω)eX /ǫ. Now if g → ω then dǫ(go, ω) → 0, so µ(Z(g)) → 0 as desired. (cid:3) In terms of the Γ-expectation, Lemma 7.2 can be interpreted as follows: if g → ω ∈ ∂X in Γ, then Eφ(g) → φ(ω) for all φ ∈ C(∂X). This means that we may extend continuous maps on the boundary ∂X to continuous maps on the compactification Γ by gluing a map to its Γ-expectation. Namely, for φ ∈ C(∂X) we define Eφ ∈ C(Γ) by Eφ =(φ on ∂X Eφ on Γ. We view E : C(∂X) → C(Γ) as a Γ-equivariant, completely positive section for the quotient map C(Γ) ։ C(∂X) given by restriction. We immediately obtain a completely positive section for the quotient map C(Γ) ⋊ Γ ։ C(∂X) ⋊ Γ. Theorem 7.3. Define (7.2) sµ : C(∂X) ⋊alg Γ → C(Γ) ⋊ Γ, sµ(cid:0)X φg g(cid:1) =X(cid:0)Eφg(cid:1) g. Then sµ extends to a completely positive section for the quotient map C(Γ) ⋊ Γ ։ C(∂X) ⋊ Γ. Consequently, (λµ, Pℓ2Γ) is a Fredholm module representing the boundary extension class [∂Γ]. Proof. Recall that π : C(Γ) ⋊ Γ → B(ℓ2Γ) is the representation induced by the multiplication representation of C(Γ) on ℓ2Γ. We claim that (7.3) for all a ∈ C(∂X) ⋊alg Γ. Indeed, for φ ∈ C(∂X) and g ∈ Γ we have πsµ(a) = Pℓ2Γλµ(a)Pℓ2Γ πsµ(φg) = π(cid:0)Eφ(cid:1)π(g) =(cid:0)Pℓ2Γλµ(φ)Pℓ2Γ(cid:1)(cid:0)Pℓ2Γλµ(g)Pℓ2Γ(cid:1) = Pℓ2Γλµ(φg)Pℓ2Γ by using (3.2), and the fact that λµΓ commutes with Pℓ2Γ. Since π is faithful, and therefore isometric, (7.3) implies that sµ extends by continuity to a completely positive section for the quotient map C(Γ) ⋊ Γ ։ C(∂X) ⋊ Γ. The Stinespring dilation πsµ = Pℓ2Γ λµ Pℓ2Γ precisely means that (λµ, Pℓ2Γ) represents [∂Γ]. (cid:3) Theorem 7.3 and Proposition 4.3 imply the following. Proposition 7.4. Let Λ be a finite-index subgroup of Γ. Then the restriction homomorphism res : K1(C(∂Γ) ⋊ Γ) → K1(C(∂Λ) ⋊ Λ) sends [∂Γ] to [Γ : Λ] · [∂Λ]. Indeed, the comparability condition required in the statement of Proposition 4.3 is satisfied in our setting. Since Γ acts by Lipschitz maps on ∂X for any choice of visual metric dǫ, the corresponding Hausdorff measure satisfies g∗µǫ ≍ µǫ for all g ∈ Γ; the same will then hold for any visual probability measure on ∂X. 8. Poincar´e duality and twisted K-cycles 8.1. Poincar´e duality. Poincar´e duality for C(∂Γ) ⋊ Γ, proved in [10], plays an essential role in this paper. The proof from [10], though most likely not Poincar´e duality itself, needs the following mild symmetry condition on the boundary. Definition 8.1. A hyperbolic group is regular if its boundary admits a continuous self-map without fixed-points. 22 HEATH EMERSON AND BOGDAN NICA Regularity in the above sense is satisfied whenever the boundary is a topological sphere, a Cantor set or a Menger compactum. We are not aware of any example where regularity fails. The 'topologically rigid' hyperbolic groups of Kapovich and Kleiner [24] come close, though, for their boundaries admit no self-homeomorphisms without fixed points (but regularity does not require a homeomorphism, merely a map). Poincar´e duality is defined by a cup-cap procedure explained in [10]. A C-algebra is Poincar´e self-dual in this sense if there are two 'fundamental classes', satisfying certain equations which we do not specify here (the zig-zag equations of the theory of adjoint functors.) Given ∆ as above, a 'Poincar´e duality' map is defined by ∆ ∈ KK1(A ⊗ A, C), b∆ ∈ KK1(C, A ⊗ A) ∆∩ : K∗(A) → K∗+1(A), ∆ ∩ x = (x ⊗C 1A) ⊗A⊗A ∆, or in other words, by composing in KK, the morphisms x ⊗ 1A ∈ KK∗(A, A ⊗ A) with ∆ ∈ KK1(A ⊗ A, C), to get a morphism in KK∗+1(A, C) = K∗+1(A). Baum - Connes machinery, it is shown that ∆∩ is an isomorphism when Γ is regular and i.e. as a cycle for the Brown - Douglas - Fillmore group Ext(C(∂Γ) ⋊ Γ ⊗ C(∂Γ) ⋊ Γ, C), whose Busby invariant is as follows. First let For C(∂Γ) ⋊ Γ, such fundamental classes ∆ and b∆ are constructed in [10], and, using the torsion-free. The inverse isomorphism comes from b∆. The class ∆ is defined as an extension, τ : C(∂Γ) ⋊ Γ → B(ℓ2Γ)/K(ℓ2Γ) be the Busby invariant of the extension (7.1). Thus τ is the integrated form of the covariant pair associated to the regular representation λ : Γ → B(ℓ2Γ) followed by the quotient map B → B/K, and the map φ 7→ M( φ) followed by the quotient map, where φ is an extension of φ ∈ C(∂Γ) to a continuous function on Γ. Let τ op : C(∂Γ) ⋊ Γ → B(ℓ2Γ)/K(ℓ2Γ), τ op(a) := Jτ (a)J where J is the symmetry J : ℓ2Γ → ℓ2Γ, J(δg) := δg−1 . Note that sop : C(∂Γ)⋊Γ → JC(Γ)⋊ΓJ ⊂ B(l2Γ), defined by sop(a) := Js(a)J is a completely positive splitting of τ op. As proved in [10], τ and τ op commute so they combine to give a unital ∗-homomorphism C(∂Γ) ⋊ Γ ⊗ C(∂Γ) ⋊ Γ → B(ℓ2Γ)/K(ℓ2Γ), a ⊗ b 7→ τ (a)τ op(b). The class ∆ is by definition the pre-image of this extension under the canonical map KK1(C(∂Γ) ⋊ Γ ⊗ C(∂Γ) ⋊ Γ, C) → Ext(C(∂Γ) ⋊ Γ ⊗ C(∂Γ) ⋊ Γ, C), which is an isomorphism since C(∂Γ) ⋊ Γ ⊗ C(∂Γ) ⋊ Γ is nuclear. However, a cycle in KK1 representing ∆ is difficult to describe because we do not know of a concrete completely positive splitting of the extension defining ∆. Nevertheless, we show that the ideas of the previous sections can be used to compute the Poincar´e duality map ∆∩ in explicit terms. Lemma 8.2. Let e ∈ C(∂Γ) ⋊ Γ be a projection. Then the Poincar´e dual ∆ ∩ [e] of the class [e] ∈ K0(C(∂Γ) ⋊ Γ) is the class in K1(C(∂Γ) ⋊ Γ) = KK1(C(∂Γ) ⋊ Γ, C) ∼= Ext(C(∂Γ) ⋊ Γ, C) of the extension with Busby invariant τe : C(∂Γ) ⋊ Γ → B(ℓ2Γ)/K(ℓ2Γ), τe(a) := τ op(e)τ (a)τ op(e). K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 23 Let u ∈ C(∂Γ) ⋊ Γ be a unitary, and let ¯u : C0(R) ⊂ C(S1) → C(∂Γ) ⋊ Γ denote the composition of the usual inclusion of C0(R) ∼= C0(S1 − {1}) into C(S1), functional calculus for u. Then the Poincar´e dual ∆ ∩ [u] of the class [u] ∈ K1(C(∂Γ) ⋊ Γ) is the class in KK1(C0(R) ⊗ C(∂Γ) ⋊ Γ, C) of the extension of C0(R) ⊗ C(∂Γ) ⋊ Γ by K(ℓ2Γ) whose Busby invariant is τu(f ⊗ a) = (τ op ◦ ¯u)(f )τ (a). Proof. Both assertions follow from the functoriality of the Kasparov and Brown - Douglas - Fillmore theories. We prove the first assertion, concerning projections. The second one, concerning unitaries, is proved similarly. If e is a projection in C(∂Γ) ⋊ Γ, then [e] = e∗([1]) where e∗ : K∗(C) ∼= Z → K∗(C(∂Γ) ⋊ Γ) is the ∗-homomorphism induced by e, and the element [e] ⊗ 1C(∂Γ)⋊Γ ∈ KK0(cid:0)C(∂Γ) ⋊ Γ, C(∂Γ) ⋊ Γ ⊗ C(∂Γ) ⋊ Γ(cid:1) is represented by the ∗-homomorphism e ⊗ 1C(∂Γ)⋊Γ, and composing with ∆ amounts to com- posing the Busby invariant for ∆ and the ∗-homomorphism e ⊗ 1. By the definitions and the fact that τ op(e) is a projection in the Calkin algebra commuting with τ (a) for all a, this yields τe. (cid:3) Note that the boundary extension class [∂Γ] ∈ K1(C(∂Γ) ⋊ Γ) is the Poincar´e dual of the unit class [1C(∂Γ)⋊Γ] ∈ K0(C(∂Γ) ⋊ Γ). On the other hand, the K-theory Gysin sequence of [11] shows that the order of [1C(∂Γ)⋊Γ] in K0(C(∂Γ) ⋊ Γ) is determined by the Euler characteristic of Γ. Consequently: Proposition 8.3 (Emerson, Emerson - Meyer). Assume that Γ is regular and torsion-free. Then [∂Γ] = 0 in K1(C(∂Γ) ⋊ Γ) if and only if χ(Γ) = ±1. Furthermore, [∂Γ] has infinite order in K1(C(∂Γ) ⋊ Γ) if and only if χ(Γ) = 0. We may pass from torsion-free to virtually torsion-free groups, and establish a version of Proposition 8.3 for this much larger class, by using Proposition 7.4. The rational Euler characteristic of a virtually torsion-free group Γ is defined by the formula χ(Γ) = χ(Λ)/[Γ : Λ] where Λ is any torsion-free subgroup of finite index. Corollary 8.4. Assume that Γ is regular and virtually torsion-free. [∂Γ] 6= 0 in K1(C(∂Γ) ⋊ Γ). If χ(Γ) = 0 then [∂Γ] has infinite order in K1(C(∂Γ) ⋊ Γ). If χ(Γ) /∈ 1/Z then The assumption of virtual torsion-freeness is a very mild one. A long-standing open problem asks whether all hyperbolic groups are virtually torsion-free. 8.2. Twisted K-cycles. We now show how to compute ∆∩ in terms of certain canonical Fredholm modules obtained by 'twisting' the basic K-cycle (λµ, Pℓ2Γ). µ of C(∂X) ⋊ Γ on ℓ2(Γ, L2(∂X, µ)) is given as follows: The right regular representation λop λop µ (φ)(cid:16)X ψhδh(cid:17) =X(h.φ)ψhδh, λop µ (g)(cid:16)X ψhδh(cid:17) =X ψhδhg−1 for φ ∈ C(∂X), g ∈ Γ. Note the covariance relation λop right and the left regular representations do not commute, but they do satisfy µ (g.φ) = λop µ (g)λop µ (φ)λop µ (g−1). The [λµ(φ), λop for all φ, φ′ ∈ C(∂X) and g, g′ ∈ Γ. µ (φ′)] = 0, [λµ(g), λop µ (g′)] = 0 24 HEATH EMERSON AND BOGDAN NICA The symmetry J on ℓ2Γ has an obvious extension J ⊗ id to ℓ2(Γ, L2(∂X, µ)), and, using the same notation for this extension, we have λop µ = JλµJ. Since s splits τ , the image in the Calkin algebra of Pℓ2Γλµ(a)Pℓ2Γ = s(a) is τ (a). Also, Pℓ2Γλop µ (a)Pℓ2Γ = sop(a), and its image in the Calkin algebra is τ op(a). Hence (8.1) Pℓ2Γλµ(a)Pℓ2Γ = τ (a) mod K, and Pℓ2Γλop µ (a)Pℓ2Γ = τ op(a) mod K. Theorem 8.5. Let Γ be regular and torsion-free. Then the following hold. • Every class in K1(C(∂Γ) ⋊ Γ) is represented by an odd Fredholm module of the form e projection in C(∂Γ) ⋊ Γ. • Every class in K0(C(∂Γ) ⋊ Γ) is represented by a balanced even Fredholm module of the form u unitary in C(∂Γ) ⋊ Γ. (cid:0)λµ, Pℓ2Γλop µ (e)Pℓ2Γ(cid:1), (cid:0)λµ, Pℓ2Γλop µ (u)Pℓ2Γ + (1 − Pℓ2Γ)(cid:1), Proof. Let Qe := Pℓ2Γλop µ (e)Pℓ2Γ. By the discussion preceding the Theorem, Qe = τ op(e) mod compact operators. Hence Qe is a self-adjoint, essential projection. Furthermore, the commutator [λµ(a), Qe] is compact for all a ∈ C(∂Γ) ⋊ Γ. For mod compacts we have λµ(a)Qe = λµ(a)Pℓ2Γλop µ (e)Pℓ2Γ = Pℓ2Γλµ(a)Pℓ2Γλop µ (e)Pℓ2Γ = τ (a)τ op(e) = τ op(e)τ (a) = Qeλµ(a). This shows that (λµ, Pℓ2Γλop µ (e)Pℓ2Γ) is a Fredholm module. The map KK1 to Ext sends its class to the extension with Busby invariant a 7→ Qeλµ(a)Qe mod K and this equals τe(a) := τ op(e)τ (a)τ op(e). Hence (λµ, Pℓ2Γλop µ (e)Pℓ2Γ) represents [τe(a)], which equals ∆∩[e] by Lemma 8.2. The second assertion is proved by combining the same observations with Bott periodicity, (cid:3) see Lemma 2 of [10]. We are implicitly using the fact that all the K-theory classes of a unital, simple and purely infinite C-algebra can be represented by non-zero projections, respectively unitaries of the C-algebra [9]. 9. Summability of the twisted K-cycles In this section we prove the following theorem. Theorem 9.1. Let Γ be regular and torsion-free. Then C(∂X)⋊Γ has uniformly D> K-homology over Lip(∂X, dǫ) ⋊alg Γ. ǫ -summable Theorem 1.3 is an immediate corollary. In turn, Theorem 1.3 and Examples 5.11, 5.12 yield the following explicit applications. Corollary 9.2. If Γ is a finitely generated free group, then the K-homology of C(∂Γ) ⋊ Γ is uniformly p-summable for every p > 2. If Γ is a torsion-free cocompact lattice in SO(n, 1), then the K-homology of C(Sn−1) ⋊ Γ is uniformly (n− 1)+-summable when n ≥ 4, respectively uniformly p-summable for every p > 2 when n = 2, 3. If Γ is a torsion-free cocompact lattice in SU(n, 1), then the K-homology of C(S2n−1) ⋊ Γ is uniformly (2n)+-summable. If Γ is a torsion-free cocompact lattice in Sp(n, 1), then the K-homology of C(S4n−1) ⋊ Γ is uniformly (4n + 2)+-summable. To prove Theorem 9.1, we start with an integral estimate. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 25 Lemma 9.3. Let o ∈ X be a basepoint, and assume that Dǫ > 2. Then there exists C > 0 such that, for all g, h ∈ Γ, we have Z dǫ(gξ, hξ) dµ(ξ) ≤ C exp(−ǫ (go, ho)o). Proof. The proof is similar to that of Lemma 6.1. First, we assume that the parameter ǫ is in the small visual range and we let α > 0. We claim the following: if Dǫ > 2α, then there exists C > 0 such that for all g, h ∈ Γ we have (9.1) Z dǫ(gξ, hξ)α dµ(ξ) ≤ C exp(−αǫ (go, ho)o). Pick ξ ∈ ∂X. Observe that (gx, hx)o + (g−1o, x)o + (h−1o, x)o ≥ (go, ho)o for x ∈ X; indeed, this amounts to d(gx, hx) − d(go, ho) ≤ 2d(o, x). Letting x → ξ and using (5.1) and (5.2), we get (gξ, hξ)o + (g−1o, ξ)o + (h−1o, ξ)o ≥ (go, ho)o − 4δ. Hence, there is C1 ≥ 0 such that, for all ξ ∈ ∂X, we have (9.2) dǫ(gξ, hξ) ≤ C1 exp(−ǫ (go, ho)o) dǫ(g−1o, ξ)−1 dǫ(h−1o, ξ)−1. Recall from the proof of Lemma 6.1 that Z dǫ(g−1o, ξ)−2α dµ(ξ) ≤ C2 independently of g ∈ Γ. By the Cauchy-Schwartz inequality, it follows that (9.3) Z dǫ(g−1o, ξ)−α dǫ(h−1o, ξ)−α dµ(ξ) ≤ C2 independently of g, h ∈ Γ. Now (9.2) and (9.3) yield (9.1). The remainder of the proof goes just like the last step in the proof of Lemma 6.1. Let ǫ be an arbitrary visual parameter, let ǫ0 be in the small visual range, and let dǫ0 be a visual metric as in Fact 5.4. We have Z dǫ(gξ, hξ) dµ(ξ) ≍Z dǫ0(gξ, hξ)ǫ/ǫ0 dµ(ξ) by Fact 5.3. As hdim(∂X, dǫ0) > 2ǫ/ǫ0, the previous part says that a constant multiple of exp(−(ǫ/ǫ0)ǫ0 (go, ho)o) = exp(−ǫ (go, ho)o) is an upper bound for the right hand side. (cid:3) Lemma 9.4. For all a, b ∈ Lip(∂X, dǫ) ⋊alg Γ, the commutator (cid:2)λµ(a), Pℓ2Γλop D> ǫ -summable. Proof. Assume that Dǫ > 2. The case when Dǫ ≤ 2 is deduced as in the proof of Theorem 6.2. We recall from Theorem 6.2 that λµ(a) commutes mod LDǫ+ with Pℓ2Γ. It follows that µ (b)Pℓ2Γ(cid:3) is (cid:2)λµ(a), Pℓ2Γλop µ (b)Pℓ2Γ(cid:3) =(cid:2)Pℓ2Γλµ(a)Pℓ2Γ, Pℓ2Γλop µ (b)Pℓ2Γ(cid:3) mod LDǫ+ and the right-hand commutator is, with our notations, [sµ(a), sop µ (b)]. Clearly, the property µ (b)] ∈ LDǫ+ is additive in a and b. Now sµ(aa′) = sµ(a)sµ(a′) mod LDǫ+ for that [sµ(a), sop µ = JsµJ is multiplicative mod LDǫ+ on Lip(∂X, dǫ) ⋊alg Γ a, a′ ∈ Lip(∂X, dǫ) ⋊alg Γ, hence sop µ (b)] ∈ LDǫ+ is also multiplicative in a and b. as well. Therefore, the property that [sµ(a), sop We thus see that it suffices to treat the case when a and b are either Lipschitz functions or group elements. For all g, g′ ∈ Γ and φ, φ′ ∈ Lip(∂X, dǫ) we have [s(g), sop(g′)] = 0, [s(φ), sop(φ′)] = 0, [s(g), sop(φ)] = −J[s(φ), sop(g)]J. 26 HEATH EMERSON AND BOGDAN NICA It therefore suffices to analyze the summability of the commutator [s(φ), sop(g)] or, more conve- niently, the summability of sop(g−1)[s(φ), sop(g)] which is readily verified to be multiplication by h 7→ Eφ(hg−1) − Eφ(h) on ℓ2Γ. For h1, h2 ∈ Γ and φ ∈ Lip(∂X, dǫ) we have (cid:12)(cid:12)Eφ(h1) − Eφ(h2)(cid:12)(cid:12) ≤Z (cid:12)(cid:12)φ(h1ξ) − φ(h2ξ)(cid:12)(cid:12) dµ(ξ) ≤ kφkLipZ dǫ(h1ξ, h2ξ) dµ(ξ) so, by Lemma 9.3, there exists a constant C > 0 such that (cid:12)(cid:12)Eφ(h1) − Eφ(h2)(cid:12)(cid:12) ≤ CkφkLip exp(−ǫ (h1o, h2o)o). Put h1 = hg−1 and h2 = h. As (hg−1o, ho)o = (g−1o, o)h−1o ≥ d(o, ho) − d(o, go), we obtain (cid:12)(cid:12)Eφ(hg−1) − Eφ(h)(cid:12)(cid:12) ≤ CkφkLip exp(ǫ d(o, go)) exp(−ǫ d(o, ho)). Finally, we recall from the proof of Theorem 6.2 that multiplication by h 7→ exp(−ǫ d(o, ho)) is in LDǫ+(ℓ2Γ). Theorem 9.5. There is a smooth subalgebra A ⊆ C(∂X) ⋊ Γ, containing Lip(∂X, dǫ) ⋊alg Γ, such that the odd, respectively even, Fredholm modules (cid:3) (cid:0)λµ, Pℓ2Γλop (cid:0)λµ, Pℓ2Γλop µ (e)Pℓ2Γ(cid:1), µ (u)Pℓ2Γ + (1 − Pℓ2Γ)(cid:1), e projection in A are D> ǫ -summable Fredholm modules over Lip(∂X, dǫ) ⋊alg Γ. u unitary in A We recall that a subalgebra A of a C-algebra A is said to be smooth if it is dense and stable under holomorphic calculus. Then the projections of A are dense in the projections of A, and the unitaries of A are dense in the unitaries of A. In particular, the K-theory classes of A can be represented by projections, respectively unitaries, from A. In light of this fact, Theorem 9.1 follows by combining Theorem 9.5 and Theorem 8.5. Proof of Theorem 9.1. To fix ideas, let us assume that Dǫ > 2. Consider the ∗-subalgebra A =(cid:8)a ∈ C(∂X) ⋊ Γ :(cid:2)λµ(a), Pℓ2Γλop C(∂X) ⋊ Γ :(cid:2)λµ(a), Pℓ2Γλop Then A contains Lip(∂X, dǫ)⋊algΓ, by Lemma 9.4, in particular A is dense in C(∂X)⋊Γ. To see that A is stable under holomorphic calculus, consider for a moment the subalgebra Ab = {a ∈ Then Ab is stable under holomorphic calculus in C(∂X) ⋊ Γ, therefore the same holds true for A, the intersection of the family of subalgebras {Ab : b ∈ Lip(∂X, dǫ) ⋊alg Γ}. µ (b)Pℓ2Γ(cid:3) ∈ LDǫ+ for all b ∈ Lip(∂X, dǫ) ⋊alg Γ(cid:9). µ (b)Pℓ2Γ(cid:3) ∈ LDǫ+} corresponding to a fixed b ∈ Lip(∂X, dǫ) ⋊alg Γ. (cid:2)λµ(a), Pℓ2Γ(cid:3) ∈ LDǫ+, Next, we prove the summability claim. Note first that µ (a), Pℓ2Γ(cid:3) ∈ LDǫ+ for all a ∈ A. It follows that Pℓ2Γλop µ (e)Pℓ2Γ is a projection mod LDǫ+ whenever e ∈ A is a projection, and that Pℓ2Γλop µ (u)Pℓ2Γ + (1 − Pℓ2Γ) is a unitary mod LDǫ+ whenever u ∈ A is a unitary. ǫ -summable for every b ∈ Lip(∂X, dǫ) ⋊alg Γ. As already hinted in the proof of Lemma 9.4, the summabil- Indeed, using the fact that ity of the above commutators is in fact symmetric in a and b. Then also(cid:2)λop By definition, if a ∈ A then the commutator (cid:2)λµ(a), Pℓ2Γλop (cid:2)λµ(a), Pℓ2Γ(cid:3) ∈ LDǫ+ for all a ∈ A, we may write µ (b)Pℓ2Γ(cid:3) =(cid:2)Pℓ2Γλµ(a)Pℓ2Γ, Pℓ2Γλop µ (a)Pℓ2Γ(cid:3) =(cid:2)Pℓ2Γλµ(b)Pℓ2Γ, Pℓ2Γλop µ (b)Pℓ2Γ(cid:3) is D+ µ (b)Pℓ2Γ(cid:3) mod LDǫ+, µ (a)Pℓ2Γ(cid:3) mod LDǫ+. (cid:2)λµ(a), Pℓ2Γλop (cid:2)λµ(b), Pℓ2Γλop for all a ∈ A. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 27 shows that, if b ∈ Lip(∂X, dǫ) ⋊alg Γ, then the commutator (cid:2)λµ(b), Pℓ2Γλop Now observe that the right-hand side commutators are conjugate by the symmetry J. This ǫ - summable for every a ∈ A. We conclude that the indicated Fredholm modules are D+ ǫ - summable over Lip(∂X, dǫ) ⋊alg Γ. (cid:3) µ (a)Pℓ2Γ(cid:3) is D+ 10. The K-homology Gysin sequence for boundary actions In this section, we attack the problem of proving that the reduced C-algebra of a hyperbolic group has uniformly summable K-homology. This involves some tools from KK-theory. We start by summarizing the basic facts we will need about 'γ-elements' and the Dirac dual-Dirac method. 10.1. Descent, γ-elements. For any discrete group (or more generally locally compact group) Γ, 'descent,' in Kasparov theory, refers to a natural map j : KKΓ ∗ (A, B) → KK∗(A ⋊ Γ, B ⋊ Γ) which extends to equivariant KK-cycles (and homotopies) the process of integrating a Γ- equivariant ∗-homomorphism A → B to an ordinary ∗-homomorphism A ⋊ Γ → B ⋊ Γ. Either the maximal or the reduced crossed-product can be used; thus there is also a 'reduced' descent map jr : KKΓ ∗ (A, B) → KK∗(A ⋊r Γ, B ⋊r Γ) in which the reduced is used. Descent j (respectively reduced descent jr) makes the abelian group KK(A ⋊ Γ, B ⋊ Γ) (respectively KK(A ⋊r Γ, B ⋊r Γ) a left module over the ring KKΓ ∗ (C, C), and likewise a right module, using the structure of KKΓ(A, B) as a module over KKΓ(C, C), for any Γ-C-algebras A, B. The γ-element is defined contingent on the existence of a proper Γ-C-algebra P and classes η ∈ KKΓ(C, P ) and D ∈ KKΓ(C, P ) such that D ⊗C η = 1P ∈ KKΓ(P, P ), as the idempotent η ⊗P D ∈ KKΓ(C, C). The existence of a γ-element is not guaranteed for arbitrary discrete groups, but a group can have at most one γ-element, as one can argue without difficulty. The existence issue involves the existence of η, called the dual-Dirac morphism: it can be shown (see [33] and [12]) that for any Γ, there exist proper P and a morphism D ∈ KKΓ(P, C) (the Dirac morphism) such that existence of η is equivalent to a coarse geometric condition on the group, namely, that the 'coarse co-assembly map' for Γ is an isomorphism (the coarse co- assembly map is described in [12]). The coarse co-assembly map is, however, an isomorphism for all hyperbolic groups, and more generally, for groups which uniformly embed in a Hilbert space, so all such groups have γ-elements. The first explicit construction of them in the case of hyperbolic groups is due to Kasparov and Skandalis [27]. It is not true that γ = 1 ∈ KKΓ 0 (C, C) for general hyperbolic groups, 1 being the class 1 := [ǫ] ∈ KKΓ 0 (C, C) of the trivial representation ǫ : C∗Γ → C. An argument of Skandalis [39] even gives examples where jr(γ) 6= 1C ∗ For cocompact lattices in SO(n, 1) or SU(n, 1), or free groups, γ = 1 is true due to results of Kasparov [25], and Kasparov and Julg [20]. These groups are also known to be a-T-menable, so γ = 1 follows from the Higson - Kasparov theorem (see [18]) as well. r Γ ∈ KK0(C∗ r Γ, C∗ r Γ). For our purposes, we are mostly concerned about whether γ acts as the identity on various r Γ). When Γ is hyperbolic, recent work of Lafforgue and others r Γ, but nothing seems to be KK-groups, especially K∗(C∗ [29, 34] shows that γ does act as the identity on the K-theory of C∗ known at present about the case of K-homology. 28 HEATH EMERSON AND BOGDAN NICA The point of the γ-part, is that it is the 'topologically accessible' part of the K-homology, in the sense of the following theorem which is essentially due to Kasparov. See Theorem 23 of [12] and Theorem 7.1 of [33]. Lemma 10.1 (Kasparov). Let Γ be a discrete group with a γ-element and EΓ its classifying space for proper actions. Then the canonical inflation map of [26] is an isomorphism from the γ-part of KKΓ EΓ : KKΓ p∗ ∗ (A, B) → RKKΓ ∗ (EΓ; A, B) ∗ (A, B) onto its target. Here RKKΓ(EΓ; A, B) is the Γ-equivariant representable K-theory of EΓ. If A = B = C and G\EΓ is compact, then it agrees with the ordinary K-theory of C0(EΓ) ⋊ Γ, and if in addition G is torsion-free, then it is isomorphic to K∗(Γ\EΓ) – see [13] for information on equivariant representable K-theory. Finally, we remind the reader that since a Gromov hyperbolic group acts amenably on its boundary, γ acts as the identity on KKΓ ∗ (C(∂Γ) ⊗ A, B) for any A, B. (The Dirac dual-Dirac method gives a KKΓ-equivalence between C(∂Γ) and a proper Γ-C-algebra, while γ acts as the identity on any KKΓ(P, B)-group when P is proper, by properties of γ – see [33].) Corollary 10.2. For any Γ-C-algebras A, B, ∗ (C(∂Γ) ⊗ A, B) ∼= RKKΓ KKΓ ∗ (C(Γ) ⊗ A, B) ∼= RKKΓ KKΓ ∗ (EΓ; C(∂Γ) ⊗ A, B) ∗ (EΓ; C(Γ) ⊗ A, B) by the inflation map p∗ EΓ. 10.2. The γ-element regarded as a K-homology class for C∗ r Γ be the projection from the maximal group C-algebra to the reduced group C-algebra, and let us call any element γr ∈ K0(C∗ r Γ. Let λ : C∗Γ → C∗ r Γ) such that λ∗(γr) = γ ∈ K0(C∗Γ) ∼= KKΓ 0 (C, C) a reduced γ-element for Γ. Proposition 10.3. The map λ∗ : K∗(C∗ ∗ (C, C) induces an isomorphism between the γ-parts of these two rings. In particular, if Γ has a γ-element, then it has a reduced γ-element. r Γ) → K∗(C∗Γ) ∼= KKΓ The (or any such) element γr as in the Proposition will play a role in the 'Gysin sequence' developed in the next subsection. r Γ) inverting λ∗ as follows. We first recall that Proof. We produce a map γK∗(C∗Γ) → γK∗(C∗ the standard identification KKΓ ∗ (A, C) → KK∗(A ⋊ Γ, C), coming from the fact that the groups have the same cycles when Γ is discrete, can be expressed in terms the 'descent' construction and the trivial representation of the group in the following way: it is the composition of the descent map j : KKΓ ∗ (A, C) → KK∗(A ⋊ Γ, C∗Γ), and ǫ∗ : KK∗(A ⋊ Γ, C∗Γ) → KK∗(A ⋊ Γ, C). In particular, taking γ ∈ KKΓ 0 (C, C), its image under the isomorphism with KK0(C∗Γ, C) = K0(C∗Γ) is j(γ)⊗C ∗Γ [ǫ]. With these formalities aside, we next factor the γ-element, or rather, its image in KK0(C∗Γ, C) as follows. Let η ∈ KKΓ(C, P ) be the dual-Dirac morphism, and let D ∈ KKΓ(P, C) be the Dirac morphism for Γ. Then j(η)⊗P ⋊Γ j(D)⊗C ∗Γ [ǫ] factors the image of γ in KK0(C∗Γ, C). This is because γ = η ⊗P D, and the naturality of the descent map j. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 29 More generally, any a ∈ γKKΓ factored as ∗ (C, C), interpreted as an element of K∗(C∗Γ), can be thus (10.1) j(a) ⊗C ∗Γ [ǫ] = j(a ⊗C γ) ⊗C ∗Γ [ǫ] = j(a ⊗C η ⊗P D) ⊗C ∗Γ [ǫ] = j(a) ⊗C ∗Γ j(η) ⊗P ⋊Γ j(D) ⊗C ∗Γ [ǫ] where the first equality is due to γ ⊗C a = a for a in the γ-part, the third by the naturality of the descent map. Now to obtain an element a′ such that λ∗(a′) = a, consider the element b′ = jr(a) ⊗C ∗ r Γ jr(η) ∈ KK∗(C∗ r Γ, P ⋊r Γ) defined using the reduced descent map. Now P being proper implies P ⋊r Γ ∼= P ⋊ Γ. Applying this isomorphism to b′ gives a class b ∈ KK∗(C∗ r Γ, P ⋊ Γ). Then the required element a′ such that λ∗(a′) = a is a′ := b ⊗P ⋊Γ j(D) ⊗P ⋊Γ [ǫ]. (cid:3) Remark 10.4. In particular, Kasparov's Theorem (Lemma 10.1) can be alternately phrased in terms of the K-homology of the reduced C-algebra: the γ-part of the K-homology of C∗ r Γ is isomorphic to the topological group RKKΓ(EΓ; C, C) (by the composition of λ∗ and the inflation map.) 10.3. The Gysin sequence. Let Γ be a hyperbolic group, ∂Γ its Gromov boundary, etc. Let iΓ : C → C(∂Γ) be the natural inclusion of C as constant functions on ∂Γ, defining a morphism [iΓ] ∈ KKΓ 0 (C, C(∂Γ)) and then, by reduced descent, a morphism [i] := jr([iΓ]) ∈ KK0(C∗ r Γ, C(∂Γ) ⋊ Γ), which is nothing but the Kasparov morphism determined by the C- algebra injection i : C∗ r Γ → C(∂Γ)⋊Γ of the reduced C-algebra in the reduced crossed-product. r Γ). The aim of this section is to compute this map. We first observe that the range of this map is contained in the γ-part of K∗(C∗ Then composition with [i] induces a map i∗ : K∗(C(∂Γ) ⋊ Γ) → K∗(C∗ r Γ). More generally: Lemma 10.5. Let A be any Γ-C-algebra and αΓ : A → C(∂Γ) a Γ-equivariant -homomorphism. Let α : A ⋊r Γ → C(∂Γ) ⋊ Γ be the induced -homomorphism. Then the range of the induced map α∗ : K∗(cid:0)C(∂Γ) ⋊ Γ(cid:1) → K∗(A ⋊r Γ) is contained in the γ-part of K∗(A ⋊r Γ). ∗ (C(∂Γ), C) → KKΓ Γ : KKΓ Proof. Since γ acts as the identity on K∗(C(∂Γ) ⋊ Γ) and α∗ a KKΓ γKKΓ ∗ (C, C)-module map, for any x ∈ KKΓ ∗ (A, C). The result follows. ∗ (A, C) is ∗ (C(∂Γ), C) it holds that i∗(x) = i∗(γx) = γi∗(x) ∈ (cid:3) Let X be a Rips complex for Γ which models EΓ (see [32]). Let • r : C(X) → C(∂X) ∼= C(∂Γ) be the Γ-equivariant map of restriction to the boundary, • u : C → C(X) be the inclusion as constant functions. Both maps are Γ-equivariant. By Lemma 10.5, the range of r∗ is contained in the γ-part of KKΓ ∗ (C(X), C). Lemma 10.6. The map u∗ : KKΓ ∗ (C(X), C) → KKΓ ∗ (C, C) 30 HEATH EMERSON AND BOGDAN NICA on KKΓ-theory induced by composition with u ∈ KKΓ between the γ-parts of the domain and co-domain. Moreover, the composition 0 (C, C(X)), restricts to an isomorphism KKΓ ∗ (C(∂Γ), C) r∗ −→ γKKΓ ∗ (C(X), C) u∗ −→ γKKΓ ∗ (C, C) equals i∗ Γ. Proof. Recalling that X = EΓ, Lemma 10.1 says that the inflation map X : KKΓ p∗ ∗ (A, B) → RKKΓ ∗ (X; A, B) is an isomorphism from the γ-part of KKΓ ∗ (X; A, B). On the other hand, X is H-equivariantly contractible for every finite subgroup H of Γ. In other worlds C(X) is H-equivariantly homotopy equivalent to C for every finite H ⊂ Γ, equivalently, u : C → C(Γ) regarded as an element of KKH X (u) is invertible. So u∗ is invertible between the γ-parts of KKΓ 0 (C, C(X)) is invertible for every such H. Hence by [33], p∗ ∗ (A, B) to RKKΓ ∗ (C(X), C) and KKΓ ∗ (C, C). The last statement is left to the reader. (cid:3) The basis of the arguments to follow is the Γ-exact sequence (10.2) 0 → C0(X) → C(X) → C(∂X) ∼= C(∂Γ) → 0 of Γ-C-algebras. Let us make a few preliminary observations regarding excision in KK-theory for this situation. Firstly, corresponding to the sequence (10.2) is a commutative diagram (10.3) 0 0 / C0(X) ⋊max Γ C(X) ⋊max Γ / C(∂Γ) ⋊max Γ / C0(X) ⋊r Γ / C(X) ⋊r Γ / C(∂Γ) ⋊r Γ 0 / 0 of crossed-products, with exact rows. The vertical maps are the natural ones, from the maximal crossed-products to the reduced. The top row is exact because we are using the maximal crossed-product; the lower row is exact by exactness of Γ. As the first and third vertical maps are isomorphisms, so is the middle one. This give a quick proof that C(X) ⋊max Γ ∼= C(X) ⋊r Γ, which in turn implies that Next, by nuclearity of C(∂Γ) ⋊ Γ, the exact sequence KKΓ ∗ (C(X), C) ∼= K∗(cid:0)C(X) ⋊r Γ). in which all crossed-products are reduced, induces a long exact sequence of K-homology groups 0 → C0(X) ⋊ Γ → C(X) ⋊ Γ → C(∂Γ) ⋊ Γ → 0 (10.4) ··· ← K∗(C0(X) ⋊ Γ) ← K∗(C(X) ⋊ Γ) ← K∗(C(∂Γ) ⋊ Γ) ← ··· Since all the K-homology groups in this sequence are isomorphic to their equivariant counter- parts, we can view this, and prefer to do so, as an exact sequence of equivariant K-homology groups (10.5) ··· ← KKΓ ∗ (C0(X), C) ← KKΓ ∗ (C(X), C) ← KKΓ ∗ (C(∂X), C) ← ··· The map KKΓ ∗ (C(X), C) → KKΓ ∗ (C0(X), C) in this sequence will be denoted ϕ!: it is the map on K-homology induced by the equivariant ∗-homomorphism C0(X) → C(X), while the map KKΓ ∗ (C(X), C) is the map r∗ on equivariant K-homology induced by the restriction homomorphism r : C(X) → C(∂X). ∗ (C(∂Γ), C) → KKΓ / / / / / / / / / / K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 31 Now γ is an idempotent and all maps in this exact sequence are KKΓ 0 (C, C)-module maps so commute with γ. It follows that, taking γ-parts, gives an exact sequence (10.6) ··· ← γKKΓ ∗ (C0(X), C) ←− γKKΓ ϕ! ∗ (C(X), C) r∗ ←− γKKΓ ∗ (C(∂X), C) ← ··· Combining this sequence with the sequence (10.5), which maps to it, and applying the Five Lemma gives that actually γ acts as the identity on all the groups in (10.5). So (10.6) can be used in place of (10.5), as they are exactly the same sequence. By Lemma 10.6 we can replace the middle term γKKΓ ∗ (C(X), C) in (10.6) by γKKΓ(C, C). With this replacement, the map r∗ is replaced by (the map induced on γ-parts by) i∗ Γ since r ◦ u = iΓ. Exactly as in [11] one computes that the map ϕ! becomes the map induced on γ-parts by the map KKΓ ∗ (C0(X), C) of external product with the Euler class defined to be ∗ (C, C) → KKΓ EulΓ := (p∗ X )−1(∆X ) ∈ KKΓ 0 (C0(X), C) where ∆X ∈ RKKΓ 0 (X; C0(X), C) is the morphism induced by the diagonal embedding X → X × X. (See §3 of [11]). Now if Γ is torsion-free then, since Γ\X is compact and models BΓ, we have KKΓ(C0(X), C) ∼= K0(Γ\X) ∼= K0(BΓ), and under this identification, the Euler class for Γ is just the ordinary Euler characteristic of Γ (an integer, equal to the Euler characteristic of BΓ) multiplied by the class of a point in K-homology (see [11]). So we can insert this into the sequence (10.5) in the torsion-free case. Finally, using the fact that γ acts as the identity on both KKΓ ∗ (C0(X), C) and on KKΓ ∗ (C(∂X), C), we obtain the following. Theorem 10.7 (Gysin sequence for K-homology). Let Γ be a torsion-free hyperbolic group. Then there is an exact sequence 0 → K1(BΓ) → K0(C(∂Γ) ⋊ Γ) i∗ −→ γKKΓ 0 (C, C) Eul−−→ K0(BΓ) → K1(C(∂Γ) ⋊ Γ) i∗ −→ γKKΓ 1 (C, C) → 0 r Γ) is the map induced by the inclusion i : C∗ where i∗ : K∗(C(∂Γ) ⋊ Γ) → K∗(C∗ r Γ → C(∂Γ) ⋊ Γ, and where Eul is the map Eul(a) = χ(Γ) index(a) [pnt] ∈ K0(BΓ), with index the ordinary Fredholm index map KKΓ(C, C) → Z, and [pnt] is the class in K-homology of a point in BΓ. Corollary 10.8. The restriction homomorphism i∗ : K∗(C(∂Γ) ⋊ Γ) → γK∗(C∗ r Γ) is a surjec- tion in dimension ∗ = 1, and a surjection in both dimensions if χ(Γ) = 0. When χ(Γ) 6= 0, let γr ∈ γK0(C∗ r Γ) there exists b ∈ K0(C(∂Γ) ⋊ Γ) such that r Γ) be a reduced γ-element. Then for each a ∈ γK0(C∗ a = index(a)γr + i∗(b). Proof. The statement regarding ∗ = 1 and the one when the Euler characteristic is zero are both obvious from the Gysin sequence. For the second statement, let a ∈ γK0(C∗ r Γ), then since index(γr) = 1, a− index(a)γr has index zero and hence is in the kernel of the map Eul. Hence it is in the range of i∗, by the Gysin sequence. Thus a = index(a)γr + i∗(b) for b ∈ ran(i∗) as claimed. (cid:3) The results of this section show that, up to a cyclic summand, the K-homology of the reduced C-algebra of Γ comes entirely from restricting Γ-equivariant K-homology classes from the boundary. 32 HEATH EMERSON AND BOGDAN NICA 11. Uniformly summable K-cycles over the reduced group C-algebra Let Γ be a regular and torsion-free hyperbolic group. Recall that every class in K1(C(∂X) ⋊ Γ) can be represented by an odd Fredholm module of the form for some projection e ∈ C(∂X) ⋊ Γ, and every class in K1(C(∂X) ⋊ Γ) can be represented by a balanced even Fredholm module of the form µ (e)Pℓ2Γ(cid:1) (cid:0)λµ, Pℓ2Γλop µ (u)Pℓ2Γ + (1 − Pℓ2Γ)(cid:1) (cid:0)λµ, Pℓ2Γλop Φ(a) :=(cid:0)λ, Pℓ2Γλop µ (a)Pℓ2Γ(cid:1) r Γ → C(∂Γ) ⋊ Γ merely restricts the representation of C(∂Γ) ⋊ Γ to the subalgebra C∗ for some unitary u ∈ C(∂X) ⋊ Γ. At the level of cycles, the map i∗ on K-homology induced by i : C∗ r Γ. Thus we restrict the representation λµ to C∗ r Γ. Then, as each λµ(g) commutes with Pℓ2Γ, we can remove the degenerate summand (1 − Pℓ2Γ) · ℓ2(Γ, L2(∂X, µ)). Note that the restriction of λµ to the remaining summand Pℓ2Γ · ℓ2(Γ, L2(∂X, µ)) = ℓ2Γ is the regular representation λ. Thus, over C∗ r Γ the above Fredholm modules take the form where a is a projection or a unitary in C(∂X) ⋊ Γ. If a is a projection or a unitary in A, where A is as in Theorem 9.5, then Φ(a) is D> Lemma 11.1. Assume that Γ is regular and torsion-free, and let A be the smooth subalgebra of Theorem 9.5. Then every class in the image of the restriction map i∗ : K∗(C(∂X) ⋊ Γ) → K∗(C∗ r Γ) is represented by a Fredholm module of the form [Φ(a)] for some projection, respec- tively unitary a ∈ A. In particular, every class in i∗K∗(C∗ r Γ) has a representative which is D> ǫ -summable over CΓ. ǫ -summable over the group algebra CΓ. Combining Lemma 11.1 and Corollary 10.8, we obtain: Theorem 11.2. Assume that Γ is regular and torsion-free, and let A be the smooth subalgebra of Theorem 9.5. Then the following hold. • Every class in γK1(C∗ every class in γK1(C∗ CΓ. r Γ) is of the form [Φ(e)] for some projection e ∈ A. In particular, r Γ) is represented by a Fredholm module which is D> ǫ -summable over • If χ(Γ) = 0, then every class in γK0(C∗ In particular, every class in γK0(C∗ summable over CΓ. r Γ) is of the form [Φ(u)] for some unitary u ∈ A. r Γ) is represented by a Fredholm module which is D> ǫ - • If χ(Γ) 6= 0, and γr is a reduced γ-element, then every class in γK0(C∗ r Γ) is of the form kγr + [Φ(u)] for some integer k and some unitary u ∈ A. In particular, if γr is represented by a Fredholm module which is p(γr)-summable over CΓ, then every class in γK0(C∗ r Γ) is represented by a Fredholm module which is max{p(γr), D> We now specialize Theorem 11.2 to four families of a-T-menable groups; note that γ = 1 ǫ }-summable over CΓ. for a-T-menable groups by [18], so γK∗(C∗ r Γ) = K∗(C∗ r Γ). 11.1. Free groups. Let Γ be a finitely generated free group of rank at least 2. Given any p > 2, every class in i∗K∗(C(∂Γ) ⋊ Γ) has a p-summable representative over CΓ. On the other hand, the Julg - Valette model for the γ-element [21] is 1-summable over CΓ, hence the same holds true for the reduced γ-element γr. We conclude that C∗ r Γ has uniformly p-summable K-homology over CΓ for every p > 2. K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 33 If n is odd, then χ(Γ) = 0 so i∗K∗(C(∂Γ) ⋊ Γ) covers in fact all the K-homology of C∗ 11.2. Real uniform lattices. Let Γ be a torsion-free uniform lattice in SO(n, 1). Then classes in i∗K∗(C(∂Γ) ⋊ Γ) are (n− 1)+-summable over CΓ when n ≥ 4, respectively p-summable over CΓ for every p > 2, when n = 2, 3. r Γ. On the other hand, but still in this odd case, Kasparov shows in [25] that the γ-element for SO(n, 1) is represented by a Fredholm module in which the operator F is the phase of a degree 1 elliptic operator on the sphere Sn−1. Moreover, the unitary action of the group Γ commutes with F modulo pseudodifferential operators of order −1 because the action is conformal (and so the operators F and gF g−1 have the same symbol). Hence the commutators [g, F ] have singular values satisfying sk ≍ k−1/(n−1), that is, Kasparov's Fredholm module is (n − 1)+-summable over CΓ. Finally, Kasparov's equivariant Fredholm module supplies a reduced γ-element, that is, the group representations involved are weakly contained in the regular representation, since they factor through representations of C(S2n−1) ⋊ Γ, and Γ acts amenably on S2n−1. 0 (C, C) to an element of KKΓ If n is even, we can make small adjustments to this argument. The pull-back of the γ- element in KKSO(n+1,1) 0 (C, C) under the inclusion Γ ⊂ SO(n, 1) ⊂ SO(n + 1, 1) of Γ as a closed subgroup of SO(n + 1, 1), is the γ-element for Γ. By Kasparov's constructions described in the previous paragraph, we have therefore a description of the γ- element for Γ as a Fredholm module in which the group representations are weakly contained in the regular representation, and which is n+-summable over CΓ. We obtain therefore a reduced γ-element with the same properties. We conclude that the K-homology of C∗ r Γ is uniformly n+-summable over CΓ when n ≥ 3, respectively p-summable over CΓ for every p > 2, when n = 2. 11.3. Complex uniform lattices. Let Γ be a torsion-free uniform lattice in SU(n, 1). Then classes in i∗K∗(C(∂Γ)⋊Γ) are (2n)+-summable over CΓ. A model for the γ-element of SU(n, 1) has been given by Julg and Kasparov in [20]. In this case, the method involves construction of an appropriate hypoelliptic operator on the contact manifold S2n−1 (the contact structure is SU(n, 1)-invariant.) Inspection of the article [20] reveals that the relevant commutators [g, F ] are pseudodifferential operators in the class Ψ−1 H (S2n−1), and it is well-known that the singular values in this case satisfy sk ≍ k−1/(2n). We conclude that the K-homology of C∗ r Γ is uniformly (2n)+-summable over CΓ. 11.4. Small-cancellation groups. Let Γ be a torsion-free group given by a finite presentation hS Ri satisfying the C′(1/6) small-cancellation condition. As a geometric model for Γ we take the Cayley graph with respect to S, denoted Γ(S). Firstly, let us point out an explicit estimate for the visual dimension of the boundary of Γ(S). Combining Fact 5.4 and Fact 5.8, we get the coarse estimate visdim ∂Γ(S) ≤ 5δ eΓ(S) ≤ In a C′(1/6) situation, it 5δ log(2S − 1), where δ is the hyperbolicity constant of Γ(S). is possible to give a combinatorial estimate for δ, namely, δ ≤ 3 max{r : r ∈ R} by [15, Appendix, Thm.36]. We thus get the explicit, though far from optimal, bound visdim ∂Γ(S) ≤ 15 log(2S − 1) max{r : r ∈ R} =: κ(SR). Secondly, let us argue that Γ is regular, in the sense of Definition 8.1. As Γ is torsion-free, the 2-complex defined by the C′(1/6) presentation hSRi is aspherical. Hence χ(Γ) = 1−S+R, and Γ has cohomological dimension at most 2. If cd Γ = 1 then, by a well-known theorem of Stallings, Γ is a free group, and this is a case we have already discussed. So let us assume that cd Γ = 2. A theorem of Bestvina and Mess (see, e.g., [23, Thm.6.5]) implies that ∂Γ(S) has topological dimension 1. We then have the following fact, due to Misha Kapovich (personal communication): 34 HEATH EMERSON AND BOGDAN NICA Lemma 11.3. If Γ is a hyperbolic group whose boundary has topological dimension 1, then Γ is regular. Proof. A result of Bonk and Kleiner [5] says that the boundary of a hyperbolic group contains a quasi-circle provided that the group is not virtually free. In particular, ∂Γ contains a topological circle. The proof is completed by the following general claim: if Z is a d-dimensional compact space containing a d-dimensional topological sphere Sd, then Z admits a continuous self-map without fixed points. To prove the claim, recall the following alternative definition of topological dimension: a compact space X has dimension at most n if and only if every continuous map X0 → Sn, defined on a compact subset X0 ⊆ X, can be continuously extended to the entire X. Applying this fact to the space Z and the identity map Sd → Sd, we obtain a retraction ρ : Z → Sd. The composition τ ρ, where τ : Sd → Sd is the antipodal involution, is clearly fixed-point free. (cid:3) r Γ) is uniformly p-summable over CΓ for every p > κ(SR), and that the same is true for the even K-homology K0(C∗ Thus Γ meets the conditions of Theorem 1.6. It follows that the odd K-homology K1(C∗ r Γ) provided that χ(Γ) = 0, i.e., Γ has deficiency S − R = 1. References [1] S. Adams: Boundary amenability for word hyperbolic groups and an application to smooth dynamics of simple groups, Topology 33 (1994), no. 4, 765–783 [2] C. Anantharaman-Delaroche: Purely infinite C ∗-algebras arising from dynamical systems, Bull. Soc. Math. France 125 (1997), no. 2, 199–225 [3] C. Anantharaman-Delaroche: Amenability and exactness for dynamical systems and their C ∗-algebras, Trans. Amer. Math. Soc. 354 (2002), no. 10, 4153–4178 [4] U. Bader, A. Furman, T. Gelander, N. Monod: Property (T) and rigidity for actions on Banach spaces, Acta Math. 198 (2007), no. 1, 57–105 [5] M. Bonk, B. Kleiner: Quasi-hyperbolic planes in hyperbolic groups, Proc. Amer. Math. Soc. 133 (2005), no. 9, 2491–2494 [6] A. Connes: Compact metric spaces, Fredholm modules, and hyperfiniteness, Ergodic Theory Dynam. Systems 9 (1989), no. 2, 207–220 [7] A. Connes: Noncommutative Geometry, Academic Press 1994 [8] M. Coornaert: Mesures de Patterson-Sullivan sur le bord d'un espace hyperbolique au sens de Gromov, Pacific J. Math. 159 (1993), no. 2, 241–270 [9] J. Cuntz: K-theory for certain C ∗-algebras, Ann. of Math. (2) 113 (1981), no. 1, 181–197 [10] H. Emerson: Noncommutative Poincar´e duality for boundary actions of hyperbolic groups, J. Reine Angew. Math. 564 (2003), 1–33 [11] H. Emerson, R. Meyer: Euler characteristics and Gysin sequences for group actions on boundaries, Math. Ann. 334 (2006), no. 4, 853–904 [12] H. Emerson, R. Meyer: A descent principle for the Dirac-dual-Dirac method, Topology 46 (2007), 185–209 [13] H. Emerson, R. Meyer: Equivariant representable K-theory, Journal of Topology 46 (2009), 123–156 [14] H. Emerson, B. Nica: Finitely summable Fredholm modules for boundary actions of hyperbolic groups, Preprint 2012, arXiv:1208.0856v1 [15] ´E. Ghys, P. de la Harpe (eds.): Sur les groupes hyperboliques d'apr`es Mikhael Gromov, Progress in Math- ematics 83, Birkhauser 1990 [16] M. Goffeng, B. Mesland: Spectral triples and finite summability on Cuntz-Krieger algebras, Doc. Math. 20 (2015), 89–170 [17] M. Gromov: Hyperbolic groups, in Essays in group theory, 75–263, Publ. MSRI 8, Springer 1987 [18] N. Higson, G. Kasparov: E-theory and KK-theory for groups which act properly and isometrically on Hilbert space, Invent. Math. 144 (2001), no. 1, 23–74 [19] N. Higson, J. Roe: Analytic K-homology, Oxford Mathematical Monographs, Oxford University Press 2000 [20] P. Julg, G. Kasparov: Operator K-theory for the group SU(n,1), J. Reine Angew. Math. 463 (1995), 99–152 [21] P. Julg, A. Valette: K-theoretic amenability for SL2(Qp), and the action on the associated tree, J. Funct. Anal. 58 (1984), no. 2, 194–215 [22] V. Kaimanovich: Double ergodicity of the Poisson boundary and applications to bounded cohomology, Geom. Funct. Anal. 13 (2003), no. 4, 852–861 K-HOMOLOGICAL FINITENESS AND HYPERBOLIC GROUPS 35 [23] I. Kapovich, N. Benakli: Boundaries of hyperbolic groups, in Combinatorial and geometric group theory (New York, 2000/Hoboken, NJ, 2001), 39–93, Contemp. Math. 296, Amer. Math. Soc. 2002 [24] M. Kapovich, B. Kleiner: Hyperbolic groups with low-dimensional boundary, Ann. Sci. ´Ecole Norm. Sup. (4) 33 (2000), no. 5, 647–669 [25] G. Kasparov: Lorentz groups: K-theory of unitary representations and crossed products, Dokl. Akad. Nauk SSSR 275 (1984), 541–545 [26] G. Kasparov: Equivariant KK-theory and the Novikov conjecture, Invent. Math. 91 (1988), 147–201 [27] G. Kasparov, G. Skandalis: Groups acting properly on "bolic" spaces and the Novikov conjecture, Ann. of Math. (2) 158 (2003), no. 1, 165–206 [28] B. Kleiner: The asymptotic geometry of negatively curved spaces: uniformization, geometrization and rigidity, International Congress of Mathematicians. Vol. II, 743–768, Eur. Math. Soc. 2006 [29] V. Lafforgue: K-th´eorie bivariante pour les alg`ebres de Banach et conjecture de Baum-Connes, Invent. Math. 149 (2002), no. 1, 1–95 [30] M. Laca, J. Spielberg: Purely infinite C ∗-algebras from boundary actions of discrete groups, J. Reine Angew. Math. 480 (1996), 125–139 [31] J. Lott: Limit sets as examples in noncommutative geometry, K-Theory 34 (2005), no. 4, 283–326 [32] D. Meintrup, T. Schick: A model for the universal space for proper actions of a hyperbolic group, New York J. Math. 8 (2002), 1–7 [33] R. Meyer, R. Nest: The Baum–Connes conjecture via localisation of categories, Topology 45 (2006), no. 2, 209–259 [34] I. Mineyev, G. Yu: The Baum-Connes conjecture for hyperbolic groups, Invent. Math. 149 (2002), no. 1, 97–122 [35] B. Nica: Proper isometric actions of hyperbolic groups on Lp-spaces, Compos. Math. 149 (2013), no. 5, 773–792 [36] M. Puschnigg: Finitely summable Fredholm modules over higher rank groups and lattices, J. K-Theory 8 (2011), no. 2, 223–239 [37] S. Rave: On finitely summable K-homology, Dissertation, University of Munster 2012 [38] B. Simon: Trace ideals and their applications (Second edition), Mathematical Surveys and Monographs 120, American Mathematical Society 2005 [39] G. Skandalis: Une notion de nucl´earit´e en K-th´eorie (d'apr`es J. Cuntz), K-Theory 1 (1988), no. 6, 549–573 [40] J. Vaisala: Gromov hyperbolic spaces, Expo. Math. 23 (2005), no. 3, 187–231 [41] D.T. Wise: Cubulating small cancellation groups, Geom. Funct. Anal. 14 (2004), no.1, 150–214 [42] G. Yu: Hyperbolic groups admit proper affine isometric actions on lp-spaces, Geom. Funct. Anal. 15 (2005), no. 5, 1144–1151 (H.E.) Department of Mathematics and Statistics, University of Victoria, Victoria (Canada) E-mail address: [email protected] (B.N.) Mathematisches Institut, Georg-August Universitat Gottingen, Gottingen (Germany) E-mail address: [email protected]
1109.1860	1	1109	2011-09-08T23:53:17	Real interpolation between row and column spaces	[ "math.OA", "math.FA" ]	We give an equivalent expression for the $K$-functional associated to the pair of operator spaces $(R,C)$ formed by the rows and columns respectively. This yields a description of the real interpolation spaces for the pair $(M_n(R), M_n(C))$ (uniformly over $n$). More generally, the same result is valid when $M_n$ (or $B(\ell_2)$) is replaced by any semi-finite von Neumann algebra. We prove a version of the non-commutative Khintchine inequalities (originally due to Lust--Piquard) that is valid for the Lorentz spaces $L_{p,q}(\tau)$ associated to a non-commutative measure $\tau$, simultaneously for the whole range $1\le p,q< \infty$, regardless whether $p<2 $ or $p>2$. Actually, the main novelty is the case $p=2,q\not=2$. We also prove a certain simultaneous decomposition property for the operator norm and the Hilbert-Schmidt one.	math.OA	math	Real interpolation between row and column spaces by Gilles Pisier Texas A&M University College Station, TX 77843, U. S. A. and Universit´e Paris VI Equipe d'Analyse, Case 186, 75252 Paris Cedex 05, France June 11, 2018 Abstract We give an equivalent expression for the K-functional associated to the pair of operator spaces (R, C) formed by the rows and columns respectively. This yields a description of the real interpolation spaces for the pair (Mn(R), Mn(C)) (uniformly over n). More generally, the same result is valid when Mn (or B(ℓ2)) is replaced by any semi-finite von Neumann algebra. We prove a version of the non-commutative Khintchine inequalities (originally due to Lust -- Piquard) that is valid for the Lorentz spaces Lp,q(τ ) associated to a non-commutative measure τ , simultaneously for the whole range 1 ≤ p, q < ∞, regardless whether p < 2 or p > 2. Actually, the main novelty is the case p = 2, q 6= 2. We also prove a certain simultaneous decomposition property for the operator norm and the Hilbert-Schmidt one. 1 Introduction Let B(ℓ2) denote the space of all bounded operators on ℓ2. Let R ⊂ B(ℓ2) (resp. C ⊂ B(ℓ2)) be the row (resp. column) operator spaces, defined by R = span[e1j j ≥ 1] (resp. C = span[ei1 i ≥ 1]). The couple (R, C) plays an important role in Operator space theory. In particular, it is known that the complex interpolation space (R, C)1/2 coincides with the (self-dual) operator space OH. See [24] for details. We refer to [30] for the real interpolation method in the operator space framework. In particular, Xu proved in [30] that (R, C)1/2,2 is completely isomorphic to OH. This paper studies three problems concerning real interpolation for several pairs of Banach spaces associated to (R, C). 1 1 0 2 p e S 8 ] . A O h t a m [ 1 v 0 6 8 1 . 9 0 1 1 : v i X r a In §3, we consider the pair (M (R), M (C)) when M = B(ℓ2). The space M (R) consists of n converges in the weak operator topology (w.o.t. n)1/2k. Then M (C) is formed of those x = (xn) such that those x = (xn) with xn ∈ B(ℓ2) such thatP xnx∗ def in short) and kxkM (R) (x∗ n) ∈ M (R) with norm kxkM (C) = k(P xnx∗ def = k(P x∗ nxn)1/2k. The main result of §3 is an equivalent expression for the K-functional for this pair (M (R), M (C)). Our result extends to more general (semi-finite) von Neumann algebras. As an application we can describe the interpolation space X(θ) = (M (R), M (C))θ,∞ for 0 < θ < 1. We find that if x is in the latter space, then kxk2 X(θ) is equivalent to the norm of the associated completely positive map n as an operator of "very weak type (p, p)" on the Lp-space associated to the Tx : T 7→ P xnT x∗ 1 trace of M with p = 1/θ. The analogous result for the complex interpolation method was obtained in our previous works (see [20, 21]). Our result can be interpreted as a description of the operator space structure of (R, C)θ,∞ in the sense of [30]. Our approach is based on a non-commutative ver- sion of a lemma originally due to Varopoulos, that we extended with a different proof in a separate paper [25]. In §4, we present a version of the non-commutative Khintchine inequalities (originally due to Lust -- Piquard [14]) that is valid for the Lorentz spaces Lp,q(τ ) associated to (M, τ ). This provides an equivalent for the average over all signs of the norm in Lp,q(τ ) of a series of the form P ±xn (xn ∈ Lp,q(τ )). The main interest of our result is the case of L2,q(τ ) which seemed out of reach of previous works (see [9]). Here again our study concentrates on a pair of Banach spaces, but this time it is the pair (A0, A1) where A0 = M (R) ∩ M (C) and where A1 is the natural predual of M (R) ∩ M (C), that we describe as the sum of the preduals of M (R) and M (C) and we denote it by A1 = M∗(R) + M∗(C). In §5, we study another pair, namely the pair (A0, A2) where A2 = (A0, A1)1/2,2. When M = B(ℓ2), the space A2 is nothing but ℓ2(S2) where S2 is the Hilbert -- Schmidt class. We formulate our result using the notions of "K-closed" and "J-closed" introduced in [18], that are isomorphic versions of Peetre's notion of "subcouple". To give a more concrete statement, the following can be viewed as the main point of §5: There is a constant c such that for any x in (M (R) + M (C)) ∩ ℓ2(S2) there is a decomposition x = x1 + x2 such that we have simultaneously (1.1) (1.2) kx1kM (R) + kx2kM (C) ≤ ckxkM (R)+M (C) kx1kℓ2(S2) + kx2kℓ2(S2) ≤ ckxkℓ2(S2). In our more abstract terminology, this says that the pair (A0, A2) is J-closed when viewed as sitting inside (via the diagonal embedding) the pair (M (R) ⊕ M (C), ℓ2(S2) ⊕ ℓ2(S2)). This is [18]) for the pair (H ∞, H 2) inside (L∞, L2) (resp. analogous to what was proved in [11] (resp. (H ∞(B(ℓ2)), H 2(S2)) inside (L∞(B(ℓ2)), L2(S2)) In §6, we briefly include a comparative discussion of what the non-commutative Khintchine inequalities become in free probability and how it relates to our interpolation problems. Acknowledgment. I am grateful to Quanhua Xu and Yanqi Qiu for stimulating discussions. 2 Notation and background We will use the real interpolation method. We refer to [1] for all undefined terms. We just recall that if (A0, A1) is a compatible couple of Banach spaces, then for any x ∈ A0 + A1 the K-functional is defined by ∀t > 0 Kt(x; A0, A1) = inf(cid:0)kx0kA0 + tkx1kA1 x = x0 + x1, x0 ∈ A0, x1 ∈ A1). Recall that the ("real" or "Lions-Peetre" interpolation) space (A0, A1)θ,q is defined, for 0 < θ < 1 and 1 ≤ q ≤ ∞, as the space of all x in A0 + A1 such that kxkθ,q < ∞ where kxkθ,q = (Z (t−θKt(x, A0, A1))qdt/t)1/q, with the usual convention when q = ∞. 2 Let M∗ be the predual of M . As is well known (see e.g. Let M be a von Neumann algebra equipped with a semi-finite faithful normal trace τ . The basic example is M = B(ℓ2), equipped with the usual trace, and, to improve readability, we will present most of our results in this special case with mere indications for the extension to the general case. [26]) M∗ can be identified with the non-commutative L1-space associated to τ usually denoted by L1(τ ). When M = B(ℓ2), M∗ is the classical "trace class" S1. More generally, for any 1 ≤ p < ∞ we denote by Lp(τ ) the associated non-commutative Lp-space. By convention we set L∞(τ ) = M . When M = B(ℓ2), M∗ (resp. Lp(τ )) is the classical "trace class" S1 (resp. the Schatten class Sp). See e.g. [4] or [26] for more information on non-commutative Lp-spaces. We will first mainly use the case p = 2 and we denote its norm simply by k.k2. We always denote by p′ the conjugate of 1 ≤ p ≤ ∞ defined by p−1 + p′−1 = 1. We denote by P(M ) or simply by P the set of all (self-adjoint) projections in M . We denote by M (R) (resp. M (C)) the space of sequences x = (xn) with coefficients xn ∈ M nxn)1/2 ∈ M ). Here we implicitly assume that the nxn) converge in (say) the w.o.t. . We equip M (R) (resp. M (C)) with 1 x∗ such that (P∞ seriesP∞ the natural norm 1 xnx∗ 1 x∗ 1 xnx∗ n)1/2 ∈ M (resp. (P∞ n (resp.P∞ kxkM (R) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X xnx∗ n(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)M(cid:18)resp. kxkM (C) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X x∗ nxn(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)M(cid:19) . Note that M (R) (resp. M (C)) can be identified with M ⊗R (resp. M ⊗C), i.e. the weak-∗ closure of M ⊗ R (resp. M ⊗ C) in the (von Neumann sense) tensor product M ⊗B(ℓ2). We will consider (M (R), M (C)) as an interpolation couple in the obvious way using the inclu- sions M (R) ⊂ M N, M (C) ⊂ M N. (2.1) Similarly, we denote by M∗(R) (resp. M∗(C)) the space of sequences x = (xn) with coefficients nxn)1/2 ∈ M∗). Here we assume that the nxn)1/2) norm-converges in M∗ when N → ∞. We equip 1 xnx∗ 1 x∗ M∗(R) (resp. M∗(C)) with the natural norm xn ∈ M∗ such that (P∞ sequence (PN 1 xnx∗ 1 x∗ n)1/2 (resp. (PN kxkM∗(R) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X xnx∗ n)1/2 ∈ M∗ (resp. (P∞ n(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)M∗ resp. kxkM∗(C) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X x∗ Note that M (R) = M∗(C)∗ and M (C) = M∗(R)∗ isometrically with respect to the duality defined by hx, yi =P τ (xnyn). (Equivalently, M (R) = M∗(R)∗ and M (C) = M∗(C)∗ with respect to the duality defined by hx, yi =P τ (xny∗ n), with the bar denoting complex conjugation.) nxn(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)M∗! . 3 K-functional between R and C Our main result is: Theorem 3.1. For any a = (an) ∈ M (R) + M (C) we have ∀t > 0 where kt(a) ≤ Kt(a; M (R), M (C)) ≤ 2kt(a) kt(a) = sup(cid:26)(cid:16)X kP anQk2 2(cid:17)1/2 max{τ (P )1/2, t−1τ (Q)1/2}−1 (cid:12)(cid:12)(cid:12) P, Q ∈ P(cid:27) . 3 Remark. The following was pointed out by Q. Xu (by a modification of the proof of Lemma 3.2 below). Let bkt(a) be the same as kt(a) except that, when t > 1 (resp. t < 1) we restrict to pairs of projections P, Q such that P ≤ Q (resp. Q ≤ P ), and when t = 1 we restrict to pairs such that P = Q. Then We merely indicate a quick argument for t > 1. Let Q′ = P ∨ Q. Then P ≤ Q′ and τ (Q′) ≤ bkt(a) ≤ kt(a) ≤ 21/2bkt(a). τ (P ) + τ (Q). With the above notation we have (cid:0)P kP anQk2 since (t−2 + 1)1/2 ≤ 21/2 we obtain kt(a) ≤ 21/2bkt(a). We leave the other cases to the reader. First part of the proof of Theorem 3.1. Consider a0 ∈ M (R), P, Q ∈ P. We have 2(cid:1)1/2 ≤ (cid:0)P kP anQ′k2 τ (P )1/2 ∨ t−1τ (Q′)1/2 ≤ τ (P )1/2 ∨ t−1(τ (P ) + τ (Q))1/2 ≤ (t−2 + 1)1/2(τ (P )1/2 ∨ t−1τ (Q)1/2 and 2(cid:1)1/2 and also X kP a0 nQk2 2 =X τ (P a0 ≤(cid:13)(cid:13)(cid:13)X a0 nQa0∗ n P ) ≤X τ (P a0 n(cid:13)(cid:13)(cid:13) τ (P ) na0∗ na0∗ n P ) and hence Similarly for any a1 ∈ M (C) we have nQk2 (cid:16)X kP a0 2(cid:17)1/2 (cid:16)X kP a1 2(cid:17)1/2 (cid:16)X kP anQk2 Therefore if a = a0 + a1 we find by the triangle inequality ≤ ka0kM (R)τ (P )1/2. nQk2 ≤ ka1kM (C)τ (Q)1/2. 2(cid:17)1/2 ≤ ka0kM (R)τ (P )1/2 + ka1kM (C)τ (Q)1/2 ≤ (ka0kM (R) + tka1kM (C)) max{τ (P )1/2, t−1τ (Q)1/2}. So we obtain kt(a) ≤ Kt(a; M (R), M (C)). To prove the converse we will use duality, via the following Lemma. Lemma 3.2. Let x ∈ M∗(R) ∩ M∗(C) be such that Jt(x) def= max(cid:26)(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X x∗ nxn(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)1 , 1 t(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X xnx∗ n(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)1(cid:27) ≤ 1. Let Ct be the subset of M∗(R) ∩ M∗(C) formed of all sequences χ = (χn) of the form χn = QynP · (τ (P ) + t−2τ (Q))−1/2 with P kynk2 closure is in M∗(R) ∩ M∗(C). 2 ≤ 1 and where P, Q are commuting projections. Then x ∈ 2 conv(Ct) where the We will need the following simple 4 Lemma 3.3. Let ϕ : [1, . . . , N ]2 → R+ be defined by ϕ(i, j) = g(i) ∧ f (j) where g ≥ 0 and f ≥ 0 satisfy P f (j) ≤ 1 and P g(i) ≤ t2. Then ϕ ∈ conv(Φ) where Φ is the set of functions on [1, . . . , N ]2 of the form (3.1) 1E×F 1 t2 E + F where E, F ⊂ [1, . . . , N ]2 are arbitrary subsets. Proof. We may write ϕ =Z ∞ =Z ∞ 0 0 1{ϕ>c}dc =Z ∞ 0 1{g>c}×{f >c} m(c) dc m(c) 1{g>c}×{f >c}dc where m(c) = 1 follows. t2 {g > c} + {f > c}. But sinceR ∞ 0 m(c)dc = 1 t2P g(i) +P f (j) ≤ 2 the Lemma Remark. A simple verification shows that if ϕ ∈ Φ is of the form (3.1), we have supj ϕ(i, j) = 1{i∈E}(t−2E + F )−1 and supi ϕ(i, j) = 1{j∈F }(t−2E + F )−1. Therefore we find (3.2) t−2Xi supj ϕ(i, j) +Xj supi ϕ(i, j) ≤ 1. Remark. Let (Ω, µ) and (Ω′, µ′) be measure spaces. Let f : Ω′ → R+ and g : Ω → R+ be step functions. Assume thatR f dµ′ ≤ 1 andR g dµ ≤ t2 (t > 0). Let ϕ(ω, ω′) = g(ω) ∧ f (ω′) on Ω × Ω′. Then ϕ ∈ 2conv(Φ) where Φ is the set of functions of the form where E ⊂ Ω and F ⊂ Ω′ are arbitrary measurable subsets. ϕ = 1E×F t−2µ(E) + µ′(F ) , Proof of Lemma 3.2. As is well known, if we truncate the sequence (xn) and replace it by x(N ) defined by xn(N ) = xn1{n≤N }, then kx(N ) − xkM∗(R) → 0 and similarly for M∗(C). Indeed, since for all N ≤ m the norms in M∗(R) and M∗(R) both satisfy kx(N ) − x(m)k2 + kx(N )k2 ≤ kx(m)k2 and since kx(m)k → kxk, this fact follows easily. Thus it suffices to prove the Lemma for finite sequences x = (x1, . . . , xN ). Let us first assume that M = B(ℓ2), M∗ = S1 (trace class) and τ is the ordinary trace on S1. Assume Jt(x) < 1. Let n)1/2. We have tr(f ) < 1, tr(g) < t2 and moreover if an = g−1xn, f = (P x∗ bn = xnf −1 thenP ana∗ nxn)1/2 and g = t(P xnx∗ n = g−1t−2g2g−1 = t−2, and similarlyP b∗ (cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X ana∗ nbn(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)X b∗ n(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13) ≤ t−1, with xn = gan = bnf . (3.3) 1/2 ≤ 1, nbn = 1, so we have Note that by a simple perturbation argument we may assume f > 0 and g > 0 so that f and g are invertible. 5 Let us now consider the matrix representation of xn with respect to the bases that diagonalize respectively f (for the column index) and g (for the row index). We have then Moreover we know from (3.3) that ∀ξ ∈ ℓ2 xn(i, j) = gian(i, j) = bn(i, j)fj . and hence taking for ξ either the i-th or the j-th basis vector we find nξk2 ≤ t−2kξk2 and X kbnξk2 ≤ kξk2 (3.4) an(i, j)2 ≤ t−2 and bn(i, j)2 ≤ 1. X ka∗ i XnXj sup sup j XnXi Let γn(i, j) = 1 (xn(i, j)). Then γn(i, j) ≤ max{an(i, j), bn(i, j)} ≤ an(i, j) + bn(i, j) and xn(i, j) = (gi ∧ fj)γn(i, j). By Lemma 3.3, since xn(i, j) = gi ∧ fj γn(i, j), we know that x/2 is in the convex hull of gi∧fj where ϕ ∈ Φ. We then note that if (ϕ(i, j)γn(i, j))n = (ϕ(i, j)1/2ϕ(i, j)1/2γn(i, j))n yn(i, j) = ϕ(i, j)1/2γn(i, j) then using (3.4) and (3.2) we have since γn(i, j) ≤ an(i, j) ∨ bn(i, j) Xn kynk2 ξn(i, j)2 2 =Xn,i,j ≤XnXij ≤Xi ≤ 1. ϕ(i, j)(an(i, j)2 + bn(i, j)2) supj ϕ(i, j)t−2 +Xj supi ϕ(i, j) Thus we find that x/2 can be written as a convex combination of elements of the form (ϕ(i, j)1/2yn)n with ϕ ∈ Φ andP kynk2 2 ≤ 1. Let Q, P be the projections associated to 1E and 1F . Then [ϕ(i, j)1/2yn] = (t−2 tr Q + tr P )−1/2QynP. This completes the proof in the case M = B(H). The case of a general semi-finite von Neumann algebra M ⊂ B(H) can easily be reduced (by density) to the case when M is finite. In that case, the densities f and g in the preceding argument can be replaced by fε = f + ε1 and gε = g + ε1 in order to obtain f, g invertible. We are thus left with a finite sequence x1, . . . , xN in M and f, g ≥ 0 in M∗ invertible such that τ (f ) < 1 and τ (g) < t2 and moreover (3.5) xn = gan = bnf (n ≤ N ) with an, bn ∈ M such that (3.3) holds. Just like we do for functions and step functions we may 1 giQi where (Pi) (resp. Qj)) are orthogonal approximate f, g by elements of the formPk projections in M withP Pi =P Qj = 1. This modification leads by (3.5) to a perturbation of xn the above Lemma with the measures µ =P gjτ (Qj)δj and µ′ =P fiτ (Pi)δi. so it suffices to complete the proof for this special case. If we then denote xn(i, j) = PixnQj, and similarly for an and bn we can essentially repeat the preceding argument using the remark following 1 fjPj andPk 6 End of the proof of Theorem 3.1. Assume kt(a) ≤ 1, so that ∀P, Q ∈ P This implies by Cauchy -- Schwarz X kP anQk2 2 ≤ τ (P ) ∨ t−2τ (Q). (cid:12)(cid:12)(cid:12)X τ (P anQyn)(cid:12)(cid:12)(cid:12) ≤ (τ (P ) ∨ t−2τ (Q))1/2(cid:16)X kynk2 2(cid:17)1/2 2(cid:17)1/2 ≤ (t−2τ (Q) + τ (P ))1/2(cid:16)X kynk2 (cid:12)(cid:12)(cid:12)X τ (anxn)(cid:12)(cid:12)(cid:12) ≤ 2Jt(x), Kt(a; M (R), M (C)) ≤ 2. and hence by Lemma 3.2 then by duality we conclude Remark. The preceding proof reveals the following slightly surprising fact: Consider a sequence x = (xn) of operators xn ∈ B(ℓ2), with say H = ℓ2. Assume that for any pair of orthonormal basis (ei) (fj) in H the matrix aij = hei, xnfji belongs to ℓ∞(i; ℓ2(n, j)) + ℓ∞(j; ℓ2(n, i)). Then x = (xn) ∈ M (R) + M (C) with M = B(ℓ2). Indeed, if E (resp. F ) is a finite subset of N, and if P (resp. Q) is the orthogonal projection into span(E) (resp. span(F )), then (cid:16)X kP anQk2(cid:17)1/2 (max(τ (P )1/2, τ (Q)1/2))−1 =(cid:16)XE×FX an(i, j)2 max{E, F }−1(cid:17)1/2 So Theorem 3.1 (compare with the Varopoulos Lemma in [28]) implies the preceding fact. Moreover the converse implication also holds since x ∈ M (R) (resp. M (C)) implies a ∈ ℓ∞(i; ℓ2(n, j)) (resp. ℓ∞(j; ℓ2(n, i)). Remark. Let N ⊂ M be a von Neumann subalgebra such that τN is still semi-finite, so that the conditional expectation E : M → N is well defined. It is easy to check that the main result remains valid if we replace the norms of M (R) (resp. M (C)) by their conditional versions: . 1/2 (cid:18)resp.(cid:13)(cid:13)(cid:13)X E(x∗ n)(cid:13)(cid:13)(cid:13) nxn)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)X E(xnx∗ 1/2(cid:19) . The formula for kt now has to be modified by restricting p, q to lie in N . To put the next corollary in proper perspective, recall that, according to [20], the elements a = (an) in the complex interpolation space (M (R), M (C))θ are precisely those such that the operator Ta : x 7→X anxa∗ n is bounded on Lp(τ ), where p = 1/θ. In the commutative case, a bounded operator T : Lp(Ω1, µ1) → Lp(Ω2, µ2) is called of strong type p, and the classical Riesz interpolation theorem says that, in the complex case, if 1 ≤ p0 < p1 ≤ ∞ and T is of strong type pj for j = 0, 1 then T is of strong type p for any intermediate p such that p0 < p < p1. The latter theorem is the founding result for the "complex interpolation method" (see [1]), while the classical Marcinkiewicz theorem is the basis for the "real method". In that context, an operator that is bounded from Lp,1(µ1) to Lp,∞(µ2) is called of very weak type p. 7 The generalized version of Marcinkiewicz theorem then says that, if T is of very weak type pj for j = 0, 1, then T is of strong type p for any p0 < p < p1. Let (Ω, µ) be a measure space. Recall that the "weak Lp" space Lp,∞(µ) is formed of all measurable functions f : Ω → R such that kf kp,∞ = sup c>0 (cpµ{f > c})1/p < ∞. When p > 1, the quasi-norm k · kp,∞ is equivalent to the following norm (3.6) kf k[p,∞] = sup(cid:26)ZE f dµ µ(E)1/p′ (cid:12)(cid:12)(cid:12) E ⊂ Ω(cid:27) . When p > 1, Lp′,∞(µ) is the dual of the "Lorentz space" Lp,1(µ) that can be defined (see e.g. [1]) as formed of those f such that (3.7) [f ]p,1 =Z ∞ 0 µ{f > c}1/p dc < ∞. Using the generalized s-numbers from [4], it is easy to define the spaces Lp,∞(τ ) and Lp,1(τ ) (or more generally Lp,q(τ ) for 1 ≤ q ≤ ∞) associated to (M, τ ). The simplest way to describe those is as follows. Given a τ -measurable operator x (in the sense of [4]), let Mx ⊂ M denote the von Neumann subalgebra generated by the spectral projections of x. Then Mx ≃ L∞(Ωx, µx) for some measure space (Ωx, µx) in such a way that the restriction of τ to Mx coincides with µx in this identification. Moreover the space of scalar valued measurable functions (that are bounded outside a set of finite measure) L0(Ωx, µx) can be identified with that of τ -measurable operators affiliated with Mx. The space Lp,∞(τ ) (resp. Lp,q(τ ) for 1 ≤ q ≤ ∞) is then formed of those x such that, in the latter identification, x ∈ Lp,∞(µx) (resp. Lp,q(µx). The duality extends to this setting: we have Lp′,∞(τ ) = Lp,1(τ )∗ for any 1 ≤ p < ∞, see [4, 26]. The following two statements appear as analogues for real interpolation of the complex case already treated in [22, 20]. Corollary 3.4. Let 0 < θ < 1. An element a = (an) (xn ∈ M ) belongs to the space (M (R), M (C))θ,∞ iff the mapping Tα : x →X anxa∗ n is bounded from Lp,1(τ ) to Lp,∞(τ ) where 1 (M (R), M (C))θ,∞ is equivalent to a 7→ kTa : Lp,1(τ ) → Lp,∞(τ )k1/2. p = 1−θ ∞ + θ 1 , i.e. p = 1/θ. Moreover, the norm in Proof. Recall (3.8) ∀a0, a1 > 0 a1−θ 0 aθ 1 = inf t>0 {(1 − θ)a0tθ + θa1tθ−1}. By definition kakθ,∞ = supt>0 t−θKt(a; M (R), M (C)). By Theorem 3.1 this is equivalent to supt>0 t−θkt(a). Note that (1 − θ)ξ + θη ≤ max(ξ, η) ≤ max{θ−1, (1 − θ)−1}((1 − θ)ξ + θη) ∀ξ, η > 0. Therefore is equivalent to supt>0 t−θ(max(τ (P )1/2, t−1τ (Q)1/2))−1 (inf t>0(1 − θ)t2θτ (P ) + θt2θ−2τ (Q))−1/2 8 or equivalently (by (3.8)) to inf s>0((1 − θ)sθτ (P ) + θsθ−1τ (Q))−1/2 = (τ (P )1−θτ (Q)θ)−1/2. Thus we find that kak2 θ,∞ is equivalent to supnX kP anQk2 = supnX τ (P anQa∗ 2(τ (P )1−θτ (Q)θ)−1 P, Q ∈ Po n)(τ (P )1−θτ (Q)θ)−1 P, Q ∈ Po = sup{hTa(Qτ (Q)−θ), P τ (P )−(1−θ)i P, Q ∈ P}. This last expression is equivalent to sup{hTa(x), yi x ∈ BLp,1(τ ), y ∈ BLp′,1(τ )}. Indeed, using convex combinations of elements of the form Qτ (Q)−θ we obtain the case of x ≥ 0 in BLp,1(τ ); then the decomposition x = x1 − x2 + i(x3 − x4) yields the general case up to a factor 4. The same reasoning applies to y. So we conclude that kak2 θ,∞ is equivalent to kTa : Lp,1(τ ) → Lp,∞(τ )k. Remark. In the particular when M = B(ℓ2) or Mn with n arbitrary, the preceding corollary yields a description of the operator space structure of (R, C)θ,∞ according to Xu's definition in [30]. Let 1 < p ≤ ∞. When p = ∞, by convention we identify Lp,∞(τ ) with M . Let M (R; p, ∞) n)1/2 ∈ Lp,∞(τ ), n)1/2kp,∞. Similarly, we define M (C; p, ∞) = {a = (an) denote the space of sequences a = (an) with an ∈ Lp,∞(τ ) such that (P ana∗ equipped with the "norm" kak = k(P ana∗ n) ∈ M (R; p, ∞)} with kakM (C;p,∞) = k(P a∗ Using a simple non-commutative adaptation of the results in [25] along the lines of the proof of nan)1/2kp,∞. (a∗ Theorem 3.1, we find Theorem 3.5. Let 2 < p0, p1 ≤ ∞. Let a = (an) be a sequence with an ∈ Lp0,∞(τ ) + Lp1,∞(τ ) for all n. To abbreviate, we set Kt(a) = Kt(a; M (R; p0, ∞), M (C; p1, ∞)), and kt(a) = sup(cid:26)(cid:16)X kP anQk2 2(cid:17)1/2 max{τ (P )α0 , t−1τ (Q)α1 }−1 P, Q ∈ P(cid:27) where α0 = 1 p0, p1) such that 2 − 1 p0 , α1 = 1 2 − 1 p1 . Then there are positive constants c and C (depending only on ∀t > 0 ckt(a) ≤ Kt(a) ≤ Ckt(a). Proof. Let (Ω, µ) be any measure space. For any measurable f : Ω → R+ we have obviously kf kp p/2,∞. Therefore, using (3.6) kf kp,∞ is equivalent to p,∞ = kf 2kp/2 Thus we may use on M (R; p0, ∞) the following equivalent norm sup E∈Ω (ZE sup(cid:26)(cid:16)τ(cid:16)X xnx∗ f 2 dµ)1/2µ(E) 1 p − 1 2 . nP(cid:17)(cid:17)1/2 τ (P )−α0 P ∈ P(cid:27) . 9 Similarly, we may equip M (C; p1, ∞) with the equivalent norm sup(cid:26)(cid:16)τ(cid:16)X x∗ nxnQ(cid:17)(cid:17)1/2 τ (Q)−α1 Q ∈ P(cid:27) . Using these norms we find kt(a) ≤ Kt(a) by the same reasoning as for Theorem 3.1. To prove the converse, we use duality again and mimic the proof of Theorem 3.1 using as a model the results presented in [25] for the commutative case. + θ p1 = 1−θ p0 Corollary 3.6. Consider 2 < p0, p1 ≤ ∞ and 0 < θ < 1. Let a = (an)n be a sequence in Lpθ,∞(τ ) where 1 . Then a = (an) belongs to the space (M (R; p0, ∞), M (C; p1, ∞))θ,∞ iff the pθ operator Ta is bounded from Lr,1(τ ) to Ls,∞(τ ) where r, s are determined by 1−θ r (i.e. s (i.e. s = p0(1 − θ + θp0)−1). Moreover the norm of a in that space is r = p′ equivalent to kTa : Lr,1(τ ) → Ls,∞(τ )k1/2. 1/θ) and 1−θ p0 ∞ + θ p′ 1 + θ 1 = 1 = 1 Proof. Using the equivalent of Kt found in Theorem 3.5, we obtain sup t−θKt(a) ≃ sup(cid:26)(cid:16)X kP anQk2 2(cid:17)1/2 τ (P )−α0(1−θ)τ (Q)−α1θ(cid:27) . The unit ball for this last norm is characterized by X τ (P anQa∗ n) ≤ τ (P )2α0(1−θ)τ (Q)2α1θ 0 τ (Q)θ/p′ ≤ τ (P )(1−θ)/p′ 1 or equivalently hTa(Q), P i ≤ τ (Q)1/rτ (P )1/s′ . As before, this implies for all x, y ≥ 0 hTa(x), yi ≤ kxkr,1kyks′,1 and then using x = x1 − x2 + i(x3 − x4) we can extend it to arbitrary elements up to an extra factor 4. Thus we conclude by homogeneity supt>0 t−θKt(a) ≃ kTa : Lr,1(τ ) → Ls,∞(τ )k1/2. Remark 3.7. By [16] for any interpolation pair (B0, B1) we have (B0 ∩ B1, B0 + B1)θ,q =(Bθ,q ∩ B1−θ,q Bθ,q + B1−θ,q if if θ ≤ 1/2 θ ≥ 1/2 where Bθ,q = (B0, B1)θ,q. If we apply this to the specific pair B0 = S∞[R], B1 = S∞[C], the result can be interpreted in terms of operator space interpolation in Xu's sense (see [30]). This gives us that we have completely isomorphically (R ∩ C, R + C)θ,∞ =(Cθ,∞ ∩ Rθ,∞ θ ≤ 1/2 Cθ,∞ + Rθ,∞ θ ≥ 1/2 where Cθ,q = (R, C)θ,q and Rθ,q = (C, R)θ,q = C1−θ,q. Note that in particular, we have as operator spaces: and similarly, by duality, (see [30, §4]) for (1/2, 1). (R ∩ C, R + C)1/2,∞ ≃ R1/2,∞ ≃ C1/2,∞ 10 4 Non-commutative Khintchine inequality in L2,q (1 ≤ q < ∞) This section is motivated by [9]. In [9], martingale inequalities extending those of [26] are proved for the non-commutative Lorentz spaces Lp,q(τ ) associated to a semi-finite trace with p 6= 2. However, the case p = 2, q 6= 2 cannot be treated by the interpolation arguments used in [9]. In fact, even the simpler case of the Khintchine inequality is open. The problem is to find a "nice" (similarly nice as in the case of Lp(τ ) presumably involving row and column norms) equivalent of the average over all signs εn = ±1 of (4.1) (cid:13)(cid:13)(cid:13)X εnxn(cid:13)(cid:13)(cid:13)2,q when xn ∈ L2,q(τ ). In this section we present a partial solution, which has the advantage to be indeed a deterministic equivalent of (4.1). We call it partial because there may be a more explicitly computable equivalent for (4.1). Notation: Recall that Sp denotes the Schatten p-class (1 ≤ p < ∞), S∞ the space of compact op- erators on Hilbert space, and S1 the trace class. We will denote by S∞(R) (resp. S∞(C)) the space nxn) converges 1 xnx∗ 1 x∗ of all sequences x = (xn) with xn ∈ S∞ such that the seriesP∞ in norm, and we equip it with the norm kxkR =(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)P xjx∗ A0 = S∞(R) ∩ S∞(C) We then define n (resp.P∞ j(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13) (resp. kxkC =(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)P x∗ j xj(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)). and we equip it with the norm kxkA0 = max{kxkR, kxkC }. We denote by S1(R) (resp. S1(C)) the space that was already introduced as M∗(R) (resp. M∗(C)) when M∗ = S1, M = B(ℓ2), see (2.1). We then define A1 = S1(R) + S1(C) and we equip it with the norm kxkA1 = kxkS1(R)+S1(C) = infx=y+z kykS1(R) + kzkS1(C). Clearly, the couple (A0, A1) forms a compatible couple in the sense of interpolation. We denote for any 1 < p < ∞, 1 ≤ q ≤ ∞ xp,q = kxk(A0,A1)θ,q where θ = 1 − θ ∞ + θ 1 = 1 p . Note that A1 = A∗ tyn)) and by a well known result (see [2] and references there) this implies (A0, A1)1/2,2 = ℓ2(S2) isomorphically. Moreover, the couple (A∗ 1) can be clearly viewed as compatible and we have (see [1, p. 54]) for 0 < θ < 1 and 1 ≤ q < ∞ 0 isometrically (in the usual duality defined by hx, yi =P tr(xn 0, A∗ (4.2) with equivalent norms. (A0, A1)∗ θ,q = (A∗ 0, A∗ 1)θ,q′ Theorem 4.1. Let (εn) be the usual independent ±1 valued random variable ("Rademacher func- tions"). Then for any 1 < p < ∞, 1 ≤ q < ∞ and for any finite sequence x = (xn)n of operators in Sp,q (4.3) dµ ≃ xp,q Z (cid:13)(cid:13)(cid:13)X εnxn(cid:13)(cid:13)(cid:13)Sp,q 11 and also (4.4) (cid:13)(cid:13)(cid:13)X εn ⊗ xn(cid:13)(cid:13)(cid:13)Lp,q(µ×tr) ≃ xp,q where µ is the usual probability on {−1, 1}N. Moreover, (4.4) remains valid for q = ∞. Here A ≃ B means there are positive constants cp,q and Cp,q such that cp,qA ≤ B ≤ Cp,qA. Remark. It is not difficult to extend this Theorem to the case when the trace on B(ℓ2)) is replaced by any semifinite faithful normal trace on a von Neumann algebra, but we choose for simplicity to present the details only in the case of B(ℓ2). Indeed, all the ingredients for this extension now exist in the literature (see [26]). Remark 4.2. Note that the space M (R) ∩ M (C) with M = B(ℓ2) considered in §3 is nothing but the bidual A∗∗ 0 of A0. Using this (and truncation of matrices in the most usual way), one can check that, for any x ∈ A0 + A1 and any t > 0, we have Kt(x; A∗∗ 0 , A1) = Kt(x; A0, A1) and hence the norms of the spaces (A0, A1)θ,q and (A∗∗ 0 , A1)θ,q coincide on any such x for any 0 < θ < 1 and 1 ≤ q ≤ ∞. Remark 4.3. Note that by interpolation for any 1 < p < ∞ (and 1 ≤ q ≤ ∞) the orthogonal projection onto span[εn] is bounded on Lp,q(µ × tr). Indeed, we know that it is bounded on Lp0(µ × tr), Lp1(µ × tr) for 1 < p0 < p < p1, and then we can use interpolation, together with the classical reiteration theorem (see [1, p. 48]). Remark 4.4. The equivalence between [εn] and Sidon lacunary sequences such as [z2n [17] implies that for any 1 < p < ∞ and 1 ≤ q ≤ ∞ ] proved in (cid:13)(cid:13)(cid:13)X εn ⊗ xn(cid:13)(cid:13)(cid:13)Lp,q(µ×tr) ≃(cid:13)(cid:13)(cid:13)X z2n ⊗ xn(cid:13)(cid:13)(cid:13)Lp,q(m×tr) where m is normalized Haar measure on T. Again this is true by [17] on Lpi, p0 < p < p1 (with simultaneous complementation) so this follows by interpolation. Remark 4.5. Let E be any Banach space. When 1 ≤ p < ∞, we denote by Lp(E) the space of E-valued functions p-integrable on the unit circle, in Bochner's sense, with the usual norm. We denote by H p(E) the subspace of all functions f with Fourier transform f supported on the non-negative integers. The case p = ∞ is slightly different. We denote L∞(B(ℓ2)) the space of essentially bounded weak-∗ measurable functions on the unit circle with values in B(ℓ2)), equipped ∞ with the sup-norm. We again denote by H ∞(B(ℓ2)) (resp. H 0 (B(ℓ2)) ) the subspace formed of all functions with Fourier transform supported on the non-negative (resp. negative) integers. By [18, Cor. 3.4] we know that the pair (H ∞(B(ℓ2)), H 1(S1)) is K-closed in (L∞(B(ℓ2)), L1(S1)). From this it follows that: (cid:13)(cid:13)(cid:13)X∞ 1 z2n xn(cid:13)(cid:13)(cid:13)(H∞(B(ℓ2)),H 1(S1))θ,q ≃(cid:13)(cid:13)(cid:13)X z2n ≃(cid:13)(cid:13)(cid:13)X z2n xn(cid:13)(cid:13)(cid:13)(L∞(B(ℓ2)),H 1(S1))θ,q ⊗ xn(cid:13)(cid:13)(cid:13)Lp,q(dm×tr) . Remark 4.6. From [15] we know that the mapping T : H 1(S1) → A1 defined by T f = ( f (2n))n is a bounded surjection from H 1(S1) to A1 = S1(R) + S1(C). Moreover, the (adjoint) map taking xn is bounded from A1 = S1(R)+S1(C) to H 1(S1). Using the identity H 1(S1)∗ = x = (xn) toP z2n L∞(B(ℓ2))/H ∞ 0 (B(ℓ2)), by duality, this implies that T is bounded from L∞(B(ℓ2))/H ∞ 0 (B(ℓ2)) = H1(S1)∗ onto A∗ 1. 12 By interpolating, we find T : (L∞(B(ℓ2))/H ∞ 0 (B(ℓ2)), H 1(S1))θ,q → (A∗ 1, A1)θ,q. Note that the natural "inclusion" H ∞(B(ℓ2)) → L∞/H ∞ 0 (B(ℓ2)) trivially has norm ≤ 1. Therefore: T : (H ∞(B(ℓ2)), H 1(S1))θ,q → (A∗ 1, A1)θ,q. 1 = A∗∗ Note that A∗ x = (xn) with xn ∈ Sp,q ⊂ S∞, the norms of (A∗ 1 < p < ∞, 1 ≤ q ≤ ∞ (here θ = 1/p). Thus, invoking again [18, Cor. 3.4], by the preceding two remarks, we find 0 . Therefore, using Remark 4.2, it is easy to check that any finite sequence 1, A1)θ,q and (A0, A1)θ,q coincide on x for any Lemma 4.7. Let 1 < p < ∞, 1 ≤ q ≤ ∞. Then T is bounded from the subspace of all analytic polynomials in Lp,q(dm × tr) into (A0, A1)θ,q. In particular, for some c, for all analytic polynomial f with coefficients in Sp,q = Lp,q(tr) ( f (2n))np,q ≤ ckf kLp,q(dm×tr). First part of the proof of Theorem 4.1. Taking f =P z2n xn in the preceding Lemma we get By duality we get (using Remark 4.3 and (4.2)) and since this holds for all 1 < p < ∞, 1 ≤ q ≤ ∞ (the case q = ∞, q′ = 1 requires a minor adjustment, see [1, Remark, p.55] ). We deduce xn(cid:13)(cid:13)(cid:13)Lp,q(dm×tr) . . xp′,q′ xp,q .(cid:13)(cid:13)(cid:13)X z2n xn(cid:13)(cid:13)(cid:13)Lp′,q′ (dm×tr) (cid:13)(cid:13)(cid:13)X x2n xp,q ≃(cid:13)(cid:13)(cid:13)X z2n and hence by Remark 4.4 This proves (4.4). xn(cid:13)(cid:13)(cid:13)Lp,q(dm×tr) ≃(cid:13)(cid:13)(cid:13)X εnxn(cid:13)(cid:13)(cid:13)Lp,q(dµ×tr) . To complete the proof we use the following rather standard Lemma 4.8. For any 1 < p < ∞, 1 ≤ q < ∞, any f ∈ Lp,q(µ × tr) of the form f =P εnxn with xn ∈ Sp,q = Lp,q(tr) and any r < ∞ we have kf kLr(Sp,q) . kf kLp,q(µ×tr). Proof. We will use repeatedly Kahane's inequality for which we refer to [17] or [12, p. 100]. By K-convexity, the mapping P : f ∈ Lp,p(µ × tr) 7→ Lr(µ; Sp) defined by (projection onto span[εn]) P f =Xhf, εniεn 13 is bounded for any value of 1 < p < ∞ and (by Kahane's inequality) 1 ≤ r < ∞. By interpolation, it follows that P is bounded from (Lp0,p0(µ × tr), Lp1,p1(µ × tr))α,q → (Lr(Sp0), Lr(Sp1))α,q for any 0 < α < 1, 1 ≤ q ≤ ∞. We may choose p0 < p < p1, 1 p = 1−α p0 + α p1 and r ≥ q (here we use q < ∞), we then get P : Lp,q(µ × tr) → (Lr(Sp0), Lr(S1))α,q but now r ≥ q guarantees that Lq(Lr) ⊂ Lr(Lq) and hence (Lr(Sp0), Lr(Sp1))α,q ⊂Lr((Sp0, Sp1)α,q) = Lr(Sp,q). End of proof of Theorem 4.1. By Lemma 4.8 we get Thus Lemma 4.8 follows by restricting to f =P εnxn. (cid:13)(cid:13)(cid:13)X εnxn(cid:13)(cid:13)(cid:13)Lr(µ;Sp,q) xp′,q′ .(cid:13)(cid:13)(cid:13)X εnxn(cid:13)(cid:13)(cid:13)Lr′ (µ;Sp′ ,q′ ) and hence again by duality . xp,q . Then we conclude since Kahane's inequality allows us to replace both r and r′ by, say, 2. Remark. Using Fernique's inequality in place of Kahane's (see [12]) it is easy to see that the preceding proof of (4.3) and (4.4) extends to the case when (εn) is replaced by a sequence of i.i.d. Gaussian normal random variables. This implies (4.3) and (4.4). Note however (see [12, p. 253]) that the Gaussian and Rademacher averages are not equivalent when q = ∞. Remark 4.9. Let Aθ = (A0, A1)θ. Let us denote by Rad(Sp) the closure in Lp(µ × tr) of the set of finite sumsP εnxn with xn ∈ Sp. By the description of Aθ obtained in [24] (see also [3]), we know that if we define p by 1 1 by Theorem 4.1 p = 1−θ ∞ + θ (A0, A1)θ,p = Rad(Sp) = Aθ ∩ A1−θ Aθ + A1−θ if if θ ≤ θ ≥ 1 2 1 2 . and in particular (A0, A1)θ,p ≃ (A0, A1)θ. But this can also be seen using (6.1) and the identity Lp,p = Lp relative to ϕ×tr (and a simultaneous complementation argument) where (M, ϕ) is the free group II1 factor. From Theorem 4.1, it is natural to search for an equivalent of the K-functional for (A0, A1): Problem: Find an explicit description (presumably in terms of x, R and C) of Kt(x; A0, A1). 14 5 Remarks on real interpolation and non-commutative Khintchine p,q) the space of all sequences x = (xn) with xn ∈ Sp,q such nxn)1/2 ∈ Sp,q We need more notation about the Lorentz space version of the Schatten classes. Notation: We denote by X c p,q (resp. X r that the series P x∗ (resp. (P xnx∗ n)1/2 ∈ Sp,q). We equip these spaces with the norms: nxn (resp. P xnx∗ def= (cid:13)(cid:13)(cid:13)(cid:13)(cid:16)X x∗ n) (assumed w.o.t. convergent) satisfies (P x∗ nxn(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)Sp,q and k(xn)kX r def= k(x∗ k(xn)kX c p,q n)kX c p,q . p,q Lastly, we set X c p = X c p,p and X r p = X r p,p. Recall the following facts and terminology from [18]. Consider a compatible couple (X0, X1) of Banach (or quasi-Banach) spaces. Assume given a closed subspace S ⊂ X0 + X1 and let Let Q0 = X0/S0 and Q1 = X1/S1 be the associated quotient spaces. Clearly (Q0, Q1) form a compatible couple since there are natural inclusion maps S0 = S ∩ X0, S1 = S ∩ X1. Q0 → (X0 + X1)/S and Q1 → (X0 + X1)/S. We say that the couple (S0, S1) is K-closed (relative to (X0, X1)) if there is a constant c such that (5.1) ∀t > 0 ∀x ∈ S0 + S1 Kt(x; S0, S1) ≤ cKt(x; X0, X1). In the terminology of [18], (Q0, Q1) is called J-closed if, for some constant c, any element x ∈ Q0∩Q1 admits a simultaneous lifting x satisfying ∀t > 0 Jt(x; X0, X1) ≤ cJt(x; Q0, Q1). In the present paper we make the convention to say that (S0, S1) is J-closed when the pair (Q0, Q1) is J-closed in the above sense. Equivalently, we will say that (S0, S1) is J-closed if there is a constant c such that for any x ∈ X0∩X1 there is a single bx ∈ S0 ∩ S1 satisfying simultaneously distXj (x,bx) ≤ c distXj (x, Sj) both for j = 0 and j = 1. The basic (simple but useful) fact on which [18] rests, is that, with our new terminology, K-closed is equivalent to J-closed (this statement should not be confused with the more obvious fact associated to the duality between the K and J norms or between subspaces and quotients). Note also that this is valid for quasi-normed spaces. j = X c Example: Consider 1 ≤ p0, p1 ≤ ∞. Let X c pj , for j = 0, 1. Let Xj = X c j = X r j ⊕ X r j pj , X r and Sj ⊂ Xj, Sj def = {(x, −x)}. Then Sj ≃ X c j ∩ X r j and Xj/Sj ≃ X c j + X r j . Note also the special case: if p1 = 2, X1/S1 ≃ S1 ≃ ℓ2(S2). Proposition 5.1. The above pair (S0, S1) is J-closed (and hence K-closed) in the following cases: (i) If 0 < p0 < 2 and p1 = 2. 15 (ii) If 2 < p0 ≤ ∞ and p1 = 2. Consequently, in both cases, for any p strictly between p0 and p1 = 2 and any 1 ≤ q ≤ ∞, with θ defined by 1/p = (1 − θ)/p0 + θ/2, we have (with equivalent norms) (S0, S1)θ,q = X c p,q ∩ X r p,q and (X0/S0, X1/S1)θ,q = X c p,q + X r p,q. Proof. To show (S0, S1) is J-closed is the same as showing the following Claim: ∃c > 0 such that the following holds. Given x = (xn), xn ∈ Sp0 there are y = (yn) and z = (zn) such that x = y + z and we have simultaneously for j = 0 and j = 1 kykX r pj + kzkX c pj ≤ ckxkX r pj +X c pj . To prove this, we fix ε > 0 and choose y0, z0 such that x = y0 + z0 and ky0kX r kxkX r + ε. p0 p0 +X c p0 Let then ξ = (P y0(n)y0(n)∗)1/2, η = (P z0(n)∗z0(n))1/2. We have kξkp0 = ky0kX r , kηkp0 = . We can assume (by perturbation) that ξ > 0 and η > 0 so that ξ, η are invertible. We p0 + kz0kX c p0 < kz0kX c introduce the states p0 Let a be defined by 1 a = 1 p0 f = ξp0(tr(ξp0))−1, g = ηp0(tr(ηp0))−1. − 1 2 . We use the decomposition: x = (kξkp0f 1 a ) 1 a )by0 +bz0 · (kξkp0 g 1 whereby0 = (kξkp0 f 1 a )−1y0 andbz0 = z0 · (kξkp0 g X kby0(n)k2 2 = trXby0(n)by0(n) a (ξ/kξkp0 )2f − 1 ∗ = tr(f − 1 = tr(f ) = 1. a )−1. We note that f = (ξkξk−1 p0 )p0 we have = tr(kξk−2 p0 f − 1 − 2 aX y0(n)y0(n)∗f − 1 2 p0 a ) a ) = tr(f a ) SimilarlyP kbz0(n)k2 2 = 1. To simplify notation let α = kξkp0f T be the map (Schur multiplier with respect to the bases in which α, β are diagonal) defined by T (t) = αt + tβ, t ∈ S∞. Consider the decomposition a . We have x = α ·by0 +bz0 · β. Let 1 a and β = kξkp0g 1 x = T (T −1(x)) = αT −1(x) + T −1(x)β. We set yn = αT −1(xn) and zn = T −1(xn)β. Clearly (since 0 ≤ αi αi+βj ≤ 1) we have kynkS2 ≤ kxnkS2, kznkS2 ≤ kxnkS2 and hence kykℓ2(S2) ≤ kxkℓ2(S2), kzkℓ2(S2) ≤ kxkℓ2(S2). But also: xn = αby0(n) +bz0(n)β 1 αi + βj kT −1(xn)kS2 =(cid:13)(cid:13)(cid:13)(cid:13) S2(cid:17)1/2 [αiby0(n)ij + z0(n)ijβj](cid:13)(cid:13)(cid:13)(cid:13)S2 ≤ kby0(n)kS 2 + kbz0(n)kS2 ≤(cid:16)X kby0(n)k2 S2(cid:17)1/2 +(cid:16)X kbz0(n)k2 S2(cid:17)1/2 ≤ 2. ⇒ (cid:16)X kT −1(xn)k2 16 Therefore we conclude that (we set tn = T −1(xn)) kykX r p0 = k(αT −1(xn))nkX r p0 =(cid:13)(cid:13)(cid:13)(cid:13)α(cid:16)X tnt∗ n(cid:17)1/2(cid:13)(cid:13)(cid:13)(cid:13)Sp0 ≤ 2kαkSa ≤ 2kξkp0 = 2ky0kX r p0 and similarly This proves our claim, whence (i) and (ii) follows by duality. kzkX c p0 ≤ 2kz0kX c p0 . Corollary 5.2. Recall we set A0 = S∞(R) ∩ S∞(C), A1 = S1(R) + S1(C). Then: (A0, A1)θ,q = X r p,q ∩ X c p,q if θ < (A0, A1)θ,q = X r p,q + X c p,q if θ > 1 2 1 2 . Proof. As already mentioned, by a well known self-duality result (note A1 ≃ A∗ 0) we have (A0, A1)1/2,2 ≃ ℓ2(S2), ≃ X r 2,2 ∩ X c 2,2 ≃ X r 2,2 + X c 2,2. So the corollary follows immediately by reiteration (see [1, p. 48]). Remark. The complex analogue of the preceding statement was proved in [23, p. 109] (see also [3]). [8] or [10]) that the maps of the form t → αT −1(t), t → T −1(t)β Remark. It is known (see e.g. (corresponding to the Schur multipliers [αi(αi + βj)−1] and [βj(αi + βj)−1]) are c.b. on SQ for any 1 < Q < ∞. In particular, in the preceding proof, we have for some constant cQ kykX r Q ≤ cQkxkX r Q kzkX c Q ≤ cQkxkX c Q and hence kykX r selection for the pair (1, Q), one can show that we have for any 1 < p < Q and any 1 ≤ q ≤ ∞ . Now let us fix 2 < Q < ∞. Using this simultaneous ≤ cQkxkX r + kzkX c Q∩X c Q Q Q kxkX r 1 +X c 1,X r Q∩X c Q)θ,q p,q +X c p,q . kxk(X r and hence (reiteration again) where kxkX r p,q+X c p,q . kxkAp,q Ap,q def = (A0, A1)θ,q A0 = X r 1 + X c 1 A1 = X r ∞ ∩ X c ∞. By duality, we also have kxkAp,q . kxkX r p,q∩X c p,q . Remark. In particular we recover a result claimed in [13, remark after Corollary 4.3] (see also [3]) kxkX r 2,q+X c 2,q . kxkA2,q . kxkX r 2,q ∩X c 2,q . Exchanging the roles of 1 and ∞, we obtain 17 Corollary 5.3. Let B0 = X r 1 ∩ X c 1, B1 = X r ∞ + X c ∞. Then (B0, B1)θ,q = X r p,q ∩ X c p,q X r p,q + X c p,q if if θ < θ > 1 2 1 2 . Remark. It is not difficult to extend the results of this section to a general non-commutative Lp- space (associated to a semi-finite faithful normal trace) in place of Sp. Remark. I do not know whether Proposition 5.1 is valid for arbitrary p0 6= p1. 6 Connection with free probability We refer to [29] for all undefined notions in this section. Let (M, ϕ) be a "non-commutative probability space", i.e. (M, ϕ) is as (M, τ ) was before but we impose ϕ(1) = 1. Let (ξn) be a free family in M. In what follows we assume that (ξn) is either a free semi-circular (sometimes called "free Gaussian") family, or a free circular one (this corresponds to complex valued Gaussians) or a (free) family of Haar unitaries. The latter are the free analogues of Steinhaus random variables. They can be realized as the free generators of the free group F∞ in the associated von Neumann algebra (the so-called "free group factor"). We could also include the free analogue of the Rademacher functions, i.e. a free family of copies of a single random choice of sign ε = ±1. See [27] for a discussion of more general free families. Consider now a finite sequence x = (xn) with xn ∈ Lp,q(M, τ ). The free analogue of Khintchine's inequality is the following fact that was (essentially) observed in [5]. There are absolute positive constants c, C such that for any 1 ≤ p, q ≤ ∞ (6.1) (6.2) cxp,q ≤(cid:13)(cid:13)(cid:13)X ξn ⊗ xn(cid:13)(cid:13)(cid:13)Lp,q(ϕ⊗τ ) ≤ Cxp,q. In [5] only the cases p = 1 and p = ∞ are considered, but it is also pointed out there that the orthogonal projection onto span[ξn] is completely bounded on Lp(ϕ) both for p = 1 and p = ∞. From that simultaneous complementation, it is then routine to deduce (6.1). more explicitly x2,q can be tackled by a more detailed study of the distribution of the "non- Remark. In particular, we have x2,q ≃ kP ξn ⊗ xnkL2,q(ϕ×τ ). Perhaps our problem to compute commutative variable" P ξn ⊗ xn. But while, in Voiculescu's theory of the free Gaussian case, in case ξn, xn and hence P ξn ⊗ xn are all self-adjoint) how to use it to estimate the spectral distribution ofP ξn ⊗ xn or, say, its L2,q norm. the "R-transform" (free analogue of the Fourier transform) can be calculated, it is not clear (even Combining this with Theorem 4.1 (with a general M in place of B(H)) we find: Proposition. For any 1 ≤ p < ∞ and 1 ≤ q ≤ ∞ and for any sequence x = (xn) of operators in Lp,q(τ ) we have (cid:13)(cid:13)(cid:13)X εn ⊗ xn(cid:13)(cid:13)(cid:13)Lp,q(µ×τ ) ≃(cid:13)(cid:13)(cid:13)X ξn ⊗ xn(cid:13)(cid:13)(cid:13)Lp,q(ϕ×τ ) . Note that this clearly fails for p = ∞ since spanL∞(µ){εn} ≃ ℓ1 while spanL∞(ϕ){ξn} ≃ ℓ2. We refer the reader to [6] for far reaching extensions of (6.1) or (6.2) involving random matrices. It is tempting to look for a direct, more conceptual proof of (6.2) but this has always eluded us (see however [3]). Note also that analogues of (6.1) and (6.2) (as well as (4.3)) are entirely open for 0 < p < 1. 18 Remark 6.1. Proposition 5.1 (ii) yields some information on the distribution function of sums such as in L1(ϕ) and L∞(ϕ) (a fortiori they form a K-closed pair), it is easy to check that, if we set S =P ξn ⊗ xn. Since the spans of the variables (ξn) are completely complemented simultaneously S =P cn ⊗ xn or S =P λ(gn) ⊗ xn, uniformly over t Kt(x; A1, A0) ≃ Kt(S; L1(τ × tr), L∞(τ × tr)) =Z t 0 S†(s)ds, where S†(s) denotes the generalized s-number of S in the sense of [4]. Similarly, by a well known fact (see [1, p. 109]), we have uniformly over t Kt(x; A1/2,2, A0) ≃ Kt(S; L2(τ × tr), L∞(τ × tr)) ≃ Z t2 0 S†(s)2ds!1/2 . The fact that the pair (X r 2) is K-closed, seems to yield some further information. Indeed, the Kt-functional for that pair can be estimated simply from the corresponding result for the (easy) pairs (X r ∞ ∩ X c 2 ∩ X c ∞, X r ∞, X r ∞) and (X c 2). So we have ∞, X c (6.3) Kt(x; X r 2 ∩ X c 2, X r ∞ ∩ X c ∞) ≃ max{Kt(x; X r 2 , X r ∞), Kt(x; X c 2, X c ∞)}, and by a well known fact (see [1, p. 109]) (s) ds!1/2 and Kt(x; X c 2, X c ∞)} ≃ Z t2 0 (cid:16)X x∗ nxn(cid:17)† (s) ds!1/2 . ≃ max{Kt(x; X r 2 , X r ∞), Kt(x; X c 2, X c ∞)}, Kt(x; X r 2 , X r ∞) ≃ Z t2 Therefore, we find both (6.3) and n(cid:17)† 0 (cid:16)X xnx∗ Z t2 S†(s)2ds!1/2 0 (6.4) that where, of course, the equivalences are meant with constants independent of t. However, a very short and direct proof was recently given in [3] that there is a constant c such ∀t > 0 S†(ct) ≤(cid:16)(X xnx∗ n)1/2(cid:17)† (t) +(cid:16)(X x∗ nxn)1/2(cid:17)† (t), from which (6.4) (and hence (6.3)) follows easily. References [1] J. Bergh and J. Lofstrom. Interpolation spaces. An introduction. Springer Verlag, New York. 1976. [2] F. Cobos and T. Schonbek, On a theorem by Lions and Peetre about interpolation between a Banach space and its dual. Houston J. Math. 24 (1998), 325-344. [3] S. Dirksen and ´E. Ricard, Some remarks on noncommutative Khintchine inequalities, preprint arxiv, Aug. 2011. [4] T. Fack and H. Kosaki, Generalized s-numbers of τ -measurable operators, Pacific J. Math. 123 (1986), 269 -- 300. 19 [5] U. Haagerup and G. Pisier, Bounded linear operators between C ∗−algebras. Duke Math. J. 71 (1993) 889-925. [6] U. Haagerup and S. Thorbjørnsen, Random matrices and K-theory for exact C ∗-algebras, Doc. Math. 4 (1999), 341-450 (electronic). [7] A. Hess and G. Pisier, The Kt-functional for the interpolation couple (L∞(dµ; L1(dν), L∞(dν; L1(dµ)), Quart. J. Math. Oxford Ser. (2) 46 (1995), 333-344. [8] F. Hiai and H. Kosaki, Means for matrices and comparison of their norms. Indiana Univ. Math. J. 48 (1999), 899-936. [9] Y. Jiao, Non-commutative martingale inequalities on Lorentz spaces, Proc. Amer. Math. Soc. 138 (2010) 2431-2441. [10] M. Junge and J. Parcet, Rosenthal's theorem for subspaces of noncommutative Lp, Duke Math. J. 141 (2008), 75-122. [11] V.Kaftal, D.Larson and G.Weiss, Quasitriangular subalgebras of semifinite von Neumann al- gebras are closed, J. Funct. Anal. 107 (1992), 387-401. [12] M. Ledoux and M. Talagrand, Probability in Banach spaces. Isoperimetry and processes. Springer-Verlag, Berlin, 1991. [13] C. Le Merdy and F. Sukochev, Rademacher averages on noncommutative symmetric spaces, J. Funct. Anal. 255 (2008), 3329-3355. [14] F. Lust-Piquard, In´egalit´es de Khintchine dans Cp (1 < p < ∞), C. R. Acad. Sci. Paris S´er. I Math. 303 (1986), 289 -- 292. [15] F. Lust-Piquard and G. Pisier, Non-Commutative Khintchine and Paley inequalities, Arkiv for Math. 29 (1991), 241-260. [16] L. Maligranda, Interpolation between sum and intersection of Banach spaces. J. Approx. Th. 47 (1986) 42-53. [17] G. Pisier, Les inegalit´es de Khintchine-Kahane, d'apr´es C. Borell. S´eminaire sur la G´eom´etrie des Espaces de Banach 1977-78, exp. 7, ´Ecole Polytechnique, Palaiseau. [18] G. Pisier, Interpolation Between H p Spaces and Non-Commutative Generalizations I. Pacific Math. J. 155 (1992) 341-368. [19] G. Pisier, Complex interpolation and regular operators between Banach lattices. Archiv der Math. (Basel) 62 (1994) 261-269. [20] G. Pisier, Projections from a von Neumann algebra onto a subalgebra. Bull. Soc. Math. France 123 (1995) 139-153. [21] G. Pisier, Regular operators between non-commutative Lp-spaces. Bull. Sci. Math. 119 (1995) 95-118. [22] G. Pisier, The operator Hilbert space OH, complex interpolation and tensor norms. Memoirs Amer. Math. Soc. vol. 122 , 585 (1996) 1-103. 20 [23] G. Pisier, Noncommutative vector valued Lp-spaces and completely p-summing maps. Ast´erisque (Soc. Math. France) 247 (1998) 1-131. [24] G. Pisier, Introduction to operator space theory, London Mathematical Society Lecture Note Series, 294, Cambridge University Press, Cambridge, 2003. viii+478 pp. [25] G. Pisier, Real interpolation and transposition of certain function spaces, Preprint to appear. [26] G. Pisier and Q. Xu, Non-commutative Lp-spaces, Handbook of the geometry of Banach spaces, Vol. 2, 1459 -- 1517, North-Holland, Amsterdam, 2003. [27] ´E. Ricard and Q. Xu, Khintchine type inequalities for reduced free products and applications. J. Reine Angew. Math. 599 (2006), 2759. [28] N. Varopoulos, On an inequality of von Neumann and an application of the metric theory of tensor products to operator theory. J. Funct. Anal. 16 (1974) 83-100. [29] D. Voiculescu, K. Dykema, and A. Nica, Free random variables. Amer. Math. Soc., Providence, RI, 1992. [30] Q. Xu, Interpolation of operator spaces, J. Funct. Anal. 139 (1996) 500-539. 21
1707.06186	2	1707	2017-08-31T19:04:47	The Delta Game	[ "math.OA", "quant-ph" ]	We introduce a game related to the $I_{3322}$ game and analyze a constrained value function for this game over various families of synchronous quantum probability densities.	math.OA	math	THE DELTA GAME KEN DYKEMA, VERN I. PAULSEN, AND JITENDRA PRAKASH Abstract. We introduce a game related to the I3322 game and analyze a constrained value function for this game over various families of synchronous quantum probability densities. 1. Introduction A subject of a great deal of current research has been the study of various mathematical models for what should constitute the set of quantum correlations beginning with Tsirelson [17, 18] and continuing with [6, 4, 11, 14, 3, 13, 15, 16]. In particular, the Tsirelson conjectures ask whether or not several different math- ematical models for these conditional quantum probabilities yield the same sets of probability densities. Whether or not two of these models yield the same sets of densities is now known to be equivalent to Connes' embedding problem [6, 4, 11]. Recently, the work of W. Slofstra [15, 16] has shown that these different models generally yield different sets. Equality of the two sets of probabilities corresponding to Connes' problem is the only unresolved case. One way to try and distinguish between these various sets of probability densities is by studying the values of finite input-output games. Indeed, this is the approach that Slofstra uses successfully. However, his games are rather large and his results rely on some very deep results in the theory of finitely presented groups. So it is always interesting to see if any simpler games can illuminate these differences or if in fact their potentially different quantum values all coincide. Another topic of current interest is attempting to compute the quantum value of the I3322 game and decide if it is actually attained over the standard set of quantum probability densities. This has been the subject of a great deal of research [5], [2], [12], [19]. The game that we study is a simplification of the I3322 game. In [3] the concept of the synchronous value of a finite input-output game was introduced. In this paper we compute the synchronous quantum values, for each of these different models, for a simplification of the I3322 game that was brought to our attention by R. Cleve that we call the ∆ game. Our results show that the synchronous values of this game corresponding to the different mathematical models of quantum probabilities that we consider all Date: September 19, 2018(Last revised). 2010 Mathematics Subject Classification. Primary 46L05; Secondary 47L90. Key words and phrases. finite input-output game, values of games, Connes' embedding conjecture. The second and third authors were supported in part by NSERC. 1 2 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH coincide. However, we introduce an extra constraint to the value function that allows us to compute the value at each constraint parameter. In this case the value of the game is the supremum of this constrained value function over the constraint parameter. We show that although the four values of the game are all equal, only three of the four functions coincide. We show that the graphs of the three quantum value function coincide, but that they lie strictly below the function that one obtains by computing values over the set of vector correlations. This gives another way to see that the set of 3 input 2 output vector correlations is strictly larger than the corresponding sets of probabilistic quantum correlations. It was already known fact that these sets differ at the 2 input 2 output level -- for example using the CHSH game. In addition, earlier work of [1] used quantum chromatic numbers to show that there is a 15 input 7 output game with a perfect strategy in the set of vector correlations, but no perfect strategy among the quantum correlations. The main new insight from our results is that the process of studying constrained value functions of games can allow one to see separation when the unconstrained value sees no separation. 2. Preliminaries Recall that a general two person finite input-output game G, involves two non- communicating players, Alice (A) and Bob (B), and a Referee (R). The game is described by G = (IA, IB, OA, OB, λ) where IA, IB , OA, OB are nonempty finite sets, representing Alice's inputs, Bob's inputs, Alice's outputs and Bob's outputs, respectively, and with λ : IA × IB × OA × OB → {0, 1} a function. For each round of the game, Alice receives input v ∈ IA and Bob receives input w ∈ IB from the Referee and then Alice and Bob produce outputs i ∈ OA and j ∈ OB, respectively. They win if λ(v, w, i, j) = 1 and lose if λ(v, w, i, j) = 0. The function λ is called the rule or predicate function. Suppose that Alice and Bob have a random way to produce outputs. This is informally what is meant by a strategy. If we observe a strategy over many rounds we will obtain conditional probabilities (p(i, jv, w)), where p(i, jv, w) is the joint conditional probability that Alice outputs i on input v and Bob outputs j on input w. For this reason, any tuple (p(i, jv, w))i∈OA ,j∈OB ,v∈IA,w∈IB satisfying p(i, jv, w) ≥ 0 and Xi∈OA,j∈OB p(i, jv, w) = 1, ∀ v ∈ IA, w ∈ IB , will be called a correlation. A correlation (p(i, jv, w)) is called a winning or perfect correlation for G if λ(v, w, i, j) = 0 ⇒ p(i, jv, w) = 0, that is, it produces disallowed outputs with zero probability. THE DELTA GAME 3 If we also assume that the Referee chooses inputs according to a known proba- bility distribution π : IA × IB → [0, 1], that is, π(v, w) ≥ 0 and X(v,w)∈IA×IB π(v, w) = 1, then it is possible to assign a number to each correlation that measures the prob- ability that Alice and Bob will win a round given their correlation. The value of the correlation p = (p(i, jv, w)), corresponding to the distribution π on inputs, is given by V (p, π) = Xi,j,v,w λ(v, w, i, j)π(v, w)p(i, jv, w). Note that a perfect correlation always has value 1 and, provided that π(v, w) > 0 for all v and w a correlation will have value 1 if and only if it is a perfect correlation. The value of the game G with respect to a fixed probability density π on the inputs over a given set F of correlations is given by ωF (G, π) = sup{V (p, π) : p ∈ F}. Because the set of all correlations is a bounded set in a finite dimensional vector space, whenever F is a closed set, it will be compact and so this supremum over F will be attained. A finite input-output game as above is called synchronous provided that IA = IB := I, OA = OB := O and for all v ∈ I, λ(v, v, i, j) = 0 whenever i 6= j. This condition can be summarized as saying that whenever Alice and Bob receive the same input then they must produce the same output. A correlation (p(i, jv, w)) is called synchronous provided that p(i, jv, v) = 0 for all v ∈ IA and for all i 6= j. Note that when G is a synchronous game, then any perfect correlation must be synchronous. In this paper we are interested in studying the ∆ game, which is a synchronous game, and computing ω(∆,F ) as we let F vary over the various mathematical models for synchronous quantum correlations. We now introduce these various models for quantum densities. Recall that a set, {Rk}n k=1 Rk = I. Also a set of projections, {Pk}n k=1, of operators on some Hilbert space H is called a positive operator valued measure (POVM) provided Rk ≥ 0, for each k, and k=1, on some Hilbert space H is k=1 Pk = I. Thus every Pn called a projection valued measure (PVM) provided Pn PVM is a POVM. A quantum correlation for a game G means that Alice and Bob have finite di- mensional Hilbert spaces HA and HB, respectively. For each input v ∈ I, Alice has a PVM -- {Pv,i}i∈O on HA, and similarly for each input w ∈ I, Bob has a PVM -- {Qw,j}j∈O on HB. They also share a state h ∈ HA ⊗ HB (khk = 1) such that p(i, jv, w) = h(Pv,i ⊗ Qw,j)h, hi . The set of all (p(i, jv, w)) arising from all choices of finite dimensional Hilbert spaces HA,HB, all PVMs and all states h is called the set of quantum correlations denoted by Cq(n, m). 4 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH Another family of correlations are the commuting quantum correlations. In this case there is a single (possibly infinite dimensional) Hilbert space H and for each input v ∈ I, Alice has a PVM {Pv,i}i∈O, and similarly for each input w ∈ I, Bob has a PVM {Qw,j}j∈O, satisfying Pv,iQw,j = Qw,jPv,i (hence the name commuting). They share a state h ∈ H (khk = 1) such that p(i, jv, w) = h(Pv,iQw,j)h, hi . The set of all (p(i, jv, w)) arising this way is denoted by Cqc(n, m) and is called the set of commuting quantum correlations. Remark 2.1. In the above definitions one could replace the PVM's with POVM's throughout, and this is used as the definitions of these sets in many references. Since there are more POVM's then PVM's one might obtain larger sets, say eCq(n, m) and eCqc(n, m). But, in fact, eCq(n, m) = Cq(n, m) and eCqc(n, m) = Cqc(n, m). The fact that eCq(n, m) = Cq(n, m) follows by a simple dilation trick. On the Hilbert space The proof that eCqc(n, m) = Cqc(n, m) is somewhat more difficult and can be found HA, one simply uses a Naimark dilation to enlarge the space to KA and dilate the set of POVM's to a set of PVM's on KA. One similarly dilates Bob's POVM's to PVM's on KB and then considers the tensor products of these PVM's on KA ⊗KB. in [4, Proposition 3.4], and also as Remark 10 of [6]. A third proof appears in [14]. We shall sometimes refer to this as the disambiguation of the two possible definitions. Remark 2.2. By Theorem 5.3 in [13], Cq(n, m) ⊆ Cqc(n, m), with (p(i, jv, w)) ∈ Cq(n, m) if and only if (p(i, jv, w)) ∈ Cqc(n, m) such that the Hilbert space H in its realization is finite dimensional. There is yet another correlation set denoted by Cvect(n, m) that is often called the set of vector correlations. It is the set of all (p(i, jv, w)) such that p(i, jv, w) = hxv,i, yw,ji for sets of vectors {xv,i : v ∈ I, i ∈ O},{yw,j : w ∈ I, j ∈ O} in a Hilbert space H and a unit vector h ∈ H, which satisfy (1) xv,i ⊥ xv,j and yw,i ⊥ yw,j for all i 6= j in O. (2) Pi∈O xv,i = h =Pj∈O yw,j for all v, w ∈ I. (3) hxv,i, yw,ji ≥ 0 for all v, w ∈ I and i, j ∈ O. These correlations have been studied at other places in the literature, see for ex- ample [9] where they are referred to as almost quantum correlations and they can be interpreted as the first level of the NPA hierarchy [10]. The above correlation sets are related in the following way Cq(n, m) ⊆ Cqc(n, m) ⊆ Cvect(n, m) ⊂ Rn2m2 (2.1) for all n, m ∈ N and they are all convex sets. It is known that the sets Cqc(n, m) and Cvect(n, m) are closed sets in Rn2m2 . Set Cqa(n, m) = Cq(n, m) so that , (2.2) for all n, m ∈ N. W. Slofstra [16] recently proved that there exists an n and m such that Cq(n, m) is not a closed set. Hence Cq(n, m) is in general a proper subset of Cq(n, m) ⊆ Cqa(n, m) ⊆ Cqc(n, m), THE DELTA GAME 5 Cqa(n, m), but whether or not they are different for all values of n, m is unknown. It also remains an open question to determine whether Cqa(n, m) = Cqc(n, m) for all n and m or not. In [6], it was proven that if Connes' embedding problem is true then Cqa(n, m) = Cqc(n, m) for all n and m. The converse was proven in [11]. Thus we know that Cqa(n, m) = Cqc(n, m), ∀n, m is equivalent to Connes' embedding conjecture. For t ∈ {q, qa, qc, vect}, let C s relations. The synchronous sets C s The set of synchronous commuting quantum correlations, C s acterized in the following way. t (n, m) denote the subset of all synchronous cor- t (n, m) are also convex for t ∈ {q, qa, qc, vect}. qc(n, m), may be char- Let A be a unital C∗-algebra. Recall that a linear functional ϕ : A → C is called a tracial state if ϕ is positive, ϕ(1) = 1, and τ (ab) = τ (ba) for all a, b ∈ A. Theorem 2.3 (Theorem 5.5, [13]). Let (p(i, jv, w)) ∈ C s qc(n, m) be realized with PVMs {Pv,i : v ∈ I}i∈O and {Qw,j : w ∈ I}j∈O in some B(H) satisfying Pv,iQw,j = Qw,jPv,i and with some unit vector h ∈ H so that p(i, jv, w) = hPv,iQw,jh, hi. Then (1) Pv,ih = Qv,ih for all v ∈ I, i ∈ O; (2) p(i, jv, w) = h(Pv,iPw,j)h, hi = h(Qw,jQv,i)h, hi = p(j, iw, v); (3) Let A be the C∗-algebra in B(H) generated by the family {Pv,i : v ∈ I, i ∈ O} and define τ : A → C by τ (X) = hXh, hi. Then τ is a tracial state on A and p(i, jv, w) = τ (Pv,iPw,j). Conversely, let A be a unital C∗-algebra equipped with a tracial state τ and with {ev,i : v ∈ I, i ∈ O} ⊂ A a family of projections such that Pi∈O ev,i = 1 for all v ∈ I. Then (p(i, jv, w)) defined by p(i, jv, w) = τ (ev,iew,j) is an element of C s qc(n, m). That is, there exists a Hilbert space H, a unit vector h ∈ H and mutually commuting PVMs {Pv,i : v ∈ I}i∈O and {Qw,j : w ∈ I}j∈O on H such that p(i, jv, w) = h(Pv,iQw,j)h, hi = h(Pv,iPw,j)h, hi = h(Qw,jQv,i)h, hi This theorem and Remark 2.2 lead to the following characterization of C s q (n, m): Proposition 2.4. We have that (p(i, jv, w)) ∈ C s q (n, m) if and only if there ex- ists a finite dimensional C∗-algebra A with a tracial state τ and with a family of projections {ev,i : v ∈ I, i ∈ O} ⊂ A such that Pi∈O ev,i = 1 for all v ∈ I and p(i, jv, w) = τ (ev,iew,j) for all i, j, v, w. The set of synchronous vector correlations is described in the next proposition. Proposition 2.5. We have (p(i, jv, w)) ∈ C s vect(n, m) if and only if p(i, jv, w) = hxv,i, xw,ji for a set of vectors {xv,i : v ∈ I, i ∈ O} ⊂ H with xv,i ⊥ xv,j when i 6= j, i=1 xv,i = h for some unit vector h ∈ H, and hxv,i, xw,ji ≥ 0. The synchronous subsets satisfy inclusions as in expression 2.2, Pm q (n, m) ⊆ C s C s qa(n, m) ⊆ C s qc(n, m) ⊆ C s vect(n, m) ⊆ Rn2m2 , 6 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH t (n, m) = C s and since Cqa(n, m), Cqc(n, m), and Cvect(n, m) are closed sets it is easy to see that their synchronous subsets are also closed. We can also ask the synchronous analogues of the questions described before. It is easy to see that, Ct(n, m) = Ct′ (n, m) =⇒ C s t′ (n, m), but there is no a priori reason that the converses should hold. It is shown in [3] that C s qc(n, m) for all n, m ∈ N is equivalent to Connes' embedding conjecture. In [7, 4] it is shown that C s q (n, m) = C s qa(n, m). The questions described above can be formulated in terms of values of games. If we restrict ωF (G, π) to the synchronous subset F s of F , we obtain the synchronous value of the game G given the probability density π defined by q (n, m) = C s As before we write this as ωs relates the synchronous values of a game to Connes' embedding conjecture. (G, π) = sup{V (p, π) : p ∈ F s}. t (G, π) when F = Ct(n, m). The following proposition ωs F Proposition 2.6 (Proposition 4.1, [3]). If Connes' embedding conjecture is true then ωq(G, π) = ωqc(G, π) and ωs qc(G, π) hold for every game G and every distribution π. Remark 2.7. It is not known if the converse of any of these above implications is true. That is, for example, if ωq(G, π) = ωqc(G, π),∀G, ∀π, then must Connes' embedding conjecture be true? qa(G, π) = ωs q (G, π) = ωs We now introduce the ∆ game. 3. The ∆ Game The ∆ game is a nonlocal game with three inputs and two outputs. We have I = {0, 1, 2} as the input set and O = {0, 1} as the output set (thus n = 3, m = 2). Out of the 36 possible tuples (v, w, i, j), allowed rules (v, w, i, j) ∈ I × I × O × O are (0, 0, 0, 0), (0, 1, 0, 1), (1, 1, 0, 0), (1, 2, 0, 1), (2, 2, 0, 0), (2, 0, 0, 1), (0, 0, 1, 1), (0, 1, 1, 0), (1, 1, 1, 1), (1, 2, 1, 0), (2, 2, 1, 1), (2, 0, 1, 0), whereas the disallowed rules are (0, 0, 0, 1), (0, 1, 0, 0), (1, 1, 0, 1), (1, 2, 0, 0), (2, 2, 0, 1), (2, 0, 0, 0), (0, 0, 1, 0), (0, 1, 1, 1), (1, 1, 1, 0), (1, 2, 1, 1), (2, 2, 1, 0), (2, 0, 1, 1). The remaining 12 tuples (v, w, i, j) are all allowed. The first 12 allowed rules may be visualized as in Figure 1. The allowed edges (0, 0), (1, 1), (2, 2) are shown with dashed lines while (0, 1), (1, 2), (2, 0) are shown with solid lines. The dashed lines are even while the solid lines are odd. This means that if Alice and Bob are given inputs joined by dashed lines then they return outputs with even sum; and in the other case they return outputs with odd sum. THE DELTA GAME 7 Alice 0 Bob 0 1 2 1 2 Figure 1. ∆ game rule function. Alice and Bob receive inputs according to the uniform distribution π = (π(v, w)) on the set of inputs E = {(0, 0), (1, 1), (2, 2), (0, 1), (1, 2), (2, 0)}, that is, π(v, w) = 1 6 for all (v, w) ∈ E (and zero otherwise). To compute the synchronous value of the game given the distribution π we first compute the value of a single correlation p = (p(i, jv, w)), which is, 1 1Xi=0 p(i, iv, v) + p(i, i + 1v, v + 1)! , 6 2Xv=0 p(i, iv, v) + p(i, i + 1v, v + 1)! : p(i, jv, w) ∈ C s t (3, 2)) , so that the value of the game becomes, t (G, π) = sup( 1 ωs where t ∈ {q, qa, qc, vect}. Denote the expression inside the braces by, V (p, π) = 6 2Xv=0 1Xi=0 6 2Xv=0 eθ = 1 p(i, iv, v) + p(i, i + 1v, v + 1)! . 1Xi=0 dimensional Hilbert spaces. involving operators and vectors in the case of t = qc and t = vect, respectively. Moreover, when t = q, by Remark 2.2 it suffices to proceed as in the case t = qc We will use Theorem 2.3 and Proposition 2.5 to simplifyeθ and to obtain expressions using Theorem 2.3 to simplify eθ, but restricting to the case of operators on finite We first handle the t = qc case. By Theorem 2.3, a correlation (p(i, jv, w)) is in C s qc(3, 2) if and only if there exists a C∗-algebra A of B(H) generated by a family of projections {Av,i : i = 0, 1 and v = 0, 1, 2} satisfying Av,0 + Av,1 = IH for v ∈ {0, 1, 2} and a tracial state τ : A → C such that p(i, jv, w) = τ (Av,iAw,j) = h(Av,iAw,j)h, hi, for some unit vector h ∈ H. For notational convenience we define A0 = A0,0, A1 = A1,0, A2 = A2,0. We now define a "parameter" θ by setting θ = 1 3 τ (A0 + A1 + A2), + 1 3 1 2 (3.1) which enables us to write eθ as 2Xv=0 eθ = previous paragraph, writing xi for xi,0, we see that eθ is given by τ (A0 + A1 + A2) − τ (Av Av+1) = 2Xv=0 + θ − 1 2 1 3 1 3 1 2 + 1 3 hx0 + x1 + x2, hi − 1 3 2Xv=0 hxv , xv+1i, eθ = Similarly, in the t = vect case, using Proposition 2.5 and proceeding as in the τ (Av Av+1). 8 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH Then Av,1 = IH − Av = IH − Av,0 for v ∈ {0, 1, 2}. Using this we can rewrite eθ as p(i, iv, v) + p(i, i + 1v, v + 1) 2Xv=0 2Xv=0 1Xi=0 1Xi=0 eθ = = = 1 6 1 6 1 2 τ (Av,iAv,i) + τ (Av,iAv+1,i+1) + 1 3 τ (A0 + A1 + A2) − 1 3 2Xv=0 τ (Av Av+1). for some set of vectors {x0, x1, x2, h} in some Hilbert space H satisfying khk = 1 and, for all v and w, xv ⊥ (h − xv), Again letting θ = 1 hh − xv, h − xwi ≥ 0. hxv, h − xwi ≥ 0, hxv, xwi ≥ 0, 3 hx0 + x1 + x2, hi, we may write 1 3 1 2 (3.2) + θ − eθ = For each t ∈ {q, qa, qc, vect}, let Θs hxv, xv+1i. 2Xv=0 t denote the set of all points (θ,eθ) ∈ R2 that t (n, m) in the manner described t behaves under different values of t. It is easy to t (n, m). can be obtained from correlations (p(i, jv, w)) ∈ C s above. We want to see how Θs verify that Θs To find Θs t is a convex set since it is the affine image of the convex set C s t , it is enough to compute the following two functions for each θ, f u t (θ) = sup{eθ : (θ,eθ) ∈ Θs where u and l stand for upper and lower, respectively. We also need to determine if the supremum and the infimum are attained or not. Notice that in the qc case, v=0 τ (AvAv+1), v=0 τ (AvAv+1). t (θ) = inf{eθ : (θ,eθ) ∈ Θs 3P2 in order to find the supremum (resp., infimum) of eθ = 1 2 + θ − 1 we need to find the infimum (resp., supremum) of the quantityP2 0 ≤ 1 In the qc case, notice that since Av's are projections and τ is a state we get, 3 τ (A0 + A1 + A2) ≤ 1. Similarly in the vect case, by the Cauchy-Schwarz A similar statement holds for the vect case. t}, t}, f l THE DELTA GAME 9 inequality we get 0 ≤ 1 3hx0 + x1 + x2, hi ≤ 1. Hence 0 ≤ θ ≤ 1. Conversely, if θ ∈ [0, 1], then we can always find projections A0, A1, A2 in some C∗-algebra with a tracial state τ , such that 1 q ⊆ Θs 3 τ (A0 + A1 + A2) = θ. qa ⊆ Θs It is evident that Θs vect. qc ⊆ Θs Theorem 3.1. For t ∈ {q, qa, qc}, we have f u t (θ) =  vect(θ) = f u 1 2 + θ 3+θ 4 4−θ 4 3 2 − θ 3 for 0 ≤ θ ≤ 1 for 1 3 ≤ θ ≤ 1 2 ≤ θ ≤ 2 for 1 for 2 3 ≤ θ ≤ 1. 2 3 1 2 + θ 1+3θ−3θ2 2 − θ 3 2 3 for 0 ≤ θ ≤ 1 3 ≤ θ ≤ 2 for 1 for 2 3 ≤ θ ≤ 1. 3 (3.3) f l t (θ) = 1 2 , Moreover, we have (3.4) f l vect(θ) = 1 2 , In all of the these cases, the infimum and supremum are attained by both f u t . Since (θ,eθ) ∈ Θs f l t is a closed set in R2 for each t ∈ {q, qa, qc, vect}. In particular, we have Θs (3.5) t if and only if 0 ≤ θ ≤ 1 and f l t (θ) ≤ eθ ≤ f u t and t (θ), we see that Θs q = Θs qa = Θs qc ( Θs vect. The functions as obtained in Theorem 3.1 are shown in Figure 2. 0.9 eθ 0.7 0.5 0 1 3 1 2 θ 2 3 1 Figure 2. Plots of f l t = f l vect, f u t and f u vect from Theorem 3.1 The fact that the functions f u vect and f u qc are different allows us to deduce the following. Corollary 3.2. We have that C s qc(3, 2) ( C s vect(3, 2). 10 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH Remark 3.3. There is another larger set of correlations that we could have consid- ered, the nonsignalling correlations. For a definition, see [8]. If we let C s ns(n, k) denote the set of synchronous nonsignalling correlations, then it is shown in [8] that the set C s ns denote the analogous func- tion obtained by taking the supremum over the set of synchronous nonsignalling correlations, then the fact that the set of such correlations is a polytope implies that f u ns and we can conclude that C s ns would be piecewise linear. Hence, f u ns(n, 2) is a polytope. vect 6= f u If we let f u vect(3, 2) ( C s ns(3, 2). 4. The case of t = vect. In this section, we compute f l vect and f u vect to prove (3.4) in Theorem 3.1. We will employ the symmetrization provided by the next lemma. Lemma 4.1. Θs exist vectors x0, x1, x2, h in a Hilbert space with the properties: vect is equal to the set of pairs (θ,eθ) with 0 ≤ θ ≤ 1, such that there • khk = 1, • ∀v hxv , hi = hxv, xvi = θ, • ∀v hxv , xv+1i = β, where eθ = 1 2 + θ − β and 2θ − 1 ≤ β ≤ θ. Proof. By Proposition 2.4, and the discussion in Section 3, Θs vect is the set of pairs (θ,eθ) such that there exist vectors x0, x1, x2, h in a Hilbert space H with the properties that khk = 1, for all v and w, we have hxv , hi = hxv , xvi, hxv, xwi ≥ 0, hxv, h − xwi ≥ 0, hh − xv, h − xwi ≥ 0 and, moreover, 1 3 2Xv=0 hxv, hi = θ, hxv, xv+1i = β, 1 3 2Xv=0 where eθ = 1 2 + θ − β. The conditions appearing in the lemma are precisely these, but with the additional requirement that the quantities hxv , hi and hxv, xv+1i are the same for all v ∈ {0, 1, 2}. However, given x0, x1, x2, h satisfying these weaker conditions and considering eh = 1 √3 (h ⊕ h ⊕ h), 1 √3 (xv ⊕ xv+1 ⊕ xv+2) exv = in the Hilbert space H⊕3, we see that ex0,ex1,ex2,eh satisfy the stronger conditions and yield the same pair (θ,eθ). We now prove the part of Theorem 3.1 involving the case t = vect. (cid:3) Theorem 4.2. The functions (4.1) are given by f l vect(θ) = inf{eθ : (θ,eθ) ∈ Θs vect}, (4.2) f l vect(θ) = 1 2 , f u vect(θ) = f u vect(θ) = sup{eθ : (θ,eθ) ∈ Θs vect} 1 2 + θ 1+3θ−3θ2 2 − θ 3 2 3 for 0 ≤ θ ≤ 1 for 1 3 ≤ θ ≤ 2 for 2 3 ≤ θ ≤ 1. 3 THE DELTA GAME 11 Moreover, both the infimum and supremum are attained, for all values of θ ∈ [0, 1]. Proof. Fix θ ∈ [0, 1]. By Lemma 4.1, we are interested in the set of β such that there exist vectors x0, x1, x2, h in some Hilbert space satisfying the conditions listed there. Let yv = h − xv. Consider the Gramian matrix G associated with the seven vectors h, x0, x1, x2, y0, y1, y2. The conditions of Lemma 4.1 imply that this is the 7 × 7 matrix 1 θ θ θ 1 − θ θ − β θ − β 1 − θ θ − β 1 − θ 0 θ β θ β θ β β θ G = 0 0 θ θ β β 0  1 − θ 1 − θ 1 − θ  θ − β θ − β θ − β 1 + β − 2θ 1 + β − 2θ and furthermore, that G is positive semidefinite and θ − β θ − β θ − β 0 0 θ − β θ − β 1 − θ θ − β 1 + β − 2θ 1 − θ 1 + β − 2θ 1 + β − 2θ 1 + β − 2θ 1 − θ (4.3) max(0, 2θ − 1) ≤ β ≤ θ. Conversely, given any such 7×7 positive semidefinite matrix and with the additional condition (4.3), we can construct seven such vectors in a Hilbert space. Thus, we are interested in the set of β that satisfy (4.3) and yield a positive semidefinite matrix G given above. We apply one step of the Cholesky algorithm, and conclude that the 7× 7 matrix G is positive semidefinite if and only if the following 6 × 6 matrix G′ is positive semidefinite:  G′ = θ − θ2 β − θ2 β − θ2 θ2 − β θ2 − β θ2 − θ θ2 − β θ2 − β θ2 − θ β − θ2 θ − θ2 β − θ2 θ2 − β θ2 − β θ2 − θ β − θ2 β − θ2 θ − θ2 θ2 − β θ2 − β θ − θ2 β − θ2 β − θ2 θ2 − θ θ − θ2 β − θ2 θ2 − β β − θ2 θ2 − β θ − θ2 θ2 − β θ2 − β θ2 − θ β − θ2 β − θ2 θ − θ2 This matrix G′ partitions into a block matrix of the form(cid:20) A −A .  −A A (cid:21) , where A = a x x x a x x x a  , with a = θ − θ2 and x = β − θ2. Thus the matrix G′ is positive semi-definite if and only if A ≥ 0. Using the determinant criteria we see that A ≥ 0 if and only if x ≤ a and 2x3 − 3ax2 + a3 ≥ 0. Simplifying we see that A ≥ 0 if and only if − a 2 ≤ x ≤ a. Substituting the values of a and x, we find that the Gramian matrix G is positive semidefinite if and only if 3θ2 − θ 2 ≤ β ≤ θ. 12 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH Thus, the set of all possible β is the set satisfying max(cid:26) 3θ2 − θ 2 This becomes , 2θ − 1, 0(cid:27) ≤ β ≤ θ. for 0 ≤ θ ≤ 1 3 ≤ θ ≤ 2 for 1 for 2 3 ≤ θ ≤ 1. 3 3 0 ≤ β ≤ θ 3θ2−θ 2 ≤ β ≤ θ 2θ − 1 ≤ β ≤ θ Thus, we obtain the values (4.2) and we have that the infimum and supremum in (4.1) are attained. (cid:3) 5. The cases t ∈ {q, qa, qc}. In this section, we compute f l t and f u t when t ∈ {q, qa, qc} to prove (3.3) in Theorem 3.1. We begin with a symmetrization lemma, analogous to Lemma 4.1 qc (resp., Θs Lemma 5.1. The set Θs θ ≤ 1, such that there exists a C∗-algebra A (resp., a finite dimensional C∗-algebra, A) with a faithful tracial state τ and with projections A0, A1, A2 ∈ A such that for all v, q), is equal to the set of pairs (θ,eθ) with 0 ≤ (5.1) τ (Av) = θ, τ (Av Av+1) = β, 2 + θ − β. where eθ = 1 Proof. By Theorem 2.3 and the discussion in Section 3, (θ,eθ) belongs to Θs qc (respec- tively, Θs q) if and only if there is a C∗-algebra A (respectively, a finite dimensional C∗-algebra A), with a faithful tracial state τ and projections A0, A1, A2 such that 1 3 τ (Av) = θ, 1 3 2Xv=0 2Xv=0 2 + θ − β. But if such exist, then we can consider the C∗-algebra eA = 3 τ , and projections eAv = Av⊕ Av+1⊕ Av+2 where eθ = 1 A⊕A⊕A with the traceeτ = 1 that satisfy the stronger requirements of the lemma that include (5.1). 3 τ ⊕ 1 3 τ ⊕ 1 τ (Av Av+1) = β, (cid:3) We now have some C∗-algebra results. Proposition 5.2. Let A be a unital C∗-algebra with a faithful tracial state τ . Let A and P be hermitian elements in A. If AP − P A 6= 0, then there exists H = H∗ ∈ A such that, letting f (t) = τ (A(eiHtP e−iHt)) for t ∈ R , we have f′(0) > 0. Proof. If H ∈ A is hermitian, then f′(0) = iτ (AHP − AP H) = iτ ((P A − AP )H), where we used the fact that τ is a tracial state. Supppose AP − P A 6= 0. Let H = i(P A − AP ). Then H is hermitian and f′(0) = τ (P A − AP2) > 0, where the strict inequality follows beacuse AP − P A 6= 0 and τ is a faithful state. (cid:3) THE DELTA GAME 13 Corollary 5.3. Let A be a unital C∗-algebra with a faithful tracial state τ . Fix θ ∈ [0, 1]. Let β = inf(cid:26) 1 3 τ (AB + BC + CA) : A, B, C ∈ A projections, τ (A) = τ (B) = τ (C) = θ(cid:27). If there exist projections A0, B0, C0 in A such that τ (A0) = τ (B0) = τ (C0) = θ and β = 1 3 τ (A0B0 + B0C0 + C0A0), then [A0, B0 + C0] = [B0, C0 + A0] = [C0, A0 + B0] = 0. Proof. We will show that A0 commutes with B0 + C0 and the other commutation relations follow by symmetry. Let P = B0 + C0. Suppose, for contradiction, that [A0, P ] 6= 0. Then, by Proposition 5.2, there exists H = H∗ ∈ A such that if f (t) = τ (A0(eiHtP e−iHt)), then f′(0) > 0. Fix some small and negative t such that f (t) < f (0). Letting Bt = eiHtB0e−iHt and Ct = eiHtC0e−iHt, we see that Bt and Ct are themselves projections in A and τ (Bt) = τ (Ct) = θ. But then for our value of t, τ (A0Bt + BtCt + CtA0) = τ (A0(Bt + Ct) + BtCt) = τ (A0(eiHtP e−iHt)) + τ ((eiHtB0e−iHt)(eiHtC0e−iHt)) = f (t) + τ (B0C0) < f (0) + τ (B0C0) = 3β, which implies that β is not the infimum, contrary to hypothesis. Thus, A0 com- mutes with B0 + C0. (cid:3) We now consider the universal unital C∗-algebra A generated by self-adjoint projections A, B, and C satisfying the commutator relations (5.2) [A, B + C] = [B, A + C] = [C, A + B] = 0. By definition, this is obtained by separation and completion of the universal uni- tal complex algebra generated by noncommuting variables A, B and C, under the seminorm that is the supremum of seminorms obtained from Hilbert space repre- sentations whereby A, B and C are sent to self-adjoint projections satisfying the above relations. 14 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH Proposition 5.4. The universal C∗-algebra A described above is isomorphic to C8 ⊕ M2, wherein A = 0 ⊕ 0 ⊕ 0 ⊕ 0 ⊕ 1 ⊕ 1 ⊕ 1 ⊕ 1 ⊕(cid:18)1 B = 0 ⊕ 0 ⊕ 1 ⊕ 1 ⊕ 0 ⊕ 0 ⊕ 1 ⊕ 1 ⊕ 1 C = 0 ⊕ 1 ⊕ 0 ⊕ 1 ⊕ 0 ⊕ 1 ⊕ 0 ⊕ 1 ⊕ 1 4 √3 4 0 0 √3 4 0(cid:19) , 4 ! , 4 ! . √3 4 − 3 3 4 √3 4 − Proof. We will describe all irreducible ∗-representations of A on Hilbert spaces. Let Y = 2(B + C) − (B + C)2 ∈ A. By the commutation relations (5.2), Y commutes with A. We also note that Y = B + C − BC − CB and BY = B − BCB = Y B, namely, that Y commutes with B. Similarly, Y commutes with C. Hence Y lies in the center of A. Thus, under any irreducible ∗-representation π, Y must be sent to a scalar multiple of the identity operator. In other words, we have π(B + C − BC − CB) = π(Y ) = λπ(1) for some λ ∈ C, so that Similarly, we have π(CB) ∈ span π(cid:0){1, B, C, BC}(cid:1). π(CA) ∈ span π(cid:0){1, A, C, AC}(cid:1), π(BA) ∈ span π(cid:0){1, A, B, AB}(cid:1). Since A is densely spanned by the set of all words in the idempotents A, B and C, we see π(A) = span π(cid:0){1, A, B, C, AB, AC, BC, ABC}(cid:1). This implies that dim π(A) ≤ 8. Since π(A) is finite dimensional and acts irre- ducibly on a Hilbert space Hπ, it must be equal to a full matrix algebra. Consid- ering dimensions, we must have dimHπ ≤ 2. The irreducible representations π of A for which dimHπ = 1 are easy to describe. They are the eight representations that send A, B and C variously to 0 and 1. We will now characterize the irreducible representations π of A for which dimHπ = 2, up to unitary equivalence. Let π be such a representation. From the commutation relations (5.2), we see that, if π(A) and π(B) commute, then also π(C) commutes with π(A) and with π(B), and the entire algebra π(A) is commutative. This would require dimHπ = 1. By symmetry we conclude that no two of π(A), π(B) and π(C) can commute. In particular, each must be a projection of rank 1. After conjugation with a unitary, we must have π(A) =(cid:18)1 0 0 0(cid:19) , π(B) =(cid:18) t pt(1 − t) pt(1 − t) 1 − t (cid:19) for some 0 < t < 1. Since π(B) + π(C) must commute with π(A), we must have THE DELTA GAME 15 π(C) =(cid:18) c11 −pt(1 − t) −pt(1 − t) c22 (cid:19) , for some c11, c22 ≥ 0. Since π(C) is a projection, the only possible choices are (i) c11 = t and c22 = 1 − t and (ii) c11 = 1 − t and c22 = t. But in Case (ii), we have π(C) = IHπ − π(B), which violates the prohibition against π(C) and π(B) commuting. Thus, we must have π(C) =(cid:18) t −pt(1 − t) −pt(1 − t) 1 − t (cid:19) . Now, using that π(A) + π(B) and π(C) commute, we see that we must have t = 1 4 and we easily check that this does provide an irreducible representation of A. To summarize, up to unitary equivalence, there are exactly nine different irre- ducible representations of A, one of them is two-dimensional and the others are one-dimensional. Thus, A is finite dimensional and is isomorphic to the direct sum of the images of its irreducible representations, namely to C8 ⊕ M2, with A, B and C as indicated. (cid:3) We now prove Theorem 3.1 for the cases t ∈ {q, qa, qc}. Theorem 5.5. For t ∈ {q, qa, qc}, the functions (5.3) f u f l t (θ) = inf{eθ : (θ,eθ) ∈ Θs t}, are given by t (θ) = sup{eθ : (θ,eθ) ∈ Θs t} (5.4) f l t (θ) = 1 2 , f u t (θ) = 1 2 + θ 3+θ 4 4−θ 4 3 2 − θ 3 for 0 ≤ θ ≤ 1 3 ≤ θ ≤ 1 for 1 for 1 2 ≤ θ ≤ 2 for 2 3 ≤ θ ≤ 1. 2 3  Moreover, both the infimum and supremum are attained, for all values of θ ∈ [0, 1]. Proof. Fix θ ∈ [0, 1]. From the inclusions (3.5), we conclude qc(θ) ≤ f l f l qa(θ) ≤ f l q(θ) ≤ f u q (θ) ≤ f u qa(θ) ≤ f u qc(θ). To find f l qc(θ), by Lemma 5.1, we should find the supremum of values β such that there exists a C∗-algebra A with faithful tracial state τ and with projections A0, A1, A2 such that ∀v, τ (Av) = θ, (5.5) By Cauchy-Schwarz, β ≤ θ. But taking A = C ⊕ C with Av = 1 ⊕ 0 and an appropriate trace τ shows that β = θ occurs, and in a finite dimensional example. Thus, we find f l τ (AvAv+1) = β. qc(θ) = f l q(θ) = 1 2 . To find f u qc(θ), again using Lenma 5.1, we should find the infimum β0 of values β as described above. Since Θs qc is closed, this infimum is attained. Thus, there exists a C∗-algebra A with tracial state τ and projections A0, A1, A2 such that (5.5) holds with β = β0. Morover, by the proof of Lemma 5.1, we have that β0 equals 16 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH the infimum of 1 3 τ (AB + BC + CA) over all projections A, B, C in some C∗-algebra with faithful tracial state τ such that τ (A) = τ (B) = τ (C) = θ. Thus, Corollary 5.3 applies and the commutation relations [A0, A1 + A2] = [A1, A0 + A2] = [A2, A0 + A1] = 0 hold. Thus, there is a representation of the universal C∗-algebra A considered in Proposition 5.4, sending A to A0, B to A1 and C to A2. So, using Gelfand -- Naimark -- Segal representations, in order to find β0, it suffices to consider tracial states (faithful or not) on A. In particular, β0 is the minimum of all values of β ≥ 0 for which there exists a tracial state τ on A satisfying (5.6) τ (A) = τ (B) = τ (C) = θ, τ (AB) = τ (AC) = τ (BC) = β. Since A is finite dimensional, we get f u qc(θ) = f u An arbitrary tracial state of A is of the form q (θ). τ(cid:18)λ1 ⊕ ··· ⊕ λ8 ⊕(cid:18)x11 x12 x21 x22(cid:19)(cid:19) = 8Xj=1 tjλj + s 2 (x11 + x22), for some t1, . . . , t8, s ≥ 0 satisfying t1 +···+t8 +s = 1. The conditions (5.6) become t5 + t6 + t7 + t8 + = t3 + t4 + t7 + t8 + = t2 + t4 + t6 + t8 + = θ, s 2 s 2 s 2 t7 + t8 + These are equivalent to s 8 = t6 + t8 + s 8 = t4 + t8 + s 8 = β. s 8 − t8 t1 = 1 + 3β − 3θ + + t8 t2 = t3 = t5 = θ − 2β − t4 = t6 = t7 = β − s 4 s 8 − t8. Thus, writing t = t8, β0 is the minimum value of β such that there exist s, t ≥ 0 such that the inequalities 1 + 3β − 3θ + s 8 − t ≥ 0, θ − 2β − s 4 + t ≥ 0, β − s 8 − t ≥ 0. hold. This is a linear programming problem. We solved it by hand using the simplex method and also (to check) by using the Mathematica software platform [20]. The solution is, β0 = , 4 0, 3θ−1 5θ−2 , 2θ − 1, 4  1 3 0 ≤ θ ≤ 1 3 ≤ θ ≤ 1 2 ≤ θ ≤ 2 3 ≤ θ ≤ 1, 2 2 1 3 which, using f u qc(θ) = 1 2 + θ − β0, yields the values given in (5.4). (cid:3) The authors would like to thank Tobias Fritz for several useful remarks. Acknowledgements THE DELTA GAME 17 References [1] Cameron, P. J., Montanaro, A., Newman, M. W., Severini, S., and Winter, A. On the quantum chromatic number of a graph, The Electronic Journal of Combinatorics, 14 (2007), R81. [2] Collins, D., and Gisin, N. A relevant two qubit Bell inequality inequivalent to the CHSH inequality. Journal of Physics A: Mathematical and General 37, 5 (2004), 1775 -- 1787. [3] Dykema, K. J., and Paulsen, V. Synchronous correlation matrices and Connes' embedding conjecture. Journal of Mathematical Physics 57, 1 (2016). [4] Fritz, T. Tsirelson's problem and Kirchberg's conjecture. Reviews in Mathematical Physics 24, (2012), no. 05, 1250012. [5] Froissart, M. Constructive generalization of Bell's inequalities. Il Nuovo Cimento B (1971- 1996) 64, 2 (1981), 241 -- 251. [6] Junge, M., Navascues, M., Palazuelos, C., Perez-Garcia, D., Scholz, V. B., and Werner, R. F. Connes' embedding problem and Tsirelson's problem. Journal of Mathematical Physics 52, 1 (2011), 012102. [7] Kim, S.-J., Paulsen, V. I., and Schafhauser, C., A synchronous game for binary constraint systems. arxiv:1707.01016, (2017). [8] Lackey, B., and Rodrigues, N. Nonlocal games, synchronous correlations, and Bell inequal- ities preprint, July 10, 2017 [9] Navascu´es, M., Guryanova, Y., Hoban, M. J., and Ac´ın, A. Almost quantum correlations. Nat. Commun. 6:6288, (2015) [10] Navascu´es, M., Pironio, S., and Ac´ın, A. A convergent hierarchy of semidefinite programs characterizing the set of quantum correlations New Journal of Physics 10, (2008), no. 7, 073013. [11] Ozawa, N. About the Connes' embedding conjecture. Japanese Journal of Mathematics 8, 1 (2013), 147 -- 183. [12] P´al, K. F., and V´ertesi, T. Maximal violation of a bipartite three-setting, two-outcome bell inequality using infinite-dimensional quantum systems. Phys. Rev. A 82, (2010), 022116. [13] Paulsen, V. I., Severini, S., Stahlke, D., Todorov, I. G., and Winter, A. Estimating quantum chromatic numbers. Journal of Functional Analysis 270, 6 (2016), 2188 -- 2222. [14] Paulsen, V. I., and Todorov, I. G. Quantum chromatic numbers via operator systems. The Quarterly Journal of Mathematics 66, 2 (2015), 677. [15] Slofstra, W., Tsirelson's problem and an embedding theorem for groups arising from non- local games. arXiv:1606.03140., (2016). [16] Slofstra, W., The set of quantum correlations is not closed. arXiv:1703.08618., (2017). [17] Tsirelson, B. S. Some results and problems on quantum Bell-type inequalities. Hadronic Journal Supplement 8, 4 (1993), 329 -- 345. [18] Tsirelson, B. S. Bell inequalities and operator algebras. http://www.imaph.tu-bs.de/qi/problems/33.html, (2006). [19] Vidick, T. and Wehner, S. More nonlocality with less entanglement. Phys. Rev. A 83, 5 (2011), 052310. [20] Wofram Research Inc. Mathematica, Version 11.0. Wolfram Research Inc., Champaign, IL, 2016. 18 K. DYKEMA, V. I. PAULSEN, AND J. PRAKASH Department of Mathematics, Texas A& M University, College Station, TX E-mail address: [email protected] Institute for Quantum Computing and Department of Pure Mathematics, University of Waterloo, Waterloo, ON, Canada N2L 3G1 E-mail address: [email protected] Institute for Quantum Computing and Department of Pure Mathematics, University of Waterloo, Waterloo, ON, Canada N2L 3G1 E-mail address: [email protected]
1501.03027	2	1501	2015-03-17T14:49:40	Amenability of Groupoids Arising from Partial Semigroup Actions and Topological Higher Rank Graphs	[ "math.OA" ]	We consider the amenability of groupoids $G$ equipped with a group valued cocycle $c:G\to Q$ with amenable kernel $c^{-1}(e)$. We prove a general result which implies, in particular, that $G$ is amenable whenever $Q$ is amenable and if there is countable set $D\subset G$ such that $c(G^{u})D=Q$ for all $u\in G^{(0)}$. We show that our result is applicable to groupoids arising from partial semigroup actions. We explore these actions in detail and show that these groupoids include those arising from directed graphs, higher rank graphs and even topological higher rank graphs. We believe our methods yield a nice alternative groupoid approach to these important constructions.	math.OA	math	AMENABILITY OF GROUPOIDS ARISING FROM PARTIAL SEMIGROUP ACTIONS AND TOPOLOGICAL HIGHER RANK GRAPHS JEAN N. RENAULT AND DANA P. WILLIAMS Abstract. We consider the amenability of groupoids G equipped with a group valued cocycle c : G → Q with amenable kernel c−1(e). We prove a general result which implies, in particular, that G is amenable whenever Q is amenable and if there is countable set D ⊂ G such that c(Gu)D = Q for all u ∈ G(0). We show that our result is applicable to groupoids arising from partial semigroup actions. We explore these actions in detail and show that these groupoids include those arising from directed graphs, higher rank graphs and even topological higher rank graphs. We believe our methods yield a nice alternative groupoid approach to these important constructions. 1. Introduction It is often important to establish the amenability of groupoids that arise in applications. For example, amenability implies the equality of the reduced and universal norms on the associated groupoid algebras. This is important in the study of the Baum-Connes conjecture for groupoid C∗-algebras as it is the reduced algebra that plays the key role, while the universal algebra has the better functorial properties. For example, Tu has shown that the C∗-algebra of an amenable groupoid with a Haar system satisfies the Baum-Connes conjecture and the UCT [24]. In the classification program, amenability implies nuclearity which is typically a crucial hypothesis. The sort of groupoids we wish to focus on are those arising from the much studied C∗-algebras associated to higher-rank graphs, and more recently, to topological higher-rank graphs. As a specific example, we recently considered the C∗-algebras of groupoids associated to k-graphs (see [20]). Such groupoids have a canonical Zk-valued cocycle c, and it is not hard to see that c−1(0) is amenable. In some cases, c is not only surjective, but strongly surjective in that c(Gu) = Zk for all u ∈ G(0). Then the amenability of G is a consequence of [1, Theorem 5.3.14]. However, there are interesting examples in which c need not be strongly surjective, and examples show that the problem of the amenability of G turns out to be very subtle. A result of Spielberg [22, Proposition 9.3] asserts that if G is ´etale and if c : G → Q is a continuous cocycle into a countable amenable group, then G is amenable whenever c−1(eQ) is. Although this result is satisfactory for most k- graphs, the proof is unsatisfying in that it circumvents groupoid theory by invoking Date: January 15, 2015; Revised March 16, 2015. 1991 Mathematics Subject Classification. Primary 22A22, 46L55, 46L05. Key words and phrases. Groupoids, amenable groupoids, cocycle, semigroup actions, higher rank graphs, topological higher rank graphs. The second author was supported by a grant from the Simons Foundation. 1 2 JEAN N. RENAULT AND DANA P. WILLIAMS the nontrivial result that for ´etale groupoid, C∗(G) is nuclear if and only if G is amenable [15, Corollary 2.17]. In particular, this result is valid only for ´etale groupoids. Here we want to prove a general result that subsumes both cocycle results from [1] and from [22]. That such a result will have delicate hypotheses is foreshadowed by the observation that one can have a surjective continuous cocycle c from a groupoid into an amenable group Q such that c−1(eG) amenable and still have G fail to be amenable. For example, let Q be an amenable group with a nonamenable subgroup N (which obviously is not closed in Q). Let G be the group bundle G = Q` N and let c be the identity map: c(γ) = γ. Nevertheless, we obtain a quite general result: Theorem 4.2. It implies in par- ticular, that if c : G → Q is a continuous cocycle into an amenable group Q with amenable kernel such that there is a countable set D so that c(Gu)D = Q for all u ∈ G(0), then G is amenable. Even in this form, we recover both results above and remove the hypothesis that G be ´etale from Spielberg's result. Although our results apply to topological groupoids with Haar systems, it is interesting that our proof relies on the notions of Borel groupoids, Borel amenability and Borel equivalence. At a crucial juncture, we are able to show that our groupoid is Borel equivalent to a Borel amenable groupoid. Since we also show that Borel equivalence preserves Borel amenability, we can appeal to the result from [19] which demonstrates that topological amenability is equivalent to Borel amenabiltiy -- at least for locally compact groupoids with Haar systems. Having proved our cocycle result, it is crucial to show how it can be applied to groupoids which are currently being studied in the literature. To do this, we show that our results can be applied to groupoids arising from (partial) semigroup actions. A key role is played by the notion of directed action (Definition 5.2) which gives a partition of the space into orbits. We explore these actions in detail. Then, making significant use of hard work due to Nica and Yeend, we are able to show that such groupoids include those arising from directed graphs, higher rank graphs and even topological higher rank graphs. We think our methods using semigroup actions yield a nice alternative groupoid approach to these important constructions. To prove our cocycle result, we work with locally Hausdorff, locally compact groupoids which are always assumed to be second countable and to possess a Haar system. Note that a second countable, locally Hausdorff, locally compact groupoid is the countable union of compact Hausdorff sets. Hence its underlying Borel struc- ture is standard. In Section 2 we review the notion of Borel amenability and some of the basic properties of Borel groupoids we need in the sequel. In Section 3 we recall the notion of Borel equivalence and prove that it preserves Borel amenability. In Section 4, we prove the main amenability result. In Section 5 we introduce semi- group actions. When the action is directed, there is a semi-direct product groupoid. This condition puts into light two classes of semigroups: Ore semigroups and quasi- lattice ordered semigroups. Under reasonable hypotheses, our main result applies and the semi-direct product groupoid is amenable. In Section 6, we show how the groupoid corresponding to a topological higher rank graph (and therefore to the many subcases of topological higher rank graphs) can be realized as the semi-direct product groupoid of a directed semigroup action of P = Nd. If fact, we consider P - graphs, where P is an arbitrary semigroup rather than simply Nd as in the original definition. The natural assumption is that the semigroup be quasi-lattice ordered. AMENABILITY OF GROUPOIDS 3 If moreover the semigroup is a subsemigroup of an amenable group and the graph satisfies a properness condition, then this groupoid is amenable as well. Besides the amenability problem, the section reveals a tight connection between higher rank C- algebras and Wiener-Hopf C-algebras. In fact, we use the techniques introduced by Nica in [14]; in particular the Wiener-Hopf groupoid of a quasi-lattice ordered semigroup defined by Nica appears as a particular case of our general construction for topological higher rank graphs. Acknowledgments. We are very grateful to an anonymous referee for a thorough reading of our paper and for pointing out the relevance of the work of Exel [8] and Brownlowe-Sims-Vittadello [4]. In addition the referee pointed out a gap in the original proof of Theorem 5.13, and suggested a line of attack which resulted in an improved version of the result. 2. Borel Amenability For the details on Borel groupoids, proper Borel amenability and proper Borel G-spaces, we refer to [1] and to [1, §2.1a] in particular. Recall that in order for a groupoid to act on (the left) a space X, we require a map rX : X → G(0). If there is no ambiguity, we write simply r in place of rX and call it the moment map.1 As in [1], to avoid pathologies we will always assume that our Borel spaces are analytic Borel spaces. We recall some of the basics here. If X and Y are Borel spaces and π : X → Y is Borel surjection, then a π-system is a family of measures m = {my}y∈Y such that each my is supported in π−1(y) and such that y 7→ ZX f (x) dmy(x) is Borel for any nonnegative Borel function f on X. If G is a Borel groupoid, X and Y are Borel G-spaces and π is G-invariant, then we have that m is invariant if γ · my = mγ·y whenever s(γ) = r(y). (By definition, γ · my(E) = my(γ−1 · E).) Definition 2.1 ([1, Definition 2.1.1(b)]). Suppose that G is a Borel groupoid and that X and Y are Borel G-spaces. A surjective Borel G-map π : X → Y is G- properly amenable if there is a invariant Borel π-system m = {my}y∈Y of probability measures on X. Then we say that m is an invariant mean for π. Remark 2.2. Notice that if π : X → Y is a Borel G-map, then the induced map π : G\X → G\Y is Borel with respect to the quotient Borel structures. Suppose that π is G-properly amenable and that {λy} is an invariant mean for π. Let p : X → G\X be the quotient map. Then the push forward p∗λy is a probability measure supported on π−1( y). By invariance, it depends only on y and we can denote this measure by λ y. If b is a bounded Borel function on G\Y , then ZG\Y b( z) d λ y( z) = ZX b(p(x)) dλy(x). Hence { λ y} y∈G\Y is a Borel π-system of probability measures on G\X. Definition 2.3 ([1, Defintion 2.1.2]). A Borel groupoid G is proper if the range map r : G → G(0) is G-properly amenable. A Borel G-space X is proper if the projection p : X ∗ G → X is G-properly amenable. 1The term projection map, anchor map and structure map have also been used in the literature. 4 JEAN N. RENAULT AND DANA P. WILLIAMS Remark 2.4. Thus X is a proper G-space if and only if there is a family {mx}x∈X of probability measures mx on G supported in Gr(x) such that x 7→ mx(f ) is Borel for all non-negative Borel functions on G and such that γ · mx = mγ·x. If X and Y are (left) G-spaces, then the fibered product X ∗ Y is a G-space with respect to the diagonal action. If X and Y are Borel G-spaces, then we give X ∗ Y the Borel structure as a subset of X × Y with Borel structure generated by the Borel rectangles. In particular, if X is a G-space, then X ∗ G = { (x, γ) ∈ X × G : r(x) = r(γ) } is a G-space with respect to the diagonal action.2 Definition 2.5. Suppose that G is a Borel groupoid and that X and Y are Borel G-spaces. A surjective Borel G-map π : X → Y is G-amenable (or simply amenable if there is no ambiguity about G) if there is a sequence {mn}n∈N of Borel systems of probability measures y 7→ my n on X which is approximately invariant in the sense n − mγ·y that for all γ ∈ G, kγ · my n k1 converges to 0, where k · k1 is the total variation norm. We say that {mn}n∈N is an approximate invariant mean for π. The notion of Borel amenability for groupoids was first formalized in [19, Defini- tion 2.1] -- however, here we use the formulation in which conditions (ii) and (iii) of [19, Definition 2.1] have been replaced by (ii′). Definition 2.6. A Borel groupoid G is (Borel) amenable if the range map r : G → G(0) is G-amenable. A Borel G-space X is amenable if the projection p : X ∗G → X is G-amenable. A key result about Borel amenable maps we need is the following. Lemma 2.7. Let G be a Borel groupoid. If there is a proper Borel G-space X such that the moment map r : X → G(0) is Borel amenable, then G is Borel amenable. Proof. Since, by assumption, the map πX : X ∗ G → X is properly amenable, there is an invariant system m = {mx}x∈X of probability measures for πX . We can view each mx as a measure on GrX (x) such that γ · mx = mγ·x. Since rX is amenable, there is an approximate invariant mean {µn}. Then µu n is a probability measure on X (u), and for all γ ∈ G, limn kγ · µs(γ) r−1 n = ZX k1 = 0. Define n − µr(γ) mx dµu n(x). λu n Then λu is certainly Borel, and n is a probability measure supported in Gu. The system λn = {λu n}u∈G(0) kγ · λs(γ) n − λr(γ) n γ · my dµs(γ) n mx µr(γ) n ZX k1 = (cid:13)(cid:13)(cid:13) ZX = (cid:13)(cid:13)(cid:13) ZX = (cid:13)(cid:13)(cid:13) (y) −ZX (y) −ZX n − µr(γ) n (x)(cid:13)(cid:13)(cid:13) (x)(cid:13)(cid:13)(cid:13) mγ·x dµs(γ) n mx dµr(γ) n mx d(γ · µs(γ) ≤ kγ · µs(γ) n − µr(γ) n k1. )(x)(cid:13)(cid:13)(cid:13) 2In fact, X ∗ G is a groupoid -- sometimes called the transformation groupoid for G acting on X. This groupoid is denoted by X ⋊ G in [1]. After identifying its unit space with X, the range map (x, γ) 7→ x is clearly G-equivariant. Hence, in Defintion 2.3, we defined a G-space X to be properly amenable exactly when the groupoid X ⋊ G is properly amenable. AMENABILITY OF GROUPOIDS 5 It follows that λ is an approximate invariant mean for r : G → G(0). Thus G is Borel amenable. (cid:3) Of course, a proper Borel groupoid is Borel amenable. Our next lemma should be compared with [1, Proposition 5.3.37]. Lemma 2.8. Suppose that G is a Borel groupoid which is the increasing union of a sequence of proper Borel subgroupoids Gn. Then G is Borel amenable. Proof. By assumption, we can find an system mm = (mu bility measures on Gn; that is, γ · ms(γ) for all γ ∈ Gn. Using the standard theory of disintegration of measures for example (see [25, Theorem I.5]), we can extend each mm to a Borel system of probability measures on all of G(0). Then that m = (mn) is an approximate invariant mean for r : G → G(0). (cid:3) of invariant proba- n)u∈G(0) n = mr(γ) n n 3. Borel Equivalence The definition of equivalence for Borel groupoids is given in the appendix of [1]. In light of [19], it turns out to be a much more significant notion than originally thought. Definition 3.1 ([1, Definition A.1.11]). Let G and H be Borel groupoids. A (G, H)-Borel equivalence is a Borel space Z such that (a) Z is a free and proper left Borel G-space, (b) Z is a free and proper right Borel H-space, (c) The G- and H-actions commute, (d) r : Z → G(0) induces a Borel isomorphism between Z/H and G(0), and (e) s : Z → H (0) induces a Borel isomorphism between G\Z and H (0). In this case, we say that G and H are Borel equivalent. It is proved in [1, Theorem 2.2.17] that equivalence of locally compact groupoids preserves (topological) amenability. Here we show a similar result holds in the Borel case using virtually the same argument. Theorem 3.2. Suppose that G and H are equivalent Borel groupoids. If H is Borel amenable (resp., properly amenable), then so is G. Proof. Let Z be a (G, H)-equivalence. Consider the commutative diagram G pG Z ∗ G p Z r pZ q / G(0) r Z q / G\Z, where p(z, γ) := γ−1 · z. First, assume that G is properly amenable. To see that H is properly amenable it will suffice, by [1, Corollary 2.1.7], to see that that H-map s : Z → H (0) is properly amenable. / O O / / O O / 6 JEAN N. RENAULT AND DANA P. WILLIAMS Let λ = {λu}u∈G(0) be an invariant mean for r : G → G(0). We first build an invariant mean λZ for pZ : Z ∗ G → Z: let λz Z be given by λz Z (f ) = ZG f (z, λ) dλrZ (z)(γ). Consider the measure on Z obtained by the push forward q∗λz implies η−1 · λr(η) = λs(η) for all η ∈ G, it follows that Z . Since invariance q∗λη·z Z (b) = ZG = ZG = q∗λz Z (b). Z (γ) b(p(η · z, γ)) dλη·z b(γ−1η · z) dλrZ (η·z)(γ) = ZG b(γ−1 · z) d(η−1 · λr(η))(γ) Hence the push forward q∗λz Z depends only on z in G\Z. Since Z is a Borel equivalence, we can identify the quotient map q : Z → G\Z with the moment map s : Z → H (0). Thus we get an s-system of measures ν = {νv}v∈H(0) where νv = q∗λz Z for any z with s(z) = v. It will suffice to see that ν is invariant. But if s(z) = r(h), then ZZ b(z) d(νr(h) · h)(z) = ZG = ZZ b((γ−1 · z) · h) dλr(z)(γ) = ZG b(γ−1 · (z · h)) dλr(z·h(γ) b(z) dνs(h)(z). This completes the proof for proper amenability. Now we assume that G is Borel amenable and that λ = {λn} is an approximately invariant mean for r : G → G(0) with λn = {λu n}u∈G(0) . By Lemma 2.7, it will suffice to show that the H-map s : Z → H (0) is Borel amenable. The argument parallels that above argument for proper amenability. We begin by lifting each λn to a system λn,Z = {λz n,Z}z∈Z of probability mea- sures for pZ : Z ∗ G → Z: λz n,Z (f ) = ZG f (z, γ) dλr(z) n (γ). We let µz µz n depends only on z. However n = q∗λz n,Z . Since λn,Z is not necessarily invariant, we can't assert that ZZ b(w) dµz n(w) = ZG b(γ−1 · z) dλr(z) n (γ), and µz n is supported on q−1(z) = G · z. Since Z is a proper G-space, by definition there is a G-invariant system of prob- ability measures ρ = {ρz}z∈Z for the Borel G-map pZ : Z ∗ G → Z which we view as a family of measures on G such that supp ρz ⊂ Gr(z). As in Remark 2.2, ρ drops to a system ρ = { ρ z} z∈G\Z for the quotient map q : Z → G\Z. Since (z, γ) 7→ γ−1 · z identifies G\(Z ∗ G) with Z, we can view these as measures on Z. Explicitly, ρ z = p∗ρz and ρ z(b) = ZG b(γ−1 · z) dρz(γ). AMENABILITY OF GROUPOIDS 7 Using the invariance of ρ we get measures depending only on z supported on q−1( z) by averaging the µz n with respect to ρ z: (3.1) ν z n = ZZ µz n d ρ z(z) = ZG µγ−1·z n dρz(γ). As in the first part of the proof, we use the fact that Z implements an equivalence to identify the quotient map q : Z → G\Z with the moment map s : Z → H (0). Thus we get an s-system {νv}v∈H(0) where νv = ν z for any z such that s(z) = v. We will complete the proof by showing that ν is an approximately invariant mean for s. Therefore we need to see that for all h ∈ H, kνr(h) k1 tends to zero with n. Notice that if (z, γ) ∈ Z ∗ G, then · h − νs(h) n n kµz n − µγ−1·z n k1 ≤ kλr(γ) n − λ · λs(γ) n k1. Next we claim that for fixed z ∈ Z, limn kν z − µz (3.1), view ρz as a measure on Gr(z), and deduce that nk1 = 0. To see this we employ (3.2) kν z n − µz nk1 ≤ ZG kµγ−1·z n − µz nk dρz(γ) ≤ ZG kλr(γ) n − γ · λs(γ) n k1 dρz(γ). Since for each γ, limn kλr(γ) Lebesgue Dominated Convergence Theorem. n − γ · λs(γ) n k1 = 0, we see that (3.2) goes to zero by the Now fix h ∈ H and ǫ > 0. Let z ∈ Z be such that s(z) = r(h) and let z′ = z · h and observe that µz n · h = µz ′ n . Let M be such that n ≥ M implies that Then for n ≥ M , we have kν z ′ n − µz ′ n k < ǫ 2 . kνr(h) n · h − νs(h) n k1 ≤ kνr(h) · h − µz n · hk1 + kµz nk + kµ z n − νs(h) n k n = kν z ′ n − µz ′ n k1 + 0 + kµz ′ n − ν z ′ n · h − µ z n k < ǫ. Thus s : Z → H (0) is Borel amenable. This completes the proof. (cid:3) We close this section with two technical results which will be of use in the next section. Recall from [19, Definition 2.3] that a Borel approximate invariant density on G is a sequence (gn) of non-negative Borel functions on G such that ZG gn(γ) dλu(γ) ≤ 1 for all n, ZG gn(γ) dλu(γ) → 1 for all u ∈ G(0) and ZG(cid:12)(cid:12)gn(γ−1γ′) − gn(γ′)(cid:12)(cid:12) dλr(γ)(γ′) → 0 for all γ ∈ G. Lemma 3.3. Let G be a Borel groupoid with a Borel Haar system {λu}u∈G(0) . Let {Yi}∞ i=1 be a countable cover of G(0) by invariant Borel subsets such that each GYi is Borel amenable. Then G is Borel amenable. Proof. Since each Yi is invariant, u ∈ Yi implies that (GYi )u = Gu. Hence λ restricts to a Borel Haar system on GYi. Hence by [19, Proposition 2.4], the Borel amenability of GYi implies that there is a Borel approximate invariant density), {gi n}∞ Let B1 = Y1 and if i ≥ 2, let Bi = Yi \ Si−1 j=1 Yj. Then the {Bi} are a pairwise disjoint cover of G(0) by invariant Borel sets (some of which might be empty). Let bi be the characteristic function of Bi. Then each bi is a Borel function on G(0) n=1, on GYi. 8 JEAN N. RENAULT AND DANA P. WILLIAMS (taking the values 0 and 1) such that bi vanishes off Yi and for each u ∈ G(0), there is one and only one i such that bi(u) = 1. In particular, we trivially have ∞ Xi=1 bi(u) = 1 for all u ∈ G(0). For each n, define gn on G by gn(γ) = Xi gi n · bi, where gi n · bi(γ) = gi n(γ)bi(s(γ)). By invariance, ZG gn(γ) dλu(γ) = ZG gi n(γ) dλu(γ), where i is such that bi(u) = 1. Now it is clear that {gn} is a Borel approximate invariant density for G. Hence G is Borel amenable as claimed. (cid:3) Recall that by assumption, all of our Borel spaces, and our Borel groupoids in particular, are analytic Borel spaces. Lemma 3.4. Let G be a Borel groupoid acting freely on (the right of ) a Borel space X such that the quotient map q : X → X/G has a Borel cross section. Then X is a proper Borel G-space. Proof. Let X ⋊ G be the transformation groupoid for the right G-action.3 The map (x, γ) 7→ (x, x · γ) is a Borel as well as an algebraic isomorphism of X ⋊ G onto the equivalence relation groupoid X ∗q X = { (x, y) ∈ X × X : q(x) = q(y) }. Since X ∗q X and X ⋊ G are both Borel subsets of the corresponding product spaces, they are themselves analytic Borel spaces. Hence X ⋊G and X ∗q X are Borel isomorphic by Corollary 2 of [2, Theorem 3.3.4]. Hence it suffices to see that X ∗q X is proper. We can identify X with the unit space of X ∗q X. For each x ∈ X, let mx = δx × δc(q(x)) where c : X/G → X is our Borel cross section for q. Since c(q(x)) = c(q(y)) if (x, y) ∈ X ∗q X, we have (y, x)·mx = my. Hence X is a proper Borel G-space. (cid:3) 4. Cocycles In this paper, by a cocycle on a groupoid G we mean a homomorphism c : G → Q into a group Q. As in [17, Definition 1.1.9],4 the corresponding skew-product groupoid is G(c) = { (a, γ, b) ∈ Q × G × Q : b = ac(γ) }. Then multiplication is given by (a, η, b)(b, γ, d) = (a, ηγ, d) and inversion by (a, γ, b)−1 = (b, γ−1, a). We can identify the unit space of G(c) with G(0) × Q, and then the range and source maps are given as expected: r(b, γ, a) = (r(γ), b) while s(b, γ, a) = (s(γ), a). If G is a locally Hausdorff, locally compact groupoid with a Haar system { λu }u∈G(0) and if Q is a locally compact group, then provided c 3Here X ⋊ G = { (x, γ) ∈ X × G : s(x) = r(γ) } has multiplication defined by (x, γ)(x · γ, η) = (x, γη). 4In [17], G(c) is described as pairs (γ, a) ∈ G × Q. The groupoids are isomorphic via the map (a, γ, b) 7→ (γ, a). AMENABILITY OF GROUPOIDS 9 is continuous, G(c) is a locally Hausdorff, locally compact groupoid with Haar system β = { β(u,a) }(u,a)∈G(0)×Q where β(u,a)(g) = ZG g(a, γ, ac(γ)) dλu(γ). Later it will be useful to note that Q acts on the left of G(c) by (groupoid) automorphisms: q · (b, γ, a) = (qb, γ, qa). The action of Q on G(0) × Q, identified with the unit space of G(c), is given by q · (u, a) = (u, qa). Part of our interest in G(c) is due to the following result. Proposition 4.1 ([17, Proposition II.3.8]). Suppose that G is a second countable locally Hausdorff, locally compact groupoid with a Haar system and that c is a continuous homomorphism from G into a locally compact group Q. If G(c) and Q are both amenable, then so is G. Proof. By [19, Definition 2.7], we need to construct a topological approximate in- variant mean {gn} in C (G). It will suffice to show that for each pair of compact sets K ⊂ G and L ⊂ G(0), and each ǫ > 0, there is a nonnegative function f ∈ C (G) such that (A) ZG (B) ZG (C) ZG f (γ) dλu(γ) ≤ 1 for all u ∈ G(0), f (γ) dλu(γ) ≥ 1 − ǫ for all u ∈ L and f (γ−1γ′) − f (γ′) dλr(γ)(γ′) ≤ ǫ for all γ ∈ K. Since Q is amenable and c(K) is compact, there is a nonnegative function k ∈ Cc(Q) such that with respect to a fixed right Haar measure α on Q we have ZQ k(a) dα(a) = 1 and ZQ k(ab) − k(a) dα(a) ≤ ǫ 2 for all b ∈ c(K). On the other hand, the (topological) amenability of G(c) allows us to find a non- negative function g ∈ C (G(c)) such that (a) ZG (b) ZG (c) ZG g(a, γ, ac(γ)) dλu ≤ 1 for all (u, a) ∈ G(0) × Q, g(a, γ, ac(γ)) dλu ≥ 1 − ǫ for all (u, a) ∈ L × supp k and g(ac(γ), γ−1γ′, ac(γ′)) − g(a, γ′, ac(γ′)) dλr(γ) ≤ ǫ for all (a, γ) ∈ supp k · c(K)−1 × K. Now we define a nonnegative function on G via f (γ) := ZG k(a)g(a, γ, ac(γ)) dα(a). If g is continuous and supported in a compact Hausdorff subset, then f is too. Since in general, g a finite sum of such functions, we have f ∈ C (G). Then it is easy to check f satisfies (A) and (B) above. On the other hand, ZG f (γ−1γ′) − f (γ′) dλr(γ)(γ′) 10 JEAN N. RENAULT AND DANA P. WILLIAMS = ZGZQ = ZGZQ k(a)(cid:0)g(a, γ−1γ′, ac(γ−1γ′)) − g(a, γ′, ac(γ′))(cid:1) dλr(γ)(γ′) k(ac(γ))(cid:0)g(ac(γ), γ−1γ′, a(c(γ′)) − g(a, γ′, ac(γ′))(cid:1) dλr(γ)(γ′) dα(a) +ZGZQ(cid:0)k(ac(γ)) − k(a)(cid:1)g(a, γ′, ac(γ′)) dλr(γ)(γ′) dα(a) which, in view of our assumptions on k and g, is ≤ ǫ 2 + ǫ 2 = ǫ. (cid:3) Theorem 4.2. Suppose that G is a locally Hausdorff, locally compact second count- able groupoid with a Haar system and that c : G → Q is a continuous homomor- phism into a locally compact group Q such that the kernel c−1(e) is amenable. Let s : G → G(0) × Q be the map given by s(γ) = (s(γ), c(γ)). Let Y be the image of s in G(0) × Q. Then Y is G(c)-invariant. Suppose that there are countably many qn ∈ Q such that G(0) × Q = [n qn · Y, then G(c) is amenable. amenable. If in addition, Q is amenable, then so is G. In particular, if Y is open in G(0) × Q, then G(c) is It is not hard to check that Y is invariant: if s(a, γ, ac(γ)) = (s(γ), ac(γ)) ∈ Y , then there is an η ∈ G such that s(η) = (s(γ), ac(γ)). Hence s(η) = s(γ) and c(η) = ac(γ). But then s(ηγ−1) = (r(γ), c(η)c(γ)−1) = (r(γ), a) = r(a, γ, ac(γ)). Our key tool is the following observation. Proposition 4.3. Let G, Q, c and Y = s(G) be as in the statement of Theorem 4.2. Then the Borel groupoids G(c)Y and c−1(e) are Borel equivalent. Remark 4.4. The proof is complicated by the fact that we are not assuming G is Hausdorff, nor is it clear whether c−1(e) has a Haar system. Thus we cannot apply [1, Corollary 2.1.17] to establish that the actions are proper. Instead, we resort to the definition. Proof. We want to show that G is a (c−1(e), G(c)Y )-Borel equivalence. We let c−1(e) act on the left by multiplication in G. (Hence the moment map for the left action is just r.) For the right action of G(c)Y we use the moment map s. Thus η · (a, γ, ac(γ)) makes sense only when (η, γ) ∈ G(2) and c(η) = a. Then we define (4.1) η · (c(η), γ, c(ηγ)) = ηγ. To see that G is a proper right G(c)Y -space, we need to see that the transfor- mation groupoid G ⋊ G(c)Y for the right G(c)Y -action is proper. But the map (η, (c(η), γ, c(ηγ)) 7→ (η, γ) is clearly a bijection of G ⋊ G(c)Y onto G ⋊ G. It is easy to see that it is a groupoid homomorphism as well. Since the right action of any Borel groupoid on itself is proper (by a left to right modification of [1, Example 2.1.4(1)]), it follows that G is a proper G(c)Y -space. The action is clearly free. To see that the left action is proper in the Borel sense, we first note that it is certainly free and proper topologically. Hence the orbit space is itself a locally AMENABILITY OF GROUPOIDS 11 Hausdorff, locally compact space by [13, Lemma 2.6]. Hence it is a standard Borel space with the Borel structure coming from its topology. Let q : G → c−1(e)\G be the quotient map. Since q is continuous and c−1(e) is closed, the inverse image of points in c−1(e)\G under q are closed. Since open sets in G are σ-compact, the forward image of open sets are σ-compact. Since a compact subset of a locally Hausdorff, locally compact space is Borel, so is the forward image of any open set. Hence q has a Borel cross section w by [2, Theorem 3.4.1]. A priori c−1(e)\G has two Borel structures: the one we've been using coming from the quotient topology, and also the quotient Borel structure. The quotient map q is Borel in both cases and w must also be a cross section for the potentially finer quotient Borel structure. Hence both spaces are Borel isomorphic to the image of w by [2, Proposition 3.4.2]. Hence the two Borel structures coincide on the orbit space. Therefore Lemma 3.4 applies and G is a free and proper Borel c−1(e)-space as required. If η ∈ c−1(e), then s(ηγ) = (s(γ), c(ηγ)) = (s(γ), c(γ)). It follows that s factors through G/c−1(e). On the other hand, suppose that s(γ) = (s(γ), c(γ)) = s(η) = (s(η), c(η)). Then c(γ) = c(η) and γ−1η ∈ c−1(e). Of course γ · γ−1η = η. Thus s induces a bijection of G/c−1(e) with Y . Since c−1(e) is closed, it acts freely and properly on G. Thus the quotient G/c−1(e) is a standard Borel space (with respect to the quotient Borel structure) and s induces a Borel bijection of G/c−1(e) with the analytic Borel space Y . Hence s is a Borel isomorphism by [2, Corollary 2 of Theorem 3.3.4]. It is clear from (4.1) that r = r factors through G/G(c)Y . In fact, the homeo- morphism of G ⋊ G(c)Y with G ⋊ G induces a homemorphism of G/G(c)Y onto G/G ∼= G(0). Hence r certainly induces a Borel isomorphism as required. (cid:3) Proof of Theorem 4.2. If Y is open, then {q · Y }q∈Q is an open cover of the second countable set G(0) × Q. By a theorem of Lindelof, it has a countable subcover; so it suffices to prove the first statement. Since G is σ-compact and s is continuous, Y is also σ-compact and hence Borel. This ensures that G(c)Y is standard as a Borel space. Since c−1(e) is assumed to be amenable, it is Borel amenable. Hence G(c)Y is Borel amenable by Proposition 4.3. On the other hand G(c)qn·Y = qn ·(cid:0)GY(cid:1). Hence each G(c)qn·Y is Borel amenable. Since Q acts by automorphisms, each qn·Y is invariant. Now G(c) is Borel amenable from Lemma 3.3. Since G(c) has a Haar system if G does, it follows from [19, Corollary 2.15] that G is amenable. The last assertion follows from Proposition 4.1. (cid:3) As special cases of Theorem 4.2, we obtain the cocycle results from [1] and [22] mentioned in the introduction. The following result strengthens [22, Theorem 9.3] by removing the hypotheses that G be ´etale. Corollary 4.5. Suppose that G is a locally Hausdorff, locally compact second count- able groupoid with a Haar system and that c : G → Q is a continuous homomor- phism into a discrete group Q such that c−1(e) is amenable. Then G(c) is amenable. If Q is amenable, then so is G. Proof. It suffices to see that Y = s(G) is open in G(0) × Q. But s is open and s(G) = [a∈Q s(c−1(a)) × {a}, 12 JEAN N. RENAULT AND DANA P. WILLIAMS which is open since each c−1(a) is open. (cid:3) The result [1, Theorem 5.3.14] mentioned in the introduction establishes the amenability of a measured groupoid G which admits a strongly surjective Borel homomorphism onto a Borel groupoid with amenable kernel and range. The fol- lowing result is similar, but concerns topological instead of measure amenability. Moreover, it is more restrictive since it assumes that the range is a group rather than a groupoid. Corollary 4.6. Suppose that G is a locally Hausdorff, locally compact second count- able groupoid with a Haar system and that c : G → Q is a strongly surjective con- tinuous homomorphism into a group Q such that c−1(e) is amenable. Then G(c) is amenable. If Q is amenable, then so is G. Proof. Since c is strongly surjective, we must have c(Gu) = Q for all u. It follows that s is surjective, and the result follows from Theorem 4.2. (cid:3) 5. Application to Semigroup Actions Our results apply nicely to semigroup action groupoids. Definition 5.1. Let X be a set and P be a semigroup with identity e. A right (partial) action of P on X consists of a subset X ∗ P of X × P and a map T : X ∗ P → X sending (x, m) to x · m, such that (a) for all x ∈ X, (x, e) ∈ X ∗ P and x · e = x; (b) for all (x, m, n) ∈ X × P × P , (x, mn) ∈ X ∗ P if and only if (x, m) ∈ X ∗ P and (x · m, n) ∈ X ∗ P ; if this holds, (x · m) · n = x · (mn). For all m ∈ P , we define U (m) = {x : (x, m) ∈ X ∗ P } and V (m) = {x·m : (x, m) ∈ X ∗ P } and Tm : U (m) → V (m) such that Tmx = x · m. The triple (X, P, T ) will be called a semigroup action. This notion generalizes that of singly generated dynamical system (SGDS) in the sense of [18], which is the case P = N. A SGDS is given by a single map T from a subset dom(T ) of X to another subset ran(T ) of X. For n ∈ N, we let U (n) = dom(T n) and define x · n = T n(x) for x ∈ U (n). Let us define the binary relation on P : m ≤ m′ if and only if there exists n ∈ P such that m′ = mn. Our axioms imply that m ≤ m′ ⇒ U (m′) ⊂ U (m). The next notion is important because it will allow a simple construction of a groupoid from a semigroup action. Definition 5.2. Let us say that a semigroup action (X, P, T ) is directed if for all pairs (m, n) ∈ P × P such that U (m) ∩ U (n) is non-empty, there exists an upper bound r ∈ P of m and n such that U (m) ∩ U (n) = U (r). Note that the equality can be replaced by the inclusion U (m) ∩ U (n) ⊂ U (r). Here are two rather different situations where the action is directed. Example 5.3. Assume that the action is everywhere defined, i.e. X ∗ P = X × P . Then the action is directed if and only if the semigroup P is directed, in the sense that for all m, n ∈ P , there exists r ∈ P such that m, n ≤ r; we then say that r is AMENABILITY OF GROUPOIDS 13 a common upper bound, or c.u.b. for short, of m and n. If P is a subsemigroup of a group Q, meaning that (5.1) P ⊂ Q, P P ⊂ P and e ∈ P. then, P is directed if and and only if it satisfies the Ore condition P −1P ⊂ P P −1. In such a case, P is called an Ore semigroup. Example 5.4. Assume that P is quasi-lattice ordered. This means that P is a subsemigroup of a group Q such that P ∩ P −1 = {e} and whenever two elements m, n ∈ P have a c.u.b. in P , they have a least upper bound, or l.u.b. for short, m ∨ n (the original definition of A. Nica in [14] contains the additional assumption that every element of Q which has an upper bound in P has a least upper bound in P ). Then, the action of P on X = P given by U (n) = { x ∈ P : n ≤ x } and x · n = n−1x is directed. If (X, P, T ) is directed, then the relation x ∼ y if and only if there exist (m, n) ∈ P × P such that x · m = y · n is transitive and hence an equivalence relation. We denote by [x] the equivalence class of x. Given P a subsemigroup of a group Q and a directed semigroup action (X, P, T ), we define G(X, P, T ) as the set of triples (x, q, y) in X × Q × X such that there exist m, n ∈ P with q = mn−1, x ∈ U (m), y ∈ U (n) and x · m = y · n. Lemma 5.5. Assume that P is a subsemigroup of a group Q and that (X, P, T ) is a directed action. Then G(X, P, T ) is a subgroupoid of X × Q × X, equipped with its natural structure of a groupoid over X. Proof. Suppose that (x, q, y) belongs to G: there exist m, n ∈ P such that q = mn−1 and x · m = y · n. Then (y, q−1, x) also belongs to G because q−1 = nm−1 and y · n = x · m. Suppose that (x, s, y) and (y, t, z) belong to G: there exist m, n, p, q ∈ P such that s = mn−1, t = pq−1, x · m = y · n and y · p = z · q. Since the action is directed, there exist (a, b) ∈ P × P such that na = pb. Then x · ma = y · na = y · pb = z · qb. Since (ma)(qb)−1 = (mn−1)(pq−1), (x, st, y) belongs to G. (cid:3) Definition 5.6. Let (X, P, T ) be a directed semigroup action, where P is a sub- semigroup of a group Q. The groupoid G(X, P, T ) associated with (X, P, T ) is called the semidirect product groupoid (or semidirect product for short) of the action. It carries the canonical cocycle c : G(X, P, T ) → Q given by c(x, q, y) = q. Remark 5.7. When X is reduced to a point, the lemma says that P P −1 is a group if P satisfies the Ore condition P −1P ⊂ P P −1; clearly, this condition is also necessary. We next make the following topological assumptions: Definition 5.8. We shall say that a semigroup action (X, P, T ), where P is a subsemigroup of a group Q, is locally compact if (a) X is a locally compact Hausdorff space; (b) Q is a discrete group; (c) for all m ∈ P , U (m) and V (m) are open subsets of X and Tm : U (m) → V (m) is a local homeomorphism. 14 JEAN N. RENAULT AND DANA P. WILLIAMS In the following, we always assume that the semigroup action (X, P, T ) is locally compact. Given m, n ∈ P and A, B subsets of X, we define Z(A, m, n, B) := { (x, mn−1, y) ∈ G : x ∈ A, y ∈ B and x · m = y · n } Let B be the family of subsets Z(U, m, n, V ), where U and V are open subsets of X. Lemma 5.9. The family B is a base for a topology T on G. Proof. First, it is clear that this family covers G. Second, let (x, q, y) be a point in the intersection of Z(U, m, n, V ) and Z(U ′, m′, n′, V ′). We are going to find Z(U ′′, m′′, n′′, V ′′) containing the point (x, q, y) and contained in the intersection. There exists (a, a′) ∈ P × P such that ma = m′a′ is an upper bound r of m and m′ and U (r) = U (m) ∩ U (m′). Then na = n′a′ is an upper bound s of n and n′ and U (s) = U (n) ∩ U (n′). Let W be an open neighborhood of x · m = y · n on which Ta is injective. Similarly, let W ′ be an open neighborhood of x · m′ = y · n′ m′ (W ′), V ′′ = on which Ta′ V ∩ V ′ ∩ T −1 n′ (W ′), m′′ = r and n′′ = s. Since x ∈ U ′′, y ∈ V ′′ and x · r = y · s, (x, q, y) belongs to Z(U ′′, m′′, n′′, V ′′). If (x′, q, y′) belongs to Z(U ′′, m′′, n′′, V ′′), x′ ∈ U ∩ U ′, y′ ∈ V ∩ V ′ and x′ · ma = y′ · na. Since x′ · m and y′ · n belong to W on which Ta is injective, x′ · m = y′ · n. Therefore (x′, q, y′) belongs to Z(U, m, n, V ). Similarly, the equality x′ · m′a′ = y′ · n′a′ implies the equality x′ · m′ = y′ · n′ because x′ · m′ and y′ · n′ belong to W ′ on which Ta′ is injective. (cid:3) is injective. We let U ′′ = U ∩ U ′ ∩ T −1 m (W ) ∩ T −1 n (W ) ∩ T −1 Remark 5.10. The sets Z(U, m, n, V ), where U and V are open subsets of X such that U ⊂ U (m), V ⊂ U (n) and TmU and TnV injective form a subbase of B. Lemma 5.11. The topology T of G is finer than the product topology of X × Q× X but it agrees with it on the sets Z(A, m, n, B), where A, B are subsets of X and m, n ∈ P . Proof. The intersection of a rectangle U × {q} × V with G, where U, V are open subsets of X is a union of elements of B. Therefore, the topology T is finer than the product topology. Let A, B be subsets of X and m, n ∈ P . Let (x, q, y) be a point of Z(A, m, n, B). Let Z(U, m′, n′, V ) ∩ Z(A, m, n, B) be a basic neighborhood of (x, q, y) in Z(A, m, n, B). Just as in the proof of the previous lemma, we introduce (a, a′) ∈ P × P such that ma = m′a′ (denoted by r) and U (r) = U (m) ∩ U (m′), s = na = n′a′, and we choose and open neighborhood W of x·m = y ·n on which Ta is injective, and an open neighborhood W ′ of x · m′ = y · n′ on which Ta′ is injective. We define U ′ = U ∩T −1 n′ (W ′). Suppose that (x′, q, y′) belongs to (U ′ × {q} × V ′) ∩ Z(A, m, n, B). Then x′ · m′a′ = x′ · ma = y′ ·na = y′ ·n′a′. We deduce as before that x′m′ = y′n′. Therefore (x′, q, y′) belongs to Z(U, m′, n′, V ). (cid:3) m′ (W ′) and V ′ = V ∩T −1 n (W )∩T −1 m (W )∩T −1 Proposition 5.12. Let (X, P, T ) be a semigroup action as above. We endow G(X, P, T ) with the topology T . Then, (a) G(X, P, T ) is an ´etale locally compact Hausdorff groupoid; (b) the canonical cocycle c : G(X, P, T ) → Q is continuous. AMENABILITY OF GROUPOIDS 15 Proof. The topology T is Hausdorff because it is finer than the product topology. Let (x, q, y) ∈ G. Pick (m, n) ∈ P × P such that q = mn−1 and x · m = y · n. Let A, B be compact neighborhoods of x, y contained respectively in U (m) and in U (n). Then Z(A, m, n, B) is a compact neighborhood of (x, q, y), and T is a locally compact topology. The injection map i(x) = (x, e, x) is a homeomorphism from X onto G(0). 7→ (y, q−1, x) transforms Z(U, m, n, V ) it is a homeomorphism. The inverse map (x, q, y) into Z(V, n, m, U ). Therefore, Suppose that (xα, sα, yα) converges to (x, s, y) and (yα, tα, zα) converges to (y, t, z). Pick basic open sets Z(U, m, n, V ) and Z(V ′, p, q, W ) containining respectively (x, s, y) and (y, t, z). By definition, for α large enough (xα, sα, yα) belongs to Z(U, m, n, V ) and (yα, tα, zα) belongs to Z(V ′, p, q, W ). As in Lemma 5.11, pick (a, b) ∈ P × P such that na = pb. Then (x, st, z) belongs to Z(U, ma, qb, W ) and so does (xα, sαtα, zα) for α large enough. We apply Lemma 5.11 to conclude that (xα, sαtα, zα) converges to (x, st, z). Thus G is a topological locally compact Hausdorff groupoid. A basis element S = Z(U, m, n, V ) such that Tm is injective on U and Tn is injective on V is a bisection, in the sense that the restrictions of the range and source maps rS and sS are homeomorphisms onto open subsets of X. Since G admits a cover of open bisections, it is an ´etale groupoid. Since the canonical cocycle is continuous with respect to the product topology, (cid:3) it is continuous with respect to T . Here is our main application of our Theorem 4.2 (in the form of Corollary 4.5); it gives the amenability of the semidirect product by a subsemigroup of an amenable group. Theorem 5.13. Let (X, P, T ) be a directed locally compact semigroup action. As- sume that P is a subsemigroup of a countable amenable group Q. Then the semi- direct product groupoid G(X, P, T ) is topologically amenable. Proof. To prove this result, we apply Corollary 4.5 to the continuous cocycle c : G(X, P, T ) → Q. The only missing point is the amenability of the equivalence relation R = c−1(e). As is standard, we try to write R as the increasing union of well-behaved equivalence relations (for example, see [21, Lemma 3.5]). To see how to do this, we call a subset F ⊂ P action-directed if e ∈ F and given n, m ∈ F with U (n) ∩ U (m) 6= ∅, then there is an r ∈ F dominating n and m such that U (r) = U (n) ∩ U (m). For example, by hypotheses, P itself is action-directed. More to the point, if F is action-directed, then RF = { (x, y) : there is a m ∈ F such that x · m = y · m } is a Borel equivalence relation: relation since F is action-directed. We just need to specify suitable sets F . it is an Fσ subset of X × X and an equivalence Let F be the collection of finite subsets F of P such that Tm∈F U (m) 6= ∅. A a simple induction argument implies the following. Claim 1. If F ∈ F , then there is an r ∈ P such that n ≤ r for all n ∈ F and U (r) = Tn∈F U (n). Claim 2. There is a map F 7→ rF from F to P such that r∅ = e, r{n} = n, and such that (a) n ≤ rF for all n ∈ F , 16 JEAN N. RENAULT AND DANA P. WILLIAMS (b) U (rF ) = Tn∈F U (n), and (c) F ′ ⊂ F implies rF ′ ≤ rF . Proof of Claim 2. We start by defining r∅ and r{n} as above. Suppose that we have defined rF for all F ∈ F with F ≤ k for some k ≥ 1 such that (a), (b) and (c) hold. Let us define rF for F ∈ F with k + 1 elements. Note that F ′ = { S ⊂ F : S = k } is a subset of F . By Claim 1, we can define rF ∈ P so that rS ≤ rF for all S ∈ F ′ and such that U (rF ) = TS∈F ′ U (rS). Now consider the set { rF : F ∈ F and F ≤ k + 1 }. Then (a) and (b) hold by assumption if F ≤ k. But if F = k + 1 and n ∈ F , then there is an S ∈ F ′ such that n ∈ S. Hence n ≤ rS ≤ rF . Similarly, U (rF ) = \S∈F ′ U (rS) = \S∈F ′ \n∈S U (n) = \n∈F U (n). Hence (a) and (b) hold for sets of k + 1 or fewer elements. Now suppose F ′ is a proper subset of F . We have rF ′ ≤ rF by assumption if F ≤ k. If F = k + 1, then there exists S ⊂ F with S = k. Then rF ′ ≤ rS ≤ rF . (cid:3) As pointed out by the referee, the possibility that F might be directed is sug- gested by [10] where the authors work with quasi-lattice ordered subgroups.That the same is true for directed actions is implied by the following claim. Claim 3. Every finite subset S ⊂ P is contained in a finite action-directed subset F . Proof of Claim 3. Let F 7→ rF be the map defined in Claim 2. Define F = { rS ′ : S′ ⊂ S and S′ ∈ F }. Clearly e ∈ F and S ⊂ F . We claim that F is action-directed. Let k, l ∈ F be such that U (k) ∩ U (l) 6= ∅. We can assume that k = rS ′ and l = rS ′′ for appropriate subsets of S. Then F ′ = S′ ∪ S′′ satisfies Tn∈F ′ U (n) = U (k) ∩ U (l) 6= ∅. But then rF ′ ∈ F . Since F 7→ rF is monotonic, we have rS ′ ≤ rF ′ and rS ′′ ≤ rF ′ . This completes the proof of the claim. (cid:3) Claim 4. There is a sequence (Fi) of finite action-directed sets such that Fi ⊂ Fi+1 and such that c−1(e) = [i RFi . Proof of Claim 4. Let P = {p1, p2, . . . } with p1 = e. Now we can employ Claim 3 to inductively construct the Fi where Fi+1 is an action-directed set containing Fi and pi+1. (cid:3) The key observation is that if F is finite and action-directed, then RF is proper. To see this, note that the equivalence class [x]F is the finite union over m ∈ F of the fibres T −1 m (x · m). Since Tm is a local homeomorphism, the fibres are discrete. Hence each orbit is discrete and therefore locally closed. The Mackey-Glimm- Ramsay dichotomy [16] then implies that the orbit space X/RF is a standard Borel space. Then [1, Example 2.1.4(2)] implies that RF is a proper Borel groupoid. AMENABILITY OF GROUPOIDS 17 It follows from Claim 4 and Lemma 2.8 that c−1(e) is Borel amenable. Since it is open in G(X, P, T ), it is ´etale. Hence it is amenable by [19, Corollary 2.15]. (cid:3) Remark 5.14. It is well known that there are interesting amenable actions (in the sense that the semi-direct product groupoid of the action is amenable) of non- amenable groups. We will encounter in the next section an amenable action of a free semigroup (appearing in the work of Nica [14] on the Wiener-Hopf algebra of a semigroup). In this case, the amenability of the action cannot be deduced from the above theorem as the free group is not amenable. 6. Application to Topological Higher Rank Graphs Higher-rank graphs provide interesting semigroup actions which generalize one- sided subshifts of finite type. We recall some definitions but refer to [26] for a complete exposition. We introduce two changes with respect to [26]. First, we define P -graphs for an arbitrary subsemigroup P of a group Q while Yeend considers the case P = Nd ⊂ Q = Zd. In order to develop the theory smoothly, we shall often need the assumption that P is quasi-lattice ordered, as defined in Example 5.4. Such P - graphs have already been introduced in [4]. Second, we construct the path space Ω as a closure in the space of closed subsets of the P -graph with respect to the Fell topology. The use of directed hereditary subsets to construct the path space goes back to [14] and is present in [4, 8, 22]. Here are our definitions. A small category Λ is given by its set of arrows Λ(1) (usually denoted by Λ), its set of vertices Λ(0) (viewed as a subset of Λ through the identity map i : Λ(0) → Λ), range and source maps r, s : Λ → Λ(0) and composition map ◦ : Λ(2) → Λ where Λ(2) is the set of composable pairs of arrows, i.e., (λ, µ) ∈ Λ × Λ such that s(λ) = r(µ). Given A, B ⊂ Λ, we write A ∗ B = (A × B) ∩ Λ(2). We make the following topological assumptions. (a) Λ and Λ(0) are locally compact Hausdorff spaces; (b) r, s : Λ → Λ(0) are continuous and s is a local homeomorphism; (c) i : Λ(0) → Λ is continuous; (d) composition ◦ : Λ(2) → Λ is continuous and open. The following definition is a topological version of [4, Definition 2.1]. Definition 6.1. Let P be a semigroup with unit element e. A higher-rank topo- logical graph graded by P , or P -graph for short, is a topological small category Λ as above endowed with a map, called the degree map, d : Λ → P which satisfies the following properties (a) the degree map d : Λ → P is continuous (where P has the discrete topology); (b) for all (µ, ν) ∈ Λ(2), d(µν) = d(µ)d(ν) and for all v ∈ Λ(0), d(v) = e; (c) it has the unique factorization property: for all m, n ∈ P , the composition map Λm ∗ Λn → Λmn is a homeomorphism. As a basic example of P -graph, we consider the graph of a semigroup action. If (X, P, T ) is a locally compact semigroup action as in the previous section, set Λ(0) = X, Λ = X ∗ P, r(x, n) = x, s(x, n) = x · n and d(x, n) = n. Composition is necessarily given by (x, m)(x · m, n) = (x, mn). It results from our axioms of a semigroup action that Λ is a P -graph which we call the graph of the 18 JEAN N. RENAULT AND DANA P. WILLIAMS action. For example, if (X, T ) is a singly generated dynamical system as defined in the previous section, it is useful to think of (x, n) ∈ Λ = X ∗ N as a finite path (x, T x, . . . , T n−1x) when n ≥ 1. More generally, we think of (x, n) ∈ X ∗ P as a finite path { (x · m) : m ≤ n }, although this latter set need not be finite. Of course, not all P -graphs arise as graphs of a semigroup action. The topological N-graphs are exactly the usual topological graphs. A topological graph is given by a pair of locally compact Hausdorff spaces (E, V ) with two maps r, s : E → V , r continuous and s local homeomorphism. The space Λ of finite paths is the disjoint union over N of the spaces E(n) of paths of length n, where E(0) = V , E(1) = E and E(n) = { e1e2 . . . en : ei ∈ E, s(ei) = r(ei+1) for i = 1, . . . , n − 1 }. endowed with the product topology, for n ≥ 2. It is a topological category with Λ(0) = V , the obvious range and source maps and composition given by concate- nation. It has an obvious degree map d : Λ → N where d(λ) = n if and only if λ ∈ E(n). This definition includes the graphs which appear in the theory of graph C-algebras, which is the case when E and V are discrete spaces. A singly gener- ated dynamical system (X, T ) can be viewed as a topological graph with V = X, E = dom(T ), r(x) = x and s(x) = T x. Its space of finite paths Λ agrees with X ∗ N. As another example of topological graph, consider (E = T, V = T) where T is the circle z = 1 and the range and source maps are respectively z 7→ z2 and z 7→ z3. Topological graphs where the range and source maps are both local homeomorphisms are called polymorphisms in [3]. By analogy with the above examples, the elements of a higher-rank graph Λ are called finite paths. We define µ ≤ λ if there exists ν such that λ = µν. This is a pre- order relation which shares some of the properties of the pre-order relation we have defined on the semigroup P in the previous section. In particular, suppose that λ and µ have a common upper bound ν. Then d(λ) and d(µ) have d(ν) as a common upper bound. If P is quasi-lattice ordered, d(λ) and d(µ) have a least common upper bound p. Therefore, there exists ν′ ≤ ν c.u.b. of λ and µ with d(ν′) = p. We say that ν′ is a l.u.b. of λ and µ. Such a l.u.b. need not be unique. Given A, B subsets of Λ, we denote by A ∨ B the set of elements which are l.u.b. of some pair (λ, µ) ∈ A × B. Given µ ≤ λ, we define the segment [µ, λ] := {ν : µ ≤ ν ≤ λ}. We shall use the following notation. If λ ∈ Λ and n ∈ P are such that n ≤ d(λ), then λ can be written uniquely λ = µν where d(µ) = n. Then we define λ · n := ν. Conversely, given µ ∈ Λ, we can define µν for all ν ∈ r−1(s(µ)). We shall need some further assumptions on our higher-rank graphs. Definition 6.2. One says that the P -graph Λ is (a) (r, d)-proper if the map (r, d) : Λ → Λ(0) × P is proper; (b) compactly aligned if P is quasi-lattice ordered and for all compact subsets A, B ⊂ Λ, the subset A ∨ B is compact. In the setting of a discrete N-graph, condition (a) means that a vertex emits finitely many edges while condition (b) is always satisfied. The graph of the action of a semigroup always satisfies (a) since (r, d) is the injection of Λ = X ∗ P into X ×P . When P is quasi-lattice ordered, condition (a) implies condition (b). Indeed, if ν belongs to A ∨ B where A, B are subsets of Λ, then r(ν) belongs to r(A) ∩ r(B) and d(ν) belongs to d(A) ∨ d(B). If A and B are compact, r(A) ∩ r(B) is compact and d(A) ∨ d(B) is finite. AMENABILITY OF GROUPOIDS 19 Let Λ be a P -graph. Set Λ ∗ P = { (λ, m) ∈ Λ × P : m ≤ d(λ) } and define T : Λ ∗ P → Λ by T (λ, m) := λ · m = ν if d(λ) = mn and λ = µν with d(µ) = m and d(ν) = n. Proposition 6.3. Let Λ be a P -graph. Define T as above. Then T is a directed action of P on Λ by partial local homeomorphisms. Proof. The domain of Tm is the open set U (m) = {λ ∈ Λ : m ≤ d(λ)}. Its range V (m) = { ν ∈ Λ : there exists µ ∈ Λm such that (µ, ν) ∈ Λ(2) } = r−1(s(Λm)) is also open because s is open. Let us show that Tm is continuous and open. Since U (m) is the union of the open subsets Λma when a runs over P , it suffices to study its restriction to Λma. This restriction factors as Λma → Λm ∗ Λa → Λ The first map is a homeomorphism by assumption. The second map is the re- striction of the projection onto the second factor, and is open because s is open. Therefore the composition is continuous and open. Moreover s : Λ → Λ(0) is a local homeomorphism. On an open subset U of Λ on which s is injective, TmU : U (m) ∩ U → Λ is a homeomorphism onto an open subset. To see that the action is directed, suppose that U (m) ∩ U (n) 6= ∅. Then m, n ≤ (cid:3) d(λ) for some λ ∈ U (m) ∩ U (n). Then U (m) ∩ U (n) ⊂ U (d(λ)). If P is a subsemigroup of a discrete group Q and Λ is a P -graph, then we can form the semi-direct product groupoid G(Λ, P, T ). This groupoid is proper, which means that the map (r, s) : G(Λ, P, T ) → Λ×Λ is proper. The fine structure of such groupoids can be interesting (in the group case, see for example [6,7]). In the theory of graph algebras, Λ is the space of finite paths. It is fruitful (and necessary) to construct a larger space, which is called the path space and which includes infinite paths. This is what we do in the next subsection. 6.1. The path space Ω. Just as in the case of a graph, we want to define a space of paths, both finite and infinite. Our construction is directly inspired by a construction of A. Nica in [14] which we recall below. In fact, our construction agrees with Nica's when Λ = X ∗ P and when X is reduced to one point. The same idea of defining paths as hereditary directed subsets of the graph appears also in [4, Section 3]. Let us first summarize the exposition given in [14]. There P is a subsemigroup of a discrete group Q and it is assumed to be quasi-lattice ordered. One embeds P into the space {0, 1}P of all subsets of P endowed with the product topology by sending m ∈ P to the segment j(m) = [e, m]. The Wiener-Hopf closure of P is the closure of j(P ) in {0, 1}P . It is denoted by Ω(P ). Nica remarks that the elements of Ω(P ) are exactly the non-empty hereditary and directed subsets of P . As we will see below, a similar construction can be employed to define a closure Ω of a topological P -graph Λ. 20 JEAN N. RENAULT AND DANA P. WILLIAMS Recall that we have defined on Λ the pre-order µ ≤ λ if there exists ν such that λ = µν. Note that this implies that r(µ) = r(λ) and d(µ) ≤ d(λ). Following [14], we call a subset A of Λ hereditary if µ ≤ λ ∈ A implies µ ∈ A, and directed if any two elements of A have a c.u.b. in A. Hereditary and directed subsets are called filters in [4, 8]. Our terminology is the same as in [22]. We denote by Ω the set of all hereditary and directed closed subsets of Λ and by Ω the set of all non-empty hereditary and directed closed subsets of Λ. We view Ω and Ω as subsets of the space C(Λ) of all closed subsets of Λ endowed with the Fell topology [25, §H.1]. Recall that a basis for the Fell topology on C(Λ) is given by sets of the form U(K; U1, . . . , Um) = { F ∈ C(Λ) : F ∩ K = ∅ and F ∩ Ui 6= ∅ } where K ⊂ Λ is compact and each Ui ⊂ Λ is open. Then as in [25, Lemma H.2]5 a net (Fβ ) converges to F in C(Λ) if and only if every subnet (Fi) of (Fβ ) is such that (F1) given λi ∈ Fi such that λi → λ, then λ ∈ F , and (F2) if λ ∈ F , then there is a subnet (Fij ) and λj ∈ Fij such that λj → λ. Lemma 6.4. (a) Let A be a non-empty directed subset of Λ. Then A is con- tained in xΛ := r−1(x) for some (necessarily unique) x ∈ Λ(0), which will be written r(A); (b) the map r : Ω → Λ(0) is continuous. Proof. (a) By definition, the relation µ ≤ λ implies that r(µ) = r(λ). Let µ, ν ∈ A. Since there exists λ such that µ, ν ≤ λ, we must have r(µ) = r(ν). (b) Suppose that Aα, A ∈ Ω and Aα tends to A. Let xα = r(Aα) and x = r(A). Let U be an open neighborhood of x. Since A ∩ r−1(U ) 6= ∅, there exists α0 such that Aα ∩ r−1(U ) 6= ∅ for all α ≥ α0. Then, xα ∈ U . (cid:3) Lemma 6.5. Assume (r, d) is proper and that P is contained in a group Q. If λi converges to λ, µi converges to µ and for all i, µi ≤ λi, then µ ≤ λ. Proof. There exists a net νi such that λi = µiνi. Since d(λi) = d(µi)d(νi), d(νi) is eventually constant. Since r(νi) = s(µi), it is contained in some compact subset of Λ(0). Because of (r, d)-properness, there is a subnet (νi) converging to some ν. Then λ = µν. (cid:3) Lemma 6.6. Let A be a subset of Λ. (a) If A is directed (resp., hereditary), then d(A) is directed (resp., hereditary). (b) If A is directed, the restriction to A of the degree map dA : A → P is a bijection onto d(A). Proof. Let m, n ∈ d(A). There exist µ, ν ∈ A such that m = d(µ) and n = d(ν). If A is directed, there exists λ ∈ A such that µ, ν ≤ λ. Then m, n ≤ d(λ). Therefore d(A) is directed. Suppose that m ≤ n and that n = d(λ) with λ ∈ A. We write n = mp. By unique factorization, we can write λ = µπ, where d(µ) = m and If A is hereditary, then µ ∈ A and m ∈ d(A). Therefore, d(A) is d(π) = p. hereditary. Suppose that A is directed. Let µ, ν ∈ A such that d(µ) = d(ν). Let λ be a (cid:3) c.u.b. of (µ, ν). By unique factorization of λ, we have the equality µ = ν. 5Unfortunately, the statement of [25, Lemma H.2] omits the observation that (F1) and (F2) must hold for every subnet of the original net. AMENABILITY OF GROUPOIDS 21 Lemma 6.7. Let Λ be a P -graph. (a) If A ⊂ Λ is hereditary (resp., directed) and n ∈ P , the subset A · n = { ν ∈ Λ : there exists µ ∈ Λn such that µν ∈ A } is hereditary (resp., directed). If (r, d) is proper and A is directed and closed, then A · n is directed and closed. (b) Assume (r, d) is proper and that P is a subsemigroup of a group Q. If B ⊂ Λ is directed, hereditary and closed and if µ ∈ r(B)Λ, then µB = [ν∈B { λ ∈ Λ : λ ≤ µν } is directed, hereditary and closed. Moreover, if A · n is non-empty, there is a unique µ ∈ Λn ∩ A such that A = µ(A · n). Proof. (a) Suppose that A is hereditary . Let b ≤ (a · n) where a ∈ A. Then, a = µ(a · n) where d(µ) = n. Moreover, a · n = b · c for some c ∈ P . Thus, a = µ(b · c) = (µb) · c, and µb ≤ a. Since A is hereditary, µb ∈ A, hence b ∈ A · n. We have shown that A · n is hereditary. Suppose that A is directed. Consider a · n and b · n where a, b ∈ A. We want a c.u.b. for a · n, b · n in A · n. Let c be a c.u.b. for a, b in A. Then c · n is defined and is a c.u.b. for a · n and b · n. Suppose that A is directed and closed, that ai belongs to A and that ai · n converges to some b ∈ Λ. We write ai = µi(ai · n), where d(µi) = n. Since we also have r(µi) = r(A), by (r, d)-properness, there is a subnet (µj) converging to some µ such that r(µ) = r(a), s(µ) = r(b) and d(µ) = n Then aj converges to a = µ(b). Since A is closed, a belongs to A and b = a · n belongs to A · n. (b) To see that µB is directed, suppose that λ ∈ B is a c.u.b. for ν1, ν2 ∈ B. Then µλ is a c.u.b. for µν1, µν2. The set µB is hereditary by construction. To see that it is closed, consider a net λα in µB converging to λ. We distinguish two cases: if d(λ) ≤ d(µ), then d(λα) ≤ d(µ) for α large enough. This implies λα ≤ µ, hence λ ≤ µ. If d(λ) ≤ d(µ) does not hold, there is a subnet λβ for which d(λβ ) ≤ d(µ) does not hold. Then, we can write λβ = µνβ with νβ ∈ B. Since d(λβ ) = d(λ) for β large enough, this fixes d(νβ) (if P ⊂ Q). We also have r(νβ ) = s(µ). By (r, d)-properness, there is a converging subnet νγ. Its limit ν belongs to B because B is closed. We have λ = µν, hence λ is in µB. If A · n is non-empty, there is λ ∈ A such that d(λ) ≥ n. We write λ = µν with µ ∈ Λn. Since µ ≤ λ, µ belongs to A. Let λ′ be another element of A such that d(λ′) ≥ n. We can write λ′ = µ′ν′. The unique factorization of a c.u.b. of (λ, λ′) gives µ = µ′. Let us compare A and µ(A · n). If λ ∈ A, the existence of a c.u.b. for (λ, µ) shows that λ ∈ µ(A · n). Conversely, suppose that λ ≤ µν for ν ∈ A · n. There exists µ′ ∈ Λn such that µ′ν ∈ A. By the above, µ′ = µ, therefore µν ∈ A, hence λ ∈ A. (cid:3) Proposition 6.8. Let Λ be a P -graph, where P is a quasi-lattice ordered subsemi- group of a group Q and Λ is (r, d)-proper. (a) the set Ω of all hereditary and directed closed subsets of Λ is a closed subset of the space C(Λ) of closed subsets of Λ equipped with the Fell topology. (b) for all λ ∈ Λ, F (λ) := {µ ∈ Λ : µ ≤ λ} belongs to Ω; (c) F (Λ) is dense in Ω; (d) if n ≤ d(λ), F (λ) · n = F (λ · n); if s(µ) = r(λ), µF (λ) = F (µλ); 22 JEAN N. RENAULT AND DANA P. WILLIAMS (e) the map F : Λ → C(Λ) is injective and continuous. Proof. (a) Suppose that Aα converges to A in C(Λ) and that the Aα's are hereditary and directed. Let us show that A is hereditary. Let λ ∈ A and µ ≤ λ. Suppose that µ /∈ A. Since A is closed, there is an open set U and a compact set K such that µ ∈ U ⊂ K ⊂ Λ \ A. The set U Λ = { αβ : α ∈ U and β ∈ Λ } is open (because multiplication is assumed to open and continuous), and contains λ. Thus we have A ∩ K = ∅ and A ∩ U P 6= ∅. There is α0 such that for α ≥ α0, Aα ∩ K = ∅ and Aα ∩ U Λ 6= ∅. If λα belongs to Aα ∩ U Λ, there exists µα ∈ U such that µα ≤ λα. Since Aα is hereditary, µα belongs to Aα. This contradicts Aα ∩ K = ∅. Let us show that A is directed. Let µ, ν be in A. There exist nets µα and να in Aα converging respectively to µ and ν. Let λα be a l.u.b. of (µα, να) belonging to Aα. (Such a l.u.b. exists because Aα is directed and hereditary and P is quasi- lattice ordered). Since P is quasi-lattice ordered and (r, d) is proper, Λ is compactly aligned. Hence the net λα has a convergent subnet. Let λ be its limit. Since Aα converges to A, λ is in A (by (F1)). Since µα, να ≤ λα, Lemma 6.5 gives µ, ν ≤ λ. This shows that A is directed. (b) The set F (λ) is obviously hereditary and directed. According to Lemma 6.5, it is closed. (c) Let A be a hereditary and directed closed subset of Λ. It will suffice to see that the net (F (λ))λ∈A converges to A. Let K be a compact set in Λ and U1, . . . , Un such that A ∩ K = ∅ and A ∩ Ui 6= ∅. Pick λi ∈ A ∩ Ui and let λ be a c.u.b. of λ1, . . . , λn in A. If λ ≥ λ, F (λ) ∩ Ui 6= ∅ because it contains λi. Since F (λ) is contained in A, its intersection with K is empty. (d) Straightforward. (e) Since we assume that P ∩ P −1 = {e}, the relation ≤ is an order relation, hence the injectivity of F . Suppose that λα → λ. Let K be a compact subset of Λ and let U1, . . . , Un be open subsets of Λ such that F (λ) ∩ K = ∅ and F (λ) ∩ Ui 6= ∅. If there is no α0 such that F (λα) ∩ K = ∅ for all α ≥ α0, there are subnets µβ ≤ λβ with µβ → µ in K. This is not possible since we have then µ ≤ λ. By assumption, λ belongs to the open set U1Λ ∩ . . . ∩ UnΛ, therefore there exists α1 ≥ α0 such that for all α ≥ α1, λα ∈ U1Λ ∩ . . . ∩ UnΛ. Hence we eventually have F (λα) ∩ Ui 6= ∅. (cid:3) We shall see later that the map F : Λ → C(Λ) is not necessarily a homeomor- phism onto its image (Remark 6.21(c)). Recall that at the beginning of the subsection we defined Ω = Ω \ {∅}. We now define Ω ∗ P = { (A, n) ∈ Ω × P : A · n 6= ∅ } = { (A, n) ∈ Ω × P : there exists λ ∈ A such that d(λ) ≥ n }, and T : Ω ∗ P → Ω sending (A, n) to A · n. As before, we define U (n) = { A ∈ Ω : (A, n) ∈ Ω ∗ P } and V (n) = { A · n : (A, n) ∈ Ω ∗ P } and we denote by Tn : U (n) → V (n) the map sending A to A · n. Definition 6.9. We define a path in Λ as a non-empty hereditary closed subset of Λ. The space Ω is called the path space of Λ. The above map T is called the shift on the path space. AMENABILITY OF GROUPOIDS 23 Theorem 6.10. Let Λ be a P -graph, where P is quasi-lattice ordered and Λ is (r, d)-proper. Let (Ω, P, T ) be as above. Then, T is a directed action of P on Ω by partial local homeomorphisms. Proof. First, we observe that U (n) and V (n) are open subsets. Indeed, we have U (n) = {A ∈ Ω : A ∩[m≥n Λm 6= ∅} and V (n) = r−1(s(Λn)), where r : Ω → Λ(0) is the range map defined in Lemma 6.4. Let us show that Tn : U (n) → V (n) is a local homeomorphism. To show that Tn is continuous, suppose that Aα converges to A. We need to show that Aα · n converges to A · n. It suffices to show that every subnet (Aβ · n) satisfies (F1) and (F2). Suppose that νβ ∈ Aβ · n converges to ν. Then there is µβ ∈ Λn such that µβνβ ∈ Aβ. Then r(µβ) = r(Aβ ) converges to r(A). By (r, d)-properness, there is a subnet µγ ∈ Λn which converges to some µ ∈ Λn. Then µν belongs to A and ν belongs to A · n. Condition (F2) is clear: let ν ∈ A · n. There exists µ ∈ Λn such that µν ∈ A. There is a subnet Aβ and λβ ∈ Aβ such that λβ converges to µν. Then λβ · n ∈ Aβ · n converges to ν. Let U be an open subset of Λn such that sU : U → s(U ) is a homeomorphism; we denote by σ the inverse of sU . Then U = { A ∈ U (n) : A ∩ U 6= ∅ } is open. Each A ∈ U contains a unique µ ∈ U which is given by µ = σ(r(A · n)). Thus, the restriction of Tn to U is a bijection of U onto Tn( U ) = r−1(s(U )) having as inverse map B 7→ σ◦r(B)B. To show that this inverse map is continuous, it suffices to show that the product map Λn ∗ Ω → Ω sending (µ, B) to µB is continuous. Consider a net (µα, Bα) converging to (µ, B). We will show that µαBα converges to µB by checking that every subnet µβBβ satisfies (F1) and (F2). For (F1), we proceed as in the proof of Lemma 6.7(b). Consider a net λβ ∈ µβBβ converging to λ. We distinguish two cases: if d(λ) ≤ d(µ), then d(λβ) ≤ d(µβ ) for β large enough. This implies λβ ≤ µβ, hence λ ≤ µ. If d(λ) ≤ d(µ) does not hold, there is a subnet λγ for which d(λγ ) ≤ d(µγ) does not hold. Then, we can write λγ = µγνγ with νγ ∈ Bγ. Since d(λγ) = d(λ) and d(µγ ) = d(µ) for γ large enough, this fixes d(νγ). We also have r(νγ ) = s(µγ). By (r, d)-properness, there is a converging subnet νδ. Its limit ν belongs to B because B is closed. We have λ = µν, hence λ is in µB. Let us check (F2). Suppose that λ belongs to µB. Suppose first that d(λ) ≤ d(µ). We have n = pq with p = d(λ). For β large enough, d(µβ) = n and we have a unique factorization µβ = λβρβ where d(λβ ) = p. By (r, d)-properness, λβ has a subnet converging to some λ′; since ρβ = µβ · p converges to µ · p, we have µ = λ′(µ · p); by unique factorization, λ′ = λ. Therefore λβ converges to λ. Since λβ belongs to µβBβ, we are done if d(λ) ≤ d(µ). Suppose now that λ = µν where ν ∈ B. There exists a subnet Bγ and νγ ∈ Bγ converging to ν. Then µγνγ belongs to µγBγ and converges to µν; therefore, (µ, B) 7→ µB is continuous. To see that the action is directed, consider U (n) = { A ∈ Ω : there exists λ ∈ A such that d(λ) ≥ n }. Assume that U (m) ∩ U (n) 6= ∅ and let A ∈ U (m) ∩ U (n). There exist µ, ν ∈ A such that d(µ) ≥ m and d(ν) ≥ n. Since A is directed, there exists λ ∈ A greater than µ and ν. Then d(λ) is greater than m and n. (cid:3) Thus, under our assumptions on Λ and P , we can construct the semi-direct product groupoid G(Ω, P, T ) according to Proposition 5.12. 24 JEAN N. RENAULT AND DANA P. WILLIAMS Definition 6.11. Let Λ be a P -graph, where P is quasi-lattice ordered and Λ is (r, d)-proper and let T be the action of P on the path space Ω. The groupoid G(Ω, P, T ) is called the Toeplitz groupoid of the topological higher rank graph Λ. Its C-algebra is called the Toeplitz algebra of Λ and denoted by C∗(Λ). This construction is the same as in the work of T. Yeend, for example [26]. The main differences are that we consider an arbitrary quasi-lattice ordered semigroup P rather than Nd and that we make an explicit (rather than implicit) use of the Fell topology on a space of closed subsets to define the topological path space Ω. In [26], the path space Ω is denoted by XΛ and the groupoid G(Ω, P, T ), called the path groupoid, is denoted by GΛ. 6.2. The boundary path space ∂Ω. We continue to assume that P is quasi- lattice ordered and that Λ is (r, d)-proper. In particular, Λ is compactly aligned. The Cuntz-Krieger algebra of the P -graph Λ is described in [26] as the C-algebra of the reduction of G(Ω, P, T ) to a closed invariant subset ∂Ω called the boundary path space. Let us describe the boundary path space in our presentation. Recall that the elements of Ω are the non-empty closed hereditary and directed subsets of Λ. Definition 6.12. Let Λ be a P -graph. We say that E ⊂ Λ is exhaustive if for all λ ∈ Λ such that r(λ) ∈ r(E), there exists µ ∈ E such that (λ, µ) has a c.u.b. Definition 6.13. Let Λ be a P -graph. (a) Given A ∈ Ω, we say that λ ∈ A is extendable in A if for all E ⊂ Λ which are exhaustive, compact and such that r(E) is a neighborhood of s(λ), there exists µ ∈ E such that λµ ∈ A. (b) We say that A ∈ Ω is a boundary path if all its elements are extendable in A. We define the boundary path space ∂Ω as the subspace of all boundary paths. Example 6.14 (Singly Generated Systems). Recall that this means a local homeo- morphism T : U → V , where U, V are open subsets of a locally compact Hausdorff space X and that Λ = X ∗ N = { (x, n) ∈ X × N : x ∈ U (n) := dom(T n) } The non-empty closed hereditary directed subsets of Λ are: F (x, n) = { (x, m) : m ≤ n } where n ∈ N := N∪{∞}, x ∈ U (n) if n is finite and x ∈ U (∞) = T U (n) if n = ∞. The boundary paths are: F (x, τ (x)) where τ (x) = sup{n ∈ N : x ∈ U (n)} Note that here the boundary paths are exactly the maximal paths. Lemma 6.15. Let A ∈ Ω. If λ′ ∈ A is extendable in A, then every λ ≤ λ′ is extendable in A. Proof. Let E ⊂ Λ be exhaustive, compact and such that r(E) is a neighborhood of s(λ). We write λ′ = λξ and we let n = d(ξ). We choose U open relatively compact neighborhood of ξ contained in Λn and such that sU is injective and r(U ) ⊂ s(E). We define E′ to be the the set of elements µ′ of Λ of minimal degree for which there AMENABILITY OF GROUPOIDS 25 exist µ ∈ E and η ∈ U such that µ ≤ ηµ′. One checks that E′ is closed. Since r(µ) ∈ r(U ) and d(µ) lies in a finite set by the compact alignment property, E′ is compact. We are going to show that E′ is exhaustive and that r(E′) = s(U ). By construction, r(E′) ⊂ s(U ). Consider ν ∈ Λ such that r(ν) ∈ s(U ). Pick η ∈ U such that s(η) = r(ν). Then r(ην) = r(η) ∈ r(U ) ⊂ s(E). Since E is exhaustive, there exists µ ∈ E such that (ην, µ) has a c.u.b. This means the existence of α, β ∈ Λ such that µβ = (ην)α. We have then µ ≤ η(να), hence the existence of µ′ ∈ E′ such that µ′ ≤ να. In particular, r(ν) ∈ r(E′), which shows that r(E′) = s(U ). We have found µ′ ∈ E′ such that (µ′, ν) has a c.u.b. This shows that E′ is exhaustive. Since λ′ is extendable in A, there exists µ′ ∈ E′ such that λ′µ′ ∈ A. By definition of E′, there is (µ, η) ∈ E × U such that µ ≤ ηµ′. Since η and ξ both belong to U and have same source, η = ξ. Since λµ ≤ λξµ′ = λ′µ′, λµ ∈ A. This shows that λ is extendable in A. (cid:3) Proposition 6.16. Let ∂Ω be the boundary path space of a P -graph Λ, where P is a quasi-lattice ordered subsemigroup of a group Q and Λ is (r, d)-proper. Then (a) ∂Ω is a closed subset of Ω. (b) If A ∈ ∂Ω and n ∈ P such that A · n 6= ∅, then A · n ∈ ∂Ω. (c) If B ∈ ∂Ω and ρ ∈ r(B)Λ, then ρB ∈ ∂Ω. Proof. (a) Suppose that Aα → A and Aα ∈ ∂Ω. If A is not a boundary path, there exists λ ∈ A and E ⊂ Λ exhaustive, compact with r(E) a neighborhood of s(λ) such that λµ 6∈ A for all µ ∈ E. Let V be an open neighborhood of s(λ) contained in r(E). Let U be an open relatively compact neighborhood of λ such that s(U ) ⊂ V and U ⊂ Λn, where n = d(λ). We have A ∩ U 6= ∅ and A ∩ U E = ∅. if λ′µ ∈ A, with λ′ ∈ U and µ ∈ E , then Let us check this second assertion: λ′ ∈ A. Since d(λ′) = d(λ), λ = λ′. This is a contradiction. There exists α0 such that for all α ≥ α0, Aα ∩ U 6= ∅ and Aα ∩ U E = ∅. Let λα ∈ Aα ∩ U . Then r(E) is a neighborhood of s(λα) and for all µ ∈ E such that r(µ) = s(λα), λαµ does not belong to A. This contradicts the fact that Aα is a boundary path. (b) Let A be a boundary path and n ∈ P such that A · n is non-empty. Let us show that A · n is a boundary path. Recall that ν ∈ A · n if and only if there exists ρ ∈ Λn such that ρν ∈ A. Let ν ∈ A · n and E ⊂ Λ exhaustive, compact with r(E) a neighborhood of s(ν). There exists ρ ∈ Λn such that λ = ρν ∈ A. Since A is a boundary path, there exists µ ∈ E such that λµ ∈ A. Therefore νµ ∈ A · n. (c) We first show that every λ ∈ ρB of the form λ = ρν, where ν ∈ B, satisfies the property. Indeed, let E be a subset of Λ which is exhaustive, compact and such that r(E) is a neighborhood of s(λ). Since s(λ) = s(ν) and B is a boundary path, there exists µ ∈ E such that νµ ∈ B. Then λµ = ρ(νµ) belongs to ρB. We apply Lemma 6.15 to conclude that ρB ∈ ∂Ω. (cid:3) Corollary 6.17. Let Λ be a P -graph, where P is quasi-lattice ordered and Λ is (r, d)-proper. Let (Ω, P, T ) and G(Ω, P, T ) be as above. Then the boundary path space ∂Ω is a closed invariant subspace of Ω with respect to G(Ω, P, T ). Proof. The equivalence relation on Ω induced by G(Ω, P, T ) is precisely A ∼ B if and only if there exist m, n ∈ P such that A · m and B · n are non-empty and equal. If A is a boundary path, so is A · m = B · n by Proposition 6.16(b). Since B = ρ(B · n) for some ρ, B is a boundary path by part (c) of the same proposition. 26 JEAN N. RENAULT AND DANA P. WILLIAMS Therefore, the space of boundary paths is invariant. We have shown above that it is closed. (cid:3) Definition 6.18. The reduction G(∂Ω, P, T ) of the Toeplitz groupoid G(Ω, P, T ) is called the Cuntz-Krieger groupoid of the topological higher rank graph Λ. Its C∗-algebra is called the Cuntz-Krieger algebra of Λ and denoted by C∗(∂Λ). Note that G(∂Ω, P, T ) is the semi-direct product groupoid of the semigroup action (∂Ω, P, T ). Since the action on Ω is directed, so is the action on ∂Ω. Hence Theorem 5.13 (and [1, Corollary 6.2.14]) give us the following. Corollary 6.19. Let P a quasi-lattice ordered subsemigroup of a group Q and Λ be a P -graph which is is (r, d)-proper. If the group Q is amenable, then the Toeplitz groupoid G(Ω, P, T ) and the Cuntz-Krieger groupoid G(∂Ω, P, T ) are amenable. Therefore the Toeplitz algebra C∗(Λ) and the Cuntz-Krieger algebra C∗(∂Λ) are nuclear. 6.3. Topological higher rank graphs coming from semigroup actions. We have seen that a semigroup action (X, P, T ) gives the topological higher rank graph Λ = X ∗ P . If P is quasi-lattice ordered, we can construct the groupoids G(Ω, P, T ) and G(∂Ω, P, T ). (Recall that the graph of an action is always (r, d)-proper.) On the other hand, if the action is directed, we can construct the groupoid G(X, P, T ). It is then natural to compare these groupoids when the action is directed and P is quasi-lattice ordered. An important case is when P is both directed and quasi- lattice ordered, which means that P is lattice ordered. Proposition 6.20. Let (X, P, T ) be a directed semigroup action, where P is a quasi-lattice ordered subsemigroup of a group Q. Then there is a P -equivariant homeomorphism of X onto ∂Ω which implements a groupoid isomorphism of G(X, P, T ) and G(∂Ω, P, T ). Proof. Given x ∈ X, A(x) = xΛ = {(x, n) ∈ X × P : x ∈ U (n)} is a closed subset of Λ. It is hereditary. It is directed because the action is directed. This defines a map J : X → Ω sending x to J(x) = xΛ which is injective. The map J is continuous. Indeed, let K = L ∗ F , where L is a compact subset of X and F a finite subset of P and Ui = Vi ∗ {pi}, i = 1, . . . , n, where Vi is an an open subset of X and pi ∈ P . Then, J(x) ∩ K = ∅ and J(x) ∩ Ui 6= ∅ for all i = 1, . . . , n if and only if x ∈ (Lc ∪ r(d−1(F c)) ∩ V1 ∩ . . . ∩ Vn, which is open. The inverse map is continuous since it is the restriction to J(X) of the range map r : Ω → X. We claim that J(x) is a boundary path. Because the action is directed, every subset E of Λ is exhaustive. Let us show that (x, m), where x ∈ U (m), is extendable in J(x). Let E be a subset of Λ such that x · m = s(x, m) ∈ r(E). There exists n ∈ P such that (x · m, n) ∈ E. Then x ∈ U (mn) and (x, mn) = (x, m)(x · m, n). We also claim that every boundary path A is of the form J(x), where x = r(A). We have A ⊂ J(x) by definition. Conversely, let (x, n) ∈ J(x). Let E = L ∗ {n} where L is a compact neighborhood of x. Since A is extendable in A at λ = (x, e) ∈ A, there exists µ ∈ E such that λµ ∈ A. Necessarily µ = (x, n) and (x, n) = (x, e)(x, n) belongs to A. Let us show that the map J : X → Ω is P -equivariant: for all (x, m) ∈ X ∗ P , (J(x), m) ∈ Ω ∗ P and J(x · m) = J(x) · m. Let (x, m) ∈ X ∗ P . Since AMENABILITY OF GROUPOIDS 27 (x, m) = (x, m)(x · m, e) belongs to J(x), J(x) · m is non-empty. In fact, we have J(x) · m = J(x · m): given that x ∈ U (m), (y, n) ∈ J(x) · m means exactly that y = xm and x ∈ U (mn) while (y, n) ∈ J(xm) means that y = xm and xm ∈ U (n). Thus, the semigroup actions (X, P, T ) and (∂Ω, P, T ) are isomorphic. One de- duces that the semi-direct product groupoids G(X, P, T ) and G(∂Ω, P, T ) are iso- morphic. (cid:3) Remark 6.21. (a) Consider a locally compact semigroup action (X, P, T ). We are able to construct the semi-direct product groupoid G(X, P, T ), hence the C-algebra C∗(G(X, P, T )), when the action is directed. If the action is not directed but P is quasi-lattice ordered, we can introduce the topological higher rank graph Λ = X ∗P and consider instead the semigroup action (∂Ω, P, T ), the groupoid G(∂Ω, P, T ) and the Cuntz-Krieger C-algebra C∗(∂Λ). Both constructions agree when they are possible. (b) It is instructive to specialize the situation (a) to the case where X is reduced to a point. It turns out that this leads us to Wiener-Hopf (also called Toeplitz) C-algebras of semigroups. We have seen earlier that, in this case, the semidirect product G(X, P, T ) is P P −1. If P is an Ore semigroup, then P P −1 is a group and the corresponding C-algebra is its group C-algebra. If P is not an Ore semigroup, we cannot even define a C-algebra. However, if P is quasi-lattice ordered, we can perform the construction of (a) which introduces the higher rank graph Λ = P . Nica shows in [14] that the corresponding path space Ω is the spectrum of a canonical diagonal C-subalgebra of the Wiener-Hopf algebra W(Q, P ). In fact, he defines in Section 9 of the preprint version of [14] an ´etale groupoid G, which he calls the Wiener-Hopf groupoid and which is exactly (up to an obvious isomorphism) our groupoid G(Ω, P, T ). This Wiener-Hopf groupoid is only briefly mentioned in the subsection 1.5 of the published version [14]. The reduced C-algebra of this groupoid is the Wiener-Hopf C-algebra W(Q, P ) while the non-degenerate representations of its full C-algebra are exactly the Nica-covariant representations of (Q, P ). He also shows that, when P ⊂ Q is the free semigroup SFn ⊂ Fn, the reduction to the boundary of the Wiener-Hopf groupoid, which we have denoted earlier by G(∂Ω, P, T ), is the Cuntz groupoid On of [17]. Therefore its reduced and its full C-algebras coincide and are isomorphic to the Cuntz algebra On. One can also note that, in this particular case (b) of the general theory of higher rank graphs, the boundary path space ∂Ω is also studied by Crisp and Laca [5]. Their definition agrees with ours. We refer the reader to the work on semigroup C*-algebras by X. Li [11, 12] and by Sundar [23] for recent developments. (c) Let us justify the assertion made earlier that the map F : Λ → C(Λ) is not necessarily a homeomorphism onto its image. Let (X, T ) be a singly generated dynamical system in the sense of [18]: X is a locally compact space and T is a local homeomorphism from an open subset dom(T ) of X onto another open subset ran(T ) of X. For n ∈ N, let U (n) = dom(T n). We view T as an action of N on X, with X ∗ N = {(x, n) ∈ X × N : x ∈ U (n)}and xn = T nx. The associated N-graph is Λ = X ∗ N. Let xα be a net converging to x in X. Assume that there exists n such that xα ∈ U (n) for all α and x ∈ U (n − 1) \ U (n). Then F (xα, n) converges to F (x, n − 1) in C(Λ) but (xα, n) does not converge to (x, n − 1) in Λ. 28 JEAN N. RENAULT AND DANA P. WILLIAMS References [1] Claire Anantharaman-Delaroche and Jean Renault, Amenable groupoids, Monographies de L'Enseignement Math´ematique [Monographs of L'Enseignement Math´ematique], vol. 36, L'Enseignement Math´ematique, Geneva, 2000. With a foreword by Georges Skandalis and Appendix B by E. Germain. [2] William Arveson, An Invitation to C ∗-algebras, Springer-Verlag, New York, 1976. Graduate Texts in Mathematics, No. 39. [3] Victor Arzumanian and Jean Renault, Examples of pseudogroups and their C ∗-algebras, Operator algebras and quantum field theory (Rome, 1996), Int. Press, Cambridge, MA, 1997, pp. 93 -- 104. [4] Nathan Brownlowe, Aidan Sims, and Sean T. Vittadello, Co-universal C ∗-algebras associated to generalised graphs, Israel J. Math. 193 (2013), 399 -- 440. [5] John Crisp and Marcelo Laca, Boundary quotients and ideals of Toeplitz C ∗-algebras of Artin groups, J. Funct. Anal. 242 (2007), 127 -- 156. [6] Siegfried Echterhoff and Heath Emerson, Stucture and K-theory for crossed products by proper actions, Expo. Math. 29 (2011), 300 -- 344. [7] Siegfried Echterhoff and Dana P. Williams, Structure of crossed products by strictly proper actions on continuous-trace algebras, Trans. Amer. Math. Soc. 366 (2014), 3649 -- 3673. [8] R. Exel, Tight representations of semilattices and inverse semigroups, Semigroup Forum 79 (2009), 159 -- 182. [9] Alex Kumjian and David Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20. [10] Marcelo Laca and Iain Raeburn, Semigroup crossed products and the Toeplitz algebras of nonabelian groups, J. Funct. Anal. 139 (1996), 415 -- 440. [11] Xin Li, Semigroup C∗-algebras and amenability of semigroups, J. Funct. Anal. 262 (2012), 4302 -- 4340. [12] , Nuclearity of semigroup C ∗-algebras and the connection to amenability, Adv. Math. 244 (2013), 626 -- 662. [13] Paul S. Muhly and Dana P. Williams, Renault's equivalence theorem for groupoid crossed products, NYJM Monographs, vol. 3, State University of New York University at Albany, Albany, NY, 2008. Available at http://nyjm.albany.edu:8000/m/2008/3.htm. [14] Alexandru Nica, C ∗-algebras generated by isometries and Wiener-Hopf operators, J. Operator Theory 27 (1992), 17 -- 52. [15] John C. Quigg, Discrete C ∗-coactions and C ∗-algebraic bundles, J. Austral. Math. Soc. Ser. A 60 (1996), 204 -- 221. [16] Arlan Ramsay, The Mackey-Glimm dichotomy for foliations and other Polish groupoids, J. Funct. Anal. 94 (1990), 358 -- 374. [17] Jean Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, vol. 793, Springer-Verlag, New York, 1980. [18] , Cuntz-like algebras, Operator theoretical methods (Timi¸soara, 1998), Theta Found., Bucharest, 2000, pp. 371 -- 386. [19] [20] Jean Renault, Aidan Sims, Dana P. Williams, and Trent Yeend, Uniquess theorems for too- , Topological amenability is a Borel property, 2013. (arXiv:math.OA.1302.0636). logical higher-rank graph C ∗-algebras, 2012. (arXiv math.OA.0906.0829v3). [21] Aidan Sims and Dana P. Williams, The primitive ideals of some ´etale groupoid C ∗-algebras, 2015. (arXiv.math.OA.1501.02302). [22] Jack Spielberg, Groupoids and C ∗-algebras for categories of paths, Trans. Amer. Math. Soc. 366 (2014), 5771 -- 5819. [23] S. Sundar, C ∗-algebras associated to Ore semigroups, 2015. (arXiv.math.OA.1408.4242). [24] Jean-Louis Tu, La conjecture de Baum-Connes pour les feuilletages moyennables, K-Theory 17 (1999), 215 -- 264. [25] Dana P. Williams, Crossed products of C ∗-algebras, Mathematical Surveys and Monographs, vol. 134, American Mathematical Society, Providence, RI, 2007. [26] Trent Yeend, Groupoid models for the C ∗-algebras of topological higher-rank graphs, J. Op- erator Theory 57 (2007), 95 -- 120. AMENABILITY OF GROUPOIDS 29 D´epartment de Math´ematiques, Universit´e d'Orl´eans et CNRS (UMR 7349 et FR 2964), 45067 Orl´eans Cedex 2, FRANCE E-mail address: [email protected] Department of Mathematics, Dartmouth College, Hanover, NH 03755-3551 USA E-mail address: [email protected]
1609.06754	1	1609	2016-09-21T21:04:53	Kadison's Pythagorean Theorem and essential codimension	[ "math.OA" ]	Kadison's Pythagorean theorem (2002) provides a characterization of the diagonals of projections with a subtle integrality condition. Arveson (2007), Kaftal, Ng, Zhang (2009), and Argerami (2015) all provide different proofs of that integrality condition. In this paper we interpret the integrality condition in terms of the essential codimension of a pair of projections introduced by Brown, Douglas and Fillmore (1973), or, equivalently of the index of a Fredholm pair of projections introduced by Avron, Seiler, and Simon (1994). The same techniques explain the integer occurring in the characterization of diagonals of selfadjoint operators with finite spectrum by Bownik and Jasper (2015).	math.OA	math	KADISON'S PYTHAGOREAN THEOREM AND ESSENTIAL CODIMENSION VICTOR KAFTAL AND JIREH LOREAUX Abstract. Kadison's Pythagorean theorem (2002) provides a characteriza- tion of the diagonals of projections with a subtle integrality condition. Arveson (2007), Kaftal, Ng, Zhang (2009), and Argerami (2015) all provide different proofs of that integrality condition. In this paper we interpret the integrality condition in terms of the essential codimension of a pair of projections intro- duced by Brown, Douglas and Fillmore (1973), or, equivalently of the index of a Fredholm pair of projections introduced by Avron, Seiler, and Simon (1994). The same techniques explain the integer occurring in the characterization of diagonals of selfadjoint operators with finite spectrum by Bownik and Jasper (2015). 1. Introduction In his seminal papers on the Pythagorean Theorem ([Kad02a, Kad02b]), Kadison characterizes the diagonals of projections, that is the sequences that can appear on the diagonal of a matrix representation of a projection. The main assertion of his Theorem 15 is by now usually paraphrased as follows: Theorem 1.1 ([Kad02b, Theorem 15]). A sequence {dn} with 0 ≤ dn ≤ 1 is the diagonal of a projection B(H) if and only if for a = Xdn≤1/2 dn and b = Xdn>1/2 (1 − dn), either and referred to that integer as "curious". (i) a + b = ∞, or (ii) a + b < ∞ and a − b ∈ Z. Kadison proved that a−b is arbitrarily close to an integer and hence is an integer Let us first express a and b in operator theoretic terms. Call p the projection, {ej} the orthonormal basis of H used for the matrix representation, and q the projection on span{ej dj > 1/2}. Then (1.1) hence if a+b < ∞, we have q⊥(p−q)q⊥ = q⊥pq⊥ ∈ L1, q(p−q)q = −(q−qpq) ∈ L1, and a = Tr(q⊥pq⊥) and b = Tr(q − qpq), (1.2) a − b = Tr(cid:0)q(p − q)q + q⊥(p − q)q⊥(cid:1). 2010 Mathematics Subject Classification. Primary:47B15, 47A53; Secondary: 46C05, 42C15. Key words and phrases. Essential codimension, Fredholm pairs of projections, diagonals of projections, diagonals of selfadjoint operators. This work was partially supported by the Simons Foundation grant No 245660 to Victor Kaftal. 1 2 VICTOR KAFTAL AND JIREH LOREAUX If we knew that p − q ∈ L1, then we would have a − b = Tr(p − q) and then, by [Eff89, Lemma 4.1], we could conclude that Tr(p − q) ∈ Z. However, since p − q is not necessarily positive, the fact that its corners are trace-class does not imply that p − q itself is trace-class. In fact, Argerami proved in [Arg15] that p − q ∈ L2 and by modifying Effros' argument, he showed that this is sufficient to guarantee that a − b is an integer. However, neither Kadison's nor Argerami's proof shed much light on the origin of that integer itself. One of Bill Arveson's sayings was that if you find an integer in operator theory you should look for a Fredholm operator. Arveson partially extended Kadison's work on the Pythagorean Theorem in [Arv07] where he studied the diagonals of normal operators with finite spectrum with infinite multiplicity that forms the ver- tices of a convex polygon in C, infinite co-infinite projections being a degenerate special case. He also found an "index obstruction" for their diagonals which de- pended on the following result. Theorem 1.2 ([Arv07, Theorem 3]). Let p, q be projections in B(H) with p−q ∈ L2. Then p ∧ q⊥ and p⊥ ∧ q are finite, q(p − q)q and q⊥(p − q)q⊥ belong to L1, and Tr(cid:0)q(p − q)q + q⊥(p − q)q⊥(cid:1) = Tr(p ∧ q⊥) − Tr(p⊥ ∧ q) ∈ Z. Whenever we have two projections, p and q, we denote by q pH the operator in B(pH, qH). Then a key step in Arveson's proof is the fact that if p − q ∈ L2, then (1.3) q pH is Fredholm and ind(q pH) = Tr(cid:0)q(p − q)q + q⊥(p − q)q⊥(cid:1). Although Arveson did not state so explicitly, embedded in his proofs one can also find the fact that using the notations established above, if a + b < ∞, then indeed p − q ∈ L2 and hence a − b = ind(q pH), which explains why a − b is an integer. What remains to be explained is the role of q pH and the significance of its index. Note that ind(q pH) = − ind(p qH) since (q pH)∗ = p qH. A similar question arises from another proof that a − b is an integer which was obtained in [KNZ09, Corollary 3.6]. Let us briefly sketch the original computa- tion (reformulated in new notation) as it introduces the connections we want to illustrate. Let w be an isometry with range p, let Λ := {j dj > 1/2}, dj 6= 0 dj = 0 w∗ej e1 ej ⊗ fj fj :=( 1√dj f :=Xj∈Λ f = vf ta :=Xj6∈Λ tb :=Xj∈Λ the polar decomposition of f djfj ⊗ fj Then kfjk = 1 for all j, ta, tb ∈ L1 (1 − dj)fj ⊗ fj +, and 1 =Pj djfj ⊗ fj, hence ta − tb = 1 − f∗f = v∗v(ta − tb) + 1 − v∗v E(q − vv∗) = −E(v(ta − tb)v∗). KADISON'S PYTHAGOREAN THEOREM AND ESSENTIAL CODIMENSION 3 Hence a − b = Tr(ta − tb) = − Tr(q − vv∗) + Tr(1 − v∗v) ∈ Z. It is then immediate to see (but was not remarked explicitly in [KNZ09]), that a − b = − ind(v∗ qH). Notice that f can be interpreted as the frame transform (the analysis operator) of the Bessel sequence {fj}j∈Λ and v as the frame transform of the associated Parseval frame. While this construction provides indeed a proof that v∗ qH is Fredholm, a natural question is why ind(v∗ qH) = ind(p qH) as can be obtained from Arveson's work. To answer it, notice first that since f∗f is a trace-class perturbation of the identity, it is Fredholm and hence so are f and f∗ = fv∗. Furthermore ind(f∗ qH) = ind(v∗ qH) since ind(f) = 0. Next w∗q =Xj∈Λ w∗ej ⊗ ej =Xj∈Λpdjfj ⊗ ej = f d q is invertible in B(qH). Thus w∗ qH is also where d := Pj∈Λpdjej ⊗ ej ≥ 1√2 Fredholm in B(qH,H) and (1.4) It is then immediate to verify (see also (2.3) below) that ind(w∗ qH) = ind(p qH) as obtained by Arveson. However, neither the proof due to Arveson nor the one in [KNZ09]) provides a a − b = − ind(w∗ qH). natural explanation of the role of w∗ qH or p qH . The goal of our paper is to provide an explanation of that role in the context of the notion of essential codimension [p : q] of a pair of projections p and q with p − q ∈ K that was introduced in the BDF theory (see [BDF73] and Section 2), or of the more general notion of index of a Fredholm pair of projections, introduced by Avron, Seiler, and Simon in [ASS94]. Combining Arveson's work with the study of Fredholm pairs and essential codi- mension, one can provide a natural identification of Kadison's integer with the essential codimension of a pair of projections. In the notations of Theorem 1.1 we have: Theorem 1.3. Let p ∈ B(H) be a projection such that a + b < ∞ and let q be the projection on span{ej dj > 1/2}. Then p − q ∈ L2 and a − b = [p : q]. To understand the simple proof of this result, and for the convenience of the readers not familiar with the notions of Fredholm pairs, essential codimension, and the work of Arveson in [Arv07], we will provide in Section 2 a self-contained short presentation of the relevant results of the theory of Fredholm pairs. We have strengthened several results and generalized them to the case when p − q belongs to an arbitrary (two-sided) operator ideal J rather than just the Hilbert -- Schmidt ideal L2. Since Fredholm pairs have found most of their applications in the theory of spectral flows in type I or type II von Neumann algebras, we will conclude Section 2 with a very brief foray into the case when the notion of Fredholm operators and indices are taken relative to a semifinite von Neumann algebra (also called Breuer -- Fredholm, or more precisely τ -Breuer -- Fredholm operators). In Section 3 we will assemble the results previously collected into a proof of Theorem 1.3 that is inspired by, but independent of, the work by Arveson. Then 4 VICTOR KAFTAL AND JIREH LOREAUX we will extend part of Proposition 2.8 to positive contractions. We will then use the same techniques to identify an integer appearing in the study by Bownik and Jasper of the diagonals of selfadjoint operators with finite spectrum and also to simplify the proof of one of the key results of that paper (see [BJ15]). We thank R. Douglas for having suggested to the first named author of this paper to consider a possible connection between the frame transform approach originally used and essential codimension. 2. Essential codimension and Fredholm pairs In this paper H denotes a separable infinite dimensional complex Hilbert space The notion of essential codimension of two projections was first introduced in and K the C*-algebra of compact operators on H. ([BDF73, Remark 4.9]). Definition 2.1 ([BDF73]). Given projections p, q ∈ B(H) for which p − q ∈ K(H), the essential codimension [p : q] of p and q is defined by: [p : q] :=(Tr(p) − Tr(q) Tr(p) < ∞, Tr(q) < ∞ Tr(p) = Tr(q) = ∞, w, v isometries, ww∗ = p, vv∗ = q. This definition depends on the fact that (v∗w)∗(v∗w) = 1 + w∗(q − p)w and similarly for w∗v. Thus setting π to be the projection onto the Calkin algebra, we see that π(v∗w) is unitary and hence v∗w is Fredholm. Also, if w and v are another pair of isometries with ranges p and q respectively, then ind(v∗w) ind(v∗w) = ind(v∗vv∗ w w∗w) = ind(v∗ w) since w∗ w and v∗v are unitaries. This shows that [p : q] does not depend on the choice of the isometries w and v. Some properties of the essential codimension were presented without proof in [Bro81] and a complete exposition can be found in [BL12], together with an in- teresting application to liftability of projections in the corona algebra of, among others, C([0, 1]) ⊗ K. Independently, and without reference to essential codimension, Avron, Seiler, and Simon defined in [ASS94] the more general notion of Fredholm pairs of projections. Definition 2.2 ([ASS94]). A pair of projections (p, q) in B(H) is said to be Fred- holm if q pH is a Fredholm operator as an element of B(pH, qH). The index of the pair is defined to be ind(q pH). Notice that if v ∈ B(H4,H3) and w ∈ B(H1,H2) are unitaries, then (2.1) g ∈ B(H2,H3) is Fredholm ⇔ v∗gw ∈ B(H1,H4) is Fredholm. and then (2.2) ind(v∗gw) = ind g. Thus if w and v are isometries with ranges p and q respectively, then (2.3) q pH∈ B(pH, qH) is Fredholm ⇔ v∗w = v∗ pH w ∈ B(H) is Fredholm ⇔ v∗ pH∈ B(pH,H) is Fredholm KADISON'S PYTHAGOREAN THEOREM AND ESSENTIAL CODIMENSION 5 Recall that v∗w is Fredholm if and only if π(v∗w) is invertible. We have seen above that if p − q ∈ K, then π(v∗w) is unitary and hence v∗w is Fredholm, that is (p, q) is a Fredholm pair and by (2.2), [p : q] = ind(q pH). For consistency we will henceforth write [p : q] := ind(q pH) whenever (p, q) is a Fredholm pair even when p − q 6∈ K. Soon after [ASS94], W. Amrein, K. Sinha [AS94] realized that the proofs in [ASS94] could be considerably simplified by reducing to the case of projections in generic position. This notion was first introduced by Dixmier [Dix48] (he called them in "position p") and independently by Krein, Kranosleskii and Milman [KKM48], and further studied by Davis [Dav58], Halmos[Hal69] (he called them "generic pairs"), and others. Definition 2.3. Two projections p, q ∈ B(H) are said to be in generic position if p ∧ q = p ∧ q⊥ = p⊥ ∧ q = p⊥ ∧ q⊥ = 0. When just the first three projections are zero, the pair p, q is in generic position in B((p ∨ q)H) and when there is no risk of confusion we will simply call them in generic position. For the readers' convenience we will collect here some results on projections in generic position. Good references can be found in the texts of Strătilă [Str81, 17.15], Takesaki [Tak02, pages 306-308] and in the article of Amrein and Sinha [AS94]. Theorem 2.4. Let p, q ∈ B(H) be two projections. (i) Then the projections p0 := p − p ∧ q − p ∧ q⊥ and q0 := q − p ∧ q − p⊥ ∧ q Suppose p, q are in generic position and let N := {p, q}′′ be the von Neumann algebra generated by them. Then are in generic position (in the Hilbert space (p0 ∨ q0)H.) (ii) Nq(= (qN q qH) is masa of N and N can be identified with M2(Nq). (iii) There are two positive injective contractions c and s in Nq with c2 + s2 = 1 (the identity operator of Nq) such that p and q can be identified with 0 0 cs (2.4) q =(cid:18)1 0(cid:19) and p =(cid:18)c2 s2(cid:19) . (iv) c = (qpq)1/2 qH, s = (qp⊥q)1/2 qH, and ksk = kp − qk. In this section we will often use the representation (2.4) without further reference. Notice that for projections in generic position, the equality ksk = kp − qk ([Str81, 17.15 (8)], see also [Dix48, pg 391] and [Tak02, pg 308]) follows also from the identity: cs c 0 0 s(cid:19) =(cid:18)s 0 0 s(cid:19)(cid:18)−s c c s(cid:19) c cs cs (2.5) s2(cid:19) =(cid:18)−s p − q =(cid:18)−s2 s(cid:19)(cid:18)s Notice also that if we set v := 1√2(cid:18)(1 − s)1/2 p − q = v(cid:18)s (2.6) and 0 0 −s(cid:19) v∗. (1 + s)1/2 (1 + s)1/2 −(1 − s)1/2(cid:19), then v = v∗ is unitary It is well known that projections in generic position are unitarily equivalent in N , and over the years various authors (e.g., [Kat66], [BL12], [Dix48], [ASS94]) have 6 VICTOR KAFTAL AND JIREH LOREAUX constructed different unitaries in N implementing the equivalence. We will use the following unitary: (2.7) p = uqu∗ for the unitary u :=(cid:18)c −s c (cid:19) . s As shown by Amrein and Sinha in [AS94], reduction to generic position makes the analysis of Fredholm pairs simpler and more transparent. For the convenience of the readers, we will provide here a short self-contained presentation of the main results on Fredholm pairs that we will need in the sequel, completing and gener- alizing results obtained in [BL12], [ASS94], [AS94], [Arv07]. The starting point is the analysis of the case when the projections are in generic position. Recall that Fredholm operators are characterized by being invertible modulo the compact operators. Lemma 2.5. Let p, q be projections in generic position. Then the following are equivalent. (i) p qH is a Fredholm operator (ii) c is invertible (iii) p qH is invertible (iv) kp − qk < 1 (v) kp − qkess < 1. If these conditions are satisfied, then ind(p qH) = 0. Proof. (i) ⇒ (ii). Since p qH = c is injective, and since p qH has closed range, by the Inverse Mapping Theorem p qH is invertible and hence so is c. cs(cid:19)−1 (ii) ⇒ (iii). Immediate since p qH=(cid:18)c2 (iii) ⇒ (i). Obvious. (ii) ⇔ (iv). Immediate since kp − qk = ksk < 1 and c is invertible iff c2 ≥ δ1 for some δ > 0 iff s2 ≤ (1 − δ)1 for some δ > 0. (iv) ⇔ (v). By (2.6), kp − qkess = ksk and since s is positive and injective this implies that kp − qk = ksk < 1. The other direction is trivial. cs(cid:19) and (1, c−1s) =(cid:18)c2 (cid:3) . The equivalence of (i) and (iv) and the fact that then the index is zero were obtained in [AS94, Proposition 4]. Using this lemma it is now easy to obtain a characterization of Fredholm pairs also when the projections are not in generic position. Proposition 2.6. Let p, q be projections in B(H). Then (p, q) is a Fredholm pair if and only if kp − qkess < 1 and then p ∧ q⊥ and p⊥ ∧ q are finite and [p : q] = Tr(p ∧ q⊥) − Tr(p⊥ ∧ q). Proof. In the notations of Theorem 2.4 and (2.5) and (2.6) we have p = (p ∧ q + p ∧ q⊥) ⊕ p0 and q = (p ∧ q + p⊥ ∧ q) ⊕ q0 and hence (2.8) p − q = (p ∧ q⊥ − p⊥ ∧ q) ⊕ (p0 − q0). As in [AS94, Theorem 2], it is easy to see that (p, q) is a Fredholm pair if and only if both(cid:0)p∧ q + p∧ q⊥, p∧ q + p⊥ ∧ q(cid:1) and (p0, q0) are Fredholm pairs, and then the index of (p, q) is the sum of the indices of the other two pairs. It is obvious that KADISON'S PYTHAGOREAN THEOREM AND ESSENTIAL CODIMENSION 7 the first pair is Fredholm if and only if both p ∧ q⊥ and p⊥ ∧ q are finite and then its index is Tr(p ∧ q⊥) − Tr(p⊥ ∧ q). By Lemma 2.5, the pair (p0, q0) is Fredholm if and only if kp0 − q0kess < 1 and then its index is zero. Thus to conclude the proof one just notices that p − q is a finite rank perturbation of p0 − q0 and hence it has the same essential norm. (cid:3) The implication that if (p, q) is a Fredholm pair then p∧ q⊥ and p⊥ ∧ q are finite and the formula for the index of the pair was obtained in [AS94, Theorem 2]. The necessity and sufficiency of the condition kp − qkess < 1 (albeit not expressed in terms of essential norm) was only implicit in [ASS94], and was obtained explicitly and with more generality in [BCP+06]. We consider now the cases when the difference p − q belongs to some proper operator ideal J . Proposition 2.7. Let J be an operator ideal and p, q ∈ B(H) be projections. Then (i) p − q ∈ J if and only if p⊥ ∧ q and p ∧ q⊥ are finite and s ∈ J (where p0 := p − p ∧ q − p ∧ q⊥ =(cid:18)c2 cs cs s2(cid:19) as by (2.4)). (ii) If p − q ∈ J , then [p : q] = 0 if and only if there is a unitary u ∈ 1 + J such that uqu∗ = p. Proof. (i). By (2.8) we see that p − q ∈ J if and only if p ∧ q⊥ ∈ J , p⊥ ∧ q ∈ J and p0 − q0 ∈ J . By (2.6), the last condition holds if and only if s ∈ J , and the conclusion then follows from the fact that a projection belongs to a proper ideal if and only if it is finite. (ii). First notice that from part (i) and Proposition 2.6 it follows that (2.9) p − q ∈ J and [p : q] = 0 ⇔ Tr(p ∧ q⊥) = Tr(p⊥ ∧ q) < ∞ and s ∈ J . s u2q0u∗2. Furthermore, Therefore if there is a unitary in 1 +J such that uqu∗ = p , then it is immediate to see that p − q ∈ J and [p : q] = 0. Conversely, assume that p − q ∈ J and [p : q] = 0. Then by (2.9) s ∈ J , p⊥ ∧ q + p ∧ q⊥ is finite, and Tr(p ∧ q⊥) = Tr(p⊥ ∧ q) and hence p ∧ q⊥ ∼ p⊥ ∧ q. Then there is a unitary u1 on (cid:0)p ∧ q⊥ + p⊥ ∧ q)H such that u1(p⊥ ∧ q)u∗1 = p ∧ q⊥. Let u2 := (cid:18)c −s c (cid:19). Then by (2.7), u2 is a unitary on (p0 ∨ q0)H) and p0 = u2 − 1 (p0∨q0)H=(cid:18)c − 1 −s c − 1 = −s2(cid:0)c + 1(cid:1)−1 Let u3 be the identity on (cid:0)p ∧ q⊥ + p⊥ ∧ q + p0 ∨ q0(cid:1)⊥. Then u := u1 ⊕ u2 ⊕ u3 is unitary and uqu∗ = p. Since u1 has finite rank, we conclude that u − 1 ∈ J . Property (ii) in the above proposition was obtained for J = K in [BL12, Theorem Using Proposition 2.7 (i) we obtain an independent proof of the first part of Theorem 1.2 which extends it to arbitrary proper operator ideals, and establishes the sufficiency of the conditions listed. c − 1(cid:19) ∈ J ∈ J 2 ⊂ J . because 2.7]. (cid:3) s 8 VICTOR KAFTAL AND JIREH LOREAUX Proposition 2.8. Let J be a proper ideal of B(H) and let p, q be projections in B(H). Then the following are equivalent: (i) p − q ∈ J (ii) q(p − q)q ∈ J 2 and q⊥(p − q)q⊥ ∈ J 2 (iii) The projections p ∧ q⊥ and p⊥ ∧ q are finite and at least one the conditions q(p − q)q ∈ J 2 and q⊥(p − q)q⊥ ∈ J 2 holds. Furthermore if p − q ∈ L2, then Proof. From (2.8) we see that [p : q] = Tr(q(p − q)q + q⊥(p − q)q⊥). 0 (2.10) q(p − q)q = −p⊥ ∧ q ⊕(cid:18)−s2 s2(cid:19) . Thus q(p − q)q ∈ J 2 (resp., q⊥(p − q)q⊥ ∈ J 2) if and only if p⊥ ∧ q is finite and s ∈ J (resp., p ∧ q⊥ is finite and s ∈ J ). By Proposition 2.7 (i) it is now obvious that the conditions (i)-(iii) are equivalent. Assume now that p − q ∈ L2, then by the implication (i) ⇒ (ii) we see that 0(cid:19) and q⊥(p − q)q⊥ = p ∧ q⊥ ⊕(cid:18)0 0 0 0 q(p − q)q ∈ L1 and q⊥(p − q)q⊥ ∈ L1. Then by (2.10) Tr(q(p − q)q + q⊥(p − q)q⊥) = Tr(p ∧ q⊥) − Tr(p⊥ ∧ q) Tr(cid:18)s2 0 0 0(cid:19) = Tr(cid:18)0 0 0 s2(cid:19) < ∞ as s2 ∈ L1. The conclusion then follows from Proposition 2.6. The equivalence of just (i) and (ii) follows immediately from the identity (cid:3) (p − q)2 = −q(p − q)q + q⊥(p − q)q⊥ because whence (2.11) ((p − q)+)2 = q⊥(p − q)q⊥ and ((p − q)−)2− = −q(p − q)q. To conclude this survey, we observe that every Fredholm operator can be asso- ciated in a natural way to a Fredholm pair of projections (p, q) so that the index of the operator equals the index of the pair. To this end, consider any Fredholm operator x : H → K and scale x to have norm 1. After choosing an arbitrary infinite, co-infinite projection q and identifying K with qH, we have the following proposition. Proposition 2.9. Suppose x : H → qH is a contraction with q an infinite, co- infinite projection. Then x can be completed to an isometry w : H → H (i.e., x = qw), and for any such completion, if x is Fredholm then (p := ww∗, q) is a Fredholm pair with [p : q] = ind x. Proof. Consider the defect operator (1 − x∗x)1/2 and a partial isometry v taking R1−x∗x to vv∗ ≤ q⊥. Then define w = x + v(1 − x∗x)1/2. Since v∗x = 0 = x∗v, a simple computation shows w∗w = 1, which establishes that x can be completed to an isometry. Now suppose w is any such completion, and hence x = qw. Define p := ww∗ and notice x = qpw = (q pH)w when the operators are viewed on the appropriate spaces. Then from (2.1) and (2.2), if x is Fredholm, so is q pH and ind x = ind(q pH) = [p : q]. (cid:3) KADISON'S PYTHAGOREAN THEOREM AND ESSENTIAL CODIMENSION 9 Remark 2.10. We note that any completion of a contraction x : H → qH to an isometry arises in the manner above. Indeed, suppose w′ is such a completion. Set y := w′ − x and note that qy = 0 since x = qw′. Thus y∗y = (w′)∗w′ − (w′)∗x − x∗w′ + x∗x = 1 − 2(w′)∗qw′ + x∗x = 1 − x∗x. In particular, y = v′(1− x∗x)1/2 for some partial isometry v′ with (v′)∗v′ = R1−x∗x and v′(v′)∗ = Ry ≤ q⊥. Moreover, u = q + v(v′)∗ is a partial isometry for which w = uw′. Another perspective of Proposition 2.9 is that p is a dilation of xx∗ to H for which ind x = [p : q]. Indeed, p := ww∗ is a dilation of xx∗ because if y = w − x, then with respect to the decomposition q + q⊥ = 1 ww∗ =(cid:18)xx∗ xy∗ yy∗(cid:19) . yx∗ 2.1. Breuer Fredholm. As mentioned in the introduction, the essential codimen- sion/relative index of projections has found its main application in the study of spectral flows in Fredholm modules. However, in many cases of interest the Fred- holm modules are with respect to a semifinite von Neumann algebra (see [CPS03], [CP04], [BCP+06],[CPRS06]). The following short summary may be of interest to the reader. Let M be a semifinite von Neumann algebra with separable predual (but not necessarily a factor), τ a faithful semifinite normal trace, and let Jτ (M ) the ideal of τ -compact operators Jτ (M ) := span{x ∈ M(A ⊗ K)+ τ (x) < ∞} (norm closure). Let π : M → M/Jτ (M ) be the canonical quotient map and let kxkess := kπ(x)k be the essential norm. Then and element x ∈ M is called τ -Breuer Fredholm (also called just τ -Fredholm) if π(x) is invertible. A necessary and sufficient condition is that τ (Nx) < ∞ (where Nx is the projection on the kernel of x) and that there exists a projection e ∈ M with τ (e) < ∞ such that (1− e)H ⊂ xH. Then the index is defined as and satisfies the expected properties of an index, but of course it is no longer integer valued. ind(x) = τ (Nx) − τ (Nx∗ ) ∈ R The original definition by Breuer was given in terms of the ideal J (M ) of com- pact operators on M which received considerable attention over the years, J (M ) := span{x ∈ M(A ⊗ K)+ Rx is finite} (norm closure). When M is a factor and hence has a unique trace (up to normalization), the notions of τ -Breuer -- Fredholm and Breuer -- Fredholm coincides, but for global algebras they do not and so their theory had to be partially re-derived in [BCP+06]. With these definitions almost all of the results listed here for B(H) hold with the same statements and mostly with the same proofs. So we will briefly list here only the properties that fail or that require a different proof. Proposition 2.6 holds with the same statements and a natural modification of the proof of [ASS94, Proposition 3.1] in the case that M is a factor, but required more work for the general case [BCP+06, Lemma 4.1] less the trace condition which is only relevant when M is a factor. q2n+1 q2n + ∞ 2n c := s := (r 1 X1 X1 r1 − q =(cid:18)1 0 0 0(cid:19) and 1 2n 2n + 1 ∞ X0 r1 − X0 r 1 p =(cid:18)c2 s2(cid:19) . cs cs 2n + 1 q2n + q2n+1 10 VICTOR KAFTAL AND JIREH LOREAUX It is still true that if p, q ∈ M are in generic position and form a Fredholm pair then [p : q] = 0, but contrary to Lemma 2.5, we can have kp − qk = 1 and g and c are only invertible modulo Jτ (M ) as the following example shows. Example 2.11. Let M be a type II∞ factor. Let q ∼ q⊥ ∼ 1 be a projection in M which we decompose into a sum P∞1 qn = q of mutually orthogonal finite projections such that τ(cid:0)P∞1 q2n(cid:1) < ∞. Let ∞ ∞ 1 Then q and p are in generic position and kp− qk = ksk = 1 while s ∈ J (M ), hence p − q ∈ J (M ), and (p, q) is a Fredholm pair with respect to M . Proposition 2.7 (ii) holds without any changes if M is a factor, but does not hold for global algebras. Consider for instance M := B(H1) ⊕ B(H2) and τ := Tr⊕ Tr. Let q, p be rank one projections in B(H1) and B(H2) respectively. Then q and p are not equivalent (with respect to M ) and hence a fortiori they are not unitarily equivalent. On the other hand p − q =∈ J (M ) = K(H1) ⊕ K(H2), hence p, q is a Fredholm pair and furthermore [p : q] = Tr(p) − Tr(q) = 0. 3. The Kadison theorem and some applications We begin by using the tools developed in Section §2 to identify the integer a − b in Kadison's theorem, that is, to prove Theorem 1.3. Proof. By (1.2) a − b = Tr(cid:0)qpq − q + q⊥pq⊥(cid:1) = Tr(cid:0)q(p − q)q + q⊥(p − q)q⊥(cid:1) and by (1.1), q(p − q)q ∈ L1 and q⊥(p − q)q⊥ ∈ L1. Thus by Proposition 2.8, p − q ∈ L2 and [p : q] = Tr(q(p − q)q + q⊥(p − q)q⊥). (cid:3) As a first consequence of Kadison's theorem and of the work in Section §2, we observe that if the diagonal of a projection p clusters sufficiently fast around 0 and 1 (that is, if a + b < ∞, or, equivalently, if p − q ∈ L2), then one can "read" from the diagonal the essential codimension [p : q]. But what if a + b = ∞? If a = ∞ and b < ∞, from q⊥(p − q)q⊥ ∈ L1 we can deduce that p ∧ q⊥ is finite and s ∈ L2, and hence from q(p − q)q 6∈ L1 it follows that p⊥ ∧ q is infinite. Similarly, if a < ∞ and b = ∞ then p ∧ q⊥ is infinite. In either case (p, q) is not a Fredholm pair and in particular, p − q 6∈ K. Less trivial is the case when a = b = ∞ and p − q ∈ K \ L2, as we see from the following proposition. We first need to introduce two notations. Given an orthonormal basis {en}, denote by E the conditional expectation on the algebra of diagonal operators, namely E(x) is the diagonal of the operator x ∈ B(H). Next, given two sequences ξ and η of non-negative numbers converging to 0, with ξ∗ and η∗ their monotone non-increasing rearrangements, we say that ξ is majorized by η (ξ ≺ η) if Pn j=1 ξ∗j ≤Pn j=1 η∗j for all n. KADISON'S PYTHAGOREAN THEOREM AND ESSENTIAL CODIMENSION 11 Proposition 3.1. Suppose p, q are projections with p−q ∈ K\L2. Then there exists a projection p′ such that p′ − q ∈ K, [p′ : q] 6= [p : q] and there is an orthonormal basis {en} that diagonalizes q such that E(p) = E(p′). Proof. By Proposition 2.7 (i), p−q ∈ K implies that p∧q⊥ and p⊥∧q are both finite. Thus p0−q0 = (p−q)+(p∧q⊥−p⊥∧q) ∈ K\L2. It suffices to prove the proposition for projections in generic position because then we simply set p′ := p∧q +p∧q⊥ +p′0 for the general case. So to simplify notation, assume henceforth that p, q are in generic position and have the form as in (2.4). In particular, q ∼ q⊥ ∼ 1 and by Lemma 2.5, [p : q] = 0. Next, choose a rank one projection r′ ≤ q⊥ and let r := q⊥−r′. After identifying B(H) ≃ B(qH ⊕ rH ⊕ C) with M2(qB(H)q) ⊕ C via a partial isometry taking qH → rH, consider the projection c2 cs 0 p :=  cs s2 0 0 0 1  . Note that p − (0 ⊕ 0 ⊕ 1) and q are in generic position relative to their join, and hence their essential codimension is zero. This implies that [p, q] = 1 6= 0 = [p, q]. Now, choose an orthonormal basis {en} that diagonalizes q, E its corresponding conditional expectation, and let ξ be the diagonal sequence of s2. Then under natural notations we have E(p) = Eq(p) ⊕ Eq⊥ (p) and furthermore, Eq(p) = Eq(p) = Eq(c2) and Eq⊥ (p) = Eq(s2) = diag ξ. Since p − q ∈ K \ L2, by Proposition 2.7 we have that s2 ∈ K \ L1, that is ξ → 0 but ξ 6∈ ℓ1. By the Schur -- Horn theorem for compact operators [KW10, Proposition 6.4] ξ is majorized by the eigenvalues sequence λ(s2) of the operator s2 and hence, ξ ≺ λ(s2) ≺ λ(cid:18)s2 0 0 1(cid:19) . 0 Now(cid:18)s2 Hence by [KW10, Proposition 6.6] there is a unitary u ∈ B(q⊥H) such that 1(cid:19) is a positive compact operator with zero kernel belonging to B(q⊥H). 0 diag ξ = Eq⊥H(cid:0)u(cid:18)s2 1(cid:19) u∗(cid:1). Let u′ := 1 qH ⊕u and p′ := u′(p)u′∗. Then E(p′) = Eq(c2) ⊕ diag ξ = E(p). 0 0 Since u′qu′∗ = q we have [p′ : q] = [upu∗ : uqu∗] = [p : q] 6= [p, q]. (cid:3) As a second application of Theorem 1.3 and of the techniques used to prove it, we will consider a recent work by Bownik and Jasper [BJ15]. Based on Kadi- son's characterization of diagonals of projections, Bownik and Jasper character- ized the diagonals of selfadjoint operators with finite spectrum and in a key part of their analysis they too encountered an index obstruction similar to the one in Theorem 1.1 (ii). Following their notations, if z ∈ B(H) is a selfadjoint operator 12 VICTOR KAFTAL AND JIREH LOREAUX with finite spectrum we let σ(z) = {aj}n+r projection corresponding to the eigenvalue aj, so that j=−m and pj = χ{aj}(z) be the spectral n+r For ease of notations perform if necessary a transformation so to have z = ajpj. Xj=−m Tr(pj) < ∞ for j < 0 and j > n + 1, a0 = 0, and an+1 = 1. Let {en} be an orthonormal basis, {dn} be the diagonal of z with respect to that basis and let as in Theorem 1.1, a = Xdn≤1/2 dn and b = Xdn>1/2 (1 − dn), Then their Theorem 4.1, which is a key component of the necessity part of their characterization, states that Theorem 3.2 ([BJ15, Theorem 4.1]). If a + b < ∞ then whether Tr(p0) is finite or not. (i) Tr(pj) < ∞ for 0 < j < n + 1; (ii) a − b −Pj6=n+1 aj Tr(pj) ∈ Z. Here of course we use the convention that 0 · ∞ = 0 and hence a0 Tr(p0) = 0 We will present an independent proof of this result and at the same time identify the integer in (ii) proving that if we set q as in Theorem 1.1 to be the projection on span{ej dj > 1/2}, then (3.1) aj Tr(pj) = [pn+1 : q]. a − b − Xj6=n+1 First we need an extension to positive elements of the equivalence of (i) and (ii) in Proposition 2.8. Lemma 3.3. Let J be a proper ideal, x ∈ B(H)+ a positive contraction, and q ∈ B(H) a projection. (i) If q − qxq ∈ J and q⊥xq⊥ ∈ J , then x − q ∈ J 1/2 and xχ[0,ǫ](x) ∈ J for (ii) Assume that x is a projection or that J is idempotent (i.e., J = J 2). If x − q ∈ J 1/2 and xχ[0,ǫ](x) ∈ J for some 0 < ǫ < 1, then q − qxq ∈ J and q⊥xq⊥ ∈ J . every 0 < ǫ < 1. Proof. (i). Since q⊥xqxq⊥ ≤ q⊥x2q⊥ ≤ q⊥xq⊥ ∈ J , it follows that qxq⊥ and q⊥xq belong to J 1/2. But then x − q = (qxq − q) + q⊥xq⊥ + qxq⊥ + q⊥xq ∈ J 1/2. Let ǫ > 0 and let xǫ := xχ[0,ǫ](x). Then 0 ≤ q⊥xǫq⊥ ≤ q⊥xq⊥ ∈ J , whence q⊥xǫq⊥ ∈ J . Furthermore, 1 − x ≥ (1 − x)χ[0,ǫ](x) ≥ (1 − ǫ)χ[0,ǫ](x) ≥ (1 − ǫ)xǫ and hence q − qxq ≥ (1 − ǫ)qxǫq. Thus qxǫq ∈ J and since 0 ≤ xǫ ≤ 2(cid:0)qxǫq + q⊥xǫq⊥) ∈ J it follows that xǫ ∈ J . KADISON'S PYTHAGOREAN THEOREM AND ESSENTIAL CODIMENSION 13 (ii). The case when x is a projection is given by (2.11). Assume then that J is idempotent. Then x − q ∈ J 1/2 = J implies that q − qxq = −q(x − q)q ∈ J . Furthermore, (x− q)2 = x2 − xq − qx + q hence q⊥x2q⊥ = q⊥(x− q)2q⊥ ∈ J . Then q⊥(x − xǫ)q⊥ ≤ and hence 1 ǫ q⊥(x − xǫ)2q⊥ ≤ q⊥x2q⊥ ∈ J 1 ǫ q⊥xq⊥ = q⊥(x − xǫ)q⊥ + q⊥xǫq⊥ ∈ J . (cid:3) 1/2 Notice that if J is not idempotent and k ∈ J Now we can proceed with the proof of Theorem 3.2 and (3.1). + \J is a positive contraction, then x := 1−k and q := 1 satisfy both hypotheses of Lemma 3.3 (ii) but k = q−qxq 6∈ J . Proof. Set x =Pn+1 z − x = j=1 ajpj. Then 0 ≤ x ≤ 1 and ajpj has finite rank. ajpj + −1 n+r Xj=−m Xj=n+2 As in (1.2) we have that a = Tr(q⊥xq⊥) + Tr(q⊥(z − x)q⊥) = Tr(q⊥zq⊥) and b = Tr(q − qxq) − Tr(q(z − x)q) = Tr(q − qzq), hence q(z − q)q ∈ L1 , q⊥(z − q)q⊥ ∈ L1, and (3.2) a − b = Tr(cid:0)q(z − q)q + q⊥(z − q)q⊥(cid:1). We also have q(x − q)q ∈ L1 and q⊥(x − q)q⊥ ∈ L1, hence by Lemma 3.3, it follows that x − q ∈ L2 and Xj=1 j=1 ajpj has finite rank and in particular, Tr(pj) < ∞ for ajpj = xχ[0,an](x) ∈ L1. 0 < j < n + 1, thus proving (i). As a consequence, But then x − pn+1 = Pn n and hence by Proposition 2.8, pn+1 − q = pn+1 − x + x − q ∈ L2 (3.3) [pn+1 : q] = Tr(cid:0)q(pn+1 − q)q + q⊥(pn+1 − q)q⊥(cid:1). Furthermore, y := z − pn+1 = Pj6=n+1 ajpj has finite rank and in particular is in L1, so that (3.4) aj Tr(pj) = Tr(y) = Tr(qyq + q⊥yq⊥). Xj6=n+1 Finally from (3.2) and (3.3), a − b = Tr(cid:0)q(pn+1 − q)q + q⊥(pn+1 − q)q⊥ + qyq + q⊥yq⊥(cid:1) = [pn+1 : q] + Tr(y). Thus by (3.4), a − b −Pj6=n+1 aj Tr(pj) = [pn+1 : q] ∈ Z. (cid:3) 14 VICTOR KAFTAL AND JIREH LOREAUX References [Arg15] Martín Argerami, Majorisation and the carpenter's theorem, Integral Equations Oper. Theory 82 (2015), no. 1, 33 -- 49 (English). [Arv07] William Arveson, Diagonals of normal operators with finite spectrum, Proc. Natl. Acad. Sci. USA 104 (2007), no. 4, 1152 -- 1158 (English). [AS94] Werner Amrein and Kalyan B. Sinha, On pairs of projections in a Hilbert space, Linear [ASS94] Algebra Appl. 208/209 (1994), 425 -- 435. MR 1287363 Joseph E. Avron, Rudolf Seiler, and Barry Simon, The index of a pair of projections, J. Funct. Anal. 120 (1994), no. 1, 220 -- 237. MR 1262254 [BCP+06] Moulay-Tahar Benameur, Alan L. Carey, John Phillips, Adam Rennie, Fyodor A. Sukochev, and Krzysztof P. Wojciechowski, An analytic approach to spectral flow in von Neumann algebras, Analysis, geometry and topology of elliptic operators (Matthias Lesch, Bernhelm Boo-Bavnbek, Slawomir Klimek, and Weiping Zhang, eds.), World Sci. Publ., Hackensack, NJ, 2006, pp. 297 -- 352. MR 2246773 [BJ15] [BL12] [BDF73] Lawrence G. Brown, Ronald George Douglas, and Peter Arthur Fillmore, Unitary equivalence modulo the compact operators and extensions of C ∗-algebras, Proceedings of a Conference on Operator Theory (Peter Arthur Fillmore, ed.), Lecture Notes in Mathematics, vol. 345, Springer, Berlin, 1973, pp. 58 -- 128. MR 0380478 (52 #1378) Marcin Bownik and John Jasper, The Schur-Horn theorem for operators with finite spectrum, Trans. Amer. Math. Soc. 367 (2015), no. 7, 5099 -- 5140. MR 3335412 Lawrence G. Brown and Hyun Ho Lee, Homotopy classification of projections in the corona algebra of a non-simple C ∗-algebra, Canad. J. Math. 64 (2012), no. 4, 755 -- 777. MR 2957229 Lawrence G. Brown, Ext of certain free product C ∗-algebras, J. Operator Theory 6 (1981), no. 1, 135 -- 141. MR 637007 Alan Carey and John Phillips, Spectral flow in Fredholm modules, eta invariants and the JLO cocycle, K-Theory 31 (2004), no. 2, 135 -- 194. MR 2053481 [Bro81] [CP04] [CPRS06] Alan L. Carey, John Phillips, Adam Rennie, and Fyodor A. Sukochev, The local index formula in semifinite von Neumann algebras. II. The even case, Adv. Math. 202 (2006), no. 2, 517 -- 554. MR 2222359 [CPS03] Alan Carey, John Phillips, and Fyodor A. Sukochev, Spectral flow and Dixmier traces, [Dav58] [Dix48] [Eff89] [Hal69] Adv. Math. 173 (2003), no. 1, 68 -- 113. MR 1954456 Chandler Davis, Separation of two linear subspaces, Acta Sci. Math. Szeged 19 (1958), 172 -- 187. MR 0098980 Jacques Dixmier, Position relative de deux variétés linéaires fermées dans un espace de Hilbert, Revue Sci. 86 (1948), 387 -- 399. MR 0029095 (10,546e) Edward G. Effros, Why the circle is connected: an introduction to quantized topology, Math. Intelligencer 11 (1989), no. 1, 27 -- 34. MR 979021 (90d:46084) Paul R. Halmos, Two subspaces, Trans. Amer. Math. Soc. 144 (1969), 381 -- 389. MR 0251519 (40 #4746) [Kad02a] Richard V. Kadison, The Pythagorean Theorem I: the finite case, Proc. Natl. Acad. Sci. USA 99 (2002), no. 7, 4178 -- 4184. [Kad02b] , The Pythagorean Theorem II: the infinite discrete case, Proc. Natl. Acad. Sci. [Kat66] USA 99 (2002), no. 8, 5217 -- 5222. Tosio Kato, Perturbation theory for linear operators, Die Grundlehren der mathema- tischen Wissenschaften, Band 132, Springer-Verlag New York, Inc., New York, 1966. MR 0203473 [KKM48] M. G. Kreın, M. A. Kranoselskiı, and D. P. Milman, Defect numbers of linear operators in banach space and some geometrical problems, Sobor. Trudov. Insst. Mat. Akad. Nauk SSSR (1948), 97 -- 112 (Russian). [KNZ09] Victor Kaftal, Ping Wong Ng, and Shuang Zhang, Strong sums of projections in von Neumann factors, J. Funct. Anal. 257 (2009), no. 8, 2497 -- 2529. MR 2555011 (2011b:46096) Victor Kaftal and Gary Weiss, An infinite dimensional Schur -- Horn Theorem and majorization theory, J. Funct. Anal. 259 (2010), no. 12, 3115 -- 3162. [KW10] KADISON'S PYTHAGOREAN THEOREM AND ESSENTIAL CODIMENSION 15 [Str81] Şerban Strătilă, Modular theory in operator algebras, Editura Academiei Republicii Socialiste România, Bucharest; Abacus Press, Tunbridge Wells, 1981, Translated from the Romanian by the author. MR 696172 [Tak02] Masamichi Takesaki, Theory of operator algebras. I, Encyclopaedia of Mathematical Sciences, vol. 124, Springer-Verlag, Berlin, 2002, Reprint of the first (1979) edition, Operator Algebras and Non-commutative Geometry, 5. MR 1873025 (2002m:46083) Department of Mathematics, University of Cincinnati, P. O. Box 210025, Cincin- nati, OH, 45221-0025, USA E-mail address: [email protected] Department of Mathematics and Statistics, Southern Illinois University Edwardsville, 1 Hairpin Dr, Edwardsville, IL, 62026-1653, USA E-mail address: [email protected]
1708.03987	2	1708	2017-08-25T17:17:57	The pure extension property for discrete crossed products	[ "math.OA" ]	Let $G$ be a discrete group acting on a unital $C^$-algebra $\mathcal{A}$ by $$-automorphisms. In this note, we show that the inclusion $\mathcal{A} \subseteq \mathcal{A} \rtimes_r G$ has the pure extension property (so that every pure state on $\mathcal{A}$ extends uniquely to a pure state on $\mathcal{A} \rtimes_r G$) if and only if $G$ acts freely on $\mathcal{\widehat{A}}$, the spectrum of $\mathcal{A}$. The same characterization holds for the inclusion $\mathcal{A} \subseteq \mathcal{A} \rtimes G$. This generalizes what was already known for $\mathcal{A}$ abelian.	math.OA	math	THE PURE EXTENSION PROPERTY FOR DISCRETE CROSSED PRODUCTS VREJ ZARIKIAN Abstract. Let G be a discrete group acting on a unital C ∗-algebra A by ∗-automorphisms. In this note, we show that the inclusion A ⊆ A ⋊ rG has the pure extension property (so that every pure state on A extends uniquely r G) if and only if G acts freely on bA, the spectrum of A. to a pure state on A ⋊ The same characterization holds for the inclusion A ⊆ A ⋊G. This generalizes what was already known for A abelian. . A O h t a m [ 2 v 7 8 9 3 0 . 8 0 7 1 : v i X r a 1. Introduction Let A ⊆ B be a C ∗-inclusion, i.e., an inclusion of unital C ∗-algebras, with 1A = 1B. We say that A ⊆ B has the pure extension property (PEP) if every pure state on A extends uniquely to a (pure) state on B. The PEP was first considered by Kadison and Singer in [13], where they showed that L∞[0, 1] ⊆ B(L2[0, 1]) fails the PEP, and asked whether or not ℓ∞ ⊆ B(ℓ2) has the PEP. The latter question, famously unsolved for over 50 years, was settled affirmatively by Marcus, Spielman, and Srivastava in [14]. Initially, the study of the PEP focused on the case of A abelian [3, 5, 7]. More recently, the general case has received attention [8, 4]. Thanks to these efforts, we have various characterizations of the PEP, and know that it entails significant struc- tural consequences for the inclusion. Very recently, several authors have advanced our understanding of the PEP for specific classes of inclusions [18, 2, 1, 17, 16]. This note continues the aforementioned line of inquiry, by characterizing (in terms of the dynamics) when the inclusion A ⊆ A ⋊rG (resp. A ⊆ A ⋊G) has the PEP. Here G is a discrete group acting on a unital C ∗-algebra A by ∗- automorphisms (abbreviated G y A), and A ⋊rG (resp. A ⋊G) is the resulting reduced (resp. full) crossed product. When needed, αg ∈ Aut(A) will denote the ∗-automorphism corresponding to g ∈ G. If A is abelian, then A ∼= C(X), the continuous complex-valued functions on a compact Hausdorff space X. In that case, the answer is already known. Indeed, an action G y C(X) by ∗-automorphisms corresponds to an action G y X by homeomorphisms. The inclusion C(X) ⊆ C(X) ⋊ G has the PEP if and only if G y X is free (meaning that {x ∈ X : g · x = x} = ∅, for all e 6= g ∈ G) [7, Cor. 2010 Mathematics Subject Classification. Primary: 47L65, 46L55, 46L30 Secondary: 46L07. Key words and phrases. Pure state extension property, crossed product C ∗-algebra, free action, discrete group. 1 2 VREJ ZARIKIAN 6.2] (see also [19, Thm. 3.3.7] and [18, Prop. 5.11]). In order to state our result, we remind the reader that the spectrum of a unital C ∗-algebra A is the set bA of all unitary equivalence classes of non-zero irreducible representations of A, equipped with the final topology induced by the surjection P S(A) → bA : φ 7→ [πφ] arising from the GNS construction. An action G y A by ∗-automorphisms determines an action G y bA by homeomorphisms. Namely g· [π] = [π◦ αg], for g ∈ G and π : A → B(H) a non-zero irreducible representation. Our result (Theorem 2.4 below) says that A ⊆ A ⋊rG (resp. A ⊆ A ⋊G) has the PEP if and only if G y bA is free (meaning that {[π] ∈ bA : [π ◦ αg] = [π]} = ∅, for all e 6= g ∈ G). If A = C(X), then there is canonical homeomorphism bA ∼= X, and so we have generalized the abelian case. It should be noted that when moving from the abelian to the general case, many inequivalent notions of a "free action" present themselves, each with its own ben- efits and limitations [15]. Theorem 2.4 singles out one of these notions as being harmonious with the PEP. This paper can be regarded as a companion to our earlier paper [21], with which it shares many techniques. There we determine when the inclusion A ⊆ A ⋊rG (resp. A ⊆ A ⋊G) has a unique conditional expectation. This happens if and only if G y A is free1 [21, Thm. 3.1.2]. Likewise, we determine when the inclusion A ⊆ A ⋊rG (resp. A ⊆ A ⋊G) has a unique pseudo-expectation, in the sense of Pitts [16]. This happens if and only if G y A is properly outer2 [21, Thm. 3.2.2]. For an arbitrary C ∗-inclusion A ⊆ B, we have that PEP =⇒ unique pseudo-expectation =⇒ at most one conditional expectation. Thus, in passing, we re-prove the known implications G y bA is free =⇒ G y A is properly outer =⇒ G y A is free. 2. The Main Result In this section we prove our main result, Theorem 2.4. Before doing so, we will need a few preparatory lemmas. Our first preliminary result will make it slightly easier to determine when G y bA is free. Lemma 2.1. Let A be a unital C ∗-algebra and α ∈ Aut(A). Then {[π] ∈ bA : [π ◦ α] = [π]} = ∅ if and only if for every non-zero irreducible representation π : A → B(H) and every T ∈ B(H), T π(a) = π(α(a))T, a ∈ A =⇒ T = 0. 1We say that an action G y A of a discrete group on a unital C ∗-algebra by ∗-automorphisms is free if αg ∈ Aut(A) has no non-zero dependent elements, for all e 6= g ∈ G. That is, if b ∈ A and ba = αg(a)b for all a ∈ A, then b = 0, unless g = e. 2We say that G y A is properly outer if for all e 6= g ∈ G, the only αg -invariant ideal J ⊆ A such that αgJ is quasi-inner is J = {0}. PEP FOR CROSSED PRODUCTS 3 Proof. (⇒) Let π : A → B(H) be a non-zero irreducible representation, and sup- pose there exists T ∈ B(H) such that Arguing as in [9], we see that T π(a) = π(α(a))T, a ∈ A . T ∗T = T T ∗ ∈ π(A)′ = C I. U π(a) = π(α(a))U, a ∈ A . Thus, either T = 0 or there exists a unitary U ∈ B(H) such that Since {[π] ∈ bA : [π ◦ α] = [π]} = ∅, we conclude that T = 0. (⇐) Conversely, suppose there exists a non-zero irreducible representation π : A → B(H) such that [π◦ α] = [π]. Then there exists a unitary U ∈ B(H) such that π(α(a)) = U π(a)U ∗, a ∈ A . It follows that U π(a) = π(α(a))U, a ∈ A . (cid:3) Our second preliminary result is a minor variation of [20, Thm. 4.5], on decom- posing a completely bounded (CB) bimodule map into completely positive (CP) bimodule maps. The main difference is that π in Lemma 2.2 need not be faithful. The proof consists of an elementary reduction to the faithful case, so that [20, Thm. 4.5] can be invoked. Lemma 2.2. Let A ⊆ B be a C ∗-inclusion, π : A → B(H) be a unital ∗- homomorphism, and θ : B → B(H) be a CB map. Assume that θ is A-bimodular with respect to π, meaning that for all a ∈ A and x ∈ B, we have that θ(ax) = π(a)θ(x) and θ(xa) = θ(x)π(a). Then θ = (θ1 − θ2) + i(θ3 − θ4), where θj : B → B(H) is a CP map which is A-bimodular with respect to π, for all 1 ≤ j ≤ 4. Proof. We may assume that B ⊆ B(K), for some Hilbert space K. Define π : A → B(K) ⊕ B(H) by the formula and θ : B → B(K) ⊕ B(H) by the formula π(a) = a ⊕ π(a), a ∈ A, θ(x) = x ⊕ θ(x), x ∈ B . Then π is a faithful ∗-homomorphism, and θ is a CB map which is A-bimodular with respect to π. Since B(K) ⊕ B(H) is injective, [20, Thm. 4.5] applies to show that θ = (θ1 − θ2) + i(θ3 − θ4), where θj : B → B(K) ⊕ B(H) is a CP map which is A-bimodular with respect to π, for all 1 ≤ j ≤ 4. For each 1 ≤ j ≤ 4, we have that θj = αj ⊕ θj, where αj : B → B(K) and θj : B → B(H) are CP maps. One easily verifies that θj is A-bimodular with respect to π, for all 1 ≤ j ≤ 4, and that θ = (θ1 − θ2) + i(θ3 − θ4). (cid:3) Our third and last preliminary result is a bimodule version of [11, Lemma 5.1.6], on factoring a completely positive (CP) map as a unital completely positive (UCP) map followed by a conjugation. The proof is nearly identical. 4 VREJ ZARIKIAN Lemma 2.3. Let A ⊆ B be a C ∗-inclusion, π : A → B(H) be a unital ∗- homomorphism, and θ : B → B(H) be a CP map which is A-bimodular with respect to π. Then there exists a UCP map θ : B → B(H) which is A-bimodular with respect to π, and such that θ(x) = θ(1)1/2 θ(x)θ(1)1/2, x ∈ B . In particular, θA = π. Proof. Since θ is A-bimodular with respect to π, we have that θ(1) ∈ π(A)′. We claim that SOT− lim n→∞ (θ(1) + 1/n)−1/2θ(x)(θ(1) + 1/n)−1/2 (θ(1) + 1/n)−1/2θ(x)1/2 = SOT− lim exists for all x ∈ B. Indeed, if 0 ≤ x ≤ 1, then 0 ≤ θ(x) ≤ θ(1), which implies θ(x)1/2 = tθ(1)1/2 = θ(1)1/2t∗ for some t ∈ B(H). Then SOT− lim where p ∈ π(A)′ is the support projection of θ(1). Likewise SOT− lim It follows that θ(x)1/2(θ(1) + 1/n)−1/2 = SOT− lim (θ(1) + 1/n)−1/2θ(1)1/2t∗ = pt∗, tθ(1)1/2(θ(1) + 1/n)−1/2 = tp. n→∞ n→∞ n→∞ n→∞ SOT− lim n→∞ (θ(1) + 1/n)−1/2θ(x)(θ(1) + 1/n)−1/2 = pt∗tp, which proves the claim. Now for x ∈ B, define θ(x) = SOT− lim n→∞ (θ(1) + 1/n)−1/2θ(x)(θ(1) + 1/n)−1/2 + p⊥Φ(x)p⊥, where Φ : B → B(H) is any fixed UCP extension of π. Then θ is a UCP map which is A-bimodular with respect to π, and such that θ(1)1/2 θ(x)θ(1)1/2 = θ(x), x ∈ B . Indeed, if 0 ≤ x ≤ 1, then (keeping the notation from above) we have that θ(1)1/2 θ(x)θ(1)1/2 = θ(1)1/2(pt∗tp + p⊥Φ(x)p⊥)θ(1)1/2 = θ(1)1/2t∗tθ(1)1/2 = θ(x). (cid:3) Finally we state and prove the main result. Theorem 2.4. Let G be a discrete group acting on a unital C ∗-algebra A by ∗- automorphisms. Then the following are equivalent: i. A ⊆ A ⋊G has the PEP; ii. A ⊆ A ⋊rG has the PEP; iii. G y bA is free. Proof. (i =⇒ ii) The PEP passes to quotients, by [5, Lemma 3.1]. (ii =⇒ iii) Suppose A ⊆ A ⋊rG has the PEP. Fix e 6= g ∈ G, and assume that π : A → B(H) is a non-zero irreducible representation, such that [π ◦ αg] = [π]. Then there exists a unitary U ∈ B(H) such that π(αg(a)) = U π(a)U ∗, a ∈ A . Define a CB map θ : A ⋊rG → B(H) by the formula θ(x) = π(E(xg−1))U, x ∈ A ⋊rG, PEP FOR CROSSED PRODUCTS 5 where E : A ⋊rG → A is the canonical conditional expectation. Then for all a ∈ A and x ∈ A ⋊rG, we have that θ(ax) = π(E(axg−1))U = π(a E(xg−1))U = π(a)π(E(xg−1))U = π(a)θ(x). Likewise, θ(xa) = π(E(xag−1))U = π(E(xg−1gag−1))U = π(E(xg−1αg(a)))U = π(E(xg−1)αg(a))U = π(E(xg−1))π(αg(a))U = π(E(xg−1))U π(a) = θ(x)π(a). That is, θ is A-bimodular with respect to π. Furthermore, θ(g) = U 6= 0. By Lemma 2.2, θ = (θ1 − θ2) + i(θ3 − θ4), where θj : A ⋊rG → B(H) is a CP map which is A-bimodular with respect to π, for 1 ≤ j ≤ 4. Without loss of generality, θ1(g) 6= 0. By Lemma 2.3, there exists a UCP map θ1 : A ⋊rG → B(H) which is A-bimodular with respect to π, and such that θ1(x) = θ1(1)1/2 θ1(x)θ1(1)1/2, x ∈ A ⋊rG. Obviously θ1(g) 6= 0. Let ξ ∈ H be a unit vector such that hθ1(g)ξ, ξi 6= 0. Then a 7→ hπ(a)ξ, ξi is a pure state on A with distinct state extensions x 7→ hπ(E(x))ξ, ξi and x 7→ hθ1(x)ξ, ξi on A ⋊rG, a contradiction. (iii =⇒ i) Suppose G y bA is free. Let φ ∈ P S(A). Then the corresponding GNS representation πφ : A → B(Hφ) is a non-zero irreducible representation. Let Φ ∈ S(A ⋊G) be an extension of φ, and let πΦ : A ⋊G → B(HΦ) be the resulting GNS representation. For all a ∈ A, we have that πΦ(a)V = V πφ(a), where V : Hφ → HΦ is the unique isometry such that Taking adjoints, V πφ(a)ξφ = πΦ(a)ξΦ, a ∈ A . V ∗πΦ(a) = πφ(a)V ∗, a ∈ A . Now fix e 6= g ∈ G. For all a ∈ A, we have that gag−1 = αg(a) =⇒ ga = αg(a)g =⇒ πΦ(g)πΦ(a) = πΦ(αg(a))πΦ(g) =⇒ V ∗πΦ(g)πΦ(a)V = V ∗πΦ(αg(a))πΦ(g)V =⇒ V ∗πΦ(g)V πφ(a) = πφ(αg(a))V ∗πΦ(g)V. Since {[π] ∈ bA : [π ◦ αg] = [π]} = ∅, we have that V ∗πΦ(g)V = 0, by Lemma 2.1. Thus for any a ∈ A, V ∗πΦ(ag)V = V ∗πΦ(a)πΦ(g)V = πφ(a)V ∗πΦ(g)V = 0. Therefore Φ(ag) = hπΦ(ag)ξΦ, ξΦi = hπΦ(ag)V ξφ, V ξφi = hV ∗πΦ(ag)V ξφ, ξφi = 0. Since a ∈ A and e 6= g ∈ G were arbitrary, Φ = φ ◦ E, where E : A ⋊G → A is the canonical conditional expectation. (cid:3) 6 VREJ ZARIKIAN 3. An Example In this final section, we use Theorem 2.4 to analyze the inclusion O2 ⊆ O2 ⋊ Z2, showing that it fails the PEP, although it has many of the features of a PEP in- clusion. Here O2 is the Cuntz algebra, i.e., the universal C ∗-algebra generated by isometries s0, s1 satisfying the relation s0s∗ 1 = 1, and the action of Z2 = {0, 1} on O2 switches the generators. More precisely, α0 = id and α1 is the unique ∗- automorphism of O2 such that α1(s0) = s1. This action was studied in detail by Choi and Latr´emoli`ere in [10], and we benefit tremendously from their insights. In particular, they show that O2 ⋊ Z2 ∼= O2 [10, Thm. 2.1]. Thus O2 ⋊ Z2 is simple, and we need not distinguish between O2 ⋊ Z2 and O2 ⋊r Z2. 0 +s1s∗ By [10, Prop. 2.2], Z2 y O2 is outer (i.e., not inner). In fact, since O2 is simple, Z2 y O2 is properly outer, and therefore free (see [21, Rmk. 4.1.3]). Also, by [12, Ex. 5.7], Z2 y O2 has the Rokhlin property. On the other hand, Z2 y bO2 is not free. Indeed, as shown in [10, App. A], there is a non-zero irreducible representation π : O2 → B(L2[−1, 1]) such that [π ◦ α1] = [π]. Namely, for f ∈ L2[−1, 1] and t ∈ [−1, 1], (π(s0)f )(t) = √2f (2t − 1)χ[0,1](t) (π(s1)f )(t) = √2f (2t + 1)χ[−1,0](t). and Then π(α1(x)) = U π(x)U ∗, x ∈ O2, where U ∈ B(L2[−1, 1]) is the unitary operator defined by (U f )(t) = f (−t), f ∈ L2[−1, 1], t ∈ [−1, 1]. (It would be interesting to know if Z2 y bO2 is topologically free, in the sense of [6].) In light of the previous paragraph, we draw the following conclusions: • O2 ⊆ O2 ⋊ Z2 admits a unique conditional expectation [21, Thm. 3.1.2]; • O2 ⊆ O2 ⋊ Z2 admits a unique pseudo-expectation, in the sense of Pitts • O2 ⊆ O2 ⋊ Z2 fails the PEP, by Theorem 2.4. [21, Thm. 3.2.2]; References [1] Akemann, Charles A.; Sherman, David Conditional expectations onto maximal abelian ∗- subalgebras. J. Operator Theory 68 (2012), no. 2, 597-607. [2] Akemann, Charles; Wassermann, Simon; Weaver, Nik Pure states on free group C ∗-algebras. Glasg. Math. J. 52 (2010), no. 1, 151-154. [3] Anderson, Joel Extensions, restrictions, and representations of states on C ∗-algebras. Trans. Amer. Math. Soc. 249 (1979), no. 2, 303-329. [4] Archbold, R. J. Extensions of pure states and projections of norm one. J. Funct. Anal. 165 (1999), no. 1, 2443. [5] Archbold, R. J.; Bunce, J. W.; Gregson, K. D. Extensions of states of C ∗-algebras. II. Proc. Roy. Soc. Edinburgh Sect. A 92 (1982), no. 1-2, 113-122. [6] Archbold, R. J.; Spielberg, J. S. Topologically free actions and ideals in discrete C ∗-dynamical systems. Proc. Edinburgh Math. Soc. (2) 37 (1993), no. 1, 119-124. [7] Batty, C. J. K. Simplexes of extensions of states of C ∗-algebras. Trans. Amer. Math. Soc. 272 (1982), no. 1, 237-246. [8] Bunce, L. J.; Chu, C.-H. Unique extension of pure states of C ∗-algebras. J. Operator Theory 39 (1998), no. 2, 319-338. PEP FOR CROSSED PRODUCTS 7 [9] Choda, Marie; Kasahara, Isamu; Nakamoto, Ritsuo Dependent elements of an automorphism of a C ∗-algebra. Proc. Japan Acad. 48 (1972), 561-565. [10] Choi, Man-Duen; Latr´emoli`ere, Fr´ed´eric Symmetry in the Cuntz algebra on two generators. J. Math. Anal. Appl. 387 (2012), no. 2, 1050-1060. [11] Effros, Edward G.; Ruan, Zhong-Jin Operator spaces. London Mathematical Society Mono- graphs. New Series, 23. The Clarendon Press, Oxford University Press, New York, 2000. xvi+363 pp. [12] Izumi, Masaki Finite group actions on C ∗-algebras with the Rohlin property. I. Duke Math. J. 122 (2004), no. 2, 233-280. [13] Kadison, Richard V.; Singer, I. M. Extensions of pure states. Amer. J. Math. 81 (1959), 383-400. [14] Marcus, Adam W.; Spielman, Daniel A.; Srivastava, Nikhil Interlacing families II: Mixed characteristic polynomials and the Kadison-Singer problem. Ann. of Math. (2) 182 (2015), no. 1, 327-350. [15] Phillips, N. Christopher Freeness of actions of finite groups on C ∗-algebras. Operator struc- tures and dynamical systems, 217-257, Contemp. Math., 503, Amer. Math. Soc., Providence, RI, 2009. [16] Pitts, David R. Structure for regular MASA inclusions. I. J. Operator Theory, to appear. [17] Popa, Sorin A II1 factor approach to the Kadison-Singer problem. Comm. Math. Phys. 332 (2014), no. 1, 379-414. [18] Renault, Jean Cartan subalgebras in C ∗-algebras. Irish Math. Soc. Bull. No. 61 (2008), 29-63. [19] Tomiyama, Jun Invitation to C ∗-algebras and topological dynamics. World Scientific Ad- vanced Series in Dynamical Systems, 3. World Scientific Publishing Co., Singapore, 1987. x+167 pp. [20] Wittstock, Gerd Ein operatorwertiger Hahn-Banach Satz. (German) [An operator-valued Hahn-Banach theorem] J. Funct. Anal. 40 (1981), no. 2, 127-150. [21] Zarikian, Vrej Unique expectations for discrete crossed products, preprint, arXiv:1707.09339v1 [math.OA], 28 July 2017. Department of Mathematics, U. S. Naval Academy, Annapolis, MD 21402 E-mail address: [email protected]
1811.02692	1	1811	2018-11-06T22:11:23	A second countable locally compact transitive groupoid without open range map	[ "math.OA" ]	Dana P. Williams raised in [Proc. Am. Math. Soc., Ser. B, 2016] the following question: Must a second countable, locally compact, transitive groupoid have open range map? This paper gives a negative answer to that question. Although a second countable, locally compact transitive groupoid G may fail to have open range map, we prove that we can replace its topology with a topology which is also second countable, locally compact, and with respect to which G is a topological groupoid whose range map is open. Moreover, the two topologies generate the same Borel structure and coincide on the fibres of G.	math.OA	math	A second countable locally compact transitive groupoid without open range map Madalina Roxana Buneci Abstract. Dana P. Williams raised in [Proc. Am. Math. Soc., Ser. B, 2016] the following question: Must a second countable, locally compact, transitive groupoid have open range map? This paper gives a negative answer to that question. Although a second countable, locally compact transitive groupoid G may fail to have open range map, we prove that we can replace its topology with a topology which is also second countable, locally compact, and with respect to which G is a topological groupoid whose range map is open. Moreover, the two topologies generate the same Borel structure and coincide on the fibres of G.. 1. Introduction In order to construct convolution algebras associated to a locally compact topo- logical groupoid one needs an analogue of the Haar measure on a locally compact group. Starting with the work of Jean Renault [7], this analogue is a system of mea- sures, called Haar system, subject to suitable support, invariance and continuity conditions. According to a result of Anthony Seda [8], the continuity assumption is essential for the Renault's construction [7] of the C ∗-algebra associated to a locally compact groupoid. This continuity assumption entails that the range map is open ([9, Proposition I.4] or [8, p. 118]). As Williams pointed out [10, Question 3.5], while there certainly exist groupoids that fail to have open range maps, most of these examples are group bundles which are as far as from been principal groupoids or transitive groupoids as possible. This led him to the next question [10, Question 3.5]: must a second countable, locally compact, transitive groupoid have open range map? In this paper we construct a second countable, locally compact, transitive and principal groupoid that fails to have open range map (and hence open domain map). In addition, we prove that for every second countable, locally compact topology τ on a transitive groupoid G making G a topological groupoid, there is a topology τ which is also second countable, locally compact, and with respect to which G is a topological groupoid with open range map. Moreover, τ and τ generate the same Borel structure and coincide on the fibres of G. 2010 Mathematics Subject Classification. Primary 22A22; Secondary 54E15. Key words and phrases. transitive groupoid, principal groupoid, locally compact groupoid, range map. 1 2 M AD ALINA ROXANA BUNECI We shall use the definition of a topological groupoid given by Jean Renault in [7]. For a groupoid G, G(2) will denote the set of the composable pairs and G(0) its unit space. As usual the inverse map will be written x → x−1 [: G → G] and the product map will be written (x, y) → xy (cid:2): G(2) → G(cid:3). For each x ∈ G, r (x) = xx−1, respectively, d (x) = x−1x will denote the range, respectively the domain (source) of x in G(0) (thus r : G → G(0), respectively d : G → G(0) will be the range map, respectively the domain/source map). For each u ∈ G(0), the fibre of the range, respectively domain map over u is denoted Gu = r−1 ({u}), respectively, Gu = d−1 ({u}). A groupoid G is said to be transitive if for every u,v ∈ G(0), there is x ∈ G such that r (x) = u and d (x) = v. A groupoid is called principal if the map (r, d) : G → G(0) × G(0), defined by (r, d) (x) = (r (x) , d (x)) for all x ∈ G, is injective. If A, B ⊂ G and x ∈ G, one may form the following subsets of G: A−1 = (cid:8)x ∈ G : x−1 ∈ A(cid:9) AB = nxy : (x, y) ∈ G(2) ∩ (A × B)o xA = {x}A and Ax = A {x} . A topological groupoid consists of a groupoid G and a topology compatible with the groupoid structure. This means that the inverse map x 7→ x−1 [: G → G] is continuous, as well as the product map (x, y) 7→ xy (cid:2): G(2) → G(cid:3) is continuous, where G(2) has the induced topology induced from G × G. By a locally compact groupoid we mean a topological groupoid whose topology is locally compact (Haus- dorff) 2. A second countable locally compact transitive groupoid without open range map Let us modify the usual topology of the space of real numbers R in the points of the form 3 2n+2 and 5 3 2n+2 with n ∈ N. For every x ∈ R, let 2n+2 , 2n+2 − ε, 2n+2 + ε(cid:1) , ε > 0(cid:9) , if x = 3 2n+2(cid:3) , ε > 0(cid:9) , if x = 5 {(x − ε, x + ε) , ε > 0} , if x ∈ R \ (cid:8)(cid:2) 3 (cid:8)(cid:0) 5 2n+2 (n ∈ N) 2n+2 (n ∈ N) 5 2n+2 , ∞ 5 Sn=0(cid:8) 3 2n+2(cid:9) . Fx = {V ⊂ R : there is U ∈ Bx such that U ⊂ V } Bx =  and Let τ0 = {O ⊂ R: if x ∈ O, then O ∈ Fx} be the unique topology on X = R with the property that for every x ∈ X, Fx is the family of neighborhoods of x. Then (X, τ0) is a Hausdorff second countable 2n+2 , small enough such that [x − ε, x + ε] ∩ topological space. Moreover if x ∈ R \ (cid:18) ∞ Sn=0(cid:8) 3 Sn=0(cid:8) 3 2n+2(cid:3) is a compact neighborhood of 2n+1 , then (cid:2) 3 is a compact neighborhood of x with respect to the topology τ0. Also if n ∈ N and 0 < ε < 1 2n+2 and 2n+2 − ε, 2n+2 + ε(cid:3) is a compact neighborhood of 2n+2(cid:9) ∪ {0}(cid:19) and ε > 0 is 2n+2(cid:9) = ∅, then [x − ε, x + ε] (cid:2) 5 2n+2 . 2n+2 , 2n+2 , ∞ 5 5 3 5 5 3 A LOCALLY COMPACT TRANSITIVE GROUPOID WITHOUT OPEN RANGE MAP 3 Let G = X × X = R × R be the pair groupoid (product: (x, y) (y, z) = (x, z), inverse: (x, y)−1 = (y, x)). For every (x, y) ∈ G let B(x,y) = {A × B : (x, y) ∈ A × B, A, B ∈ τ0} , if x 6= 0 and y 6= 0 {{0} × B : y ∈ B, B ∈ τ0} , if x = 0 and y 6= 0 {A × {0} : x ∈ A, A ∈ τ0} , if x 6= 0 and y = 0 {{(0, 0)} ∪ Un : n ∈ N} , if (x, y) = (0, 0) ,   2k+2(cid:3) for all n ∈ N. F(x,y) =(cid:8)V ⊂ R × R : there is U ∈ B(x,y) such that U ⊂ V(cid:9) . 2k+2(cid:3) ×(cid:2) 3 Sk=n(cid:2) 3 where Un = 2k+2 , 2k+2 , ∞ 5 5 Let us endow G = R × R with the unique topology τ1 with the property that for every (x, y) ∈ G, F(x,y) is the family of neighborhoods of (x, y). It is easy to see that the inverse map is continuous with respect to τ1. Since for all subsets A, B, C ⊂ X = R, and (A × C) (C × B) ⊂ A × B it follows that the product map is continuous in the points of the form ((x, y) , (y, z)) ∈ G(2) with x 6= 0 and z 6= 0. For every n ∈ N, let Un = ∞ 5 5 2k+2 , Sk=n(cid:2) 3 2k+2 , 2k+2(cid:3)×(cid:2) 3 2k+2(cid:3). The continuity of the product map in ((0, 0) , (0, 0)) is the consequence of the fact that ({(0, 0)} ∪ Un) ({(0, 0)} ∪ Un) = {(0, 0)} ∪ Un. Since for all subsets B ⊂ X = R, ({0} × B) (B × {0}) = {(0, 0)} ⊂ {(0, 0)} ∪ Un it follows that the product map is continuous is continuous in the points of the form ((0, y) , (y, 0)) such that y 6= 0. The fact that for all subsets B ⊂ X = R, ({(0, 0)} ∪ Un) ({0} × B) = {(0, 0)} ({0} × B) = {0} × B implies that the product map is continuous in the points of the form ((0, 0) , (0, y)) such that y 6= 0. Similarly, the product map is continuous in the points of the form ((y, 0) , (0, 0)) such that y 6= 0. Therefore (G, τ1) is a topological groupoid. For every x ∈ R \ {0}, let Kx be a compact neighborhood of x with respect to τ0. Then Kx × Ky is a compact neighborhood of (x, y) with respect to the topology τ1. For every x ∈ R \ {0}, {0} × Kx, respectively Kx × {0} is a compact neighborhood of (0, x), respectively (x, 0) with respect to τ1. Let us prove that for each m ∈ N, Km = {(0, 0)}[ ∞ [k=m(cid:20) 3 2k+2 , 5 2k+2(cid:21) ×(cid:20) 3 2k+2 , 5 2k+2(cid:21)! is a compact neighborhood of (0, 0). Let ((xn, yn))n ∈ Km. If {n ∈ N : (xn, yn) = (0, 0)} is infinite, then ((xn, yn))n has a subsequence converging to (0, 0) ∈ Km. If {n ∈ N : (xn, yn) = (0, 0)} is finite, then there is an integer n0 ≥ m such that (xn, yn) ∈ ∞ 2k+2 , Sk=n(cid:2) 3 2k+2 , 5 2k+2(cid:3) ×(cid:2) 3 5 2k+2(cid:3) for all n ≥ n0. In this case for every 4 M AD ALINA ROXANA BUNECI n ≥ n0, there are un, vn ∈ [3, 5] and kn ∈ N, such that xn = un 2kn +2 . If (kn)n is unbounded, then (kn)n has a subsequence that diverges to ∞ and hence (xn)n and (yn)n have subsequences which converge to 0 in the usual topology on R. Since for all n, 2kn +2 and yn = vn Km ∩(cid:18)(cid:18)− 5 2n+2 , 5 2n+2(cid:19) ×(cid:18)− 5 2n+2 , 5 2n+2(cid:19)(cid:19) ⊂ [k=n(cid:20) 3 ∞ 2k+2 ⊂ {(0, 0)}[ , 5 2k+2(cid:21) ×(cid:20) 3 2k+2 , 5 2k+2(cid:21) , it follows that ((xn, yn))n has a subsequence converging to (0, 0) with respect to τ1. If (kn)n is bounded, then it has a convergent subsequence in the usual topology on R, or equivalently a stationary subsequence. Also (un)n and (vn)n have convergent subsequences in the usual topology on R. Thus (xn)n (respectively, (yn)n) has v 2p+2 ) in the usual topology on R. a subsequence converging to u 2p+2 (respectively, Since u, v ∈ [3, 5] , it follows that every neighborhood of(cid:0) u the topology τ1 contains a set of the form A × B, with A, B ∈ τ0 and(cid:0) u A × B. Therefore ((xn, yn))n has a subsequence converging to (cid:0) u topology τ1 of G = X × X = R × R. Thus Km is a compact neighborhood of (0, 0). Therefore (G, τ1) is a second countable locally compact groupoid. Since for any 2p+2(cid:1) with respect to 2p+2(cid:1) ∈ 2p(cid:1) in the 2p+2 , 2p , v 2p+2 , v v B ∈ τ0, r ({0} × B) = {(0, 0)} is not open in G(0) = {(x, x) : x ∈ X}, it follows that the range map is not open (and hence the domain map d is not open). 3. Replacing the topology of a locally compact transitive groupoid with a locally transitive topology We prove that for every second countable, locally compact, transitive groupoid G there is a second countable, locally compact topology τ making G a topological groupoid with open range map. Moreover, the original topology on G and τ gen- erate the same Borel structure and coincide on the r-fibres and d-fibres of G. The topology τ is in fact the topology τW introduced in [2, Definition 3.1], where W is a suitable G-uniformity. Let us recall that a G-uniformity (in the sense of [2, Definition 2.1]) is a collection {W }W ∈W of subsets of a groupoid G satisfying the following conditions: (1) G(0) ⊂ W ⊂ G for all W ∈ W. (2) If W1, W2 ∈ W, then there is W3 ⊂ W1 ∩ W2 such that W3 ∈ W. (3) For every W1 ∈ W there is W2 ∈ W such that W2W2 ⊂ W1. (4) W = W −1for all W ∈ W. If G is groupoid endowed with a topology, then a G-uniformity W is said to be compatible with the topology of the r-fibres (in the sense of [2, Definition 3.4]) if for every u ∈ G(0) and every open neighborhood U of u, there is W ∈ W such that W ∩ Gu ⊂ U ∩ Gu and u is in the interior of W ∩ Gu with respect to the topology on Gu coming from (G, τ ) . A subset K of G is diagonally compact (in the sense of [5, p. 10]) if K ∩ r−1 (L) and K ∩ d−1 (L) are compact whenever L is a compact subset of G(0). A LOCALLY COMPACT TRANSITIVE GROUPOID WITHOUT OPEN RANGE MAP 5 Proposition 1. If G is a second countable, locally compact Hausdorff groupoid, then G admits a countable G-uniformity {Wn}n∈N compatible with the topology of the r-fibres such that for every n ∈ N, Wn is a diagonally compact neighborhood of G(0) Proof. Since G is a second countable, locally compact Hausdorff space, it follows that G is metrizable. Let us denote the metric by d. Also since G is a second countable, locally compact Hausdorff space it follows that G as well as G(0) are paracompact spaces. Thus G has a funda- mental system of diagonally compact neighborhoods of G(0) (by [5, Lemma 2.10/p. 10] or the proof of [7, Proposition 1.9]). For each n ∈ N \ {0} let us write Dn =(cid:26)x ∈ G : d (x, r (x)) < 1 n + 1(cid:27) . Let W0 be a diagonally compact symmetric neighborhood of G(0) such that W0 ⊂ D0. Inductively we construct a G-uniformity {Wn}n∈N consisting in diago- nally compact symmetric neighborhoods of G(0) such that Wn ⊂ Dn for all n ∈ N. Suppose a symmetric neighborhood Wn of G(0) has already been built. Let Vn be a diagonally compact neighborhood of G(0) such that Vn ⊂ Dn+1 ∩ Wn. Since G is paracompact, according to [6, p. 361-362], there is a neighborhood Un of G(0) such that UnUn ⊂ Vn. Let Wn+1 be a diagonally compact neighborhood of G(0) such that Wn+1 ⊂ Un. Replacing Wn+1 with Wn+1 ∩ W −1 n+1, we may assume that Wn+1 = W −1 n+1. Thus we obtain a diagonally compact symmetric neighborhood of G(0) such that Wn+1 ⊂ Wn+1Wn+1 ⊂ UnUn ⊂ Vn ⊂ Dn+1 ∩ Wn. Let us remark that for every u ∈ G(0) we have Wn ∩ Gu ⊂ Dn ∩ Gu =(cid:26)x ∈ Gu : d (x, u) < 1 n + 1(cid:27) for all n ∈ N. Consequently, {Wn}n∈N is compatible with the topology of the (cid:3) r-fibres. Let us recall that A ⊂ G is open with respect to the topology τW [2, Definition 3.1] (respectively, τ r W [1, p. 59]) associated to a G-uniformity W [2, Definition 2.1] if and only if for every x ∈ A there is Wx ∈ W such that WxxWx ⊂ A (respectively, xWx ⊂ A). If G is a topological groupoid and if W is a G-uniformity compatible with the topology of the r- fibres, then by [2, Proposition 3.6] for every W1 ∈ W and x ∈ G there is W2 ∈ W such that W2 ∩ Gd(x) d(x) ⊂ x−1W1x and by [2, Proposition 3.7], G endowed with τW is a topological groupoid. Moreover the topologies induced by τ r W and τW on r-fibres coincide. According [2, Proposition 3.8], the compatibility of the G-uniformity W with the topology of G ensures that the topologies induced by τW and the original topology of G on the r-fibres Gu coincide. Thus for each u ∈ G(0) and each x ∈ Gu, {xW }W ∈W is a neighborhood basis (local basis) for x with respect to the topology induced from G on Gu. Proposition 2. Let G be a topological transitive groupoid endowed with a G- uniformity W compatible with the topology of the r-fibres and let u ∈ G(0). If S ⊂ Gu is a dense subset of Gu with respect to the topology induced from G, then S−1S is a dense subset of G with respect to the topology τW 6 M AD ALINA ROXANA BUNECI Proof. Let W ∈ W and y ∈ G. Since G is transitive, it follows that there is x ∈ Gu such that d (x) = r (y). The density of S implies that there is s ∈ S such that s ∈ xW or equivalently, x ∈ sW −1 = sW . Using again the density of S in Gu and the fact that xy ∈ Gu, it follows that there is t ∈ S such that t ∈ xyW . Consequently, xy ∈ tW −1 = tW . Therefore y ∈ x−−1tW ⊂ W s−1tW or equivalently, s−1t ∈ W yW . Thus any neighborhood (with respect to τW ) of y contains at least one point from S−1S. (cid:3) A topological groupoid is said to be locally transitive [8, p. 119] (or groupoıde microtransitif [3]) if for every u ∈ G(0) the map ru is open, where ru : Gu → G(0) is defined by ru (x) = r (x) for all x ∈ Gu and Gu is endowed with the topology coming from G. Hence the maps du are open, where du : Gu → G(0), du (x) = d (x) for all x ∈ Gu. Obviously, every locally transitive groupoid has open range and domain maps. Conversely, according [4, Theorem 2.2 A amd Theorem 2.2 N] or [6, Theorem 3.2], if G is a second countable, locally compact, transitive groupoid with open range map, then G is locally transitive. The example constructed in Section 2 demonstrates that there are second countable, locally compact, transitive groupoids which are not locally transitive. The following theorem shows that if G is a second countable, locally compact, transitive groupoid G, then we can eventually replace the topology of G with a second countable, locally compact topology τ making G a locally transitive groupoid. In addition, the original topology on G ant τ generate the same Borel structure and coincide on the r-fibres (hence on d-fibres) of G. Theorem 1. Let G be a transitive groupoid endowed with second countable locally compact Hausdorff topology τ making G a topological groupoid. Then the topology τ of G can be replaced with a topology τ such that: (1) G is a (topological) locally transitive groupoid with respect to the topology τ (hence G has open range map with respect to τ ). (2) The topology τ is in general finer than τ . However τ and τ coincide if for every u ∈ G(0), rGu is open with (G, τ ) is locally transitive (i.e. respect to the topology induced by τ on Gu). (3) The topologies induced by τ and τ on r-fibres (respectively, on d-fibres) of G coincide. (4) The topology τ is second countable and locally compact Hausdorff. (5) The topologies τ and τ generate the same Borel structure on G (the Borel sets of a topological space are taken to be the σ-algebra generated by the open sets). Proof. According Proposition 1, G admits a countable G-uniformity W = {Wn}n∈N compatible with the topology of the r-fibres such that for every n ∈ N, Wn is a diagonally compact neighborhood of G(0). Let τ = τW . Then [2, Proposition 3.6] and [2, Proposition 3.7] show 1 and [2, Proposition 3.8] implies 2 and 3. Let us fix u ∈ G(0) and let {xn, n ∈ N} be a dense subset of Gu . ◦ 4. For each n, let Wn denote the interior of Wn with respect to τ . We claim W kx−1 m xn (cid:26) ◦ ◦ W k : m, n, k ∈ N(cid:27) is a countable base for τ = τW . that for every n, m, k ∈ N and x ∈ G, ◦ W kx ◦ W kis an open set with respect to Let us prove τW . Let s ∈ W k ∩ Gr(x)and t ∈ W k ∩ Gd(x). Since r (s) s ∈ ◦ ◦ ◦ W kand (G, τ ) is a A LOCALLY COMPACT TRANSITIVE GROUPOID WITHOUT OPEN RANGE MAP 7 topological groupoid, there is an open neighborhood V of r (s) (with respect to τ ) ◦ such that V s ⊂ W k. The fact that W = {Wn}n∈N is compatible with the topology of the r-fibres implies that there is p ∈ N such that Wp ∩ Gr(s) ⊂ V −1 ∩ Gr(s) and consequently, Wp ∩ Gr(s) ⊂ V ∩ Gr(s). Hence Wps ⊂(cid:0)Wp ∩ Gr(s)(cid:1) s ⊂ V s ⊂ ◦ W k. ◦ W k. If r ∈ N is such that Wr ⊂ Wp ∩ Wq, Similarly, there q ∈ N such that tWq ⊂ then sxt ∈ ◦ W rsxt ◦ W r ⊂ ◦ W kx ◦ W k. Thus ◦ W kx ◦ W kis an open set with respect to τW . Let us prove that(cid:26) ◦ W kx−1 m xn ◦ W k : m, n, k ∈ N(cid:27) is a base for τ = τW . Indeed, let x ∈ G and let A be an open subset of G with respect to τW such that x ∈ A. Then there is k ∈ N such that x ∈ WkxWk ⊂ A. Let r ∈ N such that WrWr ⊂ Wk . Proposition 2 implies that (cid:8)x−1 to τW . Hence there are m, n ∈ N such that x−1 m xn : m, n ∈ N(cid:9) is dense in G with respect ◦ W r, or equivalently, m xn ∈ ◦ W rx ◦ x ∈ W rx−1 m xn ◦ W r. Therefore x ∈ ◦ W rx−1 m xn ◦ W r ⊂ Wrx−1 m xnWr ⊂ Wr ◦ W rx ◦ W rWr ⊂ WkxWk ⊂ A. 5. Since the topology τ is finer than τ , it suffices to prove that each open set with respect to τ belong to the Borel structure generated by τ . But as we have proved every open set with respect to τ is a countable union of sets of the form Wrx−1 m xnWr which are compact with respect to τ (because Wr is diagonally compact). (cid:3) References [1] M. Buneci, Various notions of amenability for not necessarily locally compact groupoids, Surveys in Mathematics and its Applications, 9 (2014), 55-78. [2] M. Buneci, A Urysohn type lemma for groupoids, Theory and Applications of Categories 32 (28) (2017), 970-994. [3] C. Ehresmann, Cat´egories topologiques. I, II, III, (French) Nederl. Akad. Wetensch. Proc. Ser. A, Indag. Math. 28 1966. [4] P. Muhly, J. Renault and D. Williams, Equivalence and isomorphism for groupoid C ∗- algebras, J. Operator Theory 17 (1987), 3 -- 22. [5] P. Muhly and D. Williams, Renault's Equivalence Theorem for groupoid crossed products, NYJM Monographs 3, 2008. [6] A. Ramsay, The Mackey-Glimm dichotomy for foliations and other Polish groupoids, J. Funct. Anal. 94 (1990), 358-374. [7] J. Renault, A groupoid approach to C ∗- algebras, Lecture Notes in Math., Springer-Verlag, 793, 1980. [8] A. K. Seda, On the continuity of Haar measure on topological groupoids, Proc. Amer. Math. Soc., 96 (1986), 115-120. [9] J. Westman, Nontransitive groupoid algebras, Univ. of California at Irvine, 1967. [10] D. P. Williams, Haar systems on equivalent groupoids, Proc. Am. Math. Soc., Ser. B 3 (2016), 1-8. University Constantin Brancus¸i of Targu-Jiu, Calea Eroilor No. 30, 210135 Targu- Jiu, Romania. E-mail address: [email protected] URL: http://www.utgjiu.ro/math/mbuneci/
1604.07147	1	1604	2016-04-25T07:09:44	Local derivations on measurable operators and commutativity	[ "math.OA" ]	We prove that a von Neumann algebra $M$ is abelian if and only if the square of every derivation on the algebra $S(M)$ of measurable operators, affiliated with $M$, is a local derivation. We also show that for general associative unital algebras this is not true.	math.OA	math	LOCAL DERIVATIONS ON MEASURABLE OPERATORS AND COMMUTATIVITY SHAVKAT AYUPOV AND KARIMBERGEN KUDAYBERGENOV ABSTRACT. We prove that a von Neumann algebra M is abelian if and only if the square of every derivation on the algebra S(M ) of measurable operators affiliated with M is a local derivation. We also show that for general associative unital algebras this is not true. Keywords: von Neumann algebra, measurable operators, derivation, local derivation. AMS Subject Classification: 17A36, 46L57. 1. INTRODUCTION Let A be an algebra over the field of complex numbers. A linear operator D : A → A is called a derivation if it satisfies the identity D(xy) = D(x)y + xD(y) for all x, y ∈ A (Leibniz rule). Each element a ∈ A defines a derivation Da on A given by Da(x) = ax − xa, x ∈ A. Such derivations Da are said to be inner. If the element a implementing the derivation Da on A, belongs to a larger algebra B, containing A (as a proper ideal as usual) then Da is called a spatial derivation. One of the main problems considered in the theory of derivations is to prove the automatic continuity, innerness or spatialness of derivations, or to show the existence of non inner and discontinuous derivations on various topological algebras. In particular, it is a general algebraic problem to find algebras which admit only inner derivations. Such problems in the framework of algebras of measurable operators affiliated with von Neumann algebras has been considered in several papers (e.g. [1, 3 -- 6]). A linear operator ∆ on an algebra A is called local derivation if given any x ∈ A there ex- ists a derivation D (depending on x) such that ∆(x) = D(x). The main problem concerning these operators is to find conditions under which local derivations become derivations and to present examples of algebras which admit local derivations that are not derivations (see e.g. [12], [14]). In particular Kadison in [12] proves that every continuous local derivation from a von Neumann algebra M into a dual M-bimodule is a derivation. This theorem gave rise to studies and several results on local derivations on C∗-algebras, culminating with a de- finitive contribution due to Johnson [11], who proved that every (not necessary continuous) local derivation of a C∗-algebra is a derivation. A comprehensive survey of recent results concerning local derivations on C∗- and von Neumann algebras is presented in [8]. In [2] we have studied local derivations on the algebra S(M, τ ) of all τ-measurable op- erators affiliated with a von Neumann algebra M and a faithful normal semi-finite trace τ. 2 AYUPOV AND KUDAYBERGENOV One of our main results in [2] (Theorem 2.1) presents an unbounded version of Kadison's Theorem A from [12], and it asserts that every local derivation on S(M, τ ) which is continu- ous in the measure topology automatically becomes a derivation. In particular, for type I von Neumann algebras M all such local derivations on S(M, τ ) are inner derivations. We also proved that for type I finite von Neumann algebras without abelian direct summands, as well as for von Neumann algebras with the atomic lattice of projections, the continuity condition on local derivations in the above theorem is redundant. For survey of results concerning local derivations on algebras of measurable operators we refer to [7]. The most interesting effects appear when we consider the problem of existence of local derivations which are not derivations. The consideration of such examples on various finite- and infinite dimensional algebras was initiated by Kadison, Kaplansky and Jensen (see [12]). We have considered this problem on a class of commutative regular algebras, which includes the algebras of measurable operators affiliated with abelian von Neumann algebras. Unlike C∗- and von Neumann algebras cases, when all derivations, and hence all local derivations, are trivial, the abelian algebras S(M) may admit non-zero derivations (see [9]). In [2, Theorem 3.2] the authors obtained necessary and sufficient conditions for the algebras of measurable and τ-measurable operators affiliated with a commutative von Neumann algebra to admit local derivations which are not derivations. In the latter paper we have proved some properties of derivations and local derivations which seem to be specific for the abelian case as it was noted by Professor E. Zelmanov (UCSD). In the present paper we confirm this conjecture and give some criteria for a von Neumann algebra M to be abelian in terms of properties of derivations and local derivations on the algebra S(M) of measurable operators affiliated with M. The following theorem is the main result of the paper. 2. THE MAIN RESULT Theorem 2.1. Let S(M) be the algebra of measurable operators affiliated with a von Neu- mann algebra M. The following conditions are equivalent: (a) M is abelian; (b) For every derivation D on S(M) its square D2 is a local derivation; (c) A linear map ∆ : S(M) → S(M) is a local derivation if and only if ∆(p) = 0 for all projection p in M, and s(∆(x)) ≤ s(x) for all x in S(M), where s(x) is the support projection of the element x. Let us first present some necessary definitions and facts from the theory of measurable operators, which will be used in the proof of this theorem. Let H be a Hilbert space over the field C of complex numbers and let B(H) be the algebra of all bounded linear operators on H. Denote by 1 the identity operator on H and let P(H) = {p ∈ B(H) : p2 = p2 = p∗} be the lattice of projections in B(H). Consider a von Neumann algebra M on H, i.e. a weakly closed -subalgebra in B(H) containing the operator 1 and denote by k · kM the operator norm on M. The set P(M) = P(H) ∩ M is a complete LOCAL DERIVATIONS AND COMMUTATIVITY 3 orthomodular lattice with respect to the natural partial order on Mh = {x ∈ M : x∗ = x}, generated by the cone M+ of positive operators from M. Two projections e, f ∈ P(M) are said to be equivalent (denoted by e ∼ f) if there exists a partial isometry v ∈ M with initial projection e and final projection f, i.e. v∗v = e, vv∗ = f. The relation " ∼ " is equivalence relation on the lattice P(M). A linear subspace D in H is said to be affiliated with M (denoted as DηM), if u(D) ⊂ D for every unitary u in the commutant M ′ = {y ∈ B(H) : xy = yx, ∀x ∈ M} of the von Neumann algebra M in B(H). A linear operator x on H with the domain D(x) is said to be affiliated with M (denoted as xηM) if D(x)ηM and u(x(ξ)) = x(u(ξ)) for all ξ ∈ D(x) and for every unitary u in M ′. A linear subspace D in H is said to be strongly dense in H with respect to the von Neu- mann algebra M, if 1) DηM; 2) there exists a sequence of projections {pn}∞ n = 1 − pn is finite in M for all n ∈ N. and p⊥ n=1 in P(M) such that pn ↑ 1, pn(H) ⊂ D A closed linear operator x acting in the Hilbert space H is said to be measurable with respect to the von Neumann algebra M, if xηM and D(x) is strongly dense in H. Denote by S(M) the set of all measurable operators with respect to M. It is well-known (see e.g. [15]) that the set S(M) is a unital -algebra when equipped with the algebraic operations of the strong addition and multiplication and taking the adjoint of an operator. For every x ∈ S(M) we set s(x) = l(x) ∨ r(x), where l(x) and r(x) are left and right supports of x respectively (see [15]). Let ∆ : S(M) → S(M) be a local derivation. Then (2.1) for every idempotent p ∈ M. p∆(p)p = 0 Indeed, for any idempotent p ∈ M, we have ∆(p) = Dp(p) = D(p2) = Dp(p)p + pDp(p) = ∆(p)p + p∆(p). Thus ∆(p) = ∆(p)p + p∆(p) for every idempotent p ∈ M. Multiplying both sides of the last equality by p we obtain (2.1). Let M be an abelian von Neumann algebra and let D be a derivation on S(M). By [9, Proposition 2.3] or [3, Theorem] we have that s(D(x)) ≤ s(x) for any x ∈ S(M) and also D(p) = 0 for any p ∈ P(M). Therefore by the definition, each local derivation ∆ on S(M) satisfies the following two conditions: (2.2) for all x ∈ S(M) and (2.3) s(∆(x)) ≤ s(x) ∆P(M ) = 0. 4 AYUPOV AND KUDAYBERGENOV This means that (2.2) and (2.3) are necessary conditions for a linear operator ∆ to be a local derivation on the algebra S(M). We are going to show that these two condition are in fact also sufficient. The following Lemma in a more general setting of regular commutative algebras was obtained in [2, Lemma 3.2]. For the sake of completeness below we an give an alternative and straightforward proof of this result. Lemma 2.2. Let M be an abelian von Neumann algebra. Then each linear operator ∆ on the algebra S(M) satisfying conditions (2.2) and (2.3) is a local derivation on S(M). Proof. Let us first assume that M is an abelian von Neumann algebra with a faithful normal finite trace τ. In this case the algebra S(M) is a complete metrizable regular algebra with respect to the metric ρ(x, y) = τ (s(x − y)), x, y ∈ S(M) (see [9, Example 2.6]). It is well-known that the abelian von Neumann algebra M can be identified with the al- gebra C(Ω) of all complex valued continuous functions on a hyperstonean compact space Ω. Let a ∈ S(M) be a fixed element. In order to proof that the linear map ∆ is a local derivation we must find a derivation D on S(M) such that ∆(a) = D(a). Let P [t] be the algebra of all polynomials at variable t over C. Consider the following subalgebra of S(M) : A(a) = {f (a) : f ∈ P [t], f (0) = 0} . Consider the linear operator D from A(a) into S(M) defined as (2.4) D(f (a)) = f ′(a)∆(a), where f ′ is the usual derivative of the polynomial f. It is clear that D is a derivation. Let us show that (2.5) s(D(f (a))) ≤ s(f (a)) for all f ∈ P [t] with f (0) = 0. Case 1. Suppose first that a is a bounded element, i.e. a ∈ M ≡ C(Ω). Take an arbitrary λktk ∈ P [t]. Denote e = 1 − s(f (a)). There exists a closed subset S of Ω such n f (t) = Pk=1 Pk=0 that e equals to the characteristic function χS of the set S. Since ef (a) = 0, it follows that n λka(t)k = 0 for all t ∈ S, that is the complex numbers a(t) are roots of f for all t ∈ S. Since the polynomial f may have at most n roots the set {a(t) : t ∈ S} is finite. This means that ea is a simple (step) function, i.e. it is a linear combination of projections. Since by (2.3) ∆(p) = 0 for all p ∈ P(M), it follows that ∆(ea) = 0. Further from the linearity of ∆ and properties of the support we have s (∆(a)) = s (∆((1 − e)a + ea)) = s (∆((1 − e)a) + ∆(ea)) = = s (∆((1 − e)a)) ≤ s(1 − e) = s(f (a)). Therefore s (D(f (a))) = s (f ′(a)∆(a)) ≤ s (∆(a)) ≤ s(f (a)). LOCAL DERIVATIONS AND COMMUTATIVITY 5 Case 2. Let a ∈ S(M) be an arbitrary element. There exists an increasing sequence en = 1. Taking into of projections {en} in M such that ena ∈ M for all n ∈ N and Wn∈N account the equalities enD(x) = D(enx), enf (a) = f (ena) (the second equality follows from f (0) = 0) and the Case 1, we obtain that ens (D(f (a))) = s (enD(f (a))) = s (D(enf (a))) = s (D(f (ena))) ≤ ≤ s (f (ena)) = s(enf (a)) = ens(f (a)) for all n. Thus s (D(f (a))) ≤ s(f (a)). Thus, the derivation D defined by (2.4) satisfies the condition (2.5). Since S(M) is a com- plete metrizable regular algebra, by [9, Theorem 3.1] the derivation D from A(a) into S(M) can be extended to a derivation on the whole algebra S(M), which we also denote by D. Taking the polynomial p(t) = t in the the definition (2.4) of D, we obtain that ∆(a) = D(a). This means that ∆ is a local derivation. Now let M be an arbitrary abelian von Neumann algebra. It is well-known that it has a faithful normal semifinite trace τ. There exists a family of mutually orthogonal projections {zi}i∈I in M such that Wi∈I Take x = zix ∈ ziS(M) ≡ S(ziM). The condition (2.2) implies that zi = 1 and τ (zi) < +∞ for all i ∈ I. s(∆(x)) = s(∆(zix)) ≤ s(zix) ≤ zi. Thus ∆(x) = zi∆(x) ∈ S(ziM). This means that ∆ maps S(ziM) into itself for all i ∈ I. Therefore, since ziM is an abelian von Neumann algebra with a faithful normal finite trace, from above it follows that the restriction ∆S(ziM ) of ∆ onto S(ziM) is a local derivation. Thus ∆ is also a local derivation. The proof is complete. (cid:3) Proof of Theorem 2.1. (a) ⇒ (c). Let M be abelian. Before Lemma 2.2 we already men- tioned that any local derivation on S(M) satisfies the conditions (2.2) and (2.3). Conversely, suppose that ∆ is a linear operator on S(M) such that s(∆(x)) ≤ s(x) for any x ∈ S(M) and ∆(p) = 0 for all p ∈ P(M). By Lemma 2.2 it follows that ∆ is a local derivation. (c) ⇒ (b). Let D be a derivation on S(M). Then D is a local derivation. By the assumption (c) we have that for all x in S(M) and s(D2(x)) = s(D(D(x)) ≤ s(D(x)) ≤ s(x) D2(p) = D(D(p)) = D(0) = 0 for all projection p in M. Applying (c) to a linear operator D2, we conclude that it is a local derivation. (b) ⇒ (a). Suppose that M be a noncommutative von Neumann algebra and a ∈ M be a non central element. By the assumption of Theorem a linear operator ∆ : S(M) → S(M) defined by ∆(x) = [a, [a, x]], x ∈ S(M) 6 AYUPOV AND KUDAYBERGENOV is a local derivation. By (2.1), we obtain that for every projection p ∈ M. Thus p[a, [a, p]]p = 0 i.e., (2.6) p(a2p + pa2 − 2apa)p = 0, pa2p = papap. Since M is noncommutative, there exists a non central projection e ∈ M. Then z(e)z(1 − e) 6= 0, where z(x) denotes the central support of the element x. By [13, Propo- sition 6.1.8] we can find mutually orthogonal, equivalent nonzero projections p, q ∈ M such that p ≤ e and q ≤ 1 − e. Take a partial isometry u ∈ M such that uu∗ = p and u∗u = q. Let us consider the element Using the equality u = uu∗u, we obtain that a = u + u∗. u2 = uu = (uu∗u)(uu∗u) = u(u∗u)(uu∗)u = uqpu = 0. Then and therefore (2.7) Further and therefore (2.8) a2 = uu∗ + u∗u = p + q, pa2p = p. pap = uu∗(u + u∗)uu∗ = uu∗uuu∗ + uu∗u∗uu∗ = 0, papap = 0. Combining (2.7), (2.8) and (2.6), we obtain a contradiction. This contradiction implies that M is commutative. The proof is complete. (cid:3) Remark. In [9] the authors have proved that if M is an abelian von Neumann algebra with a non-atomic lattice of projections P(M), then the algebra S(M) admits non-zero derivations. If D is such a derivation then from the above theorem we have that D2 is a non-zero local derivation. Moreover D2 is not a derivation, because from [2, Lemma 3.3] it follows that D2 is a derivation if and only if D is identically zero. λijeij! = [(a11 − a22)λ12 − a12(λ11 − λ22)] e12. P1≤i≤j D 2 X1≤i≤j λijeij! =(cid:2)(a11 − a22)2λ12 − (a11 − a22)a12(λ11 − λ22)(cid:3) e12 D2 2 X1≤i≤j Thus LOCAL DERIVATIONS AND COMMUTATIVITY 7 3. COUNTEREXAMPLE FOR GENERAL ASSOCIATIVE ALGEBRAS The following example shows that Theorem 2.1 and the last Remark fail in the general case of unital associative algebras. Example 3.1. Let T2(C) be the algebra of all upper triangular 2 × 2 matrices over C, i.e., T2(C) =(cid:26)(cid:18) λ11 λ12 λ22 (cid:19) : λij ∈ C, 1 ≤ i ≤ j ≤ 2(cid:27) . 0 Let e11, e12, e22 be the matrix units in T2(C). It is known [10] that any derivation on the algebra T2(C) is inner. Direct computations show that the inner derivation generated by an element a = aijeij ∈ T2(C) acts on T2(C) as 2 and D2 is an inner derivation generated by the matrix 11 + a2 22 0 (cid:18) a2 (a11 − a22)a12 2a11a22 (cid:19) . So, the square of every derivation on the algebra T2(C) is a derivation (and hence a local derivation), but T2(C) is non commutative. ACKNOWLEDGEMENTS The idea of this paper came during the visit of the first author to the University of Califor- nia, San Diego. The first author would like to thank Professor Efim Zelmanov for hospitality and useful discussions during the stay in UCSD. REFERENCES [1] S. Albeverio, Sh. A. Ayupov and K. K. Kudaybergenov, Structure of derivations on various algebras of measurable operators for type I von Neumann algebras, J. Funct. Anal. 256 (2009) 2917 -- 2943. [2] S. Albeverio, Sh. A. Ayupov, K. K. Kudaybergenov and B. O. Nurjanov, Local derivations on algebras of measurable operators, Comm. in Contem. Math. 13 (2011), 643 -- 657. [3] Sh.A.Ayupov, Derivations on algebras of measurable operators, Dokl.Akad. Nauk R.Uzbekistan 3 (2000), 14-17. [4] Sh. A.Ayupov and K.K.Kudaybergenov, Additive derivations on algebras of measurable operators. Jour- nal of Operator Theory, 67:2 (2012), 495-510. [5] Sh. A. Ayupov and K. K. Kudaybergenov, Innerness of continuous derivations on algebras of measurable operators affiliated with finite von Neumann algebras, J. Math. Anal. Appl. 408 (2013) 256-267. 8 AYUPOV AND KUDAYBERGENOV [6] Sh. A. Ayupov and K. K. Kudaybergenov, Spatiality of derivations on the algebra of τ-compact operators, Integr. Equ. Oper. Theory, 77 (2013), 581 -- 598 [7] Sh. A. Ayupov and K. K. Kudaybergenov, Derivations, local and 2-local derivations on algebras of measurable operators, Contemporary Mathematics AMS, (to appear). [8] Sh. A. Ayupov, K. K. Kudaybergenov and A. M. Peralta, A survey on local and 2-local derivations on C∗- and von Neumann algebras, Contemporary Mathematics, AMS, (to appear). [9] A. F. Ber, V. I. Chilin, and F. A. Sukochev, Non-trivial derivation on commutative regular algebras, Extracta Math. 21 (2006) 107 -- 147. [10] S. P. Coelho, C. P. Milies, Derivations of upper triangular matrix rings, Linear Algebra and its Applica- tions, 187 (1993) 263 -- 267. [11] B. E. Johnson, Local derivations on C∗-algebras are derivations, Trans. Amer. Math. Soc. 353 (2001) 313 -- 325. [12] R. V. Kadison, Local derivations, J. Algebra, 130 (1990) 494 -- 509. [13] R.V. Kadison, J.R. Ringrose, Fundamentals of the theory of operator algebras, Vol. II, Birkhauser Boston, 1986. [14] D. R. Larson and A. R. Sourour, Local derivations and local automorphisms of B(X), Operator theory: operator algebras and applications, part 2 (Durham,NH, 1988), 187 -- 194, Proc. Sympos. Pure Math. 51, Part 2, Amer.Math.Soc., Providence, RI, (1990). [15] M. A. Muratov and V. I. Chilin, Algebras of measurable and locally measurable operators, Institute of Mathematics Ukrainian Academy of Sciences 2007. E-mail address: sh−[email protected] DORMON YOLI 29, INSTITUTE OF MATHEMATICS, NATIONAL UNIVERSITY OF UZBEKISTAN, 100125 TASHKENT, UZBEKISTAN E-mail address: [email protected] CH. ABDIROV 1, DEPARTMENT OF MATHEMATICS, KARAKALPAK STATE UNIVERSITY, NUKUS 230113, UZBEKISTAN
1409.1107	3	1409	2016-09-15T14:49:13	Self-similar graphs, a unified treatment of Katsura and Nekrashevych C*-algebras	[ "math.OA" ]	Given a graph $E$, an action of a group $G$ on $E$, and a $G$-valued cocycle $\phi$ on the edges of $E$, we define a C-algebra denoted ${\cal O}_{G,E}$, which is shown to be isomorphic to the tight C-algebra associated to a certain inverse semigroup $S_{G,E}$ built naturally from the triple $(G,E,\phi)$. As a tight C-algebra, ${\cal O}_{G,E}$ is also isomorphic to the full C-algebra of a naturally occurring groupoid ${\cal G}_{tight}(S_{G,E})$. We then study the relationship between properties of the action, of the groupoid and of the C-algebra, with an emphasis on situations in which ${\cal O}_{G,E}$ is a Kirchberg algebra. Our main applications are to Katsura algebras and to certain algebras constructed by Nekrashevych from self-similar groups. These two classes of C-algebras are shown to be special cases of our ${\cal O}_{G,E}$, and many of their known properties are shown to follow from our general theory.	math.OA	math	SELF-SIMILAR GRAPHS A UNIFIED TREATMENT OF KATSURA AND NEKRASHEVYCH C-ALGEBRAS Ruy Exel and Enrique Pardo Given a graph E, an action of a group G on E, and a G-valued cocycle ϕ on the edges of E, we define a C-algebra denoted OG,E , which is shown to be isomorphic to the tight C-algebra associated to a certain inverse semigroup SG,E built naturally from the triple (G, E, ϕ). As a tight C-algebra, OG,E is also isomorphic to the full C-algebra of a naturally occurring groupoid Gtight(SG,E). We then study the relationship between properties of the action, of the groupoid and of the C-algebra, with an emphasis on situations in which OG,E is a Kirchberg algebra. Our main applications are to Katsura algebras and to certain algebras constructed by Nekrashevych from self-similar groups. These two classes of C-algebras are shown to be special cases of our OG,E , and many of their known properties are shown to follow from our general theory. 1. Introduction. The purpose of this paper is to give a unified treatment to two classes of C-algebras which have been studied in the past few years from rather different points of view, namely Katsura algebras [18], and certain algebras constructed by Nekrashevych [24], [26] from self-similar groups. The realization that these classes are indeed closely related, as well as the fact that they could be given a unified treatment, came to our mind as a result of our attempt to understand Katsura's algebras OA,B from the point of view of inverse semigroups. The fact, proven by Katsura in [18], that all Kirchberg algebras in the UCT class may be described in terms of his OA,B was, in turn, a strong motivation for that endeavor. While studying OA,B, it slowly became clear to us that the two matricial parameters A and B, present in Katsura's construction, play very different roles. The reader acquainted with Katsura's work will easily recognize that the matrix A is destined to be viewed as the edge matrix of a graph, but it took us much longer to realize that B should be thought of as providing parameters for an action of the group Z on the graph given by A. In trying to understand these different roles, some interesting arithmetic popped up sparking a connection with the work done by Nekrashevych [26] on the C-algebra O(G,X) associated to a self-similar group (G, X). Date: 27 September 2014. 2010 Mathematics Subject Classification: 46L05, 46L55. Key words and phrases: Kirchberg algebra, Katsura algebra, Nekrashevych algebra, tight represen- tation, inverse semigroup, groupoid, groupoid C-algebra. The first-named author was partially supported by CNPq. The second-named author was partially supported by PAI III grants FQM-298 and P11-FQM-7156 of the Junta de Andaluc´ıa and by the DGI- MICINN and European Regional Development Fund, jointly, through Project MTM2011-28992-C02-02. 2 r. exel and e. pardo While Nekrashevych's algebras contain a Cuntz algebra, Katsura's algebras contain a graph C-algebra. This fact alone ought to be considered as a hint that self-similar groups lie in a much bigger class, where the group action takes place on the path space of a graph, rather than on a rooted tree (which, incidentally, is the path space of a bouquet of circles). One of the first important applications of the idea of self-similarity in group theory is in constructing groups with exotic properties [15], [16]. Many of these are defined as subgroups of the group of all automorphisms of a tree. Having been born from automor- phisms, it is natural that the theory of self-similar groups generally assumes that the group acts faithfully on its tree (see, e.g. [26: Definition 2.1]). However, based on the example provided by Katsura's algebras, we decided that perhaps it is best to view the group on its own, the action being an extra ingredient. The main idea behind self-similar groups, namely the equation g(xw) = yh(w) (1.1) appearing in [26: Definition 2.1], and the subsequent notion of restriction, namely gx := h, depend on faithfulness, since otherwise the group element h appearing in (1.1) would not be unique and therefore will not be well defined as a function of g and x. Working with non-faithful group actions we were forced to postulate a functional dependence h = ϕ(g, x), and we were surprised to find that the natural properties expected of ϕ are that of a group cocycle. To be precise, the ingredients needed in our generalization of self-similar groups are: a countable discrete group G, an action G × E → E of G on a finite graph E = (E0, E1, r, d), and a one-cocycle ϕ : G × E1 → G for the action of G on the edges of E. Starting with this data (which we assume satisfies a few other natural axioms) we construct an action of G on the space of finite paths E∗ which satisfies the "self-similarity" equation g(αβ) = (gα)(cid:0)ϕ(g, α)β(cid:1), ∀ g ∈ E, ∀ α, β ∈ E∗. Adopting a philosophy similar to that embraced by Katsura and Nekrashevych, we define a C-algebra, denoted OG,E, inverse semigroup actions and self-similar graphs 3 in terms of generators and relations inspired by the above group action. The study of OG,E is, thus, the purpose of this paper. Given a self-similar group (G, X), if we consider X as the set of edges of a graph with a single vertex, and if we define ϕ(g, x) = gx, then our OG,E coincides with Nekrashevych's O(G,X). On the other hand, if we are given two integer N ×N matrices A and B, with Ai,j ≥ 0, for all i and j, we may form a graph E with vertex set E0 = {1, 2, . . . , N } and with Ai,j edges from vertex i to vertex j. We may then use B to define an action of Z on E, by fixing all vertices and acting on the set of edges as follows: denote the set of edges in E from i to j by {ei,j,n : 0 ≤ n < Ai,j}. Given m ∈ Z, and given an edge ei,j,n, in order to define σm(ei,j,n), we first perform the Euclidean division of mBi,j + n by Ai,j, say with 0 ≤ n < Ai,j. We then put mBi,j + n = kAi,j + n σm(ei,j,n) := ei,j,n, so that the group element m permutes the Ai,j edges from i to j in the same way that addition by mBi,j, modulo Ai,j, permutes the integers {0, 1, . . . , Ai,j − 1}. The quotient k in the above Euclidean division also plays an important role, namely in the definition of the cocycle: ϕ(m, ei,j,n) := k. In possession of the graph, the action of Z, and the cocycle ϕ constructed above, we apply our construction and we find that OG,E is isomorphic to Katsura's OA,B. So, both Nekrashevych's and Katsura's algebras become special cases of our construc- tion. We therefore believe that the project of studying such group actions on path spaces as well as the corresponding algebras is of great importance. Taking the first few steps in this direction we have been able to describe OG,E as the C-algebra of an ´etale groupoid GG,E, whose construction is remarkably similar to the groupoid associated to the relation of "tail equivalence with lag" on the path space, as described by Kumjian, Pask, Raeburn and Renault in [20]. The first similarity is that our groupoid GG,E has the exact same unit space as the corresponding graph groupoid, namely the infinite path space. The second, and most surprising similarity is that GG,E is also described by a lag function, except that the values of the lag are not integer numbers, as in [20], but lie in a slightly more complicated group, namely the semi-direct product of the corona group of G by the right shift automorphism (see below for precise definitions). We would like to stress that, like Nekrashevych's groupoid [26: Theorem 5.1], our groupoid GG,E is constructed as a groupoid of germs. However, departing from Nekra- shevych's techniques, we use Patterson's [27] notion of "germs", rather than the one 4 r. exel and e. pardo employed in [26: Section 5]. While agreeing in many cases, such as when the action is topologically free (see below for the precise definition), the former has a much better chance of producing Hausdorff groupoids and, in our case, we are able to give a precise characterization of Hausdorffness in terms of a property we call pseudo freeness (see below for the precise definition). The techniques we use to give OG,E a groupoid model bear heavily on the theory of tight representations of inverse semigroups developed by the first named author in [6]. In particular, from our initial data we construct an abstract inverse semigroup SG,E and show that OG,E is the universal C-algebra for tight representations of SG,E. In another direction we again take inspiration from Nekrashevych [24] and give a description of OG,E as a Cuntz-Pimsner algebra for a very natural correspondence M over the algebra C(E0) ⋊ G. As a result we are able to prove that OG,E is nuclear when G is amenable. We briefly study the natural representation of the graph C-algebra C∗(E) [30] into OG,E, which turns out to be faithful. Also, we study the natural representation of the group G into OG,E, which turns out to be faithful when the triple (G, E, ϕ) satisfies pseudo freeness, but fails in general. Simplicity of OG,E is also discussed by using our description of this algebra as a groupoid C-algebra and employing results from [4]. In doing so, it is crucial to determine when is Gtight(SG,E) a Hausdorff, minimal essentially principal groupoid. To this end, we strongly rely on results obtained by both authors in [12] about characterization of minimal- ity and essential irreducibility for the groupoid of germs of a general ∗-inverse semigroup. We then specialize these results to the particular context of the inverse semigroup SG,E. Hence, we characterize Hausdorffness of Gtight(SG,E) in terms of the existence of finitely many minimal strongly fixed paths (see below for a precise definition). Also, we characterize minimality of Gtight(SG,E) in terms of weak G-transitivity of the graph (see Section 13 for a definition of this concept). We then obtain a natural generalization of the analog result obtained in [9] for Exel-Laca algebras. We also show that being essentially principal is related to the topological freeness of the action of SG,E on the infinite path space. In this sense, we obtain a characterization that relies on the existence of entries for any circuit of the graph, plus a formal condition which forces any element of G fixing open sets to be tighly related to the existence of suitable minimal strongly fixed paths. Moreover, we give sufficient conditions on Gtight(SG,E) to guarantee its local contrac- tiveness (see e.g. [1] for a definition); this property turns out to be a consequence of essential principality, so that any simple algebra in the class OG,E will be purely infinite simple. With the machinery developped we are then able to give a characterization of sim- plicity (and so pure infinite simplicity) for OG,E when Gtight(SG,E) is Hausdorff. Finally, we revisit the case of Nekrashevych and Katsura algebras, giving a picture of the properties enjoyed by these algebras that turns out to be more general than the ones given by Nekrashevych or Katsura. inverse semigroup actions and self-similar graphs 5 Some of the results in the present paper appeared in the preprints [10] and [11], which in turn are to be replaced by the present work. We would also like to mention [13] and [32], which are strongly related to the algebras we study here. In [13] conditions are given for OG,E to be a partial crossed product and in [32] an interesting connection with Zappa-Sz´ep products is made. Part of this work was done during visits of the second named author to the Departa- mento de Matem´atica da Universidade Federal de Santa Catarina (Florian´opolis, Brasil) and he would like to express his thanks to the host center for its warm hospitality. Both authors thank Benjamin Steinberg for interesting discussions on topological freeness of actions, and Hausdorffness of groupoids. 2. Groups acting on graphs. Let E = (E0, E1, r, d) be a directed graph, where E0 denotes the set of vertices, E1 is the set of edges, r is the range map, and d is the source, or domain map. By definition, a source in E is a vertex x ∈ E0, for which r−1(x) = ∅. Thus, when we say that a graph has no sources, we mean that r−1(x) 6= ∅, for all x ∈ E0. By an automorphism of E we shall mean a bijective map σ : E0 ∪ E1 → E0 ∪ E1 such that σ(Ei) ⊆ Ei, for i = 0, 1, and moreover such that r ◦ σ = σ ◦ r, and d ◦ σ = σ ◦ d, on E1. It is evident that the collection of all automorphisms of E forms a group under composition. By an action of a group G on a graph E we shall mean a group homomorphism from G to the group of all automorphisms of E. If X is any set, and if σ is an action of a group G on X, we shall say that a map is a one-cocycle for σ, when ϕ : G × X → G ϕ(gh, x) = ϕ(cid:0)g, σh(x)(cid:1)ϕ(h, x), (2.1) for all g, h ∈ G, and all x ∈ X. In this case, plugging g = h = 1, above, we see that necessarily ϕ(1, x) = 1, (2.2) for every x. 2.3. ◮ Standing Hypothesis. Throughout this work we shall let G be a countable discrete group, E be a finite graph with no sources, σ be an action of G on E, and ϕ : G × E1 → G be a one-cocycle for the restriction of σ to E1, which moreover satisfies σϕ(g,e)(x) = σg(x), ∀ g ∈ G, ∀ e ∈ E1, ∀ x ∈ E0. (2.3.1) 6 r. exel and e. pardo The assumptions that E is finite and has no sources will in fact only be used in the next section and it could probably be removed by using well known graph C-algebra techniques. By a path in E of length n ≥ 1 we shall mean any finite sequence of the form α = α1α2 . . . αn, where αi ∈ E1, and d(αi) = r(αi+1), for all i (this is the usual convention when treating graphs from a categorical point of view, in which functions compose from right to left). The range of α is defined by while the source of α is defined by r(α) = r(α1), d(α) = d(αn). A vertex x ∈ E0 will be considered a path of length zero, in which case we set r(x) = d(x) = x. For every integer n ≥ 0 we denote by En the set of all paths in E of length n (this being consistent with the already introduced notations for E0 and E1). Finally, we denote by E∗ the sets of all finite paths, and by E≤n the set of all paths of length at most n, namely E∗ = Sk≥0 Ek, and E≤n = Ek. nSk=0 We will often employ the operation of concatenation of paths. That is, if (and only if) α and β are paths such that d(α) = r(β), we will denote by αβ the path obtained by juxtaposing α and β. In the special case in which α is a path of length zero, the concatenation αβ is allowed if and only if α = r(β), in which case we set αβ = β. Similarly, when β = 0, then αβ is defined iff d(α) = β, and then αβ = α. We would now like to describe a certain extension of σ and ϕ to finite paths. g = σg, on E≤1, 2.4. Proposition. Under the assumptions of (2.3) there exists a unique pair (σ∗, ϕ∗), formed by an action σ∗ of G on E∗ (viewed simply as a set), and a one-cocycle ϕ∗ for σ∗, such that, for every n ≥ 0, every g ∈ G, and every x ∈ E0, one has that: (i) σ∗ (ii) ϕ∗(g, x) = g, (iii) ϕ∗ = ϕ, on G × E1, (iv) σ∗ (v) r ◦ σ∗ (vi) d ◦ σ∗ (vii) σϕ∗(g,α)(x) = σg(x), for all α ∈ En, g = σg ◦ r, on En, g = σg ◦ d, on En, g (En) ⊆ En, inverse semigroup actions and self-similar graphs 7 (viii) σ∗ (ix) σ∗ 1 is the identity 1 on En, g (αβ) = σ∗ g (α) σ∗ ϕ∗(g,α)(β), provided α and β are finite paths with αβ ∈ En, (x) ϕ∗(g, αβ) = ϕ∗(cid:0)ϕ∗(g, α), β(cid:1), provided α and β are finite paths with αβ ∈ En. Proof. Initially notice that, once (v), (vi) and (vii) are proved, the concatenation of the paths "σ∗ ϕ∗(g,α)(β)", appearing in (ix), is permitted because g (α)" and "σ∗ r(cid:0)σ∗ ϕ∗(g,α)(β)(cid:1) (v) = σϕ∗(g,α)(r(β)) (vii) = σg(r(β)) = σg(cid:0)d(α)(cid:1) (vi) = d(cid:0)σ∗ g (α)(cid:1). For every g in G, define σ∗ g on E≤1 to coincide with σg. Also, define ϕ∗ on G × E≤1 by (ii) and (iii). It is then clear that (i -- iii) hold and it is easy to see that the remaining properties (iv -- x) hold for all n ≤ 1. We shall complete the definitions of σ∗ and ϕ∗ by induction, so we assume that m ≥ 1, that is defined for all g in G, that σ∗ g : E≤m → E≤m ϕ∗ : G × E≤m → G, is defined, and that (i -- x) hold for all n ≤ m. We then define for all g in G, and σ∗ g : Em+1 → Em+1 ϕ∗ : G × Em+1 → G, by induction as follows. Given α ∈ Em+1, write α = α′α′′, with α′ ∈ E1, and α′′ ∈ Em, and put σ∗ g (α) = σg(α′)σ∗ ϕ(g,α′)(α′′), and ϕ∗(g, α) = ϕ∗(cid:0)ϕ(g, α′), α′′(cid:1). (2.4.1) A quick analysis, as done in the first paragraph of this proof, shows that the concatenation ϕ(g,α′)(α′′)", appearing above, is permitted. We next verify (iv -- x), of "σg(α′)" and "σ∗ substituting m + 1 for n. We have that the length of σ∗ With respect to (v) we have that g (α), as defined above, is clearly 1 + m, thus proving (iv). As for (vi), notice that r(cid:0)σ∗ g (α)(cid:1) = r(cid:0)σg(α′)(cid:1) = σg(cid:0)r(α′)(cid:1) = σg(cid:0)r(α)(cid:1). ϕ(g,α′)(α′′)(cid:1) = σϕ(g,α′)(cid:0)d(α′′)(cid:1) = σg(cid:0)d(α′′)(cid:1) = σg(cid:0)d(α)(cid:1). d(cid:0)σ∗ g (α)(cid:1) = d(cid:0)σ∗ 1 This is evidently already included in the statement that σ∗ is an action, but we repeat it here to aid our proof by induction. 8 r. exel and e. pardo Given x ∈ E0, we have that σϕ∗(g,α)(x) = σϕ∗(ϕ(g,α′),α′′)(x) = σϕ(g,α′)(x) = σg(x), taking care of (vii). The verification of (viii) is done as follows: for α = α′α′′, as in (2.4.1), one has 1(α) = σ∗ σ∗ 1(α′α′′) = σ1(α′)σ∗ ϕ(1,α′)(α′′) (2.2) = σ1(α′)σ∗ 1(α′′) = α′α′′ = α. In order to prove (ix), pick paths α in Ek and β in El, where k + l = m + 1, and such that d(α) = r(β). We leave it for the reader to verify (ix) in the easy case in which k = 0, that is, when g given in α is a vertex. The case k = 1 is also easy as it is nothing but the definition of σ∗ (2.4.1). So we may assume that k ≥ 2. Writing α = α′α′′, with α′ ∈ E1, and α′′ ∈ Ek−1, we then have that αβ = α′α′′β, and hence, by definition, σ∗ g (αβ) = σg(α′)σ∗ ϕ(g,α′)(α′′β) = σg(α′)σ∗ ϕ(g,α′)(α′′) σ∗ ϕ∗(ϕ(g,α′),α′′)(β) = = σ∗ g (α′α′′) σ∗ ϕ∗(g,α′α′′)(β). We remark that, in last step above, one should use the induction hypothesis in case k ≤ m, and the definitions of σ∗ and ϕ∗, when k = m + 1. To verify (x) we again pick paths α in Ek and β in El, where k + l = m + 1, and such that d(α) = r(β). We once more leave the easy case k = 0 to the reader and observe that the case k = 1 follows from the definition of ϕ∗. We may then suppose that k ≥ 2, so we write α = α′α′′, with α′ ∈ E1, and α′′ ∈ Ek−1. Then ϕ∗(g, αβ) = ϕ∗(g, α′α′′β) = ϕ∗(cid:0)ϕ(g, α′), α′′β(cid:1) = ϕ∗(cid:16)ϕ∗(cid:0)ϕ(g, α′), α′′(cid:1), β(cid:17) = = ϕ∗(cid:16)ϕ∗(g, α′α′′(cid:1), β(cid:17) = ϕ∗(cid:16)ϕ∗(g, α(cid:1), β(cid:17). σ∗ g σ∗ Let us now prove that σ∗ is in fact an action of G on En. We begin by proving that h = σ∗ This follows immediately from the hypothesis for n ≤ 1, so let us assume that n ≥ 2. gh on En, for every g and h in G, which we do by induction on n. Given α ∈ En, write α = α′α′′, with α′ ∈ E1, and α′′ ∈ En−1. Then σ∗ g(cid:0)σ∗ h(α)(cid:1) = σ∗ g(cid:0)σ∗ h(α′α′′)(cid:1) = σ∗ ϕ(g,σh(α′))(cid:0)σϕ(h,α′)(α′′)(cid:1) = σgh(α′)σ∗ = σg(cid:0)σh(α′)(cid:1)σ∗ g(cid:0)σh(α′)σϕ(h,α′)(α′′)(cid:1) = ϕ(gh,α′)(α′′) = σ∗ gh(α′α′′) = σ∗ = σgh(α′)σ∗ gh(α). ϕ(g,σh(α′))ϕ(h,α′)(α′′) = inverse semigroup actions and self-similar graphs 9 That α∗ g is bijective on each En then follows2 from (viii), so α∗ is indeed an action of G on En. Finally, let us show that ϕ∗ is a cocycle for σ∗ on En. For this fix g and h in G and let α ∈ En. Then, with α = α′α′′, as before, ϕ∗(gh, α) = ϕ∗(gh, α′α′′) = ϕ∗(ϕ(gh, α′), α′′) = ϕ∗(cid:16)ϕ(cid:0)g, σh(α′)(cid:1)ϕ(h, α′), α′′(cid:17) = On the other hand, focusing on the right-hand-side of (2.1), notice that ϕ∗(g, σ∗ = ϕ∗(cid:16)ϕ(cid:0)g, σh(α′)(cid:1), σ∗ ϕ(h,α′)(α′′)(cid:17)ϕ∗(cid:0)ϕ(h, α′), α′′(cid:1) =: (⋆). h(α′α′′)(cid:1)ϕ∗(h, α′α′′) = ϕ(h,α′)(α′′)(cid:1)ϕ∗(cid:0)ϕ(h, α′), α′′(cid:1) = ϕ(h,α′)(α′′)(cid:17)ϕ∗(cid:0)ϕ(h, α′), α′′(cid:1), h(α))ϕ∗(h, α) = ϕ∗(cid:0)g, σ∗ = ϕ∗(cid:0)g, σh(α′)σ∗ = ϕ∗(cid:16)ϕ(cid:0)g, σh(α′)(cid:1), σ∗ which coincides with (⋆) above. This concludes the proof. (cid:3) The only action of G on E∗ to be considered in this paper is σ∗ so, from now on, we will adopt the shorthand notation gα = σ∗ g (α). Moreover, since ϕ∗ extends ϕ, we will drop the star decoration and denote ϕ∗ simply as ϕ. The group law, the cocycle condition, and properties (ii), (v), (vi), (vii), (ix) and (x) of Proposition (2.4) may then be rewritten as follows: 2.5. Equations. For every g and h in G, for every x ∈ E0, and for every α and β in E∗ such that d(α) = r(β), one has that (a) (gh)α = g(hα), (b) ϕ(gh, α) = ϕ(cid:0)g, hα(cid:1)ϕ(h, α), (ii) ϕ(g, x) = g, (v) r(gα) = gr(α), (vi) d(gα) = gd(α), (vii) ϕ(g, α)x = gx, (ix) g(αβ) = (gα) ϕ(g, α)β, (x) ϕ(g, αβ) = ϕ(cid:0)ϕ(g, α), β(cid:1). 2 This is why it is useful to include (viii) as a separate statement, since we may now use it to prove bijectivity. 10 r. exel and e. pardo It might be worth noticing that if ϕ(g, α) = 1, then (2.5.ix) reads "g(αβ) = (gα)β", which may be viewed as an associativity property. However associativity does not hold in general as ϕ is not always trivial, and hence parentheses must be used. On the other hand parentheses are unnecessary in expressions of the form αgβ, when α, β ∈ E∗, and g ∈ G, since the only possible interpretation for this expression is the concatenation of α with gβ. Another useful property of ϕ is in order. 2.6. Proposition. For every g ∈ G, and every α ∈ E∗, one has that ϕ(g−1, α) = ϕ(g, g−1α)−1. Proof. We have 1 = ϕ(1, α) = ϕ(gg−1, α) = ϕ(g, g−1α)ϕ(g−1, α), from where the conclusion follows. (cid:3) 3. The universal C-algebra OG,E. As in the above section we fix a graph E, an action of a group G on E, and a one-cocycle ϕ satisfying (2.3). It is our next goal to build a C-algebra from this data but first let us recall the following notion from [30]: 3.1. Definition. A Cuntz-Krieger E-family consists of a set {px : x ∈ E0} of mutually orthogonal projections and a set {se : e ∈ E1} of partial isometries, all lying in some C-algebra, and satisfying (i) s∗ ese = pd(e), for every e ∈ E1, (ii) px = Xe∈r−1(x) ses∗ e, for every x ∈ E0 for which r−1(x) is finite and nonempty. 3.2. Definition. We define OG,E to be the universal unital C-algebra generated by a set {px : x ∈ E0} ∪ {se : e ∈ E1} ∪ {ug : g ∈ G}, subject to the following relations: (a) {px : x ∈ E0} ∪ {se : e ∈ E1} is a Cuntz-Krieger E-family, (b) the map u : G → OG,E, defined by the rule g 7→ ug, is a unitary representation of G, ( c ) ugse = sgeuϕ(g,e), for every g ∈ G, and e ∈ E1, (d) ugpx = pgxug, for every g ∈ G, and x ∈ E0. inverse semigroup actions and self-similar graphs 11 Observe that, under our standing assumptions (2.3), for every x ∈ E0 we have that r−1(x) is finite and nonempty. So (3.1.ii) and (3.2.a) imply that ugpxu∗ ugses∗ eu∗ g = Xr(e)=x = Xr(e)=x g = Xr(e)=x ge = Xr(f )=gx sges∗ sgeuϕ(g,e)u∗ ϕ(g,e)s∗ ge = sf s∗ f = pgx, which says that (3.2.d) follows from the other conditions. We have nevertheless included it in (3.2) in the belief that our theory may be generalized to graphs with sources. Our construction generalizes some well known constructions in the literature as we would now like to mention. 3.3. Example. Let (G, X) be a self similar group as in [26: Definition 2.1]. We may then consider a graph E having only one vertex and such that E1 = X. If we define ϕ(g, x) = gx, where, in the terminology of [26], gx is the restriction (or section) of g at x, then the triple (G, E, ϕ) satisfies (2.3) and one may easily show that OG,E is isomorphic to the algebra O(G,X) introduced by Nekrashevych in [26]. 3.4. Example. As in [18], let us assume we are given two N × N matrices A and B with integer entries, and such that Ai,j ≥ 0, for all i and j. We may then consider the graph E with vertex set E0 = {1, 2, . . . , N }, and such that, for each pair of vertices i, j ∈ E0, the set of edges from vertex j to vertex i is a set with Ai,j elements, say {ei,j,n : 0 ≤ n < Ai,j}. Assuming moreover that A has no identically zero rows, it is easy to see that E has no sources. Define an action σ of Z on E, which is trivial on E0, and which acts on edges as follows: given m ∈ Z, and ei,j,n ∈ E1, let (k, n) be the unique pair of integers such that mBi,j + n = kAi,j + n, and 0 ≤ n < Ai,j. That is, k is the quotient and n is the remainder of the Euclidean division of mBi,j + n by Ai,j. We then put σm(ei,j,n) = ei,j,n. In other words, σm corresponds to the addition of mBi,j to the variable "n" of "ei,j,n", taken modulo Ai,j. In turn, the one-cocycle is defined by ϕ(m, ei,j,n) = k. 12 r. exel and e. pardo Observe that if Ai,j = 0, then there are no edges from j to i, so the value Bi,j is entirely irrelevant for the above construction. Therefore it makes no difference to assume that Ai,j = 0 ⇒ Bi,j = 0. It may then be proved without much difficulty that OZ,E is isomorphic to Katsura's [18] algebra OA,B, under an isomorphism sending each um to the mth power of the unitary u := ui NXi=1 in OA,B, and sending sei,j,n to si,j,n. When N = 1, the relevant graph for Katsura's algebras is the same as the one we used above in the description of Nekrashevych's example. However the former is not a special case of the latter because, contrary to what is required in [26], the group action might not be faithful. 3.5. Example. Given any finite graph E, and any action σ of a group G on E, the map ϕ : G × E1 → G defined by ϕ(g, a) = g, ∀ g ∈ G, ∀ a ∈ E1 is a one-cocycle, and the triple (G, E, ϕ) satisfies (2.3). By (3.2.c), we have that ugsau∗ g = sga, for any g in G, and every a in E1. It is therefore easy to see that OG,E is isomorphic to the crossed product of the graph C-algebra C∗(E) [30] by G, relative to the natural action of G on C∗(E) induced by σ. In particular, if σ is the trivial action, we have that OG,E is the maximal tensor product of C∗(E) by the full group C-algebra of G. 3.6. Example. Given any finite graph without sources, and any action σ of a group G on E fixing the vertices, consider the map ϕ : G × E1 → G defined by ϕ(g, a) = 1, ∀ g ∈ G, ∀ a ∈ E1. It is easy to see that ϕ is a one-cocycle, and that the triple (G, E, ϕ) satisfies (2.3). Since E has no sources we have, for any g in G, that ug = Xx∈E0 ugpx = Xx∈E0 Xa∈r−1(x) ugsas∗ a (3.2.c) = Xx∈E0 Xa∈r−1(x) sgas∗ a, which therefore lies in the copy of C∗(E) within OG,E. Since the natural representation of C∗(E) in OG,E is faithful by (11.1), the conclusion is that OG,E ∼= C∗(E). inverse semigroup actions and self-similar graphs 13 We now return to the general case of a triple (G, E, ϕ) satisfying (2.3). We initially recall the usual extension of the notation "se" to allow for paths of arbitrary length. 3.7. Definition. Given a finite path α in E∗, we shall let sα denote the element of OG,E given by: (i) when α = x ∈ E0, we let sα = px, (ii) when α ∈ E1, then sα is already defined above, (iii) when α ∈ En, with n > 1, write α = α′α′′, with α′ ∈ E1, and α′′ ∈ En−1, and set sα = sα′sα′′, by recurrence. Commutation relation (3.2.c) may then be generalized to finite paths as follows: 3.8. Lemma. Given α ∈ E∗, and g ∈ G, one has that ugsα = sgαuϕ(g,α). Proof. Let n be the length of α. When n = 0, 1, this follows from (3.2.d&c), respectively. When n > 1, write α = α′α′′, with α′ ∈ E1, and α′′ ∈ En−1. Using induction, we then have ugsα = ugsα′ sα′′ = sgα′uϕ(g,α′)sα′′ = sgα′sϕ(g,α′)α′′ uϕ(ϕ(g,α′),α′′) = = s(gα′)ϕ(g,α′)α′′ uϕ(g,α′α′′) = sg(α′α′′)uϕ(g,α′α′′) = sgαuϕ(g,α). (cid:3) Our next result provides a spanning set for OG,E. 3.9. Proposition. Let S =(cid:8)sαugs∗ β : α, β ∈ E∗, g ∈ G, d(α) = gd(β)(cid:9) ∪ {0}. Then S is closed under multiplication and adjoints and its closed linear span coincides with OG,E. Proof. That S is closed under adjoints is clear. With respect to closure under multiplica- tion, let sαugs∗ δ be elements of S. β and sγuhs∗ From (3.2.a) we know that s∗ βsγ = 0, unless either γ = βε, or β = γε, for some ε ∈ E∗. If γ = βε, then and hence s∗ βsγ = s∗ βsβε = s∗ βsβsε = sε, (sαugs∗ β)(sγuhs∗ δ) = sαugsεuhs∗ δ = sαsgεuϕ(g,ε)uhs∗ δ = sαgεuϕ(g,ε)hs∗ δ. (3.9.1) Moreover, since d(αgε) = d(gε) = gd(ε) = ϕ(g, ε)d(ε) = ϕ(g, ε)d(γ) = ϕ(g, ε)hd(δ), we deduce that the element appearing in the right-hand-side of (3.9.1) indeed belongs to S. 14 r. exel and e. pardo In the second case, namely if β = γε, then the adjoint of the term appearing in the left-hand-side of (3.9.1) is (sδuh−1s∗ γ)(sβug−1s∗ α), and the case already dealt with implies that this belongs to S. The result then follows from the fact that S is self-adjoint. In order to prove that OG,E coincides with the closed linear span of S, let A denote the latter. Given that S is self-adjoint and closed under multiplication, we see that A is a closed -subalgebra of OG,E. Since A evidently contains sα for every α in E≤1, and since it also contains ug for every g in G, we deduce that A = OG,E. (cid:3) 4. The inverse semigroup SG,E. As before, we keep (2.3) in force. In this section we will give an abstract description of the set S appearing in (3.9) as well as its multiplication and adjoint operation. The goal is to construct an inverse semigroup from which we will later recover OG,E. 4.1. Definition. Over the set SG,E =(cid:8)(α, g, β) ∈ E∗ × G × E∗ : d(α) = gd(β)(cid:9) ∪ {0}, consider a binary multiplication operation defined by (α, g, β)(γ, h, δ) = (αgε, ϕ(g, ε)h, δ), if γ = βε, (α, gϕ(h−1, ε)−1, δh−1ε), if β = γε, 0, otherwise, and a unary adjoint operation defined by (α, g, β)∗ := (β, g−1, α). Furthermore, the subset of SG,E formed by all elements (α, g, β), with g = 1, will be denoted by SE. It is easy to see that SE is closed under the above operations, and that it is isomorphic to the inverse semigroup generated by the canonical partial isometries in the graph C- algebra of E. Let us begin with a simple, but useful result: 4.2. Lemma. Given (α, g, β) and (γ, h, δ) in SG,E, one has β = γ ⇒ (α, g, β)(γ, h, δ) = (α, gh, δ). Proof. Focusing on the first clause of (4.1), write γ = βε, with ε = d(β). Then (α, g, β)(γ, h, δ) = (αgd(β), ϕ(cid:0)g, d(β)(cid:1)h, δ) = (αd(α), gh, δ) = (α, gh, δ). 4.3. Proposition. SG,E is an inverse semigroup with zero. (cid:3) inverse semigroup actions and self-similar graphs 15 Proof. We leave it for the reader to prove that the above operations are well defined and the multiplication is associative. In order to prove the statement it then suffices [22: Theorem 1.1.3] to show that, for all y, z ∈ SG,E, one has that (i) yy∗y = y, and (ii) yy∗ commutes with zz∗. Given y = (α, g, β) ∈ SG,E, we have by the above Lemma that yy∗y = (α, g, β)(β, g−1, α)(α, g, β) = (α, 1, α)(α, g, β) = (α, g, β) = y, proving (i). Notice also that yy∗ = (α, 1, α) (4.3.1) is an element of the idempotent semi-lattice of SE, which is a commutative set because SE is an inverse semigroup. Point (ii) above then follows immediately, concluding the proof. (cid:3) As seen in (4.3.1), the idempotent semi-lattice of SG,E, henceforth denoted by E, is given by E =(cid:8)(α, 1, α) : α ∈ E∗(cid:9) ∪ {0}. Evidently E is also the idempotent semi-lattice of SE. For simplicity, from now on we will adopt the short-hand notation fα = (α, 1, α), ∀ α ∈ E∗. The following is a standard fact in the theory of graph C-algebras: 4.6. Proposition. If α, β ∈ E∗, then (4.4) (4.5) fαfβ = fα, fβ, if there exists γ such that α = βγ, if there exists γ such that αγ = β, 0, otherwise. Recall that if α and β are in E∗, we say that α (cid:22) β, if α is a prefix of β, i.e. if there exists γ ∈ E∗, such that αγ = β. It therefore follows from (4.6) that fα ≤ fβ ⇐⇒ β (cid:22) α. (4.7) Another easy consequence of (4.6) is that, for any two elements e, f ∈ E, one has that either e ⊥ f , or e and f are comparable. It follows that e ⋓ f ⇒ e ≤ f , or f ≤ e. (4.8) 16 r. exel and e. pardo 5. Pseudo freeness and E-unitarity. Again working under (2.3), suppose we are given g in G and a finite path α such that gα = α, and ϕ(g, α) = 1. (5.1) Then, given any finite path α′ extending α, that is a path of the form α′ = αβ, where β is another finite path, we have gα′ = g(αβ) = (gα)ϕ(g, α)β = αβ = α′, and ϕ(g, α′) = ϕ(g, αβ) = ϕ(cid:0)ϕ(g, α), β(cid:1) = ϕ(cid:0)1, β(cid:1) = 1. This says that any path α′ extending α also satisfies (5.1) so, in particular, every extension of α is fixed by g. 5.2. Definition. If g ∈ G and α is a finite path satisfying (5.1), we will say that α is strongly fixed by g. In addition, if no proper prefix of α is strongly fixed by g, we will say that α is a minimal strongly fixed path for g. The following result is an easy consequence of the discussion above: 5.3. Proposition. Given g in G, let Mg be the set of all minimal strongly fixed paths for g. Then the set of all strongly fixed paths for g is given by {µγ : γ ∈ E∗, d(µ) = r(γ)}, Fµ∈Mg where the square cup stands for disjoint union. Let us now introduce terminology to describe situations in which nontrivial strongly fixed paths do not exist. 5.4. Definition. We will say that (G, E, ϕ) is pseudo free 3 if, whenever (g, e) ∈ G × E1, is such that ge = e, and ϕ(g, e) = 1, then g = 1. Notice that pseudo freeness is equivalent to the fact that an edge is never a strongly fixed path for a nontrivial group element. In fact we may boost this up to finite paths as follows: 5.5. Proposition. Suppose that (G, E, ϕ) is pseudo free and that a finite path α of nonzero length is strongly fixed for some g in G. Then g = 1. 3 In a preprint version of this work we have used the term residually free to refer to the concept presently being defined, but this apparently conflicts with a well established notion in group theory. inverse semigroup actions and self-similar graphs 17 Proof. Arguing by contradiction, assume that there is a counter-example to the statement, meaning that there is a strongly fixed path α for a nontrivial group element g. Then, as already mentioned, α has a minimal strongly fixed prefix, so we may assume without loss of generality that α itself is minimal. By (2.5.ii), α can't be a vertex, and neither can it be an edge, by hypothesis. So α ≥ 2, and we may then write α = βγ, with β, γ ∈ E∗, and β, γ < α. Then βγ = α = gα = g(βγ) = (gβ)ϕ(g, β)γ, whence β = gβ, and γ = ϕ(g, β)γ, by length considerations. Should ϕ(g, β) = 1, the pair (g, β) would be a smaller counter-example to the statement, violating the minimality of α. So we have that ϕ(g, β) 6= 1. In addition, ϕ(cid:0)ϕ(g, β), γ(cid:1) = ϕ(g, βγ) = ϕ(g, α) = 1. It follows that(cid:0)ϕ(g, β), γ(cid:1) is a counter-example to the statement, again violating the min- imality of α. This is a contradiction and hence no counter-example exists whatsoever, concluding the proof. (cid:3) An apparently stronger version of pseudo freeness is in order. 5.6. Proposition. Suppose that (G, E, ϕ) is pseudo free. Then, for all g1, g2 ∈ G, and α ∈ E∗, one has that g1α = g2α and ϕ(g1, α) = ϕ(g2, α) ⇒ g1 = g2. Proof. Defining g = g−1 2 g1, observe that gα = α, and we claim that ϕ(g, α) = 1. In fact, 2 g1, α(cid:1) = ϕ(cid:0)g−1 2 , g1α(cid:1)ϕ(cid:0)g1, α(cid:1) (2.6) ϕ(g1, α) = ϕ(g2, α)−1ϕ(g1, α) = 1, = ϕ(g, α) = ϕ(cid:0)g−1 = ϕ(cid:0)g2, g−1 2 g1α(cid:1)−1 so it follows that g = 1, which is to say that g1 = g2. (cid:3) We will now determine conditions under which SG,E is E-unitary. In order to do so we first need to understand when does an element s of SG,E dominate a nonzero idempotent e, which in turn must necessarily have the form e = (γ, 1, γ), as seen in (4.3.1). If s indeed dominates a nonzero idempotent, it is clear that s is itself nonzero, so s must have the form (α, g, β). 5.7. Proposition. Let α, β and γ be finite paths in E, and let g ∈ G be such that d(α) = gd(β), so that s := (α, g, β) is a general nonzero element of SG,E and e := (γ, 1, γ) is a general nonzero idempotent element of SG,E. Then e ≤ s, if and only (i) α = β, (ii) γ = ατ , for some finite path τ , (iii) τ is strongly fixed by g. 18 r. exel and e. pardo Proof. In order to prove the "if" part, we have se = (α, g, β)(γ, 1, γ) = (α, g, α)(ατ, 1, γ) = (αgτ, ϕ(g, τ ), γ) = = (ατ, 1, γ) = (γ, 1, γ) = e, proving that e ≤ s. Conversely, assuming that e ≤ s, we have se = e, so in particular se 6= 0, and hence by the definition of the multiplication on SG,E, either γ is a prefix of β or vice versa. In case β is a prefix of γ, we may write γ = βτ , for some finite path τ , and then (γ, 1, γ) = e = se = (α, g, β)(βτ, 1, γ) = (αgτ, ϕ(g, τ ), γ), so we conclude that αgτ = γ = βτ , and ϕ(g, τ ) = 1. So α = β, gτ = τ and the statement is proved. On the other hand, if γ is a prefix of β, we may write β = γε and, again according to the definition of the multiplication on SG,E, the third coordinate of the product se will be γε, from where we conclude that γ = γε. So ε = 0 and then γ = β, which in particular means that β is a prefix of γ, and the proof follows as above. (cid:3) 5.8. Proposition. SG,E is an E-unitary inverse semigroup if and only if (G, E, ϕ) is pseudo free. Proof. Let s be an element of SG,E which dominates a nonzero idempotent element e. As discussed above, we necessarily have s = (α, g, β), and e = (γ, 1, γ), where α, β and γ are finite paths in E, and d(α) = gd(β). Then, by (5.7) we conclude that gτ = τ , and ϕ(g, τ ) = 1 so, assuming that (G, E, ϕ) is pseudo free, we have g = 1. Moreover by (5.7.i) we see that α = β, so s = (α, g, β) = (α, 1, α), which is idempotent as desired. In order to prove the converse, let (g, e) ∈ G × E1, be such that ge = e, and ϕ(g, e) = 1. Then the element lies in SG,E because s :=(cid:0)d(e), g, d(e)(cid:1) gd(e) = d(ge) = d(e). Moreover observe that s dominates the nonzero idempotent element (e, 1, e), since s(e, 1, e) =(cid:0)d(e), g, d(e)(cid:1)(cid:0)d(e)e, 1, e(cid:1) = (d(e)ge, ϕ(g, e), e) = (e, 1, e). So, under the hypothesis that SG,E is E-unitary, we conclude that s is idempotent, which is to say that g = 1. This proves that (G, E, ϕ) is pseudo free. (cid:3) inverse semigroup actions and self-similar graphs 19 6. Tight representations of SG,E. As before, we keep (2.3) in force. It is the main goal of this section to show that OG,E is the universal C-algebra for tight representations of SG,E. Recall from (4.6) that fα ≤ fd(α), for every α ∈ E∗, so we see that the set {fx : x ∈ E0} (6.1) is a cover [6: Definition 11.5] for E. 6.2. Proposition. The map defined by π(0) = 0, and π : SG,E → OG,E, π(α, g, β) = sαugs∗ β, is a tight [6: Definition 13.1] representation. Proof. We leave it for the reader to show that π is in fact multiplicative and that it preserves adjoints. In order to prove that π is tight, we shall use the characterization given in [6: Propo- sition 11.8], observing that π satisfies condition (i) of [6: Proposition 11.7] because, with respect to the cover (6.1), we have that _x∈E0 π(fx) = _x∈E0 π(x, 1, x) = _x∈E0 px = Xx∈E0 px = 1, by (3.2.a). So we assume that {fα1 , . . . , fαn} is a cover for a given fβ, where α1, . . . αn, β ∈ E∗, and we need to show that π(fαi) ≥ π(fβ). (6.2.1) n_i=1 In particular, for each i, we have that fαi ≤ fβ, so by (4.7) there exists γ i ∈ E∗ such that αi = βγ i. We shall prove (6.2.1) by induction on the variable L = min 1≤i≤n γ i. If L = 0, we may pick i such that γ i = 0, and then necessarily γ i = d(β), in which case αi = β, and (6.2.1) is trivially true. Assuming that L ≥ 1, let x := d(β). Observe that x is not a source either because this is part of our standing assumptions (2.3), or simply because x is the range of every γ i. In any case let us write r−1(x) = {e1, . . . , ek}, 20 r. exel and e. pardo and observe that π(fβ) = sβs∗ β = sβpxs∗ β (3.2.a) = kXj=1 sβsej s∗ ej s∗ β = π(fβej ). kXj=1 In order to prove (6.2.1) it is therefore enough to show that π(fαi) ≥ π(fβej ), (6.2.2) n_i=1 for all j = 1, . . . , k. Fixing j we claim that fβej is covered by the set Z =(cid:8)fαi : 1 ≤ i ≤ n, fαi ≤ fβej(cid:9). In order to see this let y be a nonzero element in E such that y ≤ fβej . Then y ≤ fβ, and so y ⋓ fαi for some i. Thus, to prove the claim it is enough to check that fαi lies in Z. Observe that yfβej fαi = yfαi 6= 0, which implies that fβej ⋓ fαi. By (4.8) we have that fβej and fαi are comparable, so either βej (cid:22) αi or αi (cid:22) βej , by (4.7). Since we are under the hypothesis that L ≥ 1, and hence that αi = βi + γ i ≥ β + 1 = βej, we must have that βej (cid:22) αi, from where we deduce that fαi ≤ fβej , proving our claim. Employing the induction hypothesis we then deduce that π(z) ≥ π(fβej ), _z∈Z verifying (6.2.2), and thus concluding the proof. (cid:3) We would now like to prove that the representation π above is in fact the universal tight representation of SG,E. 6.3. Theorem. Let A be a unital C-algebra and let ρ : SG,E → A be a tight represen- tation. Then there exists a unique unital -homomorphism ψ : OG,E → A, such that the diagram SG,E π −→ OG,E ................................................................................................................ .......... ρ yψ A commutes. inverse semigroup actions and self-similar graphs 21 Proof. We will initially prove that the elements px := ρ(x, 1, x), ∀x ∈ E0, ∀e ∈ E1, se := ρ(cid:0)e, 1, d(e)(cid:1), ug := Xx∈E0 ρ(x, g, g−1x), ∀g ∈ G, satisfy relations (3.2.a -- d). Since the fx defined in (4.5) are mutually orthogonal idempo- tents in SG,E, it is clear that the px are mutually orthogonal projections. Evidently the se are partial isometries so, in order to check (3.2.a), we must only verify (3.1.i) and (3.1.ii). With respect to the former, let e ∈ E1. Then s∗ e se = ρ(cid:0)(d(e), 1, e)(e, 1, d(e)(cid:1) = ρ(cid:0)d(e), 1, d(e)(cid:1) = pd(e), proving (3.1.i). In order to prove (3.1.ii), let x be a vertex such that r−1(x) is nonempty and write Putting qi = (ei, 1, ei), we then claim that the set r−1(x) =(cid:8)e1, . . . , en(cid:9). is a cover for q := (x, 1, x). idempotent f is dominated by q, then f ⋓ qi for some i. In order to prove this we must show that, if the nonzero (cid:8)q1, . . . , qn(cid:9) Let f = (α, 1, α) by (4.4) and notice that 0 6= f = f q = (α, 1, α)(x, 1, x). So α and x are comparable, and this can only happen when x = r(α). If α = 0 then necessarily α = x, so f = q, and it is clear that f ⋓ qi for all i. On the other hand, if α ≥ 1, we write α = α′α′′, with α′ ∈ E1, so that r(α′) = r(α) = x, and hence α′ = ei, for some i. Therefore f qi = (α, 1, α)(ei, 1, ei) = (α, 1, α)(α′, 1, α′) = (α, 1, α) 6= 0, so f ⋓ qi, proving the claim. Since ρ is a tight representation, we deduce that ρ(q) = ρ(qi), n_i=1 but since the qi are easily seen to be pairwise orthogonal, their supremum coincides with their sum, whence px = ρ(q) = ρ(qi) = ρ(ei, 1, ei) = nXi=1 nXi=1 22 r. exel and e. pardo = nXi=1 ρ(cid:0)(ei, 1, d(ei)) (d(ei), 1, ei)(cid:1) = thus verifying (3.1.ii), and hence proving (3.2.a). sei s∗ ei, nXi=1 With respect to (3.2.b), let us first prove that u1 = 1. Considering the subsets of E given by X = ∅, Y = ∅, and Z =(cid:8)(x, 1, x) : x ∈ E0(cid:9), notice that, according to [6: Definition 11.4], one has that E X,Y = E, and that Z is a cover for E X,Y , as seen in (6.1). By the tightness condition [6: Definition 11.6] we have _z∈Z ρ(z) ≥ ^x∈X ρ(x) ∧ ^y∈Y ¬ρ(y). As explained in the discussion following [6: Definition 11.6], the right-hand-side above must be interpreted as 1 because X and Y are empty. On the other hand, since the ρ(z) are pairwise orthogonal, the supremum in the left-hand-side above becomes a sum, so 1 =Xz∈Z ρ(z) = Xx∈E0 ρ(x, 1, x) = u1. In order to prove that u is multiplicative, let g and h be in G. Then ug uh = Xx,y∈E0 ρ(cid:0)(x, g, g−1x)(y, h, h−1y)(cid:1) = ρ(cid:0)x, gh, (gh)−1x(cid:1) = ugh. We next claim that u∗ g = ug−1, for all g in G. To prove it we compute = Xx∈E0 ρ(cid:0)(x, g, g−1x)(g−1x, h, h−1g−1x)(cid:1) = Xx∈E0 g = Xx∈E0 ρ(x, g, g−1x)∗ = Xx∈E0 u∗ ρ(g−1x, g−1, x) = · · · which, upon the change of variables y = g−1x, becomes · · · = Xy∈E0 ρ(y, g−1, gy) = ug−1. This shows that u is a unitary representation, verifying (3.2.b). Turning now our attention to (3.2.c), let g ∈ G and e ∈ E1. Then ug se = Xx∈E0 ρ(x, g, g−1x) ρ(cid:0)e, 1, d(e)(cid:1) = ρ(cid:0)gr(e), g, r(e)(cid:1) ρ(cid:0)e, 1, d(e)(cid:1) = inverse semigroup actions and self-similar graphs 23 On the other hand = ρ(cid:0)r(ge)ge, ϕ(g, e), d(e)(cid:1) = ρ(cid:0)ge, ϕ(g, e), d(e)(cid:1) = (⋆). sge uϕ(g,e) = ρ(cid:0)ge, 1, d(ge)(cid:1) Xx∈E0 ρ(cid:0)x, ϕ(g, e), ϕ(g, e)−1x(cid:1) = = ρ(cid:0)ge, 1, d(ge)(cid:1) ρ(cid:0)d(ge), ϕ(g, e), g−1d(ge)(cid:1) = = ρ(cid:0)ge, 1, d(ge)(cid:1) ρ(cid:0)d(ge), ϕ(g, e), d(e)(cid:1) = ρ(cid:0)ge, ϕ(g, e), d(e)(cid:1), which coincides with (⋆) and hence proves (3.2.c). We leave the proof of (3.2.d) to the reader after which the universal property of OG,E intervenes to provide us with a - homomorphism ψ : OG,E → A sending px 7→ px, se 7→ se, and ug 7→ ug. Now we must show that ψ(cid:0)π(γ)(cid:1) = ρ(γ), We will first do so for the following special cases: (i) γ = (x, 1, x), for x ∈ E0, (ii) γ =(cid:0)e, 1, d(e)(cid:1), for e ∈ E1, (iii) γ = (x, g, g−1x), for x ∈ E0, and g ∈ G. ∀ γ ∈ SG,E. (6.3.1) In case (i) we have ψ(π(γ)) = ψ(π(x, 1, x)) = ψ(px) = px = ρ(x, 1, x) = ρ(γ). As for (ii), Under (iii), ψ(π(γ)) = ψ(cid:0)π(cid:0)e, 1, d(e)(cid:1)(cid:1) = ψ(se) = se = ρ(cid:0)e, 1, d(e)(cid:1) = ρ(γ). ψ(π(γ)) = ψ(cid:0)π(x, g, g−1x)(cid:1) = ψ(pxugpg−1x) = ψ(pxug) = px ug = = ρ(x, 1, x) Xy∈E0 ρ(y, g, g−1y) = Xy∈E0 ρ(cid:0)(x, 1, x)(y, g, g−1y)(cid:1) = ρ(x, g, g−1x) = ρ(γ). In order to prove (6.3.1), it is now clearly enough to check that the -sub-semigroup of SG,E generated by the elements mentioned in (i -- iii), above, coincides with SG,E. 24 r. exel and e. pardo Denoting this -sub-semigroup by T , we will first show that(cid:0)α, 1, d(α)(cid:1) is in T , for every α ∈ E∗. This is evident for α ≤ 1, so we suppose that α = α′α′′, with α′ ∈ E1, and r(α′′) = d(α′). We then have by induction that T ∋(cid:0)α′, 1, d(α′)(cid:1)(cid:0)α′′, 1, d(α′′)(cid:1) =(cid:0)α′α′′, 1, d(α′′)(cid:1) =(cid:0)α, 1, d(α)(cid:1). Considering a general element (α, g, β) ∈ SG,E, let x = d(α), so that g−1x = d(β), and notice that T ∋(cid:0)α, 1, d(α)(cid:1)(x, g, g−1x)(cid:0)β, 1, d(β)(cid:1)∗ =(cid:0)α, 1, d(α)(cid:1)(cid:0)d(α), g, d(β)(cid:1)(cid:0)d(β), 1, β(cid:1) = (α, g, β), = which proves that T = SG,E, and hence that (6.3.1) holds. To conclude we observe that the uniqueness of ψ follows from the fact that OG,E is (cid:3) generated by the px, the se, and the ug. Given an inverse semigroup S with zero, recall from [6: Theorem 13.3] that Gtight(S) (denoted simply as Gtight in [6]) is the groupoid of germs for the natural action of S on the space of tight filters over its idempotent semi-lattice. Moreover the C-algebra of Gtight(S) is universal for tight representations of S. 6.4. Corollary. Under the assumptions of (2.3) one has that OG,E is isomorphic to the C-algebra of the groupoid Gtight(SG,E). Proof. Follows from [6: Theorem 13.3] and the uniqueness of universal C-algebras. (cid:3) We should notice that our requirement that G be countable in (2.3) is only used in the above proof, since the application of [6: Theorem 13.3] depends on the countability of SG,E. 7. The Lag Group. It is our next goal to give a concrete description of Gtight(SG,E), similar to the description given of the groupoid associated to a row-finite graph in [20: Definition 2.3]. The crucial ingredient there is the notion of tail equivalence with lag. In this section we will construct a group where our generalized lag function will take values. Let G be a group. Within the infinite cartesian product4 consider the infinite direct sum G G∞ = Yn∈ G(∞) =Mn∈ G formed by the elements g = (gn)n∈ ∈ G∞ which are eventually trivial, that is, for which there exists n0 such that gn = 1, for all n ≥ n0. It is clear that G(∞) is a normal subgroup of G∞. 4 For the purpose of this cartesian product we adopt the convention that = {1, 2, 3, . . .}. inverse semigroup actions and self-similar graphs 25 7.1. Definition. Given a group G, the corona of G is the quotient group G = G∞/G(∞). Consider the left and right shift endomorphisms of G∞ λ, ρ : G∞ → G∞ given for every g = (gn)n∈ ∈ G∞, by and λ(g)n = gn+1, ∀ n ∈ , ρ(g)n =(cid:26) 1, gn−1, if n = 0, if n ≥ 1. It is readily seen that G(∞) is invariant under both λ and ρ, so these pass to the quotient providing endomorphisms λ, ρ : G → G. For every g = (gn)n∈ ∈ G∞, we have that λ(ρ(g)) = g, and ρ(λ(g)) = (1, g2, g3, . . .) ≡ g, (7.2) (7.3) where we use "≡" to refer to the equivalence relation determined by the normal subgroup G(∞). Therefore both λρ and ρλ coincide with the identity, and hence λ and ρ are each other's inverse. In particular, they are both automorphisms of G. Iterating ρ therefore gives an action of Z on G. 7.4. Definition. Given any countable discrete group G, the lag group associated to G is the semi-direct product group G ⋊ρ Z. The reason we call this the "lag group" is that it will play a very important role in the next section, as the co-domain for our lag function. 8. The tight groupoid of SG,E. We would now like to give a detailed description of the groupoid Gtight(SG,E). As already mentioned this is the groupoid of germs for the natural action of SG,E on the space of tight filters over the idempotent semi-lattice E of SG,E. See [6: Section 4] for more details. By an infinite path in E we shall mean any infinite sequence of the form ξ = ξ1ξ2 . . . , where ξi ∈ E1, and d(ξi) = r(ξi+1), for all i. The set of all infinite paths will be denoted by E∞. Given an infinite path ξ = ξ1ξ2 . . . ∈ E∞, 26 r. exel and e. pardo and an integer n ≥ 0, denote by ξn the finite path of length n given by ξ1ξ2 . . . ξn, if n ≥ 1, r(ξ1), if n = 0. ξn = 8.1. Proposition. There is a unique action of G on E∞ such that, (g, ξ) ∈ G × E∞ 7→ gξ ∈ E∞ (gξ)n = g(ξn), for every g ∈ G, ξ ∈ E∞, and n ∈ . Proof. Left to the reader. (cid:3) Recall from (4.5) that, for any finite path α ∈ E∗, we denote by fα the idempotent element (α, 1, α) in E. Thus, given an infinite path ξ ∈ E∞, we may look at the subset which turns out to be an ultra-filter [6: Definition] over E. Denoting the set of all ultra- Fξ = {fξn : n ∈ } ⊆ E, that the correspondence ξ ∈ E∞ 7→ Fξ ∈ bE∞ may be proven to be a homeomorphism if E∞ is equipped with the product topology. filters over E by bE∞, as in [6: Definition 12.8], one may also show [6: Proposition 19.11] is bijective, and we will use it to identify E∞ and bE∞. Furthermore, this correspondence Since E is finite, E∞ is compact by Tychonov's Theorem, and consequently so is bE∞. Being the closure of bE∞ within bE [6: Theorem 12.9], the space bEtight formed by the tight filters therefore necessarily coincides with bE∞. Identifying bEtight with E∞, as above, we may transfer the canonical action of SG,E from the former to the latter resulting in the following: to each element (α, g, β) ∈ SG,E, we associate the partial homeomorphism of E∞ whose domain is the cylinder Z(β) := {η ∈ E∞ : η = βξ, for some ξ ∈ E∞}, (8.2) and which sends each η = βξ ∈ Z(β) to αgξ, where the meaning of "gξ" is as in (8.1). As before we will not use any special symbol to indicate this action, using module notation instead: (α, g, β)η = αgξ, ∀ (α, g, β) ∈ SG,E, ∀ η = βξ ∈ Z(β). (8.3) Before we proceed let us at least check that αgξ is in fact an element of E∞, which is to say that d(α) = r(gξ). Firstly, for every element (α, g, β) ∈ SG,E, we have that d(α) = gd(β). Secondly, if η = βξ ∈ E∞, then d(β) = r(ξ). Therefore r(gξ) = gr(ξ) = gd(β) = d(α). This leads to a first, more or less concrete description of Gtight(SG,E). inverse semigroup actions and self-similar graphs 27 8.4. Proposition. Under (2.3), one has that Gtight(SG,E) is isomorphic to the groupoid of germs for the above action of SG,E on E∞. Our aim is nevertheless a much more precise description of this groupoid. Recall from [6: Definition 4.6] that the germ of an element s ∈ SG,E at a point ξ in the domain of s is denoted by [s, ξ]. If s = (α, g, β), this would lead to the somewhat awkward notation [(α, g, β), ξ], which from now on will instead be written as Thus the groupoid Gtight(SG,E), consisting of all germs for the action of SG,E on E∞, (cid:2)α, g, β; ξ(cid:3). is given by Gtight(SG,E) =n(cid:2)α, g, β; ξ(cid:3) : (α, g, β) ∈ SG,E, ξ ∈ Z(β)o. Let us now prove a useful criterion for equality of germs. (8.5) 8.6. Proposition. Suppose that (G, E, ϕ) is pseudo free and let us be given elements (α1, g1, β1) and (α2, g2, β2) in SG,E, with β1 ≤ β2, as well as infinite paths η1 in Z(β1), and η2 in Z(β2). Then (cid:2)α1, g1, β1; η1(cid:3) =(cid:2)α2, g2, β2; η2(cid:3) if and only if there is a finite path γ and an infinite path ξ, such that (i) α2 = α1g1γ, (ii) g2 = ϕ(g1, γ), (iii) β2 = β1γ, (iv) η1 = η2 = β1γξ. Proof. Assuming that the germs are equal, we have by [6: Definition 4.6] that and there is an idempotent (δ, 1, δ) ∈ E, such that η ∈ Z(δ), and η1 = η2 =: η, (α1, g1, β1)(δ, 1, δ) = (α2, g2, β2)(δ, 1, δ). (8.6.1) It follows that η = δζ, for some ζ ∈ E∞. Upon replacing δ by a longer prefix of η, we may assume that δ is as large as we want. Furthermore the element of SG,E represented by the two sides of (8.6.1) is evidently nonzero because the partial homeomorphism associated to it under our action has η in its domain. So, focusing on (4.1), we see that β1 and δ are comparable, and so are β2 and δ. Assuming that δ exceeds both β1 and β2, we may then write δ = β1ε1 = β2ε2, for suitable ε1 and ε2 in E∗. But since β1 ≤ β2, this in turn implies that β2 = β1γ, for some γ ∈ E∗, hence proving (iii). Therefore δ = β1γε2, so η = δζ = β1γε2ζ, 28 r. exel and e. pardo and (iv) follows once we choose ξ = ε2ζ. Moreover, equation (8.6.1) reads (α1, g1, β1)(β1γε2, 1, β1γε2) = (α2, g2, β1γ)(β1γε2, 1, β1γε2). Computing the products according to (4.1), we get (cid:0)α1g1(γε2), ϕ(g1, γε2), β1γε2(cid:1) =(cid:0)α2g2ε2, ϕ(g2, ε2), β1γε2(cid:1), from where we obtain α2g2ε2 = α1g1(γε2) = α1(g1γ)ϕ(g1, γ)ε2, and Since g2ε2 = ε2 = ϕ(g1, γ)ε2, we deduce from (8.6.2) that ϕ(g2, ε2) = ϕ(g1, γε2) = ϕ(cid:0)ϕ(g1, γ), ε2(cid:1). and hence also that g2ε2 = ϕ(g1, γ)ε2, α2 = α1g1γ, (8.6.2) (8.6.3) (8.6.4) proving (i). In view of (8.6.3) and (8.6.4), point (ii) follows from (5.6). Conversely, assume (i -- iv) and let us prove equality of the above germs. Setting δ = β1γ, we have by (iv) that η := η1 = η2 ∈ Z(δ), so it suffices to verify (8.6.1), which the reader could do without any difficulty. (cid:3) Proposition (8.6) then says that the typical situation in which an equality of germs takes place is Our next two results are designed to offer convenient representatives of germs. (cid:2)α, g, β; βγξ(cid:3) =(cid:2)αgγ, ϕ(g, γ), βγ; βγξ(cid:3). 8.7. Proposition. Given any germ u, there exists an integer n0, such that for every n ≥ n0, (i) there is a representation of u of the form u =(cid:2)α1, g1, β1; β1ξ1(cid:3), with α1 = n. (ii) there is a representation of u of the form u =(cid:2)α2, g2, β2; β2ξ2(cid:3), with β2 = n. Proof. Write u =(cid:2)α, g, β; η(cid:3), and choose any n0 ≥ max{α, β}. Then, for every n ≥ n0 we may write η = βγξ, with γ ∈ E∗, ξ ∈ E∞, and such that γ = n − α (resp. γ = n − β). Therefore u =(cid:2)α, g, β; βγξ(cid:3) =(cid:2)αgγ, ϕ(g, γ), βγ; βγξ(cid:3), and we have αgγ = α + gγ = α + γ = n (resp. βγ = β + γ = n). (cid:3) inverse semigroup actions and self-similar graphs 29 8.8. Corollary. Given u1 and u2 in Gtight(SG,E), such that (u1, u2) ∈ Gtight(SG,E)(2), that is, such that the multiplication u1u2 is allowed or, equivalently, such that d(u1) = r(u2), then there are representations of u1 and u2 of the form u1 =(cid:2)α1, g1, α2; α2g2ξ(cid:3), and u2 =(cid:2)α2, g2, β; βξ(cid:3), and in this case Proof. Using (8.7), write u1u2 =(cid:2)α1, g1g2, β; βξ(cid:3). ui =(cid:2)αi, gi, βi; βiξi(cid:3), with β1 = α2. By virtue of (u1, u2) lying in Gtight(SG,E)(2), we have that β1ξ1 = (α2, g2, β2)(β2ξ2) = α2g2ξ2, so in fact β1 = α2, and ξ1 = g2ξ2. Then u1 =(cid:2)α1, g1, β1; β1ξ1(cid:3) =(cid:2)α1, g1, α2; α2g2ξ2(cid:3), and it suffices to put ξ = ξ2, and β = β2. With respect to the last assertion we have that u1u2 = [s; βξ], where s is the element of SG,E given by concluding the proof. s = (α1, g1, α2)(α2, g2, β) (4.2) = (α1, g1g2, β), (cid:3) Having extended the action of G to the set of infinite paths in Proposition (8.1), one may ask whether it is possible to do the same for the cocycle ϕ. The following is an attempt at this which however produces a map taking values in the infinite product G∞, rather than in G. 8.9. Definition. We will denote by Φ, the map defined by the rule for g ∈ G, ξ ∈ E∞, and n ≥ 1. Φ : G × E∞ → G∞ Φ(g, ξ)n = ϕ(g, ξn−1), Recall that we are indexing the elements of G∞ on the set {1, 2, 3, . . .}, so the first coordinate of Φ(g, ξ) is Φ(g, ξ)1 = ϕ(g, ξ0) = ϕ(cid:0)g, r(ξ)(cid:1) (2.5.ii) = g. We wish to view Φ as some sort of cocycle but, unfortunately, property (2.5.x) does not hold quite as stated. On the fortunate side, a suitable modification of this relation, involving the left shift endomorphism λ of G∞, is satisfied: 30 r. exel and e. pardo 8.10. Proposition. Let α be a finite path and let ξ be an infinite path such that d(α) = r(ξ). Then, for every g in G, one has that Proof. For all n ≥ 1, we have Φ(cid:0)ϕ(g, α), ξ(cid:1) = λα(cid:0)Φ(cid:0)g, αξ)(cid:1). Φ(cid:0)ϕ(g, α), ξ(cid:1)n = ϕ(cid:0)ϕ(g, α), ξn−1(cid:1) = ϕ(cid:0)g, α(ξn−1)(cid:1) = = ϕ(cid:0)g, (αξ)n−1+α(cid:1) = λα(cid:0)Φ(g, αξ)(cid:1)n. Another reason to think of Φ as a cocycle is as follows: 8.11. Proposition. For every ξ ∈ E∞, and every g, h ∈ G, we have that Proof. We have for all n ∈ , that Φ(gh, ξ) = Φ(cid:0)g, hξ(cid:1)Φ(h, ξ). (8.1) = (cid:3) (cid:3) Φ(gh, ξ)n = ϕ(gh, ξn−1) (2.5.b) = ϕ(cid:0)g, h(ξn−1)(cid:1)ϕ(h, ξn−1) = ϕ(cid:0)g, (hξ)n−1(cid:1)ϕ(h, ξn−1) = Φ(cid:0)g, hξ(cid:1)nΦ(h, ξ)n. The following elementary fact might perhaps justify the choice of "n − 1" in the definition of Φ. 8.12. Proposition. Given g ∈ G, and ξ ∈ E∞, one has that (gξ)n = Φ(g, ξ)n ξn. Proof. By (8.1) we have that (gξ)n = g(ξn), so the nth coordinate of gξ is also the nth coordinate of g(ξn). In addition we have that g(ξn) = g(ξn−1ξn) (2.5.ix) = g(ξn−1)ϕ(g, ξn−1)ξn, so (gξ)n = ϕ(g, ξn−1)ξn = Φ(g, ξ)n ξn. (cid:3) We now wish to define a homomorphism (sometimes also called a one-cocycle) from Gtight(SG,E) to the lag group G ⋊ρ Z, by means of the rule (cid:2)α, g, β; βξ(cid:3) 7→(cid:16)ρα(cid:0)Φ(g, ξ)(cid:1), α − β(cid:17). As it is often the case for maps defined on groupoid of germs, the above tentative definition uses a representative of the germ, so some work is necessary to prove that the definition does not depend on the choice of representative. The technical part of this task is the content of our next result. inverse semigroup actions and self-similar graphs 31 8.13. Lemma. Suppose that (G, E, ϕ) is pseudo free. For each i = 1, 2, let us be given (αi, gi, βi) in SG,E, as well as ηi = βiξi ∈ Z(βi). If then modulo G(∞). (cid:2)α1, g1, β1; η1(cid:3) =(cid:2)α2, g2, β2; η2(cid:3), ρα1(cid:0)Φ(g1, ξ1)(cid:1) ≡ ρα2(cid:0)Φ(g2, ξ2)(cid:1) Proof. Assuming without loss of generality that β1 ≤ β2, we may use (8.6) to write α2 = α1g1γ, g2 = ϕ(g1, γ), β2 = β1γ, and η1 = η2 = β1γξ, for suitable γ ∈ E∗ and ξ ∈ E∞. Then necessarily ξ1 = γξ, and ξ2 = ξ, and ρα2(cid:0)Φ(g2, ξ2)(cid:1) = ρα1+γ(cid:16)Φ(cid:0)ϕ(g1, γ), ξ(cid:1)(cid:17) (8.10) ≡ ρα1(cid:0)Φ(cid:0)g1, ξ1)(cid:1). = ρα1ργλγ(cid:0)Φ(cid:0)g1, γξ)(cid:1) (7.3) = (cid:3) ◮ Due to our reliance on Proposition (8.6) and Lemma (8.13), from now on and until the end of this section we will assume, in addition to (2.3), that (G, E, ϕ) is pseudo free. If g is in G∞, we will denote by g its class in the quotient group G. Likewise we will denote by Φ the composition of Φ with the quotient map from G∞ to G. Φ G × E∞ −→ G∞ −→ G ................................................................................................................................................................................................................................... .......... Φ It then follows from (8.13) that the correspondence (cid:2)α, g, β; βξ(cid:3) ∈ Gtight(SG,E) 7→ ρα(cid:0) Φ(g, ξ)(cid:1) ∈ G is a well defined map. This is an important part of the one-cocycle we are about to introduce. 8.14. Proposition. The correspondence gives a well defined map ℓ :(cid:2)α, g, β; βξ(cid:3) 7→(cid:16)ρα(cid:0) Φ(g, ξ)(cid:1), α − β(cid:17) ℓ : Gtight(SG,E) → G ⋊ρ Z, which is moreover a one-cocycle. From now on ℓ will be called the lag function. 32 r. exel and e. pardo Proof. By the discussion above we have that the first coordinate of the above pair is well defined. On the other hand, in the context of Proposition (8.6) one easily sees that α1 − β1 = α2 − β2, so the second coordinate is also well defined. In order to show that ℓ is multiplicative, pick (u1, u2) ∈ Gtight(SG,E)(2). We may then use (8.8) to write So u2 =(cid:2)α2, g2, β; βξ(cid:3). u1 =(cid:2)α1, g1, α2; α2g2ξ(cid:3), and ℓ(u1)ℓ(u2) =(cid:16)ρα1(cid:0)Φ(g1, g2ξ)(cid:1), α1 − α2(cid:17)(cid:16)ρα2(cid:0)Φ(g2, ξ)(cid:1), α2 − β(cid:17) = α1 − α2 + α2 − β(cid:17) = =(cid:16)ρα1(cid:0)Φ(g1, g2ξ)(cid:1) ρα1(cid:0)Φ(g2, ξ)(cid:1), =(cid:16)ρα1(cid:16)Φ(g1, g2ξ)Φ(g2, ξ)(cid:17), α1 − β(cid:17) (8.11) = (cid:16)ρα1(cid:0)Φ(g1g2, ξ)(cid:1), α1 − β(cid:17) = = ℓ(cid:0)(cid:2)α1, g1g2, β; βξ(cid:3)(cid:1) (8.8)  d(u1) = d(u2) ℓ(u1) = ℓ(u2) r(u1) = r(u2) ⇒ u1 = u2. Proof. Using (8.7), write ui =(cid:2)αi, gi, βi; βiξi(cid:3), for i = 1, 2, with β1 = β2. Since β1ξ1 = d(u1) = d(u2) = β2ξ2, The main relevance of this one-cocycle is that, together with the domain and range maps, it uniquely describes the elements of Gtight(SG,E), as we will now show. 8.15. Proposition. Given u1, u2 ∈ Gtight(SG,E), one has that = ℓ(u1u2). (cid:3) we conclude that β1 = β2, and ξ1 = ξ2 =: ξ. By focusing on the second coordinate of ℓ(ui), we see that α1 − β1 = α2 − β2, and hence α1 = α2. Moreover, since α1g1ξ = α1g1ξ1 = r(u1) = r(u2) = α2g2ξ2 = α2g2ξ, we see that α1 = α2, and g1ξ = g2ξ. (8.15.1) The fact that ℓ(u1) = ℓ(u2) also implies that ρα1(cid:0) Φ(g1, ξ)(cid:1) = ρα2(cid:0) Φ(g2, ξ)(cid:1), and since α1 = α2, we conclude that Φ(g1, ξ) = Φ(g2, ξ), and hence that there exists an integer n0 such that By (8.15.1) we also have that g1(ξn) = g2(ξn), so (5.6) gives g1 = g2, whence u1 = u2. (cid:3) ϕ(g1, ξn) = ϕ(g2, ξn), ∀ n ≥ n0. inverse semigroup actions and self-similar graphs 33 As a consequence of the above result we see that the map F : Gtight(SG,E) → E∞ × ( G ⋊ρ Z) × E∞ defined by the rule is one-to-one. F (u) =(cid:0)r(u), ℓ(u), d(u)(cid:1), (8.16) Observe that the co-domain of F has a natural groupoid structure, being the cartesian product of the lag group G ⋊ρ Z by the graph of the transitive equivalence relation on E∞. Putting together (8.14) and (8.15) we may now easily prove: 8.17. Corollary. F is a groupoid homomorphism (functor), hence establishing an iso- morphism from Gtight(SG,E) to the range of F . The range of F is then the concrete model of Gtight(SG,E) we are after. But, before giving a detailed description of it, let us make a remark concerning notation: since the co-domain of F is a mixture of cartesian and semi-direct products, the standard notation for its elements would be something like (cid:0)η, (u, p), ζ(cid:1), for η, ζ ∈ E∞, u ∈ G, and p ∈ Z. As part of our effort to avoid heavy notation we will instead denote such an element by (cid:0)η; u, p; ζ(cid:1). 8.18. Proposition. The range of F is precisely the subset of E∞ × ( G ⋊ρ Z) × E∞, formed by the elements (η; g, p − q; ζ), where η, ζ ∈ E∞, g ∈ G∞, and p, q ∈ , are such that, for all n ≥ 1, (i) gn+p+1 = ϕ(gn+p, ζn+q), (ii) ηn+p = gn+pζn+q. Proof. Pick a general element(cid:2)α, g, β; βξ(cid:3) ∈ Gtight(SG,E) and, recalling that let η = αgξ, g = ρα(cid:0)Φ(g, ξ)(cid:1), p = α, q = β, and ζ = βξ, so that the element depicted F ((cid:2)α, g, β; βξ(cid:3)) =(cid:16)αgξ; ρα(cid:0) Φ(g, ξ)(cid:1), α − β; βξ(cid:17), in (8.18.1) becomes (η; g, p − q; ζ), and we must now verify (i) and (ii). For all n ≥ 1, one has that (8.18.1) so gn+α = Φ(g, ξ)n = ϕ(g, ξn−1), ηn+p = (αgξ)n+α = (gξ)n (8.12) = ϕ(g, ξn−1)ξn = gn+α(βξ)n+β = gn+pζn+q, proving (ii). Also, gn+p+1 = gn+α+1 = ϕ(cid:0)g, ξn(cid:1) = ϕ(g, ξn−1ξn) = ϕ(cid:0)ϕ(g, ξn−1), ξn(cid:1) = = ϕ(cid:0)gn+α, (βξ)n+β(cid:1) = ϕ(cid:0)gn+p, ζn+q(cid:1), 34 r. exel and e. pardo proving (i) and hence showing that the range of F is a subset of the set described in the statement. Conversely, pick η, ζ ∈ E∞, g ∈ G∞, and p, q ∈ satisfying (i) and (ii), and let us show that the element (η; g, p − q; ζ) lies in the range of F . Let g = gp+1, α = ηp, and β = ζq, order to see this notice that so ζ = βξ for a unique ξ ∈ E∞. We then claim that(cid:2)α, g, β; βξ(cid:3) lies in Gtight(SG,E). In so (α, g, β) ∈ SG,E, and therefore (cid:2)α, g, β; βξ(cid:3) is indeed a member of Gtight(SG,E). The proof will then be concluded once we show that gd(β) = gd(ζq) = gr(ζq+1) = r(gp+1ζq+1) (ii) = r(ηp+1) = d(ηp) = d(α), F ((cid:2)α, g, β; βξ(cid:3)) = (η; g, p − q; ζ), which in turn is equivalent to showing that (a) αgξ = η, (b) ρα(cid:0) Φ(g, ξ)(cid:1) = g, ( c ) α − β = p − q, (d) βξ = ζ. Before proving these points we will show that ϕ(gp+1, ξn) = gn+p+1, ∀ n ≥ 0. (†) This is obvious for n = 0. Assuming that n ≥ 1 and using induction, we have ϕ(gp+1, ξn) = ϕ(gp+1, ξn−1ξn) = ϕ(cid:0)ϕ(gp+1, ξn−1), ξn(cid:1) = = ϕ(gn+p, ζn+q) (i) = gn+p+1, verifying (†). Addressing (a) we have to prove that (αgξ)k = ηk, for all k ≥ 1, but given that α is defined to be ηp, this is trivially true for k ≤ p. On the other hand, for k = n + p, with n ≥ 1, we have (αgξ)k = (αgξ)n+p = (gξ)n (8.12) = ϕ(g, ξn−1)ξn = = ϕ(gp+1, ξn−1)ξn (†) = gn+pζn+q (ii) = ηn+p = ηk, proving (a). Focusing on (b) we have for all n ≥ 1 that ρα(cid:0)Φ(g, ξ)(cid:1)p+n = Φ(g, ξ)n = ϕ(gp+1, ξn−1) (†) = gn+p, proving that ρα(cid:0)Φ(g, ξ)(cid:1) ≡ g, modulo G(∞), hence taking care of (b). The last two points, namely (c) and (d) are trivial and so the proof is concluded. (cid:3) inverse semigroup actions and self-similar graphs 35 As an immediate consequence we get a very precise description of the algebraic struc- ture of Gtight(SG,E): 8.19. Theorem. Suppose that (G, E, ϕ) satisfies the conditions of (2.3) and is moreover pseudo free. Then Gtight(SG,E) is isomorphic to the sub-groupoid of E∞ × ( G ⋊ρ Z) × E∞ given by GG,E = (η; g, p − q; ζ) ∈ E∞ × ( G ⋊ρ Z) × E∞ : g ∈ G∞, p, q ∈ , gn+p+1 = ϕ(gn+p, ζn+q), ηn+p = gn+pζn+q, for all n ≥ 1  .  Recall from [20] that the C-algebra of every graph is a groupoid C-algebra for a certain groupoid constructed from the graph, and informally called the groupoid for the tail equivalence with lag. Viewed through the above perspective, our groupoid may also deserve such a denomi- nation, except that the lag is not just an integer as in [20], but an element of the lag group G ⋊ρ Z precisely described by the lag function ℓ introduced in Proposition (8.14). 9. The topology of Gtight(SG,E). It is now time we look at the topological aspects of Gtight(SG,E). In fact what we will do is simply transfer the topology of Gtight(SG,E) over to GG,E via F . Not surprisingly F will turn out to be an isomorphism of topological groupoids. Recall from [6: Proposition 4.14] that, if S is an inverse semigroup acting on a locally compact Hausdorff topological space X, then the corresponding groupoid of germs, say G, is topologized by means of the basis consisting of sets of the form Θ(s, U ), where s ∈ S, and U is an open subset of X, contained in the domain of the partial homeomorphism attached to s by the given action. Each Θ(s, U ) is in turn defined by See [6: 4.12] for more details. Θ(s, U ) =n[s, x] ∈ G : x ∈ Uo. (9.1) If we restrict the choice of the U 's above to a predefined basis of open sets of X, e.g. the collection of all cylinders in E∞ in the present case, we evidently get the same topology on the groupoid of germs. Therefore, referring to the model of Gtight(SG,E) presented in (8.5), we see that a basis for its topology consists of the sets of the form Θ(α, g, β; γ) :=n(cid:2)α, g, β; ξ(cid:3) ∈ Gtight(SG,E) : ξ ∈ Z(γ)o, (9.2) where (α, g, β) ∈ SG,E, and γ ∈ E∗. We may clearly suppose that γ ≥ β and, since Θ(α, g, β; γ) = ∅, unless β is a prefix of γ, we may also assume that γ = βε, for some ε ∈ E∗. 36 r. exel and e. pardo In this case, given any(cid:2)α, g, β; ξ(cid:3) ∈ Θ(α, g, β; γ), notice that ξ ∈ Z(γ), and from where one concludes that for all ξ ∈ Z(γ), and hence also that (α, g, β)(γ, 1, γ) =(cid:0)αgε, ϕ(g, ε), γ(cid:1), (cid:2)α, g, β; ξ(cid:3) =(cid:2)αgε, ϕ(g, ε), γ; ξ(cid:3), Θ(α, g, β; γ) = Θ(cid:0)αgε, ϕ(g, ε), γ; γ(cid:1). This shows that any set of the form (9.2) coincides with another such set for which β = γ. We may therefore do away with this repetition and redefine Θ(α, g, β) :=n(cid:2)α, g, β; ξ(cid:3) ∈ Gtight(SG,E) : ξ ∈ Z(β)o. We have therefore shown: (9.3) 9.4. Proposition. The collection of all sets of the form Θ(α, g, β), where (α, g, β) range in SG,E, is a basis for the topology of Gtight(SG,E). We may now give a precise description of the topology of Gtight(SG,E), once it is viewed from the alternative point of view of Theorem (8.19): 9.5. Proposition. For each (α, g, β) in SG,E, the image of Θ(α, g, β) under F coincides with the set Ω(α, g, β) := (η; g, k; ζ) ∈ GG,E : η ∈ Z(α), g ∈ G∞, k = α − β, ζ ∈ Z(β), g1+α = g, gn+α+1 = ϕ(gn+α, ζn+β), ηn+α = gn+αζn+β, for all n ≥ 1 ,  and hence the collection of all such sets form the basis for a topology on GG,E, with respect to which the latter is isomorphic to Gtight(SG,E) as topological groupoids. Proof. Left for the reader. (cid:3) We may now summarize the main results obtained so far: 9.6. Theorem. Suppose that (G, E, ϕ) satisfies the conditions of (2.3) and is moreover pseudo free. Then OG,E is -isomorphic to the C-algebra of the groupoid GG,E described in (8.19), once the latter is equipped with the topology generated by the basis of open sets Ω(α, g, β) described in (9.3), for all (α, g, β) in SG,E. inverse semigroup actions and self-similar graphs 37 10. OG,E as a Cuntz-Pimsner algebra. Inspired by Nekrashevych's paper [24], we will now give a description of OG,E as a Cuntz- Pimsner algebra [29]. With this we will also be able to prove that OG,E is nuclear and that Gtight(SG,E) is amenable when G is an amenable group. As before, we will work under the conditions of (2.3). We begin by introducing the algebra of coefficients over which the relevant Hilbert bimodule, also known as a correspondence, will later be constructed. Since the action of G on E preserves length by (2.4.iv), we see that the set of vertices of E is G-invariant, so we get an action of G on E0 by restriction. By dualization G acts on the algebra C(E0) of complex valued functions5 on E0. We may therefore form the crossed-product C-algebra A = C(E0) ⋊ G. Since C(E0) is a unital algebra, there is a canonical unitary representation of G in the crossed product, which we will denote by {vg}g∈G. On the other hand, C(E0) is also canonically isomorphic to a subalgebra of A and we will therefore identify these two algebras without further warnings. For each x in E0, we will denote the characteristic function of the singleton {x} by qx, so that {qx : x ∈ E0} is the canonical basis of C(E0), and thus A coincides with the closed linear span of the set (10.1) (cid:8)qxvg : x ∈ E0, g ∈ G(cid:9). For later reference, notice that the covariance condition in the crossed product reads vgqx = qgxvg, ∀ x ∈ E0, ∀ g ∈ G. (10.2) Our next step is to construct a correspondence over A. In preparation for this we denote by Ae the right ideal of A generated by qd(e), for each e ∈ E1. In technical terms Ae = qd(e)A. With the obvious right A-module structure, and the inner product defined by hy, zi = y∗z, ∀ y, z ∈ Ae, one has that Ae is a right Hilbert A-module. Notice that this is not necessarily a full Hilbert module since hAe, Aei is the two-sided ideal of A generated by qd(e), which might be a proper ideal in some cases. As already seen in (10.1), A is spanned by the elements of the form qxvg. Therefore Ae is spanned by the elements of the form qd(e)qxvg, but, since the q's are mutually orthogonal, this is either zero or equal to qd(e)vg. Therefore we see that Ae = span{qd(e)vg : g ∈ G}. 5 Notice that, since E0 is a finite set, C(E0) is nothing but C E0. 38 r. exel and e. pardo Introducing the right Hilbert A-module which will later be given the structure of a correspondence over A, we define M = Me∈E1 Ae. Observe that if x is a vertex which is the source of many edges, say then d−1(x) = {e1, e2, . . . , en}, Aei = qd(ei)A = qxA, for all i, so that qxA appears many times as a direct summand of M . However these copies of qxA should be suitably distinguished, according to which edge ei is being considered. On the other hand, notice that if d−1(x) = ∅, then qxA does not appear among the summands of M , at all. Addressing the fullness of M , observe that hM, M i = Xx∈E0 d−1(x)6=∅ AqxA, so, when E has no sinks, that is, when d−1(x) is nonempty for every x, one has that M is full. Given e ∈ E1, the element qd(e), when viewed as an element of Ae ⊆ M , will play a very special role in what follows, so we will give it a special notation, namely (10.3) There is a small risk of confusion here in the sense that, if e1, e2 ∈ E1 are such that te := qd(e). x := d(e1) = d(e2), then (10.3) assigns qx to both te1 and te2. However the coordinate in which qx appears in tei is determined by the corresponding ei, so if e1 6= e2, then te1 6= te2 . In order to completely dispel any confusion, here is the technical definition: where We should notice that te = (mf )f ∈E1 , mf =(cid:26) qd(e), 0, if f = e, otherwise. teqd(e) = te, and that any element y ∈ M may be written uniquely as y = Xe∈E1 teye, where each ye ∈ Ae. As the next step in constructing a correspondence over A, we would now like to define a certain -homomorphism from A to the algebra L(M ) of adjointable linear operators on M . Since A is a crossed product algebra, this will be accomplished once we produce a covariant representation (ψ, V ) of the C-dynamical system(cid:0)C(E0), G(cid:1) on M . We begin with the group representation V . (10.4) (10.5) inverse semigroup actions and self-similar graphs 39 10.6. Definition. For each g ∈ G, let Vg be the linear operator on M given by Vg(cid:16) Xe∈E1 teye(cid:17) = Xe∈E1 tgevϕ(g,e)ye, whenever ye ∈ Ae, for each e in E1. By the uniqueness in (10.5), it is clear that Vg is well defined. 10.7. Proposition. Each Vg is a unitary operator on M . Moreover, the correspondence g 7→ Vg is a unitary representation of G. Proof. Pick g in G. We begin by claiming that the two sides in the expression defining Vg, above, coincide whenever the ye are in A, even if ye does not belong to Ae. Since Vg is clearly additive, we only need to check that Observing that te = teqd(e), we have Vg(tey) = tgevϕ(g,e)y, ∀ y ∈ A. Vg(tey) = Vg(teqd(e)y) = tgevϕ(g,e)qd(e)y = = tgeqϕ(g,e)d(e)vϕ(g,e)y (2.5.vii) = tgeqd(ge)vϕ(g,e)y = tgevϕ(g,e)y, proving the claim. One therefore concludes that Vg is right-A-linear. We next claim that, for all e, f ∈ E1, one has We have hVg(te), tf i = hte, Vg−1(tf )i. (10.7.1) hVg(te), tf i = htgevϕ(g,e), tf i = v∗ ϕ(g,e)htge, tf i = [ge=f] v−1 ϕ(g,e)qd(ge) = = [ge=f] qϕ(g,e)−1d(ge)v−1 ϕ(g,e) (2.5.vii) = [ge=f] qd(e)v−1 ϕ(g,e) = (⋆). Starting from the right-hand-side of (10.7.1), we have hte, Vg−1(tf )i = hte, tg−1f vϕ(g−1,f )i = [e=g−1f] qd(e)vϕ(g−1,f ) = = [ge=f] qd(e)vϕ(g,g−1f )−1 = [ge=f] qd(e)v−1 ϕ(g,e), which agrees with (⋆) above, and hence proves claim (10.7.1). If y, z ∈ A, we then have that hVg(tey), tf zi = y∗hVg(te), tf iz = y∗hte, Vg−1(tf )iz = htey, Vg−1(tf z)i, from where one sees that hVg(ξ), ηi = hξ, Vg−1(η)i, for all ξ, η ∈ M , hence proving that Vg is an adjointable operator with V ∗ g = Vg−1. Let us next prove that VgVh = Vgh, ∀ g, h ∈ G. By A-linearity it is enough to prove that these operators coincide on the set formed by the te's, which is a generating set for M . We thus compute Vg(cid:0)Vh(te)(cid:1) = Vg(cid:0)thevϕ(h,e)(cid:1) = tghevϕ(g,he)vϕ(h,e) = tghevϕ(gh,e) = Vgh(te). Since it is evident that V1 is the identity operator on M we obtain, as a consequence, (cid:3) g , so each Vg is unitary and the proof is concluded. that V −1 g = Vg−1 = V ∗ 40 r. exel and e. pardo In order to complete our covariant pair we must now construct a -homomorphism from C(E0) to L(M ). With this in mind we give the following: 10.8. Definition. For every x in E0, let Mx = Me∈r−1(x) Ae, which we view as a complemented sub-module of M . orthogonal projection from M to Mx, so that In addition, we let Qx be the Qx(tey) = [r(e)=x] tey, ∀ e ∈ E1, ∀ y ∈ A. (10.8.1) Observe that the Qx are pairwise orthogonal projections and thatPx∈E0 Qx = 1. 10.9. Definition. Let ψ : C(E0) → L(M ) be the unique unital -homomorphism such that ψ(qx) = Qx, ∀ x ∈ E0. From our working hypothesis that E has no sources, we see that for every x in E0, there is some e ∈ E1 such that r(e) = x. So whence Qx 6= 0. Consequently ψ is injective. Qx(te) = te, 10.10. Proposition. The pair (ψ, V ) is a covariant representation of the C-dynamical system(cid:0)C(E0), G(cid:1) in L(M ). Proof. All we must do is check the covariance condition Vgψ(y) = ψ(cid:0)σg(y)(cid:1)Vg, ∀ g ∈ G, ∀ y ∈ C(E0), where σ is the name we temporarily give to the action of G on C(E0). Since C(E0) is spanned by the qx, it suffices to consider y = qx, in which case the above identity becomes VgQx = QgxVg. (10.10.1) Furthermore M is generated, as an A-module, by the te, for e ∈ E1, so we only need to verify this on the te. We have while Vg(cid:0)Qx(te)(cid:1) = [r(e)=x] Vg(te) = [r(e)=x] tgevϕ(g,e), Qgx(cid:0)Vg(te)(cid:1) = Qgx(cid:0)tgevϕ(g,e)(cid:1) = [r(ge)=gx] tgevϕ(g,e), verifying (10.10.1) and concluding the proof. (cid:3) inverse semigroup actions and self-similar graphs 41 It follows from [28: Proposition 7.6.4 and Theorem 7.6.6] that there exists a -homo- morphism such that and Ψ : C(E0) ⋊ G → L(M ), Ψ(qx) = Qx, ∀ x ∈ E0, Ψ(vg) = Vg, ∀ g ∈ G. Equipped with the left-A-module structure provided by Ψ, we then have that M is a correspondence over A. For later reference we record here a few useful calculations involving the left-module structure of M . 10.11. Proposition. Let g ∈ G, e ∈ E1, and x ∈ E0. Then (i) vgte = tgevϕ(g,e), (ii) qxvgte = [r(ge)=x] tgevϕ(g,e). Proof. We have proving (a). Also vgte = Ψ(vg)te = Vg(te) = tgevϕ(g,e), qxvgte = Ψ(qx)(vgte) = Qx(tgevϕ(g,e)) = [r(ge)=x] tgevϕ(g,e). (cid:3) It is our next goal to prove that OG,E is naturally isomorphic to the Cuntz-Pimsner algebra associated to the correspondence M , which we denote by OM . As a first step, we identify a certain Cuntz-Krieger E-family. 10.12. Proposition. The following relations hold within OM . (a) For every x ∈ E0, one has thatPe∈r−1(x) tet∗ (b) Pe∈E1 tet∗ e = 1. e = qx. ( c ) The set {qx : x ∈ E0} ∪ {te : e ∈ E1} is a Cuntz-Krieger E-family. Proof. We first claim that, for every x ∈ E0, and every m ∈ M , one has that Xe∈r−1(x) tet∗ em = qxm. To prove it, it is enough to consider the case in which m = tf , for f ∈ E1, since these generate M . In this case we have Xe∈r−1(x) tet∗ etf = [r(f )=x] tf t∗ f tf = [r(f )=x] tf (10.8.1) = Qx(tf ) = qxtf , 42 r. exel and e. pardo proving the claim. This says that the pair(cid:0)qx,Pe∈r−1(x) tet∗ the terminology of [29], that the generalized compact operator e(cid:1) is a redundancy or, adopting Xe∈r−1(x) Ωte,te is mapped to Ψ(qx) via Ψ(1). Therefore qx = Xe∈r−1(x) tet∗ e, in OM , proving (a). Point (b) then follows from the fact thatPx∈E0 qx = 1. Focusing now on (c), it is evident that {qx : x ∈ E0} is a family of mutually orthogonal projections. Moreover, for each e ∈ E1, we have t∗ ete = hte, tei = qd(e), proving (3.1.i) and also that te is a partial isometry. Property (3.1.ii) also holds in view of (a), so the proof is concluded. (cid:3) 10.13. Proposition. There exists a unique surjective -homomorphism Λ : OG,E → OM such that Λ(px) = qx, Λ(se) = te, and Λ(ug) = vg. Proof. By the universal property of OG,E, in order to prove the existence of Λ it is enough to check that the qx, te, and vg satisfy the conditions of Definition (3.2). Condition (3.2.a) has already been proved above while (3.2.b) is evidently true since v is a representation of G in C(E0) ⋊ G ⊆ OM . Condition (3.2.c) is precisely (10.11.i), while (3.2.d) was taken care of in (10.2). Since A is spanned by the qx and the vg by (10.1), and since M is generated over A by the te, we see that OM is spanned by the set {qx, te, vg : x ∈ E0, e ∈ E1, g ∈ G}, so Λ is surjective. (cid:3) Let us now prove that Λ is invertible by providing an inverse to it. Since A is the crossed product C-algebra C(E0) ⋊ G, one sees that (3.2.a&d) guarantees the existence of a -homomorphism θA : A → OG,E, sending the qx to the px, and the vg to the ug. For each e in E1, consider the linear mapping θM : M → OG,E, inverse semigroup actions and self-similar graphs 43 given, for every m = (me)e∈E1 ∈ M , by θM (m) = Xe∈E1 seθA(me) ∈ OG,E. Notice that θM (te) = se, for all e ∈ E1, because θM (te) = seθA(qd(e)) = sepd(e) = se. 10.14. Lemma. The pair (θA, θM ) is a representation of the correspondence M in the sense of [29: Theorem 3.4], meaning that for all y ∈ A and all ξ, ξ′ ∈ M , (i) θM (ξ)θA(y) = θM (ξy), (ii) θA(y)θM (ξ) = θM (yξ), (iii) θM (ξ)∗θM (ξ′) = θA(hξ, ξ′i). Proof. Considering the various spanning sets at our disposal, we may assume that y = qxvg, that ξ = tez, and ξ′ = te′ z′, with x ∈ E0, g ∈ G, e, e′ ∈ E1, z ∈ qd(e)A, and z′ ∈ qd(e′)A. We then have θM (ξ)θA(y) = θM (tez)θA(y) = seθA(z)θA(y) = seθA(zy) = θM (tezy) = θM (ξy), proving (i). As for (ii), we have θA(y)θM (ξ) = θA(qxvg)θM (tez) = pxugseθA(z) = pxsgeuϕ(g,e)θA(z) = = [r(ge)=x] sgeθA(vϕ(g,e)z) = [r(ge)=x] θM (tgevϕ(g,e)z(cid:1) (10.11.ii) proving (ii). Focusing now on (iii), we have = θM (qxvgtez) = θM (yξ), θM (ξ)∗θM (ξ′) = (seθA(z)(cid:1)∗ se′θA(z′) = [e=e′] θA(z)∗pd(e)θA(z′) = = [e=e′] θA(z∗qd(e)z′) = θA(hξ, ξ′i). (cid:3) It is well known [29: Theorem 3.4] that the Toeplitz algebra for the correspondence M , usually denoted TM , is universal for representations of M , so there exists a -homo- morphism coinciding with θA on A and with θM on M . Θ0 : TM → OG,E, 10.15. Theorem. The map Θ0, defined above, factors through OM , providing a - isomorphism Θ : OM → OG,E, such that Θ(qx) = px, Θ(te) = se, and Θ(vg) = ug, for all x ∈ E0, e ∈ E1, and g ∈ G. 44 r. exel and e. pardo Proof. The factorization property follows immediately from (10.12.b) and an easy modifi- cation of [14: Proposition 7.1] to Cuntz-Pimsner algebras. In order to prove that Θ is an isomorphism, observe that Θ ◦ Λ coincides with the identity map on the generators of OG,E, by (10.13), and hence Θ ◦ Λ = id. The result then follows from the fact that Λ is surjective. (cid:3) 10.16. Corollary. If G is amenable then OG,E is nuclear. Proof. The amenability of G ensures that C(E0) ⋊ G is nuclear. The result then follows from Theorem (10.15), the fact that Toeplitz-Pimsner algebras over nuclear coefficient al- gebras is nuclear [5: Theorem 4.6.25], and so are quotients of nuclear algebras [5: Theorem 9.4.4]. (cid:3) 10.17. Remark. Since E0 is finite, the nuclearity of C(E0) ⋊ G is equivalent to the amenability of G. However, if the present construction is generalized for infinite graphs, one could produce examples of non amenable groups acting amenably on E0, in which case C(E0) ⋊ G would be nuclear. The proof of Corollary (10.16) could then be adapted to prove that OG,E is nuclear. 10.18. Corollary. If G is amenable, then Gtight(SG,E) is an amenable groupoid. If more- over (G, E, ϕ) is pseudo free, then its sibling GG,E is an amenable groupoid. Proof. For Gtight(SG,E), it follows from (10.16), (6.4) and [5: Theorem 5.6.18]. For GG,E, it follows from (10.16), (9.6) and [5: Theorem 5.6.18]. (cid:3) Nekrashevych has proven in [26: Theorem 5.6], that a certain groupoid of germs, denoted DG, constructed in the context of self-similar groups, is amenable under the hy- pothesis that the group is contracting and self-replicating. Even though there are numerous differences between DG and GG,E, including a different notion of germs and Nekrashevych's requirement that group actions be faithful, we believe it should be interesting to try to generalize Nekrashevych's result to our context. 11. Representing C∗(E) and G into OG,E . In this section we will study natural representations of the graph C-algebra C∗(E) and of the group G in OG,E. As before, we keep (2.3) in force. Given that {px : x ∈ E0} ∪ {se : e ∈ E1} is a Cuntz-Krieger E-family, the universal property of the graph C-algebra C∗(E) [30] provides for the existence of a -homomorphism ι : C∗(E) → OG,E, sending the canonical Cuntz-Krieger E-family of C∗(E) to the corresponding one within OG,E. 11.1. Proposition. The -homomorphism ι above is injective. inverse semigroup actions and self-similar graphs 45 Proof. Using the universal property of OG,E, it is easy to see that, for each complex number z, with z = 1, there is a -homomorphism γz : OG,E → OG,E, satisfying γz(px) = px, γz(se) = zse, and γz(ug) = ug, for all x ∈ E0, e ∈ E1 and g ∈ G. It is also easy to see that the correspondence z → γz defines an action of the circle group on OG,E, and moreover that ι is covariant relative to this action on OG,E, on the one hand, and the standard gauge action on C∗(E), on the other. In order to prove the injectivity of ι we may then apply the gauge invariant uniqueness Theorem [30: Theorem 2.2], which requires, in addition, that we verify that the px are nonzero. To prove this we observe that, in the groupoid model of OG,E given by (9.6), for each x in E0, the element px is the characteristic function of the cylinder Z(x), seen as a subset of E∞, which in turn is the unit space of the groupoid GG,E. Since E has no sources, we have that Z(x) is nonempty, whence px is nonzero, as required. This concludes the proof. (cid:3) In particular, (11.1) implies that SE is ∗-isomorphic to the inverse semigroup of OG,E generated by {sa : a ∈ E1}. With respect to the injectivity of the representation of G into OG,E, we have to work a bit more to obtain a result in the line of (11.1). 11.2. Lemma. Let π : SG,E → OG,E and u : G → OG,E be the natural maps. If π is injective, the so is u. Proof. Let g ∈ G such that ug = u1. For any x ∈ E0 we have π(x, g, g−1x) = pxug and π(x, 1, x) = pxu1 = px. Since ug = u1, we get (x, g, g−1x) = (x, 1, x) ∈ SG,E, whence g = 1. (cid:3) We need to recall some extra definitions. Let G be an ´etale groupoid, i.e. a topological groupoid whose unit space G(0) is locally compact and Hausdorff in the relative topology, and such that the range map r : G → G(0) is a local homeomorphism (and then so is the source map d : G → G(0)). An open set U ⊂ G is a slice if the restrictions of r and d to [27]). In particular, G(0) is a slice [6: Proposition 3.4], and the U are injective (see e.g. collection of all slices forms a basis for the topology of G [6: Proposition 3.5]. 11.3. Definition. We denote by Sℓ(G, E) the set of all compact slices. It is well known (see e.g. [27: Proposition 2.2.4]) that Sℓ(G, E) forms a ∗-inverse semigroup with the op- erations U V = {uv : u ∈ U, v ∈ V, (u, v) ∈ G(2)}, and U ∗ = {u−1 : u ∈ U }. Moreover, if UG,E = {1U : U is a compact slice} ⊆ C∗(Gtight(SG,E)) is the semigroup formed by their characteristic functions, then UG,E ∼= Sℓ(G, E). (11.3.1) 46 r. exel and e. pardo Fix the canonical action θ of SG,E on E∞. Given any (α, g, β) ∈ SG,E, notice that the domain Dom(θ(α,g,β)) of the partial homeomorphism of E∞ given by the action of (α, g, β) is Z(β). Now, given (α, g, β) ∈ SG,E and any open set U ⊆ Z(β), set (see Section 9) Θ((α, g, β), U ) = {[α, g, β; η] : η ∈ U }. According to [6: Proposition 4.18], for every (α, g, β) ∈ SG,E and every open set U ⊆ Z(β), Θ((α, g, β), U ) is a slice (in fact, they form a basis for the topology of Gtight(SG,E)). Then, we have the following result 11.4. Lemma. Sℓ(G, E) = hΘ((α, g, β), Z(β)) : (α, g, β) ∈ SG,Ei. Proof. By [33: Proposition 5.13(7)], C∗(Gtight(SG,E)) is generated by {1Θ((α,g,β),Z(β)) : (α, g, β) ∈ SG,E}. Thus, the result holds by (11.3.1). (cid:3) The next result is the key point for proving the injectivity of u : G → OG,E. 11.5. Lemma. If (G, E, ϕ) is pseudo free, then the map ρ : SG,E → Sℓ(G, E) (α, g, β) 7→ Θ((α, g, β), Z(β)) is a ∗-semigroup isomorphism. Proof. The surjectivity of ρ derives from (11.4). Now, let (α, g, β), (γ, h, η) ∈ SG,E such that Θ((α, g, β), Z(β)) = Θ((γ, h, η), Z(η)). Then, for any ω ∈ Z(β) we have [α, g, β; ω] = [γ, h, η; ω]. Since (G, E, ϕ) is pseudo free, by (8.6) there exists τ ∈ E∗ such that γ = α · gτ , η = βτ and h = ϕ(g, τ ). If whence ρ is injective, as desired. (cid:3) 1Θ((x,g,g−1x),Z(g−1x)) for every g ∈ G. 11.6. Proposition. There exists a ∗-isomorphism φ : OG,E → C∗(Gtight(SG,E)) such that φ(px) = 1Θ((x,1,x),Z(x)) for every x ∈ E0, φ(sa) = 1Θ((a,1,d(a)),Z(a)) for every a ∈ E1, (α, g, β) 6= (γ, h, η), then we can pick δ 6= τ in E∗ and bω = βδeω ∈ Z(β). Thus, bω ∈ Z(η) but [(γ, h, η),bω] is not defined, contradicting the hypothesis. Hence, (α, g, β) = (γ, h, η), and φ(ug) = Px∈E0 Proof. Notice that ug = Px∈E0 ∪(Xx∈E0 ugpx. Then it is direct but tedious to check that (cid:8)1Θ((x,1,x),Z(x)) : x ∈ E0(cid:9) ∪(cid:8)1Θ((a,1,d(a)),Z(a)) : a ∈ E1(cid:9) ∪ 1Θ((x,g,g−1x),Z(g−1x)) : g ∈ G) satisfy the defining relations for OG,E. Thus, by the Universal Property of OG,E, the map ϕ is an ∗-homomorphism. Notice that ϕ is the homomorphism given in [6: Theorem 13.3], and so is injective. Surjectivity is due to [33: Proposition 5.13(7)]. (cid:3) inverse semigroup actions and self-similar graphs 47 11.7. Corollary. If (G, E, ϕ) is pseudo free, then π : SG,E → OG,E is injective. Proof. The composition map SG,E → OG,E → C∗(Gtight(SG,E)) 7→ 1Θ((α,g,β),Z(β)) 7→ sαugs∗ β (α, g, β) is injective by (11.4) and (11.5). By (11.6), OG,E → C∗(Gtight(SG,E)) is injective. Thus, π is injective. Hence, we conclude 11.8. Proposition. If (G, E, ϕ) is pseudo free, then u : G → OG,E is injective. Proof. By Corollary (11.7), π is injective. Then, so is u by (11.2). (cid:3) (cid:3) 11.9. Remark. In particular, (11.8) implies that, if (G, E, ϕ) is pseudo free, then SG,E is ∗-isomorphic to the inverse semigroup of OG,E generated by {sa : a ∈ E1} ∪ {pxug : x ∈ E0, g ∈ G}. Proposition (11.8) provides the best situation possible, as the next example shows: 11.10. Example. Let E be the graph with only one vertex and one edge, and let G be any noncommutative group. Fix the trivial action of G on E, and let ϕ be the one- cocycle of G defined by ϕ(g, a) = 1 for every g ∈ G, a ∈ E1. Then, it is easy to see that OG,E ∼= C∗(E) ∼= C(T), which is a commutative C-algebra, so that it cannot contain any faithful copy of G. 12. The Hausdorff property for Gtight(SG,E). Again considering a triple (G, E, ϕ) satisfying (2.3), we will now give a characterization of the Hausdorff property for the tight groupoid of SG,E. The first result we may present in this direction is: 12.1. Proposition. If (G, E, ϕ) is pseudo free, then Gtight(SG,E) is a Hausdorff groupoid. Proof. If (G, E, ϕ) is pseudo free, then SG,E is E-unitary by (5.8), so Gtight(SG,E) is Haus- dorff by [12: Corollary 3.17]. This could also be obtained from [6: Propositions 6.4 and 6.2]. (cid:3) The converse of the above result is not true: as we will see in Example (18.15), there are examples in which (G, E, ϕ) fails to be pseudo free but still Gtight(SG,E) is Hausdorff. This may be interpreted as saying that the above assumption that (G, E, ϕ) is pseudo free is a much too strong hypothesis which one would therefore like to relax. On the other hand, recall from (5.5) that the failure of pseudo freeness for (G, E, ϕ) is equivalent to the existence of strongly fixed paths for nontrivial group elements. The result below consists in allowing a limited amount of minimal strongly fixed paths, and hence a limited number of counter-examples for pseudo freeness, without harming Hausdorffnes. 48 r. exel and e. pardo 12.2. Theorem. Assuming that (G, E, ϕ) satisfies (2.3), the following are equivalent: (a) for every g in G, there are at most finitely many minimal strongly fixed paths for g, (b) Gtight(SG,E) is Hausdorff. Proof. We will of course use [12: Theorem 3.16]. So, given s in SG,E, we must provide a finite cover for Js. Since such a cover exists by trivial reasons when Js is empty, let us assume that s dominates at least one nonzero idempotent element. By (5.7) we then have that s necessarily has the form s = (α, g, α), and the set of nonzero idempotent elements dominated by s is given by Using (5.3) we may further describe Js as Js =(cid:8)(ατ, 1, ατ ) : τ ∈ E∗, d(α) = r(τ ), τ is strongly fixed by g(cid:9). Js =(cid:8)(αµγ, 1, αµγ) : µ ∈ Mg, γ ∈ E∗, d(α) = r(µ), d(µ) = r(γ)(cid:9), where Mg is the set of all minimal strongly fixed paths for g. Assuming (a), we have that Mg is finite and then it is clear that (12.2.1) (cid:8)(αµ, 1, αµ) : µ ∈ Mg, d(α) = r(µ)(cid:9) is a finite cover for Js, whence Gtight(SG,E) is Hausdorff by [12: Theorem 3.16]. Assuming (b), let g ∈ G, and for each vertex x in E0, denote by M x g the set of all minimal strongly fixed paths for g whose range coincides with x. Since E is finite, in order to prove that Mg is finite, it is enough to check that each M x g is finite. If M x g is empty, there is nothing to do, so let us assume the contrary. Given any µ in M x g , we then have that x = r(µ) = r(gµ) = gr(µ) = gx, so x is fixed by g. Consequently s := (x, g, x) lies in SG,E and, assuming (b), we have by [12: Theorem 3.16] that Js admits a finite cover which, in view of (12.2.1), must necessarily be of the form (cid:8)(µiγi, 1, µiγi)}n i=1, where the µi ∈ M x apearing above exhaust M x g , meaning that g , and the γi are paths with d(µi) = r(γi). We then claim that the µi M x g = {µi : i = 1, . . . , n}. (12.2.2) To see this, let µ ∈ M x g , so that (µ, 1, µ) ≤ s, by (5.7), and hence (µ, 1, µ) ∈ Js. For some i, one would then have that (µiγi, 1, µiγi)(µ, 1, µ) 6= 0, in which case either µ is a prefix of µiγi, or vice versa. This implies that either µ is a prefix of µi, or vice versa, but since both µ and µi are minimal, we must have µ = µi, proving (12.2.2), and hence that M x g is finite. Consequently Mg, which decomposes as the disjoint union of the M x (cid:3) g , is also finite. This verifies (a) and hence concludes the proof. inverse semigroup actions and self-similar graphs 49 13. Minimality for Gtight(SG,E). In this section we will study conditions under which Gtight(SG,E) is minimal. Some of the results we obtain here are analog to those proved in [9] for the case of partial actions of groups. Given a triple (G, E, ϕ) satisfying (2.3), there are two relations among vertices in E0 which are relevant for the question at hand. First of all let us say that x ⇀ y provided there exists a path α in E∗ such that d(α) = x and r(α) = y. Notice that this relation is reflexive (take α to be x) and transitive (take the concatenation of the relevant paths). However this is neither symmetric nor antisymmetric, hence it is not an equivalence relation nor an order relation. The other relation we have in mind is simply the orbit relation, defined by x ∼ y when there exists g in G such that gx = y. Unlike "⇀", it is well known that "∼" is an equivalence relation. We may then consider the smallest transitive relation extending both "⇀" and "∼", by saying that vertices x and y are related when one may find a sequence of vertices x0, x1, . . . , x2n such that x = x0 ⇀ x1 ∼ x2 ⇀ x3 ∼ . . . ∼ x2n−2 ⇀ x2n−1 ∼ x2n = y. (13.1) The situation is in fact not so complicated due to the following: 13.2. Proposition. Let x and y be vertices in E0. Then the following are equivalent; (i) there exists a vertex u such that x ⇀ u ∼ y, (ii) there exists a vertex v such that x ∼ v ⇀ y. Proof. The fact that x ⇀ u ∼ y means that there exists a path α in E∗ such that d(α) = x, and r(α) = u, and there exists some g in G such that gu = y. Considering the path β = gα, and the vertex v = gx, notice that while d(β) = d(gα) = gd(α) = gx = v, r(β) = r(gα) = gr(α) = gu = y, so x ∼ v ⇀ y. Conversely, assuming (ii) we have that gx = v = d(β) and r(β) = y, for suitable g in G and β in E∗. Defining u = g−1y, and α = g−1β, we then have that and so x ⇀ u ∼ y. d(α) = g−1d(β) = x, r(α) = g−1r(β) = g−1y = u, (cid:3) 50 r. exel and e. pardo 13.3. Definition. Given x and y in E0, we will say that x ≫ y if the equivalent conditions of (13.2) are satisfied. Observe that "≫" coincides with the relation defined in (13.1), thanks to (13.2), and It is also evident that "≫" is reflexive but, again, it is hence it is clearly transitive. neither symmetric nor antisymmetric. Nevertheless we will view it as a defective order relation, in the sense that it satisfies all of the postulates of a (partial) order relation but for antisymmetry. Anytime we have such a defective order relation, it is possible to turn it into a bona fide partial order by identifying elements whenever antisymmetry fails. By this we mean that two vertices x and y in E0 will be called equivalent, in symbols whenever x ≫ y and y ≫ x. Writing [x] for the equivalence class of each x in E0, the set of all equivalent classes, namely x ≈ y becomes a partially ordered set via the well defined order relation E0 ≈ =(cid:8)[x] : x ∈ E0(cid:9) [x] ≥ [y] ⇐⇒ x ≫ y. 13.4. Definition. Under the assumptions of (2.3), we will say that: (i) E is G-transitive if, for any two vertices x and y in E0, one has that x ≫ y, (ii) E is weakly G-transitive if, given any infinite path ξ, and any vertex x in E0, there is some vertex v along ξ such that v ≫ x. The notion of G-transitivity generalizes the well known notion of transitivity in graphs. When it holds, E0 has a single equivalence class. On the other hand, weak G-transitivity is inspired by the notion of cofinality intro- duced in [20: Section 3], (see also [8: Definition 37.16]). The reader is however warned that the notions of weak G-transitivity and cofinality may only be reconciled upon a reversal of the direction of the edges in E1, following the new trend in graph algebras started by Katsura (see the penultimate paragraph of the introduction in [17]). It is evident that every G-transitive graph is weakly G-transitive, but these are some- times equivalent notions as we will now show: 13.5. Proposition. In addition to the assumptions in (2.3), suppose that E has no sinks, meaning that d−1(x) is nonempty for every x in E0. Then, if E is weakly G-transitive, it must also be G-transitive. inverse semigroup actions and self-similar graphs 51 Proof. Since E0 is finite, we may choose a minimal element [x] in E0/ ≈. Using that E has no sinks, we may find an infinite sequence of edges . . . , α−i−1, α−i, . . . , α−2, α−1, α0 ∈ E1, such that d(α0) = x, and d(αi−1) = r(αi), for every i ≤ 0. Since E1 is also finite, there must be repetitions among the αi, say αm = αn, where m < n ≤ 0. The finite path γ = αmαm+1 . . . αn−1, therefore satisfies d(γ) = d(αn−1) = r(αn) = r(αm) = r(γ), and hence γ may be concatenated with itself infinitely many times producing the infinite path ξ = γγγ . . . Given any y in E0, and assuming weak G-transitivity, there is some vertex v along ξ, such that v ≫ y. Since ξ is made of repetitions of γ, one has that v = r(αk), for some k in the integer interval [m, n]. We then have x = d(α0) = d(αk . . . α−2α−1α0) ⇀ r(αk) = v ≫ y, so x ≫ y, but since [x] is minimal, we deduce that [x] = [y], which is to say that x ≈ y. The conclusion is that E0/≈ is a singleton, from where G-transitivity follows. (cid:3) Of course the above result has taken advantage of the fact that E is a finite graph in an essential way, so nothing like this is to be expected for infinite graphs. Regardless of the absence of sinks, we have: 13.6. Theorem. Given (G, E, ϕ) satisfying (2.3), one has that the following are equiva- lent: (i) the standard action of SG,E on E∞ defined in (8.3) is irreducible, (ii) Gtight(SG,E) is minimal, (iii) E is weakly G-transitive. Proof. The equivalence between (i) and (ii) is a consequence of [12: Proposition 5.4]. We will next show that the above condition (iii) is equivalent to condition (iii) of [12: Theorem 5.5], from where the result will follow. In doing so, it is useful to understand how do idempotents in E behave under conjugation by elements in SG,E, and we leave it for the reader to verify that, given (α, g, β) in SG,E and (γ, 1, γ) ∈ E, one has that (α, g, β)(γ, 1, γ)(α, g, β)∗ = (αgε, 1, αgε), if γ = βε, (α, 1, α), if γε = β, (13.6.1) 0, otherwise . 52 r. exel and e. pardo (iii)⇒[12: Theorem 5.5.iii]: Given any two nonzero idempotent elements in E, necessarily of the form fα = (α, 1, α), and fβ = (β, 1, β), to employ the notation introduced in (4.5), we must find an outer cover of fα (in the sense of [12: Definition 2.9]) formed by a finite number of conjugates of fβ. As a first step, notice that s :=(cid:0)d(β), 1, β(cid:1) lies in SG,E and sfβs∗ =(cid:0)d(β), 1, β(cid:1) (β, 1, β)(cid:0)β, 1, d(β)(cid:1) =(cid:0)d(β), 1, d(β)(cid:1) = fd(β). Thus, anything that may be obtained by conjugating fd(β) by an element t ∈ SG,E, may also be obtained by conjugating fβ by ts. It therefore suffices to find an outer cover of fα formed by conjugates of fd(β). On the other hand, observe that fα ≤ fr(α), so any outer cover of fr(α) is necessarily also an outer cover of fα. This said we see that we may assume, without loss of generality, that α and β are vertices. Our task thus gets simplifyed in the sense that we now need to find an outer cover of fx made of conjugates of fy, for any given vertices x and y in E0. Recall from (8.2) that the set of all infinite paths with a given prefix γ is denoted Z(γ). In case we take γ = x, we then have that Z(x) is the set of all infinite paths with range x. Thanks to weak G-transitivity, for each ξ in Z(x), we may choose a vertex vξ along ξ such that vξ ≫ y. This is to say that we may write ξ = αξηξ, where αξ is a finite path, ηξ is an infinite path and d(αξ) = vξ. uξ ................................................................................................................................................................................................................................................... • • y • ................................................................................................................................................................................................................................................... vξ • x .............. ................ gξ .............. ................ ηξ .............................................................................. ............................................................................................................................................................................................................................................................ {z ξ βξ ................ .............. ................ .............. αξ } (cid:8)Z(αξ)(cid:9)ξ∈Z(x) Z(x) ⊆ Sξ∈F Z(αξ), The fact that vξ ≫ y may be expressed by saying that vξ ∼ uξ ⇀ y, for some vertex uξ, so there exists gξ in G, and a finite path βξ, such that gξuξ = vξ, d(βξ) = uξ, and r(βξ) = y. Speaking of the cylinders Z(αξ), it is obvious that ξ ∈ Z(αξ), so we see that the collection of cylinders is an open cover (in the topological sense of the word) for Z(x). Since Z(x) is compact, we may extract a finite subcover, say (13.6.2) inverse semigroup actions and self-similar graphs 53 where F is a finite subset of Z(x). We next claim that {fαξ}ξ∈F is an outer cover6 of fx. To see this, let e be a nonzero idempotent in E, with e ≤ fx. Then e is necessarily given by e = (γ, 1, γ), for some finite path γ such that r(γ) = x. Using our standing hypothesis (2.3) according to which E has no sources, we may prolong γ to an infinite path η, which will then share ranges with γ, whence η ∈ Z(x). By (13.6.2) we then have that η lies in Z(αξ), for some ξ ∈ F . This implies that both αξ and γ are prefixes of η, from where it is easy to see that αξ ⋓ fγ, proving our claim. is a prefix of γ or vice-versa. In particular we conclude that fαξ Incidentally this could also be obtained from [12: Proposition 3.8]. We next claim that each fαξ is a conjugate of fy. To see this, observe that, since gξd(βξ) = gξuξ = vξ = d(αξ), one has that sξ := (αξ, gξ, βξ) lies in SG,E, and sξfys∗ ξ = (αξ, gξ, βξ)(y, 1, y)(βξ, gξ, αξ) = (αξ, 1, αξ) = fαξ . This concludes the proof of condition (iii) of [12: Theorem 5.5]. [12: Theorem 5.5.iii]⇒(iii): Given any infinite path ξ, and any vertex y in E0, we must show that there is some vertex v along ξ such that v ≫ y. Letting x = r(ξ), let us use the hypothesis regarding the nonzero idempotents fx = (x, 1, x), and fy = (y, 1, y). This is to say that there are s1, s2, . . . , sn in SG,E, such that {sifys∗ cover for fx. For each i, write si = (αi, gi, βi), so that i }1≤i≤n is an outer sifys∗ i = (αi, gi, βi)(y, 1, y)(βi, gi, αi). Observe that, unless r(βi) = y, the element displayed above vanishes, so it cannot possibly have any use as a member of a cover. We may then safely discard it, being left only with those βi such that that r(βi) = y. In this case, by the second clause in (13.6.1) we have sifys∗ i = (αi, 1, αi) = fαi. Unless r(αi) = x, notice that fαi ⊥ fx, in which case fαi may again be discarded as it plays no role in an outer cover for fx. We may therefore suppose, without loss of generality that r(αi) = x, for all i. ..............gi ................ .............................................................................. βi ................ .............. ................ .............. αi ................................................................................................................................................................................................................................................... • • y • ................................................................................................................................................................................................................................................... • x 6 This is in fact a cover but we do not need to worry about this right now. 54 r. exel and e. pardo Given that (αi, g, βi) lies in SG,E, we necessarily have that d(αi) = gid(βi). Recalling that the infinite path ξ, chosen at the beginning of the present argument, has range x, we claim that ξ is necessarily of the form ξ = αiξ′, for some i and some infinite path ξ′. To see this, write ξ = δξ′′, where ξ′′ is an infinite path and δ is a finite path whose length exceeds the length of all of the αi. Observing that r(δ) = x, we then have that fδ := (δ, 1, δ) ≤ (x, 1, x) = fx. So, by the covering property we must have fδ ⋓fαi, for some i, which implies that either δ is a prefix of αi or vice versa. However, due to the fact that δ > αi, by construction, the first alternative cannot hold, meaning that αi is a prefix of δ, and hence also of ξ, proving the claim. It follows that d(αi) is a vertex along ξ, and it is clear from the above diagram that (cid:3) d(αi) ≫ y. This concludes the proof. Combining the above result with (13.5), we immediately deduce: 13.7. Corollary. If, in addition to the assumptions of (13.6) we have that E has no sinks, then conditions (13.6.i -- iii) are also equivalent to: (iv) E is G-transitive. It is interesting to observe that G-transitivity, when it holds, is the result of a joint effort by the action of G and the edges, both of which may be seen as pushing vertices around. However, sometimes only one of the players bear the responsibility to do the pushing around: (1) If G acts transitively on E0, then E is G-transitive regardless of the graph. Easy examples of this situation may be built on a graph formed by a disjoint union of loops, for instance. (2) If G fixes all vertices, then E is (weakly) G-transitive if and only if E is (weakly) transitive [8: Definition 37.16]. This is the case of Katsura algebras, when seen in the present framework. 14. Essentially principal groupoids. In this section we will discuss conditions under which Gtight(SG,E) is an essentially prin- cipal groupoid, a condition which is intimately tied to the action of SG,E on E∞ being topologically free. The reader is referred to [12: Section 4] for the definition of the notion of topologically free actions of inverse semigroups, as well as some of the main tools we shall use here. inverse semigroup actions and self-similar graphs 55 14.1. Definition. (1) A circuit 7 is a finite path γ ∈ E∗ of nonzero length such that d(γ) = r(γ). (2) A G-circuit is a pair (g, γ), where g ∈ G, and γ ∈ E∗ is a finite path of nonzero length such that d(γ) = gr(γ). γ5 • ................................................................................ .................................................................... • ......................................................................... γ6 γ4 ............ ................................................................................. ........................................................ ..................................................................................... ............ • γ3 ............................................................ g • ...................... .................... • ................................................................................................................. ............ ............................................................................. ......................................................................................................... γ1 • ...................................................................................... ............ ............................................................. • γ2 Thus, a G-circuit needs a little help from the group to close it up. Notice that a (usual) circuit γ may be concatenated with itself infinitely many times producing an infinite path A G-circuit. ξ = γγγ . . . Moreover, if s =(cid:0)γ, 1, d(γ)(cid:1), then, regarding the standard action of SG,E on E∞ defined in (8.3), it is easy to see that sξ = ξ, which is to say that ξ is a fixed point for s. It is also possible to create fixed points from G-circuits as follows: 14.2. Proposition. Given a G-circuit (g, γ), define a sequence {γ n}n≥1 of finite paths, and a sequence {gn}n≥1 of group elements, recursively by γ1 = γ, g1 = g, and ( γ n+1 = gnγ n gn+1 = ϕ(gn, γ n), for all n ≥ 1. Then (i) d(γ n) = r(γ n+1), for all n ≥ 1, (ii) the concatenation ξ = γ1γ2γ3 . . . (iii) for every finite path β such that d(β) = r(γ), one has that s := (βγ, g, β) lies in SG,E, is a well defined infinite path, and βξ is a fixed point for s. 7 Circuits are also called loops or cycles in the graph C-algebra literature. Our preference for circuits comes from the fact that it is the terminology of choice among graph theorists and also because in the established graph theory terminology the word loop refers to a single edge whose source and range coincide. 56 r. exel and e. pardo Proof. In order to prove the case n = 1 of (i), we have d(γ1) = d(γ) = gr(γ) = r(gγ) = r(g1γ1) = r(γ2). For n ≥ 1, we have d(γ n+1) = d(gnγ n) = gnd(γ n) (2.5.vii) = ϕ(gn, γ n)d(γ n) = gn+1d(γ n) (⋆) = = gn+1r(γ n+1) = r(gn+1γ n+1) = r(γ n+2), where we have used induction in the step marked with (⋆) above. This proves (i), which in turn immediately implies (ii). In order to show that the element s defined in (iii) indeed lies in SG,E, it is enough to observe that gd(β) = gr(γ) = d(γ) = d(βγ). Before proving the last part of (iii), we claim that gn(γ nγ n+1 . . . γ n+k) = γ n+1γ n+2 . . . γ n+k+1, ∀ n ≥ 1, ∀ k ≥ 0. In case k = 0, this is true by the recursive definition above, and if k ≥ 1, we have gn(γ nγ n+1 . . . γ n+k) = (gnγ n)ϕ(gn, γ n)(γ n+1 . . . γ n+k) = = γ n+1gn+1(γ n+1 . . . γ n+k), and the claim then follows easily by induction. A useful consequence is that g1(γ1γ2 . . . γ n) = γ2γ3 . . . γ n+1, ∀ n ≥ 1, from where we further deduce that gξ = g1(γ1γ2γ3 . . .) = γ2γ3γ4 . . . (14.2.1) With this we may now tackle the final task: s(βξ) = (βγ, g, β)(βξ) (8.3) = βγgξ (14.2.1) = βγγ2γ3 . . . γ n+1 = βξ. (cid:3) The above method does not give us all fixed points of every single element s in SG,E, but in certain cases it does: 14.3. Proposition. Given s := (α, g, β) in SG,E, suppose that α > β. Then, regarding the standard action of SG,E on E∞, one has that: (i) s admits at most one fixed point, (ii) if s admits a fixed point ζ, then there is a G-circuit (g, γ) such that α = βγ, and ζ coincides with the fixed point βξ mentioned in (14.2.iii), constructed from (g, γ). inverse semigroup actions and self-similar graphs 57 Proof. Assuming that ζ is a fixed point for s, it must lie in Z(β), so necessarily ζ = βξ, for a suitable infinite path ξ. We then have βξ = ζ = sζ = (α, g, β)(βξ) = αgξ. (14.3.1) This imples that both α and β are prefixes of ζ, so one must be a prefix of the other, but since α > β, the only alternative is that β is a prefix of α. We may therefore write for some finite path γ, which necessarily satisfies α = βγ, In other words, (g, γ) is a G-circuit. From (14.3.1) we also deduce that gr(γ) = gd(β) = d(α) = d(γ). βξ = αgξ = βγgξ, so ξ = γgξ. Let us now write ξ = γ1γ2γ3 . . ., where each γ i is a finite path with γ i = γ. Then γ1γ2γ3 . . . = ξ = γgξ = γg(γ1γ2γ3 . . .) = γ(g1γ1)(g2γ2)(g3γ3) . . . , where the gi are recursively defined by g1 = g, and gn+1 = ϕ(gn, γ n). It then follows that γ1 = γ, and γ n+1 = gnγ n, for all n ≥ 1, so we see that the γ n and the gn are precisely defined as in (14.2). This concludes the proof. (cid:3) As already announced we will eventually be interested in determining conditions under which the standard action of SG,E on E∞ is topologically free, so the fixed points that will really interest us are the interior ones. Under the conditions of the above result, when there is at most one fixed point, the existence of interior fixed points hinges on whether or not the unique fixed point is isolated in E∞. We will now introduce certain concepts designed to study isolated fixed points. Recall from (2.3) that our graph E has no sources, meaning that r−1(x) is nonempty for every vertex x. 14.4. Definition. (1) We shall say that a vertex x in E0 is a simple vertex if r−1(x) is a singleton. (2) Given a path γ = γ1γ2 . . . γn in E∗, where each γi is in E1, we will say that γ has no entry if d(γi) is a simple vertex for every i = 1, . . . , n. (3) If the condition above fails, we will say that γ has an entry. Thus, if a path γ = γ1γ2 . . . γn has no entry, then r−1(cid:0)d(γi)(cid:1) is a singleton for every i, and we may obviously guess which is the edge forming this singleton, namely provided i < n. However the same cannot be said when i = n, unless (g, γ) is a G-circuit, in which case The notion of entryless paths will only be useful when applied to G-circuits. r−1(cid:0)d(γi)(cid:1) = {γi+1}, r−1(cid:0)d(γn)(cid:1) = {gγ1}. 58 r. exel and e. pardo 14.5. Proposition. Under the conditions of (14.3.ii), let (g, γ) be the G-circuit and ζ be the fixed point for s mentioned there. Then the following are equivalent: (i) ζ is an isolated point in E∞, (ii) γ has no entry. Proof. In case γ has no entry, writing γ = γ1γ2 . . . γn, where the γi are edges, notice that the only infinite path extending γ1 is the path ξ referred to in (14.2.ii). The fixed point ζ = βξ mentioned in (14.3.ii) is therefore the only infinite path extending βγ1, whence Z(βγ1) = {ζ}, which implies that ζ is isolated. Conversely, assuming that ζ is isolated, there exists a sufficiently long prefix ε of ζ, such that Z(ε) = {ζ}. This means that ζ is the only infinite path extending ε. Writing ζ = ζ1ζ2ζ3 . . . , where the ζi are edges, one then has that, for sufficiently large i, there is only one edge whose range is d(ζi). Letting {γ n}n≥1 and {gn}n≥1 be the sequences defined in (14.2), we then have that, for sufficiently large n, the G-circuit (gn, γ n) has no entry. Since G acts on E by graph automorphisms, we may easily prove by induction that all G-circuits (gk, γ k) have no entry, including (g1, γ1) = (g, γ). (cid:3) Since we are interested in topologically free actions, we would like to avoid isolated fixed points and hence we will be interested in situations when every G-circuit has an entry. However, given that we are working with finite graphs only, the action of G on E turns out not to be relevant in this respect. In precise terms, what we mean is that: 14.6. Proposition. Under the conditions of (2.3), the following are equivalent: (i) every G-circuit has an entry, (ii) every circuit has an entry. Proof. Since every circuit γ gives rise to the G-circuit (1, γ), it is evident that (i) implies (ii). Conversely, assume (ii) and let γ be a G-circuit. Leting {γ n}n≥1 and {gn}n≥1 be as in (14.2), consider the infinite path ξ = γ1γ2γ3 . . . mentioned in (14.2.iii). Notice that the γ i are all in the orbit of γ under the action of G, and hence the length of γi coincides with that of γ. As E is finite, there is only a finite number of paths of this length, so there must necessarily be repetitions among the γ i, say γ i = γ j, where i < j. Then so the path r(γ i+1) = d(γ i) = d(γ j), γ i+1γ i+2 . . . γ j is a circuit, which by hypothesis has an entry. It is now easy to see that some γ k must have an entry. Finally, since γ k is in the orbit of γ under the action of G, then γ likewise has an entry, concluding the proof. (cid:3) inverse semigroup actions and self-similar graphs 59 Observe that we have used the finiteness of E in a very strong way above. Thus, should our theory ever be extended to infinite graphs, one might have to distinguish between conditions (14.6.i) and (14.6.ii). We should point out that a graph in which every circuit has an entry is usually said to satisfy condition (L). The above results, mainly (14.3) and (14.5), may also be used to study the fixed points for elements s := (α, g, β) when α < β, since such fixed points are precisely the same as the fixed points of s∗, and s∗ clearly satisfies the hypothesis of (14.3). However we still have work to do in order to treat the remaining case α = β. 14.7. Proposition. Let s := (α, g, β) ∈ SG,E, whith α = β, and suppose that s admits a fixed point. Then (i) α = β, (ii) the fixed points of s in E∞ are precisely the elements of the form ζ = βξ, where ξ is an infinite path such that r(ξ) = d(β), and gξ = ξ. Proof. Left for the reader. (cid:3) The conclusion of the previous Proposition is that when α = β, understanding the fixed points for s requires understanding the fixed points for the action of g on E∞. One may easily describe such fixed points in terms of the action of G on E and the cocycle ϕ, but apparently there is no smart way to control each and every one of them. However, since our main interest is in studying topological freeness, we need only focus on large (meaning open) sets of fixed points: 14.8. Proposition. Suppose that s := (α, g, α) lies in SG,E, and that ζ is an interior fixed point for s. Then there is a finite path γ, such that: (i) gγ = γ, (ii) d(α) = r(γ), (iii) ζ ∈ Z(αγ), (iv) the group element h := ϕ(g, γ) pointwise fixes8 the cylinder Z(d(γ)). Conversely, if γ is any finite path satisfying (i), (ii) and (iv), then every ζ ∈ Z(αγ) is a (necessarily interior) fixed point for s. Proof. In particular ζ a fixed point for s so, by (14.7) we have that ζ = αξ, with gξ = ξ. Moreover there exists a neighborhood U of ζ consisting of fixed points for ζ. Since the cylinders form a basis for the topology of E∞, we may assume without loss of generality that U = Z(β), for some finite path β, which we may assume is as long as we wish, and our wish in this case is simply that β > α. Since ζ lies in Z(β), we have that β is a prefix of ζ, so we may write ζ = βη, for some infinite path η. We then have βη = ζ = αξ. 8 By this we mean that every point in Z(d(γ)) is fixed by h. 60 r. exel and e. pardo Given that β > α, this implies that α is a prefix of β, so we write β = αγ, for a suitable finite path γ, obviously satisfying (ii). Consequently ζ = βη = αγη ∈ Z(αγ), proving (iii). Given any infinite path µ ∈ Z(d(γ)), we may form the path αγµ, which necessarily lies in Z(αγ) = Z(β), and hence is fixed under s. Therefore αγµ = (α, g, α)(αγµ) = αg(γµ) = α (gγ)(cid:0)ϕ(g, γ)µ(cid:1) = α (gγ) (hµ), whence γ = gγ, proving (i), and µ = hµ, in turn proving (iv). In order to prove the last sentence in the statement it is enough to notice that any el- ement in Z(αγ) is necessarily of the form αγµ, where µ ∈ Z(d(γ)), and the last calculation displayed above could be used to check that αγµ is fixed under s. (cid:3) Searching for conditions under which the standard action of SG,E on E∞ is topo- logically free, one should probably worry about group elements fixing whole cylinders, as in (14.8.iv). The following notion is designed to pinpoint situations under which whole cylinders of the form Z(x) are in fact fixed. 14.9. Definition. Given g ∈ G, and x ∈ E0, we shall say that g is slack at x, if there is a non-negative integer n such that all finite paths γ with r(γ) = x, and γ ≥ n, are strongly fixed by g, as defined in (5.1). As already discussed at the begining of section (5), if γ is strongly fixed by g, then g fixes any finite path extending γ, and hence also all infinite paths in Z(γ). If g is slack at x, and if n is as in (14.9), notice that Z(x) = Sr(γ)=x γ=n Z(γ), and since each γ occuring above is strongly fixed by g, we have that g pointwise fixes Z(γ), and hence also the whole cylinder Z(x). Notice that a path of length zero, namely a vertex x, is never strongly fixed by a nontrivial group element g, because ϕ(g, x) (2.5.ii) = g 6= 1. The concept of slackness above should therefore be seen as the best replacement for the notion of being strongly fixed in case of a vertex. We are now ready for a main result: 14.10. Theorem. Under the conditions of (2.3), the standard action of SG,E on E∞ is topologically free if and only if the following two conditions hold: (i) every G-circuit has an entry 9, (ii) given a vertex x, and a group element g fixing every infinite path in Z(x), then necessarily g is slack at x. 9 Recall that this is the same as saying that every circuit has an entry by (14.6). inverse semigroup actions and self-similar graphs 61 Proof. Suppose (i) and (ii) hold and let ζ be an interior fixed point for some s = (α, g, β) in SG,E. In order to prove topological freeness, we need to prove that ζ is a trivial fixed point for s. Case 1: Let us first assume that α = β. Letting γ and h as in (14.8), we then have that h pointwise fixes the cylinder Z(d(γ)). By (ii) we then conclude that h is slack at d(γ), so there is n such that every finite path of length n and range d(γ) is strongly fixed by h. By (14.8.iii) we have that ζ lies in Z(αγ), so we may write ζ = αγξ, for some infinite path ξ with d(γ) = r(ξ). Denoting by ε the path formed by the first n edges of ξ, we then have that so ε is strongly fixed by h, and we may further write r(ε) = r(ξ) = d(γ), ζ = αγεξ′, for a suitable infinite path ξ′. idempotent If follows that ζ ∈ Z(αγε), which is the domain of the fαγε = (αγε, 1, αγε). In addition sfαγε = (α, g, α)(αγε, 1, αγε) =(cid:0)αg(γε), ϕ(g, γε), αγε(cid:1), (14.10.1) and we claim that the element at the end of the above calculation coincides with fαγε. To see this notice that while g(γε) = (gγ)(cid:0)ϕ(g, γ)ε(cid:1) (14.8.i) = γhε = γε, ϕ(g, γε) (2.5.x) = ϕ(cid:0)ϕ(g, γ), ε(cid:1) = ϕ(cid:0)h, ε(cid:1) = 1. Plugging the last two identities at the end of (14.10.1) leads to sfαγε = fαγε, thus proving that ζ is a trivial fixed point, as needed. Case 2: Let us now assume that α > β. By (14.3) we have that ζ is the only fixed point for s, necessarily given in terms of a G-circuit (g, γ), as in (14.3.iii). Being the unique fixed point, as well as an interior member of the set of fixed points, we see that ζ is isolated in E∞. So (g, γ) has no entry by (14.5), contradicting (i). This implies that in fact s has no interior fixed points, so there is nothing to do. Case 3: The last remaining alternative, namely when α < β, may be treated by simply observing that the fixed points for s are the same as the fixed points for s∗ = (β, g−1, α), and that s∗ fits the previous case studied, so there are no interior fixed points for s∗, either. This concludes the proof that (i) and (ii) imply topological freeness. In order to prove that topological freeness implies (i), assume the former and suppose by contradiction that a G-circuit (g, γ) exists with no entry. Let x = r(γ), and notice that gx = gr(γ) = d(γ), 62 r. exel and e. pardo so the triple s := (γ, g, x) is seen to lie in SG,E. We may then use (14.2) to obtain a fixed point ζ for s, and by (14.5) we have that ζ is an isolated point of E∞, hence also an interior fixed point. Working under the assumption of topological freeness, we deduce that ζ is a trivial fixed point, which is to say that there is an idempotent e in E, whose domain contains ζ, and such that se = e. Observing that e cannot possibly be zero, we deduce that e = (ε, 1, ε), for some finite path ε. We then have that (ε, 1, ε) = e = se = (γ, g, x)(ε, 1, ε) =(cid:0)γgε, ϕ(g, ε), ε(cid:1). In particular this implies that ε = γgε, so ε = γgε = γ + gε = γ + ε, whence γ = 0, contradicting the fact that G-circuits have nonzero length by definition. This shows that there are no G-circuit without an entry, hence proving (i). We next show that topological freeness implies (ii). So we suppose that some g in G pointwise fixes a whole cylinder Z(x), where x is a vertex. In particuler we have that gx = x, so the element s := (x, g, x) belongs to SG,E, and it clearly also fixes every point in Z(x). Each ζ in Z(x) is therefore an interior fixed point for s, hence necessarily a trivial one by hypothesis. This means that there exists an idempotent e = (γ, 1, γ) ∈ E, such that ζ lies in the domain of e, also known as Z(γ), and moreover se = e. Therefore (γ, 1, γ) = e = se = (x, g, x)(γ, 1, γ) =(cid:0)xgγ, ϕ(g, γ), γ(cid:1), from where we deduce that gγ = γ, and ϕ(g, γ) = 1, which is to say that γ is strongly fixed by g. Given that ζ ∈ Z(γ), we have that γ is a prefix of ζ, whence r(γ) = r(ξ) = x, so ζ ∈ Z(γ) ⊆ Z(x). We then deduce that Z(x) is the union of the Z(γ), where γ range in the set of all finite paths strongly fixed by g, with r(γ) = x. By compactness we may find a finite collection of such finite paths, say γ1, γ2, . . . , γk, such that Z(x) = Z(γi). (14.10.2) We next wish to argue that the above γi's may be taken so that their length is constant. To see this let n be the length of the longer γi, and observe that, for each i, one has that kSi=1 Z(γi) = Sr(ε)=d(γi) ε=n−γi Z(γiε). inverse semigroup actions and self-similar graphs 63 Moreover, each γiε occuring above is also strongly fixed by g, as seen in the discussion near the beginning of section (5). Thus, if we replace each γi by the set of all γiε, where ε is as above, all of the properties so far mentioned of the original γi's will be preserved, and now γiε = γi + ε = n. Therefore we may and will assume, from now on, that the γi have a constant length, say n. From (14.10.2) it is now easy to conclude that the γi exhaust the set of all finite paths with range x and length n. In fact, if α is such a path, we may extend it to an infinite path of the form ξ = αη. Since ξ ∈ Z(x), then ξ ∈ Z(γi), for some i, whence γi is a prefix of ξ and, by considering lengths, we see that α = γi. The conclusion is that every finite path with length n and range x is strongly fixed (cid:3) by g, which is to say that g is slack at x. 14.11. Remark. If for any g ∈ G \ {1} and for any x ∈ E0 there exists η ∈ Z(x) such that gη 6= η, then (14.10.ii) holds trivially. This fact will be used in (18.9) and subsequent examples. 14.12. Remark. Regarding [12: Theorem 4.10.ii], and letting γi be as in (14.10.2), one may show that {fγi}i is a cover of fx consisting of idempotents fixed under s (in the sense of [12: Definition 4.8.1]). In case (G, E, ϕ) is pseudo free, and if g is a nontrivial group element, then g admits no strongly fixed paths by (5.5), so g will never be slack at any vertex. Condition (14.10.ii) can therefore only be satisfied if no nontrivial group element pointwise fixes a cylinder Z(x), and hence we have the following immediate consequence of (14.10): 14.13. Corollary. In addition to the conditions of (2.3), suppose that (G, E, ϕ) is pseudo free. Then the standard action of SG,E on E∞ is topologically free if and only if the following two conditions hold: (i) every G-circuit has an entry (which is the same as saying that every circuit has an entry by (14.6)), (ii) for every g in G, with g 6= 1, and for every x in E0, there is at least one ζ in Z(x) such that gζ 6= ζ. An important case for the theory of self-similar groups is when G acts faithfully10 on E∞, and E is a graph with a single vertex. 14.14. Corollary. Under the conditions of (2.3), suppose moreover that: (a) E has a single vertex, and at least two edges, (b) G acts faithfully on E∞. Then the standard action of SG,E on E∞ is topologically free. 10 Meaning that if gξ = ξ, for all ξ in E∞, then g = 1. 64 r. exel and e. pardo Proof. In the present situation the conditions of (14.10) become trivially true because: (i) all path are circuits and all circuits have entries, and (ii) there is only one Z(x) to consider, namely the whole space E∞, and by faithfulness no nontrivial group element acts trivially on E∞. (cid:3) As the title of the present section suggests, our main interest is in determining condi- tions for Gtight(SG,E) to be an essentially principal groupoid. Having understood topolog- ical freeness, an immediate consequence of [12: Theorem 4.7] is: 14.15. Corollary. Under the assumptions of (2.3), one has that Gtight(SG,E) is essentially principal if and only if (14.10.i&ii) hold. Two other similar results could be stated giving conditions for Gtight(SG,E) to be essentially principal, by combining [12: Theorem 4.7] with either (14.13) or (14.14), but we will refrain from doing it here since the reader can easily guess them. 15. Local contractivity for SG,E. In [12: Section 6] local contractivity for groupoids and for actions of inverse semigroups is studied. We will now use these results to characterize local contractivity for the tight groupoid associated to an inverse semigroup S. 15.1. Theorem. Under the conditions of (2.3), one has that the following are equivalent: (i) SG,E is a locally contracting inverse semigroup, (ii) the standard action θ : SG,E y E∞ is locally contracting, (iii) Gtight(SG,E) is a locally contracting groupoid, (iv) every circuit in E has an entry. Proof. As already mentioned in section (8), every tight filter in E is an ultra-filter, so the equivalence between (i) and (ii) follows from [12: Theorem 6.5]. (ii)⇒(iii): Follows immediately from [12: Proposition 6.3]. (iii)⇒(iv): We will prove this by contraposition, that is, assuming the existence of a circuit γ without an entry, we will show that Gtight(SG,E) is not locally contracting. Our task is actually very easy. Given an entryless circuit γ, the path ξ = γγγ . . . is an isolated point, whence U := {ξ} is an open subset of E∞. Viewing the latter as the unit space of Gtight(SG,E), as usual, and plugging U into [12: Definition 6.1], clearly there can be no open set V , and bissection S, as mentioned there, simply because a chain of nonempty subsets SV S−1 $ V ⊆ U cannot possibly exist withing a singleton such as U . This shows that Gtight(SG,E) is not locally contracting, as desired. (iv)⇒(i): Assuming that every circuit has an entry, we will show local contractivity of SG,E via [12: Proposition 6.7]. Given a nonzero idempotent e in E, write e = (µ, 1, µ) for some finite path µ. Using that E has no sources, we may find an infinite path ξ = ξ1ξ2ξ3 . . ., inverse semigroup actions and self-similar graphs 65 such that d(µ) = r(ξ). Since E is a finite graph, there must be repetitions amongst the ξi, say ξi = ξj, for some i < j. Letting α = ξ1ξ2 . . . ξi, and γ = ξi+1ξi+2 . . . ξj, notice that d(γ) = d(ξj) = d(ξi) = r(ξi+1) = r(γ), so γ is a circuit. It is also clear that µα and µαγ are well defined paths. Noticing that d(α) = r(γ) = d(γ), we have that s := (µαγ, 1, µα) lies in SG,E. Moreover, setting and using the notation introduced in (4.5), we have β1 = µα, sfβ1 s∗ = (µαγ, 1, µα)(µα, 1, µα)(µαγ, 1, µα)∗ (13.6.1) = (µαγ, 1, µαγ) ≤ fβ1 , thus verifying [12: Proposition 6.7.ii]. By hypothesis γ has an entry, so we may find a path γ′, with r(γ′) = r(γ), which is not a prefix of γ, or vice versa. Setting we then have β0 = µαγ′, 0 6= fµαγ′ ≤ fµα ≤ fµ ⇒ 0 6= fβ0 ≤ fβ1 = s∗s ≤ e, verifying [12: Proposition 6.7.i]. Focusing now on [12: Proposition 6.7.iii] notice that fβ0 s = (µαγ′, 1, µαγ′)(µαγ, 1, µα) = 0, precisely because γ and γ′ are not each other's prefix. So evidently fβ0 sfβ1 = 0, proving the last condition in [12: Proposition 6.7], and hence that SG,E is locally contracting, thus proving (i). (cid:3) It is worth noticing that many results of [12] used in the above proof, such as [12: Proposition 6.3], [12: Theorem 6.5] and [12: Proposition 6.7], comparing local contractivity for groupoids, inverse semigroups, and actions, are either one way implications only, or the converse depends on special conditions. Nevertheless, the situation in which we are working has fortunately allowed for a downright equivalence of the various manifestations of contractivity. However, this result should be taken with a certain skepticism. First of all it is well known that the above condition on circuits is not sufficient for local contractivity for the groupoid associated to infinite graphs [21]. Considering that finite graphs are special cases of our theory (just take the acting group to be the trivial group), it is not unreasonable 66 r. exel and e. pardo to believe that our results admit natural generalizations to infinite graphs, but then a characterization of local contractivity for the corresponding groupoid will certainly not follow from the fact that every circuit has an entry, since this is false for infinite graphs, as mentioned above. Secondly, observe that the condition on the existence of entries for circuits completely ignores the group G, but, again, a generalization to infinite graphs will probably depend on the action. A hypothesis such as "every vertex connects to a G-circuit with an entry", to paraphrase the main hypothesis of [21: Lemma 3.8], is probably more realistic in the conjectured infinite graph scenario. 16. Simplicity and pure infiniteness for OG,E. In this section we use the results in the previous sections to characterize when OG,E is simple and purely infinite. The central results are the following: 16.1. Theorem. Assume that (G, E, ϕ) satisfies (2.3), that G is amenable, and that for every g ∈ G there are at most finitely many minimal strongly fixed paths for g. Then OG,E is simple if and only if the following conditions are satisfied: (a) E is weakly-G-transitive. (b) Every G-circuit has an entry. ( c ) Given a vertex x, and a group element g fixing Z(x) pointwise, then necessarily g is a slack at x. Proof. By (12.2), the groupoid Gtight(SG,E) is Hausdorff. Clearly Gtight(SG,E) is ´etale with second countable unit space. By (10.18), Gtight(SG,E) is amenable. Then, by (6.4), OG,E ∼= C∗(Gtight(SG,E)) = C∗ r (Gtight(SG,E)). By [4: Theorem 5.1], OG,E is simple if and only if Gtight(SG,E) is minimal and essentially principal. Since minimality of Gtight(SG,E) is equivalent to (a) by (13.6), and essential principality of Gtight(SG,E) is equivalent to (b&c) by (14.15), the result holds. (cid:3) With respect to pure infiniteness, we have: 16.2. Theorem. Let (G, E, ϕ) be under (2.3), and let G be an amenable group. If Gtight(SG,E) is essentially principal, then every hereditary subalgebra of OG,E contains an infinite projection. Proof. By the same argument as in (16.1), OG,E ∼= C∗(Gtight(SG,E)) = C∗ r (Gtight(SG,E)). By (14.10.i), every circuit of E has an entry. Thus, Gtight(SG,E) is locally contracting by (15.1). Hence, by [1: Proposition 2.4], every nonzero hereditary sub-C-algebra of OG,E contains an infinite projection, as desired. (cid:3) As an immediate consequence we have 16.3. Corollary. If (G, E, ϕ) satisfies (2.3), the group G is amenable, and Gtight(SG,E) is Hausdorff then, whenever OG,E is simple, it is necessarily also purely infinite (simple). Proof. By (14.10) and (16.1), Gtight(SG,E) is essentially principal. Thus, by (16.2), every nonzero hereditary sub-C-algebra of OG,E contains an infinite projection. Hence, OG,E is purely infinite simple, as desired. (cid:3) inverse semigroup actions and self-similar graphs 67 17. Revisiting Nekrashevych algebras. In this section we will analyze Nekrashevych algebras from our point of view. The Nekrashevych C-algebra O(G,X), associated to a self-similar action of a group G on a finite alphabet X [26], is a direct example of our definition (see (3.3)). Here, the graph E is the rose of n petals for n = X ≥ 2, so that the action on vertices is trivial, and the action is faithful. Since E0 = 1, we have the following facts: (1) E is G-transitive, whence Gtight(SG,E) is minimal by (13.6). (2) Gtight(SG,E) is essentially principal by (14.14) and [12: Theorem 14.7]. In particular, Gtight(SG,E) is locally contracting by (14.10) and (15.1). Thus, if Gtight(SG,E) is Hausdorff, we conclude that C∗ r (Gtight(SG,E)) is a purely infinite simple C-algebra by [12: Theorem 6.8]. Hausdorffness of Gtight(SG,E) is equivalent, ac- cording to (12.2), to the existence of at most finitely many minimal strongly fixed paths for every g ∈ G. In this sense, it is interesting to remark that Nekrashevych also gave a presentation of its algebra as a groupoid C-algebra associated to a groupoid of germs of an inverse semigroup S [26: Section 5]. While S turns out to be SG,E, the notion of germ that he used is the one adopted by Arzumanian and Renault [3], which differs from the one we used, due to Patterson [27: Page 140]. Luckily, both definitions coincide when the action of SG,E on E∞ is topologically free, which is the case of Nekrashevych triples, as we noticed above. So, Nekrashevych's groupoid and Gtight(SG,E) coincide, and the characterization of Hausdorffness we obtained in (12.2) coincide with that given by Nekrashevych [26: Lemma 5.4]. In order to obtain a characterization of (pure infinite) simplicity for O(G,X), we need to keep control of whether O(G,X) is nuclear. So, it only remains to determine when Gtight(SG,E) is amenable, which implies that C∗(Gtight(SG,E)) = C∗ r (Gtight(SG,E)). By (10.18), if G is an amenable group, then Gtight(SG,E) is an amenable groupoid. Thus, we obtain 17.1. Proposition. If (G, X, ϕ) is a Nekrashevych triple, with G an amenable group and Gtight(SG,E) a Hausdorff groupoid, then O(G,X) is a nuclear, separable, purely infinite simple C-algebra. Here, Nekrashevych's approach differs from ours. In [26] he stated a sufficient con- dition for the amenability of Gtight(SG,E), which apparently does not require the group G to be amenable. The condition relies on two concepts associated to self-similar groups: self-replication and contractiveness (see [25] or [26] for definitions of these concepts). Nekrashevych [26: Theorem 5.6] proved that if Gtight(SG,E) is Hausdorff and (G, X) is self-replicating and contractive, then Gtight(SG,E) is of polynomial growth [25], and thus it its amenable by [2: Proposition 3.2.32]. 18. Revisiting Katsura algebras. In this section we will analyze Katsura algebras from our point of view. We will quickly recall the definition and basic properties of Katsura algebras that will be needed in the sequel. This is borrowed from [18]. 68 r. exel and e. pardo 18.1. Definition. Let N ∈ ∪ {∞}, let A ∈ MN (Z matrices. Define a set ΩA by +) and B ∈ MN (Z) be row-finite ΩA := {(i, j) ∈ {1, 2, . . . , N } × {1, 2, . . . , N } : Ai,j ≥ 1}. For each i ∈ {1, 2, . . . , N }, define a set ΩA(i) ⊂ {1, 2, . . . , N } by ΩA(i) := {j ∈ {1, 2, . . . , N } : (i, j) ∈ ΩA}. Notice that, by definition, ΩA(i) is finite for all i. Finally, fix the following relation: (0) ΩA(i) 6= ∅ for all i, and Bi,j = 0 for (i, j) 6∈ ΩA. With these data we can define Katsura algebras 18.2. Definition. Define OA,B to be the universal C-algebra generated by mutually orthogonal projections {qi}N i ui = qi, and partial isometries {si,j,n}(i,j)∈ΩA,n∈Z satisfying the relations: (i) si,j,nuj = si,j,n+Ai,j and uisi,j,n = si,j,n+Bi,j for all (i, j) ∈ ΩA and n ∈ Z. (ii) s∗ i,j,nsi,j,n = qj for all (i, j) ∈ ΩA and n ∈ Z. i = u∗ i=1, partial unitaries {ui}N i=1 with uiu∗ (iii) qi = Pj∈ΩA(i) Ai,jPn=1 si,j,ns∗ i,j,n for all i. 18.3. Remark. Now, the following facts holds: (1) The C-algebra OA,B is separable, nuclear, in the UCT class [18: Proposition 2.9]. (2) If the matrices A, B satisfy the following additional properties: (a) A is irreducible, and (b) Aii ≥ 2 and Bi,i = 1 for every 1 ≤ i ≤ N , then the C-algebra OA,B is purely infinite simple, and hence a Kirchberg algebra [18: Proposition 2.10]. (3) The K-groups of OA,B are [18: Proposition 2.6]: (a) K0(OA,B) ∼= coker(I − A) ⊕ ker(I − B), and (b) K1(OA,B) ∼= coker(I − B) ⊕ ker(I − A). (4) Every Kirchberg algebra can be represented, up to isomorphism, by an algebra OA,B for matrices A, B satisfying the conditions in (18.3.2) [19: Proposition 4.5]. As we have seen in (3.4), unital Katsura algebras are natural examples of our con- struction. So, it is easy to use our results in order to characterize some properties, like simplicity or pure infinite simplicity, in terms of matrices A and B. This work has been previously done in [10], but the approach we chose there was fairly more direct and com- putational, so that the conditions appearing there were less elegant and clear than the ones we will present here. Across this section, we will say that a triple (Z, E, ϕ) is a Katsura triple if there exist finite matrices A, B satisfying (18.1) such that the triple associated to the algebra OA,B is inverse semigroup actions and self-similar graphs 69 (Z, E, ϕ); in particular, E is the graph whose adjacency matrix is A. Also, we will fix the following agreement: let ξ be either in E∗ or in E∞, i.e. ξ = ei1,i2,n1ei2,i3,n2 · · · eik,ik+1,nk or ξ = ei1,i2,n1 ei2,i3,n2 · · · eik ,ik+1,nk · · · , then, for any r ∈ we define Bξr := rYt=1 Bit,it+1 and Aξr := Ait,it+1 . rYt=1 The first step to work out the corresponding results to the ones we obtained for the general setting is to determine when a finite path α ∈ E∗ is fixed under the action of an element l ∈ Z. 18.4. Lemma. Let (Z, E, ϕ) be a Katsura triple. Given an element α of E∗ of length r and an integer l ∈ Z, the following are equivalent: (1) α is fixed under the action of l. (2) For every 1 ≤ j ≤ r the element Kj := l Bαj Aαj belongs to Z. Proof. Set α = ei1,i2,n1 ei2,i3,n2 · · · eir ,ir+1,nr . By definition of (Z, E, ϕ), α = lα if and only if there exists a sequence (Kj)j≥0 ⊆ Z such that: (i) K0 = l. (ii) For every 1 ≤ j ≤ r, nj−1 + Kj−1Bij ,ij+1 = nj−1 + KjAij ,ij+1 . Notice that (ii) is equivalent to ask Kj−1Bij ,ij+1 = KjAij ,ij+1 for every j ≥ 1. Now, for j = 1 we have K0Bi1,i2 = lBi1,i2 = K1Ai1,i2 , so that K1 = l suppose that for 1 ≤ t ≤ j − 1 we have proved that Kt := l Bαt Aαt . Hence Bi1,i2 Ai1,i2 . Now, KjAij ,ij+1 = Kj−1Bij ,ij+1 = l Bαj−1 Aαj−1 · Bij ,ij+1 , so that Kj = l Bαj Aαj . This completes the proof. (cid:3) Now, we are ready to characterize pseudo freeness for a Katsura triple (Z, E, ϕ). 18.5. Lemma. Let (Z, E, ϕ) be a Katsura triple. Then, the following are equivalent: (1) (Z, E, ϕ) is pseudo free. (2) Bi,j = 0 if and only if (i, j) 6∈ ΩA. 70 r. exel and e. pardo Proof. Let α = ei1,i2,n1 ei2,i3,n2 · · · eir ,ir+1,nr of E∗, and let l ∈ Z. By (18.4), lα = α exactly Bαj Aαj when the elements Kj := l belongs to Z for every 1 ≤ j ≤ r. Since ϕ(l, αj) = Kj, ϕ(l, α) = 0 exactly when Kj = 0 for some j ≤ r. Thus, the situation reduces to Kj−1eij ,ij+1,nl = eij ,ij+1,nl and Kj = 0 for some 1 ≤ j ≤ r, which corresponds to the equa- tion nj +Kj−1Bij ,ij+1 = nj. And this occurs exactly when Bij ,ij+1 = 0, so we are done. (cid:3) Which these results in mind, we are ready to characterize when Gtight(SZ,E) is Haus- dorff 18.6. Theorem. Let (Z, E, ϕ) be a Katsura triple. Then, the following are equivalent: (1) Gtight(SZ,E) is Hausdorff. (2) Whenever (i, j) ∈ ΩA with Bi,j = 0, then for any l ∈ Z there exist finitely many finite paths α ∈ E∗ with d(α) = i such that l Bαt Aαt ∈ Z for every 1 ≤ t ≤ r − 1. Proof. The result holds by (12.2) and (18.5). (cid:3) The next step is to determine the minimality of Gtight(SZ,E). 18.7. Theorem. Let (Z, E, ϕ) be a Katsura triple. Then, the following are equivalent: (1) Gtight(SZ,E) is minimal. (2) The adjacency matrix A of E is irreducible. Proof. First notice that E has no sinks by (18.1.(0)). Moreover, the action of Z on E fixes all the vertices. Then, by (13.6), Gtight(SZ,E) is minimal if and only if E is transitive, which is equivalent to the matrix A being irreducible, so we are done. (cid:3) Now, we will give a characterization of when Gtight(SZ,E) is essentially principal. 18.8. Theorem. Let (Z, E, ϕ) be a Katsura triple. Then, the following are equivalent: (1) Gtight(SZ,E) is essentially principal. (2) (a) Every circuit in E has an entry. (b) If 1 ≤ i ≤ N , l ∈ Z, and for any ξ ∈ Z(i) the elements l Bξn Aξn ∈ Z for all n ∈ , then there exists m ∈ such that Bξm = 0 for all ξ ∈ Z(i). Proof. Since the action of Z fixes all the vertices of E, (2a) is (14.10.i). On the other side, (2b) is exactly (14.10.ii) because of (18.4) and (18.5). Thus, the result is consequence of (14.15). (cid:3) We can obtain an easy sufficient condition for Gtight(SZ,E) being essentially principal. 18.9. Corollary. Let (Z, E, ϕ) be a Katsura triple. If (1) Every circuit of E has an entry, and (2) For every 1 ≤ i ≤ N and every l ∈ Z there exists η ∈ Z(i) such that lim n→∞ l Bηn Aηn = 0, then Gtight(SZ,E) is essentially principal. inverse semigroup actions and self-similar graphs 71 Proof. By (18.4), condition (2) implies that lη 6= η for any l ∈ Z, whence the triple (Z, E, ϕ) trivially satisfies (14.10.ii), as remarked in (14.11). (cid:3) Corollary (18.9) applies when we have a pair of finite matrices A, B under (18.1), such that for every 1 ≤ i ≤ N we have Aii ≥ 2 and Bii < Aii. In particular, Gtight(SZ,E) is essentialy principal for Katsura systems (Z, E, ϕ) satisfying (18.3.2). Also, it is immediate to characterize when Gtight(SZ,E) is locally contracting. 18.10. Theorem. Let (Z, E, ϕ) be a Katsura triple. Then, the following are equivalent: (1) Gtight(SZ,E) is locally contracting. (2) Every circuit of E has an entry. Proof. This is (15.1). Finally, we have the following fact (cid:3) 18.11. Proposition. If (Z, E, ϕ) is a Katsura triple, then Gtight(SZ,E) is an amenable groupoid. Proof. Since Z is an amenable group, (10.18) applies. (cid:3) Now, we are ready to characterize simplicity of the algebra OA,B, as follows 18.12. Theorem. Let (Z, E, ϕ) be a Katsura triple such that Gtight(SZ,E) is Hausdorff (see (18.6)). Then, the following are equivalent: (1) (a) The matrix A is irreducible. (b) Every circuit of E has an entry. ( c) If 1 ≤ i ≤ N , l ∈ Z, and for any ξ ∈ Z(i) the elements l n ∈ , then there exists m ∈ such that Bξm = 0. Bξn Aξn ∈ Z for all (2) OA,B is simple. Proof. This is exactly (16.1) for the Katsura triple (Z, E, ϕ), because of (18.7), (18.8) and (18.11). (cid:3) In particular, when Gtight(SZ,E) is Hausdorff and OA,B is simple, the Gtight(SZ,E) is locally contracting by (18.10) and (18.12.1b). Hence, we have 18.13. Corollary. If (Z, E, ϕ) is a Katsura triple such that Gtight(SZ,E) is Hausdorff and OA,B is simple, then OA,B is purely infinite simple. Proof. This is by (16.3). (cid:3) 18.14. Remark. Notice that, because of (18.9), Katsura's condition (18.3.2) for OA,B being a purely infinite simple C-algebra derive directly from (18.12) and (18.13) when Gtight(SZ,E) is Hausdorff. Moreover, when Gtight(SZ,E) is Hausdorff, (18.12) provides a characterization of simplicity for OA,B, improving Katsura's results on that direction, where only sufficient conditions are given [18]. 72 r. exel and e. pardo We close this section by presenting a couple of examples. The first one illustrates the difference between (G, E, ϕ) being pseudo free and Gtight(SG,E) being Hausdorff, and also the difference between the action of SG,E on E∞ being topologically free and the action of G on E∞ being topologically free. 18.15. Example. Set N = 2, and consider the matrices A =(cid:18) 2 1 1 2(cid:19) and B =(cid:18) 1 0 0 1(cid:19). Let (Z, E, ϕ) be the associated Katsura triple. Then, we have the following: (1) Since Z is amenable, then so is Gtight(SZ,E). (2) Since A is irreducible, Gtight(SZ,E) is minimal. (3) Every circuit in E has an entry. (4) Since A1,2 6= 0 and B1,2 = 0, (Z, E, ϕ) is not pseudo free by (18.5). B2,2 A2,2 (5) Notice that the only possible quotient values B1,1 A1,1 Bi,j Ai,j are = = 1 2 and B1,2 A1,2 = = B2,1 A2,1 minimal strongly fixed paths for l. Thus, Gtight(SZ,E) is Hausdorff by (18.6). = 0. Then, for any l ∈ Z, it is clear that there exists only finitely many 0 1 (6) Moreover, by the argument in point (5), the only infinite paths fixed by the action of Z are the ones associated to minimal strongly fixed paths, and thus trivial. Hence, the action of SZ,E on E∞ is topologically free. But the action of Z is not topologically free, since every element of Z fix the cylinders Z(e1,2,1) and Z(e2,1,1). Notice that OA,B is purely infinite simple by (18.12) and (18.13). The second example shows that (G, E, ϕ) being pseudo free do not imply that the action of SG,E is topologically free. 18.16. Example. Set N = 1, and set A = B = (n) for any n ≥ 2. Let (Z, E, ϕ) be the associated Katsura triple. Then, we have the following: (1) Since Z is amenable, then so is Gtight(SZ,E). (2) Since A is irreducible, Gtight(SZ,E) is minimal. (3) Every circuit in E has an entry. (4) Since A = B, (Z, E, ϕ) is pseudo free, whence in particular Gtight(SZ,E) is Hausdorff. (5) Since A = B, the action of Z on E is trivial. Thus, the action of SZ,E on E∞ (and of Z) cannot be topologically free, because there are no slacks. References [1] C. Anantharaman-Delaroche, "Purely infinite C∗-algebras arising form dynamical systems", Bull. Soc. Math. France, 125 (1997), no. 2, 199 -- 225. [2] C. Anantharaman-Delaroche and J. Renault, "Amenable groupoids", Monogr. Enseign. Math. 36, Universit´e de Gen`eve, 2000. [3] V. Arzumanian, J. Renault, "Examples of pseudogroups and their C∗-algebras", Operator algebras and quantum field theory (Rome, 1996), 93-104, Int. Press, Cambridge, MA, 1997. [4] J. Brown, L. O. Clark, C. Farthing and A. Sims, "Simplicity of algebras associated to ´etale groupoids", Semigroup Forum, 88 (2014), 433 -- 452. inverse semigroup actions and self-similar graphs 73 [5] N. P. Brown and N. Ozawa, "C-algebras and finite-dimensional approximations", Graduate Studies in Mathematics, 88, American Mathematical Society, 2008. [6] R. Exel, "Inverse semigroups and combinatorial C-algebras", Bull. Braz. Math. Soc., 39 (2008), no. 2, 191 -- 313. [7] R. Exel, "Non-Hausdorff ´etale groupoids", Proc. Amer. Math. Soc., 139 (2011), no. 3, 897 -- 907. [8] R. Exel, "Partial Dynamical Systems, Fell Bundles and Applications", Licensed under a Creative Available online from Commons Attribution-ShareAlike 4.0 International License, 351pp, 2014. mtm.ufsc.br/∼exel/publications. PDF file md5sum: bc4cbce3debdb584ca226176b9b76924. [9] R. Exel and M. Laca, "Cuntz-Krieger algebras for infinite matrices", J. reine angew. Math., 512 (1999), 119 -- 172. [10] R. Exel and E. Pardo, "Representing Kirchberg algebras as inverse semigroup crossed products", arXiv:1303.6268 [math.OA], 2013. [11] R. Exel and E. Pardo, "Graphs, groups and self-similarity", arXiv:1307.1120 [math.OA], 2013. [12] R. Exel and E. Pardo, "The tight groupoid of an inverse semigroup", arXiv:1408.5278 [math.OA], 2014. [13] R. Exel and C. Starling, "Self-similar graph C-algebras and partial crossed products", arXiv:1406. 1086 [math.OA], 2014. [14] R. Exel and A. Vershik, "C-algebras of irreversible dynamical systems", Canadian Mathematical Journal, 58 (2006), 39 -- 63. [15] R. I. Grigorchuk, "On Burnside's problem on periodic groups", Funct. Anal. Appl., 14 (1980), 41 -- 43. [16] N. D. Gupta and S. N. Sidki, "On the Burnside problem for periodic groups", Math. Z., 182 (1983), 385 -- 388. [17] T. Katsura, "A class of C-algebras generalizing both graph algebras and homeomorphism C- algebras. I. Fundamental results", Trans. Amer. Math. Soc., 356 (2004), no. 11, 4287 -- 4322. [18] T. Katsura, "A construction of actions on Kirchberg algebras which induce given actions on their K-groups", J. reine angew. Math., 617 (2008), 27 -- 65. [19] T. Katsura, "A class of C∗-algebras generalizing both graph algebras and homeomorphism C∗- algebras IV, pure infiniteness", J. Funct. Anal., 254 (2008), 1161 -- 1187. [20] A. Kumjian, D. Pask, I. Raeburn and J. Renault, "Graphs, groupoids, and Cuntz-Krieger algebras", J. Funct. Anal., 144 (1997), 505 -- 541. [21] A. Kumjian, D. Pask and I. Raeburn, "Cuntz-Krieger algebras of directed graphs", Pacific J. Math., 184 (1998), no. 1, 161 -- 174. [22] M. V. Lawson, "Inverse semigroups, the theory of partial symmetries", World Scientific, 1998. [23] M. V. Lawson, "Compactable semilattices", Semigroup Forum, 81 (2010), no. 1, 187 -- 199. [24] V. Nekrashevych, "Cuntz-Pimsner algebras of group actions", J. Operator Theory, 52 (2004), 223 -- 249. [25] V. Nekrashevych, "Self-similar groups", Mathematical Surveys and Monographs, 117, Amer. Math. Soc., Providence, RI, 2005. [26] V. Nekrashevych, "C-algebras and self-similar groups", J. reine angew. Math., 630 (2009), 59 -- 123. [27] A. L. T. Paterson, "Groupoids, inverse semigroups, and their operator algebras", Birkhauser, 1999. [28] G. K. Pedersen, "C-algebras and Their Automorphism Groups", Academic Press, 1979. [29] M. V. Pimsner, "A class of C-algebras generalizing both Cuntz-Krieger algebras and crossed products by Z", Fields Inst. Commun., 12 (1997), 189 -- 212. [30] I. Raeburn, "Graph algebras", CBMS Regional Conference Series in Mathematics, 103 (2005), pp. vi+113. [31] J. Renault, "Cartan subalgebras in C∗-algebras", Irish Math. Soc. Bull., 61 (2008), 29 -- 63. [32] C. Starling, "Boundary quotients of C*-algebras of right LCM semigroups", arXiv:1409.1549 [math.OA]. 74 r. exel and e. pardo [33] B. Steinberg, "A groupoid approach to discrete inverse semigroup algebras", Adv. Math., 223 (2010), 689 -- 727. Departamento de Matem´atica; Universidade Federal de Santa Catarina; 88010-970 Florian´opolis SC; Brazil ([email protected]) Departamento de Matem´aticas, Facultad de Ciencias; Universidad de C´adiz, Campus de Puerto Real; 11510 Puerto Real (C´adiz); Spain ([email protected])
1311.4638	2	1311	2015-09-06T22:29:11	Factoriality and type classification of \textsf{k}-graph von Neumann algebras	[ "math.OA", "math.FA" ]	Let $\Fth$ be a single vertex \textsf{k}-graph, and $\pi_\omega(\O_\theta)"$ be the von Neumann algebra induced from the GNS representation of a distinguished state $\omega$ of its $\textsf{k}$-graph C*-algebra $\O_\theta$. In this paper, we prove the factoriality of $\pi_\omega(\O_\theta)"$ and further determine its type, when either $\Fth$ has the little pull-back property, or the intrinsic group of $\Fth$ has rank $0$. The key step to achieve this is to show that the fixed point algebra of the modular action corresponding to $\omega$ has a unique tracial state.	math.OA	math	FACTORIALITY AND TYPE CLASSIFICATION OF k-GRAPH VON NEUMANN ALGEBRAS DILIAN YANG Abstract. Let F+ θ be a single vertex k-graph, and πω(Oθ)′′ be the von Neumann algebra induced from the GNS representation of a distinguished state ω of its k-graph C-algebra Oθ. In this paper, we prove the factoriality of πω(Oθ)′′ and further determine its type, when either F+ θ has the little pull-back property, or the intrinsic group of F+ θ has rank 0. The key step to achieve this is to show that the fixed point algebra of the modular action corresponding to ω has a unique tracial state. 1. Introduction Since the work [KP00] of Kumjian and Pask, higher rank graph (or k-graph) algebras have been extensively studied. See, for example, [FS02, HRSW13, HLRS13, KP00, KPS11, KPS12, PRRS06, Rae05, RSY03, RSY04] to mention but a few and the references therein for self-adjoint algebras, and [DPY08, DY09, DPY10, FY13, KP06, Pow07] for non-self-adjoint algebras. Being a special case, single vertex k- graph algebras provide a very important and intriguing class [DY09]. In particular, single vertex 2-graph algebras are systematically studied in [DPY08, DPY10, DY10, FY13, Yan10, Yan13]. Roughly speaking, a single vertex k-graph is a unital semigroup F+ θ having k types of generators. The generators from the same type form a unital free semigroup, while the generators from distinct types satisfy a family θ of permutations. It is often useful to think of a k-graph as a k-coloured directed graph, whose type i-generators are edges with i-th colour (1 ≤ i ≤ k). The graph C-algebra of F+ θ is denoted by Oθ, which is the universal C-algebra for the Cuntz-type representations of F+ θ . The universal property of Oθ yields a family of gauge automorphisms. Integrating those automorphisms over the k-torus Tk gives a faithful expectation Φ0 of Oθ onto the core F (i.e. the fixed point algebra of the 2010 Mathematics Subject Classification. 46L36, 47L65, 46L10, 46L05. Key words and phrases. k-graph, k-graph C-algebra, k-graph von Neumann algebra, factoriality, type classification, intrinsic group. The research was supported in part by an NSERC Discovery grant. 1 2 D. YANG gauge action) of Oθ. It is known that F is a UHF algebra. Particularly, F has a unique tracial state, denoted by τ . Then τ and Φ0 induce a distinguished faithful state ω of Oθ, namely, ω := τ ◦ Φ0. Let πω(Oθ)′′ denote the von Neumann algebra induced from the GNS representation of ω. Slightly abusing the terminology, we will simply call it the k- graph von Neumann algebra associated to F+ θ (induced from ω). As in [Yan10] for k = 2, one can obtain an explicit formula for the modular automorphism σt (t ∈ R) of πω(Oθ)′′ corresponding to ω. Identifying Oθ with πω(Oθ) yields a C-dynamical system (Oθ, R, σ). In [Yan10, Yan13], we initiate the study of the factoriality and type of πω(Oθ)′′ for a single vertex 2-graph F+ θ . Let m and n be the numbers of the blue and red edges of F+ θ , respectively. It is shown in [Yan10] that if ln m ln n 6∈ Q, then πω(Oθ)′′ is a type III1 (AFD) factor. In order to obtain the factoriality of πω(Oθ)′′ when ln m ln n ∈ Q for (aperiodic) 2-graph F+ θ , we carefully study the structure of the fixed point algebra Oσ θ of the modular action σ. It is proved in [Yan13] that Oσ θ is a crossed product of the core F by Z. Under the assumption that Oσ θ has a unique tracial state, we are able to show that πω(Oθ)′′ is a type IIIλ factor, where 0 < λ < 1 is completely determined by m and n. It is asked in [Yan13] if this assumption that Oσ θ has a unique tracial state is redundant. In this paper, we answer this affirmatively when F+ θ has the little pull-back property (see Section 2.1 below for its definition). Actually we are able to obtain much more. Before going further, let us introduce an important group G associated to F+ θ : k G =(cid:8)g = (g1, . . . , gk) ∈ Zk Yi=1 mgi i = 1(cid:9), where mi is the number of i-th coloured edges of F+ θ (1 ≤ i ≤ k). G is called the intrinsic group of F+ θ . In general, 0 ≤ rank G ≤ k − 1. In this paper, we prove that πω(Oθ)′′ is a factor and further determine its type, when either F+ θ has the little pull-back property, or the intrinsic group of F+ θ has rank 0. To this end, we carefully study the structure of the fixed point algebra Oσ θ has the property of possessing a unique tracial state. θ , and show that Oσ The remainder of this paper is organized as follows. In Section 2, some backgrounds needed later are briefly given. We first investigate the modular theory of k-graph algebras in Section 3. Since it is similar to the case of k = 2 studied in [Yan10], we sketch it here and only give somehow more unified proofs when necessary. Then we study the structure of the fixed point algebra Oσ θ is a C- algebra generated by the standard generators of Oθ with degrees in G. θ . It is proved that Oσ FACTORIALITY AND TYPE CLASSIFICATION 3 θ has the Dixmier property if F+ When rank G = 1, it is a crossed product of F by Z. In Section 4, we show that Oσ θ has the little pull-back property, and so it has a unique tracial state. This is obtained as a consequence of the fact that Oθ enjoys a property stronger than the Dixmier property. To this end, it needs some careful analysis about a class of averaging operators in Oθ. Those averaging operators are induced from unitaries in the algebraic part of the core F. The idea of this part is strongly influenced by the work [Arc80] of Archbold. Our mains theorems are stated and proved in Section 5. We should mention that after the first version of this paper was cir- culated in 2013, very recently the main results of this paper have been generalized in [LLNSW15] by using completely different approaches. Let us finish off this introduction by fixing our notation and conven- tions. Notation and Conventions. There is a natural coordinate-wise par- tial ordering on Zk: p, q ∈ Zk with p ≤ q ⇐⇒ pi ≤ qi for all 1 ≤ i ≤ k. For p, q ∈ Zk, p ∨ q and p ∧ q denote the coordinate-wise maximum and minimum of p and q, respectively. Let p+ = p ∨ 0, and p− = (−p) ∨ 0. Set Nk = {p ∈ Zk : p ≥ 0}. Then p = p+ − p− with p+, p− in Nk and p+ ∧ p− = 0. For a nonzero n ∈ N we use the following notation: n = (n, . . . , n) ∈ Nk, n = {1, . . . , n}. In this paper, every C-algebra A is assumed to be unital. The group of its unitaries is denoted by U(A), and its centre is written as Z(A). By an endomorphism of A, we always mean a unital, -endomorphism. Let εi (1 ≤ i ≤ k) be the standard generators for Zk. From now on, θ is a fixed single vertex k-graph with mi i-th we always assume that F+ coloured generators (i.e., of degree ǫi) (1 ≤ i ≤ k). If F+ θ is periodic (equivalently, Oθ is not simple), then it is known that the centre of Oθ is isomorphic to C(Ts) for some s ≥ 1 [DY09, DY10]. So πω(Oθ)′′ is not a factor in this case. Hence, throughout the rest of this paper, we always assume that F+ is aperiodic, unless otherwise θ specified. In particular, this implies that mi > 1 for all 1 ≤ i ≤ k [DY09]. Set m := (m1, . . . , mk), and we use the multi-index notation: for p ∈ Zk, mp :=Qk Finally, if p ∈ Nk, the set of words in F+ θ of degree p is denoted by i=1 mpi i . Λp = {w ∈ F+ θ d(w) = p}. 4 D. YANG 2. Preliminaries The main sources of this section are [Arc80, DPY08, DY09, Exe08, KP00, Rae05, Yan10]. 2.1. k-graphs. Kumjian and Pask [KP00] define a k-graph as a small category Λ with a degree map d : Λ → Nk satisfying the factorization property: for every λ ∈ Λ and p, q ∈ Nk with d(λ) = p + q, there are unique elements µ, ν ∈ Λ such that λ = µν with d(µ) = p and d(ν) = q. Recall that a k-graph Λ is finitely aligned if for every µ, ν ∈ Λ one has Λmin(µ, ν) := {(ξ, η) ∈ Λ × Λ : µξ = νη and d(µξ) = d(µ) ∨ d(ν)} is finite. Definition 2.1. We say that Λ is singly aligned, if Λmin(µ, ν) ≤ 1 for every (µ, ν) ∈ Λ × Λ. An obvious weaker notion is the following. Definition 2.2. A k-graph Λ is said to have the little pull-back property if Λmin(e, f ) ≤ 1, whenever e, f are edges of Λ with distinct degrees. To my best knowledge, the little pull-back property was first in- troduced by Exel in [Exe08] in order to apply his tight representation theory to k-graphs. This property has nice and clear geometrical mean- ing in the commuting-square representation. It turns out that that the little pull-back property and single alignment are equivalent ([Exe08]). In this paper, we focus on single vertex k-graphs. Their definition can be equivalently stated as follows. Definition 2.3. A single vertex k-graph F+ erated by {ei ties: θ is a unital semigroup gen- s : s ∈ mi, 1 ≤ i ≤ k}, which satisfies the following proper- (fsg). {ei s : s ∈ mi} generates a unital free semigroup F+ mi; (θ-CR). There is a family of permutations θ = {θij : θij ∈ Smi×mj for 1 ≤ i < j ≤ k}, such that the following θ-commutation relations hold sej ei t = ej t′ei s′ where θij(s, t) = (s′, t′). (Cubic). Every three sets of generators {ei t1, ej t2, ek t3} satisfy the follow- t1ej ei ing cubic condition: ej t1ek t′ t′ 3 2 t1′ ek t2′ ei t3 = ei = ej t2ek ej ei = ek t′′ t′ t′ 3 1 2 t3 = ej t3′ ei t2′ ek ej = ek ei t′′ t′′ t′′ 3 1 2 t3′′ ej t1′′ = ek t2′′ ei t1′′ FACTORIALITY AND TYPE CLASSIFICATION 5 =⇒ ei t′′ 1 = ei t1′′ , ej t′′ 2 = ej t2′′ , ek t′′ 3 = ek t3′′ . t , ek s, ej By the above definition, for k = 2, every permutation completely determines a 2-graph. But for k ≥ 3, not every family θ of permutations yields a k-graph. The above cubic condition requires, very roughly speaking, that a word of degree (1, 1, 1) should be well defined. Thus F+ θ is a k-graph if and only if the restriction of F+ θ to every triple family of edges {ei u} is a 3-graph. Return to the little pull-back property. One can see that F+ θ has s, ej the little pull-back property if and only if so is every 2-graph {ei t }. By [DY09], if F+ θ has the little pull-back property, then it has to be aperiodic; however, the converse does not hold in general. The little pull-back property seems to be restrictive. But it is actually not very so. It is known in [Pow07] that there are essentially only nine single vertex 2-graphs with 2 blue and 2 red edges. One can easily check that all aperiodic graphs have the little pull-back property (and so are singly aligned) except only one(!) given by θ = ((1, 1), (2, 1), (1, 2)). 2.2. k-graph algebras. Recall from [FY13] that an n-tuple V = (V1, ..., Vn) of operators in a C-algebra A is said to be of Cuntz-type if n V ∗ i Vj = δi,j and ViV ∗ i = I. Xi=1 Namely, Vi's are orthogonal isometries, and satisfy the defect free prop- erty in the terminology of [DPY08, DPY10, DY10, DY09]. Let F+ θ be a k-graph, and Ei = (Ei 1, . . . , Ei mi) (1 ≤ i ≤ k). We call [E1, . . . , E k] a representation of F+ Ei sEj t = Ej t′Ei θ if it satisfies (θ-CR): s′ when θij(s, t) = (s′, t′) for all 1 ≤ i < j ≤ k, s ∈ mi, t ∈ mj. This representation is simply denoted as E. E is said to be of Cuntz-type, if every Ei is of Cuntz-type (1 ≤ i ≤ k). Definition 2.4. The k-graph C-algebra of F+ universal C-algebra for Cuntz-type representations of F+ θ . θ , denoted as Oθ, is the So Oθ is the C-algebra generated by F+ that every Cuntz-type representation of F+ type representation of Oθ. Also, we reserve the notation sej written as sj θ with the universal property θ extends uniquely a Cuntz- , simply t (1 ≤ j ≤ k, t ∈ mj), for the generators of Oθ. For w in F+ θ , by θ-commutation relations (θ-CR), one can always write w = u1 · · · uk with ui being a word of ei's. The degree of w is t 6 D. YANG defined by d(w) = (u1, . . . , uk), where ui is the length of ui. For w = ej1 θ , we use the multi-index notation t1 · · · ejn tn ∈ F+ sw = se · · · sejn tn j1 t1 = sj1 t1 · · · sjn tn . The degree map d of F+ θ can be extended as follows. By the universal property of Oθ, there is a family of gauge automorphisms γt for t ∈ Tk given by γt(sw) = td(w)sw for all w ∈ F+ θ . For each n ∈ Zk, define a mapping Φn on Oθ via Φn(A) =ZTk t−nγt(A)dt for all A ∈ Oθ. So Ran Φn is a spectral subspace of the gauge action γ. If A ∈ Ran Φn, we say that the degree of A is n, and write d(A) = n. So in particular d(sus∗ v) = d(u) − d(v) for u, v ∈ F+ θ . It turns out that Φ0 is of particular importance to us. It is a faithful i -UHF θ . Furthermore, Oγ i=1 m∞ expectation onto the core Oγ algebra: θ is a Qk F := Ran Φ0 = Oγ θ = [n≥1 where Fn, Fn = span{sus∗ v : d(u) = d(v) = n} . (See [Yan10] for the explanation is the full matrix algebra MQk why one can only consider such special elements n in Nk.) Let τ be the (unique) tracial state of F. With Φ0, τ induces a distinguished faithful state ω on Oθ by i=1 mn i ω(A) = τ (Φ0(A)) for all A ∈ Oθ. Let πω(Oθ)′′ be the von Neumann algebra generated by the GNS rep- resentation of ω. When no confusion is caused, we shall omit the subscript ω and write π(Oθ)′′ for short. Abusing the terminology, we give the following definition. Definition 2.5. The von Neumann algebra π(Oθ)′′ is called the k-graph von Neumann algebra associated to F+ θ (induced from ω). Also, since π is faithful, we will often identify Oθ and π(Oθ). Refer to [DY09, Yan10] and the references therein for more details about this subsection. FACTORIALITY AND TYPE CLASSIFICATION 7 2.3. Endomorphisms of C-algebras arising from Cuntz-type tuples. Let A be a C-algebra. Let V = (V1, . . . , Vn) be an arbitrary Cuntz-type tuple in A and C∗(V ) be the C-algebra generated by Vi's. For p ∈ N, one has a Cuntz-type np-tuple: Vp = (Vw : w ∈ F+ n , w = p). As in [Yan10], we define γp(A) = Xw=p VwAV ∗ w for all A ∈ C∗(V ). Then γp is an endomorphism of C∗(V ). Clearly, γp can be extended to A, which is also denoted by γp, sometimes more precisely by γV p in order to emphasize the tuple V which gives rise to it. Return to the k-graph C-algebra Oθ. Let E = [E1, ..., E k] be a Cuntz-type representation of Oθ. One obtains an endomorphism γp on C∗(E) for p ∈ Nk. If Ei's are from Oθ, then we have an endomorphism γ E p on Oθ: γ E p (A) = Xw∈Λp EwAE∗ w for all A ∈ Oθ. (1) Here the multi-index notation Ew is similar to sw as above. 2.4. The Dixmier property. Let A be a C-algebra and V be a subgroup of the unitary group U(A) of A. For given U1, ..., Un ∈ V and i=1 λi = 1, we define an averaging positive numbers λ1, ..., λn with Pn operator α on A via n α(A) = Xi=1 λiUiAU ∗ i for all A ∈ A. We let Ave(A, V) denote the set of all such averaging operators. It is easy to see that Ave(A, V) is a unital semigroup with composition as the semigroup multiplication, and that every α ∈ Ave(A, V) is a unital positive map. In particular, every averaging operator is contractive and self-adjoint. Definition 2.6. A C-algebra A is said to have the Dixmier property if, for every A ∈ A, {α(A) : α ∈ Ave(A, U(A))} k·k ∩ Z(A) 6= ∅. We say that A satisfies the weak Dixmier property if, for every A ∈ A+ \ {0}, span{α(A) : α ∈ Ave(A, U(A))} k·k ⊇ CI. 8 D. YANG The Dixmier property has been extensively studied. See, for ex- ample, [Arc77, Arc79, Arc80, HZ84, Rie82], amongst others. It is well- known that a simple C-algebra with the Dixmier property has at most one tracial state. Haagerup and Zsid´o [HZ84] showed the converse for simple C-algebras: every simple C-algebra with at most one tracial state has the Dixmier property. The weak Dixmier property is intro- duced in [Rie82] in order to partially answer a question posed in [Arc79]. Clearly, the weak Dixmier property is weaker than the Dixmier prop- erty. It is proved in [Rie82] that A satisfies the weak Dixmier property if and only if A is simple and has a most one tracial state. So it is now apparent that the Dixmier property and the weak Dixmier property are equivalent for simple C-algebras. One advantage of the weak Dixmier property is that it implies simplicity. 3. The modular theory of Oθ and Structure of Oσ θ In this section, we first briefly discuss the modular theory of k-graph C-algebras. This is very similar to [Yan10] for k = 2. The associated modular objects are explicitly given. They will be useful in calculating the Connes spectrum of the modular operator. In the second part of this section, we investigate the structure of Oσ θ , the fixed point algebra of the modular action corresponding to the distinguished state ω. Its structure plays a critical role for what follows. 3.1. The modular theory of Oθ. Let ◦Oθ denote the algebraic alge- bra generated by su's. So ◦Oθ = span{sus∗ θ }. In the sequel, for n ∈ Zk, we use Xn to denote the intersection of Ran Φn and ◦Oθ: v : u, v ∈ F+ Xn = span{sus∗ v : u, v ∈ F+ θ , d(u) − d(v) = n}. So every element of Xn has degree n. Recall from Section 2 that the distinguished faithful state ω is given by ω = τ ◦Φ0, where τ is the tracial state on F and Φ0 is the expectation of Oθ onto the core F of Oθ. Let L2(Oθ) be the GNS Hilbert space determined by ω. So the inner product on L2(Oθ) is given by hA Bi = ω(A∗B) for all A, B ∈ L2(Oθ). For A ∈ Oθ, we denote the left action of A by πω(A), that is, πω(A)B = AB for all B in Oθ. We now begin to give the modular objects in the celebrated Tomita- Takesaki modular theory. Define an anti-linear operator S◦ on ◦Oθ ⊂ L2(Oθ) by S◦(A) = A∗ for all A ∈ ◦Oθ, and another one F◦ on ◦Oθ, which acts on generators by F◦(sus∗ v) = md(u)−d(v)svs∗ u, FACTORIALITY AND TYPE CLASSIFICATION 9 and then extend it anti-linearly. As in [Yan10], we shall show that F◦ is indeed the adjoint of S◦. The following proof is essentially the same as that of [Yan10, Lemma 5.2], but in a slightly more unified way (even in the case of k = 2). A lemma first, whose proof can be easily modified from that of [Yan10, Lemma 5.1] and so we omit it. Lemma 3.1. Suppose u, v ∈ F+ θ with d(u) = d(v). Then τ (suAs∗ v) = δu,v m−d(u)τ (A) for all A ∈ F. Lemma 3.2. Let S◦ and F◦ be defined as above. Then F◦ = S∗ ◦ : hS◦(A) Bi = hF◦(B) Ai for all A, B in ◦Oθ. Proof. It suffices to prove this lemma for standard generators A, B in ◦Oθ. By the definitions of S◦ and F◦, we have hS◦(A) Bi = ω(AB), hF◦(B) Ai = md(B)ω(BA). If d(A) + d(B) 6= 0, then neither AB nor BA belongs to F. So hS◦(A) Bi = hF◦(B) Ai = 0. We now assume that d(A) + d(B) = 0. Assume d(A) = n ∈ Zk. Then, by the θ-commutation relations, there are u1, v1, u2, v2 ∈ F+ θ with d(u1) = d(u2) = n+ and d(v1) = d(v2) = n−, and A′, B′ in F such that A = su1A′s∗ v1, B = sv2B′s∗ u2. Then by Lemma 3.1 we have ω(AB) = ω(su1A′s∗ v1sv2B′s∗ u2) = δv1, v2ω(su1A′B′s∗ u2) = δv1,v2δu1,u2 md(u1)ω(A′B′), ω(BA) = ω(sv2B′s∗ u2su1A′s∗ v1) = δu1, u2ω(sv2B′A′s∗ v1) = δu1,u2δv1,v2 md(v1)ω(B′A′). Thus ω(AB) = md(v1)−d(u1)ω(BA) = md(v2)−d(u2)ω(BA) = md(B)ω(BA). This ends the proof. By Lemma 3.2, we particularly obtain that both F◦ and S◦ are clos- able. We use F and S to denote their corresponding closures. The closure S is called the Tomita operator. The following is well-known: S and F have polar decompositions S = J∆ 1 2 = ∆− 1 2 J, F = J∆− 1 2 = ∆ 1 2 J, 10 D. YANG where ∆ = F S is the modular operator, and J is the modular conjuga- tion. By Lemma 3.2 we have and for z ∈ C J(sus∗ v) = m d(u)−d(v) 2 svs∗ u, ∆z(sus∗ v) = mz(d(v)−d(u))sus∗ v. Some obvious modifications of the proof of [Yan10, Theorem 5.3] shows that the algebra ◦Oθ, with the inner product h· ·i: hA Bi = ω(A∗B), is a modular Hilbert algebra. Note that πω(Oθ)′′ is nothing but the left von Neumann algebra of ◦Oθ ([Tak03II]). The Tomita- Takesaki modular theory says that ∆itπ(Oθ)′′∆−it = π(Oθ)′′ for all t ∈ R. For z ∈ C, let σz(π(A)) = ∆izπ(A)∆−iz for all A ∈ ◦Oθ. The one-parameter group {σt : t ∈ R} is called the modular automor- phism group. As in [Yan10], it is not hard to obtain the formula of σt on generators: σt(π(sus∗ v)) = mit(d(v)−d(u))π(sus∗ v). (2) Refer to [KR97II, Str81, Tak03II] for more information on the Tomita- Takesaki modular theory. Identifying Oθ with π(Oθ), one obtains a C-dynamical system (Oθ, R, σ), or simply (Oθ, σ). The main purpose of the next subsection is to look closer at the structure of the fixed point algebra Oσ θ . 3.2. The structure of Oσ 1 ≤ i ≤ k, mi is the number of edges of F+ θ . Recall that m = (m1, . . . , mk), where, for θ of degree ǫi. Define G = {g ∈ Zk : mg = 1}. (3) It is easy to check that G is a subgroup of Zk. This group is important in this section and in classifying k-graph von Neumann algebras later. So it deserves having its own name. Definition 3.3. The group G defined in (3) is called the intrinsic group of F+ θ . In the following, let r be the rank of G and g1, ..., gr be the generators of G. Consider the set {ln mi : 1 ≤ i ≤ k}. There are two cases for this set: it is either rationally dependent or independent. FACTORIALITY AND TYPE CLASSIFICATION 11 Suppose first that it is rationally dependent. Then it is equivalent to mg = 1 for some g in Zk \ {0}. Let us notice that for every g ∈ G \ {0}, one has g+ 6= 0 and g− 6= 0. For 1 ≤ p ≤ r, consider the generator gp. Since gp ∈ G, we have mgp+ = mgp−. So there is a bijection, say p (once chosen and fixed), from the subset onto the subset Λgp+ = {w ∈ F+ θ : d(w) = gp+} Λgp− = {w ∈ F+ θ : d(w) = gp−}. Define an operator Up in Oθ by Up = Xu∈Λgp+ sus∗ p(u), which is easily seen to be unitary. We now define an action ρ of the free group Fr with the generators a1, ..., ar on F as follows: ρap = AdUp (1 ≤ p ≤ r). Then we obtain a C-dynamical system (F, Fr, ρ). If {ln mi : 1 ≤ i ≤ k} is rationally independent, equivalently, G = {0} and so r = 0, , we make the convention that U0 = I. The following two results generalize [Yan13, Theorem 3.1] where G = Z(a, −b), and unifies some proofs of [Yan10, Yan13]. Proposition 3.4. Keep the same notation as above. Then Oσ θ = C∗(X ∈ Oθ : d(X) ∈ G) = C∗(F, U1, . . . , Ur). Proof. Similar to the proof of [Yan13, Theorem 3.1], to show the first θ ⊆ C∗(X ∈ Oθ : d(X) ∈ G). To this "=", it suffices to verify Oσ end, let us consider the generator X = sus∗ θ . Since σt(X) = X for all t ∈ R, one has eit ln md(v)−d(u) = 1 for all t ∈ R. This implies md(v)−d(u) = 1, namely, d(v) − d(u) ∈ G, as required. v in Oσ To prove the second "=", one needs to check C∗(X ∈ Oθ : d(X) ∈ G) ⊆ C∗(F, U1, . . . , Ur). For this, take X ∈ Oθ with d(X) ∈ G and without loss of generality assume that X = sus∗ v. Rewrite it 'in the 12 D. YANG order of a1, . . . , ar', and then use the definition of Up's to obtain X = su1,1 · · · su1,n1 · · · sur,1 · · · sur,nr s∗ v = U1s1(u1,1) · · · U1s1(u1,n1 ) · · · Ursr(ur,1) · · · Ursr(ur,nr )s∗ = U1s1(u1,1) · · · U1s1(u1,n1 ) · · · Ursr(ur,1) · · · Ur sr(ur,nr ) · s∗ v v · (U ∗nr r · · · U ∗n1 1 ) · (U n1 1 ∈ F U n1 1 · · · U nr r ) · · · U nr r , where up,j ∈ Λgp+ for all 1 ≤ p ≤ r and 1 ≤ j ≤ np. With the aid of [Hop05], the following lemma can be easily proved as [Yan10, Lemma 4.1] where k = 2 , and so we omit its proof here. Lemma 3.5. The following statements are equivalent. θ is aperiodic. (i) F+ (ii) The relative commutant of F is trivial: F′ ∩ Oθ = C. Proposition 3.6. Suppose rank G = 1. Then Oσ algebra, which is isomorphic to the crossed product of F by Z: θ is a simple C- Oσ θ ∼= F ⋊ρ Z. Proof. If r = 0, namely, {ln mi : 1 ≤ i ≤ k} is rationally independent, then Oσ θ coincides with F by Proposition 3.4 and so clearly has all required properties. Suppose r = 1. Since F+ θ is aperiodic, as shown in [Yan13, Theorem 3.1] using Lemma 3.5, ρ is aperiodic meaning that ρn is outer for every n ∈ Z. Then it follows from, e.g. [Kis81], that the reduced crossed product F ⋊ρ, r Z is simple. But since Z is amenable, one has F ⋊ρ Z ∼= F ⋊ρ, r Z. So the universal property of F ⋊ρ Z gives the required isomorphism by Proposition 3.4. It would be interesting to know the dynamical system structure of θ when rank G > 1. Oσ 4. The Dixmier Property of Oσ θ In [Arc80], Archbold studied the Dixmier property of Cuntz algebras On. But actually more was obtained there:1 the averaging operators α in Definition 2.6 can be chosen from Ave(On, U(X0)), where X0 is the algebraic part of the core of On. It is this stronger property of On that motivates this section. The aim of this section is to show that the fixed point algebra Oσ θ has the little θ has a unique tracial state if either F+ 1This was kindly informed to the author by Professor Robert Archbold. FACTORIALITY AND TYPE CLASSIFICATION 13 θ has rank 0. To achieve pull-back property, or the intrinsic group of F+ this, we first prove that Oσ θ is simple and has the Dixmier property. However, the higher rank flavour of k-graphs causes some compli- cation here. For instance, the degree of a standard generator of On and 0 are always comparable; but this is not the case for k-graphs. So it seems that some key results in [Arc80] can not be generalized to k-graph algebras. Fortunately, in our case, we can make full use of the little pull-back property of F+ θ and the defect free property of Cuntz-type representations to achieve our goal. Let E be a Cuntz-type representation of F+ θ . For p ∈ Nk, let Ωp be the set of all functions from Λp, the set of all words in F+ θ with degree p, to {0, 1}, and Pp be the permutation group of Λp. For f ∈ Ωp and ς ∈ Pp, define the unitary operator by U(f,ς) = Xw∈Λp (−1)f (w)Eς(w)E∗ w, and the averaging operator on C∗(E) via αp(A) = 1 2mp(mp)! Xf ∈Ωp, ς∈Pp U(f,ς)AU ∗ (f,ς) for all A ∈ C∗(E). (4) p in Section 2, when the operators of E are from Oθ, we use αE p to As γ E denote its extension to Oθ. 4.1. Some auxiliary lemmas. In this subsection, we prove some technical lemmas, which will be useful later. Lemma 4.1. Let E be a Cuntz-type representation of Oθ and p ∈ Nk. Suppose that u, v ∈ F+ θ are such that u = u1u2 and v = v1v2 with ui, vi in F+ θ (i = 1, 2) and d(u1) = d(v1) = p. Then αp(EuE∗ v) = δu1,v1 mp γp(Eu2)γp(Ev2)∗. Proof. The following proof is essentially the same as that of [Arc80, Proposition 2 (ii)], so we only sketch it here. 14 D. YANG Indeed, we have mp αp(EuE∗ v) = = = 1 2mp(mp − 1)! Xf ∈Ωp, ς∈Pp Xµ∈Λp v1 · Xν∈Λp Eu1Eu2E∗ v2 E∗ (−1)f (µ)Eς(µ)E∗ (−1)f (ν)EνE∗ µ! · ς(ν)! (−1)f (u1)+f (v1)Eς(u1)Eu2E∗ v2 E∗ ς(v1) 1 2mp(mp − 1)! Xf ∈Ωp, ς∈Pp (mp − 1)! Xς∈Pp δu1,v1 = δu1,v1γp(Eu2E∗ = δu1,v1γp(Eu2)γp(Ev2)∗ v2) Eς(u1)Eu2E∗ v2 E∗ ς(u1) (as γp is a C-homomorphism). We are done. The following lemma is really of higher rank flavour, and it is crucial in the proof of Lemma 4.4 below. Lemma 4.2. Let u, v ∈ F+ d(u) + d(v). Suppose that E is a Cuntz-type representation of F+ we have the following. θ be such that d(u) ∧ d(v) = 0 and p = θ . Then (i) αp(EuE∗ v) = 1 mp X(v′,u′)∈Λmin(u,v) γp(E∗ v′ Eu′). Here we use the convention: if Λmin(u, v) = 0, then E∗ 0. v′ Eu′ := (ii) If F+ θ has the little pull-back property, then αp(EuE∗ v) = 0 or 1 mp γp(Eu′′ E∗ θ × F+ v′′), for a unique pair (u′′, v′′) ∈ F+ d(v′′) = d(v). θ with d(u′′) = d(u) and Proof. Suppose that E is a Cuntz-type representation of F+ θ . (i) Similar to Lemma 4.1, we prove it from the following calculations. αp(EuE∗ v) = = 1 2mp(mp)! Xf ∈Ωp, ς∈Pp Xµ∈Λp (mp)!Xς Xµ∈Λp Eς(µ)E∗ 1 µ · EuE∗ v · EµE∗ ς(µ)! (−1)f (µ)Eς(µ)E∗ µ · EuE∗ (−1)f (ν)EνE∗ ς(ν)! v · Xν∈Λp FACTORIALITY AND TYPE CLASSIFICATION 15 We now write µ as µ1µ2 with d(µ1) = d(u) and d(µ2) = d(v). Contin- uing above, one has αp(EuE∗ v) v · Eµ1µ2E∗ µ2 · E∗ vEu · Eµ2E∗ Eς(uv′) · E∗ v′Eu′ · E∗ Eς(uv′) · E∗ v′ Eu′ · E∗ = = = = = = 1 1 1 Eς(µ)E∗ Eς(uµ2)E∗ µ1µ2 · EuE∗ (mp)!Xς Xµ∈Λp  (mp)!Xς  Xµ2∈Λd(v)  (mp)!Xς  Xµ2∈Λd(v) X(v′,u′)∈Λmin(u,v)  (mp)!Xς  X(v′,u′)∈Λmin(u,v) (mp)! X(v′,u′)∈Λmin(u,v) Xς mp X(v′,u′)∈Λmin(u,v) v′Eu′). γp(E∗ 1 1 1 ς(µ)! ς(uµ2)  ς(uv′)  ς(uv′)! Eς(uµ2)E∗ µ2 · Eu′E∗ v′ · Eµ2E∗ ς(uµ2)  (ii) Suppose that F+ θ has the little pull-back property. Then we have a most one pair (v′, u′) in Λmin(u, v), and a most one pair (u′′, v′′) in Λmin(v′, u′). Thus αp(EuE∗ v′′) if Λmin(u, v) = Λmin(v′, u′) = 1 and (u′′, v′′) ∈ Λmin(v′, u′), and αp(EuE∗ v) = γp(Eu′′E∗ v) = 0, otherwise. For n ∈ Nk, recall that Xn is the algebraic part of the spectral sub- space Ran Φn of the gauge action of Oθ. It follows from the definition of αp directly that Xn is invariant for αp. We record it below for later reference. Lemma 4.3. For p ∈ Nk and n ∈ Zk, we have αp(Xn) ⊆ Xn. In the rest of this section, let us fix the endomorphism γp and the averaging operator αp on Oθ as follows: suAs∗ u for all A ∈ Oθ, γp(A) = Xu∈Λp 1 αp(A) = 2mp(mp)! Xf ∈Ωp, ς∈Pp where Uf,ς =Pu∈Λp(−1)f (u)sς(u)s∗ u. U(f,ς)AU ∗ (f,ς) for all A ∈ Oθ, 16 D. YANG Lemma 4.4. Suppose that F+ ǫ > 0. θ has the little pull-back property, and let (i) If A ∈ Xn with n 6= 0, then there exists α ∈ Ave(Oθ, U(X0)) such that kα(A)k < ǫ. (ii) If A = Pn6=0 An in ◦Oθ with An ∈ Xn, then there is α ∈ Ave(Oθ, U(X0)) such that kα(A)k < ǫ. Proof. Let ǫ > 0 be given. (i) Let A ∈ Xn with n 6= 0. We divide the proof into three steps. Step 1. Clearly, A is spanned by the standard generators susw1s∗ w2s∗ v, where u, v, w1, w2 ∈ F+ θ , d(u) ∧ d(v) = 0, and d(w1) = d(w2) =: q, say. Thanks to the defect free property of s, we can (and do) always assume that all summands of A have a uniform such common degree q. (Actually, one can take it as the join ∨q of q's in the summands.) Notice that for a fixed n ∈ Zk, the equation k − l = n with k, l ∈ Nk and k ∧ l = 0 has a unique solution k = n+ and l = n−. Thus one can write A as follows: A = au,v,w1,w2susw1s∗ w2s∗ v. u∈Λn+ , v∈Λn− , w1,w2∈Λq X For convenience, set S := span{γq(su)γq(sv)∗ : u ∈ Λn+, v ∈ Λn−}. From Lemma 4.1 it follows that αq(A) ∈ A. Let us assume that αq(A) =Xu,v au,vγq(su)γq(sv)∗, where of course there are only finitely many coefficients au,v, which are not equal to 0. Step 2. Let p = n+ + n−. Applying Lemma 4.2 (ii) to the Cuntz- type representation E1 := γq(s) and using what we obtain in Step 1, we conclude that αE1 p (αq(A)) = 1 mp γ E1 p (A1), where A1 ∈ S and its coefficients either come from those of αq(A) or vanish. Straightforward calculations show that γ E1 p ◦ γq = γp+q. Since γp+q(s) =: E2 is a Cuntz-type representation of F+ θ , applying Lemma 4.2 (ii) again to E2 yields p (αE1 αE2 p (αq(A))) = 1 mp αE2 p (γ E1 p (A1)) = 1 m2p γ E2 p (γ E1 p (A2)) = 1 m2p γ2p(A2), kα(A)k = 1 mnp kγnp(An)k ≤ 1 mnp kAnk ≤ 1 mnp X au,v < ǫ. FACTORIALITY AND TYPE CLASSIFICATION 17 where A2 shares the same properties of A1. Step 3. Continuing the above process in an obvious way, we ob- tain Cuntz-type representations E1, . . . , En and averaging mappings αE1 p , . . . , αEn in Ave(Oθ, U(X0)) such that p p ◦ · · · ◦ αE1 αEn p ◦ αq(A) = 1 mnp γnp(An), where An ∈ S, and its coefficients either come from those of αq(A) or vanish. Let α = αEn p ◦ αq, which is clearly in Ave(Oθ, U(X0)). p ◦ · · · ◦ αE1 By the definition of αq(A), one has Pu,v au,v is finite. It follows from the properties of An that kAnk ≤P au,v < ∞ as kγm(susv)∗k ≤ 1 for all m ∈ Nk. Hence, when n is large enough, we have Here we use the facts that γnp is isometric since it is an endomorphism on Oθ and Oθ is simple. (ii) This is completely similar to the proof of [Arc80, Proposition 4(iii)]. Let A = Pn6=0 An in ◦Oθ with An ∈ Xn. We label the set of non-zero summands {An 6= 0 : n 6= 0} of A as {A1, . . . , AL}. So L is finite as A ∈ ◦Oθ. By (i) above, there is α1 ∈ Ave(Oθ, U(X0)) such that kα1(A1)k < ǫ L . Apply (i) above to α1(A2) by using Lemma 4.3, and we get α2 ∈ Ave(Oθ, U(X0)) such that kα2 ◦ α1(A2)k < ǫ L . Continuing this way, we obtain averaging operators α1, . . . , αL in Ave(Oθ, U(X0)) such that kαL ◦ · · · ◦ α1(AL)k < ǫ L . Hence kαL ◦ · · · ◦ α1(Ai)k < ǫ L (1 ≤ i ≤ L), as every αi is contractive. Therefore, α = αL ◦ · · · ◦ α1 belongs to Ave(Oθ, U(X0)) and satisfies as required. kα(A)k < ǫ, Remark 4.5. Lemma 4.4 does not hold true in general if F+ θ does not satisfy the little pull-back property. For instance, consider a periodic k-graph F+ It is shown in [DY09] that there is a central unitary θ . γ(u) in Oθ, where γ is a bijection from Λa to Λb with a, b > 0 in Nk and a ∧ b = 0. It is rather easy to check that α(W ) = W for every averaging operator α. So one always has kα(W )k = 1. W = Pu∈Λa sus∗ 18 D. YANG 4.2. The Dixmier property of Oσ θ . We already know that Oθ has the Dixmier property since it is simple and has no tracial state (cf. Section 2). The following shows that Oθ has a stronger property when F+ θ has the little pull-back property. Proposition 4.6. Suppose that F+ Then for each A ∈ Oθ, θ has the little pull-back property. {α(A) : α ∈ Ave(Oθ, U(X0))} k·k ∩ CI 6= ∅. Proof. The proof here is borrowed from [Arc80, Theorem 5]. Let A ∈ ◦Oθ and arbitrarily choose ǫ > 0. It is known that A = Pn An with An ∈ Xn (cf., e.g., [DY09, Rae05]). By Lemma 4.4, there is α1 ∈ Ave(Oθ, U(X0)) such that kα1(A − A0)k < ǫ/2. Also, there is p ∈ N such that α1(A0) ∈ Fp, which is isomorphic to MΠk i=1mp has the Dixmier property (cf., e.g., [Arc80]). So there are λ ∈ C and α2 ∈ Ave(Oθ, U(X0)) such that kα2 ◦ α1(A0 − λ)k < ǫ/2. Thus kα2 ◦ α1(A) − λk < ǫ. Then apply [Arc77, Lemma 2.8] to finish our proof. . But MΠk i=1mp i i Corollary 4.7. Suppose that F+ θ has the little pull-back property. Let A be a unital C-subalgebra of Oθ such that X0 ⊂ A. Then A is simple and has the Dixmier property. Proof. Since X0 ⊂ A, clearly U(X0) ⊂ U(A). Hence the restriction αiA is an averaging mapping on A for every αi's in the proof of Proposi- tion 4.6. It follows from Proposition 4.6 that A has the weak Dixmier property. Therefore, A is simple and has the Dixmier property by [HZ84, Rie82] (also cf. Section 2.4), Corollary 4.8. Suppose that either F+ that the intrinsic group of F+ state. θ has rank 0. Then Oσ θ has the pull-back property, or θ has a unique tracial Proof. If the intrinsic group of F+ holds as Oσ θ = F by Proposition 3.4. θ has rank 0, clearly the conclusion Suppose F+ θ has the little pull-back property. On one hand, it follows from Proposition 3.4 and Corollary 4.7 that Oσ θ has at most one tracial state. On the other hand, it is not hard to check that it has a tracial state, which is extended from the tracial state τ of F. 5. The factoriality and type of π(Oθ)′′ We first recall the definition of KMS-states for C-dynamical systems from [BR97II, Chapter 5]. Let (A, R, ) be a C-dynamical system. FACTORIALITY AND TYPE CLASSIFICATION 19 The state ϕ over A is said to be a -KMS state at value β ∈ R, or a (, β)-KMS state, if ϕ(AB) = ϕ(Biβ(A)) (5) for all A, B entire for ρ. A -KMS state at value β = −1 is simply called a -KMS state. Recently, Huef, Laca, Raeburn and Sims [HLRS13] study in detail the KMS states of some dynamical systems associated to arbitrary finite k-graphs without sources. (Also refer to it for a nice introduction to KMS states.) In particular, it is shown that if {ln mℓ : 1 ≤ ℓ ≤ k} is rationally independent, then Oθ has a unique σ-KMS state ([HLRS13, Theorem 7.2], also see [Yan10] for k = 2). Moreover, they remark in [HLRS13, Example 7.3] that the condition of {ln mℓ : 1 ≤ ℓ ≤ k} being rationally independent is necessary. But one should notice that the example provided there is a periodic single vertex 2-graph. Actually, it is the flip algebra using the terminology of [DPY08, DPY10, DY10]. The following result generalizes [HLRS13, Theorem 7.2] in the case of single vertex k-graphs and [Yan13, Proposition 4.4]. The special case of θ = id is proven in [Yan13] using a more direct method. Proposition 5.1. Suppose that either F+ erty, or the intrinsic group of F+ unique σ-KMS state over Oθ. θ has the little pull-back prop- θ has rank 0. Then the state ω is the Proof. As in [Yan10, Proposition 5.4], it is easy to see that ω is a σ-KMS over Oθ. It is left to show that ω is the only one. If the intrinsic group of F+ We now suppose that F+ θ has rank 0, namely, {ln mℓ : 1 ≤ ℓ ≤ k} is rationally independent, this case is done by [HLRS13, Theorem 7.2]. θ has the little pull-back property. Consider the set {ln mℓ : 1 ≤ ℓ ≤ k}. Assume that it has exactly n rationally independent elements. Clearly 1 ≤ n ≤ k as all mℓ's are > 1. WLOG, say ln m1, . . . , ln mn are rationally independent. Therefore, one can simply rewrite the action σ as follows: σt = γ(m−it 1 ,...,m−it k ) = γ(z1,...,zn, f1,..., fk−n), (6) ℓ where zℓ = m−it (1 ≤ ℓ ≤ n), and fj = fj(z1, . . . , zn) (1 ≤ j ≤ k − n) is a T-valued function of z1, ..., zn (with rational coefficients). Since {ln mℓ : 1 ≤ ℓ ≤ n} is rationally independent, σ yields an action of Tn on Oθ. So there is a faithful conditional expectation E from Oθ onto Oσ θ (cf. [BO08]). From Section 2, it is not hard to see that E is given by E(A) =ZTn γ(z1,...,zn, f1,..., fk−n)(A) dz for all A ∈ Oθ. (7) 20 D. YANG Now we assume that ϕ is also a σ-KMS over Oθ. On one hand, by Corollary 4.8 we have ϕ ◦ E = τ ◦ Ψ ◦ E = ω ◦ E, (8) where Ψ is the canonical faithful expectation of Oσ θ onto F, and τ ◦ Ψ is the unique tracial state of Oσ θ . On the other hand, every σ-KMS sate is invariant with respect to σ ([BR97II]). So ω = ω ◦ σ and ϕ = ϕ ◦ σ. It now follows from (6) and (7) that ϕ = ϕ ◦ E and ω = ω ◦ E. This implies ϕ = ω from (8). Remark 5.2. As in [Yan13, Lemma 4.1], every KMS state over Oθ for σ is a σ-KMS state. One can also borrow the proof from [Yan10, Proposition 5.4 (iii)] in the case of rational independence. Before giving our main result of this paper, we need to introduce more notation. Let M be a σ-finite von Neumann algebra. The Connes spectrum S(M) is the intersection over all faithful normal states of the spectra of their corresponding modular operators [BR97II]. A type III factor M is said to be of type III0 type III1 type IIIλ if S(M) = {0, 1}; if S(M) = [0, ∞); if S(M) = {0, λn : n ∈ Z} (0 < λ < 1). See, for example, [BR97II, KR97II, Tak03II] for more information. Theorem 5.3. Let G be the intrinsic group of F+ θ . (i) If rank G = 0, then π(Oθ)′′ is an AFD type III1 factor. (ii) If F+ θ has the little pull-back property, then π(Oθ)′′ is an AFD type III factor. Moreover, the type of π(Oθ)′′ is determined as follows: (III1) If rank G 6= k − 1, then π(Oθ)′′ is of type III1. (IIIλ) If rank G = k − 1, then π(Oθ)′′ is of type IIIλ, where 0 < λ < 1 is determined by m1, . . . , mk. Proof. By Proposition 5.1 and [BR97II, Theorem 5.3.30], π(Oθ)′′ is factor. Clearly, it is of type III. Since Oθ is known to be amenable, π(Oθ)′′ has to be AFD. Hence π(Oθ)′′ is an AFD factor. We now determine the type of π(Oθ)′′. The following proof is similar to [Yan13, Theorem 5.2]. By Corollary 4.8, π(Oσ θ )′′ is a factor (cf., e.g., [Was91]). It follows from [Str81, Section 28.3] that the Connes spectrum S(π(Oθ)′′) coincides with the spectrum Sp(∆) of the modular operator ∆ of ω. That is, S(π(Oθ)′′) = Sp(∆). FACTORIALITY AND TYPE CLASSIFICATION 21 (III1) Suppose that rank G 6= k − 1. Then ln mi ln mj 6∈ Q for some i, j. So Sp(∆) = cl {mn : n ∈ Zk} i mnj j ⊇ cl {mni = [0, ∞), : ni, nj ∈ Z} where, for a set A ⊆ R, cl A means the closure of A. Hence π(Oθ)′′ is of type III1. (IIIλ) Suppose that rank G = k − 1. Equivalently, one has ln m1 ln mj ∈ Q for 2 ≤ j ≤ k. So there are natural numbers aj and bj (2 ≤ j ≤ k) such that maj j with gcd(ai, bj) = 1 for 2 ≤ j ≤ k. Then 1 = mbj Sp(∆) = cl {mn : n ∈ Zk} 1 Πk j=2m(aj /bj )nj 1 n1+Pk 1 j=2(aj nj/bj ) : n1, ..., nk ∈ Zo : n1, ..., nk ∈ Z(cid:27) j=2 aj njΠℓ6=jbℓ = clnmn1 = cl(cid:26)m = cl((cid:18)m =(0,(cid:18)m − 1 1 b2···bk − 1 1 b2···bk (cid:19)−b2···bkn1−Pk : N ∈ Z) (cid:19)N : n1, ..., nk ∈ Z) − 1 So π(Oθ)′′ is of type IIIλ with λ = m 1 b2···bk . Acknowledgement. The author would like to thank the referee for his/her useful comments. References [Arc77] [Arc79] [Arc80] [BR97II] [BO08] [DPY08] R.J. Archbold, An averaging process for C-algebras related to weighted shifts, Proc. London Math. Soc. 35 (1977), 541 -- 554. R.J. Archbold, On the Dixmier property of certain algebras, Math. Proc. Cambridge Philos. Soc. 86 (1979), 251 -- 259. R.J. Archbold, On the simple C-algebras of J. Cuntz, J. London Math. Soc. 21 (1980), 517 -- 526. O. Bratteli, D.W. Robinson, Operator Algebras and Quantum Statis- tical Mechanics 2, Texts and Monographs in Physics. Springer-Verlag, Berlin, 1997. N.P. Brown and N. Ozawa, C* -- algebras and finite-dimensional ap- proximations. Graduate Studies in Mathematics, 88. American Math- ematical Society, Providence, RI, 2008. K.R. Davidson, S.C. Power and D. Yang, Atomic representations of rank 2 graph algebras, J. Funct. Anal. 255 (2008), 819 -- 853. 22 [DPY10] [DY09] [DY10] [Eva08] [Exe08] [FS02] [FY13] [HZ84] D. YANG K.R. Davidson, S.C. Power and D. Yang, Dilation theory for rank 2 graph algebras, J. Operator Theory 63 (2010), 245 -- 270. K.R. Davidson and D. Yang, Representations of higher rank graph algebras, New York J. Math. 15 (2009), 169 -- 198. K.R. Davidson and D. Yang, Periodicity in rank 2 graph algebras, Canad J. Math. 61 (2010), 1239 -- 1261. G. Evans, On the K-theory of higher rank graph C-algebras, New York J. Math. 14 (2008), 1 -- 31. R. Exel, Inverse semigroups and combinatorial C-algebras, Bull. Braz. Math. Soc. (N.S.) 39 (2008), 191 -- 313. N.L. Fowler and A. Sims, Product systems over right-angled Artin semigroups, Trans. Amer. Math. Soc. 354 (2002), 1487 -- 1509. A.H. Fuller and D. Yang, Nonself-adjoint 2-graph algebras, Trans. Amer. Math. Soc. 367 (2015), 6199 -- 6224. U. Haagerup and L. Zsid´o, Sur la propri´et´e de Dixmier pour les C- alg`ebres, (French) [On the Dixmier property for C-algebras], C. R. Acad. Sci. Paris S´er. I Math. 298 (1984), 173 -- 176. [HRSW13] R. Hazlewood, I. Raeburn, A. Sims and S.B. Webster, Remarks on some fundamental results about higher-rank graphs and their C- algebras, Proc. Edinb. Math. Soc. (2) 56 (2013), 575 -- 597. A. Hopenwasser, The spectral theorem for bimodules in higher rank graph C-algebras, Illinois J. Math. 49 (2005), 993 -- 1000. [Hop05] [Kis81] [Kis96] [KR97II] [HLRS13] A.a. Huef, M. Laca, I. Raeburn and A. Sims , KMS states on C- algebras associated to higher-rank graphs, J. Funct. Anal. 266 (2014), 265 -- 283. R.V. Kadison and J.R. Ringrose, Fundamentals of the theory of oper- ator algebras. Vol. II. Advanced theory. Graduate Studies in Mathe- matics, 16. American Mathematical Society, Providence, RI, 1997. A. Kishimoto, Outer automorphisms and reduced crossed product of simple C-algebras, Comm. Math. Phys. 81 (1981), 429 -- 435. A. Kishimoto, The Rohlin property for shifts on UHF algebras and automorphisms of Cuntz algebras, J. Funct. Anal. 140 (1996), 100 -- 123. D.W. Kribs and S.C. Power, Analytic algebras of higher rank graphs, Math. Proc. Royal Irish Acad. 106A (2006) 199 -- 218. A. Kumjian and D. Pask, Higher rank graph C* -algebras, New York J. Math. 6 (2000), 1 -- 20. A. Kumjian, D. Pask and A. Sims, Generalised morphisms of k-graphs: k-morphs, Trans. Amer. Math. Soc. 363 (2011), 2599 -- 2626. A. Kumjian, D. Pask and A. Sims, Homology for higher-rank graphs and twisted C-algebras, J. Funct. Anal. 263 (2012), 1539 -- 1574. [KPS11] [KP06] [KP00] [KPS12] [LLNSW15] M. Laca, N.S. Larsen, S. Neshveyev, A. Sims, S.B.G. Webster, von Neuman algebras of strongly connected higher-rank graphs, to appear in Math. Ann. [PRRS06] D. Pask, I. Raeburn, M. Rordam and A. Sims, Rank-two graphs whose C-algebras are direct limits of circle algebras, J. Funct. Anal. 239 (2006), 137 -- 178. [Pow07] [Rae05] [RSY03] [RSY04] [Rie82] [RS07] [RS09] [Str81] FACTORIALITY AND TYPE CLASSIFICATION 23 S.C. Power, Classifying higher rank analytic Toeplitz algebras, New York J. Math. 13 (2007), 271 -- 298. I. Raeburn, Graph Algebras, CBMS 103, American Mathematical So- ciety, Providence, RI, 2005. I. Raeburn, A. Sims and T. Yeend, Higher-rank graphs and their C- algebras, Proc. Edinb. Math. Soc. 46 (2003), 99 -- 115. I. Raeburn, A. Sims and T. Yeend, The C-algebras of finitely aligned higher-rank graphs J. Funct. Anal. 213 (2004), 206 -- 240. N. Riedel, On the Dixmier property of simple C-algebras, Math. Proc. Cambridge Philos. Soc. 91 (1982), 75 -- 78. D. Robertson and A. Sims, Simplicity of C-algebras associated to higher-rank graphs, Bull. Lond. Math. Soc. 39 (2007), 337 -- 344. D. Robertson and S. Sims, Simplicity of C-algebras associated to row- finite locally convex higher-rank graphs, Israel J. Math. 172 (2009), 171 -- 192. S. Stratila, Modular theory in operator algebras. Editura Academiei Republicii Socialiste Romnia, Bucharest; Abacus Press, Tunbridge Wells, 1981. [Was91] [Yan10] [Tak03II] M. Takesaki, Theory of operator algebras. II. Encyclopaedia of Math- ematical Sciences, 125. Operator Algebras and Non-commutative Ge- ometry, 6. Springer-Verlag, Berlin, 2003. A. Wassermann, Operators on Hilbert space (1991) (notes). D. Yang, Endomorphisms and modular theory of 2-graph C-algebras, Indiana Univ. Math. J. 59 (2010), 495 -- 520. D. Yang, Type III von Neumann algebras associated with 2-graphs, Bull. Lond. Math. Soc. 44 (2012), 675 -- 686. [Yan13] Dilian Yang, Department of Mathematics & Statistics, University of Windsor, Windsor, ON N9B 3P4, CANADA E-mail address: [email protected]
1210.6066	2	1210	2013-01-14T21:49:02	C-algebras and Equivalences for C-correspondences	[ "math.OA" ]	We study several notions of shift equivalence for C-correspondences and the effect that these equivalences have on the corresponding Pimsner dilations. Among others, we prove that non-degenerate, regular, full C-correspondences which are shift equivalent have strong Morita equivalent Pimsner dilations. We also establish that the converse may not be true. These results settle open problems in the literature. In the context of C-algebras, we prove that if two non-degenerate, regular, full C-correspondences are shift equivalent, then their corresponding Cuntz-Pimsner algebras are strong Morita equivalent. This generalizes results of Cuntz and Krieger and Muhly, Tomforde and Pask. As a consequence, if two subshifts of finite type are eventually conjugate, then their Cuntz-Krieger algebras are strong Morita equivalent. Our results suggest a natural analogue of the Shift Equivalence Problem in the context of C*-correspondences. Even though we do not resolve the general Shift Equivalence Problem, we obtain a positive answer for the class of imprimitivity bimodules.	math.OA	math	C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES EVGENIOS T.A. KAKARIADIS AND ELIAS G. KATSOULIS Dedicated to our mentor and friend Aristides Katavolos Abstract. We study several notions of shift equivalence for C- cor- respondences and the effect that these equivalences have on the corre- sponding Pimsner dilations. Among others, we prove that non- degen- erate, regular, full C∗-correspondences which are shift equivalent have strong Morita equivalent Pimsner dilations. We also establish that the converse may not be true. These results settle open problems in the literature. In the context of C∗-algebras, we prove that if two non-degenera- te, regular, full C- correspondences are shift equivalent, then their cor- responding Cuntz-Pimsner algebras are strong Morita equivalent. This generalizes results of Cuntz and Krieger and Muhly, Tomforde and Pask. As a consequence, if two subshifts of finite type are eventually conjugate, then their Cuntz-Krieger algebras are strong Morita equivalent. Our results suggest a natural analogue of the Shift Equivalence Prob- lem in the context of C-correspondences. Even though we do not re- solve the general Shift Equivalence Problem, we obtain a positive answer for the class of imprimitivity bimodules. 1. Introduction In [21] Williams introduced three relations for the class of matrices with non-negative integer entries, which are successively weaker. Two such ma- trices E and F are said to be elementary strong shift equivalent (symb. s E ∼ F ) if there exist matrices R and S with non-negative integer entries such that E = RS and F = SR. The transitive closure of the elementary SSE ∼ ) and strong shift equivalence is called strong shift equivalence (symb. SE E is shift equivalent to F (symb. E ∼ F ) if there exist matrices R, S with non-negative integer entries such that En = RS, F n = SR and ER = SF , F R = SE for some n ∈ N. The main goal of Williams was to characterize the topological conjugacy of subsifts of finite type using algebraic criteria. In [21], he proved that two subsifts σE and σF of finite type are topologically conjugate if and only if ESSE ∼ are ∼ F . Williams also claimed that the relations SSE ∼ and SE 2010 Mathematics Subject Classification. 47L25, 46L07. Key words and phrases: C∗-correspondences, Shift Equivalence, Morita Equivalence. First author partially supported by the Fields Institute for Research in the Mathemat- ical Sciences. 1 2 E.T.A. KAKARIADIS AND E.G. KATSOULIS equivalent, thus providing a more manageable criterion for the conjugacy of shifts. Unfortunately, an error in [21] made invalid the proof of that last assertion, and the equivalence of SSE ∼ remained an open problem for over than 20 years, known as Williams' Conjecture. The breakthrough came with the work of Kim and Roush [9] who proved that the Williams Conjecture for the class of non-negative integral matrices is false. Their work reshaped the Williams' Conjecture into what is known today as the Shift Equivalence Problem, which for a particular class S of matrices with entries in a certain ring R asks whether SSE ∼ is equivalent to SE ∼ within S. ∼ and SE Williams' notions of shift equivalence carry over to the class of C- corre- spondences if one replaces in the above definitions the matrices E and F with C-correspondences and the multiplication of matrices with the internal ten- sor product. (See Section 4 for the precise definitions.) This introduces three notions of relation between C∗-correspondences, which will be denoted again as s ∼. There exists also a fourth equivalence relation, named strong Morita equivalence (symb. SME ∼ ), which generalizes the concept of unitary conjugacy for matrices to the realm of C∗-correspondences. ∼ and SE ∼, SSE The concept of strong Morita equivalence for C∗-correspondences was first developed and studied by Abadie, Eilers and Exel [1] and Muhly and Solel [16]. Among others these authors show that if E SME ∼ F then the associated Cuntz-Pimsner algebras OE and OF are (strong) Morita equiva- lent as well. The notion of elementary and strong shift equivalence for C∗- correspondences was first studied by Muhly, Tomforde and Pask [14]. These authors also prove that strong shift equivalence of C∗-correspondences im- plies the Morita equivalence of the associated Cuntz-Pimsner algebras, thus extending classical results of Cuntz and Krieger [3], Bates [2] and Drinen and Sieben [5] for graph C∗-algebras. In their study of strong shift equiva- lence Muhly, Tomforde and Pask [14] raise two conjectures, which turn out to be important for the further development of the theory [14, Remark 5.5]: Conjecture 1. Let E and F be two non-degenerate, regular C∗-correspondences and let E∞ and F∞ be their associated Pimsner dilations. If ESSE Conjecture 2. Let E and F be two non-degenerate, regular C∗-correspondences and let E∞ and F∞ be their associated Pimsner dilations. If E∞ SME ∼ F∞, then ESSE ∼ F , then E∞ SME ∼ F∞. ∼ F . The concept of shift equivalence has been studied extensively from both the dynamical and the ring theoretic viewpoint. (See [20] for a comprehen- sive exposition.) In general, shift equivalence has been recognized to be a more manageable invariant than strong shift equivalence, as it is decidable over certain rings [10]. Unlike strong shift equivalence, the study of shift equivalence, from the viewpoint of C∗-correspondences, has been met with C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 3 limited success [13]. been quite successful [12].) (Other operator theoretic viewpoints however have There are three major objectives that are being met in this work. First we complete the study of strong shift equivalence of Muhly, Tomforde and Pask [14] by settling both of their conjectures: with the extra requirement of fullness, Conjecture 1 is settled in the affirmative (see Theorem 5.3 and the remarks preceding it for the fullness condition that has to be considered), while Conjecture 2 has a negative answer (Theorem 5.13). A second objective is the detailed study of the shift equivalence for C∗- correspondences. First, we raise the analogues of Conjectures 1 and 2 for shift equivalence (instead of strong shift equivalence) and we discover that the answers are the same as in the case of strong shift equivalence. Us- ing that information we prove Theorem 5.10, which states that if two non- degenerate, regular, full C-correspondences E and F are shift equivalent, then their corresponding Cuntz-Pimsner algebras OE and OF are (strong) Morita equivalent. This generalizes results of Cuntz and Krieger [3], and Muhly, Pask and Tomforde [14] and appears to be new even for Cuntz- Krieger algebras, where the two notions of shift and strong shift equivalence are known to be different. Combined with the work of Williams our result says that if two subshifts of finite type are eventually conjugate [20], then their Cuntz-Krieger algebras are strong Morita equivalent. Our third goal in this paper is the introduction of the Shift Equivalence Problem in the context of C∗-correspondences. In light of our previous dis- cussion, it seems natural to ask whether strong shift equivalence and shift equivalence are two different notions of equivalence for C∗-correspondences. We coin this problem as the Shift Equivalence Problem for C∗- correspon- dences. The work of Kim and Roush [9] shows that the Shift Equivalence Problem has a negative answer within the class of graph correspondences, but it leaves open the option for a positive answer within the whole class of C∗-correspondences. In general, we do not know the answer even though the work of Kim and Roush hints that it should be negative. In spite of this, we show that the Shift Equivalence Problem has a positive answer for imprimitivity bimodules: all four notions of "equivalence" described in this paper coincide for imprimitivity bimodules, Theorem 6.1. There are more things accomplished in this paper, and we describe each of them within the appropriate sections. Most notably, we settle a third conjecture of Muhly, Pask and Tomforde coming from [14] by showing that [14, Theorem 3.14] is valid without the assumption of non-degeneracy (The- orem 7.1). That is, regular (perhaps degenerate) strong shift equivalent C∗-correspondences have strong Morita equivalent Cuntz-Pimsner algebras. Finally we discuss on the importance of Conjectures 1 and 2 regarding the behaviour of the shift relations when dilating. Pimsner dilations are often the natural objects arising in the C∗-literature for invertible relations, and these latter objects have been under thorough investigation. For example the Pimsner dilation of an injective C∗-dynamical system is an automorphic 4 E.T.A. KAKARIADIS AND E.G. KATSOULIS dynamical system [19], thus the Cuntz-Pimsner algebra is a usual crossed product. (Under this prism our work here can be used in combination to provide further invariants for the Shift Equivalence Problem.) From this point of view one may wonder whether there is a general feature of the Pimsner dilation that makes it unique in a certain sense. In Theorem 3.6 we answer this by showing that the Pimsner dilation X∞ of X is the unique essential Hilbert bimodule that shares the same Cuntz-Pimsner algebra with X. Moreover it is the minimal essential Hilbert bimodule containing X (Theorem 3.5). Further applications of this result are of independent interest and are to be pursued elsewhere. The paper is organized as follows. In Section 2 we establish the notation and terminology to be used throughout this paper. In Section 3 we explore the concept of a dilation for a C∗-correspondence. In addition we prove the minimality of Pimsner dilation. In Sections 4, 5 and 6 we present the main results of this paper. 2. Preliminaries We use [11] as a general reference for Hilbert C∗-modules and C∗- cor- respondences. An inner-product right A-module over a C∗-algebra A is a linear space X which is a right A-module together with an A-valued inner product. For ξ ∈ X we define kξkX := khξ, ξiAk1/2 A . The A-module X will be called a right Hilbert A-module if it is complete with respect to the norm k·kX. In this case X will be denoted by XA. It is straightforward to prove that if (ai) is an approximate unit in A or in the closed ideal hX, XiX, then (ai) is also a right contractive approximate unit (c.a.i.) for X. Dually we call X a left Hilbert A-module if it is complete with respect to the norm induced by a left A-module inner-product [·, ·]X . The term Hilbert module is reserved for the right Hilbert modules, whereas the left case will be clearly stated. Given a Hilbert A-module X over A, let X ∗ = {ξ∗ ∈ L(X, A) ξ∗(ζ) = hξ, ζiX } be the dual left Hilbert A-module, with a · ξ∗ = (ξa∗)∗ and [ξ∗, ζ ∗]X ∗ = hξ, ζiX , for all ξ, ζ ∈ X and a ∈ A. For X, Y Hilbert A-modules let L(X, Y ) be the (closed) linear space of the adjointable maps. For ξ ∈ X and y ∈ Y , let Θy,ξ ∈ L(X, Y ) such that Θy,ξ(ξ′) = y hξ, ξ′iX , for all ξ′ ∈ X. We denote by K(X, Y ) the closed linear subspace of L(X, Y ) spanned by {Θy,ξ : ξ ∈ X, y ∈ Y }. If X = Y then K(X, X) ≡ K(X) is a closed ideal of the C∗-algebra L(X, X) ≡ L(X). Lemma 2.1. Let X, Y, Z be Hilbert A-modules. If hX, XiX provides a right c.a.i. (ai) for Y , then K(X, Y )K(Z, X) = K(Z, Y ). Proof. The existence of the right c.a.i. (ai) implies that Y hX, Xi is dense in Y , hence K(X, Y )K(Z, X) = K(Z, Y hX, Xi) = K(Z, Y ). C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 5 Definition 2.2. An A-B-correspondence X is a right Hilbert B-module together with a ∗-homomorphism φX : A → L(X). We denote this by AXB. When A = B we refer to X as a C∗-correspondence over A. A submodule Y of X is a subcorrespondence of AXB, if it is a C-D- correspondence for some C∗-subalgebras C and D of A and B, respectively. A C∗-correspondence X is called non-degenerate (resp. strict) if the closed linear span of φX(A)X is equal to X (resp. complemented in X). We say that X is full if hX, XiX is dense in A. Finally, X is called regular if both it is injective, i.e., φX is injective, and φX (A) ⊆ K(X). Two A-B-correspondences X and Y are called unitarily equivalent (symb. X ≈ Y ) if there is a unitary u ∈ L(X, Y ) such that u(φX (a)ξb) = φY (a)(uξ)b, for all a ∈ A, b ∈ B, ξ ∈ X. Example 2.3. Every Hilbert A-module X is a K(X)-A-correspondence when endowed with the left multiplication φX ≡ idK(X) : K(X) → L(X). A left inner product over K(X) can be defined by [ξ, η]X = Θξ,η, for all ξ, η ∈ X. Also X ∗ is an A-K(X)-correspondence, when endowed with the following operations hξ∗, η∗iX ∗ = [ξ, η]X , ξ∗ · k = (k∗ξ)∗, and φX ∗ (a)ξ∗ = a · ξ∗ = (ξ · a∗)∗, for all ξ, η ∈ X, k ∈ K(X) and a ∈ A. Example 2.4. For Hilbert A-modules X and Y , L(X, Y ) becomes an L(Y )- L(X)-correspondence by defining hs, ti := s∗t, t · a := ta and b · t := bt, for every s, t ∈ L(X, Y ), a ∈ L(X) and b ∈ L(Y ). Trivially, K(X, Y ) is a K(Y )-K(X)-subcorrespondence of L(X, Y ). Note that, when hX, XiX provides a right c.a.i. for Y , then K(Y ) acts faithfully on K(X, Y ). When X = Y this is automatically true. For two C∗-correspondences AXB and BYC, the interior or stabilized ten- sor product, denoted by X ⊗B Y or simply by X ⊗ Y , is the quotient of the vector space tensor product X ⊗alg Y by the subspace generated by elements of the form ξb ⊗ y − ξ ⊗ φ(b)y, for all ξ ∈ X, y ∈ Y, b ∈ A. It becomes a Hilbert C-module when equipped with (ξ ⊗ y)c := ξ ⊗ (yc), hξ1 ⊗ y1, ξ2 ⊗ y2iX⊗Y := hy1, φ(hξ1, ξ2iX )y2iY , (ξ ∈ X, y ∈ Y, c ∈ C), (ξ1, ξ2 ∈ X, y1, y2 ∈ Y ). For s ∈ L(X) we define s ⊗ idY ∈ L(X ⊗ Y ) be the map ξ ⊗ y 7→ (sξ) ⊗ y. Then X ⊗B Y becomes an A-C-correspondence by defining φX⊗Y (a) := φX(a) ⊗ idY . The interior tensor product plays the role of a generalized associative multiplication of C∗-correspondences and the following lemmas will be useful in the sequel. Lemma 2.5. Let the C∗-correspondences AXB and BYC . If (ci) is an ap- proximate identity of hY, Y iY , then (ci) is a right c.a.i. for the interior tensor product X ⊗B Y . 6 E.T.A. KAKARIADIS AND E.G. KATSOULIS i.e., Proof. The norm on X ⊗B Y is a submultiplicative tensor norm, kξ ⊗ yk ≤ kyk kξk for all ξ ∈ X, y ∈ Y . Thus limi yci = y implies limi(ξ ⊗ y)ci = limi ξ ⊗ (yci) = ξ ⊗ y, for a c.a.i. (ci) as above. Lemma 2.6. Let AXB and BYC be two C∗-correspondences. non-degenerate then X ⊗B Y is non-degenerate. If AXB is Proof. Immediate since the tensor norm is submultiplicative. Lemma 2.7. Let X, Y be Hilbert A-modules and AZB be a regular C∗- correspondence. Then the mapping ⊗idZ : L(X, Y ) → L(X ⊗A Z, Y ⊗A Z) : t 7→ t ⊗ idZ is isometric and maps K(X, Y ) inside K(X ⊗A Z, Y ⊗A Z). Proof. Set Ψ ≡ ⊗idZ : L(X, Y ) → L(X ⊗A Z, Y ⊗A Z), ψ ≡ ⊗idZ : L(X) → L(X ⊗A Z). The exact analogue of [11, Equation 4.6] shows that the mapping Ψ is well defined and contractive. Since Ψ(t1)∗Ψ(t2) = ψ(t∗ 1t2), for all t1, t2 ∈ L(X, Y ), and ψ is isometric [11, Proposition 4.7], we obtain that Ψ is also isometric. Finally, let (si) be a right approximate unit for K(X, Y ) inside K(X). Then by the previous, k ⊗ idZ = limi(k ⊗ idZ ) · ψ(si). However, [11, Proposition 4.7] shows that ψ(si) ∈ K(X ⊗A Z) and the conclusion follows by noting that (k ⊗ idZ ) · K(X ⊗A Z) ⊆ K(X ⊗A Z, Y ⊗A Z). Example 2.8. When a Hilbert A-module X is considered as the K(X)-A- correspondence, then X ⊗A X ∗ ≈ K(X) as C∗-correspondences over K(X), via the mapping u1 : ξ ⊗ ζ ∗ 7→ Θξ,ζ, and X ∗ ⊗K(X) X ≈ hX, XiX, as C∗- correspondences over A, via the mapping u2 : ξ∗ ⊗ ζ 7→ hξ, ζi . In particular X ∗ ⊗K(X) X ≈ A, when X is full. Definition 2.9. A Hilbert A-B-bimodule is a C∗-correspondence AXB to- gether with a left inner product [·, ·]X : X × X → A, which satisfies [ξ, η]X · ζ = ξ · hη, ζiX , (ξ, η, ζ ∈ X). An A-B-imprimitivity bimodule or equivalence bimodule is an A-B- bi- module M which is simultaneously a full left and a full right Hilbert A- module, i.e., [M, M ]M = A and hM, M iM = B. An imprimitivity bimodule AMB is automatically non-degenerate and reg- ular because A ≃φM K(M ). It is immediate that M is an imprimitivity bimodule if and only if M ∗ is a B-A-imprimitivity bimodule and Example 2.8 induces the following. C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 7 Lemma 2.10. If M is an A-B-imprimitivity bimodule, then M ⊗B M ∗ ≈ A and M ∗ ⊗A M ≈ B, where A and B are the trivial C∗-correspondences over themselves. There is a number of ways of considering a direct sum of Hilbert mod- ules and a direct sum of C∗-correspondences; these constructions are sub- correspondences of the matrix C∗-correspondence. Given AEA, ARB, BSA and BFB, the matrix C∗-correspondence X = (cid:20) E R Hilbert (A ⊕ B)-module of the linear space of the "matrices" (cid:20) e S F (cid:21) over A ⊕ B is the s f (cid:21), r e ∈ E, r ∈ R, s ∈ S, f ∈ F , with s (cid:20) e f1 (cid:21) ,(cid:20) e2 s2 r1 r f (cid:21) · (a, b) := (cid:20) ea f2 (cid:21)(cid:29)X r2 rb sa f b (cid:21) , := (cid:0) he1, e2iE + hs1, s2iS , hr1, r2iR + hf1, f2iF (cid:1), s1 (cid:28)(cid:20) e1 and the ∗-homomorphism φ : A ⊕ B → L(cid:18)(cid:20) E R s f (cid:21) := (cid:20) φE(a)e φR(a)r φ(a, b)(cid:20) e φS(b)s φF (b)f (cid:21) . r S F (cid:21)(cid:19) is defined by The E, R, S and F imbed naturally as subcorrespondences in (cid:20) E R S F (cid:21), two interior direct sum C∗-correspondences (cid:20) E since the latter is exactly the exterior direct sum C∗-correspondence of the F (cid:21). S (cid:21) and (cid:20) R The following Lemma explains the use of the terminology "matrix C∗- correspondence", as tensoring is really "matrix multiplication". Lemma 2.11. Let E, F, R, S be C∗-correspondences as above. Then S F (cid:21) ⊗A⊕B (cid:20) E R (cid:20) E R S F (cid:21) ≈ R ⊗B S (cid:21) (cid:20) E ⊗A R (cid:20) E ⊗A E R ⊗B F (cid:21) F ⊗B S (cid:21) (cid:20) S ⊗A R (cid:20) S ⊗A E F ⊗B F (cid:21)   .   Proof. Note that all entries in the second matrix make sense. It is a matter of routine calculations to show that the mapping r1 ⊗ s2 (cid:21) (cid:20) e1 ⊗ r2 r1 ⊗ f2 (cid:21) (cid:20) e1 ⊗ e2 (cid:20) s1 ⊗ e2 f1 ⊗ s2 (cid:21) (cid:20) s1 ⊗ r2 f1 ⊗ f2 (cid:21) .   defines the unitary element that gives the equivalence. (cid:20) e1 s1 f1 (cid:21) ⊗(cid:20) e2 s2 f2 (cid:21) 7→ r1 r2   8 E.T.A. KAKARIADIS AND E.G. KATSOULIS It will be convenient to omit the zero entries, when possible. For example, we write K(E, S) instead of K(cid:18)(cid:20) E 0 0 (cid:21) ,(cid:20) 0 0 S 0 (cid:21)(cid:19). 0 A (Toeplitz ) representation of AXA into a C∗-algebra B, is a pair (π, t), where π : A → B is a ∗-homomorphism and t : X → B is a linear map, such that π(a)t(ξ) = t(φX (a)(ξ)) and t(ξ)∗t(η) = π(hξ, ηiX), for a ∈ A and ξ, η ∈ X. An application of the C∗-identity shows that t(ξ)π(a) = t(ξa) is also valid. A representation (π, t) is said to be injective if π is injective; in that case t is an isometry. The C∗-algebra generated by a representation (π, t) equals the closed lin- ear span of tn(¯ξ)tm(¯η)∗, where for simplicity ¯ξ ≡ ξ1 ⊗ · · · ⊗ ξn ∈ X ⊗n and tn(¯ξ) ≡ t(ξ1) . . . t(ξn). For any representation (π, t) there exists a ∗- homomorphism ψt : K(X) → B, such that ψt(ΘX ξ,η) = t(ξ)t(η)∗. Let J be an ideal in φ−1 X (K(X)); we say that a representation (π, t) is J-coisometric if ψt(φX (a)) = π(a), for any a ∈ J. Following [8], the JX - coisometric representations (π, t), for JX = ker φ⊥ X ∩ φ−1 X (K(X)), are called covariant representations. The Toeplitz-Cuntz-Pimsner algebra TX is the universal C∗-algebra for "all" representations of X, and the Cuntz-Pimsner algebra OX is the univer- sal C∗-algebra for "all" covariant representations of X. The tensor algebra T + X is the norm-closed algebra generated by the universal copy of A and X in TX. If X is an A-B-correspondence, then we may identify X with the (A⊕B)- 0 (cid:21) and thus define the Toeplitz-Cuntz-Pimsner, the correspondence (cid:20) 0 X of the (A ⊕ B)-correspondence (cid:20) 0 X 0 (cid:21). 0 0 Cuntz-Pimsner and the tensor algebra of X as the corresponding algebras If X is a C∗-correspondence, then X ∗ may not be a C∗-correspondence (in the usual right-side sense) but it can be described as follows: if (πu, tu) is the universal representation of AXA, then X ∗ is the closed linear span of t(ξ)∗, ξ ∈ X with the left multiplication and inner product inherited by the correspondence C∗(πu, tu). Nevertheless, X ∗ is an imprimitivity bimodule, whenever X is. Example 2.12. If XG is the C∗-correspondence coming from a graph G, then XG is an imprimitivity bimodule if and only if G is either a cycle or a double-infinite path. In that case, X ∗ G is the correspondence coming from the graph having the same edges as G but its arrows reversed. If α : A → B is a ∗-homomorphism, then the induced correspondence Bα is an imprimitivity bimodule if and only if α is a ∗-isomorphism. In that case, B∗ α = Aα−1. C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 9 3. Dilations of C∗-correspondences For C∗-correspondences AXA and BYB we write X . Y when X is uni- tarily equivalent to a subcorrespondence of Y . In other words there is a ∗-injective representation π : A → B, a π(A)-π(A)- subcorrespondence Y0 of Y and a unitary u ∈ L(X, Y0) such that u : X → Y0, where X is (now) considered a C∗-correspondence over π(A). Here we do not impose u to be an isometry in L(X, Y ), i.e., that u has a complemented range in Y [11, Theorem 3.2]. What we ask is that u is an isometric map and X ≈(π,u) Y0 ⊆ Y , in terms of the representation theory of C∗-correspondences. Definition 3.1. A C∗-correspondence BYB is said to be a dilation of the correspondence AXA, if AXA . BYB and the associated Cuntz-Pimsner algebras OX and OY are ∗-isomorphic. 3.1. Pimsner Dilations of C∗-correspondences. Given an injective C∗- correspondence X there is a natural way to pass to an injective Hilbert bimodule X∞, such that X∞ is a dilation of X. This construction was first introduced by Pimsner in [17]. In [7, Appendix A] we revisited such a con- struction by using direct limits. As it appears both constructions produce the same essential Hilbert bimodule, something that is not proved in [7, Appendix A]. Below we give a brief description (slightly corrected) and the appropriate identification. As we will show, Pimsner dilation is uniquely characterized by a minimal condition, analogously to minimal maximal rep- resentations of non-selfadjoint operator algebras. For Pimsner's dilation as in [17] fix an injective covariant pair (π, t) of X that admits a gauge actions, and let A∞ := Oβ X = E(OX ) ≡ π(A) + {ψtn (K(X ⊗n)) n ≥ 1}. Define X∞ := t(X) · A∞, where "·" denotes the usual multiplication of operators. Then X∞ is an A∞-subcorrespondence of C∗(π, t) and JX∞ = {ψtn(K(X ⊗n)) n ≥ 1}. An approximate identity in ψt(K(X)) is an approx- imate identity of JX∞ , and one can show that JX∞ is essential in A. Even more X∞ is an (essential) Hilbert bimodule with OX∞ ≃ OX . This is mainly [17, Theorem 2.5], when using the identification X∞ ≈ X ⊗A A∞. Note that A∞ and X∞ do not depend on the choice of (π, t). For the dilation as in [7, Appendix A], let the isometric mapping τ : X → L(X, X ⊗2), such that τξ(η) = ξ ⊗ η, and consider the direct limits X τ −→ L(X, X ⊗2) A φX−→ L(X) ⊗idX−→ L(X ⊗2) ⊗idX−→ . . . −→ lim−→(L(X ⊗n, X ⊗n+1), ⊗idX ) =: Y ⊗idX−→ . . . −→ lim−→(L(X ⊗n), ⊗idX ) =: B If r ∈ L(X ⊗n), s ∈ L(X ⊗n, X ⊗n+1) and [r], [s] are their equivalence classes in B and Y respectively, then we define [s] · [r] := [sr]. From this, it is easy to define a right B-action on Y . Similarly, we may define a B-valued right 10 E.T.A. KAKARIADIS AND E.G. KATSOULIS inner product on Y by setting (cid:10)[s′], [s](cid:11)Y ≡ [(s′)∗s] ∈ B. for s, s′ ∈ L(X ⊗n, X ⊗n+1), n ∈ N, and then extending to Y × Y . Finally, we define a ∗-homomorphism φY : B → L(Y ) by setting φY ([r])([s]) ≡ [rs], r ∈ L(X ⊗n), s ∈ L(X ⊗n−1, X ⊗n), n ≥ 0, and extending to all of B by continuity. We therefore have a left B-action on Y and thus Y becomes a C∗-correspondence over B. Let Y0 be the subspace of Y generated by the copies of K(X ⊗n, X ⊗n+1), for n ∈ Z+ and B0 be the C∗-subalgebra of B that is generated by the copies of K(X ⊗n), for n ∈ Z+. Then Y0 is a B0-subcorrespondence of Y , that contains AXA. In particular, when φX (A) ⊆ K(X) then τ (X) ⊆ K(X, X ⊗2) and Lemma 2.7 implies that Y0 = lim −→ (K(X ⊗n, X ⊗n+1), σn) and B0 = lim −→ (K(X ⊗n), ρn), where σ0 = τ , ρ0 = φX and σn = ρn = ⊗idX for n > 1. Let us sketch a proof for the unitary equivalence of Y0 with X∞. Assume first that AXA is regular, so JX = A. Let (π, t) be the covariant Fock representation in L(F(X))/K(F(X)). For k ∈ K(X ⊗n) observe that ψtn(k) = 0 ⊕ · · · ⊕ k ⊕ (k ⊗ idX) ⊕ · · · + K(F(X)). Then the family {ψtn } is compatible with the directed system since ψtn+1(k ⊗ idX) − ψtn (k) = 0 ⊕ · · · ⊕ k ⊕ 0 ⊕ · · · ∈ K(F(X)), (K(X ⊗n), σn) into thus it defines a representation of the direct limit lim −→ A∞. It is straightforward that it is injective and onto. Hence it induces a (K(X ⊗n, X ⊗n+1), σn), since the latter representation of the direct limit lim −→ can be written as the closure of X · lim −→ In order to show that Y0 ≈ X∞ for the general case we have to prove that the family {ψtn } induces a ∗-isomorphism between B0 and A∞. What is required is to prove that {ψtn} is compatible with the equivalence classes in the direct limit; hence to show that if k ∈ K(X ⊗n) with k⊗idX ∈ K(X ⊗n+1) then k ∈ K(X ⊗nJX ). This is derived by a similar argument as in the proof of [8, Proposition 5.9] and [6, Lemma 2.6]. (K(X ⊗n), σn). Theorem 3.2. [17, Theorem 2.5], [7, Theorem 6.6] Let X be an injective C∗-correspondence and let X∞ be the A∞-correspondence constructed above. Then X∞ is an essential Hilbert bimodule and the Cuntz-Pimsner algebras OX∞ and OX coincide. In the sequel, the A∞-correspondence X∞ appearing in the Theorem above will be called the Pimsner dilation of X. When X is non-degenerate and full, then its Pimsner dilation X∞ is also full, hence an imprimitivity bimodule. Indeed, it suffices to prove that K(X ⊗n+1, X ⊗n)K(X ⊗n, X ⊗n+1) = K(X ⊗n), C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 11 for all n ≥ 1. By Lemma 2.1, it suffices to show that (cid:10)X ⊗n+1, X ⊗n+1(cid:11) provides a c.a.i. for X ⊗n. For n = 2, (cid:10)X ⊗2, X ⊗2(cid:11) = hX, φX (hX, Xi)Xi = hX, φX (A)Xi = hX, Xi = A, and an inductive argument completes the claim. However, when X is not full then the previous argument fails (even when X is non-degenerate). For example, let X be the non-degenerate, injective correspondence coming from the infinite tail graph e0 e1 e2 •v0 •v1 •v2 · · · Then hX, Xi = A \ Cδv0, thus X0 := φX (hX, Xi)X = X \ Cξe0, therefore (cid:10)X ⊗2, X ⊗2(cid:11) = hX, X0i = hX0, X0i = A \ span{δv0 , δv1 }. Hence we obtain that X(cid:10)X ⊗2, X ⊗2(cid:11) = X hX0, X0i = X0 ·(cid:0)A \ span{δv0 , δv1 }(cid:1) = X0 X. Example 3.3. Let (A, α) denote a dynamical system where α is a unital ∗-injective endomorphism of A. We can define the direct limit dynamical system (A∞, α∞) by A α A α A α · · · α A α α α / A α / A α / · · · / A∞ α∞ / A∞ The limit map α∞ is an automorphism of A∞ and extends α (note that A imbeds in A∞ since α is injective). Then the A∞-A∞-correspondence Xα∞ , is the Pimsner dilation of Xα. Moreover, Xα∞ satisfies a minimal condition, i.e., A∞ = ∪n∈Zα−n(A) [19]. In analogy, we introduce a "minimality condition" that AXA has as a subcorrespondence of an injective BYB. This will be that X · B is dense in Y . Note that this assumption excludes trivial extensions. Indeed, if there was a submodule Z of Y such that Z ⊥ X, then Z ⊥ (X · B), therefore Z ⊥ Y , hence Z = {0}. Lemma 3.4. Let AXA be an injective sub-correspondence of an injective correspondence BYB such that Y = X · B. Then JX ⊆ JY . Consequently, every covariant pair (π, t) of BYB defines a covariant pair (πA, tX ) for AXA, therefore OX ⊆ OY . Proof. Since X and Y are injective we obtain that JX = φ−1 X (K(X)) and JY = φ−1 Y (K(Y )). Let a ∈ JX . Then φX (a) = k ∈ K(X) and we can view k as an operator in K(Y ) of the same form. Then for every ξ ∈ X and b ∈ B we obtain that φY (a)(ξ · b) = φY (a)(ξ)b = φX(a)(ξ)b = k(ξ)b = k(ξ · b). By assumption X · B is dense in Y , hence φY (a)(y) = k(y) for all y ∈ Y . Therefore φY (a) = k ∈ K(Y ), which implies that a ∈ JY . t t t t t t / / / / / / / / / / / 12 E.T.A. KAKARIADIS AND E.G. KATSOULIS For the second part of the proof, let (π, t) be a covariant pair for Y and note that ψtX (k) = ψt(k) for every k ∈ K(X) ⊆ K(Y ). Indeed, for ξ, η ∈ X, ψtX (ΘX ξ,η) = tX(ξ)tX (η)∗ = t(ξ)t(η)∗ = ψt(ΘY ξ,η). Thus for a ∈ JX we get πA(a) = π(a) = ψt(φY (a)) = ψt(k) = ψtX (k) = ψtX (φX (a)) therefore (πA, tX ) is a covariant pair for X. Moreover, a gauge action {βz} for (π, t) is inherited to (πA, tX ) and the proof is complete. Theorem 3.5. The Pimsner dilation X∞ of an injective correspondence X is a minimal Hilbert bimodule that contains X as a subcorrespondence. Proof. Let BYB be a Hilbert bimodule such that X (cid:22) Y (cid:22) X∞. Up to unitary equivalence, we can assume that X ⊆ Y ⊆ X∞ as subcorrespon- dences, hence A ⊆ B ⊆ A∞ as subalgebras. In particular φ∞Y = φY and K(X) ⊆ K(Y ) ⊆ K(X∞). It suffices to prove that B = A∞, because then X∞ = X · A∞ ⊆ Y · A∞ = Y · B = Y ⊆ X∞, hence Y = X∞, where the symbol "·" denotes the right action for X∞. Claim. The ideal JY of B is contained in the ideal J∞ of A∞. Proof of Claim. Recall that X∞ is injective, thus J∞ = φ−1 ∞ (K(X∞)). Let b ∈ JY = φ−1 Y ; in particular there is a kY ∈ K(Y ) such that φY (b) = kY . The compact operator kY can be also viewed as a compact operator k∞ in K(X∞) (of the same form) and it suffices to show that φ∞(b) = k∞. To this end let an element η ∈ X∞; by the form of X∞ we can assume that η = ξ · a, for some ξ ∈ X and a ∈ A∞. Then Y (K(Y )) ∩ ker φ⊥ φ∞(b)(ξ · a) = φ∞(b)(ξ) · a = φ∞(b)Y (ξ) · a = φY (b)(ξ) · a = kY (ξ) · a = k∞(ξ) · a = k∞(ξ · a), where we use the fact that ξ ∈ Y . Fix an injective covariant pair (π, t) of X∞ that admits a gauge action. Since JX , JY ⊆ J∞, then OX = C∗(πA, tX ), OY = C∗(πB, tY ) and OX∞ = C∗(π, t). Therefore π(A) ⊆ π(B) ⊆ π(A∞) = π(A) + span{ψ(tX )n(K(X ⊗n)) n ≥ 1}. Showing that ψ(tX )n(K(X ⊗n)) ⊆ π(B), for every n ≥ 1, will complete the proof. By assumption Y is a Hilbert bimodule, therefore ψtY (K(Y ⊗n)) ⊆ π(B) for all n ≥ 1. But then ψ(tX )n (K(X ⊗n)) = ψtn(K(X ⊗n)) ⊆ ψtn(K(Y ⊗n)) = ψ(tY )n(K(Y ⊗n)) ⊆ π(B), for every n ≥ 1, which completes the proof. Theorem 3.6. Let AXA be an injective correspondence. Then the Pimsner dilation is contained in any correspondence BYB that satisfies: C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 13 (1) X is a subcorrespondence of Y , (2) Y is an essential Hilbert bimodule, (3) Y = X · B, where "·" denotes the right action in Y that extends the right action of A. Proof. As proved, X∞ is such a correspondence. In order to show that X∞ is a subcorrespondence of a Y that satisfies the above, it suffices to prove that A∞ ⊆ B. Indeed, then X∞ = X · A∞ ⊆ X · B = Y, where "·" denotes the right action in Y . Fix an injective covariant pair (π, t) of Y that admits a gauge action. Then (πA, tX ) is an injective covariant pair for X that admits a gauge action by Lemma 3.4, and A∞ ≃ π(A) + span{ψ(tX )n(K(X ⊗n)) n ≥ 1}. Since Y is a Hilbert bimodule, then φY (JY ) = K(Y ). Thus ψtX (K(X)) = ψt(K(X)) ⊆ ψt(K(Y )) = ψt(φY (JY )) = π(JY ) ⊆ π(B). Moreover, ψ(tX )2(K(X ⊗2)) ⊆ t(X)ψtX (K(X))t(X)∗ ⊆ t(X)π(B)t(X)∗ = t(X · B)t(X)∗ ⊆ t(Y )t(Y )∗ = ψt(K(Y )) ⊆ π(B). Inductively, we get that ψ(tX )n(K(X ⊗n)) ⊆ π(B), for all n ≥ 1. Trivially π(A) ⊆ π(B) since X is a subcorrespondence of Y , therefore A∞ is (∗- isomorphic to) a subalgebra of π(B), hence A∞ ⊆ B. Moreover the Pimsner dilation of an injective C∗-correspondence is the unique minimal dilation of X. Theorem 3.7. Let AXA be an injective subcorrespondence of some BYB. Then Y ≈ X∞ if and only if (1) Y is an essential Hilbert bimodule, (2) Y = X · B, (3) OY = OX . Then YB is minimal as a Hilbert bimodule containing XA, automatically. Proof. As in the proof of Theorem 7.1 it suffices to show that B ≃ A∞. To this end fix an injective covariant pair (π, t) of Y that admits a gauge action. Then (πA, tX ) is an injective covariant pair of X that inherits the gauge action of (π, t) by restriction, and we can assume that C∗(πA, tX) = OX = OY = C∗(π, t). Therefore A∞ is ∗-isomorphic to the fixed point algebra Oβ X . Since OX = OY and YB is a Hilbert bimodule we obtain A∞ ≃ Oβ X = E(OX ) = E(OY ) = π(B). Therefore A∞ ≃ B and the proof is complete. 14 E.T.A. KAKARIADIS AND E.G. KATSOULIS 3.2. Dilations for Compacts. The construction of the Pimsner dilation applies to more general settings. Indeed, let X, Y be Hilbert A-modules and let AZB be a regular correspondence. By Lemma 2.7, one can form the following directed systems K(X, Y ) ⊗idZ−→ K(X ⊗ Z, Y ⊗ Z) K(X) ⊗idZ−→ K(X ⊗ Z) K(Y ) ⊗idZ−→ K(Y ⊗ Z) ⊗idZ−→ . . . −→ lim−→(K(X ⊗ Z⊗n, Y ⊗ Z⊗n), ⊗idZ) ⊗idZ−→ . . . −→ lim−→(K(X ⊗ Z⊗n), ⊗idZ) ⊗idZ−→ . . . −→ lim−→(K(Y ⊗ Z⊗n), ⊗idZ ). For simplicity, we will write (K(X, Y ), idZ)∞ for the direct limit lim −→(cid:0)K(X ⊗ Z ⊗n, Y ⊗ Z ⊗n), idZ(cid:1) . By imitating the proofs in [7, Appendix A] and Lemma 2.7, we see that (K(X, Y ), idZ )∞ is a K(Y )∞-K(X)∞-correspondence, which will be called the dilation of K(X, Y ) by Z, or more simply the Z-dilation of K(X, Y ). If, in addition, hX ⊗ Z ⊗n, X ⊗ Z ⊗ni provides a right c.a.i. for Y ⊗ Z ⊗n, for every n ≥ 0, then by Example 2.4 and Lemma 2.1 (K(X, Y ), idZ )∞ is a regular C∗-correspondence. In the case of a regular C∗-correspondence AXA, the Pimsner dilation X∞ is simply the X-dilation of K(X, X ⊗2), which is always regular. Proposition 3.8. Let X, Y be Hilbert A-modules and Z, W be regular C∗- correspondences over A. If Z ⊗A W ≈ W ⊗A Z, then (K(X, Y ), idZ )∞ . (K(X ⊗A W, Y ⊗A W ), idZ )∞. Proof. We will identify Z ⊗ W with W ⊗ Z. Since Z commutes with W the diagram K(X, Y ) idK(X,Y ) / K(X, Y ) ⊗idZ / K(X ⊗ Z, Y ⊗ Z) ⊗idZ / . . . idK(X,Y ) ⊗idW ⊗idW K(X, Y ) ⊗idW / / K(X ⊗ W, Y ⊗ W ) ⊗idZ / / K(X ⊗ W ⊗ Z, Y ⊗ W ⊗ Z) ⊗idZ / / . . . is commutative and defines a linear map s : (K(X, Y ), idZ)∞ → (K(X ⊗A W, Y ⊗A W ), idZ )∞. That is, if [k] ∈ (K(X, Y ), idZ )∞, such that k ∈ K(X ⊗ Z ⊗n, Y ⊗ Z ⊗n) then k ⊗ idW ∈ K(X ⊗ Z ⊗n ⊗ W, Y ⊗ Z ⊗n ⊗ W ) = K(X ⊗ W ⊗ Z ⊗n, Y ⊗ Z ⊗n), and we define s[k] = [k ⊗ idW ]. Note that s is defined on (K(X), idZ )∞ and on (K(Y ), idZ )∞ in the analogous way. For example we get the commutative diagram idK(X) K(X) / K(X) ⊗idZ / K(X ⊗ Z) ⊗idZ / / . . . idK(X) ⊗idW ⊗idW K(X) ⊗idW/ / K(X ⊗ W ) ⊗idZ / / K(X ⊗ W ⊗ Z) ⊗idZ / / . . . / / / / / C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 15 It is a matter of routine computations to show that s is a (K(Y ), idZ )∞- (K(X), idZ )∞-mapping. What is left to show is that s is also a unitary onto its range. It suffices to prove this at the n-th level. For k1, k2 ∈ K(X ⊗ Z ⊗n, Y ⊗ Z ⊗n), then hs[k1], s[k2]i = h[k1 ⊗ idW ], [k2 ⊗ idW ]i = [k1 ⊗ idW ]∗[k2 ⊗ idW ] = [(k∗ 1 ◦ k2) ⊗ idW ] = [(k∗ 1 ◦ k2)] = h[k1], [k2]i , where we have identified (K(X), idZ)∞ with its image via s. For the proof of the following Proposition, recall that there is a major difference between isometric mappings of C∗-correspondences and mappings of C∗-correspondences that are isometries (of Hilbert modules). However, when an isometric mapping is also onto then it is a unitary between Hilbert modules. Proposition 3.9. Let X, Y, W be Hilbert A-modules and Z be a regular C∗- correspondence over A. If the ideal hX ⊗ Z ⊗n, X ⊗ Z ⊗ni of A provides a right c.a.i. for Y ⊗ Z ⊗n, for all n ≥ 0, then (K(X, Y ), idZ)∞ ⊗(K(X),idZ )∞ (K(W, X), idZ )∞ ≈ (K(W, Y ), idZ )∞. Proof. In order to "visualize" the proof, imagine that we multiply "verti- cally" and term by term the correspondences K(X, Y ) K(W, X) ⊗idZ−→ K(X ⊗A Z, Y ⊗A Z) ⊗idZ−→ K(W ⊗A Z, X ⊗A Z) ⊗idZ−→ . . . −→ (K(X, Y ), idZ )∞ ⊗idZ−→ . . . −→ (K(W, X), idZ )∞ in the order K(X ⊗A Z ⊗n, Y ⊗A Z ⊗n) · K(W ⊗A Z ⊗n, X ⊗A Z ⊗n). We consider K(X ⊗Z ⊗n), K(Y ⊗Z ⊗n) and K(W ⊗Z ⊗n) as C∗-subalgebras of (K(X), idZ)∞, (K(Y ), idZ )∞ and (K(W ), idZ )∞, respectively, for every n ∈ Z+. The proof is divided into two parts. For the first part of the proof note that for every n ∈ Z+, the Banach space K(X ⊗A Z ⊗n, Y ⊗A Z ⊗n) (resp. K(W ⊗A Z ⊗n, X ⊗A Z ⊗n)) is an injective K(Y ⊗A Z ⊗n)-K(X ⊗A Z ⊗n)-correspondence (resp. a K(X ⊗A Z ⊗n)- K(W ⊗A Z ⊗n)-correspondence) in the obvious way. Set An := K(X ⊗A Z ⊗n, Y ⊗A Z ⊗n) ⊗K(X⊗AZ⊗n) K(W ⊗A Z ⊗n, X ⊗A Z ⊗n) For every n ∈ Z+, the mapping ρn : An → An+1 : k ⊗ t 7→ (k ⊗ idZ ) ⊗ (t ⊗ idZ ), is isometric on sums of elementary tensors and therefore extends to the clo- sures. Thus An . An+1, and we can form the direct limit C∗-correspondence A0 ρ0−→ A1 ρ1−→ A2 ρ2−→ . . . −→ lim −→ (An, ρn), which is a (K(Y ), idZ )∞-(K(W ), idZ )∞-correspondence. Now the mappings φn : An → (K(X, Y ), idZ )∞ ⊗(K(X),idZ )∞ (K(W, X), idZ)∞ : k ⊗ t 7→ [k] ⊗ [t] 16 E.T.A. KAKARIADIS AND E.G. KATSOULIS are compatible with the directed system and define a C∗-correspondence mapping φ : lim −→ (An, ρn) → (K(X, Y ), idZ )∞ ⊗(K(X),idZ )∞ (K(W, X), idZ )∞, which is isometric since every φn+1 is. Thus lim −→ (An, ρn) . (K(X, Y ), idZ)∞ ⊗(K(X),idZ )∞ (K(W, X), idZ)∞, via φ. However (K(X, Y ), idZ)∞ ⊗(K(X),idZ )∞ (K(W, X), idZ)∞ is spanned by the elements [k] ⊗ [t] for k ⊗ t ∈ An. Therefore φ is onto and so lim −→ (An, ρn) ≈ (K(X, Y ), idZ )∞ ⊗(K(X),idZ )∞ (K(W, X), idZ )∞ For the second part of the proof we construct an isometric map from (An, ρn) onto (K(W, Y ), idZ)∞. Start by defining the maps lim −→ un : An → (K(W, Y ), idZ )∞ : k ⊗ t 7→ [kt]. These are well defined since kt ∈ K(W ⊗ Z ⊗n, Y ⊗ Z ⊗n) and un((k · a) ⊗ t) = [(k · a)t] = [k(a · t)] = un(k ⊗ (a · t)), due to the associativity of the multiplication. Also, the maps un are iso- metric and compatible with the direct sequence. Therefore the family {un} defines an isometric map u from the direct limit into (K(W, Y ), idZ )∞, which extends to a mapping of C∗-correspondences. Thus An . (K(W, Y ), idZ)∞ and so lim (An, ρn) . (K(W, Y ), idZ)∞. Finally note that the assumption −→ that the ideals hX ⊗ Z ⊗n, X ⊗ Z ⊗ni provide a c.a.i. for each Y ⊗ Z ⊗n, com- bined with Lemma 2.1, implies that the isometric maps un are onto and hence u is onto. 4. Relations associated with C∗-correspondences In [16] Muhly and Solel introduced the notion of Morita equivalence of C∗-correspondences. This concept generalizes the notion of outer conjugacy for C∗-dynamical systems [16, Proposition 2.4]. Definition 4.1. The C∗-correspondences AEA and BFB are called strong Morita equivalent if there is an imprimitivity bimodule AMB such that E ⊗A M ≈ M ⊗B F . In that case we write E SME ∼ F . Muhly and Solel [16] examined this relation under the assumption that the C∗-correspondences are both non-degenerate and injective. Neverthe- SME less, non-degeneracy is automatically implied. Indeed, if E ∼ F via M , then E ≈ E ⊗A A ≈ E ⊗A M ⊗B M ∗ ≈ M ⊗B F ⊗B M ∗ ≈ M ⊗B (F ⊗B M ∗), and Lemma 2.6 implies that E is non-degenerate. A symmetrical argument applies for F . C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 17 Remark 4.2. In contrast to non-degeneracy, injectivity is not automatically implied by SME ∼ . For example, pick your favorite non-degenerate and non- injective C∗-correspondence AEA and let E = F . Then E SME ∼ E via the trivial imprimitivity bimodule AAA, but both E and F are not injective. In [14] Muhly, Pask and Tomforde introduced the notion of elementary strong shift equivalence between C∗-correspondences that generalizes the corresponding notion for graphs. Definition 4.3. Let AEA and BFB be C∗-correspondences. Then E and F s will be called elementary strong shift equivalent (symb. E ∼ F ) if there are C∗-correspondences ARB and BSA such that E ≈ R ⊗B S and F ≈ S ⊗A R as C∗-correspondences. Elementary strong shift equivalence is obviously symmetric. Moreover it is reflexive for non-degenerate C∗-correspondences; indeed, E s ∼ E via E and A. But, it may not be transitive as [21, Example 2] shows1. Nevertheless, we have the following proposition. Proposition 4.4. Let AEA, BFB and CGC be C∗-correspondences. Assume s that E ∼ G via T, Z. If either Z or R is an imprimitivity bimodule, then E s ∼ F via R, S and F s ∼ G. Proof. For E, F and G as above we have that E ≈ R ⊗B S , F ≈ S ⊗A R , F ≈ T ⊗C Z , G ≈ Z ⊗B T. Assume that Z is an imprimitivity bimodule (a symmetric argument can be used if R is an imprimitivity bimodule). Then by Lemma 2.10 Z ∗ ⊗C Z ≈ B and Z ⊗B Z ∗ ≈ C. Hence, (Z ⊗B S) ⊗A (R ⊗B Z ∗) ≈ Z ⊗B (S ⊗A R) ⊗B Z ∗ ≈ Z ⊗B F ⊗B Z ∗ ≈ Z ⊗B T ⊗C Z ⊗B Z ∗ ≈ Z ⊗B T ⊗C C ≈ Z ⊗B (T · C) = Z ⊗B T ≈ G. On the other hand, (R ⊗B Z ∗) ⊗C (Z ⊗B S) ≈ R ⊗B B ⊗B S = (R · B) ⊗B S = R ⊗B S ≈ E, hence E s ∼ G, which completes the proof. Following Williams [21] we denote by SSE ∼ the transitive closure of the ∼ F if there are n C∗-correspondences Ti, i = 0, . . . , n s ∼. That is, ESSE relation such that T0 = E, Tn = F and Ti s ∼ Ti+1. There is also another relation between C∗-correspondences inspired by Williams' work [21]. 1 Note that [21, Example 2] shows that s of non-negative integral matrices, which doesn't mean that s whole class of C-correspondences. ∼ is not transitive when restricted to the class ∼ is not transitive for the 18 E.T.A. KAKARIADIS AND E.G. KATSOULIS Definition 4.5. Let AEA and BFB be C∗-correspondences. Then E and F will be called shift equivalent (symb. E SE ∼ F ) if there are C∗-correspondences ARB and BSA and a natural number m so that (i) E⊗m ≈ R ⊗A S, F ⊗m ≈ S ⊗B R, (ii) S ⊗A E ≈ F ⊗ S, E ⊗A R ≈ R ⊗B F. The number m is called the lag of the equivalence. If E SE ∼ F with lag m, then by replacing S with S ⊗A E⊗k we obtain another shift equivalence of lag m + k, i.e., there is a shift equivalence of lag L for every L ≥ m. In contrast to elementary strong shift equivalence, shift equivalence is an equivalence relation for the class of arbitrary C∗-correspondences. Proposition 4.6. Shift equivalence is an equivalence relation. Proof. Note that E shift equivalence is symmetric. If E with lag n via V, U , then E SE ∼ E with lag 2 and R = S = E, and it is clear that SE ∼ G SE ∼ F with lag m via R, S and F SE ∼ G with lag mn + m via V is repeated m times {z which shows that SE ∼ is transitive. R ⊗ V ⊗ U ⊗ · · · ⊗ V and U ⊗ S, } Theorem 4.7. Let AEA and BFB be C∗-correspondences. Then E SME ∼ F ⇒ E s ∼ F ⇒ E SSE ∼ F ⇒ E SE ∼ F. Proof. Recall that when E E ⊗ M and S ≡ M ∗. By Lemma 2.11, SME ∼ F then E, F are non-degenerate. Let R ≡ R ⊗B S ≈ E ⊗A M ⊗B M ∗ ≈ E ⊗A A ≈ E, and S ⊗A R ≈ M ∗ ⊗A E ⊗A M ≈ M ∗ ⊗A M ⊗B F ≈ B ⊗B F ≈ F. ∼. Trivially s ∼ implies SSE ∼ implies s ∼ F , i.e., there are Ti, i = 0, . . . , n, such that Ti Hence SME that ESSE T0 = E and Tn = F . Then one can directly verify that E SE via R = R1 ⊗ · · · ⊗ Rn and S = Sn ⊗ · · · ⊗ S1. ∼ . To complete the proof assume s ∼ Ti+1, where ∼ F with lag n, Muhly, Pask and Tomforde [14] provide a number of examples to show that strong Morita equivalence differs from elementary strong shift equiva- lence. In Theorem 4.7 above we prove that in fact it is stronger. Neverthe- less, the following result shows that under certain circumstances, the two notions coincide. C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 19 Proposition 4.8. Let AEA and BFB be C∗-correspondences and assume s that E ∼ F via R, S. If either R or S is an imprimitivity bimodule, then SME E ∼ F . Proof. When E (analogously S ⊗B E ≈ F ⊗B S). s ∼ F via R, S then E ⊗A R ≈ R ⊗B S ⊗A R ≈ R ⊗B F 5. Passing to Pimsner Dilations In this section we show that if ∼ is any of the four relations defined in the previous section, then E ∼ F implies E∞ ∼ F∞, under the standing hypothesis that E and F are regular C∗-correspondences. Our study is based on the concept of a bipartite inflation, an insightful construct originating in the work of Muhly, Pask and Tomforde [14]. Assume that AEA and BFB are C∗-correspondences which are elementary strong shift equivalent via ARB and BSA. Let X = (cid:20) 0 R S 0 (cid:21) be the bipartite inflation of S by R. By Lemma 2.11, we obtain2 X ⊗2 = (cid:20) R ⊗B S 0 By induction and Lemma 2.11, 0 S ⊗A R (cid:21) = (cid:20) E 0 0 F (cid:21) . X ⊗2k = (cid:20) E⊗k 0 0 F ⊗k (cid:21) , X ⊗2k+1 = (cid:20) 0 F ⊗k ⊗B S E⊗k ⊗A R 0 (cid:21) , for all k ∈ Z+. In particular, if E and F are regular (resp. non-degenerate) then R and S, and consequently X and X ⊗2, are regular (resp. non- degenerate) as shown in [14]. Proposition 5.1. Let E ∼ F via R, S. Then, for k ∈ Z+, s (1) (cid:10)E⊗k ⊗ R, E⊗k ⊗ R(cid:11) provides a right c.a.i. for F ⊗k+1, (2) (cid:10)F ⊗k ⊗ S, F ⊗k ⊗ S(cid:11) provides a right c.a.i. for E⊗k+1. Lemma 2.5 we have that (cid:10)X ⊗2k+1, X ⊗2k+1(cid:11) provides a right c.a.i. Proof. Let X be the bipartite inflation of S by R and let k ∈ Z+. By for X ⊗2k+2 = X ⊗ X ⊗2k+1. However, X ⊗2k+2 = (cid:20) E⊗k+1 0 0 F ⊗k+1 (cid:21) 2 In order to ease notation, unitary equivalence of C∗-correspondences will be simply denoted as equality in this section. 20 and E.T.A. KAKARIADIS AND E.G. KATSOULIS DX ⊗2k+1, X ⊗2k+1E = = (cid:28)(cid:20) = DF ⊗k ⊗ S, F ⊗k ⊗ SE ⊕DE⊗k ⊗ R, E⊗k ⊗ RE . E⊗k ⊗A R F ⊗k ⊗B S F ⊗k ⊗B S (cid:21) ,(cid:20) 0 0 0 E⊗k ⊗A R 0 (cid:21)(cid:29) This completes the proof. We will make use of the dilation X∞ of the bipartite inflation of S by R. Let (A ⊕ B)∞ be the direct limit C∗-algebra of the following directed system (cid:20) A 0 0 B (cid:21) ρ0−→ K(cid:18)(cid:20) 0 R 0 (cid:21)(cid:19) ρ1−→ K(cid:18)(cid:20) E 0 S 0 F (cid:21)(cid:19) ρ2−→ (cid:21)(cid:19) ρ3−→ K(cid:18)(cid:20) E⊗2 0 0 F ⊗2 (cid:21)(cid:19) ρ4−→ · · · , ρ2−→ K(cid:18)(cid:20) 0 E ⊗ R F ⊗ S 0 where ρ0 = φX : A = L(A) −→ L(X), ρn = ⊗idX : L(X ⊗n) −→ L(X ⊗n+1) : r 7−→ r ⊗ idX , n ≥ 1. Note that, since X and X ⊗2 are regular, (A ⊕ B)∞ is ∗-isomorphic to the direct limit C∗-algebra of the following directed system K(cid:18)(cid:20) A 0 0 B (cid:21)(cid:19) ρ1◦ρ0−→ K(cid:18)(cid:20) E 0 0 F (cid:21)(cid:19) ρ3◦ρ2−→ K(cid:18)(cid:20) E⊗2 0 0 F ⊗2 (cid:21)(cid:19) ρ5◦ρ4−→ · · · . Since E and F are orthogonal subcorrespondences of X ⊗2 we get that (A ⊕ B)∞ = A∞ ⊕ B∞ and we can write A∞ and B∞ via the following directed systems A∞ : K(cid:18)(cid:20) A 0 B∞ : K(cid:18)(cid:20) 0 0 0 (cid:21)(cid:19) ρ1◦ρ0−→ K(cid:18)(cid:20) E 0 0 (cid:21)(cid:19) ρ3 ◦ρ2−→ K(cid:18)(cid:20) E⊗2 0 0 0 B (cid:21)(cid:19) ρ1 ◦ρ0−→ K(cid:18)(cid:20) 0 0 F (cid:21)(cid:19) ρ3◦ρ2−→ K(cid:18)(cid:20) 0 0 0 0 0 (cid:21)(cid:19) ρ5 ◦ρ4−→ · · · , 0 F ⊗2 (cid:21)(cid:19) ρ5 ◦ρ4−→ · · · , 0 The C∗-correspondence X∞ is defined by the directed system S S 0 (cid:21) τ (cid:20) 0 R −→ K(cid:18)(cid:20) 0 R σ1−→ K(cid:18)(cid:20) E 0 σ2−→ K(cid:18)(cid:20) where τξ(η) = ξ ⊗ η, and F ⊗ S 0 0 (cid:21) ,(cid:20) E 0 0 F (cid:21)(cid:19) σ1−→ 0 F (cid:21) ,(cid:20) 0 E ⊗ R F ⊗ S 0 E ⊗ R 0 (cid:21) ,(cid:20) E⊗2 0 (cid:21)(cid:19) σ2−→ F ⊗2 (cid:21)(cid:19) σ3−→ · · · 0 σn = ⊗idX : L(X ⊗n, X ⊗n+1) → L(X ⊗n+1, X ⊗n+2) : s 7→ s ⊗ idX , n ≥ 1, Let R∞ be the subcorrespondence of X∞ that is generated by the copies of R , K(cid:18)(cid:20) 0 0 F ⊗k (cid:21) ,(cid:20) 0 E⊗k ⊗ R 0 0 0 (cid:21)(cid:19) , C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 21 and K(cid:18)(cid:20) 0 0 S 0 (cid:21) ,(cid:20) E 0 0 (cid:21)(cid:19) , K(cid:18)(cid:20) 0 0 F ⊗k ⊗ S 0 (cid:21) ,(cid:20) E⊗k+1 0 0 (cid:21)(cid:19) , 0 0 for k ≥ 1. Because of Lemma 2.7, R∞ can be written alternatively as a direct limit in the following two forms (i) R τ ◦(⊗idX ) −→ K (F, E ⊗ R) (ii) R τ−→ K (S, E) ⊗idX⊗2 −→ K(cid:0)F ⊗2, E⊗2 ⊗ R(cid:1) ⊗idX⊗2 −→ · · · , ⊗idX⊗2 −→ K(cid:0)F ⊗ S, E⊗2(cid:1) ⊗idX⊗2 −→ · · · , where we omit the zero entries for convenience. The first form of R∞ shows that it is a Hilbert B∞-module, whereas the second form shows that the left multiplication by elements of A∞ defines a left action of A∞ on R∞. Hence R∞ is a A∞-B∞-correspondence. In a dual way we define S∞ as the subcorrespondence of X∞ generated by the copies of S , K(cid:18)(cid:20) E⊗k 0 0 (cid:21) ,(cid:20) 0 0 F ⊗k ⊗ S (cid:21)(cid:19) 0 0 and K(cid:18)(cid:20) 0 R 0 (cid:21) ,(cid:20) 0 0 F (cid:21)(cid:19) , K(cid:18)(cid:20) 0 E⊗k ⊗ R 0 0 0 0 (cid:21) ,(cid:20) 0 0 F ⊗k+1 (cid:21)(cid:19) , 0 for k ≥ 1. By using Lemma 2.7 and writing S∞ as a direct limit in the following two forms (i) S τ ◦(⊗idX ) −→ K (E, F ⊗ S) (ii) S τ−→ K (R, F ) ⊗idX⊗2 −→ K(cid:0)E⊗2, F ⊗2 ⊗ S(cid:1) ⊗idX⊗2 −→ · · · , ⊗idX⊗2 −→ K(cid:0)E ⊗ R, F ⊗2(cid:1) ⊗idX⊗2 −→ · · · , we get that S∞ becomes a B∞-A∞-correspondence. We must remark here on our use of the subscript ∞. For the correspon- dences E, F we denote by E∞, F∞ their Pimsner dilations, whereas for R, S we denote by R∞, S∞ the subcorrespondences in the Pimsner dilation X∞ of the (injective) bipartite inflation X of S by R. Nevertheless, when the correspondences are non-degenerate, R∞ is the X-dilation (K(B, R), idX)∞ and S∞ is the X-dilation (K(A, S), idX)∞. Remark 5.2. Note that R∞ and S∞ are both full left Hilbert bimodules (thus injective). In order to prove this for, say S∞, it is enough (by its second form) to show that K(E⊗n ⊗ R, F ⊗n+1)K(E⊗n ⊗ R, F ⊗n+1)∗ = K(F ⊗n+1, F ⊗n+1) for all n ≥ 1. By Lemma 2.1 it suffices to show that hE⊗n ⊗ R, E⊗n ⊗ Ri contains a right c.a.i. for F ⊗n+1, for all n ≥ 1, a fact established in Propo- sition 5.1. A symmetrical argument can be used for R∞. 22 E.T.A. KAKARIADIS AND E.G. KATSOULIS Moreover, if S is an imprimitivity bimodule, then S∞ is also right full, hence an imprimitivity bimodule as well. Indeed, by the first form of S∞ it suffices to show that K(E⊗n, F ⊗n ⊗ S)∗K(E⊗n, F ⊗n ⊗ S) = K(E⊗n, E⊗n), for all n. By Lemma 2.1, it is enough to show that hF ⊗n ⊗ S, F ⊗n ⊗ Si provides a right c.a.i. for E⊗n. Note that F ⊗ S = S ⊗ E, hence F ⊗n ⊗ S = S ⊗ E⊗n, thus E⊗n is non-degenerate. Therefore (cid:10)F ⊗n ⊗ S, F ⊗n ⊗ S(cid:11) = (cid:10)S ⊗ E⊗n, S ⊗ E⊗n(cid:11) = (cid:10)E⊗n, φE⊗n(hS, Si)E⊗n(cid:11) = (cid:10)E⊗n, φE⊗n(A)E⊗n(cid:11) = (cid:10)E⊗n, E⊗n(cid:11) , and the latter ideal provides a right c.a.i. for E⊗n. Theorem 5.3. Let AEA, BFB be regular C∗-correspondences. then E∞ ∼ F∞. s If E s ∼ F , Proof. Assume that E, F are elementary strong shift equivalent via R, S. It suffices to prove that the interior tensor product S∞ ⊗A∞ R∞ is (unitarily equivalent to) F∞. Then, by duality, R∞ ⊗B∞ S∞ is (unitarily equivalent s ∼ F∞. Towards this end we view S∞ as the X ⊗2-dilation to) E∞, hence E∞ of K(E ⊗ R, F ⊗2), i.e., K(cid:0)E ⊗ R, F ⊗2(cid:1) ⊗idX⊗2 and R∞ as the X ⊗2-dilation of K(F, E ⊗ R), i.e., −→ · · · , −→ K(cid:0)E⊗2 ⊗ R, F ⊗3(cid:1) ⊗idX⊗2 −→ K(cid:0)F ⊗2, E⊗2 ⊗ R(cid:1) ⊗idX⊗2 ⊗idX⊗2 −→ · · · . K (F, E ⊗ R) Note that A∞ can be written as (K(E ⊗ R), idX ⊗2)∞. By Proposition 5.1, for F ⊗k+1, for every the ideal (cid:10)E⊗k ⊗ R, E⊗k ⊗ R(cid:11) provides a right c.a.i. k ≥ 0. Therefore, (cid:10)E ⊗ R ⊗ (X ⊗2)⊗n, E ⊗ R ⊗ (X ⊗2)⊗n(cid:11) provides a right for F ⊗2 ⊗ (X ⊗2)⊗n, for every n ≥ 0. Thus, Proposition 3.9 applies, c.a.i. and we obtain that S∞ ⊗A∞ R∞ is (K(F, F ⊗2), idX ⊗2)∞ = F∞. Remark 5.4. In the above proof we have actually shown that (R ⊗B S)∞ = E∞ = R∞ ⊗B∞ S∞, and (S ⊗A R)∞ = F∞ = S∞ ⊗A∞ R∞. An immediate consequence of Theorem 5.3 is the following. SSE ∼ F∞. Theorem 5.5. Let AEA, BFB be regular C∗-correspondences. then E∞ Proof. Assume that ESSE Ti for i = 0, . . . , n. Since T0 = E and Tn = F , then E∞ s ∼ Ti+1, for i = 0, . . . , n. By Theorem 5.3, we get that (Ti)∞ ∼ F via a sequence of Ti, for i = 0, . . . , n. Then s ∼ (Ti+1)∞ SSE ∼ F∞. If E SSE ∼ F , C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 23 We use Theorem 5.3 to understand the passage to dilations for the strong Morita equivalence. Theorem 5.6. Let AEA, BFB be regular C∗-correspondences. If E then E∞ ∼ F∞. SME SME ∼ F , SME ∼ F ; then E and F are non-degenerate and E s Proof. Assume that E ∼ F via E ⊗A M ∗ and M (see Theorem 4.7). Therefore, by Theorem 5.3, s ∼ F∞ via (E ⊗ M ∗)∞ and M∞. Since M is assumed an imprimitivity E∞ bimodule, then M∞ is also an imprimitivity bimodule by Remark 5.2. By Proposition 4.8 we conclude that E∞ SME ∼ F∞. Remark 5.7. An alternative and more direct way of proving Theorem 5.6 SME would be to show that E∞ ∼ F∞ via the imprimitivity bimodule X = A∞ ⊗A X ⊗B B∞. It may take no effort to show that E ⊗A A∞ = A∞ ⊗A E, as sets, and conclude that E∞ ⊗A∞ X = X ⊗B∞ F∞. However there is a problem with the definition of X. It is easy to define a left and right action of A on A∞ (simply by multiplication), but in order to get the tensor product A∞ ⊗A X (and so X) an inner product of A∞ taking values in A is needed. The existence of such an inner product is not obvious to us. We have arrived to the last relation to be examined when passing to Pimsner dilations. Our next Theorem is one of the central results in this paper and will enable us to obtain new information even for concrete classes of operator algebras, i.e., Cuntz-Krieger C∗-algebras. Theorem 5.8. Let AEA, BFB be regular C∗-correspondences. with lag m, then E∞ ∼ F∞ with lag m. SE If E SE ∼ F Proof. Let R, S such that (i) E⊗m = R ⊗A S, F ⊗m = S ⊗A R, (ii) S ⊗A E = F ⊗B S, E ⊗A R = R ⊗B F. Thus E⊗m s ∼ F ⊗m. Since E is regular, we have E∞ = (K(E, E⊗2), idE)∞ = (K(E⊗2, E⊗3), idE)∞. Proposition 3.9 implies that (E∞)⊗2 = (K(E⊗2, E⊗3), idE)∞ ⊗A∞ (K(E, E⊗2), idE)∞ = (K(E, E⊗3), idE)∞ = (E⊗2)∞. A repetitive use of this argument shows that (E⊗m)∞ = (E⊗m−1)∞ ⊗ E∞ = (E∞)⊗m. Thus, by the remark following Theorem 5.3, we obtain (E∞)⊗m = (E⊗m)∞ = (R ⊗B S)∞ = R∞ ⊗B∞ S∞, and in a similar fashion (F∞)⊗m = S∞ ⊗A∞ R∞. 24 E.T.A. KAKARIADIS AND E.G. KATSOULIS What remains to be proved, in order to complete the proof, is that E∞ ⊗A∞ R∞ = R∞ ⊗B∞ F∞. Since E is regular, we get that E∞ coincides with (K(E⊗m, E⊗m+1), idE)∞ = (K(E⊗m, E⊗m+1), idX ⊗2)∞, so E∞ = (K(E⊗m, E⊗m+1), idX ⊗2)∞ , R∞ = (K(S, E⊗m), idX ⊗2)∞, where X is the bipartite inflation of R and S. Now E⊗m+mn = E⊗m ⊗ (X ⊗2)n, n ≥ 0, and, by Lemma 2.5, the ideal hE⊗m+mn, E⊗m+mni = (cid:10)E⊗m ⊗ (X ⊗2)n, E⊗m ⊗ (X ⊗2)n(cid:11) provides a right c.a.i. for E⊗E⊗m+mn = E⊗1+m+mn = E⊗m+1⊗(X ⊗2)n. Hence Proposition 3.9 implies that E∞ ⊗A∞ R∞ = (K(S, E⊗m+1), idX ⊗2)∞. Similarly, we express R∞ = (K(F ⊗m, E⊗m ⊗ R), idX ⊗2)∞ , F∞ = (K(F ⊗m−1, F ⊗m), idX ⊗2)∞; then R∞ ⊗B∞ F∞ = (K(F ⊗m−1, E⊗m ⊗ R), idX ⊗2)∞, since F ⊗m+mn = F ⊗m ⊗ (X ⊗2)n and hF ⊗m+mn, F ⊗m+mni provides a right c.a.i. for R ⊗ F ⊗m+mn = E⊗m ⊗ R ⊗ F ⊗mn = E⊗m ⊗ R ⊗ (X ⊗2)n. To prove that E∞ ⊗ R∞ = R∞ ⊗ F∞ we show that each one is unitarily equivalent to a submodule of the other. First we observe that for U = (cid:20) 0 E ⊗ R F ⊗ S 0 (cid:21) we get X ⊗ U = (cid:20) 0 R S 0 (cid:21) ⊗(cid:20) = (cid:20) R ⊗ F ⊗ S = (cid:20) E ⊗ R ⊗ S 0 0 0 E ⊗ R F ⊗ S 0 (cid:21) 0 S ⊗ E ⊗ R (cid:21) F ⊗ S ⊗ R (cid:21) = U ⊗ X, 0 Thus X ⊗2 ⊗ U = U ⊗ X ⊗2 and, by Proposition 3.8, we have that the correspondence R∞ ⊗B∞ F∞ = (K(F ⊗m−1, E⊗m ⊗ R), idX ⊗2)∞ C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 25 is unitarily equivalent, via some unitary u, to a submodule of (K(F ⊗m−1 ⊗ U, E⊗m ⊗ R ⊗ U ), idX ⊗2)∞ = = (K(F ⊗m−1 ⊗ F ⊗ S, E⊗m+1 ⊗ R ⊗ S), idX ⊗2)∞ = (K(S ⊗ R ⊗ S, E⊗m+1 ⊗ R ⊗ S), idX ⊗2)∞ = (K(S ⊗ X ⊗2, E⊗m+1 ⊗ X ⊗2), idX ⊗2)∞ = (K(S, E⊗m+1), idX ⊗2)∞ = E∞ ⊗ R∞. Also define V = (cid:20) 0 E⊗m−1 ⊗ R F ⊗m−1 ⊗ S 0 (cid:21) and verify that X ⊗ V = V ⊗ X. Thus, again by Proposition 3.8, the correspondence E∞ ⊗ R∞ = (K(S, E⊗m+1), idX ⊗2)∞ is unitarily equivalent, via some unitary v, to a submodule of (K(S ⊗ V, E⊗m+1 ⊗ V ), idX ⊗2) = = (K(F ⊗2m−1, E⊗2m ⊗ R), idX ⊗2)∞ = (K(F ⊗m−1 ⊗ X ⊗2, E⊗m ⊗ X ⊗2 ⊗ R), idX ⊗2)∞ = (K(F ⊗m−1 ⊗ X ⊗2, E⊗m ⊗ R ⊗ X ⊗2), idX ⊗2)∞ = (K(F ⊗m−1, E⊗m ⊗ R), idX ⊗2)∞ = R∞ ⊗ F∞. It remains to show that uv = id and vu = id. Note that uv maps an element [q] of any compact shift considered above (either it is in a C∗-algebra, either it is an element of the module) to [q ⊗ idU ⊗V ] = [q ⊗ idX ⊗4], since E⊗m−1 ⊗ R E ⊗ R 0 (cid:21) F ⊗m−1 ⊗ S 0 0 F ⊗ S ⊗ E⊗m−1 ⊗ R (cid:21) F ⊗ F ⊗m−1 ⊗ S ⊗ R (cid:21) 0 U ⊗ V = (cid:20) 0 0 F ⊗ S (cid:21) ⊗(cid:20) = (cid:20) E ⊗ R ⊗ F ⊗m−1 ⊗ S = (cid:20) E ⊗ E⊗m−1 ⊗ R ⊗ S = (cid:20) E⊗m ⊗ E⊗m 0 0 0 0 F ⊗m ⊗ F ⊗m (cid:21) = X ⊗4. However, throughout the proof we have used exclusively X ⊗2-dilations and so [q ⊗ idX ⊗4] = [q]. Therefore uv = id. In a similar way vu = id, since V ⊗ U = X ⊗4, and the proof is complete. As mentioned in the Introduction, we were strongly motivated by [14, Remark 5.5] and one of our aims was to check whether the alternative way proposed there to prove [14, Theorem 3.14] could be achieved, i.e., to show 26 E.T.A. KAKARIADIS AND E.G. KATSOULIS ∼ F then E∞ that if E s SME ∼ F∞. There is, though, a delicate point in this ap- proach. In [14, Remark 5.5] the authors claim that if E is a non-degenerate and regular C∗-correspondence, then the dilation E∞ as introduced by Pim- sner [17] is an imprimitivity bimodule. This is true, but only in the context of [17], because the C∗-correspondences there are always assumed full [17, Remark 1.2(3)]. This is a consequence of another delicate point in Pim- sner's theory, as his version of C∗-algebras, unfortunately denoted by the same symbols TE and OE, are not what have eventually become the usual C∗-algebras generated by the images of X and A, but they are only gen- erated by the image of X; hence there is no reason to make a distinction between full and non-full correspondences in Pimsner's theory. Note that when X is regular one can recover A in Pimsner's C∗-algebra OE, but this is not the case for TE. Of course, if one adds this extra element, then the scheme in [14, Remark 5.5] can be implemented, as we are about to show. Note that the previ- ous discussion and the next results settles Conjecture 1 appearing in the Introduction. Theorem 5.9. Let AEA and BFB be full, non-degenerate and regular C∗- correspondences. Then the following scheme holds E SME ∼ F 3 E s ∼ F 3 E SSE ∼ F 3 E SE ∼ F E∞ SME ∼ F∞ k 3 E∞ s ∼ F∞ k 3 E∞ SSE ∼ F∞ k 3 E∞ SE ∼ F∞ Proof. E and F are full and non-degenerate, thus E∞ and F∞ are imprim- itivity bimodules, and Theorem 6.1 (that will follow) applies. An immediate consequence of Theorem 5.9 and Theorem 3.2 is the fol- lowing. Theorem 5.10. Let AEA and BFB be full, non-degenerate and regular C∗- SE correspondences. If E ∼ F , then the corresponding Cuntz-Pimsner algebras are strong Morita equivalent. Proof. Suppose that E F∞. Therefore [16, Theorem 3.5] implies that OE∞ conclusion follows from Theorem 3.2. SE ∼ F . Then by Theorem 5.9 we have that E∞ SME ∼ SME ∼ OF∞ and the In particular we obtain the following result for Cuntz-Krieger C∗-algebras mentioned in the Introduction. Corollary 5.11. Let G and G′ be finite graphs with no sinks or sources and let AG and AG ′ be their adjacent matrices. If AG ∼ AG ′ , in the sense of Williams, then the Cuntz-Krieger C∗-algebras OG and OG ′ are strong Morita equivalent. SE + + + s + s + s + C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 27 There is also an application to unital injective dynamical systems. Corollary 5.12. Let (A, α) and (B, β) be unital injective dynamical sys- ∼ Yβ∞ and the crossed products A∞ ⋊α∞ Z tems. If Xα and B ⋊β∞ Z are strong Morita equivalent. ∼ Xβ , then Xα∞ SE SE We close this section by settling Conjecture 2 of the Introduction. This conjecture asserts that the vertical arrows in Theorem 5.9 are actually equiv- alences. We will use the following. Theorem 5.13. Let AEA be a full, non-degenerate and regular C∗- corre- spondence. If E ∼ E∞ then E is an imprimitivity bimodule. SE Proof. By assumption there exist non-degenerate, regular C∗- correspon- dences ARA∞, A∞ SA and a positive integer m such that E⊗m = R ⊗ S, (E∞)⊗m = (E⊗m)∞ = S ⊗ R and E ⊗ R = R ⊗ E∞, S ⊗ E = E∞ ⊗ S. Then A = (cid:10)E⊗m, E⊗m(cid:11) = hR ⊗ S, R ⊗ Si = hS, hR, Ri Si ⊆ hS, Si ⊆ A, since E⊗m is also full. Hence S is full. Now, let k ∈ K(S). Then k ⊗ idR ∈ K((E∞)⊗m), and since (E∞)⊗m is an imprimitivity bimodule, there is an x ∈ A∞ such that k ⊗ idR = φ(E∞)⊗m(x) = φS(x) ⊗ idR. Thus φS(x) = k, since R is regular; therefore A∞ ≃ K(S), hence S is an imprimitivity bimodule. Going back to the definition of E SE ∼ E∞, we see that S ⊗ E = E∞ ⊗ S, hence E = S∗ ⊗ E∞ ⊗ S. The fact that S, S∗ and E∞ are imprimitivity bimodules implies that E is also an imprimitivity bimodule. Let us see now why Conjecture 2 has a negative answer. Let E be any full, non-degenerate and regular C∗-correspondence, which is not an imprimitiv- ity bimodule. If Conjecture 2 was true, then E SE ∼ E∞, since both E and E∞ have unitarily equivalent Pimsner dilations. But then Theorem 5.13 would imply that E is an imprimitivity bimodule, a contradiction. 6. Imprimitivity Bimodules and the Shift Equivalence Problem Our work on shift equivalences suggests the following generalization of the Williams Problem in the context of C∗-correspondences. Shift Equivalence Problem for C∗-correspondences. If E, F are non- degenerate, regular C∗-correspondences and E SE ∼ F , then E SSE ∼ F . One might be tempted to say that the work of Kim and Roush [9] readily shows that the answer to the above conjecture is negative. However, it is not known to us whether two graph C∗-correspondences which fail to be strong shift equivalent via non-negative integral matrices remain inequivalent if one considers arbitrary C∗-correspondences in order to implement the strong shift equivalence. In other words, we do not know the answer to the Williams 28 E.T.A. KAKARIADIS AND E.G. KATSOULIS Problem even for the class of graph C∗-correspondences. If it does have a positive answer then it will provide an alternative route for proving our Corollary 5.11, but we conjecture that it doesn't in general. Nevertheless, in our next result we show that Shift Equivalence Problem has an affirmative answer for the class of imprimitivity bimodules. Recall that this result is essential for the proof of Theorem 5.9. Theorem 6.1. Strong Morita equivalence, elementary strong shift equiva- lence, strong shift equivalence and shift equivalence are equivalent for the class of imprimitivity bimodules. Proof. In view of Theorem 4.7, it suffices to show that if AEA and BFB are imprimitivity bimodules such that E SE ∼ F . Assume that E SE ∼ F via R, S with lag m. Then E⊗m = R ⊗ S and F ⊗m = S ⊗ R. Since E, F are imprimitivity bimodules, then E⊗m and F ⊗m are also imprimitivity bimodules. Hence S is regular and, due to It suffices to prove that S is an imprimitivity bimodule. First, it is full right since SE ∼, intertwines E and F . ∼ F , then E SME A = (cid:10)E⊗m, E⊗m(cid:11) = hR ⊗ S, R ⊗ Si = hS, hR, Ri · Si ⊆ hS, Si ⊆ A, thus hS, Si = A. In order to prove that it is full left, it suffices to prove that φS : B → L(S) is onto K(S) (since φS(B) ⊆ K(S), by regularity of S). To this end, let k ∈ K(S). Then, due to regularity of R, we obtain that k ⊗ idR ∈ K(S ⊗ R) = K(F ⊗m). Since F ⊗m is an imprimitivity bimodule, there is a b ∈ B such that φF ⊗m(b) = k ⊗ idR; in particular φS(b) ⊗ idR = k ⊗ idR. Thus φS(b) = k, since R is regular. Combining [16, Theorem 3.2, Theorem 3.5] with Theorem 6.1 we obtain the following corollary. SSE ∼ F or E s Corollary 6.2. If AEA, BFB are imprimitivity bimodules and E ∼ SE F, E ∼ F , then the corresponding Toeplitz-Cuntz-Pimsner al- gebras and Cuntz-Pimsner algebras are strong Morita equivalent as C∗- algebras, and the corresponding tensor algebras are strong Morita equivalent in the sense of [4]. SME ∼ F, E 7. Other Applications 7.1. Extension of [14, Theorem 3.14]. A natural question raised in [14] was whether [14, Theorem 3.14] is valid without the extra assumption of non-degeneracy. This can be established now with the theory we have devel- oped in Section 5. However, one cannot dispose of regularity, as mentioned explicitly in [14] and [2, Example 5.4]. Theorem 7.1. Let AEA and BFB be regular C∗-correspondences. If E then OE ∼ OF . SME SSE ∼ F C-ALGEBRAS AND EQUIVALENCES FOR C-CORRESPONDENCES 29 s Proof. First suppose that E s ∼ F∞. But E∞ and F∞ are regular Hilbert bimodules hence they are non-degenerate. SME ∼ OF∞ and Theorem 3.2 com- Therefore [14, Theorem 3.14] implies OE∞ pletes the proof in the case of elementary strong shift equivalence. ∼ F . By Theorem 5.3 we have E∞ Let ESSE ∼ F via a sequence of Ti, i = 0, . . . , n. Then OTi SME ∼ OTi+1, by the previous arguments, for every i = 0, . . . , n − 1. Strong Morita equivalence of C∗-algebras is transitive, hence OE = OT0 SME ∼ OTn = OF . 7.2. [14, Proposition 4.2] Revisited. The results in [14] and here, con- cerning strong Morita equivalence of the Cuntz-Pimsner algebras, can be generalized for degenerate correspondences over unital C∗- algebras because of [14, Proposition 4.2], i.e., if X is a correspondence over a unital C∗- algebra A, then OXess is a full corner of OX , where Xess := φX(A)X. In fact the proofs in [14, Proposition 4.2] apply in general for strict correspon- dences X. The key observation is that, if (ai) is a c.a.i. in A, then φX(ai) converges in the s-topology to a projection, say p, in L(X). As a conse- quence JX = JXess and, if X is regular, then so is Xess, in the same way as in [14, Proposition 4.1]. Theorem 7.2. Let AXA be a strict C∗-correspondence. Then OXess is a full corner of OX . Proof. Fix a covariant injective representation (π, t) of X (that admits a gauge action); then (π, tXess ) is a covariant injective representation of Xess (that admits a gauge action). Let P be the projection in M(OX ) that is defined by limi π(ai) for a c.a.i. (ai) of A; for example P t(ξ) := lim i π(ai)t(ξ) = t(lim i φ(ai)ξ) = t(pξ), for all ξ ∈ X. Then, the rest of the proof goes as the proof of [14, Proposition 4.2]. References [1] B. Abadie, S. Eilers and R. Exel, Morita equivalence for crossed products by Hilbert C∗- bimodules, Trans. Amer. Math. Soc. 350 (1998), 3043 -- 3054. [2] T. Bates, Applications of the Gauge-Invariant Uniqueness Theorem, Bulletin of the Australian Mathematical Society 65 (2002), 57 -- 67. [3] J. Cuntz and W. Krieger, A class of C-algebras and topological Markov chains, Inventiones Math. 56 (1980), 251 -- 268. [4] D. Blecher, P. Muhly and V. Paulsen, Categories of operator modules -- Morita equiv- alence and projective modules, Mem. Amer. Math. Soc. 143 (2000), no 681. [5] D. Drinen and N. Sieben, C-equivalences of graphs, J. Operator Theory 45 (2001), 209 -- 229. [6] N.J. Fowler, P.S. Muhly and I. Raeburn, Representations of Cuntz -- Pimsner algebras, Indiana Univ. Math. J. 52(3) (2003) 569 -- 605. [7] E. T.A. Kakariadis and E. G. Katsoulis, Contributions to the theory of C∗-envelopes with applications to multivariable dynamics, Trans. Amer. Math. Soc, to appear. [8] T. Katsura, On C ∗-algebras associated with C∗-correspondences, J. Funct. Anal. 217(2) (2004), 366 -- 401. 30 E.T.A. KAKARIADIS AND E.G. KATSOULIS [9] K. H. Kim and F. W. Roush, The Williams conjecture is false for irreducible subshifts, Ann. of Math 149(2) (1999), 545 -- 558. [10] K. H. Kim and F. W. Roush, Decidability of shift equivalence, Dynamical systems (College Park, MD, 198687), 374424, Lecture Notes in Math. 1342, Springer, Berlin, 1988. [11] C. Lance, Hilbert C∗-modules. A toolkit for operator algebraists, London Mathematical Society Lecture Note Series, 210 Cambridge University Press, Cambridge, 1995. [12] W. Krieger, On dimension functions and topological Markov chains, Invent. Math. 56 (1980), 239 -- 250. [13] K. Matsumoto, Actions of symbolic dynamical systems on C*-algebras, J. Reine Angew. Math. 605 (2007), 23 -- 49. [14] P. S. Muhly, D. Pask and M. Tomforde, Strong Shift Equivalence of C∗- correspondences, Israel J. Math. 167 (2008), 315 -- 345. [15] P. S. Muhly and B. Solel, Tensor algebras over C∗-correspondences: representations, dilations and C∗-envelopes J. Funct. Anal. 158 (1998), 389 -- 457. [16] P. S. Muhly and B. Solel, On the Morita Equivalence of Tensor Algebras, Proc. London Math. Soc. 3 (2000), 113 -- 168. [17] M. Pimsner, A class of C ∗-algebras generalizing both Cuntz-Krieger algebras and crossed products by Z, Free probability theory (Waterloo, ON, 1995), 189 -- 212, Fields Inst. Commun., 12, Amer. Math. Soc., Providence, RI, 1997. [18] I. Raeburn, Graph algebras, CBMS Regional Conference Series in Mathematics 103, 2005. [19] P.J. Stacey, Crossed products of C∗-algebras by ∗-endomorphisms, J. Austral. Math. Soc. Ser. A 54 (1993), 204 -- 212. [20] J.B. Wagoner, Strong shift equivalence theory and the shift equivalence problem, Bull. Amer. Math. Soc. (N.S.) 36 (1999), 271 -- 296. [21] R. Williams, Classification of subshifts of finite type, Annals of Math. 98 (1973), 120 -- 153; erratum, Annals of Math. 99 (1974), 380 -- 381. Pure Mathematics Department, University of Waterloo, Ontario N2L-3G1, Canada E-mail address: [email protected] Department of Mathematics, University of Athens, 15784 Athens, Greece Alternate address: Department of Mathematics, East Carolina University, Greenville, NC 27858, USA E-mail address: [email protected]
1102.3823	1	1102	2011-02-18T12:44:00	Convex polytopes and the index of Wiener-Hopf operators	[ "math.OA", "math.KT" ]	We study the C$^*$-algebra of Wiener-Hopf operators $A_\Omega$ on a cone $\Omega$ with polyhedral base $P$. As is known, a sequence of symbol maps may be defined, and their kernels give a filtration by ideals of $A_\Omega$, with liminary subquotients. One may define $K$-group valued 'index maps' between the subquotients. These form the $E^1$ term of the Atiyah-Hirzebruch type spectral sequence induced by the filtration. We show that this $E^1$ term may, as a complex, be identified with the cellular complex of $P$, considered as CW complex by taking convex faces as cells. It follows that $A_\Omega$ is $KK$-contractible, and that $A_\Omega/\mathbb K$ and $S$ are $KK$-equivalent. Moreover, the isomorphism class of $A_\Omega$ is a complete invariant for the combinatorial type of $P$.	math.OA	math	CONVEX POLYTOPES AND THE INDEX OF WIENER -- HOPF OPERATORS ALEXANDER ALLDRIDGE Abstract. We study the C∗-algebra of Wiener -- Hopf operators AΩ on a cone Ω with polyhedral base P . As is known, a sequence of symbol maps may be defined, and their kernels give a filtration by ideals of AΩ , with liminary subquotients. One may define K-group valued 'index maps' between the sub- quotients. These form the E 1 term of the Atiyah -- Hirzebruch type spectral sequence induced by the filtration. We show that this E 1 term may, as a com- plex, be identified with the cellular complex of P , considered as CW complex by taking convex faces as cells. It follows that AΩ is KK-contractible, and that AΩ/K and S are KK-equivalent. Moreover, the isomorphism class of AΩ is a complete invariant for the combinatorial type of P . INTRODUCTION The classical Wiener -- Hopf algebra AR+ , also known as the reduced Toeplitz algebra, is an object of basic interest in Operator Theory. It may be defined as the C∗- algebra of bounded operators on L2(R+) generated by all Wiener -- Hopf operators f (x − y)g(y) dy for all g ∈ L2(R+) , f ∈ L1(R) , x ∈ R+ . (Wf g)(x) =Z ∞ 0 The symbol map σ : AR+ → C0(R) is the surjective ∗-morphism defined on generators by σ(Wf ) = f , where the latter denotes the Fourier transform of f ∈ L1(R) . It gives rise to a short exact sequence (∗) 0 / K / AR+ σ / S = C0(R) / 0 . Let ∂ : K1(S) → K0(K) = Z be the connecting map in K-theory induced by this exact sequence. The following theorem is well-known. Theorem. If T is a Fredholm operator contained in the unitisation of AR+ , then ∂[σ(T )] = index T ∈ Z = K0(K) . Moreover, ∂ is a group isomorphism, and for any n ∈ Z , there exists a Fredholm element T as above such that index T = n . The six-term exact sequence immediately gives the following corollary. Corollary. We have Ki(AR+ ) = 0 for i = 0, 1 . In fact, the Theorem might also be deduced from the statement of the Corollary, by using the six term exact sequence. Moreover, the Theorem can be used to prove Bott periodicity and thus the existence of the six-term exact sequence. This is the approach taken by Cuntz [5]. (Cuntz defines AR+ algebraically, and his elegant deduction of the K-triviality of this C∗-algebra is quite different from the one we shall propose below.) 1 / / / / 2 A. ALLDRIDGE From the point of view of analysis and index theory, it seems natural to consider the multivariate generalisation of the Wiener -- Hopf algebra and to study its K-theory. Thus, let Ω ⊂ Rn be a closed convex cone, which we assume to be pointed, i.e. Ω contains no affine line, and solid, i.e. Ω generates Rn as a vector space. Then Wiener -- Hopf operators shall be the bounded operators on L2(Ω) given by (Wf g)(x) =ZΩ f (x − y)g(y) dy for all g ∈ L2(Ω) , f ∈ L1(Rn) , x ∈ Ω . The C∗-algebra AΩ of bounded operators generated by the Wf will be called the Wiener -- Hopf algebra. This C∗-algebra and its relatives are the object of study of quite an extensive literature, and we refer the interested reader to the introduction of our joint paper with Troels Johansen [1], for a partial overview. Just as in the n = 1 case, there is an obvious symbol map σ : AΩ → C0(Rn) , which continues to be a surjective ∗-morphism for n > 1 . However, it is not to be expected that the kernel of σ (the commutator ideal) equals the ideal of compact operators in this case. Rather, AΩ has a composition series whose length is at most n . For the remainder of the paper, let us assume that Ω is polyhedral, i.e. finitely generated as a convex cone. Then the length of the composition series is exactly n , cf. [1]. However, on the level of K- and even KK-theory, this distinction is invisible. Indeed, we shall prove in this paper the following theorem. Theorem 0.1. Let Ω be a polyhedral cone. Then AΩ is KK-contractible, and AΩ/K and S are KK-equivalent. This Theorem was previously known only for a particular class of polyhedral cones called exhaustible, and is due to Buyukliev [4] in this case. He exploits the particular combinatorial structure of these cones to prove the Theorem via Mayer -- Vietoris sequences and the exact six-term sequence. Arguably, in the general polyhedral case, the proof of KK-contractibility must take the whole combinatorial structure of an arbitrary polyhedral cone into account. In fact, the following result comes about as spin-off of our proof the above Theorem. Theorem 0.2. Let Ω be a cone with polyhedral base P . Then the isomorphism class of AΩ completely determines the combinatorial type of P , i.e., the lattice isomorphism class of its lattice of faces. This is turn relies on the fact that the cellular differential of P , considered as a CW complex by considering each j-face as a j-cell, may be identified with the d1 differential of the E1 term of the Atiyah -- Hirzebruch spectral sequence induced by the composition series alluded to above, a result which may be interesting in itself. We will describe this in detail below. 1. STRUCTURE OF THE WIENER -- HOPF ALGEBRA In this section, we review some results on the composition series of the Wiener -- Hopf algebra AΩ , in particular, the construction and computation of certain 'index maps'. These results are actually valid far beyond the case of polyhedral cones. However, restricting to polyhedral cones simplifies matters considerably, so we state them in this case only. The interested reader is referred to our joint papers with Troels Johansen [1, 2], for the general case. POLYTOPES AND INDEX 3 1.1. Composition Series and Analytical Index Formula. Let Ω ⊂ Rn be a pointed and solid polyhedral cone. Ω is spanned by its exposed rays, and one may choose a set E of generators of exposed rays contained in an affine hyperplane H . There exists a linear automorphism L of Rn such that L(H) = 1 × Rn−1 . Let P be the convex hull of all x such that (1, x) ∈ L(E) . Then P is a convex polyhedron in Rn−1 , and L(Ω) = R+ · (1 × P ) =(cid:8)(λ, λ · x)(cid:12)(cid:12) λ > 0 , x ∈ P ⊂ Rn−1(cid:9) . Henceforth, we will omit reference to the linear automorphism L , and assume that Ω = R+ ·(1×P ) where the set P ⊂ Rn−1 is a convex polyhedron. This assump- tion is no loss of generality, since the C∗-algebras AΩ and AL(Ω) are isomorphic. For j = −1, . . . , n − 1 , let fj be the number of j-dimensional convex faces of P (where the empty set is considered as the unique face of dimension −1). This somewhat annoying index shift is an artefact introduced by considering j-faces of P as (j + 1)-faces of Ω , and will continue to trouble us in the following. Theorem (Muhly -- Renault [7]). There exists a finite filtration of AΩ by ideals I0 = 0 ⊂ I1 = K ⊂ · · · ⊂ In+1 = AΩ such that Ij+1/Ij is a liminary C∗-algebra with spectrum {1, . . . , fj−1} × Rj . Both the ideals Ij and the isomorphism of the subquotients Ij+1/Ij with the algebras C0({1, . . . , fj−1} × Rj) ⊗ K , j < n , and C0(Rn) , j = n , respectively, are given quite explicitly, but we shall not need the precise formulae. Let ∂j : K i c({1, . . . , fj−1} × Rj) → K i+1 c ({1, . . . , fj−2} × Rj−1) , j = 1, . . . , n , be the K-theory connecting maps induced by the exact sequences 0 / Ij/Ij−1 / Ij+1/Ij−1 σj / Ij+1/Ij / 0 . Let Fj be the set of j-dimensional faces of P , and let ΩF , for F ∈ P , be the face of Ω spanned by F . For any A ⊂ Rn , let hAi denote the linear span. We may identify {1, . . . , fj−1} × Rj with the trivial rank j vector bundle over the finite base Fj−1 . Σj =(cid:8)(F, y) ∈ Fj−1 × Rn (cid:12)(cid:12) y ∈ hΩF i(cid:9) For any subset A ⊂ Rn , let A∗ = {y ∈ Rn(yA) > 0} be the dual cone, and let A⊥ = {y ∈ Rn(yA) = 0} be the orthogonal complement of the linear span. Then define, for any face F of P , Ω⊛ F =(cid:8)x ∈ hΩ⊥ F ∩ Ω∗i(cid:12)(cid:12) (xy) > 0 for all y ∈ Ω⊥ F ∩ Ω∗(cid:9) . (This notation differs from [1, 2].) The continuous field of Hilbert spaces (L2(Ω⊛ F ))(F,y)∈Σj naturally defines a Hilbert C0(Σj )-module Ej . The ∗-morphism σj extends to a representation of AΩ by ad- jointable endomorphisms of this Hilbert module. By these means, the map ∂j lends itself to an analytical expression, as follows. Theorem (A. -- Johansen [1]). Let a ∈ MN (I + [σj(a)] ∈ K 1 module E N j+1) represent the K-theory class c (Σj ) . Then σj−1(a) is a Fredholm operator on the Hilbert C0(Σj−1)- j−1 , and ∂j[σj(a)] = index σj−1(a) ∈ K 0 c (Σj−1) . / / / / 4 A. ALLDRIDGE 1.2. KK-Theoretical Index Formula. The finite set may be considered as bibundle w.r.t. the obvious projections ξ : Pj → Fj−2 and η : Pj → Fj−1 (j > 1). The map Pj =(cid:8)(E, F ) ∈ Fj−2 × Fj−1 (cid:12)(cid:12) E ⊂ F(cid:9) η∗Σj → ξ∗Σj−1 : (E, F, y) 7→ (E, F, pE(y)) realises η∗Σj as the trivial line bundle over the base ξ∗Σj−1 . (Here, for A ⊂ Rn , pA denotes the orthogonal projection onto hAi .) Indeed, a nowhere vanishing section is given by the map sj : ξ∗Σj−1 → η∗Σj : (E, F, u) 7→ (E, F, u + eF (E)) where the unit vector eF (E) ∈ Ω⊥ denotes the dual face of ΩF , then Ω⊥ F ∩ Ω⊛ the unique unit vector contained in this ray. E ∩ hΩF i is given as follows. If ΩF = Ω⊥ E is an extreme ray of Ω⊛ F ∩ Ω∗ E , and eF (E) is c (ξ∗Σj−1) . The trivial line bundle η∗Σj → ξ∗Σj−1 induces an isomorphism of K-groups c(η∗Σj) → K i+1 K i It is given by multiplication by an invertible KK- theory element yj ∈ KK 1(η∗Σj, ξ∗Σj−1) where for locally compact Hausdorff spaces X and Y , we write KK q(X, Y ) = KK(C0(X), C0(Rq × Y )) . Another way to think about yj is that it is 'fibre integration', i.e. the inverse of the Thom iso- morphism for the above line bundle. This depends on the choice of an orientation; we will go into detail further below. The only other ingredients needed for our index formula (at least in the poly- hedral case) are the projections pξ : ξ∗Σj−1 → Σj−1 and pη : η∗Σj → Σj induced, respectively, by ξ and η . Since ξ and η have finite domain, pξ and pη are proper, and thus induce ∗-morphisms ξξξ : C0(Σj−1) → C0(ξ∗Σj−1) and ηηη : C0(Σj) → C0(η∗Σj) , respectively. Theorem (A. -- Johansen [2]). Let 1 6 j 6 n . As elements of KK 1(Σj−1, Σj) , ∂j = ξξξ∗ηηη∗yj . 2. COHOMOLOGICAL INDEX AND CELLULAR DIFFERENTIAL 2.1. Cohomological expression of the index. If X is a locally compact space, k=0 H ∗+2k(X, Q) , called then there is a natural ring morphism ch : K ∗ the Chern character, which is rationally an isomorphism. Let π : V → X be an oriented real vector bundle. We have Thom isomorphisms c (X) → L∞ ϕV : K ∗ c (X) → K ∗+rk V c (V ) and ψV : H ∗ c (X) → H ∗+rk V c (V ) . In the special case that V is trivial, these are related via the Chern character, i.e. ch ◦ϕV = ψV ◦ ch . (In general, of course, the interplay is more subtle.) As is customary, we denote integration along the fibres of π , which is the inverse map ψ−1 V : H ∗+rk V In particular, applying this to y−1 (η∗Σj) , we immedi- c (ξ∗Σj−1) → K ∗+1 c (X) , by π∗ . (V ) → H ∗ c : K ∗ j c ately obtain the following cohomological expression of the index map ∂j . Proposition 2.1. For all u ∈ K i c(Σj) , we have ∞ ch(∂j (u)) = pη∗π∗p∗ ξ ch(u) in H 2k+i+1 c (Σj−1, Q) where π denotes the projection of the (trivial) line bundle η∗Σj → ξ∗Σj−1 . Mk=0 POLYTOPES AND INDEX 5 We remark that K i c(Σj ) = K i c(Fj−1 × Rj) =(0 Zfj−1 i + j ≡ 1 (mod 2) , i + j ≡ 0 (mod 2) , c (Σj ) has no torsion. In particular, ch is injective on K ∗ so that K ∗ c (Σj) . It is known that its image is the integral cohomology H ∗ c (Σj, Z) , cf. [6, Proposition 4.3]. In particular, ∂j is completely determined by its cohomological expression, and the latter is integral. In we take the trivialisation of η∗Σj → ξ∗Σj−1 to be given by the non-vanishing section sj defined above, then the orientation of this bundle is induced by choices of orientations of Σj−1 and Σj . These will be induced by choices of trivialisations of these bundles. (The triviality of the latter bundles is particular to the polyhedral situation.) As we shall presently see, the detailed inspection of these choices leads directly to the explicit expression of ∂j as a cellular differential. 2.2. The Wiener -- Hopf Index as a Cellular Differential. The n-dimensional convex polytope P can be considered as a finite CW complex by taking the j-faces to be the j-cells. (Topologically, P is of course an n-cell, so there are simpler ways to consider it as a CW-complex. However, our point of view captures the combinatorics of the face lattice.) The cellular complex is then (H 0(Fj ), dj) where H 0(Fj) is the free Abelian group generated by the j-faces. The vector bundle Σj → Pj is trivial, hence orientable, and we have a Thom c (Σj+1) given by the choice of an orientation. Let isomorphism ψj : H 0(Fj ) → H j us make this choice explicit. A trivialisation of Σj is given by the map Fj−1 × Rj → Σj : (F, y) 7→ (F, AF y) where for each F ∈ Fj−1 , AF : Rj → hΩF i is a linear isomorphism. An orientation of (the fibres of) Σj is given by pulling back the standard orientation σ+ = ε ◦ det of Rj to hΩF i along A−1 F to an orientation σF . The Thom isomorphism ψj is given by the cup product with the Thom class cj ∈ H j c (Σj) which is determined by the condition The line bundle π : η∗Σj → ξ∗Σj−1 is oriented by the choice of the non-vanishing section sj . Observe that for each (E, F ) ∈ Pj , there exists a unique orientation σ on hΩF i such that σ(eF (E), v1, . . . , vj−1) = σE(v1, . . . , vj−1) for all v1, . . . , vj−1 ∈ hΩEi . We denote by [E : F ] = ±1 the unique sign such that σF = [E : F ] · σE . If E 6⊂ F , we define [E : F ] = 0 . Proposition 2.2. For 0 6 j 6 d , let dj : H 0(Fj ) → H 0(Fj−1) be the map induced by the index map ∂j+1 : K j+1 c (Σj) and the Thom isomorphisms ψj+1 , ψj , via the relation (Σj+1) → K j c Then dj is given on generators by the formula ψj ◦ dj ◦ ψ−1 j+1 ◦ ch = ch ◦ ∂j+1 . dj(F ) = XE⊂F [E : F ] · E . Z(hΩF i,σF ) cj(F, ·) = 1 for all F ∈ Fj−1 . 6 A. ALLDRIDGE Proof. Consider a form cj+1 ∈ Γc(Σj+1, ∧j+1T ∗Σj+1) representing the Thom class in H j+1 ηcj+1 is represented by (Σj+1) . Then p∗ c η∗Σj+1 → ∧j+1T ∗η∗Σj+1 : (E, F, u) 7→ cj+1(F, u) . Because we have the decomposition hΩF i = R · eF (E) ⊕ hΩEi , the Fubini theorem gives Z(hΩE i,σE )ZR p∗ ηcj+1(E, F, · + te)(e, ·) dt = [E : F ] ·Z(hΩF i,σF ) cj+1(F, ·) = [E : F ] , where we write e = eF (E) . Since this condition characterises the Thom class cj up to the factor [E, F ] , π∗p∗ ηcj+1(E, F, ·) = [E : F ] · cj(E, ·) in H j c (hΩEi) . We find pξ∗π∗p∗ ηcj+1(E, u) = XF ⊂E π∗p∗ ηcj+1(E, F, u) = XF ⊂E Applying the cup product, the conclusion follows. [E : F ] · cj(E, u) . (cid:3) In order to see that the maps dj coincide with the cellular differentials dj , let us explicitly describe dj . To that end, we construct attaching maps. First, for any convex set C ⊂ Rj containing the origin in its interior, let µC be its Minkowski gauge. If, more generally, the convex set C has non-void interior and bC is its barycentre, then the map ϕC : Rj → Rj , ϕC (x) =( µC−bC (x−bC ) kx−bC k 0 · (x − bC) x 6= bC , x = bC , is a homeomorphism inducing homeomorphisms C → Bj and ∂C → Sj−1 where the latter has degree 1 . Next, for any j-face F ∈ Fj , let the linear isomorphism AF : Rj+1 → hΩF i used above to trivialise Σj+1 be chosen such that AF (0 × Rj) is the linear subspace parallel to the affine span of F , and such that F := A−1 F (F − bF ) ⊂ Bj . We then define the attaching map for the j-cell associated with F by φF : Bj → F ⊂ P : x 7→ A F ϕ−1 F (x) + bF If for any pair (E, F ) ∈ Fj−1 × Fj the number cEF ∈ Z denotes the degree of the composite Sj−1 φF / / ∂F / P/(P \ E◦) = E/∂E φ−1 E / / Bj−1/Sj−2 / Sj−1 where the rightmost map is the standard one (x 7→ (ux, 2kxk − 1) where u is suitably chosen), then the cellular differential dj : H 0(Fj ) → H 0(Fj−1) is defined on generators by dj(F ) = XE⊂F cEF · E . Proposition 2.3. We have cEF = [E, F ] for any pair (E, F ) ∈ Fj−1 × Fj . In particular, the cellular differential dj coincides with dj . / / POLYTOPES AND INDEX 7 Proof. Let H0 : ∂F → Sj−1 be the composite ∂F / P/(P \ E◦) = E/∂E φ−1 E / / Bj−1/Sj−2 / Sj−1 . Next, let 1 > t > 0 . Let e be a positive multiple of the projection of eF (E) onto the subspace 0 × Rn = hP i , and define Fs =(cid:8)x ∈ F (cid:12)(cid:12) (x − bE : e) 6 s(cid:9) for all s ∈ [0, 1] , the elements of 'height' 6 s over E . By definition of e , F0 = E . We may take e to be normalised in such a way that F1 = F . We wish to define a map Ht : ∂Ft → Sj−1 . To that end, define an affine map At : R × Rj−1 → Rn by At(2s − t, x) = s · e + AEx + bE . Then Ft = A−1 Define t (Ft) is a compact convex subset of Rj containing 0 in its interior. where ϕt = ϕ Ft Ht : Ft ∩ ∂F → Sj−1 by Ht = ft ◦ ϕ−1 t and for r ∈ [0, 1] , fr : Sj−1 → Sj−1 is given by t ◦ A−1 , for suitably chosen u . The map fr maps all points of the sphere of height > r to the 'north pole' ej = (0, . . . , 0, 1) . fr(s, x) =(cid:0)ux, min(cid:0)1, −1 + 2 s+1 r+1(cid:1)(cid:1) , The set B = ∂Ft ∩(cid:8)x (cid:12)(cid:12) (x − bE : e) = t(cid:9) bounds a flat in Ft . Since Ft ⊂ Bj , the elements of ϕ−1 (B)) have jth coordinate > t . Thus, Ht maps B to ej , and hence extends to all of ∂F by sending ∂F \ Ft to ej . Moreover, Ht , together with H0 , form a homotopy. We conclude t (A−1 t cEF = deg H0 ◦ φF = deg H1 ◦ φF = sign det((e, AE)−1AF ) = [E : F ] , which proves our claim. (cid:3) 3. PROOF THE MAIN THEOREM 3.1. Invariance of the Combinatorial Type of P . Recall that the combinato- rial type of the convex polyhedron P is the lattice isomorphism class of the lattice of convex faces of P . The f -vector of P is the vector (f0, . . . , fn) whose component fj is the numbers of j-faces. The following theorem is a somewhat surprising if simple consequence of Propo- sition 2.3. Theorem 3.1. Let Ω be a convex cone with polyhedral base P . Then the isomor- phism class of AΩ determines the combinatorial type of P . I.e., if Ω′ is another cone with polyhedral base P ′ , and AΩ and AΩ′ are isomorphic, then P and P ′ have isomorphic face lattices. Proof. The ideals in the filtration (Ij ) of AΩ from Theorem 1.1 are recursively characterised by the property that Ij+1/Ij is the largest liminary ideal of AΩ/Ij with Hausdorff spectrum. Thus, the f -vector of P and the index maps, and thus the maps dj , are uniquely determined up to a choice of orientations. In particular, the absolute values [E : F ] are uniquely determined for any pair of faces (E, F ) where E is of codimension one in F . But these numbers determine the lattice order. Hence the assertion. (cid:3) / / 8 A. ALLDRIDGE 3.2. The KK-contractibility of AΩ. As C∗-algebras with finite ideal filtrations with subquotients stably isomorphic to multiples of C0(Rn) , AΩ and AΩ/K belong to the bootstrap category N and therefore obey the UCT [3, Definition 22.3.4, Theorem 23.1.1]. In order to determine their KK equivalence class, it suffices to compute their K-theory. The C∗-algebra AΩ has a filtration by ideals (Ij ) , and since I1 = K , AΩ/K has the filtration by ideals given by (Ij /I1) . For any C∗-algebra A filtered by ideals (Ij) , Schochet has introduced an Atiyah -- Hirzebruch type homology spectral spectral sequence (Er In the case of A = AΩ , by [8, Theorem 2.1], its E1 term is p,q) which converges to the K-theory of A . E1 p,q = Kp+q(Ip/Ip−1) =(H 0(Fp−2) 0 q ≡ 1 (mod 2) , q ≡ 0 (mod 2) . Hence, every other column of E1 is zero. Hence, Er abuts to E2 . Proposition 3.2. The homology spectral sequence Er p,q in K-theory induced by the filtration (Ij) abuts to its E2 term, which is zero. In particular, K∗(AΩ) = 0 and K∗(AΩ/K) = K∗(C) . Proof. By definition, the d1 differential is the composite E1 p,1 = Kp+1(Ip/Ip−1) ∂ / Kp(Ip−1) / Kp(Ip−1/Ip−2) = E1 p−1,1 where the first map is the boundary map in the exact six-term sequence in K-theory, and the second is induced by the quotient map Ip−1 → Ip−1/Ip−2 . Considering the commutative diagram with exact rows, 0 0 / Ip−1 Ip Ip/Ip−1 / Ip−1/Ip−2 / Ip/Ip−2 / Ip/Ip−1 / 0 / 0 it follows from the naturality of connecting maps that d1 is also given by the connecting map for the lower line. This is just the map ∂p−1 . We have already noted that ∂p−1 is uniquely determined by its cohomological expression, and the latter gives the cellular differential dp−2 . Thus, (E1 p) is up to a shift, just the augmented cellular chain complex of the CW complex P . Since P is contractible, this complex is exact. Hence, E2 = 0 , and the first statement follows. p,1, d1 To complete the proof, observe that dividing by I1 = K corresponds to removing the augmentation from the cellular complex. The resulting complex has cohomology concentrated in degree zero, and H 0(F0) = Z . (Alternatively, use the exact six- term sequence.) (cid:3) Corollary 3.3. The C∗-algebra AΩ is KK-contractible, and AΩ/K and S are KK- equivalent. Corollary 3.4. The isomorphism K1(AΩ/K) → Z given by computing the numer- ical index of Fredholm Wiener -- Hopf operators, is an isomorphism. Proof. The groups K1(AΩ/K) and Z are isomorphic, and Buyukliev [4] has con- structed a Fredholm Wiener -- Hopf operator of index one. (cid:3) / / / / / / / / / / / / POLYTOPES AND INDEX 9 Acknowledgements. Part of this work was conducted while the author was a vis- itor at the Institut Henri Poincar´e, Paris. The author wishes to thank the institute for its hospitality and the organisers of the Special Trimester on 'Groupoids, Stacks in Physics and Geometry' for their support. References [1] A. Alldridge, T.R. Johansen, Spectrum and Analytical Indices of the C∗-Algebra of Wiener -- Hopf Operators, J. Funct. Anal. 249(2007), no. 2, 425 -- 453. [2] A. Alldridge, T.R. Johansen, An Index Theorem for Wiener -- Hopf Operators, Adv. Math. 218(2008), no. 1, 163 -- 201. [3] B. Blackadar, K-Theory for Operator Algebras, 2nd edition, MSRI Publications 5, Cam- bridge University Press, Cambridge, 1998. [4] N.P. Buyukliev, K-Theory of the C∗-Algebra of Multivariable Wiener -- Hopf Operators Associated with some Polyhedral Cones in Rn, Annuaire Univ. Sofia Fac. Math. Inform. 91(1997), no. 1 -- 2, 115 -- 125. [5] J. Cuntz, K-Theory and C∗-Algebras, pp. 55 -- 79 in: A. Bak (ed.), Algebraic K-Theory, Number Theory, Geometry and Analysis (Bielefeld, 1982), Lect. Notes Math. 1046, Springer- Verlag, Berlin, 1984. [6] A. Hatcher, Vector Bundles and K-Theory, Version 2.0, 2003. [7] P.S. Muhly, J.N. Renault, C∗-Algebras of Multivariable Wiener -- Hopf Operators, Trans. Amer. Math. Soc. 274(1982), no. 1, 1 -- 44. [8] C. Schochet, Topological Methods for C*-Algebras I: Spectral Sequences, Pac. J. Math., 96(1981), no. 1, 193 -- 211. A. ALLDRIDGE, Institut fur Mathematik, Universitat Paderborn, 33098 Paderborn, Germany E-mail address: [email protected]
1505.06633	3	1505	2016-02-26T06:01:20	Ergodic invariant states and irreducible representations of crossed product $C^*$-algebras	[ "math.OA", "math.DS" ]	Motivated by reformulating Furstenberg's $\times p,\times q$ conjecture via representations of a crossed product $C^$-algebra, we show that in a discrete $C^$-dynamical system $(A,\Gamma)$, the space of (ergodic) $\Gamma$-invariant states on $A$ is homeomorphic to a subspace of (pure) state space of $A\rtimes\Gamma$. Various applications of this in topological dynamical systems and representation theory are obtained. In particular, we prove that the classification of ergodic $\Gamma$-invariant regular Borel probability measures on a compact Hausdorff space $X$ is equivalent to the classification a special type of irreducible representations of $C(X)\rtimes \Gamma$.	math.OA	math	ERGODIC INVARIANT STATES AND IRREDUCIBLE REPRESENTATIONS OF CROSSED PRODUCT C ∗-ALGEBRAS HUICHI HUANG AND JIANCHAO WU Abstract. Motivated by reformulating Furstenberg's ×p, ×q con- jecture via representations of a crossed product C ∗-algebra, we show that in a discrete C ∗-dynamical system (A, Γ), the space of (ergodic) Γ-invariant states on A is homeomorphic to a subspace of (pure) state space of A ⋊ Γ. Various applications of this in topolog- ical dynamical systems and representation theory are obtained. In particular, we prove that the classification of ergodic Γ-invariant regular Borel probability measures on a compact Hausdorff space X is equivalent to the classification a special type of irreducible representations of C(X) ⋊ Γ. 1. Introduction Assume that p, q are two positive integers greater than 1 with log p log q irrational. H. Furstenberg gives the classification of closed ×p, ×q-invariant sub- sets of the unit circle T, which says such a set is either finite or T [7, Theorem IV.1.]. He also gives the following conjecture concerning the classification of ergodic ×p, ×q-invariant measures on T. Conjecture. [Furstenberg's ×p, ×q conjecture] An ergodic ×p, ×q-invariant Borel probability measure on T is either finitely supported or the Lebesgue measure. Furstenberg's conjecture is the simplest case of conjectures concern- ing classifications of invariant measures, and there are vast literatures Date: September 27, 2018. 2010 Mathematics Subject Classification. Primary 46L30, 46L55, 37B99. Key words and phrases. Invariant state, crossed product C ∗-algebra, irreducible representation. 1 about its general versions and their applications in number theory. See [6] for a survey. For Furstenberg's conjecture, the best known result is the following theorem, which is proven by D. J. Rudolph under the assumption that p, q is coprime [16, Theorem 4.9.], later improved by A. S. A. Johnson [9, Theorem A]. Theorem. [Rudolph-Johnson's Theorem] If µ is an ergodic ×p, ×q-invariant measure on T, then either hµ(Tp) = hµ(Tq) = 0 or µ is the Lebesgue measure. Here hµ(Tp) and hµ(Tq) stand for the measure-theoretic entropy of ×p and ×q with respect to µ respectively. See [18, Chapter 4] for the definition of entropy for measure preserving maps. For a ×p, ×q-invariant measure on T, denote the two isometries on L2(T, µ) induced by continuous maps ×p, ×q : T → T by Vp, Vq. By Rudolph-Johnson's Theorem, to classify ergodic ×p, ×q-invariant measures on T, it suffices to classify such ergodic measures with zero entropy for Tp or Tq. J. Cuntz notices that when hµ(Tp) = hµ(Tq) = 0, the operators Vp and Vq are two commuting unitary operators on L2(T, µ) [18, Corollary 4.14.3]. For the unitary operator Mz : L2(T, µ) → L2(T, µ) given by Mzf (z) = zf (z) for all f ∈ L2(T, µ) and z ∈ T, one have VpMz = M p z Vp and VqMz = M q z Vq. So a ×p, ×q invariant measure µ with zero entropy gives rise to a representation πµ of the universal unital C ∗-algebra C ∗(s, t, z) generated by three unitaries s, t and z with relations st = ts, sz = zps, tz = zqt in the following way: πµ(s) = Vp, πµ(t) = Vq, πµ(z) = Mz. pq ]) ⋊ Z2, where the two generators of Z2 acts on C ∗(Z[ 1 With the above observation, Cuntz suggests that one can consider er- godic ×p, ×q-invariant measures on T via representations of C ∗(s, t, z) ∼= C ∗(Z[ 1 pq ]) by automorphisms induced by ×p, ×q maps on [ 1 pq ], and the isomorphism Φ : C ∗(s, t, z) → C ∗(Z[ 1 pq ]) ⋊ Z2 is given by Φ(s) = a, Φ(t) = b and Φ(z) = 1. Here a = (1, 0) and b = (0, 1) are in Z2 and 1 is in Z[ 1 pq ] [3]. 2 Motivated by Cuntz's observation, firstly one have to answer the fol- lowing question: what kind of representation of C ∗(Z[ 1 invariant measure on T? pq ]) ⋊ Z2 is induced by a ×p, ×q- Denote the dual of Z[ 1 isomorphisms on Z[ 1 pq ] by Spq, the pq-solenoid [13, A.1]. The ×p, ×q pq ] give rise to ×p, ×q isomorphisms on Spq. We answer the above question in the following way. Firstly the space of ergodic ×p, ×q-invariant measures on T is home- omorphic to the space of ergodic ×p, ×q-invariant measures on Spq, hence the classification of ergodic ×p, ×q-invariant measures on T amounts to classification of ergodic ×p, ×q-invariant measures on Spq. Secondly ergodic ×p, ×q-invariant measures on Spq 1-1 corresponds to irreducible representations of C ∗(Z[ 1 pq ]) ⋊ Z2 whose restriction to Z2 contains the trivial representation. Moreover in a more general context, we prove the following which briefly shows how the problem of invariant states relates to crossed product C ∗-algebras. Assume that a discrete group Γ acts on a unital C ∗-algebra A as auto- morphisms. Denote this action by α, which is, a group homomorphism from Γ to the automorphism group Aut(A) of A. A state ϕ on A is Γ-invariant if ϕ(αs(a)) = ϕ(a) for all s in Γ and a in A. An extreme point of the set of Γ-invariant states on A 1 are called ergodic. Denote by A ⋊ Γ the full crossed product of the C ∗-dynamical system (A, Γ, α). Theorem 3.2. The space of (ergodic) Γ-invariant states on A is home- omorphic to the space of (pure) states on A ⋊ Γ whose restriction to Γ is the trivial character. We give some applications of Theorem 3.2 to topological dynamical systems and representation theory. Suppose a discrete group Γ acts on a compact Hausdorff space X as homeomorphisms (this is the same as Γ acting on the unital C ∗-algebra C(X), the space of continuous functions on X, by automorphisms). For 1This is a closed convex set when equipped with weak-∗ topology, hence when nonempty, the set of extreme points is also nonempty. 3 a representation π : C(X)⋊Γ → B(H), denote the space of Γ-invariant vectors in H by HΓ. Theorem 3.10. Every irreducible representation π of C(X) ⋊ Γ on a Hilbert space H satisfies that dim HΓ ≤ 1. When dim HΓ = 1, the representation π is uniquely induced by an ergodic Γ-invariant regular Borel probability measure µ on X. A special case of Theorem 3.10 is the following. Corollary 3.12. Suppose a discrete group Γ acts on a discrete abelian group G by group automorphisms. Every irreducible unitary representation π of G ⋊ Γ on a Hilbert space H satisfies that dim HΓ ≤ 1. When dim HΓ = 1, the representation π is uniquely induced by an ergodic Γ-invariant regular Borel probability measure µ on the Pon- tryagin dual bG of G. The paper is organized as follows. In the preliminary section, we recall some background of crossed prod- uct C ∗-algebras. The proof of Theorem 3.2 is given in section 3.2. At the end of this section, we include two immediate applications of Theo- rem 3.2, namely, Proposition 3.7 and Proposition 3.5, to C ∗-dynamical systems. In section 3.3, we show Theorem 3.9 and Theorem 3.10. In the last section we prove Theorem 4.2 which enables us to reformulate Furstenberg's ×p, ×q problem in terms of representation theory of the semidirect product group Z[ 1 pq ] ⋊ Z2. Acknowledgements We are grateful to Joachim Cuntz for sharing his insight for Fursten- berg's problem with us. We benefits a lot from various discussions with him. H. Huang would thank Xin Li for his suggestion to consider invari- ant measures in more general settings. He also thanks Hanfeng Li for his detailed comments. Thanks are also due to Kang Li for his pointing out the reference [19] to us which helps to prove Proposition 3.5. We also thank Sven Raum for his valuable comments which lead to much briefer proofs of Lemma 3.3 and Proposition 3.7. We also thank an anonymous referee for helpful comments. The paper was finished when we were postdoctoral fellows supported by ERC Advanced Grant No. 267079. 4 2. Preliminary In this section, we list some background for C ∗-dynamical systems. Within this article Γ stands for a discrete group and A stands for a unital C ∗-algebra whose state space and pure state space are denoted by S(A) and P (A) respectively. Denote the GNS representation of A with respect to a ϕ ∈ S(A) by πϕ : A → B(L2(A, ϕ)) where L2(A, ϕ) stands for the Hilbert space corresponding to πϕ. Let Iϕ = {a ∈ A ϕ(a∗a) = 0}. Denote a + Iϕ by a for all a ∈ A. Definition 2.1. An action of Γ on A as automorphisms is a group homomorphism α : Γ → Aut(A), where Aut(A) stands for the set of ∗-isomorphisms from A to A (this is a group under composition). We call (A, Γ, α) a dynamical system. A Γ-invariant state is a state ϕ on A such that ϕ(αs(a)) = ϕ(a) for all s ∈ Γ and a ∈ A [17]. Denote the set of Γ-invariant states on A by SΓ(A). It is clear that SΓ(A) is a convex closed set under weak-∗ topology. If SΓ(A) is nonempty, then it contains at least one extreme point. We call an extreme point of SΓ(A) an ergodic Γ-invariant state on A. The set of ergodic Γ-invariant states on A is denoted by EΓ(A). A representation of a C ∗-algebra B on a Hilbert space H is a ∗- homomorphism π : B → B(H) and it is called irreducible if the com- mutant C(π(B)) consisting of elements in B(H) commuting with every element in π(B) are scalar multiples of identity operator. A covariant representation (π, U, H) of a dynamical system (A, Γ, α) consists of a representation π of A and a unitary representation U of Γ on a Hilbert space H such that π(αs(a)) = Usπ(a)U∗ s for all a ∈ A and s ∈ Γ. Let Cc(Γ, A) be the space of finitely supported A-valued functions on Γ. For f, g ∈ Cc(Γ, A), the product f ∗ g is given by f ∗ g(t) = X f (s1)αs1(g(s2)) s1s2=t 5 and f ∗ is given by f ∗(t) = αt(f (t−1)∗) for every t ∈ Γ. Then Cc(Γ, A) is a ∗-algebra . Given a covariant repre- sentation (π, U, H) of a dynamical system (A, Γ, α), one can construct a representation of Cc(Γ, A) on H. Definition 2.2. For a dynamical system (A, Γ, α), the crossed prod- uct C ∗-algebra A ⋊ Γ is the completion of Cc(Γ, A) under the norm kf k = sup kπ(f )k for f ∈ Cc(Γ, A) where the supreme is taken over all representations of Cc(Γ, A). Denote by us the unitary in A ⋊ Γ corresponding to an s ∈ Γ. There is a one-one correspondence between representations of A ⋊ Γ and covariant representations of (A, Γ, α). We refer readers to [4, Chapter VIII] for more about discrete crossed products. 3. Main results 3.1. Covariant representations of (A, Γ, α) induced by invariant states. If ϕ ∈ SΓ(A), then there is a unitary representation (the Koopman representation) Uϕ of Γ on L2(A, ϕ) given by Uϕ(s)(a) = [αs(a) for all s ∈ Γ and a ∈ A [12, 14, 17]. We give details below for com- pleteness. The representation Uϕ is well-defined since (1) For every a ∈ Iϕ, we have ϕ(αs(a)∗αs(a)) = ϕ(αs(a∗a)) = ϕ(a∗a) = 0 for all s ∈ Γ. (2) For every s ∈ Γ, the map Uϕ(s) is surjective since its image is dense in L2(A, ϕ), and Uϕ(s) is an isometry since for all a ∈ A, h [αs(a), [αs(a)i = ϕ(αs(a)∗αs(a)) = ϕ(αs(a∗a)) = ϕ(a∗a) = ha, ai. So Uϕ(s) is a unitary. 6 (3) Uϕ(st)(a) = \αst(a) = Us(Ut(a)) for all s, t ∈ Γ and a ∈ A. Hence Uϕ is a unitary representation of Γ on L2(A, ϕ). Furthermore we have the following. Lemma 3.1. Given a Γ-invariant state ϕ on A, the triple (πϕ, Uϕ, L2(A, ϕ)) gives a covariant representation of (A, Γ, α). So there is a representa- tion of A ⋊ Γ on L2(A, ϕ), which we denote by ρϕ, given by ρϕ(X asus) = X πϕ(as)Uϕ(s) for any Ps∈Γ asus ∈ Cc(Γ, A). s∈Γ s∈Γ Proof. For all a, b ∈ A and s ∈ Γ, we have Uϕ(s)(πϕ(a))Uϕ(s−1)(b) = Uϕ(s)(πϕ(a))((αs−1(b))∧) = Uϕ(s)((aαs−1(b))∧) = αs((aαs−1(b))∧) = (αs(a)αss−1(b))∧ = (αs(a)b)∧ = πϕ(αs(a))(b). This completes the proof. (cid:3) 3.2. Γ-invariant states on A and states on A ⋊ Γ. Denote {ϕ ∈ S(A ⋊ Γ) ϕ(us) = 1 for all s ∈ Γ} by S1(A ⋊ Γ) and {ψ ∈ P (A ⋊ Γ)ψ(us) = 1 for all s ∈ Γ} by P 1(A ⋊ Γ). We have the following. Theorem 3.2. When equipped with weak-∗ topologies, the restriction maps R : SΓ(A) → S1(A ⋊ Γ) and R : EΓ(A) → P 1(A ⋊ Γ) are homoemorphisms. To prove this theorem, we first prove the following lemma which says that for every ϕ in S1(A ⋊ Γ), the restriction ϕA belongs to SΓ(A). Lemma 3.3. For any state ϕ on A ⋊ Γ such that ϕ(us) = 1 for every s ∈ Γ, we have ϕ(usaut) = ϕ(a) for all a ∈ A and s, t ∈ Γ. Consequently the restriction ϕA is a Γ-invariant state on A. Proof. The proof follows from [1, Proposition 1.5.7.] since by assump- tion every us is contained in the multiplicative domain of ϕ. For convenience of readers, we give a direct proof which is also based on Cauchy-Schwarz inequality like [1, Proposition 1.5.7.]. 7 For all a ∈ A and s, t ∈ Γ, we have ϕ(usaut) − ϕ(a) ≤ ϕ(usaut) − ϕ(aut) + ϕ(aut) − ϕ(a) = ϕ((us − 1)aut) + ϕ(a(ut − 1)) = haut, (us − 1)∗iL2(A⋊Γ,ϕ) + hut − 1, a∗iL2(A⋊Γ,ϕ) (Cauchy-Schwarz Inequality) ≤ [ϕ((ut)∗a∗aut)] (ϕ(ut) = 1 implies that ϕ((ut − 1)(ut − 1)∗) = 0 and ϕ((ut − 1)∗(ut − 1)) = 0.) = 0. 2 + [ϕ((ut − 1)∗(ut − 1))] 2 [ϕ((us − 1)(us − 1)∗)] 1 1 1 2 [ϕ(aa∗)] 1 2 (cid:3) It follows from Lemma 3.1 that for a Γ-invariant state ϕ, there is a representation ρϕ of A ⋊ Γ on L2(A, ϕ) given by ρϕ(X asus) = X s∈Γ s∈Γ πϕ(as)Uϕ(s) for every Ps∈Γ asus ∈ Cc(Γ, A). Proof. [Proof of Theorem 3.2.] By Lemma 3.3, the restriction map R : S1(A ⋊ Γ) → SΓ(A) given by R(ϕ) = ϕA for every ϕ ∈ S1(A ⋊ Γ), is well-defined. Since A is a C ∗-subalgebra of A ⋊ Γ, the map R is continuous under weak-∗ topology. If R(ϕ1) = R(ϕ2) for ϕ1, ϕ2 ∈ S1(A ⋊ Γ), then ϕ1(a) = ϕ2(a) for all a ∈ A. By Lemma 3.3, ϕ1(aus) = ϕ2(aus), for all a ∈ A and s ∈ Γ. Since every element in Cc(Γ, A) is a linear combination of aus and Cc(Γ, A) is a dense subspace of A ⋊ Γ. Hence ϕ1 = ϕ2 and R is injective. Moreover given a ϕ ∈ SΓ(A). By Lemma 3.1, ϕ gives a representation ρϕ of A ⋊ Γ on L2(A, ϕ). Let ϕ be the state of A ⋊ Γ given by (ϕ)(b) = hρϕ(b)(1), 1i for all b ∈ A ⋊ Γ. By the definition of ρϕ, we see that ϕ ∈ S1(A ⋊ Γ) and ϕA = ϕ. This shows the surjectivity of R. Above all R is a bijective continuous map between two compact Haus- dorff spaces S1(A ⋊ Γ) and SΓ(A). Therefore R is a homeomorphism. 8 Note that R is an affine map between two convex spaces S1(A ⋊ Γ) and SΓ(A), so the set of extreme points of S1(A ⋊ Γ) is homeomorphic to EΓ(A). Suppose that ϕ is an extreme point of S1(A⋊Γ) and ϕ = λϕ1+(1−λ)ϕ2 for two states ϕ1, ϕ2 on A ⋊ Γ and some 0 < λ < 1. Then 1 = ϕ(us) = λϕ1(us) + (1 − λ)ϕ2(us) for every s ∈ Γ. It follows that ϕ1(us) = ϕ2(us) = 1, that is, ϕ1, ϕ2 ∈ S1(A ⋊ Γ). Hence ϕ = ϕ1 = ϕ2 and ϕ is a pure state. (cid:3) For a character ξ on Γ (a group homomorphism from Γ to T), denote {ϕ ∈ S(A ⋊ Γ)ϕ(us) = ξ(s) for all s ∈ Γ} by Sξ(A ⋊ Γ) and {ϕ ∈ P (A ⋊ Γ)ϕ(us) = ξ(s) for all s ∈ Γ} by P ξ(A ⋊ Γ). For a representation of A ⋊ Γ on a Hilbert space H, define Hξ = {x ∈ Hπ(us)(x) = ξ(s)x for all s ∈ Γ}. We have the following improvement of Theorem 3.2. Corollary 3.4. Let ξ be a character on Γ. When equipped with weak-∗ topologies, SΓ(A) ∼= Sξ(A ⋊ Γ) and EΓ(A) ∼= P ξ(A ⋊ Γ). Proof. By Theorem 3.2, SΓ(A) ∼= S1(A ⋊ Γ). Note that S1(A ⋊ Γ) ∼= Sξ(A ⋊ Γ) (P 1(A ⋊ Γ) ∼= P ξ(A ⋊ Γ) follows) and the homeomor- phism is induced by the isomorphism Λ : A ⋊ Γ → A ⋊ Γ given by Λ(Ps∈Γ asus) = Ps∈Γ ξ(s)asus for all Ps∈Γ asus ∈ Cc(Γ, A). (cid:3) For a representation π : A ⋊ Γ → B(H), denote {x ∈ Hπ(us)(x) = x for all s ∈ Γ} by HΓ. Proposition 3.5. For a C ∗-dynamical system (A, Γ, α), the following are equivalent. (1) The set SΓ(A) is nonempty. (2) The canonical homomorphism C ∗(Γ) → A ⋊Γ is an embedding. (3) There exists a representation π : A ⋊ Γ → B(H) such that HΓ 6= 0, or equivalently, there exists a covariant representation (π, U, H) of (A, Γ, α) such that U contains the trivial represen- tation of Γ. Proof. (1) =⇒ (2). Take a ϕ ∈ SΓ(A). Let Γ act on C trivially. By the invariance, the map ϕ : A → C is a Γ-equivariant contractive completely positive map. 9 By [1, Exercise 4.1.4], there exists a contractive completely positive map ϕ : A ⋊ Γ → C ∗(Γ) such that ϕ(Ps∈Γ asus) = Ps∈Γ ϕ(as)us. Immediately one can check that the composition of maps C ∗(Γ) → A ⋊ Γ → ϕ C ∗(Γ) is the identity map. Hence the canonical homomorphism C ∗(Γ) → A ⋊ Γ is an embedding. (2) =⇒ (1). By Theorem 3.2, it suffices to show S1(A ⋊ Γ) is nonempty. Let π0 : Γ → C be the trivial unitary representation of Γ on C. Then π0 is a state on C ∗(Γ) such that π0(us) = 1 for every s ∈ Γ. Note that C ∗(Γ) is a Banach subspace of A ⋊ Γ. By the Hahn-Banach Theorem, one can extend π0 to a bounded linear functional ϕ on A ⋊ Γ without changing its norm [2, Corollary 6.5]. Hence kϕk = kπ0k = 1 = π0(1) = ϕ(1). So ϕ is a state on A ⋊ Γ [11, Theorem 4.3.2], and satisfies that ϕ(us) = 1 for all s ∈ Γ. (1) =⇒ (3). This is guaranteed by Lemma 3.1. (3) =⇒ (1). Take a unit vector x in HΓ and define a state ϕ on A ⋊ Γ by ϕ(b) = hπ(b)x, xi for all b ∈ A ⋊ Γ. It follows that ϕ ∈ S1(A ⋊ Γ). (cid:3) Remark 3.6. The equivalence of (1) and (2) may be well-known. When A is commutative and Γ is locally compact, this is mentioned in [19, Remark 7.5]. To the best of our knowledge, it does not appear elsewhere in the literature. Notice that Uϕ gives rise to an action of Γ on B(L2(A, ϕ)), also denoted by α for convenience, defined by αs(T ) = Uϕ(s)T Uϕ(s−1) for every T ∈ B(L2(A, ϕ)) and s ∈ Γ. Denote hT (1), 1i by ϕ(T ) for all T ∈ B(L2(A, ϕ)). When ϕ is a Γ- invariant state on A, it is also a Γ-invariant state on B(L2(A, ϕ)) since ϕ(αs(T )) = ϕ(Uϕ(s)T Uϕ(s−1)) = hUϕ(s)T Uϕ(s−1)(1), 1i = hT Uϕ(s−1)(1), Uϕ(s−1)(1)i = hT (1), 1i = ϕ(T ) for all s ∈ Γ. 10 We can also see that (3.A) αs(πϕ(a)) = πϕ(αs(a)) for every a ∈ A and s ∈ Γ. We call a T ∈ B(L2(A, ϕ)) is Γ-invariant if αs(T ) = T for all s ∈ Γ. Denote the set of Γ-invariant operators in B(L2(A, ϕ)) by B(L2(A, ϕ))Γ. Let πϕ(A)′ = {T ∈ B(L2(A, ϕ)) T πϕ(a) = πϕ(a)T for all a ∈ A}. Proposition 3.7. A Γ-invariant state ϕ on A is ergodic iff ϕ(T ∗T ) = ϕ(T )2 for every T ∈ B(L2(A, ϕ))Γ ∩ πϕ(A)′. Proof. Recall that R : S1(A ⋊ Γ) → SΓ(A) is the restriction map. For ϕ ∈ SΓ(A), denote R−1(ϕ) by ψ. By Theorem 3.2, ψ is in S1(A ⋊ Γ). Observe that B(L2(A, ϕ))Γ ∩ πϕ(A)′ = πψ(A ⋊ Γ)′. Again by Theorem 3.2, the state ϕ on A is an ergodic Γ-invariant state iff ψ is a pure state on A ⋊ Γ iff πψ(A ⋊ Γ)′ = C. The "only if" part follows immediately. Now suppose ϕ(T ∗T ) = ϕ(T )2 for every T ∈ B(L2(A, ϕ))Γ ∩ πϕ(A)′ = πψ(A ⋊ Γ)′. A straightforward calculation shows that T (1) = ϕ(T )1. Then for every a ∈ A, we have T (a) = T πϕ(a)(1) = πϕ(a)T (1) = πϕ(a)(ϕ(T )1) = ϕ(T )a. This means T = ϕ(T ), a scalar multiple of the identity operator. Hence πψ(A ⋊ Γ)′ = C and ψ is a pure state. (cid:3) Remark 3.8. The key observation B(L2(A, ϕ))Γ ∩πϕ(A)′ = πψ(A⋊Γ)′ in the proof is pointed out to us by Sven Raum. 3.3. Ergodic Γ-invariant states on A and irreducible represen- tations of A ⋊ Γ. We say a representation π1 : B → B(H1) of a C ∗-algebra B is unitar- ily equivalent to a representation π2 : B → B(H2) if there exists a surjective isometry U : H1 → H2 such that U π1(b)(x) = π2(b)U(x) for all b ∈ B and x ∈ H1. Theorem 3.9. A representation π : A ⋊ Γ → B(H) is unitarily equivalent to ρϕ : A ⋊ Γ → B(L2(A, ϕ)) for some ϕ ∈ EΓ(A) iff π is irreducible and HΓ 6= 0. 11 Proof. For a ϕ ∈ EΓ(A), by Theorem 3.2, there exists a ψ ∈ P 1(A ⋊ Γ) such that R(ψ) = ϕ. Next we show that ρϕ is unitarily equivalent to the GNS representation πψ : A ⋊ Γ → B(L2(A ⋊ Γ, ψ)) of A ⋊ Γ with respect to ψ. Since ψ is a pure state, this shows ρϕ is irreducible. Claim. ρϕ is unitarily equivalent to πψ. Proof. Define Φ : L2(A ⋊ Γ, ψ) → L2(A, ϕ) by Φ((X fsus)∧) = X fs + Iϕ. Here Iψ = {b ∈ A ⋊ Γψ(b∗b) = 0} and Iϕ = {a ∈ Aϕ(a∗a) = 0}. s∈Γ s∈Γ Now we check Φ is a Hilbert space isomorphism. By Lemma 3.3 we have s∈Γ h(X = ψ(X = ϕ((X s,t∈Γ fsus)∧, (X gtut)∧i = ψ((X gtut)∗(X fsus)) t∈Γ t fsus) = ψ(X fs)) = h(X ut−1g∗ gt)∗(X s,t∈Γ t∈Γ s∈Γ g∗ t fs) fs)∧, (X gt)∧i. t∈Γ s∈Γ s∈Γ t∈Γ Thus Φ is an isometry. The image of Φ is dense in L2(A, ϕ), so Φ is also surjective. These prove that Φ is an isomorphism between Hilbert spaces. Next we verify unitary equivalence between πψ and ρϕ. Φπψ(X = Φ((X fsus)((X gtut)∧) = Φ((X fsusgtus−1ust)∧) = (X s∈Γ t∈Γ s,t∈Γ s,t∈Γ s,t∈Γ fsusgtut)∧) fsαs(gt))∧. On the other hand, ρϕ((X = (X s∈Γ fsus)Φ(X gtut)∧) = ρϕ(X fsus)((X gt)∧) t∈Γ s∈Γ t∈Γ fsαs(gt))∧. s,t∈Γ So πψ(b)(x) = Φ−1ρϕ(b)Φ(x) for all b ∈ A ⋊ Γ and x ∈ L2(A ⋊ Γ, ψ). Hence πψ and ρϕ are unitarily equivalent. (cid:3) 12 Note that 0 6= 1 ∈ L2(A, ϕ) and ρϕ(us)(1) = 1 for all s ∈ Γ. Hence L2(A, ϕ)Γ 6= 0. Conversely given an irreducible representation π : A ⋊ Γ → B(H) with HΓ 6= 0, take a unit vector x ∈ HΓ and define a state ψ on A ⋊ Γ by ψ(b) = hπ(b)x, xi for all b ∈ A ⋊ Γ. Since π is irreducible, the state ψ is a pure state and the GNS representation of A ⋊ Γ with respect to ψ, πψ is unitarily equivalent to π [4, Theorem I.9.8] [5, 2.4.6]. Also x ∈ HΓ implies that ψ(us) = 1 for all s ∈ Γ. So ϕ = ψA ∈ EΓ(A). By the previous claim πψ is unitarily equivalent to ρϕ. This finishes the proof. (cid:3) Now we consider the case when A is commutative. Theorem 3.10. For any irreducible representation π : C(X) ⋊ Γ → B(H), we have dim HΓ ≤ 1. If HΓ 6= 0, then there exists a unique ergodic Γ-invariant state ϕ on C(X) (or a unique regular Borel proba- bility measure on X) such that π is unitarily equivalent to ρϕ. Proof. Suppose HΓ 6= 0 for an irreducible representation π : C(X) ⋊ Γ → B(H). Take unit vectors x, y ∈ HΓ. Define a state ψ on A ⋊ Γ by ψ(b) = hπ(b)x, xi for all b ∈ C(X) ⋊ Γ. Then ϕ = ψC(X) gives an ergodic Γ-invariant probability measure µ on X with ϕ(f ) = RX f dµ for all f ∈ C(X). Also the GNS representation πψ of C(X) ⋊ Γ with respect to ψ, is unitarily equivalent to ρϕ : C(X) ⋊ Γ → B(L2(A, ϕ)). Note that L2(A, ϕ) = L2(X, µ) and L2(A, ϕ)Γ consists of Γ-invariant functions in L2(X, µ), which are always constant functions [8, Chapter3, 3.10.]. Under surjective isometries H ∼= L2(A ⋊ Γ, ψ) ∼= L2(X, µ), both x and y are mapped to Γ-invariant functions in L2(X, µ). Since µ is ergodic, their images in L2(X, µ) are both constant functions. Hence there exists a constant λ with absolute value 1 such that x = λy. This shows that dim HΓ = 1. For the second part, the existence of ϕ follows from Theorem 3.9. To prove the uniqueness of ϕ, we show the following. Claim. If ρϕ ∼ ρψ for ϕ, ψ ∈ SΓ(C(X)), then ϕ = ψ. 13 Proof. Let Θ : L2(A, ϕ) → L2(A, ψ) be an isomorphism. It is easy i.e., Θ : L2(A, ϕ)Γ → to see that Θ preserves Γ-invariant vectors, L2(A, ψ)Γ is also an isomorphism. Hence Θ(1) = λ1 for some com- plex number λ with λ = 1. By definition of ρϕ, we have ϕ(f ) = hρϕ(f )1, 1i for all f ∈ C(X). It follows that ψ(f ) = hρψ(f )1, 1i = hΘ−1ρϕ(f )Θ1, 1i = hρϕ(f )λ1, λ1i = ϕ(f ) for all f ∈ C(X). (cid:3) Hence ϕ is uniquely determined by the unitary equivalence class of π. (cid:3) Remark 3.11. (1) Theorem 3.9 and Theorem 3.10 say that clas- sification of ergodic Γ-invariant regular Borel probability mea- sures on a compact Hausdorff space X amounts to classification of equivalence classes of irreducible representations of C(X) ⋊Γ whose restriction to Γ contains trivial representation. (2) When A is non-commutative, Theorem 3.10 fails. For instance, one can take a noncommutative C ∗-algebra A and a discrete group Γ acting on A trivially. An irreducible representation π : A → B(H) with dim H > 1 and the trivial representation Γ → B(H) give rise to an irreducible representation of ρ : A ⋊ Γ → B(H). But H = HΓ is not of dimension 1. There is an immediate application of Theorem 3.10 to representation theory of semidirect product groups. Corollary 3.12. Suppose a discrete group Γ acts on a discrete abelian group G by group automorphisms. Every irreducible unitary represen- tation π : G⋊Γ → B(H) of the semidirect product group G⋊Γ satisfies that dim HΓ ≤ 1. When dim HΓ = 1, the representation π is induced by an ergodic Γ-invariant regular Borel probability measure µ on the Pontryagin dual bG of G. Proof. Note that Γ acts on group C ∗-algebra C ∗(G) as automorphisms, C ∗(G) = C( G) for the dual group G of G and C ∗(G) ⋊ Γ ∼= C ∗(G ⋊ Γ). There exists a 1-1 correspondence between irreducible unitary repre- sentations of G⋊Γ and irreducible representations of C ∗(G⋊Γ). Apply Theorem 3.10 to the case C(X) = C( G). (cid:3) 14 4. Furstenberg's ×p, ×q problem via representation theory Let S, T : X → X be two commuting continuous maps on a compact Hausdorff space X. A Borel probability measure µ on X is called S, T - invariant if µ(S −1A) = µ(T −1A) = µ(A) for every Borel subset A of X. An S, T -invariant measure µ is called ergodic if every Borel set E with S −1E = E = T −1E satisfies that µ(E) = 0 or 1. Define maps Tp, Tq : T → T as Tp(z) = zp and Tq(z) = zq for all z ∈ T. A Borel probability measure µ on T is called ×p, ×q-invariant if it is Tp, Tq-invariant. A Borel set E ⊂ T is called ×p, ×q-invariant if A = TpA = TqA. We can define ×p, ×q maps Tp, Tq on Z[ 1 pq ] by Tp(g) = pg, Tq(g) = qg for every g ∈ Z[ 1 pq ]. Note that Tp and Tq are group automorphisms. Hence they induces group automorphisms on the dual group Spq of Z[ 1 pq ]. For convenience we also call them ×p, ×q maps on Spq. Denote the set of ×p, ×q-invariant measures on unit circle by Mp,q(T), the set of ergodic ×p, ×q-invariant measures on unit circle by EMp,q(T), the set of ×p, ×q-invariant measures on Spq by Mp,q(Spq), the set of er- godic ×p, ×q-invariant measures on Spq by EMp,q(Spq). 4.1. ×p, ×q-invariant measures on pq-solenoid and ×p, ×q-invariant measures on the unit circle. The following result is well-known for experts. For completeness we give a proof here. Proposition 4.1. When equipped with weak-∗ topologies, the re- striction map R : Mp,q(Spq) → Mp,q(T) defined by R(µ)(f ) = µ(f ) for µ ∈ Mp,q(Spq) and f ∈ C(T) is a homeomorphism. Also R restricts a homeomorphism from EMp,q(Spq) to EMp,q(T). Proof. Take µ ∈ Mp,q(Spq). Since C(T) is a C ∗-subalgebra of C(Spq), the restriction R(µ) of µ on C(T) belongs to Mp,q(T) and R is also continuous under the weak-∗ topology. Conversely, assume that µ ∈ Mp,q(T). Note that the group algebra ]) is a dense ∗-subalgebra of C(Spq) and define ν(zkpmqn) = µ(zk) CZ([ 1 pq 15 for n, m, k ∈ Z. By Bochner's Theorem [15, 1.4.3] ν is a Borel proba- bility measure on Spq iff {ν(zk)}k∈Z[ 1 pq ] is a positive definite sequence. For any finite subset F of Z[ 1 pq ], there exist positive integers k, l such that F ′ = pkqlF = {pkqlss ∈ F } is a finite subset of Z. Then we have ν((X = X λszs)∗(X ¯λsλtµ(zpkql(t−s)) = µ((X λsp−kq−lzs)∗(X λtzt)) = X λtp−kq−lzt)) > 0. s∈F t∈F s,t∈F ¯λsλtν(zt−s) s,t∈F s∈F ′ t∈F ′ Furthermore the ×p, ×q-invariance of ν follows from the definition. This shows that ν is in Mp,q(Spq). Moreover µ(zk) = ν(zk) for all k ∈ Z by the ×p, ×q-invariance of µ, hence µ is the restriction of ν on C(T), and this proves the subjectivity of R. On the other hand, if R(µ1) = R(µ2) for µ1, µ2 ∈ Mp,q(Spq), then µ1(zk) = µ2(zk) for all k ∈ Z. Since µ1 and µ2 are ×p, ×q-invariant, we have µ1(zkpmqn) = µ2(zkpmqn) for all n, m, k ∈ Z. This proves the injectivity of R. So R is a bijective continuous map between two compact Hausdorff spaces Mp,q(Spq) and Mp,q(T), this implies that R is a homeomorphism. Furthermore R is a homeomorphism from EMp,q(Spq) to EMp,q(T) since R is affine. (cid:3) Theorem 4.2. A representation π : C(Spq) ⋊ Z2 → B(H) is induced by a finitely supported ergodic ×p, ×q-invariant measure µ on T if and only if (1) π is irreducible; (2) HZ2 6= 0; (3) There exists nonzero N ∈ Z ⊂ Z[ 1 pq ] such that π(zN )x = x for every x ∈ HZ2. Proof. Suppose that π : C(Spq) ⋊ Z2 → B(H) is induced by a finitely supported ergodic ×p, ×q-invariant measure µ on T. Since both ×p and ×q maps have zero entropy with respect to µ, there is a representation πµ : C(Spq) ⋊ Z2 → B(L2(T, µ)) induced by µ (see Introduction for the definition of πµ). 16 By Proposition 4.1, ν = R−1(µ) is an ergodic ×p, ×q-invariant measure on Spq. Hence the representation ρν of C(Spq) ⋊ Z2 on L2(Spq, ν) is irreducible. Note that L2(T, µ) is a subspace of L2(Spq, ν) since µ is the restriction of ν from C(Spq) onto C(T). Also L2(T, µ) is a nonzero invariant subspace of L2(Spq, ν) under ρν. Hence L2(T, µ) = L2(Spq, ν). Hence πµ is unitarily equivalent to ρν. So π is irreducible and HZ2 6= 0. Moreover µ is finitely supported in a subset of { i i=0 ⊂ [0, 1) (here we identify T with [0, 1)). Hence µ(zN ) = 1 which implies that π(zN )x = x for every x ∈ HZ2. N }N −1 Conversely assume that π is an irreducible representation satisfying that HZ2 6= 0 and π(zN )x = x for a nonzero N ∈ Z and every x ∈ HZ2. Claim. H is finite dimensional. Proof. Take a unit y ∈ HZ2. Then Spanπ(Z[ 1 space of H under π. Since π is irreducible, we have pq ])y is an invariant sub- H = Span(π(Z[ 1 pq ])y). So it suffices to prove Span(π(Z[ 1 pq ])y) is finite dimensional. Firstly we prove that π(zM )y = y for a positive integer M coprime to pq. Without loss of generality we can assume N > 0. There exist nonneg- ative integers i, j, K, M such that KN = M piqj with M coprime to pq. Then π(zM piqj )y = π(zKN )y = y. Note that Z2 acts on Z[ 1 all m, n ∈ Z and every k ∈ Z[ 1 pq ] by ×p, ×q, that is, (m, n) · zk = zkpmqn for pq ]. Since y is in HZ2, we have π((i, j))π(zM )y = π(zM piqj )π((i, j))y = π(zM piqj )y = y, which implies π(zM )y = y. Secondly we prove that π(Z[ 1 pq ])y = π(Z)y. For all nonnegative integers i, j, there exists an integer l such that lpiqj = rM + 1 since M is coprime to pq. Hence for every positive 17 integer k, we have k pi qj )y = π(z (π(zM )y = y and y ∈ HZ2) π(z k(lpi qj −rM ) pi qj )y = π(zkl)π(z −krM piqj )y = π(zkl)y. This shows that π(Z[ 1 pq ])y ⊆ π(Z)y. Lastly we prove that Span(π(Z)y) is finite dimensional. Every k ∈ Z can be written as k = lN +r for some l ∈ Z and 0 ≤ r < N. Hence π(zk)y = π(zlN +r)y = π(zr)y. This implies that π(Z)y ⊂ Span{π(zi)y}N −1 i=0 . So dim H ≤ N. We finish proof of the claim. (cid:3) Define a state ψ on C(Spq) ⋊ Z2 by ψ(b) = hπ(b)y, yi for every b ∈ C(Spq) ⋊ Z2. We have ψ ∈ P 1(C(Spq) ⋊ Z2) since π is irreducible and y ∈ HZ2. By Theorem 3.2, ν = R(ψ) = ψC(Spq) is an ergodic ×p, ×q-invariant measure on Spq. By Theorem 3.9 and Theorem 3.10, we have π ∼ ρν. Of course H ∼= L2(Spq, ν). From the claim, L2(Spq, ν) is finite dimensional. Hence ν is finitely supported in Spq. Let µ = R(ν) = νC(Spq). As before, ρν is unitarily equivalent to πµ. Hence L2(T, µ) ∼ L2(Spq, ν) is finite dimensional. Hence µ is a finitely supported ergodic ×p, ×q- invariant measure on T and π ∼ πµ. (cid:3) Consequently we have the following. Corollary 4.3. Furstenberg's conjecture is true iff there is a unique irreducible unitary representation U : Z[ 1 pq ] ⋊ Z2 → B(H) such that HZ2 6= 0 and π(zk)x 6= x for every nonzero integer k and nonzero x ∈ HZ2. References [1] N. Brown and N. Ozawa. C ∗-algebras and Finite-dimensional Approxi- mations. Graduate Studies in Mathematics, 88. American Mathemat- ical Society, Providence, RI, 2008. [2] J. B. Conway. A Course in Functional Analysis. Second edition. Grad- uate Texts in Mathematics, 96. Springer-Verlag, New York, 1990. [3] J. Cuntz. Private communication. 18 [4] K. R. Davidson. C ∗-Algebras by Example. Fields Institute Mono- graphs, 6. American Mathematical Society, Providence, RI, 1996. [5] J. Dixmier. C ∗-Algebras. Translated from the French by Francis Jel- lett. North-Holland Mathematical Library, Vol. 15. North-Holland Publishing Co., 1977. [6] M. Einsiedler and E. Lindenstrauss. Diagonalizable flows on locally homogeneous spaces and number theory. International Congress of Mathematicians. Vol. II, 1731-1759, Eur. Math. Soc., Zurich, 2006. [7] H. Furstenberg. Disjointness in ergodic theory, minimal sets, and a problem in Diophantine approximation. Math. Systems Theory 1 (1967), 1-49. [8] E. Glasner. Ergodic Theory via Joinings. Mathematical Surveys and Monographs, 101. American Mathematical Society, Providence, RI, 2003. [9] A. S. A. Johnson. Measures on the circle invariant under multiplication by a nonlacunary subsemigroup of the integers. Israel J. Math. 77 (1992), no.1-2, 211-240. [10] Y. Katznelson, An Introduction to Harmonic Analysis. Third edition. Cambridge Mathematical Library. Cambridge University Press, Cam- bridge, 2004. [11] R. V. Kadison and J. R. Ringrose. Fundamentals of the Theory of Operator Algebras. Vol. I. Elementary Theory. Reprint of the 1983 original. Graduate Studies in Mathematics, 15. American Mathemat- ical Society, Providence, RI, 1997. [12] O. Lanford and D. Ruelle. Integral representations of invariant states on B ∗ algebras. J. Mathematical Phys. 8, (1967), 1460-1463. [13] A. M. Robert. A Course in p-adic Analysis. Graduate Texts in Math- ematics, 198. Springer-Verlag, New York, 2000. [14] D. W. Robinson and D. Ruelle. Extremal invariant states. Ann. Inst. H. Poincar Sect. A (N.S.) 6, (1967), 299-310. [15] W. Rudin, Fourier Analysis on Groups. Reprint of the 1962 original. John Wiley & Sons, Inc., New York, 1990. [16] D. J. Rudolph. ×2 and ×3 invariant measures and entropy. Ergod. Th. and Dynam. Syst. 10, (1990), 395 -- 406. [17] D. Ruelle. States of physical systems. Comm. Math. Phys. 3, (1966), 133-150. 19 [18] P. Walters, An Introduction to Ergodic Theory. Graduate Texts in Mathematics, 79. Springer-Verlag, New York-Berlin, 1982. [19] R. Willett and Y. Yu. Geometric property (T ). Chin. Ann. Math. Ser. B 35, (2014), no. 5, 761-800. E-mail address: [email protected] Jianchao Wu, Mathematisches Institut, Universitat Munster, Einste- instr. 62, Munster, 48149, Germany E-mail address: [email protected] 20
1505.02215	1	1505	2015-05-09T00:40:14	Extended de Finetti theorems for boolean independence and monotone independence	[ "math.OA", "math.PR" ]	We construct several new spaces of quantum sequences and their quantum families of maps in sense of So{\l}tan. Then, we introduce noncommutative distributional symmetries associated with these quantum maps and study simple relations between them. We will focus on studying two kinds of noncommutative distributional symmetries: monotone spreadability and boolean spreadability. We provide an example of a spreadable sequence of random variables for which the usual unilateral shift is an unbounded map. As a result, it is natural to study bilateral sequences of random objects, which are indexed by integers, rather than unilateral sequences. In the end of the paper, we will show Ryll-Nardzewski type theorems for monotone independence and boolean independence: Roughly speaking, an infinite bilateral sequence of random variables is monotonically(boolean) spreadable if and only if the variables are identically distributed and monotone(boolean) with respect to the conditional expectation onto its tail algebra. For an infinite sequence of noncommutative random variables, boolean spreadability is equivalent to boolean exchangeability.	math.OA	math	EXTENDED DE FINETTI THEOREMS FOR BOOLEAN INDEPENDENCE AND MONOTONE INDEPENDENCE WEIHUA LIU Abstract. We construct several new spaces of quantum sequences and their quantum families of maps in sense of So ltan. Then, we introduce noncommutative distributional symmetries associated with these quantum maps and study simple relations between them. We will focus on studying two kinds of noncommutative distributional symme- tries: monotone spreadability and boolean spreadability. We provide an example of a spreadable sequence of random variables for which the usual unilateral shift is an un- bounded map. As a result, it is natural to study bilateral sequences of random objects, which are indexed by integers, rather than unilateral sequences. In the end of the paper, we will show Ryll-Nardzewski type theorems for monotone independence and boolean independence: Roughly speaking, an infinite bilateral sequence of random variables is monotonically(boolean) spreadable if and only if the variables are identically distributed and monotone(boolean) with respect to the conditional expectation onto its tail algebra. For an infinite sequence of noncommutative random variables, boolean spreadability is equivalent to boolean exchangeability. 1. Introduction The characterization of random objects with distributional symmetries is an important object in modern probability and the recent context of Kallenberg [12] provides a compre- hensive treatment of distributional symmetries in classical probability. A finite sequence of random variables (ξ1, ξ2, ..., ξn) is said to be exchangeable if (ξ1, ..., ξn) d= (ξσ(1), ..., ξσ(n)), ∀σ ∈ Sn, where Sn is the permutation group of n elements and d= meas the joint distribution of the two sequences are the same. Compare with exchangeability, there is a weaker condition of spreadability: (ξ1, ..., ξn) is said to be spreadable if for any k < n, we have (1) (ξ1, ..., ξk) d= (ξl1, ..., ξlk), l1 < l2 < · · · < lk An infinite sequence of random variables is said to be exchangeable or spreadable if all its finite subsequences have this property. In the study of distributional symmetries in classical probability, one of the most important results is de Finetti's theorem which states that an infinite sequence of random variables, whose joint distribution is invariant under all finite permutations, is conditionally independent and identically distributed. Later, in [22], Ryll-Nardzewski showed that de Finetti theorem hold under the weaker condition of spreadability. Therefore, for infinite sequences of random variables in classical probability, spreadability is equivalent to exchangeability. 1 2 WEIHUA LIU Recently, Koslter [14] studied three kinds of distributional symmetries, which are sta- tionarity, contractability and exchangeablity, It was shown that exchangeability and spreadability do not characterize any universal inde- pendent relation in his framework. In addition, for infinite sequences, exchangeability is strictly stronger than spreadability in noncommutative probability. It should be pointed out that the framework in his paper is a W ∗-probability space with a faithful state. In this paper, we will consider our problems in a more general framework. in noncommutative probability. In the 1980's, Voiculescu developed his free probability theory and introduced a univer- sal independent relation, namely free independence, via reduced free products of unital C ∗-algebras[29]. For more details of free probability, the reader is referred to the mono- graph [28]. One can see that there is a deep parallel between classical probability and free probability. Recently, in [15], Koslter and Speicher extended this parallel to the aspect of distributional symmetries. In their work, by strengthening classical exchangeability to quantum exchangeability, they proved a de Finetti type theorem for free independence, i.e. for an infinite sequence of random variables, quantum exchangeability is equivalent to the fact that the the random variables are identically distributed and free with respect to the conditional expectation onto the tail algebra. The notion of quantum exchangeability is given by invariance conditions associated with quantum permutation groups As(n) of Wang [31]. This noncommutative de Finetti type theorem is an instance that free indepen- dence plays in the noncommutative world the same role as classical independence plays in the commutative world. It naturally raises a motivation for further study of noncommu- tative symmetries that "any result in classical probability should have an extension in free probability." For applications of this philosophy, see[2], [5], [6]. Especially, in [5], Curran introduced a quantum version of spreadability for free independence. It was shown that quantum spreadability is weaker than quantum exchangeability and is a characterization of free independence. More specifically, in a W ∗-probability space with a tracial faithful state, for an infinite sequence of random variables, quantum spreadability is equivalent to the fact that the the random variables are identically distributed and free with respect to the conditional expectation onto the tail algebra. In other words, quantum spreadability is equivalent to quantum exchangeability for infinite sequences of random variables in tra- cial W ∗-probability spaces. Another remarkable application of quantum exchangeability was given by Freslon and Weber [10]. They characterize Voiculescu's Bi-freeness [30] via certain invariance conditions associated with Wang's quantum groups As(n). In [26], Speicher and Woroudi introduced another independence relation which is called It was show that boolean independence is related to full free boolean independence. product of algebras [4] and boolean product is the unique non-unital universal product in noncommutative probability [25]. The study of distributional symmetries for boolean independence was started in [16]. We constructed a family of quantum semigroups in analogue with Wang's quantum permutation groups and defined their coactions on joint distributions of sequences. It was shown that the distributional symmetries associated those coactions can be used to characterize boolean independence in a proper framework. For more details about boolean independence and universal products, see [25]. It inspires us to study more distributional symmetries for boolean independence under the philos- ophy "any result in classical probability and free probability should have an extension NONCOMMUTATIVE SPREADABILITY 3 for boolean independence". In analogue with easy quantum groups in [1], we construct "easy "boolean semigroups and study their de Finetti type theorems in [17]. To apply our philosophy further, it is naturally to find an extended de Finetti type theorem for boolean independence. Specifically, we need to find the "noncommutative version of spreadability "for boolean independence and prove an extended de Finetti type theorem associated with the noncommutative spreadability. The main purpose of this paper is to study noncommutative versions of spreadability and extended de Finetti type theorems associated with them. Some other objects come into our consideration when we study spreadable sequences of random objects. It was shown in [19], there are two other universal products in noncom- mutative probability if people do not require the universal construction to be commuta- tive. We call the two universal products monotone and anti-monotone product. As tensor product, free product and boolean product, we can define monotone and anti-monotone independence associated with monotone and anti-monotone product. Monotone indepen- dence and anti-monotone independence are essentially the same but with different orders, i.e. if a is monotone with b, then b is anti-monotone with a. For more details of mono- tone independence, the reader is referred to [18], [21]. It is well known that a sequence of monotone random variables is not exchangeable but spreadable. Therefore, there should be a noncommutative spreadability which can characterize conditionally monotone inde- pendence. The first several sections devote to defining noncommutative distributional symmetries in analogue with spreadability and partial exchangeability. Recall that in [2] [5], noncom- mutative distributional symmetries are defined via invariance conditions associated with certain quantum structures. For instance, Curran's quantum spreadability is described by a family of quantum increasing sequences and their quantum family of maps in sense of Soltan. The family of quantum increasing sequences are universal C ∗-algebras Ai(n, k) generated by the entries of a n × k matrix which satisfy certain relations R. Following the idea in [16], to construct a boolean type of spaces of increasing sequences Bi(n, k), we replace the unit partition condition in R by an invariant projection condition. Recall that in In [9], Franz studied relations between freeness, monotone independence and boolean independence via Bozejko, Marek and Speicher's two-state free products[3]. In his con- struction, monotone product is something "between" free product and boolean product. Thereby, we construct the noncommutative spreadability for monotone independence by modifying quantum spreadability and our boolean spreadability. We will study simple relations between those distributional symmetries, i.e. which one is stronger. As the situation for boolean independence, there is no nontrivial pair of monotonically independent random variables in W ∗-probability spaces with faithful states. Therefore, the framework we use in this paper is a W ∗-probability space with a non-degenerated normal state which gives a faithful GNS representation of the probability space. In this framework, we will see that spreadability is too weak to ensure the existence of a con- ditional expectation. Recall that, in W ∗-probability spaces with faithful states, we can define a normal shift on a unilateral infinite sequence of spreadable random variables. 4 WEIHUA LIU Here, "unilateral"means the sequence is indexed by natural numbers N. An important property of this shift is that its norm is one. Therefore, given an operator, we can con- struct a WOT convergent sequence of bounded variables via shifts. The is the key step to construct a normal conditional expectation in previous works. But, in W ∗-probability spaces with non-degenerated normal states, the unilateral shift of spreadable random vari- ables is not necessarily norm one. An example is provided in the beginning of section 6. Actually, the sequence of random variables are monotonically spreadable which is an in- variance condition stronger than classical spreadability. Therefore, we can not construct a conditional expectation, for unilateral sequences, via shifts under the condition of spread- ability. To fix this issue, we will consider bilateral sequences of random variables instead of unilateral sequences. "bilateral"means that the sequences are indexed by integers Z. In this framework, we will see that the shift of spreadable random variables is norm one so that we can define a conditional expectation via shifts by following Kostler's construction. Notice that the index set Z has two infinities, i.e. the positive infinity and the negative infinity. Therefore, we will have two tail algebras with respect to the two infinities and will define two conditional expectations consequently. We denote by E+ the conditional expectation which shifts indices to the positive infinity and E− the conditional expecta- tion which shifts indices to negative infinity. We will see that the two tail algebras are subsets of fixed points of the shift and the conditional expectations may not be extended normally to the whole algebra. In general, the two tail algebras are different and the conditional expectation may have different properties. To noncommutative spreadability for monotone independence, we have the following: Theorem 1.1. Let (A, φ) be a non degenerated W ∗-probability space and (xi)i∈Z be a bilat- eral infinite sequence of selfadjoint random variables which generate A as a von Neumann algebra. Let A+ k be the WOT closure of the non-unital algebra generated by {xii ≥ k}. Then the following are equivalent: a) The joint distribution of (xi)i∈Z is monotonically spreadable. b) For all k ∈ Z, there exits a φ preserving conditional expectation Ek : A+ k → tail such that the sequence (xi)i≥k is identically distributed and monotonically A+ independent with respect Ek. Moreover, EkAk′ = Ek′ when k ≥ k′. In general, we can not extend E+ to the whole algebra A, but we have the following: Proposition 1.2. Let (A, φ) be a non degenerated W ∗-probability space and (xi)i∈Z be a bilateral infinite sequence of selfadjoint random variables which generate A as a von Neumann algebra. If the joint distribution of (xi)i∈Z is monotonically spreadable, then E− can be extend to the whole algebra A normally. We will see that boolean spreadability implies monotone spreadability and anti-monotone spreadability. Therefore, both E+ and E− can be extended normally to the whole alge- bra A. Moreover, for boolean spreadable sequences, E+ = E− and the two algebras are identical. In summary, we have Theorem 1.3. Let (A, φ) be a non degenerated W ∗-probability space and (xi)i∈Z be a bilat- eral infinite sequence of selfadjoint random variables which generate A as a von Neumann algebra. Then the following are equivalent: NONCOMMUTATIVE SPREADABILITY 5 a) The joint distribution of (xi)i∈N is boolean spreadable. b) The sequence (xi)i∈Z is identically distributed and boolean independent with respect to the φ−preserving conditional expectation E onto the non unital tail algebra of the (xi)i∈Z The paper is organized as follows: In section 2, we will introduce preliminaries and no- tation from noncommutative probability and recall Wang's quantum permutation groups and boolean quantum semigroups. In section 3, we briefly review distributional symme- tries for finite sequences of random variables in classical probability and we restate these symmetries in words of quantum maps. Then, we introduce noncommutative versions of these symmetries and their quantum maps. In the end of this section, we will define quantum spreadability, monotone spreadability and boolean spreadability for bilateral in- finite sequences of random variables. In section 4, we will study simple relations between our noncommutative symmetries. In particular, we will show boolean exchangeability is strictly stronger than boolean spreadability. Therefore, operator-valued boolean in- dependent random variables are boolean spreadable. In section 5, we will introduce an equivalence relation on the set of sequences of indices. With the help of the equivalence relation, we will show that operator-valued monotone independent sequences of random variables are monotonically spreadable. In section 6, we first provide an example that a monotonically spreadable unilateral sequence of bounded random variables is unbounded. Therefore, we cannot define conditional expectation for unilateral spreadable sequences via shifts in a W ∗-probability space with a non-degenerated normal state. Then we will turn to study bilateral sequences of random variables. We will introduce tail algebras associated with positive infinity and negative infinity and study elementary properties of conditional expectations associated with the two tail algebras. In section 7, we will study properties of conditional expectations under the assumption that our bilateral sequences are monotonically spreadable. In section 8, we will prove a Ryll-Nardzewski type theorem for Monotone independence. In section 9, we will prove a Ryll-Nardzewski type theorem for boolean independence. 2. Preliminaries and examples We recall some necessary definitions and notions from noncommutative probability. For further details, see contexts [15], [20], [28], [21]. Definition 2.1. A non-commutative probability space (A, φ) consists of a unital algebra A and a linear functional φ : A → C such that φ(1A) = 1. (A, φ) is called a ∗-probability space if A is a ∗-algebra and φ(xx∗) ≥ 0 for all x ∈ A. (A, φ) is called a W ∗- probability space if A is a W ∗-algebra and φ is a normal state on it. We will not assume that φ is faithful. The elements of A are called random variables. Let x ∈ A be a random variable, then its distribution is a linear functional µx on C[X]( the algebra of complex polynomials in one variable), defined by µx(P ) = φ(P (x)). Definition 2.2. Let A be a W ∗-algebra, a normal state φ on A is said to be non- degenerated if x = 0 whenever φ(axb) = 0 for all a, b ∈ A. By proposition 7.1.15 in [11], if φ is a non-degenerated normal state on A then the GNS representation associated to φ is faithful. In this paper, we will work with W ∗-probability 6 WEIHUA LIU space with a non-degenerated normal state. The reason is that there is no non-trivial pair of boolean or monotonically independent random variables in W ∗-probability spaces with faithful states. See [16]. X k2 i2 · · · X kn Definition 2.3. Let I be an index set. The algebra of noncommutative polynomials in I variables, ChXii ∈ Ii, is the linear span of 1 and noncommutative monomials of the form X k1 in with i1 6= i2 6= · · · 6= in ∈ I and all kj's are positive integers. i1 For convenience, we will denote by ChXii ∈ Ii0 the set of noncommutative polynomials without a constant term. Let (xi)i∈I be a family of random variables in a noncommutative probability space (A, φ). Their joint distribution is the linear functional µ : ChXii ∈ Ii → C defined by and µ(1) = 1. µ(X k1 i1 X k2 i2 · · · X kn in ) = φ(xk1 i1 xk2 i2 · · · xkn in ), In general, the joint distribution depends on the order of the random variables, e.g µx,y may not equal µy,x. According to our notation, µx,y(X1X2) = φ(xy), but µy,x(X1X2) = φ(yx). In this paper, our index set I is always an ordered set with order "> "e.g. N, Z. Definition 2.4. Let (A, φ) be a noncommutative probability space. A family of (not necessarily unital) subalgebras {Aii ∈ I} of A is said to be boolean independent if φ(x1x2 · · · xn) = φ(x1)φ(x2) · · · φ(xn) whenever xk ∈ Ai(k) with i(1) 6= i(2) 6= · · · 6= i(n). The family of subalgebras {Aii ∈ I} is said to be monotonically independent if φ(x1 · · · xk−1xkxk+1 · · · xn) = φ(xk)φ(x1 · · · xk−1xk+1 · · · xn) whenever xj ∈ Aij with i1 6= i2 6= · · · 6= in and ik−1 < ik > ik+1. A set of random variables {xi ∈ Ai ∈ I} is said to be boolean(monotonically) independent if the family of non- unital subalgebras Ai, which are generated by xi respectively, is boolean(monotonically) independent. One refers to [8] for more details of boolean product and monotone product of random variables. In general, the framework for boolean independence and monotone indepen- dence is a non-unital algebra. Thereby, we will use the following version of operator valued probability spaces: Definition 2.5. Operator valued probability space An operator valued probability space (A, B, E : A → B) consists of an algebra A, a subalgebra B of A and a B − B bimodule linear map E : A → B, i.e. E[b1ab2] = b1E[a]b2, E[b] = b for all b1, b2, b ∈ B and a ∈ A. According to the definition in [27], we call E a conditional expectation from A to B if E is onto, i.e. E[A] = B. The elements of A are called random variables. Remark 2.6. In free probability theory, A and B are assumed to be unital and share the same unit NONCOMMUTATIVE SPREADABILITY 7 Definition 2.7. For an algebra B, we denote by BhXi the algebra which is freely gen- erated by B and the indeterminant X. Let 1X be the identity of ChXi, then BhXi is set of linear combinations of the elements in B and the noncommutative monomials b0Xb1Xb2 · · · bn−1Xbn where bk ∈ B ∪ {C1X} and n ≥ 0. The elements in BhXi are called B-polynomials. In addition, BhXi0 denotes the subalgebra of BhXi which does not contain a constant term in B, i.e. the linear span of the noncommutative monomials b0Xb1Xb2 · · · bn−1Xbn where bk ∈ B ∪ {C1X} and n ≥ 1. Here are the operator valued versions of noncommutative independences: Definition 2.8. Let {xi}i∈I be a family of random variables in an operator valued prob- ability space (A, B, E : A → B), where A and B are not necessarily unital. {xi}i∈I is said to be boolean independent over B if E[p1(xi1)p2(xi2) · · · pn(xin)] = E[p1(xi1)]E[p2(xi2)] · · · E[pn(xin)] whenever i1, · · · , in ∈ I, i1 6= i2 6= · · · 6= in and p1, · · · , pn ∈ BhXi0. {xi}i∈I is said to be monotonically independent over B if E[p1(xi1) · · · pk−1(xik−1)pk(xik)pk+1(xik+1) · · · pn(xin)] = E[p1(xi1) · · · pk−1(xik−1)E[pk(xik)]pk+1(xik+1) · · · pn(xin)] whenever i1, · · · , in ∈ I, i1 6= i2 6= · · · 6= in, ik−1 < ik > ik+1 and p1, · · · , pn ∈ BhXi0. Notice that there is another natural order "< "on I, i.e. a < b if b > a. Therefore, we can define another noncommutative independence relation. {xi}i∈I is said to be anti- monotonically independent with respect to E and index order ">" if {xi}i∈I is said to be anti-monotonically independent with respect to E and index order "<". See more details in [19]. 2.1. Noncommutative distributional symmetries. Recall that, in [32], Wang intro- duced the following quantum analogue of permutation groups: Definition 2.9. As(n) is defined as the universal unital C ∗-algebra generated by elements (ui,j)i,j=1,···n such that we have • Each ui,j is an orthogonal projection, i.e. u∗ • The elements in each row and column of u = (uij)n ij = uij = u2 i.e. are orthogonal and sum up to 1: for each i = 1, · · · , n and k 6= l we have ij for all i, j = 1, · · · , n. i,j=1,··· ,n form a partition of unit, and for each i = 1, · · · , n we have uikuil = 0 and ukiuli = 0; . n Xk=1 uik = 1 = n Xk=1 uki. As(n) is a compact quantum group in sense of Woronowicz [33], with comultiplication, counit and antipode given by the formulas: n ∆ui,j = Xk=1 ui,k ⊗ uk,j 8 WEIHUA LIU ǫ(ui,j) = δi,j S(ui,j) = uj,i It was shown that the this quantum structure can be used to characterize conditionally free independence [15]. In [16], we modify the universal conditions of Wang's quantum permutation groups: By replacing the condition associated with partitions of the unit by a condition associated with an invariant projection, we get the following universal algebras: Quantum semigroups (Bs(n), ∆): The algebra Bs(n) is defined as the universal unital C ∗-algebra generated by elements ui,j (i, j = 1, · · · n) and a projection P such that we have • each ui,j is an orthogonal projection, i.e. u∗ • i,j = ui,j = u2 i,j for all i, j = 1, · · · , n. ui,kui,l = 0 and uk,iul,i = 0 whenever k 6= l. n • For all 1 ≤ i ≤ n, P = uk,iP. Pk=1 There is a natural comultiplication ∆ : Bs(n) → Bs(n) ⊗min Bs(n) defined by ∆ui,j = n Xk=1 ui,k ⊗ uk,j, ∆(P) = P ⊗ P, ∆(I) = I ⊗ I, where ⊗min stands for the reduced C ∗-tensor product. The existence of of these maps is given by the universal property of Bs(n). Therefore, (Bs(n), ∆)'s are quantum semigroups in sense of So ltan [24]. These quantum structures can characterize conditionally boolean independence, see more details in [16]. 3. Distributional symmetries for finite sequences of random variables In this section, we will review two kinds of distributional symmetries which are spread- ability and partial exchangeability, in classical probability. In [12], we see that the dis- tributional symmetries can be defined for either finite sequences or infinite sequences. Moreover, each kind of distributional symmetry for infinite sequences of random objects is determined by distributional symmetries on all its finite subsequences. For example, an infinite sequence of random variables is exchangeable iff all its finite subsequences are exchangeable. We will present distributional symmetries for finite sequences and then introduce their counterparts in noncommutative probability. In the first subsection, we recall notions of spreadability and partial exchangeability in classical probability and rephrase these notions in words of quantum maps. In the second subsection, we will introduce counterparts of spreadability and partial exchangeability in noncommutative probability. Even though there are many interesting properties of partial exchangeability, we are not going to study it too much in this paper because the main problem we concern is about extended de Finetti type theorems for noncommutative spreadable sequences of random variables. We will discuss relations between those noncommutative distributional symmetries in the next section. NONCOMMUTATIVE SPREADABILITY 9 3.1. Spreadability and partial exchangeability. Recall that in [13], a finite sequence of random variables (x1, ..., xn) is said to be spreadable if for any k < n, we have (2) (x1, ..., xk) d= (xl1, ..., xlk), l1 < l2 < · · · < lk For fixed natural numbers n > k, it is mentioned in [5], the above relation can be described in words of quantum family of maps in sense of Soltan [23]: Considering the space Ik,n of increasing sequences I = (1 ≤ i1 < · · · < ik ≤ n). For 1 ≤ i ≤ n, 1 ≤ j ≤ k, define fi,j : Ik,n → C by: fi,j(I) = (cid:26) 1, 0, ij = i ij 6= i . If we consider In,k as a discrete space, then the functions fi,j generate C(In,k) by the Stone- Weierstrass theorem. Let C[X1, ..., Xm] be the set of commutative polynomials in m vari- ables. The algebra C(In,k) together with an algebraic homomorphism α : C[X1, ..., Xk] → C[X1, ..., Xn] ⊗ C(Ik,n) define by: α : Xj = n Xi=1 Xi ⊗ fi,j, α(1) = 1C(Ik,n) defines a quantum family of maps from {1, ..., k} to {1, ..., n}. We can use this family of quantum maps to rephrase equation (2): Let µx1,....,xn be the joint distribution of (x1, ..., xn). For fixed natural numbers n > k, µx1,...,xk(p)1C(In,k) = µx1,...,xn ⊗ idC(In,k)(α(p)) (3) j1 · · · X im for all p ∈ C[x1, ..., xk]. For completeness, we provide a sketch of proof here: Suppose equation (2) holds. Let p = X i1 jm be a monomial in C[X1, ..., Xk] such that 1 ≤ j1 < j2 < · · · < jm ≤ k and i1, ..., im are positive integers. Let I = (1 ≤ l1 < · · · < lk ≤ n) be a point in Ik,n. Then, the I-th component of µx1,...,xk(p)1C(In,k) is E[xi1 jm]. The I-th componentµx1,...,xn ⊗ idC(In,k)(α(p)) is j1 · · · xim n Xs1,...,sm=1 E[xi1 s1 · · · xim sm](fs1,j1 · · · fsm,jm)(I). According the definition of fi,j, (fs1,j1 · · · fsm,jm)(I) is not vanished only if st = ljt for all 1 ≤ t ≤ m. Therefore, n Xs1,...,sm=1 E[xi1 s1 · · · xil sl](fs1,j1 · · · fsm,jm)(I) = E[xi1 lj1 · · · xim ljm ]. Since 1 ≤ j1 < j2 < · · · < jm ≤ k and I is an in creasing sequence, we have 1 ≤ lj1 < · · · < ljm ≤ n. Hence, the I-components of the two sides of equation (5) are equal to each other. Since I is arbitrary, equation (5) holds. By checking the I component of equation 5, we can also show that (5) implies (2). We will say that (ξ1, ..., ξn) is (n, k)-spreadable if (x1, ..., xn) satisfies equation (5). Remark 3.1. We see that the above (n, k)-spreadability describes limited relations be- tween the mixed moments of (x1, ..., xn). Once we fix n, k, the (n, k)-spreadability gives no information about mixed moments which involve k + 1 variables. For example, let 10 WEIHUA LIU n = 4, k = 2 and assume that (x1, ..., x4) is a (4, 2)-spreadable sequence. According to equation (2), we know nothing about the relation between E[x1x2x3] and E[x2x3x4]. We will call this kind of distributional symmetries partial symmetries because they just provide information of part of mixed moments but not all. By using the idea of partial symmetries, we can define another family of distributional symmetries which is stronger than (n, k)-spreadability but weaker than exchangeability. Definition 3.2. For fixed natural numbers n > k, we say a sequence of random variables (x1, ..., xn) is (n, k)-exchangeable if (x1, ..., xk) d= (xσ(1), ..., xσ(k)), ∀σ ∈ Sn, where Sn is the permutation group of n elements. This kind of exchangeability is called partial exchangeability. For more details, see [7]. As well as (n, k)-spreadability, we can rephrase partial exchangeability in words of quan- tum family of maps: Considering the space En,k of length k sequences {I = (i1, ..., ik)1 ≤ i1, ..., ik ≤ n, ij 6= ij ′ for j 6= j′}. For 1 ≤ i ≤ n, 1 ≤ j ≤ k, define gi,j : In,k → C by: gi,j(I) = (cid:26) 1, 0, ij = i, ij 6= i. Given two different sequences I = (i1, ..., ik) and I ′ = (i′ k), there must exists a num- ber j such that ij 6= i′ j. Then, we have that gi,ij (I) = 1 6= 0 = gi,ij (I). Therefore, the set of functions {gi,ji = 1, ..., n; j = 1, ..., k} separates En,k. According to Stone Weierstrass theorem, the functions gi,j generate C(En,k). Again, we can define a homomorphism α′ : C[X1, ..., Xk] → C[X1, ..., Xn] ⊗ C(En,k) by the following formulas: 1, ..., i′ α′ : Xj = n Xi=1 Xi ⊗ gi,j, α′(1) = 1C(Ik,n). Lemma 3.3. Let µx1,....,xn be the joint distribution of x1, ..., xn. Then µx1,...,xk(p)1C(In,k) = µx1,...,xn ⊗ idC(In,k)(α(p)) for all p ∈ C[X1, ..., Xk] if and only if x1, ..., xn is (n, k) exchangeable. The proof is similar the proof of (n, k)-spreadability, we just need to check the values at all components of En,k. 3.2. Noncommutative analogue of partial symmetries. Now, we turn to introduce noncommutative versions of spreadability and partial exchangeability. The pioneering work was done by Curran [5]. He defined a quantum version of C(In,k) in analogue of Wang's quantum permutation groups as following: Definition 3.4. For k, n ∈ N with k ≤ n, the quantum increasing space A(n, k) is the universal unital C ∗−algebra generated by elements {ui,j1 ≤ i ≤ n, 1 ≤ j ≤ k} such that i,j for all i = 1, ..., n; j = 1. Each ui,j is an orthogonal projection: ui,j = u∗ i,j = u2 1, ..., k. NONCOMMUTATIVE SPREADABILITY 11 2. Each column of the rectangular matrix u = (ui,j)i=1,...,n;j=1,...,k forms a partition of n unity: for 1 ≤ j ≤ k we have ui,j = 1. Pi=1 3. Increasing sequence condition: ui,jui′,j ′ = 0 if j < j′ and i ≥ i′. Remark 3.5. Our notation is different from Curran's, we use Ai(n, k) instead of his Ai(k, n) for our convenience. For any natural numbers k < n, in analogue of coactions of As(n), there is a unital ∗-homomorphism αn,k : ChX1, ..., Xki → ChX1, ..., Xni ⊗ Ai(n, k) determined by: αn,k(Xj) = n Xi=1 Xi ⊗ ui,j. The quantum spreadability of random variables is defined as the following: Definition 3.6. Let (A, φ) be a noncommutative probability space. A finite ordered sequence of random variables (xi)i=1,...,n in A is said to be Ai(n, k)-spreadable if their joint distribution µx1,...,xn satisfies: µx1,...,xn(p)1Ai(n,k) = µ ⊗ idAi(n,k)(αn,k(p)), (xi)i=1,...,n is said to be quantum spreadable if (xi)i=1,...,n is for all p ∈ ChX1, ..., Xki. Ai(n, k)-spreadable for all k = 1, ..., n − 1. Remark 3.7. In [5], Curran studied sequences of C ∗-homomorphisms which are more general than random variables. For consistency, we state his definitions in words of random variables. It is a routine to extend our work to the framework of sequences of C ∗-homomorphisms. Recall that in [16], by replacing the condition associated with partitions of the unity of Wang's quantum permutation groups, we defined a family of quantum semigroups with invariant projections. With a natural family of coactions, we defined invariance conditions which can characterize conditional boolean independence. Here, we can modify Curran's quantum increasing spaces in the same way: Definition 3.8. For k, n ∈ N with k ≤ n, the noncommutative increasing space Bi(k, n) is the unital universal C ∗−algebra generated by elements {u(b) i,j 1 ≤ i ≤ n, 1 ≤ j ≤ k} and an invariant projection P such that 1. Each u(b) i,j is an orthogonal projection:u(b) i,j = (u(b) i,j )∗ = (u(b) i,j )2 for all i = 1, ..., n; j = 1, ..., k. n 2. For 1 ≤ j ≤ k we have Pi=1 u(b) i,j P = P. i,j u(b) 3. Increasing sequence condition: u(b) i′,j ′ = 0 if j < j′ and i ≥ i′. The same as Ai(n, k), there is a unital ∗-homomorphism α(b) n,k : ChX1, ..., Xki → ChX1, ..., Xni⊗ Bi(n, k) determined by: α(b) n,k(xj) = n Xi=1 xi ⊗ u(b) i,j As boolean exchangeability defined in [16], we have 12 WEIHUA LIU Definition 3.9. A finite ordered sequence of random variables (xi)i=1,...,n in (A, φ) is said to be Bi(n, k)-spreadable if their joint distribution µx1,...,xn satisfies: µx1,...,xn(p)P = Pµ ⊗ idBi(n,k)(α(b) n,k(p))P, for all p ∈ ChX1, ..., Xki. Bi(n, k)-spreadable for all k = 1, ..., n − 1. (xi)i=1,...,n is said to be boolean spreadable if (xi)i=1,...,n is We will see that Bi(k, n) is an increasing space of boolean type, because we can derive an extended de Finetti type theorem for boolean independence. Recall that, in [9], Franz showed some relations between free independence, mono- tone independence and boolean independence via Bozejko, Marek and Speicher's two- states free products[3]. We can see that monotone product is "between" free product and boolean product. From this viewpoint of Franz's work, we may hope to define a kind of "spreadability"for monotone independence by modifying quantum spreadability and boolean spreadability. Notice that there are at least two ways to get quotient algebras of Bi(k, n)'s such that the P-invariance condition of the quotient algebras is equivalent quantum spreadability: 1. Require P to be the unit of the algebra. 2. Let Pj = ui,j, require Pj ′uij = uijPj ′ for all 1 ≤ j, j′ ≤ k and 1 ≤ i ≤ n. n Pi=1 To define the monotone increasing spaces, we modify the second condition a little: Definition 3.10. For fixed n, k ∈ N and k < n, a monotone increasing sequence space Mi(n, k) is the universal unital C ∗-algebra generated by elements {u(m) i,j }i=1,...,n;j=1,...,k 1. Each ui,j is an orthogonal projection; u(m) i,j , Pju(m) 2. Monotone condition: Let Pj = n i′j ′ = ui′j ′ if j′ ≤ j. n 3. u(m) i,j P1 = P1 for all 1 ≤ j ≤ k. Pi=1 4. Increasing condition: u(m) i,j u(m) i′,j ′ = 0 if j < j′ and i ≥ i′. Pi=1 We see that P1 plays the role as the invariant projection P in the boolean case. For consistency, we denote P1 by P. Then, we can define a P-invariance condition associated with Mi(n, k) in analogy with Bi(n, k): For fixed n, k ∈ N and k < n, there is a unique unital ∗- isomorphism α(m) n,k : ChX1, ..., Xki → ChX1, ..., Xni ⊗ Mi(n, k) such that α(m) n,k (Xj) = n Xi=1 Xi ⊗ u(m) i,j . The existence of such a homomorphism is given by the universality of ChX1, ..., Xki. Definition 3.11. A finite ordered sequence of random variables (xi)i=1,...,n in (A, φ) is said to be Mi(n, k)-invariant if their joint distribution µx1,...,xn satisfies: µx1,...,xk(p)P = Pµx1,...,xn ⊗ idMi(n,k)(α(m) n,k (p))P, for all p ∈ ChX1, ..., Xki. (xi)i=1,...,n is said to be monotonically spreadable if it is Mi(n, k)- invariant for all k = 1, ..., n − 1. NONCOMMUTATIVE SPREADABILITY 13 We will see that these invariance conditions can characterize conditionally Monotone independence in a proper framework. As remark 2.3 in [5], a first question to our definitions is whether Ai(n, k), Bi(n, k), Mi(n, k) exist. In [5], Curran has showed several nontrivial representations of Ai(n, k). In the following, we provide a family of presentations of Ai(n, k), Bi(n, k), Mi(n, k) for n > k: Fix natural numbers n > k, let l1, ..., lk ∈ N such that l1 + · · · + lk = n. We denote by Hi a li-dimensional Hilbert spaces with orthonormal basis {e(i) Let Ili be the unit of the algebra B(Hli), P jection onto Ce(li) j j = 1, ..., li}. be the one dimensional orthogonal pro- . Consider the (li) j , Pi be the one dimensional projection onto CPj e(li) j e j following matrix: P1,1 ... P1,l1 0 ... 0 0 ... 0 ... 0   0 ... 0 P2,1 ... P2,l2 0 ... 0 ... 0 · · · 0 ... . . . · · · 0 · · · 0 ... . . . · · · 0 · · · 0 ... . . . · · · Pk,1 ... . . . · · · Pk,lk .   We see that the entries of the matrix satisfy the increasing condition of spaces of increas- ing sequences. By choosing proper projections Pi,j, we will get representations for our universal algebras: Quantum family of increasing sequences: For each 1 ≤ j ≤ k, the algebra generated by {Pei Zli is a quotient algebra of Ai(n, k). One can define a C ∗-homomorphism π from Ai(n, k), such that i = 1, ..., lj} is isomorphic to C ∗(Zlj ). The reduced free product ∗k j=1 j the image of P in ∗k j=1 C ∗(Zlj ) if 0 < j′ = j − e (li) j′ if otherwise lm ≤ li i−1 Pl=m k 0 π(ui,j) =   Ni=1 π(ui,j) =   Boolean family of increasing sequences: One can define a C ∗-homomorphism π from Bi(n, k) into B( Hi) such that i−1 Nm1=1 0 Plm1 ⊗ P e (li) j′ k Nm2=i+1 Plm2 if 0 < j′ = j − if otherwise lm ≤ li i−1 Pl=m 14 WEIHUA LIU Monotone family of increasing sequences: One can define a C ∗-homomorphism π from k Mi(n, k) to B( Hi) Ni=1 π(ui,j) =   i−1 Nm1=1 0 Ilm1 ⊗ P e (li) j′ k Nm2=i+1 Plm2 if 0 < j′ = j − if otherwise lm ≤ li i−1 Pl=m The existence of these homomorphisms are given by the universal conditions for Ai(n, k), Bi(n, k) and Mi(n, k) respectively. Since the above representation of Mi(n, k) plays an important role in proving our main theorems, we summarize it as the following proposition: Proposition 3.12. For fixed natural numbers n > k. Let l1, ..., lk ∈ N such that l1 + · · · + lk = n. Let Hi be a li-dimensional Hilbert spaces with orthonormal basis {e(i) j j = 1, ..., li} and Ili be the unit of the algebra B(Hli), P be the one dimensional orthogonal projection onto Ce(li) . Then, there is a C ∗- e(li) (li) j e j homomorphism π : Mi(n, k) → B(H1 ⊗ · · · ⊗ Hk) defined as follows: j , Pi be the one dimensional projection onto CPj π(ui,j) =   Nm2=i+1 Nm1=1 Plm2 Ilm1 ⊗ P (li) j′ i−1 0 k e Also, we need the following property in the future: if 0 < j′ = j − if otherwise lm ≤ li i−1 Pl=m Lemma 3.13. Given natural numbers n1, n2, n, k ∈ N such that n > k. Let (ui,j)i=1,...,n;j=1,...,k be the standard generators of Mi(n, k) and (u′ i,j)i=1,...,n+n1+n2;j=1,...,k+n1+n2 be the standard generators of Mi(n + n1 + n2, k + n1 + n2). Then, there exists a C ∗-homomorphism π : Mi(n + n1 + n2, k + n1 + n2) → Mi(n, k) such that π(u′ i,j) = δi,jP δi,j 0 δi−n,j−kI 1 ≤ i ≤ n1 n1 + 1 ≤ i ≤ n + n1, n1 ≤ j ≤ n1 + k if if if n1 + 1 ≤ i ≤ n + n1, j ≤ n1 or j > n1 + k if i ≥ n + n1 + 1 where P = P1 = ui,1 and I is the identity of Mi(n, k). Proof. We can see that the matrix form of (π(u′ i,j))i=1,...,n+n1+n2;j=1,...,k+n1+n2 is   Pi=1 n 0 ... P · · · 0 ... ... . . . 0 · · · P 0 0 · · · 0 u1,1 ... ... ... . . . 0 · · · 0 un,1 0 · · · 0 0 ... ... ... . . . 0 · · · 0 0   0 ... 0 0 · · · · · · ... . . . . . . · · · 0 · · · · · · u1,k 0 · · · . . . . . . · · · un,k 0 · · · · · · · · · I ... . . . . . . 0 · · · · · · 0 ... 0 ... ... 0 ... 0 0 ... 0 0 ... I   NONCOMMUTATIVE SPREADABILITY 15 It is easy to check that the coordinates of the above matrix satisfy the universal conditions of Mi(n + n1 + n2, k + n1 + n2). The proof is complete. (cid:3) In analogue of the (n, k)-partial exchangeability, we can define noncommutative versions of partial exchangeability for free independence and boolean independence: Definition 3.14. For k, n ∈ N with k ≤ n, the quantum space Al(n, k) is the universal unital C ∗-algebra generated by elements {uij1 ≤ i ≤ n, 1 ≤ j ≤ k} such that 1. Each uij is an orthogonal projection:uij = u∗ 2. Each column of the rectangular matrix u = (uij) forms a partition of unity: for ij = u2 ij. n 1 ≤ j ≤ k we have uij = 1. Pi=1 Remark 3.15. Ai(n, k) is a quotient algebra of Al(n, k), because the definition of Ai(n, k) has one more restriction than Al(n, k)'s. Al(n, n) is exactly Wang's quantum permutation group As(n). There is a well defined unital algebraic homomorphism α(f p) n,k : ChX1, ...Xki → ChX1, ...Xni ⊗ Al(n, k) such that α(f p) n,k Xj = n Xi=1 Xi ⊗ ui,j where 1 ≤ j ≤ k. The distributional symmetry associated with this quantum structure is: Definition 3.16. Let x1, ..., xn ∈ (A, φ) be a sequence of n-noncommutative random variables, k ≤ n be a positive integer. We say the sequence is (n, k)-quantum exchangeable if µx1,...xk(p) = µx1,...xn ⊗ idAl(n,k)(α(f p) n,k (p)), for all p ∈ ChX1, ..., Xki, where µx1,...,xj is the joint distribution of x1, ...xj with respect to φ for j = k, n. By modifying the second universal condition of Al(n, k), we can define a boolean version of partial exchangeability: Definition 3.17. For natural numbers k ≤ n, Bl(n, k) is the non-unital universal C ∗- algebra generated by the elements {ui,j}i=1,...,n;j=1,...,k and an orthogonal projection P, such that (1) ui,j is an orthogonal projection, i.e. ui,j = u∗ (2) n Pi=1 ui,jP = P for all 1 ≤ j ≤ k. i,j = u2 i,j. Remark 3.18. Bl(n, n) is exactly the boolean exchangeable quantum semigroup Bs(n). There is a well defined unital algebraic homomorphism α(bp) n,k : ChX1, ...Xki → ChX1, ...Xni ⊗ Bl(n, k) 16 such that WEIHUA LIU α(bp) n,k Xj = n Xi=1 Xi ⊗ ui,j where 1 ≤ j ≤ k. The distributional symmetry associated with this quantum structure is: Definition 3.19. Let x1, ..., xn ∈ (A, φ) be a sequence of n-noncommutative random variables, k ≤ n be a positive integer. We say the sequence is (n, k)-boolean exchangeable if µx1,...xk(p)P = Pµx1,...xn ⊗ idBl(n,k)(α(bp) n,k (p))P for all p ∈ ChX1, ..., Xki, where µx1,...,xj is the joint distribution of x1, ...xj with respect to φ. Now, we turn to define our noncommutative distributional symmetries for infinite se- In this paper, our infinite ordered index set I would be quences of random variables. either N or Z. Definition 3.20. Let (A, φ) be a noncommutative probability space, I be an ordered in- dex set and (xi)i∈I a sequence of random variables in A. (xi)i∈I is said to be monotonically (boolean) spreadable if all its finite subsequences (xi1, ..., xil) are monotonically(boolean) spreadable. Lemma 3.21. Let (x1, ..., xn+1) be a monotonically spreadable sequence of random vari- ables in (A, φ). Then, all its subsequences are monotonically spreadable. Proof. It suffices to show that the subsequence (x1, ..., xl−1, xl+1, ..., xn+1) is monotonically spreadable for all 1 ≤ l ≤ n. If we denote (x1, ..., xl−1, xl+1, ..., xn+1) by (y1, ..., yn), then we need to show that (y1, ..., yn) is Mi(n, k)-spreadable for all k < n. Fix k < n, let {ui,j}i=1,...,n;j=1,...,k be the set of generators of Mi(n, k) and {Pi,j}i=1,...,n+1;j=1,...,k+1 be an n + 1 by k + 1 matrix with entries in Mi(n, k) such that Pi,j =   i,j u(m) u(m) i−1,j u(m) i,j−1 u(m) i−1,j−1 0 if 1 ≤ i, j < l if 1 ≤ j < l, i ≥ l if 1 ≤ i < l, j ≥ l . if i, j ≥ l otherwise It is a routine to check that the set {Pi,j}i=1,...,n+1;j=1,...,k+1 satisfies the universal condi- tions of Mi(n+1, k+1). Therefore, there exists a C ∗-homomorphism ψ : Mi(n+1, k+1) → Mi(n, k) such that where {u′ notation: i,j} is the set of generators of Mi(n + 1, k + 1). Now, we need a convenient ψ(u′ i,j) = Pi,j σ(i) = (cid:26) i i + 1 if 1 ≤ i < l if i ≥ l NONCOMMUTATIVE SPREADABILITY 17 i,j and yi = xσ(i) for all i = 1, ..., n and j = 1, ...., k + 1. For all 1 = u′ i,1 and P be the invariance projection Then, Pσ(i),σ(j) = u(m) monomial Xj1 · · · Xjm ∈ ChX1, ..., Xki, let P ′ of Mi(n, k), we have µy1,...,yn(Xj1 · · · Xjm)P = Pµx1,...,xn+1(Xσ(j1) · · · Xσ(jm))ψ(P ′ = Pψ(µx1,...,xn+1(Xσ(j1) · · · Xσ(jm))P ′ n Pi=1 1)P 1)P n+1 Pi1,...,im=1 = Pψ(µx1,...,xn+1 ⊗ idMi(n+1,k+1)( Xi1 · · · Xim ⊗ u′ i1,σ(j1) · · · u′ im,σ(jm)))P Notice that u′ l,σ(j) = 0 since σ(j) never equals l, the quality can be written as the following: µy1,...,yn(Xj1 · · · Xjm)P n = Pψ(µx1,...,xn+1 ⊗ idMi(n+1,k+1)( Pi1,...,im=1 = P = n n Pi1,...,im=1 Pi1,...,im=1 µx1,...,xn+1(Xσ(i1) · · · Xσ(im))ψ(u′ σ(i1),σ(j1) · · · u′ σ(im),σ(jm))P µy1,...,yn(Xi1 · · · Xim)Pu(m) i1,j1 · · · u(m) im,jmP Xσ(i1) · · · Xσ(im) ⊗ u′ σ(i1),σ(j1) · · · u′ σ(im),σ(jm)))P which completes the proof. Then, we have (cid:3) Proposition 3.22. Let (A, φ) be a noncommutative probability space and (xi)i∈Z be a sequence of random variables in A. Then, (xi)i∈Z is monotonically (quantum, boolean) spreadable if and only if (xi)i=−n,−n+1,...,.n−1,n is monotonically (quantum, boolean) spread- able for all n. Proof. It is sufficient to prove "⇐". Given a subsequence (xi1 , ..., xil) of (xi)i∈Z, there exits an n such that −n < i1, ..., il < n. Since (xi)i=−n,−n+1,...,.n−1,n is monotonically spreadable, by Lemma 3.21, we have that (xi1, ..., xil) is monotonically spreadable. The same to quantum spreadability and boolean spreadability. (cid:3) 4. Relations between noncommutative probabilistic symmetries In this section, we will study relations between the noncommutative distributional symmetries which are introduced in the previous section. It is well know that every C ∗-algebra admits a faithful representation. Fix n, k ∈ N, such that 1 ≤ k ≤ n − 1. Let Φ be a faithful representation of Bl(n, k) into B(H) for some Hilbert space H. For convenience, we denote Φ(ui,j) by ui,j and Φ(P) by P. According to the definition of Bl(k, n), ui,j's and P are orthogonal projections in B(H). k Let Qi = ui,j for 1 ≤ i ≤ n. In [11], we know that the set P (H) of orthogonal projections on H is a lattice with respect to the usual order ≤ on the set of selfadjoint operators, i.e. two selfadjoint operators A and B, A ≤ B iff B − A is a positive operator. Pj=1 18 WEIHUA LIU Now, we need the following notation in our construction. Given two projections E and F , we denote by E ∨ F the minimal orthogonal projection in P (H), such that E ∨ F is greater or equal to E and F . E ∨ F is well define and unique, we call it the supreme of E and F . It is easy to see that (E ∨ F )E = E and (E ∨ F )F = F We turn to define a sequence of orthogonal projections {P ′ i }i=1,...,n in P (H) as follows: P ′ 1 = I − Q1, i = I − P ′ P ′ 1 ∨ · · · ∨ P ′ i−1 ∨ Qi for 2 ≤ i ≤ n. Lemma 4.1. Given a nonzero vector v ∈ H, E and F are two orthogonal projections on H. If (E ∨ F )x = x and Ex = 0, then F x = x. According the construction of {P ′ i }1≤i≤n, we have j = δi,jP ′ P ′ i P ′ i and P ′ i ui,j = 0 for all 1 ≤ i ≤ n and 1 ≤ j ≤ k. Lemma 4.2. n Pi=1 P ′ i = I, where I is the identity in B(H). Proof. Since the orthogonal projections P ′ i are orthogonal to each other, thogonal projection which is less than or equal to the identity I. If exists a nonzero vector v ∈ H such that n n Pi=1 P ′ i v = 0. Xi=1 Then, we have or say 0 = P ′ i x = (I − P ′ 1 ∨ · · · ∨ P ′ i−1 ∨ Pi)x P ′ i is an or- n Pi=1 P ′ i < I, then there (P ′ 1 ∨ · · · ∨ P ′ i−1 ∨ Pi)x = x for all i. Since P ′ mx = 0 for all 1 ≤ m ≤ i − 1, by Lemma4.1, Pix = x. Then, we have n nx = = = n Pi=1 Pi=1 Pj=1 k Pix k n Pj=1 Pi=1 ( ui,jx , ui,jx) NONCOMMUTATIVE SPREADABILITY 19 which implies that n is in the spectrum of k n Pj=1 Pi=1 ui,j. Notice that, to every 1 ≤ j ≤ k, n Pi=1 ui,j ≤ I since they are orthogonal projections and orthogonal to each other. Therefore, 0 ≤ k n Xj=1 Xi=1 ui,j ≤ k Xj=1 I ≤ kI. It contradicts to the implication above. The proof is complete. (cid:3) Corollary 4.3. P ′ i P = P. n Pi=1 Proposition 4.4. Let (A, φ) be a noncommutative probability space, (xi)i=1,...,n is a finite ordered sequence of random variables in A. For fixed n > k, the joint distribution µx1,...,xn is Al(n, k)-invariant if it is Al(n, k + 1)-invariant Proof. Let {uij1 ≤ i ≤ n, 1 ≤ j ≤ k} the set of standard generators of Al(n, k), Φ be a faithful representation of Al(n, k) into B(H). With the above construction, we can define {u′ i,j}i=1,...,n;j=1,...,k+1 as following: u′ i,j = (cid:26) Φ(ui,j) P ′ j if if j ≤ k j = k i,j}i=1,...,n;j=1,...,k+1 satisfies the universal conditions for Al(n, k + 1). Let ij1 ≤ i ≤ n, 1 ≤ j ≤ k + 1} be the set of standard generators of Al(n, k + 1). then By Lemma 4.2, {u′ {u′′ there exists a C ∗−homomorphism Φ : Al(n, k + 1) → B(H) such that: Φ′(u′′ ij) = u′ i,j. Therefore, Φ−1Φ′ defines a unital C ∗−homomorphism Φ−1Φ′ : C ∗ − alg{u′ i,j1 ≤ i ≤ n, 1 ≤ j ≤ k} → Al(n, k) such that for all 1 ≤ i ≤ n, 1 ≤ j ≤ k. If µx1,...,xn is Al(n, k + 1)-invariant, then Φ−1Φ′(u′ i,j) = ui,j µx1,...,xk+1(p) = µx1,...,xk ⊗ idAl(n,k+1)(α(f p) n,k+1(p)) for all p ∈ ChX1, ..., Xk+1i. Let p = Xj1 · · · Xjl ∈ ChX1, ..., Xki, then we have µx1,...,xk(P )1A(n,k) Φ−1Φ′(µx1,...,xk+1(P )1A(n,k+1))) = Φ−1Φ′(µx1,...,xn ⊗ idAl(n,k+1)(α(f p) n,k+1(Xj1 · · · Xjl)) n = Φ−1Φ′(µx1,...,xn ⊗ idAl(n,k+1)( Xi1 · · · Xil ⊗ u′ Pi1,...,il n i1,j1 · · · u′ il,jl) = µx1,...,xn ⊗ idAl(n,k)( Pi1,...,il = µx1,...,xn ⊗ idAl(n,k)(α(f p) n,k (P )) Xi1 · · · Xil ⊗ ui1,j1 · · · uil,jl) 20 WEIHUA LIU Since p is an arbitrary monomial, the proof is complete. (cid:3) The same, we can show that Corollary 4.5. µx1,...,xn is Bl(n, k)-invariant if it is Bl(n, k + 1)-invariant Lemma 4.6. µx1,...,xn is (n, k)-quantum spreadable if it is Al(n, k)-invariant. Proof. Let {ui,j}i=1,...,n;j=1,...,k be generators of Ai(n, k) and {u′ i,j}i=1,...,n;j=1,...,k be genera- tors of Al(n, k). Then, there is a well defined C ∗-homomorphism β : Al(n, k) → Ai(n, k) such that β(ui,j = u′ i,j). The existence of β is given by the universality of Al(n, k). Since µx1,...,xn is Al(n, k)-invariant, for all monomials p = Xi1 · · · Xim ∈ ChX1, ..., Xki, we have µx1,...,xk(p)1Al(n,k) = µx1,...xn ⊗ idAl(n,k)(α(f p) n,k (p)) = Xj1,...,jm φ(xj1 · · · xjm)uj1,i1 · · · ujm,im. Apply β on both sides of the above equation, we have µx1,...,xk(p)1Ai(n,k) = Xj1,...,jm The proof is complete. The same, we have φ(x′ j1 · · · xjm)u′ j1,i1 · · · ujm,im = µx1,...xn ⊗ idAi(n,k)(αn,k(p)). (cid:3) Corollary 4.7. µx1,...,xn is (n, k)-boolean spreadable if it is Bl(n, k)-invariant. Corollary 4.8. (x1, ..., xn) is boolean spreadable if it is boolean exchangeable. (x1, ..., xn) is quantum spreadable if it is quantum exchangeable. In summary, for fixed n, k ∈ N such that k < n, we have the following diagrams: B(n, n)inv Bl(n, k)inv Bi(n, k)inv Mi(n, k)inv A(n, n)inv / Al(n, k)inv / Ai(n, k)inv and Booolean exchangeability Boolean spreadability M onotone spreadability Quantum exchangeability / quantum spreadability The arrow "condtion a) → condition b)" means that condition a) implies condition b). / / / / / / / / / NONCOMMUTATIVE SPREADABILITY 21 5. Monotonically equivalent sequences In order to study monotone spreadability, we need find relations between mixed mo- ments of monotonically spreadable sequences of random variables. Since all the mixed moments can be denoted by finite sequences of indices, we will turn to study finite se- quences of ordered indices. In this section, we introduce an equivalent relation, which has a deep relation with monotone spreadability, on finite sequences of ordered indices. Definition 5.1. Given two pairs of integers (a, b), (c, d) , we say these two pairs have the same order if a − b, c − d are both positive or negative or 0. For example, (1, 2) and (3, 5) have the same order but (1, 2) and (5, 3) do not have the same order. Definition 5.2. Let Z be the set of integers with natural order and ZL = Z × · · · × Z be the set of finite sequences of length L. We define a partial relation ∼m on ZL. Given two sequences of indices I = {i1, ...., iL}, J = {j1, ..., jL} ∈ ZL. If for all 1 ≤ l1 < l2 ≤ L such that il3 > max{il1, il2} for all l1 < l3 < l2, (il1, il2) and (jl1, jl2) have the same order, then we denote I ∼m J . Example: (5, 3, 4) ∼m (5, 3, 5) but (5, 6, 4) 6∼m (5, 6, 5). Remark 5.3. In general, the relation can be defined on any ordered set but not only Z. We will show this partial relation is exactly an equivalence relation on the set of finite sequences of ordered indices. It follows the definition that (il, il+1) and (jl, jl+1) have the same order for all 1 ≤ l < L if I ∼m J . Now we turn to show that ∼m is actually an equivalent relation. To achieve it, we need to show that the relation ∼m is reflexive, symmetric and transitive. (Reflexivity) First, reflexivity is obvious, because a pair (il1, il2) always has the same order with itself. Lemma 5.4. (Symmetry) Let I = {i1, ...., iL}, J = {j1, ..., jL} ∈ ZL such that I ∼m J , then J ∼m I. Proof. Suppose that J 6∼m I. Then, there exist two natural numbers 1 ≤ l1 < l2 ≤ L such that jl3 > max{jl1, jl2} for all l1 < l3 < l2, but (jl1, jl2) and (il1, il2) do not have the same order. Fix l1, we choose the smallest l2 which satisfies the above property. Notice that I ∼m J , (jl1, jl1+1) and (il1, il1+1) have the same order, then l2 6= l1 + 1. According to our assumption, we have for l1 < l′ Suppose that there exists an l3 between l1 and l2 such that 3 < l2. jl′ 3 > max{jl1, jl2} il3 ≤ max{il1, il2}. 22 WEIHUA LIU Without loss of generality, we assume that then il1 ≥ il2, il3 ≤ il1. Again, among these l3, we choose the smallest one. Then, we have il > il1 ≥ il3 for l1 < l < l3. Since I ∼m J , (il1, il3) and (jl1, jl3) must have the same order, but il1 ≥ il3 and il1 < jl3. 3 < l2. It It contradicts the existence of our l3. Hence, il′ follows that (il1, il2) and (jl1, jl2) have the same order. But, it contradicts our original assumption. Therefore, J ∼m I. > max{il1, il2} for all l1 < l′ 3 Lemma 5.5. Given two sequences I = {i1, ...., iL}, J = {j1, ..., jL} ∈ ZL such that I ∼m J . Let 1 ≤ l1 < l2 ≤ L such that il3 > max il1, il2 for all l1 < l3 < l2. Then, we have for all l1 < l3 < l2. jl3 > max{jl1, jl2} Proof. If the statement is false, then there exists l3 between l1 and l2 such that (cid:3) Suppose jl1 ≥ jl2, then jl3 ≤ max{jl1, jl2}. jl3 ≤ jl1. Among all these l3, we take the smallest one. Then, we have jl4 > max{jl1, jl3} for all l1 < l4 < l3. By Lemma5.4, J ∼m I since I ∼m J . Therefore, (jl1, jl3) and (il1, il3) must have the same order which means il1 ≥ il3. If we assume that jl1 < jl2, then we just need to consider This is a contradiction. the largest one among those l3 and we will get the same contradiction. The proof is complete. (cid:3) Lemma 5.6. (Transitivity)Given three sequences I = {i1, ...., iL}, J = {j1, ..., jL}, Q = {q1, ...qL} ∈ ZL, such that I ∼m J and J ∼m Q, then I ∼m J Proof. Given 1 ≤ l1 < l2 ≤ L such that for all l1 < l3 < l2. By Lemma 5, we have il3 > max{il1, il2} jl3 > max{jl1, jl2} for all l1 < l3 < l2. It follows the definition that (il1, il2), (jl1, jl2) have the same order and (jl1, jl2), (ql1, ql2) have the same order. Therefore, (il1, il2), (ql1, ql2) have the same order. Since l1, l2 are arbitrary, it completes the proof. (cid:3) NONCOMMUTATIVE SPREADABILITY 23 By now, we have shown that the relation ∼m is reflexive, symmetric and transitive. Therefore, we have Proposition 5.7. ∼m is an equivalence relation on ZL. As we mentioned before, Z can be replaced by any ordered set I. When there is no confusion, we always use ∼m to denote the monotone equivalence relation on I L for ordered set I and positive integers L. For example, I can be [n] = {1, ..., n}. , ..., il′ Definition 5.8. Let I = (i1, ..., iL) be a sequence of ordered indices. An ordered sub- 2) of I is called an interval if the sequence contains all the elements sequence (il′ 1 and l′ 3 is between l′ ) of I is called a crest if il′ 2+1}. In addition , we assume that i0 < i1 and iL > iL+1 > max{il′ 1 = il′ il′ 1−1, il′ even though i0, iL+1 are not in I. 2. An interval (il′ whose position l′ 1+1 · · · = il′ , ..., il′ 1 2 3 1 2 Example: (1, 2, 3, 4) has one crest of length 1, namely (4). (1, 2, 1, 3, 4, 4, 3, 5) has 3 crests (2), (4, 4), (5) and 2 is the first peak of the sequence. (1, 1, 1, 1, 1) has one crest (1, 1, 1, 1, 1) which is the sequence itself, because we assumed i0 < i1 and i6 < i5. Lemma 5.9. Given I = (i1, ..., iL) ∈ ZL, I has at least one crest. Proof. Since I consists of finite elements, it has a maximal one, i.e. il such that il ≥ il′ for 1 ≤ l′ ≤ L . It is obvious that il must be contained in an interval (il′ 2) such that , ..., il′ 1 il′ 1 = il′ 1+1 · · · = il′ 2 = il′ and Therefore, I contains a crest. il′ > max{il′ 1−1, il′ 2+1}. Lemma 5.10. Given two index sequences I, J ∈ ZL such that I ∼m J . If (il′ a crest of I, then (jl′ ) is a crest of J , ..., jl′ 1 1 2 (cid:3) , ..., il′ 2) is Proof. Since I ∼m J , all consecutive pairs (il, il+1) and (jl, jl+1) have the order. Accord- ing to the definition, we have If follows that il′ 1−1 < il′ 1 = il′ 1+1 · · · = il′ 2 jl′ 1−1 < jl′ 1 = jl′ 1+1 · · · = jl′ 2 > jl′ 2+1 > jl′ 2+1, thus (jl′ 1 , ..., jl′ 2) is a crest of J . (cid:3) Now, we will introduce some ∼m preserving operations on index sequences. The first ) be an interval of I = 2+1, ..., iL). We operation is to remove a crest from a sequence. Let (il′ (i1, ..., iL), we denote by I \ (il′ denote by the empty set ∅ = I \ I and we assume ∅ ∼m ∅. 2) the new sequence (i1, ..., il′ 1−1, il′ , ..., il′ , ..., il′ 1 1 2 Lemma 5.11. Let I = (i1, ...., il), J = (j1, ..., jL) ∈ ZL such that I 6∼m J . If (il′ is a crest of I and (jl′ , ..., jl′ 1 , ..., il′ 2 ) 1 2) is a crest of J . Then, I \ (il′ 2) 6∼m J \ (jl′ , ..., il′ 1 1 2) , ..., jl′ 24 WEIHUA LIU Proof. If I \ (il′ of I, J are the same. If I \ (il′ , ..., il′ 1 2) is empty, then J \ (jl′ 1 , ..., jl′ 2) must be empty because the lengths , ..., il′ 2 1 ) is non empty, then I can be written as and and I \ (il′ 1 , ..., il′ 2 (i1, ..., il′ 1 , ..., il′ 2 , ..., iL) ) = (i1, ..., il′ 1−1, il′ 2+1, ..., iL) = (i′ 1, ..., i′ l′ 1−1, i′ l′ 1 , ..., i′ L−l′ 2+l′ 1−1) 1 , ..., jl′ J \ (jl′ 2) = (j1, ..., jl′ For any indices 1 ≤ l1 < l2 < L − l′ If l1, l2 ≤ l′ , j′ and (j′ l1 If l1 < l′ 1 ≤ l2, then i′ 1 − 1 or l1, l2 ≥ l′ l2) have the same order . l2 = il2+l′ 2−l′ 1, then (i′ l1 1−1, jl′ 2 + l′ 1, ..., j′ 1−1, j′ 2+1, ..., jL) = (j′ l′ 1 1 − 1 such that il3 > max{i′ l1 , ..., i′ l′ , i′ l2}: , ..., j′ L−l′ 2+l′ 1−1) l2) is an interval of I. Since I ∼m J , (i′ l1 , i′ l2) 1+1. We have il3 > il′ 1−1 ≥ max{i′ l1 , i′ l2} for all l′ 1 ≤ l3 ≤ l′ 2. It follows that il3 > max{il1, il2} 2 − l′ for all l1 < l3 < l2 + l′ 1 + 1. It follows that (il1, il2+l′ the same order. Thus, (i′ , i′ l2) and (j′ l1 l1 The proof is complete. l2) have the same order. , j′ 2−l′ 1+1) and (jl1, jl2+l′ 2−l′ 1+1) have (cid:3) The same as the previous proof, by checking the definition of ∼m, we have Lemma 5.12. Let I = (i1, ...., iL) ∈ ZL and (il′ 1 , ..., il′ 2) is a crest of I, then we have for any integer K such that il′ I = (i1, ...iL) ∼m (i1, ..., il′ 1−1, il′ 1 + K > max{il′ 1 + K, ..., il′ 1−1, il′ 2+1}. 2 + K, il′ 2+1, ..., il) The following proposition shows a deep relation between the set of standard generators of M(n, k) and ∼m: Proposition 5.13. Given two sequences I = {i1, ..., iL} ∈ [k]L, J = {j1, ..., jL} ∈ [n]L, let {u(m) i,j }i=1,...,n;j=1,...,k be the set of standard generators of M(n, k), then we have X(q1,...,qL)∼mJ u(m) q1,i1 · · · u(m) qL,iLP = (cid:26) P if J ∼m I otherwise 0 Proof. We will prove the proposition by induction. When L = 1, the statement is apparently true. Suppose the statement is true for all L ≤ L′. Let us consider the case L = L′ + 1. Let (il′ Case 1: If (jl′ happens: ) is not a crest of J , then I 6∼m J and one of the following cases 2) be a crest of I: , ..., jl′ , ..., il′ 1 2 1 1. There exists an index jl′ 2. jl′ 3. jl′ ≤ jl′ 2 ≤ jl′ 1−1. 2+1. 1 3 of J such that jl′ 3 But, for all Q = (q1, ..., qL) ∼m J , we have: 6= jl′ 3+1 for some l′ 1 ≤ l′ 3 < l′ 2. 1. (ql′ 2. (ql′ 3. (ql′ 3 1 2 , ql′ , ql′ , ql′ 3−1) and (jl′ 1−1) and (jl′ 2+1) and (jl′ 1 2 3 Therefore, we have: NONCOMMUTATIVE SPREADABILITY 25 , jl′ , jl′ , jl′ 3−1) have the same order. 1−1) have the same order. 2+1) have the same order. 1. ql′ 2. ql′ 3. ql′ 3 6= ql′ 1 ≤ ql′ ≤ ql′ 3−1 and il′ 1−1 and il′ 2+1 and il′ 3 = il′ > il′ > il′ 1 2 3−1 for some l′ 1−1. 2+1. 2 1 ≤ l′ 3 < l′ 2. According to the definition of Mi(n, k), we have one of the following equations: u(m) 3+1,il′ ql′ u(m) ,il′ ql′ 1 1 3+1 −1 −1,il′ 1 1. u(m) ,il′ ql′ 3 3 2. u(m) ql′ 1 3. u(m) ,il′ ql′ 2 2 u(m) ql′ 2 +1,il′ 2 In this case, we have +1 = 0 for some l′ 1 ≤ l′ 3 < l′ 2. = 0. = 0. X(q1,...,qL)∼mJ u(m) q1,i1 · · · u(m) qL,iLP = 0. Case 2: If (jl′ 1 , ..., jl′ 2) is a crest of J , then (ql′ · · · u(m) ,il′ ql′ 2 2 u(m) ,il′ ql′ 1 1 1 , ..., ql′ 2) is a crest of Q. Therefore, = u(m) ,il′ ql′ 1 1 . By Lemma 5.12, if we fix the indices of Q \ (ql′ integers such that ql′ 1 = ... = ql′ 2 and max{ql′ 1−1, ql′ 2), then ql′ 1 , ..., ql′ 2 can be any ≤ n. Therefore, we have 1 , ..., ql′ 2+1)} < ql′ u(m) ,il′ ql′ 1 1 −1 1 u(m) ql′ 1 −1,il′ 1 P −1,ql′ u(m) ql′ 1 ≤n max{ql′ 1 2+1)}<ql′ 1 ≤n u(m) ql′ ,il′ 1 1 u(m) ql′ 2+1,il′ 2+1 −1,il′ 1 −1 u(m) ql′ 2 +1,il′ 2 +1 . = P1≤ql′ = u(m) ql′ 1 1 u(m) 2+1,il′ ql′ 2+1 −1 −1,il′ 1 The first equality holds because the extra terms are 0. The second equality uses the 1 + 1 ≤ L′, then J \ monotone universal condition of Mi(n, k). Let L′′ = L − l′ ). If we denote by (jl′ (i′ ) ∈ [n]L′′ By Lemma 5.10, Q \ (ql′ L′′) the sequence I \ (il′ , ..., jl′ 1, ..., i′ ) ∼m J \ (jl′ 2 + l′ , ..., ql′ , ..., jl′ 1 2 1 2 1 2 1 , ..., il′ 2), then we have u(m) q1,i1 · · · u(m) P(q1,...,qL)∼mJ u(m) P 1,i′ q′ 1 L′′ )∼mJ \(jl′ 1 qL,iLP ,...,jl′ 2 1,...,q′ (q′ ) = = (cid:26) P if J \ (jl′ 0 1 , ..., jl′ 2) ∼m I \ (il′ 1 otherwise · · · u(m) L′′ ,i′ q′ L′′ P 2) , ..., il′ The last equality comes from the assumption of our induction. By Lemma5.10 and Lemma5.11, J \ (jl′ The proof is complete. 1 , ..., jl′ 2) ∼m I \ (il′ 1 , ..., il′ 2) iff J ∼m I. (cid:3) 26 WEIHUA LIU 5.1. Operator valued monotone sequences are monotonically spreadable. In this subsection, we will show that operator valued monotone finite sequences of random variables are monotonically spreadable. To achieve it, we need to consider the positions of the smallest elements of indices sequences. Definition 5.14. Let I = (i1, ..., iL) be a sequence of ordered indices and a = min{i1, ..., iL}. We call the set §(I) = {lil = a} the positions of the smallest elements of I. An interval of (il′ 2, here we assume i0 = iL+1 = a for convenience. 2) is called a hill of I if il′ l3 6= a for all l′ 2+1 = a and i′ 1−1 = il′ 3 ≤ l′ 1 ≤ l′ , ..., il′ 1 Example: (1, 2, 3, 4, 1, 2, 1) has two hills (2, 3, 4) and (2). (1, 2, 1, 3, 4, ) has two hills (2) and (3, 4). (1, 1, 1, 1, 1) has no hill. Lemma 5.15. Given two sequences I = {i1, ..., iL}, J = {j1, ..., jL} ∈ [n]L such that I ∼m J , then §(I) = §(J ). Let (il′ , ..., il′ 2). Proof. Let us check the values of J one by one. Suppose , ..., il′ (il′ 1 1 2) be a hill of I, then 2) ∼m (jl′ , ..., jl′ 1 §(I) = {l′′ 1 < · · · < l′′ k′}, where k′ is the number of elements of §(I). Let b = min{j1, ...jL}, we want to show that 1 = · · · = jl′′ jl′′ Given an integer 1 ≤ p < k′, we have k′ = b and jl > b for all l 6∈ §(I). il > a = il′′ p = il′′ p+1 for all l′′ p < l < l′′ p+1. According to the definition of ∼m and Lemma5, we have and jl′′ p = jl′′ p+1 jl > max{jl′′ p , jl′′ p+1 } for all l′′ exists and l < l′′ p < l < l"p+1. The left is to check the elements jl with l < l′′ 1 or l > l′′ 1 , we chose the greatest such l. Then, we have 1 such that jl ≤ jl′′ k′. If there for all l < l′ < l′′ 1. Therefore, we have 1 } jl′ > max{jl, jl′′ which is a contradiction. It implies that for all l < l′′ 1. the same we have il ≤ il′′ 1 jl > jl′′ 1 jl > jl′′ 1 for all l > l′′ from the definition of ∼m. The proof is complete. k. Therefore, jl′′ = · · · = jl′′ k′ = min{j1, ..., jL}. The last statement is obvious (cid:3) 1 Given I = {i1, ..., iL} ∈ ZL, we will denote by xI = xi1xi2 · · · xjL for short . Then, we have NONCOMMUTATIVE SPREADABILITY 27 Proposition 5.16. Let (A, B, E) be an operator valued probability space, and (xi)i=1,...,n be a sequence of random variables in A. If (xi)i=1,...,n are identically distributed and monotonically independent. Then, for indices sequences I = {i1, ..., iL}, J = {j1, ..., jL} ∈ [n]L such that I ∼m J , L ∈ N, we have E[xI] = E[xJ ]. Proof. When L = 1, the statement is true since the sequence is identically distributed. Suppose the statement is true for all L ≤ L′ ≥ 1. Let us consider the case L = L′ + 1. If I has no hill, then i1 = · · · = iL which implies j1 = · · · = jL. The statement is true for this case, because the sequence is identically distributed. Suppose I has hills I1, ..., Il and a = min{i1, ..., iL}. Then, xI can be written as xn1 a xI1xn2 a xI2 · · · xnl a xIlxnl+1 a , where n2, ..., nl ∈ Z+ and n1, nl+1 ∈ Z ∪ {0}. Since (xi)i=1,...,n are monotonically indepen- dent, we have E[xI] = E[xn1 a E[xI1]xn2 a E[xI2] · · · xnl a E[xIl]xnl+1 a ]. Let b = min{j1, ..., jL}, by Lemma 5.15, J has hills J1, ..., Jl whose positions of elements correspond to the positions of elements of I1, ..., Il and Jl′ ∼m Jl′ for all 1 ≤ l′ ≤ k′. Therefore, we have E[xJ ] = E[xn1 = E[xn1 = E[xn1 = E[xI], b b b E[xJ1]xn2 b E[xI1]xn2 a E[xI1]xn2 b E[xJ2] · · · xnl b E[xI2] · · · xnl a E[xI2] · · · xnl b E[xJl]xnl+1 b E[xIl]xnl+1 ] a E[xIl]xnl+1 ] a ] where the second equality follows the induction and the third equality holds because xa and xb are identically distributed. The proof is complete. (cid:3) Proposition 5.17. Let (A, B, E) be an operator valued probability space, and (xi)i=1,...,n be a sequence of random variables in A. If (xi)i=1,...,n are identically distributed and monotonically independent with respect to E. Let φ be a state on A such that φ(·) = φ(E[·]). Then, (xi)i=1,...,n is monotonically spreadable with respect to φ. Proof. For fixed natural numbers n, k ∈ N, let (ui,j)i=1,...,n;j=1,...,k be standard generators of Mi(n, k). Let J = (j1, ..., jL) ∈ [k]L and denote xj1 · · · xjL by xJ . We denote the equiv- alent class of [n]L associated with ∼m by [nL]. For each I ∈ [n]L, we denote ui1,j1 · · · uiL,jL 28 WEIHUA LIU by uI,J . Then, by proposition 5.13, we have φ(xI)PuI,J P φ(E[xI])PuI,J P φ(E[xI])PuI,J P PI∈[n]L = PI∈[n]L = P¯Q∈[n]L PI∈ ¯Q = PJ 6∈ ¯Q∈[n]L PI∈ ¯Q = PJ 6∈ ¯Q∈[n]L PI∈ ¯Q = 0 + φ(E[xJ ])P = φ(xJ )P φ(E[xI])PuI,J P + PJ ∈ ¯Q∈[n]L PI∈ ¯Q φ(E[xQ])PuI,J P + PI∼mJ φ(E[xJ ])PuI,J P φ(E[xI])PuI,J P Since n, k are arbitrary, the proof is complete. (cid:3) 6. Tail algebras In the previous work on distributional symmetries, infinite sequences of objects are indexed by natural numbers. For this kind of infinite sequences of random variables, the conditional expectations in de Finetti type theorems are defined via the limit of unilateral shifts. It is shown in [14] that unilateral shift is an isometry frome A to itself if (A, φ) is a W ∗-probability space generated by a spreadable sequence of random variables and φ is faithful. Therefore, WOT continuous conditional expectations defined via the limit of unilateral shift exist in a very weak situation, i.e. the sequence of random variables just need to be spreadable. However, our works are in a more general situation that the state φ is not necessarily faithful. In our framework, we will provide an example in which the sequence of random variables is monotone spreadable but the unilateral shift is not an isometry. Therefore, we can not get an extended de Finetti type theorem for monotone independence in the usual way. The key change in this paper is that we will consider bilateral sequences of random variables. We begin with an interesting example : 6.1. Unbounded spreadable sequences. Unlike the situation in probability spaces with faithful states, an infinite spreadable sequence of random variables indexed by natural numbers needs not to be bounded. Even more, there exists an infinite monotonically spreadable unbounded sequence of bounded random variables in a non-degenerated W ∗- probability space. Example: Let H be the standard 2-dimensional Hilbert space with orthonormal basis {v = (cid:18) 1 0 (cid:19) , w = (cid:18) 0 1 (cid:19)}. Let p, A, x ∈ B(H) be operators on H with the following matrix forms: p = (cid:18) 1 0 0 0 (cid:19) , A = (cid:18) 1 0 0 2 (cid:19) , x = (cid:18) 0 1 1 0 (cid:19) . NONCOMMUTATIVE SPREADABILITY 29 H the infinite tensor product of H. Let {xi}∞ i=1 be a sequence of selfadjoint Let H = ∞ Nn=1 operators in B(H ) defined as follows: Let φ be the vector state h·v, vi on H and Φ = φ be a state on B(H ). It is obvious that Φ(xn For any x, y ∈ B(H), an elementary computation shows i ) = φ(xn) for for i. Therefore, the sequence (xi)i∈N is identically distributed. xi = i−1 On=1 A ⊗ x ⊗ p ∞ Om=1 Nn=1 ∞ φ(xpy) = φ(x)φ(y). i−1 Nn=1 ∞ Nn=1 For convenience, we will denote A⊗i−1 = A and P ⊗∞ = P . Also, we denote xi1 · · · xiL = xI for I = (i1, ..., iL) ∈ NL . We will show that the sequence {xi}i∈N is Mi(n, k)-spreadable with respect to Φ. Lemma 6.1. For indices sequences I = (i1, ..., iL), J = (j1, ..., jL) ∈ [n]L such that I ∼m J and L ∈ Z+, we have Φ(xI ) = Φ(xJ ) Proof. When L = 1, the statement is true since the sequence is identically distributed. Suppose the statement is true for all L ≤ L′. Let us consider the case L = L′ + 1. If I has no hill, then i1 = · · · = iL which implies j1 = · · · = jL. The statement is true for this case, because the sequence is identically distributed. Also, we denote by x(n) the n-the component of xi. Then, i a if n < i x if n = i p if n > i x(n) i =   I = x(n) and x(n) According to the definition of Φ, we have that · · · x(n) iL . x(n) i2 i1 Φ(xi1 xi2 · · · xjL) = ∞ Yn=1 φ( L Yl=1 x(n) i ). Notice that all the terms φ( x(n) i ) are 1 except finite terms. Suppose I has hills I1, ..., Il L Ql=1 and a = min{i1, ..., iL}, then xI can be written as a xI2 · · · xnl xn1 a xI1xn2 a xIlxnl+1 a . Therefore, φ( L Yl=1 x(n) i ) =   1 φ(xn1AI1xn2AI2 · · · xnlAIlxnl+1) φ(px(n) I1 px(n) I2 p · · · px(n) Il p) if n < a if n = a if n > a 30 WEIHUA LIU It follows that φ( L Yl=1 x(n) i ) = ∞ Yn≥min{I} φ( L Yl=1 x(n) i ). Because we have φ(px(n) I1 px(n) I2 p · · · px(n) Il p) = φ(x(n) I1 )φ(x(n) I2 ) · · · φ(x(n) Il ), Φ(xi1xi2 · · · xjL) = φ(xn1AI1xn2AI2 · · · xnlAIlxnl+1) = φ(xn1AI1xn2AI2 · · · xnlAIlxnl+1) φ(px(n) I1 px(n) I2 p · · · px(n) Il p). φ(x(n) I1 )φ(x(n) I2 ) · · · φ(x(n) Il ) ∞ ∞ Qn>a Qn>a = φ(xn1AI1xn2AI2 · · · xnlAIlxnl+1)Φ(xI1)Φ(xI2) · · · Φ(xIl) Let b = min{j1, ..., jL}, by Lemma5.15, J has hills J1, ..., Jl whose positions of elements correspond to the positions of elements of I1, ..., Il and Jl′ ∼m Jl′ for all 1 ≤ l′ ≤ k′. Therefore, we have Φ(xJ ) = Φ(xi1 xi2 · · · xiL) = φ(xn1AJ1xn2AJ2 · · · xnlAJlxnl+1)Φ(xJ1)Φ(xJ2) · · · Φ(xJl) = φ(xn1AI1xn2AI2 · · · xnlAIlxnl+1)Φ(xI1)Φ(xI2) · · · Φ(xIl) = Φ(xI) where the second equality follows the induction and the true that Jk ∼m Ik and Jk = Ik for all 1 ≤ k ≤ l. The proof is complete. (cid:3) Proposition 6.2. The joint distribution of (xi)i∈N with respect to Φ is monotonically spreadable. Proof. Fixed n > k ∈ N, let {u(m) i,j }i=1,...,n;j=1,...,k be the set of standard generators of M(n, k). For all I = (i1, ..., iL) ∈ [k]L, we denote by [n]L the ∼m equivalence class of NONCOMMUTATIVE SPREADABILITY 31 [n]L, then we have Pµx1,...,xn ⊗ (idM(n,k))(α(m) J ,IP µx1,...,xn(XJ )Pu(m) n,k (XI))P J ,IP + PJ ∼mI J ,IP + PJ ∼mI Pu(m) J ,IP + PJ ∼mI µx1,...,xn(XJ )Pu(m) J ,IP µx1,...,xn(XI)Pu(m) J ,IP µx1,...,xn(XI)Pu(m) J ,IP µx1,...,xn(XI)Pu(m) J ,IP = PJ ∈[n]L = P¯Q∈[n]L PJ ∈ ¯Q = PI6∈ ¯Q∈[n]L PJ ∈ ¯Q = PI6∈ ¯Q∈[n]L PJ ∈ ¯Q = PI6∈ ¯Q∈[n]L = PI6∈ ¯Q∈[n]L = PJ ∼mI = Φ(xI)P µx1,...,xn(XJ )Pu(m) J ,IP µx1,...,xn(XJ )Pu(m) µx1,...,xn(XQ)Pu(m) µx1,...,xn(XQ) PJ ∈ ¯Q µx1,...,xn(XQ) · 0 + PJ ∼mI µx1,...,xn(XI)Pu(m) J ,IP The proof is complete. By direct computations, we have and (4) n Yi=1 xn+1−iv⊗∞ = w⊗n ⊗ v⊗∞ xn+1w⊗n ⊗ v⊗∞ = 2nw⊗n+1 ⊗ v⊗∞ (cid:3) Let (H′, π′, ξ′) be the GNS representation of the von Neumann algebra generated by (xi)i=1,...,∞ associated with Φ. We have kπ′(xn+1)k ≤ kxn+1k = 2n, but equation 4 shows that kπ′(xn+1)k ≥ 2n. Therefore, kπ′(xn+1)k = 2n. Therefore, there is no bounded endomorphism α on A such that α(xi) = xi+1. 6.2. Tail algebras of bilateral sequences of random variables. In the last subsec- tion, we showed that, in a W ∗-probability space with a non-degenerated normal state, the unilateral shift of a spreadable unilateral sequence of random variables may not be extended to be a bounded endomorphism. Therefore, in general, we can not define a normal condition expectation by taking the limit of unilateral shifts of variables. The main reason here is that the spreadability of variables does not give enough restrictions to control the norms of the variables in our probability space. In (A, φ), a W ∗-probability space with a faithful state, the norm of a selfadjoint random variable x ∈ A is controlled by the moments of X, i.e. kxk = lim n→∞ φ(xn) 1 n . 32 WEIHUA LIU But, in our non-degenerated W ∗-probability spaces, the norm of a random variable de- pends on all mixed moments which involve it. To make the conditional expectation exist, we will consider spreadable sequences of random variables indexed by Z but not N. In this case, the sequence (xi)i∈Z is bilateral. As a consequence, we will have two choices to take limits on defining normal conditional expectations and tail algebras. Before studying properties of tail algebras of bilateral sequences, we introduce some necessary notations and assumptions first. Let (A, φ) is a W ∗−probability space generated by a spreadable bilateral sequence of bounded random variables (xi)i∈Z and φ is a non-degenerated normal state. We assume that the unit of A is contained in the WOT-closure of the non-unital algebra generated by (xi)i∈Z. Let (H, π, ξ) be the GNS representation of A associated with φ. Then, {π(P (xii ∈ Z))ξP ∈ ChXii ∈ Zi} is dense in H. For convenience, we will denote π(y)ξ by y for all y ∈ A. When there is no confusion, we will write y short for π(y). We denote by Ak+ the non-unital algebra generated by (xi)i≥k and Ak− the non-unital algebra generated by (xi)i≤k. Let A+ k be the WOT-closure of Ak+ and Ak−, respectively. k and A− Definition 6.3. Let (A, φ) be a no-degenerated noncommutative W ∗-probability space, (xi)i∈Z be a bilateral sequence of bounded random variables in A such that A is the WOT closure of the non-unital algebra generated by (xi)i∈Z. The positive tail algebra A+ tail of (xi)i∈Z is defined as following: A+ tail = \k>0 A+ k . In the opposite direction, we define the negative tail algebra A− tail of (xi)i∈Z as following: A− tail = \k<0 A− k . Remark 6.4. In general, the positive tail algebra and the negative tail algebra are differ- ent. Even though our framework looks quit different from the framework in [14], we can show that there exists a normal bounded shift of the sequence in a similar way. For completeness, we provide the details here. Lemma 6.5. There exists a unitary map U : H → H such that U(P (xii ∈ Z))ξ = P (xi+1i ∈ Z)ξ Proof. Since (xi)i∈Z is spreadable, we have φ((P (xii ∈ Z))∗P (xii ∈ Z)) = φ((P (xi+1i ∈ Z))∗P (xi+1i ∈ Z)). It implies that U(P (xii ∈ Z)ξ) = P (xi+1i ∈ Z)ξ is a well defined isometry on {π(P (xii ∈ Z))ξP ∈ ChXii ∈ Zi} . Since {π(P (xii ∈ Z))ξP ∈ ChXii ∈ Zi} is dense in H, U can be extended to the whole space H. It is obvious that {π(P (xii ∈ Z))ξP ∈ ChXii ∈ Zi} is contained in the range of U. Therefore, the extension of U is a unitary map on H. (cid:3) NONCOMMUTATIVE SPREADABILITY 33 Now, we can define an automorphism α on A by the following formula: α(y) = U yU −1. Lemma 6.6. α is the bilateral shift of (xi)i∈Z, i.e. α(xk) = xk+1 for all k ∈ Z. Proof. For all y = P (xii ∈ Z)ξ, we have α(xk)y = U xkU −1P (xii ∈ Z)ξ = U xkP (xi−1i ∈ Z)ξ = xk+1P (xii ∈ Z)ξ. By the density of {π(P (xii ∈ Z))ξP ∈ ChXii ∈ Zi}, we have α(xk) = xk+1. The proof is complete. (cid:3) Since α is a normal automorphism of A, we have k ) = A+ Corollary 6.7. For all k ∈ Z, we have α(A+ k+1. Lemma 6.8. Fix n ∈ Z. Let y1, y2 ∈ An−. Then, we have hαl(a)y1, y2i = hay1, y2i, where l ∈ N and a ∈ A+ n+1. Proof. It is sufficient to prove the statement under the assumption that l = 1. Since a ∈ A+ n+1, by Kaplansky's theorem, there exists a sequence (am)m∈N ⊂ A(n+1)+ such that kamk ≤ kak for all m and am converges to a in WOT. Then, by the spreadability of (xi)i∈Z, we have hα(a)y1, y2i = lim m→∞ hα(am)y1, y2i = lim m→∞ φ(y∗ 2am y1) = hay1, y2i (cid:3) In the following context, we fix k ∈ Z. Lemma 6.9. For all a ∈ A+ k , we have that exists. Moreover, E+[a] ∈ A+ tail E+[a] = W OT − lim l→∞ αl(a) Proof. For all y1, y2 ∈ {π(P (xii ∈ Z))ξP ∈ ChXii ∈ Zi}, there exits n ∈ Z such that y1, y2 ∈ An−. For all l > n − k, we have αl(a) ∈ A(n+1)+. By Lemma 6.8, we have Therefore, hαn+1−k(a)y1, y2i = hαn+2−k(a)y1, y2i = · · · . hαl(a)y1, y2i = hαn+1−k(a)y1, y2i. lim l→∞ αl(a) converges pointwisely to an element E+[a]. Since for all n > 0, we have αl(a) ∈ A+ n n for all n. Hence, E+[a] ∈ for all l > n − k + 1. It follows that W OT − lim l→∞ A+ (cid:3) αl(a) ∈ A+ tail. Proposition 6.10. E+ is normal on A+ k for all k ∈ Z. 34 WEIHUA LIU Proof. Let (am)m∈N ⊂ A+ k be a bounded sequence which converges to 0 in WOT. For all y1, y2 ∈ {π(P (xii ∈ Z))ξP ∈ ChXii ∈ Zi}, there exits n ∈ Z such that y1, y2 ∈ An−. Then, we have lim m→∞ hE+[am]y1, y2i = lim m→∞ hαn+1−k(am)y1, y2i = 0. The last equality holds because αl is normal for all l ∈ N. The proof is complete. (cid:3) Remark 6.11. E+ is defined on Sk∈Z to the whole algebra A. A+ k but not on A. In general, we can not extend E+ Lemma 6.12. E+[a] = a for all a ∈ A+ tail. Proof. For all y1, y2 ∈ {π(P (xii ∈ Z))ξP ∈ ChXii ∈ Zi}, there exits n ∈ Z such that y1, y2 ∈ An−. Since a ∈ A+ n+1, by Kaplansky's theorem, there exists a sequence of (am)m∈N ⊂ A(n+1)+ such that am → a in WOT and kamk ≤ kak for all m. Then we have. tail ⊂ A+ hay1, y2i = lim m→∞ hay1, y2i = lim m→∞ hα(am)y1, y2i = hα(a)y1, y2i. Since y1, y2 are arbitrary, we have a = α(a). Remark 6.13. One should be careful that A+ set of α. tail could be a proper subset of the fixed points (cid:3) Lemma 6.14. E+[a1ba2] = a1E+[b]a2 for all b ∈ A+ k , a1, a2 ∈ A+ tail. Proof. By Lemma6.12, we have E+[a1ba2] = lim l→∞ αl(a1ba2) = lim l→∞ αl(a1)αl(b)αl(a2) = lim l→∞ a1αl(b)a2 = a1E+[b]a2 (cid:3) 7. Conditional expectations of bilateral monotonically spreadable sequence In this section, we assume that the joint distribution of (xi)i∈Z is monotonically spread- able. Lemma 7.1. Fix n > k ∈ N, let (ui,j)i=1,...,n; j=1,...,k be the standard generators of Mi(n, k). Then, we have φ(a1xl1 i1 b1xl2 i2 b2 · · · bm−1xlm im a2)P = n Xj1,...,jm=1 φ(a1xl1 j1 b1xl2 j2 b2 · · · bm−1xlm jm a2)Puj1,i1 · · · ujm,imP, where 1 ≤ i1, ...im ≤ k, b1, ..., bm−1 ∈ A(n+1)+ and a1, a2 ∈ A0−. Proof. Without loss of generality, we assume that there exist n1, n2 ∈ N such that and a1, a2 ∈ A[−n1+1,0] b1, ..., bm−1 ∈ A[n+1,n2+k]. NONCOMMUTATIVE SPREADABILITY 35 Since the map is linear, we just need to consider the case that a1, a2 and b1, ..., bm−1 are products of (xi)i∈Z. Let and a1 = xs1,1 · · · xs1,t1 a2 = xs2,1 · · · xs2,t2 for some t1, t2 ∈ N and −n1 + 1 ≤ sc,d ≤ 0. Let bi = xri,1 · · · xri,t′ i 1, ..., t′ for t′ m−1 ∈ N ∪ {0} and n + 1 ≤ rc,d ≤ k + n2. Then, (x−n1+1, ..., xn+n2) is a sequence of length n + n1 + n2, we denote it by (y1, ..., yn+n1+n2). Let n′ = n + n1 + n2 and k′ = k + n1 + n2. By our assumption, a1xl1 ima2 is in the algebra generated i1 i,j)i=1,...,n′; j=1,...,k′ be the standard generators of Mi(n′, k′) and P′ by (y1, ...., yk′). Let (u′ be the invariant projection. Let π be the C ∗-homomorphism in Lemma 3.13 and id be the identity may on ChX1, ...., Xn′i. Since 1 ≤ sc,d + n1 ≤ n1, we have b2 · · · bm−1xlm b1xl2 i2 id ⊗ π(α(m) n′,k′(Xsi,1+n1 · · · Xsi,t1 + n1)) = Xsi,1+n1 · · · Xsi,t1 +n1 ⊗ P. Since n1 + n + 1 ≤ rc,d + n1 ≤ n1 + n2 + k, we have id ⊗ π(α(m) n′,k′(Xri,1+n1 · · · Xr1,t′ + n1)) = Xri,1+n1+n−k · · · Xri,t′ i +n1+n−k ⊗ I, 1 where I is the identity of Mi(n, k). According to our assumption, we have 1 ≤ it ≤ k for t = 1, ...., m. Then id ⊗ π(α(m) n′,k′(X lt it+n1) = n Xjt=1 X lt jt+n1 ⊗ ujt,it. According to the monotone spreadability of (y1, ..., yn′) and Lemma3.13, we have b1xl2 i2 φ(a1xl1 i1 b2 · · · bm−1xlm im a2)P = µy1,...,yk′ (Xs1,1+n1 · · · Xs1,t1 +n1X l1 = Pµy1,...,yn′ ⊗ π(α(m) = i1+n1 · · · X lm im+n1 n′,k′(Xs1,1+n1 · · · Xs1,t1 +n1X l1 µy1,...,yn′ (Xs1,1+n1 · · · Xs1,t1 +n1X l1 jm Xs1,1+n1 · · · Xs2,t2 +n−kX lm+n1 Pj1,...,jm=1 Xrm−1,t′ j1+n1 n m−1+n1 i1+n1 · · · X lm Xr1,1+n1+n−k · · · im+n1 )Puj1,i1 · · · ujm,imP Xs1,1+n1 · · · Xs2,t2 +n1)π(P′) Xs1,1+n1 · · · Xs2,t2 +n1))P Notice that (y1, ..., yn′) is spreadable and n + 1 ≤ r,, the above equation becomes b2 · · · bm−1xlm b1xl2 i2 µy1,...,yn′ (Xs1,1+n1 · · · Xs1,t1 +n1X l1 ima2)P j1+n1 Xr1,1+n1 · · · X lm jm+n1 Xs1,1+n1 · · · Xs2,t2 )Puj1,i1 · · · ujm,imP +n1 m−1 φ(xs1,1 · · · xs1,t1 xl1 j1 xr1,1 · · · xrm−1,t′ m−1 xlm jmxs1,1 · · · xs2,t2 ) = = = n n φ(a1xl1 i1 Pj1,...,jm=1 Xrm−1,t′ Pj1,...,jm=1 Pj1,...,jm=1 n Puj1,i1 · · · ujm,imP φ(a1xl1 j1 b1xl2 j2 b2 · · · bm−1xlm jm a2)Puj1,i1 · · · ujm,imP 36 WEIHUA LIU The proof is complete. (cid:3) Lemma 7.2. Fix n > k ∈ N, let (ui,j)i=1,...,n; j=1,...,k be the standard generators of Mi(n, k). Then, we have E+[xl1 i1 b1xl2 i2 b2 · · · bm−1xlm im] ⊗ P = n Xj1,...,jm=1 E+[xl1 j1 b1xl2 j2 b2 · · · bm−1xlm jm] ⊗ Puj1,i1 · · · ujm,imP, where 1 ≤ i1, ...im ≤ k, b1, ..., bm−1 ∈ A(n+1)+. Proof. It is necessary to check the two sides of the equation equal to each other pointwisely, i.e. (5) n φ(a1E+[xl1 i1 b1xl2 i2 b2 · · · bm−1xlm im]a2)P = Xj1,...,jm=1 φ(a1E+[xl1 j1 b1xl2 j2 b2 · · · bm−1xlm jm]a2)Puj1,i1 · · · ujm,imP for all a1, a2 ∈ A[−∞,∞]. Given a1, a2 ∈ A[−∞,∞], then there exists M ∈ N such that a1, a2 ∈ AM −. Then, α−m(a1), α−m(a2) ∈ A0− for all m > M. By Lemma 7.1, we have b1xl2 i2 b2 · · · bm−1xlm im α−m(a2))P φ(α−m(a1)xl1 j1 b1xl2 j2 b2 · · · bm−1xlm jm α−m(a2))Puj1,i1 · · · ujm,imP. = n φ(α−m(a1)xl1 i1 Pj1,...,jm=1 Therefore, for all m > M,we have b1xl2 i2 = n φ(a1αm(xl1 i1 φ(a1αm(xl1 Pj1,...,jm=1 j1 b2 · · · bm−1xlm im)a2)P b1xl2 j2 b2 · · · bm−1xlm jm)a2)Puj1,i1 · · · ujm,imP. Let m go to +∞, we get equation 5. The proof is complete since a1, a2 are arbitrary. Proposition 7.3. Let (A, φ) be a W ∗-probability space, (xi)i∈Z a sequence of selfadjoint random variables in A , E+ be the conditional expectation onto the positive tail algebra A+ tail. Assume that the joint distribution of (xi)i∈Z is monotonically spreadable, then the same is true for the joint distribution with respect to E+, i.e. for fixed n > k ∈ N and (ui,j)i=1,...,n; j=1,...,k the standard generators of Mi(n, k), we have that (cid:3) E+[xl1 i1 b1xl2 i2 b2 · · · bm−1xlm im] ⊗ P = n Xj1,...,jm=1 E+[xl1 i1 b1xl2 i2 b2 · · · bm−1xlm im] ⊗ Puj1,i1 · · · ujm,imP, 1 ≤ i1, ..., im ≤ k, l1, ..., lm ∈ N and b1, ..., bn ∈ A+ Proof. Since b1, ..., bm−1 ∈ A+ tail ∈ A+ tail. n , by Kaplansky's theorem, there exists sequences such that kbs,tk ≤ kbsk and lim n→∞ b1,t1 xl2 i2 SOT − lim t1→∞ xl1 i1 {bs,t}s=1,...m−1;t∈N ⊂ An+ bs,t = bs in SOT for each s = 1, ..., m − 1. Therefore, b2,t2 · · · bm−1,tm xlm im = xl1 i1 b1xl2 i2 b2,t2 · · · bm−1,tm xlm im. NONCOMMUTATIVE SPREADABILITY 37 By Lemma 7.2, we have E+[xl1 i1 b1,t1 xl2 i2 b2,t2 · · · bm−1,tm xlm im]⊗P = n Xj1,...,jm=1 E+[xl1 j1 b1,t1xl2 j2 b2,t2 · · · bm−1,tm−1 xlm jm]⊗Puj1,i1 · · · ujm,imP Let t1 go to +∞, by normality of E+, we have E+[xl1 i1 b1xl2 i2 b2,t2 · · · bm−1,tmxlm im]⊗P = n Xj1,...,jm=1 E+[xl1 j1 b1xl2 j2 b2,t2 · · · bm−1,tm−1 xlm jm]⊗Puj1,i1 · · · ujm,imP Again, take t2, ..., tm−1 to +∞, we have (6) n E+[xl1 i1 b1xl2 i2 b2 · · · bm−1xlm im] ⊗ P = Xj1,...,jm=1 E+[xj1b1xj2b2 · · · bm−1xjm] ⊗ Puj1,i1 · · · ujm,imP (cid:3) According to the universal conditions of Mi(n, k), if is = is+1 for some s, then the terms on the right hand side are not vanished only if js = js+1. Therefore we can shorten the product on the right hand side of 6 if is = is+1 for some s. We have Proposition 7.4. Let (A, φ) be a W ∗-probability space, (xi)i∈Z a sequence of selfadjoint random variables in A , E+ be the conditional expectation onto the positive tail algebra A+ tail. Assume that the joint distribution of (xi)i∈Z is monotonically spreadable, for fixed n > k ∈ N and (ui,j)i=1,...,n; j=1,...,k the standard generators of Mi(n, k), we have that E+[p1(xi1) · · · pm(xim)] ⊗ P = n Xj1,...,jm=1 E+[p1(xj1) · · · pm(xjm)] ⊗ Puj1,i1 · · · ujm,imP, whenever 1 ≤ i1, ..., im ≤ k, i1 6= · · · 6= im and p1, ..., pm ∈ A+ tailhXi0. Lemma 7.5. Let (A, φ) be a W ∗-probability space, (xi)i∈Z a sequence of selfadjoint ran- dom variables in A , E+ be the conditional expectation onto the positive tail algebra A+ tail. Assume that the joint distribution of (xi)i∈Z is monotonically spreadable, then E+[p1(xi1) · · · ps(xis) · · · pm(xim)] = E+[p1(xi1) · · · E+[ps(xis)] · · · pm(xim)] whenever is > it for all t 6= s, i1 6= · · · 6= im and p1, ..., pm ∈ A+ tailhXi0. Proof. Since (xi)i∈Z is spreadable, by Lemma 6.9, we have that α(pt(xit)) = pt(α(xit)) E+[αk′ (a)] = E+[a] A+ n′ and k′ ∈ Z. Therefore, it is sufficient to prove the statement under the assumption that i1, ..., im > 0. and for all a ∈ Sn′∈Z 38 WEIHUA LIU Let is = k, (ui,j)i=1,...,n+1; j=1,...,k the standard generators of Mi(n + k, k). By proposition 7.4, we have E+[p1(xi1) · · · pm(xim)] ⊗ P = n+k Xj1,...,jm=1 E+[p1(xj1) · · · pm(xjm)] ⊗ Puj1,i1 · · · ujm,imP. Now, apply proposition 3.12 by letting l1 = · · · = lk−1 = 1 and lk = n + 1, then we have E+[p1(xi1) · · · ps(xis) · · · pm(xim)] ⊗ P = 1 n + 1 n+k Xjs=k E+[p1(xi1) · · · ps(xjs) · · · pm(xim)] ⊗ P. Since n is arbitrary, and E+ is normal on A+ 0 , we have E+[p1(xi1) · · · ps(xis) · · · pm(xim)] = 1 n+1 n+k Pjs=k E+[p1(xi1) · · · ps(xjs) · · · pm(xim)] = WOT − lim n→∞ E+[p1(xi1) · · · ( 1 n+1 ps(xjs)) · · · pm(xim)] = WOT − lim n→∞ = WOT − lim n→∞ E+[p1(xi1) · · · ( 1 n+1 E+[p1(xi1) · · · E+[ps(xis)] · · · pm(xim)]. αt(ps(xis)) · · · pm(xim)] n+k n Pjs=k Pt=0 The proof is complete. (cid:3) Now, we turn to consider the case that the maximal index is not unique. Proposition 7.6. Let (A, φ) be a W ∗-probability space, (xi)i∈Z a sequence of selfadjoint random variables in A , E+ be the conditional expectation onto the positive tail algebra A+ tail. Assume that the joint distribution of (xi)i∈Z is monotonically spreadable, then E+[p1(xi1) · · · ps(xis) · · · pm(xim)] = E+[p1(xi1) · · · E+[ps(xis)] · · · pm(xim)] whenever is = max{i1, ..., in} for all t 6= s, i1 6= · · · 6= im and p1, ..., pm ∈ A+ tailhXi0. Proof. Again, we can assume that i1, ..., it > 0 and max{i1, ..., im} = k. Suppose the number k appears t times in the sequence, which are {ilj }j = 1, ..., t such that ilj = k and l1 < l2 < · · · < lt. Fix n, k and consider Mi(n + k, k), by proposition 7.4 and proposition 3.12, we have k+n E+[p1(xi1) · · · pl1(xil1 Pjl1 ) · · · pl2(xil2 E+[p1(xi1) · · · pl1(xjl1 ,...jlt=k ,jl2 ) · · · pm(xim)] ⊗ P ) · · · pl2(xjl2 ) · · · pm(xim)] ⊗ P Pjl1 ,kP Pjl2 ,kP · · · ujlt ,kP = = = 1 (n+1)t 1 (n+1)t ( + k+n ,jl2 ,...jlt=k N Pjl1 Pjls 6=jlr if s6=r P N jls =jlt for some s6=t E+[p1(xi1) · · · pl1(xjl1 ) · · · pl2(xjl2 ) · · · pm(xim)]] ⊗ P E+[p1(xi1) · · · pl1(xjl1 ) · · · pl2(xjl2 ) · · · pm(xim)] ⊗ P E+[p1(xi1) · · · pl1(xjl1 ) · · · pl2(xjl2 ) · · · pm(xim)] ⊗ P ) NONCOMMUTATIVE SPREADABILITY 39 In the first part of the sum, apply proposition 7.5 on indices jl1 , ...jllt that recursively, it follows E+[p1(xi1) · · · ps(xjl1 ) · · · ps(xjl2 ) · · · pm(xim)] = E+[p1(xi1) · · · E[pl1(xjl1 )] · · · E[pl2(xjl2 )] · · · pm(xim)]. Since E[ps(xjl1 )] = E[ps(xk)], for all jl1, ..., jlt, E+[p1(xi1) · · · pl1(xjl1 ) · · · pl2(xjl2 ) · · · pm(xim)] = E+[p1(xi1) · · · E[pl1(xk)] · · · E[pl2(xk)] · · · pm(xim)]. Then, we have 1 (n+1)t ( N Pjls 6=jlr if s6=r E+[p1(xi1) · · · pl1(xjl1 ) · · · pl2(xjl2 ) · · · pm(xim)] ⊗ P , (n+1−s) n+1t E+[p1(xi1) · · · E[pl1(xk)] · · · E[pl2(xk)] · · · pm(xim)] ⊗ P t−1 Q s=0 = which converges to E[ps(xk)] · · · E[ps(xk)] · · · pm(xim)] ⊗ P in norm as n goes to +∞. To the second part of the sum, we have kE+[p1(xi1) · · · pl1(xjl1 ) · · · pl2(xjl2 ) · · · pm(xim)]k ≤ kp1(xi1) · · · pl1(xjl1 ≤ kp1(xi1)k · · · kpl1(xjl1 ≤ kp1(x1)k · · · kpl1(x1)k · · · kpl2(x1)k · · · kpm(x1)k )k · · · kpl2(xjl2 ) · · · pm(xim)k ) · · · pl2(xjl2 )k · · · kpm(xim)k which is finite. Therefore, N jls =jlt for some s6=t P t−1 Q s=0 (n+1−s) (n+1)t E+[p1(xi1) · · · pl1(xjl1 ) · · · pl2(xjl2 ) · · · pm(xim)]k ≤ (1 − )kp1(x1)k · · · kpl1(x1)k · · · kpl2(x1)k · · · kpm(x1)k goes to 0 as n goes to +∞. Therefore, we have E+[p1(xi1) · · · pl1(xil1 ) · · · pl2(xil2 ) · · · pm(xim)] = E+[p1(xi1) · · · E[pl1(xk)] · · · E[pl2(xk)] · · · pm(xim)] The same we can show that E+[p1(xi1) · · · pl1(xk) · · · E+[ps(xis)] · · · pl2(xk) · · · pm(xim)] = E+[p1(xi1) · · · E[pl1(xk)] · · · E[pl2(xk)] · · · pm(xim)] which implies E+[p1(xi1) · · · ps(xis) · · · pm(xim)] = E+[p1(xi1) · · · E+[ps(xis)] · · · pm(xim)] (cid:3) 40 WEIHUA LIU 8. de Finetti type theorem for monotone spreadability 8.1. Proof of main theorem 1. Now, we turn to prove our main theorem for monotone independence: Theorem 8.1. Let (A, φ) be a non degenerated W ∗-probability space and (xi)i∈Z be a bilateral infinite sequence of selfadjoint random variables which generate A. Let A+ k be the WOT closure of the non-unital algebra generated by {xii ≥ k}. Then the following are equivalent: a) The joint distribution of (xi)i∈Z is monotonically spreadable. b) For all k ∈ Z, there exits a φ preserving conditional expectation Ek : A+ k → A+ tail such that the sequence (xi)i≥k is identically distributed and monotone with respect Ek. Moreover, EkAk′ = Ek′ when k ≥ k′. Proof. "b) ⇒ a) "follows corollary 5.17 We will prove "a) ⇒ b) "by induction. Since the sequence is spreadable, it is suffices k → A+ to prove a) ⇒ b) for k = 1: By the results in the previous two sections, there exists a conditional expectation Ek : A+ tail such that the sequence (xi)i≥k is identically distributed with respect to Ek and EkAk′ = Ek′ when k ≥ k′. Actually, Ek is the restriction of E+ on A+ k . Since the sequence is spreadable, we just need to show that the sequence (xi)i∈N is monotonically independent with respect to E1, i.e. E+[p1(xi1) · · · ps(xis) · · · pm(xim)] = E+[p1(xi1) · · · E+[ps(xis)] · · · pm(xim)] (7) is−1 < is > is+1, i1 6= · · · 6= im, i1, ..., im ∈ N and p1, ..., pm ∈ A+ Now, we prove this equality by induction on the maximal index of {i1, ..., im}: When max{i1, ..., im} = 1, then equality is true because is = 1 and the length of the sequence (i1, ..., im) can only be 1. Suppose the equality holds for max{i1, ..., im} = n. When max{i1, ..., im} = n + 1, we have two cases: Case 1: is = n + 1. In this case the equality follows proposition 7.6. tailhXi. is ≤ n. Suppose the number n + 1 appears t times in the sequence, which Case 2: are {ilj }j = 1, ..., t such that ilj = k and l1 < l2 < · · · < lt. Since is−1 < is > is+1, is−1, is, is+1 6= n + 1. By proposition 7.6, we have: E+[p1(xi1) · · · pl1(xil1 ) · · · ps−1(xis−1)ps(xis)ps+1(xis+1) · · · plt(xilt ) · · · pm(xim)] = E+[p1(xi1) · · · E+[pl1(xil1 )] · · · ps−1(xis−1)ps(xis)ps+1(xis+1) · · · E+[plt(xilt )] · · · pm(xim))] Notice that p1(xi1) · · · E+[pl1(xil1 by induction, we have )] · · · ps−1(xis−1)ps(xis)ps+1(xis+1) · · · E+[plt(xilt )] · · · pm(xim) ∈ A+ tailhX1, ..., Xni E+[p1(xi1) · · · E+[pl1(xil1 = E+[p1(xi1) · · · E+[pl1(xil1 = E+[p1(xi1) · · · pl1(xil1 )] · · · ps−1(xis−1)ps(xis)ps+1(xis+1) · · · E+[plt(xilt )] · · · ps−1(xis−1)E+[ps(xis)]ps+1(xis+1) · · · E+[plt(xilt ) · · · pm(xim)] ) · · · ps−1(xis−1)E+[ps(xis)]ps+1(xis+1) · · · plt(xilt ) · · · pm(xim)] )] · · · pm(xim)] NONCOMMUTATIVE SPREADABILITY The last equality follows proposition 7.6. This our desired conclusion. 41 (cid:3) 8.2. Conditional expectation E−. We do not know whether we can extend E+ to the whole space A. But, the conditional expectation E− can be extended to the whole algebra A if the bilateral sequence (xi)i∈Z is monotonically spreadable. Given a, b, c ∈ A[−∞,∞], then there exists L ∈ N such that a, b, c ∈ A[−L,L]. Therefore, α−3L(c) ∈ A[−4L,−3L]. Since (x−4L, x−4L+1, ...) is monotonically with respect to E+, we have φ(aE−[b]c) = lim n→∞ = = = lim n→∞,n>4L lim n→∞,n>4L lim n→∞,n>4L φ(aα−n(b)c) φ(aα−n(b)c) φ(E+[aα−n(b)c]) φ(E+[E+[a]α−n(b)E+[c]]) = lim n→∞ = lim n→∞ φ(E+[a]α−n(b)E+[c]) φ(E+[a]E−[b]E+[c]) Since A is generated by countablely many operators, by Kaplansky's density theorem, for all y ∈ A, there exists a sequence {yn}n∈N ⊂ A[−∞,∞] such that kynk ≤ kyk for all n and yn converges to y in WOT. Then, for all a, c ∈ A[−∞,∞] we have lim n→∞ φ(aE−[yn]c) = lim n→∞ φ(E+[a]ynE+[c]) = φ(E+[a]yE+[c]) Therefore, E−[yn] converges to an element y′ pointwisely. Moreover, y′ depends only on y. If we define E−[y] = y′, then we have Proposition 8.2. Let (A, φ) be a non-degenerated W ∗-probability space and (xi)i∈Z be a bilateral infinite sequence of selfadjoint random variables which generate A. If (xi)i∈Z is monotonically spreadable, then the negative conditional expectation E− can be extend to the whole algebra A such that for all y ∈ A and a, c ∈ A[−∞,∞]. Moreover, the extension is normal. φ(aE−[y]b) = φ(E+[a]yE+[c]) 9. de Finetti type theorem for boolean spreadability In this section, we assume that (A, φ) is a W ∗-probability space with a non-degenerated normal state and A is generated by a bilateral sequence of random variables (xi)i∈Z and (xi)i∈Z are boolean spreadable. Lemma 9.1. Let yi = x−i for all i ∈ Z, then (yi)i∈Z is also boolean spreadable. Proof. By proposition 3.22, it suffices to show that (yi)i=1,...,n is boolean spreadable for all n ∈ N. Given a natural number k < n, assume the standard generators of Bi(n, k) are {ui,j}i=1,...,n;j=1,...,k and invariant projection P. 42 WEIHUA LIU Consider the matrix {u′ the entries of the matrix are are orthogonal projections and i,j}i=1,...,n;j=1,...,k such that u′ i,j = un+1−i,k+1−j.. it is obvious that n Xi=1 u′ i,jP = n Xi=1 ui,k+1−jP = P. Given j, j′, i, i′ ∈ N such that 1 ≤ j < j′ ≤ k and 1 ≤ i ≤ i′ ≤ n. Then, we have n + 1 − i ≤ n + 1 − i′ and k + 1 − j < k + 1 − j′. Therefore, u′ i,ju′ i′,j ′ = un+1−i,k+1−jun+1−i′,k+1−j ′ = 0. It implies that {u′ follows that there exists a unital C ∗-homomorphism Φ : Bi(n, k) → Bi(n, k) such that: i,j}i=1,...,n;j=1,...,k and P satisfy the universal conditions of Bi(n, k). It Φ(ui,i) = u′ i,j, and Φ(P) = P. Let zi = xi−n−1 for i = 1, ..., n. Since (xi)i∈Z are boolean spreadable, (zi)i=1,...,n is boolean spreadable. Therefore, for i1, ..., iL ∈ [k], we have φ(yi1 · · · yiL)P = φ(yn−k+i1 · · · yn−k+iL)P = φ(x−n+k−i1 · · · xn−k−iL)P = Φ(φ(zk+1−i1 · · · zk+1−iL)P) n = Φ( φ(zj1 · · · zjL)Puj1,k+1−i1 · · · ujL,k+1−iLP) φ(zj1 · · · zjL)Pun+1−j1,i1 · · · un+1−jL,iLP n n Pj1,...,jL=1 Pj1,...,jL=1 Pj1,...,jL=1 Pj1,...,jL=1 Pj1,...,jL=1 n n = = = = φ(xj1−n−1 · · · xjL−n−1)Pun+1−j1,i1 · · · un+1−jL,iLP φ(yn+1−j1 · · · yn+1−jL)Pun+1−j1,i1 · · · un+1−jL,iLP φ(yj1 · · · yjL)Puj1,i1 · · · ujL,iLP which completes the proof. (cid:3) Proposition 9.2. (A, φ) is a W ∗-probability space with a non-degenerated normal state and A is generated by a bilateral sequence of random variables (xi)i∈Z and (xi)i∈Z are boolean spreadable. Then, E− and E+ can be extend to the whole algebra A. Moreover, E− = E+ Proof. Since (xi)i∈Z is boolean spreadable, (xi)i∈Z is monotonically spreadable. By propo- sition 8.2 E− can be extended to the whole algebra. By Lemma 9.1, (x−i)i∈Z is also boolean spreadable and its negative-conditional expectation is exactly the positive conditional ex- pectation of (xi)i∈Z. Therefore, E+ can also be extended the whole algebra A normally. Give a, b, c ∈ A[−∞,∞], by Lemma 8.2, we have NONCOMMUTATIVE SPREADABILITY 43 φ(aE−[b]c) = φ(E+[a]bE+[c]) = φ(E+[E+[a]bE+[c]]) = φ(E+[a]E+[b]E+[c]) = lim n→∞ = lim n→∞ φ(αn(a)E+[b]E+[c]) lim m→∞ φ(αn(a)E+[b]αm(c)) Notice that, for fixed n, m, φ(αn(a)E+[b]αm(c)) = φ(αn(a)αL(b)αm(c)) for L ∈ N which is large enough. Since (x−i)i∈Z is monotonically spreadable, by theorem 1.1, (x−i)i∈Z is monotonically independent with respect to E−. Therefore, we have φ(αn(a)E+[b]αm(c)) = φ(αn(a)αL(b)αm(c)) = φ(E−[αn(a)αL(b)αm(c)]) = φ(E−[αn(a)]E−[αL(b)]E−[αm(c)]) = φ(E−[a]E−[b]E−[c]) = φ(E−[E−[a]bE−[c]]) = φ(E−[a]bE−[c]) = φ(aE+[b]c) φ(aE−[b]c) = φ(E+[a]bE+[c]) = lim lim n→∞ m→∞ = lim lim n→∞ m→∞ = φ(aE+[b]c) φ(αn(a)E+[b]αm(c)) φ(aE+[b]c) It implies that E+[b] = E−[b] for all b ∈ A[−∞,∞]. Since A is the WOT closure of A[−∞,∞], the proof is complete. (cid:3) Corollary 9.3. (A, φ) is a W ∗-probability space with a non-degenerated normal state and A is generated by a bilateral sequence of random variables (xi)i∈Z and (xi)i∈Z are boolean spreadable. Then, the positive tail algebra and the negative tail algebra of (xi)i∈Z are the same. Now, we are ready to prove theorem 1.3 Theorem 9.4. Let (A, φ) be a non degenerated W ∗-probability space and (xi)i∈Z be a bilat- eral infinite sequence of selfadjoint random variables which generate A as a von Neumann algebra. Then the following are equivalent: a) The joint distribution of (xi)i∈N is boolean spreadable. b) The sequence (xi)i∈Z is identically distributed and boolean independent with respect to the φ−preserving conditional expectation E+ onto the non unital positive tail algebra of the (xi)i∈Z Proof. "b) ⇒ a)". If the sequence (xi)i∈Z is identically distributed and boolean indepen- dent with respect to a φ−preserving conditional expectation E , then sequence (xi)i∈Z is boolean exchangeable by theorem 7.1 in [16]. According the diagram in section 4, (xi)i∈Z 44 WEIHUA LIU is boolean spreadable. "a) ⇒ b)". By Lemma9.2, (xi)i∈Z is monotone with respect to E+, (x−i)i∈Z is monotone with respect to E− and E+ = E−. Therefore, E+[p1(xi1) · · · pm(xim)] = E+[p1(xi1)]E+[p2(xi2) · · · pm(xim)] = · · · = E+[p1(xi1)]E+[p2(xi2)] · · · · · · E+[pm(xim)] whenever i1 6= · · · 6= im and p1, ..., pm ∈ A+ tailhXi. The proof is complete. (cid:3) References [1] Teodor Banica, Stephen Curran, and Roland Speicher. "Classification results for easy quantum groups". In: Pacific J. Math. 247.1 (2010), pp. 1 -- 26. issn: 0030-8730. doi: 10.2140/pjm.2010.247.1. url: http://dx.doi.org/10.2140/pjm.2010.247.1. [2] Teodor Banica, Stephen Curran, and Roland Speicher. "De Finetti theorems for easy quantum groups". In: Ann. Probab. 40.1 (2012), pp. 401 -- 435. issn: 0091-1798. doi: 10.1214/10-AOP619. url: http://dx.doi.org/10.1214/10-AOP619. [3] Marek Bozejko and Roland Speicher. "ψ-independent and symmetrized white noises". In: Quantum probability & related topics. QP-PQ, VI. World Sci. Publ., River Edge, NJ, 1991, pp. 219 -- 236. [4] Vitonofrio Crismale and Francesco Fidaleo. "Exchangeable stochastic processs and symmetric states in quantum probability". In: (2014). doi: DOI10.1007/s10231-014-0407-5. [5] Stephen Curran. "A characterization of freeness by invariance under quantum spread- ing". In: J. Reine Angew. Math. 659 (2011), pp. 43 -- 65. issn: 0075-4102. doi: 10.1515/CRELLE.2011.066. url: http://dx.doi.org/10.1515/CRELLE.2011.066. [6] Stephen Curran and Roland Speicher. "Quantum invariant families of matrices in free probability". In: J. Funct. Anal. 261.4 (2011), pp. 897 -- 933. issn: 0022-1236. doi: 10.1016/j.jfa.2011.04.004. url: http://dx.doi.org/10.1016/j.jfa.2011.04.004. [7] P. Diaconis and D. Freedman. "Partial exchangeability and sufficiency". In: Statis- tics: applications and new directions (Calcutta, 1981). Indian Statist. Inst., Calcutta, 1984, pp. 205 -- 236. [8] Uwe Franz. "Monotone and Boolean convolutions for non-compactly supported probability measures". In: Indiana Univ. Math. J. 58.3 (2009), pp. 1151 -- 1185. issn: 0022-2518. doi: 10.1512/iumj.2009.58.3578. url: http://dx.doi.org/10.1512/iumj.2009.58.3578. [9] Uwe Franz. "Multiplicative monotone convolutions". In: Quantum probability. Vol. 73. Banach Center Publ. Polish Acad. Sci., Warsaw, 2006, pp. 153 -- 166. doi: 10.4064/bc73-0-10. url: http://dx.doi.org/10.4064/bc73-0-10. [10] Amaury Freslon and Moritz Weber. "On bi-free De Finetti theorems". In: arXiv:1501.05124 (). [11] Richard V. Kadison and John R. Ringrose. Fundamentals of the theory of operator algebras. Vol. II. Vol. 16. Graduate Studies in Mathematics. Advanced theory, Cor- rected reprint of the 1986 original. American Mathematical Society, Providence, RI, 1997, i -- xxii and 399 -- 1074. isbn: 0-8218-0820-6. REFERENCES 45 [12] Olav Kallenberg. Probabilistic symmetries and invariance principles. Probability and its Applications (New York). Springer, New York, 2005, pp. xii+510. isbn: 978-0387- 25115-8; 0-387-25115-4. [13] Olav Kallenberg. "Spreading-invariant sequences and processes on bounded index sets". In: Probab. Theory Related Fields 118.2 (2000), pp. 211 -- 250. issn: 0178-8051. doi: 10.1007/s440-000-8015-x. url: http://dx.doi.org/10.1007/s440-000-8015-x. [14] Claus Kostler. "A noncommutative extended de Finetti theorem". In: J. Funct. Anal. 258.4 (2010), pp. 1073 -- 1120. issn: 0022-1236. doi: 10.1016/j.jfa.2009.10.021. url: http://dx.doi.org/10.1016/j.jfa.2009.10.021. [15] Claus Kostler and Roland Speicher. "A noncommutative de Finetti theorem: in- variance under quantum permutations is equivalent to freeness with amalgama- tion". In: Comm. Math. Phys. 291.2 (2009), pp. 473 -- 490. issn: 0010-3616. doi: 10.1007/s00220-009-0802-8. url: http://dx.doi.org/10.1007/s00220-009-0802-8. [16] Weihua Liu. "A noncommutative De Finetti theorem for boolean independence". In: arXiv:1403.1772 (2014). [17] Weihua Liu. "Boolean analogue of De Finetti theorems for easy quantum groups". In: (in preparation). [18] Naofumi Muraki. "Noncommutative Brownian motion in monotone Fock space". In: Comm. Math. Phys. 183.3 (1997), pp. 557 -- 570. issn: 0010-3616. doi: 10.1007/s002200050043. url: http://dx.doi.org/10.1007/s002200050043. [19] Naofumi Muraki. "The five independences as natural products". In: Infin. Dimens. Anal. Quantum Probab. Relat. Top. 6.3 (2003), pp. 337 -- 371. issn: 0219-0257. doi: 10.1142/S0219025703001365. url: http://dx.doi.org/10.1142/S0219025703001365. [20] Alexandru Nica and Roland Speicher. Lectures on the combinatorics of free probabil- ity. Vol. 335. London Mathematical Society Lecture Note Series. Cambridge Univer- sity Press, Cambridge, 2006, pp. xvi+417. isbn: 978-0-521-85852-6; 0-521-85852-6. doi: 10.1017/CBO9780511735127. url: http://dx.doi.org/10.1017/CBO9780511735127. [21] Mihai Popa. "A combinatorial approach to monotonic independence over a C ∗- algebra". In: Pacific J. Math. 237.2 (2008), pp. 299 -- 325. issn: 0030-8730. doi: 10.2140/pjm.2008.237.299. url: http://dx.doi.org/10.2140/pjm.2008.237.299. [22] C. Ryll-Nardzewski. "On stationary sequences of random variables and the de Finetti's equivalence". In: Colloq. Math. 4 (1957), pp. 149 -- 156. issn: 0010-1354. [23] Piotr M. So ltan. "Quantum families of maps and quantum semigroups on finite quantum spaces". In: J. Geom. Phys. 59.3 (2009), pp. 354 -- 368. issn: 0393-0440. doi: 10.1016/j.geomphys.2008.11.007. url: http://dx.doi.org/10.1016/j.geomphys.2008.11.007. [24] Piotr Miko laj So ltan. "On quantum semigroup actions on finite quantum spaces". In: Infin. Dimens. Anal. Quantum Probab. Relat. Top. 12.3 (2009), pp. 503 -- 509. issn: 0219-0257. doi: 10.1142/S0219025709003768. url: http://dx.doi.org/10.1142/S0219025709003768. [25] Roland Speicher. "On universal products". In: Free probability theory (Waterloo, ON, 1995). Vol. 12. Fields Inst. Commun. Amer. Math. Soc., Providence, RI, 1997, pp. 257 -- 266. [26] Roland Speicher and Reza Woroudi. "Boolean convolution". In: Free probability the- ory (Waterloo, ON, 1995). Vol. 12. Fields Inst. Commun. Amer. Math. Soc., Prov- idence, RI, 1997, pp. 267 -- 279. 46 REFERENCES [27] S¸erban Stratila. Modular theory in operator algebras. Translated from the Romanian by the author. Editura Academiei Republicii Socialiste Romania, Bucharest; Abacus Press, Tunbridge Wells, 1981, p. 492. isbn: 0-85626-190-4. [28] D. V. Voiculescu, K. J. Dykema, and A. Nica. Free random variables. Vol. 1. CRM Monograph Series. A noncommutative probability approach to free products with applications to random matrices, operator algebras and harmonic analysis on free groups. American Mathematical Society, Providence, RI, 1992, pp. vi+70. isbn: 0-8218-6999-X. [29] Dan Voiculescu. "Symmetries of some reduced free product C ∗-algebras". In: Op- erator algebras and their connections with topology and ergodic theory (Bu¸steni, 1983). Vol. 1132. Lecture Notes in Math. Springer, Berlin, 1985, pp. 556 -- 588. doi: 10.1007/BFb0074909. url: http://dx.doi.org/10.1007/BFb0074909. [30] Dan-Virgil Voiculescu. "Free probability for pairs of faces I". In: Comm. Math. Phys. 332.3 (2014), pp. 955 -- 980. issn: 0010-3616. doi: 10.1007/s00220-014-2060-7. url: http://dx.doi.org/10.1007/s00220-014-2060-7. [31] Shuzhou Wang. "Free products of compact quantum groups". In: Comm. Math. Phys. 167.3 (1995), pp. 671 -- 692. issn: 0010-3616. url: http://projecteuclid.org/euclid.cmp/1104272163. [32] Shuzhou Wang. "Quantum symmetry groups of finite spaces". In: Comm. Math. Phys. 195.1 (1998), pp. 195 -- 211. issn: 0010-3616. doi: 10.1007/s002200050385. url: http://dx.doi.org/10.1007/s002200050385. [33] S. L. Woronowicz. "Compact matrix pseudogroups". In: Comm. Math. Phys. 111.4 (1987), pp. 613 -- 665. issn: 0010-3616. url: http://projecteuclid.org/euclid.cmp/1104159726. Department of Mathematics University of California at Berkeley Berkeley, CA 94720, USA E-MAIL: [email protected]
1005.1999	1	1005	2010-05-12T06:51:00	Semi-invertible extensions of C*-algebras	[ "math.OA" ]	We prolonge the list of C-algebras for which all extensions by any stable separable C-algebra are semi-invertible. In particular, we handle certain amalgamations, both of C-algebras and of groups. Concerning groups we consider both reduced and full group C-algebras.	math.OA	math	SEMI-INVERTIBLE EXTENSIONS OF C ∗-ALGEBRAS VLADIMIR MANUILOV AND KLAUS THOMSEN 1. Introduction and statements of results The number of examples of C ∗-algebras for which the semi-group of extensions by the compact operators is not a group was only slowly increasing during the first decades following the first example of J. Anderson,[A], but recently the pace has picked up, cf. [HT], [HS], [HLSW] and [Se], and there are now whole series of C ∗- algebras A for which it is known that there are non-invertible extensions of A by the C ∗-algebra of compact operators K. Furthermore, by considering extensions by general stable C ∗-algebras the stock of examples of non-invertible extensions grows considerably. Indeed, a non-invertible extension of a C ∗-algebra A by K gives rise to a non-invertible extension of A by B ⊗ K for any unital C ∗-algebra B.1 In a different direction the authors have shown that many of the non-invertible extensions are invertible in a slightly weaker sense, called semi-invertibility. Recall that an extension of a C ∗-algebra A by a stable C ∗-algebra B is invertible when there is another extension, the inverse, with the property that the direct sum extension of the two is a split extension. Semi-invertibility requires only that the sum is asymptotically split, in the sense that there is an asymptotic homomorphism as defined by Connes and Higson, [CH], consisting of right-inverses of the quotient map. It turns out that extensions of a suspended or a contractible C ∗-algebra are always semi-invertible, [MT3], [MT1], and in [ST] it was shown that the extensions of the reduced group C ∗-algebra of a free product of amenable groups are all semi- invertible. The main purpose of the present paper is to prolonge this list of C ∗- algebras for which all the extensions by a separable stable C ∗-algebra are semi- invertible. To explain why semi-invertibility is a natural notion which can be considered as the best alternative when invertibility fails, we recall first the central definitions. Let A and B be separable C ∗-algebras. The multiplier algebra of B will be denoted by M(B), the generalized Calkin algebra of B by Q(B) and qB : M(B) → Q(B) is then the canonical surjection. We let Ext(A, B) denote the semi-group of unitary equiva- lence classes of extensions of A by B. Thus elements of Ext(A, B) are represented by ∗-homomorphisms ϕ : A → Q(B) and two extensions ϕ, ψ : A → Q(B) are unitarily equivalent when there is a unitary u ∈ M(B) such that Ad qB(u) ◦ ϕ = ψ. The ad- dition ϕ ⊕ ψ of two extensions is defined from a choice of isometries V1, V2 ∈ M(B) such that V1V ∗ 2 = 1 to be the extension 1 + V2V ∗ (ϕ ⊕ ψ) (a) = qB (V1) ϕ(a)qB (V1)∗ + qB (V2) ψ(a)qB (V2)∗ . Version: November 15, 2018. 1Tensor the non-invertible extension with B using the maximal tensor-product, and pull back along the unital inclusion A ⊆ A⊗max B. It is easy to see that the resulting extension of A by B ⊗K does not have a completely positive section for the quotient map because the original extension does not. 1 2 VLADIMIR MANUILOV AND KLAUS THOMSEN An extension ϕ : A → Q(B) is split when there is a ∗-homomorphism π : A → M(B) such that ϕ = qB ◦ π and asymptotically split when there is an asymptotic homomorphism πt : A → M(B), t ∈ [1, ∞), such that qB ◦ πt = ϕ for all t. We say that Ext(A, B) is a group when every extension ϕ : A → Q(B) has an inverse, meaning that there is another extension ϕ′ : A → Q(B), the inverse of ϕ, such that ϕ ⊕ ϕ′ is split. An extension ϕ : A → Q(B) is semi-invertible when there is another extension ϕ′ : A → Q(B) such that ϕ ⊕ ϕ′ is asymptotically split. When the theory of C ∗-extensions was first introduced, in the work of Brown, Douglas and Fillmore, [BDF1], [BDF1], the authors had very good (operator the- oretic) reasons for wanting to trivialize the split extensions. 2 However, there are other reasons why split extensions must be trivialized in order to get a group from the semi-group Ext(A, B). For a split extension ψ it makes sense to define the direct sum ψ∞ of a countably infinite collection of copies of ψ. Since ψ ⊕ ψ∞ ⊕ 0 = ψ∞ ⊕ 0 in Ext(A, B) this shows that split extensions are trivial in any group-quotient of It is not difficult to show that ψ∞ can also be defined when the ex- Ext(A, B). tension ψ is asymptotically split. In fact, this is possible as soon as the extension splits via a discrete asymptotic homomorphism, e.g when it is quasi-diagonal. But by using the real parameter for the asymptotic section it can also be arranged that ψ ⊕ ψ∞ ⊕ 0 becomes unitarily equivalent to ψ∞ ⊕ 0. It follows that also asymp- totically split extensions must vanish in a group-quotient of Ext(A, B). In fact, any group-quotient of Ext(A, B) must factor through the cancellation semi-group of Ext(A, B). In retrospect it seems therefore not particularly surprising that it is not generally enough to trivialize only the split extensions to get a group, or even the asymptotically split extensions, as demonstrated in [MT4]. In fact, seen through the right looking-glasses it seems more surprising that Ext(A, B) actually is a group in so many cases, and that semi-invertibility prevails in many cases where invertibility fails. Complementing on the cases covered by the results in [MT3], [MT1], [M], [Th4] and [ST] we shall show in this paper that all extensions in Ext(A, B) are semi- invertible when a) A is the reduced group C ∗-algebra C ∗ r (G) and the group G is an amalgamated free product G = G1 ∗F G2 with F finite, G2 is amenable and G1 abelian, and when b) A is the amalgamated free product of C ∗-algebras, A = A1 ∗D A2, when D is nuclear and all extensions of Ai by B are semi-invertible, i = 1, 2. The result concerning a) is actually slightly more general and involves a KK-theory condition which is automatically fullfilled when G1 is abelian. Furthermore we es- tablish a few permanence properties for semi-invertibility: If all extensions of A and A′ by B are semi-invertible then so are all extensions of A ⊕ A′ by B, all extensions of C(T) ⊗ A by B and all extensions of K ⊗ A by B. It follows from this that all extensions of A by B are semi-invertible when a') A = C ∗ r (G′) provided G′ = Zk × H × G where H is a finite group and G is an amalgamated free product as in a) above, and when 2They also had good reasons for restricting the attention to essential extensions, but that's another story. EXTENSIONS OF C ∗-ALGEBRAS 3 b') A is the full group C ∗-algebra C ∗(Zk × H × G′′) where H is a finite group and G′′ is obtained through successive amalgamations G′′ = (· · · ((G1 ∗H1 G2) ∗H2 G3) ∗H3 . . . . . . ) ∗Hn−1 Gn, provided all the groups H1, H2, . . . , Hn−1 are amenable, and all extensions of C ∗(Gi) by B are semi-invertible, i = 1, 2, . . . , n. While we know from [HS], [HLSW] and [Se] that there are non-invertible extensions of A by B in many of the cases dealt with in a), our ignorance concerning invertibility of the extensions handled by b') is complete: There is no known example of an extension of a full group C ∗-algebra by a stable C ∗-algebra which is not invertible. The proof of a) above is an elaboration of the ideas developed in [M], [Th4] and [ST]. In particular, the argument uses the notion of strong homotopy of extensions and depends on Lemma 4.3 in [MT1]. In contrast the method of proof of b) is new and does not use strong homotopy of extensions. Instead a key step uses methods devised for the classification of C ∗-algebras by Lin, Dadarlat and Eilers. This difference in the proofs has consequences for the conclusions we obtain; in case a) the inverse (for semi-invertibility) can be chosen to be invertible while we do not know if this is so in case b). Acknowledgement. The main part of this work was done during a stay of both authors at the Mathematische Forchungsinstitut in Oberwolfach in January 2010 in the framework of the 'Research in Pairs' programme. We want to thank the MFO for the perfect working conditions. 2. The reduced group C ∗-algebra of free products with amalgamation over a finite subgroup Throughout A and B are separable C ∗-algebras and B is stable. Two extensions ϕ, ϕ′ : A → Q(B) are strongly homotopic when there is a path ψt, t ∈ [0, 1], of extensions ψt : A → Q(B) such that 1) t 7→ ψt(a) is continuous for all a ∈ A, and 2) ψ0 = ϕ and ψ1 = ϕ′. By Lemma 4.3 of [MT1] we have the following Theorem 2.1. Assume that two extensions ϕ, ϕ′ : A → Q(B) are strongly homo- topic. Then ϕ is asymptotically split if and only if ϕ′ is asymptotically split. In some of the cases we deal with below we show that for any extension ϕ : A → Q(B) there is an extension ψ : A → Q(B) such that ϕ ⊕ ψ is strongly homotopic to a split extension. This will be expressed by saying that ϕ is strongly homotopy invertible. Thanks to Theorem 2.1 this implies that ϕ is semi-invertible. In some cases it turns out that ψ can be taken to be invertible. We express this by saying that ϕ is strongly homotopy invertible with an invertible inverse. Lemma 2.2. Let Gi, i = 1, 2, be discrete countable amenable groups with a common finite subgroup H ⊆ Gi, i = 1, 2. Let G1 ∗H G2 be the amalgamated free product group. Let µ : C ∗(G1 ∗H G2) → C ∗ r (G1 ∗H G2) be the canonical surjection and let hτ : C ∗(G1 ∗H G2) → C be the character corresponding to the trivial one-dimensional representation of G1 ∗H G2. There are then a separable infinite-dimensional Hilbert space H, ∗-homomorphisms σ, σ0 : C ∗ r (G1 ∗H G2) → B(H), and a path ζs : C ∗(G1 ∗H G2) → B(H), s ∈ [0, 1], 4 VLADIMIR MANUILOV AND KLAUS THOMSEN of unital ∗-homomorphisms such that a) ζ0 = σ ◦ µ; b) ζ1 = hτ ⊕ σ0 ◦ µ; c) ζs(a) − ζ0(a) ∈ K, s ∈ [0, 1], and d) s 7→ ζs(a) is continuous for all a ∈ C ∗(G1 ∗H G2). Proof. Set G = G1 ∗H G2. Being amenable Gi has the Haagerup Property. See the discussion in 1.2.6 of [CCJJV]. It follows then from Propositions 6.1.1 and 6.2.3 of [CCJJV] that also G has the Haagerup Property. Since the Haagerup Property implies K-amenability by [Tu] (or Theorem 1.2 in [HK]) we conclude that G is K- amenable. We can therefore find a separable infinite-dimensional Hilbert space H and ∗-homomorphisms σ, σ0 : C ∗ r (G) → B(H) such that σ and hτ ⊕ σ0 are both unital and 1) σ ◦ µ(x) − (hτ ⊕ σ0 ◦ µ) (x) ∈ K, x ∈ C ∗(G), and 2) [σ ◦ µ, hτ ⊕ σ0 ◦ µ] = 0 in KK (C ∗(G), K), cf. [C]. By adding the same unital and injective ∗-homomorphism to σ and σ0 we can arrange that both σ and σ0 are injective and have no non-zero compact operator in their range. Since µC ∗(Gi) : C ∗(Gi) → C ∗ r (Gi) is injective because Gi is amenable, it follows that σ ◦ µC ∗(Gi) and (hτ ⊕ σ0 ◦ µ)C ∗(Gi) are admissible in the sense of Section 3 of [DE] for each i. Thus Theorem 3.12 of [DE] applies to show that there is a norm-continuous path ui s, s ∈ [1, ∞), of unitaries in 1 + K such that lim s→∞(cid:13)(cid:13)σ ◦ µC ∗(Gi)(a) − ui for all a ∈ C ∗(Gi) and s (hτ ⊕ σ0 ◦ µ) C ∗(Gi)(a)ui s ∗(cid:13)(cid:13) = 0 σ ◦ µC ∗(Gi)(a) − ui s (hτ ⊕ σ0 ◦ µ) C ∗(Gi)(a)ui s ∗ ∈ K for all a ∈ C ∗ (Gi) and all s ∈ [1, ∞). Set F = (hτ ⊕ σ0 ◦ µ) (C ∗(H)) (2.1) (2.2) which is a finite dimensional unital C ∗-subalgebra of B(H), and let P : B(H) → F ′ ∩ B(H) be the conditional expectation given by P (x) = ZU (F ) uxu∗ du, where we integrate with respect to the Haar-measure on the unitary group U(F ) of F . Note that P (1 + K) ⊆ 1 + K. It follows from (2.1) that u2 s asymptotically s commutes with elements of F and hence also that ∗u1 (2.3) ∗u1 lim s→∞(cid:13)(cid:13)P (cid:0)u2 s s(cid:1) − u2 ∗u1 s s(cid:13)(cid:13) = 0. Standard C ∗-algebra techniques provides us then with a norm-continuous path vt, t ∈ [1, ∞), of unitaries in F ′ ∩ (1 + K) such that lims→∞(cid:13)(cid:13)vs − P (cid:0)u2 combined with (2.3) implies that s ∗u1 s(cid:1)(cid:13)(cid:13) = 0, which lim s→∞(cid:13)(cid:13)u2 svs − u1 s(cid:13)(cid:13) = 0. It follows that we can work with u2 and (2.2) we have also that svs in the place of u1 s to arrange that besides (2.1) Ad u1 s ◦ (hτ ⊕ σ0 ◦ µ) C ∗(H) = Ad u2 s ◦ (hτ ⊕ σ0 ◦ µ) C ∗(H) EXTENSIONS OF C ∗-ALGEBRAS 5 for all s. It follows that the ∗-homomorphisms ψ′ s = (cid:0)Ad u1 s ◦ (hτ ⊕ σ0 ◦ µ)(cid:1) ∗C ∗(H) (cid:0)Ad u2 s ◦ (hτ ⊕ σ0 ◦ µ)(cid:1) are all defined and give us a norm-continuous path of unital ∗-homomorphisms ηs : C ∗(G) → B(H), s ∈ [0, 1], such that 1 ◦ (hτ ⊕ σ0 ◦ µ)) ∗C ∗(H) (Ad u2 a') η0 = (Ad u1 b') η1 = σ ◦ µ; c') ηs(a) − η0(a) ∈ K, a ∈ C ∗(G), s ∈ [0, 1]. 1 ◦ (hτ ⊕ σ0 ◦ µ)); The unitary group of F ′∩(C1 + K) is norm-connected; a fact which can be seen either from the spectral theory of compact operators or by observing that the algebra is AF. 1 to 1 in F ′ ∩ (1 + K) By using first a continuous path of unitaries connecting u2 1 and then a continuous path of unitaries connecting u2 1 to 1 in the unitary group of 1 + K, we obtain continuous paths w1 s, s ∈ [0, 1], of unitaries in 1 + K such that w1 1, w2 s ◦ (hτ ⊕ σ0 ◦ µ) C ∗(H) = Ad w2 s ◦ (hτ ⊕ σ0 ◦ µ) C ∗(H) for all s ∈ [0, 1]. It follows that the ∗-homomorphisms s and w2 1 = u2 1 and Ad w1 0 = 1, w1 0 = w2 1 = u1 ∗u1 η′ s = (cid:0)Ad w1 s ◦ (hτ ⊕ σ0 ◦ µ)(cid:1) ∗C ∗(H) (cid:0)Ad w2 s ◦ (hτ ⊕ σ0 ◦ µ)(cid:1) are all defined and give us a norm-continuous path of unital ∗-homomorphisms s : C ∗(G) → B(H), s ∈ [0, 1], such that η′ a") η′ b") η′ c") η′ 0 = hτ ⊕ (σ0 ◦ µ); 1 = (Ad u1 s(a) − η′ 1 ◦ (hτ ⊕ σ0 ◦ µ)) ∗C ∗(H) (Ad u2 0(a) ∈ K, a ∈ C ∗(G), s ∈ [0, 1]. 1 ◦ (hτ ⊕ σ0 ◦ µ)); The desired path ζ is then obtained by concatenation of the paths, η and η′. (cid:3) Theorem 2.3. Let Gi, i = 1, 2, be discrete countable amenable groups with a com- mon finite subgroup H ⊆ Gi, i = 1, 2, and let B be a separable stable C ∗-algebra. Let G1 ∗H G2 be the amalgamated free product group. Assume that the map i∗ 1 − i∗ 2 : KK (C ∗(G1), B) ⊕ KK (C ∗(G2), B) → KK (C ∗(H), B) , induced by the inclusions ij : C ∗(H) → C ∗(Gj), j = 1, 2, is rationally surjective, i.e. for every x ∈ KK (C ∗(H), B) there is an n ∈ N\{0} such that nx is in the range of 1 − i∗ i∗ 2. It follows that every extension of C ∗ r (G1 ∗H G2) by B is strongly homotopy invert- ible with an invertible inverse. Proof. Set G = G1 ∗H G2 and consider an extension ϕ : C ∗ r (G1 ∗H G2) → Q(B). Since C ∗(G) ≃ C ∗ (G1) ∗C ∗(H) C ∗ (G2) it follows from Proposition 2.8 of [Th2] that every extension of C ∗(G) by B is invertible. As observed in the proof of Lemma 2.2, G is K-amenable and it follows therefore from [C] that µ∗ : Ext−1 (C ∗ r (G), B) → Ext−1 (C ∗(G), B) is an isomorphism. In particular the inverse of ϕ ◦ µ is in the range of µ∗, which means that there is an invertible extension ϕ′′ : C ∗ r (G) → Q(B) such that [ϕ ◦ µ ⊕ ϕ′′ ◦ µ] = 0 (2.4) in Ext−1 (C ∗(G), B). Let β0 : C ∗ r (G) → M(B) be an absorbing homomorphism, whose existence is guaranteed by [Th1] and set ϕ′ = ϕ ⊕ qB ◦ β0. By Lemma 2.2 of r (Gi) → M(B) is absorbing for each i = 1, 2. Since Gi is amenable [Th2] β0C ∗ µC ∗(Gi) : C ∗(Gi) → C ∗ r (Gi) is a ∗-isomorphism and it follows therefore from (2.4) that (ϕ′ ◦ µ ⊕ ϕ′′ ◦ µ) C ∗(Gi) is a split extension for each i. In other words, there r (Gi) : C ∗ 6 VLADIMIR MANUILOV AND KLAUS THOMSEN are ∗-homomorphisms πi : C ∗ (Gi) → M(B) such that (ϕ′ ◦ µ ⊕ ϕ′′ ◦ µ) C ∗(Gi) = qB ◦ πi, i = 1, 2. Note that π1(x) − π2(x) ∈ B for all x ∈ C ∗(H) so that (π1, π2) represents an element of KK (C ∗(H), B). We need to change the situation to a case where this pair represents 0 in KK (C ∗(H), B). This is done as follows: β0C ∗(Gi), i = 1, 2, are both absorbing so after adding qB ◦ β0 to ϕ′′ we get a situation where there are unitaries ui ∈ M(B) such that Ad ui ◦ πi(y) − β0(y) ∈ B for all y ∈ C ∗(Gi), i = 1, 2. Then ϕ′ ◦ µ ⊕ ϕ′′ ◦ µ = AdqB(u∗ 2) ◦(cid:0)(cid:0)qB ◦ Ad u2u∗ 1 ◦ β0C ∗(G1)(cid:1) ∗C ∗(H) (cid:0)qB ◦ β0C ∗(G2)(cid:1)(cid:1) . It follows that we can choose the lifts, π1, π2, above such that (cid:2)π1C ∗(H), π2C ∗(H)(cid:3) = (cid:2)Ad w ◦ β0C ∗(H), β0C ∗(H)(cid:3) in KK (C ∗(H), B) where w = u2u∗ 1. To proceed we need a description of the KK-groups obtained in [Th1] and [Th3]: When A is a separa- ble C ∗-algebra and α : A → M(B) is an absorbing ∗-homomorphism, there is an isomorphism between K1 (Dα(A)) and KK(A, B), where Dα(A) = {m ∈ M(B) : α(a)m − mα(a) ∈ B ∀a ∈ A} . (2.5) The isomorphism sends a unitary u ∈ Dα(A) to [Ad u ◦ α, α]. Ignoring the passage to matrices in K1 our assumption implies, in this picture of KK-theory, that there is an n > 0 and a norm-continuous path of unitaries in Dβ0 (C ∗(H)) connecting wn to a product w∗ 2w1, where wi ∈ Dβ0 (C ∗(Gi)) , i = 1, 2. Then(cid:2)Ad wn ◦ β0C ∗(H), β0C ∗(H)(cid:3) = qB ◦ β0 ◦ µ = (cid:0)qB ◦ Ad w∗ (cid:2)Ad w1 ◦ β0C ∗(H), Ad w2 ◦ β0C ∗(H)(cid:3) in KK(C ∗(H), B). Note that 1 ◦ β0C ∗(G1)(cid:1) ∗C ∗(H) (cid:0)qB ◦ Ad w∗ 2 ◦ β0C ∗(G2)(cid:1) . After adding (ϕ′ ⊕ ϕ′′) ⊕ (ϕ′ ⊕ ϕ′′) ⊕ · · · ⊕ (ϕ′ ⊕ ϕ′′) ⊕ qB ◦ β0 n−1 times {z } to ϕ′′ we come in a position where the pair (π1, π2) can be chosen such that [π1, π2] = 0 in KK (C ∗(H), B). (If we take the passage to matrices in K1 into account in the previous argument, it may be necessary to add a finite direct sum of copies of qB ◦ β0 instead of a single copy.) We can then proceed as follows: Set β = qB ◦ β∞ is the direct sum of a sequence of copies of β0. By adding β to ϕ′′ we come then in a situation where Theorem 3.8 of [DE] applies to give us a continuous path ut, t ∈ [1, ∞), of unitaries in 1 + B such that 0 where β∞ 0 lim t→∞ Ad ut ◦ π1(x) = π2(x) for all x ∈ C ∗(H). Since C ∗(H) is finite dimensional we have that for t large enough there is a unitary v ∈ 1 + B such that vutπ1(x)u∗ t v∗ = π2(x) for all x ∈ C ∗(H). Hence, by exchanging π1 with Ad vut ◦ π1 we conclude that ϕ′ ◦ µ ⊕ ϕ′′ ◦ µ is split. By a standard argument, based on Kasparov's stabilization theorem, we may add a split extension to arrange that ϕ′ ◦ µ ⊕ ϕ′′ ◦ µ = qB ◦ χ ⊕ 0 where χ : C ∗(G) → M(B) is a unital ∗-homomorphism. Let γ : G → M(B) be the unitary representation of G defined by χ and let ζs be the continuous path of ∗-homomorphisms from Lemma 2.2, and νs the corresponding unitary representations. Let hγ⊗νs be the EXTENSIONS OF C ∗-ALGEBRAS 7 ∗-homomorphism C ∗(G) → M(B) defined from the tensor product representation γ ⊗ νs by use of a spatial isomorphism B ⊗ K ≃ B. Then qB ◦ hγ⊗νs, s ∈ [0, 1], is a strong homotopy of extensions of C ∗ (G) by B. By the argument used in the proof of Theorem 2.3 of [Th3] and again in the proof of Theorem 2.2 in [ST] the properties of {ζs} ensure that this homotopy factors through C ∗ r (G) and gives us a strong homotopy, as well as split extensions ψ, ψ′, of C ∗ r (G) by B connecting ϕ ⊕ qB ◦ β0 ⊕ ϕ′′ ⊕ ψ = ϕ′ ⊕ ϕ′′ ⊕ ψ to ψ′. Since qB ◦ β0 ⊕ ϕ′′ ⊕ ψ is invertible, this completes the proof. (cid:3) As in [ST] the fact that the strong homotopy inverse is invertible implies that the group Ext−1/2(C ∗ r (G1∗H G2), B) of extensions modulo asymptotically split extensions agrees with the corresponding KK-theory group and can be calculated from the universal coefficient theorem. The proof is the same as in [ST] and we omit it here. The KK-condition of Theorem 2.3 is satisfied when G1 is abelian since in this case already the map i∗ 1 : KK (C ∗ (G1) , B) → KK(C ∗(H), B) is surjective. This follows because there is in this case a ∗-homomorphism p : C ∗ (G1) → C ∗(H) which is a left-inverse for i1. We get in this way the following corollary. Corollary 2.4. Let G1 and G2 be countable discrete amenable groups with a common finite subgroup H ⊆ Gi, i = 1, 2, and B a separable stable C ∗-algebra. Let G1∗H G2 be the amalgamated free product group. Assume that G1 is abelian. It follows that every extension of C ∗ r (G1 ∗H G2) by B is strongly homotopy invertible with an invertible inverse. Example 2.5. It is known that Sl2(Z) ≃ Z4 ∗Z2 Z6, cf. p. 11 in [S]. Hence Corollary 2.4 applies. (As the generator of Z4 one can use 1 0 ), and ( 1 −1 ( 0 −1 1 0 ) can serve as the generator of Z6. The amalgamation is over the subgroup ±1.) It has been shown by Hadwin and Shen in Corollary 4.4 of [HS] that one can get an example of an non-invertible extension of C ∗ r (Sl2(Z)) by K, starting from the non-invertible extension of C ∗ r (F2) found by Haagerup and Thorbjørnsen in [HT]. This means that concerning invertibility of extensions of C ∗ r (Sl2(Z)) the r (F2): For every stabilization B of a unital separable C ∗-algebra situation is as for C ∗ there are non-invertible extensions of C ∗ r (Sl2(Z)) by B, but all are semi-invertible. And the inverse (for semi-invertibility) can be taken to be invertible. For the full group C ∗-algebra C ∗ (Sl2(Z)) the situation is also as for F2, namely that all extensions by C ∗ (Sl2(Z)) are invertible. This follows from [Br] when the ideal is K and from [Th2] when it is an arbitrary separable stable C ∗-algebra. Remark 2.6. The KK-condition of Theorem 2.3 can fail even when G1 and G2 are finite and equal, and H is abelian. Here is the simplest example. Let α be the unique non-trivial automorphism of Z3 which has order 2 and let G1 = Z3 ⋊α Z2 be the semidirect product by this automorphism. Thus G1 is a copy of the symmetric group S3. Set H = Z3 ⊂ G1. Let B = K. Then KK(C ∗(G), B) ∼= R(G) for any finite group G, where R(G) denotes the Grothendieck group of the semigroup generated by 8 VLADIMIR MANUILOV AND KLAUS THOMSEN irreducible (necessarily finite dimensional) representations of G. The functorial map KK(C ∗(G1), B) → KK(C ∗(H), B) becomes the restriction map R(G1) → R(H) after the above identification. The abelian group R(H) is freely generated by the three one-dimensional representations, ρ0, ρ1 and ρ2, that send a fixed generator of H to 1, e2πi/3 and e−2πi/3, respectively. As the number of irreducible representations equals the number of conjugacy classes by the Burnside theorem, and as the group order equals the sum of squares of the dimensions of these representations, it follows that G1 has three irreducible representations; two, σ0 and σ1, of dimension 1 and one, τ , of dimension 2. Thus, R(G1) is freely generated by three representations, σ0, σ1 and τ . One of the one-dimensional representations, σ0, is the identity one, and the other, σ1, maps H to 1 and G1 \ H to −1. Restrictions of both to H equal the trivial representation ρ0 of H. The two-dimensional representation τ is the orthogonal complement to the constant functions in the obvious representation of G1 on l2(H) ∼= C3. Then it is easy to see that τ H = ρ1 ⊕ ρ2. Thus, the restriction map R(G1) → R(H) is not surjective. This example goes only to show that the KK-condition of Theorem 2.3 is not vacuous. For all we know the conclusion of Theorem 2.3 may very well be true without this condition. 3. Amalgamated free product C ∗-algebras In this section we consider free products of C ∗-algebras with amalgamation. The first result is an application of the relative K-homology developed by the authors in [MT2]. Theorem 3.1. Let A1, A2 and B be separable C ∗-algebras, B stable. Let D be a common C ∗-subalgebra of A1 and A2, i.e. D ⊆ A1 and D ⊆ A2. Assume that 1) there is a ∗-homomorphism α0 : A1 ∗D A2 → M(B) such that also α0A1, α0A2 and α0D are absorbing, and assume that 2) Ext(A1, B) and Ext(A2, B) are both groups. It follows that every extension of A1 ∗D A2 by B is strongly homotopy invertible. Proof. Set α = qB ◦ α0 and consider an extension ϕ : A1 ∗D A2 → Q(B). By assumption 2) there is an extension ψi : Ai → Q(B) representing the inverse of ϕAi in Ext(Ai, B) both for i = 1 and i = 2. Then ψ1D and ψ2D represent the same element in Ext(D, B), namely the inverse of the element represented by ϕD. After addition of α0Ai to ϕAi we therefore assume that ψ1D and ψ2D are unitarily equivalent. Thus, after conjugating ψ2 by a unitary, we can arrange that ψ1D = ψ2D. Then ψ = ψ1 ∗D ψ2 : A1 ∗D A2 → Q(B) is defined. Set Φ = ϕ ⊕ ψ. By adding a copy of α to Φ both extensions ΦAi : Ai → Q(B), i = 1, 2, become split, i.e. there are ∗-homomorphisms Φi : Ai → M(B) such that qB ◦Φi = ΦAi, i = 1, 2. By passing to a unitarily equivalent extension, i.e. by conjugating Φ by a unitary of the form qB(u), we can arrange that in addition qB ◦ Φ2 = αA2 and that Φ2 = α0A2. Then qB ◦ Φ1 represents an element of the relative extension semi-group ExtD,αA1 (A1, B), In fact, it follows from Lemma 3.2 of [MT2] and assumption 2) that cf. qB ◦ Φ1 is invertible in this semi-group, i.e. qB ◦ Φ1 ∈ Ext−1 1 : D,αA1 A1 → Q(B) represent the inverse of qB ◦ Φ1 in Ext−1 (A1, B) and note that Φ′ 1 ∗D αA2 : A1 ∗D A2 → Q(B) is then defined. After addition by this extension to Φ we can assume that Φ1 represents 0 in Ext−1 (A1, B). By definition of (A1, B). Let Φ′ [MT2]. D,αA1 D,αA1 EXTENSIONS OF C ∗-ALGEBRAS 9 ExtD,αA1 (A1, B) this means that there is a unitary u in the connected component of 1 in the relative commutant of α(D) in Q(B) such that Ad u ◦ qB ◦ Φ1 = αA1. Let ut, t ∈ [0, 1], be a continuous path of unitaries in α(D)′ ∩ Q(B) such that u0 = 1 and u1 = u. Then ψt = (Ad ut ◦ qB ◦ Φ1) ∗D (qB ◦ Φ2) is defined for every t ∈ [0, 1], and ψt, t ∈ [0, 1], is a strong homotopy of extensions connecting Φ = ψ0 to ψ1 = qB ◦ α. This completes the proof. (cid:3) Condition 1) of Theorem 3.1 is always satisfied when D is nuclear or is the range of a conditional expectation Ai → D for both i = 1 and i = 2, but it can fail in general. See [Th2]. Condition 2) is satisfied when A1 and A2 are nuclear so Theorem 3.1 has the following corollary. Corollary 3.2. Let A1, A2 and B be separable C ∗-algebras, B stable. Let D be a common C ∗-subalgebra of A1 and A2, i.e. D ⊆ A1 and D ⊆ A2. If A1, A2 and D are all nuclear it follows that every extension of A1 ∗D A2 by B is strongly homotopy invertible. The next theorem shows that condition 2) of Theorem 3.1 can be weakened when D is nuclear, at the price of a slightly weaker conclusion. Theorem 3.3. Let A1, A2 and B be separable C ∗-algebras, B stable. Let D be a common C ∗-subalgebra of A1 and A2, i.e. D ⊆ A1 and D ⊆ A2. Assume that 1) there is a ∗-homomorphism β : A1∗DA2 → M(B) such that βD : D → M(B) is absorbing, 2) that Ext(D, B) and Ext (D, C0 ([1, ∞), B)) are both groups, and 3) that all extensions of A1 by B and all extensions of A2 by B are semi- invertible. It follows that all extensions of A1 ∗D A2 by B are semi-invertible. Proof. By adding units to A1, A2 and D if necessary, we may assume that D is unital. (There are various ways to see this; 1. step: (Finding the first candidate for the inverse.) Let ϕ : A1 ∗D A2 → Q(B) be an extension. By assumption 2) there are extensions ψi : Ai → Q(B) such that ϕAi ⊕ ψi : Ai → Q(B) are asymptotically split, i = 1, 2. By assumption 2) Ext(D, B) is a group and hence [ψ1D] = [ψ2D] = − [ϕD] in Ext(D, B). it follows for example from Lemma 4.7 of [MT1].) Furthermore, by assumption 1) there is a ∗-homomorphism β : A1 ∗D A2 → M(B) such that βD is absorbing. So after adding by qB ◦ βA1 to ψ1 and qB ◦ βA2 to ψ2 we may assume that ψ1D and ψ2D are unitarily equivalent, and hence without loss of generality that ψ1D = ψ2D. Then we have a candidate for a semi-inverse to ϕ, namely ψ1 ∗D ψ2. We will show that after addition by additional extensions (some of which may be non-trivial), ϕ ⊕ (ψ1 ∗D ψ2) becomes asymptotically split. 2. step: (Removing a KK-obstruction.) First note that ϕ ⊕ (ψ1 ∗D ψ2) is split over D. Hence, by adding a copy of qB ◦ β to ϕ and conjugating by a unitary we can arrange that ϕ ⊕ (ψ1 ∗D ψ2) D = qB ◦ βD. (3.1) 10 VLADIMIR MANUILOV AND KLAUS THOMSEN Let ξi qB ◦ ξi : Ai → M(B), be equi-continuous asymptotic homomorphisms such that t = ϕAi ⊕ ψi for all t, i = 1, 2. Note that by (3.1) we have that ξi t(d) − β(d) ∈ B (3.2) for all t ∈ [1, ∞), d ∈ D, i = 1, 2. Let β∞ denote the direct sum of a count- able infinite number of copies of β and set π = 1C0[1,∞) ⊗ β∞; i.e. 1C0[1,∞) is the unit in the multiplier algebra M (C0[1, ∞)) and π(x) = 1C0[1,∞) ⊗ β∞(x) ∈ M (C0[1, ∞), B). Then π : D → M (C0[1, ∞), B) is absorbing by Lemma 2.3 of [Th3]. Since Ext (D, C0[1, ∞), B)) is the trivial group by assumption 2), this im- plies that there is a strictly continuous path Ut, t ∈ [1, ∞), of unitaries in M(B) such that t 7→ Ut(cid:0)ξ1 t (d) ⊕ β∞(d)(cid:1) U ∗ t −(cid:0)ξ2 t (d) ⊕ β∞(d)(cid:1) (3.3) is in C0[1, ∞), B) for all d ∈ D. For each n ∈ N, Ut, t ∈ [1, n], defines a unitary Wn in M (C[1, n] ⊗ B)) in the natural way. Set πn = 1C[1,n] ⊗ β∞D and βn = 1C[1,n] ⊗ βD. Then (3.3) and (3.2) imply that Wn (βn ⊕ πn) (d)W ∗ n − (βn ⊕ πn) (d) ∈ C[1, n] ⊗ B (3.4) for all d ∈ D, i.e. Wn is a unitary in the C ∗-algebra Dβn⊕πn(D), cf. (2.5). Note that βn ⊕ πn is absorbing, again by Lemma 2.3 of [Th3], so that K1 (Dβn⊕πn(D)) = KK(D, C[1, n]⊗B) by (3.2) of [Th3]. Identifying KK(D, C[1, n]⊗B) and KK(D, B) we can say that [Ad Wn ◦ (βn ⊕ πn) , (βn ⊕ πn)] = [Ad U1 ◦ (βD ⊕ β∞D) , (βD ⊕ β∞D)] . (3.5) in KK(D, C[1, n] ⊗ B). Add then the extension (qB ◦ Ad U1 ◦ (β ⊕ β∞)A1) ∗D (qB ◦ (β ⊕ β∞)A2) to ϕ ⊕ (ψ1 ∗D ψ2). We can then exchange ξ1 ξ2 t ⊕ (β ⊕ β∞) A2, and Ut by Ut ⊕ U ∗ notation and conclude from (3.5) that t by ξ1 t by 1 . We may therefore return to the previous t ⊕ Ad U1 ◦ (β ⊕ β∞) A1, ξ2 [Ad Wn ◦ (βn ⊕ πn) , (βn ⊕ πn)] = 0 in KK(D, C[1, n] ⊗ B) for all n. It follows therefore that diag(Wn, 1, 1, . . . , 1) is in the connected component of 1 in the unitary group of Mkn (Dβn⊕πn(D)) for some kn ∈ N, kn ≥ 2. Since βn ⊕ πn is absorbing, there is an isomorphism from Mkn (Dβn⊕πn(D)) onto M2 (Dβn⊕πn(D)) which takes diag(Wn, 1, 1, . . . , 1) to diag(Wn, 1). It follows that diag(Wn, 1) is in the connected component of 1 in the unitary group of M2 (Dβn⊕πn(D)) for each n. After addition by the split extension β∞ so that we can substitute Wn ⊕ 1 for Wn, we may therefore assume that Wn is in the connected component of 1 in the unitary group of Dβn⊕πn(D) for each n ∈ N. step: (The tricky part. This is an elaboration on ideas developed by Lin, Dadarlat and Eilers, in [L], [DE], and a very similar argument was used to prove Theorem 4.1 in [Th3].) 3. Let En denote the C ∗-subalgebra of M(C[1, n]⊗B)) generated by the unit 1C[0,1]⊗B, C[1, n] ⊗ B and (βn ⊕ πn) (D). It follows from (3.4) that Ad Wn defines an automor- phism αn of En, and the path of unitaries in Dβn⊕πn(D) connecting Wn to 1 gives us a uniform norm-continuous path of automorphisms in Aut En connecting αn to the identity in Aut En. Since En is separable, it follows from 8.7.8 and 8.6.12 in [P], cf. Proposition 2.15 of [DE], that αn is asymptotically inner, i.e. there is a continuous EXTENSIONS OF C ∗-ALGEBRAS 11 s , s ∈ [1, ∞), of unitaries in En such that αn(x) = lims→∞ V n s xV n s path V n x ∈ En. ∗ for all Let F1 ⊆ F2 ⊆ F3 ⊆ · · · be a sequence of finite subsets with dense union in D. Since lim s→∞ sup t∈[1,n] kV n s (t)(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)V n s (t)∗ − Ut(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)Ut ∗k = 0 for all d ∈ D, we can find an sn ∈ [1, ∞) so big that (3.6) (3.7) (k)∗V k s (k). (3.8) for all s ≥ sn, all t ∈ [1, n] and all d ∈ Fn. Note that kV n s (t)(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)V n (n)∗V n V n+1 s lim s→∞ s (t)∗ − Ut(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)Ut ∗k ≤ 1 n s (n)xV n s (n)∗V n+1 s (n) = x for all x ∈ B∪(ξ1 It follows from (3.7) that if we increase sn we can arrange that t ⊕ β∞) (D), t ∈ [1, n]. To simplify notation, set ∆k s = V k+1 s k∆k s (cid:0)ξ1 t ⊕ β∞D(cid:1) (d)∆k s ∗ −(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)k ≤ 1 n2 for all d ∈ Fn, t ∈ [1, n], and all k = 2, 3, · · · , n, when s ≥ sn. Proceeding inductively we can arrange that sn < sn+1 for all n. Let s : [1, ∞) → [1, ∞) be a continuous increasing function such that s(n) = sn+1, n = 1, 2, 3, · · · . Define a norm-continuous path Wt, t ∈ [1, ∞), in E = C ∗(cid:0)1B,(cid:0)ξ1 1 ⊕ β∞D(cid:1) (D), B(cid:1) = C ∗ (1B, (β ⊕ β∞D) (D), B) such that Wt = V 2 s(t), t ∈ [k, k + 1], k ≥ 2. Let d ∈ Fn and consider t ∈ [k, k + 1], where k ≥ n. Since s(t) ≥ sk+1 and d ∈ Fk+1, it follows from (3.8) that s(t)(t), t ∈ [1, 2], and Wt = V k+1 s(t) (t)∆k s(t) · · · ∆3 s(t)∆2 Wt(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)W ∗ t ∼k· 1 k2 V k+1 s(t) (t)(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)V k+1 s(t) (t)∗, (3.9) where ∼δ means that the distance between the two elements is at most δ. Further- more, it follows from (3.6) that It follows from (3.10), (3.9) and (3.3) that V k+1 s(t) (t)(cid:0)ξ1 s(t) (t)∗ ∼ 1 t ⊕ β∞D(cid:1) (d)V k+1 Wt(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)W ∗ k lim t→∞ t −(cid:0)ξ2 t . Ut(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)U ∗ t ⊕ β∞D(cid:1) (d) = 0, (3.10) (3.11) first when d ∈ Fn, and then for all d ∈ D since n was arbitrary and {ξi continuous. t}i,t equi- Recall that D is unital. For each t there are unique elements xt ∈ D, λt ∈ C and bt ∈ B such that Wt = (cid:0)ξ1 t ⊕ β∞D(cid:1) (xt) + λt(cid:0)ξ1 t ⊕ β∞D(cid:1) (1) ⊥ + bt. Since qB ◦ (ξ1 continuous path of unitaries in D such that limt→∞ xtdx∗ t ⊕ β∞D) = qB ◦ (ξ1 1 ⊕ β∞D) is injective we find that {xt} must be a t = d for all d ∈ D. Set Then Ut, t ∈ [1, ∞), is a continuous path of unitaries 1 + B such that Ut = Wt(cid:0)ξ1 Ut(cid:0)ξ1 t ⊕ β∞D(cid:1) (xt)∗ + Wtλt(cid:0)ξ1 t ⊕ β∞D(cid:1) (d)U ∗ t ⊕ β∞D(cid:1) (1)⊥. t ⊕ β∞D(cid:1) (d) = 0 t −(cid:0)ξ2 lim t→∞ 12 VLADIMIR MANUILOV AND KLAUS THOMSEN for all d ∈ D. 4. step: (Conclusion.) By adding the split extension qB ◦ β∞ we can now return to the notation in the 1. step and assume that Ut, t ∈ [1, ∞), is a continuous path of unitaries 1 + B such that for all d ∈ D. Set lim t→∞ Utξ1 t (d)U ∗ t − ξ2 t (d) = 0 (3.12) A = {f ∈ Cb ([1, ∞), M(B)) : f (1) − f (t) ∈ B ∀t ∈ [1, ∞)} and note that C0 ([1, ∞), B) is an ideal in A. Let p : A → A/C0 ([1, ∞), B) be the quotient map. Define ∗-homomorphisms κ1 : A1 → A and κ2 : A2 → A such that κ1(a)(t) = Utξ1 t − ξ2 t (d) ∈ D for all t and d ∈ D, it follows from (3.12) that t (a), respectively. Since Utξ1 t and κ2(a)(t) = ξ2 t (a)U ∗ t (d)U ∗ (p ◦ κ1) ∗D (p ◦ κ2) : A1 ∗D A2 → A/C0 ([1, ∞), B) is defined. By composing this ∗-homomorphism with a continuous right-inverse for p, whose existence follows from the Bartle-Graves selection theorem, we get an asymptotic homomorphism Φ : A1 ∗D A2 → M(B) such that qB ◦Φt = ϕ⊕(ψ1 ∗D ψ2) for all t. (cid:3) Corollary 3.4. Let A1, A2 and B be separable C ∗-algebras, B stable. Let D be a common C ∗-subalgebra of A1 and A2, i.e. D ⊆ A1 and D ⊆ A2. Assume that 1) D is nuclear, and 2) that all extensions of A1 by B and all extensions of A2 by B are semi- invertible. It follows that all extensions of A1 ∗D A2 by B are semi-invertible. Proof. It is well-known that condition 2) of Theorem 3.3 is fullfilled when D is nuclear. That condition 1) also holds follows from Lemma 2.2 of [Th2]. (cid:3) One important virtue of Theorem 3.3 and Corollary 3.4 when compared with Theorem 3.1 is the improved symmetry between assumptions and conclusions which allows to use it iteratively, for example to reach the following conclusion: Let A1, A2, A3, A4 be separable C ∗-algebras, D1 ⊆ A1, D1 ⊆ A2, and D2 ⊆ A3, D2 ⊆ A4 common C ∗-algebras. Assume that the Ai's and Di's are all nuclear, and let E be a common nuclear C ∗-subalgebra of A1 ∗D1 A2 and A3 ∗D2 A4. It follows that all extensions of (A1 ∗D1 A2) ∗E (A3 ∗D2 A4) by a separable stable C ∗-algebra B are semi-invertible. 4. Full group C ∗-algebras In this section we collect some consequences of Theorem 3.1 and Theorem 3.3 for the semi-invertibility of extensions by full group C ∗-algebras. Proposition 4.1. Let G1, G2 be countable discrete groups and H ⊆ Gi, i = 1, 2, a common subgroup. Set G = G1 ∗H G2 and let B be a separable stable C ∗-algebra. Assume that Ext(C ∗(Gi), B), i = 1, 2, are both groups. It follows that every extension of C ∗(G) by B is strongly homotopy invertible. EXTENSIONS OF C ∗-ALGEBRAS 13 Proof. We can apply Theorem 3.1 because C ∗(G) = C ∗(G1) ∗C ∗(H) C ∗(G2). In- deed, there are canonical conditional expectations C ∗(G) → C ∗(H) and C ∗(G) → C ∗(Gi), i = 1, 2, so any absorbing ∗-homomorpshism α0 : C ∗(G) → M(B), whose existence is guaranteed by [Th1], will meet the requirements in 1) of Theorem 3.1 by Lemma 2.1 of [Th2]. The conclusion of the corollary follows therefore from Theorem 3.1. (cid:3) Similarly, Theorem 3.3 implies the following Proposition 4.2. Let Gi, i = 1, 2, be discrete countable groups with a common subgroup H ⊆ Gi, i = 1, 2, and B a separable stable C ∗-algebra. Let G1 ∗H G2 be the amalgamated free product group and let B be a separable stable C ∗-algebra. Assume that 1) Ext(C ∗(H), B) and Ext (C ∗(H), C0[1, ∞) ⊗ B) are both group, and 2) for both i = 1 and i = 2 every extension of C ∗(Gi) by B is semi-invertible. It follows that every extension of C ∗(G1 ∗H G2) by B is semi-invertible. As is wellknown, condition 1) in Proposition 4.2 is satisfied when H is amenable, but it is also satisfied for certain non-amenable groups, e.g. free groups or an amal- gamated free product of amenable groups over a finite subgroup. We shall finish this paper by showing that the conclusions of Propositions 4.1 and 4.2, and partly also the conclusion of Theorem 2.3, are preserved by taking the product of the group with a group of the form Zk ⊕ H, with H finite. Lemma 4.3. Let A and B be separable C ∗-algebras, B stable. There are semi-group homomorphisms µ : Ext(A, B) → Ext(A⊗K, B) and ν : Ext(A⊗K, B) → Ext(A, B) such that µ ◦ ν(x) ⊕ 0 = x ⊕ 0 for all x ∈ Ext(A ⊗ K, B) and ν ◦ µ(y) ⊕ 0 = y ⊕ 0 for all Ext(A, B). Proof. Since B is stable we can identify B and K ⊗ B. Let e be a minimal projection in K and let V ∈ M(K ⊗ K ⊗ B) be an isometry such that V V ∗ = e ⊗ 1K⊗B. Then α(x) = V ∗(e ⊗ x)V is an isomorphism α : K ⊗ B → K ⊗ K ⊗ B, giving us isomorphisms M (K ⊗ B) → M (K ⊗ K ⊗ B) and Q (K ⊗ B) → Q (K ⊗ K ⊗ B) which we also denote by α. Let s : A → K⊗A be the ∗-homomorphism s(a) = e⊗a. We can then define a map Ext (K ⊗ A, K ⊗ K ⊗ B) → Ext (A, K ⊗ B) (4.1) by ϕ 7→ α−1 ◦ ϕ ⊗ s. To get a map in the other direction note that the canonical embedding K ⊗ M(K ⊗ B) ⊆ M(K ⊗ K ⊗ B) induce a ∗-homomorphism L : K ⊗ Q(K ⊗ B) → Q(K ⊗ K ⊗ B) which we can use to define a map (4.2) by ϕ 7→ L ◦ (idK ⊗ϕ). Then α−1 ◦ (L ◦ (idK ⊗ϕ)) ◦ s = Ad qK⊗B(W ) ◦ ϕ for some isometry W ∈ M(K ⊗ B), showing that Ext (A, K ⊗ B) → Ext (K ⊗ A, K ⊗ K ⊗ B) in Ext (A, K ⊗ B). (cid:2)(cid:0)α−1 ◦ (L ◦ (idK ⊗ϕ)) ◦ s(cid:1) ⊕ 0(cid:3) = [ϕ ⊕ 0] Consider next an extension ϕ : K ⊗ A → Q (K ⊗ K ⊗ B). Note that L ◦(cid:0)idK ⊗(cid:0)α−1 ◦ ϕ ◦ s(cid:1)(cid:1) (k ⊗ a) = L(cid:0)k ⊗ α−1 (ϕ(e ⊗ a))(cid:1) 14 VLADIMIR MANUILOV AND KLAUS THOMSEN on simple tensors, k ∈ K, a ∈ A. Since the automorphism of Q(K ⊗ K ⊗ A) which interchange the two copies of K is given by a unitary in M (K ⊗ K ⊗ B), the exten- sion L ◦ (idK ⊗ (α−1 ◦ ϕ ◦ s)) is unitarily equivalent to an extension ψ : K ⊗ A → Q(K ⊗ K ⊗ B) such that ψ(k ⊗ a) = L(cid:0)e ⊗ α−1 (ϕ(k ⊗ a))(cid:1) on simple tensors. Since L (e ⊗ α−1 (ϕ(k ⊗ a))) = Ad qK⊗K⊗B(V ) (ϕ(k ⊗ a)), we see that the two maps, (4.1) and (4.2) are inverses of each other, up to addition by 0. Since both maps clearly are semi-group homomorphisms, the proof is complete. (cid:3) Corollary 4.4. Let A and B be separable C ∗-algebras, B stable. Then all extensions of A by B are semi-invertible or strongly homotopy invertible if and only if the same is true for all extensions of Mn(A) by B, for any n ∈ N. Lemma 4.5. Let A1, A2 and B be separable C ∗-algebras, B stable. Assume that all extensions of Ai by B are semi-invertible or are strongly homotopy invertible (with an invertible inverse), i = 1, 2. It follows that all extensions of A1 ⊕ A2 by B have the same property. Proof. Let pi : A1 ⊕ A2 → Ai ⊆ A1 ⊕ A2, i = 1, 2, be the canonical projections, and consider an extension ϕ : A1 ⊕ A2 → Q(B). By a standard rotation argument ϕ ⊕ 0 is strongly homotopic to the sum (ϕ ◦ p1) ⊕ (ϕ ◦ p2). The conclusion follows from this by use of Theorem 2.1. (cid:3) By combining Corollary 4.4 and Lemma 4.5 we get the following. Corollary 4.6. Let A, F and B be separable C ∗-algebras, B stable, F finite dimen- sional. Assume that all extensions of A by B are semi-invertible or are strongly homotopy invertible (with an invertible inverse). It follows that all extensions of F ⊗ A by B have the same property. In particular, it follows that if G is a countable discrete group with the property r (G) by B are semi-invertible or strongly homotopy invertible r (H × G) for any finite group that all extensions of C ∗ (with an invertible inverse), then the same is true for C ∗ H. Lemma 4.7. Let A and B be separable C ∗-algebras, B stable. Assume that all extensions of A by B are semi-invertible or strongly homotopy invertible. It follows that all extensions of C(T) ⊗ A by B have the same property. Proof. Let χ be the automorphism of C(T) ⊗ A such that χ(f )(z) = f (z) and let ev : C(T) ⊗ A → A be evaluation at 1 ∈ T. As is wellknown the ∗-homomorphism C(T) ⊗ A → M2 (C(T) ⊗ A) defined such that f 7→ (cid:16) f χ(f )(cid:17) is homotopic to a ∗-homomorphism which factorizes through ev. It follows that for any extension ϕ : C(T) ⊗ A → Q(B) the extension ϕ ⊕ ϕ ◦ χ is strongly homotopic to an extension of the form ψ ◦ ev, where ψ : A → Q(B) is an extension of A by B. By assumption there is an extension ψ′ of A by B such that ψ ⊕ ψ′ is either asymptotically split or strongly homotopic to a split extension. It follows that ϕ ⊕ ϕ ◦ χ ⊕ ψ′ ◦ ev has the same property by Theorem 2.1. Hence ϕ is semi-invertible or strongly homotopy invertible, as the case may be. (cid:3) EXTENSIONS OF C ∗-ALGEBRAS 15 Proposition 4.8. Let G be a countable discrete group, H a finite group and k ∈ N. Let B be a separable stable C ∗-algebra and assume that all extensions of C ∗ r (G), (resp. C ∗(G)), by B are semi-invertible or strongly homotopy invertible. It follows that all extensions of C ∗ same property. r (cid:0)Zk × H × G(cid:1), (resp. C ∗(cid:0)Zk × H × G(cid:1)), by B have the r (G), and that C ∗(H) It follows then from Corollary 4.6 and Lemma 4.7 that all r (cid:0)Zk × H × G(cid:1) by B are semi-invertible or strongly homotopy in- r (cid:0)Zk × H × G(cid:1) ≃ C(cid:0)Tk(cid:1) ⊗ C ∗(H) ⊗ C ∗ Proof. Note that C ∗ is finite dimensional. extensions of C ∗ vertible if C ∗ C ∗-algebra. r (G) has this property. The same argument works for the full group (cid:3) Finally, we observe that it is also possible to use Theorem 3.1 and Theorem 3.3 to prove semi-invertibility for extensions of the full group C ∗-algebra of certain HNN- extensions by using the realization obtained by Ueda in [U] of such group C ∗-algebras as amalgamated free products. References [A] J. Anderson, A C ∗-algebra for which Ext(A) is not a group, Ann. of Math. 107 (1978), 455 -- 458. [BDF1] L.G. Brown, R.G. Douglas and P.A. Fillmore, Unitary equivalence modulo the compact operators and extensions of C ∗-algebras, Proc. Conf. on Operator Theory, Lecture Notes in Mathematics 345, Springer Verlag (1973), 58 -- 128. [BDF1] [Br] [CCJJV] P-A. Cherix, M. Cowling, P. Jolissaint, P. Julg and A. Valette, Groups with the Haagerup , Extensions of C ∗-algebras and K-theory, Ann. of Math. 105 (1977), 265 -- 324. L. Brown, Ext of certain free product C ∗-algebras, J. Operator Th. 6 (1981), 135-141. [CH] [C] [DE] [HT] [HS] Property, Birkhauser Verlag, (2001). A. Connes and N. Higson, D´eformations, morphismes asymptotiques et K-th´eories bivari- ante, C.R. Acad. Sci. Paris S´er I Math. 311 (1990), 101-106. J. Cuntz, K-theoretic amenability for discrete groups, J. Reine u. Angew. Math. 344 (1983), 180-195. M. Dadarlat and S. Eilers, Asymptotic Unitary Equivalence in KK-theory, K-theory 23 (2001), 305-322. U. Haagerup and S. Thorbjørnsen, A new application of random matrices: Ext(C ∗ is not a group, Ann. of Math. 162 (2005), 711-775. D. Hadwin and J. Shen, Some examples of Blackadar and Kirchberg's MF algebras, Preprint, arXiv:0806.4712. red(F2)) [HLSW] D. Hadwin, J. Li, J. Shen, J. Wang, Reduced free products of unital AH algebras and [HK] [K] [L] [M] Blackadar and Kirchberg's MF algebras, Preprint, arXiv:0812.0189v1 N. Higson and G. Kasparov, E-theory and KK-theory for groups which act properly and isometrically on Hilbert space, Invent. Math. 144 (2001), 23-74. G. Kasparov, Equivariant KK-theory and the Novikov conjecture, Invent. Math. 91 (1988), 513-572. H. Lin, Stable approximate unitary equivalence of homomorphisms, J. Oper. Theory 47 (2002), 343-378. V. Manuilov, Asymptotic representations of the reduced C ∗-algebra of a free group: an example, Bull. London Math. Soc. 40 (2008), 838-844. [MT1] V. Manuilov and K. Thomsen, E-theory is a special case of KK-theory, Proc. London Math. Soc. 88 (2004), 455-478. [MT2] [MT3] [MT4] , Relative K-homology and normal operators, J. Operator Th. 62 (2009), 249-279. , The Connes-Higson construction is an isomorphism, J. Func. Anal. 213 (2004), 154-175. , On the lack of inverses to C ∗-extensions related to property T groups, Can. Math. Bull. 50 (2007), 268-283. 16 VLADIMIR MANUILOV AND KLAUS THOMSEN [P] [Se] [S] [ST] [Th1] [Th2] [Th3] [Th4] [Tu] [U] G. K. Pedersen, C ∗-algebras and their automorphisms group, Academic Press, New York, 1979. J.A. Seebach, On the Reduced Amalgamated Free Products of C ∗-algebras and the MF- Property, arXiv:1004.3721 J.-P. Serre, Trees, Springer Verlag, Berlin, 1977. J. A. Seebach and K. Thomsen, Extensions of the reduced group C ∗-algebra of a free product of amenable groups, Adv. Math. 223 (2010), 1845-1854. K. Thomsen,, On absorbing extensions, Proc. Amer. Math. Soc. 129 (2001), 1409-1417. , On the KK-theory and the E-theory of amalgamated free products of C ∗-algebras, J. Funct. Anal. 201 (2003), 30-56. , Homotopy invariance in E-theory, Homology, homotopy and applications 8 (2006), 29-49. , All extensions of C ∗ r (Fn) are semi-invertible, Math. Ann. 342 (2008), 273-277. J.L. Tu, La conjecture de Baum-Connes pour les feuilletages moyennables, K-theory 17 (1999), 215-264. Y. Ueda, Remarks on HNN extensions in operator algebras, Illinois J. Math. 52 (2008), 705-725. E-mail address: [email protected] Dept. of Mech. and Math., Moscow State University, Moscow, 119899, Russia Institut for matematiske fag, Ny Munkegade, 8000 Aarhus C, Denmark
1904.06816	1	1904	2019-04-15T02:51:09	II$_1$ factors with exotic central sequence algebras	[ "math.OA", "math.FA" ]	We provide a class of separable II$_1$ factors $M$ whose central sequence algebra is not the "tail" algebra associated to any decreasing sequence of von Neumann subalgebras of $M$. This settles a question of McDuff \cite{Mc69d}.	math.OA	math	II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS ADRIAN IOANA AND PIETER SPAAS Abstract. We provide a class of separable II1 factors M whose central sequence algebra is not the "tail" algebra associated to any decreasing sequence of von Neumann subalgebras of M . This settles a question of McDuff [Mc69d]. 1. Introduction and statement of main results A uniformly bounded sequence (xk) in a II1 factor M is called central if limk kxky − yxkk2 = 0, for every y ∈ M . Central sequences have played a fundamental role in the study of II1 factors since the very beginning of the subject with Murray and von Neumann's property Gamma [MvN43]. A separable II1 factor M has property Gamma if it admits a central sequence (xk) which is not trivial, in the sense that inf k kxk − τ (xk)1k2 > 0. Murray and von Neumann proved that the unique hyperfinite II1 factor has property Gamma, while the free group factor L(F2) does not, thus giving the first example of two non-isomorphic separable II1 factors [MvN43]. Over two decades later, in the late 60s, the analysis of central sequences of [MvN43] was refined to provide additional examples of non-isomorphic separable II1 factors in [Ch69, DL69, Sa68, ZM69], culminating with McDuff's construction of a continuum of such factors [Mc69a, Mc69b]. Shortly after, McDuff [Mc69c] defined the central sequence algebra of a II1 factor M as the relative commutant, M ′ ∩ M ω, of M into its ultrapower M ω ([Wr54, Sa62]), where ω is a free ultrafilter on N. This has since allowed for a more structural approach to central sequences and led to significant progress in the study of II1 factors. Indeed, the central sequence algebra was a crucial tool in Connes' famous classification of amenable II1 factors [Co76]. Furthermore, the relative commutant M ′ ∩ Mω, for some von Neumann algebra M ⊃ M , was used by Popa to formalise his influential spectral gap rigidity principle in [Po06a, Po06b]. Most recently, central sequence algebras and their subalgebras were used to provide a continuum of II1 factors with non-isomorphic ultrapowers in [BCI15] (adding to the four such factors noticed in [FGL06, FHS11, GH16]). However, despite the progress the use of central sequence algebras has allowed, their structure remains fairly poorly understood. For instance, it is open whether any II1 factor M whose central sequence algebra is abelian admits an abelian subalgebra A such that M ′ ∩ M ω ⊂ Aω (see [Ma17]). In this article, we investigate the existence of a certain "canonical form" for central sequence algebras. To make this precise, we recall the following notions introduced by McDuff in [Mc69d] in order to distil the key ideas of [Mc69b]: Definition 1.1 ([Mc69d, Definition 2]). Let M be a separable II1 factor. A von Neumann subal- gebra A of M is called residual if limk kxk − EA(xk)k2 = 0, for every central sequence (xk) in M . A sequence (An)n∈N of von Neumann subalgebras of M is called a residual sequence if (1) An+1 ⊂ An, for every n, (2) An is residual in M , for every n, and (3) if xk ∈ Ak and kxkk ≤ 1, for every k, then the sequence (xk) is central in M . The authors were supported in part by NSF Career Grant DMS #1253402. 1 2 A. IOANA AND P. SPAAS Remark 1.2. A decreasing sequence (An)n∈N of von Neumann subalgebras of M is residual if and only if M ′ ∩ M ω = ∩n∈NAω n. Thus, a separable II1 factor M admits a residual sequence if and only if its central sequence algebra is equal to the "tail" algebra, ∩n∈NAω n, associated to a decreasing sequence of von Neumann subalgebras (An)n∈N. In [Mc69d], McDuff noted that it was unknown whether every II1 factor admits a residual sequence. She gave examples of II1 factors which do not admit any strongly residual sequence (An)n∈N (i.e. ones satisfying, in addition to (1)-(3), the existence of a subalgebra An ⊂ An such that An = An+1 ¯⊗An), but left open the case of residual sequences. The main goal of this article is to provide the first examples of II1 factors with no residual sequence. Before stating our results in this direction, let us note that several large, well-studied classes of II1 factors admit a residual sequence. Examples 1.3. The following II1 factors admit a residual sequence: (Rn)n∈N is a residual sequence in R. (1) Any II1 factor without property Gamma. (2) The hyperfinite II1 factor R. If we write R = ¯⊗k∈NM2(C), and let Rn = ¯⊗k≥nM2(C), then (3) Any II1 factor M which is strongly McDuff, i.e. can be written as M = N ¯⊗R, where N is a II1 factor without property Gamma. If An = 1 ⊗ Rn, then Connes' characterization of property Gamma [Co76, Theorem 2.1] implies that (An)n∈N is a residual sequence in M . (4) Any infinite tensor product M = ¯⊗k∈NMk of II1 factors without property Gamma. If An = ¯⊗k≥nMk, then [Co76, Theorem 2.1] implies that (An)n∈N is a residual sequence in M . Note that M is McDuff, i.e. M ∼= M ¯⊗R, but not strongly McDuff [Po09a, Theorem 4.1]. (5) The II1 factors L(T0(Γ)) and L(T1(Γ)), where Γ is any countable group and the countable groups T0(Γ), T1(Γ) are defined as in [DL69, Mc69b] (see also [BCI15, Section 1.1]). Then T0(Γ) and T1(Γ) both containeΓ := ⊕i∈NΓi, where each Γi is a copy of Γ. If An = L(⊕i≥nΓi), then [BCI15, Corollary 2.11] shows that (An)n∈N is a residual sequence in both L(T0(Γ)) and L(T1(Γ)). In particular, the uncountably many II1 factors which were shown to have non-isomorphic ultrapowers in [BCI15] all admit residual sequences. (6) Any tensor product M = ¯⊗N admitting a residual sequence, (Ak,n)n∈N. If Bn = ( ¯⊗min{n,N } then [Ma17, Proposition 5.2] implies that (Bn)n∈N is a residual sequence in M . k=1Mk, where N ∈ N∪{∞}, and for every k, Mk is a II1 factor k=min{n,N }+1Mk), Ak,n) ¯⊗( ¯⊗N k=1 Remark 1.4. In [Po09a,Po09b], Popa studied the class of II1 factors M which arise as an inductive limit of subfactors (Mn) with spectral gap and noticed that M ′ ∩ M ω = ∩n(M ′ n ∩ M )ω (see [Po09a, Lemma 2.3]). Thus, every such II1 factor M admits a residual sequence, (M ′ n ∩ M )n∈N. Conversely, although it is unclear whether any II1 factor admitting a residual sequence must be an inductive limit of subfactors with spectral gap, we note that this holds for the factors in Examples 1.3 (1)-(5). We are now ready to state our first main result which gives examples of II1 factors with no residual sequences, and thereby settles McDuff's question [Mc69d]. Theorem A. Let Γ be a countable non-amenable group. For every k ∈ N, let πk : Γ → O(Hk) be an orthogonal representation such that (1) π⊗l k (2) there is an orthonormal sequence (ξm is weakly contained in the left regular representation of Γ, for some l = l(k) ∈ N, and k k → 0, k )m∈N ⊂ Hk such that supm∈N kπk(g)(ξm k ) − ξm as k → ∞, for every g ∈ Γ. Let Γ y (Bk, τk) be the Gaussian action associated to πk, and Γ y (B, τ ) := ¯⊗k∈N(Bk, τk) be the diagonal product action. Define M = B ⋊ Γ. Then the II1 factor M does not admit a residual sequence of von Neumann subalgebras. II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 3 t For the definition of Gaussian actions, we refer the reader to Section 2.6. Next, we provide a class of examples to which Theorem A applies, and discuss a connection with a problem posed in [JS85]. Example 1.5. Let Γ = Fn be the free group on n ≥ 2 generators. Denote by g the word length of an element g ∈ Γ with respect to a free set of generators. Let t > 0. By [Ha79], the function ϕt : Γ → R given by ϕt(g) = e−tg is positive definite. Let ρt : Γ → O(Ht) be the GNS orthogonal representation associated to ϕt and ξt ∈ Ht such that hρt(g)(ξt), ξti = ϕt(g), for all g ∈ Γ. Let ρt = ρt ⊗ Idℓ2(N) : Γ → O(Ht ⊗ ℓ2(N)) be the direct sum of infinitely many copies of ρt. Let (tk) be any sequence of positive numbers converging to 0 and put πk := ρtk : Γ → O(Htk⊗ℓ2(N)). Then the representations (πk)k∈N satisfy the hypothesis of Theorem A. Firstly, given t > 0, note that t ∈ ℓ2(Γ), and hence ρ⊗l ϕl is contained in a multiple of the left regular representation of Γ, whenever l > log(2n−1)/(2t). This implies that π⊗l k is contained in a multiple of the left regular representation k := ξtk ⊗ δm ∈ Htk ⊗ ℓ2(N) of Γ, for some integer l = l(k) ≥ 1. Secondly, note that the vectors ξm satisfy supm∈N kπk(g)(ξm Remark 1.6. Theorem A also sheds new light on a problem of Jones and Schmidt. In [JS85, Theorem 2.1], they proved that any ergodic but not strongly ergodic countable measure preserving equivalence relation R on a probability space (X, µ) admits a hyperfinite quotient. More specifically, there exists an ergodic hyperfinite measure preserving equivalence relation Rhyp on a probability space (Y, ν) together with a factor map π : (X, µ) → (Y, ν) such that (π × π)(R) = Rhyp, almost In [JS85, Problem 4.3], Jones and Schmidt asked whether there is always such a everywhere. quotient with the additional property that R0 := {(x1, x2) ∈ R π(x1) = π(x2)} is strongly ergodic on almost all of its ergodic components. If such a quotient exists, then following [IS18, Definition 1.3] we say that R has the Jones-Schmidt property. If R has the Jones-Schmidt property and we let M = L(R), A = L∞(X), then there exists a decreasing sequence of von Neumann subalgebras (Bn)n∈N of A such that M ′ ∩ Aω = ∩nBω n and Bn+1 ⊂ Bn has finite index for every n ∈ N (see [IS18, Proposition 5.3 and the proof of Lemma 6.1]). k k =p2(1 − ϕtk (g)) → 0, as k → ∞, for any g ∈ Γ. k ) − ξm In [IS18, Theorems E and F], the authors settled in the negative [JS85, Problem 4.3] by providing examples of equivalence relations R without the Jones-Schmidt property. This was achieved by showing that for certain R, in the above notation, M ′∩ Aω is not equal to ∩nBω n , for any decreasing sequence of von Neumann subalgebras (Bn)n∈N of A with Bn+1 ⊂ Bn of finite index for every n ∈ N. Theorem A allows us to strengthen the negative solution to [JS85, Problem 4.3] given in [IS18]. More precisely, in the context of Theorem A, assume that Γ is not inner amenable and let R be the equivalence relation associated to the action Γ y B. Since M = L(R) = B ⋊ Γ has no residual sequence by Theorem A, while M ′ ∩ Aω = M ′ ∩ M ω by [Ch82], we deduce that M ′ ∩ Aω cannot be written as ∩nBω Our second main result shows that the conclusion of Theorem A also holds if we replace Gaussian by free Bogoljubov actions (see Section 2.6). Moreover, we establish the following stronger statement: Theorem B. Let Γ be a countable non-inner amenable group. For every k ∈ N, let πk : Γ → O(Hk) be an orthogonal representation such that n , for any decreasing sequence (Bn)n∈N of von Neumann subalgebras of A. (1) π⊗l k (2) there are orthogonal unit vectors ξ1 is weakly contained in the left regular representation of Γ, for some l = l(k) ∈ N, and k k → 0, k ∈ Hk such that maxm∈{1,2} kπk(g)(ξm k ) − ξm as k → ∞, for every g ∈ Γ. k, ξ2 Let Γ y (Bk, τk) be the free Bogoljubov action associated to πk, and Γ y (B, τ ) := ¯⊗k∈N(Bk, τk) be the diagonal product action. Define M = B ⋊ Γ. Then the II1 factor M does not admit a residual sequence of von Neumann subalgebras. 4 A. IOANA AND P. SPAAS Moreover, there exists a separable von Neumann subalgebra P ⊂ M ′ ∩ M ω such that there is no sequence (An)n∈N of von Neumann subalgebras of M satisfying P ⊂Qω An ⊂ M ′ ∩ M ω. Since Γ = Fn is not inner amenable for any n ≥ 2, and the representations (πk)k∈N from Example 1.5 satisfy the hypothesis of Theorem B, its conclusion holds for those examples. Moreover, in the notation from Example 1.5, πk = ρtk ⊕ ρtk also satisfy the hypothesis of Theorem B. In order to put Theorem B into a better perspective and to contrast it with Theorem A, we note the following result: Proposition C. Let (Mn, τn), n ∈ N, be a sequence of tracial von Neumann algebras. Let P , Q be commuting separable von Neumann subalgebras of Qω Mn. Assume that P is amenable. Then there exist commuting von Neumann subalgebras Pn, Qn of Mn, for every n ∈ N, such that P ⊂Qω Pn and Q ⊂Qω Qn. Proposition C implies that for any tracial von Neumann algebra (M, τ ) and any separable amenable von Neumann subalgebra P ⊂ M ′ ∩ M ω, there is a sequence (Pn)n∈N of von Neumann subalgebras n ∩ M ), and therefore P ⊂ Qn Pn ⊂ M ′ ∩ M ω. of M such that P ⊂ Qω Pn and M ⊂ Qω(P ′ Consequently, the moreover part of Theorem B cannot hold if P is amenable. In particular, if M = B ⋊ Γ is as in Theorem A and Γ is not inner amenable, then M will not satisfy the moreover assertion of Theorem B. Indeed, in this case M ′∩M ω is abelian, being a subalgebra of Bω by [Ch82]. In recent years there has been growing interest in the study of the notion of stability for groups (see the survey [Th18]). As a byproduct of the methods developed in this article, we obtain two applications to the notion of tracial stability for countable groups, formalised recently in [HS17] (see also [HS16]): Definition 1.7 ([HS17, Definition 3]). A countable group Γ is W ∗-tracially stable if for any sequence (Mn, τn), n ∈ N, of tracial von Neumann algebras and any homomorphism ϕ : Γ → U (Qω Mn), there exist homomorphisms ϕn : Γ → U (Mn), for every n ∈ N, such that ϕ = (ϕn)n. The class of W∗-tracially stable groups contains all abelian and free groups, as well as other classes of both amenable and non-amenable groups, see [HS17]. As an immediate consequence of Proposition C, we deduce that the class of W∗-tracially stable groups is closed under taking the direct product with an amenable group. For the case of the direct product with an abelian group, this result is part of [HS17, Theorem 1]. Corollary D. Let Γ and Σ be W ∗-tracially stable groups. Assume that Σ is amenable. Then Γ× Σ is W ∗-tracially stable. In contrast to Corollary D, we show that any direct product of non-abelian free groups is not W∗- tracially stable, thereby answering a question of Atkinson in the negative (see [At18, Question 4.16]). Theorem E. Fl × Fm is not W ∗-tracially stable, for any 2 ≤ l, m ≤ +∞. Moreover, there exist a II1 factor M and a trace preserving ∗-homomorphism ϕ : L(F2× F2) → M ω such that there is no sequence of homomorphisms ϕn : F2× F2 → U (M ) satisfying ϕF2×F2 = (ϕn)n. Structure of the paper. Besides the introduction there are four other sections in this paper. In Section 2 we recall some preliminaries and prove a few useful lemmas needed in the remainder of the paper. In Section 3, inspired by Boutonnet's work [Bo12, Bo14], we prove a structural result concerning II1 factors associated to Gaussian and free Bogoljubov actions. In Section 4 this is used to prove Theorems A and B. Finally in Section 5 we prove Proposition C and use the established machinery from the previous sections to deduce Theorem E. II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 5 Acknowledgements. We are very grateful to Scott Atkinson for bringing [At18, Question 4.16] to our attention and for stimulating discussions on tracial stability. 2.1. Tracial von Neumann algebras. We begin this section by recalling several notions and constructions involving tracial von Neumann algebras. 2. Preliminaries A tracial von Neumann algebra (M, τ ) is a von Neumann algebra M equipped with a faithful normal tracial state τ : M → C. We denote by L2(M ) the completion of M with respect to the 2-norm kxk2 = pτ (x∗x) and consider the standard representation M ⊂ B(L2(M )). We also denote by U (M ) the group of unitary elements of M , by (M )1 = {x ∈ M kxk ≤ 1} the unit ball of M , and by Z(M ) = M ∩ M ′ the center of M . It follows from von Neumann's bicommutant theorem, that a self-adjoint set S ⊂ M generates M as a von Neumann algebra if and only if S′′ = M . Let P ⊂ M be a unital von Neumann subalgebra. Jones' basic construction of the inclusion P ⊂ M is defined as the von Neumann subalgebra of B(L2(M )) generated by M and the orthogonal projection eP : L2(M ) → L2(P ), and is denoted by hM, ePi. The basic construction hM, ePi carries a canonical semi-finite trace τ defined by τ (xeP y) = τ (xy), for all x, y ∈ M . We further denote by EP : M → P the conditional expectation onto P , by P ′ ∩ M = {x ∈ M xy = yx, for all y ∈ P} the relative commutant of P in M , and by NM (P ) = {u ∈ U (M ) uP u∗ = P} the normalizer of P in M . We say that P is regular in M if NM (P ) generates M as a von Neumann algebra. Any trace preserving action Γ yσ (M, τ ) extends to a unitary representation σ : Γ → U (L2(M )) called the Koopman representation of σ. Let ω be a free ultrafilter on N. Consider the C∗-algebra ℓ∞(N, M ) = {(xn) ∈ M N supkxnk < ∞} n→ω kxnk2 = 0}. Then M ω := ℓ∞(N, M )/I together with its closed ideal I = {(xn) ∈ ℓ∞(N, M ) lim is a tracial von Neumann algebra, called the ultrapower of M , whose canonical trace is given by τ (xn), for all x = (xn) ∈ M ω. If (Mn)n is a sequence of von Neumann subalgebras of τω(x) = lim n→ω M , then their ultraproduct, denoted by Qω Mn, can be realized as the von Neumann subalgebra of Lemma 2.1. Let (M, τ ) be a tracial von Neumann algebra and (An)n be a sequence of von Neumann subalgebras of M such that Qω An ⊂ M ′ ∩ M ω. Then lim n∩M (x)k2 = 0, for every x ∈ M . Proof. Let x ∈ M . If n ∈ N, we can find un ∈ U (An) such that kx − unxu∗ n∩M (x)k2 (see, e.g., the proof of [IS18, Theorem 2.5]). Since (un) ∈ Qω An and Qω An ⊂ M ′ ∩ M ω, we get that limn→ω kx − unxu∗ 2.2. Hilbert bimodules. Let (M1, τ1) and (M2, τ2) be two tracial von Neumann algebras. An M1-M2-bimodule is a Hilbert space H endowed with two normal, commuting ∗-homomorphisms π1 : M1 → B(H) and π2 : M op 2 → B(H) by πH(x⊗ yop) = π1(x)π2(yop) and write xξy = π1(x)π2(yop)ξ, for all x ∈ M1, y ∈ M2 and ξ ∈ H. We also write M1HM2 to indicate that H is an M1-M2-bimodule. Examples of bimodules include the trivial M1-bimodule M1L2(M1)M1 and the coarse M1-M2-bimodule M1L2(M1) ⊗ L2(M2)M2. Next, we recall a few notions and constructions involving bimodules (see [Co94, Appendix B] and [Po86]). If H and K are M1-M2-bimodules, we say that H is weakly contained in K and write H ⊂weak K if kπH(T )k ≤ kπK(T )k, for all T ∈ M1 ⊗ M op 2 . If H is an M1-M2-bimodule and K is an M2-M3-bimodule, then the Connes fusion tensor product of H and K is an M1-M3-bimodule denoted by H ⊗M2 K. If Φ : M1 → M2 is a unital normal completely positive map, then there 2 → B(H). We define a ∗-homomorphism πH : M1 ⊗ M op M ω consisting of x = (xn) such that lim n→ω kxn − EMn(xn)k2 = 0. nk2 = 0 and hence limn→ω kx − EA′ n∩M (x)k2 = 0. (cid:4) n→ω kx− EA′ nk2 ≥ kx − EA′ 6 A. IOANA AND P. SPAAS is a unique M1-M2-bimodule, denoted by HΦ, with a unit vector ξΦ ∈ HΦ such that M1ξΦM2 is dense in HΦ and hxξΦy, ξΦi = τ2(Φ(x)y), for all x ∈ M1 and y ∈ M2. The next result analyzes the Connes fusion tensor product of bimodules associated to completely positive maps: Lemma 2.2. Let Φ : M1 → M2 and Ψ : M2 → M3 be unital normal completely positive maps, where (M1, τ1), (M2, τ2), (M3, τ3) are tracial von Neumann algebras. Then the following hold: (1) The M1-M3-bimodule HΨ◦Ad(u)◦Φ is isomorphic to a sub-bimodule of HΦ ⊗M2 HΨ, for every (2) If U is a set of unitaries in M2 whose span is k.k2-dense in M2, then the M1-M3-bimodule u ∈ U (M2). HΦ ⊗M2 HΨ is isomorphic to a sub-bimodule of ⊕u∈UHΨ◦Ad(u)◦Φ. Proof. For u ∈ U (M2), we denote ηu := ξΦu∗ ⊗M2 ξΨ ∈ HΦ ⊗M2 HΨ. Following [Po86, Section 1.3.1], for every x ∈ M1, y ∈ M3, we have that hxηuy, ηui = hxξΦu∗ ⊗M2 ξΨy, ξΦu∗ ⊗M2 ξΨi = hxξΦu∗p, ξΦu∗i = τ2(Φ(x)u∗pu), where p ∈ M2 is such that τ2(zp) = hzξΨy, ξΨi = τ3(Ψ(z)y), for all z ∈ M2. Thus, for all x ∈ M1, y ∈ M3 we have that hxηuy, ηui = τ2(uΦ(x)u∗p) = τ3(Ψ(uΦ(x)u∗)y). This shows that the M1-M3-bimodule M1ηuM3 is isomorphic to HΨ◦Ad(u)◦Φ and proves the first assertion of the lemma. Finally, note that if the span of U ⊂ U (M2) is k.k2-dense in M2, then the span of {M1ηuM3 u ∈ U} is dense in HΦ ⊗M2 HΨ. This implies the second assertion. (cid:4) 2.3. Intertwining-by-bimodules. We next recall from [Po03, Theorem 2.1 and Corollary 2.3] the powerful intertwining-by-bimodules technique of Popa. Theorem 2.3 ([Po03]). Let (M, τ ) be a tracial von Neumann algebra and P ⊂ pM p, Q ⊂ qM q be unital von Neumann subalgebras, for some projections p, q ∈ M . Then the following conditions are equivalent: • There exist projections p0 ∈ P, q0 ∈ Q, a ∗-homomorphism θ : p0P p0 → q0Qq0 and a non-zero partial isometry v ∈ q0M p0 such that θ(x)v = vx, for all x ∈ p0P p0. • There is no net un ∈ U (P ) satisfying kEQ(x∗uny)k2 → 0, for all x, y ∈ pM q. • There exists a non-zero projection f ∈ P ′ ∩ hM, eQi with τ (f ) < ∞. If one of these conditions holds true, then we write P ≺M Q, and say that a corner of P embeds into Q inside M . If P p′ ≺M Q for any non-zero projection p′ ∈ P ′ ∩ pM p, then we write P ≺s M Q. 2.4. Amenability. A tracial von Neumann algebra (M, τ ) is called amenable if there exists a positive linear functional ϕ : B(L2(M )) → C such that ϕM = τ and ϕ is M -central, in the following sense: ϕ(xT ) = ϕ(T x), for all x ∈ M and T ∈ B(L2(M )). Equivalently, (M, τ ) is amenable if M L2(M )M is weakly contained in M L2(M ) ⊗ L2(M )M . By Connes' celebrated classification of amenable factors [Co76], M is amenable if and only if it is approximately finite dimensional. Next, we recall the notion of relative amenability introduced by Ozawa and Popa. Let p ∈ M be a projection, and P ⊂ pM p, Q ⊂ M be von Neumann subalgebras. Following [OP07, Section 2.2] we say that P is amenable relative to Q inside M if there exists a positive linear functional ϕ : phM, eQip → C such that ϕpM p = τ and ϕ is P -central. As shown in [DHI16, Lemma 2.7], relative amenability is closed under inductive limits. Here we establish the following generalization of this result, which we will need later on. Given a set I, we denote by limn a state on ℓ∞(I) which extends the usual limit. II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 7 Lemma 2.4. Let (M, τ ) be a tracial von Neumann algebra and P, Q ⊂ M be von Neumann subalgebras. Assume that Pn ⊂ M , n ∈ I, is a net of von Neumann subalgebras such that kEPn(x) − xk2 → 0, for all x ∈ P , and pn ∈ P ′ n ∩ M are projections such that Pnpn is amenable relative to Q inside M , for every n ∈ I. Then there exists a projection p ∈ P ′ ∩ M such that P p is amenable relative to Q inside M and τ (p) ≥ limn τ (pn). Proof. We may clearly assume that c := limn τ (pn) > 0 and τ (pn) > 0, for every n. For every n, let ϕn : pnhM, eQipn → C be a Pnpn-central positive linear functional such that ϕnpnM pn = τ . The Cauchy-Schwarz inequality implies that ϕn(pnT xpn) ≤ pϕn(pnT T ∗pn)ϕn(pnx∗xpn) ≤ kTkkxk2, and similarly that ϕn(pnxT pn) ≤ kTkkxk2, for all x ∈ M , T ∈ hM, eQi. We define a state ϕ : hM, eQi → C by letting ϕ(T ) = lim n ϕn(pnT pn) τ (pn) , for every T ∈ hM, eQi. We claim that ϕ is P -central. To this end, let x ∈ P , T ∈ hM, eQi and n ∈ I. Since ϕn is Pnpn-central, ϕn(pnT EPn(x)pn) = ϕn(pnEPn(x)T pn) and thus ϕn(pnT xpn) − ϕn(pnxT pn) ≤ ϕn(pnT (x − EPn(x))pn) + ϕn(pn(x − EPn(x))T pn) ≤ 2kTkkx − EPn(x)k2. Since kx− EPn(x)k2 → 0 and limn τ (pn) > 0, we get that ϕ(T x) = ϕ(xT ), and the claim is proven. Finally, note that ϕM ≤ 1 c and ϕ(x) = τ (xy), for all x ∈ M . Let p ∈ P ′∩M be the support projection of y. Then y ≤ 1 c p, hence τ (p) ≥ cτ (y) = c. Since the restriction of ϕ to p(P ′∩M )p is faithful, [OP07, Theorem 2.1] implies that P p is amenable relative to Q inside M , which finishes the proof. c τ . Thus, we can find y ∈ P ′ ∩ M such that 0 ≤ y ≤ 1 (cid:4) Corollary 2.5. Let (M, τ ) and (N, τ ′) be tracial von Neumann algebras. Assume that there exists a net of von Neumann subalgebras Pn ⊂ M , n ∈ I, and trace preserving ∗-homomorphisms πn : N → M such that kπn(x) − EPn(πn(x))k2 → 0, for every x ∈ N . For n ∈ I, let pn ∈ P ′ n ∩ M be a projection such that Pnpn is amenable. Then there is a projection z ∈ Z(N ) such that N z is amenable and τ (z) ≥ limn τ (pn). In particular, if Pn is amenable for every n, then N is amenable. Proof. For every n, let Mn = M and view Pn and N as subalgebras of Mn, via the identity map and πn, respectively. If we put M = ∗N,n∈I Mn, then we have kEPn(x) − xk2 → 0, for every x ∈ N . Since Pnpn is amenable for every n, Lemma 2.4 implies the existence of a projection p ∈ N ′ ∩ M such that N p is amenable and τ (p) ≥ limn τ (pn). Thus, if z is the support projection of EZ(N )(p), then N z is amenable. Since z ≥ p, we have that τ (z) ≥ τ (p), which finishes the proof. The next Lemma, which appears to be of independent interest, provides general conditions which guarantee that if P is amenable relative to a decreasing net of subalgebras Qn, then P is amenable relative to their intersection, ∩nQn. More generally, we have: Lemma 2.6. Let (M, τ ) be a tracial von Neumann algebra and Q ⊂ M a von Neumann subalgebra. Assume that there exist nets of von Neumann subalgebras Qn, Mn ⊂ M such that (cid:4) (1) Q ⊂ Mn ∩ Qn and QnL2(M )Mn ⊂weak QnL2(Qn) ⊗Q L2(Mn)Mn, for every n, (2) kx − EMn(x)k2 → 0, for every x ∈ M . If P ⊂ M is a von Neumann subalgebra which is amenable relative to Qn inside M , for every n, then P is amenable relative to Q inside M . 8 A. IOANA AND P. SPAAS Lemma 2.6 applies in particular if there exists un ∈ U (M ) such that unP u∗ generally, if P ≺s that P is amenable relative to Qn inside M . n ⊂ Qn, or, more M Qn, for every n. Indeed, by [DHI16, Lemma 2.6(3)], the latter condition implies Proof. Assume that P is amenable relative to Qn, for every n. Then [OP07, Theorem 2.1] gives that P L2(M )M ⊂weak P L2(M ) ⊗Qn L2(M )M , and thus P L2(M )Mn ⊂weak P L2(M ) ⊗Qn L2(M )Mn, for every n. Since QnL2(M )Mn ⊂weak QnL2(Qn) ⊗Q L2(Mn)Mn, we further get that P L2(M )Mn ⊂weak P L2(M ) ⊗Q L2(Mn)Mn, and thus P L2(M ) ⊗Mn L2(M )M ⊂weak P L2(M ) ⊗Q L2(Mn) ⊗Mn L2(M )M = P L2(M ) ⊗Q L2(M )M , for every n. On the other hand, since kx − EMn(x)k2 → 0, for every x ∈ M , we have P L2(M ) ⊗Mn L2(M )M . P L2(M )M ⊂weak Mn By combining the last two displayed inclusions, we get that P L2(M )M ⊂weak P L2(M )⊗Q L2(M )M , and therefore P is amenable relative to Q inside M . (cid:4) Remark 2.7. Several weaker versions of particular cases of Lemma 2.6 have been observed before. Indeed, conditions (1) and (2) from Lemma 2.6 are satisfied in the two following cases: (a) M = ∗Q,k∈NMk is an amalgamated free product of tracial von Neumann algebras (Mk)k∈N (b) M = ( ¯⊗k∈NMk) ¯⊗Q is an infinite tensor product of tracial von Neumann algebras (Mk)k∈N over a common subalgebra Q, Qn = ∗Q,k≥nMk and Mn = ∗Q,k<nMk. and Q, Qn = ( ¯⊗k≥nMk) ¯⊗Q and Mn = ( ¯⊗k<nMk) ¯⊗Q. Lemma 2.6 was first noticed by the first author in case (a) under the assumption that P can be unitarily conjugated into Qn, and extended in [HU15, Proposition 4.2] to cover the more general assumption that P ≺s M Qn. When Q = C1, the latter result was also noticed by R. Boutonnet and S. Vaes (personal communication), whose proof inspired our Lemma 2.6. In case (b), weaker versions of Lemma 2.6 were obtained in [Is16, Lemma 4.4] and [CU18, Proposition 2.7]. 2.5. Malleable deformations. In [Po01, Po03], Popa introduced the notion of an s-malleable deformation of a von Neumann algebra. In combination with his powerful deformation/rigidity techniques, this notion has led to remarkable progress in the theory of von Neumann algebras (see, e.g., [Po07, Va10a, Io18]). S-malleable deformations will also play an important role in this paper. Definition 2.8. Let (M, τ ) be a tracial von Neumann algebra. We say that a triple ( M , (αt)t∈R, β) is an s-malleable deformation of M if the following conditions hold: (1) ( M , τ ) is a tracial von Neumann algebra such that M ⊃ M and τM = τ , (2) (αt)t∈R ⊂ Aut( M , τ ) is a 1-parameter group with limt→0 kαt(x) − xk2 = 0, for all x ∈ M . (3) β ∈ Aut( M , τ ) satisfies β2 = Id M , βαtβ−1 = α−t for all t ∈ R, and β(x) = x, for all x ∈ M . As established in [Po06a], s-malleable deformations have the following "transversality" property: Lemma 2.9 ([Po06a, Lemma 2.1]). For any x ∈ M and t ∈ R we have kx − α2t(x)k2 ≤ 2kαt(x) − EM (αt(x))k2 . II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 9 2.6. Gaussian and free Bogoljubov actions. We next discuss two kinds of actions that will play a crucial role in this paper, Gaussian and free Bogoljubov actions. Below we describe one possible construction of these actions, following [PS09] and [VDN92]. For further properties of Gaussian and free Bogoljubov actions, we refer the reader to [Bo14] and [Ho12a], respectively. For the remainder of the preliminaries, we fix an orthogonal representation π : Γ → O(HR) of a countable group Γ on a real Hilbert space HR. Let H = HR ⊗R C be the complexified Hilbert space, H ⊗n its nth tensor power, and H ⊙n its symmetric nth tensor power. The latter is the closed subspace of H ⊗n spanned by vectors of the form n! Xσ∈Sn with the inner product normalized such that kξk2 H ⊙n = n!kξk2 Fock space H ⊙n, ξ1 ⊙ ··· ⊙ ξn := 1 S(H) := CΩ ⊕Mn≥1 ξσ(1) ⊗ ··· ⊗ ξσ(n), H ⊗n. We then consider the symmetric where the unit vector Ω is the so-called vacuum vector. Any vector ξ ∈ H gives rise to an unbounded operator ℓξ on S(H), the so-called left creation operator, defined by ℓξ(Ω) = ξ, and ℓξ(ξ1 ⊙ ··· ⊙ ξn) = ξ ⊙ ξ1 ⊙ ··· ⊙ ξn. Denoting s(ξ) = ℓξ + ℓ∗ ξ, one checks that the operators {s(ξ)}ξ∈H commute. Moreover, one can show ([PS09]) that with respect to the vacuum state h·Ω, Ωi, they can be regarded as independent random variables with Gaussian distribution N (0,kξk2). Consider the abelian von Neumann algebra Aπ ⊂ B(S(H)) generated by all operators of the form ω(ξ1, . . . , ξn) := exp(iπs(ξ1) . . . s(ξn)), together with the trace τ = h·Ω, Ωi. Any orthogonal operator T ∈ O(HR) can also be viewed as a unitary operator on its complexification H, and gives rise to a unitary operator on S(H), which we will still denote by T , defined by T (Ω) = Ω, and T (ξ1 ⊙ ··· ⊙ ξn) = (T ξ1) ⊙ ··· ⊙ (T ξn). One then checks that T ω(ξ1, . . . , ξn)T ∗ = ω(T ξ1, . . . , T ξn), hence T normalizes Aπ. Since T (Ω) = Ω, Ad(T ) is a trace preserving automorphism of Aπ. Definition 2.10. The Gaussian action associated to π is the action σ = σπ : Γ y (Aπ, τ ) defined by σg = Ad(π(g)), for every g ∈ Γ. One can easily check that the unitaries ω(ξ) satisfy the properties ω(0) = 1, ω(ξ + η) = ω(ξ)ω(η), τ (ω(ξ)) = exp(−kξk2), and σg(ω(ξ)) = ω(π(g)ξ) for all ξ, η ∈ H, g ∈ Γ. This in fact gives an equivalent description of the Gaussian action (see [Va10b]). The free Bogoljubov action arises in a similar way using the full Fock space F(H) := CΩ ⊕Mn≥1 H ⊗n. We consider the left creation operator Lξ associated to ξ ∈ H defined by Lξ(Ω) = ξ, and Lξ(ξ1 ⊗ ··· ⊗ ξn) = ξ ⊗ ξ1 ⊗ ··· ⊗ ξn. Putting W (ξ) = Lξ + L∗ ξ, one can show ([VDN92]) that the distribution of the self-adjoint operator W (ξ) with respect to the vacuum state h·Ω, Ωi is the semicircular law supported on [−2kξk , 2kξk], and that for any orthogonal set of vectors from HR, the associated family of operators is freely independent with respect to h·Ω, Ωi. 10 A. IOANA AND P. SPAAS Denote by Γ(HR)′′ the von Neumann algebra generated by {W (ξ) ξ ∈ HR}. Then Γ(HR)′′ is isomorphic to the free group factor L(Fdim(HR)). Moreover, τ = h·Ω, Ωi is a normal faithful trace on Γ(HR)′′. As for the symmetric Fock space, any operator T ∈ O(HR) induces an operator T ∈ U (F(H)), satisfying Ad(T )(W (ξ)) = W (T ξ). Definition 2.11. The free Bogoljubov action associated to π is the action ρ = ρπ : Γ y (Γ(HR)′′, τ ) defined by ρg = Ad(π(g)), for every g ∈ Γ. Since Γ(HR)′′Ω = F(H), the Koopman representation associated to ρ of Γ on L2(Γ(HR)′′) is isomorphic to the representation of Γ on F(H). This implies the following fact which will be needed later on: Lemma 2.12. Denote by ρ0 the restriction of the Koopman representation of ρ to L2(Γ(HR)′′)⊖C1. If π⊗k is weakly contained in the left regular representation of Γ, for some k ∈ N, then ρ⊗k is weakly contained in the left regular representation of Γ. 0 2.7. Deformations associated to Gaussian and free Bogoljubov actions. We will now re- call the construction of s-malleable deformations of the crossed product von Neumann algebras associated to the above actions. On HR ⊕ HR consider the orthogonal operators At =(cid:18)cos( π sin( π 2 t) − sin( π 2 t) cos( π 2 t) 2 t) (cid:19) , t ∈ R, and B =(cid:18)1 0 0 −1(cid:19) . We note that canonically, Aπ⊕π ∼= Aπ ¯⊗Aπ and Γ(HR ⊕ HR)′′ ∼= Γ(HR)′′ ∗ Γ(HR)′′. Under these identifications, we have that σπ⊕π ∼= σπ ⊗ σπ and ρπ⊕π ∼= ρπ ∗ ρπ, respectively. Associated to the operators At and B we get automorphisms αt := Ad(At), t ∈ R, and β := Ad(B) of Aπ ¯⊗Aπ and Γ(HR)′′ ∗ Γ(HR)′′, respectively. Since At and B commute with π ⊕ π, it follows that αt and β commute with σπ ⊗ σπ and ρπ ∗ ρπ, respectively. • For the Gaussian action, let M = Aπ ⋊ Γ, M = (Aπ ¯⊗Aπ) ⋊ Γ, and view M as a subalgebra of M via M ∼= (Aπ ¯⊗1) ⋊ Γ. By the discussion above, the automorphisms αt and β of Aπ ¯⊗Aπ extend to automorphisms of M by letting αt(ug) = β(ug) = ug, for all g ∈ Γ. • For the free Bogoljubov action, let M = Γ(HR)′′ ⋊ Γ, M = (Γ(HR)′′ ∗ Γ(HR)′′) ⋊ Γ, and view M as a subalgebra of M via M ∼= (Γ(HR)′′ ∗ 1) ⋊ Γ. By the discussion above, the automorphisms αt and β of Γ(HR)′′ ∗ Γ(HR)′′ extend to automorphisms of M by letting αt(ug) = β(ug) = ug, for all g ∈ Γ. In both cases, it is easy to check that ( M , (αt)t∈R, β) is an s-malleable deformation of M . 3. Spectral gap rigidity This section is devoted to the following rigidity result and its Corollary 3.2. Theorem 3.1. Let (M, τ ) be a tracial von Neumann algebra and N, P ⊂ M be von Neumann subalgebras. Assume that there exists an s-malleable deformation ( M , (αt)t∈R, β) such that (1) The M -bimodule H := L2( M ) ⊖ L2(M ) has the property that H⊗M kis weakly contained in (2) The M -bimodule L2( M ) with the bimodular structure given by x · ξ · y = xξα1(y), for every the bimodule L2(M ) ⊗N L2(M ), for some k ∈ N. x, y ∈ M, ξ ∈ L2( M ), is contained in a multiple of the bimodule L2(M ) ⊗P L2(M ). II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 11 M P . Let Q ⊂ M be a von Neumann subalgebra such that Qp is not amenable relative to N inside M , for any non-zero projection p ∈ Q′ ∩ M . Then Q′ ∩ M ≺s The proof of Theorem 3.1 relies on Popa's deformation/rigidity theory and notably uses his spectral gap rigidity principle introduced in [Po06a,Po06b]. Theorem 3.1 and Corollary 3.2 were inspired by Boutonnet's work (see [Bo12] and [Bo14, Chapter II]), whose exposition we follow closely. Finally, we note that condition (1) in Theorem 3.1 was first considered by Sinclair in [Si10]. Corollary 3.2. Let Γ be a countable group and π : Γ → O(HR) be an orthogonal representation. Assume that π⊗k is weakly contained in the left regular representation of Γ, for some k ∈ N. Let Γ y (C, τ ) be either the Gaussian action or the free Bogoljubov action associated to π. Let Γ y (D, ρ) be a trace preserving action on a tracial von Neumann algebra D, consider the diagonal product action Γ y (C ¯⊗D, τ ⊗ ρ), and denote M = (C ¯⊗D) ⋊ Γ. Let Q ⊂ M be a von Neumann subalgebra such that Qp is not amenable relative to D inside M , for any non-zero projection p ∈ Q′ ∩ M . Then Q′ ∩ M ≺s The remainder of this section is devoted to the proofs of Theorem 3.1 and Corollary 3.2. Lemma 3.3 ([Bo12]). Let ( M , τ ) be a tracial von Neumann algebra and N ⊂ M ⊂ M be von Neumann subalgebras. Assume that the M -bimodule H := L2( M ) ⊖ L2(M ) has the property that H⊗M k is weakly contained in the bimodule L2(M ) ⊗N L2(M ), for some k ∈ N. Let Q ⊂ M be a von Neumann subalgebra such that Qp is not amenable relative to N inside M , for any non-zero projection p ∈ Q′ ∩ M . Then Q′ ∩ M ω ⊂ M ω. In particular, Q′ ∩ M ⊂ M . Proof. The proof of [Bo12, Lemma 2.3], which applies verbatim for N = C1, works in general. (cid:4) M D ⋊ Γ. The following lemma is a standard application of Popa's spectral gap rigidity principle. Lemma 3.4. Let (M, τ ) be a tracial von Neumann algebra and N ⊂ M be a von Neumann sub- algebra. Assume that there exists an s-malleable deformation ( M , (αt)t∈R, β) such that the M - bimodule H := L2( M ) ⊖ L2(M ) has the property that H⊗M k is weakly contained in the bimodule L2(M ) ⊗N L2(M ), for some k ∈ N. Let Q ⊂ M be a von Neumann subalgebra such that Qp is not amenable relative to N inside M , for any non-zero projection p ∈ Q′ ∩ M . Then αt converges uniformly on (Q′ ∩ M )1. Proof. Fix ε > 0. Since Q′ ∩ M ω ⊂ M ω by Lemma 3.3, there exist x1, . . . , xn ∈ Q and δ > 0 such that for all y ∈ ( M )1: ∀i ∈ {1, . . . , n} : k[y, xi]k2 ≤ δ =⇒ ky − EM (y)k2 ≤ ε. Taking t > 0 such that kαs(xi) − xik2 ≤ δ x ∈ (Q′ ∩ M )1 2 for all 1 ≤ i ≤ n and all s ∈ [0, t], we get for any kαs(x)xi − xiαs(x)k2 = kxα−s(xi) − α−s(xi)xk2 ≤ 2kxkkα−s(xi) − xik2 + kxxi − xixk2 ≤ 2kαs(xi) − xik2 ≤ δ. Hence for all s ∈ [0, t] and x ∈ (Q′ ∩ M )1, we have kαs(x) − EM (αs(x))k2 ≤ ε and thus by Lemma 2.9, kα2s(x) − xk2 ≤ 2ε. It follows that αt converges uniformly on (Q′ ∩ M )1. Lemma 3.5. Assume the setting of Lemma 3.4 and let p ∈ (Q′∩ M )′∩ M be a non-zero projection. Then there is a non-zero element a1 ∈ p M α1(p) such that xa1 = a1α1(x) for all x ∈ (Q′ ∩ M )p. (cid:4) 12 A. IOANA AND P. SPAAS Proof. We follow closely the proof of [Po03, Theorem 4.1]. Put D = Q′ ∩ M and fix a projection p ∈ D′ ∩ M . Claim 1. For any t > 0 small enough, there exists a non-zero element at ∈ p M αt(p) such that at = uatαt(u∗) for all u ∈ U (Dp). Proof of Claim 1. By Lemma 3.4, αt → id uniformly on (Dp)1, as t → 0. Thus, for any t > 0 small enough we have that ku − αt(u)k2 (3.1) 2 ≤ τ (p) and hence τ (p) ℜτ (uαt(u∗)) ≥ , for all u ∈ U (Dp). 2 t u∗ β(u1ru∗ 1)u2ru∗ t )u∗ 2 > 0, hence at 6= 0. Consider the unique element at of minimal k.k2-norm in the k.k2-closure of the convex hull of the set {uαt(u∗) u ∈ U (Dp)}. By uniqueness, we have at = uatαt(u∗) for all u ∈ U (Dp). Moreover, by (3.1) we get ℜτ (at) ≥ τ (p) Claim 2. Let t > 0 and at ∈ p M αt(p) be a non-zero element such that at = uatαt(u∗) for all t )bat) 6= 0. Moreover, a2t ∈ p M α2t(p) u ∈ U (Dp). Then there exists b ∈ Q such that a2t := αt(β(a∗ satisfies a2t = ua2tα2t(u∗) for all u ∈ U (Dp). Proof of Claim 2. To prove the first part of the claim, assume that αt(β(a∗ 1) = u∗ β(a∗ t )bat) = 0 and thus 1, we get that t )bat = 0, for all b ∈ Q. Thus, if we let r = ata∗ t ∈ M , then since β(u∗ 1u2at)(a∗ 2 = β(u1at)(β(a∗ 2) = 0, for all u1, u2 ∈ U (Q). (3.2) Let s be the element of minimal k.k2-norm in the k.k2-closure of the convex hull of the set {uru∗ u ∈ U (Q)}. Since τ (s) = τ (r) > 0 and s ≥ 0, we get that s 6= 0 and further that s2 6= 0. By uniqueness, we have that s ∈ Q′∩ M and since Q′∩ M ⊂ M by Lemma 3.3 we conclude that s ∈ M . By combining the last two facts we get that β(s)s = s2 6= 0. This however contradicts (3.2) which implies that β(s)s = 0. The moreover assertion is now a straightforward calculation. By Claim 1, its conclusion holds for t = 2−k for some k ∈ N. Using Claim 2 and induction, we then find 0 6= a1 ∈ p M α1(p) such that a1 = ua1α1(u∗), for all u ∈ U (Dp). Proof of Theorem 3.1. Let p ∈ (Q′ ∩ M )′ ∩ M be a non-zero projection. We need to show that (Q′ ∩ M )p ≺M P . By Lemma 3.5 we can find 0 6= a1 ∈ p M α1(p) such that xa1 = a1α1(x) for all x ∈ (Q′ ∩ M )p. Thus, the pM p-bimodule pM pL2( M )α1(pM p) contains a non-zero (Q′ ∩ M )p-central vector. Since this bimodule is contained in a multiple of pL2(M ) ⊗P L2(M )p by assumption (2), we get that pL2(M ) ⊗P L2(M )p contains a non-zero (Q′ ∩ M )p-central vector. In other words, the pM p-bimodule pL2(hM, eP i)p contains a non-zero (Q′ ∩ M )p-central vector ξ. Let ε > 0 such that f = 1[ε,∞)(ξ∗ξ) 6= 0. Then we have that f ∈ ((Q′ ∩ M )p)′ ∩ phM, ePip. Since τ (f ) ≤ kξk2/ε < ∞, Theorem 2.3 implies that (Q′ ∩ M )p ≺M P , thus finishing the proof of the theorem. Proof of Corollary 3.2. In Section 2.7, we defined an s-malleable deformation ( C ⋊ Γ, (αt)t∈R, β) of C ⋊ Γ, where C = C ¯⊗C or C = C ∗ C, depending on whether Γ y C is the Gaussian action or the free Bogoljubov action associated to π, respectively. By construction, αt( C) = C, β( C) = C and αt(ug) = ug, for all t ∈ R and g ∈ Γ. Recall that M = (C ¯⊗D) ⋊ Γ and put M = ( C ¯⊗D) ⋊ Γ. We extend αt and β to automorphisms of M by letting αt(x) = β(x) = x, for all t ∈ R and x ∈ D. Then ( M , (αt)t∈R, β) is an s-malleable deformation of M . In order to derive the conclusion, it remains to verify that conditions (1) and (2) from Theorem 3.1 are satisfied with N = D and P = D ⋊ Γ. As in the proof of [Va10b, Lemma 3.5], given a unitary representation η : Γ → U (K), we define Kη = K ⊗ L2(M ) and endow it with the following M -bimodule structure: (cid:3) (cid:4) (aug) · (ξ ⊗ x) · (buh) = ηg(ξ) ⊗ augxbuh, for all a, b ∈ C ¯⊗D, g, h ∈ Γ, x ∈ M , and ξ ∈ K. (cid:3) (cid:4) II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 13 If η′ : Γ → U (K′) is another unitary representation of Γ, then Kη⊗η′ ∼= Kη⊗M Kη′ , and if η is weakly contained in η′, then Kη ⊂weak Kη′ . Case 1. Γ yσ (C, τ ) is the Gaussian action associated to π. Let σ0 : Γ → U (L2(C) ⊖ C1) be the restriction of the Koopman representation of σ to L2(C)⊖ C1. Since π⊗k is weakly contained in the left regular representation λ of Γ, the same holds for σ⊗k by [PS09, Proposition 2.7] and [Bo14, Proposition II.1.15]. Since the M -bimodule L2( M )⊖ L2(M ) is isomorphic to Kσ0 we conclude that 0 (L2( M ) ⊖ L2(M ))⊗M k ∼= K⊗M k σ0 ∼= Kσ⊗k 0 ⊂weak Kλ. Since C is abelian, hence amenable, Kλ ∼= L2(M )⊗C ¯⊗DL2(M ) is weakly contained in L2(M ) ¯⊗DL2(M ), k.k2 and τ (xα1(y)) = τ (xED⋊Γ(y)), for all x, y ∈ M , proving condition (1). Since L2( M ) = M α1(M ) the M -bimodule M L2( M )α1(M ) is isomorphic to L2(M )⊗D⋊ΓL2(M ). Thus, condition (2) also holds. Case 2. Γ yρ (C, τ ) is the free Bogoljubov action associated to π. We will denote still by ρ the diagonal product action of Γ on C ¯⊗D. Claim. Let ξ = ξ1ξ2...ξn ∈ C = C∗C, where ξ1 ∈ 1∗(C⊖C1), ξ2 ∈ (C⊖C1)∗1, ..., ξn ∈ 1∗(C⊖C1). Then the M -bimodule Lξ := M ξM satisfies L⊗M k Proof of the claim. Define ϕ : Γ → C and the completely positive map Φ : M → M by letting ϕ(g) = hρg(ξ), ξi and Φ((c ⊗ d)ug) = τ (c)ϕ(g)(1 ⊗ d)ug, for all c ∈ C, d ∈ D and g ∈ Γ. If c, c′ ∈ C ∗ 1, d, d′ ∈ D and g, g′ ∈ Γ, then hcρg(ξ)c′, ξi = τ (ξ∗cρg(ξ)c′) = τ (c)τ (c′)ϕ(g), and thus ⊂weak L2(M ) ⊗D L2(M ). ξ h(c ⊗ d)ugξug′(c′ ⊗ d′), ξi = δgg′,ehcρg(ξ)c′, ξihdd′, 1i = δgg′,eϕ(g)τ (c)τ (c′)τ (dd′) = τ (Φ((c ⊗ d)ug)ug′(c′ ⊗ d′)), In other words, using the notation from section 2.2, this means that Lξ ∼= HΦ, as M -bimodules. Note that if v ∈ U (C), w ∈ U (D), h ∈ Γ, then for all d ∈ D and g ∈ Γ we have that (3.3) [Φ ◦ Ad((v ⊗ w)uh)]((1 ⊗ d)ug) = τ (vρhgh−1(v)∗)ϕ(hgh−1)Ad((1 ⊗ w)uh)((1 ⊗ d)ug). Let U be the set of unitaries u ∈ M of the form u = (v ⊗ w)uh, with v ∈ U (C), w ∈ U (D), h ∈ Γ. ∼= H⊗M k Since the span of U is k.k2-dense in M , Lemma 2.2(2) implies that the M -bimodule L⊗M k is isomorphic to a sub-bimodule of Φ ξ Mu1,...,uk−1∈ U HΦ◦Ad(uk−1)◦Φ◦···◦Ad(u1)◦Φ. We fix u1, ..., uk−1 ∈ U and denote Ψ := Φ ◦ Ad(uk−1) ◦ Φ ◦ ··· ◦ Ad(u1) ◦ Φ : M → M . Thus, in order to prove the claim it suffices to argue that HΨ ⊂weak L2(M ) ⊗D L2(M ). To this end, for i ∈ {1, ..., k − 1}, write ui = (vi ⊗ wi)uhi, where vi ∈ U (C), wi ∈ U (D) and hi ∈ Γ. We define U = (1⊗ wk−1)uhk−1...(1⊗ w1)uh1 ∈ U (D ⋊ Γ) and a positive definite function ψ : Γ → C by letting ψ(g) = k−1Yi=1 τ (viρhi...h1gh−1 1 ...h−1 i (vi)∗), for all g ∈ Γ. By using (3.3) and induction, it follows that for all c ∈ C, d ∈ D and g ∈ Γ we have that (3.4) Ψ((c ⊗ d)ug) = τ (c)ψ(g)ϕ(g) ϕ(hi...h1gh−1 1 ...h−1 i )Ad(U )((1 ⊗ d)ug) k−1Yi=1 14 A. IOANA AND P. SPAAS Let Θ : M → M and Ω : M → M be the completely positive maps given by Θ(xug) = ψ(g)xug and Ω(xug) = ϕ(g)Qk−1 )xug, for all x ∈ C ¯⊗D and g ∈ Γ. Then (3.4) rewrites as Ψ = Ad(U ) ◦ Θ ◦ Ω ◦ ED⋊Γ. By Lemma 2.2(1) we get that (3.5) the M -bimodule HΨ is isomorphic to a sub-bimodule of HED⋊Γ ⊗M HΩ ⊗M HΘ. i=1 ϕ(hi...h1gh−1 1 ...h−1 i Let ρ0 : Γ → U (L2(C) ⊖ C1) be the restriction of the Koopman representation of ρ to L2(C) ⊖ C1. Since ϕ(g) = hρg(ξ), ξi =Qn i=1hρg(ξi), ξii and ξi ∈ C ⊖ C1, for all g ∈ Γ and i ∈ {1, ..., n}, it follows . Since π⊗k is weakly contained that the M -bimodule HΩ is isomorphic to a sub-bimodule of Kρ⊗kn in the left regular representation λ, so is ρ⊗k Hence, Kρ⊗kn with (3.5), we derive that 0 ⊂weak Kλ ∼= HEC ¯⊗D . Altogether, we conclude that HΩ ⊂weak HEC ¯⊗D . In combination 0 by Lemma 2.12. Thus, ρ⊗kn is weakly contained in λ. 0 0 (3.6) HΨ ⊂weak HED⋊Γ ⊗M HEC ¯⊗D ⊗M HΘ. Since HEN ∼= L2(M )⊗N L2(M ), for any von Neumann subalgebra N ⊂ M , and the (D ⋊Γ)-(C ¯⊗D)- bimodule L2(M ) is isomorphic to L2(D ⋊ Γ)⊗D L2(C ¯⊗D), it follows that HΨ ⊂weak L2(M )⊗D HΘ. Using that D is regular in M and ΘD = idD, it is easy to see that L2(M ) ⊗D HΘ is isomorphic to a sub-bimodule of a multiple of L2(M ) ⊗D L2(M ). Thus, HΨ ⊂weak L2(M ) ⊗D L2(M ), which finishes the proof of the claim. Since L2( M ) ⊖ L2(M ) decomposes as a direct sum of M -bimodules of the form Lξ as in the claim, condition (1) follows. To verify condition (2), let ξ ∈ C be a non-zero element of the form ξ = ξ1ξ2...ξn, where ξ1 ∈ 1 ∗ (C ⊖ C1), ξ2 ∈ (C ⊖ C1) ∗ 1, ..., ξn ∈ (C ⊖ C1) ∗ 1. Using a calculation similar to the one in the claim, it follows that the M -bimodule M M ξα1(M )α1(M ) is isomorphic to a submodule of a multiple of L2(M )⊗D⋊Γ L2(M ). This implies that condition (2) holds in case (2) and finishes the proof of Corollary 3.2. (cid:3) (cid:4) 4. Proofs of Theorems A and B k The proofs of Theorems A and B rely on the following consequence of Corollary 3.2. Lemma 4.1. Let Γ be a non-amenable group. For k ∈ N, let πk : Γ → O(Hk) be an orthogonal representation such that π⊗l(k) is weakly contained in the left regular representation of Γ, for some l(k) ∈ N. Let Γ y (Bk, τk) be either the Gaussian or the free Bogoljubov action associated to πk. Let Γ y (B, τ ) := ¯⊗k(Bk, τk) be the diagonal product action and denote M = B ⋊ Γ. Let (Mn)n∈N be a sequence of von Neumann subalgebras of M such that kx − EMn(x)k2 → 0, for every x ∈ M . Then there exist projections pn ∈ Z(M ′ n ∩ M ), for n ∈ N, such that limn→∞ τ (pn) = 1 and (M ′ Moreover, if Γ is not inner amenable, then there exist projections rn ∈ Z(M ′ such that limn→∞ τ (rn) = 1 and (M ′ M ( ¯⊗k>N Bk) ⋊ Γ, for every n, N ∈ N. n ∩ M ), for n ∈ N, n ∩ M )pn ≺s n ∩ M )rn is amenable, for every n ∈ N. Proof. Let qn ∈ Z(M ′ n ∩ M ) be the largest projection such that Mnqn is amenable relative to B. We claim that τ (qn) → 0. Otherwise, after replacing (Mn)n∈N with a subsequence, we may assume that τ (qn) → c > 0. By Lemma 2.4, this implies that there is a non-zero projection q ∈ Z(M ) such that M q is amenable relative to B. Since M is a factor, this would give that M is amenable relative to B and hence that Γ is amenable by [OP07, Proposition 2.4], which is a contradiction. Next, fix n ∈ N and put pn = 1 − qn. Then Mnp′ is not amenable relative to B, for any non-zero projection p′ ∈ (M ′ n∩M )pn. Otherwise, [DHI16, Lemma 2.6(2)] would provide a non-zero projection z ∈ Z(M ′ n ∩ M )pn such that Mnz is amenable relative to B, contradicting the maximality of qn. II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 15 Let i ∈ N and denote Ci = ¯⊗k6=iBk. Since Ci ⊂ B, Mnp′ is not amenable relative to Ci, for any non-zero projection p′ ∈ (M ′ n ∩ M )pn. Since Γ y Bi is either the Gaussian or the free Bogoljubov action associated to πi, and a multiple of πi is weakly contained in the left regular representation of Γ, we can apply Corollary 3.2 to the inclusion Mnpn ⊂ M = (Bi ¯⊗Ci) ⋊ Γ to deduce that (4.1) (M ′ n ∩ M )pn ≺s M Ci ⋊ Γ, for all i ∈ N. M ∩N n ∩ M )pn ≺s i=1 of M are regular and any two form a commuting square, i=1(Ci ⋊ Γ) = ( ¯⊗k>N Bk) ⋊ Γ. Since Let N ∈ N. Since the subalgebras {Ci}N (4.1) and [DHI16, Lemma 2.8(2)] imply that (M ′ τ (pn) → 1, this proves the main assertion. For the moreover assertion, assume that Γ is not inner amenable. Then by [Ch82] we get that M ′ ∩ M ω ⊂ Bω and henceQω(M ′ n ∩ M ) ⊂ M ′ ∩ M ω ⊂ Bω. By combining this with [IS18, Lemmas 2.2 and 2.3] we can find projections qn ∈ Z(M ′ n ∩ M )qn ≺s (4.2) Put rn = pn ∧ qn ∈ Z(M ′ M ( ¯⊗k>N Bk) ⋊ Γ, for every n, N ∈ N. Since B is regular in M , B and ( ¯⊗k>N Bk) ⋊ Γ form a commuting square and B ∩ (( ¯⊗k>N Bk) ⋊ Γ) = ¯⊗k>N Bk, [DHI16, Lemma 2.8(2)] implies that (4.3) For N ∈ N, put QN = ¯⊗k>N Bk and RN = ( ¯⊗k≤N Bk) ⋊ Γ. Then kx − ERN (x)k2 → 0, for any x ∈ M , and QN L2(M )RN ∼= QN L2(QN ) ⊗ L2(RN )RN , for any N ∈ N. These facts and (4.3) imply that we can apply Lemma 2.6 to deduce that (M ′ n ∩ M ) such that τ (qn) → 1 and M B, for every n ∈ N. M B and (M ′ M ¯⊗k>N Bk, for every n, N ∈ N. n ∩ M ). Then (M ′ n ∩ M )rn ≺s n ∩ M )rn ≺s n ∩ M )rn ≺s (M ′ (M ′ (cid:4) n ∩ M )rn is amenable, for every n ∈ N. 4.1. Proof of Theorem A. Assume by contradiction that M admits a residual sequence (An)n. n ⊂ M ′ ∩ M ω, Lemma 2.1 implies that For n ∈ N, let Mn = A′ kx − EMn(x)k2 → 0, for every x ∈ M . By Lemma 4.1 we can find projections pn ∈ Z(M ′ n ∩ M ) M ( ¯⊗l>N Bl) ⋊ Γ, for every n, N ∈ N. Since An ⊂ M ′ such that τ (pn) → 1 and (M ′ n ∩ M , we thus get that n ∩ M . Since Qω An ⊂ ∩nAω n ∩ M )pn ≺s (4.4) Anpn ≺s M ( ¯⊗l>N Bl) ⋊ Γ, for every n, N ∈ N. k = ω(ξm Let n ∈ N be fixed such that τ (pn) > 15/16. Recall that Γ y Bk is the Gaussian action associated to πk and denote U m Claim. There exists k ∈ N such that kU m Proof of the claim. Assuming the claim is false, for every k ∈ N, we can find m(k) ∈ N such that Uk := U m(k) ∈ U (Bk) satisfies kUk − EAn(Uk)k2 > 1/16. Since 1 − e−t ≤ t, for any t ≥ 0, we get k ) ∈ U (Bk), for every k, m ∈ N. k )k2 ≤ 1/16, for every m ∈ N. k − EAn(U m k kugUku∗ k k g − Ukk2 = kω(πk(g)(ξm(k) )) − ω(ξm(k) =q2(1 − exp(−kπk(g)(ξm(k) √2kπk(g)(ξm(k) ) − ξm(k) ≤ k k → 0, we deduce that kugUku∗ k k )k2 ) − ξm(k) k k2)) k k ) − ξm k, for every g ∈ Γ. Since supm∈N kπk(g)(ξm g − Ukk2 → 0, for every g ∈ Γ. Since Uk ∈ U (Bk), we also have that Ukx = xUk, for every x ∈ B. By combining the last two facts we get that U := (Uk) ∈ M ′ ∩ M ω. However, since kU − EAω n (U )k2 = limk→ω kUk − EAn(Uk)k2 ≥ 1/16, this contradicts that M ′ ∩ M ω ⊂ Aω Let k ∈ N be as in the claim and put Vm = U m kVmk2 =p1 − exp(−1) and kVm − EAn(Vm)k2 ≤ 1/16, for every m ∈ N. Since τ (V ∗ all m 6= m′, we also have that Vm → 0 weakly. k ). Then we have Vm ∈ Bk, kVmk ≤ 2, m′Vm) = 0, for n. Altogether, this proves the claim. k − τ (U m (cid:3) 16 A. IOANA AND P. SPAAS M ( ¯⊗l>kBl) ⋊ Γ. This implies that we can find By specializing (4.4) to N = k we get that Anpn ≺s a finite dimensional subspace K ⊂ ¯⊗l≤kBl such that if e denotes the orthogonal projection from L2(M ) onto the k.k2-closed linear span of {(y ⊗ z)ug y ∈ K, z ∈ ¯⊗l>kBl, g ∈ Γ}, then (4.5) Next, if m ∈ N, then kVm − EAn(Vm)k ≤ 1/16 and hence kVmpn − EAn(Vm)pnk2 ≤ 1/16. Since EAn(Vm)pn ∈ Anpn and kEAn(Vm)pnk ≤ 2, (4.5) gives kEAn(Vm)pn − e(EAn(Vm)pn)k2 ≤ 1/8. Combining the last two inequalities further implies that kx − e(x)k2 ≤ 1/16, for all x ∈ (Anpn)1. (4.6) Now, we claim that kVmpn − e(Vmpn)k2 ≤ 1/4, for every m ∈ N. m→∞kE( ¯⊗l>kBl)⋊Γ(xVmy)k2 = 0, for all x, y ∈ M . lim (4.7) Indeed, it is enough to check this when x = ug(a ⊗ b) and y = (c ⊗ d)uh, for a, c ∈ ¯⊗l≤kBl, b, d ∈ ¯⊗l>kBl and g, h ∈ Γ. Then, since Vm ∈ Bk, we have E( ¯⊗l>kBl)⋊Γ(xVmy) = τ (aVmb)ugbduh and the conclusion follows since Vm → 0 weakly. This proves (4.7). j=1 be an orthonormal basis for K. Since E( ¯⊗l>kBl)⋊Γ(ξ∗ Let {ξj}r we get that e(x) =Pr that ke(Vmpn)k2 → 0. On the other hand, since kVmk ≤ 2 and τ (pn) > 15/16, we have that kVmpnk2 ≥ kVmk2 − kVm(1 − pn)k2 ≥ kVmk2 − 2k1 − pnk2 i ξj) = δi,j, for all i, j ∈ {1, ..., r}, j x), for every x ∈ M . In combination with (4.7) it follows j=1 ξjE( ¯⊗l>kBl)⋊Γ(ξ∗ Altogether, we get that lim inf m→∞ kVmpn − e(Vmpn)k2 > 1/4, which contradicts (4.6). So M cannot have a residual sequence. (cid:4) =p1 − exp(−1) − 2p1 − τ (pn) > 1/4, for every m ∈ N. Remark 4.2. The proof of Theorem A shows that there is no sequence (An)n∈N of von Neumann n. In particular, there is no sequence (An)n∈N of von Neumann subalgebras of M which satisfies conditions (2) and (3) of Definition 1.1. subalgebras of M such that Qω An ⊂ M ′ ∩ M ω ⊂ ∩n∈NAω 4.2. Proof of Theorem B. Recall that Γ y Bk is the free Bogoljubov action associated to πk and denote Wk,m = W (ξm k ) ∈ Bk, for k ∈ N and m ∈ {1, 2}. Then for any k ∈ N, {Wk,1, Wk,2} are freely independent semicircular operators with kWk,1k = kWk,2k = 2. Moreover, if m ∈ {1, 2}, then for any g ∈ Γ we have that kugWk,mu∗ k k → 0. Since Wk,m ∈ Bk, we also have that kWk,mx − xWk,mk2 → 0, for every x ∈ B. By combining the last two facts, we get that Wm = (Wk,m)k ∈ M ′ ∩ M ω. Let us first prove the moreover assertion. To this end, let P ⊂ M ′ ∩ M ω be the von Neumann subalgebra generated by W1 and W2. Assume by contradiction that there is a sequence (An)n of von Neumann subalgebras of M such that g−Wk,mk2 = kW (πk(g)(ξm k )k2 = kπk(g)(ξm k ))−W (ξm k )−ξm P ⊂Yω An ⊂ M ′ ∩ M ω. For n ∈ N, let Mn = A′ n∩ M . Lemma 2.1 implies that limn→ω kx− EMn(x)k2 → 0, for every x ∈ M . The moreover assertion of Lemma 4.1 implies the existence of projections rn ∈ Z(M ′ n ∩ M ) such that limn→ω τ (rn) → 1 and (M ′ n ∩ M )rn is amenable, for every n ∈ N. Thus, Anrn is amenable, for every n ∈ N. If n ∈ N, then since Wm = (Wk,m)k ∈ P ⊂Qω Ak, there is kn ∈ N satisfying τ (rkn) ≥ 1− 1/n2 and kWkn,m − EAkn (Wkn,m)k2 ≤ 1/n, for every m ∈ {1, 2}. Thus, if Bn = Aknrkn ⊕ C(1 − rkn), then (4.8) kWkn,m − EBn(Wkn,m)k2 ≤ 1/n + k1 − rknk2 ≤ 2/n, for every n ∈ N and m ∈ {1, 2}. II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 17 Let N be the II1 factor generated by two freely independent semicircular operators S1, S2 with kS1k = kS2k = 2. For n ∈ N, let πn : N → M be the unique trace preserving ∗-homomorphism such that πn(Sm) = Wkn,m, for all m ∈ {1, 2}. Then (4.8) gives that kπn(x) − EBn(πn(x))k2 → 0, for every x ∈ N . Since Bn is amenable, for every n ∈ N, Corollary 2.5 implies that N is amenable. Since N ∼= L(F2) is not amenable, this gives a contradiction and thus proves the moreover assertion. To prove the main assertion, assume by contradiction that M admits a residual sequence (An)n. Then P ⊂ M ′ ∩ M ω = ∩nAω n and since P is separable, we can find an increasing sequence of positive integers (kn) such that P ⊂Qω Akn. Since Qω Ank ⊂ ∩nAω n = M ′ ∩ M ω, this contradicts the moreover assertion. (cid:4) 5. Stability 5.1. Proof of Proposition C. Since P is amenable, it is aproximately finite dimensional by Connes' theorem [Co76]. Thus, we can find an increasing sequence (Bk)k of finite dimensional von Neumann subalgebras such that P = (∪kBk)′′. If k ∈ N, then since Bk is finite dimensional, there exists Sk ∈ ω such that for every n ∈ Sk we have an embedding Bk ⊂ Mn in such a way that the embedding Bk ⊂Qω Mn is the diagonal embedding. Put S0 = N. Claim. There exists a sequence (kn) ⊂ N such that n ∈ Skn, for all n ∈ N, limn→ω kn = +∞, and Q ⊂Yω (B′ kn ∩ Mn). Proof of the claim. Since Bk is finite dimensional, Q ⊂ P ′∩Qω Mn ⊂ B′ for every k ∈ N. Hence Q ⊂ ∩k∈NQω(B′ k∩Mn(qn)(cid:13)(cid:13)(cid:13)2 Now, let {q(m)}m∈N be a k.k2-dense sequence in (Q)1. Let X0 = N and k ∩ Mn), i.e. (5.1) = 0, for all k ∈ N and q = (qn) ∈ Q. k∩Qω Mn =Qω(B′ k∩ Mn), and thus n − EB′ contradicting (5.1). By construction Q ⊂Qω(B′ Taking (kn) as in the Claim, we also have that P ⊂Qω Bkn. Thus, Pn = Bkn and Qn = B′ kn ∩ Mn), which finishes the proof of the claim. (cid:3) kn ∩ Mn verify the conclusion of Proposition C. k + 1 (cid:4) , k+1∩Mn(q(i) 1 n )(cid:13)(cid:13)(cid:13)2 ≥ For n ∈ N, define kn to be the largest k ≤ n such that n ∈ Xk. We claim that limn→ω kn = +∞. Otherwise, there exists k ∈ N such that {n ∈ N kn = k} ∈ ω. Then {n ∈ N n 6∈ Xk+1} ∈ ω. Since Sk+1 ∈ ω, this would imply the existence of i ∈ {1, . . . , k + 1} such that we have n − EB′ k∩Mn(q(i) n )(cid:13)(cid:13)(cid:13)2 ≤ , for all 1 ≤ i ≤ k(cid:27) . 1 k n − EB′ k+1∩Mn(q(i) 1 k + 1(cid:27) ∈ ω, > n )(cid:13)(cid:13)(cid:13)2 lim n→ω(cid:13)(cid:13)(cid:13)qn − EB′ Xk =(cid:26)n ∈ Sk (cid:13)(cid:13)(cid:13)q(i) (cid:26)n ∈ N (cid:13)(cid:13)(cid:13)q(i) n→ω(cid:13)(cid:13)(cid:13)q(i) lim 18 A. IOANA AND P. SPAAS 5.2. Proof of Theorem E. In the proof of Theorem E we will need the following consequence of Corollary 3.2. Recall that a tracial von Neumann algebra (M, τ ) is called solid [Oz03] if the relative commutant P ′ ∩ M is amenable, for any diffuse von Neumann subalgebra P ⊂ M . Lemma 5.1. Let Γ be a countable group and π : Γ → O(HR) be a mixing orthogonal representation. Assume that π⊗k is weakly contained in the left regular representation of Γ, for some k ∈ N. Let Γ y (C, τ ) be the free Bogoljubov action associated to π. If L(Γ) is solid, then C ⋊ Γ is solid. Proof. Assume that L(Γ) is solid. In order to prove that M = C ⋊ Γ is solid it suffices to show that if P ⊂ M is a diffuse von Neumann subalgebra, then P ′ ∩ M has an amenable direct summand. Suppose by contradiction that P ′∩M has no amenable direct summand. By applying Corollary 3.2, we get that P ≺M L(Γ). Hence there exist projections p ∈ P, q ∈ L(Γ), a ∗-homomorphism θ : pP p → qL(Γ)q, and a non-zero partial isometry v ∈ qM p such that θ(x)v = vx for all x ∈ pP p. Since π is mixing, the action Γ y C is mixing by [Ho12a, Proposition 2.6]. Since θ(pP p) ⊂ qL(Γ)q is a diffuse subalgebra and vv∗ ∈ θ(pP p)′∩qM q, [Po03, Theorem 3.1] implies that q0 := vv∗ ∈ L(Γ). Thus, P0 := vP v∗ is a diffuse subalgebra of q0L(Γ)q0. Since v(P ′ ∩ M )v∗ ⊂ q0M q0 is a subalgebra which commutes with P0, [Po03, Theorem 3.1] gives that v(P ′ ∩ M )v∗ ⊂ P ′ 0 ∩ q0L(Γ)q0. Since L(Γ) is solid, we get that v(P ′ ∩ M )v∗ is amenable and thus P ′ ∩ M has an amenable direct summand. This finishes the proof of the lemma. (cid:4) Proof of Theorem E. First, note that if W is a self-adjoint operator in a tracial von Neumann algebra whose distribution with respect to the trace is the semicircular law supported on [−2, 2], then {W}′′ is a diffuse abelian von Neumann algebra. Hence we can find a Borel function f : [−2, 2] → T such that U = f (W ) ∈ {W}′′ is a Haar unitary, i.e. τ (U n) = 0, for all n ∈ Z \ {0}. From now on, fix two freely independent self-adjoint operators W1, W2 in a tracial von Neumann algebra whose distribution is the semicircular law supported on [−2, 2]. Define U1 = f (W1) and U2 = f (W2). Then U1 and U2 are freely independent Haar unitaries and thus N = {U1, U2}′′ satisfies N = {U1}′′ ∗ {U2}′′ ∼= L(F2). Let Γ = F2 and a1, a2 ∈ Γ be free generators. Let πk : Γ → O(Hk), k ∈ N, be a sequence of mixing representations such that a tensor multiple of πk is weakly contained in the left regular representation of Γ, and there exist unit vectors ξk,m ∈ Hk such that kπk(g)(ξm k k → 0, for every m ∈ {1, 2} and g ∈ Γ. For instance, let (πk)k∈N be as in Example 1.5 and notice that by construction πk is indeed mixing, for every k ∈ N. Let Γ y Bk be the free Bogoljubov action associated to πk and denote Mk = Bk ⋊ Γ, for every k ∈ N. Then Wk,m = W (ξm supported on [−2, 2]. Moreover, kugWk,m−Wk,mugk2 = kπk(g)(ξm and g ∈ Γ. Thus, if we put Uk,m = f (Wk,m) ∈ U (Bk), then (5.2) k ) ∈ Bk is a self-adjoint operator whose distribution is the semicircular law k k → 0, for every m ∈ {1, 2} k ) − ξm k )−ξm kugUk,m − Uk,mugk2 → 0, for every m ∈ {1, 2} and g ∈ Γ. Let ρk : N → Mk be the unique trace preserving ∗-homomorphism given by ρk(U1) = Uk,1 and ρk(U2) = Uk,2. Then (5.2) rewrites as (5.3) kugρk(Um) − ρk(Um)ugk2 → 0, for every m ∈ {1, 2} and g ∈ Γ. In the rest of the proof, we treat the two assertions of Theorem E separately. Part 1. We first prove that Γ× Γ is not W∗-tracially stable. This readily implies that Fl× Fm is not W∗-tracially stable, for every 2 ≤ l, m ≤ +∞. Assume by contradiction that Γ × Γ is W∗-tracially stable. Using (5.3) we can define a homomorphism ϕ : Γ × Γ → U (Qω Mk) by letting ϕ(am, e) = (ρk(Um))k and ϕ(e, g) = ug, for all m ∈ {1, 2} and g ∈ Γ. (5.4) II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 19 Since Γ × Γ is assumed W∗-tracially stable, there must be homomorphisms ϕk : Γ × Γ → U (Mk) such that ϕ = (ϕk)k. Let Ck = ϕk(Γ × {e})′′ and Dk = ϕk({e} × Γ)′′. Then Ck and Dk are commuting von Neumann subalgebras of Mk and we have that (5.5) (5.6) k→ω kρk(Um) − ECk (ρk(Um))k2 = 0, for every m ∈ {1, 2}, and lim k→ω kug − EDk (ug)k2 = 0, for every g ∈ Γ. lim k→ω kρk(x) − ECk (ρk(x))k2 → 0, for every x ∈ N . Since N is a non- Then (5.5) implies that lim amenable II1 factor, Corollary 2.5 implies that if pk ∈ Z(Ck) is the largest projection such that Ckpk is amenable, then limk→ω τ (pk) = 0. Since L(Γ) is also a non-amenable II1 factor, by repeating this argument using (5.6), it follows that limk→ω τ (qk) = 0, where qk ∈ Z(Dk) denotes the largest projection such that Dkqk is amenable. Thus, for every k ∈ N, rk := (1− pk)(1− qk) ∈ {Ck, Dk}′′ is a projection such that Ckrk and Dkrk have no amenable direct summands, and limk→ω τ (rk) = 1. In particular, we can find k such that rk 6= 0. This implies that rkM rk and thus M is not solid, which is a contradiction by Lemma 5.1. This finishes the proof of the first assertion of Theorem E. Part 2. For the moreover assertion, put B = ¯⊗k∈NBk and M = B⋊Γ. Using the natural embeddings Mk ⊂ M , for every k ∈ N, we can view Qω Mk as a subalgebra of M ω. Thus, we may view ϕ as a homomorphism ϕ : Γ × Γ → U (M ω). Since by the definition (5.4) of ϕ we have ϕ(a, e) ∈ Bω, τ (ϕ(a, e)) = δa,e and ϕ(e, g) = ug, it follows that τ (ϕ(a, g)) = τ (ϕ(a, e)ug) = δ(a,g),(e,e), for all a, g ∈ Γ. Thus, ϕ extends to a ∗-homomorphism ϕ : L(F2 × F2) → M ω. We claim that there are no homomorphisms ϕk : Γ × Γ → U (M ) such that ϕ = (ϕk)k. Assume by contradiction that such homomorphisms (ϕk) exist. Then Ck = ϕk(Γ×{e})′′ and Dk = ϕk({e}×Γ)′′ are commuting von Neumann subalgebras of M such that (5.5) and (5.6) hold. Since Γ is non-amenable, [OP07, Proposition 2.4] implies that L(Γ) is not amenable relative to B inside M . Thus, since L(Γ)′ ∩ M = C1, there is no non-zero projection q ∈ L(Γ)′ ∩ M such that L(Γ)q is amenable relative to B inside M . Let qk ∈ D′ k∩M be the largest projection such that Dkqk is amenable relative to B inside M . Then by [DHI16, Lemma 2.6] we have that qk ∈ Z(D′ k ∩ M ). Since by (5.6) we have that limω kx − EDk (x)k2 = 0, for every x ∈ L(Γ), we can apply Lemma 2.4 to conclude that limω τ (qk) = 0. Next, fix k ∈ N. Then Dkp′ is not amenable relative to B inside M , for any non-zero projection p′ ∈ (D′ k ∩ M )(1 − qk). For i ∈ N, let Ri = ¯⊗l6=iBl. Then by applying Corollary 3.2 to the decomposition M = (Bi ¯⊗Ri) ⋊ Γ it follows that Ck(1 − qk) ≺s M Ri ⋊ Γ, for every i ∈ N. If N ∈ N, then the subalgebras {Ri ⋊ Γ}N i=1 of M are regular and any two form a commuting square. Since i=1(Ri ⋊ Γ) = ( ¯⊗l>N Bl) ⋊ Γ, [DHI16, Lemma 2.8(2)] implies that ∩N (5.7) Ck(1 − qk) ≺s M ( ¯⊗l>N Bl) ⋊ Γ, for every k, N ∈ N. Since Γ = ha1, a2i is not inner amenable, we can find a constant c > 0 such that kx − EB(x)k2 ≤ c(k[x, ua1 ]k2 + k[x, ua2 ]k2), for every x ∈ M . (5.8) For k ∈ N, denote εk = kua1 −EDk(ua1 )k2 +kua2−EDk (ua2)k2. Then (5.6) implies that limω εk = 0. Since Ck and Dk commute, we have that k[x, ua1 ]k2 + k[x, ua2 ]k2 ≤ 2εk, for all x ∈ (Ck)1. In combination with (5.8), we get that kx − EB(x)k2 ≤ 2cεk, for all x ∈ (Ck)1. By applying [IS18, Lemma 2.2] we derive the existence of a projection rk ∈ Z(C ′ k ∩ M ) such that τ (rk) ≥ 1− 2cεk and (5.9) Ckrk ≺s M B, for every k ∈ N. 20 A. IOANA AND P. SPAAS Since B and ( ¯⊗l>N Bl) ⋊ Γ are regular subalgebras of M which form a commuting square, if pk = (1 − qk)rk ∈ C ′ (5.10) k ∩ M , by combining (5.7), (5.9) and [DHI16, Lemma 2.8(2)] we get that Ckpk ≺M ¯⊗l>N Bl, for every k, N ∈ N. Using (5.10) and reasoning as at the end of the proof of Lemma 4.1, it follows that Ckpk is amenable, for every k ∈ N. Since limω τ (qk) = 0 and limω τ (rk) = 1, we get that limω τ (pk) = 1. On the other hand, (5.5) implies that limω kρk(x) − ECk (ρk(x))k2 = 0, for every x ∈ N . By applying Corollary 2.5 we derive that N is amenable, which is a contradiction. (cid:4) [At18] [BCI15] R. Boutonnet, I. Chifan, and A. Ioana: II1 factors with non-isomorphic ultrapowers, Duke Math. J. Volume S. Atkinson: Some results on tracial stability and graph products, preprint arXiv:1808.04664. References 166, Number 11 (2017), 2023-2051. R. Boutonnet: On solid ergodicity for Gaussian actions, J. Funct. Anal. 263 (2012), no. 4, 1040-1063. R. Boutonnet: Plusieurs aspects de rigidit´e des alg`ebres de von Neumann, PhD Thesis, 2014. [Bo12] [Bo14] [Ch69] W.-M. Ching: Non-isomorphic non-hyperfinite factors, Canad. J. Math. 21 (1969), 1293-1308. [Ch82] M. Choda: Inner amenability and fullness, Proc. Amer. Math. Soc. 86 (1982), 663-666. [Co76] A. Connes: Classification of injective factors, Ann. of Math. (2) 104 (1976), no. 1, 73-115. [Co94] A. Connes: Noncommutative Geometry, Academic Press, Inc., San Diego, CA, 1994, xiv+661 pp. [CU18] I. Chifan and B. Udrea: Some rigidity results for II1 factors arising from wreath products of property (T) groups, preprint arXiv:1804.04558. [DHI16] D. Drimbe, D. Hoff, and A. Ioana: Prime II1 factors arising from irreducible lattices in products of rank one simple Lie groups, preprint arXiv:1611.02209, to appear in J. Reine. Angew. Math. J. Dixmier and E. C. Lance: Deux nouveaux facteurs de type II1, Invent. Math. 7 (1969), 226-234. [DL69] [FGL06] J. Fang, L. Ge and W. Li: Central sequence algebras of von Neumann algebras, Taiwanese J. Math. 10 [FHS09] [FHS11] [Gl03] [GH16] (2006), 187-200. I. Farah, B. Hart, and D. Sherman: Model theory of operator algebras I: stability, Bull. Lond. Math. Soc. 45 (2013), 825-838. I. Farah, B. Hart, and D. Sherman: Model theory of operator algebras III: elementary equivalence and II1 factors, Bull. Lond. Math. Soc. 46 (2014), 609-628. E. Glasner: Ergodic theory via joinings, Amer. Math. Society, 2003. I. Goldbring and B. Hart: On the theories of McDuff 's II1 factors, Int. Math. Res. Not., Volume 27, Issue 18 (2017), 5609-5628. [Ha79] U. Haagerup: An example of a non-nuclear C∗-algebra, which has the metric approximation property, Invent. Math. 50 (1978/79), 279-293. [HS16] D. Hadwin and T. Shulman: Tracial Stability for C∗-Algebras, Integral Equations Operator Theory, 90(1):90:1, 2018. [HS17] D. Hadwin and T. Shulman: Stability of group relations under small Hilbert-Schmidt perturbations, J. Funct. Anal. 275 (2018), no. 4, 761-792. [Ho12a] C. Houdayer: Structure of II1 factors arising from free Bogoljubov actions of arbitrary groups, Adv. Math. 260 (2014), 414-457. [HU15] C. Houdayer and Y. Ueda: Rigidity of free product von Neumann algebras, Compos. Math. 152 (2016), [Io12] [Io18] [Is16] [IS18] [JS85] 2461-2492. A. Ioana: Cartan subalgebras of amalgamated free product II1 factors, With an appendix by Adrian Ioana and Stefaan Vaes. Ann. Sci. ´Ec. Norm. Sup´er. (4) 48 (2015), no. 1, 71-130. A. Ioana: Rigidity for von Neumann algebras, Proc. Int. Cong. of Math. 2018 Rio de Janeiro, Vol. 2 (1635- 1668). Y. Isono: On fundamental groups of tensor product II1 factors, preprint arXiv:1608.06426, to appear in J. Inst. Math. Jussieu. A. Ioana and P. Spaas: A class of II1 factors with a unique McDuff decomposition, preprint arXiv: 1808.02965. V.F.R. Jones and K. Schmidt: Asymptotically invariant sequences and approximate finiteness, Amer. J. Math. 109 (1987), no. 1, 91-114. [Ma17] A. Marrakchi: Stability of products of equivalence relations, Compos. Math. 154 (2018), no. 9, 2005-2019. [Mc69a] D. McDuff: A countable infinity of II1 factors, Ann. of Math. (2) 90 (1969), 361-371. [Mc69b] D. McDuff: Uncountably many II1 factors, Ann. of Math. (2) 90 (1969), 372-377. II1 FACTORS WITH EXOTIC CENTRAL SEQUENCE ALGEBRAS 21 [Mc69c] D. McDuff: Central sequences and the hyperfinite factor, Proc. London Math. Soc. (3) 21 (1970), 443-461. [Mc69d] D. McDuff: On residual sequences in a II1 factor, J. London Math. Soc. (2) (1971), 273-280. [MvN43] F. Murray and J. von Neumann: Rings of operators, IV, Ann. of Math. 44 (1943), 716-808. [OP07] N. Ozawa and S. Popa: On a class of II1 factors with at most one Cartan subalgebra, Ann. of Math. (2) 172 (2010), no. 1, 713-749. [Oz03] N. Ozawa: Solid von Neumann algebras, Acta Math. 192 (2004), 111-117. [PS09] Popa: J. Peterson and T. Sinclair: On cocycle superrigidity for Gaussian actions, Erg. Th. Dyn. Sys., 32 (2012), 249-272. S. www.math.ucla.edu/~popa/preprints.html. S. Popa: Some rigidity results for non-commutative Bernoulli shifts, J. Funct. Anal. 230 (2006), 273-328. S. Popa: Strong rigidity of II1 factors arising from malleable actions of w-rigid groups. I., Invent. Math. 165 (2006), no. 2, 369-408. Correspondences, available (1986), INCREST preprint 56 at [Po86] [Po01] [Po03] [Po06a] S. Popa: On the superrigidity of malleable actions with spectral gap, J. Amer. Math. Soc. 21 (2008), 981- 1000. [Po06b] S. Popa: On Ozawa's property for free group factors, Int. Math. Res. Not. 2007, no. 11, Art. ID rnm036, 10 [Po07] pp. S. Popa: Deformation and rigidity for group actions and von Neumann algebras, In Proceedings of the In- ternational Congress of Mathematicians (Madrid, 2006), Vol. I, European Mathematical Society Publishing House, 2007, p. 445-477. [Po09a] S. Popa: On spectral gap rigidity and Connes invariant χ(M ), Proc. Amer. Math. Soc. 138 (2010), no. 10, 3531-3539. [Po09b] S. Popa: On the classification of inductive limits of II1 factors with spectral gap, Trans. Amer. Math. Soc. [Sa62] [Sa68] [Si10] 364 (2012), 2987-3000. S. Sakai: The Theory of W∗-Algebras, lecture notes, Yale University, 1962. S. Sakai: Asymptotically abelian II1-factors, Publ. Res. Inst. Math. Sci. Ser. A 4 1968/1969 299-307. T. Sinclair: Strong solidity of group factors from lattices in SO(n,1) and SU(n,1), J. Funct. Anal., 260 (2011), 3209-3221. [Th18] A. Thom: Finitary approximations of groups and their applications, Proc. Int. Cong. of Math. 2018 Rio de Janeiro, Vol. 2 (1775-1796). [Va10a] S. Vaes: Rigidity for von Neumann algebras and their invariants, Proceedings of the ICM (Hyderabad, India, 2010), Vol. III, Hindustan Book Agency (2010), 1624-1650. [Va10b] S. Vaes: One-cohomology and the uniqueness of the group measure space of a II1 factor, Math. Ann., 355 (2013), 661-696. [VDN92] D.-V. Voiculescu, K. J. Dykema, and A. Nica: Free Random Variables, CRM Monogr. Ser. vol.1, AMS, Providence, RI, 1992. F. B. Wright: A reduction for algebras of finite type, Ann. of Math. 60 (1944), 560-570. [Wr54] [ZM69] G. Zeller-Meier: Deux autres facteurs de type II1, (French) Invent. Math. 7 (1969) 235-242. Department of Mathematics, University of California San Diego, 9500 Gilman Drive, La Jolla, CA 92093, USA E-mail address: [email protected] Department of Mathematics, University of California San Diego, 9500 Gilman Drive, La Jolla, CA 92093, USA E-mail address: [email protected]
1911.02070	1	1911	2019-11-05T20:20:36	Fredholm conditions for invariant operators: finite abelian groups and boundary value problems	[ "math.OA", "math.AP", "math.FA" ]	We answer the question of when an invariant pseudodifferential operator is Fredholm on a fixed, given isotypical component. More precisely, let $\Gamma$ be a compact group acting on a smooth, compact, manifold $M$ without boundary and let $P \in \psi^m(M; E_0, E_1)$ be a $\Gamma$-invariant, classical, pseudodifferential operator acting between sections of two $\Gamma$-equivariant vector bundles $E_0$ and $E_1$. Let $\alpha$ be an irreducible representation of the group $\Gamma$. Then $P$ induces by restriction a map $\pi_\alpha(P) : H^s(M; E_0)_\alpha \to H^{s-m}(M; E_1)_\alpha$ between the $\alpha$-isotypical components of the corresponding Sobolev spaces of sections. We study in this paper conditions on the map $\pi_\alpha(P)$ to be Fredholm. It turns out that the discrete and non-discrete cases are quite different. Additionally, the discrete abelian case, which provides some of the most interesting applications, presents some special features and is much easier than the general case. We thus concentrate in this paper on the case when $\Gamma$ is finite abelian. We prove then that the restriction $\pi_\alpha(P)$ is Fredholm if, and only if, $P$ is "$\alpha$-elliptic", a condition defined in terms of the principal symbol of $P$. If $P$ is elliptic, then $P$ is also $\alpha$-elliptic, but the converse is not true in general. However, if $\Gamma$ acts freely on a dense open subset of $M$, then $P$ is $\alpha$-elliptic for the given fixed $\alpha$ if, and only if, it is elliptic. The proofs are based on the study of the structure of the algebra $\psi^{m}(M; E)^\Gamma$ of classical, $\Gamma$-invariant pseudodifferential operators acting on sections of the vector bundle $E \to M$ and of the structure of its restrictions to the isotypical components of $\Gamma$. These structures are described in terms of the isotropy groups of the action of the group $\Gamma$ on $E \to M$.	math.OA	math	FREDHOLM CONDITIONS FOR INVARIANT OPERATORS: FINITE ABELIAN GROUPS AND BOUNDARY VALUE PROBLEMS ALEXANDRE BALDARE, R´EMI C OME, MATTHIAS LESCH, AND VICTOR NISTOR We dedicate this paper to Professor Dan Voiculescu on the occasion of his 70th birthday Abstract. We answer the question of when an invariant pseudodifferential operator is Fredholm on a fixed, given isotypical component. More precisely, let Γ be a compact group acting on a smooth, compact, manifold M without boundary and let P ∈ ψm(M ; E0, E1) be a Γ-invariant, classical, pseudodif- ferential operator acting between sections of two Γ-equivariant vector bundles E0 and E1. Let α be an irreducible representation of the group Γ. Then P induces by restriction a map πα(P ) : H s(M ; E0)α → H s−m(M ; E1)α between the α-isotypical components of the corresponding Sobolev spaces of sections. We study in this paper conditions on the map πα(P ) to be Fredholm. It turns out that the discrete and non-discret cases are quite different. Additionally, the discrete abelian case, which provides some of the most interesting applica- tions, presents some special features and is much easier than the general case. Moreover, some results are true only in the abelian case. We thus concentrate in this paper on the case when Γ is finite abelian. We prove then that the restriction πα(P ) is Fredholm if, and only if, P is "α-elliptic," a condition defined in terms of the principal symbol of P (Definition 1.1). If P is elliptic, then P is also α-elliptic, but the converse is not true in general. However, if Γ acts freely on a dense open subset of M , then P is α-elliptic for the given fixed α if, and only if, it is elliptic. The proofs are based on the study of the structure of the algebra ψm(M ; E)Γ of classical, Γ-invariant pseudodifferential operators acting on sections of the vector bundle E → M and of the struc- ture of its restrictions to the isotypical components of Γ. These structures are described in terms of the isotropy groups of the action of the group Γ on E → M . Contents Introduction 1. 1.1. Γ-principal symbol and α-ellipticity 1.2. Statement of the main result 1.3. Contents of the paper 2. Preliminaries 2.1. Group representations 2.2. 2.3. The primitive ideal spectrum 2.4. Lie group actions on manifolds Induction and Frobenius reciprocity 2 3 4 5 6 6 7 11 12 M.L. was partially supported by the Hausdorff Center for Mathematics, Bonn. A.B., R.C., and V.N. have been partially supported by ANR-14-CE25-0012-01 (SINGSTAR). Manuscripts available from http://www.math.psu.edu/nistor/. 1 2 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR 2.5. Pseudodifferential operators 2.6. Reduction to order-zero operators 3. The structure of regularizing operators Inner actions of Γ: the abstract case 3.1. 3.2. Inner actions of Γ: pseudodifferential operators 3.3. The case of free actions 4. The principal symbol 4.1. The primitive ideal spectrum of the symbol algebra 4.2. Factoring out the minimal isotropy 5. Applications and extensions 5.1. Fredholm conditions 5.2. Boundary value problems 5.3. The case of non-discrete groups References 13 13 14 14 15 16 17 17 19 21 21 22 23 23 1. Introduction Fredholm operators have many applications to Mathematical Physics, to Partial Differential Equations (linear and non-linear), to Geometry, and to other areas of mathematics. They have their origin in the work of several mathematicians on spectral theory and on integral equations at the end of the nineteenths century. Fredholm operators are ubiquitous in applications since on a compact manifold, a classical pseudodifferential operator is Fredholm (acting between suitable Sobolev spaces) if, and only if, it is elliptic. In this paper, we obtain an analogous result for the restriction of an invariant, classical pseudodifferential operator to a fixed isotypical component for the action of a finite abelian group Γ. Of course, a Γ-invariant operator is Fredholm if, and only if, its restrictions to all isotypical components are Fredholm. Our result, focuses on one fixed given isotypical component. Namely, the restriction to the isotypical component corresponding to an irreducible representation α of Γ is Fredholm if, and only if, the operator is α-elliptic (Definition 1.1 and Theorem 1.2 below). The reasons we assume our group to be abelian and discrete are, first, that the main result is nolonger correct as stated in the non-discrete case and, second, that the general discrete case is quite different (and much more difficult) than the abelian discrete case. Moreover, some useful intermediate results are true only in the abelian case and this is the case needed for applications to boundary value problems. Although in the formulation of our main result we do not use C∗-algebras, for its proof, we have found it convenient to use them. Recently, there were quite a few papers using C∗-algebras to obtain Fredholm conditions, see, for instance, [17, 19, 30, 32] among many others. Often groupoids were also used [13, 20, 36, 44]. Fredholm conditions play an important role in Quantum Mechanics in the study of the essential spectrum of N -body Hamiltonians [6, 24, 23, 27, 29]. A powerful related technique is that of "limit operators" [28, 34, 35, 43]. Some of the most recent papers using similar ideas include [4, 12, 13, 14, 37, 38, 39, 51], to which we refer for further references. Besides C∗-algebras, pseudodifferential operators were also often used to obtain Fredholm conditions, see [18, 31, 26] and the references therein. FREDHOLM CONDITIONS 3 Let us now explain our main result in more detail. 1.1. Γ-principal symbol and α-ellipticity. In general, for any compact group representations), as usual. G, we let bG denote the set of equivalence classes of irreducible G-modules (or G-equivariant map of G-modules and α ∈ bG, we let It is a finite set if G is finite. If T : V → W is a (1) πα(T ) : Vα → Wα denote by the induced map obtained by restricting and corestricting T to the cor- responding isotypical component. Our main result, Theorem 1.2 is stated in terms of the Γ-principal symbol and α-ellipticity, two concepts which we now introduce. Let Γ be a finite abelian group acting on a smooth, compact, boundaryless man- ifold M and let P ∈ ψm(M ; E0, E1) be a Γ-invariant classical pseudodifferential operator acting between sections of two Γ-equivariant vector bundles E0 and E1. Let α ∈bΓ and consider as above (2) πα(P ) : H s(M ; E0)α → H s−m(M ; E1)α , acting between the α-isotypical components of the corresponding Sobolev spaces of sections. The main question that we answer in this paper is to determine when πα(P ) is Fredholm in terms of its principal symbol (3) σm(P ) ∈ Γ(T ∗M r {0}; Hom(E0, E1)) , regarded as a homogeneous function on the cotangent bundle of M . Note that for Γ abelian,bΓ consists of characters, that is, of group morphisms Γ → C∗. The reason for restricting to the case Γ finite abelian is that it presents some special features, although some, but not all, of our results extend to the case Γ finite (possibly non- abelian) and even to the case Γ compact. The case Γ abelian provides some of the most important applications and is much easier than the general case, so, for the sake of the clarity and brevity of the presentation, we will assume in our main result that Γ is abelian. Moreover, some of our results are not true in the non-abelian case, in general, and the main result (Theorem 1.2) is not true for non-discrete groups (Corollary 2.5 of [2]). For simplicity, we will consider only classical pseudodifferential operators in this article [33, 48, 50]. Recall that a classical pseudodifferential operator P is called elliptic if its principal symbol is invertible (away from the zero section). If P is elliptic, then it is Fredholm, and hence πα(P ) is also Fredholm. The converse is not true, however, in general. Indeed, we introduce, for any irreducible representation α of Γ, an "α-principal symbol" σα m(P ) (Definition 1.1) and prove that πα(P ) is Fredholm if, and only if, its α-principal symbol is invertible (in which case we call P α-elliptic, see Theorem 1.2 below for the precise statement). As we have just noticed, in general, the invertibility of the α-principal symbol does not imply ellipticity. To state these results in more detail, we need to introduce some notation and terminology. The Γ-invariance of P implies that its principal symbol is also Γ invariant: σm(P ) ∈ Γ(T ∗M r {0}; Hom(E0, E1))Γ . To study the space of morphisms Γ(T ∗M r {0}; Hom(E0, E1))Γ in which the prin- cipal symbol σm(P ) lives, let (4) Γξ := {γ ∈ Γ γξ = ξ} 4 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR denote the isotropy group of a ξ ∈ T ∗ x M in Γ, x ∈ M , as usual. The isotropy Γx of x ∈ M is defined similarly. Then the groups Γξ ⊂ Γx act on E0x and on E1x, the fibers of E0, E1 → M at x. If q ∈ Γ(T ∗M r {0}; Hom(E0, E1))Γ, then q(ξ) ∈ Hom(E0x, E1x)Γξ . As we will see below, there is no loss of generality for our main result to assume that E0 = E1 = E, in which case Hom(E0, E1) = End(E). Let us consider the space (5) XM,E,Γ := {(ξ, ρ) ξ ∈ T ∗M r {0}, ρ ∈bΓξ, and Hom(ρ, Ex)Γξ 6= 0} . Let also ρ ∈bΓξ be an irreducible representation of Γξ, then q(ξ, ρ) := πρ(q(ξ)) ∈ End(Exρ)Γξ (6) denotes the restriction of q(ξ) to the isotypical component corresponding to ρ, with πρ defined in Equation (1). Thus q is a function on XM,E,Γ. Applying this construction to σm(P ) ∈ Γ(T ∗M r {0}; End(E))Γ, we obtain the function (7) m(P ) := \σm(P ) : XM,E,Γ → σΓ End(Exρ)Γξ . [(x,ρ)∈XM,E,Γ That is, (8) σΓ m(P )(ξ, ρ) := πρ(σm(P )(ξ)) ∈ End(Exρ)Γξ , ξ ∈ T ∗ x M . The characterization of Fredholm operators can be reduced to each component of the manifold. We shall therefore assume for the rest of this Introduction and beginning with Subsection 4.2 that our manifold M is connected. This simplifies also the statements and the proof of our results. Let then Γ0 be a minimal isotropy group for the connected manifold M (see Subsection 2.4.3). The α-principal symbol σα m(P ), but in order to define it, we need an additional crucial ingredient that takes α into account. m(P ) of P , α ∈bΓ, is defined in terms of σΓ For α ∈bΓ and ρ ∈bΓξ, we will say that α and ρ are Γ0-disjoint if HomΓ0(ρ, α) = 0, otherwise, we will say that they are Γ0-associated. Let M,E,Γ := {(ζ, ρ) ∈ XM,E,Γ ρ and α are Γ0-associated } . X α (9) Let us assume for the rest of this introduction that Γ is abelian. Then we have that α ∈ bΓ and ρ ∈ bΓξ are Γ0-associated if, and only if, their restrictions to Γ0 coincide, that is, if αΓ0 = ρΓ0. We can now define the α-principal symbol σα of P . Definition 1.1. The α-principal symbol σα X α m(P ) of P is the restriction of σΓ m(P ) to m(P ) M,E,Γ: m(P ) := σΓ σα m(P )X α M,E,Γ . We shall say that P ∈ ψm(M ; E) is α-elliptic if its α-principal symbol σα m(P ) is invertible everywhere on its domain of definition. This definition extends right away to operators in ψm(M ; E0, E1). 1.2. Statement of the main result. Thus P is α-elliptic if, and only if, σΓ is invertible on X α M,E,Γ. We are ready now to state our main result. m(P ) Theorem 1.2. Let Γ be a finite abelian group acting on a smooth, compact mani- fold M and let P ∈ ψm(M ; E0, E1) be a classical pseudodifferential operator acting between sections of two Γ-equivariant bundles E0, E1 → M , m ∈ R, and α ∈ bΓ. We have that πα(P ) : H s(M ; E0)α → H s−m(M ; E1)α is Fredholm if, and only if P is α-elliptic. FREDHOLM CONDITIONS 5 If Γ acts without fixed points on a dense open subset of M , then Γ0 = 1, and M,E,Γ for all α ∈ bΓ. Hence, in this case, P is α-elliptic if, and hence XM,E,Γ = X α only if, it is elliptic. The ellipticity of P can thus be checked in this case simply by looking at a single isotypical component. We stress, however, that if Γ is not discrete, this statement, as well as the statement of our main result (Theorem 1.2 above), are not true anymore. However, our main result, as well as many intermediate results hold for general finite groups and some even for compact Lie groups. The extension of our main result to general finite groups is work in progress [5], but the proof seems to be much more involved. A motivation for our result comes from index theory. Let us assume that P is Γ-invariant and elliptic. Atiyah and Singer have determined, for any γ ∈ Γ, the value at γ of the character of indΓ(P ) ∈ R(G), that is they have computed chγ(indΓ(P )) ∈ C in terms of data at the fixed points of γ on M [3]. (Here R(G) := Z bG is the representation ring of G.) Bruning [9, 10] considered the "isotypical heat trace" tr(pαe−t∆), which is noth- ing but the heat trace of πα(∆), and its short time asymptotic expansion. Clearly, carrying out Bruning's programme in full would lead to a heat equation proof of an index theorem for the α isotypical component of Dirac type operators. However, the technical obstacles for this approach are enormous. 1.3. Contents of the paper. Let us quickly describe here the contents of our paper. We start in Section 2 with some preliminaries. We thus recall some facts about groups, most notably Frobenius reciprocity (for finite groups) and the defini- tions of induced representations, of minimal isotrypy groups (for connected M ) and of the principal orbit bundle. We also review some notions concerning the primi- tive ideal spectrum of C∗-algebras, as well as basic facts concerning (equivariant) pseudodifferential operators. In Section 3, we compute the image of the algebra ψ−1(M ; E) of regularizing operators via πα. We do this by proving some general results on the structure of C∗-algebras with an inner action of our group Γ. When the action of the group Γ is inner, the results and their proofs become simpler. Let AM := C0(M ; End(E)). The main difficulties arise in Section 4. There, we identify the primitive spectrum of the C∗-algebra AΓ M of Γ-invariant symbols with the set XM,E,Γ/Γ described above. Some care is taken to describe the corresponding topology on XM,E,Γ/Γ. We then consider the projection from AΓ M to the Calkin algebra of L2(M ; E)α and show that the closed subset of Prim AΓ M associated to its kernel is X α M,E,Γ/Γ. These results are used in Section 5 to prove the main result of the paper, Theorem 1.2. We also discuss an application to mixed boundary value problems and explain why our result is not true when the group Γ is not discrete, The last named author (V.N.) thanks Max Planck Institute for support while this research was performed. Since this paper is dedicated to Professor Dan Voiculescu, the last named author would like to mention that his papers, most notably [41, 42, 52], and [53] have played an important role in this author's formation as a mathematician in his early years. Moreover, we are happy to dedicate to Voiculescu this paper in which we prove a result that does not explicitly use C∗-algebras, but whose proof uses in an essential way the theory of these algebras. This was the spirit of interdisciplinarity that Voiculescu was promoting while he was a member of 6 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR INCREST, an institute that is now called the Institute of the Roumanian Academy of Sciences (IMAR). 2. Preliminaries We begin by setting up the terminology and the notation used in this paper. We also recall some basic results that are needed in the sequel. Throughout the paper, Γ will be a compact group acting on a locally compact space M . For the most part, M will be a smooth Riemannian manifold and Γ will be a compact Lie group acting smoothly and isometrically on M . The final result holds only for discrete (thus finite) groups and M compact, but many intermediate results hold in greater generality, so we have tried to state the results in the greatest generality possible when this did not involve too much extra work. In particular, we shall start with a compact group Γ acting on a possibly non-compact Riemannian manifold1 M . Eventually, we shall assume that M is compact and that Γ is discrete, hence finite. Moreover, since the case Γ abelian is simpler and presents some special features, for the final result we will assume that Γ is abelian. 2.1. Group representations. We follow the standard conventions, see [7, 47], to which we refer for further references and unexplained concepts. 2.1.1. Group actions on sets. Let us assume now that Γ is a group that acts on a set M . If x ∈ M , then Γx denotes the Γ orbit of x and Γx denotes the isotropy group (of M ) at x, that is (10) Γx := {γ ∈ Γ gx = x} ⊂ Γ . We shall write H ∼ H ′ if the subgroups H and H ′ are conjugated in Γ. If H ⊂ Γ is a subgroup, we shall denote by M(H) the set of elements of M whose isotropy Γx is conjugated to H (in Γ), i.e. H ∼ Γx. 2.1.2. Representations and isotypical components. Let V be a locally convex space and L(V ) denote the set of continuous linear maps V → V . Let Γ now be a compact topological group and ρ : Γ → L(Hρ) be a strongly continuous representation in a complete, locally convex topological vector space Hρ, in the sense that, for each ξ ∈ Hρ, the map Γ ∋ γ → ρ(γ)ξ ∈ Hρ is continuous. We shall say then that Hρ is a Γ-module and we shall often drop ρ from the notation, thus H = Hρ and γξ := ρ(γ)ξ. If Γ is a discrete group (as it will be the case for our final results), the continuity conditions will, of course, be automatically satisfied. For any two Γ-modules H and H1, we shall denote by HomΓ(H, H1) = Hom(H, H1)Γ = L(H, H1)Γ the set of continuous linear maps T : H → H1 that commute with the action of Γ, that is, T (γξ) = γT (ξ) for all ξ ∈ H and γ ∈ Γ. Finally, we denote by Γ the set of equivalence classes of irreducible unitary Γ- modules. We shall need the following terminology. Definition 2.1. Let H ⊂ Γ1 and H ⊂ Γ2 be compact groups. Two irreducible rep- resentations αi ∈bΓi, i = 1, 2, are called H-associated if L(α1, α2)H 6= 0. Otherwise, we shall say that they are H-disjoint. 1The metric on M is, however, for convenience only. FREDHOLM CONDITIONS 7 If Γi, i = 1, 2, are both abelian, then the irreducible representations αi are characters, that is, morphisms αi : Γi → C∗, and we have that they are associated if, and only if, α1H = α2H . 2.1.3. Isotypical component. Let H be a Γ-module and α ∈ Γ. Then pα will denote the Γ-invariant projection onto the α-isotypical component Hα of H, defined as the largest (closed) Γ submodule of H that is isomorphic to a multiple of α. In other words, Hα is the sum of all Γ-submodules of H that are isomorphic to α. Equivalently, since Γ is compact, we have Hα ≃ α ⊗ HomΓ(α, H); moreover (11) Hα 6= 0 ⇔ HomΓ(α, H) 6= 0 ⇔ HomΓ(H, α) 6= 0. If T ∈ L(H) commutes with the action of Γ (i.e. it is a Γ-module morphism), then T (Hα) ⊂ Hα and we denote by (12) πα : L(H)Γ → L(Hα) , πα(T ) := T Hα , the associated morphism and operator, as in Equation (1) of the Introduction. The morphism πα will play a crucial role in what follows. 2.1.4. Convolution algebras. The algebra C(Γ) of continuous functions on Γ, en- dowed with the convolution product, will act on any Γ-module H. If, moreover, H is a Hilbert space and Γ acts by unitary operators, we shall say then that H is a unitary Γ-module. We endow Γ with a fixed Haar measure and let C∗ r (Γ) be the completion of C(Γ) acting on L2(Γ), which is a unitary Γ-module. Then C∗ r (Γ) will act on any unitary Γ-module H, since Γ is compact (and hence amenable: C∗ r (Γ) = C∗(Γ)). For any algebra A, we shall denote by Z(A) the center of A, that is, the set of elements z ∈ A that commute with all other elements a ∈ A. An element z ∈ Z(A) will be called central (in A). For instance, if α ∈ bΓ, then its character defines a central projection pα ∈ Z(C(Γ)) ⊂ Z(C∗(Γ)). More explicitely, for any γ ∈ Γ we have that pα(γ) = χα(γ) := tr(α(γ)). Given a representation ρ of Γ on H, the image of pα is then ρ(pα) =ZΓ χα(g)ρ(g)dg, where integration is against the Haar measure. We are interested in this projection since Hα = pαH. 2.2. Induction and Frobenius reciprocity. We now review some basic defi- nitions and results for induced representations. We will use induction for finite groups only, so we assume in this discussion of the Frobenius reciprocity (i.e. in this subsection) that Γ is finite. 2.2.1. Definition of the induced module. Since we are assuming in this subsection that Γ is finite, we have that C∗(Γ) = C(Γ) = C[Γ], the group algebra of Γ. We will use the standard notation V (I) := {f : I → V }, valid for I finite. If H ⊂ Γ is a subgroup (hence also finite) and V is a H-module, we let (13) IndΓ H (V ) := C[Γ] ⊗C[H] V ≃ { ξ : Γ → V f (gh−1) = hf (g) } ≃ V (Γ/H) be the induced representation from V . The last isomorphism is obtained by choosing a set of representatives of the right cosets Γ/H. The action of Γ on IndΓ H (V ) is 8 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR by left multiplication on C[Γ], and the indicated isomorphism is an isomorphism of Γ-modules. The Γ-module IndΓ H (V ) depends functorially on V . H (V ) is an algebra for the pointwise product. Remark 2.2. If V is an algebra and the group H acts on V by algebra homomor- phisms, then the isomorphism IndΓ H (V ) ≃ {ξ : Γ → V f (gh−1) = hf (g)} shows that IndΓ If V1 is a left V -module (with a structure compatible with the action of Γ), then IndΓ H (V ) module, again with the pointwise multiplication. The induction is moreover com- patible with morphisms of modules and algebras (change of scalars), again by the function representation of the induced representation. In particular, if φ : V → W is a H-morphism of algebras, if V1 and W1 are modules over these algebras, and ψ : V1 → W1 is a H-module morphism such that ψ(ab) = φ(a)ψ(b), then, if mult denotes the multiplication map, the following diagram commutes: H (V1) is a IndΓ (14) indΓ H (V ) ⊗ indΓ H (V1) φ⊗ψ −−−−→ indΓ H (W ) ⊗ indΓ H (W1) multy indΓ H (V1) ymult indΓ H (W1) . ψ −−−−→ (All maps, including the multiplications, are assumed to be compatible with the action of Γ.) 2.2.2. Explicit isomorphisms. We will use the following form of the Frobenius reci- procity: the map Φ = ΦΓ,H H,V : HomH (H, V ) → HomΓ(H, IndΓ H (V )) , (15) Φ(f )(ξ) := g ⊗C[H] f (g−1ξ) , 1 H Xg∈Γ is an isomorphism (ξ ∈ H, f ∈ HomH(H, V )). This version of the Frobenius reciprocity is not valid in general, but is valid for finite groups [7, 47]. Often one writes HomH (ResΓ H(H), V ) instead of HomH (H, V ). Let α be an irreducible representation of Γ, H ⊂ Γ be a subgroup, and β an irreducible representation of H. Frobenius reciprocity gives, in particular, that the multiplicity of α in indΓ H(β) is the same as the multiplicity of β in the restriction of α to H. In particular, α is contained in indΓ H (β) if, and only if, β is contained in the restriction of α to H i.e. α and β are H-associated. Furthemore, by taking H to be the trivial Γ-module C, we obtain an isomorphism Φ : V H = HomH (C, V ) ≃ HomΓ(C, IndΓ H (V )) = IndΓ H (V )Γ , (16) Φ(ξ) := 1 H Xg∈Γ g ⊗C[H] ξ = Xx∈Γ/H x ⊗ ξ . The chosen normalization in the definition of Φ is such that it is an isomorphism of algebras if V is an algebra. 2.2.3. The abelian case. If Γ is abelian and V is an irreducible H-module, then the action of H on V is via scalars: h · v = χV (h)v, for some group morphism (i.e. character) χV : H → C∗. The induced module IndΓ H (V ) splits then as the direct sum of the irreducible Γ modules β that are H-associated to V . If χ is a character of Γ, we shall denote by Vχ the H-module equal to C as a vector space, with the action of h ∈ H given by h · v = χ(h)v. FREDHOLM CONDITIONS 9 Lemma 2.3. Assume that Γ is a finite abelian group and that H is a subgroup of Γ. Let V be an irreducible H-module corresponding to the character χV : H → C∗. Then IndΓ H (V ) ≃ Mχ∈Γ, χH =χV IndΓ H (V )χ. Moreover, by writing IndΓ H (V ) ≃ C[Γ/H] ⊗ V as vector spaces, we obtain an action of dΓ/H on IndΓ H (V ) by the formula ρ · (γH ⊗ v) := ρ(γH)γH ⊗ v, This action maps IndΓ H (V )χ to IndΓ H (V )ρχ. γ ∈ Γ, v ∈ V, and ρ ∈ dΓ/H . in IndΓ H (V ) if, and only if, χH = χV . If A is a finite abelian group, we have the Proof. Given χ ∈ bΓ, we have by the Frobenius isomorphism that χ is contained (non-canonical) isomorphism bA ≃ A of groups. There are [Γ/H = Γ/H-many characters χ of Γ with the property χH = χV . It follows, by counting dimensions, that they all appear with multiplicity one in IndΓ H (V ). The statement about the action of [Γ/H is proved by a direct computation. This proof is complete. (cid:3) 2.2.4. Inducing endomorphism modules. Let now α be an irreducible Γ-module, let βj, j = 1, . . . , N , be non-isomorphic irreducible H-modules (with H a subgroup of Γ, as above), and (17) β := ⊕N j=1βkj j . If H is abelian, then each βj is one dimensional, and hence H acts by scalars on each βkj j . We want to study the algebra IndΓ H(β) and on its isotypical component pα(IndΓ H (β)α (see 2.1.3 for the definition of the projection pα). We have, by the Frobenius isomorphism and by the form of the H-module β, that H (End(β))Γ acting on IndΓ H (β)) = IndΓ (18) IndΓ H (End(β))Γ ≃ End(β)H ≃ ⊕N j=1 End(βkj j )H ≃ ⊕N j=1Mkj (C), which is a semi-simple algebra. Moreover, we have that IndΓ where Φ is the map of Equation (16). From the properties of the induction functor IndΓ H (End(β))Γ = Φ(End(β)H ), H , we also have that IndΓ H (β) = ⊕N j=1 IndΓ H (βkj j ). Lemma 2.4. Let β := ⊕N j=1βkj j be as in Equation (17), let T = (Tj) ∈ End(β)H ≃ ⊕N j=1 End(βkj j )H , with Tj ∈ End(βkj j )H , and let ξj ∈ IndΓ H (βkj ξ := (ξj ) ∈ ⊕N j=1 IndΓ j ). We let H (βkj j ) ≃ IndΓ H (β) . Then Φ(T )(ξ) = (Φ(Tj)ξj )j=1,...,N . In other words, the Frobenius isomorphism Φ of Equation (16) is compatible with direct sums and with the action of morphisms on modules. 10 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR Proof. This follows from the naturality of the product, the isomorphism (19) IndΓ H (⊕N j=1 End(βkj j ))Γ ∼−→ IndΓ H (End(β))Γ , and Remark 2.2 (especially Equation (14)). (cid:3) Put differently, the simple factor of the algebra IndΓ H (End(β))Γ corresponding to IndΓ i=1 IndΓ H (β). The following proposition will play a crucial role in what follows. j ))Γ acts only on the jth component of ⊕N H (End(βkj i ) = IndΓ H (βki Proposition 2.5. Let β := ⊕N j be as in Equation (17). Let J ⊂ {1, 2, . . . , N } be the set of indices j such that αH = βj (i.e. α and βj are H-associated). Then the morphism j=1βkj πα : IndΓ H (End(β))Γ ≃ ⊕N j=1 IndΓ H (End(βkj j ))Γ → End(pα IndΓ H (β))Γ is such that we have natural isomorphisms ker(πα) ≃ ⊕j /∈J IndΓ H (End(βkj j ))Γ and Im(πα) ≃ ⊕j∈J IndΓ H (End(βkj j ))Γ . Proof. By Lemma 2.4, we can assume that N = 1. By Lemma 2.3, IndΓ ⊕χ∈bΓVχ, with χH = β1. We obtain H (βk1 1 ) ≃ (20) pα IndΓ H (βk1 1 ) ≃ ( αk1 0 if αH = β1 otherwise. Thus, if αH 6= β1, J = ∅, πα = 0, and hence ker(πα) = IndΓ H (End(βk1 1 ))Γ = ⊕j /∈J IndΓ j ))Γ and Im(πα) = 0 = ⊕j∈J IndΓ H (End(βkj H (End(βkj j ))Γ , as claimed. Let us assume now that that αH = β1. The morphism πα can then be written as the composition πα : IndΓ H (End(βk1 1 ))Γ ≃ End(βk1 1 )H ≃ Mk1 (C) Ψα−→ Mk1(C) ≃ End(αk1 )Γ ≃ End(pα IndΓ H (βk1 1 ))Γ . To prove our proposition in this case (χH = β1, and hence in general, by Lemma 2.4), it thus suffices to show that Ψα = id, which equivalent to the fact that Ψα 6= 0 for αH = β1. H (End(β)) → End(IndΓ To prove this (that Ψα 6= 0 for αH = β1), let us begin by noticing that the morphism IndΓ H (End(β))Γ → End(IndΓ H (β))Γ is injective as well. This means that not all of the maps Ψρ, with ρH = β1, can be zero, by Lemma 2.3. On the other hand, it can be checked by a direct calculation that the action of [Γ/H on IndΓ H (β) of the same lemma permutes the morphisms Ψρ. It follows either that they are all zero or that they are all non-zero. As their direct sum is non-zero, it follows that all Ψρ 6= 0, ρH = β1. (cid:3) H (β)) is injective. Hence IndΓ FREDHOLM CONDITIONS 11 2.3. The primitive ideal spectrum. Let us begin by recalling a few basic facts about C∗-algebras [21]. Recall that a two-sided ideal I ⊂ A of a C∗-algebra A is called primitive if it is the kernel of an irreducible, non-zero ∗-representation of A (so A is not considered to be a primitive ideal of itself). We will denote by Prim(A) the primitive ideal spectrum of A, which, we recall, is defined as the set of primitive ideals of A. For instance, if A = C0(X), the space of continuous functions X → C vanishing at infinity on the locally compact space, then we have a natural identification (homeomorphism) (21) Prim(C0(X)) ≃ X , which is an identification that lies at the heart of non-commutative geometry [15, 16]. All the C∗-algebras considered in this paper are type I. The relevance of this fact, for us, is that, if A is type I, then Prim(A) identifies with the set of isomorphism classes of irreducible representations of A. Any C∗-algebra with only finite dimensional representations is a type I algebra [21]. All the algebras con- sidered in this paper (except the algebras of compact operators on various Hilbert spaces), have this property. j=1βkj Remark 2.6. The following example will be used several times. Let H be a finite group and β = ⊕N j be a finite dimensional H-module with βj non-isomorphic simple H-modules. Since HomH (βkj j )H = 0 for i 6= j by the assumption that the simple H-modules βi and βj are non-isomorphic, we obtain j ) ≃ Mkj (C) and Hom(βki j , βkj i , βkj L(β)H = EndH (β) ≃ HomH (⊕iβki i , ⊕jβkj j ) ≃ ⊕i,j HomH (βki ≃ ⊕j EndH (βkj i , βkj j ) j ) ≃ ⊕jMkj (C) . The algebra L(β)H = EndH (β) is thus a C∗-algebra with only finite dimensional representations and we have natural bijections Prim(EndH (β)) ↔ {β1, β2, . . . , βN } ↔ {1, 2, . . . , N } . The algebra L(β)H is thus a semi-simple complex algebra with simple factors EndH (βkj j ) ≃ Mkj (C), j = 1, 2, . . . , N . We shall need the connection between ideals and the topology of the primitive ideal spectrum, which we now recall. Let J be a closed, two-sided ideal of A. Then (22) Prim(J) = {I ∈ Prim(A) J 6⊂ I} and Prim(A/J) ∼= {I ∈ Prim(A) J ⊂ I} = Prim(J)c . The second bijection sends an ideal I ⊃ J to π−1 A,J (I), where πA,J : A → A/J is the canonical bijection. In the following, we shall identify the two sets using this bijection. In analogy with the Zariski topology, Prim(A) is endowed with the Jacobson (or hull-kernel) topology, which is the topology whose open sets are those of the form Prim(J). Conversely, given an open subset U ⊂ Prim(A), then U = Prim(JU ), where JU := ∩P∈Prim(A)rU P. For us, it will be convenient to regard these proper- ties as a one-to-one correspondence between the closed, two-sided ideals of A and the closed subsets Prim(A/J) = Prim(A) r Prim(J) of Prim(A). If I ⊂ A is an ideal such that a ∈ A and aI = 0 implies a = 0, then I is called an essential ideal of A and its associated open set Prim(J) is dense in Prim(A). 12 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR Let Z be a commutative C∗-algebra and φ : Z → M (A) be a -morphism to the multiplier algebra of A [1, 11]. Assume that φ(Z) commutes with A and φ(Z)A = A. Then Schur's lemma gives that there exists a natural continuous map φ∗ : Prim(A) → Prim(Z), which we shall call also the "central character map" (associated to φ). 2.4. Lie group actions on manifolds. We will assume from now on that Γ is a compact Lie group, that M is a Riemannian manifold, and that Γ acts smoothly and isometrically on M . 2.4.1. Slices and tubes. Let us fix x ∈ M arbitrarily. Then its orbit Γx is a smooth, compact submanifold of M . Let Nx be the orthogonal of Tx(Γx) in TxM . (If Γ is discrete, as in our main results, then, of course, Nx = TxM .) Then the isotropy group Γx acts linearly and isometrically on Nx. For r > 0, let Ux := (Nx)r denote the set of vectors of length < r in Nx. It is known then that, for r > 0 small enough, the exponential map gives a Γ-equivariant isometric diffeomorphism (23) Wx ≃ Γ ×Γx Ux := {(γ, y) ∈ Γ × Ux (γh, y) ≡ (γ, hy), h ∈ Γx} , where Wx is a Γ-invariant neighborhood of x in M . More precisely, Wx is the set of y ∈ M at distance < r to the orbit Γx, if r > 0 is small enough. The set Wx is called a tube around x (or Γx) and the set Ux is called the slice at x. The range of r depends, of course, on x, namely it must satisfy 0 < r < rx, for some rx > 0. We assume that, for each x ∈ M , such an r ∈ (0, rx) has been chosen and we will keep fixed the notation for the slice Ux and for the tube Wx associated to x and to the fixed choice of r. 2.4.2. Equivariant vector bundles. Let E → M be a Γ-equivariant smooth vector bundle. All our vector bundles will be assumed to be finite dimensional. Let us fix x ∈ M and consider the tube Wx ≃ Γ ×Γx Ux around x of Equation (23). We use this diffeomorphism to identify Ux with a subset of M , in which case, we can also asume the restriction of E to the slice Ux to be trivial. More precisely, (24) EUx ≃ Ux × β and EWx ≃ Γ ×Γx (Ux × β) , for some Γx-module β, the second isomorphism being Γ-equivariant. Let F → Y be a Hermitian vector bundle on a locally compact measure space. Then L2(Y ; F ) denotes the set of (equivalence classes) of square integrable sec- tions of the vector bundle F and C0(Y ; F ) denote the set of its continuous sections vanishing at infinity. 2.4.3. The principal orbit bundle. Let us assume from now on that M is connected, except where explicitly stated otherwise. The reason of this assumption is that it is known then [49] that there exists a minimal isotropy subgroup Γ0 ⊂ Γ, in the sense that M(Γ0) is a dense open subset of M . (Recall that M(H) denotes the set of points of M whose stabilizer is conjugated in Γ to H.) In particular, there exist minimal elements for inclusion for the set of isotropy groups of points in M and all minimal isotropy groups are conjugated (to a fixed subgroup, denoted Γ0 in what follows). The notation Γ0 will remain fixed from now on. By the definition, the set M(Γ0) consists of the points whose stabilizer is conjugated to that minimal subgroup. The set M(Γ0) is called the principal orbit bundle of M . FREDHOLM CONDITIONS 13 If x ∈ M(Γ0), then Γx acts trivially on the slice Ux at x, by the minimality of Γ0. (See 2.4.1 for notation.) In general, for x ∈ M , the isotropy of Γx acting on Ux will contain a subgroup conjugated to Γ0. If Γ is abelian, there is only one minimal isotropy group Γ0 (recall that we are assuming M to be connected). Moreover, we can then factor the action of Γ to an action of Γ/Γ0 on M , which has trivial minimal isotropy, that is, it is free on a dense, open subset of M . 2.5. Pseudodifferential operators. We continue to assume that Γ is a compact Lie group that acts smoothly and isometrically on a smooth Riemannian manifold M . We let ψm(M ; E) denote the space of order m, classical pseudodifferential operators on M with compactly supported distribution kernel. Recall that, in this article, we consider only classical pseudodifferential operators. We let ψ0(M ; E) and ψ−1(M ; E) denote the norm closures of ψ0(M ; E) and ψ−1(M ; E), respectively. The action of Γ then extends to an action on E and on ψm(M ; E), ψ0(M ; E), and ψ−1(M ; E). We will denote by K(H) the algebra of compact operators acting on a Hilbert space H. We will write K instead of K(H) when the Hilbert spaces H is clear from the context. We have (25) ψ−1(M ; E) = K(L2(M ; E)) , since we have considered only pseudodifferential operators with compactly sup- ported distribution kernels. Let S∗M denote the unit cosphere bundle of M , that is, the set of unit vectors in T ∗M , as usual. We will denote, as usual, by C0(S∗M ; End(E)) the set of continuous sections of the lift of the vector bundle End(E) → M to S∗M . Corollary 2.7. We have an exact sequence 0 → KΓ = ψ−1(M ; E)Γ → ψ0(M ; E)Γ σ0 −−→ C0(S∗M ; End(E))Γ → 0 . Proof. Recall then that the principal symbol extends to a continuous, surjective, Γ-equivariant map (26) σ0 = σM 0 : ψ0(M ; E) → C0(S∗M ; End(E)) with kernel ψ−1(M ; E) = K = K(L2(M ; E)). In other words, we have the following well known exact sequence 0 → K → ψ0(M ; E) σ0 −−→ C0(S∗M ; End(E)) → 0. The exact sequence of the corollary is obtained from the fact that the functor H → HΓ is exact on the category of Γ -- modules with continuous Γ-action, since Γ is compact. (cid:3) 2.6. Reduction to order-zero operators. We assume here that M is a compact Riemannian manifold with metric g. Let us fix a Γ-invariant metric on E → M . Let ∇ be a metric preserving, Γ-invariant connection on E. In what follows ∆ := ∆E g = ∇∗∇ will denote the (positive) Laplacian on M with coefficients in E. Since M is compact, the Sobolev space H s(M ; E) can be defined for any s ∈ R as the domain of (1 + ∆)s/2 and (1 + ∆)s/2 : H s(M ; E) → L2(M ; E) is an isomorphism. Lemma 2.8. Assume that M is compact. We have that a bounded operator P : H s(M ; E) → H s−m(M ; E) is Fredholm if, and only if, eP := (1 + ∆)(s−m)/2P (1 + ∆)−s/2 : L2(M ; E) → L2(M ; E) is Fredholm. Moreover, if P is the limit in 14 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR ψ0(M ; E). L(H s(M ; E), H s−m(M ; E)) of a sequence of operators Pn ∈ ψm(M ; E), then eP ∈ Proof. The first part follows from the fact (1 + ∆)s : H m(M ; E) → H m−2s(M ; E) is an isomorphism for all m, s ∈ R, since M is compact. The second part follows from the well known fact that the powers of the Laplace operator are classical pseudodifferential operators [46] and the continuity of the multiplication in the operator norm. (cid:3) The case of operators in ψm(M ; E, F ) acting between two vector bundles E, F → M can be reduced to the case of a single vector bundle by considering ψm(M ; E⊕F ). Moreover, P ∈ ψ (M ; E ⊕ F ) will be Fredholm if, and only if, (M ; E, F ) ⊂ ψ 0 0 (cid:20) 0 P ∗ P 0 (cid:21) ∈ ψ 0 (M ; E ⊕ F ) is Fredholm (here all operators act on L2-spaces). Therefore it is sufficient to consider only the case of order-zero operators acting on a single vector bundle. 3. The structure of regularizing operators We continue to assume that M is a complete Riemannian manifold and that Γ is a compact Lie group acting by isometries on M . From now on, all our vector bundles will be Γ-equivariant vector bundles. As explained in the Introduction, we want to identify the structure of the restric- tions of Γ-invariant pseudodifferential operators on M to the isotypical components of L2(M ; E). Let πα be this restriction morphism to the α-isotypical component. More precisely, we want to understand the structure of the algebra πα(ψ0(M ; E)Γ), for any fixed α ∈ Γ. See Equations (1) and (12) for the definition of the restriction morphism πα and of the projectors pα ∈ C∗(Γ). In this section, we study two basic cases: that of inner actions and that of free actions (of Γ). 3.1. Inner actions of Γ: the abstract case. In this subsection we deal with the case when the action of Γ is implemented by unitaries in the multiplier algebra (the case of inner actions). This allows us, in particular, to settle the case of regularizing operators. We shall need the following notion of direct sum. Recall that M (A) denotes the multiplier algebra of a C∗-algebra A. Recall the following standard definition. Definition 3.1. Let φ : Γ → Aut(A) be the action of a group Γ by automorphisms on a C∗-algebra A. We shall say that this action is inner if the morphism φ lifts to a morphism ψ : Γ → U (M (A)) to the group of unitary elements of M (A) such that φg(a) = ψ(g)aψ(g)−1 , a ∈ A, g ∈ Γ . Remark 3.2. If α : Γ → Aut(A) is an inner action as in Definition 3.1, then we obtain, in particular, a morphism C∗(Γ) → M (A). If, moreover, π : A → L(H) is a ∗-representation, then it extends to a representation of M (A). This induces natu- rally a unitary representation of Γ on H. This representation is uniquely determined if π is non-degenerate (i.e. if π(A)H is dense in H). FREDHOLM CONDITIONS 15 If An, n ≥ 1 is a sequence of C∗-algebras, we shall denote by c0 − ⊕∞ n=1An the n=1An. This definition extends immediately to countable inductive limit limN→∞ ⊕N irreducible unitary representations of Γ. We shall need then the following general result. families of C∗-algebras. Recall that bΓ denotes the set of isomorphism classes of let pα ∈ C∗(Γ) denote the central projector corresponding to α ∈bΓ. Let AΓ := {a ∈ Proposition 3.3. Let A be a C∗-algebra with a -morphism C∗(Γ) → M (A) and (27) If I ⊂ A is a closed two-sided ideal, then C∗(Γ)I ⊂ I and hence we obtain - morphisms C∗(Γ) → M (I) and C∗(Γ) → M (A/I) such that the induced sequence A af = f a, f ∈ C∗(Γ)}. Then AΓ ≃ c0 − ⊕α∈Γ pαAΓ . 0 → pαI Γ → pαAΓ → pα(A/I)Γ → 0 is exact for any α ∈bΓ. Proof. Let us arrange the elements of bΓ in a sequence ρn, n ≥ 1. Then, for any a ∈ A, limN→∞PN n=1 pρn a = a. Since, for any a ∈ AΓ, we have pαa = apα, the isomorphism AΓ ≃ c0 − ⊕α∈Γ AΓpα follows. Finally, it is known that if I is a two-sided ideal of A, then M (A)I ⊂ I. (cid:3) If A ⊂ L(H) is a sub-C∗-algebra of the algebra of bounded operators on a Hilbert space H together with a compatible Γ-module structure on H, we let πα : AΓ → L(Hα) be the restriction morphism to the α-isotypical component, as before, see Equations (1) and (12). Proposition 3.4. In addition to the assumptions of Proposition 3.3, let us suppose πα : AΓ → L(Hα) restricts to an isomorphism pαAΓ → πα(AΓ). Proof. Let us recall that, as explained in Remark 3.2, the group Γ will be repre- sented on H. The rest follows from Proposition 3.3, whose notation we shall use that A ⊂ L(H), for some Hilbert space H. Let α ∈bΓ, as before. Then the morphism freely. Indeed, if α 6= β ∈ bΓ, then πα(pβ) = 0. On the other hand πα(pα) = pα. The result follows by combining this property with Hα = pαH and with Proposition 3.3. (cid:3) 3.2. Inner actions of Γ: pseudodifferential operators. We now apply the results of the previous subsection to the algebra of pseudodifferential operators. In particular, this gives a rather complete picture for the case of negative order operators (recall that the closure of regularizing operators coincides with that of negative order operators). Since we are eventually interested only in the case Γ finite, we discuss only briefly the issues related to the non-discrete case (such as the continuity of the action of Γ). A crucial first observation is that if γ ∈ Γ and P ∈ ψ−∞(M ; E), then γP, P γ ∈ ψ−∞(M ; E). This leads to several interesting consequences. We record this as a lemma. Lemma 3.5. We have γψ−∞(M ; E) = ψ−∞(M ; E)γ = ψ−∞(M ; E), for all γ ∈ Γ, and the induced actions (to the right and to the left) of Γ on ψ−∞(M ; E) are continuous and unitary in the operator norm. Consequently, we have C∗(Γ)ψ−1(M ; E) + ψ−1(M ; E)C∗(Γ) ⊂ ψ−1(M ; E) . 16 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR Proof. Since smoothing operators have smooth kernels k(x, y) ∈ Hom(Ey, Ex), the action of Γ induced an action on ψ−∞(M ; E). Thanks to the Γ-invariance of the metric on M , we have kγP k = kP γk = kP k in the L2-operator norm, for any P ∈ ψ−∞(M ; E). The second part follows from the first part. (cid:3) We then have the following: Proposition 3.6. The multiplication by Γ on ψ−1(M ; E) defines a -morphism C∗(Γ) → U (M (ψ−1(M ; E))) to the multiplier algebra of ψ−1(M ; E). Proof. This follows from Lemma 3.5. We obtain the following corollary. (cid:3) Corollary 3.7. Let A = ψ−1(M ; E). If Γ acts trivially on M (so that the action of C∗(Γ) extends to the algebra ψ0(M ; E)), we also allow A = ψ0(M ; E) or A = C0(S∗M ; End(E)). We then have isomorphisms AΓ ≃ c0 − ⊕α∈bΓ pαAΓ . Moreover, if Γ acts trivially on M and α ∈ Γ, we have an exact sequence 0 → pαψ−1(M ; E)Γ → pαψ0(M ; E)Γ → C0(S∗M ; End(pαE)Γ) → 0 . Proof. The first part follows from Proposition 3.3 applied to the algebra ψ0(M ; E) and to its ideal K. The second part follows from the exactness of the functors V → V Γ and V → pαV on the category of Γ-modules. (cid:3) Let α ∈ Γ and let πα be the representation of ψ0(M ; E)Γ on L2(M ; E)α defined by restriction as before, Equations (1) and (12). The assumptions of Proposition 3.4 are satisfied for A = ψ−1(M ; E), so we obtain the following. Corollary 3.8. The morphism πα restricts to an isomorphism from pαψ−1(M ; E)Γ to πα(ψ−1(M ; E)Γ). We also have the following simple result, which makes the last corollary more precise. Recall that K ≃ ψ−1(M ; E). This allows us to better describe the structure of ψ−1(M ; E)Γ. Proposition 3.9. The algebra πα(KΓ) is the algebra of Γ -- equivariant compact operators on L2(M ; E)α. Proof. Let T ∈ K commute with Γ. Then its restriction to a Γ-invariant subspace is still compact and still commutes with Γ. This shows that πα(KΓ) is contained in the set Kα of Γ-invariant compact operators on L2(M ; E)α. Conversely, Kα ⊂ KΓ and πα acts as the identity on Kα. (cid:3) 3.3. The case of free actions. Let us now tackle the opposite case, that is when Γ acts freely on M . We shall assume that Γ is finite, for simplicity (we only need this case), and hence, in particular, the action of Γ is proper. We have then the following well-known result (see [16, 40] and the references therein). Proposition 3.10. Let us assume that Γ is a finite group acting freely on M and let α(γ) = 1 be the trivial representation, so πα = π1. Let us denote by F := E/Γ → M/Γ the resulting vector bundle and φ : C0(S∗M ; End(E))Γ → FREDHOLM CONDITIONS 17 C0(S∗M/Γ; End(E/Γ)) the resulting isomorphism. Then we have the following mor- phism of exact sequences, with the vertical arrows surjective. 0 0 / ψ−1(M ; E)Γ ψ0(M ; E)Γ C0(S∗M ; End(E))Γ π1 π1 φ / ψ−1(M/Γ, F ) / ψ−0(M/Γ, F ) / C0(S∗M/Γ; End(F )) 0 / 0 For G ⊂ M , let AG := C0(S∗G; End(E)) and consider the surjective map (28) RG : AΓ G := C0(S∗G; End(E))Γ ≃ ψ0(G; E)Γ/ψ−1(G; E)Γ → πα(ψ0(G; E)Γ)/πα(ψ−1(G; E)Γ) . Proposition 3.11. Let Γ be a finite group acting on a smooth compact mani- fold M (without boundary). Assume that the action of Γ is free on a dense, open subset of M . Let E → M be an equivariant vector bundle. Then the map RM : C0(S∗M ; End(E))Γ → π1(ψ0(M ; E)Γ)/π1(ψ−1(G; E)Γ) of Equation (28) is injective, and hence an isomorphism of algebras. Proof. Let M0 ⊂ M be an open, dense subset on which Γ acts freely. Proposition 3.10 for M replaced with M0 shows that RM0 is injective. Hence RM is injective on C0(S∗M0; End(E))Γ, because the restriction of RM to C0(S∗M0; End(E))Γ is RM0 . Since the later is an essential ideal in C0(S∗M ; End(E))Γ, it follows that RM is also injective (everywhere on C0(S∗M ; End(E))Γ). (cid:3) 4. The principal symbol Let us fix an irreducible representation α of Γ and consider the fundamental restriction morphism πα of Equation (1). See also Subsection 2.1.2, especially Equation (12), for more details on the morphism πα. We are mostly concerned with the morphism πα : ψ0(M ; E)Γ → L(L2(M ; E)α) and, in this section, we identify the quotient πα(ψ0(M ; E)Γ)/πα(ψ−1(M ; E)Γ) . Since πα(ψ−1(M ; E)Γ) was identified in the previous section, information on the above quotient algebra will give further insight into the structure of the algebra πα(ψ0(M ; E)Γ) and will provide us, eventually, with Fredholm conditions. In the beginning of this section, we continue to assume that M is a complete Riemannian manifold and that Γ is a compact Lie group acting by isometries on M . Since the results are different in the discrete and non-discrete case, we will assume beginning with Subsection 4.2 that Γ is finite. Moreover, for the main result, we shall assume that Γ is abelian, since the abelian case is simpler and presents some additional features. Some intermediate results are true only in the abelian case. It is also the abelian case that will be used for our application to boundary value problems. 4.1. The primitive ideal spectrum of the symbol algebra. We now turn to the description of the primitive ideal spectrum of the algebra C0(S∗M ; End(E))Γ of symbols. For simplicity, for O ⊂ M open, we denote again AO := C0(S∗O; End(E)), as in the definition of the morphism RO of Equation (28). We shall be mostly / / / / / / / / / / / 18 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR concerned with the cases O = M and O = O0 := M(Γ0). We have the following standard result. Proposition 4.1. The algebra ZM := C0(S∗M )Γ = C0(S∗M/Γ) identifies with a M := C0(S∗M ; End(E))Γ. Let zξ be the maximal ideal of ZM central subalgebra of AΓ associated to the orbit Γξ for some ξ ∈ S∗ xM , let Ex be a fiber of E corresponding to ξ, and let Ex ≃ ⊕N be its decomposition into Γξ-isotypical components, with M ≃ End(Ex)Γξ is a semi- βj simple, non-isomorphic Γξ modules. Then AΓ simple, finite-dimensional (complex) algebra with N simple factors EndΓξ (βkj j ) ≃ Mkj (C), j ∈ {1, 2, . . . , N }. M /zξAΓ j=1βkj j Proof. We have C0(M ) ⊂ C0(M ; End(E)) ⊂ End(C0(M ; E)), with f ∈ C0(M ) act- In fact, this identifies C0(M ) with the center ing as a scalar on each fiber Ex. Z(C0(M ; End(E))) of C0(M ; End(E)). By considering Γ invariant functions, we obtain that ZM := C0(S∗M )Γ = C0(S∗M/Γ) is contained in the center Z(AΓ M ) of AΓ M := C0(S∗M ; End(E))Γ. Let Γξ denote the orbit in S∗M that we consider and let J be the (non-maximal, in general) ideal of C0(S∗M ) corresponding to func- tions vanishing on this orbit. Then J is Γ invariant and J Γ = zξ. By taking the Γ invariants in the exact sequence 0 → JAM → AM → AM /JAM → 0 and using Frobenius reciprocity for AM /JAM ≃ IndΓ Γξ (End(Ex)), we obtain that (29) AΓ M /zξAΓ M ≃ (AM /JAM )Γ ≃ End(Ex)Γξ . The proof is completed using Remark 2.6 for H = Γξ. (cid:3) See [8, 22, 25, 45, 54] for similar results. Recall that if φ : Z → M (A) is a central *-morphism (i.e. φ(z)a = aφ(z), for a ∈ A and z ∈ Z) such that φ(Z)A = A, then it defines a natural "central character" map φ∗ : Prim(A) → Prim(Z) by Schur's Lemma. The same proof yields the following. Corollary 4.2. There is a one-to-one correspondence between the primitive ideals of AΓ xM M := C0(S∗M ; End(E))Γ and the Γ-orbits of the pairs (ξ, ρ), where ξ ∈ S∗ and ρ ∈ bΓξ appears in Ex (i.e. HomΓξ (ρ, Ex) 6= 0). The group Γ acts by joint conjugation on both ξ and ρ. The inclusion ZM := C0(S∗M )Γ → AΓ the associated canonical central character map of spectra M is such that Prim(AΓ M ) → Prim(ZM ) = S∗M/Γ is continuous, finite-to-one, and maps the orbit Γ(ξ, ρ) to the orbit Γξ. M := C0(S∗M ; End(E))Γ. Then π Proof. Let π be an irreducible representation of AΓ is a multiple of a character on ZM := C0(S∗M )Γ ⊂ Z(AΓ M ), by Schur's lemma. Let this character correspond to the orbit Γξ ∈ S∗M/Γ, with ξ ∈ S∗ xM and denote by zξ the corresponding maximal ideal of ZM , as in the proof of the last lemma. In other words, zξ is the value of the central character map corresponding to the inclusion ZM ⊂ AΓ M applied to π. Then π factors out through an irreducible representation of AΓ j with βj non-isomorphic simple Γξ modules, as in the statement of Proposition 4.1. Then End(Ex)Γξ ≃ ⊕N j ), a direct sum of simple algebras. Thus π factors through one of the simple algebras EndΓξ (βkj j ) This associates to π the pair (ξ, ρ) = (ξ, βj), as desired. This pair is not unique, but depends on the choice of ξ. It becomes unique modulo the action of Γ, however. Conversely, given such a pair (ξ, ρ), we obtain an M ≃ End(Ex)Γξ . Let us write Ex ≃ ⊕N j=1 EndΓξ (βkj M /zξAΓ j=1βkj FREDHOLM CONDITIONS 19 irreducible representation of AΓ order. The first part of the result follows. M following exactly the same procedure in reverse To prove that φ∗ is finite to one, we notice that, by construction, φ∗(ξ, ρ) = ξ. Since only a finite number of (isomorphism classes of) simple Γξ modules appears in Ex, the finiteness follows. (cid:3) Remark 4.3. Let us denote by XM,E,Γ the set of pairs (ξ, ρ), where ξ ∈ T ∗ x M \ {0} and ρ ∈bΓξ appears in Ex (i.e. HomΓξ (ρ, Ex) 6= 0), as in the statement of Corollary 4.2. The main result of that corollary is a natural bijection XM,E,Γ/Γ ≃ Prim(AΓ M ) . (30) This bijection can be explicitely described as follows: to an orbit Γ(ξ, ρ) in XM,E,Γ is associated ker(πξ,ρ) in Prim AΓ M we define πξ,ρ(f ) as the restriction of f (ξ) to the ρ-isotypical component of Ex. M , where for any f ∈ AΓ 4.2. Factoring out the minimal isotropy. Recall that we are assuming M to be connected; in that case there is a minimal isotropy type for the action of Γ. We shall also assume from now on that Γ is abelian, for the reasons discussed in the Introduction. In particular, it is the case needed for our applications to boundary value problems and, moreover, some results are not true in the non-abelian case. Let α ∈ bΓ as before, and recall that we want to determine the structure of the quotient πα(ψ0(M ; E)Γ)/πα(ψ−1(M ; E)Γ) of the restricted algebras to the α- isotypical component. To this end, recall the morphism RM : AΓ M := C0(S∗M ; End(E))Γ → πα(ψ0(M ; E)Γ)/πα(ψ−1(M ; E)Γ) of Equation (28). The main result of this subsection is to determine the kernel of this morphism. The main reason why the abelian case is simpler than the general case is that in the abelian case all minimal isotropy subgroups of Γ acting on M coincide. The (unique) minimal isotropy subgroup of Γ acting on M will be denoted by Γ0, as before. Recall that the set O0 := M(Γ0) of points x ∈ M with isotropy Γx = Γ0 is called the principal orbit bundle of M ; it is a dense, open subset of M . For every (other) x ∈ M , we have Γ0 ⊂ Γx. We obtain that the group Γ0 acts trivially on M . Moreover, there exists a unitary group morphism (representation) Γ0 → End(E) that implements the action of Γ0 on ψ0(M ; E). Let p(0) irreducible representations of Γ0 (the additional exponent is to differentiate them β ∈ C∗(Γ0), β ∈bΓ0, be the central projectors associated to the from the projectors pα, α ∈bΓ). Corollary 3.7 then gives the exact sequence β ψ0(M ; E)Γ0 → C0(S∗M ; End(p(0) β ψ−1(M ; E)Γ0 → p(0) β E))Γ0 → 0 . 0 → p(0) Moreover, ψ0(M ; E)Γ0 ≃ ⊕β∈bΓ0 p(0) β ψ0(M ; E)Γ0 . (Here the direct sum is finite, so there is no need to include the "c0"-specification like in Corollary 3.7.) Since the actions of Γ and Γ0 commute, we can further take the Γ-invariants to obtain: (31) 0 → p(0) β ψ−1(M ; E)Γ → p(0) β ψ0(M ; E)Γ → C0(S∗M ; End(p(0) β E))Γ → 0 . Moreover, (32) ψ0(M ; E)Γ ≃ ⊕β∈bΓ p(0) β ψ0(M ; E)Γ . 20 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR In particular, we have that Lemma 4.4. Let Γ be a finite abelian group and E → M a Γ-equivariant vector bundle over a smooth, compact, connected manifold M (thus without boundary). Let Γ0 ⊂ Γ be the minimal isotropy group. We have C0(S∗M ; End(E))Γ ≃ ⊕β∈bΓ0 C0(S∗M ; End(p(0) β E))Γ Proof. We successively have C0(S∗M ; End(E))Γ ≃ (cid:0)C0(S∗M ; End(E))Γ0(cid:1)Γ/Γ0 ≃ (cid:0)C0(S∗M ; End(E)Γ0 )(cid:1)Γ/Γ0 ≃ ⊕β∈bΓ0 (cid:0)C0(S∗M ; End(p(0) β′ E)Γ0 for β 6= β′ ∈bΓ0. Let us record now the following corollary of Proposition 3.11. where we have used that Hom(p(0) β E, p(0) ≃ ⊕β∈bΓ0 β E)Γ0 )(cid:1)Γ/Γ0 C0(S∗M ; End(p(0) β E))Γ , (cid:3) Corollary 4.5. Let Γ be a finite abelian group acting on a smooth, connected compact manifold M (without boundary). Let E → M be an equivariant vec- tor bundle. Assume that minimal isotropy is trivial: Γ0 = 1. Then the map RM : C0(S∗M ; End(E))Γ → πα(ψ0(M ; E)Γ)/πα(ψ−1(O; E)Γ) of Equation (28) is injective, and hence an isomorphism of algebras. Proof. By replacing the action π of Γ on E with π0 := πα−1, that is, with, π0(g)ξ := α−1(g)π(g)ξ, we can assume that α = 1. The action of Γ is moreover free on the dense open subset M(1) = MΓ0 of M . Proposition 3.11 then allows us to conclude. (cid:3) We now turn to the main result of this section. Theorem 4.6. Let Γ be a finite abelian group acting on a smooth compact, con- nected manifold M (without boundary) and let E → M be a Γ-equivariant vector bundle. Then the kernel of the morphism RM : ⊕β∈bΓ0 C0(S∗M ; End(p(0) β E))Γ ≃ C0(S∗M ; End(E))Γ =: AΓ M → πα(ψ0(O; E)Γ)/πα(ψ−1(M ; E)Γ) of Equation (28) is ⊕β∈bΓ0,β6=α′ C0(S∗M ; End(p(0) ticular, RM (C0(S∗M ; End(E))Γ) ≃ C0(S∗M ; End(p(0) α′ E))Γ. β E))Γ, where α′ := αΓ0 . In par- Proof. It is enough to identify the action of RM on each direct summand C0(S∗M ; End(p(0) of C0(S∗M ; End(E))Γ. We can thus study the action of the morphism RM one iso- β E))Γ The relation between the central projectors of C∗(Γ0) and C∗(Γ) is that p(0) typical component β ∈bΓ0 at a time. PαΓ0 =β pα, β ∈bΓ0. Of course, pαpα′ = 0 if α 6= α′ ∈bΓ. It follows that β P L2(M;E)α = ( pαP L2(M;E)α if αΓ0 = β otherwise. β P ) = pαp(0) πα(p(0) (33) 0 β = This shows that RM = 0 on C0(S∗M ; End(p(0) β E))Γ if αΓ0 6= β. FREDHOLM CONDITIONS 21 On the other hand, for αΓ0 = β, we shall show that RM is an isomorphism on C0(S∗M ; End(p(0) β E))Γ. By replacing the action π of Γ on E with π0 := α−1π, that is, with, π0(g)ξ := α−1(g)π(g)ξ, we can assume that Γ0 acts trivially on Eβ = p(0) β E (we already know that Γ0 acts trivially on M ). We can then factor the action of Γ to an action of Γ/Γ0, and thus assume that the minimal isotropy is trivial: Γ0 = 1. After these reductions, the orbit bundle of M is O0 := M(1), and the action of Γ on O0 is free (and proper since Γ is compact). Corollary 4.5 then shows that RM is injective on C0(S∗M ; End(p(0) (cid:3) β E))Γ. We note that the component of σm(P ) in C0(S∗M ; End(p(0) 0 (P ) to X α m(P ), M,E,Γ. In this regard, we notice that β E))Γ is σα αΓ0 = β, that is, the restriction of σΓ M,E,Γ = X α′ M,E,Γ ∩ X α′ ( X α (34) X α M,E,Γ M,E,Γ = ∅ if αΓ0 = α′Γ0 otherwise. M,E,Γ = X α M,E,Γ if αΓ0 = β. This gives the disjoint union decom- Let us denote X β position (35) XM,Γ,E = Gβ∈bΓ0 X β M,E,Γ . It would be interesting to establish an analogous relation in the nonabelian case. 5. Applications and extensions We now prove the main result of the paper on the characterization of Fredholm operators and discuss some extensions of our results. 5.1. Fredholm conditions. We now turn to the proof of our main result. We assume that M is a compact smooth manifold. We have the following Γ -- equivariant version of Atkinson's theorem. Proposition 5.1. Let V be a unitary Γ -- module and P be a Γ -- equivariant bounded operator on V . We have that P is Fredholm if, and only if, it is invertible modulo K(V )Γ, in which case, we can choose the parametrix (i.e. the inverse modulo the compacts) to also be Γ-invariant. Proof. This follows from the inclusion of C∗-algebras L(V )Γ/K(V )Γ ⊂ L(V )/K(V ). It is a standard fact that, if B ⊂ A is an inclusion of unital C∗-algebras, then an element a ∈ B is invertible in A if, and only if, it is invertible in B [21, Proposition 1.3.10]. Therefore if P ∈ L(V )Γ, then its projection in L(V )Γ/K(V )Γ is invertible if, and only if, it is invertible in the greater algebra L(V )/K(V ). By Atkinson's theorem, the latter is equivalent to P being Fredholm. (cid:3) Since πα(KΓ) = πα(K(L2(M ; E)α))Γ and ψ−1(M ; E) = K := K(L2(M ; E)α), we obtain the following corollary. Corollary 5.2. Let P ∈ ψ0(M ; E)Γ and α ∈ Γ. We have that πα(P ) is Fredholm on L2(M ; E)α if, and only if, πα(P ) is invertible modulo πα(KΓ) in πα(ψ0(M ; E)Γ). We are now in a position to prove the main result of this paper, Theorem 1.2. 22 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR Proof ot Theorem 1.2. Lemma 2.8 implies that we may assume P ∈ ψ0(M ; E)Γ. Corollary 5.2 then states that πα(P ) is Fredholm if, and only if, the image of its principal symbol σ0(P ) is invertible in the quotient algebra M ) = πα(ψ0(M ; E)Γ)/πα(KΓ). RM (AΓ We have shown the isomorphism Prim(AΓ M ) ≃ XM,E,Γ/Γ in Equation (30). Now Theorem 4.6 and the discussion following it identify the primitive spectrum of RM (AΓ M,E,Γ/Γ. Recall that M ), which is a closed subset of Prim(AΓ M ), with the set X α as defined in the introduction. X α M,E,Γ = {(ξ, ρ) ∈ T ∗M \ {0} ×bΓξ ρΓ0 = αΓ0}, Therefore RM (σ(P )) is invertible if, and only if, the endomorphism πξ,ρ(σ(P )) (cid:3) M,E,Γ, i.e. if and only if P is α-elliptic. is invertible for all (ξ, ρ) ∈ X α 5.2. Boundary value problems. In this subsection, we very briefly indicate an application to mixed boundary value problems. Let M be a smooth compact man- ifold with boundary and choose a tubular neighborhood U ≃ [0, 1) × ∂M of the boundary. Let M d be the double of M along ∂M : as a topological space, the space M d is the quotient of M × {−1, 1} by the subspace ∂M × {−1, 1}. We shall denote M± := M × {±1}. On M d we consider the smooth structure such that (1) the projections p± : M± → M are smooth maps, and (2) the map U d ≃ (0, 1) × ∂M is smooth. Thus the smooth structure on M d thus depends on our choice of tubular neighbor- hood. For any x = (x′, i) ∈ M d, we denote by −x its symmetrical counterpart, i.e. −x = (x′, −i). Then the map x 7→ −x gives a natural smooth action of Z2 on M d. If E → M is a smooth vector bundle, then we define Ed → M d as the smooth vector bundle obtained by gluing two copies of E on M+ and M− along ∂M . Then the Z2-action on M d extends to an action on Ed, which maps an element v ∈ Ed x to its copy in Ed −x, for any x ∈ M d. We generalize this construction to the case when we have a disjoint union de- composition of the boundary ∂M = ∂DM ∪∂N M into two disjoint, closed and open subsets. Then we double first with respect to the "Dirichlet" part of the boundary ∂DM and then with respect to the "Neumann" part of the boundary ∂N M . We obtain accordingly an action of Z2 2 on the resulting manifold M dd. We let this group act on the resulting vector bundle Edd such that the action of the first component of Z2 is twisted (i.e. tensored) with its only non-trival character, namely −1. We have the following standard lemma. Lemma 5.3. The restriction map r+ : C∞(M dd; Edd) → C∞(M+; E) induces a isomorphisms (1) L2(M dd; Edd)Z2 (2) H 2(M dd; Edd)Z2 2 ≃ L2(M ; E), 2 ≃ H 2(M ; E) ∩ { u∂DM = 0, ∂ν u∂N M = 0 }. An order-2, Z2 2-invariant pseudodifferential operator P on M dd will map invari- ant sections to invariant sections; this means that we consider the case α = 1 in 2 is free on a dense subset of M dd, Theorem Theorem 1.2. Because the action of Z2 1.2 implies that P is Fredholm from H 2(M dd; Edd)Z2 2 if, and only if, it is elliptic. This then yields Fredholm conditions for the restriction of P to M , with mixed Dirichlet/Neumann boundary conditions on ∂DM and ∂N M . 2 to L2(M dd; Edd)Z2 FREDHOLM CONDITIONS 23 5.3. The case of non-discrete groups. If Γ is not discrete, then it is enough for our operators to be transversally elliptic. Indeed, let us assume that M is a compact smooth manifold and that Γ is a compact Lie group acting on M . Denote by g the Lie algebra of Γ. Recall that any X ∈ g defines as usual the vector field X ∗ etX · m. Let first introduce the Γ-transversal space M given by X ∗ M (m) = d T ∗ ΓM := {α ∈ T ∗M α(X ∗ dt t=0 M (π(α))) = 0, ∀X ∈ g}. A Γ-invariant classical pseudodifferential operator P of order m is said Γ-transversally ΓM \ {0}. Let P ∈ ψm(M ; E0, E1) elliptic if its principal symbol is invertible on T ∗ be Γ-transversally elliptic. Recall the now classical result of Atiyah and Singer [2, Corollary 2.5] Theorem 5.4. Assume P is Γ-transversaly elliptic. Then for every irreducible representation α ∈bΓ, is Fredholm. πα(P ) : H s(M ; E0)α → H s−m(M ; E1)α, Note that this implies that Theorem 1.2 is not true anymore for Γ-transversally elliptic operators if Γ is non-discrete. References [1] C. Akemann, G. Pedersen, and J. Tomiyama. Multipliers of C ∗-algebras. J. Functional Anal- ysis, 13:277 -- 301, 1973. [2] M. Atiyah. Elliptic operators and compact groups. Lecture Notes in Mathematics, Vol. 401. Springer-Verlag, Berlin-New York, 1974. [3] M. Atiyah and I Singer. The index of elliptic operators III. Ann. of Math., 87:546 -- 604, 1968. [4] K. Austin and J. Zhang. Limit operator theory for groupoids. arXiv preprint, to appear in Trans. AMS, 2018. [5] A. Baldare, R. Come, M. Lesch, and V. Nistor. Fredholm conditions for invariant pseudodif- ferential operators restricted to isotypical components. (tentative title) in final preparation. [6] S. Beckus, D. Lenz, M. Lindner, and Ch. Seifert. On the spectrum of operator families on discrete groups over minimal dynamical systems. Math. Z., 287(3-4):993 -- 1007, 2017. [7] T. Brocker and T. tom Dieck. Representations of compact Lie groups, volume 98 of Grad- uate Texts in Mathematics. Springer-Verlag, New York, 1995. Translated from the German manuscript, Corrected reprint of the 1985 translation. [8] L. Brown, P. Green, and M. Rieffel. Stable isomorphism and strong Morita equivalence of C ∗-algebras. Pacific J. Math., 71(2):349 -- 363, 1977. [9] J. Bruning and E. Heintze. Representations of compact Lie groups and elliptic operators. Invent. Math., 50(2):169 -- 203, 1978/79. [10] J. Bruning and H. Schroder. On the absence of log terms in the constant curvature case. Asymptotic Anal., 1(3):193 -- 203, 1988. [11] R. Busby. Double centralizers and extensions of C ∗-algebras. Trans. Amer. Math. Soc., 132:79 -- 99, 1968. [12] C. Carvalho, R. Come, and Yu Qiao. Gluing action groupoids: Fredholm conditions and layer potentials. Rev. Roumaine Math. Pures Appl., 2019. [13] C. Carvalho, V. Nistor, and Yu Qiao. Fredholm conditions on non-compact manifolds: theory and examples. In Operator theory, operator algebras, and matrix theory, volume 267 of Oper. Theory Adv. Appl., pages 79 -- 122. Birkhauser/Springer, Cham, 2018. [14] R. Come. The Fredholm Property for Groupoids is a Local Property. Results in Mathematics, 74(4):160, August 2019. [15] A. Connes. Non commutative differential geometry. Publ. Math. IHES, 62:41 -- 144, 1985. [16] A. Connes. Noncommutative geometry. Academic Press, San Diego, 1994. [17] A. de Monvel-Berthier and V. Georgescu. Graded C ∗-algebras in the N -body problem. J. Math. Phys., 32(11):3101 -- 3110, 1991. 24 A. BALDARE, R. C OME, M. LESCH, AND V. NISTOR [18] C. Debord, J.-M. Lescure, and F. Rochon. Pseudodifferential operators on manifolds with fibred corners. Ann. Inst. Fourier (Grenoble), 65(4):1799 -- 1880, 2015. [19] C. Debord and G. Skandalis. Adiabatic groupoid, crossed product by R∗ + and pseudodiffer- ential calculus. Adv. Math., 257:66 -- 91, 2014. [20] C. Debord and G. Skandalis. Pseudodifferential extensions and adiabatic deformation of smooth groupoid actions. Bull. Sci. Math., 139(7):750 -- 776, 2015. [21] J. Dixmier. Les C ∗-alg`ebres et leurs repr´esentations. Les Grands Classiques Gauthier-Villars. [Gauthier-Villars Great Classics]. ´Editions Jacques Gabay, Paris, 1996. Reprint of the second (1969) edition. [22] S. Echterhoff and D. Williams. Structure of crossed products by strictly proper actions on continuous-trace algebras. Trans. Amer. Math. Soc., 366(7):3649 -- 3673, 2014. [23] V. Georgescu. On the structure of the essential spectrum of elliptic operators on metric spaces. J. Funct. Anal., 260(6):1734 -- 1765, 2011. [24] V. Georgescu and A. Iftimovici. Crossed products of C ∗-algebras and spectral analysis of quantum Hamiltonians. Comm. Math. Phys., 228(3):519 -- 560, 2002. [25] P. Green. The local structure of twisted covariance algebras. Acta Math., 140(3-4):191 -- 250, 1978. [26] L. Hartmann, M. Lesch, and B. Vertman. On the domain of Dirac and Laplace type operators on stratified spaces. J. Spectr. Theory, 8(4):1295 -- 1348, 2018. [27] B. Helffer and A. Mohamed. Caract´erisation du spectre essentiel de l'op´erateur de Schrodinger avec un champ magn´etique. Ann. Inst. Fourier (Grenoble), 38(2):95 -- 112, 1988. [28] B. Lange and V. Rabinovich. Pseudodifferential operators on Rn and limit operators. Mat.Sb.(N.S.), 129(171(2)):175 -- 185, 1986. [29] Y. Last and B. Simon. The essential spectrum of Schrodinger, Jacobi, and CMV operators. J. Anal. Math., 98:183 -- 220, 2006. [30] R. Lauter, B. Monthubert, and V. Nistor. Pseudodifferential analysis on continuous family groupoids. Doc. Math., 5:625 -- 655 (electronic), 2000. [31] R. Lauter and S. Moroianu. Fredholm theory for degenerate pseudodifferential operators on manifolds with fibered boundaries. Comm. Partial Differential Equations, 26:233 -- 283, 2001. [32] M. Lein, M. Mantoiu, and S. Richard. Magnetic pseudodifferential operators with coefficients in C ∗-algebras. Publ. Res. Inst. Math. Sci., 46(4):755 -- 788, 2010. [33] M. Lesch. Operators of Fuchs type, conical singularities, and asymptotic methods, volume 136 of Teubner-Texte zur Mathematik [Teubner Texts in Mathematics]. B. G. Teubner Ver- lagsgesellschaft mbH, Stuttgart, 1997. [34] M. Lindner. Infinite matrices and their finite sections. Frontiers in Mathematics. Birkhauser Verlag, Basel, 2006. An introduction to the limit operator method. [35] M. Lindner and M. Seidel. An affirmative answer to a core issue on limit operators. J. Funct. Anal., 267(3):901 -- 917, 2014. [36] B. Monthubert. Pseudodifferential calculus on manifolds with corners and groupoids. Proc. Amer. Math. Soc., 127(10):2871 -- 2881, 1999. [37] J. Mougel. Essential spectrum, quasi-orbits and compactifications: application to the heisen- berg group. Rev. Roumaine Math. Pures Appl., 2019. [38] M. Mantoiu. Essential spectrum and Fredholm properties for operators on locally compact groups. J. Operator Theory, 77(2):481 -- 501, 2017. [39] M. Mantoiu and V. Nistor. Spectral theory in a twisted groupoid setting: Spectral decompo- sitions, localization and fredholmness. to appear in Muenster J. Math, preprint 2018. [40] V. Nistor. Higher index theorems and the boundary map in cyclic cohomology. Doc. Math., 2:263 -- 295 (electronic), 1997. [41] M. Pimsner and D. Voiculescu. Exact sequences for K-groups and Ext-groups of certain cross-product C ∗-algebras. J. Operator Theory, 4(1):93 -- 118, 1980. [42] M. Pimsner and D. Voiculescu. K-groups of reduced crossed products by free groups. J. Operator Theory, 8(1):131 -- 156, 1982. [43] V. Rabinovich, S. Roch, and B. Silbermann. Limit operators and their applications in oper- ator theory, volume 150 of Operator Theory: Advances and Applications. Birkhauser, 2004. [44] J. Renault. A groupoid approach to C ∗-algebras, volume 793 of LNM. Springer, 1980. [45] J. Rosenberg. The role of K-theory in noncommutative algebraic topology. In Operator al- gebras and K-theory (San Francisco, Calif., 1981), volume 10 of Contemp. Math., pages 155 -- 182. Amer. Math. Soc., Providence, R.I., 1982. FREDHOLM CONDITIONS 25 [46] R. T. Seeley. Complex powers of an elliptic operator. In Singular Integrals (Proc. Sympos. Pure Math., Chicago, Ill., 1966), pages 288 -- 307. Amer. Math. Soc., Providence, R.I., 1967. [47] J.-P. Serre. Linear representations of finite groups. Springer-Verlag, New York-Heidelberg, 1977. Translated from the second French edition by Leonard L. Scott, Graduate Texts in Mathematics, Vol. 42. [48] M. Taylor. Partial differential equations II. Qualitative studies of linear equations, volume 116 of Applied Mathematical Sciences. Springer, New York, second edition, 2011. [49] Tammo tom Dieck. Transformation groups, volume 8 of De Gruyter Studies in Mathematics. Walter de Gruyter & Co., Berlin, 1987. [50] F. Tr`eves. Introduction to pseudodifferential and Fourier integral operators. Vol. 1. Plenum Press, New York-London, 1980. Pseudodifferential operators, The University Series in Math- ematics. [51] E. Van Erp and R. Yuncken. A groupoid approach to pseudodifferential operators. http://arxiv.org/abs/1511.01041 [math.DG], 2015. [52] D. Voiculescu. Un th´eor`eme du type Weyl-von Neumann non commutatif. C. R. Acad. Sci. Paris S´er. A-B, 281(17):Ai, A735 -- A736, 1975. [53] D. Voiculescu. A non-commutative Weyl-von Neumann theorem. Rev. Roumaine Math. Pures Appl., 21(1):97 -- 113, 1976. [54] D. Williams. Crossed products of C ∗-algebras, volume 134 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2007. E-mail address: [email protected] E-mail address: [email protected] M.L.: Mathematisches Institut, Universitat Bonn, Endenicher Allee 60, 53115 Bonn, Germany E-mail address: [email protected], [email protected] URL: www.matthiaslesch.de, www.math.uni-bonn.de/people/lesch E-mail address: [email protected] A.B, R.C., and V.N. Universit´e Lorraine, 57000 Metz, France URL: http://www.iecl.univ-lorraine.fr/~Victor.Nistor
1501.03479	2	1501	2015-01-19T13:19:35	Intrinsic Chern-Connes Characters for Crossed Products by $\mathbb Z^d$	[ "math.OA", "cond-mat.str-el", "math-ph", "math-ph" ]	We present a natural imbedding of the crossed product $\mathcal A \rtimes_\xi \mathbb Z^d$ into the $C^\ast$-algebra of adjointable operators over the standard Hilbert $\mathcal A$-module $\mathcal H_{\mathcal A}$. By replacing the representations on Hilbert spaces with this canonical imbedding, we define Fredholm modules and corresponding Chern-Connes characters that are intrinsic to the $C^\ast$-dynamical system $(\mathcal A,\xi,\mathbb Z^d)$. The compression of the Dirac operator against projectors from $\mathcal A \rtimes_\xi \mathbb Z^d$ produces generalized Fredholm operators over $\mathcal H_{\mathcal A}$ and Mingo's index defines a $KK$-map from $K_0(\mathcal A \rtimes_\xi \mathbb Z^d)$ to $K(\mathcal A)$. Using a generalized Fedosov principle and a generalized Fedosov formula, we prove an index formula for the pairing of the intrinsic Chern-Connes characters and $K_0(\mathcal A \rtimes_\xi \mathbb Z^d)$. This pairing takes values in the image of $K_0(\mathcal A)$ in $\mathbb R$ under a canonical trace. A local index formula enables new applications in condensed matter physics to the so called weak topological invariants.	math.OA	math	INTRINSIC CHERN-CONNES CHARACTERS FOR CROSSED PRODUCTS BY Zd EMIL PRODAN Abstract. By imbedding A ⋊ξ Zd into the C ∗-algebra of adjointable oper- ators over the standard Hilbert A-module HA, we define Fredholm modules and Chern-Connes characters that are intrinsic to the C ∗-dynamical system over HA and develop a generalized Fedosov principle and formula. We prove an index formula for the pairing of the characters with K0(A ⋊ξ Zd) and con- (A, ξ, Zd). We introduce a bT -index for the generalized Fredholm operators clude that the pairing is in the image of K0(A) under the trace bT . A local index formula enables new applications in condensed matter physics. INTRODUCTION (0.1) Consider a classical dynamical system (Ω, ξ), where ξ = (ξ1, . . . , ξd) is a system of d-commuting homeomorphisms (d = even). Let (cid:0)C(Ω), ξ, Zd(cid:1) be its dual C∗- dynamical system and C(Ω) ⋊ξ Zd the canonically associated crossed-product. Let πω be the standard representation of the crossed product on ℓ2(Zd): (cid:16)πω(cid:16)Pq φq · q(cid:17)ψ(cid:17)x =Pq φq(ξxω)ψx−q, and Dx0 = γ ⊗ (X + x0) be the (shifted-) Dirac operator on C2 (cid:0)C2 2 ⊗ ℓ2(Zd). Then 2 ⊗ℓ2(Zd), πω, Fx0, γ(cid:1) with Fx0 = sign(Dx0) is a natural even Fredholm module over C(Ω) ⋊ξ Zd and its Chern-Connes character, defined as the cohomology class of the cyclic cocycle: d d (0.2) τd(a0, . . . ad) = 1/2 Tr(cid:8)γFx0[Fx0 , πω(a0)] . . . [Fx0 , πω(ad)](cid:9), pairs well and integrally with the K0-group [12]: (0.3) K0(C(Ω) ⋊ξ Zd) ∋ [p]0 → τd(p, . . . p) = Index(cid:8)π+ In [3, 32], under certain optimal conditions, the local formula: ω (p)Fx0 π− (0.4) τd(a0, . . . ad) = Λd Xρ∈Sd (−1)ρT {a0∂ρ1 a1 . . . ∂ρd ad}, (cid:16)Λd = ω (p)(cid:9) ∈ Z. (2π√−1) (d/2)! (cid:17) d 2 was proven by elementary means. Above, Sd is the group of permutations and (∂,T ) is the noncommutative differential calculus over C(Ω) ⋊ξ Zd. The identity 6.19 played a key role in this computation. In d = 2, this identity is due to Connes [12], who used it to compute the 2-dimensional Chern characters 1991 Mathematics Subject Classification. 46L87, 19K35, 19K56,19L64. Key words and phrases. Chern-Connes character, discrete crossed products, local index formula. This work was supported by the U.S. NSF grant DMR-1056168. 1 2 EMIL PRODAN c (R2) and C∞ of the convolution algebras C∞ c (SL(2, R)). The local index formula of [3, 32] can be seen as a particularization of the generic Connes-Moscovici formula [14] and its later extensions [19, 10, 9, 11, 8]. A recent related work is [1], where local index formulas for Rieffel deformed crossed products are derived. Among these works, only [8] seems to cover the general settings of [3, 32] where the index formulas were shown to hold over certain noncommutative Sobolev spaces. Notice also that 0.4 gives a local formula for the character itself and not just of the pairing. There is some interest from the condensed matter physics community in the results summarized by 0.1-0.4 because all non-interacting quantum lattice models of homogeneous materials can be generated as representations of crossed products by Zd. The righthand side of 0.4 relates to the transport coefficients of real materials, in particular, it has certain relevance for topological insulators [31]. While [3, 32] explained some outstanding properties of these materials in the presence of strong disorder and magnetic fields, the approach is limited to the single particle theory of solids, where the electron-electron interaction is treated as a mean-field correction. In fact, even before considering the electron-electron interaction, one needs to address the following shortcomings: (a) The formalism works only for crossed products of commutative algebras. (b) The K-groups of many crossed-products used in condensed matter can be fully resolved by the top and the lower Chern numbers. [32] produced a non- commutative theory only for the top Chern number. (c) The pairing of the characters with the K-groups is always integral, hence not relevant for the sequences of fractional topological phases, such as the fractional Chern insulators [27]. Removing these deficiencies while maintaining the elementary character of the cal- culation was the main motivation for the present work. The root of the problem is the representation of the Fredholm modules on Hilbert spaces and a natural cure is provided by KK-theory. While exploring this path, we discovered that the crossed-product A⋊ξ Zd can be canonically imbedded in the C∗- algebra of adjointable operators over the standard Hilbert A-module HA. A Dirac operator over HA can be naturally defined from the action of Zd after tensoring with an appropriate Clifford algebra. Furthermore, if A posses a continuous trace, then this trace can be naturally promoted to a lower semicountinuous trace "T over the imbedding algebra. The definitions of the intrinsic Fredholm modules, of the notion of summability and of the intrinsic Chern-Connes characters are then fairly straightforward (see Defs. 5.2, 5.3 and 6.1, respectively). One outcome of the approach (see Th. 6.3) is that the operator replacing the one appearing inside the Index in 0.3 is now a generalized Fredholm operator over HA. As such, Mingo's index [24] for C∗-modules provides a KK-map from K0(A ⋊ξ Zd) to K0(A). One could imagine the possibility of the crossed product A ⋊ξ Zd being imbedded in the algebra of adjointable operators over HB of another C∗-algebra, in which case one will perhaps obtain a KK-map into K0(B). This could be a useful tool for the computation of the K0-groups. However, what we find interesting about our construction is that it is intrinsic, in the sense that the entire construction is natural and relying entirely on data from the C∗-dynamical system (A, ξ, Zd). the later generalizations [26], and especially [28], we define a numerical "T -index In analogy with the Breuer-Fredholm index for von Neumann algebras [5, 6] and INTRINSIC CHERN-CONNES CHARACTERS 3 for the multiplier algebra, by applying the trace "T on the Mindex of generalized Fredholm operators over HA. This numerical index takes values in the image of K0(A) under the trace "T and the structure of this sub-group of R can be far more complex than Z. For example, if A is the irrational rotational algebra Aθ, then this image is at least (Z + θZ)∩ [0, 1] [34]. This addresses point (c) above. In order to compute the "T -index, we develop a generalized Fedosov principle and a Fedosov formula (see Th. 4.6), which enables us to prove an index formula for the pairing of the intrinsic Chern-Connes characters with K0(A ⋊ξ Zd). Hence, this pairing is in the image of K0(A) under the trace "T . Lastly, using the same elementary methods as in [32], we derive a local formula for the Chern-Connes cocycle, similar to 0.4 (see Th. 6.5). The theory now covers the lower Chern numbers, though only in the regime of weak disorder. An application for disordered topological insulators in 3 space-dimensions is provided in the last Chapter, where an index formula for the so called weak topological invariants is derived and predictions about their possible values are made. 1. PRELIMINARIES Let (A, ξ, Zd) be a C∗-dynamical and A⋊ξ Zd its canonical crossed-product. The C∗-algebra A is assumed separable and to posses a unit and a faithful, continuous and ξ-invariant trace TA. This is the only input we need for the definition and characterization of the intrinsic Chern-Connes characters. Throughout our presentation, we follow closely the notation from Davidson's monograph [15]. In particular, the core algebra AZd will be represented by formal finite sequences: (1.1) a =Xq aq · q, aq ∈ A, q ∈ Zd, together with the standard algebraic operations. The algebra A is imbedded in the crossed-product as: (1.2) A ∋ a → a = a · 0 ∈ A ⋊ξ Zd, and the additive group as: (1.3) Zd ∋ q → uq = 1 · q ∈ A ⋊ξ Zd. The first imbedding is isometric and the second one is in the group of unitaries of A ⋊ξ Zd. The Fourier calculus is defined by the group of automorphisms (cid:8)ρλ(cid:9)λ∈Td on A ⋊ξ Zd [15], which act on the core algebra as: (1.4) λqaq · q, λq = λq1 1 . . . λqd d . Td ∋ λ → ρλ(a) =Xq The Fourier coefficients of a ∈ A ⋊ξ Zd are defined by the Riemann integral: (1.5) dµ(λ) λ−qρλ(a)u−1 q , dµ(λ) ρλ(au−1 Φq(a) =ZTd q ) =ZTd where µ is the Haar measure on Td. The Fourier coefficients will be seen as elements of A. 4 EMIL PRODAN Proposition 1.1 ([15], pg. 223). The Ces`aro sums: . . . (1.6) aN = Å1 − qj N + 1ã Φq(a) · q NXq1=−N NXqd=−N dYj=1 converge in norm to a ∈ A ⋊ξ Zd as N → ∞. Hence, generic elements from A ⋊ξ Zd can be representated as a Fourier series: (1.7) a = Xq∈Zd Φq(a) · q, where the infinite sum must be interpreted via 1.1. The Fourier calculus generates a canonical faithful and continuous trace on A ⋊ξ Zd [15]: (1.8) T {a} = TA{Φ0(a)}. It also generates a set of un-bounded derivations, ∂ = (∂1, . . . , ∂d), through the generators of the d-parameter group of automorphisms {ρλ}λ∈Td. The derivations act as: (1.9) ∂a = ı Xq∈Zd qΦq(a) · q, (ı = √−1). Together, (∂,T ) define the non-commutative calculus over A ⋊ξ Zd. Generically, the cocycles can be defined only on a pre C∗-sub-algebra of A ⋊ Zd [13], which can be generated by various means [37, 33]. Below we describe one such sub-algebra which we find particularly convenient for the calculations to follow. Proposition 1.2. Consider the set (A ⋊ξ Zd)loc: na ∈ A ⋊ξ Zd ∃ α < 1, s.t. sup (1.10) q (cid:0)α−qkΦq(a)k(cid:1) < ∞o. (i) The set (A ⋊ξ Zd)loc can be equivalently characterized as: Then: na ∈ A ⋊ξ Zd ρλ(a) analytic of λ in a strip around Tdo. (1.11) (ii) When equipped with the algebraic operations, (A ⋊ξ Zd)loc becomes a dense sub-algebra of A ⋊ξ Zd. (iii) This sub-algebra is stable under the holomorphic functional calculus. Proof. (i) Note that ρλ is entire on AZd. If b ∈ AZd and a ∈ A ⋊ξ Zd with ρα(a) analytic of α in a finite strip around Td, then ρα(ab) can be analytically continued in the same strip via ρα(ab) = ρα(a)ρα(b). Likewise, if λ ∈ Td, then ραλ(a) can be analytically continued in α via: (1.12) ραλ(a) = ρλ ◦ ρα(a) = ρα ◦ ρλ(a). The last equality holds because the analytic continuations are unique [4]. Then: (1.13) Φq(cid:0)ρα(a)(cid:1) =ZTd dµ(λ) ρλ(cid:0)ρα(a)u−q(cid:1) = αqZTd dµ(λ) ρα(cid:0)ρλ(au−q)(cid:1). INTRINSIC CHERN-CONNES CHARACTERS 5 (1.14) We can exchange ρα and the integral by using the dominated convergence theorem, to conclude: Φq(cid:0)ρα(a)(cid:1) = αqρα(cid:0)Φq(a)(cid:1) = αqΦq(a). This identity is valid for any α in a strip around Td, and this strip is independent of q. Taking αi = α−qi/q with α close-enough to 1, we obtain: (1.15) kΦq(a)k ≤ αqkρα(a)k, ∀q ∈ Zd. Now assume a ∈ (A ⋊ξ Zd)loc. We can apply ρα on the Ces`aro sums: N + 1ã αqΦq(a) · q, Å1 − qj ρα(aN ) = . . . (1.16) NXq1=−N NXqd=−N dYj=1 and the righthand-side and its derivatives with respect to α can be seen to be absolutely norm-convergent as N → ∞, for α in a thin-enough strip around Td. The statement then follows. (ii) (A ⋊ξ Zd)loc is closed under the algebraic operations, which can be seen directly from representation 1.11. Indeed, if ρα(a) and ρα(b) are analytic in a strip around Td, then ρα(ab) can be analytically continued over the same strip via ρα(ab) = ρα(a)ρα(b). (A ⋊ξ Zd)loc is dense in A ⋊ξ Zd because it contains AZd. (iii) Consider a ∈ (A ⋊ξ Zd)loc which is invertible in A ⋊ξ Zd. We need to show that its inverse belongs to (A ⋊ξ Zd)loc. Since the latter is dense in A ⋊ξ Zd, there is b from (A ⋊ξ Zd)loc such that kab− 1k < 1. Then ab = 1− (1− ab) is invertible, hence b is invertible and a−1 = b(ab)−1. Taking r = 1 − ab, the problem is reduced to showing that (1− r)−1 belongs to (A ⋊ξ Zd)loc. Recall that krk < 1 and r ∈ (A ⋊ξ Zd)loc. Now, ρα(r) is analytic of α in a finite strip around Td hence, by n=0(cid:0)ρα(r)(cid:1)n taking α sufficiently close to the unit circle, kρα(r)k < 1 and P∞ is converges in norm together with its ∂α-derivatives. We conclude that ρα(cid:0)(1− r)−1(cid:1) (A ⋊ξ Zd)loc is a Fr`echet algebra and (A ⋊ξ Zd)loc belongs to the domain of any simple or higher derivation ∂n. The stability under the holomorphic functional calculus ensures that the K0-groups of (A ⋊ξ Zd)loc and A ⋊ξ Zd coincide [13]. Besides, as we shall see, the canonical Fredholm module associated to the crossed product A ⋊ξ Zd is automatically (d + 1)-summable. These facts made us believe that (A ⋊ξ Zd)loc is the natural domain for the intrinsic Chern-Connes characters. is analytic in a strip around Td. (cid:3) 2. THE REPRESENTATION In this Chapter we show that the C∗-algebra of adjointable operators over the standard Hilbert A-module HA can serve as a natural imbedding algebra for A ⋊ξ Zd. This will provide a natural connection between the K-theories of A ⋊ξ Zd and of A, as formulated within the framework of the standard Hilbert A-module (see for example Chapter III in [38]). Viewing A as a right Hilbert A-module, HA is defined as the tensor product A ⊗ H of Hilbert C∗-modules, with H being an ordinary separable Hilbert space [16]. The inner product on HA is ha ⊗ φb ⊗ ψi = hφψia∗b. The C∗-algebra of adjointable compact operators ([20], Def. 4) over HA is isomorphic to A⊗ K, where K is the algebra of compact operators over separable Hilbert spaces. The C∗-algebra of adjointable operators ([20], Def. 3) over HA is isomorphic to M(A⊗ K) (cf. [20], 6 EMIL PRODAN Th. 1), the multiplier algebra or double centralizer [7] of A ⊗ K. From a standard property of the tensor products and their multiplier algebras ([38], Corollary T.6.3), we have a chain of algebra inclusions: (2.1) A ⊗ K ⊂ A ⊗ B ⊂ M(A ⊗ K), where B is the algebra of bounded operators over separable Hilbert spaces and the tensor product in A⊗ B is with the spatial C∗-norm. This observation is important for us because the crossed-product algebra can be naturally imbedded in A⊗ B and 2.1 will provide an imbedding in M(A ⊗ K), which is what we formally need. In the present context, it is convenient to make the choice H = ℓ2(Zd), in which case the elements of HA can be uniquely expressed as: (2.2) bx ⊗ δx ∈ A ⊗ ℓ2(Z2), (bx) = Xx∈Zd with {δx} being the canonical orthonormal basis in ℓ2(Zd). Similarly, the elements of A ⊗ K and A ⊗ B take the form: (2.3) a =Xx,y axy ⊗ Ex,y, with Ex,y being the system of matrix-units over ℓ2(Zd), Ex,yδz = δyzδx. Let us also mention the ideal of finite-rank elements, algebraically generated by the rank-one operators: θ(bx) (ax)(cid:0)(cx)(cid:1) = (ax)h(bx)(cx)i, (ax), (bx), (cx) ∈ HA. (2.4) In our settings: (2.5) θ(bx) (ax) = Xx,y∈Zd axb∗ y ⊗ Ex,y. The ideal of finite-rank operators is dense in A ⊗ K. Proposition 2.1. The following map (2.6) A ⋊ξ Zd ∋ a → π(a) = Xx,q∈Zd ξxΦq(a) ⊗ Ex,x−q ∈ A ⊗ B, is well defined. It provides a faithful morphism of C∗-algebras and a faithful imbed- ding of A ⋊ξ Zd in M(A ⊗ K). Proof. Our first task is to show that π is really into A⊗B, i.e. that π(a) is a bounded operator when represented on some Hilbert space K ⊗ ℓ2(Zd). For a ∈ AZd, this can be accomplished by elementary means using a covariant representation of the dynamical system (A, ξ, Zd). Furthermore, for a and b from AZd, one can explicitly verify that (note the similarity between 2.6 and 0.1): (2.7) π(a)π(b) = π(ab) and π(a∗) = π(a)∗. The map is obviously faithful. At this point we established that π is a faithful ∗-representation of AZd inside the C∗-algebra A ⊗ B. By the very definition of A ⋊ξ Zd, this representation must extend over the whole crossed-product. (cid:3) INTRINSIC CHERN-CONNES CHARACTERS 7 3. THE CANONICAL TRACE AND ITS SPECIFIC CLASSES OF ELEMENTS The trace TA on A and the ordinary trace on B provide a trace "T = TA ⊗ Tr on A ⊗ B. This trace can be characterized more directly as follows. Proposition 3.1. Let A ⊗ B+ be the positive cone of A ⊗ B. The functional: (3.1) A ⊗ B+ ∋ a → "T {a} = Xx∈Zd TA{axx} ∈ [0,∞], is a faithful, lower semicontinuous trace. We recall the following standard consequences of "T being a trace: • The set of trace-class elements: (3.2) (3.3) (3.4) (3.5) S1 = span{a ∈ A ⊗ B+ "T {a} < ∞} is a two-sided ideal in A ⊗ B. • The set of Hilbert-Schmidt elements: S2 = {a ∈ A ⊗ B "T {a∗ a} < ∞} is a two-sided ideal in A ⊗ B and S1 = S2 · S2. • The trace is cyclic: "T {ab} = "T {ba} for all a ∈ SA and b ∈ A ⊗ B. • The trace is invariant to conjugation by unitaries: "T { u∗ a u} = "T {a} for all a ∈ SA and unitary u ∈ A ⊗ B. The following characterization of the trace-class elements is essential for the generalized Fedosov principle elaborated in the next Chapter. Proposition 3.2. S1 = A⊗ S1, where S1 is the ideal of trace-class operators in B. Proof. Given that "T is the extension of TA ⊗ Tr, A ⊗ S1 automatically belongs to the domain of "T . We need to show the reverse inclusion, which follows if we can show that S2 ⊂ A ⊗ S2 where S2 is the ideal of Hilbert-Schmidt operators in B. Since A is separable and TA is faithful and continuous, A can be completed to a separable Hilbert space when equipped with the inner product habi = TA(a∗b). Let {ek} ∈ A be an orthonormal basis of this Hilbert space. Then any element in A ⊗ B can be uniquely represented as Pk ek ⊗ Tk. If such element belongs to S2, then "Tn(cid:16)Xk ek ⊗ Tk(cid:17)∗Xk ek ⊗ Tko =Xk Tr{T ∗ k Tk} < ∞, (cid:3) It will also be important to realize that the ideal of finite-rank operators belongs (3.6) (3.7) hence Pk ek ⊗ Tk ∈ A ⊗ S2. to the domain of the trace. Indeed: "Tnθ(bx) (ax)o = TA(cid:8)h(ax)(bx)i∗(cid:9). Then it is obvious that any finite combination of θ's is in S1. 8 EMIL PRODAN 4. GENERALIZED FREDHOLM INDEX AND THE FEDOSOV FORMULA Within the framework of standard modules, K0(A) is defined as ([38], pg. 264): (4.1) K0(A) = {[ p] − [ q] p, q projectors in A ⊗ K}, where [ ] indicates the classes under the Murray-von-Neumann equivalence relation in M(A ⊗ K). We recall the index map for Hilbert C∗-modules, developed over a stretch of works by Kasparov [20, 21], Miscenco and Fomenko [25], Pimsner, Popa, and Voiculescu [29] and assembled in the final form by Mingo [24]. Se also the pedagogical exposition in [38] which calls this index the Mindex. We will do the same here. In short, one defines the generalized Calkin algebra as the corona M(A ⊗ K)/A ⊗ K, and the class of generalized Fredholm elements as: (4.2) where p is the quotient map M(A ⊗ K) → M(A ⊗ K)/A ⊗ K. M(A⊗ K) is not a von-Neumann algebra, hence the polar decomposition cannot be assumed automatically.1 Nevertheless, any f ∈ FA admits a compact perturba- tion g (i.e. f − g ∈ A ⊗ K) which does accept a polar decomposition g = wg, with w a partial isometry (see [24], Proposition 1.7). This partial isometry defines two compact projectors: FA = { f ∈ M(A ⊗ K) p( f ) invertible in M(A ⊗ K)/A ⊗ K}, (4.3) and ker g = 1 − w w∗ ∈ A ⊗ K ker g∗ = 1 − w∗ w ∈ A ⊗ K, (4.4) which at their turn define an element of the K0(A)-group: (4.5) This element of K0(A) is insensitive of the compact perturbation g used in the construction, hence it is truly determined by the Fredholm element f . Mingo's index map over HA is then defined as [24]: (4.6) [ker g] − [ker g∗] ∈ K0(A). Mindex{ f} = [ker f ] − [ker f ∗] ∈ K0(A), where it is understood that possible (irrelevant) compact perturbations are used to define the kernels. The Mindex is invariant to norm-continuous deformations of f and two Fredholm operators have the same Mindex precisely when they are in the same path-component of FA. The group of the homotopy classes of the generalized Fredholm elements characterizes completely the K0-group: Theorem 4.1 ([24]). K0(A) ≃ [FA]. Our goal for this Chapter is to define a numerical index which is computationally more advantageous. We start with a statement which is standard for operators over ordinary Hilbert spaces: Proposition 4.2. Any projector in A ⊗ K is finite-rank. 1According to an argument due to W. J. Phillips, the polar decomposition is equivalent to asking that f has closed range (see [38], Th. 15.3.8). INTRINSIC CHERN-CONNES CHARACTERS 9 Proof. We will take advantage of the fact that the set of finite-rank elements is an ideal in M(A ⊗ K) and that this ideal is dense in A ⊗ K. Let (4.7) pN = Xx<N 1 ⊗ Ex,x, be the standard approximation of identity for A ⊗ K. Then, for any projector e ∈ A ⊗ K, e pN converges to e in M(A ⊗ K) as N → ∞. Consequently, for N large enough: (4.8) is invertible in M(A ⊗ K) and es = e pN , hence a finite rank operator. Therefore e = (e pN )s−1 is finite-rank. s = 1 − e(1 − pN ) (cid:3) This detail is important for our development because it shows that the compact projectors belong to the domain of the trace "T . As such, the pairing of "T with K0(A) is straightforward: Proposition 4.3. The map: (4.9) is a group morphism. K0(A) ∋ [ p] − [ q] → "T { p} −"T { q} ∈ R. Definition 4.4. The generalized Fredholm index, which we call the "T -index, is defined as: (4.10) FA ∋ f → Index{ f} = "T {ker f} −"T {ker f ∗}. properties of the Mindex mentioned earlier. The following characterization of the "T -index is a direct consequence of the Proposition 4.5. The "T -index defines a map "T : FA → R which is a locally constant homomorphism of semigroups (multiplicative for FA and additive for R). As for the Breuer-Fredholm index [11], the injectivity modulo path components present in 4.1 may be lost and the information about K0(A) group, that can be extracted using the "T -index, can be limited. One recalls that this is not the case for the classic Fredholm index. However, there is an advantage for using both the "T -index and the Mindex, the former being algorithmically computable as the following generalization of the Fedosov's work [17] shows. Theorem 4.6 (Fedosov principle and formula). Let f ∈ M(A ⊗ K) with k fk ≤ 1. Then: (i) If there exists a natural number n such that: (4.11) (1 − f ∗ f )n and (1 − f f ∗)n ∈ S1, (ii) If (1) holds, then the "T -index can be computed via: then f ∈ FA. Index{ f} = "T {(1 − f ∗ f )n} −"T {(1 − f f ∗)n}. (4.12) Proof. (i) Assume 4.11 true. Then 3.2 assures that (1 − f ∗ f )n belongs to A ⊗ S1, hence compact. Taking n′ > n of the form n′ = 2k, we can apply successively the to access 1− f ∗ f . The latter is necessarily compact due square root on (1− f ∗ f )n′ 10 EMIL PRODAN to the following generic argument. Suppose a ∈ A ⊗ K is positive. Then its square root is defined as a positive element by the continuous functional calculus and: (4.13) √a = lim ǫց0 a(cid:0)ǫ + √a(cid:1)−1 , where the limit is in M(A ⊗ K). Since A ⊗ K is an ideal, the operators inside the limit belong to A ⊗ K and, since A ⊗ K is closed, the limit is in A ⊗ K. Similarly, 1 − f f ∗ ∈ A ⊗ K. Therefore, p( f ) is invertible in M(A ⊗ K)/A ⊗ K. (ii) Assume for the beginning that f accepts a polar decomposition f = w f, with w = w w∗ w and w∗ = w∗ w w∗. We write: (4.14) and note that: (4.15) and 1 − f f ∗ = 1 − w w∗ + w(1 − f2) w∗, (1 − w w∗)(cid:0) w(1 − f2) w∗(cid:1) = 0 (cid:0) w(1 − f2) w∗(cid:1)(1 − w w∗) = 0. (4.16) When combined with the elementary fact w∗ w f = f, these lead to: (4.17) Using the cyclic property of the trace: (1 − f f ∗)n = 1 − w w∗ + w(1 − f2)n w∗. "T { w(1 − f2)n w∗} = "T { w∗ w(1 − f2)n} w∗ w(1 − f2)n = w∗ w − 1 + (1 − f2)n. (4.18) and, since w∗ w f = f, it follows that: (4.19) This together with 4.12 gives: "T {(1 − f ∗ f )n} −"T {(1 − f f ∗)n} = "T {1 − w∗ w} −"T {1 − w w∗}. (4.20) The affirmation is then proven for the particular case when f accepts a polar decomposition. For the generic case, note that we already established that f is unitary modulo A ⊗ K. Then we can apply Lemma 7.4 of [29] which, as noted in [24], extends to the present context. This Lemma assures us that f (1 − pN ) accepts the polar decomposition, with pN defined in 4.7 and N large enough. Then f (1 − pN ) is a compact perturbation of f with a polar decomposition, and since 4.11 still applies for f (1 − pN ): (4.21) Index{ f} = "T {(1 − (1 − pN ) f ∗ f (1 − pN ))n} −"T {(1 − f (1 − pN ) f ∗)n}. Using the cyclic property of the trace, one can check directly that all the terms containing pN cancel identically, hence this last equation can be reduced to 4.12. (cid:3) The above statements can be generalized in the following way. Suppose there are projections p and p′ such that p′ f = f p = f and p′ = s ps with s a symmetry (i.e. s self-adjoint and s2 = 1). Let: (4.22) By applying the generalized Fedosov principle on f , we learn that f ∈ FA provided: (4.23) f = (1 − p)s + f . ( p′ − f ∗ f )n and ( p − f f ∗)n ∈ S1. INTRINSIC CHERN-CONNES CHARACTERS 11 Furthermore, if 4.23 holds, then we can apply the generalized Fedosov formula on f , which gives: (4.24) Index{ f} = "T {( p′ − f ∗ f )n} −"T {( p − f f ∗)n}. The above index will be understood as the index of f as an operator between the right A-modules pHA and p′HA. 5. THE INTRINSIC FREDHOLM MODULE We extend π to a representation πγ = id ⊗ π on Cld ⊗ M(A ⊗ K), where Cld is the even Clifford algebra with generators: (5.1) γiγj + γjγi = 2δij i, j = 1, . . . , d. We denote by Trγ the standard normalized trace on Cld. Kasparov's stabilization theorem ([20], Th. 2) assures us that we can replace A by Cld⊗A without changing the K-theory and, by examining the arguments in the previous Chapter, one can convince himself that replacing TA by Trγ ⊗ TA will fully accommodate the new setup. This extension enables us to define the Dirac element: bdx0 = Xx∈Zd γ · (x + x0) ⊗ 1 ⊗ Ex,x, (5.2) (5.3) where x · y denotes the Euclidean scalar product x · y = x1y1 + . . . + xdyd. Note that bdx0 is an unbounded operator over HA but its phase Êdx0 = sign(bdx0 ) is part of the imbedding algebra: Êdx0 = Xx∈Zd γ · Ìx + x0 ⊗ 1 ⊗ Ex,x ∈ Cld ⊗ A ⊗ B. Throughout, we will use the notation Êx = x/x. Remark 5.1. The shift x0 is allowed to take values in the cube [0, 1]d. When x0 /∈ Zd, the shift is useful because it provides an un-ambiguous phase. However, the considerations leading to the inclusion of the shift x0 go beyond that. In particular, the average over x0 ∈ [0, 1]d is absolutely essential for the key identity 6.19. When x0 ∈ Zd then γ · Ìx + x0 needs to be modified at one point. This can be done in various ways and is not an important detail. One can verify directly that (Êdx0)∗ = Êdx0 and (Êdx0 )2 = 1. Also: d (5.4) γ = γ0 ⊗ 1 ⊗ 1, γ0 = −ı provides a grading that has the right properties: 2 γ1γ2 . . . γd, (5.5) Êdx0γ = −γÊdx0, and [πγ(a), γ] = 0 ∀ a ∈ A ⋊ξ Zd. All these lead to the following definition. Definition 5.2. The family of intrinsic Fredholm modules for A ⋊ξ Zd is defined as: (5.6) (Cld ⊗ M(A ⊗ K), πγ, Êdx0, γ)x0∈[0,1]d . 12 EMIL PRODAN Definition 5.3. The family of Fredholm modules is n-summable over a sub-algebra of A ⋊ξ Zd if: (5.7) nYi=1(cid:2)Êdx0, πγ(ai)(cid:3) ∈ Cld ⊗ S1, for any ai in that sub-algebra. Theorem 5.4. The intrinsic family of Fredholm modules is n-summable over (A⋊ξ Zd)loc, for any n ≥ d + 1. Proof. Let ai ∈ (A ⋊ξ Zd)loc. We will first show that: (5.8) (cid:13)(cid:13)(cid:13)(cid:16) kYi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17)xx(cid:13)(cid:13)(cid:13) ≤ Ax + x0−k. i=1(cid:2)Êdx0, πγ(ai)(cid:3) belongs to S2 for all k > d/2 and the main af- Consequently, Qk firmation follows for n ≥ d + 2. Additional work will be needed to cover the case n = d + 1. Returning to 5.8, it is useful to write the commutators explicitly: (cid:2)Êdx0, πγ(ai)(cid:3) = Xx,y∈Zd γ · (Ìx + x0 − Ìy + x0) ⊗ ξxΦx−y(ai) ⊗ Ex,y. (5.9) Then (5.10) and (5.11) (5.12) δx1,0δxk+1,0 (cid:16) kYi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17)xx = (id ⊗ ξx) Xxi∈Zd kYi=1 γ · (Ú xi + x + x0 − Ú xi+1 + x + x0) ⊗ ξxiΦxi−xi+1 (ai), (cid:13)(cid:13)(cid:13)(cid:16) kYi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17)xx(cid:13)(cid:13)(cid:13) ≤ A Xxi∈Zd kYi=1(cid:12)(cid:12)Ú xi + x + x0 − Ú xi+1 + x + x0(cid:12)(cid:12)kΦxi−xi+1(ai)k. δx1,0δxk+1,0 Throughout, all uninteresting constants will be denoted by A. Due to the asymp- totic behavior: as x → ∞, the supremum (5.13) Ïy + x − Ìy′ + x ∼ x−1(cid:16)y − y′ +(cid:0)Êx · (y − y′)(cid:1)Êx(cid:17), S(y, y′) = sup(cid:8)x(cid:12)(cid:12)Ïy + x − Ìy′ + x(cid:12)(cid:12), x ∈ Rd(cid:9) is finite. S(y, y′) has the scaling property: S(sy, sy′) = sS(y, y′), (5.14) hence, by taking s = (y + y′)−1, we obtain the upper bound: (5.15) S(y, y′) ≤ (y + y′) sup{S(x, x′), x + x′ = 1}. INTRINSIC CHERN-CONNES CHARACTERS 13 The conclusion is: (5.16) (cid:13)(cid:13)(cid:13)(cid:16) kYi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17)xx(cid:13)(cid:13)(cid:13) ≤ x + x0−k A Xxi∈Zd δx1,0δxk+1,0 kYi=1 (xi + xi+1)αxi−xi+1, with α < 1. The remaining sum is finite. Let us now consider the element: (5.17) e = Xx∈Zd x− 1 4 1 ⊗ 1 ⊗ Ex,x ∈ Cld ⊗ A ⊗ B. Given 5.8, we have: (5.18) (cid:16) d/2Yi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17) e ∈ S2, ∀ ai ∈ (A ⋊ξ Zd)loc and, by conjugation, also: (5.19) e(cid:16) d/2Yi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17) ∈ S2, ∀ ai ∈ (A ⋊ξ Zd)loc. Consider the sum: (5.20) b = Xx,q∈Zd x 1 4 hγ, Ìx + x0 −Ú 2 +1 ∈ (A ⋊ξ Zd)loc. Using 5.15, we have: with a d x − q + x0ix − q 1 4 ⊗ ξxΦq(a d 2 +1) ⊗ Ex,x−q, (5.21) kbx,x−qk ≤ Ax− 1 hence b ∈ Cld ⊗ A ⊗ B. Since S2 is an ideal: 2 αq, α < 1, (5.22) (cid:16) d/2Yi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17)eb ∈ S2, ∀ ai ∈ (A ⋊ξ Zd)loc and combining with 5.19: (5.23) (cid:16) d/2Yi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17)ebe(cid:16) d+1Yi=d/2+1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17) ∈ S1. A direct computation will show that: (5.24) ebe =(cid:2)Êdx0, πγ(a d 2 +1)(cid:3), and the main affirmation follows, with n = d + 1 this time. (cid:3) 14 EMIL PRODAN 6. THE INTRINSIC CHERN-CONNES CHARACTER AND ITS LOCAL FORMULA Definition 6.1. The intrinsic Chern-Connes character is defined by the cohomol- ogy class of following d + 1-cyclic cocycle: (6.1) τd(a0, . . . , ad) = 1/2 Z [0,1]d dx0 "TnγÊdx0 dYi=0(cid:2)Êdx0, πγ(ai)(cid:3)o, defined over (A ⋊ξ Zd)loc. Remark 6.2. We recall that the classic Chern-Connes characters are defined using a representation and the trace on an ordinary Hilbert space [13]. Same algebra is required to show that τd above is a cocycle. Theorem 6.3 (The KK-map). Let p ∈ (A ⋊ξ Zd)loc be a projector and let π± γ = 1/2(1 ± γ)πγ be the decomposition of the representation π with respect to the grading γ. Then the element π− γ (p) ∈ M(A ⊗ K) is Fredholm, hence it provides a map: Proof. Let us list a few useful identities: γ (p) dx0 π+ K0(A ⋊ξ Zd) ∋ [p] → Mindex{π− γ (p)Êdx0 π+ Êdx0(cid:2)Êdx0, πγ(p)(cid:3) = −(cid:2)Êdx0, πγ(p)(cid:3)Êdx0, πγ(p)(cid:2)Êdx0, πγ(p)(cid:3) = −(cid:2)Êdx0, πγ(p)(cid:3)πγ(p) γ (p)} ∈ K(A). πγ(p)(cid:2)Êdx0 , πγ(p)(cid:3)2 = πγ(p)Êdx0 πγ(p)Êdx0 πγ(p) − πγ(p). Now, if we take f = π− γ (p) such that π+ γ (p). As such, we are allowed to use the Fedosov principle in the form 4.23. From identity 6.5: γ (p) dx0 π+ γ (p) = f and (γ1 ⊗ 1 ⊗ 1)π± γ (p), then obviously we have the projectors π± γ (p)(γ1 ⊗ 1 ⊗ 1) = π∓ γ (p) f = f π− (6.2) (6.3) (6.4) and (6.5) (6.6) (6.7) γ (p) − f f ∗)n = −π− (π− γ (p) − f ∗ f )n = −π+ (π+ γ (p)(cid:2)Êdx0, πγ(p)(cid:3)2n γ (p)(cid:2)Êdx0, πγ(p)(cid:3)2n , and, given the summability result 5.4, the Fedosov principle applies if we take 2n > d + 1. (cid:3) Theorem 6.4 (The index theorem). (6.8) τd(p, . . . , p) = Index(cid:8)π− The "T -index on the right is independent of x0. Proof. Given 5.12, dx0 − dx′ take n = d 0 γ (p) dx0 π+ γ (p)(cid:9). (6.9) 2 + 1 and, aided by 6.6 and 6.7, we apply the Fedosov-formula 4.24: Index(cid:8)π+ γ (p) dx0 π− γ (p)(cid:9) = −"Tnγ πγ(p)(cid:2)Êdx0, πγ(p)(cid:3)d+2o. is compact, hence the index is independent of x0. We INTRINSIC CHERN-CONNES CHARACTERS 15 Using 6.5, the righthand side becomes: −"Tnγ(cid:16)πγ(p)Êdx0 πγ(p)Êdx0 πγ(p) − πγ(p)(cid:17)(cid:2)Êdx0, πγ(p)(cid:3)do, which can be rewritten in two different ways: (6.10) (6.11) or (6.12) (6.13) "TnγÊdx0 πγ(p)(cid:2)Êdx0, πγ(p)(cid:3)d+1o, "TnγÊdx0(cid:2)Êdx0, πγ(p)(cid:3)d+1 πγ(p)o. +(cid:2)Êdx0, πγ(p)(cid:3)d+1 Eq. 6.8 then follows from the following elementary identity: πγ(p)(cid:2)Êdx0, πγ(p)(cid:3)d+1 property 5.5 and an average over x0. πγ(p) =(cid:2)Êdx0, πγ(p)(cid:3)d+1 , (cid:3) Theorem 6.5 (The local formula). The Chern-Connes cocycle accepts the local formula: (6.14) τd(a0, . . . , ad) = Λd Xρ∈Sd (−1)ρTna0 dYi=1 ∂ρi aio, where (∂,T ) is the non-commutative calculus over A ⋊ξ Zd and Λd was defined in 0.4. Proof. Opening the first commutator in the product of definition 6.1 transforms its righthand side into: (6.15) 1/2 Xx∈Zd Trγ ⊗ TAn(cid:16)γ(cid:0)πγ(a0) − Êdx0 πγ(a0)Êdx0(cid:1) dYi=1 [Êdx0, πγ(ai)](cid:17)xxo. Given 6.3 and since d is even, we can swap Êdx0 and the product, and using the [Êdx0, πγ(ai)](cid:17)xxo. Trγ ⊗ TAn(cid:16)γ πγ(a0) τd(a0, . . . , ad) = Z cyclic property of Trγ we arrive at: dx0 Xx∈Zd [0,1]d dYi=1 (6.16) Now, using 5.10: (6.17) dYi=1(cid:2)Êdx0, πγ(ai)(cid:3)(cid:17)xx (cid:16)γ πγ(a0) = (id ⊗ ξx) Xxi∈Zd dYi=1 δxd+1,0 γ0 ⊗ Φ−x1(a0) γ · (Ú xi + x + x0 − Ú xi+1 + x + x0) ⊗ ξxi Φxi−xi+1(ai). 16 EMIL PRODAN Using the invariance of TA with respect to ξ-automorphisms and by combining the integration over x0 with the summation over x, we conclude: (6.18) δxd+1,0 τd(a0, . . . , ad) = Xxi∈Zd ZRd dx Trγnγ0 γ · (Ìx + xi − Ôx + xi+1)o dYi=1 ξxi Φxi−xi+1(ai)o. TAnΦ−x1(a0) dYi=1 At this point we use the identity discovered in [32]: (6.19) δxd+1,0Z Rd dx Trγnγ0 dYi=1 γ · (Ìx + xi − Ôx + xi+1)o = Λd Xρ∈Sd (−1)ρ dYi=1 (xi)ρi , with Λd = − (2π) (6.20) ı d 2 d 2 (d/2)! , to continue: (−1)ρ Xxi∈Zd TAnΦ−x1(a0) (xi)ρi ξxi Φxi−xi+1(ai)o. dYi=1 τd(a0, . . . , ad) = Λd Xρ∈Sd Due to the anti-symmetrizer factor, we can replace (xi)ρi by (xi)ρi − (xi+1)ρi and then (cid:0)(xi)ρi − (xi+1)ρi(cid:1)Φxi−xi+1(ai) by −ıΦxi−xi+1(∂ρi ai). After these substitu- tions we can recognize that: (6.21) τd(a0, . . . , ad) = Λd Xρ∈Sd (−1)ρTAnΦ0(cid:16)a0 dYi=1 ∂ρi ai(cid:17)o and the affirmation follows. (cid:3) 7. AN APPLICATION One of the most successful applications of the noncommutative geometry in con- densed matter is the solution to the Integer Quantum Hall Effect (IQHE) [3], which explained the quantization and homotopy invariance of the Hall conductance of a 2-dimensional electron gas subjected to a perpendicular magnetic field and strong disorder. It was Haldane [18] who first realized that certain materials can exhibit all the characteristics of the standard IQHE even in the absence of an external magnetic field. This type of materials are now called Chern insulators. Presently, there exists an entire classificaltion table of presumably all topological insulating phases of matter [36, 22, 35]. When examining this table (see for example Table III in [35]), one notices that there are no Chern insulators in 3 space-dimensions. The physical argument is that, although the transport coefficients do show some topological characteristics in the absence or at weak disorder, those features will presumably disappear in the regime of strong disorder. Translating, this means that, although K0(C(Ω) ⋊ξ Z3) = Z4 [C(Ω) ⋊ξ Z3 is the algebra of physical ob- servables, see below], the cocycles that can be paired with K0 are of degree lower than the space dimension, hence they are not stable once the spectral gap of the INTRINSIC CHERN-CONNES CHARACTERS 17 Hamiltonian closes due to strong disorder. From a mathematical point of view, as far as we know, this problem is completely open. The issue is definitely interesting, being also tied to the quantized Hall-Effect in 3-dimensions, theoretically proposed quite some time ago [23] but whose decisive experimental confirmation is still to come. The quantum dynamics of electrons in homogeneous 3-dimensional materials is generated by self-adjoint operators on ℓ2(Z3, CQ) of the form [2]: (Hψ)x = Xq∈Z3(cid:0)1 + λq(ξxω)(cid:1)Aqψx−q, (7.1) (7.2) with: (7.3) where λq : Ω → C are functions over a classical dynamical system (Ω, ξ) and Aq's are Q × Q ordinary matrices called hopping matrices. Typically, λq's are much smaller than 1, hence the particular writing in 7.1. One will recognize in 7.1 the standard representation πωh (see 0.1) of: h =Xq φq · q ∈ C(Ω) ⊗ MQ(C) ⋊ξ Z3, φq(ω) =(cid:0)1 + λq(ω)(cid:1) ⊗ Aq. Ω = I Z3 , Ω ∋ ω = {ωx}x∈Z, For disordered crystals, the prototypical Ω is a Tychonoff space: (7.4) and the action of Z3 is provided by the shift: ξy(ω) = {ωq x+y}. I is endowed with a probability measure dPI and the product measure dP(ω) =Qx∈Z3 dPI (ωx) provides an ergodic and ξ-invariant probability measure over Ω. At its turn, dP(ω) provides a natural trace over C(Ω), hence over C(Ω) ⊗ MQ(C). For topological insulators, one can distinguish two different disorder regimes [30]: 1) The weak disorder regime, where λq's are small and some spectral gaps of H remain open. The Fermi level ǫF is fixed in the middle of such a spectral gap. 2) The strong disorder regime, where λq's are large and all the spectral gaps of H are closed but there are still regions of pure-point spectrum. ǫF is fixed in middle of such a region. Although we are not yet in the position to resolve the strong disorder regime, the machinery developed in this work is quite relevant here, as it generates an index formula for the transport coefficients. The linear conductivity tensor σ is defined by the relation Ji = σijEj , where J is the electron current set in motion by a weak electric field E. The off-diagonal components of σ were computed in [23]. In the language of crossed product algebras, it takes a form similar to that of the 1-st noncommutative Chern number: (7.5) σ12 = 2πı Xρ∈S2 (−1)ρT3np 2Yi=1 ∂ρi po, in some adjusted physical units, but note the mismatch between the degree of the cycle and the dimension of the crossed product. Above, T3 is the trace 1.8 on C(Ω) ⊗ MQ(C) ⋊ξ Z3 and p = χ(−∞,ǫF ](h) is the projector onto the spectrum of h up to ǫF . The existence of a spectral gap at ǫF , i.e the weak disorder regime, ensures that p belongs to the C∗-algebra C(Ω) ⊗ MQ(C) ⋊ξ Z3. One can already see that the righthand side of 7.5 is a pairing between a cyclic- cocycle and a projector, hence a homotopy invariant [13]. We can now complete 18 EMIL PRODAN with an index formula which will tell us where this pairing takes place. We want to point out that the methods of [32] cannot be used here because the cocycle is not trace-class when represented on a Hilbert space! Within the new framework, Z and define TA via the Fourier calculus on we identify A = C(Ω) ⊗ MQ(C) ⋊ξ3 Z2. Then C(Ω) ⊗ MQ(C) ⋊ξ3 C(Ω) ⊗ MQ(C) ⋊ξ Z3 = A ⋊ξ12 (7.6) Z (cf. 1.8), and similarly for the trace T over A ⋊ξ12 Z2 and T3 = T . Hence: ∂ρi po = Index{π− (−1)ρTna0 σ12 = 2πı Xρ∈S2 γ (p)}. 2Yi=1 γ (p)Êdx0 π+ Consequently, σ12 takes values in "T(cid:0)K0(C(Ω) ⋊ξ3 about the pattern of possible experimental values for σ12. Z)(cid:1). This provides a prediction References [1] A. Andersson. pairings http://arxiv.org/abs/1406.4078, 2014. Index for Rn-actions and Rieffel deformations. [2] J. Bellissard. Noncommutative geometry of aperiodic solids. In Geometric and Topological Methods for Quantum Field Theory, pages 86 -- 156, River Edge, NJ, 2003. World Sci. Publ. [3] J. Bellissard, A. van Elst, and H Schulz-Baldes. The non-commutative geometry of the Quan- tum Hall-Effect. J. Math. Phys., 35:5373 -- 5451, 1994. [4] E. K. Blum. A theory of analytic functions in banach algebras. Trans. Amer. Math. Soc., 78:343 -- 370, 1955. [5] M. Breuer. Fredholm theories in von Neumann algebras I. Math. Annalen, 178:243 -- 254, 1968. [6] M. Breuer. Fredholm theories in von Neumann algebras II. Math. Annalen, 180:313 -- 325, 1969. [7] R. C. Busby. Double centralizer and extensions of C ∗-algebras. Trans. Amer. Math. Soc., 132:79 -- 99, 1968. [8] A. L. Carey, V. Gayral, A. Rennie, and F. A. Sukochev. Index theory for locally compact noncommutative geometries. In Memoirs of the American Mathematical Society, volume 231. Am. Math. Soc., Providence, Rhode Islands, USA, 2014. [9] A. L. Carey, J. Phillips, A. Rennie, and F. A. Sukochev. The local index formula in semifinite von Neumann algebras I: Spectral flow. Adv. Math., 202:451 -- 516, 2006. [10] A. L. Carey, J. Phillips, A. Rennie, and F. A. Sukochev. The local index formula in semifinite von Neumann algebras II: The even case. Adv. Math., 202:517 -- 554, 2006. [11] A. L. Carey, J. Phillips, A. Rennie, and F. A. Sukochev. The local index formula in noncom- mutative geometry revisited. In Noncommutative Geometry and Physics 3, volume 3, pages 3 -- 36. World Sci. Publ., 2013. [12] A. Connes. Noncommutative differential geometry. Publications Mathematiques de l'I.H.E.S., 62:257 -- 360, 1985. [13] A. Connes. Noncommutative Geometry. Academic Press, San Diego, CA, 1994. [14] A. Connes and H. Moscovici. The local index formula in noncommutative geometry. Geom. Funct. Anal., 5:174 -- 243, 1995. [15] Kenneth R. Davidson. C ∗-algebras by example. Fields Institute Monographs. Am. Math. Soc., Providence, Rhode Islands, USA, 1996. [16] F. M. Bruckler. Tensor products of C ∗-algebras, operator spaces and Hilbert C ∗-modules. Math. Comm., 4:257 -- 268, 1999. [17] B. V. Fedosov. Analytic formulas for the index of elliptic operators. Trans. Mosc. Math. Soc., 30:159 -- 240, 1974. [18] F. D. M. Haldane. Model for a Quantum Hall-Effect without Landau levels: Condensed- matter realization of the parity anomaly. Phys. Rev. Lett., 61:2015 -- 2018, 1988. [19] N. Higson. The local index formula in noncommutative geometry. In Contemporary Develop- ments in Algebraic K-Theory, volume 15 of ICTP Lecture Notes, pages 444 -- 536, 2003. [20] G. G. Kasparov. Hilbert C ∗-modules; Theorems of Stinespring and Voiculescu. J. Operator Th., 4:133 -- 150, 1980. [21] G. G. Kasparov. The operator K-functor and extensions of C ∗-algebras. Math. USSR Izv., 16:513 -- 572, 1981. INTRINSIC CHERN-CONNES CHARACTERS 19 [22] Alexei Kitaev. Periodic table for topological insulators and superconductors. In Vladimir Lebedev and Mikhail Feigel'man, editors, Adv. Theor. Phys.: Landau Memorial Conference, volume 1134, pages 22 -- 30. AIP, 2009. [23] M. Kohmoto, B. I. Halperin, and Y.-S. Wu. Quantized hall effect in 3d periodic systems. Physica B, 184:30 -- 33, 1993. [24] J. A. Mingo. K-Theory and multipliers of stable C ∗-algebras. Trans. Amer. Math. Soc., 299:397 -- 411, 1987. [25] A. S. Miscenko and A. T. Fomenko. The index of elliptic operators over C ∗-algebras. Math. USSR Izv., 15:87 -- 112, 1980. [26] C. L. Olsen. Index theory in von Neumann algebras. In Memoirs of the American Mathe- matical Society, volume 47. Am. Math. Soc., Providence, Rhode Islands, USA, 1984. [27] S. A. Parameswaran, R. Roy, and S. L.Sondhi. FractionalquantumHallphysicsintopologi- calflatbands. C. R. Physique, 14:816 -- 839, 2013. [28] J. Phillips. Spectral flow in type I and type II factors-a new approach. In Cyclic Cohomology and Noncommutative Geometry, volume 17 of Fields Institute Communications, pages 137 -- 153, Providence, Rhode Islands, USA, 1997. Am. Math. Soc. [29] M. Pimsner, S. Popa, and D. Voiculescu. Homogeneous C ∗-extensions of C(X) ⊗ K(H). Part II. J. Operator Th., 4:211 -- 249, 1980. [30] E. Prodan. Disordered topological insulators: A non-commutative geometry perspective. J. Phys. A: Math. Theor., 44:113001, 2011. [31] E. Prodan. The non-commutative geometry of the complex classes of topological insulators. Topol. Quantum Matter, 1:1 -- 16, 2014. [32] E. Prodan, B. Leung, and J. Bellissard. The non-commutative n-th Chern number (n ≥ 1). J. Phys. A: Math. Theor., 46:485202, 2013. [33] A. Rennie. Smoothness and locality for nonunital spectral triples. K-Theory, 28:127 -- 165, 2003. [34] M. A. Rieffel. C ∗-algebras associated with irrational rotations. Pacific J. Math., 93:415 -- 429, 1981. [35] S. Ryu, A. P. Schnyder, A. Furusaki, and A. W. Ludwig. Topological insulators and super- conductors: tenfold way and dimensional hierarchy. New J. Phys., 12:065010, 2010. [36] Andreas P. Schnyder, Shinsei Ryu, Akira Furusaki, and Andreas W. W. Ludwig. Classification of topological insulators and superconductors in three spatial dimensions. Phys. Rev. B, 78:195125, 2008. [37] L. B. Schweitzer. Spectral invariance of dense subalgebras of operator algebras. Int. J. Math., 4:289 -- 317, 1993. [38] N. E. Wegge-Olsen. K-Theory and C*-Algebras. Oxford University Press, Oxford, 1993. Emil Prodan, Department of Physics, Yeshiva University, New York, NY 10016, USA E-mail address: [email protected]
1209.4092	1	1209	2012-09-18T20:04:13	An Imprimitivity Theorem for Partial Actions	[ "math.OA" ]	We define proper, free and commuting partial actions on upper semicontinuous bundles of $C^-$algebras. With such, we construct the $C^-$algebra induced by a partial action and a partial actions on that algebra. Using those action we give a generalization, to partial actions, of Raeburn's Symmetric Imprimitivity Theorem.	math.OA	math	AN IMPRIMITIVITY THEOREM FOR PARTIAL ACTIONS DAMI ´AN FERRARO Abstract. We define proper, free and commuting partial actions on up- per semicontinuous bundles of C ∗ −algebras. With such, we construct the −algebra induced by a partial action and a partial actions on that algebra. C ∗ Using those action we give a generalization, to partial actions, of Raeburn's Symmetric Imprimitivity Theorem [10]. Introduction The main idea of this article appear in the following example. Let β be a continuous, free and proper action of a locally compact and Hausdorff (LCH) group G on a LCH space Y . This gives us a continuous action of G on the continuous functions vanishing at infinity of Y, C0(Y ). If Y /G is the orbit space of Y , then Green's Theorem [11] implies C0(Y /G) is strongly Morita equivalent to the crossed product C0(Y ) ⋊β G. Now consider an open subset X ⊂ Y such that ∪{βt(X) : t ∈ G} = Y . Lets call α the restriction of β to X. That is, for every t ∈ G set αt : X ∩ βt−1 (X) → X ∩βt(X), x 7→ βt(x). This is an example of a partial action. Now consider the open set Γ := {(t, x) ∈ G × X βt−1 (x) ∈ X} ⊂ G × Y. The crossed product C0(X) ⋊α G is the closure of Cc(Γ) ⊂ Cc(G, Y ) in C0(Y ) ⋊β G. It is strongly Morita equivalent to C0(Y ) ⋊β G [2, 3]. Putting all together, we conclude that C0(X) ⋊α G is strongly Morita equivalent to C0(Y /G). The objective of the present work is to generalize the previous idea to the case where we just know X, G and α. That is, α is a partial action of G on X. The outline of this work is as follows. In Section 1 we give the definitions of free, proper and commuting partial actions and prove some basic results involving those concepts, it is based on [1, 2]. In the second section we define partial actions on up- per semicontinuous C ∗−bundles and, with such, construct the induced C ∗−algebra of a partial action and partial actions on those induced algebras. Here we follow Raeburn's work [10]. Finally, we prove our main theorem which is a generalization, to partial actions, of Raeburn's Theorem [10]. On a first read, to understand the basic ideas, we suggest the reader to consider bundles of the form X × C (X is a topological space and C the complex numbers) with trivial action on C. Date: March 1, 2018. 2010 Mathematics Subject Classification. Primary 46L05. Secondary 46L55. Key words and phrases. partial actions, crossed products, Morita equivalence. This work was supported by the CSIC and started when the author was a member of the Centro de Matem´atica of the Facultad de Ciencias, Universidad de la Rep´ublica, Uruguay. 1 2 DAMI ´AN FERRARO 1. Properties of Partial Actions Through this work the letters G, H and K will denote LCH topological groups and X, Y topological spaces. When any additional topological property is required it will be explicitly mentioned (this will never happen for the groups). This section is a brief resume of some results contained in [2] and in the PHD Thesis [1], for that reason some proof will be omitted. We start by recalling the definition of partial action. Definition 1.1 ([5, 2, 1]). A pair α = ({Xt}t∈H, {αt}t∈H) is a partial action of H on X if, for every t, s ∈ H: (1) Xt is a subset of X and Xe = X (e being the identity of H). (2) αt : Xt → Xt−1 is a bijection and αe = idX (the identity on X). (3) If x ∈ Xt−1 and αt(x) ∈ Xs−1, then x ∈ X(st)−1 and αst(x) = αs ◦ αt(x). The domain of α is the set Γα := {(t, x) ∈ H × X x ∈ Xt−1}. Recall α is continuous if Γα is open in H × X and the function, also called α, Γα → X, (t, x) 7→ αt(x), is continuous. The graph of the partial action α, Gr(α), is the graph of the function α : Γα → X. We say α has closed graph if Gr(α) is closed in H × X × X. Take two continuous partial actions of H, α and β, on the spaces X and Y respectively. A morphism f : α → β is a continuous function f : X → Y such that for every t ∈ H : f (Xt) ⊂ Yt and the restriction of βt ◦ f to Xt−1 equals f ◦ αt. Given β as before and a non empty open set Z ⊂ Y, the restriction of β to Z is the continuous partial action of H on Z given by γt : Z ∩ βt−1(Z) → Z ∩ βt(Z), z 7→ βt(z). Up to isomorphism of partial actions, every continuous partial action can be obtained as a restriction of a global action. That is, given α as before there exits a global and continuous action of H on a topological space Y, β, and an open set Z ⊂ Y such that α is isomorphic to the restriction of β to Z. If in addition Y = ∪{βt(Z) t ∈ H}, we say β is an enveloping action of α. Enveloping actions exists and are unique up to isomorphism of (partial) actions [2, 1]. The enveloping1 action of α is denoted αe and the space where it acts X e, we also think X is an open set of X e and α is the restriction of αe to X. The orbit of a subset U ⊂ X by α is the set HU := ∪{αt(U ∩ Xt−1) t ∈ G}. The orbit of a point x ∈ X is the orbit of the set {x} and is denoted Hx. If we want to emphasize the name of the action we write αHx. The orbits of two points are equal or disjoint and the union of all of them is equal to X. With this partition of X we construct the quotient space X/H with the quotient topology, this is the orbit space of α. The canonical projection X → X/H is continuous, surjective and open. The function X/H → X e/H, αHx → αeHx is a homeomorphism. Raeburn's Symmetric Imprimitivity Theorem involves free, proper and commut- ing actions. We now give the corresponding definitions for partial actions. We refer the reader to [1] to a more detailed exposition of these concepts. The stabilizer of a point x ∈ X is the set Hx := {t ∈ H x ∈ Xt−1, αt(x) = x}. It is easy to see that Hx is a subgroup of H, not necessarily closed if the action is not global. A partial action is free is the stabilizer of every point is the set {e}. A partial action is free if and only if it's enveloping action is free. 1We say "the" enveloping action because it is unique up to isomorphisms. AN IMPRIMITIVITY THEOREM FOR PARTIAL ACTIONS 3 We say α and β commute if for every (s, t) ∈ H × K (i) αs(X H The next concept we define is commutativity. We will have two continuous partial actions, α and β, of H and K, on X. As we do not want any confusions, we will use the notation αs : X H t , for s ∈ H and t ∈ K. t ) = βt(X K t−1). This definition expresses the fact that we can compute αsβt(x) if and only if we can compute βtαs(x), and in that case αsβt(x) = βtαs(x). As we can see, if both actions are global, this is the usual notion of commuting actions. s ) and (ii) αs ◦ βt(x) = βt ◦ αs(x), for every x ∈ αs−1 (X H s−1 ∩ X K s ∩ X K s and βt : X K s−1 → X H t−1 → X K t−1 ∩ X H Recall [6] a subset U ⊂ X is said α−invariant if αt(X H t ∈ H. Condition (i) of the previous definition implies X K s s ∈ K. t−1 ∩ U ) ⊂ U , for every is α−invariant for every An important property of commuting global actions is that we can define an action of the product group, this is also true for partial actions. Lemma 1.2 (cf. [1] Proposi¸cao 4.35). If α and β commute then there is a contin- uous partial action, α × β, of G := H × K on X such that, for every (s, t) ∈ G, (1) X(s,t) = βt(X K (2) α × β(s,t) = αs ◦ βt. t−1 ∩ X H s ) = αs(X H s−1 ∩ X K t ) and Proof. The fact that µ := α × β is a partial action (not necessarily continuous) is an easy consequence of the fact that α and β are commuting partial actions ([1] Proposi¸cao 4.35). We just have to deal with the continuity. To show Γµ is open in G×X notice that the set Γ−1 β := {(t, x) ∈ K ×X x ∈ X K t } is open in K × X, and define: πH : H × K × X → H × X, (h, k, x) 7→ (h, x), πK : H × K × X → K × X, (h, k, x) 7→ (k, x), and F : π−1 β ), (h, k, x) 7→ (h, k, βk(x)). K (Γβ) → π−1 K (Γ−1 It is easy to see that the three functions are continuous. So, the domain and range of F are open in H × K × X and Γµ = F −1(π−1 H (Γα)) is open in G × X. Finally, the continuity of µ : Γµ → X follows from that of α : Γα → X and β : Γβ → X. (cid:3) β ) ∩ π−1 K (Γ−1 Here is another property of partial actions we will use. Lemma 1.3. If α and β commute then there is a partial action of H on X/K, s−1. s and (1) (X/K)s := KX H F : Γα → H × X/K as F (s, x) = (s, Kx). This map is open and continuous. The (2) bαs(Kx) = Kαs(x) for any x ∈ X H called bα, such that for every s ∈ H Proof. The first step is to show we can define bα as in (1) and (2). Define the function domain of bα will be the image of F , Γ bα := Im(F ), which is an open set. Now define S : Γα → X/K in such a way that (s, x) 7→ Kαs(x), and consider (on Γα) the equivalence relation u ∼ v if F (u) = F (v). The function S is constant in the classes of ∼, and the quotient space Γα/ ∼ is homeomorphic to Γ bα trough the map defined by F . So, there is a unique continuous map Γ bα → X/K such that (s, Kx) 7→ Kαs(x). This is the partial action bα we are looking for. It remains to be shown that bα is a partial action. Properties (1) and (2) of Definition 1.1 are easy to prove, for (3) recall every X H s is β−invariant. (cid:3) 4 DAMI ´AN FERRARO Assume for a moment we have a continuous global action of H on X. It is immediate that it's domain, being equal to H × X, is a closed and open (clopen) set of H × X. That is not always the case for partial actions. Definition 1.4. A partial action, α, of H on X has closed domain if Γα is closed in H × X. Lemma 1.5. If X is Hausdorff and α is continuous, the following conditions are equivalent: (1) α has closed domain. (2) The enveloping space X e is Hausdorff and X is closed in X e. Proof. We start by proving (1)⇒(2). Recall [2] X e is Hausdorff if α has closed graph. Consider the function F : H × X × X → H × X, (s, x, y) 7→ (s, x). The set F −1(Γα) is closed in H × X × X. Now, Gr(α) is closed in F −1(Γα) because it is the pre-image of the diagonal {(x, x) x ∈ X} ⊂ X × X by the continuous function F −1(Γα) → X × X, (t, x, y) 7→ (αt(x), y). This implies α has closed graph. To show X is closed in X e take a net contained in X, {xi}i∈I , converging to a t (x) ∈ X. By the continuity of αe t (x)), being the point x ∈ X e. There exists t ∈ H such that αe there is an i0 such that (t−1, αe limit of {(t−1, αe t (xi)) ∈ Γα for i > i0. Then (t−1, αe t (xi))}i>i0 , belongs to ∈ Γα. Finally x = αt−1αe t (x) ∈ X. For the converse notice three facts: the topology of H ×X is the topology relative to H × X e, (αe)−1(X) is closed in H × X e and Γα = H × X ∩ (αe)−1(X). So we clearly have that Γα is closed in H × X. (cid:3) The previous lemma characterizes the continuous partial actions with closed domain on Hausdorff spaces, as those arising as the restriction of a global action on a Hausdorff space to a clopen set. Lemma 1.6. Given two continuous and commuting partial actions, both with closed domain, the partial action of the product group (as defined on Lemma 1.2) has closed domain. Proof. In the proof of Lemma 1.2 we showed Γα×β is open, use the same arguments changing the word "open" for "closed". (cid:3) A dynamical system (DS for short) is a tern (Y, H, β) where β is a continuous action of H on Y , where H and Y are LCH. The natural extension to partial actions is the following one. Definition 1.7. The tern (X, H, α) is a partial dynamical system (PDS) if α is a continuous partial action of H on X and both (H and X) are LCH. Lemma 1.8. If (X, H, α) is a PDS then (X e, H, αe) is a DS if and only if α has closed graph. Proof. By Theorem 1.1. of [2] every point of X e has a neighbourhood homeomor- phic to X. So, every point of X e has a local basis of compact neighbourhoods. As α is continuous and H is LCH, (X e, H, αe) is a DS if and only if X e is Hausdorff. By Proposition 1.2. of [2] X e is Hausdorff if and only if α has closed graph. (cid:3) A PDS (X, H, α) is proper if the function Fα : Γα → X, (t, x) 7→ (x, αt(x)), is proper (the pre-image of a compact set is compact). This definition, and part of the next Lemma, are taken from [1]. AN IMPRIMITIVITY THEOREM FOR PARTIAL ACTIONS 5 Lemma 1.9. Given a PDS (X, H, α), the following statements are equivalent: (1) The system is proper. (2) Every net contained in Γα, {(ti, xi)}i∈I , such that {(xi, αti (xi))}i∈I con- verges to some point of X × X, has a subnet converging to a point of Γα. (3) α has closed graph and the enveloping DS (X e, H, αe) is proper. Proof. The equivalence between (1) and (3) is proved in [1] (Proposicao 4.62). The equivalence between (1) and (2) is proved as in Lemma 3.42 of [13]. (cid:3) It is a known fact that the orbit space of a proper DS is a LCH space, this is also true for PDS. Lemma 1.10. If (X, H, α) is a proper PDS then X/H is a LCH space. Proof. By the previous Lemma (X e, H, αe) is a proper DS. So, X e/H is LCH. But X/H is homeomorphic to X e/H and so is LCH. (cid:3) The next result follows immediately from the previous ones. Lemma 1.11. Let (X, H, α) and (X, K, β) be commuting PDS (that is, α and β commute). If β is proper then (X/K, H,bα) is a PDS, where bα is the partial action defined on Lemma 1.3. 2. Partial actions on bundles of C ∗−Algebras. The definition of upper semicontinuous C ∗−bundle we are going to use is Def- inition C.16 of [13] (notice that we do not require the base space to be Haus- dorff). From now on B = {Bx}x∈X and C = {Cy}y∈Y will be upper semicontin- uous C ∗−bundles. The projections of B and C will be denoted p : B → X and q : C → Y , respectively. The set of continuous and bounded sections of the bundle B will be denoted Cb(B) (this notation differs from that of [13]). Similarly, C0(B) is the set of con- tinuous sections vanishing at infinity (C.21 [13]) and Cc(B) the set of continuous sections of compact support. When X is a LCH space, Cb(B) and C0(B) are C ∗−algebras with the supremum norm and Cc(B) is a dense ∗−sub algebra of C0(B). Definition 2.1. A partial action of H on B is a pair (α, ·), where α and · are continuous partial actions of H on B and X, respectively, satisfying (1) p−1(Xt) = tB for every t ∈ H. Here tB is the range of αt. (2) p is a morphism of partial actions (Definition 1.1 of [2]). (3) The restriction of αt to a fiber is a morphism of C ∗−algebras, for each t ∈ H. Notice that · is determined by α. For that reason, with abuse of notation, we name α the pair (α, ·). We say α is global if the partial action on the total space is a global action or, what is the same, if · is global. The domain of the partial action on the total and base space will be denoted Γ(B, α) and Γ(X, α), respectively. Example 2.2. Let (X, H, ·) be a PDS and (A, H, γ) a C ∗−DS, that is, A is a C ∗−algebra and γ : H → Aut(A) is a strongly continuous action. With such define the trivial bundle p : A × X → X, where p(a, x) = x. All the fibers of this bundle, called B, are isomorphic to A by the maps A → Bx, a 7→ (a, x). We define a global action of H on B by setting αt : A × Xt−1 → A × Xt, (a, x) 7→ (γt(a), t · x). 6 DAMI ´AN FERRARO I would like to emphasize that, from now on, we are going to use the letters α and β for actions on total spaces. The actions on the base spaces will be denoted · and ⋆. We will write αt(a) and t · x, similarly with β and ⋆. If β = (β, ⋆) is a partial action of H on C, a morphism (F, f ) : α → β is a pair of continuous functions, F : B → C and f : X → Y , such that: both are morphism of partial actions, q ◦ F = f ◦ p and the restriction of F to each fiber is a morphism of C ∗−algebras. Naturally, the composition of morphisms is the composition of functions (on each coordinate). Following [2] we can define the restriction of actions. Let β = (β, ⋆) be a global action of H on B and U an open subset of X. Consider the restriction bundle BU = {Bu}u∈U with the partial action βU , which is the pair formed by the restriction of the actions of H to p−1(U ) and U . Notice we have obtained a partial action because p−1(U ) ∩ βt(p−1(U )) = p−1(U ∩ t ⋆ U ), for every t ∈ H. Rephrasing Theorem 1.1 of [2] we get Theorem 2.3. For every continuous partial action α of H on an upper semicon- tinuous C ∗−bundle B, there exists a tern (ιX , ιB, αe) such that αe is an action of H on an upper semicontinuous C ∗−bundle Be, and (ιX , ιB) : α → αe is a mor- phism, such that for any morphism ψ : α → β, where β is an action of H (on an upper semicontinuous C ∗−bundle), there exists a unique morphism ψe : αe → β such that ψe ◦ (ιX , ιB) = ψ. Moreover, the pair (ιX , ιB, αe) is unique up to canonical isomorphisms, and (1) ιX (X) is open in X e. (2) (ιX , ιB) : α → (αe)ιX (X) is an isomorphism. (3) X e is the orbit of ιX (X). (4) Be is a continuous C ∗−bundle if and only if B is. Proof. Let (ιX , ·e) and (ιB, αe) be the pairs given by Theorem 1.1 of [2] for · and α. We also have a morphism pe : αe → ·e. Notice pe is surjective because is a morphism and the orbit of ιX (X) equals X e. Again, as Be is the orbit of ιB(B) and pe is a morphism, to prove pe is open we only have to see that pe ◦ αe t ◦ ιB is open, for every t ∈ H. But this is true because, if U is open in B pe ◦ αe t ◦ ιB(U ) = t ·e (ιX (p(U ))), the last being an open set. We have proved Be fibers over X e. We now give a structure of C ∗−algebra to each fiber of Be. Let x be an element of X e, take t ∈ H such that t·e x ∈ ιX (X) and define the C ∗−structure on Be X (t·ex) → Be x an isomorphism of C ∗−algebras. This is independent of the choice of t because α acts as isomorphism of C ∗−algebras on the fibers of B. x as the unique making αe t−1 ◦ ιB : Bι−1 {b ∈ Be : kbk < ε} equals the open set St∈H αe To prove the norm of Be is semicontinuous notice that, given ε > 0, the set t ◦ ιB({b ∈ B : kbk < ε}). In fact, a similar argument shows the norm of Be is continuous if and only if the norm of B is continuous. This suffices to prove property (4) of the thesis. We now indicate how to prove the continuity of the product, for the other oper- ations there are analogous proofs. Set De := {(a, b) ∈ Be × Be : pe(a) = pe(b)}. We prove the continuity of De → Be, (a, b) 7→ ab, locally. Fix (a, b) ∈ De, we may assume p(a) = t ·e x for some x ∈ X and t ∈ H. The product is continuous on (a, b) because U := (αt ◦ ιB(B) × αt ◦ ιB(B)) ∩ De is open in De, and the restriction of AN IMPRIMITIVITY THEOREM FOR PARTIAL ACTIONS 7 the product to U is the continuous function (c, d) 7→ αe t ◦ ιB(cid:2)(ιB)−1 ◦ αe t−1(c) + (ιB)−1 ◦ αe t−1 (d)(cid:3) . Up to here we have constructed an upper semicontinuous C ∗−bundle Be = x}x∈X e. By the previous construction we also have that (αe, ιe) is a global {Be action of H on Be, and (ιX , ιB) : α → αe is a morphism. Except for property (2) of the thesis, everything follows immediately from the previous constructions and Theorem 1.1 of [2]. To prove property (2) it suffices to see that (pe)−1(ιX (X)) = ιB(B). We clearly have the inclusion ⊃, for the other one let b ∈ (pe)−1(ιX (X)). We may suppose b = αe t (ιB(c))) ∈ ιB(B), we have t (ιB(c)) for some c ∈ B and t ∈ H. As pe(b) = pe(αe pe(b) = pe(αe t (ιB(c))) = t ·e ιX (p(c)) ∈ ιX (X). So, p(c) ∈ Xt−1. This implies b = ιB(αt(c)) ∈ ιB(B). (cid:3) The non commutative analogue of PDS's are the C ∗−PDS's, they are terns (A, G, γ) formed by a C ∗−algebra A, a LCH group G and a partial action γ of G on A (Definition 2.2 of [2], for a more general definition see [5]). We know every PDS gives us a C ∗−PDS with commutative algebra [2]. Follow- ing that construction, we are going to use partial actions on upper semicontinu- ous C ∗−bundles over LCH spaces to construct partial actions on the C ∗−algebras C0(B). The ideals are of the form C0(B, U ) := {f ∈ C0(B) f (x) = 0x if x /∈ U }, for open sets U ⊂ X. Theorem 2.4. Let X be a LCH space, B = {Bx}x∈X an upper semicontinuous C ∗−PDS, where x 7→ kf (t−1 · x)k, vanishes at infinity. (1) C0(B)t = C0(B, Xt), for every t ∈ H. Proof. First of all we have to show that, given t ∈ H and f ∈ C0(B)t−1 , the C ∗−bundle and α a continuous partial action of H on B. Then (C0(B), H,eα) is a (2) If f ∈ C0(B)t−1 then eαt(f )(x) = αt(f (t−1 · x)) if x ∈ Xt and 0x otherwise. function eαt(f ) belongs to C0(B)t. It is clear that eαt(f ) is a section that vanishes outside Xt. Besides, the function Xt → R, x 7→ kbαt(f )(x)k, being equal to Xt → R, Clearly bαt(f ) is continuous on Xt and in the interior of the complement of Xt. To prove the continuity of bαt(f ) it suffices to show that given a net {xi}i∈I ⊂ Xt converging to a point x /∈ Xt, we have keαt(f )(xi)k → 0. Notice that the function outside every compact of Xt−1, we conclude keαt(f )(xi)k = kf (t−1 · xi)k → 0. The next step is to show bα is a partial action (Definition 1.1). We omit the proof Xt−1 → R, y 7→ kf (y)k, vanishes at infinity and the net {t−1 · xi}i∈I is eventually To prove {C0(B)t}t∈H is a continuous family [5], let U be an open set of C0(B) and fix t ∈ H such that C0(B)t ∩ U 6= ∅. By the Urysohn Lemma we can find g ∈ C0(B)t ∩ U with compact support. As the domain of the partial action on X is an open set, there is an open set containing t, V , such that Xr contains the support of g for every r ∈ V . Then V is an open set containing t and contained in {r ∈ H C0(B) ∩ U 6= ∅}. Now we deal with the continuity of eα. Let {(ti, fi)}i∈I be a net contained in Γ eα converging to (t, f ) ∈ Γ eα. Given ε > 0 there exists g ∈ Cc(B), with support contained in Xt−1 , such that kf − gk < ε 3 (by the Urysohn Lemma). of this fact because it is an easy task. 8 DAMI ´AN FERRARO We can find an i0 ∈ I such that supp(g) ⊂ Xt−1 i i > i0. Then, for every i > i0 and kfi − gk < ε 3 , for every keαt(f ) −eαti (fi)k 6 keαti(fi − g)k + keαti(g) −eαt(g)k + keαt(f − g)k < 2ε 3 + keαti(g) −eαt(g)k. To complete the proof it suffices to see that limi keαti (g) − eαt(g)k = 0. To this purpose let D be a compact containing t · supp(g) on it's interior, and contained in Xt. We may find i1 (larger than i0) such that ti · supp(g) ⊂ D and D ⊂ Xti, for every i > i1. Given i > i1 we have · x)) − αt(g(t−1 · x))k : x ∈ D}. keαti(g) −eαt(g)k = sup{kαti(g(t−1 i As D is compact, it suffices to prove eαti (g) converges point wise to eαt(g), which is an easy consequence of Lemma C.18 of [13] and the continuity of α and g. (cid:3) The definitions of proper, free and commuting partial actions on bundles are the following ones. Definition 2.5. Given an upper semicontinuous C ∗−bundle over a LCH space and a partial action of a LCH group on the bundle, we say the partial action is free, proper, has closed graph or has closed domain if the partial action on the base space has the respective property. Similarly, given two partial actions on an upper semicontinuous C ∗−bundle we say they commute if the partial actions on the total and base space commute. Relating the concepts of enveloping action, in the contexts of C ∗−algebras and bundles, we have the following result. Theorem 2.6. Let α be a partial action of H on the upper semicontinuous C ∗- bundle B = {Bx}x∈X. If X is LCH, α has closed graph, αe is the enveloping action of α and Be the enveloping bundle, then (C0(Be), H,fαe) is the enveloping system of (C0(B), H,eα) (Definition 2.3 of [2]). So C0(B) ⋊ eα H is a hereditary and full sub C ∗−algebra of C0(Be) ⋊fαe H. In particular those crossed products are strongly Morita equivalent. Proof. By Theorem 2.3 we may suppose X ⊂ X e, B ⊂ Be, pe(B) = X and that p is the restriction of pe to B. Now, by Lemma 1.8, X e is LCH. This considerations allows us to identify the bundle B with the restriction of Be to X, which gives C0(B) = C0(Be, X). We have identified C0(B) with an ideal of C0(Be). We also have, for every t ∈ H, = C0(Be, Xt) = C0(B)t; t(C0(Be, X)) ∩ C0(Be, X) = C0(Be, t · X) ∩ C0(Be, X) = C0(Be, X ∩ t · X) fαe and clearly the restriction of fαe fαe−orbit of C0(B) is dense in C0(Be). [3], the only thing that remains to be showed is that the space generated by the t to C0(B)t−1 equals eαt. So, using Corollary 1.3 of We show every continuous section with compact support of Be is a finite sum of points in the orbit of C0(B). Fix f ∈ Cc(Be). The support of f has an open cover by sets of the form t ·e X, t varying in H. We can find t1, . . . , tn ∈ H and h1, . . . , hn ∈ Cc(X e) such that: 0 6 h1 + · · ·+ hn 6 1, the support of hi is contained AN IMPRIMITIVITY THEOREM FOR PARTIAL ACTIONS 9 in ti · X (i = 1, . . . , n) and h1(x) + · · · + hn(x) = 1 if x ∈ supp(f ). Defining fi(x) = (fi) ∈ Cc(B) tn (gn), that gives the desired (cid:3) hi(x)f (x) (i = 1, . . . , n), we have f = f1 + · · · + fn and gi := fαe for every i = 1, . . . , n. Besides, f = fαe result. t1(gi) + · · · +fαe t−1 i We can reproduce most of the results of Section 1 in this context. For example, the next Theorem is a direct consequence of Lemma 1.2. Theorem 2.7. Given an upper semicontinuous C ∗−bundle B and commutative partial actions, (α, ·) and (β, ⋆) of H and K on B, respectively, the pair (α, ·) × (β, ⋆) := (α × β, · × ⋆) is a partial action of H × K on B. Writing α = (α, ·) and β = (β, ⋆), the product α × β is the one defined in the previous Theorem. 2.1. Orbit bundle. There is a notion of "orbit bundle", analogous to the notion of "orbit space", but to construct it we have to consider proper and free partial actions. Fix an upper semicontinuous C ∗−bundle over a LCH space, B = {Bx}x∈X, and a proper and free partial action, α, of H on B. Let B/H and X/H be the orbit spaces and πB : B → B/H and πX : X → X/H be the orbit maps. As p is a morphism of partial actions, there is a unique continuous (also open and surjective) function pα : B/H → X/H such that πX ◦ p = pα ◦ πB. We want to equip B/H with operations making B/H := (B/H, X/H, pα) an upper semicontinuous C ∗-bundle. To do this first notice that, given Hx ∈ X/H, the fiber (B/H)xH is homeomorphic to Bx trough the restriction of πB to Bx (because the partial action on X is free). Call that map hx : Bx → (B/H)xH. Define the structure of C ∗−algebra of (B/H)xH in such a way that hx is an isomorphism of C ∗−algebras. This definition is independent of the choice of x because, if Hy = Hx, then h−1 y ◦ hx : Bx → By is an isomorphism. As it is the restriction of αt to Bx, t ∈ H being the unique such that x ∈ Xt−1 and t · x = y. The norm of B/H is the function k k : B/H → R, Ha 7→ kak. To prove it is upper semicontinuous let ε be a positive number. The set {Hb ∈ B/H kHbk < ε} is open because it equals the open set πB({b ∈ B kbk < ε}). Similarly, we prove k k : B/H → R is continuous if k k : B → R is continuous. To prove the continuity of the product and the sum name D the set of points (a, b) ∈ B × B such that Ha = Hb. If (a, b) ∈ D there is a unique t ∈ H, which we name t(p(a), p(b)), such that p(a) ∈ Xt−1 and t · p(a) = p(b). Hence, αt(a) and b are in the same fiber, we define S(a, b) := αt(a) + b and P (a, b) := αt(a)b. To prove the continuity of S and P we only have to prove the continuity of the function F : {(x, y) ∈ X × X : Hx = Hy} → Γ(X, α), F (x, y) = (t(x, y), x). Call DX the domain of F . Consider the function R : Γ(X, α) → X × X given by R(t, x) = (x, t · x), this is a continuous, proper and injective function between LCH spaces. Such functions are homeomorphisms over its image, but the image of R is DX and F = R−1. So, F is continuous. Once we have proved the continuity of S and P , using the freeness of the partial action on X, we prove they are constant in the classes of the equivalence relation 10 DAMI ´AN FERRARO on D : (a, b) ∼ (c, d) if Ha = Hc and Hb = Hd. The space D/ ∼ is (homeomorphic to) D′ := {(a, b) ∈ B/H × B/H : pα(a) = pα(b)}, and the functions defined by S and P on D′ are exactly the sum and product of B/H. We have proved they are continuous, for the rest of the operations we proceed in a similar way. The last step, to show B/H is an upper semicontinuous C ∗−bundle is to prove it satisfies the following property: for every net {bi}i∈I ⊂ B/H such that kbik → 0 and pα(bi) → z, for some z ∈ X/H, we have bi → 0z. Let {bi}i∈I be a net as before, it suffices to show it has a subnet converging to 0z. There is a net in B, {ai}i∈I , such that bi = Hai for every i ∈ I. We have that Hp(ai) = pα(bi) → Hx, where x ∈ X is such that Hx = z. As the orbit map X → X/H is open and surjective, Proposition 13.2 Chapter II of [7] implies there is and tj ·p(aij ) → x. a subnet {aij }j∈J and a net {tj}j∈J ⊂ H such that p(aij ) ∈ Xt−1 (aij )k = kbij k → 0, (aij )) → x. But also kαtij This implies aij ∈ t−1 (aij ), πB(0x) = H0x = 0z and πB is so, αtij continuous, 0z is a limit point of {bij }j∈J . (aij ) → 0x. Finally, as bij = Hαtij j B and p(αtij j Definition 2.8. The orbit bundle of B by α is the upper semicontinuous C ∗−bundle B/H constructed before. Example 2.9. Consider the situation of Example 2.2 where the action of H on A is the trivial one (γt = idA for every t ∈ H) and the system (X, H, ·) is free and proper. Then the quotient bundle B/H is isomorphic to the trivial bundle A×X/H. Notice that C0(B/H) is isomorphic to C0(X/H, A), the set of continuous functions from X/H to A vanishing at infinity. Our next goal is to identify C0(B/H) with a C ∗−sub algebra of Cb(B). Ev- ery function f ∈ Cb(B), which is also a morphism of partial actions, induces a continuous and bounded section Indb(f ) : X/H → B/H, given by Hx 7→ Hf (x). The induced algebra Indb(B, α) is the subset of Cb(B) formed by all the sections which are also morphism of partial actions. There is a natural map Indb : Indb(B, α) → Cb(B/H), f 7→ Indb(f ). Similarly, the algebra Ind0(B, α) is the pre image of C0(B/H) under Indb. The function Ind0 is simply the restriction of Indb to Ind0(B, α). In fact, the induced algebras are C ∗−sub algebras of Cb(B). To prove this it suffices to show Indb(B, α) is a C ∗−sub algebra and to notice Indb is a morphism of C ∗−algebras. The non trivial fact is that Indb(B, α) is closed in Cb(B). Assume {fn}n∈N is a sequence contained in Indb(B, α) converging to f . Choose some t ∈ H and x ∈ Xt−1. Even if B is not Hausdorff, Bx and Bt·x are, so we have the following equalities αt(f (x)) = lim n αt(fn(x)) = lim n fn(t · x) = f (t · x). Theorem 2.10. The functions Indb : Indb(B, α) → Cb(B/H) and Ind0 : Ind0(B, α) → C0(B/H) are isomorphism of C ∗−algebras. Proof. The only thing to prove is that Indb is surjective (it is injective because is an isometry). Fix g ∈ Cb(B), we will construct f ∈ Indb(B, α) such that Indb(f ) = g. AN IMPRIMITIVITY THEOREM FOR PARTIAL ACTIONS 11 As the action on the base space is free, for every x ∈ X there is a unique f (x) ∈ Bx such that Hf (x) = g(Hx). Clearly f is a bounded section. To prove f is continuous, let {xi}i∈I be a net contained in X converging to It suffices to find a subnet {xij }j∈J such that f (xij ) → f (x). By the x ∈ X. continuity of g the net {Hf (xi)}i∈I has Hf (x) as a limit point. As the orbit map B → B/H is open, there is a subnet {xij }j∈J and a net {tj}j∈J such that {(tj, f (xij ))}j∈J ⊂ Γ(B, α) and αtj (f (xij )) → f (x). This implies {(tj, xij )}j∈J ⊂ Γ(X, α) and tj ·xij → x. Then tj = t(xij , tj ·xij ) → t(x, x) = e (see the construction of the orbit bundle in Section 2.1). Finally, the net {f (xij )}j∈J , being equal to {αt−1 αtj (f (xij ))}j∈J , has f (x) as a limit point. j It remains to prove f is a morphism of partial actions. Clearly f (Xt) ⊂ tB for every t ∈ H. Now take t ∈ H and x ∈ Xt−1. The points f (t · x) and αt(f (x)) are, both, the unique point of Bt·x in the class of g(Hx), so they are equal. (cid:3) Theorem 2.11. Let B = {Bx}x∈X be an upper semicontinuous C ∗−bundle over a LCH space and α and β be partial actions of H and K on B, respectively. If α is free and proper, then (Ind0(B, α), K,bβ) is a C ∗−PDS where (1) Ind0(B, α)t := {f ∈ Ind0(B, α) : x 7→ kf (x)kvanishes outside X H t }. t and 0x (2) For every f ∈ Ind0(B, α)t−1 bβt(f )(x) = βt(f (t−1 ⋆ x)) if x ∈ X K otherwise. Proof. Let B/H be the orbit bundle. As α commutes with β, using Lemmas 1.3 and 1.11, we define a partial action, µ, of K on B/H. the map Ind0 : Ind0(B, α) → C0(B/H) is an isomorphism. Notice Ind0(B, α)t is the pre image of C0(B/H)t. The partial action of the thesis is the unique making (cid:3) By Theorem 2.4, µ defines a C ∗−PDS (C0(B/H), K,eµ). Lemma 2.10 ensures Ind0 : bβ → eµ an isomorphism of partial actions. 3. Morita equivalence In our last section we prove our main theorem, which is a generalization of Raeburn's and Green's Symmetric Imprimitivity Theorems [10, 11]. The first task is to translate Raeburn's result to the language of actions on bundles. Consider two C ∗−DS (A, H, γ) and (A, K, δ), and two proper and free DS (X, H, ·) and (X, K, ⋆). Assume also that the actions on A and X commute. On the trivial bundle B = A × X define the actions of H and K as in Example 2.2, call them α and β, respectively. Let Ind γ be the induced C ∗-algebra defined as in [10]. We have an isomor- phism ρ : Ind γ → Ind0(B, α), given by ρ(f )(x) = (f (x), x). This isomorphism takes the action of K on Ind γ (as defined on [10]) into the action bβ. By using Raeburn's Theorem we conclude that Ind0(B, α) ⋊ bβ K is strongly Morita equiva- lent to Ind0(B, β) ⋊ bα H. Our purpose is to give a version of this result for partial actions. We will write A ∼M B whenever A and B are strongly Morita equivalent C ∗−algebras [12]. 3.1. The main Theorem. From now on we work with two LCH topological groups, H and K, an upper semicontinuous C ∗−bundle with LCH base space, B = {Bx}x∈X, and two continuous, free, proper and commuting partial actions, α and β, of H and K on B, respectively. 12 DAMI ´AN FERRARO We want to give conditions under which we can say that Ind0(B, α) ⋊ bβ K is strongly Morita equivalent to Ind0(B, β) ⋊ bα H. For global actions, with some additional hypotheses on the group and the base space, this is proved in [4], [9] or [8]. In fact, the proof of the next Theorem is a minor modification of Raeburn's proof the Symmetric Imprimitivity Theorem [10]. Theorem 3.1. If α and β are global actions then Ind0(B, α) ⋊ bβ K ∼M Ind0(B, β) ⋊ bα H. Proof. Define E := Cc(H, Ind0(B, β)) and F := Cc(K, Ind0(B, α)), viewed as dense ∗−sub-algebras of the respective crossed products. Define also Z := Cc(B), which will be an E − F −bimodule with inner products; whose completion implements the equivalence between Ind0(B, α) ⋊ bβ K and Ind0(B, β) ⋊ bα H. For f, g ∈ Z, b ∈ E and c ∈ F define (3.1) (3.2) (3.3) (3.4) b · f (x) :=ZH b(s)(x)eαs(f )(x)∆H (s)1/2ds, f · c(x) :=ZK eβt(f c(t−1))(x)∆K (t)−1/2dt, Ehf, gi(s)(x) := ∆H (s)−1/2ZK eβt (feαs(g∗)) (x)dt, hf, giF (t)(x) := ∆K(t)−1/2ZH eαs(cid:16)f ∗eβt(g)(cid:17) (x)ds. The integration is with respect to left invariant Haar measures; ∆H and ∆K are on Theorem 2.4. We now justify the fact that b · f ∈ Z. The function H → C0(B), given by the modular functions of the groups. Here eα and eβ are the partial actions defined s 7→ b(s)eαs(f )∆H (s)1/2, is continuous (Theorem 2.4). Besides, it's support is contained in the support of b and so we can integrate it. This integral is exactly b · f . Finally, notice supp(b · f ) ⊂ {s · x : (s, x) ∈ supp(b) × supp(f )}, the last being a compact set. To prove (3.3) defines an element of E we proceed as follows. Fixed s ∈ H and x ∈ X the function K → Bx, given by t 7→ eβt (feαs(g∗)) (x), is continuous with support contained in the compact {t ∈ K : t−1 ⋆ x ∈ supp(f )}. So, the function is integrable. The value of that integral is Ehf, gi(s)(x). We now prove Ehf, gi is continuous, what we do locally. Fix some s0 ∈ H and x0 ∈ X. Take compact neighbourhoods, V of s0 and W of x0. The bundle BW will be the restriction of B to W . Define the function F : V × H → C(BW ) by support. By integrating, with respect to the second coordinate, we get a continuous F (s, t)(x) = eβt (feαs(g∗)) (x). As the action of K on X is proper, F has compact function R ∈ C(V, C(BW )), defined by R(s) =RK F (s, t)dµK(t) ([7] II.15.19). Fixing (s, x) ∈ V × W , we have R(s)(x) = Ehf, gi(s)(x). From this follows the continuity of Ehf, gi. An easy calculation shows Ehf, gi(s)(t · x) = βt(Ehf, gi(s)(x)), for every t ∈ K and x ∈ X. Besides, if Ehf, gi(s)(x) 6= 0x, then x belongs to the K−orbit of supp(f ), and s to the compact set {s ∈ H : s · supp(g) ∩ supp(f ) 6= ∅}. We have proved that Ehf, gi ∈ E. The computations needed to prove equations (3.1)-(3.4) define an equivalence bi-module are the same as in [10] or [13]. For the construction of the approximate AN IMPRIMITIVITY THEOREM FOR PARTIAL ACTIONS 13 unit, analogous to that of Lemma 1.2 of [10], follow the proof of Proposition 4.5 of [13], recalling IndP (cid:3) H (A, β) plays the role of our Ind0(B, β). Our next step is to let α and β to be partial, but to have the same result we need additional hypotheses, which are trivially satisfied in the previous case. Let α × β be the partial action given by Theorem 2.7. Now, by Theorem 2.3, we have an enveloping action (α × β)e and an enveloping bundle Be. We can assume B is the restriction of Be to X ⊂ X e. For the action given by (α × β)e on X e we will use the notation (s, t)x, for (s, t) ∈ H × K and x ∈ X e. Define σ and τ as the restriction of (α × β)e to H and K, respectively (identify H with H ×{e} ⊂ H ×K). It is immediate that (α×β)e = σ ×τ , σ and τ commute, and that α (β) is the restriction of σ (τ ) to B. The next is the main Theorem of this article. Theorem 3.2. If α × β has closed graph and σ and τ are proper then Ind0(B, α) ⋊ bβ K ∼M Ind0(B, β) ⋊ bα H. Proof. To show that σ (and also τ ) is free. Assume (s, e)x = x for some s ∈ H and x ∈ X e. As X e is the H × K−orbit of X, there exists (h, k) ∈ H × K such that (h, k)x ∈ X. Notice that (hsh−1, e)(h, k)x = (h, k)x ∈ X ∩ (hsh−1, e)−1X, so, hsh−1 · (h, k)x = (h, k)x and hsh−1 = e. We conclude s = e. As α×β has closed graph, Be is an upper semicontinuous C ∗−bundle over a LCH space. The hypotheses, together with Theorems 2.11 and 3.1, imply Ind0(Be, σ) ⋊bτ K is strongly Morita equivalent to Ind0(Be, τ ) ⋊bσ H. The proof of our Theorem will be completed if we can show that Ind0(B, α) ⋊ bβ K is strongly Morita equivalent to Ind0(Be, σ) ⋊bτ K, because, by symmetry, the same will hold changing α for β, σ for τ and H for K. Ind0(B, α) ⋊ bβ K is isomorphic to C0(B/H) ⋊eµ K and Ind0(Be, σ) ⋊bτ K isomorphic to C0(Be/H) ⋊eν K. Here µ and ν are the partial actions of K on B/H and Be/H Tracking back the construction of bβ and bτ , to Theorem 2.11, we notice that given by Lemma 1.3, respectively. Meanwhile, eµ andeν are the one given by Theorem 2.4. Putting all together, by Theorem 2.6, it suffices to prove ν is the enveloping action of µ. Consider the map B → Be/H, given by b 7→ Hb. This is an open and continuous map, it is also constant in the α−orbits. So it defines a unique map F : B/H → Be/H, given by Hb 7→ Hb (this is not the identity map). It turns out this function is continuous, open, injective and maps fibers into fibers. In an analogous way we define f : X/H → X e/H, which has the same topological properties. Recalling the construction of µ and ν, it is easy to show (F, f ) : µ → ν is a morphism. To show that µe = ν it suffices to prove only two things. Namely, that f ((X/H)t) = f (X/H) ∩ tf (X/H) for every t ∈ K (we adopted the notation tz for the action of t ∈ K on z ∈ X e/H) and that the K−orbit of f (X/H) is X e/H. t = HX ∩ H(e, t)X = HX ∩ tHX = f (X/H) ∩ tf (X/H). For the first one notice that f ((X/H)t) = HX K The second equality of the previous formula is not immediate, but the inclusion ⊂ is. For the other one assume y ∈ HX ∩H(e, t)X. Then there exists x, z ∈ X such that y = Hx = H(e, t)z. There is some s ∈ H such that x = (s, e)(e, t)z = (s, t)z. t and y = K(s−1, e)x ∈ HX K So x ∈ X ∩ (s, t)X = s · (X H t . t ), (s−1, e)x ∈ X K s ∩ X K 14 DAMI ´AN FERRARO To show X e/H is the K−orbit of f (X/H), notice that [t∈K tf (X/H) = [t∈K tHX = H [s∈H [t∈K (s, t)X = HX e = X e/H. We have proved ν is the enveloping action of µ, by Theorem 2.6 C0(B/H) ⋊eµ K is strongly Morita equivalent to C0(Be/H) ⋊eν K. This completes the proof of our main theorem. (cid:3) The next Theorem is a consequence of the previous one, it has the advantage of not making any mention to σ nor τ . Theorem 3.3. If α and β have closed domain then Ind0(B, α) ⋊ bβ K ∼M Ind0(B, β) ⋊ bα H. Proof. We check the hypotheses of the previous theorem are satisfied. To show α × β has closed graph notice it has closed domain (Lemma 1.6) and use Lemma 1.5. Finally, we only have to show σ and τ are proper. It is enough to show σ is proper, for that purpose we use Lemma 1.9. Let {(si, xi)} be a net in H × X e such that {(xi, (si, e)xi)} converges to the point (x, y) ∈ X e × X e. It is enough to show {si}i∈I has a converging subnet. We may assume (s, t)x ∈ X and (h, k)y ∈ X, for some (h, k), (s, t) ∈ H × K. There is an i0 such that, for i > i0, (s, t)xi and (h, k)(si, e)xi belong to X. For i > i0 define ui = (s, t)xi. By the construction of α × β and because (hsis−1, kt−1)ui ∈ X, we have that ui is an element of the clopen set tk−1 ⋆(X K ssih−1 ). Defining vi := (e, k−1t)ui for every i > i0, we have that vi ∈ X. So, the limit limi vi is an element of X (recall X is clopen in X e). k−1t ∩X H The net {(hsis−1, vi)}i is contained in Γ(X, α) and {(vi, hsis−1 · vi)}i has a limit point. Then {hsis−1}i has a converging subnet, an so {si}i has a converging subnet. We conclude σ is proper, and we are done. (cid:3) References [1] F. Abadie, Sobre A¸coes Parciais, Fibrados de Fell, e Grup´oides, PhD thesis, Universidade de Sao Paulo, September 1999. [2] , Enveloping actions and Takai duality for partial actions, J. Funct. Anal., 197 (2003), pp. 14 -- 67. [3] F. Abadie and L. Mart´ı P´erez, On the amenability of partial and enveloping actions, Proc. Amer. Math. Soc., 137 (2009), pp. 3689 -- 3693. [4] A. an Huef, I. Raeburn, and D. P. Williams, An equivariant brauer semigroup and the symmetric imprimitivity theorem, Trans. Amer. Math. Soc., 352 (2010), pp. 4759 -- 4787. [5] R. Exel, Twisted partial actions: a classification of regular C ∗-algebraic bundles, Proc. London Math. Soc. (3), 74 (1997), pp. 417 -- 443. [6] R. Exel, M. Laca, and J. Quigg, Partial dynamical systems and C ∗-algebras generated by partial isometries, J. Operator Theory, 47 (2002), pp. 169 -- 186. [7] J. M. G. Fell and R. S. Doran, Representations of ∗-algebras, locally compact groups, and Banach ∗-algebraic bundles., vol. 125 -- 126 of Pure and Applied Mathematics, Academic Press Inc., Boston, MA, 1988. [8] S. Kaliszewski, P. S. Muhly, J. Quigg, and D. P. Williams, Fell bundles and imprimitivity theorems, eprint arXiv:1201.5035, (2012). [9] G. Kasparov, Equivariant kk-theory and the Novikov conjecture, Invent. Math., 91 (1988), p. 147201. [10] I. Raeburn, Induced C ∗-algebras and a symmetric imprimitivity theorem, Math. Ann., 280 (1988), pp. 369 -- 387. AN IMPRIMITIVITY THEOREM FOR PARTIAL ACTIONS 15 [11] M. A. Rieffel, Applications of strong Morita equivalence to transformation group C ∗- algebras, in Operator algebras and applications, Part I (Kingston, Ont., 1980), vol. 38 of Proc. Sympos. Pure Math., Amer. Math. Soc., Providence, R.I., 1982, pp. 299 -- 310. [12] , Morita equivalence for operator algebras, in Operator algebras and applications, Part I (Kingston, Ont., 1980), vol. 38 of Proc. Sympos. Pure Math., Amer. Math. Soc., Providence, R.I., 1982, pp. 285 -- 298. [13] D. P. Williams, Crossed products of C ∗-algebras, vol. 134 of Mathematical Surveys and Monographs, American Mathematical Society, Providence, RI, 2007. Departamento de Matem´atica y Estad´ıstica del Litoral, Universidad de la Rep´ublica, Gral. Rivera 1350. Salto. Uruguay. E-mail address: [email protected]
1105.4523	4	1105	2011-09-21T16:43:18	On Polish Groups of Finite Type	[ "math.OA", "math.GN" ]	Sorin Popa initiated the study of Polish groups which are embeddable into the unitary group of a separable finite von Neumann algebra. Such groups are called of finite type. We give necessary and sufficient conditions for Polish groups to be of finite type, and construct exmaples of such groups from semifinite von Neumann algebras. We also discuss permanence properties of finite type groups under various algebraic operations. Finally we close the paper with some questions concerning Polish groups of finite type.	math.OA	math	On Polish Groups of Finite Type Hiroshi Ando1 , 2 1 University of Copenhagen Universitetsparken 5, 2100 København Ø, Denmark 2 Research Institute for Mathematical Sciences, Kyoto University Kyoto, 606-8502, Japan E-mail: [email protected] Yasumichi Matsuzawa3 , 4 3 Mathematisches Institut, Universitat Leipzig Johannisgasse 26, 04103, Leipzig, Germany 4 Department of Mathematics, Hokkaido University Kita 10, Nishi 8, Kita-ku, Sapporo, 060-0810, Japan E-mail: [email protected] October 22, 2018 Abstract Sorin Popa initiated the study of Polish groups which are embeddable into the unitary group of a separable finite von Neumann algebra. Such groups are called of finite type or said to belong to the class Ufin. We give necessary and sufficient conditions for Polish groups to be of finite type, and construct exmaples of such groups from I∞ and II∞ von Neumann algebras. We also discuss permanence properties of finite type groups under various algebraic operations. Finally we close the paper with some questions concerning Polish groups of finite type. Keywords bi-invariant metric, class Ufin, finite type group, Polish group, pos- itive definite function, SIN-group, II1 factor Mathematics Subject Classification (2000) 46L10, 54H11, 43A35 1 Contents 1 Introduction 2 Polish Groups of Finite Type and its Characterization 2.1 Polish Groups of Finite Type . . . . . . . . . . . . . . . . . . . . 2.2 Positive Definite Functions . . . . . . . . . . . . . . . . . . . . . 2.3 The First Characterization . . . . . . . . . . . . . . . . . . . . . 2.4 SIN-groups and Bi-invariant Metrics . . . . . . . . . . . . . . . . 2.5 Unitary Representability . . . . . . . . . . . . . . . . . . . . . . . 2.6 Simple Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.7 Application to Locally Compact Groups . . . . . . . . . . . . . . 2.8 A Characterization for Amenable Groups . . . . . . . . . . . . . 3 More Examples of Finite Type Groups 3.1 L2-unitary groups U(M )2 . . . . . . . . . . . . . . . . . . . . . . 3.2 Non-isomorphic Properties of U(M )2 . . . . . . . . . . . . . . . . 3.3 Other Known Examples . . . . . . . . . . . . . . . . . . . . . . . 4 Hereditary Properties of Finite Type Groups 4.1 Closed Subgroup and Countable Direct Product . . . . . . . . . . 4.2 Extension and Semidirect Product . . . . . . . . . . . . . . . . . 4.3 Quotient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Projective Limit . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Some Questions 1 Introduction 2 3 3 4 5 7 7 8 9 10 12 12 14 16 16 17 17 17 18 18 In this paper we consider the following problem. Denote by U(M ) the unitary group of a von Neumann algebra M . Problem 1.1. Determine the necessary and sufficient condition for a Polish group G to be isomorphic as a topological group onto a strongly closed subgroup of some U(M ), where M is a separable finite von Neumann algebra. S. Popa defined a Polish group to be of finite type if it has this embedding property. Denote by Ufin the class of all finite type Polish groups. He initiated the study of this class in an attempt to enrich the study of rapidly developing cocycle superrigidity theory (cf. [8, 16, 18]). In particular, he proposed in [18] the problem of studying and characterizing the class Ufin. Secondly, this problem is motivated from our previous work [1] on infinite- dimensional Lie algebras associated with such groups: Let M be a finite von Neumann algebra on a Hilbert space H. Let G be a strongly closed subgroup of U(M ) and M be a set of all densely defined closed operators on H which are affiliated to M . It is proved that the set Lie(G) := {A∗ = −A ∈ M ; etA ∈ G for all t ∈ R} 2 is a complete topological Lie algebra with respect to the strong resolvent topol- ogy (see also the related work of D. Beltita [3]). Since these Lie algebras turn out to be non-locally convex in general when M is non-atomic, they are quite exotic as a Lie algebra and their properties are still unknown. Therefore it would be interesting to find non-trivial examples of such groups. We give an answer in Theorem 2.7 to the above Problem by the aid of positive definite functions on groups and their GNS representations, and characterize lo- cally compact groups or amenable Polish groups of finite type via compatible bi-invariant metrics in Proposition 2.20 and Theorem 2.22 (the former is known, but we give a new proof). Combining with Popa's result [18], Theorem 2.7 gives a necessary and sufficient condition for a Polish group to be isomorphic onto a closed subgroup of the unitary group of a separable II1 factor. We then give examples of Polish groups G of finite type using noncommutative integration of E. Nelson [17]. Finally we discuss some hereditary properties of finite type groups and pose some questions concerning Polish groups of finite type. Notation. In this paper we often say a von Neuman algebra M is separa- ble if it has a separable predual, especially when the Hilbert space on which M acts is implicit. This is known to be equivalent to the condition that M has a faithful representation on a separable Hilbert space. We denote by Proj(M ) the lattice of all projections in M . A von Neumann algebra is said to be finite if it admits no non-unitary isometry. When we consider a group G, its identity is denoted as eG. However, we also use 1 as the identity when we consider a concrete subgroup of the unitary group of a von Neumann algebra. We always regard the unitary group of a von Neumann algebra as a topological group with the strong operator topology. 2 Polish Groups of Finite Type and its Charac- terization In this section, we characterize Polish groups of finite type via positive definite functions. We then characterize when locally compact groups or amenable Pol- ish groups are of finite type via compatible bi-invariant metrics. To this end, we review notions of SIN-groups, bi-invariant metrics and unitary representability. 2.1 Polish Groups of Finite Type Recall that a Polish space is a separable completely metrizable topological space, and a Polish group is a topological group whose topology is Polish. We now introduce finite type groups after Popa [18]. Definition 2.1. A Hausdorff topological group is called of finite type if it is isomorphic as a topological group onto a closed subgroup of the unitary group of a finite von Neumann algebra. 3 Remark 2.2. Popa [18] requires the topological group of finite type to be Polish, whereas our definition of finiteness does not require any countability. We will show in Theorem 2.7 that a Polish group G of finite type in our sense coincides with Popa's definition of finite type group. That is, G is isomorphic onto a closed subgroup of the unitary group of a finite von Neumann algebra acting on a separable Hilbert space. All of second countable locally compact Hausdorff groups, the unitary group of a von Neumann algebra acting on a separable Hilbert space are Polish groups. Furthermore, separable Banach spaces are Polish groups as an additive group. We denote the class of all Polish groups of finite type by Ufin. Note that since a von Neumann algebra is finite if and only if its unitary group is complete with respect to the left uniform structure, Polish groups of finite type are necessarily complete. Thus we have the following simple conse- quence. Proposition 2.3. The unitary group of a von Neumann algebra M acting on a separable Hilbert space is of finite type if and only if M is finite. Another examples of Polish groups of finite type are given later. 2.2 Positive Definite Functions A complex valued function f on a Hausdorff topological group G is called positive definite if for all g1, · · ·, gn ∈ G and for all c1, · · ·, cn ∈ C, ¯cicjf (g−1 i gj) ≥ 0 n Xi,j=1 holds. Moreover if a complex valued function f is invariant under inner auto- morphisms, that is f (hgh−1) = f (g), ∀g, h ∈ G, then f is called a class function. It is well-known that there is an one-to-one correspondence between the set of all continuous positive definite functions on a topological group and the set of unitary equivalence classes of all cyclic unitary representations of it. more precisely, for each continuous positive definite function f on a topological group G, there exists a triple (πf , Hf , ξf ) consisting of a cyclic unitary representation πf in a Hilbert space Hf and a cyclic vector ξf in Hf such that f (g) = hξf , πf (g)ξf i, g ∈ G, and this triple is unique up to unitary equivalence. This triple is called the GNS triple associated to f . Note that if G is separable, then so is Hf . The GNS triple is of the following form for each continuous positive definite class function. 4 Lemma 2.4. Let f be a continuous positive definite class function on a topo- logical group G and (π, H, ξ) be its GNS triple. Then the von Neumann algebra M generated by π(G) is finite and the linear functional τ (x) := hξ, xξi, x ∈ M, is a faithful normal tracial state on M . In particular M is countably decompos- able. Proof . It is clear that τ is a normal state on M . Since f is a class function, it is easy to see that τ is tracial on the strongly dense -subalgebra of M spanned by π(G). Therefore by normality, τ is tracial on M . Therefore we have only to check the faithfulness of τ . Assume τ (x∗x) = 0. Since τ is a trace, we have kxπ(g)ξk2 = τ (π(g)∗x∗xπ(g)) = 0, for all g ∈ G. By the cyclicity of ξ, x must be 0. Example 2.5 (I. J. Schoenberg [19]). Let H be a complex Hilbert space. Note that H is an additive group. Then a function f defined by f (ξ) := e−kξk2 (ξ ∈ H) is a positive definite (class) function on H. Example 2.6 (I. J. Schoenberg [19]). For all 1 ≤ p ≤ 2 a function fp defined by fp(a) := e−kakp p (a ∈ lp) is a positive definite (class) function on a separable Banach space lp. For more details about positive definite class functions, see [12]. 2.3 The First Characterization We now characterize Polish groups of finite type. Theorem 2.7. For a Polish group G the following are equivalent. (i) G is of finite type. (ii) G is isomorphic as a topological group onto a closed subgroup of the unitary group of a finite von Neumann algebra acting on a separable Hilbert space. (iii) A family F of continuous positive definite class functions on G gen- erates a neighborhood basis of the identity eG of G. That is, for each neighborhood V of the identity, there are functions f1, · · ·, fn ∈ F and open sets O1, · · ·, On in C such that eG ∈ n \i=1 f −1 i (Oi) ⊂ V. (iv) There exists a positive, continuous positive definite class function which generates a neighborhood basis of the identity of G. 5 (v) A family F of continuous positive definite class functions on G sepa- rates the identity of G and closed subsets A with A 6∋ eG. That is, for each closed subset A with A 6∋ eG, there exists a continuous positive defi- nite class function f ∈ F such that f (x) < f (eG). sup x∈A (vi) There exists a positive continuous positive definite class function which separates the identity of G and closed subsets A with A 6∋ eG. Proof . (iv)⇔(vi)⇒(v)⇒(iii) and (ii)⇒(i) are trivial. (iii)⇒(ii). Since G is first countable, there exists a countable subfamily {fn}n of F which generates a neighborhood basis of the identity of G. Let (πn, ξn, Hn) be the GNS triple associated to fn and Mn be a von Neumann algebra generated by πn(G). Since each Mn is finite, the direct sum M :=Ln Mn is also finite and acts on a separable Hilbert space H := Ln Hn (see the remark above Lemma 2.4). Put π := Ln πn, then π is an embedding of G into U(M ). The image of π is closed in U(M ), as both G and U(M ) are Polish. (i)⇒(iii). Let π be an embedding of G into the unitary group of a finite von Neumann algebra M . Since each finite von Neumann algebra is the direct sum of countably decomposable finite von Neumann algebras, we can take of a family of countably decomposable finite von Neumann algebras {Mi}i∈I with M = Li∈I Mi. In this case π is also of the form π = Li∈I πi, where each πi : G → U(Mi) is a continuous group homomorphism. Let τi be a faithful normal tracial state on Mi and (ρi, ξi, Hi) be its GNS triple as a C∗-algebra. Here each ρi is an isomorphism from Mi into B(Hi) and τi(x) = hξi, ρi(x)ξii, x ∈ Mi, holds. Now set fi := τi ◦ πi. Then each fi is a continuous positive definite class functions on G and {fi}i∈I generates a neighborhood basis of the identity eG of G. (iii)⇒(iv). Let {fn}n be a countable family of continuous positive defi- nite class functions generating a neighborhood basis of the identity of G with fn(eG) = 1. Set f ′ n(g) := eRe(fn(g))−1 = e−1 1 k! ∞ Xk=0 [Re(fn(g))]k , g ∈ G, then {f ′ n}n is not only a family of continuous positive definite class functions generating a neighborhood basis of the identity of G with f ′ n(eG) = 1 but also a family of positive functions. Define a positive, continuous positive definite class n(g)/2n (g ∈ G). It is easy to see that f generates a function by f (g) := Pn f ′ neighborhood basis of the identity of G. 6 Remark 2.8. The proof of the above theorem is inspired by Theorem 2.1 of S. Gao [10]. Remark 2.9. Popa (Lemma 2.6 of [18]) showed that a Polish group G is of finite type if and only if it is isomorphic onto a closed subgroup of the unitary group of a separable II1 factor. Therefore Theorem 2.7 gives a necessary and sufficient condition for a Polish group to be isomorphic onto a closed subgroup of the unitary group of a separable II1 factor. 2.4 SIN-groups and Bi-invariant Metrics To discuss further properties of finite type groups, we consider the following notions, say SIN-groups, bi-invariant metrics and unitarily representability. A neighborhood V at the identity of a topological group G is called invariant if it is invariant under all inner automorphisms, that is, gV g−1 = V holds for all g ∈ G. A SIN-group is a topological group which has a neighborhood basis of the identity consisting of invariant identity neighborhoods. Note that a locally compact Hausdorff SIN-group is unimodular. A bi-invariant metric on a group G is a metric d which satisfies d(kg, kh) = d(gk, hk) = d(g, h), ∀g, h, k ∈ G. It is known that a first countable Hausdorff topological group is SIN if and only if it admits a compatible bi-invariant metric. As Popa [18] pointed out, one of the most important fact of Polish groups of finite type is an existence of a compatible bi-invariant metric. Lemma 2.10. Each Polish group of finite type has a compatible bi-invariant metric. In particular, it is SIN. Proof . It is enough to show that for every finite von Neumann algebra M acting on a separable Hilbert space H the unitary group U(M ) has a compatible bi- invariant metric. For this let τ be a faithful normal tracial state on M . Then a metric d defined by d(u, v) := τ ((u − v)∗(u − v)) 1 2 , u, v ∈ U(M ), is a compatible bi-invariant metric on U(M ). 2.5 Unitary Representability A Hausdorff topological group is called unitarily representable if it is isomorphic as a topological group onto a subgroup of the unitary group of a Hilbert space. All locally compact Hausdorff groups are unitarily representable via the left reg- ular representation. It is clear that a Polish group of finite type is necessarily unitarily representable. The following characterization of unitary representabil- ity has been considered by specialists and can be seen in e.g., Gao [10]. Lemma 2.11. For a Polish group G the following are equivalent. 7 (i) G is unitarily representable. (ii) There exists a positive, continuous positive definite function which sep- arates the identity of G and closed subsets A with A 6∋ eG. 2.6 Simple Examples All of the following examples are well-known. The first three examples are locally compact groups. Example 2.12. Any compact metrizable group is a Polish group of fintie type. This follows from the Peter-Weyl theorem. Example 2.13. Any abelian second countable locally compact Hausdorff group is a Polish group of finite type. Indeed its left regular representation is an embedding into the unitary group of a Hilbert space and the von Neumann algebra generated by its image is commutative (in particular, finite). Example 2.14. Any countable discrete group is a Polish group of finite type. For its left regular representation is an embedding into the unitary group of a finite von Neumann algebra. The following two examples suggest there are few other examples of locally compact groups of finite type. 1 0 0 0 0 (cid:19) , (cid:18) a b 1 (cid:19)(cid:18) x y ax + b group, where K = R or C. By easy computations, we have =(cid:18) x −bx + ay + b Example 2.15. Let G := (cid:26)(cid:18) x y 1 (cid:19)(cid:18) a b so that the conjugacy class C(cid:18)(cid:18) x y ( x ♯ ( 1 ( 1 1 (cid:19) ∈ GL(2, K) ; x ∈ K×, y ∈ K(cid:27) be the 1 (cid:19)−1 1 (cid:19)(cid:19) of (cid:18) x y 1 ! ; ♯ ∈ K) (x 6= 1), 1 ! ; ♯ ∈ K×) (x = 1, y 6= 0), 1 !) C(cid:18)(cid:18) x y 1 (cid:19)(cid:19) = 1 (cid:19) is (x = 1, y = 0). 0 0 ♯ 0 0 0 0 0 Thus for each n ∈ N there exists a matrix hn ∈ G such that hngnh−1 n = (cid:18) 1 0 1 1 (cid:19), where gn :=(cid:18) 1 0 1/n 1 (cid:19). Clearly, gn → 1 and hngnh−1 implies that the ax + b group does not admit a compatible bi-invariant metric. Hence it is not of finite type. n 6→ 1. This   0 8 Example 2.16. The special linear group SL(n, K) (n ≥ 2) is not of finite type since the map (cid:18) a b 0 1 (cid:19) 7→ (cid:18) a 0 a−1 (cid:19) is an embedding of the ax + b group b into SL(2, K). Thus the general linear group GL(n, K) (n ≥ 2) is also not of finite type. Next we consider abelian groups. Note that an abelian topological group is of finite type if and only if it is unitarily representable. Example 2.17. Any separable Hilbert space is a Polish group of finite type. This follows from Example 2.5 and Theorem 2.7. Example 2.18. A separable Banach space lp (1 ≤ p ≤ ∞) is a Polish group of finite type if and only if 1 ≤ p ≤ 2. The "only if" part follows from Example 2.6 and Theorem 2.7, but the "if" part is non-trivial. For details, see [15]. Here is another counter example. Example 2.19. Separable Banach space C[0, 1] of all continuous functions on the interval [0, 1] is a Polish group but not of finite type. For, since every sep- arable Banach space is isometrically isomorphic to a closed subspace of C[0, 1], if C[0, 1] is of finite type, then any separable Banach space is a Polish group of finite type. But this is a contradiction to the previous example. 2.7 Application to Locally Compact Groups It is known that a second countable locally compact group is of finite type if and only if it is a SIN-group (see e.g., Theorem 13.10.5 of J. Dixmier [4]). We give a new proof of this fact using Theorem 2.7. We thank the referee for letting us know the above literature. Proposition 2.20. A second countable locally compact Hausdorff group is of finite type if and only if it is SIN. Proof . Let G be a second countable locally compact Hausdorff SIN-group, µ be the Haar measure on it and λ be its left-regular representation. For each com- pact invariant neighborhood U of the identity, we define a continuous positive definite function ϕU on G by ϕU (g) := hχU , λ(g)χU i = µ (U ∩ gU ) , g ∈ G. Note that, for each g, h, x ∈ G, we have h−1x ∈ U ⇔ x ∈ hU = U h ⇔ xh−1 ∈ U, and (gh)−1x ∈ U ⇔ x ∈ ghU = gU h ⇔ xh−1 ∈ gU. 9 Also note that a locally compact SIN-group is unimodular. Thus we see that ϕU (h−1gh) = hλ(h)χU , λ(gh)χU i χU (h−1x)χU ((gh)−1x)dµ(x) χU (xh−1)χgU (xh−1)dµ(x) χU (x)χgU (x)dµ(x) χU (x)χU (g−1x)dµ(x) =ZG =ZG =ZG =ZG = ϕU (g). This implies ϕU is a class function. It is not hard to check that a family {ϕU }U generates a neighborhood basis of the identity of G. This completes the proof by Theorem 2.7. Remark 2.21. (1) R. V. Kadison and I. Singer [14] proved that every connected locally compact Hausdorff SIN group is isomorphic as a topological group onto a topological group of the form Rn × K, where K is a compact Hausdorff group. (2) K. Hofmann, S. Morris and M. Stroppel [13] proved that every totally dis- connected locally compact Hausdorff group is SIN if and only if it is a strict projective limit of discrete groups. 2.8 A Characterization for Amenable Groups Next, we characterize (not necessarily locally compact) amenable Polish groups of finite type. Recall that a Hausdorff topological group G is amenable if LUCB(G) admits a left-translation invariant positive functional m ∈ LUCB(G)∗ with m(1) = 1, where LUCB(G) is a complex Banach space of all left-uniformly continuous bounded functions on G. Such a m is called an invariant mean. Theorem 2.22. A unitarily representable amenable Polish group is of finite type if and only if it is SIN. Proof . Let G be a unitarily representable amenable Polish SIN-group and let f be a positive, continuous positive definite function on G which separates the identity of G and closed subsets A with A 6∋ eG (see Lemma 2.11 ). We may and do assume f (eG) = 1. For each x ∈ G, we define a positive function Ψx,f : G → [0, 1] by Ψx,f (g) := f (g−1xg), g ∈ G. We show that Ψx,f ∈ LUCB(G). Fix an arbitrary ε > 0. Since the positive definite function f is left-uniformly continuous, there exists a neighborhood V of eG such that f (g) − f (h) < ε 10 holds whenver g, h ∈ G satisfy g−1h ∈ V . There exists a neighborhood W of eG such that W = W −1 and W · W ⊂ V holds. Since G is SIN, there exists an invariant neighborhood U of eG with U ⊂ W . Let g, h ∈ G satisfy g−1h ∈ U . By the invariance of U , it holds that h ∈ gU = U g and therefore that hg−1 ∈ U . Then we see that (h−1xh)−1(g−1xg)−1 = h−1x−1hg−1xg ∈ h−1x−1U xg = h−1U g = U h−1g = U (g−1h)−1 ⊂ W · W −1 ⊂ V, which implies Ψx,f (h) − Ψx,f (g) = f (h−1xh) − f (g−1xg) < ε. Hence Ψx,f is left-uniformly continuous and we have Ψx,f ∈ LUCB(G)+. Let m ∈ LUCB(G)∗ be an invariant mean. Put ψf (x) := m(Ψx,f ), x ∈ G, then ψf (x) is clearly a positive, positive definite class function on G with ψf (eG) = 1. We show that ψf is continuous. Since m is continuous, it suf- fices to show that G ∋ x 7→ Ψx,f ∈ LUCB(G)+ is continuous. Let x, y ∈ G. By Krein's inequality, we have Ψx,f − Ψy,f 2 = sup g∈G f (g−1xg) − f (g−1yg)2 ≤ 2 sup g∈G = 2 sup g∈G 1 − Ref (g−1x−1yg) 1 − f (g−1x−1yg). Fix ε > 0. Since f is left-uniformly continuous, there exists an invariant neigh- borhood V of eG such that f (x) − f (y) < ε holds for x, y ∈ G with x−1y ∈ G. Then for x, y ∈ G with x−1y ∈ V , we have g−1x−1yg ∈ g−1V g = V . Therefore it holds that 1 − f (g−1x−1yg) = f (eG) − f (g−1x−1yg) < ε. Hence we have Ψx,f − Ψy,f 2 ≤ 2ε. Therefore G ∋ x 7→ Ψx,f ∈ LUCB(G)+ is continuous, hence so is ψf . We next show that ψf separates the identity of G and closed subsets A with A 6∋ eG. Fix such a closed set A. Since Ac = G \ A is an open neighborhood of eG, there exists an open invariant neighborhood V of eG contained in Ac. Then we have A ⊂ V c and eG /∈ V . Since f separates eG and V c, we have δ := sup g∈V c f (g) < 1. 11 It then follows, by the invariance of V c, that for x ∈ V c, Ψx,f = sup g∈G f (g−1xg) ≤ sup g∈V c f (g) ≤ δ, which implies sup x∈A ψf (x) ≤ sup x∈V c ψf (x) = sup x∈V c m(Ψx,f ) ≤ sup x∈V c Ψx,f ≤ δ < 1. Therefore ψf separates A and eG. This completes the proof by Theorem 2.7. Remark 2.23. The above proof is inspired by the proof of Theorem 2.13 of J. Galindo [9]. 3 More Examples of Finite Type Groups In this section we will give nother examples of Polish groups of finite type. To construct such examples we need to start not from finite von Neumann algebras, but from semifinite von Neumann algebras, say of type I∞ or of type II∞. In the end of this section we also review other known examples of Polish groups of finite type. 3.1 L2-unitary groups U(M)2 Let M be a semifinite von Neumann algebra on a Hilbert space H equipped with a normal faithful semifinite trace τ . A densely defined, closed operator T on H is said to be affiliated to M if for all u ∈ U(M ′), uT u∗ = T holds. Denote by M the set of all densely defined, closed operators on H which are affiliated to M . Recall that L2(M, τ ) is a Hilbert space completion of the space nτ := {x ∈ M ; τ (x∗x) < ∞} by the inner product hx, yi := τ (x∗y), x, y ∈ nτ . We define x2 := τ (x∗x) 1 2 for x ∈ L2(M, τ ). Definition 3.1. We call U(M )2 := {u ∈ U(M ); 1 − u ∈ L2(M, τ )} the L2- unitary group of (M, τ ). Note that when M is not a factor, U(M )2 depends on the choice of τ too. In the sequel we show the following theorem. Theorem 3.2. Let M be a separable semifinite von Neumann algebra with a normal faithful semifinite trace τ . Then U(M )2 is a Polish group of finite type, where the topology is determined by the following metric d, d(u, v) := u − v2, u, v ∈ U(M )2. 12 To prove the theorem, we need some preparations. In the sequel we consider M to be represented on H = L2(M, τ ) by left multiplication. Recall that a closed operator T ∈ M on L2(M, τ ) is called τ -measurable if for any ε > 0, there exists a projection p ∈ M with ran(p) ⊂ dom(T ) and τ (1 − p) < ε. Note that L2(M, τ ) can be identified with the set of closed, densely defined and τ -measurable operators T such that T 2 2 := τ (T 2) =Z ∞ 0 λ2dτ (e(λ)) < ∞, where e(·) is a spectral resolution of T = (T ∗T ) 2 and T = uT is the polar decomposition of T (for more details about non-commutative integration, see vol II of [20]). 1 Lemma 3.3. Let M be a semifinite von Neumann algebra with a normal faithful semifinite trace τ . Then U(M )2 is a topological group. Proof . This can be shown directly, using the equalities: x∗2 = x2, uxv2 = x2, for all x ∈ L2(M, τ ) and u, v ∈ U(M ). Lemma 3.4. Let M be a semifinite von Neumann algebra with a normal faithful semifinite trace τ . Let U be a densely defined closed τ -measurable operator on L2(M, τ ) affiliated to M . Then dom(U ) ∩ M is dense in L2(M, τ ). Proof . Let ε > 0. Let ξ ∈ L2(M, τ ). Since M ∩ L2(M, τ ) is dense, there exists ξ0 ∈ M ∩ L2(M, τ ) such that ξ − ξ02 < ε. On the other hand, the measurability of U implies the existence of an increasing sequence {pn}∞ n=1 of projections in M such that pnL2(M, τ ) ⊂ dom(U ) for all n and pn ր 1 strongly. Therefore there exists n0 ∈ N such that ξ0 − pn0ξ02 < ε. By the choice of ξ0, pn0ξ0 ∈ dom(U ) ∩ M and ξ − pn0 ξ02 ≤ ξ − ξ02 + ξ0 − pn0ξ02 ≤ ε + ε = 2ε. Since ε is arbitrary, it follows that dom(U ) ∩ M is dense in L2(M, τ ). Lemma 3.5. Let M be a semifinite von Neumann algebra with a normal faithful semifinite trace τ . d is a complete metric on U(M )2. Proof . Suppose {un}∞ n=1 is a d-Cauchy sequence in U(M )2. Since L2(M, τ ) is complete, there exists V ∈ L2(M, τ ) such that (1 − un) − V 2 → 0. Define U := 1 − V . Then U − un2 → 0. We show that U is bounded and moreover U ∈ U(M )2. Since U is closed and dom(U )∩M is dense by Lemma 3.4, to prove 13 the boundedness of U it suffices to show that U is isometric on dom(U ) ∩ M . Let ξ ∈ dom(U ) ∩ M . Since ξ is bounded, we have (U − un)ξ2 2 = τ (ξ∗(U − un)∗(U − un)ξ) = τ ((U − un)ξξ∗(U − un)∗) ≤ ξ2τ ((U − un)(U − un)∗) = ξ2U − un2 2 → 0, which implies U ξ2 = lim n→∞ unξ2 = ξ2, for all ξ ∈ dom(U ) ∩ M . Therefore U dom(U)∩M is isometric and U is bounded. Since U ∗ − u∗ n2 = U − un2, it holds that U ∗ is an isometry too, which means U is a unitary. Finally, it is clear that U = 1 − V ∈ U(M )2. Proof of Theorem 3.2. Since M is separable, the separability of U(M )2 follows from the separability of L2(M, τ ). Therefore by Lemma 3.5, U(M )2 is a Polish group. By Schoenberg's theorem (see Example 2.5), ϕ(u) := e−1−u2 2, u ∈ U(M )2, is a continuous, positive definite class function on U(M )2. It is easy to see that ϕ generates a neighborhood basis of the identity of U(M )2. Therefore the claim follows from Theorem 2.7. Remark 3.6. U(M )′′ 2 = M. Proof . Clearly U(M )′′ L2(M, τ ) and 1 − 2p ∈ U(M )2. Therefore p ∈ U(M )′′ is generated by finite projections. Therefore U(M )′′ 2 ⊂ M . Let p be a finite projection in M . Then 2p ∈ 2 . Since M is semifinite, M 2 = M . When M = B(H), U(M )2 is the well-known example of a Hilbert-Lie group and is denoted as U(H)2. 3.2 Non-isomorphic Properties of U(M)2 J. Feldman [7] gave a complete description of a group isomorphism between the unitary groups of type II1 von Neumann algebras. In particular, in the proof of Theorem 4 of [7], he uses the following simple observation: let p be a projection in a von Neumann algebra M , then up := 1 − 2p is a self-adjoint unitary in M . Using this correspondence, he deduced that the group isomorphism π : U(M1) → U(M2) between type II1 von Neumann algebras M1, M2 induces an order isomorphism between their projection lattices, thereby proving that the isomorphism π is lifted to a ring -isomorphism π : M1 → M2 (which may not preserve the scalar multiplication) in such a way that π(u) = θ(u)π(u), for all u ∈ U(M1) 14 holds, where θ is a multiplicative map from U(M1) to Z(U(M2)). Let H be an infinite dimensional Hilbert space. Using his idea, we show that when M is a II∞ factor and N is a finite von Neumann algebra, then U(M )2, U(H)2 and U(N ) are mutually non-isomorphic. In this subsection, no separability assumptions are required. Proposition 3.7. Let M be a II∞ factor. Then U(M )2 is not isomorphic onto U(H)2. Proof . Let τ be a normal faithful semifinite trace on M , Tr be the usual operator trace on H. We denote their corresponding trace 2-norms by · 2,τ and · 2,Tr, respectively. We prove the claim by contradiction. Suppose there exists a topological group isomorphism ϕ : U(M )2 → U(H)2. Let p be a nonzero finite-rank projection in B(H). Then 1 − 2p ∈ U(H)2 and let q := 1 2 (1 − ϕ−1(1 − 2p)). It is easy to see that q ∈ L2(M, τ ) is a nonzero finite projection in M . Let k ∈ N. Since M is a II∞ factor, there exists a projection 0 < qk ≤ q in M such that limk→∞ τ (qk) = 0. Define pk := 1−ϕ(1−2qk) . Since 2 qk2 2,τ = τ (qk) → 0 (k → ∞), 1 − 2qk → 1 holds in U(M )2, which in turn means 1 − 2pk = ϕ(1 − 2qk) → ϕ(1) = 1 in U(H)2. However, since the topology of U(H)2 is given by the operator trace 2-norm, it holds that 2 ≤ 2pk2,Tr = 1 − (1 − 2pk)2,Tr → 0 (k → ∞). This is clearly a contradiction. Therefore U(M )2 6∼= U(H)2. Proposition 3.8. Let M be a type I∞ or type II∞ factor, N be a finite von Neumann algebra. Then U(M )2 is not isomorphic onto U(N ). Proof . Let τ be a normal faithful semifinite trace on M . Let u ∈ Z(U(M )2) be an element of U(M )2 which commutes with every element in U(M )2. Then for any finite projection p ∈ M , u(1 − 2p) = (1 − 2p)u holds. Therefore u commutes with all finite projections in M . Since M is generated by its finite projections, u ∈ Z(M ) = C1 holds. Since u − 1 ∈ L2(M, τ ), this forces u = 1. Therefore the center of U(M )2 is {1}, while the center of U(N ) contains C1. Remark 3.9. We thank the referee for telling us the above simple proof and the literature [7]. 15 3.3 Other Known Examples The class Ufin has not been studied well. However, there are some known examples other than the ones presented in §2.6. Example 3.10. Normalizer groups NM (A) and N (E) Let A be an abelian von Neumann subalgebra of a separable II1 factor M . The normalizer group NM (A) of A, defined by NM (A) := {u ∈ U(M ); uAu∗ = A}, is clearly a strongly closed subgroup of U(M ) and hence belongs to Ufin. This group has been drawn much attention to specialists, especially when A is max- imal abelian and NM (A) generates M as a von Neumann algebra. In such a case, A is called a Cartan subalgebra. Similarly, the normalizer group N (E) for a normal faithful conditional expectation E : M → N onto a von Neumann subalgebra N , N (E) := {u ∈ U(M ); uE(x)u∗ = E(uxu∗), for all x ∈ M } is also of finite type. Example 3.11. The full group [R] Let R be a II1 countable equivalence relation on a standard probability space (X, µ). A. Furman showed that the full group [R] equipped with so-called uniform topology is a Polish group of finite type (see §2 of Furman [8]). 4 Hereditary Properties of Finite Type Groups In this section, we discuss several permanence properties of the class Ufin un- der several algebraic operations. In summary, we will observe the following permanence properties of finite type groups. Operation Closed subgroup H < G Countable direct product Qn≥1 Gn Semidirect product G ⋊ H Quotient G/N Extension 1 → N → G → K → 1 Projective limit lim ←− Gn Ufin? YES YES NO NO NO YES As can be seen from the above table, finiteness property is delicate and can easily be broken under natural operations. Remark 4.1. (On the ultraproduct of metric groups) Let {(Gn, dn)}∞ n=1 be a sequence of finite type Polish groups with a compatible bi-invariant metric. It is not difficult to show that the ultraproduct (Gω, dω) of {(Gn, dn)}∞ n=1 along a free ultrafilter ω ∈ βN \ N is a completely metrizable topological group of finite 16 type, but not Polish in general. We will discuss topological groups which are embeddable into the unitary group of a (not necessarily separable) finite von Neumann algebra elsewhere. 4.1 Closed Subgroup and Countable Direct Product It is clear the class Ufin is closed under taking a closed (or even Gδ) subgroup. Since a countable direct sum of separable finite von Neumann algebras is again separable and finite, the class Ufin is closed under countable direct product. 4.2 Extension and Semidirect Product The class Ufin is not closed under extension nor semidirect product. Proposition 4.2. There exits a Polish group G not of finite type, which has a closed normal subgroup N such that N and the quotient group G/N are of finite type. Proof . Let G be the ax + b group (see Example 2.15). Since G does not have a compatible bi-invariant metric, it is not of finite type. On the other hand, G can be written as a semidirect product G = K ⋊ K×, where K× acts on K as a multiplication. There fore the exact sequence 0 −→ K −→ G −→ K× −→ 1 gives a counter example for extension case. Note that the above example also shows that the class Ufin is not closed under semidirect product. 4.3 Quotient The class Ufin is not closed under quotient. Proposition 4.3. There exists an abelian Polish groups of finite type G such that the quotient G/N of G by its closed subgroup is not of finite type. Proof . Consider the separable Banach space A := l3 as an additive Polish group. As we saw in Example 2.18, lp(1 ≤ p ≤ ∞) is unitarily representable if and only if 1 ≤ p ≤ 2. On the other hand, every separable Banach space is isomorphic onto a quotient Banach space of ℓ1 (see e.g., Theorem 5.1 of [6]). In particular, although not of finite type, A = ℓ3 is a quotient of G := ℓ1 by its closed subgroup N . Remark 4.4. Note that even for abelian Polish groups, the situation can be worst possible. It is known (chapter 4 of [2]) that there exists an abelian Polish group A which has no non-trivial unitary representation. Such a group is called strongly exotic. On the other hand, S. Gao and V. Pestov [11] proved that any abelian Polish group is a quotient of ℓ1 by a closed subgroup N . Therefore, strongly exotic groups are also quotients of finite type Polish groups. 17 4.4 Projective Limit The class Ufin is closed under projective limit. Proposition 4.5. Let {Gn, jm,n : Gm → Gn(n ≤ m)}∞ system of Polish groups of finite type. Then G = lim ←− finite type. n,m=1 be a projective Gn is a Polish group of Proof . Since the connecting map {jm,n} is continuous, it is clear that G can be seen as a closed subgroup of Qn∈N Gn. Since finiteness property passes to direct product, Qn∈N Gn is also a Polish group of finite type. Therefore its closed subgroup G is also a Polish of finite type. 5 Some Questions Finally let us discuss some questions to which we do not have answers at this stage. Let Uinv denote the class of Polish groups with a compatible bi-invariant metric. As we saw in Example 2.6, Uinv is strictly larger than Ufin (l3 is in Uinv but not in Ufin). Therefore the unitarily representability is indispensable (this was also pointed out by Popa). Furthermore, there exists a more interesting example. Recently L. van den Dries and S. Gao [5] constructed a Polish group G with a compatible bi-invariant metric, which does not have Lie sum (see [5] for the definition). On the other hand, we proved in [1] that if G belongs to the class Ufin, then G has Lie sum. Thus G is not of finite type. Therefore it would be desirable to consider the following questions (the latter was posed in Popa[18], §6.5): Question 5.1. Is van den Dries-Gao's Polish group unitarily representable? Question 5.2 (Popa). Is a unitarily representable Polish SIN-group of finite type? Hopefully Theorem 2.7 will play the role for solving the above questions. Also, since lp belongs to Ufin if and only if 1 ≤ p ≤ 2, it is worth considering whether Question 5.3. Let H be a separable infinite-dimensional Hilbert space. Does U(H)p := {u ∈ U(H); 1 − u ∈ Sp(H)} belong to Ufin for some 1 ≤ p < 2 ? Here Sp(H) denotes the space of Schatten p-class operators. Finally, let us remind that there is another candidate for a counterexample to Question 5.2. Recall that a finite von Neumann algebra N equipped with a normal faithful tracial state τ is said to have property (T) if for each ε > 0, there exists a finite set F ⊂ N and δ > 0 with the property that whenever ϕ : N → N is a unital completely positive τ -preserving map satisfying ϕ(x) − x2 < δ for all x ∈ F , then ϕ(a) − a2 ≤ εa holds for all a ∈ N . Let M be a separable II1 factor with property (T), Aut(M ) be a Polish group of all *-automorphisms of M equipped with the pointwise · 2-convergence 18 topology. Due to the property (T), this topology coincides with the topology of uniform · 2-convergence on the closed unit ball M1. Since the latter topology is given by the bi-invariant metric d defined by d(α, β) := sup x∈M1 α(x) − β(x)2, α, β ∈ Aut(M ), Aut(M ) is a Polish SIN-group. By considering the standard representation, Aut(M ) is unitarily representable as well. Therefore it would be interesting to check if Aut(M ) is actually of finite type or not. Acknowledgements The authors would like to express their thanks to Professor Asao Arai, Professor Uffe Haagerup, Professor Izumi Ojima and Professor Konrad Schmudgen for their important comments, discussions and supports. We also thank to Mr. Takahiro Hasebe, Mr. Ryo Harada, Mr. Kazuya Okamura and Dr. Hayato Saigo for the discussions in a seminar and for their continual interests in our work. The second named author thanks to Mr. Abel Stolz for his kind discussions. Final version of the paper was done during authors' visit to the conference "Von Neumann algebras and ergodic theory of group actions 2011" at Institut Henri Poincar´e. They thank the organizers Professor Damien Gaboriau, Professor Sorin Popa and Professor Stefaan Vaes for their fruitful discussions and also to Professor Jesse Peterson for informing us about the Popa and his recent results concerning cocycle superrigidity and the class Ufin. Last but not least, we thank the anonymous referee for telling us the literature and suggesting a simpler proof of some of the main results. Both of the authors are supported by Research fellowships of the Japan Society for the Promotion of Science for Young Scientists. References [1] H. Ando, Y. Matsuzawa, Lie group-Lie algebra correspondences of unitary groups in finite von Neumann algebras, to appear in Hokkaido Math. J. [2] W. Banaszczyk, Additive subgroups of topological vector spaces, Lecture Notes in Mathematics, 1466. Springer-Verlag, Berlin, 1991. [3] D. Beltita, Lie theoretic significance of the measure topologies associated with a finite trace, Forum Math. 22 (2010), 241 -- 253. [4] J. Dixmier, C∗-algebras, North-Holland Mathematical Library vol. 15, 1977. [5] L. van den Dries, S. Gao, A Polish group without Lie sums, Abh. Math. Sem. Hambourg 79 (2009) 135 -- 147. [6] M. Fabian, P. Habala, P. H´ajek, V. Montesinos, V. Zizler, Banach Space Theory: The Basis for Linear and Nonlinear Analysis, CMS Books in 19 Mathematics/Ouvrages de Math´ematiques de la SMC. Springer, New York, 2011. [7] J. Feldman, Isomorphisms of finite type II rings of operators, Ann. Math. (2), 63, (1956), 565 -- 571. [8] A. Furman, On Popa's Cocycle Superrigidity Theorem, Int. Math. Res. Not. IMRN (2007), 1 -- 46. [9] J. Galindo, On group and semigroup compactifications of topological groups, Online Note http://www.mat.ucm.es/imi/documents/topologicalGroups/JGalindo.pdf [10] S. Gao, Unitary group actions and Hilbertian Polish metric spaces. in Logic and its applications, Contemp. Math. 380, AMS Providence, RI (2005), 53 -- 72. [11] S. Gao, V. Pestov, On a universality property of some abelian Polish groups, Fund. Math. 179 (2003), 1 -- 15. [12] T. Hirai, E, Hirai, Positive definite class functions on a topological group and characters of factor representations, J. Math. Kyoto. Univ. 45 (2005), 355 -- 379. [13] K. Hofmann, S. Morris and M. Stroppel, Varieties of topological groups, Lie groups and SIN-groups, Colloq. Math. 70 (1996), 151 -- 163. [14] R. V. Kadison, I. M. Singer, Some remarks on representations of connected groups, Proc. Nat. Acad. Sci. 38 (1952), 419 -- 423. [15] M. Megrelishvili, Reflexively but not unitarily representable topological groups, Topol. Proc. 25 (2000), 615 -- 625. [16] J. Peterson, T. Sinclair, On cocycle superrigidity for Gaussian actions, arXiv:0910.3958, (2009). [17] E. Nelson, Notes on non-commutative integration, J. Funct. Anal. 15 (1974), 103 -- 116. [18] S. Popa, Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups, Invent. Math. 170 (2007), 243 -- 295. [19] I. J. Schoenberg, Metric spaces and positive definite functions, Trans. Amer. Math. Soc. 44 (1938), 522 -- 536. [20] M. Takesaki, Theory of Operator Algebras. I (2002), II (2003) and III (2003). Springer-Verlag, Berlin. 20
1112.1361	1	1112	2011-12-06T18:03:46	Filtrated K-theory for real rank zero C*-algebras	[ "math.OA" ]	Using Kirchberg KK_X-classification of purely infinite, separable, stable, nuclear C-algebras with finite primitive ideal space, Bentmann showed that filtrated K-theory classifies purely infinite, separable, stable, nuclear C-algebras that satisfy that all simple subquotients are in the bootstrap class and that the primitive ideal space is finite and of a certain type, referred to as accordion spaces. This result generalizes the results of Meyer-Nest involving finite linearly ordered spaces. Examples have been provided, for any finite non-accordion space, that isomorphic filtrated K-theory does not imply KK_X-equivalence for this class of C-algebras. As a consequence, for any non-accordion space, filtrated K-theory is not a complete invariant for purely infinite, separable, stable, nuclear C-algebrass that satisfy that all simple subquotients are in the bootstrap class. In this paper, we investigate the case for real rank zero C-algebras and four-point primitive ideal spaces, as this is the smallest size of non-accordion spaces. Up to homeomorphism, there are ten different connected T_0-spaces with exactly four points. We show that filtrated K-theory classifies purely infinite, real rank zero, separable, stable, nuclear C-algebras that satisfy that all simple subquotients are in the bootstrap class for eight out of ten of these spaces.	math.OA	math	FILTRATED K-THEORY FOR REAL RANK ZERO C∗-ALGEBRAS SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ Abstract. Using Kirchberg KKX -classification of purely infinite, separable, stable, nuclear C ∗-algebras with finite primitive ideal space, Bentmann showed that filtrated K-theory classifies purely infinite, separable, stable, nuclear C ∗- algebras that satisfy that all simple subquotients are in the bootstrap class and that the primitive ideal space is finite and of a certain type, referred to as accordion spaces. This result generalizes the results of Meyer-Nest in- volving finite linearly ordered spaces. Examples have been provided, for any finite non-accordion space, that isomorphic filtrated K-theory does not imply KKX -equivalence for this class of C ∗-algebras. As a consequence, for any non-accordion space, filtrated K-theory is not a complete invariant for purely infinite, separable, stable, nuclear C ∗-algebras that satisfy that all simple sub- quotients are in the bootstrap class. In this paper, we investigate the case for real rank zero C ∗-algebras and four-point primitive ideal spaces, as this is the smallest size of non-accordion spaces. Up to homeomorphism, there are ten different connected T0-spaces with exactly four points. We show that filtrated K-theory classifies purely infinite, real rank zero, separable, stable, nuclear C ∗-algebras that satisfy that all simple subquotients are in the bootstrap class for eight out of ten of these spaces. 1. Introduction The C∗-algebra classification programme initiated by G. A. Elliott in the early seventies has seen a rapid development during the past 20 years. The notion of real rank zero C∗-algebras introduced by G. K. Pedersen and L. G. Brown in the late eighties has turn out to be of particular interest in connection with classification of C∗-algebras. Until the mid-nineties most results were concerned with the sta- bly finite algebras, when people such as M. Rørdam, N. C. Phillips, E. Kirchberg and D. Huang classified some purely infinite, nuclear, separable C∗-algebras in the bootstrap class. All these had finitely many ideals -- in fact, almost all cases were either the simple case or the one non-trivial ideal case. D. Huang was also able to classify purely infinite Cuntz-Krieger algebras with finite K-theory (implying that all the K1-groups are zero). In contrast to the stably finite case, the positive cone of purely infinite C∗-algebras carries no extra information, so it was clear from the beginning, that to classify non-simple purely infinite C∗-algebras one needs to come up with a new invariant, which also encodes the ideal structure and the K-theory of all ideals and quotients. The main ingredients of the proof of N. C. Phillips and E. Kirchberg were the UCT of J. Rosenberg and C. Schochet and a result saying that every KK-equivalence between (simple, purely infinite, stable, nuclear, separable) C∗-algebras can be lifted to a ∗-isomorphism between the algebras. Shortly after, E. Kirchberg generalized 1 2 SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ this result to X-equivariant KK-theory, where X is (homeomorphic to) the prim- itive ideal space of the C∗-algebra. The only ingredient thus missing to classify purely infinite, nuclear, separable, stable C∗-algebras seemed to be to find the right invariant and prove a UCT for X-equivariant KK-theory with this new invariant. For the case with one non-trivial ideal, A. Bonkat reproved Rørdams result by pro- viding a UCT for this class using the cyclic six-term exact sequence in K-theory. The second named author generalized this to two non-trivial ideals by including four cyclic six-term exact sequences. R. Meyer and R. Nest, and R. Bentmann recently proved that the obvious guess of an invariant gives a UCT for certain ideal lattices -- the so-called accordion spaces (including, e.g., all C∗-algebras with ex- actly three primitive ideals). In turn they also provide a series of counter-examples, where we do not have a UCT. They actually find examples of stable, purely infi- nite, nuclear, separable C∗-algebras in the bootstrap class with finitely many ideals having isomorphic invariants without being isomorphic. This result seems to be in sharp contrast to the stable classification result for all purely infinite Cuntz-Krieger algebras with finitely many ideals obtained by the second named author by use of methods from shift spaces. We find it very likely that the reason that Cuntz-Krieger algebras are classifiable, is the restrictive nature of their K-theory. In this paper we examine what happens to real rank zero algebras in the cases where the primitive ideal space has exactly four points. Moreover, we assume that the space is connected (since otherwise the algebras are direct sums of algebras with fewer than four primitive ideals). Also, all the basic counterexamples of R. Meyer, R. Nest, and R. Bentmann are formulated for algebras with four primitive ideals. Up to homeomorphism, there are ten different connected T0-spaces with exactly four points. These are O(X1) = {∅, {4}, {1, 4}, {2, 4}, {3, 4}, {1, 2, 4}, {1, 3, 4}, {2, 3, 4}, {1, 2, 3, 4}}, O(X2) = {∅, {4}, {3, 4}, {2, 3, 4}, {1, 3, 4}, {1, 2, 3, 4}}, O(X3) = {∅, {4}, {3, 4}, {2, 4}, {2, 3, 4}, {1, 2, 3, 4}}, O(X4) = {∅, {1}, {2}, {3}, {1, 2}, {1, 3}, {2, 3}, {1, 2, 3}, {1, 2, 3, 4}}, O(X5) = {∅, {1}, {2}, {1, 2}, {1, 2, 3}, {1, 2, 3, 4}}, O(X6) = {∅, {3}, {4}, {3, 4}, {1, 3, 4}, {2, 3, 4}, {1, 2, 3, 4}}, O(X7) = {∅, {1}, {1, 2}, {1, 2, 3}, {1, 2, 3, 4}}, O(X8) = {∅, {1}, {4}, {1, 2}, {1, 4}, {1, 2, 3}, {1, 2, 4}, {1, 2, 3, 4}}, O(X9) = {∅, {1}, {3}, {1, 3}, {3, 4}, {1, 2, 3}, {1, 3, 4}, {1, 2, 3, 4}}, O(X10) = {∅, {2}, {1, 2}, {2, 3}, {1, 2, 3}, {2, 3, 4}, {1, 2, 3, 4}}. R. Meyer and R. Nest, and R. Bentmann have proved that the spaces X7, X8, X9 and X10 have a UCT, and thus we can classify stable, purely infinite, nuclear, separable C∗-algebras in the bootstrap class with these spaces as primitive ideal spaces. Moreover they have provided counter-examples for classification for all the spaces X1, X2, X3, X4, X5, X6. In this paper we prove the following Theorem 1.1. Let A and B be purely infinite, nuclear, separable C∗-algebras of real rank zero in the bootstrap class of R. Meyer and R. Nest (cf. [MN09, 4.11]). Assume that the primitive ideal space of A and B both are homeomorphic to Xi for an i = 1, 2, 4, 5, 7, 8, 9, 10. FILTRATED K-THEORY FOR REAL RANK ZERO C ∗-ALGEBRAS 3 (1) If A and B are stable, then every isomorphism from FK(A) to FK(B) can be lifted to a ∗-isomorphism from A to B. (2) If A and B are unital, then every isomorphism from FK(A) to FK(B) that preserves the unit can be lifted to a ∗-isomorphism from A to B. Theorem 1.2. There exist stable, purely infinite, nuclear, separable C∗-algebras of real rank zero in the bootstrap class of R. Meyer and R. Nest (cf. [MN09, 4.11]) with the primitive ideal space homeomorphic to X3, which have isomorphic filtrated K-theory without being isomorphic. where FK denotes the functor filtrated K-theory which will be defined shortly. For the case where the primitive ideal space is isomorphic to X6 there are still no counterexamples for the real rank zero case -- however our methods do not apply as there is no known finite refinement of FK which gives a UCT. In general the unital part of Theorem 1.1 follows from the stable part by using results from [RR07]. For X7, Theorem 1.1 is proved by R. Meyer and R. Nest in [MN, 4.14], for X8, X9 and X10, it is proved by R. Bentmann in [Ben10, 5.4.2]. In Section 2 of this paper we set up notation and prove some preliminary results used later in this paper. In Sections 3 and 4 Theorem 1.1 is proved for X1, X2, X4 and X5 (cf. Corollaries 3.9 and 4.6 and Remarks 3.10 and 4.7). The proofs rely on the result [Kir00, 4.3] of E. Kirchberg that KK(X)-equivalences lift to X- equivariant isomorphisms for stable, separable, nuclear, purely infinite C∗-algebras with primitive ideal space homeomorphic to a finite T0-space X. Theorem 1.2 is proved in Section 5. 2. Preliminaries and notation In this section, we briefly discuss C∗-algebras over a topological space X and the invariant introduced by R. Meyer and R. Nest in [MN] called filtrated K-theory. We refer the reader to [MN] for details. We would like to note that there are other invariants in the literature which are closely related to filtrated K-theory. Examples are filtered K-theory and ideal related K-theory. It has been proved by R. Meyer and R. Nest in [MN] and R. Bent- mann in [Ben10, 5.4.2] that for the spaces Xi that these invariants are naturally isomorphic to filtrated K-theory. It is not known if these invariants are naturally isomorphic for all finite topological spaces. 2.1. C∗-algebras over a topological space X. A C∗-algebra over a topologi- cal space X is a pair (A, ψ) consisting of a C∗-algebra A and a continuous map ψ : Prim(A) → X where Prim(A) denotes the primitive ideal space of A. Assume from now on that X is a finite topological space satisfying the T0 separation axiom, i.e., such that {x} 6= {y} for all x, y ∈ X with x 6= y. Let O(X) denote the open subsets of X, and let I(A) denote the lattice of (two-sided, closed) ideals of A. A C∗-algebra over X can then equivalently be defined as a pair (A, ψ) consisting of a C∗-algebra A and a map ψ : O(X) → I(A) that preserves infima and suprema. We then write A(U ) for ψ(U ). The locally closed subsets of X are denoted by LC(X) = {U \V V, U ∈ O(X), V ⊆ U }, and the connected, non-empty, locally closed subsets of X are de- noted by LC(X)∗. For Y ∈ LC(X) we define A(Y ) = A(U )/A(V ) when Y = U \V for some V, U ∈ O(X) satisfying V ⊆ U . Up to canonical isomorphism, A(Y ) does not depend on the choice of U and V . 4 SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ For C∗-algebras A and B over X, we say that a ∗-homomorphism ϕ : A → B is X- equivariant if ϕ(A(U )) ⊆ B(U ) holds for all U ∈ O(X). An extension A ֒→ B ։ C is called X-equivariant if it induces an extension A(U ) ֒→ B(U ) ։ C(U ) for all U ∈ O(X). E. Kirchberg has constructed X-equivariant KK-theory KK∗(X; −, −), also called ideal related KK-theory and here referred to as KK(X)-theory. We denote by KK(X) the category of separable C∗-algebras over X with KK0(X)-classes as mor- phism groups. In [MN09, 3.11], R. Meyer and R. Nest show that the category KK(X) equipped with the suspension automorphism S and mapping cone trian- gles as distinguished triangles is triangulated; so mapping cones of X-equivariant ∗-homomorphisms give exact triangles, and so do extensions over X that split by an X-equivariant completely positive contraction. 2.2. Filtrated K-theory FK and the UCT. One defines for each Y ∈ O(X) the functor FKY by FKY (A) = K∗(A(Y )). We write FKi Y (A) for Ki(A(Y )). In [MN] R. Meyer and R. Nest construct commutative C∗-algebras RY over X such that KK∗(X; RY , −) and FKY are equivalent functors. By the Yoneda Lemma, cf. [ML98, 3.2], the set N T (Y, Z) of natural transforma- tions from the functor FKY to the functor FKZ is then given by KK∗(X; RZ, RY ). Given α ∈ KK∗(X; RZ, RY ) we denote by ¯α the corresponding element in N T (Y, Z) given by α ⊠ − where − ⊠ − denotes the X-equivariant Kasparov product. Given f ∈ N T (Y, Z), we let f denote the corresponding element in KK∗(X; RZ, RY ). The functor FK is then defined as the family of functors (FKY )Y ∈LC(X)∗ to- gether with the sets N T (Y, Z) of natural transformations. The target category of FK is the category of modules over the ring N T = LY,Z∈LC(X)∗ N T (Y, Z). A homomorphism FK(A) → FK(B) is then a family of homomorphisms (ϕY ) that respects the natural transformations. Kasparov multiplication induces a map KK∗(X; A, B) → Hom(FK(A), FK(B)), and for A = RY this map is an isomor- phism. In [MN] R. Meyer and R. Nest establish a UCT for KK(X)-theory, i.e., they establish exactness of Ext1 N T (FK(A), FK(B)) ֒→ KK∗(X; A, B) ։ HomN T (FK(A), FK(B)) for A and B separable C∗-algebras over X with A belonging to the bootstrap class B(X) defined by R. Meyer and R. Nest, cf. [MN09, 4.11], and with FK(A) having projective dimension at most 1 as a module over N T . By construction, FK(RY ) has projective dimension 0 for all Y ∈ LC(X)∗. By [MN09, 4.13], a nuclear C∗-al- gebra over X belongs to B(X) if and only if its simple subquotients belong to the bootstrap class of J. Rosenberg and C. Schochet. 2.3. Construction of RY . The C∗-algebras RY are constructed as follows. Define a partial order on X by x ≤ y when x ∈ {y}. The order complex Ch(X) is the geo- metric realisation of the simplicial set whose nondegenerate n-simplices [x0, . . . , xn] are strict chains x0 < · · · < xn. Maps m, M : Ch(X) → X is then defined by the inner of a simplex [x0, . . . , xn] being sent to x0 by m and to Xn by M . The C∗-algebras RY over X are then defined by RY (Z) = C0(m−1(Y ) ∩ M −1(Z)) for all Y, Z ∈ LC(X). For Y ∈ LC(X) and U ∈ O(Y ), we then get X-equivariant extensions RY \U ֒→ RY ։ RU . The natural transformation given by RY \U ֒→ RY is denoted rY \U Y FILTRATED K-THEORY FOR REAL RANK ZERO C ∗-ALGEBRAS 5 and called a restriction map, the natural transformation given by RY ։ RU is denoted by iY U and called an extension map, and the natural transformation given by RY \U ֒→ RY ։ RU is denoted by δU Y \U and called a boundary map. For a C∗-algebra A over X, these natural transformations are the ones appearing in the six-term exact sequence induced by the extension A(U ) ֒→ A(Y ) ։ A(Y \U ). It is unknown whether there exists finite T0-spaces X over which the ring N T is not generated by transformations of this form, but for the spaces X1, X2, . . . , X10 considered in this paper, this is not the case. 3. The counterexample of Meyer and Nest We now restrict to the space X1 = {1, 2, 3, 4} with O(X1) = {∅} ∪ {U ⊆ X1 4 ∈ U }. We abbreviate, e.g., {1, 2, 3} to 123. A C∗-algebra A over X1 is then an extension of the form A(4) ֒→ A ։ A(1) ⊕ A(2) ⊕ A(3). The ordering on X induced by its topology is then defined by i ≤ 4 for all i ∈ X1, its Hasse diagram (or, more correctly, the Hasse diagram of the inverse order relation) is 1 _❄❄❄❄ 2 4 ?⑧⑧⑧⑧ 3 , and LC(X1)∗ = {4, 14, 24, 34, 124, 134, 234, 1234, 1, 2, 3}. In [MN] it is shown that the ring N T = LY,Z∈LC(X1)∗ N T (Y, Z) is generated by natural transformations i, r and δ that are induced by six-term exact sequences, and the indecomposable transformations are of infinite order and fit into the following diagram 4 i =④④④④④④④④ i !❈❈❈❈❈❈❈❈ i 14 24 34 i i i =④④④④④④④ !❈❈❈❈❈❈❈ =④④④④④④④ i !❈❈❈❈❈❈❈ i i 124 i !❈❈❈❈❈❈❈ =④④④④④④④ 134 i i / 234 / 1234 r =④④④④④④④④ r !❈❈❈❈❈❈❈❈ r 1 2 3 ❈❈❈❈ δ ◦ δ ◦ !❈❈❈❈ / 4 =④④④④ δ ◦ ④④④④ where the six squares commute and the sum of the three transformations from 1234 to 4 vanishes. 3.1. The refined invariant. In [MN], R. Meyer and R. Nest refine the invariant FK to an invariant FK′. They prove a UCT for this refined invariant, so for A and B in the bootstrap class B(X1) one can lift isomorphisms between FK′(A) and FK′(B) to KK(X1)-equivalences, and by combining this with the classification result [Kir00, 4.3] of E. Kirchberg conclude that it strongly classifies the stable, purely infinite, separable, nuclear C∗-algebras A that are tight over X1 and whose simple subquotients A(4), A(1), A(2) and A(3) lie in the bootstrap class, see [MN, 5.14, 5.15]. In [RR07], the second and third author showed how one can strongly classify a class of unital properly infinite C∗-algebras given that the this class are strongly classified up to stable isomorphism. Since FK′(·) strongly classifies the class of stable, purely infinite, separable, nuclear C∗-algebras A that are tight over X1, by Theorem 2.1 of [RR07], FK′(·) together with class of the unit strongly classifies the class of unital, purely infinite, separable, nuclear C∗-algebras A that are tight over X1. _ O O ? / / ! ! ! = / / ! = ! / = / / ! / = / = = 6 SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ The invariant is defined by constructing a C∗-algebra R12344 over X1 and adding KK∗(X1; R12344, −) to the family of functors. The indecomposable transformations in the larger ring N T ′ = LY,Z∈LC(X1)∗∪{12344} N T (Y, Z) fit into the following diagram: 4 i =④④④④④④④④ i !❈❈❈❈❈❈❈❈ i 14 24 34 !❈❈❈❈❈❈❈ =④④④④④④④ / 12344 134 i / 1234 =④④④④④④④ !❈❈❈❈❈❈❈ i !❈❈❈❈❈❈❈ =④④④④④④④ i 124 234 r =④④④④④④④④ r !❈❈❈❈❈❈❈❈ r 1 2 3 ❈❈❈❈ δ ◦ δ ◦ !❈❈❈❈ / 4 =④④④④ δ ◦ ④④④④ (1) The C∗-algebra R12344 is the mapping cone of a generator of the cyclic free group N T (234, 14) and its filtrated K-theory is 0 0 i =④④④④④④④ i !❈❈❈❈❈❈❈ i Z Z i i i =④④④④④④④④ !❈❈❈❈❈❈❈❈ =④④④④④④④④ i !❈❈❈❈❈❈❈❈ i i i i !❈❈❈❈❈❈❈ =④④④④④④④④ i / Z2 Z Z r =④④④④④④④ r !❈❈❈❈❈❈❈❈ r ❈❈❈❈ δ ◦ δ ◦ ④④④④ !❈❈❈ δ / Z[1] ◦ / =④④④ Z[1] (2) 0 where the three maps i1234 ij4 Z3/(1, 1, 1), the three maps rk Z2/(1, 1) onto coordinate hyperplanes, and the three maps δ4 are given by the three coordinate embeddings Z → 1234 are given by the three projections Z3/(1, 1, 1) → k are the identity. Z / Z Lemma 3.1. Assume that FKY (A) and FKY (R12344) are isomorphic for all Y ∈ LC(X1)∗ and that i1234 134 : FK124(A) ⊕ FK134(A) → FK1234(A) is an isomor- phism. Then FK(A) and FK(R12344) are isomorphic as N T -modules and A and R12344 are KK(X1)-equivalent. 124 ⊕ i1234 Proof. Define for each Y ∈ LC(X)∗ an N T -module PY as PY (Z) = N T (Y, Z). Then PY is freely generated by idY ∈ PY (Y ) as an N T -module. Define j : P1234 → P124 ⊕ P134 ⊕ P234 by f 7→ f i1234 234 . Then FK(R12344) is isomorphic to coker j as N T -modules, cf. [MN, Section 5], with im j generated by i1234 134 + i1234 234 . 124 + f i1234 134 + f i1234 124 +i1234 Hence an N T -morphism FK(R12344) → FK(A) can be defined by choosing el- ements gY ∈ FKY (A), Y ∈ {124, 134, 234}, satisfying i1234 134 (g134) + i1234 234 (g234) = 0, and defining the map by idY 7→ gY for Y ∈ {124, 134, 234} and expanding by N T -linearity. 124 (g124) + i1234 If gY generates FKY (A) for all Y ∈ {124, 134, 234}, then the morphism will be it is automatically bijective FKZ (R12344) → FKZ(A) for Z ∈ an isomorphism: {124, 134, 234}, by the assumptions in the lemma it is therefore surjective and hence bijective for Z = 1234, and by exactness it then follows that it is bijective for Z ∈ {1, 2, 3} whereby bijectivity for Z = 4 also follows, cf. the Diagram (2). Let gY be a generator of FKY (A) for Y ∈ {124, 134, 234}. Since i1234 is an isomorphism, FK1234(A) is spanned by i1234 write i1234 0, FK24(A) = 0 and FK14(A) = 0, the four maps r2 234 : FK234(A) → FK3(A), r2 r3 124 ⊕ i1234 134 134 (g134) so we may 134 (g134) for some m, n ∈ Z. Since FK34(A) = 234 : FK234(A) → FK2(A), 134 : FK134(A) → 124 : FK124(A) → FK2(A) and r3 124 (g124) and i1234 234 (g234) = mi1234 124 (g124) + ni1234 ! ! ! = / / ! / = / / ! / = / / ! / = = = / / ! ! ! = / / ! = ! / = / / ! = / = = FILTRATED K-THEORY FOR REAL RANK ZERO C ∗-ALGEBRAS 7 234(g234) = mr2 FK3(A) are isomorphisms, so r2 124(g124) both generate 134(g134) and r3 FK2(A), and r3 134(g134) both generate FK3(A), so m, n ∈ {±1}. By replacing g124 with −mg124 and g134 with −ng134 the required is fulfilled. 124(g124) and r2 234(g234) = nr3 In the discussion after the proof of Lemma 5.9 in [MN], we have that the natural homomorphism from KK(X1; R12344, A) to Hom(FK(R12344), FK(A)) is an isomor- phism. Since FK(A) and FK(R12344) are isomorphic as N T -modules, we have that A and R12344 are KK(X1)-equivalent. (cid:3) Lemma 3.2. There exists an exact triangle R1234 π ◦ R12344 !❈❈❈❈❈❈❈ ϕ =④④④④④④④ ι R124 ⊕ R134 ⊕ R234 satisfying that ¯ϕ = (i1234 234 ) ∈ L N T (ij4, 1234), that ¯π generates the group N T (1234, 12344), and that ¯ι = (f 124, f 134, f 234) ∈ L N T (12344, ij4) with each f ij4 generating N T (12344, ij4) respectively. 124 , i1234 134 , i1234 Proof. Let ϕ : R1234 → R124 ⊕ R134 ⊕ R234 be given by restriction to subsets; then ¯ϕ = (i1234 234 ). Constructing the mapping cone 0 ϕ of ϕ, we get an exact triangle 134 , i1234 124 , i1234 R1234 π ◦ S 0 ϕ !❈❈❈❈❈❈❈ ϕ =④④④④④④ ι R124 ⊕ R134 ⊕ R234 and by applying KK∗(X1; RY , −) = FKY and calculating FKY (ϕ), one sees that FKY (S 0 ϕ) and FKY (R12344) are isomorphic for all Y ∈ LC(X1)∗, cf. Diagram (2). Furthermore one sees that FKij4(ι) are isomorphisms, and that FK1234(ι) is surjective as FK1234(ϕ) is injective and (using standard generators) is given by Z ∋ x 7→ (x, x, x) ∈ Z ⊕ Z ⊕ Z. Using that FKY (ι) respects the natural transfor- mations, and that the natural transformation FK124(L Rij4) ⊕ FK134(L Rij4) → FK1234(L Rij4) is given by Z ⊕ Z ∋ (x, y) 7→ (x, y, 0) ∈ Z ⊕ Z ⊕ Z (using standard generators), one can then check that FK124(S 0 ϕ)⊕ FK134(S 0 ϕ) → FK1234(S 0 ϕ) is an isomorphism. Hence S 0 ϕ and R12344 are KK(X1)-equivalent by Lemma 3.1. Therefore π and ι induce natural transformations, and since all the involved groups of natural transformations are cyclic and free, we may write ¯π = nf1234 with f1234 generating N T (1234, 12344) and ¯ι = (nij4f ij4) with f ij4 generating the group N T (12344, ij4). Then FKij4(ι) = nij4 FKij4( f ij4), since FKij4(R124⊕R134⊕R234) = FKij4(Rij4), so as FKij4(ι) is an isomorphism Z → Z, we see that nij4 = ±1. But FKY (π) = 0 for all Y . However, since FKij4(R1234) = 0 and FK1234(R1234) = Z, we get by applying KK∗(X1; −, R1234) to the exact triangle that ¯π = nf1234 on R1234 is an isomorphism Z → Z[1], hence n = ±1. (cid:3) ! o o = ! o o = 8 SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ Lemma 3.3. There exists an exact triangle R12344 ι ◦ R4 !❈❈❈❈❈❈❈ π =④④④④④④④ ϕ R14 ⊕ R24 ⊕ R34 satisfying that ¯ϕ = (i14 4 ) ∈ L N T (4, k4), that ¯ι generates N T (12344, 4), and that ¯π = (f14, f24, f34) ∈ L N T (k4, 12344) with each fk4 generating the group N T (k4, 12344) respectively. 4 , i24 4 , i34 Proof. Let ϕ : R14 ⊕ R24 ⊕ R34 → M3(R4) be given by restriction to subsets such that ¯ϕ = (i14 4 ) and construct the mapping cone 0 ϕ of ϕ. By calculating FKY (ϕ) and by applying FKY to the mapping cone triangle 4 , i24 4 , i34 0 ϕ ι ◦ R4 =④④④④④④④ ϕ !❈❈❈❈❈❈ π R14 ⊕ R24 ⊕ R34 we see that FKY (0 ϕ) ∼= FKY (R12344) for all Y ∈ LC(X1)∗. Furthermore we see that FK4(ι) and FKk(π) are isomorphisms, and that FKij4(π) and FK1234(π) are injective as FKij4(ϕ) and FK1234(ϕ) are surjective and (by using standard generators) are given by Z ⊕ Z ∋ (x, y) 7→ x + y ∈ Z respectively Z ⊕ Z ⊕ Z ∋ (x, y, z) 7→ x + y + z ∈ Z. Using that FKY (π) respects the natural transformations, and that the natural transformation FK124(L Rk4) ⊕ FK134(L Rk4) → FK1234(L Rk4) is given by (Z ⊕ Z)⊕(Z⊕Z) ∋ (x, y, z, w) 7→ (x+z, y, w) ∈ Z⊕Z⊕Z (using standard generators), one can then check that FK124(0 ϕ) ⊕ FK134(0 ϕ) → FK1234(0 ϕ) is an isomorphism. Hence 0 ϕ and R12344 are KK(X1)-equivalent by Lemma 3.1. Therefore π and ι induce natural transformations, so we may write ¯ι = nf 4 with f 4 generating N T (12344, 4) and ¯π = (nk4fk4) with fk4 generating the group N T (k4, 12344). As FK4(ι) = n FK4( f 4) is an isomorphism Z → Z[1], we see that n = ±1. And as FKk(R14 ⊕ R24 ⊕ R34) = FKk(Rk), we see that FKk(π) = nk4 FKk( fk4), so as FKk(π) is an isomorphism Z → Z, we see that nk4 = ±1. (cid:3) Lemma 3.4. There exist natural transformations f14, f24, f34, f 124, f 134, f 234 such that hfk4i = N T (k4, 12344) and (cid:10)f ij4(cid:11) = N T (12344, ij4) and such that the se- quences FK1234(A) 4 δ4 fm4im4 ◦ nrn 1234 / FK12344(A) f▲▲▲▲▲▲▲▲ (i1234 ij4 ) xrrrrrrrr (f ij4) FK124(A) ⊕ FK134(A) ⊕ FK234(A) and FK12344(A) mrm mn4f mn4 δ4 ◦ / FK4(A) f▲▲▲▲▲▲▲▲ (fk4) xrrrrrrrr (ik4 4 ) FK14(A) ⊕ FK24(A) ⊕ FK34(A) are exact for all C∗-algebras A over X1 and all m, n ∈ {1, 2, 3} with m 6= n. ! o o = ! o o = / x f / x f FILTRATED K-THEORY FOR REAL RANK ZERO C ∗-ALGEBRAS 9 Proof. This follows from Lemmas 3.2 and 3.3 by applying KK∗(X1; −, A) and using that by the Diagram (1) the transformation fm4im4 1234 generates the group N T (1234, 12344) and δ4 (cid:3) mn4f mn4 generates N T (12344, 4). 4 δ4 mrm nrn 3.2. A classification result. m(A) → FK1−n Proposition 3.5. Let A and B be C∗-algebras over X1 and assume that the maps m : FKn δ4 (B) vanish for some m ∈ {1, 2, 3} and some n ∈ {0, 1}. Then any homomorphism ϕ : FK(A) → FK(B) can be uniquely extended to a homomorphism ϕ′ : FK′(A) → FK′(B). Furthermore, if ϕ is an isomorphism then so is ϕ′. m(B) → FK1−n (A) and δ4 m : FKn 4 4 Proof. Let ϕ : FK(A) → FK(B) be a homomorphism. We may extend it by defining ϕ12344 : FK12344(A) → FK12344(B) by the following diagrams: 0 0 / FK1−n 12344(A) (f ij4) ϕ1−n 12344 / FK1−n 12344(B) (f ij4) FK1−n 124 (A) ⊕ FK1−n 134 (A) ⊕ FK1−n 234 (A) ϕ1−n 124 ⊕ϕ1−n 134 ⊕ϕ1−n 234 / FK1−n 124 (B) ⊕ FK1−n 134 (B) ⊕ FK1−n 234 (B) (i1234 ij4 ) FK1−n 1234(A) ϕ1−n 1234 (i1234 ij4 ) / FK1−n 1234(B) (ik4 4 ) FKn 4 (A) ϕn 4 (ik4 4 ) FKn 4 (B) FKn 14(A) ⊕ FKn 24(A) ⊕ FKn 34(A) ϕn 14⊕ϕn 24⊕ϕn 34 / FKn 14(B) ⊕ FKn 24(B) ⊕ FKn 34(B) (fk4) FKn 12344(A) ϕn 12344 (fk4) / FKn 12344(B) 0 / 0 By Lemma 3.4 the four horizontal sequences in the diagrams are exact, hence ϕ12344 is well-defined and is bijective if ϕ is an isomorphism. By construction ϕn 12344 respects the natural transformations fk4 and ϕ1−n spects the naturals transformations f ij4. Since (f ij4) is injective on FK1−n and since (fk4) is surjective on FKn 12344fk4 = (f ij4)fk4ϕ1−n 12344(A), it suffices to check that ij4f ij4(fk4) = f ij4ϕn (f ij4)ϕ1−n 12344(fk4). and ϕn k4 12344 re- 12344(B) And this holds by construction of ϕ12344 as f ij4fk4ϕk4 = ϕij4f ij4fk4 since f ij4fk4 ∈ N T (k4, ij4). Since the natural transformations in FK′ are generated by the natural trans- formations in FK together with the natural transformations fk4 and f ij4, we see that the extended ϕ respects all the natural transformations in FK′, hence it is an N T ′-morphism between FK′(A) and FK′(B). (cid:3) Observation 3.6. A separable, nuclear, purely infinite, tight C∗-algebra A over a finite T0-space X is of real rank zero if and only if the boundary map δU Y \U vanishes on K0(A(Y \U )) for all Y ∈ LC(X) and all U ∈ O(Y ). This follows from the fact that all Kirchberg algebras have real rank zero combined with the following result of L. G. Brown and G. K. Pedersen, cf. [BP91, 3.14]: Given an extension I ֒→ B ։ B/I of C∗-algebras, B has real rank zero if and only if I and B/I have real rank zero and projections in B/I lift to projections in B. / / / / / / / / / / / / / / / / / 10 SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ Corollary 3.7. Let A and B be C∗-algebras in the bootstrap class over X1 and with A of real rank zero. Then any isomorphism between FK(A) and FK(B) lifts to a KK(X1)-equivalence. Proof. Since A is of real rank zero, δ4 4(A) vanishes by [BP91, 3.14], and since FK(A) and FK(B) are isomorphic, δ4 4(B) also vanishes. By Proposition 3.5 the isomorphism therefore lifts to an isomorphism between FK′(A) and FK′(B), and by [MN, 5.14] this lifts to a KK(X1)-equivalence. (cid:3) 2(B) → FK1 2 : FK0 2 : FK0 2(A) → FK1 Definition 3.8. Let A and B be unital C∗-algebras over X. Then ϕ : FK(A) → FK(B) is a homomorphism that preserves the unit if ϕ is a homomorphism of N T -modules and ϕX ([1A]) = [1B] in FKX (A) = FKX (B). We say that ϕ is an isomorphism that preserves the unit if ϕ is an isomorphism of N T -modules that preserves the unit. Combining this with [Kir00, 4.3] and [RR07, 2.1,3.2], we obtain the following corollary. Corollary 3.9. Let A and B be separable, nuclear, purely infinite C∗-algebras that are tight over X1 and whose simple subquotients lie in the bootstrap class. Assume that A has real rank zero. (1) If A and B are stable, then every isomorphism from FK(A) to FK(B) can be lifted to a ∗-isomorphism from A to B. (2) If A and B are unital, then every isomorphism from FK(A) to FK(B) that preserves the unit can be lifted to a ∗-isomorphism from A to B. Remark 3.10. The space X4 = X op that the indecomposable transformations for X op 1 are 1 has been studied in [BK] where it is shown 1234 134 r =④④④④④④④ r / !❈❈❈❈❈❈❈ r 124 234 r 34 r r =④④④④④④④ !❈❈❈❈❈❈❈ =④④④④④④④ !❈❈❈❈❈❈❈ r r r 24 14 r r !❈❈❈❈❈❈❈❈ =④④④④④④④④ r / 4 δ ◦ δ ◦ ④④④④ ❈❈❈❈ =④④④④ δ ◦ !❈❈❈❈ 1 2 3 i i !❈❈❈❈❈❈❈❈ =④④④④④④④④ i . / 1234 It is straightforward to copy the methods of Meyer and Nest in [MN] to construct a refined filtrated K-theory for which there is a UCT; for X op the extra representing 1 object is the mapping cone of a generator of N T (14, 234). The methods we used for the spaces X1 apply to X op 1 as well since the boundary maps δ are placed in similar places in the structure diagrams for N T of X op 1 . 4. Another counterexample Consider the space X2 = {1, 2, 3, 4} with O(X2) = {∅, 4, 34, 234, 134, X2}. Then 1 < 3, 2 < 3 and 3 < 4, LC(X2)∗ = {4, 34, 234, 134, 1234, 3, 23, 13, 123, 1, 2}, and / / ! ! ! = / ! = ! / = / / ! / / / = = = FILTRATED K-THEORY FOR REAL RANK ZERO C ∗-ALGEBRAS 11 its Hasse diagram is 1 \✾✾✾✾ 2 B✆✆✆✆ . 3 4 The indecomposable transformations in the category N T have been studied in detail in [Ben10, 6.1.2] and are the maps in the following diagram: 34 3 1234 r / / 123 134 r 13 i r =④④④④④④④ !❈❈❈❈❈❈❈ i i i =④④④④④④④④ !❈❈❈❈❈❈❈ =④④④④④④④ !❈❈❈❈❈❈❈❈ i i r 234 23 i !❈❈❈❈❈❈❈ =④④④④④④④ i r δ ◦ =④④④④④④④④ !❈❈❈❈❈❈❈❈ r 1 4 2 ❈❈❈❈ δ ◦ δ ◦ ④④④④ i !❈❈❈ =④④④④ / 34 As with the first counterexample, there exists a refinement FK′ of FK for which there is a UCT, cf. [Ben10, 6.1], so for A and B in the bootstrap class B(X2) one can lift an isomorphism between FK′(A) and FK′(B) to a KK(X2)-equivalence. For X2 one constructs an extra representing object R12334 as the mapping cone of a generator of the cyclic free group N T (23, 134), and its filtrated K-theory is then Z[1] r =④④④④④④④ i !❈❈❈❈❈❈❈ i 0 0 0 Z Z r i i =④④④④④④④④ !❈❈❈❈❈❈❈❈ =④④④④④④④④ r !❈❈❈❈❈❈❈❈ i i / Z i r !❈❈❈❈❈❈❈ =④④④④④④④④ i / Z2 r =④④④④④④④ δ !❈❈❈❈❈❈❈❈ ◦ / r Z ❈❈❈❈ Z[1] δ ◦ ④④④④ Z δ ◦ i !❈❈❈ / Z[1] =④④④ 13 , r123 where the three maps i123 23 are given by the three coordinate embed- dings Z → Z3/(1, 1, 1), the three maps r1 123 are given by the three projections Z3/(1, 1, 1) → Z2/(1, 1) onto coordinate hyperplanes, and the three maps δ34 1234 and i123 123 and r2 123, δ4 2 are the identity. 1 , i34 4 and δ34 Since pd(FK(R12334)) = 1, we see that for any C∗-algebra A over X2 that lies in the bootstrap class over X2, A and R12334 will be KK(X2)-equivalent if and only if the groups FKY (A) and FKY (R12334) are isomorphic for all Y ∈ LC(X2)∗ and the natural transformation FK13(A) ⊕ FK1234(A) → FK123(A) is an isomorphism, cf. Lemma 3.1. The indecomposable transformations in the ring N T ′ fit into the following diagram: 34 3 / 12334 1234 r / / 123 134 i r =④④④④④④④ !❈❈❈❈❈❈❈ i !❈❈❈❈❈❈❈ =④④④④④④④ 234 =④④④④④④④ !❈❈❈❈❈❈❈ 13 23 i !❈❈❈❈❈❈❈ =④④④④④④④ i r δ ◦ =④④④④④④④④ !❈❈❈❈❈❈❈❈ r 1 4 2 (3) ❈❈❈❈ δ ◦ δ ◦ ④④④④ i !❈❈❈ =④④④④ / 34 \ B O O / / ! ! ! / / = ! = ! / / = ! / / / = = = / / ! ! ! = / / ! = ! / = / ! / = / = = ! ! ! / / = ! / = / / ! / / = ! / = = = 12 SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ 4.1. The refined invariant. Lemma 4.1. There exists an exact triangle R123 π ◦ R12334 !❈❈❈❈❈❈❈ ϕ =④④④④④④④ ι R13 ⊕ R1234 ⊕ R23 satisfying that ¯ϕ = (i123 and that ¯ι = (f 13, f 1234, f 23) with f Y generating N T (12344, Y ). 23 ), that ¯π generates the group N T (123, 12334), 1234, i123 13 , r123 Proof. Let ϕ : R123 → R13 ⊕ R1234 ⊕ R23 be given by inclusion respectively restric- tions to subspaces, such that ¯ϕ = (i123 23 ). The proof is similar to the proof of Lemma 3.2. Here FK(R123) is used to establish that ¯π is a generator, and FKY is used for f Y . (cid:3) 1234, i123 13 , r123 Lemma 4.2. There exists an exact triangle R12334 ι ◦ R34 !❈❈❈❈❈❈❈ π =④④④④④④④ ϕ R134 ⊕ R3 ⊕ R234 satisfying that ¯ϕ = (i134 (f134, f3, f234) with each fY generating the group N T (Y, 12344) respectively. 34 ), that ¯ι generates N T (12334, 34), and that ¯π = 34, i234 34 , r3 Proof. Let ϕ : R134 ⊕ R3 ⊕ R234 → M3(R34) be given by inclusions respectively restriction to a subspace, such that ¯ϕ = (i134 34 ). The proof is similar to the proof of Lemma 3.3. Here FK4 is used to establish that ¯ι is a generator, and FKY is used for fY . (cid:3) 34, i234 34 , r3 Lemma 4.3. There exist natural transformations f134, f3, f234, f 13, f 1234, f 23 such that hfY i = N T (Y, 12334) and (cid:10)f Y (cid:11) = N T (12334, Y ) and such that the sequences FK123(A) f134i134 ◦ 4 δ4 123 / FK12334(A) f▲▲▲▲▲▲▲▲ xrrrrrrrr 1234,i123 23 ) (f 13,f 1234,f 23) (i123 13 ,r123 FK13(A) ⊕ FK1234(A) ⊕ FK23(A) and r34 4 δ4 123i123 23 f 23 ◦ / FK34(A) FK12334(A) f▲▲▲▲▲▲▲▲ 34,i234 (f134 ,f3,f234) 34 ) FK134(A) ⊕ FK3(A) ⊕ FK234(A) 34 ,r3 xrrrrrrrr (i134 are exact for all C∗-algebras A over X2. Proof. This follows from Lemmas 4.1 and 4.2 by applying KK∗(X2; −, A) and using that by the Diagram (3) the transformation f134i134 123generates N T (123, 12334) and the transformation r34 (cid:3) 23 f 23generates N T (12334, 34). 123i123 4 δ4 4 δ4 ! o o = ! o o = / x f / x f FILTRATED K-THEORY FOR REAL RANK ZERO C ∗-ALGEBRAS 13 4.2. A classification result. A slightly more general result, like the result in Section 3.2, can be obtained, but we state a weaker result to ease notation. Proposition 4.4. Let A and B be C∗-algebras over X2 and assume that A and B have real rank zero. Then any homomorphism ϕ : FK(A) → FK(B) can be uniquely extended to a homomorphism ϕ′ : FK′(A) → FK′(B). Furthermore, if ϕ is an isomorphism then so is ϕ′. Proof. The proof is similar to the proof of Theorem 3.5. Since A and B have real rank zero, δ4 4(B) vanish, so by Lemma 4.3 the horizontal sequences in the following diagram are exact 123(B) → FK1 123(A) → FK1 4(A) and δ4 123 : FK0 123 : FK0 0 0 / FK1 12334(A) ϕ1 12334 / FK1 12334(B) FK1 13(A) ⊕ FK1 1234(A) ⊕ FK1 23(A) ϕ1 13⊕ϕ1 1234⊕ϕ1 23 / FK1 13(B) ⊕ FK1 1234(B) ⊕ FK1 23(B) FK1 123(A) ϕ1 123 / FK1 123(B) FK0 4(A) ϕ0 4 FK0 4(B) FK0 134(A) ⊕ FK0 3(A) ⊕ FK0 234(A) ϕ0 134⊕ϕ0 3⊕ϕ0 234 / FK0 134(B) ⊕ FK0 3(B) ⊕ FK0 234(B) FK0 12334(A) ϕ0 12334 / FK0 12334(B) 0 / 0 so we may recover FK1 cokernel of (i134 34, i234 34 , r3 12334 as the kernel of (i123 34 ), as in the proof of Theorem 3.5. 13 , r123 1234, i123 13 ) and FK0 12334 as the (cid:3) Corollary 4.5. Let A and B be C∗-algebras in the bootstrap class over X2 and assume that A has real rank zero. Then any isomorphism between FK(A) and FK(B) lifts to a KK(X2)-equivalence. Proof. Since FK(A) and FK(B) are isomorphic, δ4 4(B) van- ishes, so the proof of Proposition 4.4 applies, hence the isomorphism lifts to an isomorphism between FK′(A) and FK′(B) and by [Ben10, 6.1.22] this lifts to a KK(X2)-equivalence. (cid:3) 123(B) → FK1 123 : FK0 Corollary 4.6. Let A and B be separable, nuclear, purely infinite C∗-algebras that are tight over X2 and whose simple subquotients lie in the bootstrap class. Assume that A has real rank zero. (1) If A and B are stable, then every isomorphism from FK(A) to FK(B) can be lifted to a ∗-isomorphism from A to B. (2) If A and B are unital, then every isomorphism from FK(A) to FK(B) that preserves the unit can be lifted to a ∗-isomorphism from A to B. / / / / / / / / / / / / / / / / / 14 SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ Remark 4.7. The space X5 = X op that the indecomposable transformations for X op 2 are 2 has been studied in [BK] where it is shown 123 1234 3 / 34 r =④④④④④④④ i !❈❈❈❈❈❈❈ r r r =④④④④④④④ !❈❈❈❈❈❈❈❈ =④④④④④④④④ i !❈❈❈❈❈❈❈ r r i r !❈❈❈❈❈❈❈ =④④④④④④④ r i 23 234 13 / 134 δ ◦ r ④④④④ ❈❈❈❈ =④④④④ δ ◦ !❈❈❈❈ 1 4 2 i δ ◦ !❈❈❈❈❈❈❈❈ =④④④④④④④④ i . / 123 As with X op 1 , cf. Remark 3.10, it is straightforward to copy the methods of Meyer and Nest in [MN] to construct a refined filtrated K-theory for which there is a UCT; for X op the extra representing object is the mapping cone of a generator 2 of N T (134, 23). And as with X op 1 , the methods we used for the spaces X1 and X2 apply to X op since the boundary maps δ are placed in similar places in the 2 structure diagrams for N T of X op 2 . 5. A third counterexample Consider the space X3 = {1, 2, 3, 4} with O(X3) = {∅, 4, 24, 34, 234, X3}. Then 1 < 2, 1 < 3, 2 < 4, 3 < 4, LC(X3)∗ = {4, 24, 34, 234, 1234, 123, 12, 13, 1, 2, 3} and its Hasse diagram is ?⑧⑧⑧ _❄❄❄ 2 1 4 _❄❄❄ ?⑧⑧⑧ 3 . The indecomposable transformations in the category N T have been studied in detail in [Ben10, 6.2.2] and are displayed in the following diagram: δ ◦ 34 123 r =④④④④④④④ !❈❈❈❈❈❈❈ δ ◦ r 12 4 13 i r =④④④④④④④④ !❈❈❈❈❈❈❈❈ =④④④④④④④④ !❈❈❈❈❈❈❈❈ r i δ ◦ 1 24 i δ ◦ !❈❈❈❈❈❈❈ =④④④④④④④ i / 234 i 1234 r / / 123 r =④④④④④④④④ !❈❈❈❈❈❈❈❈ r 3 2 i !❈❈❈❈❈❈❈❈ =④④④④④④④④ i The methods used for the spaces X1 and X2 do not apply to X3 since the bound- ary maps δ are placed radically differently in the structure diagram for N T of X3. In fact, for this space X3 there does exist tight, nuclear, separable, purely infinite C∗-algebras A and B over X3 of real rank zero that are not KK(X3)-equivalent but have isomorphic filtrated K-theory. Proof of Theorem 1.2. The construction is similar to the one of R. Meyer and R. Nest in [MN, p. 27ff] and some of the details are carried out in [Ben10, 6.2.4]. Put PY (Z) = N T (Y, Z). Consider the injective map j : P234 → P24 ⊕P1[1]⊕P34 in- duced by three generators of the groups N T (24, 234), N T (1, 234) and N T (34, 234), / / ! ! ! = / / ! = ! / = / / ! / = / = = ? _ _ ? / / ! ! ! / / = ! = ! / / / = ! / / = = = FILTRATED K-THEORY FOR REAL RANK ZERO C ∗-ALGEBRAS 15 and let M denote the cokernel. Let k ≥ 2 and put Mk = M ⊗ Z/k. Then Mk is Z/k 9sssssssss %❑❑❑❑❑❑❑❑❑ ◦ 0 0 0 ◦ %❑❑❑❑❑❑❑❑❑ 9sssssssss %❑❑❑❑❑❑❑❑❑ 9sssssssss ◦ Z/k Z/k %❑❑❑❑❑❑❑❑ 9ssssssss ◦ / Z/k[1] / (Z/k)2 9ssssssss %❑❑❑❑❑❑❑❑ Z/k Z/k Z/k %❑❑❑❑❑❑❑❑ 9ssssssss / Z/k and has projective dimension 2, and 0 / P234 / P234 ⊕ P24 ⊕ P1[1] ⊕ P34 / P24 ⊕ P1[1] ⊕ P34 / Mk / 0 is a projective resolution of Mk. Notice that the boundary maps from even to odd parts in Mk vanish. There exists in the bootstrap class over X3 a C∗-algebra Ak with FK(Ak) = Mk, see [Ben10, 6.2.4] for details. Let Q2 / Q1 / Q0 / Ak be a ker FK-projective resolution which is a lift of the above projective resolution of Mk, and let Ak N0 a❈❈❈❈ / N1 ◦④④ }④④ P0 a❈❈❈❈ / N2 ◦④④ }④④ P1 a❈❈❈❈ / N3 ◦④④ }④④ P2 _❅❅❅❅ N3 · · · ◦⑧⑧ ⑧⑧ 0 · · · be its phantom tower. Then N2 ∼=KK(X3) Q2 and the composite map Ak → N2 lies in (ker FK)2. Construct B as the mapping cone of Ak → N2. Then B and Ak ⊕SN2 are not KK(X3)-equivalent but have FK(B) ∼= FK(Ak)⊕FK(N2)[1] = Mk ⊕P234[1]. See [MN, 4.10, 5.5] for more details. Since all KK(X3)-equivalence classes in the bootstrap class over X3 can be rep- resented by tight, stable, nuclear, separable, purely infinite C∗-algebras over X3, cf. [MN, 4.6], we can find such C and D with C ∼=KK(X3) B, D ∼=KK(X3) Ak ⊕ SN2 and FK(C) ∼= FK(D) ∼= FK(B). Since P234[1] is Z[1]2 ◦ Z =④④④④④④ !❈❈❈❈❈❈ Z[1] ◦ =④④④④④④④④ !❈❈❈❈❈❈❈ =④④④④④④④ !❈❈❈❈❈❈❈❈ ◦ 0 0 0 !❈❈❈❈❈❈❈ =④④④④④④④ ◦ / / Z[1] =④④④④④④ !❈❈❈❈❈❈ Z[1] Z[1] !❈❈❈❈❈❈ =④④④④④④ Z[1] / Z[1]2 Z[1] , we see that the boundary maps from even to odd parts in FK(B) vanish, so C and D will be of real rank zero as their simple subquotients are Kirchberg algebras and therefore of real rank zero, cf. Observation 3.6. (cid:3) Remark 5.1. The real rank zero counter-examples for the space X3 have torsion in both even and odd degrees. In [ABK], it is shown that for real rank zero C∗-al- gebras over X3 with free K1-groups, isomorphisms on a reduced filtrated K-theory lift to KK(X3)-equivalences. This reduced filtrated K-theory is defined in [ABK] by disregarding parts of the information in filtrated K-theory, and it is equivalent to the reduced filtered K-theory defined by the second named author in [Res06, 4.1]. It / / % % % / / 9 % 9 % / / 9 % / / / 9 9 9 / / / / / / / / / } / } / } a a o o a o o _ o o o o / / ! ! ! / / = ! = ! / / = ! / / / = = = 16 SARA ARKLINT, GUNNAR RESTORFF, AND EFREN RUIZ is unknown whether isomorphisms on FK lift to KK(X3)-equivalences under these conditions. Remark 5.2. The space X6 has been studied in [Ben10] where R. Bentmann fails to construct a finite refinement of filtrated K-theory over X6 that admits a UCT and remarks that it seems unlikely that such a finite refinement exists. So our method cannot be applied for the space X6. In [Ben10], R. Bentmann constructs separable, stable, nuclear, purely infinite, tight C∗-algebras A and B over X6 that have isomorphic filtrated K-theory and are not KK(X6)-equivalent. One can check that the boundary map FK1(A) → FK3(A) does not vanish in either degrees, so neither A and B nor the suspensions S A and S B have real rank zero. So there is so far no known real rank zero counter-example for X6. 6. Acknowledgement This research was supported by the NordForsk Research Network "Operator Al- gebras and Dynamics" (grant #11580), by the Faroese Research Council, and by the Danish National Research Foundation (DNRF) through the Centre for Symmetry and Deformation. The second and third named authors are grateful to Professor Søren Eilers and the department of mathematics at the University of Copenhagen for providing the dynamic research environment where this work was initiated during the spring of 2010. The first named author would like to thank Professor Ralf Meyer and the department of mathematics at Georg-August-Universität Göttingen for kind hospitality during the spring and early summer of 2010. References [ABK] Sara Arklint, Rasmus Bentmann, and Takeshi Katsura. Reductions of filtered K-theory and a characterization of Cuntz-Krieger algebras. Preprint. [Ben10] Rasmus Bentmann. Filtrated K-theory and classification of C ∗-algebras. Master's thesis, [BK] http://math.ku.dk/∼bentmann/thesis.pdf, 2010. Rasmus Bentmann and Manuel Köhler. Universal coefficient theorems for C ∗-algebras over finite topological spaces. arXiv:1101.5702v3 [math.OA]. [BP91] Lawrence G. Brown and Gert K. Pedersen. C ∗-algebras of real rank zero. J. Funct. Anal., 99(1):131 -- 149, 1991. [Kir00] Eberhard Kirchberg. Das nicht-kommutative Michael-Auswahlprinzip und die Klassifika- tion nicht-einfacher Algebren. In C ∗-algebras (Münster, 1999), pages 92 -- 141. Springer, Berlin, 2000. [ML98] Saunders Mac Lane. Categories for the working mathematician, volume 5 of Graduate [MN] Texts in Mathematics. Springer-Verlag, New York, second edition, 1998. Ralf Meyer and Ryszard Nest. C ∗-algebras over topological spaces: Filtrated K-theory. arXiv:0810.0096v2 [math.OA]. [MN09] Ralf Meyer and Ryszard Nest. C ∗-algebras over topological spaces: the bootstrap class. Münster J. Math., 2:215 -- 252, 2009. [Res06] Gunnar Restorff. Classification of Cuntz-Krieger algebras up to stable isomorphism. J. Reine Angew. Math., 598:185 -- 210, 2006. [RR07] Gunnar Restorff and Efren Ruiz. On Rørdam's classification of certain C ∗-algebras with one non-trivial ideal. II. Math. Scand., 101(2):280 -- 292, 2007. FILTRATED K-THEORY FOR REAL RANK ZERO C ∗-ALGEBRAS 17 Department of Mathematical Sciences, University of Copenhagen, Universitets- parken 5, DK-2100 Copenhagen, Denmark E-mail address: [email protected] Faculty of Science and Technology, University of Faroe Islands, Nóatún 3, FO- 100 Tórshavn, Faroe Islands E-mail address: [email protected] Department of Mathematics, University of Hawaii, Hilo, 200 W. Kawili St., Hilo, Hawaii, 96720-4091 USA E-mail address: [email protected]
1111.0381	1	1111	2011-11-02T04:31:13	Stacey crossed products associated to Exel systems	[ "math.OA" ]	There are many different crossed products by an endomorphism of a C-algebra, and constructions by Exel and Stacey have proved particularly useful. Here we show that every Exel crossed product is isomorphic to a Stacey crossed product, though by a different endomorphism of a different C-algebra. We apply this result to a variety of Exel systems, including those associated to shifts on the path spaces of directed graphs.	math.OA	math	STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS ASTRID AN HUEF AND IAIN RAEBURN Abstract. There are many different crossed products by an endomorphism of a C ∗- algebra, and constructions by Exel and Stacey have proved particularly useful. Here we show that every Exel crossed product is isomorphic to a Stacey crossed product, though by a different endomorphism of a different C ∗-algebra. We apply this result to a variety of Exel systems, including those associated to shifts on the path spaces of directed graphs. 1. Introduction Everybody agrees that when α is an automorphism of a unital C ∗-algebra A, the crossed product A ⋊α Z is generated by a unitary element u and a representation π : A → A ⋊α Z which satisfy the covariance relation (1.1) π(α(a)) = uπ(a)u∗ for a ∈ A. The covariance relation can be reformulated as π(α(a))u = uπ(a) or u∗π(α(a))u = π(a), and, when α is an automorphism, these reformulations are equivalent to (1.1). When α is an endomorphism, though, these reformulations are no longer equivalent, and give different crossed products. Thus there are several crossed products based on covariance relations in which u is an isometry [31, 38, 12], and still more crossed products in which u is a partial isometry [29, 3]. The crossed products of Stacey [38] and Exel [12] have proved to be particularly useful, and have been extensively studied, for example in [33, 5, 32] and [15, 8, 14]. Stacey's crossed product B ×α N is generated by an isometry s and a representation π of A satisfying π(α(a)) = sπ(a)s∗. His motivating example was the endomorphism α of the UHF core A in the Cuntz algebra On described by Cuntz in [11], for which we recover On as A ×α N, and is especially useful for corner endomorphisms which map A onto the corner α(1)Aα(1) (see also [33, 5]). Stacey's construction has been extended to semigroups of endomorphisms, and these semigroup crossed products were used to study Toeplitz algebras [2, 24]; they have since been used extensively in the analysis of C ∗-algebras arising in number theory (see [25, 6, 22, 28, 23], for example). Exel's construction depends on the choice of a transfer operator L : A → A for α, which is a positive linear map satisfying L(α(a)b) = aL(b). He uses L to build a Hilbert bimodule ML over A, and then his crossed product A ⋊α,L N is closely related Date: 2 November 2011. 2000 Mathematics Subject Classification. 46L55. Key words and phrases. C ∗-algebra, endomorphism, transfer operator, Exel crossed product, Stacey crossed product, graph algebra, gauge action. Some of the results in this paper, including most of those in Sections 4 and 5, were contained in the preliminary version [19], which we posted on the arXiv in July 2011. 1 2 AN HUEF AND RAEBURN to the Cuntz-Pimsner algebra O(ML) of this bimodule (the precise relationship is de- scribed in [8]). The motivating example for Exel's construction is the endomorphism of C({1, · · · , n}∞) induced by the backward shift, for which averaging over the n preim- ∼= C({1, · · · , n}∞) ⋊α,L N. ages of each point gives a transfer operator L such that On More generally, Exel realised each Cuntz-Krieger algebra as a crossed product by the corresponding subshift of finite type, thereby giving a very direct proof that the Cuntz- Krieger algebra is determined up to isomorphism by the subshift. Exel's construction has attracted a good deal of attention in connection with irreversible dynamics [15, 14, 10], and has also been extended to semigroups of endomorphisms with interesting conse- quences [27, 7]. Here we study a large family of Exel systems (A, α, L), and describe a general proce- dure which builds a corner endomorphism β of another C ∗-algebra B such that Exel's A ⋊α,L N is naturally isomorphic to Stacey's B ×β N. In our motivating example (C({1, · · · , n}∞), α, L), we recover Cuntz's description of On as the Stacey crossed prod- uct by an endomorphism of the UHF core (see Remark 5.4). Our investigations of this construction have led us to some interesting new examples of Exel and Stacey crossed products associated with graph algebras and Cuntz algebras. We begin with short sections in which we describe the two kinds of crossed products of interest to us, and a family of relative Cuntz-Pimsner algebras associated to Exel sys- tems. Then in §4, we give the details of our construction of corner endomorphisms, first in the generality of relative Cuntz-Pimsner algebras, then specialising to Exel crossed products and Cuntz-Pimsner algebras. Then we derive a six-term exact sequence in K-theory from a result of Paschke [34] about Stacey crossed products, and compare our construction to a previous one of Exel [13]. In §5, we consider a directed graph E and the Exel system (C ∗(E∞), α, L) introduced in [9], and obtain a new description of the graph algebra C ∗(E) as a Stacey crossed product of the core (which generalises results of Rørdam and Kwa´sniewski for finite graphs [37, 21]). This led us to revisit Exel systems involving other endomorphisms of the UHF core in Cuntz algebras, which we do in §6. In §5 we needed a detailed analysis of the core in the C ∗-algebra C ∗(E) of a column-finite graph, and we describe this analysis in an appendix. 2. Exel systems An endomorphism α of a C ∗-algebra A is extendible if it extends to a strictly con- tinuous endomorphism α of M(A). Nondegenerate endomorphisms, for example, are automatically extendible with α(1) = 1. In this paper we are interested in Exel systems (A, α, L) of the kind studied in [9], which means that α is an extendible endomorphism of a C ∗-algebra A and L : A → A is a positive linear map which extends to a positive linear map L : M(A) → M(A) such that (2.1) L(α(a)m) = aL(m) for a ∈ A and m ∈ M(A). Equation (2.1) implies that L is strictly continuous, and we then assume further that L(1M (A)) = 1M (A) (but not that α is unital). Following Exel [12], we say that L is a transfer operator for (A, α). STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 3 We write ML for the Hilbert bimodule over A used in [9], where the construction of [12, 8] was extended to non-unital A. Briefly, A is given a bimodule structure by a · m = am and m · b = mα(b), and the pairing hm , ni = L(m∗n) defines a pre-inner product on A. Modding out by the elements of norm 0 and completing gives a right- Hilbert A -- A bimodule ML (or correspondence over A). We denote by q : A → ML the canonical map of A onto a dense sub-bimodule of ML, and by φ the homomorphism of A into L(ML) implementing the left action. We are interested in several C ∗-algebras associated to Exel systems, and more specif- ically to the bimodule ML. 2.1. Relative Cuntz-Pimsner algebras. Following [18], a representation (ψ, π) of ML in a C ∗-algebra B consists of a linear map ψ : ML → B and a homomorphism π : A → B such that ψ(m · a) = ψ(m)π(a), ψ(m)∗ψ(n) = π(hm , ni), and ψ(a · m) = π(a)ψ(m) T (ML) = span(cid:8)j⊗i for a ∈ A and m, n ∈ ML. By [18, Proposition 1.8], a representation (ψ, π) of ML gives a representation (ψ⊗i, π) of M ⊗i L := ML ⊗A · · · ⊗A ML such that ψ⊗i(m1 ⊗A · · · ⊗A mi) = ψ(m1) · · · ψ(mi). The Toeplitz algebra T (ML) is the C ∗-algebra generated by a universal representation (jM , jA) of ML, and [18, Lemma 2.4] says that M (n)∗ : i, j ∈ N, m ∈ M ⊗i L , n ∈ M ⊗j M (m)j⊗j There is a strongly continuous gauge action γ : T → Aut T (ML) such that γz(jA(a)) = jA(a) and γz(jM (m)) = zjM (m) [18, Proposition 1.3]. A representation (ψ, π) of ML in B gives a homomorphism (ψ, π)(1) : K(ML) → B such that (ψ, π)(1)(Θm,n) = ψ(m)ψ(n)∗ for every rank-one operator Θm,n : p 7→ m · hn , pi. Following [17, §1], if J is an ideal of A contained in φ−1(K(ML)), then we view the relative Cuntz-Pimsner algebra O(J, ML) of [30] as the quotient of T (ML) by the ideal generated by L (cid:9). (cid:8)jA(a) − (jM , jA)(1)(φ(a)) : a ∈ J(cid:9). We write Q or QJ for the quotient map. Then (kM , kA) := (QJ ◦ jM , QJ ◦ jA) is universal for representations (ψ, π) which are coisometric on J (that is, satisfy πJ = (ψ, π)(1) ◦ φJ ). If J = {0} then O(J, ML) is just T (ML); if J = φ−1(K(ML)) then O(J, ML) is the Cuntz-Pimsner algebra O(ML) of [35], and representations that are coisometric on φ−1(K(ML)) are called Cuntz-Pimsner covariant. It follows from [18, Lemma 2.4] that each O(J, ML) carries a gauge action γ of T such that QJ is equivariant, and the fixed-point algebra or core is O(J, ML)γ = span(cid:8)k⊗i M (m)k⊗i M (n)∗ : m, n ∈ M ⊗i L , i ∈ N(cid:9). 2.2. Exel's crossed product. As in [9], a Toeplitz-covariant representation of (A, α, L) in a C ∗-algebra B consists of a nondegenerate homomorphism π : A → B and an element S ∈ M(B) such that Sπ(a) = π(α(a))S and S∗π(a)S = π(L(a)). The Toeplitz crossed product T (A, α, L) is generated by a universal Toeplitz-covariant representation (i, s) (or, more correctly, by i(A)∪i(A)s). By [9, Proposition 3.1], there is 4 AN HUEF AND RAEBURN a map ψs : ML → T (A, α, L) such that ψs(q(a)) = i(a)s, and (ψs, i) is a representation of ML. Set Kα := Aα(A)A ∩ φ−1(K(ML)). In [9, §4], the Exel crossed product A ⋊α,L N of a possibly non-unital C ∗-algebra A by N is the quotient of T (A, α, L) by the ideal generated by (cid:8)i(a) − (ψs, i)(1)(φ(a)) : a ∈ Kα(cid:9). We write Q for the quotient map of T (A, α, L) onto A ⋊α,L N, and (j, t) := (Q ◦ i, Q(s)). There is a dual action α of T on A ⋊α,L N such that αz(j(a)) = j(a) and αz(t) = zt. Theorem 4.1 of [9] says that there is an isomorphism θ of O(Kα, ML) onto A ⋊α,L N such that θ ◦ kA = Q ◦ i and θ ◦ kM = Q ◦ ψs. 3. Stacey crossed products Suppose that α is an endomorphism of a C ∗-algebra A. A Stacey-covariant representa- tion of (A, α) in a C ∗-algebra B consists of a nondegenerate homomorphism π : A → B and an isometry V ∈ M(B) such that π(α(a)) = V π(a)V ∗. Stacey showed in [38, §3] that there is a crossed product A ×α N which is generated by a universal Stacey- covariant representation (iA, v). If (π, V ) is a Stacey-covariant representation of (A, α) in B, then we write π × V for the nondegenerate homomorphism of A ×α N into B such that (π × V ) ◦ iA = π and (π × V )(v) = V . (Stacey called A ×α N "the multiplicity-one crossed product" of (A, α).) The crossed product A ×α N carries a dual action α of T, which is characterised by αz(iA(a)) = iA(a) and αz(v) = zv. The following "dual-invariant uniqueness the- orem" says that this dual action identifies A ×α N among C ∗-algebras generated by Stacey-covariant representations of (A, α). It was basically proved in [5, Proposition 2.1] (modulo the correction made in [2]). Proposition 3.1. Suppose that α is an endomorphism of a C ∗-algebra A, and (π, V ) is a Stacey-covariant representation of (A, α) in a C ∗-algebra D. If π is faithful and there is a strongly continuous action γ : T → Aut D such that γz(π(a)) = π(a) and γz(V ) = zV , then π × V is faithful on A ×α N. Proof. The conditions on γ say that (π × V )( αz(b)) = γz((π × V )(b)) for all b ∈ A ×α N. Thus (cid:13)(cid:13)(cid:13)(π × V )(cid:16)ZT αz(b) dz(cid:17)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)ZT ≤ZT (π × V )( αz(b)) dz(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)ZT kγz((π × V )(b))k dz =ZT = k(π × V )(b)k, γz((π × V )(b)) dz(cid:13)(cid:13)(cid:13) k(π × V )(b)k dz We now take (B, β) = (A ×α N, α), and apply [5, Lemma 2.2] to (B, β). We have just verified the hypothesis (2) of [5, Lemma 2.2]. The other hypothesis (1) asks for π × V to be faithful on the fixed-point algebra Bβ = (A ×α N) α. However, the proof of [2, Lemma 1.5] uses neither that A is unital nor the estimate (ii) in [2, Theorem 1.2], and hence we can deduce from that proof that π × V is faithful on Bβ. Thus [5, Lemma 2.2] applies, and the result follows. (cid:3) STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 5 4. Exel crossed products as Stacey crossed products The goal of this section is the following theorem and some corollaries. Theorem 4.1. Suppose that (A, α, L) is an Exel system, and J is an ideal of A contained in φ−1(K(ML)). There is a unique isometry V ∈ M(O(J, ML)) such that kM (q(a)) = kA(a)V for a ∈ A, and Ad V restricts to an endomorphism α′ of the core CJ := O(J, ML)γ such that (4.1) M (a · m)k⊗i (q(α(a)) ⊗A m)k⊗(i+1) M (q(α(b)) ⊗A n)∗ α′(cid:0)k⊗i M (b · n)∗(cid:1) = k⊗(i+1) M for a, b ∈ A and m, n ∈ M ⊗i L . Further, α′ is extendible with α′(1) = V V ∗, is injective and has range α′(1)CJα′(1). Finally, (id, V ) is a Stacey-covariant representation of (CJ , α′) such that id ×V is an isomorphism of the Stacey crossed product CJ ×α′ N onto O(J, ML). Proof. Since we know from [9, Corollary 3.5] that kA : A → O(J, ML) is nondegenerate1, there is at most one multiplier V satisfying kM (q(a)) = kA(a)V , and we have uniqueness. When the C ∗-algebra A has an identity, we can deduce from the results in [8, §3] that V := kM (q(1)) has the required properties. When A does not have an identity, we take an approximate identity {eλ} for A, and claim, following Fowler [16, §3], that {kM (q(eλ))} converges strictly in M(O(J, ML)) to a multiplier V . Indeed, for a, b ∈ A, m ∈ M ⊗i L and n ∈ M ⊗j kM (q(eλ))k⊗i L we have M (a · m)k⊗j (4.2) M (b · n)∗ = kM (q(eλ))kA(a)k⊗i = kM (q(eλα(a)))k⊗i → kM (q(α(a)))k⊗i M (m)k⊗j M (m)k⊗j M (m)k⊗j M (b · n)∗ M (b · n)∗ M (b · n)∗, and similarly (4.3) M (a · m)k⊗j k⊗i M (b · n)∗kM (q(eλ)) → k⊗i M (a · m)k⊗j M (n)∗kM (q(b∗)). Since L is positive and L(1) = 1, we have kLk ≤ 1, and kq(eλ)k ≤ keλk ≤ 1 for all λ. Thus, since the elements k⊗i M (b · n)∗ span a dense subspace of O(J, ML), an ǫ/3 argument using (4.2) and (4.3) shows that {kM (q(eλ))b} and {bkM (q(eλ))} converge in O(J, ML) for every b ∈ O(J, ML). Thus {kM (q(eλ))} is strictly Cauchy, and since M(O(J, ML) is strictly complete we deduce that {kM (q(eλ))} converges strictly to a multiplier V ; (4.2) implies that V satisfies M (a · m)k⊗j (4.4) V k⊗i M (a · m)k⊗j M (b · n)∗ = kM (q(α(a)))k⊗i M (m)k⊗j M (b · n)∗. To see that V is an isometry, we observe that kM (q(eλ))∗kM (q(eλ)) = kA(hq(eλ), q(eλ)i) = kA(L(e2 λ)); since L is strictly continuous, L(e2 λ) converges strictly to L(1M (A)) = 1M (A). Since kA : A → O(J, ML) is nondegenerate, kM (q(eλ))∗kM (q(eλ)) converges strictly to 1 = 1We caution that this nondegeneracy is not at all obvious, and even slightly surprising, because the representation π in a Toeplitz representation (ψ, π) is not required to be nondegenerate. 6 AN HUEF AND RAEBURN 1M (O(J,ML)), and since the multiplication in a multiplier algebra is jointly strictly con- tinuous on bounded sets (by another ǫ/3 argument), we deduce that V ∗V = 1. Thus V is an isometry. Next, we let a ∈ A and compute (4.5) kM (q(a)) = lim λ kM (q(aeλ)) = lim λ kM (a · q(eλ)) = lim λ kA(a)kM (q(eλ)) = kA(a)V, so V has the required properties. Conjugating by the isometry V ∈ M(O(J, ML)) gives an endomorphism Ad V : T 7→ V T V ∗, and two applications of (4.4) show that (4.6) M (a · m)k⊗i Ad V(cid:0)k⊗i M (b · n)∗(cid:1) = kM (q(α(a)))k⊗i = k⊗(i+1) M M (m)k⊗i M (n)∗kM (q(α(b)))∗ (q(α(a)) ⊗A m)k⊗(i+1) M (q(α(b)) ⊗A n)∗. The formula (4.6) implies that Ad V maps the core CJ into itself, and that the restriction α′ := (Ad V )CJ satisfies (4.1). The pair (id, V ) is then by definition Stacey covariant for α′ ∈ End CJ , and we can apply the dual-invariant uniqueness theorem (Proposition 3.1) to the gauge action γ on O(J, ML), finding that id ×V is a faithful representation of CJ ×α′ N in O(J, ML). The identity kM (q(a)) = kA(a)V implies that kM (ML) = kA(A)V , and since O(J, ML) is generated by kA(A) ⊂ CJ and kM (ML), the range of id ×V contains the generating set kA(A) ∪ kA(A)V , and hence is all of O(J, ML). It remains for us to prove the assertions about α′. It is injective because Ad V is. Since kA is nondegenerate the image {kA(eλ)} of an approximate identity {eλ} converges strictly to 1 in M(O(J, ML)). Hence {α′(kA(eλ))} = {V kA(eλ)V ∗} converges strictly to the projection V V ∗. Thus α′ is extendible with α′(1) = V V ∗ by, for example, [1, Proposition 3.1.1]. The range of α′ is certainly contained in the corner α′(1)CJα′(1). To see the reverse inclusion, fix T ∈ α′(1)CJ α′(1). Then T = V V ∗SV V ∗ = Ad V (V ∗SV ) for some S ∈ CJ . Since CJ is α′-invariant, to see that T is in the range of α′ = Ad V CJ it suffices to see that V ∗SV is in CJ, and, by continuity of Ad V ∗, it suffices to see this for S of the form M ((a · m1) ⊗A m′)k⊗i where a, b ∈ A, m1, n1 ∈ ML and m′, n′ ∈ M ⊗(i−1) S = k⊗i L M ((b · n1) ⊗A n′)∗ . For i ≥ 1 the calculation V ∗k⊗i M ((a · m1) ⊗A m′) = (kA(a∗)V )∗kM (m1)k⊗(i−1) = kM (q(a∗))∗kM (m1)k⊗(i−1) M = kA(hq(a∗), m1i)k⊗(i−1) (m′) = k⊗(i−1) (hq(a∗), m1i · m′) M M M (m′) (m′) gives V ∗SV = k⊗(i−1) M (hq(a∗), m1i · m′)k⊗(i−1) M (hq(b∗), n1i · n′)∗ ∈ CJ . For i = 0 we have V ∗SV = V ∗kA(a)kA(b)∗V = kM (q(a∗))∗kM (q(b∗)) = kA(hq(a∗) , q(b∗)i) ∈ CJ. Thus T = Ad V (V ∗SV ) = α′(V ∗SV ), and α′ has range α′(1)CJα′(1). (cid:3) STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 7 4.1. Exel crossed products. We will now use the isomorphism θ of O(Kα, ML) onto A ⋊α,L N to transfer the conclusions of Theorem 4.1 over to A ⋊α,L N. If {eλ} is an approximate identity for A, then θ ◦ kM (q(eλ)) = Q(ψs(q(eλ))) = Q(i(eλ)s) = Q(i(eλ))Q(s) converges by nondegeneracy of i to Q(s), and hence θ carries the isometry V of The- orem 4.1 into t := Q(s). The isomorphism θ is equivariant for the gauge action γ on O(Kα, ML) and the dual action α on A ⋊α,L N, and hence maps the core CKα = O(Kα, ML)γ onto (A ⋊α,L N) α. Next, we want a workable description of (A ⋊α,L N) α. For m = q(a1) ⊗A · · · ⊗A q(ai) ∈ L we have M ⊗i (4.7) k⊗i M (m) = kM (q(a1)) · · · kM (q(ai)) = kA(a1)V · · · kA(ai)V and hence θ takes k⊗i = kA(a1)kA(α(a2))V 2kA(a3)V · · · kA(ai)V M (m) into an element of the form j(a)ti. Now = kA(cid:0)a1α(a2)α2(a3) · · · αi−1(ai)(cid:1)V i, span(cid:8)j(a)tit∗kj(b) : a, b ∈ A, i, k ∈ N(cid:9) is a ∗-subalgebra of A ⋊α,L N containing the generating set j(A) ∪ j(A)t, and hence is dense in A ⋊α,L N. The expectation onto (A ⋊α,L N) α is continuous and kills terms with i 6= k, so (A ⋊α,L N) α = span(cid:8)j(a)tit∗ij(b) : a, b ∈ A, i ∈ N(cid:9). The next corollary says that every Exel crossed product is a Stacey crossed product. (4.8) Corollary 4.2. Suppose that (A, α, L) is an Exel system. Then there is an injective endomorphism β of (A ⋊α,L N) α such that (4.9) β(j(a)tit∗ij(b)) = j(α(a))ti+1t∗(i+1)j(α(b)). The endomorphism β is extendible with β(1) = tt∗ and has range tt∗(A ⋊α,L N) αtt∗. The pair (id, t) is a Stacey-covariant representation of (cid:0)(A ⋊α,L N) α, β(cid:1), and id ×t is an isomorphism of the Stacey crossed product (A ⋊α,L N) α ×β N onto the Exel crossed product A ⋊α,L N. Proof. Applying the isomorphism θ : O(Kα, ML) → A ⋊α,L N to the conclusion of Theorem 4.1 gives an endomorphism β := θ ◦ α′ ◦ θ−1 of (A ⋊α,L N) α and an isomorphism id ×t of (A ⋊α,L N) α ×β N onto A ⋊α,L N. It remains for us to check the formula for β. Let m = q(a1) ⊗A · · · ⊗A q(ai). Then the calculation (4.7) shows that θ carries k⊗i M (c · m) into j(ca1α(a2) · · · αi−1(ai))ti, and k⊗(i+1) (q(α(c)) ⊗A m) into M j(cid:0)α(c)α(a1)α2(a2) · · · αi(ai)(cid:1)ti+1 = j(cid:0)α(ca1α(a2) · · · αi−1(ai))(cid:1)ti+1, so (4.9) follows from (4.1). (cid:3) Remark 4.3. Since we have identified how the action α′ on the core of O(Kα, ML) pulls over to the Exel crossed product A ⋊α,L N, we will from now on freely identify O(Kα, ML) and A ⋊α,L N, and drop the isomorphism θ from our notation. 8 AN HUEF AND RAEBURN Example 4.4. We now discuss a family of Exel systems studied in [14] and [26]. Let d ∈ N and fix B ∈ Md(Z) with nonzero determinant N. (This matrix B plays the same role as the matrix B in [14, 26]; because A is already heavily subscribed in this paper, we write Bt in place of the matrix A used there.) We choose a set Σ of coset representatives for Zd/BZd. The map σBt : Td → Td characterised by σBt(e2πix) = e2πiBtx is an N-sheeted covering : f 7→ f ◦ σBt of C(Td), and L(f )(z) := map, and induces an endomorphism αBt N −1PσBt (w)=z f (w) defines a transfer operator L for αBt. Proposition 3.3 of [26] says that the Exel crossed product C(Td) ⋊αBt ,L N is the universal C ∗-algebra generated by a unitary representation u of Zd and an isometry v satisfying (E1) vum = uBmv for m ∈ Zd, (E2) v∗umv =(uB−1m if m ∈ BZd (E3) 1 =Pm∈Σ(umv)(umv)∗; 0 we then have otherwise, and C(Td) ⋊αBt ,L N = span{umvkv∗lu∗ n : k, l ∈ N and m, n ∈ Zd}. When we view C(Td) ⋊αBt ,L N as a Cuntz-Pimsner algebra (O(ML), jML, jC(Td)) as in Remark 4.3, v = jML(q(1)) = t and u is the representation m 7→ jC(Td)(χm), where χm(z) := zm. Since αBt(χm) = χBm, Corollary 4.2 gives an endomorphism β of (cid:0)C(Td) ⋊αBt ,L N(cid:1) αBt = span{umviv∗iu∗ n : i ∈ N and m, n ∈ Zd} β(umviv∗iu∗ n) = uBmvi+1v∗(i+1)u∗ Bn. satisfying (4.10) Now that we have the formula for β, we can prove directly that there is such an endomorphism. To see this, recall from [26, Proposition 5.5(b)] that for each i and Σi := {µ1 + Bµ2 + · · · Bi−1µi : µ ∈ Σi}, {umviv∗iu∗ n : m, n ∈ Σi} is a set of nonzero matrix units. Thus there is a homomorphism ζi : MΣi(C) → MΣi+1(C) such that ζi(umviv∗iu∗ Bn. Let δ denote the universal representation of Bi+1Zd in C ∗(Bi+1Zd). Then the unitary representation Bim 7→ δBi+1m induces a homomorphism ηi : C ∗(BiZd) → C ∗(Bi+1Zd). When we identify Ci := span{umviv∗iu∗ n} with MΣi(C) ⊗ C ∗(BiZd) as in [26, Proposition 5.5(c)], we get homomorphisms n) = uBmvi+1v∗(i+1)u∗ βi := ζi ⊗ ηi : Ci = MΣi(C) ⊗ C ∗(BiZd) → Ci+1 = MΣi+1(C) ⊗ C ∗(Bi+1Zd) satisfying (4.10). The Cuntz relation (E3) implies that βi+1Ci = βi, and hence the βi i=1 Ci satisfying (4.10); since the homomorphisms βi are norm-decreasing, β extends to an endomorphism of O(ML)γ = combine to give a homomorphism β : S∞ S∞ i=1 Ci. i=1 Ci → S∞ STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 9 4.2. Cuntz-Pimsner algebras. Let (A, α, L) be an Exel system, and let I be an ideal in A. Following [8, Definition 4.1], we say that L is almost faithful on I if a ∈ I and L(b∗a∗ab) = 0 for all b ∈ A imply a = 0. Since (4.11) L(b∗a∗ab) = hq(ab) , q(ab)i = hφ(a)(q(b)) , φ(a)(q(b))i, L is almost faithful on I if and only if φI : I → L(ML) is injective. Almost faithful transfer operators were introduced in [8] as a necessary and sufficient condition for the the canonical map kA : A → O(ML) to be injective (see also [30, Proposition 2.1] and [9, Theorem 4.3]). In this subsection, we suppose that L is almost faithful on φ−1(K(ML)). Then, since kA : A → O(ML) is injective, we can use [17, Corollary 4.9] to realise the core O(ML)γ as a direct limit, and obtain a different description of the Stacey system (O(ML)γ, α′). L ) ⊗A 1i−j for the We denote the identity operator on M ⊗i L by 1i, and write K(M ⊗j image of K(M ⊗j L ) under the map T 7→ T ⊗A 1i−j. Then, following [17, §4], we define Ci = (A ⊗A 1i) + (K(ML) ⊗A 1i−1) + · · · + K(M ⊗i L ), which is a C ∗-subalgebra of L(M ⊗i L ). We define φi : Ci → Ci+1 by φi(T ) = T ⊗A 11, and define (C∞, ιi) := lim (Ci, φi). Since kA is injective, we can now apply [17, Corollary 4.9] −→ to the Cuntz-Pimsner covariant representation (kM , kA), and deduce that there is an isomorphism κ of C∞ onto the core O(ML)γ such that (4.12) (the notation in [17] suppresses the maps ιi ). κ(ιi(T ⊗A 1i−j)) = (k⊗j M , kA)(1)(T ) for T ∈ K(M ⊗j L ) and i ≥ j To describe the endomorphism β := κ−1 ◦ α′ ◦ κ of C∞, we need some notation. Lemma 4.5. The map U : A → ML defined by U(a) = q(α(a)) is an adjointable isometry such that U ∗(q(a)) = L(a) for a ∈ A. Proof. The calculation hU(a) , U(b)i = hq(α(a)) , q(α(b))i = L(α(a)∗α(b)) = L(α(a∗b)) = a∗b = ha , bi shows that U is inner-product preserving. We next note that (4.13) hU(a) , q(b)i = L(α(a)∗b) = a∗L(b) = ha , L(b)i. Equation (4.13) implies that ka∗L(b)k = khU(a) , q(b)ik ≤ kU(a)k kq(b)k = kak kq(b)k; thus kL(b)k ≤ kq(b)k, and there is a well-defined bounded linear map T : ML → A such that T (q(b)) = L(b). Now (4.13) shows that U is adjointable with adjoint T . (cid:3) Corollary 4.6. Define maps Ui : M ⊗i and taking L → M ⊗(i+1) L by identifying M ⊗i L with A ⊗A M ⊗i L Ui := U ⊗A 1i : M ⊗i = A ⊗A M ⊗i Then each Ui is an adjointable isometry, and L → ML ⊗A M ⊗i L = M ⊗(i+1) L . 10 AN HUEF AND RAEBURN (a) Ui(a · m) = q(α(a)) ⊗A m; (b) U ∗ i (q(a) ⊗A m) = L(a) · m; (c) Ui+1 = Ui ⊗A 1 and U ∗ i+1 = U ∗ i ⊗A 1. With the notation of Corollary 4.6, we can now describe the endomorphism on C∞. We reconcile our results with Exel's [13, Theorem 6.5] in §4.3 below, and show there that Corollary 4.7 reduces to [13, Theorem 6.5] when A is unital. Corollary 4.7. Suppose that (A, α, L) is an Exel system such that L is almost faithful on φ−1(K(ML)). There is an endomorphism β of C∞ := lim −→ i ) for T ∈ Ci, β(ιi(T )) = ιi+1(UiT U ∗ (Ci, φi) such that (4.14) and β is extendible and injective with range β(1)C∞β(1). Let κ be the isomorphism of C∞ onto O(ML)γ satisfying (4.12), and let V be the isometry in M(O(ML)) such that kM (q(a)) = kA(a)V for a ∈ A (as given by Theorem 4.1). Then (κ, V ) is a Stacey- covariant representation of (C∞, β) in O(ML), and κ × V is an isomorphism of the Stacey crossed product C∞ ×β N onto O(ML). If A acts by compact operators on ML, then β(1) is a full projection. Proof. We define β := κ−1 ◦ α′ ◦ κ, where α′ is the endomorphism from Theorem 4.1. Let a · m, b · n ∈ M ⊗i L . Then κ ◦ ιi(Θa·m,b·n) = k⊗i M (b · n)∗, and M (a · m)k⊗i κ ◦ β ◦ ιi(Θa·m,b·n) = α′ ◦ κ ◦ ιi(Θa·m,b·n) (q(α(b)) ⊗A n)∗ M M = k⊗(i+1) (q(α(a)) ⊗A m)k⊗(i+1) = κ ◦ ιi+1(Θq(α(a))⊗Am,q(α(b))⊗An) = κ ◦ ιi+1(ΘUi(a·m),Ui(b·n)) = κ ◦ ιi+1(UiΘa·m,b·nU ∗ i ). (using (4.1)) This gives (4.14) for T ∈ K(M ⊗i L ). For j < i and S ∈ K(M ⊗j L ), we have β(ιi(S ⊗A 1i−j)) = β(ιj(S)) = ιj+1(UjSU ∗ j ) = ιi+1((Uj ⊗A 1i−j)(S ⊗A 1i−j)(U ∗ = ιi+1(Ui(S ⊗A 1i−j)U ∗ i ), j ⊗A 1i−j)) and adding over j gives (4.14). isomorphism. It now follows from Theorem 4.1 that κ × V is an Now suppose that φ : A → L(ML) has range in K(ML). Then A ⊗A 11 = φ(a) belongs L )). Now for i ≥ 1 and m = m′ ⊗A q(a) and i=1 ιi(K(M ⊗i to K(ML), and hence C∞ =S∞ n = n′ ⊗A q(b) in M ⊗i L , we have M κ(ι1(Θm,n)) = k⊗(i−1) = k⊗(i−1) = k⊗(i−1) M M (m′)kM (q(a))kM (q(b))∗k⊗(i−1) M (m′)kA(a)V V ∗kA(b)∗k⊗(i−1) (m′)kA(a)κ(β(1))kA(b)∗k⊗(i−1) M M (m′)∗ (m′)∗ (m′)∗. Thus each ιi(K(M ⊗i L )) is contained in the ideal generated by β(1), and β(1) is full. (cid:3) STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 11 The following exact sequence for the K-theory of O(ML) looks a little different from the usual ones (in [35] or [20], for example), because it involves the core rather than the coefficient algebra. Corollary 4.8. Let (A, α, L) be an Exel system such that A is unital, L is almost faithful on A, and A acts by compact operators on ML. Let (C∞, β) be the Stacey system constructed in Corollary 4.7. Then there exists an exact sequence (4.15) K0(C∞) β∗−id / K0(C∞) κ∗ / / K0(O(ML)) K1(O(ML)) κ∗ K1(C∞) β∗−id K1(C∞). Proof. Since A acts by compact operators on ML and A is unital, the identity operator on ML is compact, and C∞ is unital. The range of β is a full corner in C∞ by Corollary 4.7. Hence Theorem 4.1 of [34] gives (4.15). (cid:3) 4.3. Connections with a construction of Exel. In [13, Theorem 6.5], Exel shows that if A is unital, α is injective and unital, and there is a faithful conditional expectation E of A onto α(A), then his crossed product A⋊α,α−1◦E N is isomorphic to a Stacey crossed product A ×β ′ N. The C ∗-algebra A is by definition a subalgebra of the C ∗-algebraic direct limit of a sequence of algebras of the form L(Mi) for certain Hilbert modules Mi. Exel's hypotheses on α imply that L := α−1 ◦ E is a faithful transfer operator for α satisfying L(1) = 1, and hence that φ : A → L(ML) is injective. Thus the Exel crossed product A ⋊α,L N is the Cuntz-Pimsner algebra O(ML), and Corollary 4.7 gives an isomorphism of C∞ ×β N onto A ⋊α,L N. It is natural to ask whether Exel's system ( A, β′) is the same as the system (C∞, β) appearing in Corollary 4.7. : L(Mi) → L(Mi+1) characterised by φi(T ) ◦ ji = ji ◦ T . Exel's module Mi is the Hilbert module over αi(A) associated to the expectation αi ◦ Li of A onto αi(A) (which he denotes by Ei), and hence is a completion of a copy qi(A) of A. By restricting the action we can view Mi as a module over αi+1(A), and this induces a linear map ji : Mi → Mi+1; Lemma 4.7 of [13] says that there is a homomorphism φi (Exel writes the maps ji as inclusions.) Since Li is a transfer operator for αi, αi ◦ Li extends to a self-adjoint projection ei in L(Mi). Exel's C ∗-algebra A is the C ∗-subalgebra of lim (L(Mi), φi) generated by the images of A = L(M0) and {ei : i ∈ N}. Propositions −→ 4.2 and 4.3 of [13] say that α and L extend to isometric linear maps αi : Mi → Mi+1 and Li : Mi+1 → Mi, Proposition 4.6 of [13] says that the maps β′ i : T 7→ αi ◦ T ◦ Li are injective homomorphisms of L(Mi) into L(Mi+1), and Proposition 4.10 of [13] says (L(Mi), φi) which leaves A invariant and that they induce an endomorphism β′ of lim −→ satisfies β′(ei) = ei+1 for i ≥ 0. To compare our construction with that of [13, §4], we use the maps Vi : q(a1) ⊗A · · · ⊗A q(ai) 7→ qi(cid:0)a1α(a2)α2(a3) · · · αi−1(ai)(cid:1); the pairs (Vi, αi) then form compatible isomorphisms of (M ⊗i induce isomorphisms θi of L(M ⊗i Ui of Corollary 4.6 satisfy L , A) onto (Mi, αi(A)), and L ) onto L(Mi). One quickly checks that the isometries Vi+1UiV −1 i (qi(a)) = Vi+1Ui(q(a) ⊗A 1 · · · ⊗A 1 / O O o o o o 12 AN HUEF AND RAEBURN = Vi+1(q(1) ⊗A q(a) ⊗A · · · ⊗ q(1)) = qi+1(α(a)) = αi(qi(a)), and similiarly ViU ∗ i Vi+1(qi+1(a)) = qi(L(a)) = Li(qi+1(a)). Thus our endomorphism Ad Ui is carried into Exel's β′ i. The isomorphisms θi combine to give an injection of our (L(Mi), φi), and since tit∗i = βi(1) is carried into direct limit C∞ = lim −→ (β′)i(1) = ei, the formula (4.8) implies that the range of this injection is span{aeib : a, b ∈ A}, which by [13, Proposition 4.9] is precisely A. So (C∞, β) is indeed isomorphic to ( A, β′), and Corollary 4.7 extends [13, Theorem 6.5]. (Ci, φi) into lim −→ 5. Graph algebras as crossed products Let E = (E0, E1, r, s) be a locally-finite directed graph with no sources or sinks. We use the conventions of [36]. Briefly, we think of E0 as vertices, E1 as edges, and r, s : E1 → E0 as describing the range and source of an edge. Locally finite means that E is both row-finite and column-finite, so that both r−1(v) and s−1(v) are finite for every v ∈ E0. We write E∗ for the set of finite paths µ = µ1 . . . µn satisfying s(µi) = r(µi+1) for 1 ≤ i ≤ n − 1, and µ for the length n of this path. Similarly, we write En for the set of paths of length n and E∞ for the set of infinite paths η = η1η2 . . . . We equip the path space E∞ with the product topology inherited from n=1 E1, which is locally compact and Hausdorff and has a basis consisting of the cylinder sets Z(µ) := {η ∈ E∞ : ηi = µi for 1 ≤ i ≤ µ} parametrised by µ ∈ E∗. Now consider the backward shift σ on E∞ defined by σ(η1η2 . . . ) = η2η3 . . . . Since E has no sinks, σ is surjective. Since E is column-finite, σ is a local homeomorphism which is proper in the sense that inverse images of compact sets are compact (see [9, §2.2]). Since σ is proper, α : f 7→ f ◦ σ is a nondegenerate endomorphism of C0(E∞); since E is column-finite, σ−1(η) is finite, and Q∞ L(f )(η) = f (ξ) = 1 σ−1(η) Xσ(ξ)=η 1 s−1(r(η)) Xs(e)=r(η) f (eη) defines a transfer operator L : C0(E∞) → C0(E∞) for α [9, Lemma 2.2]. The same formula defines an operator L on Cb(E∞) = M(C0(E∞)), so (C0(E∞), α, L) is an Exel system. A partition of unity argument (as preceding [9, Corollary 4.2]) shows that the ac- tion of C0(E∞) on ML is by compact operators. Hence C0(E∞) ⋊α,L N := O(Kα, ML) = O(ML). A Cuntz-Krieger E-family in a C ∗-algebra B consists of a set {Pv : v ∈ E0} of mutually orthogonal projections and a family {Te : e ∈ E1} of partial isometries such that T ∗ e for all v ∈ E0. The C ∗-algebra C ∗(E) of E is universal for Cuntz-Krieger E-families; we write {t, p} for the universal Cuntz-Krieger E-family that generates C ∗(E). See [36] for more details. e Te = Ps(e) for all e ∈ E1 and Pv =Pr(e)=v TeT ∗ For e ∈ E1, we define me := s−1(s(e))1/2q(χZ(e)). Since E is locally finite with no sources or sinks, we know from [9, Theorem 5.1] that Te := kM (me) = s−1(s(e))1/2kM (q(χZ(e))) and Pv := kA(χZ(v)) form a Cuntz-Krieger E-family in O(ML), and that πT,P : te 7→ Te and pv 7→ Pv is an isomorphism of the graph algebra C ∗(E) onto C0(E∞) ⋊α,L N = O(ML). Note that πT,P α(χZ(µ)) = χZ(µ) ◦ σ = Xs(e)=µ χZ(e). So for paths µ and ν of length i we have πT,P (β(tµt∗ ν ) = α′(k⊗i M (mµ)k⊗i M (mν)∗) ν)) = α′(TµT ∗ M (χZ(r(µ)) · mµ)k⊗i M (χZ(r(ν)) · mν)∗(cid:1) M (q(α(χZ(r(µ))) ⊗ mµ)k⊗i+1 M (q(α(χZ(r(ν))) ⊗ mν)∗ = k⊗i+1 = α′(cid:0)k⊗i = Xs(e)=r(µ), s(f )=r(ν) = Xs(e)=r(µ), s(f )=r(ν) (s−1(s(e)) s−1(s(f )))−1/2k⊗i+1 (using (4.1)) M (me ⊗A mµ)k⊗i+1 M (mf ⊗A mν)∗ (s−1(s(e)) s−1(s(f )))−1/2k⊗i+1 M (meµ)k⊗i+1 M (mf ν)∗. STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 13 is equivariant for the gauge actions. Pulling over the endomorphism α′ of Theorem 4.1 gives an endomorphism β = π−1 T,P ◦ α′ ◦ πT,P of the core We are going to compute β. C ∗(E)γ = span(cid:8)tµt∗ ν : µ, ν ∈ E∗ satisfy µ = ν(cid:9). Let µ = µ1 . . . µi be a finite path and mµ := mµ1 ⊗A · · · ⊗A mµi. Then Tµ = Tµ1Tµ2 · · · Tµi = kM (mµ1)kM (mµ2) · · · kM (mµi) = k⊗i M (mµ). To compute β we first note that mµ = χZ(r(µ)) · mµ and W := π−1 Now recall that α′ is conjugation by an isometry V = limλ kM (q(eλ)), where {eλ} is is any approximate identity of C0(E∞). A quick calculation with, for example, the approximate identity {eF =Pe∈F χZ(e)} indexed by finite subsets F of E1, shows that T,P (V ) =Pe∈E 1 s−1(s(e))−1/2te. So Theorem 4.1 gives: Proposition 5.1. Suppose that E is a locally-finite directed graph with no sources or sinks, and {te, pv} is the universal Cuntz-Krieger E-family which generates C ∗(E). Then there is an endomorphism β of the core C ∗(E)γ such that (5.1) The series β(tµt∗ ν) = Xs(e)=r(µ), s(f )=r(ν)(cid:0)s−1(r(µ)) s−1(r(ν))(cid:1)−1/2teµt∗ f ν. s−1(s(e))−1/2te Xe∈E 1 converges strictly in M(C ∗(E)) to an isometry W satisfying β(tµt∗ νW ∗. If ι is the inclusion of the core C ∗(E)γ in C ∗(E), then the associated representation ι × W of the Stacey crossed product C ∗(E)γ ×β N is an isomorphism onto C ∗(E). ν) = W tµt∗ Although we found the endomorphism β using our general construction, and were surprised to find it, we have now learned that other authors have shown that Cuntz- Krieger algebras can be realised as a Stacey crossed product by an endomorphism of the core (see [37, Example 2.5], [21, Theorem 3.2], and Remark 5.3 below). Now that we have found our β we should be able to prove Proposition 5.1 directly. Before doing this we will revisit the need for the hypotheses on the graph E: we used row-finiteness and 14 AN HUEF AND RAEBURN the lack of sources to get that the path space E∞ is locally compact, but a direct proof should not go through C0(E∞). Our formula for β, on the other hand, makes sense when E is column-finite and has no sinks: the coefficients in (5.1) are crucial, as we will see in the proof of the next result. It seems likely that column-finiteness is necessary. There is no obvious way to adjust for sinks, either: if we try to interpret empty sums as 0, then β(tµt∗ ν) would be 0 if either µ or ν ends at a sink, but this property is not preserved by multiplication. (If ν ends at a sink but µ doesn't, then we'd have β(tµt∗ ν) = 0 but β((tµt∗ µ) 6= 0.) So we settle for the following. µ)) = β(tµt∗ ν)(tνt∗ Theorem 5.2. Suppose that E is a column-finite directed graph with no sinks, and {te, pv} is the universal Cuntz-Krieger E-family which generates C ∗(E). Then there is an endomorphism β of the core C ∗(E)γ such that β(tµt∗ (5.2) The series ν) = Xs(e)=r(µ), s(f )=r(ν)(cid:0)s−1(r(µ)) s−1(r(ν))(cid:1)−1/2teµt∗ f ν. s−1(s(e))−1/2te Xe∈E 1 converges strictly in M(C ∗(E)) to an isometry W satisfying β(tµt∗ νW ∗. If ι is the inclusion of the core C ∗(E)γ in C ∗(E), then the associated representation ι × W of the Stacey crossed product C ∗(E)γ ×β N is an isomorphism onto C ∗(E). ν) = W tµt∗ Proof. We use the notation of [36, §3] and Appendix A. For each v ∈ E0, {tµt∗ ν : µ = ν = i, s(µ) = s(ν) = v} is a set of matrix units for Fi(v) (see [4, page 312]). We claim their images eµ,ν under β in Fi+1(v), defined by the right-hand side of (5.2), are also matrix units. The product eµ,νeκ,λ contains terms like teµt∗ hλ, which is zero unless f = g and ν = κ. Since we then have r(ν) = r(κ), the two central terms in the coefficient are the same, and f νtgκt∗ eµ,νeκ,λ = s−1(r(µ))−1/2s−1(r(ν))−1s−1(r(λ))−1/2teµt∗ hλ. For fixed e and h, there are s−1(r(ν)) edges f with s(f ) = r(ν), and for each of these the summand is exactly the same. So s−1(r(µ))−1/2s−1(r(λ))−1/2seµs∗ hλ = eµ,λ. X s(e)=r(µ), s(f )=r(ν), s(h)=r(λ) eµ,νeκ,λ = Xs(e)=r(µ), s(h)=r(λ) To define β on Ci (Notice that the coefficients in the definition of eµ,ν had to be just right for this to work.) Thus {eµ,ν} is a set of matrix units as claimed, and there is a well-defined homomorphism βi : Fi(v) → Fi+1(v) satisfying (5.2) (with β replaced by βi). These combine to give a := F0 + F1 + · · · + Fi, we take the i-expansion c = Pi homomorphism βi of Fi =Lv Fi(v) into Fi+1 =Lv Fi+1(v). described in Proposition A.1(b), and define βi(c) =Pi they give a well-defined function onS∞ c ∈ Ci ⊂ Ci+1. Suppose that c ∈ Ci has i-expansion c = Pi expansion is c = (cid:0)Pi−1 j=0 cj j=0 βj(cj). The uniqueness of the i-expansion implies that this gives a well-defined function βi on each Ci; to check that i=0 Ci, we need to check that βi(c) = βi+1(c) for j=0 cj. Then the (i + 1)- i ∈ Ei and d ∈ Fi ∩ Fi+1 are uniquely j=0 cj(cid:1) + c′ i + d, where c′ STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 15 determined by ci = c′ ν belongs to Fi ∩ Fi+1, then v := s(µ) = s(ν) satisfies 0 < r−1(v) < ∞, and two applications of the Cuntz-Krieger relation at v show that i + d. Lemma A.2 implies that if tµt∗ βi+1(tµt∗ tµgt∗ νg(cid:17) ν) = βi+1(cid:16) Xr(g)=v X = r(g)=v, s(e)=r(µg), s(f )=r(νg)(cid:0)s−1(r(µ)) s−1(r(ν))(cid:1)−1/2teµgt∗ = Xs(e)=r(µ), s(f )=r(ν)(cid:0)s−1(r(µ)) s−1(r(ν))(cid:1)−1/2teµt∗ f ν f νg = βi(tµt∗ ν). i) + βi+1(d) = βi(c′ Thus βi(c′ i) + βi(d) = βi(ci). Thus βi+1(c) =(cid:16) i−1Xj=0 βi(cj)(cid:17) + βi(c′ i) + βi+1(d) =(cid:16) i−1Xj=0 βi(cj)(cid:17) + βi(ci) = βi(c). At this stage we have a well-defined map β :S∞ i=0 Ci → C ∗(E)γ satisfying (5.2). This map is certainly linear, and we need to prove that it is multiplicative. We consider tµt∗ λ ∈ Fj, and we may as well suppose i ≤ j. Then multiplying together the two formulas for βi(tµt∗ λ) gives a linear combination of things like teµt∗ hλ. So the sum collapses just as it did in the first paragraph, and we obtain the formula for hλ. Because i ≤ j, this product is 0 unless gκ = f νκ′, and then it is teµκ′t∗ ν ∈ Fi and tκt∗ ν) and βj(tκt∗ f νtgκt∗ β(tµκ′t∗ λ) = β((tµt∗ ν)(tκt∗ λ)). i=0 Ci = C ∗(E)γ. Thus β is multiplicative. Since it is clearly ∗-preserving, it is a ∗-homomorphism, and as such is automatically norm-decreasing on each Ci. Thus β extends to an endomorphism, same initial projection pv, and mutually orthogonal range projections tet∗ Next, note that for each v ∈ E0, the partial isometries {te : s(e) = v} have the e, so2 Tv := also called β, ofS∞ Ps(e)=v s−1(v)−1/2te is a partial isometry with initial projection pv and range projection Ps(e)=v=s(f ) s−1(v)−1tet∗ their sum converges strictly to a partial isometry W with initial projectionPv∈E 0 pv = f . Now the partial isometries {Tv : v ∈ E0} have mutually orthogonal initial projections and mutually orthogonal range projections, and hence νW ∗ is easy to check, and the universal prop- erty of the Stacey crossed product (C ∗(E)γ ×β N, iC ∗(E)γ , v) gives a homomorphism ι × W such that (ι × W ) ◦ iC ∗(E)γ = ι and (ι × W )(v) = W . This homomorphism satisfies (ι × W ) ◦ βz = γz ◦ (ι × W ), and since ι is faithful (being an inclusion), the dual-invariant uniqueness theorem (Proposition 3.1) implies that ι×W is injective. The range contains each te = s−1(s(e))1/2ι(tet∗ (cid:3) 1M (C ∗(E)). In other words, W is an isometry. The covariance relation β(tµt∗ e)W , and hence ι × W is surjective. ν) = W tµt∗ Remark 5.3. When the graph E is finite and has no sinks, the endomorphism β has also been found by Kwa´sniewski [21]. He proves in [21, Theorem 3.2] that C ∗(E) is isomorphic to a partial-isometric crossed product C ∗(E)γ ⋊β Z as introduced in [3]. He 2It is important here that there are only finitely many summands, so column-finiteness is crucial. 16 AN HUEF AND RAEBURN then applies general results about partial-isometric crossed products from [3] and [21, §1] to (C ∗(E)γ, β), and recovers many of the main structure theorems for graph C ∗- algebras, as they apply to finite graphs with no sinks [21, §3]. For such graphs E, the endomorphism β is conjugation by an isometry in C ∗(E) and is injective. Theorem 4.15 of [3] implies that C ∗(E)γ ⋊β Z is isomorphic to the Exel crossed product C ∗(E)γ ⋊β,K N, where K = β−1 ◦ Ad ¯β(1). We can then deduce from [12, Theorem 4.7] that C ∗(E) is isomorphic to the Stacey crossed product C ∗(E)γ ×β N, as in Theorem 5.2. rank-one projection, and k=1 ak 7→ e11 ⊗(cid:0)N∞ Example 5.4. Suppose that E is the bouquet of n loops on a single vertex, so that C ∗(E) is the Cuntz algebra On. Then β is not quite the same as the endomorphism α in the usual description of On as a Stacey crossed product of its core [11, 38, 5], which is defined in terms of the infinite tensor-product decomposition Oγ k=1 Mn(C) by i,j=1 n−1eij is also a α :N∞ k=1 ak(cid:1) = p ⊗(cid:0)N∞ If u is a unitary matrix such that ue11u∗ = p, then u ⊗ 1 is a unitary in N∞ such that (u ⊗ 1)α(a)(u ⊗ 1)∗ = β(a) for all a ∈N∞ k=2 ak−1(cid:1). But it is closely related: p =Pn β(cid:0)N∞ k=1 Mn(C) k=1 Mn(C). In particular, the crossed n ×β N are isomorphic. This argument would work for any choice products Oγ of rank-one projection p; ours has the advantage that no obvious choice is necessary. k=2 ak−1(cid:1). n =N∞ n ×α N and Oγ (5.3) One naturally asks whether it is a coincidence that our convoluted constructions yields essentially the same endomorphism, and we will show in the next section that for every projection p in Mn(C), rank-one or not, there is an endomorphism βp of On satisfying (5.3). Particularly interesting is the endomorphism associated to the identity matrix i . Since β1 is unital, the Stacey crossed product is not interesting, but there is a natural transfer operator, and hence a potentially interesting Exel crossed product. We will study this in the next section. 1n, which turns out to be the "canonical endomorphism" β1(a) 7→Pn i=1 sias∗ Remark 5.5. In a graph algebra C ∗(E), it does not seem to be so easy to write down other endomorphisms of the core, or at least ones which are not unitarily equivalent to e, which is sometimes referred to the β of Theorem 5.2. The map β1 : a 7→Pe∈E 1 seas∗ as the "canonical endomorphism," is not a homomorphism for most graphs3. 6. Endomorphisms of the UHF core of On We fix n ≥ 2, and view the UHF algebra A := UHF(n∞) as the direct limit of the tensor powers M ⊗k n of Mn(C) with bonding maps φk : a1 ⊗ · · · ⊗ ak 7→ a1 ⊗ · · · ⊗ ak ⊗ 1. We think of elements of A as infinite tensors, so we write a1 ⊗ · · · ⊗ ak ⊗ 1∞ for the image of a1 ⊗ · · · ⊗ ak ∈ M ⊗k in A. These elements span a dense ∗-subalgebra of A. n Let {eij : 1 ≤ i, j ≤ n} be the usual matrix units in Mn. Then the elementary tensors eµν := eµ1ν1 ⊗ · · · ⊗ eµkνk parametrised by multiindices µ, ν ∈ {1, . . . , n}k are matrix units which span M ⊗k n . The next lemma is standard. 3To see this, consider a pair of paths µ, ν with s(µ) = s(ν) but r(µ) 6= r(ν). Since sesµ = 0 unless s(e) = r(µ), and similarly for s∗ ν s∗ e, we have β1(sµs∗ ν) = 0, but β1((sµs∗ ν)(sµs∗ ν)∗) = β1(sµs∗ µ) 6= 0. STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 17 Lemma 6.1. Let {si : 1 ≤ i ≤ n} be the universal Cuntz family in On, and let γ : T → Aut On be the gauge action. Then there is an isomorphism ψ of A onto Oγ n such that ψ(eµν ⊗ 1∞) = sµs∗ ν := sµ1 · · · sµns∗ νn · · · s∗ ν1. We now take N ≤ n, and let p be the diagonal matrix 1N ⊕ 0n−N in Mn. Then are homomorphisms which satisfy the maps αk : a 7→ p ⊗ a of M ⊗k αk+1 ◦ φk = φk+1 ◦ αk, and hence induce an endomorphism α of A such that (6.1) α(a1 ⊗ · · · ⊗ ak ⊗ 1∞) = p ⊗ a1 ⊗ · · · ⊗ ak ⊗ 1∞. into M ⊗(k+1) n n Next, we identify the corner pMnp with MN (C) in the obvious way, and let tr : pMnp → C be the normalised trace on pMnp. Lemma 6.2. There is a positive linear map L : A → A such that (6.2) L(a1 ⊗ · · · ⊗ ak ⊗ 1∞) = tr(pa1p)(a2 ⊗ · · · ⊗ ak ⊗ 1∞). n → M ⊗k Then L is a transfer operator for α, and L is almost faithful. Proof. Define Lk : M ⊗(k+1) n by Lk(a1 ⊗ · · · ⊗ ak+1) = tr(pa1p)(a2 ⊗ · · · ⊗ ak+1). A calculation shows that φk ◦ Lk = Lk+1 ◦ φk+1, so the Lk combine to give a linear function n of A satisfying (6.2). Each Lk is positive and linear in A are n , and hence extends uniquely L on the dense subalgebraSk M ⊗k norm-preserving, the map L is norm-decreasing onSk M ⊗k with Lk(1) = 1, and hence is norm-decreasing; since the inclusions of the M ⊗k to a norm-decreasing linear map L : A → A. It is positive because each Lk is. n ⊂ A and calculate: To see that L is a transfer operator, we take a, b ∈ M ⊗k n L(α(a)b) = L(pb1 ⊗ a1b2 ⊗ · · · ⊗ ak−1bk ⊗ ak ⊗ 1∞) = tr(pb1)(a1b2 ⊗ · · · ⊗ ak−1bk ⊗ ak ⊗ 1∞) = tr(pb1p)(a1b2 ⊗ · · · ⊗ ak−1bk ⊗ ak ⊗ 1∞) = (a1 ⊗ · · · ⊗ ak ⊗ 1∞) tr(pb1p)(b2 ⊗ · · · ⊗ bk ⊗ 1∞) = aL(b). Next we use Lemma 6.1 to view A as a subalgebra of On, set X := {1, . . . , n}∞, and let πS : On → B(ℓ2(X)) be the infinite path representation. Let {δx : x ∈ X} be the usual basis for ℓ2(X). Then we claim that (6.3) (cid:0)πS(L(a))δx δy(cid:1) = N −1 NXi=1(cid:0)πS(a)δix δiy(cid:1) for a ∈ A and x, y ∈ X. It suffices by linearity and continuity to check this for a = eµν ⊗1∞. Then with µ = µ1µ′ and ν = ν1ν′, vanishes unless µ1 = ν1 ≤ N, and if so, (cid:0)πS(L(a))δx δy(cid:1) =(cid:0)πS(tr(peµ1ν1p)eµ′ν ′ ⊗ 1∞)δx δy(cid:1) The sum on the right-hand side of (6.3) is ν ′δx δy(cid:1). (cid:0)πS(L(a))δx δy(cid:1) = N −1(cid:0)Sµ′S∗ NXi=1(cid:0)SµS∗ NXi=1(cid:0)S∗ ν δix δiy(cid:1) = N −1 ν δix S∗ µδiy(cid:1); N −1 18 AN HUEF AND RAEBURN the ith summand vanishes unless ν1 = i = µ1, which in particular forces ν1 = µ1 ≤ N; then S∗ ν ′δx δy), and we have proved (6.3). ν ′δx, so the right-hand side of (6.3) also reduces to N −1(Sµ′S∗ ν δix = S∗ To see that L is almost faithful, suppose that a ∈ A has the property that L((ab)∗ab) = 0 for all b ∈ A. Then, with πS as in the previous paragraph, we have (cid:0)πS(L((ab)∗ab))δx δx(cid:1) = 0 for all b ∈ A, x ∈ X =⇒ =⇒ =⇒ NXi=1(cid:0)πS((ab)∗ab)δix δix(cid:1) = 0 for all b ∈ A, x ∈ X NXi=1(cid:0)πS(ab)δix πS(ab)δix(cid:1) = 0 for all b ∈ A, x ∈ X NXi=1 kπS(ab)δixk2 = 0 for all b ∈ A, x ∈ X =⇒ kπS(ab)δixk2 = 0 for all i ≤ N, b ∈ A, x ∈ X. We can in particular take b = eji ⊗ 1∞ for any j ≤ n, and deduce that 0 = kπS(ab)δixk = kπS(a)SjS∗ i δixk = kπS(a)δjxk for all j ≤ n and x ∈ X; since {δjx : j ≤ n, x ∈ X} = {δy : y ∈ X}, we deduce that πS(a) = 0 and a = 0. Thus L is almost faithful. (cid:3) Now we have an Exel system (A, α, L), and it is natural to ask what the Exel crossed product A ⋊α,L N is. Lemma 6.3. The elements form an orthonormal basis for the right Hilbert A-module ML. (cid:8)Eij := N 1/2q(eij ⊗ 1∞) : 1 ≤ i ≤ n, 1 ≤ j ≤ N(cid:9) the inner product ha , bi := L(a∗b). Write a1 ∈ Mn asPn Proof. The Hilbert module (ML)A is the completion of A in the seminorm defined by i,j=1 cijeij for some cij ∈ C. A calculation shows that k(a1 − a1p) ⊗ a2 ⊗ · · · ⊗ ak ⊗ 1∞k2 = 0. Since also eijp = 0 for j > N we have q(a1 ⊗ · · · ⊗ ak ⊗ 1∞) = q(a1p ⊗ a2 ⊗ · · · ⊗ ak ⊗ 1∞) = = nXi=1 nXi=1 NXj=1 NXj=1 cijq(eij ⊗ a2 ⊗ · · · ⊗ ak ⊗ 1∞) cijq(eij ⊗ 1∞) · (a2 ⊗ · · · ⊗ ak ⊗ 1∞). Thus the Eij generate ML as a right Hilbert A-module. To see that {Eij} is orthonormal, we compute hEij, Ekli = N tr(pe∗ ijekl vanishes unless i = k, and then hEij, Eili = 1A when j = l and 0 otherwise. (cid:3) ijeklp)1A: the product e∗ STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 19 Lemma 6.4. Suppose {Tij : 1 ≤ i ≤ n, 1 ≤ j ≤ N} is a Cuntz family of isometries in a C ∗-algebra C. Then there is a unital homomorphism πT : A → C such that (6.4) T ∗ ijπT (a)Tkl = πT (hEij, a · Ekli) for all a ∈ A and ij, kl. Proof. For µ, ν ∈ {1, . . . , n}k we define Tµν := Tµ1ν1 · · · Tµkνk. We claim that nFµν := Xλ∈{1,...,N }k TµλT ∗ νλ : µ, ν ∈ {1, . . . , n}ko is a family of nonzero matrix units in C such thatPµ∈{1,...,n}k Fµµ = 1. Since {Tij} is a Cuntz nN-family, {Tµλ} is a Cuntz (nN)k-family, and hencePµ∈{1,...,n}k Fµµ = 1. Since the Tµλ are isometries with orthogonal ranges, the Fµν are nonzero, and FµνFστ = Xλ,ω∈{1,...,N }k TµλT ∗ νλTσωT ∗ τ ω is either Fµτ or 0. So for each k, there is a homomorphism πk : M ⊗k n → C such that πk(eµν) = Fµν. For a = eµν ∈ M ⊗k n , we have πk+1(a ⊗ 1) = = πk+1(eµν ⊗ eii) = nXi=1 nXi=1 Xλ∈{1,...,N }k+1 nXi=1 NXj=1 = Xλ′∈{1,...,N }k Tµλ′(cid:16) nXi=1 = Xλ′∈{1,...,N }k = Xλ′∈{1,...,N }k πT (eµν ⊗ 1∞) = Xλ∈{1,...,N }k Tµλ′T ∗ πk+1(e(µi)(νi)) nXi=1 T(µi)λT ∗ (νi)λ T(µi)(λ′j)T ∗ (νi)(λ′j) NXj=1 TijT ∗ ij(cid:17)T ∗ νλ′ νλ′ = πk(eµν) = πk(a). TµλT ∗ νλ, Thus we have a well-defined unital homomorphism πT : A → C such that and it remains for us to check that (πT , {Tij}) satisfy (6.4). It suffices to check this on a of the form eµν ⊗ 1∞. So suppose eµν ∈ M ⊗(k+1) n . Then T ∗ ijπT (eµν ⊗ 1∞)Tkl = T ∗ ij(cid:16) Xλ∈{1,...,N }k+1 Tµ1λ1 · · · Tµk+1λk+1T ∗ νk+1λk+1 · · · T ∗ ν1λ1(cid:17)Tkl. 20 AN HUEF AND RAEBURN The λ-summand vanishes unless ij = µ1λ1 and ν1λ1 = kl, so all terms vanish unless j = l, i = µ1 and k = ν1. Thus if we again write µ = µ1µ′, etc, then if j = l, i = µ1 and k = ν1 otherwise. (6.5) T ∗ ijπT (eµν ⊗ 1∞)Tkl =(Pλ′∈{1,...,N }k Tµ′λ′T ∗ 0 ν ′λ′ On the other hand, we have πT(cid:0)hEij, eµν · Ekli(cid:1) = N 1/2πT(cid:0)hEij, (eµ1ν1ekl) ⊗ eµ′ν ′ ⊗ 1∞i(cid:1) =(0 N 1/2πT(cid:0)hEij, eµ1l ⊗ eµ′ν ′ ⊗ 1∞i(cid:1) =(0 ijeµ1lp)(eµ′ν ′ ⊗ 1∞)(cid:1) NπT(cid:0) tr(pe∗ =(0 NπT(cid:0)N −1(eµ′ν ′ ⊗ 1∞)(cid:1) which is the same as (6.5). unless k = ν1 if k = ν1 unless k = ν1 if k = ν1 unless k = ν1, i = µ1 and j = l if k = ν1, i = µ1 and j = l, (cid:3) To finish our analysis of A ⋊α,L N we need a general lemma. First, recall that a finite set {Fi : 1 ≤ i ≤ k} in a right Hilbert B-module M is a Parseval frame if (6.6) m = Fi · hFi, miB for every m ∈ M . kXi=1 Lemma 6.5. Let (B, β, K) be an Exel system with B unital. Suppose that {Fi}k i=1 is a Parseval frame for MK, that π : B → C is a unital homomorphism, and that {Si : 1 ≤ i ≤ k} is a Cuntz family of isometries in C such that (6.7) S∗ i π(b)Sj = π(hFi, b · FjiK) for all b ∈ B and 1 ≤ i, j ≤ k. Define ψ : MK → C by ψ(m) = Siπ(hFi, miK). kXi=1 Then (ψ, π) is a Cuntz-Pimsner covariant representation of MK in C such that ψ(Fi) = Si for all i. Proof. Let m, n ∈ MK and b ∈ B. We compute: ψ(m · b) = kXi=1 Siπ(hFi , m · biL) = kXi=1 Siπ(hFi , miKb) = ψ(m)π(b), ψ(m)∗ψ(n) = = kXi,j=1 kXi,j=1 π(hFi , miK)∗S∗ i Sjπ(hFj , niK) π(hFi , miK)∗π(hFi , FjiK)π(hFj , niK) STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 21 = kXi,j=1 π(cid:0)(cid:10)Fi · hFi , miK , Fj · hFj , niK(cid:11)K(cid:1) = π(hm , niK), and ψ(b · m) = Siπ(hFi , b · miK) = = kXi=1 kXi,j=1 =(cid:16) kXi=1 nXj=1 Fj · hFj , miKEK(cid:17) kXi=1 Siπ(cid:16)DFi , b · kXi,j=1 Siπ(hFi , b · FjiKhFj , miK) = SiS∗ i π(b)Sjπ(hFj , miK) SiS∗ i(cid:17)(cid:16) nXj=1 π(b)Sjπ(hFj , miK)(cid:17) = π(b)ψ(m). = kXi=1 kXi=1 Thus (ψ, π) is a representation of MK. Next we check the Cuntz-Pimsner covariance of (ψ, π). Using the reconstruction formula (6.6) we see that Thus Θb·Fi,Fi(m). kXi=1 φ(b)m = (b · Fi) · hFi , miK) = b · (Fi · hFi , miK) = kXi=1 kXi=1 Θb·Fi,Fi(cid:17) = (ψ, π)(1)(φ(b)) = (ψ, π)(1)(cid:16) kXi=1 kXi=1 π(b)ψ(Fi)ψ(Fi)∗ = π(b)(cid:16) kXi=1 ψ(b · Fi)ψ(Fi)∗ SiS∗ i(cid:17) = π(b), and so (ψ, π) is Cuntz-Pimsner covariant. Using (6.7) we have ψ(Fi) = since π is unital. Sjπ(hFj , FiiK) = SjS∗ j π(1)Si = Si kXi=1 (cid:3) Theorem 6.6. Let A := UHF(n∞) and (A, α, L) the Exel system described by (6.1) -- (6.2). Then there is an isomorphism ψ × π of A ⋊α,L N = O(ML) onto OnN . Proof. We parametrize the set {1, . . . , nN} by {ij : 1 ≤ i ≤ n, 1 ≤ j ≤ N}, and apply Lemma 6.4 to the canonical Cuntz family s = {sij} in OnN . This gives a unital homomorphism πs : A → OnN such that (6.4) holds. The orthonormal basis {Eij} of Lemma 6.3 is in particular a Parseval frame, and so Lemma 6.5 gives a Cuntz-Pimsner covariant representation (ψ, πs) of ML in OnN such that ψ(Eij) = sij. To see that ψ × πs is injective, we verify the hypotheses of [8, Corollary 5.3]. Since A is simple and πs is nonzero, πs is injective. The reconstruction formula for {Eij} implies that φ(a) = Pi,j Θa·Eij,Eij , and in particular that φ(A) ⊂ K(ML). Since A is simple, the 22 AN HUEF AND RAEBURN ideal Aα(A)A is all of A, and this implies that Kα = A = φ−1(K(ML)). Lemma 6.2 says that the transfer operator L is almost faithful. The gauge action γ on OnN satisfies (6.8) Thus [8, Corollary 5.3] implies that ψ ×πs is an isomorphism of A⋊α,L N into OnN . Since the range of ψ × πT contains every generator sij = πs(Eij), ψ × πs is also surjective. (cid:3) γz(ψ(Eij)) = γz(sij) = zsij = ψ(zEij) = ψ(bαz(Eij)). Remark 6.7. Since the isomorphism of Theorem 6.6 intertwines the dual action bα and the gauge action γ on OnN (see (6.8)), it carries the fixed-point algebra (A ⋊α,L N) bα onto Oγ nN = UHF((nN)∞). So Corollary 4.2 gives a description of A ⋊α,L N as a Stacey crossed product by a corner endomorphism on UHF((nN)∞). The rank N of the projection p ∈ Mn(C) in Theorem 6.6 is constrained by 1 ≤ N ≤ n. However, it is relatively easy to remove this constraint. Corollary 6.8. Let m, k ∈ N, let p ∈ M ⊗k the normalised trace on pM ⊗k lim −→j m such that M ⊗kj m be a projection of rank N, and let tr be m p. Then there are an endomorphism α of UHF(m∞) = α(a1 ⊗ · · · ⊗ aj ⊗ 1∞) = p ⊗ a1 ⊗ · · · ⊗ aj ⊗ 1∞, and a transfer operator L for α such that L(a1 ⊗ · · · ⊗ aj ⊗ 1∞) = tr(pa1p)(a2 ⊗ · · · ⊗ aj ⊗ 1∞). Then UHF(m∞) ⋊α,L N is isomorphic to OmkN . Proof. The results of this section apply with n = mk to give a system (UHF(n∞), α, L) with UHF(n∞) ⋊α,L N isomorphic to OmkN . Choosing an isomorphism of M ⊗k m with M ⊗j Mn gives an isomorphism of UHF(m∞) = lim n , and −→j pulling α and L over to UHF(m∞) gives an Exel system (UHF(m∞), α, L) with the required properties. (cid:3) m onto UHF(n∞) = lim −→j M ⊗kj Appendix A. The core in a graph algebra Suppose that E is an arbitrary directed graph, which could have infinite receivers, infinite emitters, sources and/or sinks. In Theorem 5.2, we wanted a description of the core C ∗(E)γ which did not depend on row-finiteness, and since we cannot recall seeing a suitable description in the literature, we give one here. As in the row-finite case, which is done in [4] and [36], our description uses the subspaces Fi := span{tµt∗ ν : µ = ν = i, s(µ) = s(ν)}, which one can easily check are in fact C ∗-subalgebras. (By convention, F0 = span{pv : v ∈ E0}.) We refer to vertices which are either infinite receivers or sources as "singular vertices". Recall that no Cuntz-Krieger relation is imposed at a singular vertex v, but if v is an e for every finite subset F infinite receiver then we impose the inequality pv ≥Pe∈F tet∗ of r−1(v). STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 23 Proposition A.1. Let E be a directed graph, and define Fi as above. i=0 Ci. cj ∈ Ej := span{tµt∗ (a) For i ≥ 0, Ci := F0 + F1 + · · · + Fi is a C ∗-subalgebra of C ∗(E)γ, Ci ⊂ Ci+1 and ν : µ = ν = j and s(µ) = s(ν) is singular} (b) For each i ≥ 0 and each c ∈ Ci, there are unique elements C ∗(E)γ =S∞ for 0 ≤ j < i and ci ∈ Fi such that c =Pi Proof of Proposition A.1(a). Since Si Ci is a vector space containing every element of Since Ci ⊂ Cj for i < j, an element of Ci has lots of the expansions described in (b). We refer to the one obtained by viewing c as an element of Cj as the j-expansion of c. the form tµt∗ ν with µ = ν, it is dense in C ∗(E)γ, and we trivially have Ci ⊂ Ci+1. We prove that Ci is a C ∗-subalgebra by induction on i. For i = 0, it is straightforward to see that C0 = F0 is a C ∗-subalgebra. Suppose that Ci is a C ∗-subalgebra. If κ = λ = i+1 and µ = ν ≤ i + 1, then the formula j=0 cj. (tµt∗ ν)(tκt∗ λ) =(0 tµκ′t∗ λ unless κ has the form νκ′ if κ = νκ′ shows that Fi+1 is an ideal in the C ∗-subalgebra C ∗(Ci+1) of C ∗(E)γ generated by Ci+1. Since Ci is a C ∗-subalgebra of C ∗(Ci+1), the sum Ci+1 = Ci+Fi+1 is also a C ∗-subalgebra (and in fact C ∗(Ci+1) = Ci + Fi+1 = Ci+1). (cid:3) For part (b), we need to do some preparation. The standard argument of [4, §2] or [36, Chapter 3] shows that, for fixed i and v ∈ E0, {tµt∗ ν : µ = ν = i, s(µ) = s(ν) = v} is a set of non-zero matrix units, and hence their closed span Fi(v) is a C ∗-subalgebra of Fi isomorphic to K(ℓ2(Ei ∩ s−1(v))). Since Fi(v)Fi(w) = {0} for v 6= w, Fi is the C ∗-algebraic direct sumLv∈E 0 Fi(v). For j satisfying 0 ≤ j ≤ i, we set Dj,i := Fj + · · · + Fi = span{tµt∗ ν : j ≤ µ = ν ≤ i, s(µ) = s(ν)}, which is another C ∗-subalgebra by the argument in the previous proof. It is an ideal in Ci. Now we prove a lemma. Lemma A.2. For every i ≥ 1 and every j < i, we have Fj(v) ∩ Dj+1,i =(0 Fj(v) if v is a singular vertex if 0 < r−1(v) < ∞. Proof. The result is trivially true if Fj(v) is {0}, so we suppose it isn't. Since Fj(v) ∩ Dj+1,i is an ideal in Fj(v), it is either {0} or Fj(v). First suppose that v is not a singular vertex. Then the Cuntz-Krieger relation at v implies that Fj(v) ⊂ Dj+1,i, and hence Fj(v) ∩ Dj+1,i = Fj(v). Now suppose that Fj(v) ∩ Dj+1,i = Fj(v); we will show that v is not singular. Choose µ is a non-zero element of Dj+1,i, so there exist κ and λ) 6= 0. But this implies that κ has the κ is non-zero, and µ ∈Ej with s(µ) = v. Then tµt∗ λ satisfying j + 1 ≤ κ = λ ≤ i and (tµt∗ form µκ′, and v = s(µ) cannot be a source. It also implies that tµt∗ µ)(tκt∗ µtκt∗ 24 AN HUEF AND RAEBURN hence so is the larger projection tµt∗ Fj(v) ∩ Fj+1 is an ideal in Fj(v), it follows that Fj(v) ∩ Fj+1 = Fj(v). µκj+1. Thus Fj(v) ∩ Fj+1 6= {0}, and since µtµκj+1t∗ We now know that the projection tµt∗ µ belongs to Fj+1, and hence is a projection in a C ∗-algebraic direct sumLw∈E 0 Fj+1(w). The norms of elements in this direct sum are µ =Pw∈F qw. Each arbitrarily small off finite subsets of E0, and projections have norm 0 or 1, so there are a finite subset F of E0 and projections qw ∈ Fj+1(w) such that tµt∗ qw is a projection in Fj+1(w) = K(ℓ2(Ej+1 ∩ s−1(w))), and hence has finite trace. Thus tµt∗ µ has finite trace. On the other hand, for every edge e with r(e) = s(µ) = v, we have tµt∗ µ ≥ tµet∗ µe : r(e) = s(µ)} is a family of mutually orthogonal projections of trace 1, we have µe; since {tµet∗ Tr(tµt∗ µ) ≥ Xr(e)=s(µ) Tr(tµet∗ µe) = r−1(s(µ)) = r−1(v). Since we have already eliminated the possibility that v is a source, this proves that it is not a singular vertex, as required. (cid:3) Proof of Proposition A.1(b). Lemma A.2 implies that and that Fj is the direct sum of Fj ∩ Dj+1,i and Fj ∩ Dj+1,i =M{Fj(v) : 0 < r−1(v) < ∞}, Ej =M{Fj(v) : v is a singular vertex}. Since Dj,i = Fj + Dj+1,i, we have Ci = F0 + D1,i = (E0 + (F0 ∩ D1,i)) + D1,i = E0 + D1,i = E0 + (E1 + (F1 ∩ D2,i)) = E0 + E1 + D2,i ... = E0 + · · · Ei−1 + Di,i = E0 + · · · Ei−1 + Fi, which shows that c has the claimed expansion. j=0 dj. Then c0 − d0 =Pi To establish uniqueness, suppose that cj, dj ∈ Ej for j < i, that ci, di ∈ Fi, and that j=1(dj − cj) belongs to E0 ∩ F0, because the left- hand side does, and to D1,i, because the right-hand side does. Since F0 = E0⊕(F0∩D1,i), j=1 dj. (cid:3) j=0 cj =Pi Pi we have E0 ∩ (F0 ∩ D1,i) = {0}, and we deduce that c0 = d0 and Pi Now an induction argument using Ej ∩ (Fj ∩ Dj+1,i) = {0} gives the result. j=1 cj = Pi References [1] S. Adji, Crossed products by semigroups of endomorphisms and the Toeplitz algebras of ordered groups, PhD thesis, University of Newcastle, 1995. [2] S. Adji, M. Laca, M. Nilsen and I. Raeburn, Crossed products by semigroups of endomorphisms and the Toeplitz algebras of ordered groups, Proc. Amer. Math. Soc. 122 (1994), 1133 -- 1141. [3] A.B. Antonevich, V.I. Bakhtin and A.V. Lebedev, Crossed product of a C ∗-algebra by an endo- morphism, coefficient algebras and transfer operators, arXiv:math/0502241. [4] T. Bates, D. Pask, I. Raeburn and W. Szyma´nski, The C ∗-algebras of row-finite graphs, New York J. Math. 6 (2000), 307 -- 324. [5] S. Boyd, N. Keswani and I. Raeburn, Faithful representations of crossed products by endomor- phisms, Proc. Amer. Math. Soc. 118 (1993), 427 -- 436. STACEY CROSSED PRODUCTS ASSOCIATED TO EXEL SYSTEMS 25 [6] B. Brenken, Hecke algebras and semigroup crossed product C ∗-algebras, Pacific J. Math. 187 (1999), 241 -- 262. [7] N. Brownlowe, A. an Huef, M. Laca and I. Raeburn, Boundary quotients of the Toeplitz algebra of the affine semigroup over the natural numbers, to appear in Ergodic Theory Dynam. Systems; arXiv:1009.3678. [8] N. Brownlowe and I. Raeburn, Exel's crossed product and relative Cuntz-Pimsner algebras, Math. Proc. Cambridge Philos. Soc. 141 (2006), 497 -- 508. [9] N. Brownlowe, I. Raeburn and S.T. Vittadello, Exel's crossed product for non-unital C ∗-algebras, Math. Proc. Cambridge Philos. Soc. 149 (2010), 423 -- 444. [10] T.M. Carlsen and S. Silvestrov, On the Exel crossed product of topological covering maps, Acta Appl. Math. 108 (2009), 573 -- 583. [11] J. Cuntz, The internal structure of simple C ∗-algebras, Proc. Symp. in Pure Math. 38, part 2 (1982), 85 -- 115. [12] R. Exel, A new look at the crossed-product of a C ∗-algebra by an endomorphism, Ergodic Theory Dynam. Systems 23 (2003), 1733 -- 1750. [13] R. Exel, Crossed-products by finite index endomorphisms and KMS states, J. Funct. Anal. 199 (2003), 153 -- 188. [14] R. Exel, A. an Huef and I. Raeburn, Purely infinite simple C ∗-algebras associated to integer dilation matrices, to appear in Indiana Univ. Math. J.; arXiv:1003.2097. [15] R. Exel and A. Vershik, C ∗-algebras of irreversible dynamical systems, Canad. J. Math. 58 (2006), 39 -- 63. [16] N.J. Fowler, Discrete product systems of bimodules, Pacific J. Math. 204 (2002), 335 -- 375. [17] N.J. Fowler, P.S. Muhly and I. Raeburn, Representations of Cuntz-Pimsner algebras, Indiana Univ. Math. J. 52 (2003), 569 -- 605. [18] N.J. Fowler and I. Raeburn, The Toeplitz algebra of a Hilbert bimodule, Indiana Univ. Math. J. 48 (1999), 155 -- 181. [19] A. an Huef and I. Raeburn, Exel crossed products, Stacey crossed products, and Cuntz-Pimsner algebras, preprint; arXiv:1107.1019. [20] T. Katsura, On C ∗-algebras associated with C ∗-correspondences, J. Funct. Anal. 217 (2004), 366 -- 401. [21] B. Kwa´sniewski, Cuntz-Krieger uniqueness theorem for crossed products by Hilbert bimodules, arXiv:1010.0446v2. [22] M. Laca, Semigroups of ∗-endomorphisms, Dirichlet series and phase transitions, J. Funct. Anal. 152 (1998), 330 -- 378. [23] M. Laca, S. Neshveyev and M. Trifkovi´c, Bost-Connes systems, Hecke algebras, and induction, arXiv:1010.4766. [24] M. Laca and I. Raeburn, Semigroup crossed products and the Toeplitz algebras of nonabelian groups, J. Funct. Anal. 139 (1996), 415 -- 440. [25] M. Laca and I. Raeburn, A semigroup crossed product arising in number theory, J. London Math. Soc. 59 (1999), 330 -- 344. [26] M. Laca, I. Raeburn and J. Ramagge, Phase transition on Exel crossed products associated to dilation matrices, J. Funct. Anal. 261 (2011), 3633 -- 3664. [27] N.S. Larsen, Crossed products by abelian semigroups via transfer operators, Ergodic Theory Dynam. Systems 30 (2010), 1147 -- 1164. [28] X. Li, Ring C ∗-algebras, Math. Ann. 348 (2010), 859 -- 898. [29] J. Lindiarni and I. Raeburn, Partial-isometric crossed products by semigroups of endomorphisms, J. Operator Theory 52 (2004), 61 -- 87. [30] P.S. Muhly and B. Solel, Tensor algebras over C ∗-correspondences: representations, dilations, and C ∗-envelopes, J. Funct. Anal. 158 (1998), 389 -- 457. [31] G.J. Murphy, Ordered groups and crossed products of C ∗-algebras, Pacific J. Math. 148 (1991), 319 -- 349. [32] G.J. Murphy, Simplicity of crossed products by endomorphisms, Integral Equations Operator The- ory 42 (2002), 90 -- 98. 26 AN HUEF AND RAEBURN [33] W.L. Paschke, The crossed product of a C ∗-algebra by an endomorphism, Proc. Amer. Math. Soc. 80 (1980), 113 -- 118. [34] W.L. Paschke, K-theory for actions of the circle group on C ∗-algebras, J. Operator Theory 6 (1981), 125 -- 133. [35] M. Pimsner, A class of C ∗-algebras generalizing both Cuntz-Krieger algebras and crossed products by Z, in Free Probability Theory, Fields Inst. Commun., vol. 12, Amer. Math. Soc., Providence, 1997, pages 189 -- 212. [36] I. Raeburn, Graph Algebras, CBMS Regional Conference Series in Mathematics, vol. 103, Amer. Math. Soc., Providence, 2005. [37] M. Rørdam, Classification of certain infinite simple C ∗-algebras, J. Funct. Anal. 131 (1995), 415 -- 458. [38] P.J. Stacey, Crossed products of C ∗-algebras by ∗-endomorphisms, J. Austral. Math. Soc. (Series A) 54 (1993), 204 -- 212. Department of Mathematics and Statistics, University of Otago, PO Box 56, Dunedin 9054, New Zealand E-mail address: [email protected], [email protected]
1307.5890	1	1307	2013-07-22T21:14:18	Chirality and principal graph obstructions	[ "math.OA", "math.QA" ]	Determining which bipartite graphs can be principal graphs of subfactors is an important and difficult question in subfactor theory. Using only planar algebra techniques, we prove a triple point obstruction which generalizes all known initial triple point obstructions to possible principal graphs. We also prove a similar quadruple point obstruction with the same technique. Using our obstructions, we eliminate some infinite families of possible principal graphs with initial triple and quadruple points which were a major hurdle in extending subfactor classification results above index 5.	math.OA	math	Chirality and principal graph obstructions David Penneys October 18, 2018 Abstract Determining which bipartite graphs can be principal graphs of subfactors is an important and difficult question in subfactor theory. Using only planar algebra techniques, we prove a triple point obstruction which generalizes all known initial triple point obstructions to possible principal graphs. We also prove a similar quadruple point obstruction with the same technique. Us- ing our obstructions, we eliminate some infinite families of possible principal graphs with initial triple and quadruple points which were a major hurdle in extending subfactor classification results above index 5. 1 Introduction Subfactor theory has many examples of unexpected discrete classification, begin- ning with Jones' index rigidity theorem [Jon83], which shows the index [B : A] of a subfactor A ⊂ B lies in the set {4 cos2(π/k)k ≥ 3} ∪ [4,∞]. We study a finite index subfactor by analyzing its standard invariant, which has many equivalent characterizations, such as Ocneanu's paragroups [Ocn88], Popa's λ-lattices [Pop95], and Jones' planar algebras [Jon99]. The standard invariant also encodes the index and the principal graphs, which are the bipartite induc- tion/restriction multi-graphs associated to tensoring with ABB and BBA. √ A second example of discrete classification is the ADE classification of hyper- finite subfactors with index at most 4, where Dodd and E7 do not occur [Ocn88, Pop94]. Haagerup classified principal graphs of subfactors with index in the range 3) [Haa94], and recently, subfactor planar algebras with index less than (4, 3 + 5 were completely classified [MS12b, MPPS12, IJMS12, PT12]. The recent survey [JMS13] provides an excellent introduction to the subfactor classification program, along with state-of-the-art information on its progress. These classifications have two main parts: restricting the list of possible principal graphs, and constructing examples when the graphs survive. The former task relies on principal graph obstructions, which rule out many possible principal graphs by combinatorial constraints, or by relating local structure of the principal graph to data intrinsic to the subfactor planar algebra. Several examples of the latter type of obstruction for triple points include Oc- neanu's triple point obstruction [Haa94], Jones' quadratic tangles obstruction [Jon12], 1 the triple-single obstruction [MPPS12], and Snyder's singly valent obstruction [Sny12], which mutually generalizes the quadratic tangles and triple-single obstructions. In this article, we prove a triple point obstruction which is strictly stronger than all known triple point obstructions for initial triple points, and a quadruple point obstruction of similar flavor. The statement of the main theorem for initial triple points uses the following notation. Suppose that the principal graphs (Γ+, Γ−) of a subfactor planar algebra have an initial triple point at depth n − 1 (where n ≥ 2), and that the projections one past the branch at depth n of Γ+ are labelled P, Q. ··· n − 2 n − 1 0 1 Q P n Tr(Q) P − Q is a rotational eigenvector orthogonal to TLn,+ with In this case, A = Tr(P ) rotational eigenvalue ωA = σ2 A (see Subsection 2.2 and the beginning of Section 3). We refer to ωA (and sometimes σA) as the chirality of the subfactor planar algebra. Theorem. Suppose that for each R at depth n + 1 connected to P , there is a unique vertex E(R) at depth n connected to the dual vertex R of R. Then there is an explicit formula for σA + σ−1 A in terms of the traces of the projections of Γ± with depth at most n + 1. (See Theorem 3.3 and Remark 3.4 for more details.) The importance of this formula is that it gives us the chirality, which is a priori hidden in the planar algebra structure, in terms of visible combinatorial data of the principal graph. As corollaries, we obtain the obstructions of Jones and Snyder (Corollary 3.8) and a stronger version of Ocneanu's obstruction for initial triple points. Corollary. Under the hypotheses of Ocneanu's obstruction for initial triple points, σA + σ−1 A = [n + 2] − [n]. (See Theorem 3.5 for more details.) After obtaining the chirality using the above corollary, we provide a quick proof that Dodd and E7 are not principal graphs by showing that the chirality is incom- patible with the supertransitivity (see Remark 3.7 and Examples 3.12). In Theorem 3.10, we obtain an obstruction for annular multiplicities ∗11 princi- pal graphs, which were a major hurdle in extending subfactor classification results above index 5 (see Subsection 2.2 for the definition of annular multiplicities). In particular, we prove the following theorem, which eliminates a particular example arising from running the principal graph odometer [MS12b] above index 5. Theorem. There is no subfactor with principal graphs a translated extension of (cid:17) . (cid:16) W = , (See Definition 2.6 for the definition of "translated extension," and see Subsection 3.3 for more details on eliminating W.) 2 Our techniques also apply to initial quadruple points which are even translated extensions of Q = , (cid:16) (cid:17) . In this case, there are two rotational eigenvectors A, B orthogonal to Temperley- Lieb one past the branch, and the chiralities ωA = σ2 B are distinct (see Proposition 4.1). We get three equations in Theorem 4.2: two for the chirality σA, and one involving both σA and σB. In a specific case, we obtain a quadruple point obstruction similar to Ocneanu's triple point obstruction in Theorem 4.5. We use this obstruction to eliminate two weeds in Corollary 4.7. A and ωB = σ2 This article only relies on planar algebra techniques, and the one click rotation F = n − 1 n − 1 plays a crucial role. In particular, we use a quadratic relation due to Liu (Lemma 2.12) which is a clever variant of Wenzl's relation [Wen87]. Such relations restrict the structure of the principal graph via the skein theory of the planar algebra, giving strong consequences. (For another such example of this phenomenon, see [BP13].) Liu's relation is effective because it only involves four terms, two of which are in Temperley-Lieb and can be ignored for our purposes, and the one click rotation only appears on one side of the relation. As the hypotheses of the main theorem above are quite general for initial triple and quadruple points, we expect that the obstructions obtained in this article will be very powerful in the classification of subfactors. 1.1 Acknowledgements The author would like to thank Zhengwei Liu, Scott Morrison, and Noah Snyder for helpful conversations. Most of this work was conceived and completed during a workshop on fusion categories held at the Institut de Math´ematiques de Bourgogne. The author would like to thank the organizers Peter Schauenburg and Siu-Hung Ng for their hospitality. The author was supported in part by the Natural Sciences and Engineering Research Council of Canada and DOD-DARPA grant HR0011-12- 1-0009. 2 Background We refer the reader to [Jon12, BMPS12] for the definition of a subfactor planar algebra. 3 2.1 Notation for planar algebras In this article, P• denotes a subfactor planar algebra of modulus [2] where and q ∈ (cid:110) exp (cid:16) 2πi 2j (cid:17)(cid:12)(cid:12)(cid:12)j ≥ 3 qk − q−k q − q−1 , [k] = (cid:111) ∪ [1,∞) such that [2] = q + q−1. We denote the Temperley-Lieb subfactor planar subalgebra by TL•. When we draw planar diagrams, we often suppress the external boundary disk. In this case, the external boundary is assumed to be a large rectangle whose distin- guished interval contains the upper left corner. We draw one string with a number next to it instead of drawing that number of parallel strings. Since the shading can be inferred from the distinguished interval, the number of strands, and whether an element is in Pk,±, we omit it completely. Finally, we usually draw elements of our planar algebras as rectangles with the same number of strands emanating from the top and bottom, and the distinguished interval is always on the left unless it is marked otherwise. We refer the reader to [GdlHJ89, BP13, JMS13] for the definition of the principal graphs (Γ+, Γ−) of P•. If there is only one projection P in the equivalence class [P ] corresponding to a vertex of Γ±, then we identify [P ] with P . For a projection P ∈ Pk,±, the dual projection P is given by k P = P . k For a vertex [P ] of Γ± at depth n, there is a corresponding dual vertex [P ] necessarily at depth n. If n is even then [P ] is a vertex of Γ+, but if n is odd, then [P ] is a vertex of Γ−. When we draw principal graph pairs, we use the convention that the vertical ordering of vertices at a given odd depth determines the duality; the lowest vertices in each graph at each odd depth are dual to each other, etc. When we specify the duality at even depths (sometimes we omit this data), the duality is represented by red arcs joining dual pairs of vertices. Self-dual even vertices have a small red dash above them. Example 2.1. The Haagerup subfactor [Haa94, AH99] has principal graphs (cid:16) (cid:17) . , 2.2 Supertransitivity and annular multiplicities We rapidly recall the notions of supertransitivity and annular multiplicities for subfactor planar algebras and potential principal graphs following [Jon01, Jon12, MPPS12, JMS13]. 4 Definition 2.2. A subfactor planar algebra P• is called k-supertransitive if Pj,+ = TLj,+ for all 0 ≤ j ≤ k. Equivalently, P• is k-supertransitive if the truncation Γ+(k) of Γ+ to depth k is Ak+1. This gives us the notion of the supertransitivity of a potential principal graph. Remark 2.3. When we say a subfactor planar algebra or potential principal graph Γ± is k-supertransitive, we usually also mean that TLk+1,+ (cid:40) Pk+1,+ or Γ±(k + 1) (cid:54)= Ak+2, although strictly speaking, this is an abuse of nomenclature. Every subfactor planar algebra decomposes into an orthogonal direct sum of ir- reducible annular Temperley-Lieb submodules. Each irreducible representation that appears in the direct sum is generated by a single low weight rotational eigenvector at some depth n, for which the rotational eigenvalue is an n-th root of unity [Jon01]. (Other eigenvalues are possible for 0-boxes which do not occur in subfactor planar algebras.) Hence Pn,± is the direct sum of the annular consequences ACn,± of Pn−1,± and the new low weight vectors at depth n. Definition 2.4. The sequence of annular multiplicities of a subfactor planar algebra is the sequence of multiplicities of lowest weight vectors, ignoring eigenvalues, i.e., an = dim(Pn,+) − dim(ACn,+). If P• is n − 1 supertransitive, then an is just called the multiplicity. Remark 2.5. A formula for the annular multiplicities in terms of the principal graph is given in [Jon01]. We give it here for the reader's convenience. First, dim(Pj,±) = Lj, the number of loops of length 2j on Γ± starting at (cid:63). We then have k(cid:88) j=0 ak = (−1)j−k 2k k + j (cid:19) (cid:18)k + j k − j Lj. Thus it makes sense to discuss the annular multiplicities of a subfactor planar algebra or of a candidate principal graph pair. The zeroth annular multiplicity a0 of a subfactor planar algebra is always 1, as the empty diagram generates TL• as an annular Temperley-Lieb module. If P• is k-supertransitive, then for all 1 ≤ j ≤ k, aj = 0, since Pj,± = TLj,±. For simplicity of notation, if P• is k-supertransitive, but not k + 1-supertransitive, we denote the initial chain (a0 = 1, a1 = 0, . . . , ak = 0) of annular multiplicities by an asterisk, and we begin the sequence after the asterisk with ak+1. Definition 2.6. We say Γ± is a translation of Γ0± if Γ± is obtained from Γ0± by increasing the supertransitivity. (If the even dual data of Γ0± is specified, we only allow even translations.) We say Γ± is an extension of Γ0± if Γ± is obtained from Γ0± by adding new vertices and edges at strictly greater depths than the maximum depth of any vertex in Γ0±. A weed is a graph pair W which represents an infinite family of possible principal graphs obtained from W by translation and extension. A vine is a graph pair V which represents an infinite family of possible principal graphs obtained from V by translation only. 5 Examples 2.7. (1) If P• has annular multiplicities ∗10, then Γ+ is a translated extension of (2) If P• has annular multiplicities ∗11, then Γ+ is a translated extension of or . or . Suppose P• is n − 1 supertransitive with multiplicity m. Then there are m new low weight rotational eigenvectors A1, . . . , Am ∈ Pn,+, and Pn,+ = TLn,+ ⊕ span{A1, . . . , Am}. Definition 2.8. The chiralities of P• are the rotational eigenvalues ωA1, . . . , ωAm corresponding to the new low weight rotational eigenvectors A1, . . . , Am respectively. Remarks 2.9. (1) The chirality of a subfactor planar algebra was first defined for some special cases in [Jon99, Definition 4.2.12] and subsequently studied in [Jon12]. (2) In this article, work with a (completely determined!) square root σA of the chirality ωA, and by a slight abuse of notation, we also call σA a chirality of our subfactor planar algebra. This abuse of notation was also used in [MP13]. 2.3 Liu's relation In this subsection, we derive a strong quadratic relation due to Liu, which is a clever variant of Wenzl's relation. Let P• be a subfactor planar algebra with prin- cipal graphs (Γ+, Γ−). We begin by stating Wenzl's relation. We will then state a generalization from which Liu's relation quickly follows. Lemma 2.10 (Wenzl's relation [Wen87]). Let f (k) ∈ TLk,± be the k-th Jones-Wenzl projection. Then k + 1 f (k+1) = k + 1 f (k) − [k] [k + 1] k k k f (k) k − 1 f (k) k . Below is a more general version of Wenzl's relation known to experts (e.g., see [MS12a, Lemma 3.7]). 6 Lemma 2.11 (Generalized Wenzl's relation). Suppose that P ∈ Pk,± is a projection corresponding to the vertex [P ] at depth k of Γ±. Suppose that f ∈ Pk−1,± is a projection corresponding to the vertex [f ] at depth k − 1 of Γ±. Suppose [P ] and [f ] are connected by a single edge, and [f ] is the only vertex at depth k − 1 attached to [P ]. Then the unique projection isomorphic to f under the image of P in Pk+1,± is given by k P k − 1 P Tr(f ) Tr(P ) . Hence the "new stuff" [GdlHJ89] P (cid:48) connected to P at depth k + 1 is given by k k + 1 P (cid:48) k + 1 = k P k − Tr(f ) Tr(P ) k P k − 1 P k . We now prove Liu's relation from this generalization of Wenzl's relation. Suppose that P• is (n− 1) supertransitive for some n ≥ 2, such that Γ+ is simply laced when truncated to depth n, i.e., Pn,+ (cid:9) TLn,+ is abelian. Lemma 2.12 (Liu's relation). Let P be a projection at depth n of Γ+, and let P (cid:48) be as in Lemma 2.11. Then F(P )∗F(P ) = n − 1 P P n − 1 = Tr(P )2 [n]2 n f (n) + n [n − 1] [n] n − 1 f (n−1) f (n−1) n − 1   − Tr(P ) [n] n P (cid:48) n . Proof. Start with P , apply the generalization of Wenzl's relation, rotate by 180◦, and then apply a cap on the right to see that  . n − 1 P − n − 1 n P (cid:48) n n − 1 P P n − 1 = Tr(P ) [n]  7 Liu's observation is that the left hand side is equal to F(P )∗F(P ), where F = n − 1 n − 1 is the one click rotation. The rest is a straightforward calculation, since capping P gives a multiple of f (n−1) (taking traces gives the multiple). Note that we use Wenzl's relation again for f (n−1) with a strand on the left to get the final formula. Remark 2.13. When the truncation Γ±(n + 1) of Γ± to depth n + 1 is simply laced and acyclic, which ensures that we can determine the rightmost term in Liu's relation (see Remark 3.4), we can use Liu's relation to determine the modulus of the entries of the one click rotation at depth n. 3 Multiplicity 1 - triple points When Γ+ is (n − 1) supertransitive for some n ≥ 2 and has multiplicity 1, we can identify the rotational low weight eigenvector. We use the following conventions: Tr(Q) Tr(P ) is the branch factor, • designate the projections P, Q and P , Q at depth n of Γ+ and Γ− respectively, • if n is odd, then choose so that P = P and Q = Q, • r = • A = rP − Q ∈ Pn,+ is a low-weight rotational eigenvector with eigenvalue ωA, • since P + Q = f (n), we have P = • A2 = (r − 1)A + rf (n), • the previous 4 lines have analogous formulas with checks, and • F(A) = A for a determined σ2 f (n) + A 1 + r and Q = rf (n) − A , 1 + r √ r√ r σA A = ωA. Note 3.1. By designating P, P , we have designated A, A, which completely deter- mines σA. Definition 3.2. Since P only connects to f (n−1) at depth n − 1, we may define P (cid:48) as in Lemma 2.11. Let ∩n+1(P (cid:48)) denote P (cid:48) capped off on the right as in Lemma 2.12: n ∩n+1(P (cid:48)) = . P (cid:48) n 8 Since capping ∩n+1(P (cid:48)) on the top or bottom gives zero, we must have that ∩n+1(P (cid:48)) ∈ span{A0, f (n)} ⊂ Pn,±, where (cid:40) A if n is even A if n is odd. A0 = We denote the coefficient of A0 in ∩n+1(P (cid:48)) by coeff ∈∩n+1(P (cid:48)) (A0) . Theorem 3.3. If n is even and P = P , then (r − 1) r r − (σA + σ−1 A ) [n] √ r√ r = −(1 + r) [n + 1] [n] coeff ∈∩n+1(P (cid:48)) If n is even and P = Q, then r = 1, and (r − 1) 1 r + (σA + σ−1 √ A ) r [n] = −2 [n + 1] [n] If n is odd, then r = r, and (cid:18) (cid:18) (cid:0) A(cid:1)(cid:19) (cid:0) A(cid:1)(cid:19) . coeff ∈∩n+1(P (cid:48)) (cid:18) (cid:19) . (E) (E) (O) (r − 1) − (σA + σ−1 A ) [n] = −(1 + r) [n + 1] [n] coeff ∈∩n+1(P (cid:48)) (A) . Proof. To prove Equations (E) and (E), we find the coefficient of A in both sides of Liu's relation, noting that A is orthogonal to TLn,−. Suppose P = P . Then since P = , the left hand side in Liu's relation is equal to f (n) + A 1 + r F(P )∗F(P ) = 1 (1 + r)2F(f (n) + A)∗F(f (n) + A)  A2 + r r n − 1 A f (n) n − 1 n − 1 f (n) + + A n − 1  . n − 1 f (n) f (n) n − 1 = 1 (1 + r)2 The fourth diagram is in TLn,− and does not contribute to the coefficient of A. Note that the only diagram in the f (n) which contributes to the second, respectively third, summand is n − 2 n − 2 , both of which have coefficient (−1)n−1/[n] [Mor, Rez07]. Hence the coefficient of A is , respectively (cid:18) 1 (1 + r)2 (cid:19) . √ r√ r (r − 1) r r + (σA + σ−1 A ) (−1)n−1 [n] 9 Since the first two diagrams on the right hand side of Liu's relation are in TLn,−, Equation (E) follows immediately using n is even and Tr(P ) = . When P = Q, a similar argument shows the coefficient of A in F(Q)∗F(Q) is [n + 1] 1 + r given by 1 (1 + r)2 (r − 1) − r(σA + σ−1 A ) r r (−1)n−1 [n] Now use r = 1 and n is even, and substitute for Tr(P ). For Equation (O), we find the coefficient of A in F(P )∗F(P ) = F( P )∗F( P ). A (cid:18) (cid:18) (cid:19) . √ r√ r (cid:19) . similar argument as before shows this coefficient is given by √ (−1)n−1 r√ r + (σA + σ−1 A ) (r − 1) (1 + r)2 r r [n] 1 Now use r = r and n is odd, and substitute for Tr(P ). Remark 3.4. To get useful formulas from Theorem 3.3, we need to be able to determine ∩n+1(P (cid:48)). This is always possible under the following condition: • For each R at depth n + 1 of Γ+, let mR be the number of edges connecting P and R. For each R with mR > 0, suppose there is a unique vertex E(R) at depth n connected to R. In this case, coeff ∈∩n+1(P (cid:48)) (A0) = mR Tr(R) Tr(E(R)) coeff ∈E(R) (A0) . (cid:88) R This very general condition is satisfied by most principal graphs of small index, after choosing the appropriate P . 3.1 Triple point obstructions We now recover all known triple point obstructions from Theorem 3.3, and we give a new annular multiplicities ∗11 obstruction. We start with Ocneanu's triple point obstruction for an initial triple point. The hypotheses of Theorem 3.5 are equivalent to those for Ocneanu's triple point ob- struction [Haa94] (see [MPPS12, Theorems 3.5 and 3.6] for more details). Our proposition actually determines the chirality, which is stronger than Ocneanu's re- sult, but only works for an initial triple point, which is weaker than Ocneanu's result. Theorem 3.5. Suppose that • if n is even, then r = r, and for every vertex R connected to P , R only connects to P if P = P or Q if P = Q, or • if n is odd, then for every vertex R connected to P , R only connects to P . 10 Then σA + σ−1 Proof. Both conditions for n even and n odd imply that r = r and A = [n + 2] − [n] = qn+1 + q−n−1. Tr(P (cid:48)) = [2] Tr(P ) − [n] = [2] − [n] = [n + 2] − r[n] . 1 + r Under these hypotheses, all three equations given in Theorem 3.3 specialize to give σA + σ−1 A = [n + 2] − [n]. We include the proof below for convenience. For n even and P = P , ∩n+1(P (cid:48)) = P , and Equation (E) rearranges and (cid:18)[n + 1] (cid:19) 1 + r Tr(P (cid:48)) Tr(P ) simplifies to give (cid:18) σA + σ−1 A = [n] (r − 1) + (1 + r) [n] = (r − 1)[n] + (1 + r)[n + 1] = (r − 1)[n] + (1 + r) = [n + 2] − [n]. (cid:0) A(cid:1)(cid:19)(cid:19) (cid:19) 1 1 + r [n + 1] (cid:18) (cid:18)Tr(P (cid:48)) (cid:18)[n + 2] − r[n] Tr(P ) · (cid:19) coeff ∈∩n+1(P (cid:48)) 1 + r For n even and P = Q, Tr(P ) = Tr(Q), so r = r = 1, ∩n+1(P (cid:48)) = Equation (E) rearranges and simplifies to give σA + σ−1 A = −2[n + 1] = −2[n + 1] coeff ∈∩n+1(P (cid:48)) (cid:0) A(cid:1)(cid:19) (cid:18) Tr(P (cid:48)) Tr(P ) Q, and = [n + 2] − [n]. (cid:19) (cid:18)Tr(P (cid:48)) Tr(P ) · −1 1 + r Finally, For n odd, ∩n+1(P (cid:48)) = Tr(P (cid:48)) Tr(P ) P , and Equation (O) rearranges and simplifies exactly as Equation (E) did for the case n even and P = P . Corollary 3.6 (Ocneanu [Haa94]). Under the conditions of Theorem 3.5, [2] ≤ 2. Proof. qn+1 + q−n−1 = [n + 2]− [n] = σA + σ−1 Remark 3.7. After suitably labeling the projections at depth n, every irreducible subfactor with index at most 4 satisfies the conditions of Theorem 3.5. A table of the possible chiralities for these subfactors is given in Example 3.12, as is a proof that E7 and D2k−1 (3 ≤ k < ∞) are not principal graphs of subfactors. Interestingly, all possible chiralities actually occur. A ∈ [−2, 2] only if q + q−1 = [2] ≤ 2. We now recover [Sny12, Theorem 3] which generalizes [Jon12, Theorem 5.1.11]. Corollary 3.8. Suppose [2] > 2, (Γ+, Γ−) has multiplicity 1, and P is singly valent. Then n is even, r = [n + 2]/[n], and r + 1 r = 2 + ωA + ω−1 A + 2 [n][n + 2] . Moreover, if ωA is a primitive k-th root of unity, then 2k n. 11 Proof. Let P be the singly valent vertex. If n is odd, then the hypotheses of Theorem 3.5 are satisfied, so Corollary 3.6 contradicts [2] > 2. Hence n is even, P = P (since Tr(P ) (cid:54)= Tr(Q)), and r = [n + 2]/[n]. Since P (cid:48) = 0, the right hand side in Equation (E) is zero, and the equation can be rearranged to give A =(cid:112)[n + 2][n] (cid:18)√ σA + σ−1 (cid:19) . r − 1√ r Squaring both sides and rearranging yields the formula. For the last statement, consider the 2-click rotation ρ on Pn,+. Since n = 2m for some m, we have ρm is a trace-preserving map on span{P, Q}. Since Tr(P ) (cid:54)= Tr(Q), A = 1, so k m, and thus ρm must be the identity, so ρm fixes S. In particular, ωm 2k n. Remark 3.9. Note that the eigenvalue statement in Corollary 3.8 is essentially in [Jon12, Theorem 5.1.11]. Theorem 3.10. Suppose (Γ+, Γ−) has annular multiplicities ∗11, and Γ± does not have a singly valent vertex at depth n, i.e., Γ+, Γ− are both translated extensions of (1) If n is even, (Γ+, Γ−) is a translated extension of . (cid:16) (cid:17) , and (cid:17) , , . , (r − 1) r r − (σA + σ−1 A ) [n] √ r√ r = r[n] − [n + 2] [n] (cid:16) (2) If n is odd, (Γ+, Γ−) is a translated extension of and (r − 1) − (σA + σ−1 A ) [n] = [n + 2] − r[n] r[n] . Proof. First, the principal graphs begin as claimed by [MS12b, Lemmas 6.3 and 6.4]. Let P be the bottom vertex at depth n on Γ+, and let P (cid:48) be the bottom vertex at depth n + 1 on Γ+. Let P be the bottom vertex at depth n on Γ−. We record the following traces: [n + 1] 1 + r [n + 1] 1 + r Tr(P ) = and Tr(Q) = Tr( P ) = and Tr( Q) = Tr(P (cid:48)) = [2] Tr(P ) − [n] = [2] r[n + 1] 1 + r r[n + 1] 1 + r (cid:18)[n + 1] (cid:19) 1 + r 12 provided n is even, and [n + 2] − r[n] − [n] = . 1 + r Suppose n is even. Since P = P and E(P (cid:48)) = P , the right hand side of Equation (E) is given by (cid:18)[n + 1] (cid:18)[n + 1] [n] (cid:19)(cid:18)Tr(P (cid:48)) (cid:19)(cid:18) (cid:19)(cid:18)[n + 2] − r[n] Tr( P ) coeff ∈ P (cid:0) A(cid:1)(cid:19) · 1 + r [n + 1] 1 + r (cid:19)(cid:18) 1 (cid:19) 1 + r RHS = −(1 + r) = −(1 + r) r[n] − [n + 2] = [n] [n] . RHS = −(1 + r) = −(1 + r) [n + 2] − r[n] = r[n] [n] . When n is odd, E(P (cid:48)) = Q, and the right hand side of Equation (O) is given by (cid:18)[n + 1] (cid:18)[n + 1] [n] (cid:19)(cid:18) (cid:19)(cid:18)Tr(P (cid:48)) (cid:19)(cid:18)[n + 2] − r[n] Tr(Q) coeff∈Q (A) · 1 + r r[n + 1] 1 + r (cid:19)(cid:18) −1 (cid:19) 1 + r (cid:19) Remark 3.11. Snyder can produce the equations in Theorem 3.10 using his tech- nique in [Sny12], but his technique does not generalize beyond the ∗11 case. 3.2 Finding chiralities of some examples Examples 3.12 (Index at most 4). We now use Theorem 3.5 to determine the chiralities of all non type-A irreducible subfactors with index at most 4 with an initial triple point, and we show E7 and D2k+1 are not principal graphs of subfactors. Note that the computation for index less than 4 was done previously in [Jon99, Theorem 4.2.13] using a different method. Jones' proof of the nonexistence of D2k−1 and E7 also shows the chirality is inconsistent with the supertransitivity, but he uses a different formula. In the table below, 4 ≤ k and 3 ≤ j ≤ ∞. σA + σ−1 Graph n k − 2 exp(cid:0) 2πi exp(cid:0) 2πi exp(cid:0) 2πi exp(cid:0) 2πi q 4k−4 (cid:1) (cid:1) (cid:1) (cid:1) 36 24 60 Dk E6 E7 E8 D(1) j+2 E(1) 6 E(1) 7 E(1) 8 3 4 5 2 3 4 6 A 0 1 2 sin(cid:0) 2π √ 9 (cid:1) 5) 1 2(1 + 2 2 2 2 possible σA ±i exp(cid:0)± 2πi exp(cid:0)± 5πi exp(cid:0)± 2πi 18 6 (cid:1) (cid:1) (cid:1) 10 ((cid:63) k odd) possible ωA −1 exp(cid:0)± 2πi exp(cid:0)± 5πi (cid:1) exp(cid:0)± 2πi (cid:1) (cid:1) 3 9 5 (cid:63) 1 1 1 1 1 1 1 1 1 1 1 1 The (cid:63)'s above denote contradictions. 13 • For Dk with k odd, −1 is not a (k − 2)-th root of unity. • For E7, exp(cid:0)± 5πi 9 (cid:1) is not a 4-th root of unity. (cid:17) (cid:16) , . Example 3.13 (Fuss-Catalan). For generic a, b > 2, the principal graphs of the Bisch-Jones Fuss-Catalan subfactor planar algebras [BJ97] are extensions of For these planar algebras, ωA = 1 [Jon12, Example 3.5.9], as can be verified by our formula. The traces of the projections labelled as in the proof of Theorem 3.10 are given by Tr(P ) = a2 − 1 Tr( P ) = b2(a2 − 1) Tr(P (cid:48)) = b(a3 − 2a). Tr(Q) = a2(b2 − 1) Tr( Q) = b2 − 1 Hence the branch factors are given by a2(b2 − 1) a2 − 1 r = and r = b2 − 1 b2(a2 − 1) =⇒ r r = a2b2. Noting that [n] = [2] = ab and [n + 2] = [4] = ab(−2 + ab2), we substitute these values into (1) from Theorem 3.10 which gives −(σA + σ−1 (cid:19) √ [n] A ) = − (r − 1) r√ r ab − ab(−2 + a2b2) r r ab + 2 − a2b2 − (cid:18) b2 − 1 (cid:19) b2(a2 − 1) − (cid:18) a2(b2 − 1) a2 − 1 + a2b2 (cid:19) − 1 a2b2 (cid:19) (cid:18) 1 ab ab [n] (cid:18) r[n] − [n + 2] (cid:17) (cid:16) a2(b2−1)  (cid:19) (cid:18)a2(b2 − 1) a2−1 a2 − 1 = = = 2. Hence σA = −1, and ωA = 1 as claimed. Example 3.14 (Uq(su(3))). The principal graphs of the Uq(su(3)) subfactors [Wen88] are extensions of (cid:16) For these planar algebras, ωA = 1, which follows from Kuperberg's A2 spider [Kup96], where the generator at depth 3 is a hexagon with legs alternating in and out. Hence it has a Z/3 symmetry. The author would like to thank Scott Morrison for pointing this out. We verify this with our formula for two Uq(su(3)) subfactors: (cid:17) . (cid:17) (cid:16) (cid:16) Q1 = Q2 = , , , 14 (cid:17) . (cid:16) For Q1, the branch factor is the root of x3 − 4x2 + 3x + 1 which is approximately 1.45, and q is the root of x6 − 2x5 + 2x4 − 3x3 + 2x2 − 2x + 1 which is approximately 2 − 1 1.636. For Q2, the branch factor is (2 + 1 + . 2 + A = −2. Using these values in (2) from Theorem 3.10, we get σA + σ−1 (cid:17) Proposition 3.15 (2D2). The 2D2 subfactor planar algebra with principal graphs 2)/2 and q = 1 2 (cid:112) √ 2 (cid:16) (cid:17) √ √ , has chirality ωA = 1. Proof. We work with the dual subfactor planar algebra with principal graphs (cid:17) . 1, P (cid:48) Let P be the bottom vertex at depth 3, and let P (cid:48) 2 be the bottom two vertices 2. The traces of P, P (cid:48) at depth 4, where P (cid:48) 1, P (cid:48) 1 is below P (cid:48) 2 are given by (cid:16) , 1) = √ 5, Tr(P (cid:48) 7 + 3 (cid:113) (cid:18)[n + 1] (cid:19)(cid:18)(cid:18)Tr(P (cid:48) [n] (cid:19)(cid:18)(cid:18)Tr(P (cid:48) (cid:19)(cid:18)1 1) Tr(P ) 1) Tr(P ) 2 1 2 (cid:19) (cid:19) [n] Tr(P ) = Tr(Q) = RHS = −(1 + r) (cid:18)[n + 1] = −2 The right hand side in Equation (O) is given by √ (1 + 5), and Tr(P (cid:48) 2) = √ 5). 1 2 (3 + (cid:19) (cid:18)Tr(P (cid:48) (cid:19)(cid:18)1 (cid:19)(cid:19) 2) Tr(Q) 2 (A) + (cid:18)Tr(P (cid:48) coeff∈P − 2) Tr(Q) (cid:19) coeff∈Q (A) = 2 [n] . Since the left hand side is −σA + σ−1 A [n] , we have σA + σ−1 A = −2, so ωA = 1. Proposition 3.16. The 2D2 subfactor planar algebra is singly generated at depth 3. More precisely, any 3-box not in TL3,± generates all of 2D2. Proof. Since 2D2 has annular multiplicities ∗12, if A generated a proper planar subalgebra, then the planar subalgebra would have either annular multiplicities ∗10 or ∗11. However, r = r = 1, which contradicts Corollary 3.8 and Theorem 3.10, since we know [2] = (cid:112) 5 > 2. 3 + √ 3.3 Eliminating a ∗11 weed above index 5 The following weed with annular multiplicities ∗11 appears when running the prin- cipal graph odometer [MS12b] above index 5: (cid:16) W = , 15 (cid:17) . Note that W begins as claimed in Theorem 3.10, after applying a suitable graph automorphism. We now eliminate W using Theorem 3.10 and the technique of [MPPS12, Sec- tion 4.1]. First, we determine the relative dimensions of the vertices as functions of n and q, where q > 1 such that [2] = q + q−1. Second, we calculate the relative branch factors, which are the expressions for r, r as functions of n, q. Fi- A + 2 ∈ [0, 4] as a function of n, q, and nally, we use Theorem 3.10 to write σA + σ−1 we show that this value is always negative, a contradiction. Theorem 3.17. There is no subfactor with principal graphs a translated extension of W. Proof. Suppose we had such a subfactor, and note that n ≥ 3 and d ≥ 2.27453, so q ≥ 1.6789. Let P be the 2-valent vertex on Γ+ at depth n, and let P be the bottom vertex on Γ− at depth n. The formula in (2) from Theorem 3.10 rearranges to give σA + σ−1 A = r[n] − [n + 2] r , (1) and the relative branch factors are given by r = r = (q2 + 1) (q2n+8 + 3q2n+10 + 2q2n+12 + 2q2n+14 − q6 − 3q4 − 2q2 − 2) 2q2n+8 + 4q2n+10 + 4q2n+12 + 4q2n+14 + q2n+16 − 2q8 − 4q6 − 4q4 − 4q2 − 1 . We then have σA + σ−1 A + 2 = r[n] − [n + 2] r + 2 = g(n, q)h(n, q) k(n, q) , where g, h, k are given by + (q9 + 2q8 + 4q7 + 4q6 + 5q5 + 4q4 + 4q3 + 4q2 + 2q + 1) g(n, q) =a3(cid:0)q18 + 2q17 + 4q16 + 4q15 + 4q14 + 5q13 + 4q12 + 4q11 + 2q10 + q9(cid:1) + a2(cid:0)−2q17 − q16 − 4q15 − 4q14 − 5q13 − 4q12 − 4q11 − 4q10 − q9 − 2q8(cid:1) + a(cid:0)−2q10 − q9 − 4q8 − 4q7 − 4q6 − 5q5 − 4q4 − 4q3 − q2 − 2q(cid:1) h(n, q) =a3(cid:0)q18 − 2q17 + 4q16 − 4q15 + 4q14 − 5q13 + 4q12 − 4q11 + 2q10 − q9(cid:1) + a2(cid:0)−2q17 + q16 − 4q15 + 4q14 − 5q13 + 4q12 − 4q11 + 4q10 − q9 + 2q8(cid:1) + a(cid:0)−2q10 + q9 − 4q8 + 4q7 − 4q6 + 5q5 − 4q4 + 4q3 − q2 + 2q(cid:1) k(n, q) = − qn+1(q − 1)(q + 1)(cid:0)q2 + 1(cid:1) ×(cid:0)a2(2q14 + 2q12 + 3q10 + q8) − q6 − 3q4 − 2q2 − 2(cid:1) ×(cid:0)a2(q16 + 4q14 + 4q12 + 4q10 + 2q8) − 2q8 − 4q6 − 4q4 − 4q2 − 1(cid:1) + (q9 − 2q8 + 4q7 − 4q6 + 5q5 − 4q4 + 4q3 − 4q2 + 2q − 1) using the shorthand a = qn. Note that k(n, q) is always negative. Claim. For all n ≥ 0 and q ≥ 1.6789, g(n, q) ≥ 0. 16 Proof of Claim. We break up g(n, q) into four parts: g(n, q) =(cid:80)3 j=0 ajgj(q) where g3(q) = q18 + 2q17 + 4q16 + 4q15 + 4q14 + 5q13 + 4q12 + 4q11 + 2q10 + q9 g2(q) = −2q17 − q16 − 4q15 − 4q14 − 5q13 − 4q12 − 4q11 − 4q10 − q9 − 2q8 g1(q) = −2q10 − q9 − 4q8 − 4q7 − 4q6 − 5q5 − 4q4 − 4q3 − q2 − 2q g0(q) = q9 + 2q8 + 4q7 + 4q6 + 5q5 + 4q4 + 4q3 + 4q2 + 2q + 1 One checks that for q ≥ 1.6789, 0 < g0(q) 0 < g3(q) + g2(q) + g1(q) 0 < g3(q) + g2(q) and 0 < g3(q). Hence for a = qn with q ≥ 1.6789 and n ≥ 0, we have 0 < g3(q) + g2(q) + g1(q) + g0(q) ≤ a(g3(q) + g2(q) + g1(q)) + g0(q) ≤ a(a(g3(q) + g2(q)) + g1(q)) + g0(q) ≤ a(a(ag3(q) + g2(q)) + g1(q)) + g0(q) = g(n, q). Claim. For n ≥ 2 and q ≥ 1.6789, h(n, q) ≥ 0. Proof of Claim. Since n ≥ 2, we replace n with m = n + 2. We break up h(m + 2, q) into four parts: h(m + 2, q) =(cid:80)3 j=0 bjhj(q) where h3(q) = q24 − 2q23 + 4q22 − 4q21 + 4q20 − 5q19 + 4q18 − 4q17 + 2q16 − q15 h2(q) = −2q21 + q20 − 4q19 + 4q18 − 5q17 + 4q16 − 4q15 + 4q14 − q13 + 2q12 h1(q) = −2q12 + q11 − 4q10 + 4q9 − 4q8 + 5q7 − 4q6 + 4q5 − q4 + 2q3 h0(q) = q9 − 2q8 + 4q7 − 4q6 + 5q5 − 4q4 + 4q3 − 4q2 + 2q − 1 and b = qm. The rest is identical to the previous claim replacing g's with h's, a with b, and n with m (except that the final conclusion is that h(m + 2, q) > 0). < 0 for n ≥ 2, a contradiction. Thus 0 ≤ σA + σ−1 g(n, q)h(n, q) A + 2 = k(n, q) 4 Multiplicity 2 - some quadruple points Suppose (Γ+, Γ−) has multiplicity 2, n ≥ 2 is even, and two of the vertices at depth n of Γ± are dual to each other, i.e., (Γ+, Γ−) is an even translated extension of (cid:16) Q = , 17 (cid:17) . For examples of subfactors with such principal graphs, see Examples 4.4. We use the following conventions: • P, Q, R and P , Q, R are the projections at depth n of Γ± from bottom to top, • r = • A = rP−(Q+R) and B = Q−R ∈ Pn,+ are low-weight rotational eigenvectors 2 Tr(Q) Tr(P ) , with distinct eigenvalues ωA, ωB (by Proposition 4.1 below), • row reducing • Substituting the formulas for P, Q, R yields 1 1 1 Q R 2(1 + r) B f (n) 0 −1 1 1  r −1 −1 A  yields   = P r2P + Q + R  =  =  A2 (cid:112)r(1 + r) (cid:112)r(1 + r) A and F(B) = R − Q Q + R AB B2 2A + 2f (n) −A − (1 + r)B + rf (n) −A + (1 + r)B + rf (n)  , (r − 1)A + rf (n) −B −1 1+r A + r  , σB B (also by Proposition 4.1). 1+r f (n) • the previous 4 lines have analogous formulas with checks, and • F(A) = √ r√ r σA The next result is a modified version of [IJMS12, Lemma 2.1]. We provide a short proof for convenience and completeness. Proposition 4.1. The elements A = rP −(Q+R) and B = Q−R are rotational low B = −1. The same holds with checks, weight eigenvectors such that ωn/2 and thus there are completely determined σA, σB such that σ2 B = ωB, and the above formulas hold for F(A),F(B). Proof. First, note that span{A, B} = TL⊥ and have trace zero. Since P is self-dual and Q is dual to R, n,+, since A, B are nonzero, orthogonal, A = 1 and ωn/2 A = ωA, σ2 ρn/2(A) = ρn/2(rP − (Q + R)) = rP − (Q + R) = A and ρn/2(B) = ρn/2(Q − R) = R − Q = −B. Thus {A, B} is a basis of eigenvectors for ρn/2 on TL⊥ eigenvectors for ρ on TL⊥ to scaling, we must have v1 = A and v2 = B. The same holds with checks. n,+. Let {v1, v2} be a basis of n,+. Then {v1, v2} is a basis of eigenvectors for ρn/2, so up Since we know ωA (cid:54)= ωB, we know F(A) = σAλA A for some real scalar λA and a completely determined σA, and similar for B. Taking the norm squared of each side gives the value of λA, λB. 18 Since P, Q only connect to f (n−1) at depth n − 1, we may define P (cid:48), Q(cid:48) and ∩n+1(P (cid:48)),∩n+1(Q(cid:48)) as in Definition 3.2 applied to P, Q respectively. Theorem 4.2. We have the following equations: (cid:18) (r − 1) (r − r − 2) + (cid:18) (σAσ−1 B + σ−1 √ r√ r − (σA + σ−1 (cid:19) r A ) r r [n] √ (σA + σ−1 A ) A σB) − (σB + σ−1 [n] rr r B ) [n] (cid:18) coeff ∈∩n+1(P (cid:48)) (cid:18) = − (1 + r) [n + 1] [n] = − 2r(1 + r) (cid:17)(cid:16)r (cid:17)√ √ rr r [n + 1] coeff ∈∩n+1(Q(cid:48)) [n] √ 1 + r 1 + r = − 2r(1 + r) [n + 1] [n] coeff ∈∩n+1(Q(cid:48)) (cid:18) (cid:0) A(cid:1)(cid:19) (cid:0) A(cid:1)(cid:19) (cid:0) B(cid:1)(cid:19) (QA1) (QA2) (QB) Proof. We find the coefficient of A, B in both sides of Liu's relation, noting that A, B are orthogonal to TLn,− and are orthogonal to each other. , the proof of Equation (QA1) is identical to the proof of Since P = f (n) + A 1 + r Equation (E) in Theorem 3.3. For Equations (QA2) and (QB), we note that Q = R = rf (n) − A + (1 + r)B 2(1 + r) , so expanding the left hand side of Liu's relation applied to Q gives F(R)∗F(R) = 1 4(1 + r)2F(rf (n) − A + (1 + r)B)∗F(rf (n) − A + (1 + r)B). This time there are 9 terms, 8 of which contribute some multiple of A or B. These terms are as follows, where TA, TB ∈ TLn,−, and we omit the leading coefficient of 4(1 + r)2 , which we will move to the other side of the equation: 1 n − 1 A A n − 1 n − 1 B B n − 1 (1 + r)2 (cid:18) r(r − 1) (cid:19) r A + TA = r r A2 = (cid:18) −r(1 + r) r (cid:19) A + TB = (1 + r)2 r(1 + r) r(1 + r) B2 = 19 (cid:18) −r(−1)n−1 [n] √ r√ r (cid:19) (σA + σ−1 A ) A (cid:112)r(1 + r) (cid:112)r(1 + r) (cid:33) B (σB + σ−1 B ) r(1 + r)(−1)n−1 [n]    n − 1 A f (n) n − 1 n − 1 B f (n) n − 1 n − 1 A B n − 1 f (n) A n − 1 n − 1 f (n) B n − 1 n − 1 B A + + + n − 1 n − 1  =  =  = −r r(1+r) −(1 + r) (cid:32) (cid:32) (cid:112)r(1 + r) (cid:112)r(1 + r) √ r√ r (cid:33) B. (σAσ−1 B + σ−1 A σB) (1 + r) Now collect the coefficients of A and B and substitute for Tr(Q). Remarks 4.3. (1) Once again, we need criteria similar to the criterion given in Remark 3.4 to determine ∩n+1(P (cid:48)) and ∩n+1(Q(cid:48)). (2) In specific examples, one can first use Equations (QA1) and (QA2) to com- pute σA, and then use Equation (QB) to compute σB. We will compute some examples in the next subsection. (3) Having two equations for σA + σ−1 A should give a strong constraint for translated extensions of Q. We will see in Subsection 4.2 that we only need Equation (QA1) to eliminate two such weeds. 4.1 Checking formulas on group-like subfactors at index 6 We now check Equations (QA1), (QA2), and (QB) on some group-like subfactors at index 6. For each of these examples n = 2, so we know the chiralities are ωA = 1 and ωB = −1 by Proposition 4.1. Examples 4.4. There are several group-like subfactors at index 6 which are trans- lated extensions of Q. Details on computing these principal graphs can be found in [BH96, KS00, MPPS13]. 20 First, we consider group-subgroup subfactors at index 6, which can be classified by subgroups G of S6 which act transitively on {1, 2, 3, 4, 5, 6}. The subgroup H of G is the point stabilizer Stab(1) of the action. We consider the following subgroups of S6 which act transitively, where the notation for the groups comes from [Pfe05]. • A4a = (cid:104)(123)(456), (135)(246)(cid:105) ⊂ S6 yields the principal graphs • S3×3a = (cid:104)(123)(456), (14)(26)(35), (142635)(cid:105) ⊂ S6 yields the principal graphs • A4 × 2a = (cid:104)(123)(456), (135)(246), (123456)(cid:105) ⊂ S6 yields the principal graphs We also consider the Bisch-Haagerup subfactors RH ⊂ R (cid:111) K for H = Z/2 and K = Z/3, where as usual, we let G = (cid:104)H, K(cid:105) ⊂ Out(R) [BH96]. We consider the following triples (G, H, K): • H = (cid:104)(12)(34)(cid:105), K = (cid:104)(123)(cid:105), and G = A4 yields the principal graphs • H = (cid:104)(34)(cid:105), K = (cid:104)(123)(cid:105), and G = S4 yields the principal graphs • H = (cid:104)(15)(24)(cid:105), K = (cid:104)(123)(cid:105), and G = A5 yields the principal graphs (cid:16) (cid:17) . In fact, just one calculation suffices to determine the chiralities for Examples 4.4 via our equations. If necessary, we pass to the dual subfactor so we may always assume P is the univalent self-dual vertex at depth 2, i.e., the principal graphs are (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:17) . (cid:17) . (cid:17) . (cid:17) . (cid:17) . (cid:17) (cid:17) . , , , , , , , , 21 or they begin like 222For these cases, P (cid:48) = 0 and ∩n+1(Q(cid:48)) = Tr(Q(cid:48)) Tr( P ) P . We have the following traces: Tr(P ) = 1 Tr( P ) = 3 Tr(Q) = Tr(R) = 2 Tr( Q) = Tr( R) = 1 √ Tr(Q(cid:48)) = Tr(R(cid:48)) = 6. Thus r = 4 and r = 2 3, and Equations (QA1), (QA2), and (QB) are given respec- (cid:19) 12 2 − 4 − 2 − 1 tively by(cid:18)2 (cid:32)(cid:18)2 (cid:32) 3 3 (σAσ−1 B + σ−1 Hence σA + σ−1 ωB = −1. + (cid:33) 6 √ (cid:32)√ − (σA + σ−1 √ (cid:19) A ) (σA + σ−1 A ) √ 3√ 2 = 0, √ 2 2√ 3 A σB) − (σB + σ−1 A = −2, so σA = −1 and ωA = 1, and σB + σ−1 (cid:33)(cid:18)12 12 2 √ 2√ 3 = −2(4)(5) (cid:19)(cid:114) 5√ 6 √ 6 √ 2 3 B ) 2 2 6 2 1 + 5 = 0. (cid:33)(cid:18) 1 (cid:19) 1 + 2 3 , and 6 3 B = 0, so σB = ±i and 4.2 Quadruple point obstructions We now prove a result similar to Theorem 3.5 for even translated extensions of Q. Theorem 4.5. Suppose that (Γ+, Γ−) is a translated extension of Q such that • r = r, and for every vertex R connected to P , R only connects to P . A = [n + 2] − [n] = qn+1 + q−n−1, and thus [2] ≤ 2. Then σA + σ−1 Proof. Just as for Theorem 3.5, we have Tr(P (cid:48)) = [2] Tr(P ) − [n] = [2] (cid:18)[n + 1] (cid:19) 1 + r − [n] = [n + 2] − r[n] . 1 + r Since Equation (QA1) is identical to Equation (E), the same proof for the case n A = [n + 2] − [n], and thus [2] ≤ 2 even with P = P for Theorem 3.5 shows σA + σ−1 as in Corollary 3.6. Definition 4.6. We define the following weeds with initial quadruple points which occur when running the principal graph odometer [MS12b] above index 5: (cid:32) (cid:32) Q1 = Q2 = , , (cid:33) (cid:33) , where we use the convention that if a weed has shaded and unshaded vertices, then only the unshaded vertices may connect to a vertex at depth n + 1. 22 (cid:16) (cid:16) (cid:16) (cid:17) . (cid:17) , , , , , Corollary 4.7. Any subfactor whose principal graphs are an even translated exten- sion of Q1 or Q2 is the Z/4 group subfactor with principal graphs Proof. Note that Q1,Q2 both satisfy the conditions of Theorem 4.5. By Theorem 4.5, for both Q1,Q2, [2] ≤ 2, and any translated extension of Q1,Q2 must be the trivial translated extension. Remark 4.8. In [MS12b] Morrison-Snyder showed that the principal graph pair of any non Temperley-Lieb subfactor planar algebra in the index range (4, 5) was either a translated extension of one of 5 weeds or a translation of one of 39 vines, up to taking duals. One of their weeds was given by (cid:16) Q3 = (cid:17) , which is a translated extension of the weed Q1. (In [MS12b], Q3 is denoted by Q, but we have already used this symbol.) Hence Corollary 4.7 gives another proof that there is no subfactor whose principal graphs are a translated extension of Q3. 4.3 The other quadruple point when n is even For formulas such as those in Theorem 4.2 to hold, we did not need to assume that (Γ+, Γ−) was a translated extension of Q; rather, we needed to know the formulas of the low weight rotational eigenvectors at depth n. Knowing the dual data ensured that we could determine the new low weight rotational eigenvectors in Proposition 4.1. If we assumed the same formulas for the low weight rotational eigenvectors for translated extensions of along with F(A) = σAλA would have obtained the same formulas. A and F(B) = σBλB B for some real scalars λA, λB, we Since we cannot verify these assumptions in this case, we cannot use these for- mulas to prove obstructions to principal graphs. However, we can prove that the low weight rotational eigenvectors are not given by these formulas, as in the following proposition. Proposition 4.9. Suppose (Γ+, Γ−) is a translated extension of Q4 = (cid:17) . Let P, Q, R and P , Q, R be the minimal projections at depth 4 from bottom to top. At least one of Q − R or Q − R is not a low weight rotational eigenvector. 23 where using r = r, we have F(A) = σA F(B) = σB r (cid:33) (cid:33) (cid:32)(cid:112)r(1 + r)√ (cid:32) √ r(cid:112)r(1 + r) (cid:18)[n + 2] − r[n] D = √ 1 + r) C and C = σA( σB√ 1 + r D. Proof. Suppose that B = Q − R and Q − R are both rotational eigenvectors. Then so are A = rP − (Q + R) and r P − ( Q + R), since both are orthogonal to Temperley- Lieb and to Q − R, Q − R respectively. Since Q4 has annular multiplicities ∗20, ωA (cid:54)= ωB by [Jon12, Theorem 5.2.3]. Hence F(A) = σAλA D where C, D are distinct elements in the set {r P − ( Q + R), Q− R}. (We avoid using A, B to avoid confusion.) There are two cases to consider. C and F(B) = σBλB Suppose that C = r P − ( Q + R) and D = Q − R. Using r = r, we have F(A) = σA C = σA C and F(B) = σB D = σB D. (cid:33) (cid:32)(cid:112)r(1 + r) (cid:112)r(1 + r) (cid:18)√ (cid:19) r√ r Hence we may write A = C and B = D without confusion. Replacing Q with Q4, the proofs of Equation (QA1), Theorem 4.5, and Corollary 4.7 all hold mutatis mutandis. We conclude that [2] ≤ 2. But 2 < (cid:107)Γ±(cid:107) ≤ [2], a contradiction. The second case is a bit trickier. Suppose that C = Q− R and D = r P −( Q+ R), As in Theorem 4.5, ∩n+1(P (cid:48)) = Tr(P (cid:48)) Tr(P ) P = · (1 + r) [n + 1] 1 + r (cid:19)(cid:18)f (n) + D (cid:19) 1 + r . Liu's relation applied to P as in Theorem 4.2 has the following left hand side after multiplying by (1 + r)2: LHS = F(A + f (n))∗F(A + f (n)) = (1 + r) C 2 − √ = − D − √ (cid:18)[n + 2] − r[n] 1 + r C + T (σA + σ−1 A ) 1 + r (σA + σ−1 A ) [n] C + T (cid:48) [n] (cid:19)(cid:18) D (cid:19) 1 + r · (1 + r) [n + 1] for some T, T (cid:48) ∈ TLn,−. The right hand side, multiplying by (1 + r)2 and ignoring terms in Temperley-Lieb, is given by RHS = −(1 + r) [n + 1] [n] 1 + r [n + 2] − r[n] [n] D. = Since C, D are orthogonal to Temperley-Lieb and to each other, we must have σA + σ−1 A = 0, and the coefficients of D must agree, which means [n + 2] − r[n] −1 = [n] ⇐⇒ r = [n + 2] + [n] [n] . 24 But Tr(Q) = Tr(R) = [n] [2] implies dim(P ) = [n + 2] − [n] [2] and r = 2[n] [n + 2] − [n] . Hence 2[n] [n + 2] − [n] = r = [n + 2] + [n] [n] ⇐⇒ 3[n]2 = [n + 2]2. It is easy to check that this is impossible when q > 1. Remark 4.10. By the symmetry of the graphs in Q4, it is natural to hypothesize that both Q−R and Q− R are rotational eigenvectors, although there is no particular reason this should be the case. Corollary 4.11. For the GHJ 3311 subfactor [GdlHJ89] with principal graphs (cid:16) , (cid:17) , at least one of Q − R or Q − R is not a low weight rotational eigenvector. Remarks 4.12. (1) In fact, the GHJ 3311 subfactor is the only subfactor whose principal graphs are a translated extension of Q4 [IJMS12]. (2) Corollary 4.11 also follows from [MP13, Theorem 5.10], where surprisingly Q− R is a rotational eigenvector! However, it is convenient to disprove the hypothesis in Remark 4.10 without having to do the substantial work of constructing the subfactor planar algebra. References [AH99] Marta Asaeda and Uffe Haagerup, Exotic subfactors of finite depth 17)/2, Comm. Math. Phys. with Jones indices (5 + 202 (1999), no. 1, 1 -- 63, MR1686551, DOI:10.1007/s002200050574, arXiv:math.OA/9803044. 13)/2 and (5 + √ √ [BH96] Dietmar Bisch and Uffe Haagerup, Composition of subfactors: new ex- amples of infinite depth subfactors, Ann. Sci. ´Ecole Norm. Sup. (4) 29 (1996), no. 3, 329 -- 383, MR1386923. [BJ97] Dietmar Bisch and Vaughan F. R. Jones, Algebras associated to interme- diate subfactors, Invent. Math. 128 (1997), no. 1, 89 -- 157, MR1437496. [BMPS12] Stephen Bigelow, Scott Morrison, Emily Peters, and Noah Sny- der, Constructing the extended Haagerup planar algebra, Acta Math. 209 (2012), no. 1, 29 -- 82, MR2979509, arXiv:0909.4099, DOI:10.1007/s11511-012-0081-7. 25 [BP13] Stephen Bigelow and David Penneys, Principal graph stability and the jellyfish algorithm, Math. Ann. (2013), 24 pages, arXiv:1208.1564, DOI:10.1007/s00208-013-0941-2. [GdlHJ89] Frederick M. Goodman, Pierre de la Harpe, and Vaughan F.R. Jones, Coxeter graphs and towers of algebras, Mathematical Sciences Research Institute Publications, 14. Springer-Verlag, New York, 1989, x+288 pp. ISBN: 0-387-96979-9, MR999799. [Haa94] √ Uffe Haagerup, Principal graphs of subfactors in the index range 4 < [M : 2, Subfactors (Kyuzeso, 1993), World Sci. Publ., River Edge, N ] < 3 + NJ, 1994, MR1317352 available at http://tqft.net/other-papers/ subfactors/haagerup.pdf, pp. 1 -- 38. [IJMS12] Masaki Izumi, Vaughan F. R. Jones, Scott Morrison, and Noah Sny- der, Subfactors of index less than 5, Part 3: Quadruple points, Comm. Math. Phys. 316 (2012), no. 2, 531 -- 554, MR2993924, arXiv:1109.3190, DOI:10.1007/s00220-012-1472-5. [JMS13] Vaughan F. R. Jones, Scott Morrison, and Noah Snyder, The classifica- tion of subfactors of index at most 5, 2013, arXiv:1304.6141. [Jon83] [Jon99] [Jon01] [Jon12] [KS00] Vaughan F. R. Jones, Index for subfactors, Invent. Math. 72 (1983), no. 1, 1 -- 25, MR696688, DOI:10.1007/BF01389127. , Planar algebras I, 1999, arXiv:math/9909027. , The annular structure of subfactors, Essays on geometry and related topics, Vol. 1, 2, Monogr. Enseign. Math., vol. 38, Enseignement Math., Geneva, 2001, MR1929335, pp. 401 -- 463. , Quadratic tangles 161 (2012), no. 12, DOI:10.1215/00127094-1723608. in planar algebras, Duke Math. J. 2257 -- 2295, MR2972458, arXiv:1007.1158, Vijay Kodiyalam and V. S. Sunder, The subgroup-subfactor, Math. Scand. 86 (2000), no. 1, 45 -- 74, MR1738515, available (gratis) at http: //www.mscand.dk/article.php?id=128. [Kup96] Greg Kuperberg, Spiders for rank 2 Lie algebras, Comm. Math. Phys. 180 (1996), no. 1, 109 -- 151, MR1403861, arXiv:q-alg/9712003 euclid.cmp/1104287237. [Mor] [MP13] Scott Morrison, A formula for the Jones-Wenzl projections, Unpublished, available at http://tqft.net/math/JonesWenzlProjections.pdf. Scott Morrison and David Penneys, Constructing spoke subfactors us- ing the jellyfish algorithm, 2013, arXiv:1208.3637, Accepted to Trans. Amer. Math. Soc. Jan. 18. 26 [MPPS12] Scott Morrison, David Penneys, Emily Peters, and Noah Snyder, Subfactors of index less than 5, Part 2: Triple points, Internat. J. Math. 23 (2012), no. 3, 1250016, 33, MR2902285, arXiv:1007.2240, DOI:10.1142/S0129167X11007586. [MPPS13] Scott Morrison, David Penneys, Emily Peters, and Noah Snyder, Some group-like subfactors at index 6, 2013, In preparation. [MS12a] [MS12b] [Ocn88] [Pfe05] [Pop94] [Pop95] [PT12] [Rez07] Scott Morrison and Noah Snyder, Non-cyclotomic fusion categories, Trans. Amer. Math. Soc. 364 (2012), no. 9, 4713 -- 4733, arXiv:1002. 0168, MR2922607, DOI:10.1090/S0002-9947-2012-05498-5. Scott Morrison and Noah Snyder, Subfactors of index less than 5, Part 1: The principal graph odometer, Comm. Math. Phys. 312 (2012), no. 1, 1 -- 35, MR2914056, arXiv:1007.1730, DOI:10.1007/s00220-012-1426-y. Adrian Ocneanu, Quantized groups, string algebras and Galois theory for algebras, Operator algebras and applications, Vol. 2, London Math. Soc. Lecture Note Ser., vol. 136, Cambridge Univ. Press, Cambridge, 1988, MR996454, pp. 119 -- 172. Gotz Pfeiffer, The subgroups of S6, 2005, Available at http://schmidt. nuigalway.ie/subgroups/s6.pdf. Sorin Popa, Classification of amenable subfactors of type II, Acta Math. 172 (1994), no. 2, 163 -- 255, MR1278111, DOI:10.1007/BF02392646. , An axiomatization of the lattice of higher relative commutants of a subfactor, Invent. Math. 120 (1995), no. 3, 427 -- 445, MR1334479 DOI:10.1007/BF01241137. David Penneys and James E. Tener, Subfactors of index less than 5, Part 4: Vines, Internat. J. Math. 23 (2012), no. 3, 1250017, 18, MR2902286, arXiv:1010.3797, DOI:10.1142/S0129167X11007641. Sarah A. Reznikoff, Coefficients of the one- and two-gap boxes in the Jones-Wenzl idempotent, Indiana Univ. Math. J. 56 (2007), no. 6, 3129 -- 3150, MR2375712. [Sny12] Noah Snyder, A rotational approach to triple point obstructions, 2012, arXiv:1207.5090, to appear Analysis & PDE. [Wen87] Hans Wenzl, On sequences of projections, C. R. Math. Rep. Acad. Sci. Canada 9 (1987), no. 1, 5 -- 9, MR873400. [Wen88] , Hecke algebras of type An and subfactors, Invent. Math. 92 (1988), no. 2, 349 -- 383, MR936086. 27
1507.03702	1	1507	2015-07-14T02:30:13	CP-stability and the local lifting property	[ "math.OA" ]	The purpose of this note is to discuss the local lifting property in terms of an equivalent approximation-type property, CP-stability, which was formulated by the author and Isaac Goldbring for the purposes of studying the continuous model theory of C$^*$-algebras and operator systems.	math.OA	math	CP-STABILITY AND THE LOCAL LIFTING PROPERTY THOMAS SINCLAIR Abstract. The purpose of this note is to discuss the local lifting property in terms of an equivalent approximation-type property, CP-stability, which was formulated by the author and Isaac Goldbring for the purposes of studying the continuous model theory of C∗-algebras and operator systems. 1. Statement of the Main Results The following definition first appears in [GS15a]. Definition 1.1. An operator system X is said to be CP-stable if for any finite dimensional subsystem E ⊂ X and δ > 0 there is a finite-dimensional subsystem E ⊂ S ⊂ X and k, ǫ > 0 so that for every C∗-algebra A and any unital linear map φ : S → A with kφkk < 1 + ǫ there exists a u.c.p. map ψ : E → A so that kφE − ψk < δ. Let u1, . . . , un be the canonical generators of C∗(Fn) and let Wn be the operator sys- tem spanned by the set {u∗ i uj : 1 ≤ i, j ≤ n + 1} where un+1 := 1. The first result gives a quantitative version of CP-stability for the operator systems Wn using work of Farenick and Paulsen [FP12]. Theorem A. The operator system Wn is CP-stable. In particular, for any δ > 0 there exists ǫ > 0 so that for any unital linear map φ : Wn → A into an arbitrary unital C∗-algebra with kφkn+1 < 1 + ǫ, there is a u.c.p. map ψ : Wn → A so that kψ − φk < δ. Note that by [CH85, Corollary 4.7] no such result can hold for the related generator subsystem of the reduced C∗-algebra C∗ λ(Fn). We say an operator system X has the local lifting property (LLP) of Kirchberg if for every unital C∗-algebra A, every ideal J of A, and every u.c.p. map φ : X → A/J and every finite-dimensional subsystem E ⊂ X the restricted map φE admits a u.c.p. lifting φ : E → A.1 It was shown in [GS15b] that for C∗-algebras CP-stability is equivalent to the local lifting property. Using operator system tensor product characterizations for exactness and the LLP (see [KP+13]), Kavruk showed that a finite-dimensional operator system has the LLP if and only if its dual system is exact [Ka14, Theorem 6.6]. We show that conversely, one can use the fact that the dual is exact (in the sense of admitting a nuclear embedding), i.e., that the operator system is CP-stable, to recover Kirchberg's tensor characterization of the LLP [KP+13, Ki93]. Date: July 2, 2018. 1The definition we give here is termed the operator system local lifting property (OSLLP) in [Ka14, KP+13] though for our purposes we will not make a distinction. 1 2 T. SINCLAIR Theorem B. If E is a finite-dimensional operator system which is CP-stable, then E ⊗min B(ℓ2) ∼= E ⊗max B(ℓ2) as operator systems. Using techniques from [Pi96] or [Ka14] Theorem A and Theorem B give a new proof of the fact (due to Kirchberg [Ki94]) that C∗(Fn) ⊗min B(ℓ2) ∼= C∗(Fn) ⊗max B(ℓ2). The author is grateful to Isaac Goldbring for many stimulating discussions from which these ideas arose. 2. Proofs of the Main Results The following result is due to Farenick and Paulsen [FP12]: see the remarks after Definition 2.1 therein. Lemma 2.1. The "covering" map γn : Mn+1 → Wn defined by φ(eij) = 1 i uj, where u1, . . . , un+1 are defined as above, is u.c.p. and the kernel Jn+1 consists of all diagonal matrices in Mn+1 of trace zero. n+1 u∗ Remarkably, Farenick and Paulsen [FP12, Theorem 2.4] go on to show that: Theorem 2.2 (Farenick+Paulsen). The map γn : Mn+1/Jn+1 → Wn is a complete order isomorphism where the quotient space Mn+1/Jn+1 is equipped with its canonical operator system structure as defined in [KP+13, Section 3]. The strategy of our proof of Theorem A will be to make use of the fact that matrix algebras are CP-stable. Lemma 2.3 (Proposition 2.40 in [GS15a]). Given k, for any δ > 0 there exists ǫ > 0 so that for any C∗-algebra A and any unital linear map φ : Mk → A with kφkk < 1 + ǫ, there exists a u.c.p. map φ : Mk → A so that k φ − φk < δ. For the reader's convenience we provide a streamlined proof. Proof. Suppose by contradiction that there is some δ > 0 so that for every n there is some unital linear map φn : Mk → An into some C∗-algebra An so that kφkk < 1 + 1 so that kψ − φk ≥ δ for any u.c.p. map ψ : Mk → An. Fix an nonprinciple ultrafilter ω on N and define A := (An)ω to be the ultrapower C∗-algebra associated to the sequence (An) and ω and eA := Qn An. Consider the map φ := (φ•) : Mk → A. Clearly φ is unital and kφkk = 1 whence by [Pa03, Proposition 2.11] φ is k-positive. By Choi's theorem [Pa03, Theorem 3.14] φ is therefore u.c.p. and the proof of the Choi+Effros eA. lifting theorem [BO08, Theorem C.3] shows there is thus a u.c.p. lift φ : Mk → However, this shows that the sequence (φn) is well-approximated by u.c.p. maps for n ∈ ω generic, a contradiction. (cid:3) n Proof of Theorem A. We begin by fixing δ > 0 and a unital C∗-algebra A. Suppose we have a unital linear map φ : Wn → A with kφkn+1 < 1 + ǫ for some ǫ > 0 sufficiently small and to be determined later. We will show that we can find a u.c.p. map ψ : Wn → A so that kψ − φk < δ. Let φ ′ := φ ◦ γn : Mn+1 → A which is again unital and linear with kφkn+1 < 1 + ǫ. By Lemma 2.3 we may choose ǫ > 0 sufficiently small so that there is a u.c.p. map : Mn+1 → A so that kψ ′ − φ ′k < δ/16n4. Since φ ′(eii) = 1 ψ ′ n+1 1A, we have that CP-STABILITY AND THE LOCAL LIFTING PROPERTY 3 kψ ′(eii) − 1 n+1 1Ak < δ/16n4 whence bi := ψ ′(eii) is uniformly invertible and positive. . Let Ψ ′ := [ψ ′(eij)] ∈ Let B ∈ Mn+1(A) be the diagonal matrix such that Bii := b Mn+1(A) be the Choi matrix associated to ψ ′. Since Ψ ′ is positive, so is Ψ ′′ := BΨ ′B, whence it defines a c.p. map ψ ′′ : Mn+1 → A via the reverse correspondence ψ ′′(eij) := ij. We can see manifestly that ψ ′′(eii) = 1 Ψ ′′ n+1 1A whence ψ ′′ is unital, Jn+1 ⊂ ker(ψ ′′), and kψ ′′ − ψ ′k < δ/4n2. −1/2 i Identifying Wn with the quotient operator system Mn+1/Jn+1 by Theorem 2.2, since Jn+1 ⊂ ker(ψ ′′) it follows by [KP+13, Proposition 3.6] that there is u.c.p. map ψ : Wn → i uj)k < δ/2n2. Alternatively, this is not difficult to see A so that supi,jkψ(u∗ i uj) − φ(u∗ by setting ψ := ψ ′′ ◦ γ−1 n and unravelling the definition of the quotient operator system structure via the identification given by Theorem 2.2. In any case it follows by the small perturbation argument that kψ − φk < δ, and we are done. (cid:3) Two formal weakenings of the LLP were introduced by the author and Isaac Gold- bring: the local ultrapower lifting property (LULP) [GS15a, Proposition 2.42] and the ap- proximate local lifting property (ALLP) [GS15b, Definition 7.3].2 Both definitions carry over straightforwardly to the category of operator systems. For instance, an operator system X can be said to have the LULP if every u.c.p. map φ : X → Aω admits local u.c.p. lifts to ℓ∞(A). The following proposition is essentially contained in [GS15a, GS15b]. We provide a sketch of the proof for the convenience of the reader. Proposition 2.4. For an operator system X the following statements are equivalent: (1) X has the LLP; (2) X has the ALLP; (3) X has the LULP; (4) X is CP-stable. The equivalence of the first two statements essentially appears in the work of Effros and Haagerup [EH85, Theorem 3.2]. We also remark that using the equivalence with the ALLP, it is easy to see that the LLP passes to inductive limits, noting that it suffices to check the ALLP only on a dense subalgebra. Proof. The equivalence of (3) and (4) is proved in [GS15a, Proposition 2.42]. The impli- cation (1) ⇒ (2) is straightforward. For (2) ⇒ (3), we note that by the small perturbation we can require the approximate lifts to be unital, and we may also assume they are ∗-linear. In conjunction with [BO08, Corollary B.11] which shows that we can correct such an approximate lift to a u.c.p. map a controlled distance away (depending on the dimension of the domain), we can thus assume that the approximate lifts are u.c.p. from which the implication follows easily. We include a proof of (4) ⇒ (2), though it closely follows the reasoning given in [GS15b, Proposition 7.7]. To this end, note that by the main result of [RS89] that for any finite-dimensional operator system, any u.c.p. map φ : E → A/J admits n-positive unital liftings φn : E → A for every n. Hence if E was a finite dimensional subsystem of a CP-stable system 2The ALLP is implicitly formulated in the work of Effros and Haagerup [EH85], where it is shown to be equivalent to the LLP. 4 T. SINCLAIR X and φ : X → A/J was u.c.p. it would follow that for every n there is a u.c.p. map ψ : E → A so that kπJ◦ψ−φEk < 1 n , where πJ : A → A/J is the quotient ∗-epimorphism. Hence X has the ALLP. Finally, the implication (2) ⇒ (1) follows from a foundational result of Arveson that liftable u.c.p. maps are closed in the point-norm topology: see [BO08, Lemma C.2]. (cid:3) Let OSn be the set of all complete isomorphism classes of n-dimensional operator systems. The set OSn is naturally equipped with two complete metrics, the cb-Banach distance and the weak metric: see [GS15b] for details. With the equivalence of LLP and CP-stability in hand, we give a new proof of a result of Kavruk [Ka14]. Proposition 2.5 (Kavruk). A finite-dimensional operator system E is exact if and only if the dual system E∗ is CP-stable. Proof. It is well known (see [Pi95, GS15b]) that E is an exact n-dimensional operator system if and only if any sequence of unital ∗-linear maps φα : Eα → E so that kφαkk → 1 for all k, there is a sequence of unital maps ψα : Eα → E with kψαkcb → 1 with kψα − φαk → 0. Dualizing (noting by [JP95, Proposition 2.1] or [BP91] that this is a well behaved operation) and applying a standard compactness argument, we see that this is implies that E∗ is CP-stable. The converse follows similarly by unravelling the definitions. (cid:3) In the category of operator systems, the correct treatment of tensor products has only recently appeared in the work of Kavruk, Paulsen, Todorov, and Tomforde [KP+11, KP+13]. We refer to these works for the basic definitions and properties of various operator system tensor products. Using these ideas we give a new proof of a famous and difficult theorem of Kirchberg [Ki94]. A short and particularly elegant proof of the same result in the context of operator spaces was given by Pisier [Pi96]. A second elementary proof was recently discovered by Farenick and Paulsen [FP12]. (See also [Ha14, Oz13].) Theorem 2.6 (Kirchberg). If E is a finite-dimensional operator system which is CP-stable, then E ⊗min B(ℓ2) = E ⊗max B(ℓ2). Lemma 2.7. Let E be a finite-dimensional operator system with the LLP. Then for every ǫ > 0, k n → E so that for any positive x ∈ Mk(E ⊗min B(ℓ2))+ there there is n and a u.c.p. map φ : M∗ is ^x ∈ Mk(M∗ n ⊗min B(ℓ2))+ positive so that kx − (φ ⊗ id)k(^x)k < ǫ. Proof. Let us fix ǫ, k > 0. Using [KP+13, Lemma 8.5] we may identify the positive cone Mk(E ⊗min B(ℓ2))+ with the space CP(E∗, Mk(B(ℓ2))) of completely positive maps φ : E∗ → Mk(B(ℓ2)). By Proposition 2.5 E∗ is exact, so there are matricial operator systems Em ⊂ Mℓm and u.c.p. bijections φm : E∗ → Em with kφ−1 m kcb → 1. By pre-composition each map φm induces a map Φm,k : CP(Mℓm, Mk(B(ℓ2))) → CP(E∗, Mk(B(ℓ2))) which preserves unitality and is easily identified with the u.c.p. map ℓm ⊗min B(ℓ2)) → Mk(E ⊗min B(ℓ2)). m ⊗ id)k : Mk(M∗ (φ∗ CP-STABILITY AND THE LOCAL LIFTING PROPERTY 5 (We are using that the minimal operator system tensor product is functorial: see [KP+11, Theorem 4.6].) Given a u.c.p. map ψ : E∗ → B(H) we may pre-compose with φ−1 m to obtain a unital, self-adjoint map ψm : Em → B(H) which we may isometrically extend to ^ψm : Mm → B(H). As kψmkcb ≤ kφ−1 m kcb → 1, by [BO08, Corollary B.9] there is an approximat- ing sequence to ( ^ψm) consisting of u.c.p. maps υm : Mm → B(H) with k ^ψm − υmk ≤ 2(kφ−1 m ⊗ id)k re- ⊗min B(ℓ2))+ is ǫ-surjective into Mk(E⊗min B(ℓ2))+ for m sufficiently stricted to Mk(M∗ ℓm large. m kcb − 1). Via the identification with Φm,k it therefore follows that (φ∗ Note that Mn is (trivially) exact and is CP-stable, whence M∗ n shares these properties n is a C∗-nuclear op- by Proposition 2.5. It would be interesting to know whether M∗ erator system in the sense of [Ka14, KP+13] -- as essentially noted in [Ka14, Theorem 6.7] this is predicted by Connes' embedding conjecture. (N.B. One's first instinct might be to conclude that C∗-nuclearity for all M∗ n would force exactness and CP-stability to coincide for finite-dimensional operator systems; however, it is easy to find the basic fault in this argument.) (cid:3) Lemma 2.8. For all n, M∗ n ⊗min B(ℓ2) = M∗ n ⊗max B(ℓ2). Proof. Since B(ℓ2) has the WEP, by [KP+13, Lemma 6.1 and Theorem 6.7] it suffices to check that M∗ n ⊗max E for any finite-dimensional operator system E. n ⊗min E)∗ ∼= Mn ⊗max E∗ ∼= Mn ⊗min E∗ Using [FP12, Proposition 1.9] we have that (M∗ as operator systems. By the same n ⊗min E = M∗ M∗ n ⊗min E ∼= (M∗ n ⊗min E)∗∗ ∼= (Mn ⊗min E∗)∗ ∼= M∗ n ⊗max E, and we are done. (cid:3) Proof of Theorem 2.6. Given ǫ, k > 0, by Lemma 2.7 we can find n such that there is a n ⊗min B(ℓ2))+ → Mk(E ⊗min B(ℓ2))+ is n → E so that (φ ⊗ id)k : Mk(M∗ u.c.p. map φ : M∗ n ⊗max B(ℓ2))+. n ⊗min B(ℓ2))+ = Mk(M∗ ǫ-surjective. By Lemma 2.8 we have that Mk(M∗ Since the maximal tensor norm is functorial [KP+11, Theorem 5.5], it follows that (φ ⊗ n ⊗max B(ℓ2))+ into Mk(E ⊗max B(ℓ2))+. As ǫ was arbitrary this shows id)k maps Mk(M∗ that Mk(E ⊗min B(ℓ2))+ ⊂ Mk(E ⊗max B(ℓ2))+, and we are done. (cid:3) References D.P. Blecher and V.I. Paulsen, Tensor products of operator spaces, J. Funct. Anal. 99 (1991) 262-292. [BP91] [BO08] N.P. Brown and N. Ozawa, C∗-Algebras and Finite-Dimensional Approximations, Grad. Studies in [CH85] [EH85] [FP12] [GS15a] Math. 88, AMS, Providence, RI, 2008. J. de Cannière and U. Haagerup, Multipliers of the Fourier algebras of some simple Lie groups and their discrete subgroups, Amer. J. Math. 107 (1985) 455-500. E. Effros and U. Haagerup, Lifting problems and local reflexivity for C∗ algebras, Duke Math. J. 52 (1985),103-128. D. Farenick, V. Paulsen, Operator system quotients of matrix algebras and their tensor products, Math. Scand. 111 (2012) 210 -- 243. I. Goldbring and T. Sinclair, On Kirchberg's embedding problem, J. Funct. Anal. 269 (2015) 155-198. 6 T. SINCLAIR [GS15b] [Ha14] [JP95] M. Junge and G. Pisier, Bilinear forms on exact operator spaces and B(H) ⊗ B(H), Geom. Funct. Anal. I. Goldbring and T. Sinclair, Omitting types in operator systems, preprint, arXiv:1501.06395. K.H. Han, A Kirchberg type tensor theorem for operator systems, preprint, arXiv:1409.1306. [Ka14] (GAFA) 5 (1995), 329-363. A.S. Kavruk, Nuclearity related properties in operator systems, Journal of Operator Theory 71 (2014), 95-156. [KP+11] A.S. Kavruk, V.I. Paulsen, I.G. Todorov, and M. Tomforde, Tensor products of operator systems, J. Funct. Anal. 261 (2011) 267-299. [KP+13] A.S. Kavruk, V.I. Paulsen, I.G. Todorov, and M. Tomforde, Quotients, exactness, and nuclearity in [Ki93] [Ki94] the operator system category, Adv. Math. 235 (2013) 321-360. E. Kirchberg, On non-semisplit extensions, tensor products, and exactness of group C∗-algebras, Invent. Math. 112 (1993) 449-489. E. Kirchberg, Commutants of unitaries in UHF algebras and functorial properties of exactness J. reine angew. Math. 452 (1994) 39-77. [Oz13] N. Ozawa, About the Connes embedding conjecture: algebraic approaches, Jpn. J. Math. 8 (2013) 147- [Pa03] [Pi95] [Pi96] [RS89] 183. V.I. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Advanced Mathematics 78, Cambridge University Press (2003). G. Pisier, Exact operator spaces, in Recent Advances in Operator Algebras (Orléans, 1992), Astérisque 232 (1995) 159-186. G. Pisier, A simple proof of a theorem of Kirchberg and related results on C∗-norms, J. Operator Theory 35 (1996) 317-335. A.G. Robertson and R.R. Smith, Liftings and extensions of maps on C∗-algebras, J. Operator Theory 21 (1989), 117-131. Department of Mathematics, Indiana University, Rawles Hall, 831 E 3rd St, Bloomington, IN 47405, USA E-mail address: [email protected]
1606.03156	1	1606	2016-06-10T01:41:49	On the uniqueness of injective III$_1$ factor	[ "math.OA" ]	We give a new proof of a theorem due to Alain Connes, that an injective factor $N$ of type III$_1$ with separable predual and with trivial bicentralizer is isomorphic to the Araki--Woods type III$_1$ factor $R_{\infty}$. This, combined with the author's solution to the bicentralizer problem for injective III$_1$ factors provides a new proof of the theorem that up to $*$-isomorphism, there exists a unique injective factor of type III$_1$ on a separable Hilbert space.	math.OA	math	On the uniqueness of injective III1 factor Uffe Haagerup May 1985 Abstract We give a new proof of a theorem due to Alain Connes, that an injective factor N of type III1 with separable predual and with trivial bicentralizer is isomorphic to the Araki -- Woods type III1 factor R∞. This, combined with the author's solution to the bicentralizer problem for injective III1 factors provides a new proof of the theorem that up to ∗-isomorphism, there exists a unique injective factor of type III1 on a separable Hilbert space. Preamble by Alain Connes Uffe Haagerup solved the hardest problem of the classification of factors, namely the uniqueness problem for injective factors of type III1. The present paper, taken from his unpublished notes, presents a direct proof of this uniqueness by showing that any injective factor of type III1 is an infinite tensor product of type I factors so that the uniqueness follows from the Araki -- Woods classification. The proof is typical of Uffe's genius, the attack is direct, and combines his amazing control of completely positive maps and his sheer analytical power, together with his solution to the bicentralizer problem. After his tragic death, Hiroshi Ando volunteered to type the manuscript1. Some pages were missing from the notes, but eventually Cyril Houdayer and Reiji Tomatsu suggested a missing proof of Lemma 3.4 and Theorem 3.1. We heartily thank Hiroshi, Cyril and Reiji for making the manuscript available to the community. We also thank Søren Haagerup for giving permission to publish his father's paper. Contents 1 Introduction 2 Preliminaries 2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Connes -- Woods' characterization of ITPFI factors . . . . . . . . . . . . . . . . . 2.3 Bicentralizers on type III1 factors . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Almost unitary equivalence in Hilbert N -bimodules . . . . . . . . . . . . . . . . 3 Completely positive maps from m × m-matrices into an injective factor of type III1 4 Q-stable states on III1 factors 5 Proof of Main Theorem 2 4 4 4 5 6 7 13 17 1 The manuscript is typed by Hiroshi Ando (Chiba University) in cooperation with Cyril Houdayer (Universit´e Paris-Sud), Toshihiko Masuda (Kyushu University), Reiji Tomatsu (Hokkaido University), Yoshimichi Ueda (Kyushu University) and Wojciech Szymanski (University of Southern Denmark). 1 1 Introduction The problem, whether all injective factors of type III1 on a separable Hilbert space are isomor- phic, has been settled affirmatively. The proof of the uniqueness of injective III1 factors falls in two parts, namely (see §2.3 for the definition of the bicentralizer): Theorem 1.1 ([Con85]). Let M be an injective factor of type III1 on a separable Hilbert space, such that the bicentralizer Bϕ is trivial (i.e., Bϕ = C1) for some normal faithful state ϕ on M , then M is ∗-isomorphic to the Araki -- Woods factor R∞. Theorem 1.2 ([Haa87]). For any normal faithful state ϕ on an injective factor M of type III1 on a separable Hilbert space, one has Bϕ = C1. In this paper we give an alternative proof of Theorem 1.1 above, based on the technique of our simplified proof [Haa85] of Connes' Theorem [Con76] "injective⇒hyperfinite" in the type II1 case2. The key steps in our proof of Theorem 1.1 are listed below: Step 1 By use of continuous crossed products, we prove that the identity map on an injective factor N of type III1 has an approximate factorization R Tλ ❅❅❅❅❅❅❅ Sλ >⑦⑦⑦⑦⑦⑦⑦ idN N N through the hyperfinite factor R of type II1, such that (Sλ)λ∈Λ and (Tλ)λ∈Λ are nets of normal unital completely positive maps, and for a fixed normal faithful state ϕ on N (chosen prior to Sλ and Tλ), there exist normal fatihful states (ψλ)λ∈Λ on R, such that for all t ∈ R and λ ∈ Λ, ϕ ◦ Tλ = ψλ, ψλ ◦ Sλ = ϕ, σϕ t ◦ Tλ = Tλ ◦ σψλ t ◦ Sλ = Sλ ◦ σϕ σψλ t , t , λ→∞→ 0 for all x ∈ N , where kykϕ := ϕ(y∗y) and kSλ ◦ Tλ(x) − xkϕ Step 2 From Step 1, we deduce that a certain normal faithful state ϕ (Q-stable state defined in §4) on an injective factor N of type III1 has the following property: for any finite set of unitaries u1, . . . , un in N and for every γ, δ > 0, there exists a finite-dimensional subfactor F of N such that 2 (y ∈ N ). 1 and such that there exist unitaries v1, . . . , vn in F and a unital completely positive map T : F → N such that ϕ = ϕF ⊗ ϕF c , ϕ ◦ T = ϕ, t ◦ T − T ◦ σϕF t kσϕ k ≤ γt, t ∈ R, and 2Typewriter's note: Haagerup used this technique to give a new proof of the uniqueness of injective type IIIλ (0 < λ < 1) factor. This result has been published in [Haa89]. kT (vk) − vkkϕ < δ, k = 1, . . . , n. 2 / / > Step 3 We prove that if N, ϕ, F, u1, . . . , un, v1, . . . , vn are as in Step 2, then for every σ-strong neigh- borhood V of 0 in N , there exists a finite set of operators a1, . . . , ap in N such that a∗ i ai ∈ 1 + V and a∗ i ai ≤ 1, pXi=1 aia∗ i! ∈ 1 + V and εF,ϕ pXi=1 i! ≤ 1, aia∗ (a) pXi=1 (b) εF,ϕ pXi=1 pXi=1 pXi=1 (d) (c) kaiξϕ − ξϕaik2 < δ′, kaiuk − vkaik2 ϕ < δ′, k = 1, . . . , n. Here ξϕ denotes the unique representing vector of ϕ in a natural cone. The above δ′ > 0 de- pends on γ and δ in Step 2, and δ′ is small when γ and δ are small. Here, εF,ϕ is the ϕ-invariant conditional expectation of N onto F . Moreover in (c), the standard Hilbert space H of N is regarded as a Hilbert N -bimodule, by putting ηa := Ja∗Jη (a ∈ N, η ∈ H). Assume now that the bicentralizer of any normal faithful state on N is trivial. Then by an averaging argument, we can exchange (b) by (b') pXi=1 aia∗ i ∈ 1 + V and pXi=1 aia∗ i ≤ 1. Step 4 From (a), (b'), (c) and (d) above, we derive that there exists a unitary operator w ∈ N such that and kwξϕ − ξϕwk < ε where ε is small when δ′ is small and V is a small σ-strong neighborhood of 0 in N . The key part of Step 4 is a theorem about general Hilbert N -bimodules, which was proved in [Haa89]. kwuk − vkwkϕ < ε, k = 1, . . . , n, Step 5 From Step 4, we get that for every finite set of unitaries u1, . . . , un ∈ N and every ε > 0, there exists a finite dimensional subfactor F1 (namely w∗F w) of N and n unitaries v′ n in F1 (namely w∗vkw, k = 1, . . . , n), such that 1, . . . , v′ and (i) kv′ k − vkkϕ < ε (ii) kϕ − ϕF1 ⊗ ϕF c 1 k < ε. The last inequality follows from the fact, that when w almost commutes with ξϕ, it almost commutes with ϕ too. The properties (i) and (ii) above show that ϕ satisfies the product condition of Connes -- Woods [CW85] and thus N is an ITPFI factor. But it is well-known that R∞ is the only ITPFI factor of type III1 (cf. [AW68] and [Con73]). 3 2 Preliminaries 2.1 Notation We use M, N, . . . to denote von Neumann algebras and ξ, η, . . . to denote vectors in a Hilbert space. Let M be a von Neumann algebra. U(M ) denotes the unitary group of M . For a faithful normal state ϕ on M , we denote by ∆ϕ (resp. Jϕ) the modular operator (resp. modular conjugation operator) associated with ϕ, and the modular automorphism group of ϕ is denoted by σϕ. The norm kxkϕ = ϕ(x∗x) 2 defines the strong operator topology (SOT) on the unit ball of M . The centralizer of ϕ is denoted by Mϕ. 1 2.2 Connes -- Woods' characterization of ITPFI factors Recall that a von Neumann algebra M with separable predual is called hyperfinite if there exists an increasing sequence M1 ⊂ M2 ⊂ ··· of finite-dimensional -subalgebras such that n=1 Mn)′′. A factor M is called an Araki -- Woods factor or an ITPFI (infinite tensor product of factors of type I) factor, if it is isomorphic to the factor of the form M = (S∞ (Mi, ϕi), Oi∈I where I is a countable infinite set and each Mi (resp. ϕi) is a σ-finite type I factor (resp. a faithful normal state). Araki and Woods classified most ITPFI factors: Theorem 2.1 ([AW68]). There exists a unique ITPFI factor with separable predual for each type I∞, II1, II∞ and IIIλ, λ ∈ (0, 1]. In particular, all ITPFI factors of type III1 are isomorphic to R∞ :=On∈N (M3(C), Tr(ρ · )), where ρ := 1 1+λ+µ diag(1, λ, µ) and 0 < λ, µ satisfies log λ log µ /∈ Q. It is clear that an ITPFI factor with separable predual is hyperfinite. The converese is also true for factors not of type III0, but false in general. Namely, Connes -- Woods [CW85] characterized hyperfinite factors of type III0 with separable predual which are isomorphic to ITPFI factors by the approximate transitivity of their flow of weights, while the existence of hyperfinite factors of type III0 with separable predual which are not isomorphic to ITPFI factors had been shown in [Con72]. Let N be a von Neumann algebra, and let F be a finite dimensional subfactor of N with relative commutant F c := F ′ ∩ N in N . Then it is elementary to check, that the map nXi=1 xi ⊗ yi 7→ nXi=1 xiyi, xi ∈ F, yi ∈ F c (1 ≤ i ≤ n) is an isomorphism of F ⊗ F c onto N . If ω1 is a normal state on F and ω2 is a normal state on F c, we let ω1 ⊗ ω2 denote the corresponding state on N , i.e., In our proof of (ω1 ⊗ ω2)(xy) = ω1(x)ω2(y), x ∈ F, y ∈ F c. [N injective III1 and Bϕ = C1] ⇒ N ∼= R∞, we shall need the following criterion for a factor to be ITPFI: Proposition 2.2 ([CW85, Lemma 7.6]). Let N be a factor on a separable Hilbert space. Then N is ITPFI if and only if N admits a normal faithful state ϕ with the following property: for 4 every finite set x1, . . . , xn of operators in N , for every ε > 0, and every strong neighborhood V of 0 in N , there exists a finite dimensional subfactor F of N , such that and xk ∈ F + V, k = 1, . . . , n kϕ − ϕF ⊗ ϕF ck < ε. 2.3 Bicentralizers on type III1 factors In this subsection, we recall Connes' bicentralizers. Let M be a σ-finite von Neumann algebra, and let ϕ be a normal faithful state on M . We denote by AC(ϕ) the set of all norm-bounded sequences (xn)∞ Definition 2.3 (Connes). The bicentralizer of ϕ is the set Bϕ of all x ∈ M such that limn→∞ kxan − anxkϕ = 0 holds for all (an)∞ n=1 in M such that limn→∞ kϕxn − xnϕk = 0 holds. n=1 ∈ AC(ϕ). Since Bϕ is a von Neumann subalgebra of M [Haa87, Proposition 1.3], it holds that ϕ = 0. limn→∞ kxan − anxk♯ It was conjectured by Connes that for all factors of type III1 with separable predual, the bicentralizer Bϕ of any normal faithful state ϕ on M is trivial, i.e., Bϕ = C1 holds. This is still an open problem. We will need the following result on type III1 factors, known as the Connes -- Størmer transitivity: Theorem 2.4 ([CS78]). Let M be a type III1 factor with separable predual. Then for every faithful normal states ϕ, ψ on M and ε > 0, there exists a unitary u ∈ U(M ) such that kuϕu∗ − ψk < ε holds. Connes showed that by the Connes-Størmer transitivity, for a type III1 factor M with separable predual, the triviality of Bϕ for one fixed faithful normal state ϕ on M implies the triviality of Bψ for every faithful normal state ψ (see [Haa87, Corollary 1.5] for the proof). He also showed that the triviality of the bicentralizer is equivalent to the following property (the proof is given in [Haa87, Proposition 1.3 (2)]): Proposition 2.5 (Connes). Let M be a von Neumann algebra with a normal faithful state ϕ. Then Bϕ = C1 holds, if and only if the following condition is satisfied: for every a ∈ M and δ > 0, where conv is the closure of the convex hull in the σ-weak topology. conv{u∗au; u ∈ U(M ), kuϕ − ϕuk ≤ δ} ∩ C1 6= ∅, We will use the following variant of Proposition 2.5. Proposition 2.6. Let M be a type III1 factor with separable predual, and let ϕ be a normal faithful state on M whose modular automorphism group σϕ leaves a finite-dimensional sub- factor F globally invariant. Let εF,ϕ : M → F be the normal faithful ϕ-preserving conditional expectation. Assume that Bϕ = C1. Then for every δ > 0 and a ∈ M , we have εF,ϕ(a) ∈ conv{u∗au; u ∈ U(F c), kuξϕ − ξϕuk ≤ δ}. (1) Here, ξϕ is the representing vector of ϕ in the natural cone. Proof. The proof is essentially the same as Proposition 2.5, so we only indicate the outline. Note that by Araki-Powers-Størmer inequality, for every u ∈ U(M ) one has: kξϕ − uξϕu∗k2 ≤ kϕ − uϕu∗k ≤ kξϕ − uξϕu∗k · kξϕ + uξϕu∗k. Therefore in the arguments below, we may replace the condition "kuξϕ − ξϕuk ≤ δ" in Propo- sition 2.6 with the condition "kuϕ − ϕuk ≤ δ", as we take δ > 0 to be arbitrarily small. As 5 was pointed out in [Haa87, Remark 1.4], it follows from the proof of Proposition 2.5 that the condition Bϕ = C1 is equivalent to the next condition that for all a ∈ M and δ > 0, conv{u∗au; u ∈ U(M ), kuξϕ − ξϕuk < δ}. (2) ϕ(a)1 ∈ \δ>0 Let a ∈ M . Since M ∼= F ⊗ F c with ϕ = ϕF ⊗ ϕF c , we may now apply (2) to F c(∼= M ) and ϕF c to obtain εF,ϕ(a) = idF ⊗ ϕF c (a) ∈ conv{u∗au; u ∈ U(F c), kuξϕ − ξϕuk ≤ δ}. Note that we used the fact that kϕu − uϕk = kψu − uψk, where ψ := ϕF c and u ∈ U(F c) thanks to the existence of a normal faithful ϕ-preserving conditional expectation from M onto F c. 2.4 Almost unitary equivalence in Hilbert N -bimodules We recall a result about almost unitary equivalence in Hilbert bimodules established in [Haa89] which is a generalization of [Haa85, Theorem 4.2]. Let N be a von Neumann algbera, and H be a normal Hilbert N -bimodule, i.e., H is a Hilbert space on which there are defined left and right actions by elements from N : (x, ξ) 7→ xξ, (x, ξ) 7→ ξx, x ∈ N, ξ ∈ H such that the above maps N × H → H are bilinear and x, y ∈ N, (xξ)y = x(ξy), ξ ∈ H. Moreover, x 7→ Lx, where Lxξ := xξ (ξ ∈ H) is a normal unital ∗-homomorphism, and x 7→ Rx, where Rxξ := ξx (ξ ∈ H) is a normal unital ∗-antihomomorphism. Definition 2.7. Let N be a von Neumann algebra, let (N, H) be a normal Hilbert N -bimodule, and let δ ∈ R+. Two n-tuples (ξ1, . . . , ξn) and (η1, . . . , ηn) of unit vectors in H are called δ- related, if there exists a family (ai)i∈I of operators in N , such that a∗ Xi∈I i ai =Xi∈I kaiξk − ηkaik2 < δ, aia∗ i = 1 k = 1, . . . , n. and Xi∈I We will use the following result which relates the δ-relatedness to approximate unitary equivalence in Hilbert N -bimodules: Theorem 2.8 ([Haa89, Theorem 2.3]). For every n ∈ N and ε > 0, there exists a δ = δ(n, ε) > 0, such that for all von Neumann algebra N and δ-related n-tuples (ξ1, . . . , ξn) and (η1, . . . , ηn) of unit vectors in a normal Hilbert N -bimodule H, there exists a unitary u ∈ U(N ) such that kuξk − ηkuk < ε, k = 1, . . . , n. Remark 2.9. As can be seen in the proof of [Haa89, Theorem 2.3], in order to show that the conclusion of Theorem 2.8 holds, it suffices to show the following: for every σ-strong 6 neighborhood V of 0 in N , there exist a1, . . . , ap ∈ N such that pXi=1 kaiξk − ηkaik2 < δ, pXi=1 a∗ i ai ≤ 1, pXi=1 a∗ i ai ∈ 1 + V, pXi=1 k = 1, . . . , n pXi=1 aia∗ i ≤ 1 aia∗ i ∈ 1 + V. (3) (4) (5) This is because we can obtain the conclusions of [Haa89, Lemma 2.5] out of (3), (4) and (5), which is enough to prove Theorem 2.8. We will use this variant in the proof Lemma 5.6. 3 Completely positive maps from m × m-matrices into an injective factor of type III1 The main result of this section is: Theorem 3.1. Let N be an injective factor of type III1 with separable predual, and let ϕ be a faithful normal state on N . Then for every finite set u1, . . . , un of unitaries in N , and every ε, δ > 0, there exists m ∈ N, a unital completely positive map T : Mm(C) → N , and n unitaries v1, . . . , vn in Mm(C), such that ψ = ϕ ◦ T is a normal faithful state on Mm(C), and kσϕ t ◦ T − T ◦ σψ kT (vk) − ukkϕ < ε, t k ≤ δt, t ∈ R, k = 1, . . . , n. In the following we let M = N ⋊σϕ R be the crossed product of N by σϕ with generators πσϕ (x) (x ∈ N ) and λ(s) (s ∈ R). We identify πσϕ (x) with x ∈ N . Let a be the (unbounded) self-adjoint operator for which λ(s) = exp(isa) (s ∈ R). For f ∈ L1(R), we define the Fourier transform f by f (s) = 1 √2πZ ∞ −∞ e−istf (t) dt, s ∈ R. In the sequel, von Neumann algebra-valued integrals are understood to be the σ-weak sense. Let s )s∈R be the dual action of σϕ on M . By [Haa79-2], there exists a normal faithful semifinite (θϕ operator-valued weight P : M+ → bN+ (bN+ is the extended positive part of N ) given by θϕ s (x) ds, (6) x ∈ M+. P (x) =Z ∞ −∞ Following [CT77], if we put c>0(cid:13)(cid:13)(cid:13)(cid:13)Z c m := span(cid:26)x ∈ M+; sup −c θϕ t (x) dt(cid:13)(cid:13)(cid:13)(cid:13) < ∞(cid:27) , For all x ∈ m, the σ-weak integral R c then the formula (6) for x ∈ m makes sense and P (x) ∈ N . Moreover, m ∋ x 7→ P (x) ∈ N defines a positive linear map. t (x) dt is σ-strongly convergent as c → ∞. The range of P is contained in πσϕ (N ), because πσϕ (N ) is the fixed point algebra in M under the dual action. Lemma 3.2. Let t 7→ x(t) be a σ-strongly* continuous function from R to N such that t 7→ kx(t)k is in L1(R) ∩ L∞(R). Put −c θϕ x :=Z ∞ −∞ λ(t)x(t) dt ∈ M. 7 Then x∗x ∈ m, and Proof. Note first, that x(t)∗x(t) dt. −∞ P (x∗x) = 2πZ ∞ x∗x =ZZR2 =ZZR2 x(s)∗λ(t − s)x(t) dsdt x(s)∗λ(t)x(s + t) dsdt. Put fn(s) = e−s2/(4n) (s ∈ R), and 2πZ ∞ gn(t) = 1 −∞ π(cid:17) 1 fn(s)e−its ds =(cid:16) n 2 e−nt2 (t ∈ R). Using that θϕ we have for every ψ ∈ M∗, s (y) = y (s ∈ R, y ∈ N ), θϕ s (λ(t)) = e−istλ(t) (s, t ∈ R)3 and the Fubini Theorem, hψ,Z ∞ −∞ θϕ u (x∗x)fn(u) dui =Z ∞ −∞Z ∞ −∞Z ∞ gn(t)(cid:18)Z ∞ =Z ∞ −∞ −∞ −∞ e−itufn(u)ψ(x(s)∗λ(t)x(s + t)) dsdtdu 2πψ(x(s)∗λ(t)x(s + t))ds(cid:19) dt. Since t 7→R ∞ −∞ ψ(x(s)∗λ(t)x(s + t))ds is in C0(R) and gn n→∞→ δ0 (weak∗ in C0(R)), we have n→∞hψ,Z ∞ lim −∞ θϕ u (x∗x)fn(u) dui = hψ, 2πZ ∞ −∞ x(s)∗x(s) dsi. Since ψ ∈ M∗ is arbitrary, θn(x∗x) ≥ 0 and fn ր 1 uniformly on compact sets, it follows that n→∞Z ∞ lim −∞ θϕ u (x∗x)fn(u)du = 2πZ ∞ −∞ x(s)∗x(s) ds (σ−strongly). Therefore x∗x ∈ m, and P (x∗x) = 2πZ ∞ −∞ x(t)∗x(t) dt. Lemma 3.3. Let a be the (unbounded) self-adjoint operator affiliated with M for which exp(ita) = λ(t) (t ∈ R) holds. Let α > 0, and let eα be the spectral projection of the oper- ator a corresponding to the interval [0, α]. Then for each x ∈ N , one has eαxeα ∈ m and P (eαxeα) =Z ∞ −∞ σϕ t (x) 1 − cos αt πt2 dt, x ∈ N. (7) Proof. It is sufficient to consider the case x ≥ 0, so we can assume that x = y∗y (y ∈ N ). For f ∈ L1(R) ∩ L∞(R) ∩ C(R), we have y f (a) = 1 √2πZ ∞ −∞ yλ(−t)f (t)dt = = 1 √2πZ ∞ √2πZ ∞ 1 −∞ −∞ λ(−t)σϕ λ(t)σϕ t (y)f (t) dt −t(y)f (−t) dt. 3Typewriter's note: Haagerup used the convention θϕ s (λ(t)) = eistλ(t). However, since the negative sign convention is widely accepted, we decided to change the definition. 8 Hence by Lemma 3.2, f (a)∗x f (a) ∈ m, and P ( f (a)∗x f (a)) =Z ∞ −∞ σϕ t (x)f (t)2 dt. For n > 2 α , let gn be the continuous function on R for which t ≤ 0, t ≥ α, n , α − 1 t ∈ [ 1 n ], n ] and [α− 1 gn(t) = 0, gn(t) = 1, and for which the graph is a straight line on [0, 1 n ]. Since a has no point spectrum, gn(a) ր eα (n → ∞). It is elementary to check that each gn is of the form gn = fn for a function fn ∈ L1(R) ∩ L∞(R) ∩ C(R) (use, for example, the fact that gn = n1 n ]). Hence gn(a)2 ∈ m, and by the Plancherel Theorem, we get n ] ∗ 1[0,α− 1 n , 1 [0, 1 P (gn(a)2) = P ( fn(a)∗ fn(a)) = kfnk2 21 = k fnk2 21. 2 = α < ∞, we have eα ∈ m and P (eα) = α1. Therefore eαM eα ⊆ m, and the Since supn k fnk2 restriction of P to eαM eα is a positive normal map. Hence for x ∈ N , σϕ t (x)fn(t)2 dt P (eαxeα) = lim n→∞ P (gn(a)xgn(a)) = lim (σ-strongly). n→∞Z ∞ −∞ Since kgn − 1[0,α]k2 n→∞→ 0, it follows that fn converges in L2(R) to the function f (t) = 1[0,α](s)eist ds 1 √2πZ ∞ t√2π i −∞ = − (eiαt − 1). Hence fn2 n→∞→ f2 in L1(R), with f2(t) = 1 Lemma 3.4. Let N be an injective factor of type III1 with separable predual, ϕ be a faithful normal state on N and let R be the hyperfinite II1 factor with tracial state τ . For every finite set u1, . . . , un of unitaries in N and every ε > 0, there exist x1, . . . , xn in the unit ball of R, a normal unital completely positive map T : R → N , such that ψ = ϕ ◦ T is a normal faithful state on R, and πt2 (1 − cos αt) (t ∈ R). Therefore (7) holds. σϕ t ◦ T = T ◦ σψ t , kT (xk) − ukkϕ < ε, t ∈ R, k = 1, . . . , n. (8) (9) Moreover, the spectrum of h = dψ/dτ is a closed interval [λ1, λ2], 0 < λ1 < λ2 < ∞, and h has no eigenvalues. Proof. 4 Let M = N ⋊σϕ R. By [Tak73-2], M has a normal faithful semifinite trace τ , such that τ ◦ θϕ s = e−sτ (s ∈ R). The trace τ can be constructed in the following way: Let ϕ be the dual weight of ϕ on M (cf. [Haa79]). Let a be the self-adjoint operator for which exp(ita) = λ(t) (t ∈ R). Then a is affiliated with the centralizer M ϕ of ϕ and 4Typewriter's note: Since some pages were missing from the original notes, we could not find all parts of the proofs of Lemma 3.4 and Theorem 3.1. We include the following proof for the reader's convenience. τ = ϕ(e−a · ) 9 in the sense of Pedersen-Takesaki [PT73]. By [Haa79-2], ϕ is on the subspace m given by Let α > 0, and let eα = 1[0,α](a). Then by Lemma 3.3, eα ∈ m and P (eα) = α1. Hence ϕeαMeα is a positive normal functional, and ϕ(eα) = α. Finally, ϕ(x) = ϕ ◦ P (x), x ∈ m. because where e−t dt = 1 − e−α < ∞, 0 τ (eα) = ϕ(e−aeα) =Z α e−aeα =Z α a =Z ∞ −∞ 0 e−t de(λ), λ de(λ) is the spectral resolution of a, and d ϕ(e(λ)) = dλ. Since N is of type III1, M is a type II∞ factor, and therefore eαM eα is a II1 factor. More- over, the injectivity of N implies that M is also injective, so that eαM eα is isomorphic to the hyperfinite factor R of type II1 by [Con76]. Claim. For any x ∈ N , we have lim 1 α α→∞(cid:13)(cid:13)(cid:13)(cid:13) P (eαxeα) − x(cid:13)(cid:13)(cid:13)(cid:13)ϕ t (x) − xkϕ ≤ ε for all t ∈ [−t0, t0]. Moreover, by(cid:12)(cid:12)(cid:12)(cid:12) This follows from a basic property of the Fej´er kernel (see e.g., [Kat68, Chapters I and VI]), but we include the proof for completeness. Let ε > 0. Choose t0 > 0 small enough so that kσϕ 1 t2 , we have 1 − cos(αt) 2 πα · παt2 (cid:12)(cid:12)(cid:12)(cid:12) ≤ = 0. dt = 0. By Lemma 3.3, we have lim παt2 1 − cos(αt) α→∞Zt≥t0 P (eαxeα) − x(cid:13)(cid:13)(cid:13)(cid:13)ϕ ≤ lim sup α→∞ Z ∞ 1 α lim sup α→∞ (cid:13)(cid:13)(cid:13)(cid:13) t (x) − xkϕ −∞ kσϕ α→∞ Zt≥t0 1 − cos(αt) dt παt2 1 − cos(αt) παt2 dt ≤ ε + 2kxkϕ lim sup = ε. Since ε > 0 is arbitrary, we obtain the conclusion. Let n ≥ 1, u1, . . . , un ∈ U(N ) and ε, δ > 0 be given. By the above claim, we may choose α > 0 large enough so that < ε, 1 ≤ k ≤ n. (10) (cid:13)(cid:13)(cid:13)(cid:13) 1 α P (eαukeα) − uk(cid:13)(cid:13)(cid:13)(cid:13)ϕ Define T := α−1PeαMeα : R = eαM eα → N and ψ := ϕ ◦ T = 1 α ϕ ◦ P (eα · eα) = 1 α ϕ(eα · eα). Then T is a normal unital completely positive map, and ψ is a normal faithful state on R. By (10) we have kT (xk) − ukkϕ < ε, 1 ≤ k ≤ n, 10 where xk := eαukeα (1 ≤ k ≤ n) are in the unit ball of R. Moreover, since eα ∈ M ϕ and since σϕ t ◦ P = P ◦ σ ϕ t (t ∈ R). By construction, we have t (t ∈ R), we have σϕ h := t ◦ T = T ◦ σψ 1 − e−α dψ dτ = α exp(a)eα, 1 + = (0,∞). which has no atoms and the spectrum of h is a closed bounded interval in R∗ Lemma 3.5. Let B ⊂ A be an inclusion of unital C∗-algebras and E : A → B be a unital completely positive map. Let h ∈ A be a self-adjoint element with σ(h) ⊂ [λ1, λ2], where σ( · ) denotes the spectrum and λ1 < λ2 are reals. Then σ(E(h)) ⊂ [λ1, λ2]. Proof. Let λ < λ1. Then h − λ is positive and invertible. Take a nonzero x ∈ A such that (h − λ) 2 ) = 1. The left hand side is dominated by kxk2E(h − λ), whence E(h − λ) ≥ kxk−21, showing that E(h) − λ1 is invertible. Thus λ /∈ σ(E(h)). Similarly, σ(E(h)) ∩ (λ2,∞) = ∅ holds. Therefore σ(E(h)) ⊂ [λ1, λ2]. Proof of Theorem 3.1. We may assume that 0 < ε < 1. By Lemma 3.4, there exist a normal unital completely positive map T : R → N and x1, . . . , xn in the unit ball of R satisfying σϕ t ◦ T = T ◦ σψ t , (t ∈ R), where ψ := ϕ ◦ T : R → N and 2 x = 1, so that E((h − λ) 2 xx∗(h − λ) 1 1 kT (xk) − ukkϕ < , 1 ≤ k ≤ n. (11) ε2 16 Let h := dψ/dτ ∈ R+. Then by Lemma 3.4, σ(h) = [λ1, λ2] for some positive reals λ1 < λ2 and h does not have a point spectrum. Since log( · ) is continuous on [λ1, λ2], continuous functional calculus guarantees that there exists δ′ > 0 such that for all a, b ∈ R+, we have the following implication σ(a), σ(b) ⊂ [λ1, λ2] and ka − bk < δ′ ⇒ k log a − log bk < . (12) δ 4 By using the spectral decomposition of h, we can choose a partition of unity {pi}ℓ {µi}ℓ + such that i=1 in R∗ i=1 in R and 1 ℓ τ (pi) = , hpi = pih, k(log h)pi − (log µi)pik < 1 4 δ, khpi − µipik < δ′, i=1 µipi, and we have kh − h0k < δ′ and k log h − log h0k < 1 4 δ. i=1 so that h0 =Pℓ σ(h0) ⊂ [λ1, λ2]. i=1 µipi satisfies for all (1 ≤ i ≤ ℓ). Let h0 :=Pℓ Moreover, we may arrange {µi}ℓ (13) (14) Since R is hyperfinite, there exists a type I subfactor F of R so that pi ∈ F (1 ≤ i ≤ ℓ) and kxk − EF (xk)kϕ < , 1 ≤ k ≤ n, (15) ε2 16 where EF : R → F denotes the τ -preserving conditional expectation. Put TF := TF : F → N and yk := EF (xk) (1 ≤ k ≤ n). Combining (11) and (15), for all 1 ≤ k ≤ n, we have (use the Schwarz inequality for completely positive maps) kTF (yk) − ukkϕ ≤ kT (yk) − T (xk)kϕ + kT (xk) − ukkϕ ≤ kyk − xkkψ + kT (xk) − ukkϕ < . ε2 8 (16) 11 Then we follow the argument of [Haa89, Lemma 6.2]. Take v1, . . . , vk ∈ U(F ) such that Then again by the Schwarz inequality for completely positive maps and (16), yk = vkyk, yk := (y∗ kyk) 1 2 , 1 ≤ k ≤ n. kykkψ ≥ kTF (yk)kϕ > kukkϕ − ε2 8 . Since k(y∗ kyk) 1 2 kψ = kykkψ and yk2 + (1 − yk)2 ≤ 1 (because 0 ≤ yk ≤ 1), we have kvk − ykk2 ψ ≤ 1 − k ykk2 ψ = k1 − ykk2 < 1 − (1 − ε2 8 )2 < . ψ ε2 4 Therefore since ε2 < ε, kTF (vk) − ukkϕ ≤ kTF (vk − yk)kϕ + kTF (yk) − ukkϕ < kvk − ykkψ + < ε. ε2 8 Next, set χ := τ (h0 · ) ∈ (R∗)+. Note that σχF t = σχ t F (t ∈ R), since h0 ∈ F . Then by (13), we have khit − hit d ds 0 k =(cid:13)(cid:13)(cid:13)(cid:13)Z 1 ≤Z 1 0 keist log h(t log h − t log h0)ei(1−s)t log h0k ds eist log hei(1−s)t log h0 ds(cid:13)(cid:13)(cid:13)(cid:13) 0 ≤ k log h − log h0k t ≤ δt 4 . (17) On the other hand, hF := dψF /dτF ∈ F+ is equal to EF (h). Therefore by Lemma 3.5, σ(hF ) ⊂ [λ1, λ2]. Moreover, since EF (h0) = h0, we have khF − h0k = kEF (h − h0)k ≤ kh − h0k < δ′. This shows by (12) and (14) that k log hF − log h0k < δ have 4 . Therefore by the same argument, we For all t ∈ R and x ∈ F , khit F − hit 0 k ≤ δt 4 , t ∈ R. (18) kσϕ t ◦ TF (x) − TF ◦ σψF t By (17), we have (x)k = kT (σψ t t (x) − σψF (x)k t (x) − σψF t (x) − σχ t ≤ kσψ ≤ kσψ (x))k t (x)k + kσχF t (x) − σψF t (x)k. kσψ t (x) − σχ t (x)k = khitxh−it − hit 0 xh−it 0 k 0 k + kh−it − h−it 0 k)kxk ≤ (khit − hit ≤ δtkxk. 1 2 12 Similarly, by (18), kσχF t (x) − σψF t Fk + kh−it 0 − h−it F k)kxk (x)k ≤ (khit 0 − hit δtkxk. t ◦ TF (x) − TF ◦ σψF ≤ 1 2 t These altogether imply that kσϕ (x)k ≤ δtkxk. 4 Q-stable states on III1 factors For technical reasons we shall consider a special class of normal faithful states, which we call Q-stable states, because they have nice properties with respect to certain operations involving rationals. Definition 4.1. A normal faithful state ϕ on a von Neumann algebra N is called Q-stable, if for every m ∈ N, there exist m isometries u1, . . . , um ∈ N with orthogonal range projections, such that uiu∗ i = 1, mXi=1 ϕui = 1 m uiϕ, i = 1, . . . , m. Theorem 4.2. Every factor of type III1 with separable predual has a Q-stable normal faithful state. For the proof of Theorem 4.2, we shall need two lemmas: Lemma 4.3. The Araki -- Woods factor R∞ has a Q-stable normal faithful state. Proof. Let Rλ (0 < λ < 1) be the Powers factor of type IIIλ, and let ϕλ be the product state on Rλ. Then ϕλ is normal and faithful, and σϕλ has period −2π/ log λ. Then the centralizer (Rλ)ϕλ is a type II1 factor (cf. [Con73, Th´eor`eme 4.2.6]), and there exists an isometry u ∈ Rλ such that σϕλ t (u) = λitu, t ∈ R. This implies that σϕλ 1.6], we have t (uu∗) = uu∗ (t ∈ R), i.e., uu∗ ∈ (Rλ)ϕλ. Moreover, by [Tak73, Lemma ϕλu = λuϕλ, and hence ϕλ(uu∗) = (ϕλu)(u∗) = λϕλ(u∗u) = λ. Assume now, that λ = 1 m , m ∈ N, m ≥ 2. Then we can choose m equivalent orthogonal projections p1, . . . , pm ∈ (Rλ)ϕλ with sum 1, such that p1 = uu∗. Next, choose partial isometries v1, . . . , vm ∈ (Rλ)ϕλ such that i vi = p1, viv∗ v∗ i = 1, . . . , m. i = pi, Put ui = viu, and ϕλui = λuiϕλ, i = 1, . . . , m. Put now i = 1, . . . , m. Then u1, . . . , um are m isometries in Rλ, such thatPm i=1 uiu∗ i = 1, (P, ϕ) = (R 1 m , ϕ 1 m ). ∞Om=2 Then it is clear from the above computations, that ϕ is a Q-stable normal faithful state on P (observe that it is sufficient to consider m ≥ 2 case in Definition 4.1). Moreover, P is an ITPFI factor for which the the asymptotic ratio set r∞(P ) contains { 1 m ; m ∈ N}. Since r∞(P )∩ R+ is a closed subgroup of R+, we have r∞(P ) ⊇ R+. Therefore by Araki -- Woods' Theorem [AW68, Theorem 7.6], P ∼= R∞ holds. 13 Lemma 4.4. Let N be a factor of type III1 with separable predual. Then there exists a normal faithful conditional expectation of N onto a subfactor P isomorphic to R∞. Proof. 5 We can write R∞ as an infinite tensor product R∞ = (Pk, ωk), ∞Ok=1 where each Pk is a copy of the 2 × 2 matrices M2(C) and (ωk)∞ k=1 is a sequence of normal faithful states on M2(C). Let ϕ be a fixed normal faithful state on N . Since N is properly infinite, we have N ⊗ M2(C) ∼= N . Moreover, by Connes-Størmer transitivity theorem [CS78], we can choose a ∗-isomorphism Φ : N ⊗ M2(C) → N such that k(ϕ ⊗ ω1) ◦ Φ−1 − ϕk < 1 2 . Put F1 = Φ(C ⊗ M2(C)), and ϕ1 = (ϕ ⊗ ω1) ◦ Φ−1. Then F1 is a type I2 subfactor of N . Moreover, it holds that N ∼= F1 ⊗ F c 1 ∩ N is the relative commutant, and ϕ1 = ϕ1F1 ⊗ ϕ1F c 1 . Moreover, we have 1 , where F c 1 = F ′ (F1, ϕ1F1 ) ∼= (P1, ω1). 1 , we can find a type I2-subfactor F2 ⊂ F c 1 , Using the same arguments to the type III1 factor F c a normal faithful state ϕ′ 2 on F c 1 , such that kϕ′ 1 k < 1 4 , 2(F1⊗F2)c , 2F2 ) ∼= (P2, ω2). Thus, if we put ϕ2 = ϕ1F1 ⊗ ϕ′ 2F2 ⊗ ϕ′ 2 − ϕ1F c ϕ′ 2 = ϕ′ and (F2, ϕ′ 2, we have kϕ1 − ϕ2k < 1 4 , and ϕ2 = ϕ2F1 ⊗ ϕ2F2 ⊗ ϕ2(F1⊗F2)c , (F1, ϕ2F1 ) ∼= (P1, ω1), (F2, ϕ2F2 ) ∼= (P2, ω2). Proceeding in this way, we obtain a sequence (Fk)∞ of N , and a sequence (ϕk)∞ k=1 of normal faithful states on N , such that k=1 of mutually commuting type I2-subfactors kϕk − ϕk−1k < 2−k, k ≥ 2, and such that for fixed m ∈ N: ϕm = ϕmF1 ⊗ ϕmF2 ⊗ . . . ⊗ ϕmFm ⊗ ϕm(F1⊗...⊗Fm)c, and (Fi, ϕmFi) ∼= (Pi, ωi) i = 1, . . . , m. Let ϕ be the norm limit in N∗ of the sequence (ϕk)∞ fail to be faithful. From the properties of ϕk, we have for all m ∈ N, k=1. Then ϕ is a normal state, but it can and ϕ = ϕF1 ⊗ ϕF2 ⊗ . . . ⊗ ϕFm ⊗ ϕm(F1⊗...⊗Fm)c , (Fm, ϕFm ) ∼= (Pm, ωm). 5Typewriter's note: this result has been extended by Haagerup -- Musat [HM09, Theorem 3.5], where the authors study more general embeddings of ITPFI type III factors into type III factors as the range of normal faithftul conditional expectations. 14 Let rk be the ratio between the largest and the smallest eigenvalues of dωk/dTr. Let u ∈ U(Pk). We may assume that ωk = Tr(hk · ), hk := 1 is positive, then diag(rk, 1) (rk ≥ 1). Then if a =(cid:18)x y z w(cid:19) ∈ Pk 1+rk uωku∗(a) = Tr(hku∗au) = Tr((u∗au) x + w 1 + rk Tr(u∗au) = 1 1 2 hk(u∗au) 1 2 ) ≥ 1 + rk ≥ r−1 k ( = r−1 rkx 1 + rk k ωk(a). w + ) 1 + rk Similarly, uωku∗(a) ≤ rkωk(a) holds. This shows that r−1 k ωk ≤ uωku∗ ≤ rkωk. Hence for all u ∈ U(Fk), r−1 k ϕ ≤ uϕu∗ ≤ rkϕ. Thus, ϕ and uϕu∗ have the same support projection in N , i.e., with e = supp(ϕ), we have This shows that e ∈ (S∞ k=1 is a sequence of commuting subfactors of eN e. Moreover, the restriction ϕe of ϕ to eN e is a normal faithful state on eN e, and ueu∗ = e, u ∈ U(Fk), k ∈ N. k=1 Fk)′ ∩ N . Put Gk = eFk. Then (Gk)∞ mOk=1 (G1 ⊗ . . . ⊗ Gm, ϕeG1⊗...⊗Gm) ∼= (Pk, ωk) for all m ∈ N. Let P be the von Neumann algebra generated byS∞ k=1 Gk. Then (P, ϕP ) ∼= (Pk, ωk). ∞Ok=1 In particular, P ∼= R∞. Moreover, since ϕe = ϕeG1 ⊗ . . . ϕeGm ⊗ ϕe(G1⊗...⊗Gm)c, where (G1 ⊗ ··· ⊗ Gm)c denotes the relative commutant ofSm for all m ∈ N, and hence also σϕe t (P ) = P, t ∈ R. Thus by [Tak72], there exists a normal faithful conditional expectation of eN e onto P . This completes the proof, since eN e is isomorphic to N . σϕe t (G1 ⊗ ··· ⊗ Gm) = G1 ⊗ ··· ⊗ Gm, k=1 Gk in eN e, we have t ∈ R Proof of Theorem 4.2. Let N be a type III1 factor with separable predual. By Lemmata 4.3 and 4.4, we can choose a normal faithful conditional expectation E of N onto a subfactor P of N isomorphic to R∞. Moreover, we can choose a Q-stable normal faithful state ω on P . Put ϕ = ω ◦ E. Then it follows from the bimodule property of conditional expectations [Tom58, Theorem 1] that ϕ is a Q-stable normal faithful state on N . Theorem 4.5. Let ϕ be a Q-stable normal faithful state on a von Neumann algebra N , let m ∈ N, and let q1, . . . , qm be m positive rational numbers with sum 1. Then there exists a type Im subfactor F of N , such that (a) ϕ = ϕF ⊗ ϕF c . 15 (b) ϕF c is Q-stable. (c) dϕF /dTrF has eigenvalues (q1, . . . , qm). Here, TrF denotes the trace on F for which TrF (1) = m. We prove first: Lemma 4.6. Let ϕ be a Q-stable normal faithful state on a von Neumann algebra N , and let q1, . . . , qm be positive rational numbers with sum 1. Then there exist isometries u1, . . . , um ∈ N with orthogonal ranges, such that uiu∗ i = 1, mXi=1 ϕui = qiuiϕ, i = 1, . . . , m. Proof. We can choose integers p, p1, . . . , pm ∈ N such that i = 1, . . . , m. , qi = pi p Note thatPm vi1, . . . , vipi in N with orthogonal ranges, such that i=1 pi = p. By Definition 4.1, for each i ∈ {1, . . . , m} we can choose pi isometries vij v∗ ij = 1 and ϕvij = 1 pi vijϕ, j = 1, . . . , pi. piXj=1 Moreover, since the set {(i, j); 1 ≤ i ≤ m, 1 ≤ j ≤ pi} containsPm also find isometries wij ∈ N (1 ≤ i ≤ m, 1 ≤ j ≤ pi) with orthogonal ranges, such that i=1 pi = p elements, we can mXi=1 piXj=1 wij w∗ ij = 1 and ϕwij = 1 p wij ϕ, 1 ≤ i ≤ m, 1 ≤ j ≤ pi. Put now Then ui := piXj=1 wij v∗ ij, i = 1, . . . , m. u∗ i ui = uiu∗ i = piXj=1 mXi=1 mXi=1 vij v∗ ij = 1, wij w∗ ij = 1, piXj=1 and since ϕwij = 1 p wij ϕ and v∗ ij ϕ = 1 pi ϕv∗ ij for all (i, j), we get ϕui = piXj=1 ϕwij v∗ ij = pi p piXj=1 wij v∗ ij ϕ = qiuiϕ. This proves Lemma 4.6. 16 Proof of Theorem 4.5. Choose m isometries u1, . . . , um ∈ N satisfying the conditions in Lemma 4.6. We can define a ∗-isomorphism Φ of N ⊗ Mm(C) onto N by where (eij)m i,j=1 are the matrix units in Mm(C). Then using ϕui = λiuiϕ, we get Φ mXi,j=1 (ϕ ◦ Φ) xij ⊗ eij := xij ⊗ eij = mXi,j=1 uixij u∗ j , mXi,j=1 mXi,j=1 mXi,j=1 mXi=1 = = ϕ(uixij u∗ j ) qiϕ(xij u∗ j ui) qiϕ(xi). Hence where ω is the state on Mm(C) for which ϕ ◦ Φ = ϕ ⊗ ω, dω dTr = q1 0 ... 0  0 q2 ··· ··· . . . 0 0 ... 0 qm .  N = F · F c ∼= F ⊗ F c. Put now F := Φ(C ⊗ Mm(C)). Then the relative commutant of F in N ⊗ Mm(C) is Φ(N ⊗ C). Since ϕ ◦ Φ = ϕ ⊗ ω, ϕ itself is a tensor product state with respect to the decomposition Moreover, dϕF /dTrF has eigenvalues (q1, . . . , qm). Let Φ0 be the isomorphism of N onto F c given by Then ϕF ◦ Φ0 = ϕ. Therefore ϕF c is a Q-stable normal faithful state on F c. Φ0(x) = Φ(x ⊗ 1), x ∈ N. 5 Proof of Main Theorem In this section we prove the main result of the paper: Theorem 5.1. Every injective factor N of type III1 on a separable Hilbert space is isomorphic to the Araki -- Woods factor R∞. We need preparations. In this section, for each von Neumann algebra N , we fix a standard form (N, H, J,P ♮). For each ϕ ∈ (N∗)+, we denote by ξϕ the unique representing vector in P ♮ [Haa75]. Lemma 5.2. Let N be a properly infinite factor with separable predual and with a normal faithful state ϕ, let F be a finite dimensional σϕ-invariant subfactor of N , and let T : F → N be a unital completely positive map, such that ϕ ◦ T = ϕ and kσϕ t ◦ T − T ◦ σϕF t k ≤ δt, t ∈ R, (19) where δ > 0 is a constant. Then there exists a norm-continuous map a : R → N such that 17 a(t)∗a(t) dt = 1 (σ-strongly), −∞ −∞ (a) Z ∞ (b) Z ∞ (c) Z ∞ (d) (cid:13)(cid:13)(cid:13)(cid:13)T (x) −Z ∞ −∞ e−tεF,ϕ(a(t)a(t)∗) dt = 1 (σ-strongly), −∞ ka(t)ξϕ − e−t/2ξϕa(t)k2 dt < δ 8 , a(t)∗xa(t) dt(cid:13)(cid:13)(cid:13)(cid:13) ≤ δ 1 2 kxk, x ∈ F , Hence for x ∈ F , where εF,ϕ is the normal faithful conditional expectation of N onto F that leaves the state ϕ invariant. Proof. Let f be the function f (t) := (πδ)− 1 4 exp(cid:18)− 1 2δ t2(cid:19) , t ∈ R, and let g be the Fourier-transformed of f : Note that 4 π(cid:19) 1 g(s) :=(cid:18) δ exp(cid:18)− Z ∞ f (t)2dt =Z ∞ −∞ −∞ δ 2 s2(cid:19) , s ∈ R. g(s)2 ds = 1. By [Haa85, Proposition 2.1], there exists an operator a ∈ N such that In particular, a∗a = 1, i.e., a is an isometry. Put T (x) = a∗xa, x ∈ F. a(t) = 1 √2πZ ∞ −∞ e−is(t−δ/4)g(s)σϕ s (a) ds, t ∈ R. Since t 7→ e−is(t−δ/4)g(s) is a continuous map from R to L1(R), the map t 7→ a(t) is a norm- continuous map from R to N . Using the Plancherel formula in L2(R, H), we get for all ξ ∈ H. Hence Z ∞ −∞ ka(t)ξk2 dt =Z ∞ −∞ g(s)2kσϕ s (a)ξk2 ds = kξk2 Z ∞ −∞ a(t)∗a(t) dt = 1 (σ-weakly). Since the convergence of the integral is monotone, we get (a). Using again the Plancherel formula, we get for ξ, η ∈ H and x ∈ F , s (a)ξ, σϕ s (a)ηi ds s ◦ T ◦ σϕ −s(x)ξ, ηi ds. −∞ Z ∞ −∞hxa(t)ξ, a(t)ηi dt =Z ∞ =Z ∞ a(t)∗xa(t) dt =Z ∞ Z ∞ −∞ −∞ −∞ g(s)2hxσϕ g(s)2hσϕ 18 g(s)2σϕ s ◦ T ◦ σϕ −s(x) ds. (20) Note that the left hand side of (20) converges σ-strongly, because F = span(F+) and for x ∈ F+, the integral converges σ-weakly and the convergence is monotone. Therefore by (19), for each x ∈ F we get (cid:13)(cid:13)(cid:13)(cid:13)T (x) −Z ∞ −∞ a(t)∗xa(t)dt(cid:13)(cid:13)(cid:13)(cid:13) ≤ δkxkZ ∞ π(cid:19) 1 =(cid:18) δ 2 kxk −∞ sg(s)2 ds 1 ≤ δ 2kxk. This proves (d). Since g(s) has the analytic extension to the function g : C → C, and since the integrals −∞ g(s + iu) ds =(cid:18) 4π δ (cid:19) 1 Z ∞ 4 2 u2 δ , e u ∈ R are uniformly bounded for u on bounded subsets of R, it follows that a(t) is analytic with respect to σϕ (in the sense of [PT73]) and that σϕ α(a(t)) = 1 √2πZ ∞ −∞ e−i(s−α)(t− δ 4 )g(s − α)σϕ s (a) ds, for all α ∈ C. To prove (c), we use the equality ξϕa(t) = Jϕa(t)∗ξϕ = ∆ 1 2 ϕ a(t)ξϕ = σϕ −i/2(a(t))ξϕ. Hence e−t/2ξϕa(t) = δ 8 e− √2πZ ∞ −∞ e−is(t− δ 4 )g(s + i 2 )σϕ s (a)ξϕ ds. Using the Plancherel formula, we get Z ∞ −∞ ka(t)ξϕ − e−t/2ξϕa(t)k2 dt =Z ∞ −∞ g(s) − e− δ 8 g(s + i 2 )2kaξϕk2 ds. On the other hand, g(s + i above integral is equal to 2 ) is the Fourier -- Plancherel transformed of et/2f (t). Therefore the It is easy to compute that for γ ∈ R, Therefore Z ∞ −∞ This proves (c). Put now f (t)2(cid:18)1 − e− δ 8 + t 2(cid:19)2 dt. f (t)2eγt dt = exp( 1 4 γ2δ). −∞ Z ∞ Z ∞ f (t)2(cid:18)1 − e− −∞ δ 8 + t 2(cid:19)2 dt = 2(1 − e− δ 16 ) < δ 8 . A(t) := e−tεF,ϕ(a(t)a(t)∗), t ∈ R. 19 Since εF,ϕ is a normal faithful conditional expectation of N onto F that leaves ϕ invariant, we have for x ∈ F , that ϕ(A(t)x) = e−tϕ(a(t)a(t)∗x). By the KMS-condition, it follows that if a, b ∈ N and a is σϕ-analytic, then ϕ(ab) = ϕ(bσϕ −i(a)) (cf. [Haa79, Theorem 3.2]). Hence for x ∈ F , ϕ(A(t)x) = e−tϕ(a(t)∗xσϕ −i(a(t))). Using e−tσϕ −i(a(t)) = δ 4 e− √2πZ ∞ −∞ e−is(t− δ 4 )g(s + i)σϕ s (a) ds, we get by the Plancherel formula, that Z ∞ −∞ ϕ(A(t)x) dt =Z ∞ −∞hx(e−tσϕ 4Z ∞ −∞ δ = e− −i(a(t)))ξϕ, a(t)ξϕi dt g(s + i)g(s)hxσϕ s (a)ξϕ, σϕ s (a)ξϕi ds. Since ϕ ◦ T = ϕ, it holds that Hence s (a)ξϕ, σϕ −s ◦ T ◦ σϕ s (x) = ϕ(x), ϕ(A(t)x) dt = e− g(s + i)g(s) ds. Since g(s + i) is the Fourier -- Plancherel transformed of f (t)et, we get −∞ hxσϕ Z ∞ Z ∞ −∞ −∞ δ s (a)ξϕi = ϕ ◦ σϕ 4 ϕ(x)Z ∞ g(s + i)g(s) ds =Z ∞ Z ∞ Z ∞ ψ(A(t)) dt = ψ(1), A(t) dt = 1 −∞ −∞ ψ ∈ F∗, (σ-weakly). −∞ f (t)2et dt = e δ 4 . that is, we have This proves (b). Since F is finite-dimensional and ϕ is faithful on F , every ψ ∈ F∗ is of the form ϕ( · x), x ∈ F . This shows that Lemma 5.3. Let N, ϕ, F and εF,ϕ be as in Lemma 5.2. Let λ > 0 and assume that c1, . . . , cs are operators in F c = F ′ ∩ N such that ϕci = λciϕ, i = 1, . . . , s, Then for all x ∈ N , c∗ i ci = 1. sXi=1 εF,ϕ sXi=1 i! = λεF,ϕ(x). cixc∗ 20 Proof. It is sufficient to check the formula for x ∈ N of the form x = ab, a ∈ F, b ∈ F c. For z ∈ F c, εF,ϕ(z) commutes with every element in F . Hence εF,ϕ(z) is a scalar multiple of the identity. Using that εF,ϕ leaves ϕ invariant, we get Therefore εF,ϕ(z) = ϕ(z)1, z ∈ F c. εF,ϕ sXi=1 cixc∗ i!! cibc∗ i! = εF,ϕ a sXi=1 i! a = ϕ sXi=1 i ci! a = λϕ sXi=1 cibc∗ bc∗ = λϕ(b)a = λεF,ϕ(x). Lemma 5.4. Let ϕ be a Q-stable normal faithful state on an injective factor N of type III1 with separable predual. Let u1, . . . , un ∈ U(N ), let δ > 0. Then there exist a finite dimensional σϕ-invariant subfactor F of N and unitaries v1, . . . , vn ∈ U(F ), such that for every σ-strong neighborhood V of 0 in N , there exists a finite set b1, . . . , br of operators in N with the following properties: b∗ i bi ∈ 1 + V and b∗ i bi ≤ 1. rXi=1 i! ∈ 1 + V and bib∗ εF,ϕ rXi=1 i! ≤ 1. bib∗ (a) rXi=1 (b) εF,ϕ rXi=1 rXi=1 rXi=1 (d) (c) kbiξϕ − ξϕbik2 < δ. kbiuk − vkbik2 ϕ < δ, k = 1, . . . , n. Proof. Put δ1 = min(δ2/4, δ). By Theorem 3.1, there exist m ∈ N, a unital completely positive map T0 : Mm(C) → N and unitaries w1, . . . , wn ∈ Mm(C) such that ψ := ϕ ◦ T ∈ Mm(C)∗ satisfies kσϕ t ◦ T0 − T0 ◦ σψ kT0(wk) − ukkϕ < ε, t k ≤ δ1 2 t, t ∈ R, k = 1, . . . , n. 1, . . . , q′ Let {q′ m} be the spectrum of dψ/dTr ∈ Mm(C)+ where the multiplicity is taken into account. Let {q1, . . . , qm} be positive rationals with sum 1, and let χ on Mm(C)+ such that dχ/dTr has the same spectral projections as dψ/dTr but the eigenvalues replaced by {q1, . . . , qm}. Since keia − eibk ≤ ka − bk for self-adjoint operators a, b, (cf. (17) in Theorem 3.1), we may arrange qi's so that the following inequality holds: kσψ t − σχ t kMm(C) ≤ δ1 2 t, t ∈ R. 21 Since ϕ is Q-stable and qi's are rationals, by Theorem 4.5, there exists a finite-dimensional subfactor F ⊂ N and a state-preserving ∗-isomorphism Φ : (Mm(C), χ) → (F, ϕF ) such that ϕF c is Q-stable. Define T := T0 ◦ Φ−1 : F → N and vk := Φ0(wk) ∈ U(F )(1 ≤ k ≤ n). Then if x = Φ(y) (y ∈ Mm(C)) and t ∈ R, we have kσϕ t ◦ T (x) − T ◦ σϕF t t ◦ T0(y) − T0 ◦ Φ−1 ◦ σϕF t ◦ T0(y) − T0 ◦ σχ t ◦ T0(y) − T0 ◦ σψ t (y)k t (y)k + kT (σψ t ◦ Φ(y)k (x)k = kσϕ = kσϕ ≤ kσϕ ≤ δ1tkyk. t (y) − σχ t (y))k Therefore we obtain T (1) = 1, kσϕ t ◦ T − T ◦ σϕF t ϕ ◦ T = ϕF , k ≤ δ1t, 1 , 1 2 t ∈ R, k = 1, . . . , n. kT (vk) − ukkϕ < δ Choose now a norm-continuous function t 7→ a(t) of R into N , such that the conditions (a), (b), (c) and (d) in Lemma 5.2 are satisfied with respect to δ1 instead of δ. Then using (d), we have kuk −Z ∞ −∞ a(t)∗vka(t) dtkϕ < 2δ 1 2 1 ≤ δ a(t)∗a(t) dt = 1, it follows that for k = 1, . . . , n. Using thatZ ∞ Z ∞ −∞ ka(t)uk − vka(t)k2 −∞ ϕ dt = 2 − 2ReZ ∞ −∞ha(t)ukξϕ, vka(t)ξϕi dt = 2 − 2Re(cid:18)hukξϕ,Z ∞ ≤ kuk −Z ∞ −∞ −∞ a(t)∗vka(t) dtkϕ a(t)∗vka(t)ξϕ dti(cid:19) < δ. Let now V be a σ-strong neighborhood of 0 in N . It is no loss of generality to assume that V is open. For sufficiently large γ ∈ R+, we have: (a') Z γ (b') Z γ (c') Z γ (d') Z γ a(t)∗a(t) dt ≤ 1. −γ a(t)∗a(t)dt ∈ 1 + V and Z γ e−tεF,ϕ(a(t)a(t)∗)dt ∈ 1 + V and Z γ −γ ka(t)ξϕ − e−t/2ξϕa(t)k2 dt < δ1 8 ≤ δ. −γ −γ −γ −γ ka(t)uk − vka(t)k2 ϕ dt < δ. e−tεF,ϕ(a(t)a(t)∗) dt ≤ 1. Since t 7→ a(t) is norm-continuous, we can approximate (in norm) the above N -valued Riemann integrals over [−γ, γ] to get the following statements: there exists an h0 > 0 such that when 0 < h < h0, the operators aj = h− 1 2 a(jh), j ∈ Z satisfy the following relations: 22 (a") (b") (c") (d") pXj=−p pXj=−p pXj=−p pXj=−p a∗ j aj ∈ 1 + V. e−jhεF,ϕ(aja∗ j ) ∈ 1 + V. kajξϕ − e− 1 2 jhξϕajk2 < δ. kajuk − vkajk2 ϕdt < δ, where p is the largest integer smaller than γ/h0. Moreover, since the Riemann sum is norm- convergent, by multiplying a scalar c > 0 to aj's which is sufficiently close to 1 if necessary, we may moreover assume that a∗ j aj ≤ 1 pXj=−p pXj=−p e−jhεF,ϕ(aja∗ j ) ≤ 1. (21) (22) Choose now h ∈ (0, h0), such that exp(h) ∈ Q. This implies that the numbers qj = e−jh, j ∈ Z are rational. Since the restriction of ϕ to F c is Q-stable, there exists for each j ∈ Z a finite set of operators cj1, . . . , cjs(j) in F c such that ϕcji = e−jhcjiϕ, i = 1, . . . , s(j) and c∗ jicji = 1. s(j)Xi=1 (Here we use Lemma 4.6 together with the fact that ϕ = ϕF ⊗ ϕF c , when F is σϕ-invariant). Put Then by (21), bji = cjiaj, j ≤ p, 1 ≤ i ≤ s(j). a∗ j aj ∈ 1 + V and pXj=−p s(j)Xi=1 b∗ jibji ≤ 1, e−jhεF,ϕ(aja∗ j ) ∈ 1 + V, and by (22) and Lemma 5.3, (a'") s(j)Xi=1 pXj=−p εF,ϕ s(j)Xi=1 pXj=−p and εF,ϕ s(j)Xi=1 pXj=−p (b'") b∗ jibji = pXj=−p ji = ji ≤ 1. bjib∗ pXj=−p bjib∗ The equality ϕcji = e−jhcjiϕ implies that ξϕcji = e− 1 2 jhcjiξϕ. 23 Therefore (c'") pXj=−p s(j)Xi=1 kbjiξϕ − ξϕbjik2 = = < ε. Finally, using that vk ∈ F and aji ∈ F c, we get kcji(ajξϕ − e− 1 2 jhξϕaj)k2 s(j)Xi=1 kajξϕ − e− 1 pXj=−p pXj=−p 2 jhξϕajk2 (d'") pXj=−p s(j)Xi=1 kbjiuk − vkbjik2 ϕ = = pXj=−p pXj=−p < δ. s(j)Xi=1 kcji(ajuk − vkaj)k2 ϕ kajuk − vkajk2 ϕ This completes the proof of Lemma 5.4. In the proof of the following lemma, it is essential that injective type III1 factors (on a separable Hilbert space) have trivial bicentralizers. Lemma 5.5. Let ϕ be a Q-stable normal faithful state on an injective factor N of type III1 with separable predual. Let u1, . . . , un ∈ U(N ), and let δ > 0. Then there exists a finite dimensional σϕ-invariant subfactor F of N and v1, . . . , vn ∈ U(F ), such that the for every σ-strong neighborhood V of 0 in N , there exists a finite set a1, . . . , ap of operators in N with the following properties: (a) (b) (c) (d) pXi=1 pXi=1 pXi=1 pXi=1 pXi=1 pXi=1 a∗ i ai ≤ 1. aia∗ i ≤ 1. a∗ i ai ∈ 1 + V and aia∗ i ∈ 1 + V and kaiξϕ − ξϕaik2 < δ. kaiuk − vkaik2 ϕ < δ, k = 1, . . . , n. Proof. Choose an F and v1, . . . , vn ∈ U(F ) satisfying the properties of Lemma 5.4 with respect to (u1, . . . , un, δ), and let V be a σ-strongly open neighborhood of 0 in N . By Lemma 5.4, there exists b1, . . . , br ∈ N such that (a') rXi=1 (b') εF,ϕ rXi=1 rXi=1 (c') b∗ i bi ∈ 1 + V and b∗ i bi ≤ 1. rXi=1 bib∗ i! ∈ 1 + V and εF,ϕ rXi=1 i! ≤ 1. bib∗ kbiξϕ − ξϕbik2 < δ. 24 rXi=1 (d') kbiuk − vkbik2 < δ, k = 1, . . . , n. Let δ′ > 0 and h denote the operatorPr i=1 bib∗ i . Since Bϕ = C1, by Proposition 2.6, we have εF,ϕ(h) ∈ conv{whw∗; w ∈ U(F c), kwξϕ − ξϕwk < δ′}. (23) Here, the bar in (23) denotes the σ-strong closure. Hence there exist w1, . . . , ws ∈ U(F c), and scalars λ1, . . . , λs ∈ R+, with sum 1, such that kwjξϕ − ξϕwjk < δ′, j = 1, . . . , s sXj=1 λjwjhw∗ j ∈ 1 + V. 1 2 aij := λ j wjbi, i = 1, . . . , r, j = 1, . . . , s. and Put Then (a") (b") rXi=1 rXi=1 sXj=1 sXj=1 a∗ ijaij = aij a∗ ij = rXi=1 sXj=1 b∗ i bi ∈ 1 + V and rXi=1 λj wjhw∗ j ∈ 1 + V and Moreover, using sXj=1 rXi=1 a∗ ij aij ≤ 1. sXj=1 aij a∗ ij ≤ 1. aijξϕ − ξϕaij = λ j wj (biξϕ − ξϕbi) + λ j (wj ξϕ − ξϕwj )bi, 1 2 1 2 we obtain (c")  rXi=1 sXj=1 1 2 kaijξϕ − ξϕaijk2 ≤ rXi=1 = rXi=1 1 2 λjkbiξϕ − ξϕbik2 + δ′ sXj=1 sXj=1 rXi=1 kbik2! 1 kbiξϕ − ξϕbik2! 1 + δ′ rXi=1 . 2 2 1 2 λjkbik2 Finaly, since vk ∈ F and wj ∈ F c, we have (d") rXi=1 sXj=1 kaijuk − vkaijk2 ϕ = = rXi=1 rXi=1 < δ. λjkwj(biuk − vkbi)k2 ϕ sXj=1 kbiuk − vkbik2 ϕ Since δ′ > 0 was arbitrary (independent of δ,V, and b1, . . . , br), we can assume that rXi=1 kbiξϕ − ξϕbik2! 1 2 + δ′ rXi=1 2 kbik2! 1 < δ 1 2 . This proves Lemma 5.5. 25 Lemma 5.6. Let N be an injective factor of type III1 with separable predual, and let ϕ be a Q-stable normal faithful state on N . Let u1, . . . , un ∈ U(N ) and let ε > 0. Then there exist a σϕ-invariant finite dimensional subfactor F of N , v1, . . . , vn ∈ U(F ) and a unitary w ∈ U(N ) such that and kwξϕ − ξϕwk < ε, kw∗vkw − ukkϕ < ε, k = 1, . . . , n. Proof. Let δ(n, ε) > 0 be the function in Theorem 2.8, and put δ1 = 1 16 δ(n + 1, ε/2). Choose F and v1, . . . , vn ∈ U(F ), such that the conditions of Lemma 5.5 are satisfied with respect to (u1, . . . , un, δ1). Put ξk = ukξϕ, ηk = vkξϕ, k = 1, . . . , n. For every σ-strong neighborhood V of 0 in N , there exist a1, . . . , ap ∈ N , such that (a), (b), (c) and (d) in Lemma 5.5 are satisfied. Since aiξk − ηkai = (aiuk − vkai)ξϕ + vk(aiξϕ − ξϕai), we have pXi=1 kaiξk − ηkaik2! 1 2 ≤ pXi=1 ϕ! 1 kaiuk − vkaik2 2 + pXi=1 kaiξϕ − ξϕaik2! 1 2 < 2δ 1 2 1 . Moreover, pXi=1 i ai ∈ 1 + V, Pp 2 kaiξϕ − ξϕaik2! 1 i ai ≤ 1, Pp i=1 a∗ 1 2 i=1 a∗ Since Pp i ≤ 1, the two (n + 1)-tuples (ξ1, . . . , ξn, ξϕ) and (η1, . . . , ηn, ξϕ) satisfies the conditions of Remark 2.9 with 4δ 1 instead of δ, so that the two (n + 1)-tuples are 16δ1-related, or equivalently they are δ(n + 1, ε 2 )-related in the sense of Remark 2.9. Hence by Theorem 2.8, there exists a unitary operator w ∈ U(N ), such that i ∈ 1 + V and Pp i=1 aia∗ i=1 aia∗ < δ 1 2 1 < 2δ 1 2 1 . and Therefore kwξk − ηkwk < ε 2 , k = 1, . . . , n. kwξϕ − ξϕwk < ε 2 . kw∗vkw − ukkϕ = kw∗(vkw − wuk)ξϕk = k(wuk − vkw)ξϕk = k(wξk − ηkw) + vk(ξϕw − wξϕ)k < ε, which completes the proof of Lemma 5.6. Now we are ready to prove the main theorem of the paper. 26 Proof of Theorem 5.1. By [AW68, Theorem 7.6], it is sufficient to show that N is an ITPFI- factor. Let ϕ be a Q-stable normal faithful state on N , let u1, . . . , un ∈ U(N ), and let ε > 0. Choose now F , v1, . . . , vn ∈ U(F ) and w ∈ U(N ) as in Lemma 5.6. Put wk = w∗vkw, k = 1, . . . , n. Then F1 := w∗F w is a finite-dimensional subfactor of N , w1, . . . , wn ∈ U(F1) and Since F is σϕ-invariant, ϕ = ϕF ⊗ ϕF c holds. Hence if we put ϕ1 = w∗ϕw, then kwk − ukkϕ < ε, k = 1, . . . , n. Since the representing vector of ϕ1 in P ♮ N is w∗ξϕw, we have ϕ1 = ϕ1F1 ⊗ ϕ1F c 1 . kϕ − ϕ1k ≤ kξϕ − w∗ξϕwkkξϕ + w∗ξϕwk ≤ 2kwξϕ − ξϕwk < 2ε. This shows that ϕ satisfies the product condition in Proposition 2.2, and thus N is an ITPFI factor. References [Ara74] H. Araki, Some properties of modular conjugation operator of von Neumann algebras and non-commutative Radon-Nikodym theorem with a chain rule, Pac. J. Math. 50 (1974), 309 -- 354. [AW68] H. Araki and E. J. Woods, A classification of factors, Publ. Res. Inst. Math. Sci. Ser. A. 4 (1968), 51 -- 130. [Con72] A. Connes, Une classification des facteurs de type III (French), C. R. Acad. Sci. Paris Ser. A-B 275 (1972), 523 -- 525. [Con73] A. Connes, Une classification des facteurs de type III (French), Ann. Sci. Ec. Norm. Sup. 6 (1973), 133-252. [Con74] A. Connes, Almost periodic states and factors of type III1, J. Funct. Anal. 16 (1974), 415 -- 445. [Con76] A. Connes, Classification of injective factors, Ann. Math. 104 (1976), 73 -- 115. [Con85] A. Connes, Factors of type III1, property L′ λ, and closure of inner automorphisms. J. Operator Theory 14 (1985), 189 -- 211. [CS78] A. Connes, E. Størmer, Homogeneity of the state space of factors of type III1, J. Funct. Anal. 28 (1978), 187 -- 196. [CT77] A. Connes and M. Takesaki, The flow of weights on factors of type III, Tohoku Math. J. 29 (1977), 473 -- 575. [CW85] A. Connes and E. J. Woods, Approximately transitive flows and ITPFI factors, Ergod. Th. Dynam. Sys. 5 (1985), 203 -- 236. [Haa75] U. Haagerup, The standard form of von Neumann algebras, Math. Scand. 37 (1975), 271 -- 283. 27 [Haa79] U. Haagerup, Operator-valued weights in von Neumann algebras. I, J. Funct. Anal. 32 (1979), 175 -- 206. [Haa79-2] U. Haagerup, Operator-valued weights in von Neumann algebras. II, J. Funct. Anal. 33 (1979), 339 -- 361. [Haa85] U. Haagerup, A new proof of the equivalence of injectivity and hyperfiniteness for factors on a separable Hilbert space, J. Funct. Anal. 62 (1985), 160 -- 201. [Haa87] U. Haagerup, Connes' bicentralizer problem and uniqueness of the injective factor of type III1. Acta Math. 158 (1987), 95 -- 148. [Haa89] U. Haagerup, The injective factors of type IIIλ, 0 < λ < 1, Pacific. J. Math. 137 (1989), 265 -- 310. [HM09] U. Haagerup and M., Musat, Classification of hyperfinite factors up to completely bounded isomorphism of their preduals, J. reine angew. Math. 630 (2009), 141 -- 176. [Kat68] Y. Katznelson, An introduction to harmonic analysis, John Wiley & sons (1968). [PT73] G. K. Pedersen, M. Takesaki, The Radon-Nikodym theorem for von Neumann algebras, Acta Math. 130 (1973), 53 -- 87. [Tak72] M. Takesaki, Conditional expectations in von Neumann algebras, J. Funct. Anal. 9 (1972), 306 -- 321. [Tak73] M. Takesaki, The structure of von Neumann algebras with a homogenelnous periodic state, Acta. Math. 131 (1973), 79 -- 121. [Tak73-2] M. Takesaki, Duality for crossed products and the structure of von Neumann algebras of type III, Acta. Math. 131, (1973), 249 -- 310. [Tom58] J. Tomiyama, On the projection of norm one in W∗-algebras, III, Tohoku Math. J., 10 (1958), 204 -- 209. 28
1204.6665	1	1204	2012-04-30T15:09:26	On generalized Powers-St$\o$rmer's Inequality	[ "math.OA" ]	A generalization of Powers-St$\o$rmer's inequality for operator monotone functions on $[0, +\infty)$ and for positive linear functional on general $C^$-algebras will be proved. It also will be shown that the generalized Powers-St$\o$rmer inequality characterizes the tracial functionals on $C^$-algebras.	math.OA	math	1 ON GENERALIZED POWERS-STØRMER'S INEQUALITY DINH TRUNG HOA, HIROYUKI OSAKAA, AND HO MINH TOAN Abstract. A generalization of Powers-Størmer's inequality for operator monotone functions on [0, +∞) and for positive linear functional on general C ∗-algebras will be proved. It also will be shown that the generalized Powers-Størmer inequality character- izes the tracial functionals on C ∗-algebras. 1. Introduction Powers-Størmer's inequality (see, for example, [12]) asserts that for s ∈ [0, 1] the following inequality (1) 2 Tr(AsB1−s) ≥ Tr(A + B − A − B) holds for any pair of positive matrices A, B. This is a key inequality to prove the upper bound of Chernoff bound, in quantum hypothesis testing theory [1]. This inequality was first proven in [1], using an integral representation of the function ts. After that, M. Ozawa gave a much simpler proof for the same inequality, using fact that for s ∈ [0, 1] function f (t) = ts (t ∈ [0, +∞)) is an operator monotone. Recently, Y. Ogata in [10] extended this inequality to standard von Neumann algebras. The motivation of this paper is that if the function f (t) = ts is replaced by another operator monotone function (this class is intensively studied, see [7][11]), then Tr(A + B − A − B) may get smaller upper bound that is used in quantum hypothesis testing. Based on M. Ozawa's proof we formulate Powers-Størmer's inequality for an arbitrary operator monotone function on [0, +∞) in the context of general C ∗-algebras. Finally, we will show that the Powers-Størmer's inequality character- izes the trace property for a normal linear positive functional on a von Neumann algebras and for a linear positive functional on a C ∗-algebra. Recall that a positive linear functional ϕ on a von Neumann algebra M is said to be normal if ϕ(sup Ai) = sup ϕ(Ai) for every bounded increasing net Ai of positive elements in M. A linear functional ϕ on a C ∗-algebra A is said to be tracial if ϕ(AB) = ϕ(BA) for all A, B ∈ A. 1AResearch partially supported by the JSPS grant for Scientic Research No. 20540220. Date: 29, Mar., 2012. 2000 Mathematics Subject Classification. 46L30, 15A45. Key words and phrases. Powers-Størmer's inequality, trace, positive functional, C ∗-algebras. 1 2 D. T. HOA, H. OSAKA, AND H. M. TOAN For all other notions used in the paper, we refer the reader to the monograph [8]. This article has been completed when the first author visited Rit- sumeikan University in Februrary, 2012. He is very grateful to all staffs in the department of Mathematical Sciences for their warm hospitality during his stay there. The second author would like to thank Profes- sor Jun Tomiyama for his stimulating discussion on matrix monotone functions through e-mail. 2. Main results Let n ∈ N and Mn be the algebra of n × n matrices. Let I be an interval in R and f : I → R be a continuous function. We call a func- tion f matrix monotone of order n or n-monotone in short whenever the inequality A ≤ B =⇒ f (A) ≤ f (B) for an arbitrary selfadjoint matrices A, B ∈ Mn such that A ≤ B and all eigenvalues of A and B are contained in I. Let H be a separable infinite dimensional Hilbert space and B(H) be the set of all bounded linear operators on H. We call a function f operator monotone whenever the inequality A ≤ B =⇒ f (A) ≤ f (B) for an arbitrary selfadjoint matrices A, B ∈ B(H) such that A ≤ B and all eigenvalues of A and B are contained in I. We denote the spaces of operator monotone functions by P∞(I). The spaces for n-monotone functions are written as Pn(I). We have then P1(I) ⊇ · · · ⊇ Pn−1(I) ⊇ Pn(I) ⊇ Pn+1(I) ⊇ · · · ⊇ P∞(I). Here we note that ∩∞ [7][11]. n=1Pn(I) = P∞(I) and each inclusion is proper The following result is well-known. For example see the proof in [5, Theorem 2.5]. Lemma 2.1. Let f be a strictly positive, continuous function on [0,∞). If the function f is 2n-monotone, then for any positive semidefinite A and a contraction C in Mn we have C ∗f (A)C ≤ f (C ∗AC). Lemma 2.2. Let f be a continuous function on (0,∞) such that 0 /∈ f ((0,∞)). Then, f is n-monotone if and only if the function − 1 f (t) is n-monotone. ON GENERALIZED POWERS-STØRMER'S INEQUALITY 3 Proof. For any t1, t2,· · · , tn ∈ (0,∞) we have f (tj ) 1 f (ti) − 1 ti − tj = f (tj )−f (ti) f (ti)f (tj ) ti − tj 1 = − f (ti)f (tj) Since f is n-monotone, the matrix [ f (ti)−f (tj ) ti−tj f (ti) − f (tj) . ti − tj ] is positive semidefinite by [9], hence, we have [ (− 1 f (ti)) − (− 1 f (tj )) ti − tj ] = −[ f (tj ) ] 1 f (ti) − 1 ti − tj 1 = −(−[ f (ti)f (tj) 1 f (ti)f (tj) ] ◦ [ = [ ≥ 0, f (ti) − f (tj) ]) ti − tj f (ti) − f (tj) ] ti − tj where ◦ means the Hadamard product. Therefore, the function − 1 Conversely, if − 1 f is n-monotone, we have f (t) is n-monotone by [9]. [ f (ti) − f (tj) ti − tj ] = [f (ti)f (tj)] ◦ [ ≥ 0, (− 1 f (ti)) − (− 1 f (tj ) ) ti − tj ] hence f is n-monotone. (cid:3) Proposition 2.1. Let f be a strictly positive, continuous function on [0,∞). If f is 2n-monotone, the function g(t) = t f (t) is n-monotone on [0,∞). Proof. Let A, B be positive matrixces in Mn such that 0 < A ≤ B. satisfies the Jensen type inequality from Lemma 2.1, that is, 2 . Then kCk ≤ 1. Since f is 2n-monotone, −f Let C = B− 1 2 A 1 −f (A) = −f (C ∗BC) ≤ −C ∗f (B)C −f (A) ≤ −A 1 2 f (B)B− 1 2 B− 1 2 f (B)B− 1 2 1 2 2 A −A− 1 2 ≤ −B− 1 2 f (A)A− 1 −A−1f (A) ≤ −B−1f (B) 4 D. T. HOA, H. OSAKA, AND H. M. TOAN Hence, the function − f (t) we conclude that t is n-monotone. Therefore, from Lemm 2.2 − 1 − f (t) t = t f (t) is n-monotone. (cid:3) Remark 1. The condition of 2n-monotonicity of f is needed to guaran- tee the n-monotonicity of g. Indeed, it is well-known that t3 is mono- tone, but not 2-monotone. In this case the function g(t) = t t2 is obviously not 1-monotone. t3 = 1 Proposition 2.2. Let h : [0,∞) → [0,∞) be a Borel function such that h is a continuous, n-monotone on (0,∞), and h(0) = 0. Then for any A, B ∈ M + n with A ≤ B we have h(A) ≤ h(B). Proof. Let B = Ps µsqs be a spectral decomposition. Set 1 − q as a projection on Ker(B). Then B = Bq = qB = Ps′ µs′qs′ and q = Ps′ qs′. Similarly, let A = Pt λtpt be a spectral projection and (1 − p) be a projection on Ker(A). Since A ≤ B, p ≤ q and A = Pt′ λt′pt′ and p = Pt′ pt′. Note that since h(0) = 0, by the function calculus we have h(A) = Pt′ h(λt′)pt′ and h(B) = Ps′ h(µs′)qs′. For any ε > 0 since 0 < Xt′ ≤ Xs′ λt′pt′ + ε1 µs′qs′ + ε1 and h is n-monotone on (0,∞), we have (λt′ + ε)pt′ + ε(1 − p)) ≤ h(Xs′ h(Xt′ (µs′ + ε)qs′ + ε(1 − q)). Since Xt′ h(λt′ + ε)pt′ + h(ε)(1 − p) = h(Xt′ ≤ h(Xs′ = Xs′ (λt′ + ε)pt′ + ε(1 − p)) (µs′ + ε)qs′ + ε(1 − q)) h(µs′ + ε)qs′ + h(ε)(1 − q) ON GENERALIZED POWERS-STØRMER'S INEQUALITY 5 and p ≤ q, it follows that Xt′ h(λt′ + ε)pt′ ≤ Xt′ ≤ Xs′ h(λt′ + ε)pt′ + h(ε)q(1 − p)q h(µs′ + ε)qs′. Therefore, since h is continuus on (0,∞), as ε → 0 we have h(A) = Xt′ ≤ Xs′ h(λt′)pt′ h(µs′)qs′ = h(B). (cid:3) f (t) 0 (t = 0) (t ∈ (0,∞)) Corollary 2.1. Let f be a 2n-monotone, continuous function on [0,∞) such that f ((0,∞)) ⊂ (0,∞), and let g be a Borel function on [0,∞) defined by g(t) = (cid:26) t . Then for any pair of positive matrices A, B ∈ Mn with A ≤ B, g(A) ≤ g(B). Proof. Since f is 2n-monotone, continuous function on [0,∞) such that f ((0,∞)) ⊂ (0,∞), from Proposition 2.1 g is n-monotone on (0,∞). Hence, since g is a Borel function on [0,∞) with g(0) = 0, from Proposition 2.2 it follows that g(A) ≤ (B). (cid:3) Theorem 2.1. Let Tr be a canonical trace on Mn and f be a 2n- monotone function on [0,∞) such that f ((0,∞)) ⊂ (0,∞). Then for any pair of positive matrices A, B ∈ Mn (2) Tr(A) + Tr(B) − Tr(A − B) ≤ 2 Tr(f (A) where g(t) = (cid:26) t f (t) 0 (t ∈ (0,∞)) (t = 0) . 1 2 g(B)f (A) 1 2 ), Proof. Let A, B be any positive matrices in Mn. For operator (A − B) let us denote by P = (A − B)+ and Q = (A − B)− its positive and negative part, respectively. Then we have (3) A − B = P − Q and A − B = P + Q, from that it follows that (4) A + Q = B + P. 6 D. T. HOA, H. OSAKA, AND H. M. TOAN On account of (4) the inequality (2) is equivalent to the following Tr(A) − Tr(f (A) 1 2 g(B)f (A) 1 2 ) ≤ Tr(P ). Since B + P ≥ B ≥ 0 and B + P = A + Q ≥ A ≥ 0, we have g(A) ≤ g(B + P ) by Corollary 2.1 and 1 1 2 ) 2 ) − Tr(f (A) 1 1 2 g(B)f (A) 1 2 ) 1 2 g(B)f (A) 1 2 ) 1 2 g(B + P )f (A) 2 (g(B + P ) − g(B))f (A) 2 ) − Tr(f (A) 2 ) 1 2 (g(B + P ) − g(B))f (B + P ) 2 g(B + P )f (B + P ) 2 ) 1 1 1 2 ) 1 1 1 2 g(A)f (A) 2 g(B)f (A) Tr(A) − Tr(f (A) = Tr(f (A) ≤ Tr(f (A) = Tr(f (A) ≤ Tr(f (B + P ) = Tr(f (B + P ) − Tr(f (B + P ) ≤ Tr(B + P ) − Tr(f (B) = Tr(B + P ) − Tr(B) = Tr(P ). 1 1 2 g(B)f (B + P ) 1 2 ) 1 2 g(B)f (B) 1 2 ) Hence, we have the conclusion. (cid:3) Remark 2. (i) When given positive matrices A, B in Mn satisfies the condition A ≤ B, the inequality (2) becomes Tr(A) ≤ Tr(f (A) 1 2 g(B)f (A) 2 ). 1 (ii) As pointed in Proposition 2.1, 2-monotonicity of f is needed to guarantee the inequality (2). Indeed, let f (t) = t3 and n = 1. Then, for any a, b ∈ (0,∞), the inequality (2) would imply a ≤ f (a) that is, 1 2 g(b)f (a) 1 2 , a f (a) ≤ b f (b) . t f (t) is, however, not 1-monotone, the latter inequality is Since impossible. As an application we get Powers-Størmer's inequality. Corollary 2.2. [1, Theorem 1] Let A and B be positive matrices, then for all s ∈ [0, 1] Tr(A + B − A − B) ≤ Tr(AsB1−s). ON GENERALIZED POWERS-STØRMER'S INEQUALITY 7 Proof. Let f (t) = ts (s ∈ [0, 1]). Then f is operator monotone with f (0,∞) ⊂ (0,∞) and g(t) = t1−s. Hence, we have the conclusion from Theorem 2.1. (cid:3) Since any C-algebra can be realized as a closed selfadjoint ∗-algebra of B(H) for some Hilbert space H. We can generalize Theorem 2.1 in the framework of C-algebras. Theorem 2.2. Let τ be a tracial functional on a C ∗-algebra A, f be a strictly positive, operator monotone function on [0,∞). Then for any pair of positive elements A, B ∈ A (5) where g(t) = tf (t)−1. τ (A) + τ (B) − τ (A − B) ≤ 2τ (f (A) 1 2 g(B)f (A) 1 2 ), Proof. Since the function t f (t) is operator monotone on (0,∞) by [5, Corollary 6], we can get the conclusion through the same steps in the proof of Theorem 2.1. (cid:3) Remark 3. For matrices A, B ∈ M + (6) Tr(A(1−s)/2BsA(1−s)/2) n let us denote Q(A, B) = min s∈[0,1] and (7) QF2n(A, B) = inf f ∈F2n Tr(f (A) 1 2 g(B)f (A) 1 2 ), where F2n is the set of all 2n-monotone functions on [0, +∞) satisfy condition of the Theorem 2.1 and g(t) = tf (t)−1 (t ∈ [0, +∞)). Note that the function f (t) = ts (t ∈ [0, +∞)) satisfies the conditions of Theorem 2.1. Since the class of 2n-monotone functions is large enough [11], we know that QF2n(A, B) ≤ Q(A, B). Hence, we hope on finding another 2n-monotone function f on [0, +∞) such that (8) 2 ) < Q(A, B). 2 g(B)f (A) Tr(f (A) 1 1 If we can find such a function, then we can refine the quantum Chernoff bound used in quantum hypothesis testing [1]. 3. Characterizations of the trace property In this section the generalized Powers-Størmer inequality in the pre- vious section implies the trace property for a positive linear functional on operator algebras. Lemma 3.1. Let ϕ be a positive linear functional on Mn and f be a continuous function on [0,∞) such that f (0) = 0 and f ((0,∞)) ⊂ (0,∞). If the following inequality (9) ϕ(A + B) − ϕ(A − B) ≤ 2ϕ(f (A) 2 g(B)f (A) 2 ) 1 1 8 D. T. HOA, H. OSAKA, AND H. M. TOAN holds true for all A, B ∈ M + of the canonical trace Tr on Mn, where g(t) = (cid:26) t f (t) 0 n , then ϕ should be a positive scalar multiple (t ∈ (0,∞)) (t = 0) . Proof. As is well known, every positive linear functional ϕ on Mn can be represented in the form ϕ(·) = Tr(Sϕ·) for some Sϕ ∈ M + n . It is easily seen that without loss of generality we can assume that Sϕ = diag(α1, α2,· · · , αn), and we have to prove that αi = αj for all i, j = 1,· · · , n. Clearly, it is sufficient to prove that α1 = α2. By assumption, the inequality (9) holds true, in particular, for any positive matirices matrices X = [xij]n n such that 0 = xij = yij if 3 ≤ i ≤ n or 3 ≤ j ≤ n. Thus, it suffices to consider the case n = 2. Assume that Sϕ = diag(d, 1) (d ∈ [0, 1]) and ϕ(D) = Tr(SϕD),∀D ∈ M2. We show that d = 1. For arbitrary positive numbers λ, µ such that λ < µ we consider the following matrices i,j=1, Y = [yij]n i,j=1 from M + and √λµ A = (cid:18) λ √λµ µ (cid:19) −√λµ B = (cid:18) −√λµ µ λ (cid:19) . It is clear that these are positive scalar multiple of projections of rank one. In addition, f (A) 1 2 g(B)f (A) 1 µ + λ(cid:19)2 2 = (cid:18)µ − λ A. We have 2ϕ(f (A) 1 2 g(B)f (A) 1 µ + λ(cid:19)2 2 ) = 2(cid:18)µ − λ µ + λ(cid:19)2 = 2(cid:18)µ − λ Tr(SϕA) (dλ + µ). By direct calculation, A − B = (cid:18) 2√λµ 0 0 2√λµ (cid:19) . Consequently, ϕ(A + B) − ϕ(A − B) = d(2λ − 2pλµ) + 2µ − 2pλµ. Then the inequality (9) becomes µ + λ(cid:19)2 (cid:18)µ − λ (dλ + µ) ≥ d(λ −pλµ) + µ −pλµ. Dividing two side by √λ(√µ − √λ), we get (√µ − √λ)(√µ + √λ)2 √λ(µ + λ)2 d + (dλ + µ) ≥ rµ λ . ON GENERALIZED POWERS-STØRMER'S INEQUALITY 9 Tending λ to µ from the left we obtain d ≥ 1. Since d ∈ [0, 1], d = 1. This means that ϕ is a positive scalar multiple of the canonical trace Tr on Mn. Remark 4. Let ϕ be a positive linear functional on Mn and s ∈ [0, 1]. From Lemma 3.1 it is clear that if the following inequality (cid:3) ϕ(A + B) − ϕ(A − B) ≤ 2ϕ(A (10) holds true for any A, B ∈ M + s = 0 the following inequality characterizes the trace property n , then ϕ is a tracial. In particular, when 2 ) 1−s 1−s 2 BsA (11) ϕ(B) − ϕ(A) ≤ ϕ(A − B) (A, B ∈ M + n ). Corollary 3.1 ([13]). Let ϕ be a positive linear functional on Mn and the following inequality (12) holds true for any self-adjoint matrices A, B ∈ Mn. Then ϕ is a tracial. Proof. From the assumption, we have ϕ(A + B) ≤ ϕ(A) + ϕ(B) ϕ(B − A) ≥ ϕ(B) − ϕ(A) for any pair of self-adjoint matrices A, B in Mn. Moreover, for any pair of positive matrices A, B ∈ Mn we have ϕ(B − A) ≥ ϕ(B) − ϕ(A). On account of Remark 4, it follows that ϕ should be a tracial. (cid:3) Corollary 3.2 ([4]). Let ϕ be a positive linear functional on Mn and the following inequality ϕ(A) ≤ ϕ(A) (13) holds true for any self-adjoint matrix A ∈ Mn. Then ϕ is a tracial. Proof. Let A, B ∈ Mn be arbitrary positive matrices. Then C = B − A is a self-adjoint matrix. Since A, B ≥ 0, the values ϕ(A) and ϕ(B) are real. From the assumption, we have ϕ(B) − ϕ(A) ≤ ϕ(B) − ϕ(A) = ϕ(B − A) ≤ ϕ(B − A). On account of Remark 4, it follows that ϕ should be a tracial. (cid:3) By analogy with a number of other similar cases (see [4] or [14]), the proof for the trace property of a positive normal functional satisfying the inequality (9) on a von Neumann algebra can be reduced to the case of the algebra M2 of all matrices of order 2 × 2. But for self-contained we will give a sketch of its proof. 10 D. T. HOA, H. OSAKA, AND H. M. TOAN Theorem 3.1. Let ϕ be a positive normal linear functional on a von Neumann algebra M and f be a continuous function on [0,∞) such that f (0) = 0 and f ((0,∞)) ⊂ (0,∞). If the following inequality (14) holds true for any pair A, B ∈ M+, then ϕ is a trace, where g(t) = (cid:26) t ϕ(A) + ϕ(B) − ϕ(A − B) ≤ 2ϕ(f (A) (t ∈ (0,∞)) (t = 0) 2 g(B)f (A) f (t) 0 2 ) . 1 1 Proof. Let P1, P2 be a pair of nonzero mutually orthogonal equiva- lent projections in M, that means P1 = V ∗V and P2 = V V ∗ for some nonzero partial isometry V ∈ M. Consider the ∗-algebra N in (P1 + P2)M(P1 + P2) generated by the partial isometry V . Then N is isomorphic to M2 and inequality (14) still holds true for the opera- tors in N and for the restriction of the functional ϕ to N . According to Lemma 3.1, this restriction is a tracial functional on N , and hence ϕ(P1) = ϕ(P2). By [8, Vol2, Proposition 8.1.1] it follows that ϕ is a trace. (cid:3) 1 1 2 ) 2 g(B)f (A) Corollary 3.3. Let ϕ be a positive linear functional on a C ∗-algebra A and f be a continuous function on [0,∞) such that f (0) = 0 and f ((0,∞)) ⊂ (0,∞). If the following inequality ϕ(A) + ϕ(B) − ϕ(A − B) ≤ 2ϕ(f (A) (15) holds true for any pair A, B ∈ A+, then ϕ is a tracial functional, where g(t) = (cid:26) t Proof. Let π be the universal representation of C ∗-algebra A and M = π(A)′′. Let ϕ be the positive normal functional on M such that ϕ(π(A)) = ϕ(A) for A ∈ A. By the Kaplansky density theorem, for any pair A, B ∈ M+ there exist bounded nets {Aα} and {Bα} in A+ such that π(Aα) → A and π(Bα) → B in the strong operator topology. Using (15) and the continuity of the corresponding operations in the strong operator topology, we have (t ∈ (0,∞)) (t = 0) f (t) 0 . ϕ( A) + ϕ( B) − ϕ( A − B) ≤ 2 ϕ(f ( A) 1 2 g( B)f ( A) 1 2 ). By Theorem 3.1, ϕ is a tracial functional M, and hence ϕ is a tracial functional on A. (cid:3) Remark 5. Let A be a von Neumann algebra and ϕ be a positive linear functional on A. The set P (A) of all orthogonal projections from A is enough as a testing space for some inequlity to characterize the trace property of ϕ (see [3]). But, in the case of the inequality (14) the set P (A) is not enough as a testing set. ON GENERALIZED POWERS-STØRMER'S INEQUALITY 11 Indeed, let p, q be arbitrary orthogonal projections from a von Neu- mann algebra M. Since q ≥ p∧q it follows that pqp ≥ p(p∧q)p = p∧q. So pqp ≥ p ∧ q holds for any pair of projections. From that it follows ϕ(p + q − p − q) = 2ϕ(p ∧ q) ≤ 2ϕ(pqp) = 2ϕ(f (p) whenever ϕ is an arbitrary positive linear functional on M. 2 g(q)f (p) 1 1 2 ) References [1] K.M.R.Audenaert, J.Calsamiglia, LI.Masanes, R.Munoz-Tapia, A.Acin, E.Bagan, F.Verstraete, The Quantum Chernoff Bound, Phys. Rev. Lett. 98 (2007) 16050. [2] R.Bhatia, Matrix analysis, Graduate texts in mathematics, Springer New York, 1997. [3] A.M.Bikchentaev, Commutation of projections and characterization of traces on von Neumann algebras, Siberian Math. J., 51 (2010) 971-977. [Translation from Sibirskii Matematicheckii Zhurnal, 51 (2010) 1228-1236] [4] L.T.Gardner, An inequality characterizes the trace, Canad. J. Math. 31 (1979) 1322-1328. [5] F.Hansen, G.K.Pedersen, Jensen's inequality for operator and Lowner's theo- rem, Math. Ann., 258 (1982) 229-241. [6] F.Hansen, Some operator monotone functions, Linear Algebra and its Applica- tions, 430 (2009) 795-99. [7] F.Hansen, G.Ji, J.Tomiyama, Gaps between classes of matrix monotone func- tions, Bull. London Math. Soc. 36 (2004) 53-58. [8] R.V.Kadison, J.R.Ringrose, Fundamentals of the Theory of Operator Algebras, Vols I, II, Academic Press., 1983, 1986. [9] K.Loewner, U ber monotone Matrixfunktionen, Math. Z. 38 (1934) 177-216. [10] Y.Ogata, A Generalization of Powers-Størmer Inequality, Letters in Mathe- matical Physics, 97:3 (2011) 339-346. [11] H.Osaka, S.Silvestrov, J.Tomiyama, Monotone operator functions, gaps and power moment problem, Math. Scand. 100:1 (2007) 161-183. [12] R.T.Powers, E.Størmer, Free States of the Canonical Anticommutation Rela- tions, Commun. math. Phys., 16 (1970) 1-33. [13] A.I.Stolyarov, O.E.Tikhonov, A.N.Sherstnev, Characterization of normal traces on von Neumann algebras by inequalities for the modulus, Mathemat- ical Notes, 72:3 (2002) 411-416. [Translation from Mat. Zametki., 72:3 (2002) 448-454.] [14] O.E.Tikhonov, Subadditivity ineqalities in von Neumann algebras and charac- terization of tracial functional, Positivity. 9 (2005) 259-264. Research Center for Sciences and Technology, Duy Tan University, 182 Nguyen Van Linh, Danang, Vietnam E-mail address: [email protected] Department of Mathematical Sciences, Ritsumeikan University, Kusatsu, Shiga 525-8577, Japan E-mail address: [email protected] Mathematical Institute, 18 Hoang Quoc Viet, Hanoi, Vietnam E-mail address: [email protected]
1705.09588	2	1705	2017-06-27T11:50:30	Blowup constructions for Lie groupoids and a Boutet de Monvel type calculus	[ "math.OA", "math.DG", "math.KT" ]	We present natural and general ways of building Lie groupoids, by using the classical procedures of blowups and of deformations to the normal cone. Our constructions are seen to recover many known ones involved in index theory. The deformation and blowup groupoids obtained give rise to several extensions of $C^*$-algebras and to full index problems. We compute the corresponding K-theory maps. Finally, the blowup of a manifold sitting in a transverse way in the space of objects of a Lie groupoid leads to a calculus, quite similar to the Boutet de Monvel calculus for manifolds with boundary.	math.OA	math	Blowup constructions for Lie groupoids and a Boutet de Monvel type calculus by Claire Debord and Georges Skandalis Universit´e Clermont Auvergne LMBP, UMR 6620 - CNRS Campus des C´ezeaux, 3, Place Vasarely TSA 60026 CS 60026 63178 Aubi`ere cedex, France [email protected] Universit´e Paris Diderot, Sorbonne Paris Cit´e Sorbonne Universit´es, UPMC Paris 06, CNRS, IMJ-PRG UFR de Math´ematiques, CP 7012 - Batiment Sophie Germain 5 rue Thomas Mann, 75205 Paris CEDEX 13, France [email protected] 7 1 0 2 n u J 7 2 ] Abstract We present natural and general ways of building Lie groupoids, by using the classical proce- dures of blowups and of deformations to the normal cone. Our constructions are seen to recover many known ones involved in index theory. The deformation and blowup groupoids obtained give rise to several extensions of C∗-algebras and to full index problems. We compute the cor- responding K-theory maps. Finally, the blowup of a manifold sitting in a transverse way in the space of objects of a Lie groupoid leads to a calculus, quite similar to the Boutet de Monvel calculus for manifolds with boundary. . A O h t a m [ 2 v 8 8 5 9 0 . 5 0 7 1 : v i X r a Contents 1 Introduction 2 Some quite classical constructions involving groupoids 2.1 Some classical notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Transversality and Morita equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Semi-direct products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Index maps for Lie groupoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 3 Remarks on exact sequences, Connes-Thom elements, connecting maps and index maps 3.1 A (well known) remark on exact sequences . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Saturated open subsets, connecting maps and full index map . . . . . . . . . . . . . 3.2.1 Connecting map and index . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Connecting maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.3 A general remark on the index . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.4 Full symbol algebra and index . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.5 Fredholm realization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.6 Relative K-theory and full index . . . . . . . . . . . . . . . . . . . . . . . . . + . . . . . . . . . . . . . . . 3.3.1 Proper action on a manifold . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Proper action on a groupoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.3 Closed saturated subsets and connecting maps . . . . . . . . . . . . . . . . . 3.3.4 Connes-Thom invariance of the full index . . . . . . . . . . . . . . . . . . . . 3.3 Connes-Thom elements and quotient of a groupoid by R∗ The authors were partially supported by ANR-14-CE25-0012-01 (SINGSTAR). AMS subject classification: Primary 58H05, 19K56. Secondary 58B34, 22A22, 46L80,19K35, 47L80. 3 8 8 9 10 11 11 11 12 12 13 13 15 17 18 19 19 20 21 21 1 4 Two classical geometric constructions: Blowup and deformation to the normal cone 4.1 Deformation to the normal cone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Blowup constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 22 23 5 Constructions of groupoids 5.2 Normal groupoids, deformation groupoids and blowup groupoids 25 5.1 Linear groupoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 5.1.1 The linear groupoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 5.1.2 The projective groupoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 5.1.3 The spherical groupoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 5.1.4 Bundle groupoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 5.1.5 VB groupoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 . . . . . . . . . . . 28 5.2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 5.2.2 Algebroid and anchor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 5.2.3 Morita equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 5.2.4 Groupoids on manifolds with boundary . . . . . . . . . . . . . . . . . . . . . 29 5.3 Examples of normal groupoids, deformation groupoids and blowup groupoids . . . . 29 Inclusion F ⊂ E of vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.1 29 Inclusion G(0) ⊂ G: adiabatic groupoid . . . . . . . . . . . . . . . . . . . . . 5.3.2 30 5.3.3 Gauge adiabatic groupoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 5.3.4 Inclusion of a transverse submanifold of the unit space . . . . . . . . . . . . . 30 V ⊂ G for a transverse hypersurface V of G: b-groupoid . . . . . . Inclusion GV 5.3.5 31 V ⊂ G for a saturated submanifold V of G: shriek map for immersion 32 Inclusion GV 5.3.6 Inclusion G1 ⊂ G2 with G(0) 32 5.3.7 . . . . . . . . . . . . . . . . . . . . . . . 5.3.8 Wrong way functoriality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 6 The C∗-algebra of a deformation and of a blowup groupoid, full symbol and index 1 = G(0) 2 map 6.1 "DNC" versus "Blup" . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.1 The connecting element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.2 The full symbol index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1.3 When V is AG-small . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 The KK-element associated with DNC . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 The case of a submanifold of the space of units . . . . . . . . . . . . . . . . . . . . . 6.3.1 Connecting map and index map . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.2 The index map via relative K-theory . . . . . . . . . . . . . . . . . . . . . . . 7 A Boutet de Monvel type calculus 7.1 The Poisson-trace bimodule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.1.1 The SBlupr,s(G, V ) − (GV V )ga-bimodule E (G, V ) . . . . . . . . . . . . . . . . 7.1.2 The Poisson-trace bimodule EP T . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 A Boutet de Monvel type algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 A Boutet de Monvel type pseudodifferential algebra . . . . . . . . . . . . . . . . . . 7.4 K-theory of the symbol algebras and index maps . . . . . . . . . . . . . . . . . . . . 7.4.1 K-theory of ΣBM and computation of the index . . . . . . . . . . . . . . . . 7.4.2 Index in relative K-theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 Appendix . . . . . . . . . . 8.1 A characterization of groupoids via elements composable to a unit 8.2 VB groupoids and their duals ([46, 31]) . . . . . . . . . . . . . . . . . . . . . . . . . 8.3 Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 33 33 34 35 36 37 37 38 39 39 39 40 41 41 42 42 43 44 44 45 46 2 Introduction 1 Let G ⇒ M be a Lie groupoid. The Lie groupoid G comes with its natural family of elliptic pseudodifferential operators. For example • if the groupoid G is just the pair groupoid M × M , the associated calculus is the ordinary (pseudo)differential calculus on M ; • if the groupoid G is a family groupoid M ×B M associated with a fibration p : M → B, the associated (pseudo)differential operators are families of operators acting on the fibers of p (those of [7]); • if the groupoid G is the holonomy groupoid of a foliation, the associated (pseudo)differential operators are longitudinal operators as defined by Connes in [11]; • if the groupoid G is the monodromy groupoid i.e. the groupoid of homotopy classes (with fixed endpoints) of paths in a (compact) manifold M , the associated (pseudo)differential operators are the π1(M )-invariant operators on the universal cover of M ... The groupoid G defines therefore a class of partial differential equations. Our study will focus here on the corresponding index problems on M . The index takes place naturally in the K-theory of the C∗-algebra of G. Let then V be a submanifold of M . We will consider V as bringing a singularity into the problem: it forces operators of G to "slow down" near V , at least in the normal directions. Inside V , they should only propagate along a sub-Lie-groupoid Γ ⇒ V of G. This behavior is very nicely encoded by a groupoid SBlupr,s(G, Γ) obtained by using a blow-up construction of the inclusion Γ → G. The blowup construction and the deformation to the normal cone are well known constructions in algebraic geometry as well as in differential geometry. Let X be a submanifold of a manifold Y . Denote by N Y X the normal bundle. • The deformation to the normal cone of X in Y is a smooth manifold DN C(Y, X) obtained by naturally gluing N Y X × {0} with Y × R∗. • The blowup of X in Y is a smooth manifold Blup(Y, X) where X is inflated to the projective space PN Y X in a natural way. We will mainly X . consider its variant the spherical blowup SBlup(Y, X) (which is a manifold with boundary) in which the sphere bundle SN Y It is obtained by gluing Y \ X with PN Y X replaces the projective bundle PN Y X . The first use of deformation groupoids in connection with index theory appears in [12]. A. Connes showed there that the analytic index on a compact manifold V can be described using a groupoid, called the "tangent groupoid". This groupoid is obtained as a deformation to the normal cone of the diagonal inclusion of V into the pair groupoid V × V . Since Connes' construction, deformation groupoids were used by many authors in various contexts. • This idea of Connes was extended in [25] by considering the same construction of a deformation to the normal cone for smooth immersions which are groupoid morphisms. The groupoid obtained was used in order to define the wrong way functoriality for immersions of foliations ([25, section 3]). An analogous construction for submersions of foliations was also given in a remark ([25, remark 3.19]). 3 • In [40, 45] Monthubert-Pierrot and Nistor-Weinstein-Xu considered the deformation to the normal cone of the inclusion G(0) → G of the space of units of a smooth groupoid G. This generalization of Connes' tangent groupoid was called the adiabatic groupoid of G and denoted by Gad. It was shown that this adiabatic groupoid still encodes the analytic index associated with G. • Many other important articles use this idea of deformation groupoids. We will briefly discuss some of them in the sequel of the paper. It is certainly out of the scope of the present paper to review them all... Let us briefly present the objectives of our paper. The groupoids DN C(G, Γ) and SBlupr,s(G, Γ). In the present paper, we give a systematic construction of deformations to the normal cone and define the blow-up deformations of groupoids. More precisely, we use the functoriality of these two constructions and note that any smooth subgroupoid Γ ⇒ V of a Lie groupoid G ⇒ M gives rise to a deformation to the normal cone Lie groupoid DN C(G, Γ) ⇒ DN C(M, V ) and to a blowup Lie groupoid Blupr,s(G, Γ) ⇒ Blup(M, V ) as well as its variant the spherical blowup Lie groupoid SBlupr,s(G, Γ) ⇒ SBlup(M, V ). Connecting maps and index maps These groupoids give rise to connecting maps and to index problems that will be the main object of our study here. Connecting maps. The (restriction DN C+(G, Γ) to R+ of the) deformation groupoid DN C(G, Γ) is the disjoint union of an open subgroupoid G × R∗ The blowup groupoid SBlupr,s(G, Γ) is the disjoint union of an open subgroupoid which is the Γ which is fibered over restriction G Γ, i.e. a VB groupoid (in the sense of Pradines cf. [46, 31]). This decomposition gives rise to exact sequences of C∗-algebras that we wish to "compute": of G to M = M \ V and a boundary which is a groupoid SN G + and a closed subgroupoid N G Γ × {0}. M M 0 −→ C∗(G M M ) −→ C∗(SBlupr,s(G, Γ)) −→ C∗(SN G Γ ) −→ 0 and 0 −→ C∗(G × R∗ +) −→ C∗(DN C+(G, Γ)) −→ C∗(N G Γ ) −→ 0 E∂ SBlup E∂ DN C+ Full index maps. Denote by Ψ∗(DN C+(G, Γ)) and Ψ∗(SBlupr,s(G, Γ)) the C∗-algebra of order 0 pseudodifferential operators on the Lie groupoids DN C+(G, Γ) and SBlupr,s(G, Γ) respectively. The above decomposition of groupoids give rise to extensions of groupoid C∗-algebras of pseudod- ifferential type E(cid:102)ind E(cid:102)ind DN C+ where ΣDN C+(G, Γ) and ΣSBlup(G, Γ) are called the full symbol algebra, and the morphisms σf ull the full symbol maps. σf ull−−−→ ΣDN C+(G, Γ) −→ 0 +) −→ Ψ∗(DN C+(G, Γ)) σf ull−−−→ΣSBlup(G, Γ) −→ 0 0 −→ C∗(G × R∗ and 0 −→ C∗(G M M ) −→ Ψ∗(SBlupr,s(G, Γ)) SBlup 4 The full symbol maps. The full symbol algebras are naturally fibered products: and ΣSBlup(G, Γ) = C(S∗ASBlupr,s(G, Γ)) ×C(S∗ASN G Γ ) Ψ∗(SN G Γ ) ΣDN C+(G, Γ) = C(S∗ADN C+(G, Γ)) ×C(S∗AN G Γ ) Ψ∗(N G Γ ). Thus, the full symbol maps have two components: • The usual commutative symbol of the groupoid. They are morphisms: Ψ∗(DN C+(G, Γ)) → C(S∗ADN C+(G, Γ)) and Ψ∗(SBlupr,s(G, Γ)) → C(S∗ASBlupr,s(G, Γ)). The commutative symbol takes its values in the algebra of continuous fonctions on the sphere bundle of the algebroid of the Lie groupoids (with boundary) DN C+(G, Γ)) and SBlupr,s(G, Γ). • The restriction to the boundary: σ∂ : Ψ∗(SBlupr,s(G, Γ)) → Ψ∗(SN G Γ ) and Ψ∗(DN C+(G, Γ)) → Ψ∗(N G Γ ) . Γ are also amenable, and exact sequences E∂ Associated KK-elements. Assume that the groupoid Γ is amenable. Then the groupoids N G Γ and SN G DN C+ give rise to connecting elements Γ ), C∗(G × R∗ SBlup ∈ KK1(C∗(SN G ∂G,Γ +)) (cf. [27]). Also, the (cid:102)ind full symbol C∗-algebras ΣSBlup(G, Γ) and ΣDN C+(G, Γ) are nuclear and we also get KK-elements SBlup ∈ KK1(ΣSBlup(G, Γ), C∗(G DN C+ ∈ KK1(ΣDN C+(G, Γ), C∗(G × R∗ SBlup and E∂ ∈ KK1(C∗(N G )) and (cid:102)ind Γ ), C∗(G )) and ∂G,Γ +)). DN C+ M M G,Γ G,Γ If Γ is not amenable, these constructions can be carried over in E-theory (of maximal groupoid C∗-algebras). M M Connes-Thom elements We will establish the following facts. Γ ), C∗(N G a) There is a natural Connes-Thom element β ∈ KK1(C∗(SBlupr,s(G, Γ)), C∗(DN C+(G, Γ))). +)) and β(cid:48)(cid:48) ∈ This element restricts to very natural elements β(cid:48) ∈ KK1(C∗(G M M KK1(C∗(SN G These elements extend to elements βΨ ∈ KK1(Ψ∗(SBlupr,s(G, Γ)), Ψ∗(DN C+(G, Γ))) and βΣ ∈ KK1(ΣSBlup(G, Γ)), ΣDN C+(G, Γ))). We have ∂G,Γ (fact 6.3). (cf. facts 6.1 and 6.2) and (cid:102)ind SBlup⊗β(cid:48) = −βΣ⊗(cid:102)ind ), C∗(G × R∗ G,Γ DN C+ Γ )). DN C+ G,Γ b) If M \ V meets all the orbits of G, then β(cid:48) is KK-invertible. Therefore, in that case, ∂G,Γ DN C+ SBlup⊗β(cid:48) = −β(cid:48)(cid:48)⊗∂G,Γ SBlup and (cid:102)ind G,Γ DN C+ determines (cid:102)ind G,Γ SBlup. determines ∂G,Γ c) We will say that V is AG-small if the transverse action of G on V is nowhere 0, i.e. if for every x ∈ V , the image by the anchor of the algebroid AG of G is not contained in TxV (cf. definition 6.5). In that case, β(cid:48), β(cid:48)(cid:48), βΣ are KK-invertible: the connecting elements ∂G,Γ and ∂G,Γ each other SBlup determine each other, and the full index maps (cid:102)ind DN C+ and (cid:102)ind DN C+ G,Γ SBlup determine G,Γ If Γ = V , then C∗(N G V ) is KK-equivalent to C0(N G Computation. V ) using a Connes-Thom isomorphism and the element ∂G,Γ V in the algebroid M of G (using a tubular neighborhood) and the index element indG ∈ KK(C0(A∗G), C∗(G)) AG = N G of the groupoid G (prop. 6.11.d). Of course, if V is AG-small, we obtain the analogous result for ∂G,Γ SBlup. The computation of the corresponding full index is also obtained in the same way in prop. 6.11.e). is the Kasparov product of the inclusion of N G DN C+ 5 Full index and relative K-theory. Assume that Γ is just a AG-small submanifold V of M . We will actually obtain a finer construction by using relative K-theory. It is a general fact that relative K-theory gives more precise index theorems than connecting maps (cf. e.g. [6, 48, 35, 36, 5]). In particular, the relative K-theory point of view allows to take into account symbols from a vector bundle to another one. Let ψ : C0(SBlup(M, V )) → Ψ∗(SBlupr,s(G, Γ)) be the natural inclusion and consider the morphism µSBlup = σf ull ◦ ψ : C0(SBlup(M, V )) → ΣSBlup(G, V ). The relative index theorem computes the map indrel : K∗(µSBlup ◦ ψ) → K∗(C∗(G )): M M • the relative K-group K∗(µSBlup) is canonically isomorphic to K∗(C0(A∗G )); • under this isomorphism indrel identifies with the index map of the groupoid G M M M M . We prove an analogous result for the morphism µDN C : C0(DN C+(M, V ) → ΣDN C+(G, V ) In fact, most of the computations involved come from a quite more general situation studied in section 3. There we consider a groupoid G and a partition of G(0) into an open and a closed saturated subset and study the connecting elements of the associated exact sequences. A Boutet de Monvel type calculus. Let H be a Lie groupoid. In [17], extending ideas of Aastrup, Melo, Monthubert and Schrohe [1], we studied the gauge adiabatic groupoid Hga: the crossed product of the adiabatic groupoid of H by the natural action of R∗ +. We constructed a bimodule EH giving a Morita equivalence between the algebra of order 0 pseudodifferential operators on H and a natural ideal in the convolution C∗-algebra C∗(Hga) of this gauge adiabatic groupoid. The gauge adiabatic groupoid Hga is in fact a blowup groupoid, namely SBlupr,s(H × (R×R), H (0)) (restricted to the clopen subset H (0) × R+ of SBlup(H (0) × R, H (0)) = H (0) × (R− (cid:116) R+)). Let now G ⇒ M be a Lie groupoid and let V be a submanifold M which is transverse to the action it is a C∗(SBlupr,s(G, V )), Ψ∗(GV of G (see def. 2.2). We construct a Poisson-trace bimodule: V ) bimodule EP T (G, V ), which is a full Ψ∗(GV V ) Hilbert module. When G is the direct product of GV V with the pair groupoid R × R the Poisson-trace bimodule coincide with E constructed in [17]. In the general case, thanks to a convenient (spherical) blowup construction, we construct a linking V × (R×R), V ). This linking space between the groupoids SBlupr,s(G, V ) and (GV space defines a C∗(SBlupr,s(G, V )), C∗((GV V )ga)-bimodule E (G, V ) which is a Morita equivalence of groupoids when V meets all the orbits of G. The Poisson-trace -bimodule is then the composition of E Denote by Ψ∗ T Q Φ ∈ Ψ∗(SBlupr,s(G, V )), P ∈ EP T (G, V ), T ∈ E ∗ BM (G; V ) its ideal - where Φ ∈ C∗(SBlupr,s(G, V )). This algebra has obvious similarities with the one involved in the Boutet de Monvel calculus for manifold with boundary [8]. We will examine its relationship with these two algebras in a forthcoming paper. BM (G, V ) → C(S∗AG) given by We still have two natural symbol maps: the classical symbol σc : Ψ∗ σc BM (G; V ) the Boutet de Monvel type algebra consisting of matrices R = = σc(Φ) and the boundary symbol rV which is restriction to the boundary. P T (G, V ) and Q ∈ Ψ∗(GV (cid:18)Φ P (cid:18)Φ P V )ga = SBlupr,s(GV V ), and C∗ (cid:19) with E (G, V ). GV V (cid:19) with GV V T Q We have an exact sequence: M(cid:116)V 0 → C∗(G M(cid:96) V BM (G; V )/C∗(G M(cid:116)V M(cid:116)V where ΣBM (G, V ) = Ψ∗ ) → Ψ∗ BM (G; V ) σBM−−−→ ΣBM (G, V ) → 0 ) and σBM is defined using both σc and rV . 6 (cid:18)Φ 0 (cid:19) 0 0 We may note that Ψ∗(SBlupr,s(G, V )) identifies with the full hereditary subalgebra of Ψ∗ consisting of elements of the form . We thus obtain Boutet de Monvel type index theorems BM (G, V ) for the connecting map of this exact sequence - as well as for the corresponding relative K-theory. The paper is organized as follows: • In section 2 we recall some classical facts, constructions and notation involving groupoids. • Section 3 is devoted to the description and computation of various KK-elements associated with groupoid C∗-algebras. The first and second type are encountered in the situation where a Lie groupoid G can be cut in two pieces G = GW (cid:116) GF where W is an open saturated subset of the units G(0) and F = G(0) \ W . They are respectively built from exact sequences of C∗-agebras of the form: 0 −→ C∗(GW ) −→ C∗(G) −→ C∗(GF ) −→ 0 E∂ and E(cid:102)indf ull The other KK-elements are Connes-Thom type elements arising when a R-action is involved in those situations. 0 −→ C∗(GW ) −→ Ψ∗(G) −→ ΣW (G) −→ 0 • In section 4 we review two geometric constructions: deformation to the normal cone and blowup, and their functorial properties. • In section 5, using this functoriality, we study deformation to the normal cone and blowup in the Lie groupoid context. We present examples which recover groupoids constructed previously by several authors. • In section 6, applying the results obtained in section 3, we compute the connecting maps and index maps of the groupoids constructed in section 5. • In section 7, we describe the above mentioned Boutet de Monvel type calculus. • Finally, in the appendix, we make a few remarks on VB groupoids. In particular, we give a presentation of the dual VB groupoid E∗ of a VB groupoid E and show that C∗(E) and C∗(E∗) are isomorphic. • Our constructions involved a large amount of notation, that we tried to choose as coherent as possible. We found it however helpful to list several items in an index at the end of the paper. Notation 1.1. We will use the following notation: • If E is a real vector bundle over a manifold (or over a locally compact space) M , the cor- responding projective bundle P(E) is the bundle over M whose fiber over a point x of M is the projective space P(Ex). The bundle P(E) is simply the quotient of E \ M by the natural action of R∗ by dilation. The quotient of E \ M under the action of R∗ + by dilation is the (total space of the) sphere bundle S(E). • If E is a real vector bundle over a manifold (or a locally compact space) M , we will denote by B∗E, B∗E and S∗E the total spaces of the fiber bundles of open balls, closed ball and spheres of the dual vector bundle E∗ of E. If F ⊂ M is a closed subset of M , we will denote by B∗ F E the quotient of B∗E where we identify two points (x, ξ) and (x, η) for x ∈ F and S∗ F E the image of S∗E in B∗ F E. Acknowledgements. We would like to thank Vito Zenobi for his careful reading and for pointing out quite a few typos in an earlier version of the manuscript. 7 2 Some quite classical constructions involving groupoids 2.1 Some classical notation Let G be a Lie groupoid. We denote by G(0) its space of objects and r, s : G → G(0) the range and source maps. The algebroid of G is denoted by AG, and its anchor by (cid:92) : AG → T G(0) its anchor. Recall that (the total space of) AG is the normal bundle N G We denote by A∗G the dual bundle of AG and by S∗AG the sphere bundle of A∗G. G(0) and the anchor map is induced by (dr − ds). • We denote by C∗(G) its (full or reduced) C∗-algebra. We denote by Ψ∗(G) the C∗-algebra of order ≤ 0 (classical, i.e. polyhomogeneous) pseudodifferential operators on G vanishing at infinity on G(0) (if G(0) is not compact). More precisely, it is the norm closure in the multiplier algebra of C∗(G) of the algebra of classical pseudodifferential operators on G with compact support in G. We have an exact sequence of C∗-algebras 0 → C∗(G) → Ψ∗(G) → C0(S∗AG) → 0. As mentioned in the introduction, our constructions involve connecting maps associated to short exact sequences of groupoid C∗-algebras, therefore they make sens a priori for the full C∗-algebras, and give rise to E-theory elements ([13]). Nevertheless, in many interesting situations, the quotient C∗-algebra will be the C∗-algebra of an amenable groupoid, thus the corresponding exact sequence is semi-split as well as for the reduced and the full C∗-algebras and it defines moreover a KK-element. In these situations C∗(G) may either be the reduced or the full C∗-algebra of the groupoid G and we have preferred to leave the choice to the reader. • For any maps f : A → G(0) and g : B → G(0), define Gf = {(a, x) ∈ A × G; r(x) = f (a)}, Gg = {(x, b) ∈ G × B; s(x) = g(b)} and g = {(a, x, b) ∈ A × G × B; r(x) = f (a), s(x) = g(b)} . Gf In particular for A, B ⊂ G(0), we put GA = {x ∈ G; r(x) ∈ A} and GA = {x ∈ G; s(x) ∈ A}; we also put GB Notice that A is a saturated subset of G(0) if and only if GA = GA = GA A. A = GA ∩ GB. • We denote by Gad the adiabatic groupoid of G, ([40, 45]), it is obtained by using the deforma- tion to the normal cone construction for the inclusion of G(0) as a Lie subgroupoid of G (see section 5.2 and 5.3 below for a complete description). Thus: Gad = G × R∗ ∪ AG × {0} ⇒ G(0) × R . If X is a locally closed saturated subset of M × R, we will denote sometimes by Gad(X) the restriction (Gad)X In the sequel of the paper, we let G[0,1] ad = G×(0, 1]∪AG×{0} ⇒ G(0)×[0, 1] and G[0,1) G[0,1] ad = Gad(G(0) × [0, 1)) i.e. ad = G×(0, 1)∪AG×{0} ⇒ G(0)×[0, 1). ad = Gad(G(0) × [0, 1]) and G[0,1) X of Gad to X: it is a locally compact groupoid. Remark 2.1. Many manifolds and groupoids that occur in our constructions have boundaries or corners. In fact all the groupoids we consider sit naturally inside Lie groupoids without boundaries as restrictions to closed saturated subsets. This means that we consider subgroupoids GV V = GV r,s⇒ G(0) where V is a closed subset of G(0). Such groupoids, have a natural of a Lie groupoid G algebroid, adiabatic deformation, pseudodifferential calculus, etc. that are restrictions to V and GV of the corresponding objects on G(0) and G. We chose to give our definitions and constructions for Lie groupoids for the clarity of the exposition. The case of a longitudinally smooth groupoid over a manifold with corners is a straightforward generalization using a convenient restriction. 8 2.2 Transversality and Morita equivalence Let us recall the following definition (see e.g. [50] for details): r,s⇒ M be a Lie groupoid with set of objects G(0) = M . Let V be a manifold. Definition 2.2. Let G A smooth map f : V → M is said to be transverse to (the action of the groupoid) G if for every x ∈ V , dfx(TxV ) + (cid:92)f (x)Af (x)G = Tf (x)M . An equivalent condition is that the map (γ, y) (cid:55)→ r(γ) defined on the fibered product Gf = G × is a submersion Gf → M . A submanifold V of M is transverse to G if the inclusion V → M is transverse to G - equivalently, if for every x ∈ V , the composition qx = px ◦ (cid:92)x : AxG → (N M V )x = TxM/TxV is onto. s,f V V . Choosing a subbundle W (cid:48) of the restriction AGV such that W (cid:48) → N M r,s⇒ M . Remark 2.3. Let V be a (locally) closed submanifold of M transverse to a groupoid G Denote by N M the (total space) of the normal bundle of V in M . Upon arguing locally, we can V assume that V is compact. By the transversality assumption the anchor (cid:92) : AGV → T MV induces a surjective bundle morphism AGV → N M is an isomorphism and using an exponential map, we thus obtain a submanifold W ⊂ G such that r : W → M is a diffeomorphism onto an open neighborhood of V in M and s is a submersion from W onto V . Replacing W by a an open subspace, we may assume that r(W ) is a tubular V ×V W → G defined by neighborhood of V in M , diffeomorphic to N M (γ1, γ2, γ3) (cid:55)→ γ1 ◦ γ2 ◦ γ−1 is a diffeomorphism and a groupoid isomorphism from the pull back groupoid (see next section) (GV V . The map W ×V GV V ×V W onto the open subgroupoid Gr(W ) s = W ×V GV V )s r(W ) of G. V 3 Pull back If f : V → M is transverse to a Lie groupoid G a Lie groupoid (a submanifold of V × G × V ). If fi : Vi → M are transverse to G (for i = 1, 2) then we obtain a Lie groupoid Gf1(cid:116)f2 f1(cid:116)f2 The linking manifold Gf1 f2 of the space of functions (half densities) with support in Gf1 f2 Fact 2.4. The bimodule C∗(Gf1 is a clopen submanifold. We denote by C∗(Gf1 r,s⇒ M , then the pull back groupoid Gf f is naturally ⇒ V1 (cid:116) V2. ) the closure in C∗(Gf1(cid:116)f2 ) f1(cid:116)f2 ) bimodule. ) is full if all the G orbits meeting f2(V2) meet also f1(V1). ; it is a C∗(Gf1 ) − C∗(Gf2 f2 f2 f1 f2 Morita equivalence r,s⇒ M1 and G2 r,s⇒ M2 are Morita equivalent if there exists a groupoid G Two Lie groupoids G1 and smooth maps fi : Mi → M transverse to G such that the pull back groupoids Gfi Gi and fi(Mi) meets all the orbits of G. Equivalently, a Morita equivalence is given by a linking manifold X with extra data: surjective smooth submersions r : X → G(0) 2 and compositions G1 ×s,r X → X, X ×s,r G2 → X, X ×r,r X → G2 and X ×s,s X → G1 with natural associativity conditions (see [42] for details). In the above situation, X is the manifold Gf1 and the extra data are the range and source maps f2 and the composition rules of the groupoid Gf1(cid:116)f2 f1(cid:116)f2 If the map r : X → G(0) is not necessarily surjective, then G1 is Morita equivalent to the restriction of G2 to the open saturated subspace s(X). We say that G1 is sub-Morita equivalent to G2. is surjective but s : X → G(0) ⇒ M1 (cid:116) M2 (see [42]). 1 and s : X → G(0) r,s⇒ M identify to fi 1 2 9 2.3 Semi-direct products Action of a groupoid on a space. Recall that an action of a groupoid G r,s⇒ G(0) on a space V is given by a map p : V → G(0) and the action G ×s,p V → V denoted by (g, x) (cid:55)→ g.x with the requirements p(g.x) = r(g), g.(h.x) = (gh).x and u.x = x if u = p(x). In that case, we may form the crossed product groupoid V (cid:111) G: • as a set V (cid:111) G is the fibered product V ×p,r G; • the unit space (V (cid:111) G)(0) is V . The range and source maps are r(x, g) = x and s(x, g) = g−1.x; • the composition is given by (x, g)(y, h) = (x, gh) (with g.y = x). If G is a Lie groupoid, M is a manifold and if all the maps defined are smooth and p is a submersion, then V (cid:111) G is a Lie groupoid. Action of a group on a groupoid. Let Γ be a Lie group acting on a Lie groupoid G r,s⇒ M by r(cid:111),s(cid:111)⇒ M we Lie groupoid automorphisms. The set G × Γ is naturally a Lie groupoid G (cid:111) Γ put r(cid:111)(g, γ) = r(g), s(cid:111)(g, γ) = γ−1(s(g)) and, when (g1, γ1) and (g2, γ2) are composable, their product is (g1, γ1)(g2, γ2) = (g1 γ1(g2), γ1γ2). Note that the semi-direct product groupoid G (cid:111) Γ is canonically isomorphic to the quotient G/Γ of the product G = G × (Γ × Γ) of G by the pair groupoid Γ × Γ where the Γ action on G is the diagonal one: γ · (g, γ1, γ2) = (γ(g), γ−1γ1, γ−1γ2). Free and proper action of a group on a groupoid. When the action of Γ on G (and therefore on its closed subset M = G(0)) is free and proper, we may define the quotient groupoid G/Γ r,s⇒ M/Γ. In that case, the groupoid G/Γ acts on M and the groupoid G identifies with the action groupoid M (cid:111) (G/Γ). Indeed, let p : M → M/Γ and q : G → G/Γ be the quotient maps. If x ∈ M and h ∈ G/Γ are such that s(h) = p(x), then there exists a unique g ∈ G such that q(g) = h and s(g) = x; we put then h.x = r(g). It is then immediate that ϕ : G → M×p,r(G/Γ) given by ϕ(g) = (r(g), q(g)) is a groupoid isomorphism. The groupoid G/Γ is Morita equivalent to G (cid:111) Γ: indeed one easily identifies G (cid:111) Γ with the pull back groupoid (G/Γ)q q where q : M → M/Γ is the quotient map. Note also that in this situation the action of Γ on G leads to an action of Γ on the Lie algebroid AG and A(G/Γ) identifies with AG/Γ. Remark 2.5. As the Lie groupoids we are considering need not be Hausdorff, the properness condition has to be relaxed. We will just assume that the action is locally proper, i.e. that every point in G has a Γ-invariant neighborhood on which the action of Γ is proper. Action of a groupoid on a groupoid. Recall that an action of a groupoid G rH ,sH⇒ H (0) is by groupoid automorphisms (cf. groupoid H a map p0 : H (0) → G(0), we have p = p0 ◦ rH = p0 ◦ sH and g.(xy) = (g.x)(g.y). In that case, we may form the crossed product groupoid H (cid:111) G = G: r,s⇒ G(0) on a [9]) if G acts on H (0) through • as a set H (cid:111) G is the fibered product H ×p,r G; • the unit space G(0) of G = H(cid:111)G is H (0). The range and source maps are rG(x, g) = rH (x) • the composition is given by (x, g)(y, h) = (x(g.y), gh). and sG(x, g) = g−1.sH (x); If G and H are Lie groupoids and if all the maps defined are smooth and p is a submersion, then G = H (cid:111) G is a Lie groupoid. 10 2.4 Index maps for Lie groupoids Recall (cf. [40, 45]) that if G is any Lie groupoid, the index map is an element in KK(C0(A∗G), C∗(G)) which can be constructed thanks to the adiabatic groupoid G[0,1] ad of G as indG = [ev0]−1 ⊗ [ev1] where ev0 : C∗(G[0,1] ad ) → C∗(Gad(0)) (cid:39) C0(A∗G) and ev1 : C∗(G[0,1] ad ) → C∗(Gad(1)) (cid:39) C∗(G) are the evaluation morphisms (recall that [ev0] is invertible). It follows quite immediately that the element (cid:102)indG ∈ KK1(C(S∗AG), C∗(G)) corresponding to the pseudodifferential exact sequence is the composition (cid:102)indG = indG ⊗ qA∗G where qA∗G ∈ KK1(C(S∗AG), C0(A∗G)) corresponds to the 0 → C∗(G) → Ψ∗(G) → C(S∗AG) → 0 EΨ∗(G) pseudodifferential exact sequence for AG which is 0 → C0(A∗G) → C(B∗AG) → C(S∗AG) → 0 This connecting element is immediately seen to be the element of KK(C0(S∗AG × R∗ associated to the inclusion of S∗AG × R∗ the zero section. EΨ∗(AG) +), C0(A∗G)) + as the open subset A∗G \ G(0) - where G(0) sits in A∗G as 3 Remarks on exact sequences, Connes-Thom elements, connect- ing maps and index maps The first part of this section is a brief reminder of some quite classical facts about connecting elements associated to short exact sequences of C∗-algebras. The second part is crucial for our main results of section 6: given a Lie groupoid and an open saturated subset of its unit space, we consider connecting maps and full index maps, compare them, compute them in some cases... In particular, we study a Fredholm realizability problem generalizing works of Albin and Melrose ([2, 3]) and study index maps using relative K-theory. In the last part we study a proper action of R∗ wich is R∗ G/R∗ + on a Lie groupoid G with an open saturated subset +-invariant. We compare the connecting maps and the index maps of G with those of +, using Connes-Thom morphisms. 3.1 A (well known) remark on exact sequences We will use the quite immediate (and well known) result: Lemma 3.1. Consider a commutative diagram of semi-split exact sequences of C∗-algebras 0 0 / J1 fJ / J2 A1 fA / A2 q1 q2 B1 fB / B2 0 / 0 a) We have ∂1 ⊗ [fJ ] = [fB] ⊗ ∂2 where ∂i denotes the element in KK1(Bi, Ji) associated with the exact sequence 0 / Ji / Ai / Bi / 0. b) If two of the vertical arrows are KK-equivalences, then so is the third one. 11 / / / / / / / / / / / / / / / Notation 3.2. When f : A → B is a morphism of C∗-algebra, we will denote the corresponding mapping cone by Cf = {(x, h) ∈ A ⊕ B[0, 1) ; h(0) = f (x)}. Proof. [14]. Let Cqi be the mapping cone of qi and ji : Bi(0, 1) → Cqi and ei : a) See e.g. Ji → Cqi the natural (excision) morphisms. The excision morphism ei is K-invertible and ∂i = [ji] ⊗ [ei]−1. b) For every separable C∗-algebra D, by applying the "five lemma" to the diagram KK∗+1(D, J1) / KK∗(D, J1) KK∗(D, B1) KK∗(D, A1) ... ... / KK∗(D, J2) / KK∗(D, A2) / KK∗(D, B2) / KK∗+1(D, J2) ... / ... we find that all vertical arrows are invertible. Applying this to D = J2 (resp. A2, B2) we find a one sided inverse to [fJ ] (resp. fA, fB). Applying this again to D = J1 (resp. A1, B1), it follows that this inverse is two-sided. 3.2 Saturated open subsets, connecting maps and full index map In this section, we let G ⇒ M be a Lie groupoid and F be a closed subset of M saturated for G. Put W = M \ F . Denote by GW the open subgroupoid GW = GW W of G and GF its complement. If F is not a submanifold, then GF is not a Lie groupoid, but as explained in remark 2.1, we still can define Ψ∗(GF ) (it is the quotient Ψ∗(G)/Ψ∗(GW )) the symbol map, etc. Define the full symbol algebra ΣW (G) to be the quotient Ψ∗(G)/C∗(GW ). In this section we will be interested in the description of elements ∂W G ∈ KK1(C∗(GF ), C∗(GW )) W f ull(G) ∈ KK1(ΣW (G), C∗(GW )) associated to the exact sequences and (cid:102)ind and 0 −→ C∗(GW ) −→ C∗(G) −→ C∗(GF ) −→ 0 0 −→ C∗(GW ) −→ Ψ∗(G) −→ ΣW (G) −→ 0. E∂ E(cid:102)indf ull To that end, it will be natural to assume that the restriction GF of G to F is amenable - so that the above sequences are exact and semi-split for the reduced as well as the full groupoid algebra. At some point, we wish to better control the K-theory of the C∗-algebras C∗(GF ) and ΣW (G). We will assume that the index element indGF ∈ KK(C0((A∗G)F ), C∗(GF )) is invertible. This assumption is satisfied in our main applications in section 6. 3.2.1 Connecting map and index Assume that the groupoid GF is amenable. We have a diagram 12 / / / / / / / / / / / / / / 0 0 E∂ : E(cid:102)indf ull : 0 0 / C∗(GW ) / C∗(G) / C∗(GW ) / Ψ∗(G) C∗(GF ) j / ΣW (G) 0 0 It follows that ∂W W f ull(G)) (proposition 3.1). G = j∗((cid:102)ind C0(S∗AG) C0(S∗AG) 0 0 3.2.2 Connecting maps Proposition 3.3. Let ∂W quence G ∈ KK1(C∗(GF ), C∗(GW )) be the element associated with the exact se- 0 −→ C∗(GW ) −→ C∗(G) −→ C∗(GF ) −→ 0. Similarly, let ∂W AG ∈ KK1(C0((A∗G)F ), C0((A∗G)W )) be associated with the exact sequence 0 −→ C0((A∗G)W ) −→ C0(A∗G) −→ C0((A∗G)F ) −→ 0. We have ∂W In particular, if the index element indGF ∈ KK(C0((A∗G)F ), C∗(GF )) is invertible, then the ele- ment ∂W AG ⊗ indGW = indGF ⊗ ∂W G . ⊗ ∂W AG ⊗ indGW . G is the composition ind−1 GF Proof. Indeed, we just have to apply twice proposition 3.1 using the adiabatic deformation G[0,1] ad and the diagram: 0 0 0 / C0((A∗G)W )) / C0(A∗G) / C0((A∗G)F )) ev0 / C∗(Gad(W × [0, 1]) ev1 / C∗(GW ) ev0 C∗(Gad) ev1 / C∗(G) ev0 C∗(Gad(F × [0, 1])) ev1 / C∗(GF ) / 0 / 0 / 0 3.2.3 A general remark on the index In the same way as the index indG ∈ KK(C0(A∗G), C∗(G)) constructed using the adiabatic groupoid is more primitive and to some extent easier to handle than (cid:102)indG ∈ KK1(C0(S∗AG), C∗(G)) con- structed using the exact sequence of pseudodifferential operators, there is in this "relative" situation a natural more primitive element. Denote by AW G = Gad(F × [0, 1) ∪ W × {0}) the restriction of Gad to the saturated locally closed subset F × [0, 1) ∪ W × {0}. Note that, since we assume that GF is amenable, and since AG is also amenable (it is a bundle groupoid), the groupoid AW G is amenable. 13 / / / / / / / / / / / / / / / / O O / / O O / / O O / / / / / Similarly to [15, 16], we define the noncommutative algebroid of G relative to F to be C∗(AW G). Note that by definition we have: C∗(G[0,1) ad )/C∗(Gad(W × (0, 1)) = C∗(Gad(F × [0, 1) ∪ W × {0}) = C∗(AW G) We have an exact sequence 0 → C∗(GW × (0, 1]) −→ C∗(Gad(F × [0, 1) ∪ W × [0, 1])) ev0−→ C∗(AW G) → 0, where ev0 : C∗(Gad(F × [0, 1) ∪ W × [0, 1])) → C∗(Gad(F × [0, 1) ∪ W × {0}) = C∗(AW G) is the restriction morphism. As C∗(GW × (0, 1]) is contractible the KK-class [ev0] ∈ KK(C∗(Gad(F × [0, 1) ∪ W × [0, 1])), C∗(AW G)) is invertible. Let as usual ev1 : C∗(Gad(F × [0, 1) ∪ W × [0, 1])) → C∗(GW ) be the evaluation at 1. We put: indW G = [ev0]−1 ⊗ [ev1] ∈ KK(C∗(AW G), C∗(GW )) . Recall from [17, Rem 4.10] and [18, Thm. 5.16] that there is a natural action of R on Ψ∗(G) such that Ψ∗(G) (cid:111) R is an ideal in C∗(G[0,1) ad ) (using a homeomorphism of [0, 1) with R+). This ideal is ad ) ev0−→ C0(A∗G) → C(M ). the kernel of the composition C∗(G[0,1) Recall that the restriction to C∗(G) of the action of R is inner. It follows that C∗(GW ) ⊂ Ψ∗(G) is invariant by the action of R - and C∗(GW ) (cid:111) R = C∗(GW ) ⊗ C0(R) = C∗(Gad(W × (0, 1))). We thus obtain an action of R on ΣW (G) = Ψ∗(G)/C∗(GW ) and an inclusion i : ΣW (G) (cid:111) R (cid:44)→ C∗(AW G). W f ull ∈ KK1(ΣW (G), C∗(GW )) corresponding to the exact se- E(cid:102)indf ull 0 −→ C∗(GW ) −→ Ψ∗(G) −→ ΣW (G) −→ 0. Proposition 3.4. The element (cid:102)ind quence is the Kasparov product of: • the Connes-Thom element [th] ∈ KK1(ΣW (G), ΣW (G) (cid:111) R); • the inclusion i : ΣW (G) (cid:111) R (cid:44)→ C∗(AW G); • the index indW G = [ev0]−1 ⊗ [ev1] defined above. Proof. By naturality of the Connes Thom element, it follows that (cid:102)ind f ull ⊗ [B] = −[th] ⊗ [∂] W where ∂ ∈ KK1(C∗(AW G), C∗(GW × (0, 1))) is the KK1-element corresponding with the exact sequence 0 / C∗(GW ) (cid:111) R / Ψ∗(G) (cid:111) R / ΣW (G) (cid:111) R / 0 and [B] ∈ KK1(C∗(GW ), C∗(GW ) (cid:111) R) is the Connes-Thom element. Note that, since the action is inner, [B] is actually the Bott element. By the diagram 0 0 / C∗(GW ) (cid:111) R Ψ∗(G) (cid:111) R ΣW (G) (cid:111) R / C∗(GW × (0, 1)) / C∗(G[0,1) ad ) i / C∗(AW G) / 0 / 0 we deduce that [∂] = i∗[∂(cid:48)] where ∂(cid:48) corresponds to the second exact sequence. 14 / / / / / / / / / / / / / / Finally, we have a diagram 0 0 0 0 / C∗(GW × (0, 1)) C∗(G[0,1) ad ) C∗(AW G) / C∗(GW × (0, 1]) / C∗(Gad(F × [0, 1) ∪ W × [0, 1])) ev0 / C∗(AW G) / 0 / 0 C∗(GW ) 0 ev1 C∗(GW ) 0 where exact sequences are semisplit. Now the connecting element corresponding to the exact se- quence / C∗(Gad(F × [0, 1) ∪ W × [0, 1])) ev0⊕ev1/ / C∗(AW G) ⊕ C∗(GW ) / 0 W f ull⊗[B] = −[th]⊗[i]⊗[∂(cid:48)] and [∂(cid:48)] = −[ev0]−1⊗[ev1]⊗[B], the result follows from invertibility [ev0] ⊗ [∂(cid:48)] + [ev1] ⊗ [B] = 0. / C∗(GW × (0, 1)) 0 is [∂(cid:48)] ⊕ [B] and it follows that As (cid:102)ind of the Bott element. 3.2.4 Full symbol algebra and index Denote by Ψ∗ F (G) the subalgebra C0(M ) + Ψ∗(GW ) of Ψ∗(G). It is the algebra of pseudodifferential operators which become trivial (i.e. multiplication operators) on F . Let ΣF (G) = Ψ∗ F (G)/C∗(GW ) be the algebra of the corresponding symbols. It is the subalgebra C0(M )+C0(S∗AGW ) of C0(S∗AG) of symbols a(x, ξ) with x ∈ M and ξ ∈ (S∗AG)x whose restriction on F does not depend on ξ. Lemma 3.5. Assume that the index element indGF ∈ KK(C0((A∗G)F ), C∗(GF )) is invertible, i.e. that the C∗-algebra of the adiabatic groupoid C∗(Gad(F × [0, 1))) is K-contractible. a) The inclusion jψ : C0(F ) → Ψ∗(GF ) is a KK-equivalence. b) The inclusion jσ : ΣF (G) = Ψ∗ a) Consider the diagram / C0((A∗G)F ) / C0((B∗AG)F ) 0 Proof. F (G)/C∗(GW ) → ΣW (G) is also a KK-equivalence. / C0((S∗AG)F ) / 0 / 0 / 0 ev0 ev0 ev0 / C∗(Gad(F × [0, 1])) Ψ∗(Gad(F × [0, 1])) C((S∗AG)F × [0, 1]) ev1 / C∗(GF ) ev1 / Ψ∗(GF ) ev1 / C((S∗AG)F ) 0 0 where the horizontal exact sequences are the pseudodifferential exact sequences EΨ∗(AG)F , EΨ∗(Gad(F×[0,1])) and EΨ∗(GF ). Since indGF is invertible ev1 : C∗(Gad(F × [0, 1]) → C∗(GF ) is a KK-equivalence. Hence, the left and right vertical arrows are all KK-equivalences, and therefore so are the middle ones. The inclusion C0(F ) in C0((B∗AG)F ) is a homotopy equiv- alence and therefore the inclusions C0(F ) → Ψ∗(Gad(F × [0, 1])) and C0(F ) → Ψ∗(GF ) are KK-equivalences. 15 / / / / / / / / / / / / / / / / / / O O / / O O / / O O / / / / / b) Apply Lemma 3.1 to the diagrams 0 0 0 0 / Ψ∗(GW ) / Ψ∗(GW ) / C∗(GW ) / C∗(GW ) / Ψ∗ F (G) Jψ / Ψ∗(G) / Ψ∗ F (G) Jψ / Ψ∗(G) C0(F ) jψ / Ψ∗(GF ) ΣF (G) jσ / ΣW (G) 0 / 0 0 / 0 we find that JΨ and jσ are K-equivalences. The diagram in lemma 3.5.b) shows that ∂F = j∗ the KK-element associated with the exact sequence σ((cid:102)ind W f ull(G)) where ∂F ∈ KK1(ΣF (G), C∗(GW )) is 0 / C∗(GW ) / Ψ∗ F (G) / ΣF (G) / 0. So, let's compute the KK-theory of ΣF (G) and the connecting element ∂F . Consider the vector bundle AG as a groupoid (with objects M ). It is its own algebroid - with anchor 0. With the notation in 1.1, • C∗(AG) identifies with C0(A∗G) and C∗(AGW ) with C0(A∗GW ); • Ψ∗(AG) identifies with C0(B∗AG); it is homotopy equivalent to C0(M ); • the spectrum of Ψ∗ and (x, η) for x ∈ F ; it is also homotopy equivalent to C0(M ). F (AG) is B∗ F AG the quotient of B∗AG where we identify two points (x, ξ) • the algebroid of the groupoid AG is AG itself; therefore, ΣF (AG) = ΣF (G); its spectrum is S∗ F AG which is the image of S∗AG in B∗ F AG. We further note. a) Let k : C0(A∗GW ) → C0(M ) be given by k(f )(x) = (cid:40) f (x, 0) 0 if x ∈ W if x ∈ F . We find a com- mutative diagram C0( B∗AGW ) C0(B∗ F AG) where the vertical arrows are homotopy C0(A∗GW ) k / C0(M ) equivalences. b) the exact sequence 0 → C∗(AGW ) → Ψ∗ C0(B∗ F AG) → C0(S∗ F AG) → 0. F (AG) → ΣF (AG) → 0, reads 0 → C0( B∗AGW ) → We deduce using successively (b) and (a): a) The algebra C0(S∗ F AG). inclusion C0( B∗AGW ) → C0(B∗ Proposition 3.6. F AG) is KK1-equivalent with the mapping cone of the b) This mapping cone is homotopy equivalent to the mapping cone of the morphism k. (cid:3) 16 / / / / / / / / / / / / / / / / / / / / / / / / / / / Note finally that we have a diagram / C0((A∗G)W ) 0 ev0 / Ψ∗ F (AG) ev0 / ΣF (AG) ev0 0 0 / C∗(Gad(W × [0, 1])) Ψ∗ F×[0,1](Gad) ΣF×[0,1](Gad) ev1 / C∗(GW ) ev1 / Ψ∗ F (G) ev1 / ΣF (AG) / 0 / 0 / 0 The right vertical arrows are KK-equivalences, and therefore we find ∂⊗ [ev0]−1⊗ [ev1] = ∂F , where ∂ is the connecting element of the first horizontal exact sequence. To summarize, we have proved: Proposition 3.7. Assume that the index element indGF ∈ KK(C0((A∗G)F ), C∗(GF )) is invertible. b) The analytic index (cid:102)ind f ull(G) ∈ KK1(ΣW (G), C∗(GW )) corresponding to the exact sequence a) The inclusion jσ : ΣF (G) → ΣW (G) is a KK-equivalence. W 0 / C∗(GW ) / Ψ∗(G) / ΣW (G) / 0 is the Kasparov product of • the element [jσ]−1 ∈ KK(ΣW (G), ΣF (G)); • the connecting element ∂ ∈ KK1(ΣF (AG), C0((A∗G)W )) associated with the exact se- quence of (abelian) C∗-algebras 0 / C0((A∗G)W ) / Ψ∗ F (AG) / ΣF (AG) / 0; • the analytic index element indGW of GW , i.e. the element [ev0]−1 ⊗ [ev1] ∈ KK(C0((A∗G)W ), C∗(GW )). (cid:3) 3.2.5 Fredholm realization Let σ be a classical symbol which defines an element in K1(C0(S∗AG)). A natural question is: when can this symbol be lifted to a pseudodifferential element which is invertible modulo C∗(GW )? In particular, if GW is the pair groupoid W × W , this question reads: when can this symbol be extended to a Fredholm operator? Particular cases of this question were studied in [2, 3]. Consider the exact sequences: 0 0 E : 0 0 / C∗(GF ) ΣW (G) q C0(S∗AG) / C∗(G) Ψ∗(G) / C0(S∗AG) / 0 / 0 C∗(GW ) C∗(GW ) 0 0 17 / / / / / O O / / O O / / O O / / / / / / / / / / / / / / / / O O / / O O / / / / O O O O / / O O O O O O O O The element σ is an invertible element in Mn(C0(S∗AG)+) (where C0(S∗AG)+ is obtained by ad- joining a unit to C0(S∗AG) - if G(0) is not compact). The question is: when can σ be lifted to an invertible element of Mn(ΣW (G)+). By the K-theory exact sequence, if this happens then the class of σ is in the image of K1(ΣW (G)) and therefore its image via the connecting map of the exact sequence E is 0 in K0(C∗(GF )). Conversely, if the image of σ via the connecting map of E vanishes, then the class of σ in K1(C0(S∗AG)) is in the image of K1(ΣW (G)). This means that there exists p ∈ N and an invertible element x ∈ Mn+p(ΣW (G)+) such that q(x) and σ ⊕ 1p are in the same path connected component of GLn+p(C0(S∗AG)+). Now the morphism q : Mn+p(ΣW (G)+) → Mn+p(C0(S∗AG)+) is open and therefore the image of (cid:17) the connected component GLn+p(ΣW (G)+)(0) of 1n+p in GLn+p(ΣW (G)+) is an open (and therefore also closed) subgroup of GLn+p(C0(S∗AG)+). It follows immediately that q = GLn+p(C0(S∗AG)+)(0). Finally (σ ⊕ 1p)x−1 is in the image of GLn+p(ΣW (G)+)(0), therefore σ ⊕ 1p can be lifted to an invertible element of Mn(ΣW (G)+). Let us make a few comments: GLn+p(ΣW (G)+)(0) (cid:16) a) Considering the diagram 0 0 / C∗(GF ) / ΣW (G) C0(S∗AG) / C∗(GF ) / Ψ∗(GF ) / C0(S∗AGF ) / 0 / 0 we find that the image of σ in K0(C∗(GF )) is the index ind(σF ) of the restriction σF of σ to F . b) Considering the diagram 0 0 0 0 0 0 / C∗(GW ) Ψ∗(GW ) C0(S∗AGW ) / C∗(G) Ψ∗(G) C0(S∗AG) / C∗(GF ) Ψ∗(GF ) C0(S∗AGF ) 0 / 0 0 0 0 0 we could also say that our question is: when is the index ind(σ) ∈ K0(C∗(G)) in the image of K0(C∗(GW )), and of course this happens if and only if the image of ind(σ) in K0(C∗(GF )) vanishes. Again we may notice that the image of ind(σ) in K0(C∗(GF )) is ind(σF ). c) Of course, the same remark holds if we start with a symbol defining a class in K0(C0(S∗AG)). 3.2.6 Relative K-theory and full index It is actually better to consider the index map in a relative K-theory setting. Indeed, the starting point of the index problem is a pair of bundles E± over M together with a pseudodifferential operator P from sections of E+ to sections of E− which is invertible modulo C∗(GW ). Consider the morphism 18 / / / / / / / / / / / / / / / / / / / / / / / / / / / / / of the morphism µ. The index · ⊗ (cid:102)ind ψ : C0(M ) → Ψ∗(G) which associates to a (smooth) function f the order 0 (pseudo)differential operator multiplication by f and σf ull : Ψ∗(G) → ΣW (G) the full symbol map. Put µ = σf ull ◦ ψ. By definition, for any P ∈ Ψ∗(G), the triple (E±, σf ull(P )) is an element in the relative K-theory W f ull(G) considered in the previous section is the composition of the morphism K1(ΣW (G)) → K0(µ) 1 with the index map indrel : K0(µ) → K0(C∗(GW )) which to (E±, σf ull(P )) associates the class of P . The morphism indrel can be thought of as the composition of the obvious morphism K0(µ) → K0(σf ull) (cid:39) K0(ker(σf ull)) = K0(C∗(GW )). Let us now compute the group K∗(µ) and the morphism indrel when the index element indGF ∈ KK(C0((A∗G)F ), C∗(GF )) is invertible. Proposition 3.8. Assume that the index element indGF ∈ KK(C0((A∗G)F ), C∗(GF )) is invertible. Then K∗(µ) is naturally isomorphic to K∗(C0(A∗GW )). Under this isomorphism, indrel identifies with indGW . Proof. We have a diagram 0 0 / C0(W ) C0(M ) C0(F ) i / C0(S∗AGW ) µ / ΣW (G) jΨ / Ψ∗(GF ) / 0 / 0 As jΨ is an isomorphism in K-theory, the map K∗(i) → K∗(µ) induced by the first commutative square of this diagram is an isomorphism. As K∗(i) = K∗(Ci) and Ci = C0(A∗GW ), we obtained the desired isomorphism K∗(C0(A∗GW )) (cid:39) K∗(µ). Comparing the diagrams C0(W ) C0(M ) i C0(S∗AGW ) µ / ΣW (G) we find that the composition K∗(i) K∗(i)−→K∗(σW ) ∼−→ K∗(σf ull). Ψ∗(G) σf ull and C0(W ) i Ψ∗(GW ) σW C0(S∗AGW ) ΣW (G) ∼−→ K∗(µ) −→ K∗(σf ull) coincides with the index C0(S∗AGW ) Ψ∗(G) σf ull / ΣW (G) Remark 3.9. We wrote the relative index map in terms of morphisms of K-groups. One can also write everything in terms KK-theory, by replacing relative K-theory by mapping cones, i.e. C([e]−1) where e : construct the relative index as the element of KK(Cµ, C∗(GW )) given as ψ∗ C∗(GW ) → Cσf ull is the (KK-invertible) "excision map" associated with the (semi-split) exact σf ull−→ ΣF (G) → 0 and ψC : Cµ → Cσf ull is the morphism associated sequence 0 → C∗(GW ) → Ψ∗(G) with ψ. 3.3 Connes-Thom elements and quotient of a groupoid by R∗ 3.3.1 Proper action on a manifold Remark 3.10 (Connes-Thom elements). Let R∗ M . We have a canonical invertible KK-element α = (H, D) ∈ KK1(C0(M ), C0(M/R∗ + +)). + act smoothly (freely and) properly on a manifold 1Recall that if f : A → B is a morphism of C ∗ -algebras, we have a natural morphism u : K∗+1(B) → K∗(f ) corresponding to the inclusion of the suspension of B in the cone of f . 19 / / / / / / / / / / / / / / / / / / / / • The Hilbert module H is obtained as a completion of Cc(M ) with respect to the C0(M/R∗ +) for ξ, η ∈ Cc(M ), where p : M → ξ(t.x) η(t.x) valued inner product (cid:104)ξη(cid:105)(p(x)) = M/R∗ + is the quotient map. dt t (cid:90) +∞ 0 • The operator D is 1 t ∂ ∂t . The inverse element β ∈ KK1(C0(M/R∗ +), C0(M )) is constructed in the following way: C0(M/R∗ +) sits in the multipliers of C0(M ). One may define a continuous function f : M → [−1, 1] such that, t→±∞ f (et.x) = ±1. The pair (C0(M ), f ) is then an element in uniformly on compact sets of M , KK1(C0(M/R∗ +), C0(M )). To construct f , one may note that, by properness, we actually have a + → M defined + → M and we may thus construct a homeomorphism R∗ section ϕ : M/R∗ by (t, x) (cid:55)→ t.ϕ(x). Then put f (et, x) = t(1 + t2)−1/2. As an extension of C∗-algebras the element β is given by considering P = (M × R+)/R∗ + (where R∗ + acts -properly - diagonally). Then M sits as an open subset (M × R∗ + and we have an exact sequence 0 → C0(M ) → C0(P ) → C0(M/R∗ + × M/R∗ +)/R∗ +) → 0. lim G/R∗ + (C0(M/R∗ + act smoothly (locally cf. remark 2.5) properly on a Lie groupoid G ⇒ M . The + acts on M , and the element α is G invariant - and β is almost G invariant in +)) and β ∈ +), C0(M )) which are inverses of each other in Le Gall's equivariant KK-theory (α) ∈ +), C∗(G)) that are also inverses of each 3.3.2 Proper action on a groupoid Let now R∗ groupoid G/R∗ the sense of [30]. In other words, we obtain elements α ∈ KK1 KK1 for groupoids. Using the descent morphism of Kasparov ([28]) and Le Gall ([30]), we obtain elements jG/R∗ KK1(C∗(G), C∗(G/R∗ other. Note also that the element βG = jG/R∗ (β) is the connecting element of the extension of groupoid C∗-algebras 0 → C∗(G) −→ C∗(G) ev0−→ C∗(G/R∗ + and ev0 comes from the evaluation at G×{0}. Using the pseudodifferential operators on the groupoid G, we obtain a KK-element βΨ +) → 0, where G = (G × R+)/R∗ +), Ψ∗(G)). We obtain a commutative diagram (β) ∈ KK1(C∗(G/R∗ G ∈ KK1(Ψ∗(G/R∗ (C0(M ), C0(M/R∗ +)) and jG/R∗ G/R∗ + + + + 0 0 0 / C∗(G) / Ψ∗(G) C∗(G) Ψ∗(G) C∗(G/R∗ +) Ψ∗(G/R∗ +) / C0(S∗AG) C0(S∗AG) C0(S∗A(G/R∗ +)) 0 0 0 0 0 0 0 0 0 The third horizontal exact sequence corresponds to the proper action of R∗ + on S∗AG. In fact S∗AG +)) × R+. As the connecting elements of is homeomorphic (using a cross section) to (S∗A(G/R∗ the first and third horizontal (semi-split) exact sequences are invertible, it follows that C∗(G) and C0(S∗AG) are K-contractible, whence so is Ψ∗(G) and therefore βΨ G is a KK-equivalence. Hence we have obtained: 20 / / / / / / / / / / / / / / / / / / / / / Proposition 3.11. If R∗ ing elements βG ∈ KK1(C∗(G/R∗ KK1(Ψ∗(G/R∗ +), Ψ∗(G)) are KK-equivalences. + acts smoothly (locally) properly on the Lie groupoid G, the connect- G ∈ +), C∗(G)), βS∗AG ∈ KK1(C0(S∗A(G/R∗ +), C0(S∗AG)) and βΨ 3.3.3 Closed saturated subsets and connecting maps If W is an open saturated subset in M for the actions of G and of R∗ the corresponding elements. We then obtain a diagram C∗(G/R∗ +) C∗(GF 0 + and F = M \ W , one compares / C∗(GW W /R∗ +) β(cid:48) / C∗(GW W ) 0 F /R∗ +) β(cid:48)(cid:48) / C∗(GF F ) 0 / 0 β / C∗(G) F is amenable, ∂W F ), C∗(GW G/R∗ W )) and ∂W F × [0, 1) ∪ G × {0} which is the restriction of the groupoid where the horizontal arrows are morphisms and the vertical ones KK1-equivalences. Using the deformation groupoid G = GF G × [0, 1) ⇒ M × [0, 1] to the closed saturated subset F × [0, 1) ∪ M × {0}, we obtain: F /R∗ +), C∗(GW W )) +)) denote the Proposition 3.12. If GF G ∈ KK1(C∗(GF where ∂W KK-elements associated with the above exact sequences. p→ A/J → 0 is Proof. Indeed, the connecting map of a semi-split exact sequence 0 → J → A obtained as the KK-product of the morphism A/J(0, 1) → Cp with the KK-inverse of the morphism J → Cp. The − sign comes from the fact that we have naturally elements of KK(C∗(GF + × (0, 1)2), C∗(GW Note also that the same holds for Ψ∗ in place of C∗. W )) which are equal but with opposite orientations of (0, 1)2. G ∈ KK(C∗(GF W /R∗ F /R∗ ⊗ β(cid:48) = −β(cid:48)(cid:48) ⊗ ∂W ∈ KK1(C∗(GF +), C∗(GW F /R∗ G/R∗ + + 3.3.4 Connes-Thom invariance of the full index Let W be as above: an open subset of M saturated for G and invariant under the action of R∗ +. One compares the corresponding (cid:102)indf ull elements. Indeed, we have a diagram 0 0 / C∗(GW W /R∗ +) Ψ∗(G/R∗ +) ΣW/R∗ +(G/R∗ +) βGW / C∗(GW W ) βG Ψ / Ψ∗(G) β(G,W ) Σ / ΣW (G) 0 / 0 E(cid:102)indf ull : where the horizontal arrows are morphisms and the vertical ones KK1-elements. As βGW and βG Ψ are invertible, we deduce as in prop. 3.12: Proposition 3.13. b) We have β(G,W ) Σ a) The element β(G,W ) W/R∗ f ull f ull(G) = −(cid:102)ind ⊗ (cid:102)ind W Σ + is invertible. (G/R∗ +) ⊗ βGW . (cid:3) 4 Two classical geometric constructions: Blowup and deformation to the normal cone One of the main object in our study is a Lie groupoid G based on a groupoid restricted to a half space. This corresponds to the inclusion of a hypersurface V of G(0) into G and gives rise to the "gauge adiabatic groupoid" gaG. The construction of gaG is in fact a particular case of the blowup 21 / / / / / / / / / / / / / / / / / / / / / / construction corresponding to the inclusion of a Lie subgroupoid into a groupoid. In this section, we will explain this general construction. We will give a more detailed description in the case of an inclusion V → G when V is a submanifold of G(0). Let Y be a manifold and X a locally closed submanifold (the same constructions hold if we are given an injective immersion X → Y ). Denote by N Y X the (total space) of the normal bundle of X in Y . 4.1 Deformation to the normal cone X ×{0} with Y × R∗. The The deformation to the normal cone DN C(Y, X) is obtained by gluing N Y smooth structure of DN C(Y, X) is described by use of any exponential map θ : U(cid:48) → U which is a diffeomorphism from an open neighborhood U(cid:48) of the 0-section in N Y X to an open neighborhood U of X. The map θ is required to satisfy θ(x, 0) = x for all x ∈ X and px ◦ dθx = p(cid:48) x where X )x ⊕ (TxX) → (N Y px : TxY → (N Y X )x are the projections. The manifold structure of DN C(Y, X) is described by the requirement that: X )x = (TxY )/(TxX) and p(cid:48) X (cid:39) (N Y x : TxN Y a) the inclusion Y × R∗ → DN C(Y, X) and b) the map Θ : Ω(cid:48) = {((x, ξ), λ) ∈ N Y ((x, ξ), 0) and Θ((x, ξ), λ) = (θ(x, λξ), λ) ∈ Y × R∗ if λ (cid:54)= 0. X ×R; (x, λξ) ∈ U(cid:48)} → DN C(Y, X) defined by Θ((x, ξ), 0) = are diffeomorphisms onto open subsets of DN C(Y, X). It is easily shown that DN C(Y, X) has indeed a smooth structure satisfying these requirements and that this smooth structure does not depend on the choice of θ. (See for example [10] for a detailed description of this structure). In other words, DN C(Y, X) is obtained by gluing Y × R∗ with Ω(cid:48) by means of the diffeomorphism Θ : Ω(cid:48) ∩ (N Y X × R∗) → U × R∗. Let us recall the following facts which are essential in our construction. Definitions 4.1. The gauge action of R∗. The group R∗ acts on DN C(Y, X) by λ.(w, t) = (w, λt) and λ.((x, ξ), 0) = ((x, λ−1ξ), 0)(cid:0)with λ, t ∈ R∗, w ∈ Y , x ∈ X and ξ ∈ (N Y (cid:1). X )x Functoriality. Given a commutative diagram of smooth maps X fX X(cid:48) Y fY / Y (cid:48) where the horizontal arrows are inclusions of submanifolds, we naturally obtain a smooth map DN C(f ) : DN C(Y, X) → DN C(Y (cid:48), X(cid:48)). This map is defined by DN C(f )(y, λ) = (fY (y), λ) for y ∈ Y and λ ∈ R∗ and DN C(f )(x, ξ, 0) = (fX (x), fN (ξ), 0) for x ∈ X and ξ ∈ (N Y X )x = TxY /TxX where fN : Nx → (N Y (cid:48) X(cid:48) )fX (x) = TfX (x)Y (cid:48)/TfX (x)X(cid:48) is the linear map induced by the differential (dfY )x at x. This map is of course equivariant with respect to the gauge action of R∗. Remarks 4.1. Let us make a few remarks concerning the DNC construction. a) The map equal to identity on X × R∗ and sending X × {0} to the zero section of N Y X leads to an embedding of X × R into DN C(Y, X), we may often identify X × R with its image in DN C(Y, X). As DN C(X, X) = X × R, this corresponds to the naturality of the diagram X X X / Y 22 / / / / / / b) We have a natural smooth map π : DN C(Y, X) → Y × R defined by π(y, λ) = (y, λ) (for X )x a normal vector). y ∈ Y and λ ∈ R∗) and π((x, ξ), 0) = (x, 0) (for x ∈ X ⊂ Y and ξ ∈ (N Y This corresponds to the naturality of the diagram X Y Y / Y c) If Y1 is an open subset of Y2 such that X ⊂ Y1, then DN C(Y1, X) is an open subset of DN C(Y2, X) and DN C(Y2, X) is the union of the open subsets DN C(Y1, X) and Y2 × R∗. This reduces to the case when Y1 is a tubular neighborhood - and therefore to the case where Y is (diffeomorphic to) the total space of a real vector bundle over X. In that case one gets DN C(Y, X) = Y × R and the gauge action of R∗ on DN C(Y, X) = Y × R is given by λ.((x, ξ), t) = ((x, λ−1ξ), λt) (with λ ∈ R∗, t ∈ R, x ∈ X and ξ ∈ Yx). d) More generally, let E be (the total space of) a real vector bundle over Y . Then DN C(E, X) identifies with the total space of the pull back vector bundle π∗(E) over DN C(Y, X), where π is the composition of π : DN C(Y, X) → Y × R (remark b) with the projection Y × R → Y . The gauge action of R∗ is λ.(w, ξ) = (λ.w, λ−1ξ) for w ∈ DN C(Y, X) and ξ ∈ Eπ(w). e) Let X1 be a (locally closed) smooth submanifold of a smooth manifold Y1 and let f : Y2 → Y1 be identifies by the restriction X2 → X1 of f . It follows that DN C(Y2, X2) a smooth map transverse to X1. Put X2 = f−1(X1). Then the normal bundle N Y2 with the pull back of N Y1 identifies with the fibered product DN C(Y1, X1) ×Y1 Y2. X1 X2 f) More generally, let Y, Y1, Y2 be smooth manifolds and fi : Yi → Y be smooth maps. Assume that f1 is transverse to f2. Let X ⊂ Y and Xi ⊂ Yi be (locally closed) smooth submanifolds. Assume that fi(Xi) ⊂ X and that the restrictions gi : Xi → X of fi are transverse also. We thus have a diagram X1 g1 / X Y1 f1 / Y g2 f2 X2 Y2 Then the maps DN C(fi) : DN C(Yi, Xi) → DN C(Y, X) are transverse and the deformation to the normal cone of fibered products DN C(Y1 ×Y Y2, X1 ×X X2) identifies with the fibered product DN C(Y1, X1) ×DN C(Y,X) DN C(Y2, X2). 4.2 Blowup constructions The blowup Blup(Y, X) is a smooth manifold which is a union of Y \X with the (total space) P(N Y X ) of the projective space of the normal bundle N Y X of X in Y . We will also use the "spherical version" SBlup(Y, X) of Blup(Y, X) which is a manifold with boundary obtained by gluing Y \ X with the X ). We have an obvious smooth onto map SBlup(Y, X) → (total space of the) sphere bundle S(N Y Blup(Y, X) with fibers 1 or 2 points. These spaces are of course similar and we will often give details in our constructions to the one of them which is the most convenient for our purposes. We may view Blup(Y, X) as the quotient space of a submanifold of the deformation to the normal cone DN C(Y, X) under the gauge action of R∗. Recall that the group R∗ acts on DN C(Y, X) by λ.(w, t) = (w, λt) and λ.((x, ξ), 0) = ((x, λ−1ξ), 0) (with λ, t ∈ R∗, w ∈ Y , x ∈ X and ξ ∈ (N Y X )x). This action is easily seen to be free and (locally cf. remark 2.5) proper on the open subset DN C(Y, X) \ X × R (see remark 4.4 below). Notation 4.2. For every locally closed subset T of R containing 0, we define DN CT (Y, X) = X × {0} = π−1(Y × T ) (with the notation of remark 4.1.b). It is the restriction of Y × (T \ {0}) ∪ N Y DN C(Y, X) to T . We put DN C+(Y, X) = DN CR+(Y, X) = Y × R∗ X × {0}. + ∪ N Y 23 / / / _ / _ _ o o / o o Definition 4.3. We put DN C(Y, X) \ X × R(cid:17) Blup(Y, X) = (cid:16) (cid:16) /R∗ (cid:17) and SBlup(Y, X) = DN C+(Y, X) \ X × R+ /R∗ +. Remark 4.4. With the notation of section 4.1, Blup(Y, X) is thus obtained by gluing Y \ X = ((Y \ X) × R∗)/R∗, with (Ω(cid:48) \ (X × R))/R∗ using the map Θ which is equivariant with respect to the gauge action of R∗. X . Let S = {((x, ξ), λ) ∈ Ω(cid:48); (cid:107)ξ(cid:107) = 1}. The map Θ induces a Choose a euclidean metric on N Y X ) in Blup(Y, X) and τ is the map ((x, ξ), λ) (cid:55)→ ((x,−ξ),−λ). In this way, with a Riemannian metric on Y , we may naturally associate a Riemannian metric on diffeomorphism of S/τ with an open neighborhood (cid:101)Ω of P(N Y Blup(Y, X) (using a partition of the identity to glue the metric of Y \ X with that of (cid:101)Ω). Since π : DN C(Y, X) → Y is invariant by the gauge action of R∗, we obtain a natural smooth map π : Blup(Y, X) → Y whose restriction to Y \ X is the identity and whose restriction to P(N Y X ) is the canonical projection P(N Y X ) → X ⊂ Y . This map is easily seen to be proper. Remark 4.5. Note that, according to remark 4.1.d), DN C(Y, X) canonically identifies with the open subset Blup(Y × R, X × {0}) \ Blup(Y × {0}, X × {0}) of Blup(Y × R, X × {0}). Thus, one may think at Blup(Y × R, X × {0}) as a "local compactification" of DN C(Y, X) (since the map Blup(Y × R, X × {0}) → Y × R is proper). Example 4.6. In the case where Y is a real vector bundle over X, Blup(Y, X) identifies non canonically with an open submanifold of the bundle of projective spaces P(Y × R) over X. Indeed, in that case DN C(Y, X) = Y × R; choose a euclidian structure on the bundle Y . Consider the smooth involution Φ from (Y \ X) × R onto itself which to (x, ξ, t) associates (x, (cid:107)ξ(cid:107)2 , t) (for x ∈ X, ξ ∈ Yx, t ∈ R). This map transforms the gauge action of R∗ on DN C(Y, X) into the action of R∗ by dilations on the vector bundle Y ×R over X and thus defines a diffeomorphism of Blup(Y, X) into its image which is the open set P(Y ×R)\ X where X embeds into P(Y ×R) by mapping x ∈ X to the line {(x, 0, t), t ∈ R}. ξ Remark 4.7. Since we will apply this construction to morphisms of groupoids that need not be proper, we have to relax properness as in remark 2.5: we will say that f : Y → X is locally proper if every point in X has a neighborhood V such that the restriction f−1(V ) → V of f is proper. In particular, if Y is a non Hausdorff manifold and X is a locally closed submanifold of Y , then the map Blup(Y × R, X × {0}) → Y × R is locally proper Functoriality Definition 4.8 (Functoriality). Let X fX X(cid:48) Y fY / Y (cid:48) be a commutative diagram of smooth maps, where the horizontal arrows are inclusions of closed submanifolds. Let Uf = DN C(Y, X) \ DN C(f )−1(X(cid:48) × R) be the inverse image by DN C(f ) of the complement in DN C(Y (cid:48), X(cid:48)) of the subset X(cid:48) × R. We thus obtain a smooth map Blup(f ) : Blupf (Y, X) → Blup(Y (cid:48), X(cid:48)) where Blupf (Y, X) ⊂ Blup(Y, X) is the quotient of Uf by the gauge action of R∗. In particular, 24 / / / a) If X ⊂ Y1 are (locally) closed submanifolds of a manifold Y2, then Blup(Y1, X) is a submanifold of Blup(Y2, X). b) Also, if Y1 is an open subset of Y2 such that X ⊂ Y1, then Blup(Y1, X) is an open subset of Blup(Y2, X) and Blup(Y2, X) is the union of the open subsets Blup(Y1, X) and Y2 \ X. This reduces to the case when Y1 is a tubular neighborhood. Fibered products Let X1 be a (locally closed) smooth submanifold of a smooth manifold Y1 and let f : Y2 → Y1 be a smooth map transverse to X1. Put X2 = f−1(X1). Recall from remark 4.1.e that in this situation DN C(Y2, X2) identifies with the fibered product DN C(Y1, X1)×Y1 Y2. Thus Blup(Y2, X2) identifies with the fibered product Blup(Y1, X1) ×Y1 Y2. 5 Constructions of groupoids 5.1 Linear groupoids We will encounter groupoids with an extra linear structure which are special cases of VB groupoids/ in the sense of Pradines [46, 31]. We will also need to consider the spherical and projective analogues. Let E be a vector space over a field K and let F be a vector sub-space. Let r, s : E → F be linear retractions of the inclusion F → E. 5.1.1 The linear groupoid The space E is endowed with a groupoid structure E with base F . The range and source maps are r and s and the product is (x, y) (cid:55)→ (x · y) = x + y − s(x) for (x, y) composable, i.e. such that s(x) = r(y). One can easily check: • Since r and s are linear retractions: r(x · y) = r(x) and s(x · y) = s(y). • If (x, y, z) are composable, then (x · y) · z = x + y + z − (r + s)(y) = x · (y · z). • The inverse of x is (r + s)(x) − x. Remarks 5.1. a) Note that, given E and linear retractions r and s on F , E ⇒ F is the only possible linear groupoid structure(2) on E . Indeed, for any x ∈ E one must have x · s(x) = x and r(x) · x = x. By linearity, it follows that for every composable pair (x, y) = (x, s(x)) + (0, y − s(x)) we have x · y = x · s(x) + 0 · (y − s(x)) = x + y − s(x). b) The morphism r − s : E/F → F gives an action of E/F on F by addition. The groupoid associated to this action is in fact E. c) Given a linear groupoid structure on a vector space E, we obtain the "dual" linear groupoid structure E∗ on the dual space E∗ given by the subspace F ⊥ = {ξ ∈ E∗; ξF = 0} and the two retractions r∗, s∗ : E∗ → F ⊥ with kernels (ker r)⊥ and (ker s)⊥: for ξ ∈ E∗ and x ∈ E, r∗(ξ)(x) = ξ(x − r(x)) and similarly s∗(ξ)(x) = ξ(x − s(x)). 2A linear groupoid is a groupoid G such that G(0) and G are vector spaces and all structure maps (unit, range, source, product) are linear. 25 5.1.2 The projective groupoid The multiplicative group K∗ acts on E by groupoid automorphisms. This action is free on the restriction (cid:101)E = E \ (ker r ∪ ker s) of the groupoid E to the subset F \ {0} of E (0) = F . The projective groupoid is the quotient groupoid PE = (cid:101)E/K∗. It is described as follows. As a set PE = P(E) \ (P(ker r) ∪ P(ker s)) and P (0) = P(F ) ⊂ P(E). The source and range maps r, s : PE → P(F ) are those induced by r, s : E → F . The product of x, y ∈ PE with s(x) = r(y) is the line x · y = {u + v − s(u); u ∈ x, v ∈ y; s(u) = r(v)}. The inverse of x ∈ PE is (r + s − id)(x). Remarks 5.2. a) When F is just a vector line, PE is a group. Let us describe it: we have a canonical morphism h : PE → K∗ defined by r(u) = h(x)s(u) for u ∈ x. The kernel of h is P(ker(r − s)) \ P(ker r). Note that F ⊂ ker(r − s) and therefore ker(r − s) (cid:54)⊂ ker r, whence ker r ∩ ker(r − s) is a hyperplane in ker(r − s). The group ker h is then easily seen to be isomorphic to ker(r) ∩ ker(s). Indeed, choose a non zero vector w in F ; then the map which assigns to u ∈ ker(r) ∩ ker(s) the line with direction w + u gives such an isomorphism onto ker h. Then: • If r = s, PE is isomorphic to the abelian group ker(r) = ker(s). • If r (cid:54)= s, choose x such that r and s do not coincide on x and let P be the plane F ⊕ x. The subgroup P(P )\{ker r∩ P, ker s∩ P} of PE is isomorphic through h with K∗. It thus defines a section of h. In that case PE is the group of dilations (ker(r) ∩ ker(s)) (cid:111) K∗. b) In the general case, let d ∈ P(F ). Put Ed d = r−1(d) ∩ s−1(d). d ) \ (P(ker r) ∪ P(ker s)) described above. • The stabilizer (PE)d • The orbit of a line d is the set of r(x) for x ∈ PE such that s(x) = d. It is therefore d is the group PEd d = P(Ed P(d + r(ker s)). c) the following are equivalent: (i) (r, s) : E → F × F is onto; (ii) r(ker s) = F ; (iii) (r − s) : E/F → F is onto; (iv) the groupoid PE has just one orbit. d) When r = s, the groupoid PE is the product of the abelian group E/F by the space P(F ). When r (cid:54)= s, the groupoid (cid:101)E is Morita equivalent to E since F \ {0} meets all the orbits of E. groupoid crossed-product (cid:101)E (cid:111) K∗. If K is a locally compact field and r (cid:54)= s, the smooth groupoid PE is Morita equivalent to the In all cases, when K is a locally compact field, PE is amenable. 5.1.3 The spherical groupoid If the field is R, we may just take the quotient by R∗ + instead of R∗. We then obtain similarly the spherical groupoid SE = S(E) \ (S(ker r) ∪ S(ker s)) where S (0)(E) = S(F ) ⊂ S(E). The involutive automorphism u (cid:55)→ −u of E leads to a Z/2Z action, by groupoid automorphisms on SE. Since this action is free (and proper!), it follows that the quotient groupoid PE and the crossed product groupoid crossed product SE (cid:111) Z/2Z are Morita equivalent. Thus SE is also amenable. As for the projective case, if (r, s) : E → F × F is onto, the groupoid SE has just one orbit. The +, and therefore the groupoid SE stabilizer of d ∈ S(F ) identifies with the group (ker r ∪ ker s) (cid:111) R∗ is Morita equivalent to the group (ker r ∪ ker s) (cid:111) R∗ +. 26 5.1.4 Bundle groupoids We may of course perform the constructions of section 5.1 (with say K = R) when E is a (real) vector bundle over a space V , F is a subbundle and r, s are bundle maps. We obtain respectively vector bundle groupoids, projective bundle groupoids and spherical bundle groupoids: E, (PE, r, s) and (SE, r, s) which are respectively families of linear, projective and spherical groupoids. Remarks 5.3. a) A vector bundle groupoid is just given by a bundle morphism α = (r − s) : E/F → F . It is isomorphic to the semi direct product F (cid:111)α E/F . b) All the groupoids defined here are amenable, since they are continuous fields of amenable groupoids (cf. [4, Prop. 5.3.4]). The analytic index element indG ∈ KK(C0(A∗G), C∗(G)) of a vector bundle groupoid G is a KK- equivalence. The groupoid G is a vector bundle E over a locally compact space X, G(0) is a vector subbundle F and G is given by a linear bundle map (r − s) : E/F → F . Proposition 5.4 (A Thom-Connes isomorphism). Let E be a vector bundle groupoid. Then C∗(E) is KK-equivalent to C0(E). More precisely, the index indE : KK(C0(A∗E), C∗(E)) is invertible. Proof. Put F = E(0) and H = E/F . Then H acts on C0(F ) and C∗(E) = C0(F ) (cid:111) H. We use the equivariant KK-theory of Le Gall (cf. [30]) KKH (A, B). The thom element of the complex bundle H ⊕ H defines an invertible element tH ∈ KKH (C0(X), C0(H ⊕ H)). We deduce that, for every pair A, B of H algebras, the morphism τC0(H) : KKH (A, B) → KKH (A ⊗C0(X) C0(H), B ⊗C0(X) C0(H)) is an isomorphism. Its inverse is x (cid:55)→ tH ⊗ τC0(H)(x) ⊗ t−1 H . Denote by A0 the C0(X) algebra A endowed with the trivial action of H. We have an isomorphism of H-algebras uA : C0(H) ⊗C(X) A (cid:39) C0(H) ⊗C(X) A0. It follows that the restriction map KKH (A, B) to KKX (A, B) (associated to the groupoid morphism X → H) is an isomorphism - compatible of course with the Kasparov product. Let vA ∈ KKH (A0, A) be the element whose image in KKX (A0, A) is the identity. The descent of jH (vA) ∈ KK(C0(H∗) ⊗C(X) A, A (cid:111) H) is a KK-equivalence. The proposition follows by letting A = C0(F ). 5.1.5 VB groupoids Recall from [46, 31] that a VB groupoid is a groupoid which is a vector bundle over a groupoid G. More precisely: rG,sG/ rE ,sE/ G(0) be a groupoid. A VB groupoid over G is a vector bundle p : Definition 5.5. Let G E → G with a groupoid structure E E(0) such that all the groupoid maps are linear vector bundle morphisms. This means that E(0) ⊂ E is a vector subbundle of the restriction of E to G(0) and that rE, sE, x (cid:55)→ x−1 and the composition are linear bundle maps. We also assume that the bundle maps rE : E → r∗ We will come back to VB groupoids in the appendix. G(E(0)) and sE : E → s∗ G(E(0)) are surjective. 27 / / / / / / 5.2 Normal groupoids, deformation groupoids and blowup groupoids 5.2.1 Definitions Let Γ be a closed Lie subgroupoid of a Lie groupoid G. Using functoriality (cf. Definition 4.8) of the DN C and Blup construction we may construct a normal and a blowup groupoid. a) The normal bundle N G Γ carries a Lie groupoid structure with objects N G(0) Γ(0) . We denote by Γ → Γ is a groupoid morphism and it follows N G that N G Γ ⇒ N G(0) Γ(0) this groupoid. The projection N G Γ is a VB groupoid over Γ. b) The manifold DN C(G, Γ) is naturally a Lie groupoid (unlike what was asserted in remark 3.19 of [25]). Its unit space is DN C(G(0), Γ(0)); its source and range maps are DN C(s) and DN C(r); the space of composable arrows identifies with DN C(G(2), Γ(2)) and its product with DN C(m) where m denotes both products G(2) → G and is Γ(2) → Γ. c) The subset (cid:94)DN C(G, Γ) = Ur ∩ Us of DN C(G, Γ) consisting of elements whose image by it is the 1 × R is an open subgroupoid of DN C(G, Γ): DN C(r) and DN C(s) is not in G(0) restriction of DN C(G, Γ) to the open subspace DN C(G(0), G(0) 1 ) \ G(0) 1 × R. d) The group R∗ acts on DN C(G, Γ) via the gauge action by groupoid morphisms. Its action on (cid:94)DN C(G, Γ) is (locally) proper. Therefore the open subset Blupr,s(G, Γ) = (cid:94)DN C(G, Γ)/R∗ of Blup(G, Γ) inherits a groupoid structure as well: its space of units is Blup(G(0) 1 ); its source and range maps are Blup(s) and Blup(r) and the product is Blup(m). 2 , G(0) e) In the same way, we define the groupoid SBlupr,s(G, Γ). It is the quotient of the restriction +. Similarly DSBlupr,s(G, Γ) will be +. This is the "double" of the Lie groupoid with (cid:94)DN C+(G, Γ) of (cid:94)DN C(G, Γ) to R+ by the action of R∗ the quotient of (cid:94)DN C(G, Γ) by the action of R∗ boundary SBlupr,s(G, Γ). An analogous result about the groupoid structure on Blupr,s(G, Γ) in the case of Γ(0) being a hypersurface of G(0) can be found in [21, Theorem 2.8] (cf. also [20]). 5.2.2 Algebroid and anchor The (total space of the) Lie algebroid AΓ is a closed submanifold (and a subbundle) of AG. The Lie algebroid of DN C(G, Γ) is DN C(AG, AΓ). Its anchor map is DN C((cid:92)G) : DN C(AG, AΓ) → DN C(T G(0), T Γ(0)). The groupoid DN C(G, Γ) is the union of its open subgroupoid G × R∗ with its closed Lie sub- groupoid N G Γ . The algebroid of G × R∗ is AG × R∗ and the anchor is just the map (cid:92)G × id : AG × R∗ → T (G(0) × R∗ +). 5.2.3 Morita equivalence 2 be Lie groupoids, Γ1 ⊂ G1 and Γ2 ⊂ G2 Lie subgroupoids. A Morita Let G1 ⇒ G(0) 1 and G2 ⇒ G(0) equivalence of the pair (Γ1 ⊂ G1) with the pair (Γ2 ⊂ G2) is given by a pair (X ⊂ Y ) where Y is a linking manifold which is a Morita equivalence between G1 and G2 and X ⊂ Y is a submanifold of Y such that the maps r, s and products of Y (see page 9) restrict to a Morita equivalence X between Γ1 and Γ2. Then, by functoriality, • DN C(Y, X) is a Morita equivalence between DN C(G1, Γ1) and DN C(G2, Γ2), • DN C+(Y, X) is a Morita equivalence between DN C+(G1, Γ1) and DN C+(G2, Γ2), 28 • Blupr,s(Y, X) is a Morita equivalence between Blupr,s(G1, Γ1) and Blupr,s(G2, Γ2), • SBlupr,s(Y, X) is a Morita equivalence between SBlupr,s(G1, Γ1) and SBlupr,s(G2, Γ2)... Note that if Y and X are sub-Morita equivalences, the above linking spaces are also sub-Morita equivalences. 5.2.4 Groupoids on manifolds with boundary + Let M be a manifold and V an hypersurface in M and suppose that V cuts M into two manifolds with boundary M = M− ∪ M+ with V = M− ∩ M+. Then by considering a tubular neighborhood of V × {0} identifies with M × R, the quotient (cid:94)DN C(M, V )/R∗ V in M , DN C(M, V ) = M × R∗ ∪ N M identifies with two copies of M and SBlup(M, V ) identifies with the disjoint union M−(cid:116)M+. Under V \ V ×{0} pointing this last identification, the class under the gauge action of a normal vector in N M Let Mb be manifold with boundary V . A piece of Lie groupoid is the restriction G = (cid:101)GMb in the direction of M+ is an element of V ⊂ M+. a Lie groupoid (cid:101)G ⇒ M where M is a neighborhood of Mb and G is a groupoid without boundary. Note that when the boundary V is transverse to the groupoid (cid:101)G, G is in fact a manifold with corners. Let then Γ ⇒ V be a Lie subgroupoid of (cid:101)G. We may construct SBlupr,s((cid:101)G, Γ) and consider its restriction to the open subset Mb of SBlup(M, V ). which Mb is saturated. Indeed SBlupr,s(G, Γ) is an open subgroupoid of SBlupr,s((cid:101)G, Γ) ⇒ Mb(cid:116)M− which is a piece of the Lie groupoid (cid:94)DN C((cid:101)G, Γ)/R∗ + (cid:39) M (cid:116) M . We may then With the above notation, since V is of codimention 1 in M , SBlup(M, V ) = Mb (cid:116) M− where M− = M \ M is the complement in M of the interior M = Mb \ V of Mb in M . We thus obtain a longitudinally smooth groupoid that will be denoted SBlupr,s(G, Γ). Note that the groupoid SBlupr,s(G, Γ) ⇒ Mb is the restriction to Mb of a Lie groupoid G ⇒ M for let G be the restriction of (cid:94)DN C(M, V )/R∗ In this way, we may treat by induction a finite number of boundary components i.e. a groupoid on a manifold with corners. + to one of the copies of M . to Mb of Mb ⇒ (cid:94)DN C(M, V )/R∗ + Remarks 5.6. a) If M is a manifold with boundary V and G = M×M is the pair groupoid, then SBlupr,s(G, V ) is in fact the groupoid associated with the 0 calculus in the sense of Mazzeo (cf. [32, 37, 34]), i.e. the canonical pseudodifferential calculs associated with SBlupr,s(G, V ) is the Mazzeo-Melrose's 0-calculus. Indeed, the sections of the algebroid of SBlupr,s(G, V ) are exactly the vector fields of M vanishing at the boundary V , i.e. those generating the 0-calculus. b) In a recent paper [44], an alternative description of SBlupr,s(G, V ) is given under the name of edge modification for G along the "AG-tame manifold" V , thus in particular V is transverse to G. This is essentially the gluing construction described in 5.3.4 below. 5.3 Examples of normal groupoids, deformation groupoids and blowup groupoids We examine some particular cases of inclusions of groupoids G1 ⊂ G2. The various constructions of deformation to the normal cone and blow-up allow us to recover many well known groupoids. As already noted in the introduction, our constructions immediately extend to the case where we restrict to a closed saturated subset of a smooth groupoid, in particular for manifolds with corners. 5.3.1 Inclusion F ⊂ E of vector spaces Let E be a real vector space - considered as a group - and F a vector subspace of E. The inclusion of groups F → E gives rise to a groupoid DN C(E, F ). Using any supplementary subspace of F in 29 E, we may identify the groupoid DN C(E, F ) with E × R ⇒ R. Its C∗-algebra identifies then with C0(E∗ × R). More generally, if F is a vector-subbundle of a vector bundle E over a manifold M (considered as a family of groups indexed by M ), then the groupoid DN C(E, F ) ⇒ M × R identifies with E × R and its C∗-algebra is C0(E∗ × R). Let pE : E → M be a vector bundle over a manifold M and let V be a submanifold of M . Let pF : F → V be a subbundle of the restriction of E to V . We use a tubular construction and find an open subset U of M which is a vector bundle π : Q → V . Using π, we may extend F to a subbundle FU of the restriction to F on U . Using that, we may identify DN C(E, F ) with the open subset E × R∗ ∪ p−1 E (U ) × R of E × R. Its C∗-algebra identifies then with C0(E∗ × R∗ ∪ p−1 E∗(U ) × R). 5.3.2 Inclusion G(0) ⊂ G: adiabatic groupoid The deformation to the normal cone DN C(G, G(0)) is the adiabatic groupoid Gad ([40, 45]), it is obtained by using the deformation to the normal cone construction for the inclusion of G(0) as a Lie subgroupoid of G. The normal bundle N G G(0) is the total space of the Lie algebroid A(G) of G. Note that its groupoid structure coincides with its vector bundle structure. Thus, DN C(G, G(0)) = G × R∗ ∪ A(G) × {0} ⇒ G(0) × R . The particular case where G is the pair groupoid M × M is the original construction of the "tangent groupoid" of Alain Connes ([12]). Note that Blup(G(0), G(0)) = ∅ = Blupr,s(G, G(0)). 5.3.3 Gauge adiabatic groupoid Start with a Lie groupoid G ⇒ V . Let G × (R × R) First notice that since V × {0} is a codimension 1 submanifold in V × R, SBlup(V × R, V × {0}) is canonically isomorphic to V × (R− (cid:116) R+). Then SBlupr,s(G × (R × R), V × {(0, 0)})V ×R+ V ×R+ is the semi-direct product groupoid Gad(V × R+) (cid:111) R∗ +: r,s⇒ V × R be the product groupoid of G with the pair groupoid over R. SBlupr,s(G × (R × R), V × {(0, 0)})V ×R+ V ×R+ = Gad(V × R+) (cid:111) R∗ ⇒ V × R+ . is the gauge adiabatic groupoid used in In other words, SBlupr,s(G × (R × R), V × {(0, 0)})V ×R+ V ×R+ [17]. Indeed, as G× (R× R) is a vector bundle over G, DN C(G× (R× R), V ×{(0, 0)}) (cid:39) DN C(G, V )× R2 (remark 4.1.d). Under this identification, the gauge action of R∗ is given by λ.(w, t, t(cid:48)) = (λ.w, λ−1t, λ−1t(cid:48)). The maps DN C(s) and DN C(r) are respectively (w, t, t(cid:48)) (cid:55)→ (DN C(s)(w), t(cid:48)) and (w, t, t(cid:48)) (cid:55)→ (DN C(r)(w), t). It follows that SBlupr,s(G × (R × R), V × {(0, 0)}) is the quotient by the diagonal action of R∗ According to the description of the groupoid of a group action on a groupoid given in section 2.3 it is + ×{−1, +1}2 where {−1, +1}2 is the pair groupoid over {−1, +1}. isomorphic to DN C(G, V )+ (cid:111) R∗ + of the open subset DN C(G, V ) × (R∗)2 of DN C+(G, V ) × R2. 5.3.4 Inclusion of a transverse submanifold of the unit space Let G be a Lie groupoid with set of objects M = G(0) and let V be a submanifold of M . We now study the special case of normal and blowup groupoids DN C(G, V ) and Blupr,s(G, V ) (as well as SBlupr,s(G, V )) associated to the groupoid morphism V → G. Put M = M \ V . Let N = N G V be the normal bundles. We identify N(cid:48) with a subbundle of N by means of the inclusion M ⊂ G. The submersions r, s : G → M give rise to bundle morphisms rN , sN : N → N(cid:48) that are sections of the inclusion N(cid:48) → N . By construction, V and N(cid:48) = N M 30 M M using remark 5.1.a), the groupoid DN C(G, V ) is the union of G × R∗ with the family of linear with the family (PN, rN , sN ) groupoids NrN ,sN (N ). It follows that Blupr,s(G, V ) is the union of G of projective groupoids. If V is transverse to G, the bundle map rN − sN : N = N G that • LrN ,sN (N ) identifies with the pull-back groupoid(cid:0)A(GV V )(cid:1)q V ) (cid:111) R∗(cid:1)ρ • (PN, rN , sN ) with the pull-back groupoid(cid:0)A(GV q where q : N(cid:48) → V is the projection, ρ where ρ : P(N(cid:48)) → V is the projec- • (SN, rN , sN ) with the pull-back groupoid(cid:0)A(GV (cid:1)p p where p : S(N(cid:48)) → V is the projec- V → N(cid:48) = N M is surjective; it follows V ) (cid:111) R∗ tion, + V tion. V → V . open subset(cid:0)(GV M\V . Upon arguing locally, we can assume that V is compact. Let us give a local description of these groupoids in the neighborhood of the transverse submanifold V . Put G = GM\V By Remark 2.3, V admits a tubular neighborhood W (cid:39) N M by the retraction q : W → V . The normal groupoid DN C(GW adiabatic deformation DN C(GV The (spherical) blowup groupoid SBlupr,s(GW R∗ V , V ) (cid:111) R∗ +)p In order to get SBlupr,s(G, V ), we then may glue (DN C+(GV W , V ) identifies with the pull back groupoid (DN C+(GV V → V . p with G in their common W , V ) identifies with the pull back groupoid (DN C(GV V , V ) = (GV p of the gauge adiabatic deformation DN C+(GV V )ga by the map p : SN M W is the pull back of GV V V )ad by the map q : N M V such that GW (cid:1)W\V W\V (cid:39) GW\V W\V . V ⊂ G for a transverse hypersurface V of G: b-groupoid + = (GV V , V ) (cid:111) R∗ +)p 5.3.5 Inclusion GV If V is a hypersurface of M , the blowup Blup(M×M, V ×V ) is just the construction of Melrose of the b-space. Its open subspace Blupr,s(M ×M, V ×V ) is the associated groupoid of Monthubert [38, 39]. Moreover, if G is a groupoid on M and V is transverse to G we can form the restriction groupoid V ⊂ G which is a submanifold of G. The corresponding blow up construction Blupr,s(G, GV GV V ) identifies with the fibered product Blupr,s(M × M, V × V ) ×M×M G (cf. remark 4.1.e). Iterating (at least locally) this construction, we obtain the b-groupoid of Monthubert for manifolds with corners - cf. [38, 39]. V )q q V , V ))q q of the V , V )(cid:111) V ) is a kind of cylindric deformation groupoid which is obtained by Remark 5.7. The groupoid Blupr,s(G, V ) corresponds to inflating all the distances when getting close to V . The groupoid Blupr,s(G, GV pushing the boundary V at infinity but keeping the distances along V constant. Remark 5.8. Intermediate examples between these two are given by a subgroupoid Γ ⇒ V of GV V . In the case where G = M × M , such a groupoid Γ is nothing else than the holonomy groupoid Hol(V,F) of a regular foliation F of V (with trivial holonomy groups). The groupoid SBlupr,s(M × M, Hol(V,F)) is a holonomy groupoid of a singular foliation of M : the sections of its algebroid. Its leaves are M \ V and the leaves of (V,F). The corresponding calculus, when M is a manifold with a boundary V is Rochon's generalization ([47]) of the φ calculus of Mazzeo and Melrose ([33]). Iterating (at least locally) this construction, we obtain the holonomy groupoid associated to a stratified space in [16]. 31 5.3.6 Inclusion GV V ⊂ G for a saturated submanifold V of G: shriek map for immersion V = GV = GV . V acts on the normal bundle N G GV V Suppose now that V is saturated thus GV V ) ⇒ In such a situation the groupoid GV V V → G. This construction DN C(G(0), V ) coincides with the normal groupoid of the immersion ϕ : GV was defined in the case of foliation groupoids in [25, section 3] and was used in order to define ϕ! as associated KK-element. 5.3.7 Inclusion G1 ⊂ G2 with G(0) This is the case for the tangent and adiabatic groupoid discussed above. Two other kinds of this situation3 can be encountered in the literature: ) and DN C(G, GV = r∗(N G(0) 1 = G(0) 2 a) The case of a subfoliation F1 of a foliation F2 on a manifold M : shriek map for submersion. As pointed out in remark 3.19 of [25] the corresponding deformation groupoid DN C(G2, G1) gives an alternative construction of the element ϕ! where ϕ : M/F1 → M/F2 is a submersion of leaf spaces. b) The case of a subgroup of a Lie group. • If K is a maximal compact subgroup of a reductive Lie group G, the connecting map associated to the exact sequence of DN C(G, K) is the Dirac extension mapping the twisted K-theory of K to the K-theory of C∗ • In the case where Γ is a dense (non amenable) countable subgroup of a compact Lie group K, the groupoid DN C(K, Γ) was used in [23] in order to produce a Hausdorff groupoid for which the Baum-Connes map is not injective. r (G) (see [22]). 5.3.8 Wrong way functoriality Let f : G1 → G2 be a morphism of Lie groupoids. If f is an (injective) immersion the construction of DN C+(G2, G1) gives rise to a short exact sequence 0 −→ C∗(G2 × R∗ +) −→ C∗(DN C+(G2, G1)) −→ C∗(N G2 G1 ) −→ 0. and consequently to a connecting map from the K-theory of the C∗-algebra of the groupoid N G2 , which is a VB groupoid over G1, to the K-theory of C∗(G2). This wrong way functoriality map will be discussed in the next section. More generally let G = G(0) that the map x (cid:55)→ (r(x), f (x), s(x)) is an immersion from G1 → G. The above construction gives a map from K∗(C∗(N G K∗(C∗(G2)) since the groupoids G2 and G are canonically equivalent. 6 The C∗-algebra of a deformation and of a blowup groupoid, full 1 . Assume )) to K∗(C∗(G)) which is isomorphic to 1 be the product of G2 by the pair groupoid of G(0) 1 × G2 × G(0) G1 G1 symbol and index map Let G ⇒ M be a Lie groupoid and Γ ⇒ V a Lie subgroupoid of G. The groupoids DN C+(G, Γ) V × {0} and SN M and SBlupr,s(G, Γ) that we constructed admit the closed saturated subsets N M respectively. In order to shorten the notation we put M = M \ V and the corresponding full symbol algebras V • ΣDN C+(G, Γ) = ΣM×R∗ 3Note that in this case Blup(G(0) +(DN C+(G, Γ)); 2 , G(0) 1 ) = ∅, whence Blupr,s(G2, G1) = ∅. 32 • Σ(cid:94)DN C+ (G, Γ) = Σ • ΣSBlup(G, Γ) = Σ M×R∗ +((cid:94)DN C+(G, Γ)); M (SBlupr,s(G, Γ)). They give rise to the exact sequences 0 −→ C∗(G M M ) −→ C∗(SBlupr,s(G, Γ)) −→ C∗(SN G Γ ) −→ 0 and 0 −→ C∗(G × R∗ +) −→ C∗(DN C+(G, Γ)) −→ C∗(N G Γ ) −→ 0 of groupoid C∗-algebras as well as index type ones 0 −→ C∗(G M M ) −→ Ψ∗(SBlupr,s(G, Γ)) −→ ΣSBlup(G, Γ) −→ 0 and 0 −→ C∗(G × R∗ +) −→ Ψ∗(DN C+(G, Γ)) −→ ΣDN C+(G, Γ) −→ 0 E∂ SBlup E∂ DN C+ SBlup E(cid:102)ind E(cid:102)ind DN C+ We will compare the exact sequences given by DN C and by SBlup. If V is AG-small (see notation 6.5 below), we will show that, in a sense, DN C and SBlup give rise to equivalent exact sequences - both for the "connecting" ones and for the "index" ones. We will then compare these elements with a coboundary construction. We will compute these exact sequences when Γ = V ⊂ M . Finally, we will study a refinement of these constructions using relative K-theory. 6.1 "DNC" versus "Blup" Let Γ be a submanifold and a subgroupoid of a Lie-groupoid G. We will further assume that the groupoid Γ is amenable. Put M = G(0) and V = Γ(0). Put also M = M \ V and let N G Γ be the restriction of the groupoid N G V \ V of its unit space N M V . Γ to the open subset N M V = N M 6.1.1 The connecting element As the groupoid Γ is amenable we have exact sequences both for the reduced and for the maximal C∗-algebras: 0 −→ C∗(G M M ) −→ C∗(SBlupr,s(G, Γ)) −→ C∗(SN G Γ ) −→ 0 E∂ SBlup and 0 −→ C∗(G × R∗ +) −→ C∗(DN C+(G, Γ)) −→ C∗(N G Γ ) −→ 0 M SBlup = ∂ Γ ), C∗(G × R∗ DN C+ By amenability, these exact sequences admit completely positive cross sections and therefore de- M×R∗ DN C+(G,Γ) ∈ fine elements ∂G,Γ KK1(C∗(N G With the notation of section 5.2, write DN C+ for DN C restricted to R+ and (cid:94)DN C+ for (cid:94)DN C restricted to R+. By section 3.3.3, we have a diagram where the vertical arrows are KK1-equivalences and the squares commute in KK-theory. SBlupr,s(G,Γ) ∈ KK1(C∗( N G +)). +), C∗(G )) and ∂G,Γ Γ /R∗ DN C+ = ∂ M M + E∂ 0 0 ) / C∗(G M M β(cid:48) × R∗ +) / C∗(G M M C∗(SBlupr,s(G, Γ)) β / C∗((cid:94)DN C+(G, Γ)) C∗( N G Γ /R∗ +) β(cid:48)(cid:48) / C∗( N G Γ ) 0 / 0 E∂ SBlup E∂ (cid:94)DN C+ Denote by ∂G,Γ according to proposition 3.12: (cid:94)DN C+ = ∂ M×R∗ (cid:94)DN C+(G,Γ) + the connecting element associated to E∂ (cid:94)DN C+ . We thus have, 33 / / / / / / / / / / / Fact 6.1. ∂G,Γ SBlup ⊗ β(cid:48) = −β(cid:48)(cid:48) ⊗ ∂G,Γ (cid:94)DN C+ ∈ KK1(C∗(SN G Γ ), C∗(G × R∗ +)). M M We also have a commutative diagram where the vertical maps are inclusions: 0 0 / C∗(G M M × R∗ +) j(cid:48) / C∗(G × R∗ +) C∗((cid:94)DN C+(G, Γ)) j / C∗(DN C+(G, Γ)) C∗( N G Γ ) j(cid:48)(cid:48) / C∗(N G Γ ) 0 / 0 (6.1) We thus find Fact 6.2. (j(cid:48)(cid:48))∗(∂G,Γ DN C+ ) = j(cid:48) ∗(∂G,Γ (cid:94)DN C+ ) ∈ KK1(C∗( N G Γ ), C∗(G × R∗ +)). 6.1.2 The full symbol index We now compare the elements (cid:102)ind and (cid:102)ind DN C+ = (cid:102)ind M×R∗ f ull G,Γ + split exact sequences: SBlup = (cid:102)ind G,Γ M f ull(SBlupr,s(G, Γ)) ∈ KK1(ΣSBlup(G, Γ), C∗(G M M )) (DN C+(G, Γ)) ∈ KK1(ΣDN C+(G, Γ), C∗(G × R∗ +)) defined by the semi- SBlup E(cid:102)ind E(cid:102)ind 0 −→ C∗(G M M ) −→ Ψ∗(SBlupr,s(G, Γ)) −→ ΣSBlup(G, Γ) −→ 0 and 0 −→ C∗(G × R∗ +) −→ Ψ∗(DN C+(G, Γ)) −→ ΣDN C+(G, Γ) −→ 0 DN C+ By prop. 3.13, we have a diagram where the vertical arrows are KK1-equivalences and the squares commute in KK-theory. 0 0 ) / C∗(G M M β(cid:48) × R∗ +) / C∗(G M M Ψ∗(SBlupr,s(G, Γ)) ΣSBlup(G, Γ) βΨ βΣ / Ψ∗((cid:94)DN C+(G, Γ)) / Σ(cid:94)DN C+ (G, Γ) 0 / 0 We let (cid:102)ind Fact 6.3. (cid:102)ind M×R∗ f ull = (cid:102)ind SBlup ⊗ β(cid:48) = −βΣ ⊗ (cid:102)ind G,Γ + G,Γ (cid:94)DN C+ ((cid:94)DN C+(G, Γ)) ∈ KK1(Σ(cid:94)DN C+ G,Γ (cid:94)DN C+ ∈ KK1(ΣSBlup(G, Γ), C∗(G × R∗ (G, Γ), C∗(G M M × R∗ M M +)). +)). We thus have: We also have a commutative diagram where the vertical maps are inclusions: 0 0 / C∗(G M M × R∗ +) j(cid:48) / C∗(G × R∗ +) Ψ∗((cid:94)DN C+(G, Γ)) Σ(cid:94)DN C+ (G, Γ) jΨ jΣ / Ψ∗(DN C+(G, Γ)) / ΣDN C+(G, Γ) 0 / 0 (6.2) We thus find: Fact 6.4. j∗ Σ((cid:102)ind ∗((cid:102)ind G,Γ DN C+) = j(cid:48) G,Γ (cid:94)DN C+ ) ∈ KK1(Σ(cid:94)DN C+ (G, Γ), C∗(G × R∗ +)). 34 / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / 6.1.3 When V is AG-small and ∂G,Γ DN C+ SBlup. V )x is not the zero map DN C+ (cid:92)x−→ if the Lebesgue measure (in the manifold Gx) of Gx SBlup correspond to each other under these isomorphisms. V is 0 for ) (cid:44)→ C∗(G) is an isomorphism. If V is small in each G orbit, i.e. every x, it follows from prop. 6.6 below that the inclusion i : C∗(G M M Also, if M meets all the orbits of G, the inclusion i is a Morita equivalence. In these cases ∂G,Γ determines ∂G,Γ Definition 6.5. We will say that V is AG-small if for every x ∈ V , the composition AGx TxM −→ (N M If V is AG-small, then the orbits of the groupoid N G Γ are never contained in the 0 section, i.e. V , and in fact the set V × {0} is small in every orbit of the groupoid they meet the open subset N M DN C(G, Γ). It follows that the map j is an isomorphism - as well of course as j(cid:48) and j(cid:48)(cid:48) of diagram (6.1). In that case, ∂G,Γ Proposition 6.6. (cf. [24, 19]) Let G ⇒ Y be a Lie groupoid and let X ⊂ Y be a (locally closed) submanifold. Assume that, for every x ∈ X, the composition AGx X )x is not the zero map. Then the inclusion C∗(GY \X Proof. For every x ∈ V , we can find a neighborhood U of x ∈ M , a section X of AG such that, for every y ∈ U , (cid:92)y(X(y)) (cid:54)= 0 and, if y ∈ U ∩ V , (cid:92)y(X(y)) (cid:54)∈ Ty(V ). Denote by F the foliation of U associated with the vector field X. It follows from [24] that C0(U \ V )C∗(U,F) = C∗(U,F); as C∗(U,F) acts in a non degenerate way on the Hilbert-C∗(G) module C∗(GU ), we deduce that C0(U \ V )C∗(GU ) = C∗(GU ). We conclude using a partition of the identity argument that Cc(M \ V )C∗(G) = Cc(M )C∗(G), whence C0(M \ V )C∗(G) = C0(M )C∗(G) = C∗(G). Proposition 6.7. We assume that Γ is amenable and that V is AG-small. Then, the inclusions jΣ : Σ(cid:94)DN C+ and jsymb : C0(S∗A((cid:94)DN C+(G, Γ))) → C0(S∗A(DN C+(G, Γ))) are KK-equivalences. Proof. We have a diagram (G, Γ) → ΣDN C+(G, Γ), jΨ : Ψ∗((cid:94)DN C+(G, Γ)) → Ψ∗(DN C+(G, Γ)) Y \X ) → C∗(G) is an isomorphism. (cid:92)x−→ TxY −→ (N Y 0 0 0 / C∗((cid:94)DN C+(G, Γ)) Ψ∗((cid:94)DN C+(G, Γ)) C0(S∗A((cid:94)DN C+(G, Γ)) j jΨ jsymb / C∗(DN C+(G, Γ)) / Ψ∗(DN C+(G, Γ)) / C0(S∗A(DN C+(G, Γ)) 0 0 As j is an equality, we find an exact sequence C0(S∗AGV × R+) 0 As j(cid:48) : C∗(G M M sequence 0 −→ Ψ∗((cid:94)DN C+(G, Γ)) +) → C∗(G × R∗ × R∗ 0 −→ Σ(cid:94)DN C+ (G, Γ)) jΨ−→ Ψ∗(DN C+(G, Γ)) −→ C0(S∗AGV × R+) −→ 0. +) is also an equality, we find (using diagram (6.2)) an exact jΣ−→ ΣDN C+(G, V ) −→ C0(S∗AGV × R+) −→ 0. As the algebra C0(S∗AGV × R+) is contractible, we deduce that jsymb and then jΨ and jΣ are KK-equivalences. 35 / / / / / / / / / / / / As a summary of these considerations, we find: Theorem 6.8. Let G ⇒ M be a Lie groupoid and Γ ⇒ V a Lie subgroupoid of G. Assume that Γ ∗ (β(cid:48)(cid:48)) ∈ is amenable and put M = M \ V . Let i : C∗(G M M KK1(C∗(SN G Γ )) and βΣ = (jΣ)∗(βΣ) ∈ KK1(ΣSBlup(G, Γ), ΣDN C+(G, V )). ) → C∗(G) be the inclusion. Put β(cid:48)(cid:48) = j(cid:48)(cid:48) Γ ), C∗(N G a) We have equalities • ∂G,Γ SBlup ⊗ [i] = β(cid:48)(cid:48) ⊗ ∂G,Γ • (cid:102)ind SBlup ⊗ [i] = βΣ ⊗ (cid:102)ind G,Γ DN C+ ∈ KK1(C∗(SN G DN C+ ∈ KK1(C∗(ΣSBlup(G, Γ), C∗(G)) Γ ), C∗(G)) and G,Γ b) If V is AG-small, then i is an isomorphism and the elements β(cid:48)(cid:48) and βΣ are invertible. (cid:3) 6.2 The KK-element associated with DNC The connecting element ∂G,Γ DN C(G, Γ) to [0, 1], i.e. G = DN C[0,1](G, Γ) = N G sequence: DN C+ can be expressed in the following way: let G be the restriction of Γ × {0} ∪ G × (0, 1]. We have a semi-split exact 0 → C∗(G × (0, 1]) → C∗(G) ev0−→ C∗(N G Γ ) → 0 . As C∗(G × (0, 1]) is contractible, ev0 is a KK-equivalence. Let ev1 : C∗(G) → C∗(G) be evaluation at 1 and let δG +)) be the Bott element. We find Γ ), C∗(G)). Let [Bott] ∈ KK1(C, C0(R∗ Γ = [ev0]−1 ⊗ [ev1] ∈ KK(C∗(N G Fact 6.9. ∂G,Γ DN C+ = δG Γ ⊗ C [Bott]. Consider now the groupoid G[0,1] ad . It is a family of groupoids indexed by [0, 1] × [0, 1]: • its restriction to {s} × [0, 1] for s (cid:54)= 0 is G[0,1] ad ; • its restriction to {0} × [0, 1] is (N G • its restriction to [0, 1] × {s} for s (cid:54)= 0 is G = DN C[0,1](G, Γ); • its restriction to [0, 1] × {0} is the algebroid AG = DN C[0,1](AG, AΓ) of G. Γ )[0,1] ad ; For every locally closed subset X ⊂ [0, 1] × [0, 1], denote by GX For every closed subset X ⊂ [0, 1] × [0, 1], denote by qX : C∗(G[0,1] We thus have the following commutative diagram: ad the restriction of G[0,1] ad ) → C∗(GX ad to X. ad) the restriction map. C∗(N G Γ ) indN G Γ C∗((N G Γ )[0,1] ad ) C∗(N AG AΓ ) C∗(DN C[0,1](AG, AΓ)) / C0(A∗G) C∗(G) δG Γ DN C[0,1](G, Γ) q(0,1) iSSSSSSSSSSSSSSS ukkkkkkkkkkkkkkk q{0}×[0,1] q(0,0) q[0,1]×{1} C∗(G[0,1] ad ) q[0,1]×{0} q(1,1) q{1}×[0,1] 6llllllllllllll (RRRRRRRRRRRRRRR q(1,0) C∗(G[0,1] ad ) indG δAG AΓ 36 , , / / o o O O u i 6 ( / / o o O O O O ; ; 2 2 o o / c c ad ) are null It follows that q{0}×[0,1], q[0,1]×{0} and q{(0,0)} are KK- For every locally closed subset T ⊂ [0, 1], the C∗-algebras C∗(G(0,1]×T homotopic as well as C∗(G[0,1]2\{0,0)} equivalences. Now [q(0,0)]−1 ⊗ [q(0,1)] = indN G In the same way, [q(0,0)]−1 ⊗ [q(1,0)] = δAG Finally, it follows from example 5.3.1 that δAG C0(A∗G) corresponding to an inclusion of A∗(N G We thus have established: ⊗ δG Γ . AΓ and it follows that [q(0,0)]−1 ⊗ [q(1,1)] = δAG AΓ ⊗ indG. AΓ is associated with a morphism ϕ : C0(A∗(N G Γ ) in A∗G as a tubular neighborhood. and it follows that [q(0,0)]−1 ⊗ [q(1,1)] = indN G ) and C∗(GT×(0,1] V )) (cid:44)→ ). ad ad Γ Γ Fact 6.10. indN G Γ ⊗ δG Γ = [ϕ] ⊗ indG. 6.3 The case of a submanifold of the space of units Let G be a Lie groupoid with objects M and let Γ = V ⊂ M be a closed submanifold of M . In this section, we push further the computations the connecting maps and indices i.e. the connecting maps of the exact sequences E∂ SBlup and E(cid:102)ind DN C+, E(cid:102)ind SBlup, E∂ DN C+. 6.3.1 Connecting map and index map From propositions 3.3, 3.6, 3.7 and fact 6.10, we find Proposition 6.11. a) The index element indN G V )) is invertible. V ×{0}(DN C+(G, V )) (cid:44)→ ΣDN C+(G, V ) is invertible in KK-theory. b) The inclusion j : ΣN M c) The C∗-algebra ΣDN C+(G, V ) is naturally KK1-equivalent with the mapping cone of the map V ∈ KK(C0(A∗N G V ), C∗(N G (cid:40) χ : C0(A∗G × R∗ +) → C0(DN C+(M, V )) defined by χ(f )(x) = d) The connecting element ∂G,V DN C+ V = ind−1N G δG neighborhood construction. V ⊗ [ϕ] ⊗ indG where ϕ : C0(A∗N G ∈ KK1(C∗(N G f (x, 0) 0 if x ∈ M × R∗ if x ∈ N M V . +)) = KK(C∗(N G V ), C∗(G)) is V ) → C0(A∗G) is the inclusion using the tubular V ), C∗(G × R∗ + e) Under the KK1 equivalence of c), the full index element (cid:102)ind DN C+ ∈ KK1(ΣDN C+(G, V ), C∗(G × R∗ G,V +)) = KK1(Cχ, C∗(G)) is q∗([Bott] ⊗ C indG) where q : Cχ → C0(A∗G × R∗ +) is evaluation at 0. (cid:3) The element [χ] ∈ KK(C0(A∗G×R∗ +), C0(DN C+(M, V ))) is the Kasparov product of the "Euler ele- +), C0(M× ment" of the bundle A∗G which is the class in KK(C0(A∗G), C0(M )) = KK(C0(A∗G×R∗ +) → C0(DN C+(M, V )). It follows that +)) of the map x (cid:55)→ (x, 0) with the inclusion C0(M × R∗ R∗ [χ] is often the zero element of KK(C0(A∗G × R∗ +), C0(DN C+(M, V ))). In particular, this is the case when the Euler class of the bundle A∗G vanishes. In that case, the algebra ΣDN C+(G, V ) is KK-equivalent to C0(A∗G) ⊕ C0(DN C+(M, V )). If V is AG small, then, by theorem 6.8, ∂G,V sition 6.11. Remark 6.12. Let Mb be a manifold with boundary and V = ∂Mb. Put M = Mb \ V . Let G be a piece of Lie groupoid on Mb in the sense of section 5.2.4. Thus G is the restriction of a Lie groupoid where M = Mb ∪ M− and M ∩ M− = V , and we let SBlupr,s(G, V ) ⇒ Mb be the restriction of (cid:101)G ⇒ M , where M is a neighborhood of Mb. Recall that in this situation, SBlup(M, V ) = Mb(cid:116) M−, SBlupr,s((cid:101)G, V ) to Mb. G,V SBlup are immediately deduced from propo- SBlup and (cid:102)ind 37 Let us denote by N G the differential of the source and range maps of (cid:101)G are non vanishing elements of N M V made of (normal) tangent vectors whose image under V pointing in V the open subset of N(cid:101)G the direction of Mb. The groupoid SBlupr,s(G, V ) is the union N G We have exact sequences V /R∗ + ∪ G M M . 0 → C∗(G M M 0 → C∗(G M M +) → 0 ) → C∗(SBlupr,s(G, V )) → C∗( N G ) → Ψ∗(SBlupr,s(G, V )) → ΣSBlup(G, V ) → 0. V /R∗ As V is of codimension 1, we find that V is A(cid:101)G-small if and only if it is transverse to (cid:101)G. In that case, +) and of ΣSBlup(G, V ) and the KK-class Proposition 6.11 computes the KK-theory of C∗( N G of the connecting maps of these exact sequences. In particular, we obtain a six term exact sequence V /R∗ K0(C(Mb)) / K0(ΣSBlup(G, V )) χ K0(C0(A∗G M M )) K1(ΣSBlup(G, V )) K0(C(Mb)) / K1(C0(A∗G M M )) χ This holds, in particular, if G = Mb× Mb since the boundary V = ∂Mb is transverse to (cid:101)G = M × M . and the index map K∗(ΣSBlup(G, V )) → K∗+1(G K∗+1(C0(A∗G M M Note that in that case, χ = 0 (in KK(C0(T ∗ M ), C0(Mb))) so that we obtain a (noncanonically) split short exact sequence: ) is the composition of K∗(ΣSBlup(G, V )) → )) with the index map of the groupoid G M M M M . 0 / K∗(C0(Mb)) / K∗(ΣSBlup(G, V )) / K∗+1(C0(A∗G M M )) / 0. 6.3.2 The index map via relative K-theory It follows now from prop. 3.8: Proposition 6.13. Let ψDN C : C0(DN C+(M, V )) → Ψ∗(DN C+(G, V )) be the inclusion map which associates to a (smooth) function f the order 0 (pseudo)differential operator multiplication by f and σf ull : Ψ∗(DN C+(G, V )) → ΣDN C+(G, V ) the full symbol map. Put µDN C = σf ull◦ψDN C. Then the relative K-group K∗(µDN C) is naturally isomorphic to K∗+1(C0(A∗G)). Under this isomorphism, indrel : K∗(µDN C) → K∗(C∗(G × R∗ (cid:3) Let us say also just a few words on the relative index map for SBlupr,s(G, V ), i.e. for the map µSBlup : C0(SBlup+(M, V )) → ΣSBlup(G, V ) which is the composition of the inclusion ψSBlup : C0(SBlup(M, V ) → Ψ∗(SBlupr,s(G, V )) with the full index map σf ull : Ψ∗(SBlupr,s(G, V )) → )). ΣSBlup(G, V )), and the corresponding relative index map indrel Equivalently we wish to compute the relative index map indrel : K∗(µ(cid:94)DN C where µ(cid:94)DN C We have a diagram M M M M (G, V ). We restrict to the case when V is AG small. : K∗(µSBlup) → K∗(C∗(G ) → K∗+1(C∗(G +)) = K∗+1(C∗(G)) identifies with indG. : C0((cid:94)DN C+(M, V )) → Σ(cid:94)DN C+ )), 0 / C0((cid:94)DN C+(M, V )) / C0(DN C+(M, V )) / C0(V × R+) / 0 and it follows that the inclusion C0((cid:94)DN C+(M, V )) → C0(DN C+(M, V )) is a KK-equivalence. Since the inclusions Ψ∗((cid:94)DN C+(G, V )) → Ψ∗(DN C+(G, V )) and Σ(cid:94)DN C+ are also KK-equivalences (prop. 6.7), it follows that the inclusion Cµ(cid:94)DN C equivalence - and therefore the relative K-groups K∗(µ(cid:94)DN C phic. Using this, together with the Connes-Thom isomorphism, we deduce: (G, V ) → ΣDN C+(G, V ) → CµDN C is a KK- ) and K∗(µDN C) are naturally isomor- 38 / / O O o o o o / / / / / / / / Corollary 6.14. We assume that V is AG small a) The relative K-group K∗(µ(cid:94)DN C morphism, indrel : K∗(µ(cid:94)DN C ) is naturally isomorphic to K∗+1(C0(A∗G)). Under this iso- ) → K∗(C∗(G × R∗ +)) = K∗+1(C∗(G)) identifies with indG. b) The relative K-group K∗(µSBlup) is naturally isomorphic to K∗(C0(A∗G)). Under this iso- (cid:3) morphism, indrel : K∗(µSBlup) → K∗(C∗(G)) identifies with indG. 7 A Boutet de Monvel type calculus From now on, we suppose that V is a transverse submanifold of M with respect to the Lie groupoid G ⇒ M . In particular V is AG-small - of course, we assume that (in every connected component of V ), the dimension of V is strictly smaller than the dimension of M . 7.1 The Poisson-trace bimodule As V is transverse to G, the groupoid GV adiabatic groupoid" (GV In [17], we constructed a bi-module relating the C∗-algebra of the groupoid (GV algebra of pseudodifferential operators of GV V . V is a Lie groupoid, so that we can construct its "gauge V )ga and the C∗- V )ga (see section 5.3.3). In this section, • We first show that the groupoid (GV V )ga, is (sub-) Morita equivalent to SBlupr,s(G, V ) (cf. also section 5.3.4 for a local construction). • Composing the resulting bimodules, we obtain the "Poisson-trace" bimodule relating C∗(SBlupr,s(G, V )) and Ψ∗(GV V ). 7.1.1 The SBlupr,s(G, V ) − (GV Define the map j : M (cid:116) (V × R) → M by letting j0 : M → M be the identity and j1 : V × R → M the composition of the projection V × R → V with the inclusion. Let G = Gj j. As V is assumed to be transverse, the map j is also transverse, and therefore G is a Lie groupoid. It is the union of four clopen subsets V )ga-bimodule E (G, V ) • the groupoids Gj0 • the linking spaces Gj0 = G = GM = GM j0 j1 = GV M and Gj1 j1 V ×R = GV × R and Gj1 j0 V × (R × R) = GV ×R V ×R . = GV ×R M = GV × R. By functoriality, we obtain a sub-Morita equivalence of SBlupr,s(GV (see section 5.2.3). Let us describe this sub-Morita equivalence in a slightly different way: Let also Γ = V × {0, 1}2, sitting in G: V ×R×R, V ) and SBlupr,s(G, V ) V × {(0, 0)} ⊂ G = Gj0 ; V × {(1, 0)} ⊂ GV × {0} ⊂ Gj1 j0 j0 ; V × {(0, 1)} ⊂ GV × {0} ⊂ Gj0 ; V × {(1, 1)} ⊂ GV V × {(0, 0)} ⊂ Gj1 j1 j1 . It is a subgroupoid of G. The blowup construction applied to Γ ⊂ G gives then a groupoid SBlupr,s(G, Γ) which is the union of: SBlupr,s(G, V ) ; SBlupr,s(GV × R, V ) ; SBlupr,s(GV × R, V ) ; SBlupr,s(GV V × R × R, V ) . Recall that SBlup(V ×R, V ×{0}) (cid:39) V ×(R−(cid:116)R+). Thus SBlupr,s(G, Γ) is a groupoid with objects SBlup(M, V ) (cid:116) V × R− (cid:116) V × R+. 39 V ×R+ V (cf. section 5.3.3). V ×R×R, V ) V )ga groupoid of GV . It is a linking space between the groupoids The restriction of SBlupr,s(G, Γ) to V ×R+ coincides with the restriction of SBlupr,s(GV to V × R+: it is the gauge adiabatic (GV Put SBlupr,s(GV × R, V )+ = SBlupr,s(G, Γ)SBlup(M,V ) V )ga. Put also SBlupr,s(GV × R, V )+ = SBlupr,s(G, Γ)V ×R+ SBlupr,s(G, V ) and (GV With the notation used in fact 2.4, we define the C∗(SBlupr,s(G, V )) − C∗((GV E (G, V ) to be C∗(SBlupr,s(GV × R, V )+). C∗(SBlupr,s(G, Γ)). It is a full Hilbert-C∗(SBlupr,s(G, V )) − C∗((GV The Hilbert-C∗((GV where Ω = r(GV ) is the union of orbits which meet V . Notice that Ω = M \ V (cid:116) V × R∗ and F = SN M and closed satured subsets of the units of SBlupr,s(G, Γ). Furthermore SBlupr,s(G, Γ)Ω C∗(GΩ of C∗-algebras. Ω, V )) V (cid:116) V (cid:116) V gives a partition by respectively open Ω and Ω ) = C∗(G) according to proposition 6.6. This decomposition gives rise to an exact sequence V )ga)-module E (G, V ) is full and K(E (G, V )) is the ideal C∗(SBlupr,s(GΩ V )ga)-bimodule It is the closure of Cc(SBlupr,s(GV × R, V )+) in V )ga)-module. Ω = GΩ SBlup(M,V ). 0 / C∗(G) / C∗(SBlupr,s(G, Γ)) / C∗(SNG Γ ) / 0 This exact sequence gives rise to an exact sequence of bimodules: 0 0 E (G,V ) V ×R∗ V ×R∗ ) + / C∗(G where E (G, V ) = C∗(GM\V V ×R∗ + / C∗(G) C∗(SBlupr,s(G, V )) C∗(SN G V ) E (G,V ) + / C∗((GV ) and E ∂(G, Γ) = C∗(cid:0)(SNG V )ga) SN M Γ ) V V E ∂ (G,Γ) (cid:111) R∗ +) / C∗(AGV (cid:1) = E (G, V )/ E (G, V ). V 0 / 0 7.1.2 The Poisson-trace bimodule EP T In [17], we constructed, for every Lie groupoid H a C∗(Hga) − Ψ∗(H)-bimodule EH . Recall that the Hilbert Ψ∗(H)-module EH is full and that K(EH ) ⊂ C∗(Hga) is the kernel of a natural ∗-homomorphism C∗(Hga) → C0(H (0) × R). We also showed that the bimodule EH gives rise to an exact sequence of bimodule as above: 0 0 / C∗(H × R∗ + × R∗ +) EH / C∗(H) 0 C∗(AH (cid:111) R∗ +) C∗(Hga) EH / Ψ∗(H) we obtain a C∗(SBlupr,s(G, V ))−Ψ∗(GV / C0(S∗AH) / 0 E ∂ H Putting together the bimodule E (G, V ) and E E (G, V ) ⊗C∗((GV just EP T . It leads to the exact sequence of bimodule: V )ga) GV V GV V E V ) bimodule that we call the Poisson-trace bimodule and denote by EP T (G, V ) - or 0 / C∗(G) EP T (G,V ) C∗(SBlupr,s(G, V )) EP T (G,V ) C∗(SN G V ) 0 E ∂ P T (G,V ) 0 / C∗(GV V ) / Ψ∗(GV V ) V )-module and K(EP T (G, V )) is a two sided ideal of The Poisson-trace bimodule is a full Hilbert Ψ∗(GV V )−C∗(SBlupr,s(G, V ))- C∗(SBlupr,s(G, V )). Denote by EP T (G, V )∗ its dual module, i.e. the Ψ∗(GV bimodule K(EP T (G, V ), Ψ∗(GV / C0(S∗AGV V ) / 0 V )). 40 / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / 7.2 A Boutet de Monvel type algebra The C∗-algebra C∗ C∗(SBlupr,s(G, V ))⊕EP T (G, V )∗(cid:17) BM (G, V ) = K(cid:16) (cid:18)K P (cid:19) is an algebra made of matrices where K ∈ C∗(SBlupr,s(G, V )), P ∈ EP T (G, V ), T ∈ EP T (G, V )∗, Q ∈ of the form Ψ∗(GV We have an exact sequence (where M (cid:116) V (cid:54)= M denotes the topological disjoint union of M with V ): T Q V ). 0 → C∗(G where the quotient ΣC∗ It is the algebra of matrices of the form P T (G, V ) := EP T (G, V ) ⊗Ψ∗(GV E V rC∗ V−→ ΣC∗ bound(G, V ) → 0, M(cid:116)V M(cid:116)V ) → C∗ BM (G, V ) (cid:18)k p (cid:19) (cid:18)r(cid:71) (cid:19) (cid:18)K P V ) C(S∗AGV V ). The map rC∗ rC∗ = q t V V T Q V (K) rV (P ) rV (T ) σV (Q) is of the form (cid:19) bound(G, V ) is the algebra of the Boutet de Monvel type boundary symbols. V ), p, t∗ ∈ V ), q ∈ C(S∗AGV where k ∈ C∗(SN G where: • the quotient map σV is the ordinary order 0 principal symbol map on the groupoid GV V ; • the quotient maps r(cid:71) V , rV , rV are restrictions to the boundary N M V : V : C∗(SBlupr,s(G, V )) → C∗(SN G r(cid:71) V ) = C∗(SBlupr,s(G, V ))/C∗(G M M ), rV : EP T (G, V ) → E V P T (G, V ) = EP T (G, V )/C∗(G M V ), and rV (T ) = rV (T ∗)∗. The map rC∗ V is called the zero order symbol map of the Boutet de Monvel type calculus. 7.3 A Boutet de Monvel type pseudodifferential algebra We denote by Ψ∗ EP T (G, V ), T ∈ EP T (G, V )∗ and Q ∈ Ψ∗(GV V ). Such an operator R = BM (G, V ) the algebra of matrices (cid:18)Φ P has two symbols: (cid:19) T Q T Q (cid:18)Φ P (cid:19) • the classical symbol σc : Ψ∗ • the boundary symbol rBM V BM (G, V ) → C0(S∗ASBlupr,s(G, V )) given by σc : Ψ∗ BM (G, V ) → ΣΨ∗ bound(G, V ) defined by (cid:18)Φ P (cid:19) T Q rV (cid:18)rψ = V (Φ) rV (P ) rV (T ) σV (Q) where rψ V : Ψ∗(SBlupr,s(G, V )) → Ψ∗(SN G V ) is the restriction. bound(G, V ) denotes the algebra of matrices of the form Here ΣΨ∗ P T (G, V ) and q ∈ C(S∗AGV E V V ). The full symbol map is the morphism (cid:19) with φ ∈ Ψ∗(SN G V ), p, t∗ ∈ σBM : Ψ∗ BM (G, V ) → ΣBM (G, V ) := C0(S∗ASBlupr,s(G, V )) ×C0(S∗ASN G V ) ΣΨ∗ bound(G, V ) 41 with Φ ∈ Ψ∗(SBlupr,s(G, V )), P ∈ (cid:18)Φ P (cid:19) T Q = σc(Φ); (cid:19) (cid:18)φ p t q defined by σBM (R) = (σc(R), rV (R)). We have an exact sequence: 0 → C∗(G M(cid:116)V M(cid:116)V ) → Ψ∗ BM (G, V ) We may note that Ψ∗(SBlupr,s(G, V )) (resp. Ψ∗(SN G of Ψ∗ BM (G, V ) (resp. of ΣBM (G, V )) consisting of elements of the form (cid:18)x 0 (cid:19) 0 0 . σBM−→ ΣBM (G, V ) → 0. EBM V )) identifies with the full hereditary subalgebra 7.4 K-theory of the symbol algebras and index maps In this section we examine the index map corresponding to the Boutet de Monvel type calculus and in particular to the exact sequence EBM . We compute the K-theory of the symbol algebra ΣBM and the connecting element (cid:102)indBM ∈ KK1(ΣBM , C∗(G)) (4). We then extend this computation by including bundles into the picture i.e. by computing a relative K-theory map. 7.4.1 K-theory of ΣBM and computation of the index As the Hilbert Ψ∗(GV V ) module EP T (G, V ) is full, ; K ∈ C∗(SBlupr,s(G, V )) is a full hereditary subalgebra of • the subalgebra C∗ BM (G, V ); • the subalgebra • the subalgebra • the subalgebra • the subalgebra (cid:19) 0 0 (cid:110)(cid:18)K 0 (cid:19) (cid:110)(cid:18)Φ 0 (cid:19) (cid:110)(cid:18)x 0 (cid:19) (cid:110)(cid:18)k 0 (cid:19) (cid:110)(cid:18)φ 0 0 0 0 0 0 0 0 0 (cid:111) (cid:111) (cid:111) (cid:111) (cid:111) ; Φ ∈ Ψ∗(SBlupr,s(G, V )) is a full hereditary subalgebra of Ψ∗ BM (G, V ); ; x ∈ ΣSBlup(G, V ) is a full hereditary subalgebra of ΣBM (G, V ); ; k ∈ C∗(SN G V ) ; φ ∈ Ψ∗(SN G V ) is a full hereditary subalgebra of ΣC∗ bound(G, V ); is a full hereditary subalgebra of ΣΨ∗ bound(G, V ). We have a diagram of exact sequences where the vertical inclusions are Morita equivalences: 0 0 / C∗(G M M ) Ψ∗(SBlupr,s(G, V )) σf ull ΣSBlup(G, V ) / C∗(G M(cid:116)V M(cid:116)V ) / Ψ∗ BM (G, V ) σBM / ΣBM (G, V ) 0 / 0 We thus deduce immediately from theorem 6.8 and prop. 6.11: this K-equivalence, the index (cid:102)indBM is q∗([Bott]⊗CindG) where q : Cχ → C0(A∗G×R∗ Corollary 7.1. The algebra ΣBM (G, V )) is KK-equivalent with the mapping cone Cχ and, under +) is evaluation at 0. 4We use the Morita equivalence of C ∗ (G) with C ∗ M(cid:116)V M(cid:116)V ). (G 42 / / / _ / / _ / / _ / / / / 7.4.2 Index in relative K-theory One may also consider more general index problems, which are concerned with generalized boundary value problems in the sense of [49, 35, 36]: those are concerned with index of fully elliptic operators of , where we are given hermitian complex vector bundles E± over SBlup(M, V ) (cid:18)Φ P (cid:19) the form R = T Q and F± over V , and • Φ is an order 0 pseudodifferential operator of the Lie groupoid SBlupr,s(G, V ) from sections of E+ to sections of E−; • P is an order 0 "Poisson type" operator from sections of F+ to sections of E−; • T is an order 0 "trace type" operator from sections of E+ to sections of F−; • Q is an order 0 pseudodifferential operator of the Lie groupoid GV V from sections of F+ to sections of F−. M(cid:116)V M(cid:116)V )). BM (G, V ))(p+ ⊕ q+). BM (G, V ))(p− ⊕ q−), such that (p+ ⊕ q+) − R(cid:48)R ∈ MN (C∗(G M(cid:116)V M(cid:116)V In other words, writing E± as associated with projections p± ∈ MN (C∞(SBlup(M, V ))) and F± as associated with projections q± ∈ MN (C∞(V )), then R ∈ (p− ⊕ q−)MN (Ψ∗ Full ellipticity for R means just that the full symbol of R is invertible, i.e. that there is a quasi- inverse R(cid:48) ∈ (p+ ⊕ q+)MN (Ψ∗ )) and (p− ⊕ q−) − RR(cid:48) ∈ MN (C∗(G In other words, we wish to compute the morphism indrel : K∗(µBM ) → K∗(C∗ is the natural morphism µBM : C0(SBlup(M, V )) ⊕ C0(V ) → ΣBM (G, V ). Let us outline here this computation. We start with a remark. Remark 7.2. Let H ⇒ V be a Lie groupoid. The bimodule E ∂ H is a Morita equivalence of an ideal C∗(AH(cid:111)R∗ +)). Let µH : C0(V ) → C0(S∗AH) be the inclusion (given by the map S∗AH → V ). The composition H (ζH ) is the zero element in KK(C0(V ), C∗(AH (cid:111) R∗ µ∗ +)). Indeed µ∗ H (ζH ) can be decomposed as • the Morita equivalence C0(V ) ∼ C0((V × R∗ +) (cid:111) R∗ +), + ⊂ C0(V × R+) (cid:111) R∗ • the inclusion C0(V × R∗ +) (cid:111) R∗ +, • the inclusion C0(V × R+) (cid:111) R∗ + → C0(A∗H) (cid:111) R∗ +) with C0(S∗AH) and therefore defines an element ζH ∈ KK(C0(S∗AH), C∗(AH(cid:111)R∗ + corresponding to the map (x, ξ) (cid:55)→ (x,(cid:107)ξ(cid:107)) BM (G, V )) where µBM from A∗H to V × R+. Now, the Toeplitz algebra C0(R+) (cid:111) R∗ From this remark, we immediately deduce: Proposition 7.3. The inclusion C0(V ) → ΣBM (G, V ) is the zero element in KK-theory. We have a diagram + is K-contractible. (cid:3) C0(SBlup(M, V )) ⊕ C0(V ) µSBlup⊕µV / / ΣSBlup(G, V ) ⊕ C0(S∗AGV V ) C0(SBlup(M, V )) ⊕ C0(V ) ψBM / ΣBM (G, V ) The mapping cone CµSBlup of the morphism µSBlup : C0(SBlup(M, V ))⊕ 0 → Ψ∗ BM (G, V ) is Morita- equivalent to the mapping cone of the morphism µSBlup : C0(SBlup+(M, V )) → ΣSBlup(G, V )) and therefore it is KK-equivalent to C0(A∗G × R) by Cor. 6.14. We then deduce: Theorem 7.4. a) The relative K-theory of µBM is naturally isomorphic to K∗(A∗G)⊕K∗+1(C0(V )). b) Under this equivalence, the relative index map identifies with indG on K∗(A∗G) and the zero (cid:3) map on K∗+1(C0(V ). 43 / 8 Appendix 8.1 A characterization of groupoids via elements composable to a unit Remark 8.1. Let G be a groupoid. For n ∈ N, we may define the subsets U n(G) = {(x1, . . . , xn) ∈ Gn; s(xi) = r(xi+1); x1· x2 . . . xn ∈ G(0). We have U 1(G) = G(0), the sets U k(G) are invariant under cyclic permutations; moreover, we have natural maps δk : U n(G) → U n+1(G) (1 ≤ k ≤ n) defined by δk(x1, . . . , xn) = (x1, . . . , xk, s(xk), xk+1, . . . , xn) and boundaries bk : U n+1(G) → U n(G) defined by bk(x1, . . . , xn+1) = (x1, . . . , xk−1, xkxk+1, xk+2, . . . xn+1) if k (cid:54)= n + 1 and bn+1(x1, . . . , xn+1) = (xn+1x1, x2, . . . , xn). Let G be a set. For (x, y, z) ∈ G3, put q1(x, y, z) = x and q1,2(x, y, z) = (x, y). Proposition 8.2. Let G be a set and U 1(G) ⊂ G, U 3(G) ⊂ G3 be subsets satisfying the following conditions. a) The subset U 3(G) of G3 is invariant under cyclic permutation. b) For all x ∈ U 1(G), (x, x, x) ∈ U 3(G). c) The map q12 : (x, y, z) (cid:55)→ (x, y) from U 3(G) to G2 is injective. d) Let R = {(x, y, z) ∈ U 3(G); z ∈ U 1(G)}. The map q1 : R → G is a bijection and U 2(G) = q12(R) is invariant under (cyclic) permutation. e) The subset U 4(G) = {(x, y, z, t) ∈ G4; ∃(u, v) ∈ U 2(G); (x, y, v) ∈ U 3(G) and (u, z, t) ∈ U 3(G)} is invariant under cyclic permutation in G4. Then there is a unique groupoid structure on G such that U 1(G) is its set of units, and U 3(G) = {(x, y, z) ∈ G3; (x, y) ∈ G(2), z = (xy)−1}. Proof. Uniqueness is easy: one defines the range and the inverse of x by saying that (x, x−1, r(x)) is the unique element w in R such that q1(w) = x; the source is defined by s(x) = r(x−1); the product for composable elements (x, y) is then defined by the fact (x, y, (xy)−1) ∈ U 3(G). Let us pass to existence. • By condition (d), U 2(G) is the graph of an involution that we denote by x (cid:55)→ x−1. By condition (b), if x ∈ U 1(G), then x−1 = x. • Define also r : G → U 1(G) to be the (unique) element in U 1(G) such that (x, x−1, r(x)) ∈ U 3(G) and put s(x) = r(x−1). • Put G(2) = q12(U 3(G)). If (x, y) ∈ G(2), then there exists z such that (x, y, z) ∈ U 3(G). As (x, s(x), x−1) ∈ U 3(G), it follows that (x, s(x), y, z) ∈ U 4(G) and thus (z, x, s(x), y) ∈ U 4(G), and therefore (s(x), y) ∈ G(2), and r(y) = s(x). Conversely, if (x, y) ∈ G2 satisfy s(x) = r(y), as (x−1, x, s(x)) ∈ U 3(G) and (s(x), y, y−1) ∈ U 3(G), it follows that (x−1, x, y, y−1) ∈ U 4(G), whence (x, y, y−1, x−1) ∈ U 4(G) and (x, y) ∈ G(2). In other words, G(2) = {(x, y) ∈ G2; s(x) = r(y)}. • For (x, y) ∈ G(2), we may define thanks to condition (c) the element xy ∈ G, by the requirement (x, y, (xy)−1) ∈ U 3(G). • Since (y, (xy)−1) ∈ G(2) and ((xy)−1, x) ∈ G(2), it follows that s(xy) = r((xy)−1) = s(y) and r(xy) = s((xy)−1) = r(x). 44 • For x ∈ G, since (r(x), x, x−1) ∈ U 3(G) and (x, s(x), x−1) ∈ U 3(G), it follows that r(x)x = x and xs(x) = x - thus units are units. As (x, x−1, r(x)) ∈ U 3(G) and (x−1, x, s(x)) ∈ U 3(G) we find xx−1 = r(x) and x−1x = s(x) and thus x−1 is the inverse of x. • Finally, let (x, y, z) ∈ G3 be such that (x, y) ∈ G(2) and (y, z) ∈ G(2). We saw that s(xy) = s(y) = r(z). Put w = ((xy)z)−1. Then (x, y, (xy)−1) ∈ U 3(G) and (xy, z, w) ∈ U 3(G), and thus (x, y, z, w) ∈ U 4(G), whence (y, z, w, x) ∈ U 4(G) and therefore (yz, w, x) ∈ U 3(G) and finally, (x, yz, w) ∈ U 3(G), which means that x(yz) = w−1 = (xy)z. 8.2 VB groupoids and their duals ([46, 31]) A VB groupoid over a groupoid G is a vector bundle E over G with a groupoid structure such that E(0) ⊂ E is a vector subbundle of the restriction of E to G(0) and such that all the structure maps of the groupoid E (rE, sE, x (cid:55)→ x−1 and the composition) are linear bundle maps and rE : E → r∗ G(E(0)) is surjective. Proposition 8.3. Let E → G be a VB groupoid. For all k ∈ N (k ≥ 1) U k(E) → U k(G) is a subbundle of the restriction to U k(G) ⊂ Gk of the bundle Ek → Gk. We identify the dual bundle of Ek → Gk with (E∗)k. Then the dual bundle E∗ is a VB-groupoid over G with U k(E∗) = U k(E)⊥ for all k. Proof. We prove that U 1(E∗) = (E(0))⊥ and U 3(E∗) = U 3(E)⊥ satisfy the conditions of prop. 8.2. (a) Condition (a) is obvious: since U 3(E) is invariant under cyclic permutations, so is U 3(E)⊥. (b) Taking the restriction of E over a point of G(0), we have a linear groupoid, and we have already proved that its orthogonal, is a linear groupoid. Condition (b) follows immediately. (c) By condition (d) for E, it follows that q1 : U 3(E) → E is onto, whence (by condition a) q3 : U 3(E) → E is onto. Therefore q1,2 : U 3(E)⊥ → E∗ × E∗ is injective. (d) Since q2,3 : U 3(E) → E × E injective, it follows that q1 : U 3(E)⊥ → E∗ is onto. Since U 2(E) is the graph of an involution, the same holds for U 2(E)⊥. Note also that condition (d) ensures that q1 : U 2(E)⊥ → E∗ is an isomorphism. We then just have to show that {(x, y, z) ∈ U 3(E)⊥; z ∈ (U 1(E))⊥} = {(x, y, z) ∈ U 3(E)⊥; (x, y) ∈ U 2(E)⊥}. The first term is the orthogonal of U 3(E) + F1 where F1 = {(0, 0, z); z ∈ U 1(E)} and the second the orthogonal of U 3(E)+F2 where F2 = {(x, y, 0); (x, y) ∈ U 2(E)}. Now, for every (x, y) ∈ U 2(E) there exists z ∈ U 1(E), namely z = rE(x) = sE(y) such that (x, y, z) ∈ U 3(E); by surjectivity of rE, it follows that for every γ ∈ G and every z ∈ E(0) there exists x ∈ Eγ such that rE(x) = z, thus (x, x−1, z) ∈ U 3(E). In other words, U 3(E) + F1 = U 3(E) + F2. Condition (d) follows. γ (e) We just need to show that U 4(E)⊥ = {(w, x, y, z) ∈ (E∗)4; ∃(u, u(cid:48)) ∈ U 2(E)⊥, (w, x, u) ∈ U 3(E)⊥ and (u(cid:48), y, z) ∈ U 3(E)⊥}. As U 4(E)⊥ is cyclicly invariant, condition (e) will follow. If there exists (u, u(cid:48)) ∈ U 2(E)⊥ such that (w, x, u) ∈ U 3(E)⊥ and (u(cid:48), y, z) ∈ U 3(E)⊥, then, for every (a, b, c, d) ∈ U 4(E), there exists (v, v(cid:48)) ∈ U 2(E) such that (a, b, v) ∈ U 3(E) and (v(cid:48), c, d) ∈ U 3(E). It follows that (w, x, y, z) ∈ U 4(E)⊥. The inclusion {(w, x, y, z) ∈ (E∗)4; ∃(u, u(cid:48)) ∈ U 2(E)⊥, (w, x, u) ∈ U 3(E)⊥ and (u(cid:48), y, z) ∈ U 3(E)⊥} ⊂ U 4(E)⊥ follows. Now, as vector bundles, if dim E = n and dim E(0) = p, it follows that dim(U k(E)) = (k − 1)n − (k − 2)p (for all k ≥ 1); therefore dim U 3(E)⊥ = n + p and dim U 4(E)⊥ = n + 2p. As the projection q1 : U 3(E)⊥ → E∗ is onto, we find that for (γ1, γ2, γ3, γ4) ∈ U 4(G), we have dim{(w, x, u), (v, y, z) ∈ U 3(E)⊥ (γ1γ2,γ3,γ4); v = u−1} is (n + p) + (n + p) − n and we find the desired equality by dimension equality. (γ1,γ2,γ3γ4) × U 3(E)⊥ It is then quite immediately seen, using induction and dimension equality, that, for every k, we have U k(E∗) = U k(E)⊥. 45 8.3 Fourier transform Remark 8.4. Let F1, F2 be real vector spaces and let H be a subspace of F1 × F2. Assume that p1 : H → F1 is injective and p2 : H → F2 is surjective. For f ∈ S (F1) we put qH (f ) = (p2)!(p∗ 1)(f ). Then, for every f ∈ S (F1), (cid:92)qH (f ) = qH⊥( f ) (taking a good normalization for the Fourier transform). (cid:90) Indeed we can write F1 = F2 × L × K, H = {((x, y, 0), x), x ∈ F2, y ∈ L}. Then H⊥ = 2 , η ∈ K∗}. {((ξ, 0, η),−ξ), ξ ∈ F ∗ For x ∈ F2, we have qH (f )(x) = For ξ ∈ F ∗ f (x, y, 0) dy. L 2 , we have f (x, y, 0)e−i(cid:104)xξ(cid:105) dx dy (cid:90) (cid:92)qH (f )(ξ) = (cid:90) F1×L (cid:16)(cid:90) (cid:17) (cid:90) K∗ (cid:17) dη. and (cid:90) (cid:16)(cid:90) qH⊥( f )(ξ) = f (ξ, 0, η) dη = K∗ F2×L×K f (x, y, z)e−i((cid:104)xξ(cid:105)+(cid:104)zη(cid:105))dx dy dz f (x, y, z)e−i(cid:104)zη(cid:105)dz K K∗ dη = f (x, y, 0). But Remark 8.5. Let E → G be a VB groupoid. For γ ∈ G, Ex is a linking space between the groupoids Es(x) and Er(x). The family of Hilbert bimodules C∗(Ex)x∈G is a Fell bundle and C∗(E) is the C∗-algebra associated with this Fell bundle ([29, 41, 43, 26]). Proposition 8.6. Let E → G be a VB groupoid, and let E∗ be the dual VB groupoid. Then C∗(E) (cid:39) C∗(E∗) - via Fourier transform. Proof. For (γ, γ(cid:48)) ∈ G(2), let F1 ⊂ Eγ × Eγ(cid:48) be the set of composable elements; let F2 = E(γγ(cid:48))−1 and H ⊂ F1 × F2 the set of (x, y, z) ∈ Eγ × Eγ(cid:48) × E(γγ(cid:48))−1 that compose to a unit. Remark 8.4 implies that for f ∈ S (Eγ) and g ∈ S (Eγ), we have (cid:92)(f · g) = f · g - where f · g ∈ Eγγ(cid:48) is the "Fell bundle product". In other words, the Fourier transform map is an isomorphism of the Fell-bundles and therefore the corresponding C∗-algebras are isomorphic. List of Symbols Fiber bundles N M V P(E), S(E) The normal bundle of a submanifold V of a manifold M , page 8 The projective and sphere bundles associated to a real vector bundle E over M , whose fiber over x ∈ M are respectively the projective space P(Ex) and the sphere S(Ex), page 7 B∗E, B∗E, S∗E The total spaces of the fiber bundles of open balls, closed ball and spheres of the dual vector bundle E∗ of E, page 7 The quotient of B∗E where we identify two points (x, ξ) and (x, η) for x ∈ F , F being a closed subset of the zero section M of the bundle E, page 7 The image of S∗E in B∗ F E, page 7 B∗ F E S∗ F E Groupoids, deformation and blowup spaces r,s⇒ G(0) G A Lie groupoid with source s, range r and space of units G(0), page 8 AG The Lie algebroid of the groupoid G, page 8 46 GA, GB, GA B Gf , Gg, Gf g A = GA ∩ GB , page 8 If A and B are subsets of G(0), GA = {x ∈ G; r(x) ∈ A}, GB = {x ∈ G; s(x) ∈ B} and GB If f : A → G(0) and g : B → G(0) are maps, Gf = {(a, x) ∈ A× G; r(x) = f (a)}, Gg = {(x, b) ∈ G × B; s(x) = g(b)} and Gf g = Gf ∩ Gg, page 8 Gad, G[0,1] ad , G[0,1) Gad(X) AW G DN C(Y, X) DN CT (Y, X) ad The adiabatic groupoid of G and its restriction respectively to G(0) × [0, 1] and to G(0) × [0, 1) , page 8 The restriction of Gad to a locally closed saturated subset X of G(0) × [0, 1], page 8 The restriction of the adiabatic groupoid Gad to F × [0, 1) ∪ W × {0} where F is a closed subset of G(0) saturated for G and W = G(0) \ F , page 13 The deformation to the normal cone of the inclusion of a submanifold X in a manifold Y , DN C(Y, X) = Y × R∗ ∪ N Y The restriction Y × (T \ {0}) ∪ N Y R containing 0, page 24 X × {0} of DN C(Y, X) to a closed subset T of X , page 22 DN C+(Y, X) The restriction DN CR+(Y, X), page 24 Blup(Y, X) SBlup(Y, X) Blupf (Y, X) The blowup of the inclusion of a submanifold X in a manifold Y , Blup(Y, X) = Y \ X ∪ P(N Y X ), page 23 The spherical blowup of the inclusion of a submanifold X in a manifold Y , SBlup(Y, X) = Y \ X ∪ S(N Y The subspace of Blup(Y, X) on which Blup(f ) : Blupf (Y, X) → Blup(Y (cid:48), X(cid:48)) can be defined for a smooth map f : Y → Y (cid:48) (with f (X) ⊂ X(cid:48)), page 24 X ), page 23 DN C(G, Γ) ⇒ DN C(G(0) 2 , G(0) 1 ) The deformation groupoid where Γ is a closed Lie subgroupoid of a Lie groupoid G, page 28 Blupr,s(G, Γ) ⇒ Blup(G(0), Γ(0)) The blowup groupoid Blupr(G, Γ) ∩ Blups(G, Γ) where Γ is a closed Lie subgroupoid of a Lie groupoid G, page 28 SBlupr,s(G, Γ) (cid:94)DN C(G, Γ) (cid:94)DN C+(G, Γ) C∗-Algebras C∗(G) Ψ∗(G) Ψ∗(GF ) Ψ∗ F (G) Cf The spherical version of Blupr,s(G, Γ), page 28 The open subgroupoid of DN C(G, Γ) of DN C(G, Γ) consisting of elements whose image by DN C(r) and DN C(s) is not in G(0) The restriction of (cid:94)DN C(G, Γ) to R+, page 28 1 × R, page 28 The (either maximal or reduced) C∗-algebra of the groupoid G, page 8 The C∗-algebra of order ≤ 0 pseudodifferential operators on G vanishing at infinity on G(0), page 8 The quotient Ψ∗(G)/Ψ∗(GW ) where F is a closed subset of G(0) saturated for G and W = G(0) \ F , page 12 The subalgebra C0(M ) + Ψ∗(GW ) of Ψ∗(G), page 15 The mapping cone of a morphism f : A → B of C∗-algebra, page 12 47 ΣW (G) ΣF (G) ΣDN C+(G, Γ) Σ(cid:94)DN C+ (G, Γ) The quotient Ψ∗(G)/C∗(GW ), page 12 The algebra Ψ∗ F (G)/C∗(GW ), page 15 The algebra ΣM×R∗ M×R∗ +(DN C+(G, Γ)), page 33 +((cid:94)DN C+(G, Γ)), page 33 The algebra Σ ΣSBlup(G, Γ) The algebra Σ M (SBlupr,s(G, Γ)), page 33 KK-elements [f ] indG (cid:102)indG ∂W G (cid:102)ind W f ull(G) indW G ∂F The KK-element, in KK(A, B) associated to a morphism of C∗-algebra f : A → B, page 12 The KK-element [ev0]−1⊗ [ev1], which belongs to KK(C0(A∗G), C∗(G)), associ- ad = G× (0, 1]∪ A(G)×{0} ⇒ G(0) × [0, 1], ated to the deformation groupoid G[0,1] page 11 The connecting element, which belongs to KK1(C(S∗AG), C∗(G)) associated to the short exact sequence 0 → C∗(G) → Ψ∗(G) → C(S∗AG) → 0, page 11 The connecting element, which belongs to KK1(C∗(GF ), C∗(GW )), associated / C∗(GF ) / 0 to the short exact sequence 0 where W is a saturated open subset of G(0) and F = G(0) \ W , page 12 The connecting element, which belongs to KK1(ΣW (G), C∗(GW )) associated to the short exact sequence 0 −→ C∗(GW ) −→ Ψ∗(G) −→ ΣW (G) −→ 0, page 12 The KK-element [ev0]−1 ⊗ [ev1], which belongs to KK(C∗(AW G), C∗(GW )), associated to the groupoid Gad(F × [0, 1) ∪ W × [0, 1]) = GW × (0, 1] (cid:116) AW G, page 14 The connecting element, which belongs to KK1(ΣF (G), C∗(GW )) associated / 0 , to the short exact sequence 0 page 16 / C∗(GW ) / C∗(GW ) / C∗(G) / ΣF (G) F (G) / Ψ∗ DN C+ ∂G,Γ SBlup, ∂G,Γ (cid:102)ind SBlup, (cid:102)ind G,Γ , ∂G,Γ G,Γ (cid:94)DN C+ DN C+, (cid:102)ind and (cid:102)ind G,Γ (cid:94)DN C+ Respectively the element ∂ Respectively the elements (cid:102)ind M×R∗ f ull + ((cid:94)DN C+(G, Γ)), page 34 (DN C+(G, Γ)) M SBlupr,s(G,Γ), ∂ + M×R∗ DN C+(G,Γ) and ∂ f ull(SBlupr,s(G, Γ)), (cid:102)ind M M×R∗ (cid:94)DN C+(G,Γ) + M×R∗ f ull + , page 33 References [1] Johannes Aastrup, Severino T. Melo, Bertrand Monthubert, and Elmar Schrohe, Boutet de Monvel's calculus and groupoids I., J. Noncommut. Geom. 4 no. 3 (2010), 313 -- 329. [2] Pierre Albin and Richard B. Melrose, Fredholm realizations of elliptic symbols on manifolds with boundary., J. Reine Angew. Math. 627 (2009), 155 -- 181. [3] , Fredholm realizations of elliptic symbols on manifolds with boundary II: Fibered bound- ary., Clay Math. Proc. 12 (2010), 99 -- 117. 48 / / / / / / / / [4] Claire Anantharaman and Jean Renault, Amenable groupoids, Groupoids in analysis, geometry, and physics (Boulder, CO, 1999), Contemp. Math., vol. 282, Amer. Math. Soc., Providence, RI, 2001, pp. 35 -- 46. [5] Iakovos Androulidakis and Georges Skandalis, The analytic index of elliptic pseudodifferential operators on a singular foliation, J. K-Theory 8 (2011), no. 3, 363 -- 385. [6] Michael F. Atiyah and Isadore M. Singer, The index of elliptic operators. I, Ann. of Math. (2) 87 (1968), 484 -- 530. [7] , The index of elliptic operators. IV, Ann. of Math. (2) 93 (1971), 119 -- 138. [8] Louis Boutet de Monvel, Boundary problems for pseudo-differential operators, Acta Math. 126 (1971), no. 1-2, 11 -- 51. [9] Jonathan Henry Brown, Proper actions of groupoids on C∗-algebras., J. Oper. Theory 67 (2012), no. 2, 437 -- 467 (English). [10] Paulo Carrillo Rouse, A Schwartz type algebra for the tangent groupoid, K-theory and noncom- mutative geometry, EMS Ser. Congr. Rep., Eur. Math. Soc., Zurich, 2008, pp. 181 -- 199. [11] Alain Connes, Sur la th´eorie non commutative de l'int´egration, Alg`ebres d'op´erateurs (S´em., Les Plans-sur-Bex, 1978), Lecture Notes in Math., vol. 725, Springer, Berlin, 1979, pp. 19 -- 143. [12] , Noncommutative geometry, Academic Press Inc., San Diego, CA, 1994. [13] Alain Connes and Nigel Higson, D´eformations, morphismes asymptotiques et K-th´eorie bivari- ante, C. R. Acad. Sci. Paris S´er. I Math. 311 (1990), no. 2, 101 -- 106. [14] Joachim Cuntz and Georges Skandalis, Mapping cones and exact sequences in KK-theory, J. Operator Theory 15 (1986), no. 1, 163 -- 180. [15] Claire Debord and Jean-Marie Lescure, K-duality for pseudomanifolds with isolated singulari- ties, J. Funct. Anal. 219 (2005), no. 1, 109 -- 133. [16] Claire Debord, Jean-Marie Lescure, and Fr´ed´eric Rochon, Pseudodifferential operators on man- ifolds with fibred corners, Ann. Inst. Fourier (Grenoble) 65 (2015), no. 4, 1799 -- 1880. [17] Claire Debord and Georges Skandalis, Adiabatic groupoid, crossed product by R∗ + and pseudod- ifferential calculus, Adv. Math. 257 (2014), 66 -- 91. [18] [19] , Pseudodifferential extensions and adiabatic deformation of smooth groupoid actions, Bull. Sci. Math. 139 (2015), no. 7, 750 -- 776. , Stability of Lie groupoid C∗-algebras, J. Geom. Phys. 105 (2016), 66 -- 74. [20] Marco Gualtieri and Songhao Li, The Stokes groupoids, arXiv:1305.7288. [21] , Symplectic groupoids of log symplectic manifolds, Int. Math. Res. Not. IMRN (2014), no. 11, 3022 -- 3074. [22] Nigel Higson, On the analogy between complex semisimple groups and their Cartan motion groups, Contemporary Mathematics 137 (2011), 1 -- 33. [23] Nigel Higson, Vincent Lafforgue, and Georges Skandalis, Counterexamples to the Baum-Connes conjecture, Geom. Funct. Anal. 12 (2002), no. 2, 330 -- 354. MR 1911663 (2003g:19007) [24] Michel Hilsum and Georges Skandalis, Stabilit´e des C∗-alg`ebres de feuilletages, Ann. Inst. Fourier (Grenoble) 33 (1983), no. 3, 201 -- 208. 49 [25] , Morphismes K-orient´es d'espaces de feuilles et fonctorialit´e en th´eorie de Kasparov (d'apr`es une conjecture d'A. Connes), Ann. Sci. ´Ecole Norm. Sup. (4) 20 (1987), no. 3, 325 -- 390. [26] Marius Ionescu and Dana P. Williams, Irreducible induced representations of Fell bundle C∗- algebras, Trans. Amer. Math. Soc. 367 (2015), no. 7, 5059 -- 5079. [27] Gennadi G. Kasparov, The operator K-functor and extensions of C∗-algebras, Izv. Akad. Nauk SSSR Ser. Mat. 44 (1980), no. 3, 571 -- 636, 719. [28] , Equivariant KK-theory and the Novikov conjecture, Invent. Math. 91 (1988), no. 1, 147 -- 201. [29] Alex Kumjian, Fell bundles over groupoids, Proc. Amer. Math. Soc. 126 (1998), no. 4, 1115 -- 1125. [30] Pierre-Yves Le Gall, Th´eorie de Kasparov ´equivariante et groupoıdes. I, K-Theory 16 (1999), no. 4, 361 -- 390. [31] Kirill C. H. Mackenzie, General theory of Lie groupoids and Lie algebroids, London Mathemat- ical Society Lecture Note Series, vol. 213, Cambridge University Press, Cambridge, 2005. [32] Rafe Mazzeo, The Hodge cohomology of a conformally compact metric, J. Differential Geom. 28 (1988), no. 2, 309 -- 339. [33] Rafe Mazzeo and Richard B. Melrose, Pseudodifferential operators on manifolds with fibred boundaries, Asian J. Math. 2 (1998), no. 4, 833 -- 866, Mikio Sato: a great Japanese mathemati- cian of the twentieth century. [34] Rafe R. Mazzeo and Richard B. Melrose, Meromorphic extension of the resolvent on complete spaces with asymptotically constant negative curvature, J. Funct. Anal. 75 (1987), no. 2, 260 -- 310. [35] Severino T. Melo, Thomas Schick, and Elmar Schrohe, A K-theoretic proof of Boutet de Mon- vel's index theorem for boundary value problems, J. Reine Angew. Math. 599 (2006), 217 -- 233. , C∗-algebra approach to the index theory of boundary value problems, Analysis, geometry and quantum field theory, Contemp. Math., vol. 584, Amer. Math. Soc., Providence, RI, 2012, pp. 129 -- 146. [36] [37] Richard B. Melrose, The Atiyah-Patodi-Singer index theorem, Research Notes in Mathematics, vol. 4, A K Peters Ltd., Wellesley, MA, 1993. [38] Bertrand Monthubert, Pseudodifferential calculus on manifolds with corners and groupoids, Proc. Amer. Math. Soc. 127 (1999), no. 10, 2871 -- 2881. [39] , Groupoids and pseudodifferential calculus on manifolds with corners, J. Funct. Anal. 199 (2003), no. 1, 243 -- 286. [40] Bertrand Monthubert and Fran¸cois Pierrot, Indice analytique et groupoıdes de Lie, C. R. Acad. Sci. Paris S´er. I Math. 325 (1997), no. 2, 193 -- 198. [41] Paul S. Muhly, Bundles over groupoids, Groupoids in analysis, geometry, and physics (Boulder, CO, 1999), Contemp. Math., vol. 282, Amer. Math. Soc., Providence, RI, 2001, pp. 67 -- 82. [42] Paul S. Muhly, Jean N. Renault, and Dana P. Williams, Equivalence and isomorphism for groupoid C∗-algebras, J. Operator Theory 17 (1987), no. 1, 3 -- 22. [43] Paul S. Muhly and Dana P. Williams, Equivalence and disintegration theorems for Fell bundles and their C∗-algebras, Dissertationes Math. (Rozprawy Mat.) 456 (2008), 1 -- 57. 50 [44] Victor Nistor, Desingularization of lie groupoids and pseudodifferential operators on singular spaces, arXiv:1512.08613, 2016. [45] Victor Nistor, Alan Weinstein, and Ping Xu, Pseudodifferential operators on differential groupoids, Pacific J. Math. 189 (1999), no. 1, 117 -- 152. [46] Jean Pradines, Remarque sur le groupoıde cotangent de Weinstein-Dazord, C. R. Acad. Sci. Paris S´er. I Math. 306 (1988), no. 13, 557 -- 560. [47] Fr´ed´eric Rochon, Pseudodifferential operators on manifolds with foliated boundaries, Microlo- cal methods in mathematical physics and global analysis, Trends Math., Birkhauser/Springer, Basel, 2013, pp. 77 -- 80. [48] Thomas Schick, Modern index theory, Lectures held at CIRM - Rencontre "Thorie de l'Indice", 2006. [49] Elmar Schrohe, A short introduction to Boutet de Monvel's calculus, Approaches to singular analysis (Berlin, 1999), Oper. Theory Adv. Appl., vol. 125, Birkhauser, Basel, 2001, pp. 85 -- 116. [50] Vito Felice Zenobi, Adiabatic groupoids and secondary invariants in k-theory, arXiv:1609.08015. 51
1703.06546	3	1703	2018-02-28T16:32:21	Partial actions of C*-quantum groups I: Restriction and Globalization	[ "math.OA" ]	Partial actions of groups on C-algebras and the closely related actions and coactions of Hopf algebras received much attention over the last decades. They arise naturally as restrictions of their global counterparts to non-invariant subalgebras, and the ambient eveloping global (co)actions have proven useful for the study of associated crossed products. In this article, we introduce the partial coactions of C-bialgebras, focussing on C-quantum, and prove existence of an enveloping global coaction under mild technical assumptions. The construction of the latter provides a left adjoint to the forgetful functor from coactions to partial coactions. We also show that partial coactions of the function algebra of a discrete group correspond to partial actions on direct summands of a C-algebra, and relate partial coactions of a compact or its dual discrete C*-quantum group to partial coactions or partial actions of the dense Hopf subalgebra.	math.OA	math	PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS FRANZISKA KRAKEN1, PAULA QUAST2, AND THOMAS TIMMERMANN3 ∗ Abstract. Partial actions of groups on C ∗-algebras and the closely related actions and coactions of Hopf algebras received much attention over the last decades. They arise naturally as restrictions of their global counterparts to non-invariant subalge- bras, and the ambient eveloping global (co)actions have proven useful for the study of associated crossed products. In this article, we introduce the partial coactions of C ∗-bialgebras, focussing on C ∗-quantum groups, and prove existence of an enveloping global coaction under mild technical assumptions. We also show that partial coactions of the function algebra of a discrete group correspond to partial actions on direct sum- mands of a C ∗-algebra, and relate partial coactions of a compact or its dual discrete C ∗-quantum group to partial coactions or partial actions of the dense Hopf subalge- bra. As a fundamental example, we associate to every discrete C ∗-quantum group a quantum Bernoulli shift. 8 1 0 2 b e F 8 2 ] . A O h t a m [ 3 v 6 4 5 6 0 . 3 0 7 1 : v i X r a 1. Introduction Partial actions of groups on spaces and on C ∗-algebras were gradually introduced in [14], [15], [21], with more recent study of associated crossed products shedding new light on the inner structure of many interesting C ∗-algebras; see [16] for a comprehensive introduction and an overview. In the purely algebraic setting, the corresponding notion of a partial action or a partial coaction of a Hopf algebra on an algebra was introduced in [12]. Naturally, such partial (co)actions arise by restricting global (co)actions to non-invariant subspaces or ideals, and in these cases, all the tools that are available for the study of global situation can be applied to the study of the partial one. Therefore, it is highly desirable to know, given a partial group action or a partial Hopf algebra (co)action, whether it can be identified with some restriction of a global one, whether there exists a minimal global one -- called a globalization -- and whether the latter, if it exists, can be constructed explicitly. For partial actions of groups on locally compact Hausdorff spaces, such a globalization can always be constructed, but the underlying space need no longer be Hausdorff [1], [2]. As a consequence, partial actions of groups on C ∗-algebras can not always be identified with the restriction of a global action [2]. In the purely algebraic setting, partial (co)actions of Hopf algebras always have a globalization [5], [6]; see also [3], [4], [13]. Date: October 9, 2018 ∗Corresponding author. 2010 Mathematics Subject Classification. Primary 46L55; Secondary 16T20. Key words and phrases. C ∗-algebra, partial action, quantum group, Hopf algebra, globalization. Partially supported by the SFB 878 "Groups, geometry and actions" funded by the DFG and by the grant H2020-MSCA-RISE-2015-691246-QUANTUM DYNAMICS. 1 2 F. KRAKEN, P. QUAST, AND T. TIMMERMANN In this article, we introduce partial coactions of C ∗-bialgebras, in particular, of C ∗- quantum groups, on C ∗-algebras, and relate them to the partial (co)actions discussed above. In case of the function algebra of a discrete group, partial coactions correspond to partial actions of groups where for every group element, the associated domain of definition is a direct summand of the total C ∗-algebra, and these are precisely the partial If the C ∗-bialgebra is a actions for which existence of a globalization can be proven. discrete C ∗-quantum group, then every partial coaction gives rise to a partial action of the Hopf algebra of matrix coefficients of the dual compact quantum group. Finally, in case of a compact C ∗-quantum group, partial coactions restrict, under a natural condition, to partial coactions of the Hopf algebra of matrix coefficients on a dense subalgebra. Partial coactions appear naturally as restrictions of ordinary coactions to ideals or, more generally, to C ∗-subalgebras that are weakly invariant in a suitable sense. An identification of a partial coaction with such a restriction will be called a dilation of the partial coaction. The main result of this article is the existence and a construction of a minimal dilation, also called a globalization, under mild assumptions. We follow the approach for coactions of Hopf algebras [6], but face new technical difficulties. To deal with these, we assume that the C ∗-algebra of the quantum group under consideration has the slice map property, which follows, for example, from nuclearity [31], and is automatic if the quantum group is discrete. Briefly, the main result can be summarised as follows. Theorem. Let (A, ∆) be a C ∗-quantum group, where A has the slice map property. Then every injective, weakly continuous, regular partial coaction of (A, ∆) has a minimal dilation and the latter is unique up tio isomorphism. Presently, we do not see whether this slice map assumption is just convenient or genuinely necessary. Parts of the results in this article were obtained in the Master's theses of the first and the second author. In following articles, we plan to study crossed products for partial coactions, and partial corepresentations of C ∗-bialgebras. The article is organized as follows. In Section 2, we recall background on C ∗-quantum groups, strict ∗-homomorphisms and the slice map property. In Section 3, we introduce partial coactions of C ∗-bialgebras and discuss a few desirable properties like weak and strong continuity. In Section 4, we show that partial actions of a discrete group Γ on a C ∗- algebra correspond to counital partial coactions of the function algebra C0(Γ) if and only if the domains of definition are direct summands of the C ∗-algebra. In Section 5, we relate partial coactions of compact and of discrete C ∗-quantum groups to coactions and actions of the Hopf algebra of matrix elements of the compact quantum group. In Section 6, we show how partial coactions arise from global ones by restriction, and discuss the closely related notion of weak or strong morphisms between partial coactions. In Section 7, we construct for every discrete quantum group a quantum a quantum Bernoulli shift and obtain, by restriction, a partial coaction that is initial in a suitable sense. In Section 8, we consider the situation where a partial coaction can be identified with the restriction of a global coaction, and study a few preliminary properties of such identifications. Finally, in Section 9, we prove the main result stated above. PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 3 Let us fix some notation and recall some background. 2. Preliminaries Conventions and notation. Given a locally compact Hausdorff space X, we denote by Cb(X) and C0(X) the C ∗-algebra of continuous functions that are bounded or vanish at infinity, respectively. For a subset F of a normed space E, we denote by [F ] ⊆ E its closed linear span. Given a C ∗-algebra A, we denote by A∗ the space of bounded linear functionals on A, by M (A) the multiplier algebra and by 1A ∈ M (A) the unit of M (A). Given a Hilbert space K, we denote by 1K the identity on H. Let A and B be C ∗-algebras. A ∗-homomorphism ϕ : A → M (B) is called nonde- generate if [ϕ(A)B] = B. Each nondegenerate ∗-homomorphism ϕ : A → M (B) extends uniquely to a unital ∗-homomorphism from M (A) to M (B), which we denote by φ again. By a representation of a C ∗-algebra A on a Hilbert space H we mean a ∗-homomorphism π : A → B(H). All tensor products of C ∗-algebras will be minimal ones. We write σ for the tensor flip isomorphism A ⊗ B → B ⊗ A, a ⊗ b 7→ b ⊗ a. C ∗-bialgebras and C ∗-quantum groups. A C ∗-bialgebra is a C ∗-algebra A with a non-degenerate ∗-homomorphism ∆ : A → M (A ⊗ A), called the comultiplication, that is coassociative in the sense that (∆ ⊗ idA) ◦ ∆ = (idA ⊗∆) ◦ ∆. It satisfies the cancellation conditions if (2.1) Given a C ∗-bialgebra (A, ∆), the dual space A∗ is an algebra with respect to the [∆(A)(1A ⊗ A)] = A ⊗ A = [(A ⊗ 1A)∆(A)]. convolution product defined by υω := (υ ⊗ ω) ◦ ∆. A counit for a C ∗-bialgebra (A, ∆) is a character ε on A satisfying (ε ⊗ idA) ◦ ∆ = idA = (idA ⊗ε) ◦ ∆. If it exists, such a counit is a unit in the algebra A∗ and thus unique. A morphism of C ∗-bialgebras (A, ∆A) and (B, ∆B) is a non-degenerate ∗-homomorphism f : A → M (B) satisfying ∆B ◦ f = (f ⊗ f ) ◦ ∆A. A C ∗-quantum group is a C ∗-bialgebra that arises from a well-behaved multiplicative unitary as follows [26, 27, 32]. Suppose that H is a Hilbert space and that W ∈ B(H ⊗H) is a multiplicative unitary [8] that is manageable or modular [32, 27]. Then the spaces A := [(ω ⊗ idH)W : ω ∈ B(H)∗] and A := [(idH ⊗ω)W : ω ∈ B(H)∗] are separable, nondegenerate C ∗-subalgebras of B(H), the unitary W is a multiplier of A ⊗ A ⊆ B(H ⊗ H), and the formulas ∆(a) = W (a ⊗ 1H)W ∗, ∆(a) = σ(W ∗(1H ⊗ a)W ) (2.2) define comultiplications on A and A, respectively, such that (A, ∆) and ( A, ∆) become C ∗-bialgebras. A C ∗-bialgebra (A, ∆) is a C ∗-quantum group if it arises from a modular multiplicative unitary W as above. Let (A, ∆) be a C ∗-quantum group arising from a unitary W as above. Denote by Σ multiplicative unitary and the associated C ∗-quantum group is ( A, ∆). The latter only depends on (A, ∆) and not on the choice of W , and is called the dual of (A, ∆). The the flip on H ⊗ H. Then also the dual cW := ΣW ∗Σ of W is a modular or manageable images of W and cW in M ( A⊗ A) or M (A⊗ A), respectively, do not depend on the choice 4 F. KRAKEN, P. QUAST, AND T. TIMMERMANN of W but only on (A, ∆). We call them the reduced bicharacters of (A, ∆) and ( A, ∆) and denote them by W A and cW A, respectively. We will need an anti-Heisenberg pair for (A, ∆), which consists of non-degenerate, faithful representations π of A and π of A on a Hilbert space K such that the unitary regarded as an element of M (A ⊗ K(K)), satisfies V := (idA ⊗π)(cW A) ∈ M (A ⊗ π( A)), V (1A ⊗ π(a))V ∗ = (idA ⊗π)∆(a) for all a ∈ A; (2.3) (2.4) see [22, §3] and [24, §3.1] . Every locally compact quantum group or, more precisely, every reduced C ∗-algebraic quantum group in the sense of Kustermans and Vaes [18], is a C ∗-quantum group. We shall use regularity of C ∗-quantum groups, which was studied for multiplicative unitaries in [8] and for reduced C ∗-algebraic quantum groups in [9, §5(b)]. We follow the approach of [25, Definition 5.37] and call a C ∗-quantum group (A, ∆) regular if its reduced bicharacter satisfies [( A ⊗ 1A)W A(1 A ⊗ A)] = A ⊗ A in M ( A ⊗ A). This is equivalent to the condition [(1 A ⊗ A)W A( A ⊗ 1A)] = A ⊗ A, see [25, proof of Corollary 5.39]. For the unitary (2.3), this translates into [(1A ⊗ π( A))V (A ⊗ 1π( A))] = A ⊗ π( A) in M (A ⊗ π( A)). (2.5) In [25], this condition is referred to as weak regularity. However, every reduced C ∗- algebraic quantum (A, ∆) is regular in the sense above if and only if it is regular in the sense of [9, §5(b)]. One implication is contained in [8, Proposition 3.6], and the other follows easily from [9, Proposition 5.6]. A compact C ∗-quantum group is, by definition, a unital C ∗-bialgebra G = (A, ∆) that satisfies the cancellation conditions, and is indeed a weakly regular C ∗-quantum group [33]. Associated to such a compact quantum group is a rigid C ∗-tensor category of unitary finite-dimensional corepresentations [23]. We denote by Irr(G) the equivalence classes of irreducible corepresentations. Their matrix elements span a dense Hopf subalgebra O(G). The dual ( A, ∆) is called a discrete C ∗-quantum group, and the underlying C ∗-algebra A is a direct sum of matrix algebras, indexed by Irr(G). We also denote the underlying C ∗-algebra A of G by C0( G). Strict ∗-homomorphisms of C ∗-algebras. Recall from [19, §5, Corollary 5.7] that a ∗-homomorphism π : B → M (C) is strict if it is strictly continuous on the unit ball, and that in that case, it extends to a ∗-homomorphism M (B) → M (C) that is strictly continuous on the unit ball. We denote this extension by π again. Using this extension, we define the composition of strict ∗-homomorphisms, which evidently is strict again. Hence, C ∗-algebras with strict ∗-homomorphisms form a category. Recall that a corner of a C ∗-algebra B is a C ∗-subalgebra of the form pBp for some projection p ∈ M (B). Strict ∗-homomorphisms are just non-degenerate ∗-homomorphisms in the usual sense Indeed, if π : B → M (C) is a strict ∗- from the domain to a corner of the target. homomorphism, then p := π(1B) ∈ M (C) is a projection, pCp ⊆ C is a corner, and the co-restriction π : B → M (pCp) is non-degenerate. Conversely, given a corner C0 ⊆ C and a non-degenerate ∗-homomorphism π : B → M (C0), we get a strict extension M (B) → PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 5 M (C0), a natural strict map M (C0) → M (C) [10, II.7.3.14], and the composition is a strict ∗-homomorphism. This description of strict ∗-homomorphisms immediately implies that the minimal tensor product of strict morphisms is a strict morphism again, and that an embedding of C ∗-algebras B ֒→ C is a strict ∗-homomorphism if and only if B is a non-degenerate C ∗-subalgebra of a corner of C. We shall call such embeddings strict. In the commutative case, partial morphisms correspond to partially defined continuous maps with clopen domain of definition. Indeed, let X and Y be locally compact Hausdorff spaces. Then every continuous map F from a clopen subset D ⊆ Y to X induces a strict ∗-homomorphism F ∗ : C0(X) → M (C0(Y )) = Cb(Y ) defined by (F ∗(f ))(y) = 0 if y 6∈ D, (F ∗(f ))(y) = f (F (y)) if y ∈ D. Conversely, if π : C0(X) → M (C0(Y )) is a strict ∗-homomorphism, then π(1X ) is the characteristic function of a clopen subset D ⊆ Y and the corestriction π : C0(X) → M (C0(D)) is the pull-back along a continuous function F : D → X. 2.1. The slice map property. In sections 8 and 9, we need the following property. A C ∗-algebra A has the slice map property if for every C ∗-algebra B and every C ∗- subalgebra C ⊆ B, every x ∈ B ⊗ A satisfying (id ⊗ω)(x) ∈ C for all ω ∈ A∗ lies in C ⊗A [31]. This property holds if A is nuclear, or, more generally, if A has the completely bounded approximation property or the strong operator approximation property; see [30] for a survey. In particular, this condition holds whenever (A, ∆) is a discrete quantum group, or, more generally, whenever (A, ∆) is a reduced C ∗-algebraic quantum group whose dual is amenable [11, Theorem 3.3]. 3. Partial coactions of C ∗-bialgebras The definition of a partial coaction given for Hopf algebras in [?] carries over to C ∗- bialgebras as follows. Definition 3.1. A partial coaction of a C ∗-bialgebra (A, ∆) on a C ∗-algebra C is a strict ∗-homomorphism δ : C → M (C ⊗ A) satisfying the following conditions: (1) δ(C)(1C ⊗ A) ⊆ C ⊗ A; (2) δ is partially coassociative in the sense that (δ ⊗ idA)δ(c) = (δ(1C ) ⊗ 1A)(idC ⊗∆)δ(c) for all c ∈ C, or, equivalently, the following diagram commutes: C δ δ M (C ⊗ A) δ⊗id M (C ⊗ A) (δ(1C )⊗1A)(idC ⊗∆)δ / M (C ⊗ A ⊗ A) (3.1) (3.2) Let δ be a partial coaction of a C ∗-bialgebra (A, ∆) on a C ∗-algebra C. For every functional ω ∈ A∗ and every multiplier T ∈ M (C), we define a multiplier ω ⊲ T := (idC ⊗ω)δ(T ) ∈ M (C), / / / 6 F. KRAKEN, P. QUAST, AND T. TIMMERMANN where we use the fact that we can write ω = aυ or ω = υ′a′ with a, a′ ∈ A and υ, υ′ ∈ A∗ by Cohen's factorization theorem. Let c ∈ C and ω ∈ A∗. Then conditions (1) and (2) in Definition 3.1 imply ω ⊲ c ∈ C and δ(ω ⊲ c) = (idC ⊗ idA ⊗ω)(δ ⊗ idA)δ(c) = δ(1C )(idC ⊗ idA ⊗ω)(idC ⊗∆)δ(c). (3.3) In particular, for every character χ ∈ A∗, χ ⊲ (ω ⊲ c) = (χ ⊲ 1C)(idC ⊗(χ ⊗ ω)∆)δ(c) = (χ ⊲ 1C)(χω ⊲ c). (3.4) The following conditions on a partial coaction are straightforward generalizations of the corresponding conditions on coactions, and will play an equally important role: Definition 3.2. We say that a partial coaction δ of a C ∗-bialgebra (A, ∆) on a C ∗- algebra C • satisfies the Podleś condition if [δ(C)(1C ⊗ A)] = [δ(1C )(C ⊗ A)]; • is weakly continuous if [A∗ ⊲ C] = C; • is counital if (A, ∆) has a counit ε and (idC ⊗ε) ◦ δ = id. Remark 3.3. If δ is a partial coaction as above and X ⊆ A∗ is a subset that separates the points of A, then a standard application of the Hahn-Banach theorem shows that [X ⊲ C] = [A∗ ⊲ C]. Every counital partial coaction evidently is weakly continuous. A coaction satisfying the Podleś condition is automatically weakly continuous, and is usually called (strongly) continuous. For partial coactions, this implication does no longer hold in general, and so we avoid this terminology. Lemma 3.4. Let δ be a partial coaction of a C ∗-bialgebra (A, ∆) on a C ∗-algebra C that satisfies the Podleś condition. Then: (1) δ is weakly continuous if and only if [(A∗ ⊲ 1C)C] = C; (2) δ is counital if and only if (A, ∆) has a counit ε and ε ⊲ 1C = 1C . Proof. (1) By assumption, the closed linear span of all elements of the form aω ⊲ c = (idC ⊗ω)(δ(c)(1C ⊗ a)), where ω ∈ A∗, a ∈ A and c ∈ C, is equal to the closed linear span of all elements of the form (idC ⊗ω)(δ(1C )(c ⊗ a)) = (aω ⊲ 1C )c. Now, use Cohen's factorization theorem. (2) If ε ⊲ 1C = 1C , then elements of the form aε ⊲ c, where a ∈ A and c ∈ C, are linearly dense in C, and for every ω ∈ A∗ and c ∈ C, (3.4) implies ε ⊲ (ω ⊲ c) = 1C · (ω ⊲ c). (cid:3) For regular reduced C ∗-algebraic quantum groups, weakly continuous coactions auto- matically satisfy the Podleś condition [9, Proposition 5.8]. More generally, we show the following: Proposition 3.5. Let (A, ∆) be a regular C ∗-quantum group. Then every weakly con- tinuous partial coaction of (A, ∆) satisfies the Podleś condition. Proof. We proceed similarly as in the proof of [9, Proposition 5.8], and use an anti- Heisenberg pair (π, π) for (A, ∆) on some Hilbert space K and the unitary V in (2.3). PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 7 Let δ be a weakly continuous partial coaction of (A, ∆) on a C ∗-algebra C. By (3.3) and Remark 3.3, [δ(C)(1C ⊗ A)] = [δ(ω ◦ π ⊲ C)(1C ⊗ A) : ω ∈ B(K)∗] = [δ(1C ) · (idC ⊗ idA ⊗ω ◦ π)((idC ⊗∆)(δ(C))) · (1C ⊗ A) : ω ∈ B(K)∗]. To shorten the notation, let δπ := (idC ⊗π) ◦ δ. We use the relations (2.4), (2.5) and [π( A)B(K)∗] = B(K)∗, and find [(idC ⊗ idA ⊗ ω ◦ π)((idC ⊗∆)(δ(C))(1C ⊗ A ⊗ 1A)) : ω ∈ B(K)∗] 23(A ⊗ π( A))23) : ω ∈ B(K)∗] = [(idC ⊗ idA ⊗ω)(V23δπ(C)13V ∗ = [(idC ⊗ idA ⊗ω)(V23δπ(C)13(A ⊗ π( A))23) : ω ∈ B(K)∗] = [(idC ⊗ idA ⊗ω)((1A ⊗ π( A))23V23(A ⊗ 1K )23δπ(C)13) : ω ∈ B(K)∗] = [(idC ⊗ idA ⊗ω)((A ⊗ π( A))23δπ(C)13) : ω ∈ B(K)∗] = [A∗ ⊲ C] ⊗ A, whence [δ(C)(1C ⊗ A)] = [δ(1C )(C ⊗ A)]. (cid:3) Partial coactions on C correspond to certain projections: Lemma 3.6. Partial coactions of a C ∗-bialgebra (A, ∆) on C correpond bijectively with projections p ∈ M (A) satisfying (3.5) Proof. Projections p ∈ M (A) correspond to strict ∗-homomorphisms δ : C → M (C⊗A) ∼= M (A) via p = δ(1), and under this correspondence, (δ ⊗ idA)δ(λ) = λ ⊗ p ⊗ p and (δ(1) ⊗ 1A)(idC ⊗∆)(δ(λ)) = λ ⊗ (p ⊗ 1A)∆(p). (cid:3) (p ⊗ 1A)∆(p) = p ⊗ p. Note that if (A, ∆) is co-commutative, for example, if A = C ∗(G) or A = C ∗ r (G) for a locally compact group G, then (3.5) just means that p is group-like in the sense that (p ⊗ 1A)∆(p) = p ⊗ p = (1A ⊗ p)∆(p). Group-like projections were also studied in connection with idempotent states, see [17, §2]. Elementary examples related to groups are as follows. Example 3.7. Let G be a locally compact group. (2) Consider the reduced group C ∗-bialgebra (C ∗ (1) Consider the C ∗-bialgebra (C0(G), ∆). A projection p ∈ M (C0(G)) is just the characteristic function of a clopen subset H ⊆ G, and satisfies (3.5) if and only if p(g)p(gg′) = p(g)p(g′) for all g, g′ ∈ G, that is, if and only if H ⊆ G is a subgroup. Thus, partial coactions of (C0(G), ∆) on C correspond to open subgroups of G. r (G), ∆). For every finite normal r (G)) satisfying (3.5), where λg denotes the left translation by g ∈ G. More information r (G) and C ∗(G) can be found in [20, Proposition on group-like projections in C ∗ 7.6] and [29]. subgroup N ⊆ G, the sum p = Pg∈N λg is a central projection in M (C ∗ Every central projection satisfying (3.5) gives rise to a quotient C ∗-bialgebra (Ap, ∆p) of (A, ∆) whose coactions can be regarded as partial coactions of (A, ∆): 8 F. KRAKEN, P. QUAST, AND T. TIMMERMANN Lemma 3.8. Suppose that (A, ∆) is a C ∗-bialgebra with a central projection p ∈ M (A) satisfying (3.5). Let Ap = pA and define ∆p : Ap → M (Ap ⊗ Ap) by a 7→ (p ⊗ p)∆(a). Then (Ap, ∆p) is a C ∗-bialgebra, the map A → Ap, a 7→ pa, is a morphism of C ∗- bialgebras, and every coaction of (Ap, ∆p) can be regarded as a partial coaction of (A, ∆). Proof. All of these assertions are easily verified, for example, if δ is a coaction of (Ap, ∆p) on a C ∗-algebra C, then for all c ∈ C, (δ(1C ) ⊗ 1A)(idC ⊗∆)δ(c) = (1C ⊗ p ⊗ 1A)(1C ⊗ ∆)((1C ⊗ p)δ(c)) = (1C ⊗ p ⊗ p)(idC ⊗∆)δ(c) = (1C ⊗ ∆p)δ(c) = (idC ⊗δ)δ(c). (cid:3) Example 3.9. Let G = (A, ∆) be a discrete quantum group, so that A is a c0-sum of matrix algebras indexed by Irr( G). Consider a central projection p ∈ M (A) supported on J ⊆ Irr( G). Then (p ⊗ 1)∆(p) = p ⊗ p if and only if the following condition holds: If α ∈ J , β, γ ∈ Irr( G) and α ⊗ β contains γ, then β ∈ J if and only if γ ∈ J . (3.6) If (Ap, ∆p) is a discrete quantum subgroup of (A, ∆), then J is closed under taking duals and summands of tensor products, and then Frobenius duality implies (3.6). Conversely, suppose that (3.6) holds. Taking γ = α, we see that J contains the trivial representation, and taking this for γ, we see that J contains the dual of α. Thus, finite sums of representations in J form a rigid tensor subcategory, and (Ap, ∆p) is a discrete quantum subgroup of (A, ∆). 4. The relation to partial actions of groups We now relate partial actions of a (discrete) group Γ to counital partial coactions of the C ∗-bialgebra C0(Γ). Recall that a partial action of Γ on a C ∗-algebra C is a family (Dg)g∈Γ of closed ideals of C together with a family (θg)g∈Γ of isomorphisms θg : Dg−1 → Dg such that (G1) De = C and θe = idC, where e ∈ Γ denotes the unit, (G2) θg−1θgθh = θg−1θgh and θgθhθh−1 = θghθh−1 for all g, h ∈ Γ as partially defined maps; see [16, 21]. We show that partial coactions of C0(Γ) correspond to partial actions of Γ as above, where each ideal Dg is a direct summand, and adopt the following terminology: Definition 4.1. A disconnected partial action of Γ on a C ∗-algebra C is given by a family (pg)g∈Γ of central projections in M (C) and a family (θg)g∈Γ of isomorphisms θg : pg−1C → pgC such that ((pgC)g∈Γ, (θg)g∈Γ, ) is a partial action. Remark 4.2. (1) Let X be a locally compact Hausdorff space. Then partial actions of Γ on C0(X) correspond bijectively to partial actions of Γ on X [16, Corollary 11.6], and a partial action on C0(X) is disconnected if and only if for every group element g ∈ Γ, the domain of definition of its action on X is not only open but also closed. This condition also implies that the partial action on X admits a globalization that is Hausdorff [16, Proposition 5.7]. (2) A partial action of Γ on an algebra C admits a globalization if and only if for every group element g ∈ Γ, its domain of definition is not just a two-sided ideal of C but also unital, that is, a direct summand [16, Theorem 6.13]. PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 9 We denote by Cb(Γ; C) the C ∗-algebra of norm-bounded C-valued functions on Γ, and identify this C ∗-algebra with a subalgebra of M (C ⊗ C0(Γ)) in the canonical way. For each g ∈ Γ, we denote by evg ∈ C0(Γ)∗ the evaluation at g. Proposition 4.3. Let Γ be a group and let C be a C ∗-algebra. (1) Let δ be a counital partial coaction of C0(Γ) on C. Then the projections pg := evg ⊲ 1C are central and the maps θg : pg−1C → pgC given by θg(c) := evg ⊲ c form a disconnected partial action of Γ on C. (2) Let ((pg)g∈Γ, (θg)g∈Γ) be a disconnected partial action of Γ on C. Then the map δ : C → Cb(Γ; C) ֒→ M (C ⊗ C0(Γ)) defined by (δ(c))(g) := θg(pg−1c) (c ∈ C, g ∈ Γ) is a counital partial coaction of C0(Γ) on C. Proof. (1) For each g ∈ Γ, the map Θg : C → C given by c 7→ evg ⊲ c is a strict endomor- phism. Since δ is counital, Θe is the identity on C. Let g, h ∈ Γ. Then by (3.4), Θg(Θh(c)) = (evg ⊲ 1C )(evgevh ⊲ c) = pgΘgh(c), (4.1) in particular, Θg(ph) = pgpgh, Θg(Θg−1(c)) = pgc, Θg−1(Θg(c)) = pg−1c. (4.2) Since Θg ◦ Θg−1 is a ∗-homomorphism, the second equation implies pgc = cpg for all c ∈ C, that is, pg is central and Dg := pgC is a direct summand of C. The second and third equations imply that Θg and Θg−1 restrict to mutually inverse isomorphisms Dg−1 θg ⇄ θg−1 Dg. It remains to show that θg−1θgh = θg−1θgθh. But the relations (4.1) and (4.2) imply that (Θg−1 ◦ Θgh)(c) = pg−1Θgh(c) = (Θg−1 ◦ Θg ◦ Θh)(c) for all c ∈ C, and that the compositions θg−1θgh and θg−1θgθh have the domain Θh−1g−1(pg)C = ph−1g−1ph−1C = Θh−1(pg−1)C. (2) For each g ∈ Γ, denote by δg ∈ C0(Γ) the characteristic function of {g} ⊂ Γ. Then δ(c)(1C ⊗ δg) = θg(pgc) ⊗ δg (g ∈ Γ, c ∈ C). We conclude that δ(C)(1C ⊗ C0(Γ)) is contained in C ⊗ C0(Γ), and that δ co-restricts to a non-degenerate ∗-homomorphism from C to q(C ⊗ C0(Γ)), where q =Pg∈Γ pg ⊗ δg, so that δ is strict. To verify that δ is partially coassociative, it suffices to check that for all g, h ∈ Γ and c ∈ C, the element (idC ⊗evg ⊗ evh)(δ ⊗ idA)δ(c) = θg(pg−1θh(ph−1c)) 10 F. KRAKEN, P. QUAST, AND T. TIMMERMANN is equal to the element (idC ⊗evg ⊗ evh)((δ(1C ) ⊗ 1A)(idC ⊗∆)δ(c)) = θg(pg−1)θgh(ph−1g−1c), and this follows easily from the definition of a partial action. (cid:3) The following example shows that the correspondence between partial coactions of C0(Γ) and partial actions of Γ does not easily extend from groups to inverse semigroups. Example 4.4. Denote by Γ the inverse semigroup consisting of the 2 × 2-matrices 0, v =(cid:18)0 0 1 0(cid:19) , v∗ =(cid:18)0 1 0 0(cid:19) , vv∗ =(cid:18)0 0 0 1(cid:19) , v∗v =(cid:18)1 0 0 0(cid:19) with matrix multiplication as composition. Then C(Γ) is a C ∗-bialgebra with respect to the transpose ∆ of the multiplication. For x ∈ Γ, define δx ∈ C(Γ) by y 7→ δx,y. Then, for example, ∆(δv∗v) = δv∗ ⊗ δv + δv∗v ⊗ δv∗v, ∆(δv) = δvv∗ ⊗ δv + δv ⊗ δv∗v. Now, the ∗-homomorphism δ : C2 → C2 ⊗ C(Γ), (α, β) 7→ (α, 0) ⊗ δv∗v + (0, α) ⊗ δv, is a partial coaction. Indeed, for all α, β ∈ C, (δ ⊗ idC(Γ))δ((α, β)) = (α, 0) ⊗ δv∗v ⊗ δv∗ v + (0, α) ⊗ δv ⊗ δv∗v is equal to the product of δ((1, 0)) ⊗ 1C(Γ) = (1, 0) ⊗ δv∗ v ⊗ 1C(Γ) + (0, 1) ⊗ δv ⊗ 1C(Γ) with (idC2 ⊗∆)δ((α, β)) = (α, 0) ⊗ (δv∗ v ⊗ δv∗ + δv∗ ⊗ δvv∗ ) + (0, α) ⊗ (δv ⊗ δv∗v + δvv∗ ⊗ δv). But the maps Θw := (id ⊗evw) ◦ δ, where w ∈ Γ, are given by Θ0 = Θv∗ = Θvv∗ = 0, Θv((α, β)) = (0, α), Θv∗v((α, β)) = (α, 0); in particular, ΘvΘv∗ Θv = 0 and ΘvΘv∗v = Θv. 5. Partial coactions of discrete and of compact C ∗-quantum groups Let G = (A, ∆) be a compact C ∗-quantum group and denote by O(G) ⊆ A the dense Hopf subalgebra of matrix elements of finite-dimensional corepresentations. We now relate partial (co)actions of G and of the discrete dual G to partial coactions and partial actions of the Hopf algebra O(G), respectively. Note that (A, ∆) and ( A, ∆) are regular, so that weakly continuous partial coactions automatically satisfy the Podleś condition by Proposition 3.5. Recall that a partial action of a Hopf algebra H on a unital algebra C is a map H ⊗ C → C, h ⊗ c 7→ h ⊲ c, satisfying the following conditions: (H1) 1H ⊲ c = c for all c ∈ C; (H2) h ⊲ (cd) = (h(1) ⊲ c)(h(2) ⊲ d) for all h ∈ H and c, d ∈ C; PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 11 (H3) h ⊲ (k ⊲ c) = (h(1) ⊲ 1C)(h(2)k ⊲ c) for all h, k ∈ H and c ∈ C; see [?], and that such a partial action is symmetric if additionally (H4) h ⊲ (k ⊲ c) = (h(1)k ⊲ c)(h(2) ⊲ 1C ) for all h, k ∈ H and c ∈ C; see [7]. If additionally h ⊲ 1C = ε(h) for all h ∈ H, we have a genuine action; in that case, (H3) and (H4) reduce to h ⊲ (k ⊲ c) = hk ⊲ c. Recall that the C ∗-algebra A of the discrete C ∗-quantum group G is a c0-direct sum of matrix algebras Aα indexed by α ∈ Irr(G). The Hopf algebra O(G) can be identified with the subspace of all functionals ω ∈ A∗ that vanish on Aα for all but finitely many α ∈ Irr(G), and then ∆(ω)(a ⊗ b) = ω(ab) and (υω)(a) = (υ ⊗ ω)(a) for all υ, ω ∈ O(G) and a, b ∈ A. Theorem 5.1. Let G = (A, ∆) be a compact quantum group and let δ be a counital partial coaction of the discrete dual G = ( A, ∆) on a unital C ∗-algebra C. Then the formula υ ⊗ c 7→ υ ⊲ c = (idC ⊗υ)(δ(c)) (υ ∈ O(G), c ∈ C) defines a symmetric partial action of the Hopf algebra O(G) on C. Proof. Condition (H1) holds because the unit of O(G), regarded as a functional on A, is the counit. Let υ, ω ∈ O(G) and c, d ∈ C. Choose central projections p, q ∈ A such that υ(pa) = υ(a), ω(a) = ω(qa) and υ(1)(a)υ(2)(b) = υ(1)(pa)υ(2)(pb) for all a, b ∈ A. Then υ ⊲ cd = (idC ⊗υ)((1C ⊗ p)δ(c)δ(d)(1C ⊗ p)). Since (1C ⊗ p)δ(c) and δ(d)(1C ⊗ p) are contained in the tensor product of C with the finite-dimensional C ∗-algebra p A + q A, this expression is equal to (id ⊗υ(1))((1 ⊗ p)δ(c)) · (id ⊗υ(2))(δ(d)(1 ⊗ p)) = (υ(1) ⊲ c)(υ(2) ⊲ d). Thus, condition (H2) is satisfied. Likewise, υ ⊲ (ω ⊲ c) = (idC ⊗υ ⊗ ω)((δ ⊗ id A)δ(c)) = (idC ⊗υ ⊗ ω)((1C ⊗ p ⊗ q)(δ(1C ) ⊗ 1 A)(idC ⊗ ∆)(δ(c))), and a similar argument as above shows that this expression is equal to (idC ⊗υ(1))(δ(1C ))(idC ⊗(υ(2) ⊗ ω) ◦ ∆)(δ(c)) = (υ(1) ⊲ 1C ) · (υ(2)ω ⊲ c). Therefore, condition (H3) holds as well, and a similar argument proves (H4). (cid:3) Next, we consider partial coactions of the compact C ∗-quantum group (A, ∆), and relate them to partial coactions of the Hopf algebra O(G). Recall that a partial coaction of a Hopf algebra H on a unital algebra C is a homomorphism δ : C 7→ C ⊗ H satisfying the following conditions, (CH1) (δ ⊗ idH)(δ(c)) = (δ(1C ) ⊗ 1H ) · (idC ⊗∆H)(δ(c)) for all c ∈ C, and (CH2) (idC ⊗εH)(δ0(c)) = c for all c ∈ C; 12 F. KRAKEN, P. QUAST, AND T. TIMMERMANN see [12]. Theorem 5.2. Let δ be a partial coaction of a compact C ∗-quantum group G = (A, ∆) on a unital C ∗-algebra C. Then the following conditions are equivalent: (1) δ is weakly continuous, δ(1C ) lies in the algebraic tensor product C ⊗ O(G) and (idC ⊗ε)(δ(1C )) = 1C, where ε denotes the counit of O(G). (2) δ restricts to a partial coaction of O(G) on a unital dense ∗-subalgebra C0 of C. Proof. Denote by O( G) ⊆ A the algebraic direct sum of the matrix algebras Aα associated to all α ∈ Irr(G), and recall that we can canonically identify O(G) with a subspace of A∗. (1)⇒(2): By Remark 3.3, the subspace C0 = O( G) ⊲ C of C is dense. We show that C0 ⊆ C is a subalgebra. Let c, d ∈ C and υ, ω ∈ O(G). Then (υ ⊲ c)(ω ⊲ d) = (idC ⊗υ ⊗ ω)(δ(c)12δ(d)13), where we use the leg notation on δ(c) and δ(d). Now, we find finitely many υ′ such that i, ω′ i ∈ O( G) υ(a)ω(b) =Xi (υ′ i ⊗ ω′ i)((a ⊗ 1A)∆(b)) for all a, b ∈ A, and then (υ ⊲ c)(ω ⊲ d) =Xi =Xi (idC ⊗υ′ i ⊗ ω′ i)((δ(c) ⊗ 1A)(id ⊗∆)(δ(d))) (idC ⊗υ′ i ⊗ ω′ i)(δ ⊗ idA)((c ⊗ 1A)δ(d)) =Xi υ′ i ⊲ (c(ω′ i ⊲ d)) ∈ C0. Next, we show that δ(C0) is contained in the algebraic tensor product C ⊗ O(G). i,j)α,i,j satisfying kj [33, Proposition 5.1], we can find finitely many υ1, . . . , υn ∈ O( G) Let ω ∈ O( G) and c ∈ C. Since O(G) has a basis of elements (uα ∆(uα and a1, . . . , an ∈ O(G) such that i,j) =Pk uα ik ⊗ uα (idA ⊗ω)(∆(b)) = nXi=1 υi(b)ai for all b ∈ O(G), and then δ(ω ⊲ c) = (idC ⊗ idA ⊗ω)(δ ⊗ idA)δ(c) = δ(1C )(idC ⊗(idA ⊗ω)∆)δ(c) = δ(1C ) · nXi=1 (υi ⊲ c) ⊗ ai lies in the algebraic tensor product of C with O(G)). Using a basis for O( G) consisting of functionals (φα k,l) = δα,βδi,kδj,l, see [33, §6], we see that δ(C0) is contained in the algebraic tensor product C0 ⊗ O(G). i,j)α,i,j such that φα i,j(uβ To finish the proof, note that with ω, c as above, (3.4) implies ε ⊲ (ω ⊲ c) = (idC ⊗ε)(δ(1C )) · (ω ⊲ c) = ω ⊲ c. (2)⇒(1): Since C0 ⊆ C is dense, the unit of C0 has to be 1C , whence δ(1C ) lies in the algebraic tensor product C ⊗ O(G) and (id ⊗ε)δ(1C ) = 1C. To prove weak continuity, PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 13 c ∈ C0 and write δ(c) =Pn we show that for every c ∈ C0, there exists some ω ∈ A∗ such that ω ⊲ c = c. So, take i=1 di ⊗ ai with di ∈ C0 and ai ∈ O(G). By Hahn-Banach, the restriction of ε to the finite-dimensional subspace of A spanned by a1, . . . , an extends to a bounded linear functional ω ∈ A∗ that satisfies ω ⊲ c = ε ⊲ c = c. (cid:3) 6. Restriction Like partial actions of groups and partial (co)actions of Hopf algebras, partial coactions of C ∗-bialgebras can be obtained from non-partial ones by restriction. Definition 6.1. Let δB be a partial coaction of a C ∗-bialgebra (A, ∆) on a C ∗-algebra B. We call a C ∗-subalgebra C ⊆ B weakly invariant if δB(C)(C ⊗ A) ⊆ C ⊗ A, and strongly invariant if the embedding C ֒→ B strict and δB(C) ⊆ M (C ⊗ A) ⊆ M (B ⊗ A). Note here that if the embedding C ֒→ B is strict, then the embedding C ⊗ A ֒→ B ⊗ A is strict as well and extends to an embedding M (C ⊗ A) ֒→ M (B ⊗ A). Remark 6.2. (1) Every ideal C ⊆ B is weakly invariant, but not necessarily strongly invariant. (2) A corner C ⊆ B is strongly invariant if and only if 1C ∈ M (C) ⊆ M (B) is strongly invariant in the sense that δB(1C) = δB(1C )(1C ⊗ 1A), as one can easily check. If one thinks of elements of M (B) and M (B ⊗A) as 2×2- matrices with respect to the Peirce decomposition B = 1C B + (1B − 1C )B, then strong invariance of C means that δB(C) is contained in the upper left corner, while weak invariance of C means that the off-diagonal part of δB(C) vanishes. Example 6.3. Suppose that δB is the partial coaction corresponding to a disconnected partial action ((pg)g∈Γ, (θg)g∈Γ) of a discrete group Γ on a C ∗-algebra B as in Proposition 4.3, and that C ⊆ B is a direct summand. Then C is automatically weakly invariant, but strongly invariant if and only if θg(pg−1C) ⊆ C for all g ∈ Γ. Evidently, partial coactions can be restricted to strongly invariant C ∗-subalgebras. Re- striction to weakly invariant C ∗-subalgebras is a bit more delicate unless the embedding of the C ∗-subalgebra is strict. Proposition 6.4. Let δB be a partial coaction of a C ∗-bialgebra (A, ∆) on a C ∗-algebra B and let C ⊆ B be a weakly invariant C ∗-subalgebra. Then: (1) δB restricts to a ∗-homomorphism δC : C → M (C ⊗ A). (2) If the embedding C ֒→ B is strict, then the composition of δC with the embedding of M (C ⊗ A) into M (B ⊗ A) is strict and δC (c) = δB(c)(1C ⊗ 1A) (c ∈ C). (3) If δC is strict, then it is a partial coaction of (A, ∆) on C. 14 F. KRAKEN, P. QUAST, AND T. TIMMERMANN Proof. (1) This follows immediately from the definition. (2) Suppose that the embedding C ֒→ B is strict. Then so is its composition with δB and hence also δC . To prove the formula given for δC(c), choose a bounded approximate unit (uν)ν for C, and note that δC(c)(uν ⊗ 1A) = δB(c)(uν ⊗ 1A) converges strictly to δC(c) in M (C ⊗ A) and to δB(c)(1C ⊗ 1A) in M (B ⊗ A). (3) Let (uν )ν be as above and let c, c′ ∈ C. Then by definition of δC, (c′ ⊗ 1A ⊗ 1A)·(δC ⊗ idA)(δC (c)(uν ⊗ 1A)) = (c′ ⊗ 1A ⊗ 1A) · (δC ⊗ idA)(δB(c)(uν ⊗ 1A)) = (c′ ⊗ 1A ⊗ 1A) · (δB ⊗ idA)(δB(c)) · (δC (uν ) ⊗ 1A) = (idC ⊗∆)((c′ ⊗ 1A)δB(c)) · (δC (uν) ⊗ 1A) = (c′ ⊗ 1A ⊗ 1A) · (idC ⊗∆)(δC(c)) · (δC (uν ) ⊗ 1A). Since c′ ∈ C was arbitrary, we can conclude that (δC ⊗ idA)(δC (c)(uν ⊗ 1A)) = (idC ⊗∆)(δC (c)) · (δC (uν) ⊗ 1A). As ν tends to infinity, δC (c)(uν ⊗ 1A) converges strictly to δC(c), and since δC and hence also δC ⊗ idA are strict, the left hand side converges to (δC ⊗ idA)δC(c) and the right hand side converges to (idC ⊗∆)(δC (c))(δC (1C ) ⊗ 1A). (cid:3) Remark 6.5. (1) As a corollary, a (partial) coaction on a C ∗-algebra C restricts to a partial coaction on every direct summand of C because every direct summand is weakly invariant by Remark 6.2 (1). (2) The restriction δC can be strict without the embedding C ֒→ B being strict, for example, this is the case if δB is the trivial coaction b 7→ b ⊗ 1A and C ⊆ B is a closed ideal that is not a direct summand. Example 6.6. Let G = (A, ∆) be a discrete quantum group, so that A is a c0-sum of matrix algebras Aα with α ∈ Irr( G). Then for every subset J ⊆ Irr( G), the restriction of ∆ to the c0-sum AJ := Lα∈J Aα yields a partial coaction. But if J is non-trivial, if α 6∈ J and γ ∈ J , then α ⊗ (α† ⊗ γ), where α† then AJ is not strongly invariant: denotes the dual of α, contains γ, and hence ∆(Aγ)(Aα ⊗ 1) 6= 0. Closely related to the concept of restriction is the notion of a morphism of partial coactions. Definition 6.7. Let δB and δC be partial coactions of a C ∗-bialgebra (A, ∆) on C ∗- algebras B and C, respectively. A strong morphism from δC to δB is a strict ∗-homo- morphism π : C → M (B) satisfying (π ⊗ idA)δC (c) = δB(π(c)) (c ∈ C). A weak morphism from δC to δB is a ∗-homomorphism π : C → M (B) satisfying (π ⊗ idA)(δC (c)(c′ ⊗ a)) = δB(π(c))(π(c′) ⊗ a) (c, c′ ∈ C, a ∈ A). We call such a weak or strong morphism π proper if π(C) ⊆ B. Evidently, partial coactions with strong morphisms or with proper weak morphisms as above form categories. PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 15 Remark 6.8. (1) Clearly, π is a strong or a weak morphism if and only if π(ω ⊲ c) = ω ⊲ π(c) or π(ω ⊲ c)π(c′) = (ω ⊲ π(c))π(c′) (6.1) respectively, for all ω ∈ A∗ and c, c′ ∈ C. (2) If π is a weak or a strong morphism and proper, then its image is weakly or strongly invariant, respectively. (3) Suppose that δB is a partial coaction of (A, ∆) on a C ∗-algebra B and that C ⊆ B is a C ∗-subalgebra that is weakly or strongly invariant. If the embedding C ֒→ B is strict, then this embedding is a weak or a strong morphism with respect to the restriction of δB to C defined above. Let us look at the special case of partial coactions associated to disconnected partial group actions. Proposition 6.9. Let B and C be two C ∗-algebras with disconnected partial actions ((pg)g, (βg)g) and ((qg)g, (γg)g), respectively, of a discrete group Γ. With respect to the associated partial coactions of C0(Γ), a strict ∗-homomorphism π : B → M (C) is a strong morphism if and only if π(pg) = qgπ(1C ) and π ◦ βg ⊆ γg ◦ π for all g ∈ Γ, (6.2) and a weak morphism if and only if π(1C )γg(qg−1π(1C )) = π(pg) = γg(π(pg−1 )) and π ◦ βg ⊆ γg ◦ π for all g ∈ Γ. (6.3) Proof. Denote the partial coactions by δB and δC . (1) Suppose that π is a strong morphism. Then the definition of δB and δC implies (π ◦ βg)(pg−1b) = (π ⊗ evg)δB(b) = (idC ⊗evg)δC (π(b)) = γg(qg−1π(b)) (6.4) for all g ∈ Γ and b ∈ B. Taking b = 1C or b = pg−1, we conclude that γg(qg−1π(pg−1)) = π(pg) = γg(qg−1π(1C )), in particular, π(pg)qg = π(pg). We use this relation on the left hand side above, apply γg−1, and get π(pg) = qgπ(1C ). Moreover, π(pgB) ⊆ qgC, and (6.4) implies π◦βg ⊆ γg ◦π. Conversely, the first relation in (6.2) implies qg−1π(1C − pg−1) = 0, whence both sides in (6.4) are zero for all b ∈ (1 − pg−1)B, and the second relation in (6.2) implies that (6.4) holds for all b ∈ pg−1B. Combined, (6.2) implies (π ⊗ id)δB = δC ◦ π. (2) Suppose that π is a strict weak morphism. As in (1), we find that (π ◦ βg)(pg−1b) = π(1C )γg(qg−1π(b)) (6.5) for all g ∈ Γ and b ∈ B, and similar arguments as in (1) yield the first equation in (6.3). Now, we apply γg−1 to this relation and find that γg−1(π(pg)) = γg−1(qgπ(1c))π(1C ) = π(pg−1). In particular, this relation and (6.5) imply the second relation in (6.3). Conversely, (6.3) implies that both sides of (6.5) coincide for all b ∈ pg−1B, and that for all b ∈ (1C − pg−1)B, π(1C )γg(qg−1π(1C − pg−1)) = π(pg) − π(pg) = 0, whence both sides of (6.5) are zero for all b ∈ (1C − pg−1)B. But this implies that (π ⊗ id) ◦ δB = (π(1C ) ⊗ 1A)(δC ◦ π). (cid:3) 16 F. KRAKEN, P. QUAST, AND T. TIMMERMANN 7. The Bernoulli shift of a discrete quantum group The Bernoulli shift of a discrete group Γ is its action on the power set P(Γ), which we identify with the infinite product {0, 1}Γ, by left translation. Restriction to the subsets containing the unit eΓ yields an important example of a partial action. To a discrete quantum group, we now associate a quantum Bernoulli shift and obtain, by restriction, a partial coaction that is initial in a natural sense. The space {0, 1}Γ ∼= P(Γ) parametrizes all maps from Γ to {0, 1} or, equivalently, all subsets of Γ, which correspond to projections in M (C0(Γ)). Given a discrete quantum group G, it is natural to define its quantum power set as a universal quantum family of maps from G to {0, 1} in the sense of [28] or, equivalently, as the unital C ∗-algebra C that comes with a universal projection in M (C ⊗ C0(G)). However, we need an additional commutativity assumption. Let G = (C0(G), ∆) be a discrete C ∗-quantum group with counit ε and compact dual G. Definition 7.1. Let C be a C ∗-algebra. We call a projection p ∈ M (C ⊗ C0(G)) admissible if in M (C ⊗ C0(G) ⊗ C0(G)), (p ⊗ 1) · (id ⊗∆)(p) = (id ⊗∆)(p) · (p ⊗ 1). (7.1) Remark 7.2. For every partial coaction δ of C0(G) on a C ∗-algebra C, the projection δ(1C ) ∈ M (C ⊗ C0(G)) is admissible. Proposition 7.3. Let G be a discrete C ∗-quantum group. Then there exists a unital C ∗- algebra C(BG) with an admissible projection p ∈ M (C(BG) ⊗ C0(G)) that is universal in the following sense: for every C ∗-algebra C with an admissible projection q ∈ M (C(BG)⊗ C0(G)), there exists a unique unital ∗-homomorphism π : C(BG) → M (C) such that q = (π ⊗ id)(p). Proof. Write C0(G) ∼=Lα Iα, where α varies in Irr( G) and each Iα is a matrix algebra. ij)i,j for each Iα. Denote by C(BG) the universal unital C ∗-algebra Choose matrix units (eα with generators 1 and (pα ij)α,i,j satisfying the following relations: (1) the finite sum pα :=Pi,j pα ij ⊗ eα ij is a projection for every α ∈ Irr( G); (2) (pα ⊗ 1)(id ⊗∆)(pβ) = (id ⊗∆)(pβ)(pα ⊗ 1) for all α, β ∈ Irr( G). Then the sum p =Pα pα ∈ M (C(BG)⊗C0(G)) converges strictly because each summand pα lies in a different summand of C(BG) ⊗ C0(G) ∼=Lα(C(BG) ⊗ Iα) and has norm at most 1. By (1) and (2), this p is an admissible projection, and by construction, C(BG) has the desired universal property by construction. (cid:3) We denote by C0(B× G) ⊂ C(BG) the non-unital C ∗-subalgebra generated by all pα i,j. Example 7.4. If G is a classical discrete group Γ, we can identify C(BΓ) with C({0, 1}Γ). Indeed, in that case, Irr(Γ) can be identified with Γ so that C(BΓ) is generated by 1 and a family of projections pγ, where γ ∈ Γ. Denote by δγ ∈ C0(Γ) the Dirac delta function PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 17 at γ ∈ Γ. Then p =Pγ pγ ⊗ δγ and the admissibility condition takes the form  Xγ pγ ⊗ δγ ⊗ 1,Xγ,γ ′ pγγ ′ ⊗ δγ ⊗ δγ ′  = 0 or, equivalently, [pγ, pγ ′′ ] = 0 for all γ, γ′′ ∈ Γ. Thus, C(BΓ) is commutative. Therefore, the map that sends pγ to the projection of {0, 1}Γ onto the γth component induces an isomorphism C(BΓ) ∼= C({0, 1}Γ). Under this isomorphism, the C ∗-subalgebra C0(B× G) corresponds to C0(P(Γ) \ {∅}). The quantum space BG comes with a natural action of G: Proposition 7.5. There exists a unique coaction δ of C0(G) on C(BG) such that (δ ⊗ id)(p) = (id ⊗∆)(p). (7.2) This coaction is counital and restricts to a coaction on C0(B× G). Proof. The projection q := (id ⊗∆)(p) ∈ M ((C(BG) ⊗ C0(G)) ⊗ C0(G)) is admissible because (id ⊗ id ⊗∆)(q) = (id ⊗∆(2))(p) = (id ⊗∆ ⊗ id)(id ⊗∆)(p) commutes with q ⊗ 1 = (id ⊗∆ ⊗ id)(p ⊗ 1). The universal property of p yields a unital ∗- homomorphism δ : C(BG) → M (C(BG)⊗ C0(G)) such that (δ ⊗ id)(p) = q = (id ⊗∆)(p). We have (δ ⊗ id)δ = (id ⊗∆)δ because by definition of δ, ((δ ⊗ id)δ ⊗ id)(p) = (δ ⊗ id ⊗ id)(id ⊗∆)(p) = (id ⊗ id ⊗∆)(δ ⊗ id)(p) = (id ⊗∆ ⊗ id)(id ⊗∆)(p) = ((id ⊗∆)δ ⊗ id)(p). Next, ((id ⊗ε)δ ⊗ id)(p) = (id ⊗(ε ⊗ id)∆)(p) = p and hence (id ⊗ε)δ = id. Finally, (7.2) implies that δ(pα ij)(1 ⊗ C0(G)) ⊆ C0(B× G) ⊗ C0(G). (cid:3) We shall restrict the coaction δ to the direct summand of C(BG) that is given by the following projection: Lemma 7.6. The projection pε := (id ⊗ε)(p) ∈ C(BG) is central and δ(pε) = p. Proof. We apply id ⊗ε ⊗ id to (7.1) and obtain (pε ⊗ 1)p = p(pε ⊗ 1). Thus, pε commutes with (id ⊗ω)(p) ∈ C(BG) for every ω ∈ C0(G)∗ and hence with C(BG). Moreover, δ(pε) = (δ ⊗ ε)(p) = (id ⊗ id ⊗ε)(δ ⊗ id)(p) = (id ⊗ id ⊗ε)(id ⊗∆)(p) = p. (cid:3) Example 7.7. If G is a classical group Γ (see Example 7.4), ε is evaluation at the unit eΓ and pε = peΓ. Therefore, restriction of the coaction δ above to the direct summand pεC(BG) of C(BG) corresponds to restriction of the Bernoulli shift on P(Γ) to the subsets containing the unit eΓ. We can now define the quantum analogue of the partial Bernoulli shift: 18 F. KRAKEN, P. QUAST, AND T. TIMMERMANN Definition 7.8. Let G be a discrete C ∗-quantum group and write C(Bε G) for the direct summand pεC(BG) of C(BG). Then the partial Bernoulli action of G is the partial coac- tion δε of C0(G) on C(Bε G) obtained as the restriction of the coaction δ as in Proposition 6.4, that is, δε(b) = δ(b)(pε ⊗ 1) for all b ∈ C(Bε G). Proposition 7.3 immediately implies: Corollary 7.9. Let C be a C ∗-algebra and let q ∈ M (C ⊗ C0(G)) be an admissible projection such that (id ⊗ε)(q) = 1C ∈ M (C). Then there exists a unique unital ∗- homomorphism π : C(Bε G) → M (C) such that q = (π ⊗ id)(p). The partial Bernoulli action is initial in the following sense: Proposition 7.10. Let δC be a counital partial coaction of C0(G) on a C ∗-algebra C. Then there exists a unique unital ∗-homomorphism π : C(Bε G) → M (C) such that (π ⊗ id)(p(pǫ ⊗ 1)) = δC (1C ), (7.3) and this π is a strong morphism of partial coactions, that is, (π ⊗ id) ◦ δε = δC ◦ π. Proof. The projection δC (1C ) ∈ M (C ⊗ C0(G)) is admissible and δC is counital. Hence, Corollary 7.9 yields a unique unital ∗-homomorphism π : C(Bε G) → M (C) such that (π ⊗ id)(p) = δC(1C ). We show that (π ⊗ id) ◦ δε = δC ◦ π. First, (7.2) and Lemma 7.6 imply (δε ⊗ id)(p(pε ⊗ 1)) = (δ ⊗ id)(p(pǫ ⊗ 1)) · (pε ⊗ 1) = (id ⊗∆)(p) · (p ⊗ 1) · (pε ⊗ 1 ⊗ 1). We apply π ⊗ id ⊗ id, use (7.3), and find that ((π ⊗ id)δε ⊗ id)(p(pε ⊗ 1)) = (π ⊗ ∆)(p(pε ⊗ 1)) · (π ⊗ id)(p(pε ⊗ 1)) = (id ⊗∆)(δC (1C)) · (δC (1C ) ⊗ 1) = (δC ⊗ id)δC (1C ) = (δC ◦ π ⊗ id)(p(pε ⊗ 1)). But this relation implies (π ⊗ id) ◦ δε = δC ◦ π. (cid:3) This partial Bernoulli shift will be studied further in a forthcoming article. In par- ticular, the partial coaction δε should give rise to a partial crossed product that can be regarded as a quantum counterpart to the partial group algebra of a discrete group, see [16, §10], and as a C ∗-algebraic counterpart to the Hopf algebroid Hpar associated to a Hopf algebra H in [7]. 8. Dilations Let (A, ∆) be a C ∗-bialgebra. Given a partial coaction of (A, ∆), a natural and important question is whether it can be identified with as the restriction of a coaction to a weakly invariant C ∗-subalgebra as in Proposition 6.4. PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 19 Definition 8.1. Let δC be a partial coaction of (A, ∆) on a C ∗-algebra C. A dilation of δC consists of a C ∗-algebra B, a coaction δB of (A, ∆) on B, and an embedding ι : C ֒→ B that is a weak morphism from δC to δB, that is, satisfies δB(ι(c))(ι(c′) ⊗ a) = (ι ⊗ idA)(δC (c)(c′ ⊗ a)) (c, c′ ∈ C, a ∈ A). Example 8.2 (Disconnected partial actions of groups). Let C be a C ∗-algebra with a disconnected partial action ((pg)g, (θg)g) of a discrete group Γ, and consider the associated partial coaction δC of C0(Γ) as in Proposition 4.3. A dilation of δC is given by a C ∗-algebra B with a coaction of C0(Γ), that is, by an action (αg)g∈Γ of Γ on B, and an embedding C ֒→ B that is a weak morphism. Suppose that this embedding is strict. By Proposition 6.9, it is a weak morphism if and only if pg = 1Cαg(1C ) and θg = αg(cid:12)(cid:12)pgC (g ∈ Γ). In particular, 1C commutes with αg(1C ) for each g ∈ Γ. We claim that our partial action coincides with the set-theoretic restriction ((Dg)g, (αgDg )g) of α to C, where Dg = αg(C) ∩ C for each g ∈ Γ. Indeed, for every element c ∈ Dg with 0 ≤ c ≤ 1C , we have c ≤ 1C and α−1 g (c) ≤ 1C, whence c ≤ αg(1C )1C = pg and c ∈ pgC. On the other hand, if c ∈ pgC, then αg−1(c) = θg−1(c) ∈ C and hence c ∈ αg(C) ∩ C = Dg. Conversely, suppose that α is an action of Γ on a C ∗-algebra B that contains C and that α is a dilation in the usual sense, so that C ⊆ B is an ideal, pgC = αg(C) ∩ C and θg = αgpgC for each g ∈ Γ. If the embedding C ⊆ B is strict, then C is a direct summand, that is, C = 1C B, and then αg(C) ∩ C = αg(1C )1C for each g ∈ Γ, so that the coaction δB corresponding to α is a dilation of δC . The main question is, of course, which partial coactions do have a dilation. We start with a necessary condition. Definition 8.3. We call a partial coaction δC of (A, ∆) on a C ∗-algebra C regular if (idC ⊗∆)(δC (C)) · (1C ⊗ 1A ⊗ A) ⊆ M (C ⊗ A) ⊗ A. (8.1) Example 8.4. (1) Every coaction is easily seen to be regular. (2) The question of regularity arises only if C is non-unital, because every partial coaction on a unital C ∗-algebra is regular. (3) If A is a direct sum of matrix algebras, for example, if (A, ∆) is a discrete quantum group, then every partial coaction of (A, ∆) is regular. Regularity is necessary for the existence of a dilation with a strict embedding: Lemma 8.5. If a partial coaction has a dilation (B, δB, ι), where ι is strict, then the partial coaction is regular. Proof. Suppose that δC is a partial coaction of (A, ∆) on a C ∗-algebra C with a dilation (B, δB, ι). It suffices to show that the product (ι ⊗ idA ⊗ idA)((idC ⊗∆)δC(C)) · (1B ⊗ 1A ⊗ A) lies in M (B ⊗ A) ⊗ A. Since ι is a weak morphism, this product is equal to (idC ⊗∆)(δB(ι(C))) · (ι(1C ) ⊗ 1A ⊗ A), 20 F. KRAKEN, P. QUAST, AND T. TIMMERMANN which by coassociativity of δB can be rewritten as (δB ⊗ idA)(δB(ι(C))(1B ⊗ A)) · (ι(1C ) ⊗ 1A ⊗ 1A), and this product lies in M (B ⊗ A) ⊗ A because δB(ι(C))(1B ⊗ A) ⊆ B ⊗ A. (cid:3) If (A, ∆) is a regular C ∗-quantum group, for example, a compact one, and if δC is weakly continuous, then regularity of δC can be tested on the unit: Lemma 8.6. Let (A, ∆) be a regular C ∗-quantum group and let δC be a weakly continuous partial coaction of (A, ∆) on a C ∗-algebra C such that (idC ⊗∆)(δC(1C )) · (1C ⊗ 1A ⊗ A) ⊆ M (C ⊗ A) ⊗ A. Then δC is regular. Proof. We use the same notation and a similar argument as in the proof of Proposition 3.5. By (3.3), (idC ⊗∆)(δC (ω ⊲ c)) = (idC ⊗∆)(δC (1C )) · (idC ⊗ idA ⊗ idA ⊗ω)(idC ⊗∆(2))δC (c)) for all ω ∈ A∗ and c ∈ C, where ∆(2) = (∆ ⊗ idA)∆ = (idA ⊗∆)∆. Since δC is weakly continuous, we can conclude that [(idC ⊗∆)δC(C) · (1C ⊗ 1A ⊗ A)] is equal to the product of (idC ⊗∆)(δC (1C )) with [(idC ⊗ idA ⊗ idA ⊗ω)((idC ⊗∆(2))(δC (C))(1C ⊗ 1A ⊗ A ⊗ 1A)) : ω ∈ A∗]. Similarly as in the proof of Proposition 3.5, we rewrite this space in the form [(idC ⊗ idA ⊗ idA ⊗ω)(V34(idC ⊗(idA ⊗π)∆)(δC (C))124V ∗ 34(A ⊗ π( A))34) : ω ∈ B(K)∗] = [(idC ⊗ idA ⊗ idA ⊗ω)((A ⊗ π( A))34(idC ⊗(idA ⊗π)∆)(δC (C))124) : ω ∈ B(K)∗] = [((idC ⊗ idA ⊗ω ◦ π)(idC ⊗∆)δC(C)) ⊗ A : ω ∈ B(K)∗] ⊆ M (C ⊗ A) ⊗ A. Summarising, we find that (idC ⊗∆)δC (C) · (1C ⊗ 1A ⊗ A) ⊆ (idC ⊗∆)δC (1C ) · (M (C ⊗ A) ⊗ A). By assumption on δC (1C ), the right hand side lies in M (C ⊗ A) ⊗ A. (cid:3) For partial actions of a group G on a space X, a canonical dilation can be constructed as a certain quotient of the product X × G; see [2] or [16, Theorem 3.5, Proposition 5.5]. We now give a dual construction. Although this one will be improved upon in the next section, we decided to include it for instructive purpose, see also Example 8.8. From now on, we almost always assume the C ∗-algebra underlying our C ∗-bialgebra to have the slice map property, which holds, for example, if it is nuclear; see 2.1. Proposition 8.7. Let δC be an injective, regular partial coaction of a C ∗-bialgebra (A, ∆) on a C ∗-algebra C and suppose that A has the slice map property. Denote by C ⊠ A ⊆ M (C ⊗ A) the subset of all x satisfying the following conditions: (1) [x, δC (1C )] = 0; (2) (δC ⊗ idA)(x) = (δC (1C ) ⊗ 1A)(idC ⊗∆)(x) = (idC ⊗∆)(x)(δC (1C ) ⊗ 1A); (3) x(1C ⊗ A) and (1C ⊗ A)x lie in C ⊗ A; (4) (idC ⊗∆)(x)(1C ⊗ 1A ⊗ A) and (1C ⊗ 1A ⊗ A)(idC ⊗∆)(x) lie in M (C ⊗ A) ⊗ A. PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 21 Then C ⊠ A is a C ∗-algebra, idC ⊗∆ restricts to a coaction of (A, ∆) on C ⊠ A, and (C ⊠ A, idC ⊗∆, δC) is a dilation of δC. Proof. Clearly, C ⊠ A is a C ∗-algebra. It contains δC (C) by (3.1) and regularity of δC . Next, we need to show that (idC ⊗∆)(C ⊠ A)(1C ⊗ 1A ⊗ A) ⊆ (C ⊠ A) ⊗ A. Condition (4) implies that the left hand side is contained in M (C ⊗ A) ⊗ A. Since A has the slice map property, it suffices to show that for every y ∈ C ⊠ A and ω ∈ A∗, the element x := (idC ⊗ idA ⊗ω)(idC ⊗∆)(y) = (idC ⊗(idA ⊗ω)∆)(y) lies in C ⊠ A, that is, satisfies conditions (1) -- (4) above. In case of (2) -- (4), we only prove the first halfs of the statements, the others follow similarly. (1) The element x commutes with δC (1C) because (idC ⊗∆)(y) commutes with (δC (1C )⊗ 1A) by (2), applied to y. (2) We use (1) for y and coassociativity of ∆ to see that (δC ⊗ idA)(x) = (idC ⊗ idA ⊗(idA ⊗ω)∆)(δC ⊗ idA)(y) = (idC ⊗ idA ⊗(idA ⊗ω)∆)((δC (1) ⊗ 1A)(idC ⊗∆)(y)) = (δC (1C ) ⊗ 1A)(idC ⊗(idA ⊗ idA ⊗ω)∆(2))(y) = (δC (1C ) ⊗ 1A)(idC ⊗∆)(idA ⊗(idA ⊗ω)∆)(y) = (δC (1C ) ⊗ 1A)(idC ⊗∆)(x). (3) Write ω = aυ with a ∈ A and υ ∈ A∗ using Cohen's factorisation theorem, and let a′ ∈ A. Then x(1C ⊗ a′) = (idC ⊗ idA ⊗υ)((idC ⊗∆)(y)(1C ⊗ a′ ⊗ a)). We use the relation A⊗A = [∆(A)(A⊗A] and condition (3) on y and find that x(1C ⊗a′) lies in C ⊗ A as desired. (4) With a, a′, υ as above, (idC ⊗∆)(x) · (1C ⊗ 1A ⊗ a′) = (idC ⊗ idA ⊗ idA ⊗υ)((idC ⊗∆(2))(y) · (1C ⊗ 1A ⊗ a′ ⊗ a)). We use the relation A ⊗ A = [∆(A)(A ⊗ A] again and find that (idC ⊗∆(2))(y) · (1C ⊗ 1A ⊗ a′ ⊗ a) ∈ (idC ⊗ idA ⊗∆)((idC ⊗∆)(y) · (1C ⊗ 1A ⊗ A)) · (1C ⊗ 1A ⊗ A ⊗ A). Condition (4), applied to y, implies that the expression above lies in M (C ⊗ A) ⊗ A ⊗ A. Slicing the last factor with υ, we get (idC ⊗∆)(x) · (1C ⊗ 1A ⊗ a′) ∈ M (C ⊗ A) ⊗ A. (cid:3) Example 8.8 (Case of a partial group action). Consider the partial coaction δC associ- ated to a disconnected partial action ((pg), (θg)g) of a discrete group Γ on a C ∗-algebra C. Identify M (C ⊗ C0(Γ)) with Cb(Γ; M (C)) and let f ∈ Cb(Γ; M (C)). Then conditions (1) and (4) in Proposition 8.7 are automatically satisfied by f , condition (3) is equivalent to f ∈ Cb(Γ; C), and condition (2) corresponds to the invariance condition θg(pg−1f (h)) = pgf (gh) (g, h ∈ Γ). 22 F. KRAKEN, P. QUAST, AND T. TIMMERMANN In particular, if C = C0(X) for some locally compact, Hausdorff space X, then each pg is the characteristic function of some clopen Dg ⊆ X, each θg is the pull-back along some homeomorphism αg−1 : Dg → Dg−1, and the invariance condition above translates into f (x, gh) = f (αg−1(x), h) (g, h ∈ Γ, x ∈ Dg), so that f descends to the quotient space of X × Γ with respect to the equivalence relation given by (x, gh) ∼ (αg−1(x), h) for all g, h ∈ Γ and x ∈ Dg. This space is, up to the reparameterization (x, g) 7→ (g−1, x), the globalization of the partial action ((Dg)g, (αg)g) of Γ on X, see [16, Theorem 3.5, Proposition 5.5], and C0(X) ⊠ C0(Γ) can be identified with a C ∗-subalgebra of Cb((X × Γ)/∼). 9. Minimal dilations Among all dilations of a fixed partial coaction δC of a C ∗-bialgebra (A, ∆), we now single out a universal one, which we call the globalization of δC. More precisely, we show that (1) every dilation of δC contains one that is minimal in a natural sense, and (2) that all such minimal dilations are isomorphic. We need to assume, however, that δC is regular and injective, that A has the slice map property, and, for (2), that (A, ∆) is a C ∗-quantum group. Definition 9.1. Let δC be a partial coaction of (A, ∆) on a C ∗-algebra C. We call a dilation (B, δB, ι) of δC minimal if ι(C) and A∗ ⊲ ι(C) generate B as a C ∗-algebra. Remark 9.2. Let (B, δB, ι) be a minimal dilation of a partial coaction δC of (A, ∆) on some C ∗-algebra C. Then ι(C) ⊆ B is an ideal because ι(C)(A∗ ⊲ι(C)) = ι(C)ι(A∗ ⊲C) ⊆ ι(C) by (6.1). If, moreover, ι is strict, then ι(C) is a direct summand of B. Example 9.3. If, in the situation above, (A, ∆) is the C ∗-bialgebra of functions on a discrete group Γ, then the coaction δB corresponds to an action α of Γ on B, and the dilation is minimal if and only if Pg∈Γ αg(ι(C)) generates B as a C ∗-algebra. Example 9.4 (Partial Bernoulli shift). Let G = (C0(G), ∆) be a discrete C ∗-quantum group. Denote by δ the coaction of C0(G) on C(BG), see Section 7, by δε its restriction to a partial coaction on C(Bε G) obtained as the restriction of δ, see Proposition 7.5, and by ι : C(Bε G) the inclusion. Then (C0(B× G), δ×, ι) is a dilation of δε because δε is a restriction of δ×, and this dilation is minimal. Indeed, δ×(pε) = p by Lemma 7.6, whence C0(G)∗ ⊲ ι(C0(Bε ij for every α, i, j, and these elements generate C0(B× G), by δ× the coaction of C0(G) on C0(B× G)) contains pα G) ֒→ C0(B× G). Every dilation contains a minimal one: Proposition 9.5. Let δC be a partial coaction of (A, ∆) on a C ∗-algebra C with a dilation (B, δB, ι), and suppose that A has the slice map property. Denote by B0 ⊆ B the C ∗-subalgebra generated by ι(C) and A∗ ⊲ ι(C). (1) δB restricts to a coaction δB0 on B0, and (B0, δB0 , ι) is a minimal dilation of δC . (2) If (A, ∆) is a regular C ∗-quantum group and δC is weakly continuous, then [A∗ ⊲ ι(C)] ⊆ B is a C ∗-algebra. If additionally ι is strict, then B0 = [(A∗⊲ι(C))(C1B + Cι(1C ))]. PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 23 Proof. (1) To prove the first assertion, we only need to show that δB(ι(C))(1B ⊗ A) ⊆ B0 ⊗ A and δB(A∗ ⊲ ι(C))(1B ⊗ A) ⊆ B0 ⊗ A. But for all c ∈ C, υ, ω ∈ A∗, both (id ⊗ω)(δB(ι(c))) = ω ⊲ι(c) and (id ⊗ω)(δB(υ ⊲ι(c))) = ωυ ⊲ ι(c) lie in B0. Since A has the slice map property, the desired inclusions follow. (2) We follow the proof of [9, Proposition 5.7], using the same notation and manipula- tions as in the proof of Proposition 3.5. To shorten the notation, let U := (π ⊗ idπ( A))(V ) and δπ := (idB ⊗π) ◦ δB ◦ ι. Then by (3.3), [A∗ ⊲ ι(C)] = [A∗ ⊲ ι(C(A∗ ⊲ C))] = [(idB ⊗υ ⊗ ω)((δB ⊗ idA)((C ⊗ 1A)δC (C))) : υ, ω ∈ A∗] = [(idB ⊗υ ⊗ ω)((δB ⊗ idA)((C ⊗ 1A)δB(C))) : υ, ω ∈ A∗] = [(idB ⊗υ ◦ π ⊗ ω ◦ π)((δB(C) ⊗ 1A)(idB ⊗∆)δB(C)) : υ, ω ∈ B(K)∗] = [(idB ⊗υ ⊗ ω)(δπ(C)12U23δπ(C)13U ∗ = [(idB ⊗υ ⊗ ω)(δπ(C)12U23δπ(C)13(π(A) ⊗ π( A))23) : υ, ω ∈ B(K)∗] = [(idB ⊗υ ⊗ ω)(δπ(C)12(π(A) ⊗ π( A))23δπ(C)13) : υ, ω ∈ B(K)∗] = [(A∗ ⊲ ι(C))(A∗ ⊲ ι(C))]. 23) : υ, ω ∈ B(K)∗] Thus, [A∗ ⊲ ι(C)] is a C ∗-algebra. If ι is strict so that ι(1C ) is well-defined, then this C ∗-algebra commutes with ι(1C ), and by (6.1) the product is [ι(A∗ ⊲ C)] = ι(C). This proves the last assertion concerning B0. (cid:3) If we apply Proposition 9.5 to the canonical dilation (C ⊠ A, idC ⊗∆, δC ) constructed in Proposition 8.7, we obtain the following dilation: Theorem 9.6. Let (A, ∆) be a C ∗-bialgebra, where A has the slice map property, and let δC be an injective, regular partial coaction of (A, ∆) on a C ∗-algebra C. Denote by G(C) ⊆ M (C ⊗ A) the C ∗-subalgebra generated by {(idC ⊗ idC ⊗ω)(idC ⊗∆)δC(c) : ω ∈ A∗, c ∈ C} and δC (C). Then idC ⊗∆ restricts to a partial coaction on G(C) and G(δC ) := (G(C), idC ⊗∆, δC ) is a minimal dilation of δC . Proof. By a similar argument as in the proof of Proposition 8.7, we only need to show that for every c ∈ C and υ, ω ∈ A∗, the elements (idC ⊗ idA ⊗υ)((idC ⊗∆)δC (c)) and (idC ⊗ idA ⊗υ)((idC ⊗∆)((idC ⊗ idA ⊗ω)(idC ⊗∆)δC (c))) lie in G(C). In the first case, this is trivially true, and in the second case, one finds that the element is equal to d = (idC ⊗ idA ⊗υω)((idC ⊗∆)δC(C)) ∈ G(C). (cid:3) Remark 9.7. Suppose that (A, ∆) and δC are as above. 24 F. KRAKEN, P. QUAST, AND T. TIMMERMANN (1) Beware that δC is strict as a map from C to M (C ⊗ A), but this does not imply that δC is strict as a map from C to G(C). (2) If δC is weakly continuous, then (3.3) implies that G(C) ⊆ M (C ⊗ A) is equal to the C ∗-subalgebra generated by {(idC ⊗ idC ⊗ω)(idC ⊗∆)δC(c) : ω ∈ A∗, c ∈ C} and δC(1C ). Example 9.8 (Case of a partial group action). Consider the partial coaction δC associ- ated to a disconnected partial action ((pg), (θg)g) of a discrete group Γ on a C ∗-algebra C, and identify M (C ⊗ C0(Γ)) with Cb(Γ; M (C)). In that case, G(C) is the C ∗-algebra generated by all functions of the form fc,h = (idC ⊗ idC0(Γ) ⊗evh)(idC ⊗∆)δC(c) : g 7→ θgh(ph−1g−1c), where c ∈ C and g, h ∈ Γ. The action ρ of Γ corresponding to the coaction idC ⊗∆ is given by right translation of functions, whence ρh′(fc,h) = fc,h′h for all h′ ∈ Γ. We shall use the following notion of a morphism between dilations: Definition 9.9. Let δC be a partial coaction of a C ∗-bialgebra (A, ∆) on some C ∗- algebra C. A morphism between dilations B = (B, δB, ιB) and D = (D, δD, ιD) of δC is a ∗-homomorphism φ : B → D satisfying φ(ιB(c)) = ιD(c) and δD(φ(b))(1D ⊗ a) = (φ ⊗ idA)(δB(b)(1B ⊗ a)) (9.1) for all c ∈ C, b ∈ B and a ∈ A. Evidently, all dilations of a fixed partial coaction δC form a category; we denote this category by Dil(δC ). Remark 9.10. The second equation in (9.1) is equivalent to the condition that φ is a morphism of left A∗-modules, that is, ω ⊲ φ(b) = φ(ω ⊲ b) for all b ∈ B and ω ∈ A∗. If δC is injective and regular, then the dilation G(δC ) is terminal among the minimal ones: Proposition 9.11. Let δC be an injective, regular partial coaction of a C ∗-bialgebra (A, ∆) on a C ∗-algebra C, let B = (B, δB, ι) be a minimal dilation of δC, and suppose that A has the slice map property. Then there exists a unique morphism φB from B to G(δC ), and on the level of C ∗-algebras, φB is surjective. For each b ∈ B, the image φB(b) is the restriction of δB(b) to the ideal ι(C) ⊗ A ∼= C ⊗ A in B ⊗ A. Proof. Uniqueness follows from the fact that B is generated by ι(C) and A∗ ⊲ ι(C). To prove existence, define φB as in (3). Since ι is a weak morphism, φB ◦ ι = δC . The relation (idB ⊗∆)δB = (δB ⊗ idA)δB implies that (φB ⊗ idA)(δB(b)(b′ ⊗ a)) = (idC ⊗∆)(φB(b))(φB(b′) ⊗ a) for all b, b′ ∈ B and a ∈ A; in particular, φB(ω ⊲ ι(c))φB(b) = (idC ⊗ idA ⊗ω)((idC ⊗∆)δC(C))φB(b) for all c ∈ C and ω ∈ C ∗. Now, the definition of G(C) and minimality of B imply φB(B) = G(C). (cid:3) If (A, ∆) is a C ∗-quantum group, then the morphism above is injective and hence an isomorphism. To show this, we use the following observation: PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 25 Lemma 9.12. Let δB be a coaction of a C ∗-quantum group (A, ∆) on a C ∗-algebra B and let b, b′ ∈ M (B). Then δB(b)(b′ ⊗ 1A) = 0 if and only if (b ⊗ 1A)δB(b′) = 0. Proof. Choose a modular multiplicative unitary W for (A, ∆) so that ∆(a) = W (a⊗1)W ∗ for all a ∈ A. Then (δB ⊗ idA)(δB(b)) · (δB(b′) ⊗ 1A) = (idB ⊗∆)(δB(b)) · (δB(b′) ⊗ 1A) = W23(δB(b) ⊗ 1A)W ∗ 23(δB(b′) ⊗ 1A). Since δB ⊗ idA is injective and W is unitary, we can conclude that δB(b)(b′ ⊗ 1A) = 0 if (9.2) (δB(b) ⊗ 1A)W ∗ 23(δB(b′) ⊗ 1A) = 0. A similar argument shows that (b ⊗ 1A)δB(b′) = 0 if and only if (9.3) Now, both (9.2) and (9.3) are equivalent to the condition δB(b)(1B ⊗ A)δB(b′) = 0. (cid:3) (δB(b) ⊗ 1A)W23(δB(b′) ⊗ 1A) = 0. We can now prove claim (2) stated in the introduction to this section: Proposition 9.13. Let δC be an injective, regular partial coaction of a C ∗-quantum group (A, ∆) on a C ∗-algebra C, suppose that A has the slice map property, and let B be a minimal dilation of δC . Then the morphism φB from B to G(δC ) is an isomorphism. Proof. Write B = (B, δB, ι). It suffices to show that φB is injective on the level of C ∗- algebras. On the direct summand ιC (C) ⊆ B, the morphism φB is given by ιC (c) 7→ δC (c) and hence injective. Since B is minimal, the direct summand (1B − ι(1C ))B of B is generated by (1B − ι(1C ))(A∗ ⊲ ι(C)). Given a non-zero b ∈ (1B − ι(1C ))B, we therefore find some c ∈ C such that δB(ι(c))(b ⊗ 1A) is non-zero, and then (ι(c) ⊗ 1A)δB(b) is non-zero by the lemma above, whence φB(b) is non-zero. (cid:3) Corollary 9.14. Let (A, ∆) be a C ∗-quantum group and suppose that A has the slice map property. Then all minimal dilations of an injective, regular partial coaction of (A, ∆) are isomorphic. Acknowledgements We thank the referees for helpful comments that improved the quality of the manu- script. References 1. Fernando Abadie, Sobre ações parciais, fibrados de Fell e grupóides, Ph.D. thesis, Sáo Paulo, 1999. , Enveloping actions and Takai duality for partial actions, J. Funct. Anal. 197 (2003), no. 1, 2. 14 -- 67. 3. Edson R. Alvares, Marcelo M. S. Alves, and Eliezer Batista, Partial Hopf module categories, J. Pure Appl. Algebra 217 (2013), no. 8, 1517 -- 1534. 4. Marcelo M. S. Alves, Eliezer Batista, Michael Dokuchaev, and Antonio Paques, Globalization of twisted partial Hopf actions, J. Aust. Math. Soc. 101 (2016), no. 1, 1 -- 28. 5. Marcelo Muniz S. Alves and Eliezer Batista, Enveloping actions for partial Hopf actions, Comm. Algebra 38 (2010), no. 8, 2872 -- 2902. 6. , Globalization theorems for partial Hopf (co)actions, and some of their applications, Groups, algebras and applications, Contemp. Math., vol. 537, Amer. Math. Soc., Providence, RI, 2011, pp. 13 -- 30. 26 F. KRAKEN, P. QUAST, AND T. TIMMERMANN 7. Marcelo Muniz S. Alves, Eliezer Batista, and Joost Vercruysse, Partial representations of Hopf algebras, J. Algebra 426 (2015), 137 -- 187. 8. Saad Baaj and Georges Skandalis, Unitaires multiplicatifs et dualité pour les produits croisés de C ∗-algèbres, Ann. Sci. École Norm. Sup. (4) 26 (1993), no. 4, 425 -- 488. 9. Saad Baaj, Georges Skandalis, and Stefaan Vaes, Non-semi-regular quantum groups coming from number theory, Comm. Math. Phys. 235 (2003), no. 1, 139 -- 167. 10. B. Blackadar, Operator algebras, Encyclopaedia of Mathematical Sciences, vol. 122, Springer-Verlag, Berlin, 2006, Theory of C ∗-algebras and von Neumann algebras, Operator Algebras and Non- commutative Geometry, III. 11. E. Bédos and L. Tuset, Amenability and co-amenability for locally compact quantum groups, Inter- national Journal of Mathematics 14 (2003), no. 08, 865 -- 884. 12. S. Caenepeel and K. Janssen, Partial (co)actions of Hopf algebras and partial Hopf-Galois theory, Communications in Algebra 36 (2008), no. 8, 2923 -- 2946. 13. Felipe Castro, Antonio Paques, Glauber Quadros, and Alveri Santána, Partial actions of weak Hopf algebras: smash product, globalization and Morita theory, J. Pure Appl. Algebra 219 (2015), no. 12, 5511 -- 5538. 14. Ruy Exel, Circle actions on C ∗-algebras, partial automorphisms, and a generalized Pimsner- Voiculescu exact sequence, J. Funct. Anal. 122 (1994), no. 2, 361 -- 401. 15. , Twisted partial actions: a classification of regular C ∗-algebraic bundles, Proc. London Math. Soc. (3) 74 (1997), no. 2, 417 -- 443. 16. Ruy Exel, Partial Dynamical Systems, Fell Bundles and Applications, Operator Algebras, arXiv:1511.04565 (to appear). 17. Uwe Franz and Adam Skalski, On idempotent states on quantum groups, J. Algebra 322 (2009), no. 5, 1774 -- 1802. 18. Johan Kustermans and Stefaan Vaes, Locally compact quantum groups, Ann. Sci. École Norm. Sup. (4) 33 (2000), no. 6, 837 -- 934. 19. E. C. Lance, Hilbert C ∗-modules. a toolkit for operator algebraists, London Mathematical Society Lecture Note Series, vol. 210, Cambridge University Press, Cambridge, 1995. 20. Magnus B. Landstad and A. Van Daele, Groups with compact open subgroups and multiplier Hopf ∗-algebras, Expo. Math. 26 (2008), no. 3, 197 -- 217. 21. K. McClanahan, K-theory for partial crossed products by discrete groups, Journal of Functional Analysis 130 (1995), no. 1, 77 -- 117. 22. Ralf Meyer, Sutanu Roy, and Stanisław Lech Woronowicz, Quantum group-twisted tensor products of C∗-algebras, Internat. J. Math. 25 (2014), no. 2, 1450019, 37. 23. Sergey Neshveyev and Lars Tuset, Compact quantum groups and their representation categories, Cours Spécialisés, vol. 20, Société Mathématique de France, Paris, 2013. 24. S. Roy and T. Timmermann, The maximal quantum group-twisted tensor product of C*-algebras, Operator Algebras, arXiv:1511.06100, (to appear in J. Noncomm. Geometry). 25. Sutanu Roy, C ∗-quantum groups with projection, PhD Thesis, Georg-August Universität Göttingen, 2013. 26. Piotr M. Sołtan and Stanisław L. Woronowicz, From multiplicative unitaries to quantum groups. II, J. Funct. Anal. 252 (2007), no. 1, 42 -- 67. 27. P. M. Sołtan and S. L. Woronowicz, A remark on manageable multiplicative unitaries, Lett. Math. Phys. 57 (2001), no. 3, 239 -- 252. 28. Piotr M. Sołtan, Quantum families of maps and quantum semigroups on finite quantum spaces, J. Geom. Phys. 59 (2009), no. 3, 354 -- 368. 29. Masamichi Takesaki and Nobuhiko Tatsuuma, Duality and subgroups, Ann. of Math. (2) 93 (1971), 344 -- 364. 30. O. Uuye and J. Zacharias, The Fubini product and its applications, ArXiv e-prints (2016). 31. Simon Wassermann, The slice map problem for C ∗-algebras, Proc. London Math. Soc. (3) 32 (1976), no. 3, 537 -- 559. 32. S. L. Woronowicz, From multiplicative unitaries to quantum groups, Internat. J. Math. 7 (1996), no. 1, 127 -- 149. PARTIAL ACTIONS OF C ∗-QUANTUM GROUPS 27 33. , Compact quantum groups, Symétries quantiques (Les Houches, 1995), North-Holland, Am- sterdam, 1998, pp. 845 -- 884. E-mail address: 1 E-mail address: 2 [email protected] [email protected] 3FB Mathematik und Informatik, University of Muenster, Einsteinstr. 62, 48149 Muen- ster, Germany E-mail address: [email protected]
0812.3601	2	0812	2010-07-23T08:47:23	A Horizontal Categorification of Gelfand Duality	[ "math.OA", "math.CT" ]	In the setting of C-categories, we provide a definition of "spectrum" of a commutative full C-category as a one-dimensional unital saturated Fell bundle over a suitable groupoid (equivalence relation) and prove a categorical Gelfand duality theorem generalizing the usual Gelfand duality between the categories of commutative unital C*-algebras and compact Hausdorff spaces. Although many of the individual ingredients that appear along the way are well-known, the somehow unconventional way we "glue" them together seems to shed some new light on the subject.	math.OA	math	A Horizontal Categorification of Gel'fand Duality Paolo Bertozzini ∗@, Roberto Conti ∗‡, Wicharn Lewkeeratiyutkul ∗§ @ e-mail: [email protected] ‡ Mathematics, School of Mathematical and Physical Sciences, University of Newcastle, Callaghan, NSW 2308, Australia e-mail: [email protected] § Department of Mathematics, Faculty of Science, Chulalongkorn University, Bangkok 10330, Thailand e-mail: [email protected] 22 July 2010 This paper is dedicated to J. E. Roberts, the "pioneer" of C-categories. Abstract In the setting of C-categories, we provide a definition of spectrum of a commutative full C-category as a one-dimensional unital Fell bundle over a suitable groupoid (equiv- alence relation) and prove a categorical Gel'fand duality theorem generalizing the usual Gel'fand duality between the categories of commutative unital C-algebras and com- pact Hausdorff spaces. Although many of the individual ingredients that appear along the way are well known, the somehow unconventional way we "glue" them together seems to shed some new light on the subject. MSC-2000: 46L87, 46M15, 46L08, 46M20, 16D90, 18F99. Keywords: C-category, Fell Bundle, Duality, Non-commutative Geometry. 1 Introduction There is no need to explain why the notions of geometry and space are fundamental both in mathematics and in physics. Typically, a rigorous way to encode at least some basic geometrical content into a mathematical framework makes use of the notion of a topological space, i.e. a set equipped with a topological structure. Although being just a preliminary step in the process of developing a more sophisticated apparatus, this way of thinking has been very fruitful for both abstract and concrete purposes. In a very important development, I. M. Gel'fand looked not at the topological space itself but rather at the space of all continuous functions on it, and realized that these seem- ingly different structures are in fact essentially the same. In slightly more precise terms, he found a basic example of anti-equivalence between certain categories of spaces and al- gebras (see for example [Bl, Theorems II.2.2.4, II.2.2.6] or [L, Section 6]). Since on the ∗Partially supported by the Thai Research Fund: grant n. RSA4780022. 1 analytic side C(X; C) is a special type of a Banach algebra called a C-algebra, the study of possibly non-commutative C-algebras has been often regarded as a good framework for "non-commutative topology". The duality aspect has been later enforced by the Serre-Swan equivalence [K, Theorem 6.18] between vector bundles and suitable modules (see also [FGV] for a Hermitian version of the theorem and [T1, T2, W] for generalizations involving Hilbert bundles). By then, break- through results have continued to emerge both in geometry and functional analysis, based on Gel'fand's original intuition, for about four decades. In connection with physical ideas, L. Crane-D. Yetter [CY] and J. Baez-J. Dolan [BD] have recently proposed a process of categorification of mathematical structures, in which sets and functions are replaced by categories and functors. From this perspective, in this paper, we wish to discuss a categorification of the notion of space extending and merging together Gel'fand duality and Serre-Swan equivalence. On one side of the extended duality we have a horizontal categorification (a terminology that we introduced in [BCL2, Section 4.2]) of the notion of commutative C-algebra, namely a commutative C-category, or commutative C-algebroid (see definition 2.1), whilst the corresponding replacement of spaces, the spaceoids (see definition 3.2), are supposed to parametrize their spectra. Spaceoids could be described in several different albeit equivalent ways. In this paper we have decided to focus on a characterization based on the notion of Fell bundle. Originally Fell bundles were introduced in connection with the study of representations of locally compact groups, but we argue that they come to life naturally on the basis of purely topological principles. Rather surprisingly, to the best of our knowledge, the notions of commutative C-category and its spectrum have not been discussed before, despite the fact that (mostly highly non- commutative) C-categories have been somehow intensively exploited over the last 30 years in several areas of research, including Mackey induction, superselection structure in quantum field theory, abstract group duality, subfactors and the Baum-Connes conjecture. At any rate, we make frequent contact with the related notions that can be found in the literature, hoping that our approach sheds new light on the subject by approaching the matter from a kind of unconventional viewpoint. Of course, once we have a running definition, it seems quite challenging in the next step to look for some natural occurrence of the notion of spaceoid in other contexts. For instance, we are not aware of any connection with the powerful concepts that have been introduced in algebraic topology to date. Also, the appearance of bundles in the structure of the spectrum suggests an intriguing connection to local gauge theory but we have not developed these ideas yet. Some of our considerations have been motivated by a categorical approach to non-commutative geometry [BCL2], and it is rewarding that some of its relevant tools (e.g., Serre-Swan theorem, Morita equivalence) appear naturally in our context. More structure is expected to emerge when our categories are equipped with a differentiable structure. In the case of usual spaces, in the setting of A. Connes' non-commutative geometry [C], this has been achieved by means of a Dirac operator, and then axiomatized using the concept of spectral triple. Here below we present a short description of the content of the paper. In section 2 we mention, mainly for the purpose of fixing our notation, some basic definitions on C-categories. Section 3 opens recalling the notion of a Fell bundle in the case of involutive inverse base categories and then proceeds to introduce the definition of the category of spaceoids that will eventually subsume that of compact Hausdorff spaces in our duality theorem. The construction of a small commutative full C-category starting from a spaceoid is undertaken in section 4, while the spectral analysis of a commutative full C-category is 2 the subject of the more technical section 5 where a spectrum functor from the category of full commutative C-categories to our category of spaceoids is defined. Section 6 presents the main result of this paper in the form of a duality between a certain category of commutative full C-categories and the category of their spectra (spaceoids). A categorified version of Gel'fand transform is introduced and used to prove a Gel'fand spectral reconstruction theorem for full commutative C-categories. Similarly a categorified evalu- ation transform is defined for the purpose of proving the representativity of the spectrum functor. Section 7 is devoted to examples and applications. Here we mention several natural exam- ples of commutative full C-categories and we produce explicit constructions of spaceoids, either reassembling the Hermitian line bundles obtained in [BCL3] as spectra of imprim- itivity Hilbert C-bimodules, or (in a way completely independent from C-categories) as associated line bundles for a categorified version of T-torsors. Among the several possi- ble future applications of such categorified Gel'fand duality, we describe in some detail a categorified continuous functional calculus. The paper ends with an outlook. While in the usual Gel'fand duality theory a spectrum is just a compact topological space, in the situation under consideration it comes up equipped with a natural bundle structure. In particular, the spectrum of a commutative full C-category can be identified with a kind of "groupoid of Hermitian line bundles"1 that can be conveniently described using the language of Fell bundles (or equivalently as a continuous field of one-dimensional C-categories). Along the way, we also discuss several categorical versions of some well-known concepts like the Gel'fand transform that we think are of independent interest. Notice that a notion of Fourier transform in the setting of compact groupoids has been discussed by M. Amini [A]. Our duality is reminiscent of an interesting but widely ignored duality result of A. Taka- hashi [T1, T2]. Takahashi's duality can be essentially understood as a duality of categories equipped with a partially defined weakly associative tensor product and a (weak) involution, although he does not explicitly examine such natural structures on his categories of Hilbert bundles and Hilbert C-modules.2 The duality presented in this paper is essentially a version of the former, where we exploit a strict realization of these tensor products and involutions, considering commutative full C-categories and spaceoids instead of Hilbert C-modules and Hilbert bundles. Most of the results presented here have been announced in our survey paper [BCL2] and have been presented in several seminars in Thailand, Australia, Italy, UK since May 2006. Note added in proof. When the present work was under preparation, we became aware of some related results in T. Timmermann's PhD dissertation [Ti] where, in the context of Hopf algebraic quantum groupoids, a very general non-commutative Pontryagin duality theory is developed by means of pseudomultiplicative unitaries in C-modules; and also in V. Deaconu-A. Kumjian-B. Ramazan [DKR], where a notion of Abelian Fell bundle (which contains our commutative C-categories as a special case) is introduced and a structure theorem for them (in terms of "twisted coverings of groupoids") is proved. In the framework of T -duality, a Pontryagin type duality between commutative principal bundles and gerbes has been proposed by C. Daenzer [D]; while a generalization of Pontryagin duality for locally compact Abelian group bundles has been provided by G. Goehle [G]. 1To be more precise, the spectrum is uniquely associated to a connected groupoid (actually an equiva- lence relation) of Hermitian line bundles, over a given compact Hausdorff space X, with compositions and involutions provided by a strict realization of fiberwise tensor products and fiberwise duals (the spectrum being the ∗-category whose class of arrows is obtained as union of the Hom sets of this groupoid). 2With more precision, Takahashi's result can be formally expressed as a duality between (a weak involutive form of) "2-fold-categories" (also called double categories in the literature). 3 2 Category A of full commutative C-categories The notion of C-category, introduced by J. Roberts (see P. Ghez-R. Lima-J. Roberts [GLR] and also P. Mitchener [M2]) has been extensively used in algebraic quantum field theory: Definition 2.1. A C-category is a category C such that: the sets CAB := HomC(B, A) are complex Banach spaces; the compositions are bilinear maps such that kxyk ≤ kxk · kyk, for all x ∈ CAB, y ∈ CBC ; there is an involutive antilinear contravariant functor ∗ : C → C, acting identically on the objects, such that kx∗xk = kxk2, for all x ∈ CBA and such that x∗x is a positive element in the C-algebra CAA, for every x ∈ CBA (i.e. x∗x = y∗y for some y ∈ CAA). In a C-category C, the "diagonal blocks" CAA := HomC(A, A) are unital C-algebras and the "off-diagonal blocks" CAB := HomC(B, A) are unital Hilbert C-bimodules on the C-algebras CAA and CBB. We say that C is full if all the bimodules CAB are imprimitivity bimodules. In practice, every full C-category is uniquely associated to a "strictification" of a sub-equivalence relation of the Picard-Rieffel groupoid of unital C-algebras,3 where the term sub-equivalence relation of a category denotes a subcategory that is itself an equivalence relation. It is also very useful to see a C-category as an involutive category fibered over the equivalence relation of its objects: in this way, a (full) C-category be- comes a special case of a (saturated) unital Fell bundle over an involutive (discrete) base category as described in definition 3.1 below. We say that C is one-dimensional if all the bimodules CAB are one-dimensional and hence Hilbert spaces. Clearly, all one-dimensional C-categories are full. The first problem that we have to face is how to select a suitable full subcategory A of "com- mutative" full C-categories playing the role of horizontal categorification of the category of commutative unital C-algebras. Since we are working in a completely strict categorical environment, our choice is to define a C-category C to be commutative if all its diagonal blocks CAA are commutative C-algebras. If C, D ∈ A are two full commutative small C-categories (with the same cardinality of the set of objects), a morphism in the category A is an object bijective ∗-functor Φ : C → D. For later usage, recall from [GLR, Definition 1.6] and [M2, Section 4] that a closed two- sided ideal I in a C-category C is always a ∗-ideal and that the quotient C/I has a natural structure as a C-category with a natural quotient functor π : C → C/I. We have this "first isomorphism theorem", whose proof is standard. Theorem 2.2. Let Φ : C → D be a ∗-functor between C-categories. The kernel of Φ defined by ker Φ := {x ∈ C Φ(x) = 0} is a closed two-sided ideal in C and there exists a unique ∗-functor Φ : C/ ker Φ → D such that Φ ◦ π = Φ. The functor Φ is faithful and it is full if and only if Φ is full. Recall (see [GLR, Definition 1.8]) that a representation of a C-category C is a ∗-functor Φ : C → H with values in the C-category H of bounded linear maps between Hilbert spaces. We end this section with a simple observation, whose proof is omitted. Lemma 2.3. A one-dimensional C-category C, admits at least one ∗-functor γ : C → C. 3By this we mean that, in a full C-category C, the family of HomC(·, ·) spaces is itself a strict subcategory of 1-arrows in the weak 2-C-category of Hilbert C-bimodules (with Rieffel tensor products and duals as compositions and involutions), that "projects onto" a sub-equivalence relation of the Picard-Rieffel groupoid (see [BCL3] for details on these categories). 4 3 Category T of full topological spaceoids We now proceed to the identification of a good category T of "spaceoids" playing the role of horizontal categorification of the category of continuous maps between compact Hausdorff topological spaces. Making use of Gel'fand duality (see e.g. [L, Section 6]) for the diagonal blocks CAA and (Hermitian) Serre-Swan equivalence (see e.g. [BCL2, Section 2.1.2] and references therein) for the off-diagonal blocks CAB of a commutative full C-category C, we see that the spectrum of C identifies a sub-equivalence relation embedded in the Picard groupoid of Hermitian line bundles over the Gel'fand spectra of the diagonal C-algebras CAA. Finally, reassembling such block-data, we recognize that, globally, the spectrum of a commutative full C-category can be described as a very special kind of a Fell bundle that we call a full topological spaceoid. Fell bundles over topological groups were first introduced by J. Fell [FD, Section II.16] and later generalized to the case of groupoids by S. Yamagami (see A. Kumjian [Ku] and references therein) and to the case of inverse semigroups by N. Sieben (see R. Exel [E, Section 2]). These notions admit a natural extension to that of a Fell bundle over an involutive inverse category4 that we describe in definition 3.1 below. For the definition of a Banach bundle, we refer to J. Fell-R. Doran [FD, Section I.13]. We recall, from A. Kumjian [Ku, Section 2], that a Fell bundle over a groupoid is a Banach bundle (E, π, X) over a topological groupoid X whose total space E is equipped with a continuous involution ∗ : E → E, denoted here by ∗ : e 7→ e∗, and with a continuous multiplication ◦ : E2 → E, denoted here by ◦ : (e1, e2) 7→ e1e2, defined on the set E2 := {(e1, e2) (π(e1), π(e2)) ∈ X2}, where Xn denotes the family of n-paths in the groupoid X, satisfying the following properties: π(e1e2) = π(e1)π(e2), for all (e1, e2) ∈ E2, for all (x1, x2) ∈ X2, the multiplication map is bilinear when restricted to the sets Ex1 × Ex2, where Ex := π−1(x), (e1e2)e3 = e1(e2e3), for all (e1, e2, e3) ∈ E such that (π(e1), π2(e2), π(e3)) ∈ X3, (1) ke1e2k ≤ ke1k · ke2k, for all (e1, e2) ∈ E2, π(e∗) = π(e)∗, for all e ∈ E, where π(e)∗ denotes the inverse of π(e) in X, for all x ∈ X, the involution map is conjugate-linear when restricted to the set Ex, (e∗)∗ = e, for all e ∈ E, (e1e2)∗ = e∗ 2e∗ 1, for all (e1, e2) ∈ E2, (2) ke∗ek = kek2, for all e ∈ E, (3) the element e∗e is positive in the C-algebra Eπ(e∗e), for all e ∈ E. This definition can be recasted in the following concise and slightly generalized form, that we systematically adopt in the sequel. Definition 3.1. A unital Fell bundle (E, π, X) over an involutive inverse category X is a Banach bundle that is also an involutive category E fibered over the involutive category X with continuous fiberwise bilinear compositions and fiberwise conjugate-linear involutions satisfying properties (1), (2) and (3) as above. We say that the Fell bundle is rank-one if Ex is one-dimensional for all x ∈ X. 4By involutive category we mean a category X equipped with an involution i.e. an object preserving contravariant functor ∗ : X → X such that (x∗)∗ = x for all x ∈ X. If X has a topology we also require composition and involution to be continuous. X is an involutive inverse category if xx∗x = x for all x ∈ X. 5 Note that a C-category C can always be seen as a Fell bundle over the maximal equivalence relation ObC × ObC of its objects, with fibers CAB, for all (A, B) ∈ ObC × ObC. Conversely, a Fell bundle whose base is such an equivalence relation can always be seen as a C-category. Definition 3.2. A topological spaceoid (or simply a spaceoid, for short) (E, π, X) is a unital rank-one Fell bundle over the product involutive topological category X := ∆X × RO where ∆X := {(p, p) p ∈ X} is the minimal equivalence relation of a compact Hausdorff space X and RO := O × O is the maximal equivalence relation of a discrete space O. With a slight abuse of notation, the arrows of the base involutive category X of a full spaceoid will simply be denoted by pAB := ((p, p), (A, B)) ∈ ∆X × RO. Note that, since a constant finite-rank Banach bundle over a locally compact Hausdorff space is locally trivial [FD, Remark I.13.9] and hence a vector bundle, a topological spaceoid is a Hermitian line bundle over X and is a disjoint union of the Hermitian line bundles (EAB, πEAB , XAB), with XAB := ∆X × {AB} and EAB := π−1(XAB). Furthermore, a topological spaceoid can be seen as a one-dimensional C-category that is a coproduct (in the category of small C-categories) of the "continuous field" of the one- dimensional C-categories Ep := π−1(Xp), where Xp := {(p, p)} × RO, for all p ∈ X. A morphism of spaceoids5 (f, F ) : (E1, π1, X1) → (E2, π2, X2) is a pair (f, F ) where: • f := (f∆, fR) with f∆ : ∆1 → ∆2 being a continuous map of topological spaces and fR : R1 → R2 an isomorphism of equivalence relations; • F : f •(E2) → E1 is a fiberwise linear continuous ∗-functor such that π1 ◦ F = πf 2 , where (f •(E2), πf 2 , X1) denotes the standard f -pull-back6 of (E2, π2, X2). Topological spaceoids constitute a category if compositions and identities are given by (g, G) ◦ (f, F ) := (g ◦ f, F ◦ f •(G) ◦ ΘE3 g,f ) and ι(E, π, X) := (ιX, ιπ X), where ΘE3 g,f : (g ◦ f )•(E3) → f •(g•(E3)) is the natural isomorphisms between standard pull- backs given by ΘE3 g,f (x1, e3) := (x1, (f (x1), e3)), for all (x1, e3) ∈ (g ◦ f )•(E3), and f •(G), thanks to the functorial properties of pull-backs, is defined on the standard pull-back by f •(G)(x1, (x2, e3)) := (x1, G(x2, e3)), for all (x1, (x2, e3))) ∈ f •(g•(E3)). 4 The section functor Γ Here we are going to define a section functor Γ : T → A that to every spaceoid (E, π, X), with X := ∆X × RO, associates a commutative full C-category Γ(E) as follows: • ObΓ(E) := O; • for all A, B ∈ ObΓ(E), HomΓ(E)(B, A) := Γ(XAB; E), where Γ(XAB; E) denotes the ∈ EpAB of the set of continuous sections σ : ∆X × {(A, B)} → E, σ : pAB 7→ σAB restriction (EAB, πEAB , XAB) of (E, π, X) to the base space XAB. p 5Morphisms of spaceoids can be seen as examples of J. Baez notion of spans (in this case, a span of the Fell bundles of the spaceoids). 6Recall that f •(E2) := {(pAB , e) ∈ X1 × E2 f (pAB ) = π2(e)} with f ◦ πf 2 and f π2 2 (pAB , e) := pAB and f π2 (pAB , e) := e. If E2 is a Fell bundle over X2, f •(E2) is 2 = π2 ◦ f π2 , where πf are defined on f •(E2) by πf a Fell bundle over X1. 6 • for all σ ∈ HomΓ(E)(B, A) and ρ ∈ HomΓ(E)(C, B): σ ◦ ρ : pAC 7→ (σ ◦ ρ)AC σ∗ : pBA 7→ (σ∗)BA p kσAB p kE, kσk := sup p∈∆X := σAB p )∗, p := (σAB p ◦ ρBC p , with operations taken in the total space E of the Fell bundle. In the following, since for all σ ∈ Γ(E) = UAB Γ(E)AB, the discrete indeces AB are already implicit in the specification of the section σ ∈ Γ(E)AB , we will simply use the shorter notation σp := σAB to denote the evaluation of the section σ at the point pAB ∈ X. p By construction, the commutative C-category Γ(E) so obtained has sections of Hermitian line bundles on a compact Hausdorff space as Hom spaces, and thus it is full. We extend now the definition of Γ to the morphisms of T . Let (f, F ) be a morphism in T from (E1, π1, X1) to (E2, π2, X2). Given σ ∈ Γ(E2), we consider the unique section f •(σ) : X1 → f •(E2) such that f π2 ◦ f •(σ) = σ ◦ f and the composition F ◦ f •(σ). In this way we have a map Γ(f, F ) : Γ(E2) → Γ(E1), Γ(f, F ) : σ 7→ F ◦ f •(σ), ∀σ ∈ Γ(E2). Proposition 4.1. For any morphism (E1, π1, X1) map Γ(f, F ) : Γ(E2) → Γ(E1) is a morphism in the category A . The pair of maps Γ : (E, π, X) 7→ Γ(E) and Γ : (f, F ) 7→ Γ(f, F ) gives a contravariant functor from the category T of spaceoids to the category A of small full commutative C-categories. (f, F ) −−−−→ (E2, π2, X2) in the category T , the Proof. Let (E1, π1, X1) (f, F ) −−−−→ (E2, π2, X2) and (E2, π2, X2) (g, G) −−−→ (E3, π3, X3) be two com- (ιX, ιπ −−−−−→ (E, π, X) be the identity X) posable morphisms in the category T and let (E, π, X) morphism of (E, π, X). To complete the proof we must show that Γ(g, G)◦(f, F ) = Γ(f, F ) ◦ Γ(g, G), Γ(ιX, ιπ X) = ιΓ(E), and these are obtained by tedious but straightforward calculations. 5 The spectrum functor Σ This section is devoted to the construction of a spectrum functor Σ : A → T that to every commutative full C-category C associates its spectral spaceoid Σ(C). Letting C be a C-category, we denote by RC the topologically discrete ∗-category C/C ≃ RObC and by CRC := ρ•(C) the one-dimensional C-category pull-back of C (considered as a C-category with only one object •) under the constant map ρ : RC → {•}. Note that, via the canonical projection ∗-functor C → RC := C/C, from the defining property of pull-backs there is a bijective map ω 7→ eω between the set of C-valued ∗-functors [C; C] and the object-preserving elements in the set of CRC-valued ∗-functors [C; CRC]. By definition, two ∗-functors ω1, ω2 in [C; C] are unitarily equivalent, see P. Mitchener [M1, Section 2], if there exists a "unitary" natural transformation A 7→ νA ∈ T between them. Note that the set Iω := {x ∈ C ω(x) = 0}, which is also equal to {x ∈ C ω(x∗x) = 0}, is an ideal in C and Iω1 = Iω2 if (and only if) the equivalence classes [ω1] and [ω2] coincide. We also need the following lemmas whose routine proofs are omitted:7 7Note that, for ω ∈ [C; C] and A, B ∈ ObC, we denote by ωAB the restriction of ω to CAB . 7 Lemma 5.1. If ω, ω′ ∈ [C; C] are unitarily equivalent, there is a unique map ψ : RC → T such that ω′ A , where ν is a unitary natural transformation from ω to ω′, and the map ψ : AB 7→ ψAB is a ∗-functor: AB = ψAB · ωAB for all AB ∈ RC given by ψAB = νBν−1 ψABψBC = ψAC , ψAB = ψ−1 BA, ψAA = 1C. (5.1) Conversely, given a ∗-functor ψ ∈ [RC; T], two ∗-functors ω, ω′ such that ω′ are unitarily equivalent. AB = ψABωAB Lemma 5.2. Every object preserving ∗-automorphism γ of the C-category CRC is given by the multiplication by an element ψ ∈ [RC; T] i.e. γ(x) = ψAB · x for all x ∈ (CRC)AB. Proposition 5.3. Two ∗-functors ω, ω′ ∈ [C; C] are unitarily equivalent if and only if ωAA = ω′ AA for all A ∈ ObC. Proof. By lemma 5.1, if [ω] = [ω′], then ω′ Let ω, ω′ ∈ [C; C] and suppose that ωAA = ω′ AA = ψAA · ωAA = ωAA, for all objects A. AA, for all A ∈ ObC. Consider the corresponding object-preserving CRC-valued ∗-functors eω, eω′ ∈ [C; CRC]. Note that Ker(eω) = Iω = Iω′ = Ker(eω′) and hence, ωAB, eωAB are nonzero if and only if ω′ nonzero for all AB ∈ RC, by theorem 2.2 we have two ∗-isomorphisms AB are nonzero. If ωAB is AB, eω′ C/ Ker(ω) ω−→ CRC ω′ ←−− C/ Ker(ω′). From lemma 5.2 there is a ψ ∈ [RC; T] such that ω′ = ψ · ω and hence also ω′ = ψ · ω so that the proposition follows from lemma 5.1. In order to complete the proof, notice that the images of ω and ω′ coincide and then apply the above argument to the connected components of the image category. Given x ∈ CAA for some object A, by evaluation in x we mean the map evx : [C; C] → C defined by ω 7→ ω(x). Proposition 5.4. The set [C; C] of C-valued ∗-functors ω : C → C, with the weakest topology making all evaluations continuous, is a compact Hausdorff topological space. Proof. Note that for all ω ∈ [C; C] and for all x ∈ CAB, ω(x) = qω(x)ω(x) = pω(x∗x) = pωAA(x∗x) ≤ pkx∗xk = pkxk2 = kxk, because ωAA is a state over the C-algebra CAA. Hence [C; C] is a subspace of the compact Hausdorff space Qx∈C Dkxk, where Dkxk is the closed ball in C of radius kxk, and it is easy to check that it is closed. Let Spb(C) := {[ω] ω ∈ [C; C]} denote the base spectrum of C, defined as the set of unitary equivalence classes of ∗-functors in [C; C]. It is a compact space with the quotient topology induced by the map ω 7→ [ω]. To show that Spb(C) is Hausdorff it is enough to note that, by proposition 5.3, if [ω] 6= [ω′], there exists at least one object A such that ωAA 6= ω′ AA and so there exists at least one evaluation evx with x ∈ CAA such that evx(ω) 6= evx(ω′). Since, for x ∈ CAA, evx induces a well-defined map on the quotient space Spb(C) by [ω] 7→ ω(x), the result follows. Proposition 5.5. Let C be a full commutative C-category. For all A ∈ ObC, there exists a natural bijective map, between the base spectrum of C and the usual Gel'fand spectrum Sp(CAA) of the C-algebra CAA, given by the restriction AA : ω 7→ ωCAA. In particular, for all objects A ∈ ObC, one has Spb(C)AA = Sp(CAA) ≃ Spb(CAA). 8 Proof. By proposition 5.3, the correspondence [ω] 7→ ωAA is well defined. We show that the map [ω] 7→ ωAA is injective. Given ω, ω′ ∈ [C; C] with ωAA = ω′ AA, we know from [BCL3, Proposition 2.30], that ωBB(x) = ωAA(φAB(x)), for all x ∈ CBB, for all B ∈ ObC, where φAB : CBB → CAA is the canonical isomorphism associated to the imprimitivity bimodule CBA. It follows that ωBB = ωAA ◦ φAB = ω′ BB, for all B ∈ ObC and, by proposition 5.3, we see that [ω] = [ω′]. AA ◦ φAB = ω′ We show that the function [ω] 7→ ωAA is surjective. Let ωo ∈ Sp(CAA). Define J to be the involutive ideal in C generated by Ker(ωo). One can see that J ∩ CAA = JAA = Ker(ωo). Since the quotient ∗-functor π : C → C/J is bijective on the objects and C is full, C/J is full. Since CAA/JAA is one-dimensional, the quotient C-category C/J is one-dimensional. If γ : C/J → C is a C-valued ∗-functor as in lemma 2.3, γ ◦ π restricted to CAA must be ωo since it vanishes on J ∩ CAA. Theorem 5.6. Let C be a full commutative C-category. For every A ∈ ObC, the bijective map AA : Spb(C) → Sp(CAA) given by [ω] 7→ ωAA is a homeomorphism between Spb(C) and the Gel'fand spectrum Sp(CAA) of the unital C-algebra CAA. Proof. Since both Spb(C) and Sp(CAA) are compact Hausdorff spaces, and the map AA is bijective, it is enough to show that AA: Spb(C) → Sp(CAA) is continuous. Since Spb(C) is equipped with the quotient topology induced by the projection map π : [C; C] → Spb(C), the map AA is continuous if and only if AA ◦π : [C; C] → Sp(CAA) is continuous. The spaces [C; C] and Sp(CAA) are equipped with the weakest topology making the evaluation maps continuous. It follows that the continuity of AA ◦π is equivalent to the continuity of evx = evx ◦ AA ◦π : [C; C] → C for all x ∈ CAA. Since evx : [C; C] → C is continuous, the result is established. Let XC := ∆C × RC be the direct product equivalence relation of the compact Hausdorff ∗-category ∆C := ∆Spb(C) and the topologically discrete ∗-category RC := C/C ≃ RObC. With a slight abuse of notation, we write AB ∈ RC for the arrow CAB/CAB in RC and denote ([ω], AB) = ([ω], CAB/CAB) ∈ XC simply by ωAB. We define EC as the disjoint union over [ω] ∈ ∆C of the quotients C/Iω. In formulae: EC ω := C Iω , EC := ] EC ω = ] EC ωAB , πC : EC → XC, ω∈∆C πC : e 7→ ωAB, ωAB∈XC ∀e ∈ EC ωAB , where EC ωAB := CAB/IωAB , with IωAB := Iω ∩ CAB. Proposition 5.7. The triple (EC, πC, XC) is naturally equipped with the structure of a unital rank-one Fell bundle over the topological involutive inverse category XC. e0 Proof. Define on EC the topology whose fundamental system of neighborhoods are the sets U O, x0, ε := {e ∈ EC πC(e) ∈ O, ∃x ∈ C : x(πC(e)) = e, kx − x0k∞ < ε}, where e0 ∈ EC, O is open in XC, ε > 0, x0 ∈ C with x0(πC(e0)) = e0 and where x denotes the Gel'fand transform of x defined in section 6.1. This topology entails that a net (eµ) is convergent to the point e in EC if and only if the net πC(eµ) converges to πC(e) in XC and, for all possible Gel'fand transforms x0 "passing" through e0, there exists a net of Gel'fand transforms xµ, "passing" through eµ, that uniformly converges, on every neighborhood of πC(e0), to x0. With such a topology the (partial) operations on EC i.e. sum, scalar multiplication, product, involution, inner product (and hence norm) become continuous and (EC, πC, XC) becomes a Banach bundle. 9 Since every sub-equivalence relation of XC is a disjoint union of "grids" {[ω]} × RC whose inverse image under πC is the one-dimensional C-category C/ Ker(ω), (EC, πC, XC) is a rank-one unital Fell bundle over the equivalence relation XC and hence a spaceoid. To a commutative full C-category C we have associated a topological spectral spaceoid Σ(C) := (EC, πC, XC). We extend now the definition of Σ to the morphisms of A . Let Φ : C → D be an object-bijective ∗-functor between two small commutative full C-categories with spaceoids Σ(C), Σ(D) ∈ T and define a morphism ΣΦ : Σ(D) category T as follows. We set λΦ : XD (λΦ R : RD → RC is the isomorphism of equivalence relations given by λΦ ∆ : ∆D → ∆C (since ω 7→ ω ◦Φ is continuous and preserves equivalence by unitary natural transformations) is the well-defined continuous map given by λΦ R(AB) := Φ−1(A)Φ−1(B), for AB ∈ RD, and where λΦ ∆([ω]) := [ω ◦ Φ] ∈ ∆C, for all [ω] ∈ ∆D. (λΦ,ΛΦ) −−−−−→ Σ(C) in the −−−−−→ XC where λΦ ∆,λΦ R) Since, for ω ∈ [D; C], the ∗-functor Φ : C → D induces a continuous field of quotient ∗-functors Φω : C/IΦ◦ω → D/Iω between one-dimensional C-categories, we can define8 ΛΦ : (λΦ)•(EC) → ED as the disjoint union of the ∗-functors Φω, for [ω] ∈ ∆D and note that it is a continuous fiberwise linear ∗-functor. Proposition 5.8. For any morphism C −−→ Σ(C) is a morphism of spectral spaceoids. The pair of maps Σ : C 7→ Σ(C) and Σ : Φ 7→ ΣΦ gives a contravari- ant functor Σ : A → T , from the category A of object-bijective ∗-functors between small commutative full C-categories to the category T of spaceoids. Φ−→ D in A , the map Σ(D) ΣΦ Proof. We have to prove that Σ is antimultiplicative and preserves the identities. If Φ : C1 → C2 and Ψ : C2 → C3 are two ∗-functors in A , by definition, ΣΨ◦Φ = (λΨ◦Φ, ΛΨ◦Φ) = (cid:16)λΦ◦λΨ, ΛΨ◦(λΨ)•(ΛΦ)◦ΘEC1 λΦ,λΨ(cid:17) = (λΦ, ΛΦ)◦(λΨ, ΛΨ) = ΣΦ◦ΣΨ. if ιC : C → C is the identity functor of the C-category C, then the morphism Also, ΣιC = (λιC, ΛιC) is the identity morphism of the spaceoid Σ(C). 6 Horizontal Categorification of Gel'fand Duality 6.1 Gel'fand Transform For a given C-category C in A , we define a horizontally categorified version of Gel'fand x[ω] := x + IωAB , for all transform as GC : C → Γ(Σ(C)) given by GC : x 7→ x where x ∈ CBA. Clearly GC : C → Γ(Σ(C)) is an object bijective ∗-functor. Lemma 6.1. Let C be a commutative C-category and Co a subcategory of C which is a full C-category such that Co AB = CAB for all A, B ∈ ObC. AA = CAA for all A ∈ ObC = ObCo. Then Co Proof. By the fullness of the bimodule ACo B there is a sequence of pairs uj, vj ∈ ACo j=1(xu∗ j )vj ∈ ACo B such B for all that ιB = P∞ x ∈ ACB, because xu∗ j=1 u∗ j vj. We have x = xιB = xP∞ A and so (xu∗ j ∈ ACA = ACo j vj = P∞ B for all j. j=1 u∗ j )vj ∈ ACo Theorem 6.2. The Gel'fand transform GC : C → Γ(Σ(C)) of a commutative full C-category C is a full faithful isometric ∗-functor. 8Note that (λΦ)•(EC) is the disjoint union of the continuous field of one-dimensional C-categories C/IΦ◦ω. 10 Proof. The proof follows from the fact that the Gel'fand transform GC, when restricted to any "diagonal" commutative unital C-algebra CAA can be "naturally identified" with the usual Gel'fand transform of CAA via the homeomorphism [ω] 7→ ωAA (see proposition 5.5 and theorem 5.6). To prove the faithfulness of GC, let x ∈ CBA with bx = 0. We have dx∗x = bx∗bx = 0 so that the usual Gel'fand transform of x∗x ∈ CAA is zero and, by Gel'fand isomorphism theorem applied to the C-algebra CAA, we have x∗x = 0 and hence x = 0. The "image" GC(C) of GC is a subcategory of the commutative full C-category Γ(Σ(C)) that is clearly a commutative full C-category on its own. By lemma 6.1, the ∗-functor GC is full as long as GC(CAA) = Γ(Σ(C))AA, for all objects A ∈ ObC and this follows again by the usual Gel'fand isomorphism theorem applied to the C-algebra CAA. For the isometry of GC we note that for all x ∈ CBA, since the Gel'fand transform restricted to the C-algebra CAA is isometric, we have kGC(x)k2 = kdx∗xk = kx∗xk = kxk2. 6.2 Evaluation Transform Given a topological spaceoid (E, π, X), we define a horizontally categorified version of eval- uation transform EE : (E, π, X) (ηE,ΩE) −−−−−→ Σ(Γ(E)) as follows: R : RO → RΓ(E) is the canonical isomorphism RO = RObΓ(E) ≃ Γ(E)/Γ(E), explicitly: • ηE ηE R(AB) := Γ(E)AB/Γ(E)AB, ∀AB ∈ RO that is, according to the running notation, written as an identity map ηE R(AB) = AB ∈ RΓ(E). • ηE ∆ : ∆X → ∆Γ(E) is given by ηE ∆ : p 7→ [γp ◦ evp] ∀p ∈ ∆X , where the evaluation map evp : Γ(E) → ⊎(AB)∈RO EpAB given by evp : σ 7→ σp is a ∗-functor with values in a one-dimensional C-category that determines9 a unique point [γp ◦ evp] ∈ ∆Spb(Γ(E)). • ΩE : (ηE)•(EΓ(E)) → E is defined by ΩE : (pAB, σ + IηE(pAB )) 7→ σp, ∀σ ∈ Γ(E)AB, ∀pAB ∈ X. In particular, with such definitions we can prove that the spectrum functor is representative: Theorem 6.3. The evaluation transform EE : (E, π, X) → Σ(Γ(E)), for all spaceoids (E, π, X), is an isomorphism in the category of spaceoids. Proof. Note that (EAA, π, X) is naturally isomorphic to the trivial C-bundle over X and thus there is an isomorphism of the C-algebras Γ(E)AA and C(X) that "preserves" evaluations. Clearly the map ζA : ∆X → Sp(Γ(E)AA) given by ζA(p) := AA ◦ ηE ∆(p) = γpAA ◦ evpAA , coincides with the usual Gel'fand evaluation homeomorphism for the diagonal C-algebra Γ(E)AA and hence, by proposition 5.5, ηE AA ◦ ζA is also a homeomorphism. ∆ = −1 For every element e ∈ E, we have π(e) ∈ ∆X × RO and, since a spaceoid is actually a vector bundle, it is always possible to find a section σ ∈ Γ(E) such that σπ(e) = e. For any such section we consider the element σ + IηE(π(e)) ∈ Γ(E)/IηE(π(e)) =: EΓ(E) ηE(π(e)) (note that the element does not depend on the choice of σ ∈ Γ(E) such that σπ(e) = e) and in this way we have a map Θ : E → EΓ(E) by Θ : e 7→ σ + IηE(π(e)). The map Θ uniquely induces a function ΞE : E → (ηE)•(EΓ(E)) with the standard ηE-pull-back of EΓ(E) given by 9By lemma 2.3, there is always a C-valued ∗-functor γp : Ep → C and by proposition 5.3 any two compositions of evp with such ∗-functors are unitarily equivalent because they coincide on the diagonal C-algebras EpAA . 11 ΞE(e) := (π(e), Θ(e)). By direct computation the map ΞE is an "algebraic isomorphism" of Fell bundles10 whose inverse is ΩE. ηE(pj The continuity of ΩE is equivalent to that of eΩE : EΓ(E) → E, eΩE(σ + IηE(pAB )) := σp, with ABj ) in EΓ(E) converging to the point σ + IηE(pAB ) σ ∈ Γ(E)AB. Given a net j → σj + I in the topology defined in proposition 5.7, without loss of generality we can assume that j → σj is uniformly convergent to σ in a neighborhood U of ηE(pAB). This means that, for all ǫ > 0, kσj([ω]AB) − σ([ω]AB)k < ǫ for [ω]AB ∈ U for all sufficiently large j. Since RΓ(E) is discrete, the net ABj is eventually equal to AB and since ηE is a homeomorphism, pj ABj )) = (σj)pj eventually lies in any neighborhood of pAB and hence the net eΩE(σj + I converges to eΩ(σ + IηE(pAB )) = σp in the Banach bundle topology of E. Since ΩE is an isometry, it follows from [FD, Proposition 13.17] that its inverse is continuous too and hence the evaluation transform EE := (ηE, ΩE) is an isomorphism of spaceoids. AB ηE(pj 6.3 Duality Theorem 6.4. The pair of functors (Γ, Σ) provides a duality between the category T of morphisms between spaceoids that are object-bijective on the discrete part of the objects and the category A of object-bijective ∗-functors between small commutative full C-categories. Proof. To see that the map G : C 7→ GC (that to every C ∈ ObA associates the Gel'fand transform of C) is a natural isomorphism between the identity endofunctor IA : A → A and the functor Γ ◦ Σ : A → A we have to show that, given an object-bijective ∗-functor Φ : C1 → C2, the identity ΓΣΦ(GC1 (x)) = GC2 (Φ(x)) holds for any x ∈ C1, i.e. the commutativity of the diagram: C1 Φ C2 GC1 GC2 Γ(Σ(C1)) ΓΣΦ / Γ(Σ(C2)), that follows from this direct computation: ΓΣΦ(GC1 (x))[ω2] = ΛΦ(cid:16)(λΦ)•(x)[ω2](cid:17) = ΛΦ(cid:16)[ω2]A2B2 , x(λΦ([ω2]A2B2 ))(cid:17) = ΛΦ(cid:16)[ω2]A2B2 , x + IλΦ([ω2]A2B2 )(cid:17) = (cid:16)[ω2]A2B2 , Φ(x) + I[ω2]A2B2(cid:17) = GC2 (Φ(x))[ω2]. To see that the map E : E 7→ EE (that to every spaceoid (E, π, X) associates its evaluation transform EE) is a natural isomorphism between the identity endofunctor IT : T → T and the functor Σ ◦ Γ : T → T we must prove, for any given morphism of spaceoids (f, F ) from (E1, π1, X1) to (E2, π2, X2), the commutativity of the diagram: (E1, π1, X1) (f,F ) (E2, π2, X2) EE1 =(ηE1 ,ΩE1 ) Σ(Γ(E1)) ΣΓ(f,F) =(λ Γ(f,F) ,ΛΓ(f,F) ) EE2 =(ηE2 ,ΩE2 ) / Σ(Γ(E2)). 10By this we mean that ΞE : E → (ηE)•(EΓ(E)) is a fiber preserving map between bundles, over the same base space X, that is also a bijective fiberwise linear ∗-functor between the total spaces. 12 / / / / / / The proof amounts to showing the equalities λΓ(f,F) ◦ ηE1 = ηE2 ◦ f, ΩE1 ◦ (ηE1 )•(ΛΓ(f,F) ) ◦ Θ1 = F ◦ f •(ΩE2 ) ◦ Θ2, (6.1) Γ(f,F) ,ηE1 λ , Θ2 := ΘEΓ(E2 ) ηE2 ,f . where Θ1 := ΘEΓ(E2 ) Since for every point pAB ∈ X1, we have λΓ(f,F) ◦ηE1(pAB) = ([γp ◦evp ◦Γ(f,F )], fR(AB)) and ηE2 ◦f (pAB) = ([γf (p)◦evf (p)], fR(AB)), the first equation is a consequence of proposition 5.3. The second equation is then proved by a lengthy but elementary calculation. Remark 6.5. Finally, note that, although for simplicity we only described a spectral theory for commutative full C-categories, it is perfectly viable and there are no substantial obstacles to the development of a spectral theory for commutative full "non-unital" C-categories11 (as defined by P. Mitchener [M2]). In this case the base spectrum is only locally compact and we have to deal with a locally compact version of topological spaceoids (so, for example, only sections "going to zero at infinity" are considered in the definition of the section functor). 6.4 Horizontal Categorification The usual Gel'fand-Naımark duality theorem is easily recovered from our result identifying a compact Hausdorff topological space X with the trivial spaceoid TX with total space XX × C and base category XX := ∆X ×ROX where OX := {X} is a discrete space with only one point X; and similarly, identifying a unital commutative C-algebra A with the full commutative C-category CA with one object via HomCA := A and ObCA := {A}. More precisely, the duality (Γ, Σ) between the categories T and A is a "horizontal categorification" of the usual Gel'fand-Naımark duality in the sense specified by the following result whose proof is absolutely elementary: Theorem 6.6. Let T (1) denote the full subcategory of T consisting of those trivial spaceoids TX := XX × C, where XX := ∆X × ROX , OX := {X} and X is a compact Hausdorff space. Let A (1) denote the full subcategory of A consisting of those full commutative small one- object C-categories CA with morphisms HomCA := A, objects ObCA := {A} and composition involution and norm induced from those in the commutative unital C-algebra A. The natural duality (Γ, Σ) between the categories T , A restricts to a duality (Γ(1), Σ(1)) between the categories T (1), A (1) i.e. the following pair of diagrams of functors is commutative: T (1) T Γ(1) Σ(1) Γ Σ / A (1) / A , where Γ(1) and Σ(1) denote the restrictions of the functors Γ and Σ and the vertical arrows denote the inclusion functors of the respective categories. The category T1 of continuous maps between compact Hausdorff spaces is isomorphic to the category T (1) via the functor T : T1 → T (1) defined as follows: • to every compact Hausdorff space X, T associates the spaceoid TX := (EX , πX , XX ) that is the trivial bundle with fiber C over the space XX := ∆X × {(X, X)}, 11Strictly speaking these are not categories, since they are lacking identities, but they otherwise satisfy all the other properties listed in the definition of a C-category. 13 _ o _ o • to every continuous map f : X → Y between compact Hausdorff spaces, T associates the morphism of spaceoids T(f ) : TX → TY defined by T(f ) := (T(f )X, T(f )E) where T(f )X : pXX 7→ f (p)Y Y , for all p ∈ X, and T(f )E : T(f )• X(EY ) → TX denotes the canonical isomorphism between trivial line bundles over X.12 The category A1 of unital ∗-homomorphisms of unital commutative C-algebras is isomor- phic to the category A (1) via the functor C : A1 → A (1) that to every unital commuta- tive C-algebra A associates the C-category CA and that to every unital ∗-homomorphism φ : A → B associates the ∗-functor C(φ) : CA → CB given on arrows by C(φ)(x) := φ(x), for all x ∈ A, and on objects by C(φ)o : A 7→ B. The functors C ◦ Γ and Γ(1) ◦ T are naturally equivalent via the natural transformation that to every X associates the canonical isomorphism between CC(X;C) and Γ(TX ). The functors T ◦ Σ and Σ(1) ◦ C are naturally equivalent via the natural transformation that to every A associates the canonical isomorphism between TSp(A) and Σ(CA). 7 Examples and Applications Commutative full C-categories are abundant, just to mention a few examples: • Every Abelian unital C-algebra A gives a commutative full C-category CA with only one object (as already mentioned in subsection 6.4). • Examples of commutative full C-categories with two objects can be obtained via the following construction (see L. Brown-P. Green-M. Rieffel [BGR]) of the "linking C-category" L(M) of an imprimitivity Hilbert C-bimodule AMB over two commu- tative unital C-algebras A, B. Let AMB be an imprimitivity C-bimodule over unital commutative C-algebras A, B. Denote by BM+ A its Rieffel dual and by ι : M → M+ the canonical bijective map such that ι(a·x·b) = b∗ ·ι(x)·a∗ for all a ∈ A, b ∈ B, x ∈ M (see for example [BCL3, Proposition 2.19] for more details). The linking C-category L(M) of M is the full commutative C-category with two objects A, B and morphisms given by L(M)AA := A, L(M)BB := B, L(M)AB := M, L(M)BA := M+ where the only non-elementary operations are the involutions of elements of M and M+, given by x∗ := ι(x) and y∗ := ι−1(y) for all x ∈ M, y ∈ M+ and the compositions between elements of M and M+ that are given via their respective A-valued and B-valued inner products as follows: x ◦ y := Ahx y∗i and y ◦ x := hy∗ xiB, for all x ∈ M, y ∈ M+. Making use of the canonical isomorphisms of imprimitivity bimodules (M ⊗ N) ⊗ T ≃ (M ⊗ N ⊗ T) ≃ M ⊗ (N ⊗ T) and (M ⊗ N)+ ≃ N+ ⊗ M+ for the definition of the compositions via tensor products and "contractions", we can generalize the previous construction of linking C-category to the case of an arbi- trary (finite) collection of objects. In practice, given a (finite) family of commutative unital C-algebras A1, . . . , An and a family of imprimitivity Hilbert C-bimodules A1MA2 , . . . , An−1MAn , the "linking C-category" L(A1 MA2, . . . , An−1MAn) is the full commutative C-category with n objects B1, . . . , Bn, where L(A1 MA2, . . . , An−1MAn )Bj Bk := line bundle on XX . 14   Aj MAj+1 ⊗ · · · ⊗ Ak−1 MAk , Aj, (Ak MAk+1 ⊗ · · · ⊗ Aj−1 MAj )+, for j = k, for j < k, for k < j, 12Remember that the pull-back of the trivial line bundle TY under the homeomorphism T(f )X is a trivial The examples in the next item are rather natural, especially for those who are familiar with the Doplicher-Roberts abstract duality theory for compact groups. • Let G be a compact group, and consider the C-category Rep(G) with objects the unitary representations of G on Hilbert spaces and arrows their intertwiners (actually, Rep(G) is a W-category). Then the full subcategory of Rep(G) whose objects are the multiplicity-free representations is commutative. Moreover, a category of multiplicity- free representations is full whenever all the objects are equivalent. Let A be a unital C-algebra. A category of nondegenerate ∗-representations of A on Hilbert spaces is commutative whenever all the objects are multiplicity-free (see In addition, such a commutative W-category is full when- e.g. [Ar, Chapter 2]). ever all the objects are equivalent. If A is commutative with metrizable spectrum, a category of nondegenerate representations of A on separable Hilbert spaces that is both commutative and full can be interpreted in terms of a family of equivalent finite Borel measures on the spectrum of A [Ar, Theorems 2.2.2 and 2.2.4]. This fact can be generalized to GCR algebras [Ar, Chapter 4]. Let A be a unital C-algebra, and consider the C-category End(A) with objects the unital ∗-endomorphisms of A and Banach spaces of arrows (ρ, σ) = {x ∈ A xρ(a) = σ(a)x, ∀a ∈ A}. The category Aut(A), i.e. the full subcategory of End(A) with objects the unital ∗-automorphisms of A, is clearly commutative, as (ρ, ρ) equals the center of A, for every automorphism ρ.13 Again, a subcategory of Aut(A) is full whenever all the objects are equivalent. • P. Mitchener [M2] associates C-categories C∗(G) and C∗ r (G) to a discrete groupoid G. It is easy to see that these categories are commutative exactly when all the stabilizer subgroups of the groupoid G are Abelian (i.e. GAA is an Abelian group for all objects A of the groupoid). In that case they are full if the groupoid G is transitive (i.e. GAB 6= ∅, for every pair of objects A, B ∈ ObG). This example admits an immediate generalization to the case of involutive categories. Given an involutive category X , the set of C-valued maps on X with finite support contained in any one of the sets XAB, with A, B ∈ ObX , is the family of morphisms of a ∗-category C∗ o (X ), with objects ObX , where the composition is the usual convolution of finite sequences and the involution is defined via (αx)∗ := αx∗. The ∗-category o (X ) has a natural continuous left-regular action on L2(X ) (that is the family of C∗ Hilbert spaces, indexed by A ∈ ObX , obtained by completing ⊕B∈ObX C∗ o (X )AB under the inner product h(αx) (βy)i := Px,y αx∗ · βy) and its C-completion in the induced operator norm is the C-category C∗ o (X ) under the supremum of all the C-norms induced by its continuous representations we obtain the C-category C∗(X ). The categories C∗(X ) and C∗ r (X ) are commutative whenever XAA is commutative for all objects A ∈ ObX and they are full if and only if the category X is saturated in the following sense XAB ◦ XBC = XAC for all objects A, B, C ∈ ObX . r (X ). Taking the C-completion of C∗ • Given any non-diagonal arrow x in a C-category C, the C-subcategory C(x) of C generated by x is full and commutative, see theorem 7.1 and the related discussion (notice that C(x) might well be a non-unital C-category if x is not invertible). 13Of course, this is still true for all irreducible morphisms. 15 • The category of Hermitian line bundles over a compact Hausdorff space X with line bundle morphisms as arrows is a C-category, which turns out to be full and commu- tative. We now deal with specific examples and constructions of spaceoids. Note that (although every topological spaceoid is of course isomorphic to the spectrum of a commutative full C-category) the examples mentioned here below have in principle no direct relation with C-categories and arise from some well-known constructs in (algebraic) topology. • As already described in detail in subsection 6.4, the most elementary examples of spaceoid are those associated to every compact Hausdorff topological spaces X via the trivial Hermitian line bundles TX := (EX , πX , XX) over the topological space XX := ∆X × {(X, X)} with total space EX := XX × C and projection πX onto the first factor. AA := (∆X × {AA}) × C, EE BB := (∆X × {BB}) × C, EE • Every (possibly non-trivial) Hermitian line bundle (E, π, X) over a compact Hausdorff topological space X uniquely determines a spaceoid L(E) := (EE, πE, XE), called its "linking spaceoid", in the following canonical way. Define the base topological involutive category as XE := ∆X × RO with O := {A, B}, and as total space con- sider EE AB := E × {AB}, BA := E+ × {BA}, where by (E+, π+, X) we denote the Hermitian line bundle dual EE to (E, π, X) (this is the line bundle with fibers E+ p := (Ep)+ given by the dual of the inner product vector space Ep). Define on the total space EE the operations of involution as usual fiberwise conjugation on EE BB and by the canonical anti- linear map induced between EE BA by the natural fiberwise anti-isomorphism between E and E+. Finally define the composition on the total space as the usual fiberwise product on EE BA, via the canon- ical contraction between E and E+ as eAB ◦ e′ BA ◦ eAB := e′(e)BB, for all e ∈ Ep and e′ ∈ E+ BB and, between elements in EE BA := e′(e)AA and e′ AB and EE AA and EE AB and EE AA, EE p , with p ∈ X. • In perfect parallel with the construction of the linking C-category for a family of Hilbert C-bimodules, the previous construction of the linking spaceoid of a Hermitian line bundle can be generalized in order to define the "linking spaceoid" L(E1, . . . , En) of a family of (possibly non-trivial) Hermitian line bundles (E1, π1, X), . . . , (En, πn, X) over the same compact Hausdorff space X. For this purpose, denoting L(E1, . . . , En) := (EE1...En base topological ∗-category as XE1...En set of n + 1 elements. We then define the total space EE1...En specifying its blocks on the topological space XE1...En ), we take the := ∆X × RO, where O := {A1, . . . , An+1} is a of the linking spaceoid = ∆X × {AjAk} as follows: , XE1...En , πE1...En Aj Ak EE1...En Aj Ak :=   for j = k − 1, for j < k − 1, Ej ⊗ · · · ⊗ Ek−1, Ej C × (∆X × {AjAk}), (Ej )+ for k = j − 1, (Ek−1)+ ⊗ · · · ⊗ (Ej)+ for k < j − 1, for j = k, where ⊗ denotes the fiberwise tensor products of line bundles over the same space X. On the total space EE1...En the fiberwise involution and the composition are defined making use of the canonical isomorphisms of line bundles (Ej ⊗Ek)+ ≃ (Ek)+ ⊗(Ej )+ and Ei ⊗ (Ej ⊗ Ek) ≃ Ei ⊗ Ej ⊗ Ek ≃ (Ei ⊗ Ej) ⊗ Ek, for all i, j, k = 1, . . . n, via the 16 contraction dualities between Ej and (Ej)+ for j = 1, . . . , n and the tensor products Ej × Ek → Ej ⊗ Ek for all j, k = 1, . . . , n. Examples of linking spaceoids of Hermitian line bundles, that stay in perfect duality with those of the linking C-categories of imprimitivity Hilbert C-bimodules previ- ously described, can be obtained via a "bivariant" notion of Hermitian line bundle (i.e. a Hermitian line bundle on a base space that is the graph f ⊂ X × Y of a home- omorphism f : X → Y between two compact Hausdorff topological spaces X, Y ) that is developed in more detail in our companion work [BCL3]. • Finally, we briefly introduce here another natural way to produce spaceoids via "associ- ated line bundles" to a suitable categorification of T-torsors. In more details: given an equivalence relation R, consider the family [R; T] of homomorphisms of the groupoid R with values in the torus group T := {α ∈ C α = 1}. Clearly [R; T] is itself an Abelian group with the operation of pointwise multiplication between homomor- phisms. For any compact Hausdorff space X consider a [R; T]-torsor (T, π, X). Since the set [R; C] of C-valued homomorphisms has a natural structure of [R; T]-space with action given by pointwise multiplication, we can construct the "associated bundle" T ×[R;T] [R; C] over the space X whose elements are equivalence classes [(φ, v)] of pairs (φ, v) ∈ T × [R; C] under the equivalence relation (φ, v) ≃ (ψ, w) if and only if there exists g ∈ [R; T] such that φ · g = ψ and v = g · w. Every such "associated bundle" can be seen, simply by rearrangement of variables, as a spaceoid over the base ∆X × R. To obtain the spectral spaceoid of a full commutative C-category C, it is sufficient to take T := [C; C], the set of C-valued ∗-functors on C. Spectral spaceoids can be easily "assembled" starting from the spectra of imprimitivity C-bimodules developed (via Serre-Swan theorem) in [BCL3, Theorem 3.1]. In every full commutative C-category C and for every pair of objects A, B ∈ ObC, the spectrum of the imprimitivity C-bimodule CAB is a Hermitian line bundle (EBA, πBA, RBA) on the graph14 of a unique homeomorphism RBA : Sp(CAA) → Sp(CBB) between the Gel'fand spectra of the two unital commutative C-algebras CAA and CBB. Now the (necessarily disjoint) union of all the graphs RBA ⊂ Sp(CAA) × Sp(CBB) of the homeomorphisms RBA, with A, B ∈ ObC, can be seen as the graph of a new relation (cid:16)SA,B∈ObC RBA(cid:17) ⊂ (cid:0)SB∈ObC Sp(CBB)(cid:1) ×(cid:0)SA∈ObC Sp(CAA)(cid:1) in the set SA∈ObC Sp(CAA) that is the "disjoint union" of the Gel'fand spectra of the diagonal C-algebras CAA with A ∈ ObC. Since the homeomorphisms RBA are given by RBA = BB ◦ −1 AA in terms of the restric- tion homeomorphisms AA : Spb(C) → Sp(CAA) defined in propositions 5.5 and theorem 5.6, of all such Hermitian line bundles (EBA, πBA, RBA) becomes a new Hermitian line bun- the relation SA,B∈ObC RBA is an equivalence relation.15 Furthermore, the "disjoint union" dle UA,B∈ObC(EAB, πAB, RAB) := (cid:16)SA,B∈ObC EBA,SA,B∈ObC πBA,SA,B∈ObC RBA(cid:17) with total space SB,A∈ObC EBA and base space SA,B∈ObC RBA. gories between ∆C × RC and SAB∈ObC RBA. Since, as Hilbert spaces, the fibers E(ωAA,ωBB ) The map τ : [ω]AB 7→ (ωAA, ωBB) provides an isomorphism of topological involutive cate- 14Note that here we are using RBA to denote the homeomorphism RBA : Sp(CAA) → Sp(CBB ) and together its graph RBA ⊂ Sp(CAA) × Sp(CBB ). More generally, we use the same letter R to denote a relation from the set A to the set B and its graph R ⊂ A × B. 15Note that the new equivalence relation SA,B∈ObC RBA is a relation between elements of the union SA∈ObC Sp(CAA) of the Gel'fand spectra not to be confused with the "coarse grained" equivalence relation {(Sp(CAA), Sp(CBB )) A, B ∈ ObC} between the Gel'fand spectra themselves or equivalently with the sub-equivalence relation {RAB A, B ∈ ObC} of the groupoid of homeomorphisms of compact Hausdorff spaces. 17 are given by E(ωAA,ωBB ) := CAB/(CAB ·Ker(ω)) and so coincide with the fibers of the spectral spaceoid Σ(C) on the elements [ω]AB, we see that SA,B∈ObC E(ωAA,ωBB ) can be naturally equipped with the structure of C-category and so the bundle UA,B∈ObC(EAB, πAB, RAB) τ -pull-back τ •(SA,B∈ObC EAB) of the rank-one Fell bundle UA,B∈ObC(EAB, πAB, RAB). is a rank-one Fell bundle. Finally we see that the spectral spaceoid Σ(C) coincides with the The classical Gel'fand-Naımark duality for commutative C-algebras had a number of im- portant applications and, in a parallel way, its extension for commutative full C-categories described here, will provide interesting "horizontal categorifications" of those applications. Among them there are, for example: • a Fourier transform in the context of Pontryagin duality for commutative groupoids, • a continuous functional calculus for bounded linear operators between Hilbert spaces, • a spectral theorem for bounded linear operators between Hilbert spaces. Most of these ideas will actually require a serious amount of additional work that deserves a separate detailed treatment elsewhere and so, in order to exemplify the "capabilities" of our result, we limit ourselves to the development of a "horizontally categorified continuous functional calculus" which is the most immediate and straightforward of the previous topics. Let C be a C-category, not necessarily commutative or full, and let x ∈ CAB be one of its morphisms. Consider now C(x), the (non-necessarily unital) C-category generated by x. By definition, this is the C-subalgebra of CAA generated by x if A = B, and the C-subcategory of C with two objects A and B and arrow spaces C(x)AA = span{(x ◦ x∗)n n = 1, 2, 3, . . .}−, C(x)BB = span{(x∗ ◦ x)n n = 1, 2, 3, . . .}− and C(x)AB = x ◦ C(x)BB = C(x)∗ BA otherwise. Notice that C(x)(= C(x∗)) is always full. If A 6= B the C-category C(x) is always commutative and for A = B it is commutative if and only if x is normal and in these two cases we can immediately apply our spectral results on the horizontal categorified Gel'fand transform to realize that: every morphism in the category C(x) is uniquely described by a continuous section of a "block" of the spectral spaceoid of C(x). In more detail we have: Theorem 7.1 (Horizontally categorified continuous functional calculus). Let x ∈ CAB be an element of a C-category C and let C(x) denote the (non-necessarily unital) C-category generated by x inside C. If either the objects A and B are different, or A = B and the element x ∈ CAA is normal, then the C-category C(x) is full and commutative. In that case, for every continuous section σ ∈ Γ(Σ(C(x)))AB of the block AB of the spectral spaceoid of C(x), there is an associated element σ(x) ∈ C(x)AB. Moreover, the resulting map Fx : σ 7→ σ(x) is an isometric ∗-functor from the (possibly non-unital) C-category Γ(Σ(C(x))) onto C(x) ⊂ C. Proof. Given x ∈ CAB, we simply define Fx : Γ(Σ(C(x))) → C as the map given by the inverse of the Gel'fand transform G : C(x) → Γ(Σ(C(x))) i.e. Fx := G−1 C(x). We call the ∗-functor Fx the continuous functional calculus of x. Note that the previous result might open the way to obtaining a spectral representation (and hence a spectral theorem) also for bounded linear operators that are not normal. In fact if T : H → H is an arbitrary bounded linear operator on a Hilbert space H, we can always regard T as a morphism in an off-diagonal block of the C-category, with two objects, of bounded linear operators between HA := H =: HB. 18 8 Outlook We have introduced commutative C-categories and started a program for their "topological description" in terms of their spectra, here called spaceoids. In particular, we have obtained a Gel'fand-type theorem for full commutative C-categories. Although the statement of the main result (theorem 6.4) looks extremely natural, our proofs mostly rely on a "brute force" exploitation of the underlying structure and more streamlined arguments are likely to be found. Also, the result by itself is not as general as possible and certainly it leaves room for extensions in several directions, still hopefully we have provided some insight about how to achieve them. For instance, we have only considered the case of ∗-functors between (full, commutative) C-categories that are bijective on the objects. (Of course, this trivially includes morphisms between commutative C-algebras). In the next step, one would like to treat the case of ∗-functors that are not bijective on the objects. We tend to believe that this should not require significant modifications of our treatment and possibly could be dealt with using relators (that we introduced in [BCL1]).16 Perhaps a more important point would be to remove the condition of fullness. At present we have not discussed the issue in detail, but certainly the information that we have already acquired should significantly simplify the task. Also, along the way, we have somehow taken advantage of our prior knowledge of the Gel'fand and Serre-Swan theorems. Eventually one would like to provide more intrinsic proofs directly in the framework of C-categories (possibly unifying and extending both Gel'fand and Serre- Swan theorems in a "strict ∗-monoidal" version of Takahashi theorem [T1, T2]). In this respect, it looks promising to work directly with module categories. Besides, it is somehow disappointing that to date, for X and Y compact Hausdorff spaces, there seems to be no available general classification result for C(X)-C(Y )-bimodules. Hilbert C-bimodules that are not-necessarily imprimitivity bimodules should definitely play a role when discussing a classification result for generally non-commutative C-categories, possibly along the lines of a generalization to C-categories of the Dauns-Hofmann the- orem for C-algebras [DH]. One might also explore possible connections with the non- commutative Gel'fand spectral theorem of R. Cirelli-A. Mani`a-L. Pizzocchero [CMP] and the subsequent non-commutative Serre-Swan duality by K. Kawamura [Ka] and E. Elliott- K. Kawamura [EK]. Similarly, it might be very interesting to investigate the connections be- tween our spectral spaceoids and other spectral notions such as locales and topoi already used in the spectral theorems by B. Banachewki-C. Mulvey [BM] and C. Heunen-K. Landsmann- B. Spitters [HLS]. In the long run, one would like to (define and) classify commutative Fell bundles over suitable involutive categories. The notion of a Fell bundle could be even generalized to that of a fibered category enriched over another (∗-monoidal) category. Needless to say, one should analyze more closely the mathematical structure of spaceoids, introduce suitable topological invariants, study their symmetries, . . . , and investigate re- lations to other concepts that are widely used in other branches of mathematics, e.g. in algebraic topology/geometry as well as in gauge theories. Some geometric structures could become apparent when considering the representation of spaceoids as continuous fields of (one-dimensional commutative) C-categories as discussed by E. Vasselli in [V]. 16For this purpose it should be enough to introduce a category of spaceoids in which the morphisms f : X1 → X2 between the two base ∗-categories X1 = ∆1 × R1 and X2 = ∆2 × R2 are given by ∗-relators f := (f∆, fR) where now the ∗-functor fR : R2 → R1 is acting in the "reverse direction". 19 The Gel'fand transform for general commutative C-categories raises several questions (un- doubtedly it could be defined for more general Banach categories, leading to a wide range of possibilities for further studies). In particular, an immediate application would yield a Fourier transform and accordingly a reasonable concrete duality theory for commutative discrete groupoids (see M. Amini [A] for another approach that applies to compact but-not-necessarily-commutative-groupoids, T. Timmermann [Ti] for a more abstract setup and G. Goehle [G] for a discussion of duality for locally compact Abelian group bundles). As far as we are concerned, our main motivation to work with C-categories comes from analyzing the categorical structure of non-commutative geometry (where morphisms of "non- commutative spaces" are given by bimodules) and one is naturally led to speculating about the possible evolution of the notion of spectra and morphisms in A. Connes' non-commutative geometry (cf. [BCL1, BCL2, CCM]). In this direction, some of the first questions that come to mind are: Is there a suitable notion of spectral triple over a C-category? Is it possible to consider a horizontal categorification of a spectral triple? Of course this represents only the starting point for a much more ambitious program aiming at a "vertical categorification" of the notion of spectral triple17 and from several fronts (see for example [DTT] and also the very detailed discussion by J. Baez [B] on the weblog "The n-Category Caf´e") it is mounting the evidence that a suitable notion of non-commutative calculus necessarily require a higher (actually ∞) categorical setting. In this respect, it seems reasonable to look for a Gel'fand theorem that applies to (strict) com- mutative higher categories (cf. [Ko]). A suitable definition of strict n-C-categories (cf. [Z] for the case n = 2) and the proof of a categorical Gel'fand duality (at least for "commutative" full strict n-C-categories) are topics that have recently attracted our attention [BCLS]. Finally, in this line of thoughts, one could envisage potential applications of a notion of Gromov-Hausdorff distance (cf. [R]) for C-categories. Acknowledgments. We acknowledge the support provided by the Thai Research Fund, grant n. RSA4780022. The main part of this work has been done in the two-year visiting period of R. Conti to the Department of Mathematics of Chulalongkorn University. P. Bertozzini acknowledges the "weekly hospitality" offered by the Department of Mathematics in Chulalongkorn University, where most of the work leading to this publication has been done, from June 2005. He also desires to thank A. Carey and B. Wang at ANU in Canberra, R. Street, A. Davy- dov and M. Batanin at Macquarie University in Sydney, K. Hibbert and J. Links at the Center for Mathematical Physics in Brisbane, A. Vaz Ferreira at the University of Bologna, F. Cipriani at the "Politecnico di Milano", G. Landi and L. Dabrowski at SISSA in Trieste, R. Longo, C. D'Antoni and L. Zsido at the University of "Tor Vergata" in Rome, W. Szymanski at the University of Newcastle in Australia, E. Beggs and T. Brzezinski at Swansea University, D. Evans at Cardiff University, A. Doring and B. Coecke at the CLAP workshop in Imperial College, for the kind hospitality, the much appreciated partial support and most of all for the possibility to offer the talks/seminars where preliminary versions of the results contained in this paper have been announced in October 2006, March/October 2007 and May 2008. Finally we thank the two anonymous referees for a very careful reading of the manuscript and for suggesting several improvements. 17The need for a notion of "higher spectral triple" has been already advocated by U. Schreiber [S]. 20 References [A] [Ar] [B] [BD] [BM] Amini M. (2007). Tannaka-Krein Duality for Compact Groupoids: Adv. Math. 214, n. 1, 78-91, arXiv:math.OA/0308259v1; II Fourier Transform, arXiv:math/0308260v1; III Duality Theory, arXiv:math/0308261v1. I Representation Theory, Arveson, W. (1976). An Invitation to C-algebras, GTM 39, Springer-Verlag. Baez J., Infinitely Categorified Calculus, in: "The n-Category Caf´e", weblog, http://golem.ph.utexas.edu/category/2007/02/infinitely categorified calcul.html, 09 February 2007. Baez J., Dolan J. (1998). Categorification, in: Higher Category Theory, eds. Getzler E., Kapranov M., Contemp. Math. 230, 1-36, arXiv:math/9802029v1. Banaschewski B., Mulvey C. (2006). A Globalisation of the Gelfand Duality Theorem, Ann. Pure Appl. Logic 137 (1-3), 62-103. [BCL1] Bertozzini P., Conti R., Lewkeeratiyutkul W. (2006). A Category of Spectral Triples and Discrete Groups with Length Function, Osaka J. Math. 43, n. 2, 327-350, arXiv:math/0502583v1. [BCL2] Bertozzini P., Conti R., Lewkeeratiyutkul W. (2007). Non-commutative Geometry, Categories and Quantum Physics, Proceedings of the International Conference on Mathematics and Its Appli- cations, ICMA-MU 2007, Bangkok, Thailand, August 15-17, 2007, Mahidol University, 135-178, appeared in: Contributions in Mathematics and Applications II, East-West J. Math. 2007, Spe- cial Volume, 213-259, arXiv:0801.2826v1. [BCL3] Bertozzini P., Conti R., Lewkeeratiyutkul W., Spectral Theorem for Imprimitivity C-bimodules, arXiv:0812.3596v1. [BCLS] Bertozzini P., Conti R., Lewkeeratiyutkul W., Suthichitranont N., Strict Higher C-categories, in preparation. [Bl] Blackadar B. (2006). Operator Algebras, Springer. [BGR] [CMP] Brown L., Green P., Rieffel M. (1977). Stable Isomorphism and Strong Morita Equivalence of C-algebras, Pacific J. Math. 71, n. 2, 349-363. Cirelli R., Mani`a A., Pizzocchero L. (1994). A Functional Representation for Non-commutative C-algebras, Rev. Math. Phys. 6, n. 5, 675-697. [C] Connes A. (1994). Noncommutative Geometry, Academic Press. [CCM] Connes A., Consani C., Marcolli M. (2007). Noncommutative Geometry and Motives: the Ther- modynamics of Endomotives, Adv. Math. 214, n. 2, 761-831, arXiv:math/0512138v2. [CY] [D] [DH] [DKR] [DTT] Crane L., Yetter D. (1998). Examples of Categorification, Cahiers de Topologie et Geom´etrie Diff´erentielle Cat´egoriques 39, n. 1, 3-25. Daenzer C. (2009). A Groupoid Approach to Noncommutative T -duality, Comm. Math. Phys. 288, n. 1, 55-96, arXiv:0704.2592v2. Dauns J., Hofmann K.-H. (1968). Representations of Rings by Sections, Mem. Amer. Math. Soc., AMS. Deaconu V., Kumjian A., Ramazan B. (2008). Fell Bundles Associated to Groupoid Morphisms, Math. Scand. 102, n. 2, 305-319, arXiv:math/0612746v2. Dolgushev V., Tamarkin D., Tsygan B. (2007). The Homotopy Gerstenhaber Algebra of Hochschild Cochains of a Regular Algebra is Formal, J. Noncommut. Geom. 1, n. 1, 1-25, arXiv:math/0605141v1. [EK] Elliott E., Kawamura K. (2008). A Hilbert Bundle Characterization of Hilbert C-modules, Trans. Amer. Math. Soc. 360, n. 9, 4841-4862. [E] Exel R., Noncommutative Cartan Sub-algebras of C-algebras, arXiv:0806.4143. 21 [FD] [FGV] Fell J., Doran R. (1998). Representations of C-algebras, Locally Compact Groups and Banach ∗-algebraic Bundles, Vol. 1, 2, Academic Press. Figueroa H., Gracia-Bondia J.-M., Varilly J.-C. (2000). Elements of Noncommutative Geometry, Birkhauser. [GLR] Ghez P., Lima R., Roberts J. (1985). W-categories, Pacific J. Math. 120, 79-109. [G] Goehle G. (2008). Group Bundle Duality, Illinois J. Math. 52, n. 3, 951-956, arXiv:0711.4322v1. [HLS] Heunen C., Landsman N., Spitters B. (2009). A Topos for Algebraic Quantum Theory, Comm. Math. Phys. 291, n. 1, 63-110, arXiv:0709.4364v2. [K] [Ka] [Ko] [Ku] [L] [M1] [M2] [R] [S] [T1] [T2] [Ti] [V] [W] [Z] Karoubi M. (1978). Introduction to K-theory, Springer. Kawamura K. (2003). Serre-Swan Theorem for Non-commutative C-algebras, J. Geom. Phys. 48, n. 2-3, 275-296, arXiv:math/0002160v2. Kondratiev G., Concrete Duality for Strict Infinity Categories, arXiv:0807.4256. Kumjian A. (1998). Fell Bundles over Groupoids, Proc. Amer. Math. Soc. 126, n. 4, 1115-1125, arXiv:math/9607230. Landsman N., Lecture Notes on C-algebras and K-theory, Part 1, Master Course 2003-2004, http://www.science.uva.nl/~npl/CK.pdf. Mitchener P. (2001). Symmetric K-theory Spectra of C-categories, K-theory 24, 157-201. Mitchener P. (2002). C-categories, Proceedings of the London Mathematical Society 84, 375-404. Rieffel M. (2010). Vector Bundles and Gromov-Hausdorff Distance, J. K-Theory 5, n. 1, 39-103, arXiv:math/0608266v3. Schreiber U., Connes' Spectral Geometry and the Standard Model II, in: "The n-Category Caf´e", weblog, http://golem.ph.utexas.edu/category/, 06 September 2006. Takahashi A. (1979). Hilbert Modules and their Representation, Rev. Colombiana Mat. 13, 1-38. Takahashi A. (1979). A Duality between Hilbert Modules and Fields of Hilbert Spaces, Rev. Colombiana Mat. 13, 93-120. Timmermann T. (2005). Pseudo-multiplicative Unitaries and Pseudo-Kac Systems on C-modules, Ph.D. dissertation, Universitat Munster. Vasselli E. (2007). Bundles of C-categories and Duality, J. Funct Anal. 247, 351-377, arXiv:math/0510594v3. Weaver N. (2001). Mathematical Quantization, Chapmann and Hall. Zito P. (2007). 2-C*-categories with Non-simple Units, Adv. Math. 210, n. 1, 122-164, arXiv:math/0509266v1. 22
1407.5715	2	1407	2015-02-22T14:36:01	Absence of algebraic relations and of zero divisors under the assumption of finite non-microstates free Fisher information	[ "math.OA" ]	We show that in a tracial and finitely generated $W^\ast$-probability space existence of conjugate variables in an appropriate sense exclude algebraic relations for the generators. Moreover, under the assumption of finite non-microstates free Fisher information, we prove that there are no zero divisors in the sense that the product of any non-commutative polynomial in the generators with any element from the von Neumann algebra is zero if and only if at least one of those factors is zero.	math.OA	math	ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS UNDER THE ASSUMPTION OF FINITE NON-MICROSTATES FREE FISHER INFORMATION TOBIAS MAI, ROLAND SPEICHER, AND MORITZ WEBER Abstract. We show that in a tracial and finitely generated W ∗- probability space existence of conjugate variables in an appropriate sense exclude algebraic relations for the generators. Moreover, under the assumption of finite non-microstates free Fisher information, we prove that there are no zero divisors in the sense that the product of any non-commutative polynomial in the generators with any element from the von Neumann algebra is zero if and only if at least one of those factors is zero. 1. Introduction In a groundbreaking series of papers [Voi93, Voi94, Voi96, Voi97, Voi98, Voi99] (see also the survey [Voi02]), Voiculescu transferred the notion of entropy and Fisher information to the world of noncommu- tative probability theory. One of the most striking results which came out of this program is certainly the proof of the fact that the free group factors do not possess Cartan subalgebras in [Voi96], which gives in par- ticular the solution of the by then longstanding open question, whether every separable II1-factor contains Cartan subalgebras. We should note that Voiculescu gave in fact two different approaches to entropy and Fisher information in the non-commutative setting. The first one is based on the notion of matricial microstates and defines free entropy χ first and deduces free Fisher information Φ from χ, the second is based on the notion of conjugate variables with respect to certain non-commutative derivatives and defines free Fisher informa- tion Φ∗ first and then deduces free entropy χ∗ from this quantity. We want to note that both constructions lead independently to objects which are, compared to the classical theory, justifiably called entropy and Fisher information, but it is still not known whether they coincide, respectively. In the classical case, as well as in the one-variable free case, finite- ness of entropy or of Fisher information imply some regularity of the corresponding distribution of the variables; in particular, they have a density (with respect to Lebesgue measure). In the non-commutative Date: October 15, 2018. 1 2 T. MAI, R. SPEICHER, AND M. WEBER situation, the notion of a density does not make any direct sense, but still it is believed that the existence of finite free entropy or finite free Fisher information (in any of the two approaches) should correspond to some regularity property of the considered non-commutative distri- butions. Thus one expects many "smooth" properties for random vari- ables X1, . . . , Xn for which either one of the quantities χ(X1, . . . , Xn), χ∗(X1, . . . , Xn), Φ(X1, . . . , Xn), or Φ∗(X1, . . . , Xn) is finite. In particu- lar, it is commonly expected that such a finiteness excludes non-trivial algebraic relations between the considered random variables. Up to now there is no proof of such a general statement. We will show here such a result in the case of finite non-microstates free Fisher informa- tion Φ∗(X1 . . . , Xn). Actually, we will not only prove the absence of algebraic relations between the X1, . . . , Xn, but also show that such cannot hold locally on non-trivial Hilbert subspaces; more formally we show that there are no non-zero divisors in the affiliated von Neumann algebra. In the second named approach, the so-called non-microstates ap- proach, Voiculescu's definition of conjugate variables in [Voi98] depends fundamentally on the existence of non-commutative derivatives. But this makes it necessary to assume right from the beginning that there are no algebraic relations between the generators of the underlying von Neumann algebra. Although this does not cause any serious trouble, this assumption might seem a bit unnatural and it could be -- depending on the particular situation -- quite hard to check. The first part of our article is devoted to a slightly different approach to conjugate variables. We will give a definition, which will circumvent the initial assumption of absence of algebraic relations. In fact, we will show that absence of algebraic relations is actually a consequence of the existence of conjugate variables. Thus, a posteriori, it will turn out that our notion of conjugate variables actually agrees with Voiculescu's definition. Nevertheless, the assumption of algebraic freeness becomes redundant and the defining conditions are much easier to handle than those for the algebraic freeness of the generators. Inspired by the observation that the existence of conjugate variables excludes algebraic relations, the second part of our article is concerned with similar but more advanced consequences of the finiteness of non- microstates Fisher information. More precisely, we will show that it excludes zero divisors in the sense that the product of any non- commutative polynomial in the generators with any element from the von Neumann algebra is zero if and only if at least one of those factors is zero. In particular, this result allows to conclude that the distribution of any (of course, non-trivial) self-adjoint non-commutative polynomial in the generators does not have atoms, if the generators have finite free ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS 3 Fisher information Φ∗. This extends the previous work of Shlyakhtenko and Skoufranis [SS13]. 2. Existence of conjugate variables and absence of algebraic relations Let ChZ1, . . . , Zni be the ∗-algebra of non-commutative polynomials in n self-adjoint (formal) variables Z1, . . . , Zn. For j = 1, . . . , n, we denote by ∂j the non-commutative derivative with respect to Zj, i.e. ∂j is the unique derivation ∂j : ChZ1, . . . , Zni → ChZ1, . . . , Zni ⊗ ChZ1, . . . , Zni that satisfies ∂jZi = δi,j1 ⊗ 1 for i = 1, . . . , n. Recall that being a derivation means for ∂j that (2.1) ∂j(P1P2) = (∂jP1)(1 ⊗ P2) + (P1 ⊗ 1)(∂jP2) holds for all P1, P2 ∈ ChZ1, . . . , Zni. More explicitly, ∂j acts on mono- mials P as ∂jP = XP =P1ZjP2 P1 ⊗ P2, where the sum runs over all decompositions P = P1ZjP2 of P with monomials P1, P2. Throughout the following, let (M, τ ) be a tracial W ∗-probability space, which means that M is a von Neumann algebra and τ is a faithful normal tracial state on M. For selfadjoint X1, . . . , Xn ∈ M we denote by vN(X1, . . . , Xn) ⊂ M the von Neumann subalgebra of M which is generated by X1, . . . , Xn and by L2(X1, . . . , Xn, τ ) ⊂ L2(M, τ ) the L2-space which is generated by X1, . . . , Xn with respect to the inner product given by hP, Qi := τ (P Q∗). In the following, we will denote by evX, for a given n-tuple X = (X1, . . . , Xn) of self-adjoint elements of M, the homomorphism evX : ChZ1, . . . , Zni → ChX1, . . . , Xni ⊂ M given by evaluation at X = (X1, . . . , Xn), i.e. the homomorphism evX is determined by Zi 7→ Xi. For reasons of clarity, we put P (X) := evX(P ) for any P ∈ ChZ1, . . . , Zni and Q(X) := (evX ⊗ evX )(Q) for any Q ∈ ChZ1, . . . , Zni⊗2. Definition 2.1. Let X1, . . . , Xn ∈ M be self-adjoint elements. If there are elements ξ1, . . . , ξn ∈ L2(M, τ ), such that (2.2) (τ ⊗ τ )((∂jP )(X1, . . . , Xn)) = τ (ξjP (X1, . . . , Xn)) is satisfied for each non-commutative polynomial P ∈ ChZ1, . . . , Zni and for j = 1, . . . , n, then we say that (ξ1, . . . , ξn) satisfies the conjugate relations for (X1, . . . , Xn). If, in addition, ξ1, . . . , ξn belong to L2(X1, . . . , Xn, τ ), we say that (ξ1, . . . , ξn) is the conjugate system for (X1, . . . , Xn). 4 T. MAI, R. SPEICHER, AND M. WEBER Like in the usual setting, we have the following. Remark 2.2. Let π be the orthogonal projection from L2(M, τ ) to L2(X1, . . . , Xn, τ ). It is easy to see that if (ξ1, . . . , ξn) satisfies the con- jugate relations for (X1, . . . , Xn), then (π(ξ1), . . . , π(ξn)) satisfies the conjugate relations for (X1, . . . , Xn) as well, and is therefore a conju- gate system for (X1, . . . , Xn). It is an easy consequence of its defining property (2.2) that a conju- gate system (ξ1, . . . , ξn) for (X1, . . . , Xn) is unique if it exists. Note that our notion of conjugate relations and conjugate variables differs from the usual definition which was given by Voiculescu in [Voi98], roughly speaking, just by the placement of brackets. To be more precise, in (2.2), we first apply the derivative ∂j to the given non- commutative polynomial P before we apply the evaluation at X = (X1, . . . , Xn), instead of applying the evaluation first, which conse- quently makes it necessary to have in the second step a well-defined derivation on ChX1, . . . , Xni corresponding to ∂j. From a more abstract point of view, this idea is in the same spirit as [Shl00, Lemma 3.2] but only on a purely algebraic level. In fact, we used the surjective homomorphism evX : ChZ1, . . . , Zni → ChX1, . . . , Xni in order to pass from (ChX1, . . . , Xni, τ ) to the non-commutative proba- bility space (ChZ1, . . . , Zni, τX), where τX := τ ◦ evX . Due to this lift- ing, the algebraic relations between the generators disappear whereas the relevant information about their joint distribution remains un- changed. Our aim is to show that the existence of a conjugate system guaran- tees that X1, . . . , Xn do not satisfy any algebraic relations. Obviously, we can rephrase this in more algebraic terms by saying that the two sided ideal I 1 X := {P ∈ ChZ1, . . . , Zni P (X1, . . . , Xn) = 0} of ChZ1, . . . , Zni is the zero ideal. But this exactly means that the evaluation homomorphism evX is in fact an isomorphism. Hence, if this is shown, we can immediately define a non-commutative derivation ∂j : ChX1, . . . , Xni → ChX1, . . . , Xni ⊗ ChX1, . . . , Xni, where the terminology derivation has to be understood with respect to the ChX1, . . . , Xni-bimodule structure of ChX1, . . . , Xni⊗2 like in (2.1). X = {0}, it is helpful to consider this question at once together with the question of the existence of well-defined derivations ∂j on ChX1, . . . , Xni. Surprisingly, it turns out that, in order to prove I 1 Proposition 2.3. As before, let (M, τ ) be a tracial W ∗-probability space and let self-adjoint X1, . . . , Xn ∈ M be given. Assume that there are elements ξ1, . . . , ξn ∈ L2(M, τ ), such that (ξ1, . . . , ξn) satisfies the conjugate relations (2.2) for (X1, . . . , Xn). ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS 5 Corresponding to X = (X1, . . . , Xn), we introduce the following two- sided ideals of ChZ1, . . . , Zni and of ChZ1, . . . , Zni⊗2, respectively: I 1 X := {P ∈ ChZ1, . . . , Zni P (X1, . . . , Xn) = 0} and I 2 X := {Q ∈ ChZ1, . . . , Zni⊗2 Q(X1, . . . , Xn) = 0}. Then, for each j = 1, . . . , n, P + I 1 X 7→ ∂jP + I 2 X induces a well-defined derivation ∂j : ChZ1, . . . , Zni/I 1 X → ChZ1, . . . , Zni⊗2/I 2 X. Before starting the proof, let us introduce a binary operation ♯ on the algebraic tensor product M ⊗ M by bilinear extension of (a1 ⊗ a2)♯(b1 ⊗ b2) := (a1b1) ⊗ (b2a2). Proof of Proposition 2.3. Obviously, it is sufficient to show that P ∈ I 1 X implies ∂jP ∈ I 2 X be given. If we take arbitrary P1, P2 ∈ ChZ1, . . . , Zni, we have that X for all j = 1, . . . , n. For seeing this, let P ∈ I 1 ∂j(P1P P2) = (∂jP1)P P2 + P1P (∂jP2) + P1(∂jP )P2 and therefore, since P (X) = 0, (τ ⊗ τ )(∂j(P1P P2)(X)) = (τ ⊗ τ )(P1(X)∂jP (X)P2(X)). Furthermore, according to (2.2), we may deduce that (τ ⊗ τ )(∂j(P1P P2)(X)) = τ (ξj(P1P P2)(X)) = 0. Thus, we observe that (τ ⊗ τ )((P1 ⊗ P2)(X)♯∂jP (X)) = (τ ⊗ τ )(P1(X)∂jP (X)P2(X)) = 0 for all P1, P2 ∈ ChZ1, . . . , Zni and hence by linearity (τ ⊗ τ )(Q(X)♯∂jP (X)) = 0 for all Q ∈ ChZ1, . . . , Zni⊗2. If we apply this observation to Q = (∂jP )∗, we easily see that ∂jP (X) = 0 (recall that τ was assumed to X. This shows that ∂j is indeed be faithful), which means that ∂jP ∈ I 2 well-defined. Due to the obvious fact that I 1 X is a two-sided ideal in ChZ1, . . . , Zni X is a two-sided ∗-ideal in ChZ1, . . . , Zni⊗2, we see that as well as I 2 ChZ1, . . . , Zni/I 1 X, where the multiplication and the involution are just defined via representa- tives. Using this, it is easy to check that ∂1, . . . , ∂n are indeed deriva- tions. (cid:3) X is a ∗-algebra as well as ChZ1, . . . , Zni⊗2/I 2 6 T. MAI, R. SPEICHER, AND M. WEBER Basic linear algebra shows that ChZ1, . . . , Zni/I 1 X ∼= ChX1, . . . , Xni and ChZ1, . . . , Zni⊗2/I 2 X ∼= ChX1, . . . , Xni⊗2. Hence, Proposition 2.3 immediately implies the following corollary. Corollary 2.4. As before, let (M, τ ) be a tracial W ∗-probability space and let self-adjoint X1, . . . , Xn ∈ M be given. Assume that there are elements ξ1, . . . , ξn ∈ L2(M, τ ), such that (ξ1, . . . , ξn) satisfies the con- jugate relations (2.2) for (X1, . . . , Xn). Then, for each j = 1, . . . , n, there is a unique derivation ∂j : ChX1, . . . , Xni → ChX1, . . . , Xni ⊗ ChX1, . . . , Xni such that the following diagram commutes. (2.3) ChZ1, . . . , Zni evX ChX1, . . . , Xni ∂j ∂j ∂j ChZ1, . . . , Zni⊗2 evX ⊗ evX / ChX1, . . . , Xni⊗2 The fact that the diagram in (2.3) commutes immediately implies that the derivations ∂j, j = 1, . . . , n, satisfy ∂j(Xi) = δj,i1 ⊗ 1 for all j, i = 1, . . . , n. Indeed, ∂j(Xi) = ∂j(evX (Zi)) = (evX ⊗ evX)(∂jZi) = (evX ⊗ evX)(δj,i1 ⊗ 1) = δj,i1 ⊗ 1. For reasons of completeness, we want to mention that the converse is also true, so that both statements are in fact equivalent: If we assume that the derivations ∂j, j = 1, . . . , n, satisfy ∂j(Xi) = δj,i1 ⊗ 1 for all j, i = 1, . . . , n, then dj := ∂j ◦ evX : ChZ1, . . . , Zni → ChX1, . . . , Xni⊗2. obviously defines a derivation, where we consider ChX1, . . . , Xni⊗2 as ChZ1, . . . , Zni-bimodule via evaluation evX . Thus, we have for each P ∈ ChZ1, . . . , Zni and j = 1, . . . , n that dj(P ) = n Xi=1 (evX ⊗ evX)(∂iP )♯dj(Zi). / / / / / ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS 7 Since dj(Zi) = ∂j(Xi) = δj,i1 ⊗ 1 holds for i = 1, . . . , n by assumption, we see that dj(P ) = (evX ⊗ evX)(∂jP ). By definition of dj, this gives ( ∂j ◦ evX)(P ) = dj(P ) = ((evX ⊗ evX) ◦ ∂j)(P ), which precisely means that the diagram in (2.3) commutes. However, this allows us to rephrase Corollary 2.4 as follows. Corollary 2.5. As before, let (M, τ ) be a tracial W ∗-probability space and let self-adjoint X1, . . . , Xn ∈ M be given. Assume that there are elements ξ1, . . . , ξn ∈ L2(M, τ ), such that (2.2) is satisfied for j = 1, . . . , n. Then, for each j = 1, . . . , n, there is a unique derivation ∂j : ChX1, . . . , Xni → ChX1, . . . , Xni ⊗ ChX1, . . . , Xni, such that ∂j(Xi) = δj,i1 ⊗ 1 holds for all i = 1, . . . , n. This final conclusion shows that even without assuming algebraic freeness of the generators X1, . . . , Xn, it is possible to define non- commutative derivations that behave exactly like if the generators would be algebraically free. In fact, this is the key observation to reach our desired result. Proposition 2.6. Let M be a tracial W ∗-probability space and let self- adjoint elements X1, . . . , Xn ∈ M be given. Assume that there are derivations ∂j : ChX1, . . . , Xni → ChX1, . . . , Xni ⊗ ChX1, . . . , Xni for j = 1, . . . , n, such that ∂j(Xi) = δj,i1 ⊗ 1 holds for all i = 1, . . . , n. Then there is no non-zero polynomial P ∈ ChZ1, . . . , Zni such that P (X1, . . . , Xn) = 0 holds. Proof. First of all, we observe that for any P ∈ ChZ1, . . . , Zni the following implication holds true: P (X1, . . . , Xn) = 0 =⇒ ∀j = 1, . . . , n : (∂jP )(X1, . . . , Xn) = 0 This follows immediately from the assumption that the diagram (2.3) commutes. Thus, we may define ∆j := ((τ ◦ evX) ⊗ id) ◦ ∂j, which is a linear mapping ∆j : ChZ1, . . . , Zni → ChZ1, . . . , Zni, such that for any P ∈ ChZ1, . . . , Zni the implication P (X1, . . . , Xn) = 0 =⇒ ∀j = 1, . . . , n : (∆jP )(X1, . . . , Xn) = 0 holds true. Now, let us assume that there is a non-zero polynomial P ∈ ChZ1, . . . , Zni such that P (X1, . . . , Xn) = 0 holds, for instance P (Z1, . . . , Zn) = a0 + d Xk=1 n Xi1,...,ik=1 ai1,...,ikZi1 . . . Zik, 8 T. MAI, R. SPEICHER, AND M. WEBER where d ≥ 1 denotes the total degree of P . We choose any summand of highest degree ai1,...,idZi1 . . . Zid of P . It is easy to see that ∆id . . . ∆i1P = ai1,...,id. Hence, we deduce ai1,...,in = (∆id . . . ∆i1P )(X1, . . . , Xn) = 0, which finally leads to a contradiction. Therefore, we must have P = 0, which concludes the proof. (cid:3) Combining the above Proposition 2.6 with Corollary 2.5 leads us directly to the following theorem. Theorem 2.7. As before, let (M, τ ) be a tracial W ∗-probability space. Let X1, . . . , Xn ∈ M be self-adjoint and assume that there are elements ξ1, . . . , ξn ∈ L2(M, τ ), such that (ξ1, . . . , ξn) satisfies the conjugate rela- tions for (X1, . . . , Xn), i.e. (2.2) holds for j = 1, . . . , n. Then we have the following statements: (a) X1, . . . , Xn do not satisfy any non-trivial algebraic relation, i.e. there exists no non-zero polynomial P ∈ ChZ1, . . . , Zni such that P (X1, . . . , Xn) = 0. (b) For j = 1, . . . , n, there is a unique derivation ∂j : ChX1, . . . , Xni → ChX1, . . . , Xni ⊗ ChX1, . . . , Xni which satisfies ∂j(Xi) = δj,i1 ⊗ 1 for i = 1, . . . , n. Since Theorem 2.7 clarifies which consequences the existence of the conjugate system has, we may proceed now by defining (non-microstates) free Fisher information. Definition 2.8. Let (M, τ ) be a tracial W ∗-probability space and let self-adjoint elements X1, . . . , Xn ∈ M be given. We define their (non- microstates) free Fisher information Φ∗(X1, . . . , Xn) by Φ∗(X, . . . , Xn) := n Xj=1 kξjk2 2 if a conjugate system (ξ1, . . . , ξn) for (X1, . . . , Xn) in the sense of Defi- nition 2.1 exists, and we put Φ∗(X1, . . . , Xn) := ∞ if no such conjugate system for (X1, . . . , Xn) exists. The quantity Φ∗(X1, . . . , Xn) is obviously well-defined, since as soon as a conjugate system in the sense of Definition 2.1 exists, Theorem 2.7 implies that it is also a conjugate system in the usual sense. Thus, Φ∗(X1, . . . , Xn) is just the usual non-microstates free Fisher informa- tion which was defined in [Voi98]. Nevertheless, it has the advantage that it can be defined even without assuming the algebraic freeness of X1, . . . , Xn right from the beginning. Let (M, τ ) be a W ∗-probability space and let self-adjoint elements X1, . . . , Xn ∈ M be given such that the condition Φ∗(X1, . . . , Xn) < ∞ ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS 9 is fulfilled. Theorem 2.7 tells us that, for j = 1, . . . , n, there is a unique derivation ∂j : ChX1, . . . , Xni → ChX1, . . . , Xni ⊗ ChX1, . . . , Xni that satisfies ∂j(Xi) = δj,i1 ⊗ 1 for i = 1, . . . , n. But furthermore, it tells us that X1, . . . , Xn do not satisfy any algebraic relation, which in other words means that the evaluation homomorphism evX induces an isomorphism between the abstract polynomial algebra ChZ1, . . . , Zni and the subalgebra ChX1, . . . , Xni of M. We have seen in Corollary 2.4 that under this identification the derivations ∂j on ChZ1, . . . , Zni and ∂j on ChX1, . . . , Xni correspond to each other. We will therefore not distinguish anymore between ∂j and ∂j. We finish this section by noting that Φ∗(X1, . . . , Xn) < ∞ moreover excludes analytic relations. More precisely, this means that there is no non-zero non-commutative power series P , which is convergent on a polydisc DR := {(Y1, . . . , Yn) ∈ M n ∀j = 1, . . . , n : kYjk < R} for some R > 0, such that (X1, . . . , Xn) ∈ DR and P (X1, . . . , Xn) = 0. Based on Voiculescu's original definition of the non-microstates free Fisher information and hence under the additional assumption that X1, . . . , Xn are algebraically free, this was shown by Dabrowski in [Dab10a, Lemma 37]. 3. Finite free Fisher information and zero divisors Inspired by the methods used in the proof of our Theorem 2.7, we address now the more general question of existence of zero divisors under the assumption of finite non-microstates free Fisher information. First, we shall make more precise, what we mean by this. Our aim is to prove the following statement. Theorem 3.1. Let (M, τ ) be a tracial W ∗-probability space. Further- more, let X1, . . . , Xn ∈ M be self-adjoint elements and assume that Φ∗(X1, . . . , Xn) < ∞ holds. Then, for any non-zero non-commutative polynomial P , there exists no non-zero self-adjoint element w ∈ vN(X1, . . . , Xn) such that P (X1, . . . , Xn)w = 0. Recall that to each element X = X ∗ ∈ M, there corresponds a unique probability measure µX on the real line R, which has the same moments as X, i.e. it satisfies τ (X k) = ZR tk dµX(t) for k = 0, 1, 2, . . . . 10 T. MAI, R. SPEICHER, AND M. WEBER It is an immediate consequence of Theorem 3.1 that the distribution µP (X1,...,Xn) of P (X1, . . . , Xn) for any non-constant self-adjoint poly- nomial P cannot have atoms, if Φ∗(X1, . . . , Xn) < ∞. The precise statement reads as follows. Corollary 3.2. Let (M, τ ) be a tracial W ∗-probability space and let X1, . . . , Xn ∈ M be self-adjoint with Φ∗(X1, . . . , Xn) < ∞. Then, for any non-constant self-adjoint non-commutative polynomial P , the dis- tribution µP (X1,...,Xn) of P (X1, . . . , Xn) does not have atoms. 3.1. Collecting the ingredients for the proof of Theorem 3.1. Corresponding to the assumptions made in Theorem 3.1, let (M, τ ) be a tracial W ∗-probability space and let X1, . . . , Xn ∈ M be self-adjoint with Φ∗(X1, . . . , Xn) < ∞. As we have seen in Section 2, those conditions guarantee that, for j = 1, . . . , n, there exists a unique derivation ∂j : ChX1, . . . , Xni → ChX1, . . . , Xni ⊗ ChX1, . . . , Xni, which is determined by the condition ∂jXi = δi,j1 ⊗ 1 for i = 1, . . . , n. For each j = 1, . . . , n, we may consider ∂j as a densely defined un- bounded operator ∂j : L2(M, τ ) ⊇ D(∂j) → L2(M⊗M, τ ⊗ τ ) with domain D(∂j) := ChX1, . . . , Xni, where we denote by M⊗M the von Neumann algebra tensor product of M with itself. Since due to Φ∗(X1, . . . , Xn) < ∞ a conjugate system (ξ1, . . . , ξn) for (X1, . . . , Xn) exists, we see by (2.2) that 1 ⊗ 1 belongs to the domain of definition of the adjoints ∂∗ j (1 ⊗ 1) holds for j = 1, . . . , n. n and that ξj = ∂∗ 1 , . . . , ∂∗ The proof of Theorem 3.1 will be based on several well-knowns facts about those operators ∂j, which we collect here for reader's convenience. Lemma 3.3 (Corollary 4.2 and Proposition 4.3 in [Voi98]). Under the above conditions, we have for j = 1, . . . , n that ChX1, . . . , Xni ⊗ ChX1, . . . , Xni ⊆ D(∂∗ j ), j is densely defined as well and ∂j is closable. More explicitly, we i.e. ∂∗ have for each Y ∈ ChX1, . . . , Xni⊗2 the formula ∂∗ j (Y ) = mξj (Y ) − m1(id ⊗τ ⊗ id)(∂j ⊗ id + id ⊗∂j)(Y ). Here, for any η ∈ L2(M, τ ), we denote by mη the linear operator mη : M ⊗ M → L2(M, τ ) defined on the algebraic tensor product M ⊗ M, which is given by mη(a1 ⊗ a2) := a1ηa2. And thus of course, m1(a1 ⊗ a2) = a1a2. The lemma above allows us to conclude that in particular (3.1) ∂∗ j (P ⊗ 1) = P ξj − (id ⊗τ )(∂jP ), ∂∗ j (1 ⊗ P ) = ξjP − (τ ⊗ id)(∂jP ) ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS 11 holds for j = 1, . . . , n and any P ∈ ChX1, . . . , Xni. Lemma 3.4 (Lemma 12 in [Dab10b]). Under the above conditions, we have for each P ∈ ChX1, . . . , Xni that (3.2) and therefore (3.3) kP ξj − (id ⊗τ )(∂jP )k2 ≤ kξjk2kP k, kξjP − (τ ⊗ id)(∂jP )k2 ≤ kξjk2kP k k(id ⊗τ )(∂jP )k2 ≤ 2kξjk2kP k, k(τ ⊗ id)(∂jP )k2 ≤ 2kξjk2kP k. Note that Lemma 3.4 is actually a slight extension of Lemma 12 in [Dab10b], since we added in (3.2) and (3.3) each time the second named estimates. In fact, they can be easily deduced from the first named estimates by using the more general identity (3.4) (τ ⊗ id)(P1(∂iP2))∗ = (id ⊗τ )((∂iP ∗ 2 )P ∗ 1 ) for all P1, P2 ∈ ChX1, . . . , Xni, which can itself easily be checked on monomials. Moreover, we note that thanks to (3.1), the inequalities in (3.2) can be rewritten in the following way: (3.5) k∂∗ k∂∗ j (P ⊗ 1)k2 ≤ kξjk2kP k j (1 ⊗ P )k2 ≤ kξjk2kP k 3.2. Proof of Theorem 3.1. Inspired by the proof of Theorem 2.7, we try to find a certain reduction argument that allows us to lower the degree of P . This will need some preparation. Lemma 3.5. For any w = w∗ ∈ vN(X1, . . . , Xn), there exists a se- quence (wk)k∈N of elements in ChX1, . . . , Xni such that: (i) wk = w∗ k (ii) sup k∈N kwkk < ∞ (iii) kwk − wk2 → 0 for k → ∞ Proof. First of all, we note that in order to prove the statement of the lemma above, it suffices to prove existence of a sequence (wk)k∈N of elements in ChX1, . . . , Xni, which satisfy only conditions (ii) and (iii). Indeed, if we replace in this case wk by its real part ℜ(wk) = 1 k), conditions (ii) and (iii) are still valid, but we have achieved condition (i) in addition. 2(wk+w∗ For proving existence under these weaker conditions, we may apply Kaplansky's density theorem. This guarantees the existence of a se- quence (wk)k∈N of elements in ChX1, . . . , Xni, such that kwkk ≤ kwk for all k ∈ N, which particularly implies (ii), and which converges to 12 T. MAI, R. SPEICHER, AND M. WEBER w in the strong operator topology. It remains to note that, with re- spect to the weak operator topology, w∗ kw → w∗w, w∗wk → w∗w, and w∗ kwk → w∗w as k → ∞, such that according to the continuity of τ kwk − wk2 2 = τ ((wk − w)∗(wk − w)) = τ (w∗ kwk) − τ (w∗ kw) − τ (w∗wk) + τ (w∗w) tends to 0 as k → 0, which shows (iii) and thus concludes the proof. (cid:3) A key observation for the absence of zero divisors is contained in the following proposition. Proposition 3.6. Let P ∈ ChX1, . . . , Xni be given. For any u = u∗, v = v∗ ∈ vN(X1, . . . , Xn), the following implication holds true: P u = 0 and P ∗v = 0 =⇒ ∀i = 1, . . . , n : v ⊗ 1(∂iP )1 ⊗ u = 0. Proof. Let (uk)k∈N and (vk)k∈N be sequences as described in Lemma 3.5 for u and v, respectively. Choose arbitrary non-commutative polyno- mials Q1, Q2 ∈ ChX1, . . . , Xni. Then, for i = 1, . . . , n and any k ∈ N, we can check that (i) { z } i (vkQ1 ⊗ Q2)i hP uk, ∂∗ = h∂i(P uk), vkQ1 ⊗ Q2i = h(∂iP )1 ⊗ uk, vkQ1 ⊗ Q2i + hP ⊗ 1(∂iuk), vkQ1 ⊗ Q2i = h(∂iP )1 ⊗ (uk − u), vkQ1 ⊗ Q2i + h(∂iP )1 ⊗ u, vkQ1 ⊗ Q2i + hP ⊗ 1(∂iuk), vkQ1 ⊗ Q2i = h(∂iP )1 ⊗ (uk − u), vkQ1 ⊗ Q2i + h(∂iP )1 ⊗ u, (vk − v)Q1 ⊗ Q2i + h(∂iP )1 ⊗ u, vQ1 ⊗ Q2i + hP ⊗ 1(∂iuk), vkQ1 ⊗ Q2i = h(∂iP )1 ⊗ (uk − u), vkQ1 ⊗ Q2i + h(∂iP )1 ⊗ u, (vk − v)Q1 ⊗ Q2i + hv ⊗ 1(∂iP )1 ⊗ u, Q1 ⊗ Q2i + h(∂iuk)1 ⊗ Q∗ 2, P ∗vkQ1 ⊗ 1i = h(∂iP )1 ⊗ (uk − u), vkQ1 ⊗ Q2i + h(∂iP )1 ⊗ u, (vk − v)Q1 ⊗ Q2i } (iv) {z {z (ii) } } (iii) {z + hv ⊗ 1(∂iP )1 ⊗ u, Q1 ⊗ Q2i + h(id ⊗τ )((∂iuk)1 ⊗ Q∗ 2), P ∗vkQ1i . Now, we will separately discuss the terms appearing in the above cal- culation: ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS 13 (i) According to Voiculescu's formula, which we recalled in Lemma 3.3, we have for all Y1, Y2 ∈ ChX1, . . . , Xni that ∂∗ i (Y1 ⊗ Y2) = Y1ξiY2 − m1(id ⊗τ ⊗ id)(∂i ⊗ id + id ⊗∂i)Y1 ⊗ Y2 = Y1ξiY2 − (id ⊗τ )(∂iY1)Y2 − Y1(τ ⊗ id)(∂iY2) = ∂∗ = ∂∗ i (Y1 ⊗ 1)Y2 − Y1(τ ⊗ id)(∂iY2) i (Y1 ⊗ 1)Y2 − Y1(id ⊗τ )(∂iY ∗ 2 )∗ and thus, by applying the estimates (3.3) and (3.5), that i (Y1 ⊗ Y2)k2 k∂∗ ≤ k∂∗ ≤ 3kξik2kY1kkY2k. i (Y1 ⊗ 1)k2kY2k + kY1kk(id ⊗τ )(∂iY ∗ 2 )k2 Therefore, we may deduce that i (vkQ1 ⊗ Q2)i hP uk, ∂∗ ≤ kP ukk2k∂∗ = kP (uk − u)k2k∂∗ ≤ 3kξik2kP kkQ1kkQ2kkvkkkuk − uk2. i (vkQ1 ⊗ Q2)k2 i (vkQ1 ⊗ Q2)k2 (ii) We apply partial integration in order to obtain (id ⊗τ )((∂iY1)1 ⊗ Y2) = (id ⊗τ )(∂i(Y1Y2)) − (id ⊗τ )(Y1 ⊗ 1(∂iY2)) = (id ⊗τ )(∂i(Y1Y2)) − Y1(id ⊗τ )(∂iY2) for arbitrary Y1, Y2 ∈ ChX1, . . . , Xni. From this, we can easily deduce by using (3.3) that k(id ⊗τ )((∂iY1)1 ⊗ Y2)k2 ≤ k(id ⊗τ )(∂i(Y1Y2))k2 + kY1kk(id ⊗τ )(∂iY2)k2 ≤ 4kξik2kY1kkY2k. Hence, it follows that 2), P ∗vkQ1i h(id ⊗τ )((∂iuk)1 ⊗ Q∗ ≤ k(id ⊗τ )((∂iuk)1 ⊗ Q∗ = k(id ⊗τ )((∂iuk)1 ⊗ Q∗ ≤ 4kξik2kP kkQ1kkQ2kkukkkvk − vk2 2)k2kP ∗vkQ1k2 2)k2kP ∗(vk − v)Q1k2 (iii) Note that, according to (3.4), our calculations from (ii) imply k(τ ⊗ id)(Y1 ⊗ 1(∂iY2))k2 ≤ 4kξik2kY1kkY2k 14 T. MAI, R. SPEICHER, AND M. WEBER for all Y1, Y2 ∈ ChX1, . . . , Xni. This allows us to deduce from h(∂iP )1 ⊗ (uk − u), vkQ1 ⊗ Q2i = hQ∗ = h(τ ⊗ id)(Q∗ 1vk ⊗ 1(∂iP ), 1 ⊗ Q2(uk − u)i 1vk ⊗ 1(∂iP )), Q2(uk − u)i that h(∂iP )1 ⊗ (uk − u), vkQ1 ⊗ Q2i ≤ k(τ ⊗ id)(Q∗ ≤ 4kξik2kP kkQ1kkQ2kkvkkkuk − uk2. 1vk ⊗ 1(∂iP ))k2kQ2(uk − u)k2 (iv) It remains to observe that h(∂iP )1 ⊗ u, (vk − v)Q1 ⊗ Q2i ≤ k∂iP k2kQ1kkQ2kkukkvk − vk2. In the limit k → ∞, a combination of all estimates proved in (i) up to (iv) shows that hv ⊗ 1(∂iP )1 ⊗ u, Q1 ⊗ Q2i = 0. Since Q1, Q2 ∈ ChX1, . . . , Xni were arbitrarily chosen, it follows v ⊗ 1(∂iP )1 ⊗ u = 0 for all i = 1, . . . , n as claimed. (cid:3) Remark 3.7. Let P ∈ ChX1, . . . , Xni be given and assume that there are u = u∗, v = v∗ ∈ vN(X1, . . . , Xn) such that P u = P ∗v = 0 holds. Then, according to Proposition 3.6, we know that v ⊗ 1(∂jP )1 ⊗ u = 0 If we replace now P by P ∗, the statement of for any j = 1, . . . , n. Proposition 3.6 also gives u ⊗ 1(∂jP ∗)1 ⊗ v = 0 for j = 1, . . . , n. But we want to point out that this does not lead to any new information. Indeed, if we take adjoints in the initial statement we get v ⊗ 1(∂jP )1 ⊗ u = 0, 1 ⊗ u(∂jP )∗v ⊗ 1 = 0. Then, if we apply the flip σ : M⊗M → M⊗M, i.e. the ∗-homomorphism induced by σ(a1 ⊗ a2) := a2 ⊗ a1, it follows that u ⊗ 1σ((∂jP )∗)1 ⊗ v = 0. An easy calculation on monomials shows σ((∂jP )∗) = ∂jP ∗, such that the above result reduces exactly to the statement obtained by replacing P with P ∗. Before doing the final step, we first want to test in two examples how strong the result in Proposition 3.6 is. ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS 15 Example 3.8. For the self-adjoint polynomial P = X1X2X1, we calcu- late ∂2P = X1⊗X1, such that P w = 0 implies according to Proposition 3.6 that wX1 ⊗ X1w = 0 and therefore X1w = 0 holds. Applying now Proposition 3.6 once again with respect to ∂1, we end up with w ⊗ w = 0, such that w = 0 follows. Example 3.9. Take P = X1X2 + X2X1. We have ∂1P = 1 ⊗ X2 + X2 ⊗ 1, ∂2P = X1 ⊗ 1 + 1 ⊗ X1 and thus according to Proposition 3.6 (X2w)∗(X2w) = mX2(w ⊗ 1(∂1P )1 ⊗ w) = 0, (X1w)∗(X1w) = mX1(w ⊗ 1(∂2P )1 ⊗ w) = 0. We conclude X1w = X2w = 0, from which we may deduce like above by a second application of Proposition 3.6 that w = 0. Although the above examples might give the feeling that Proposition 3.6 is strong enough to allow directly a successive reduction of any polynomial, the needed algebraic manipulations turn out to be obscure in general. Anyhow, in contrast to Theorem 3.1, any result like this would need a symmetric starting condition. Thus, we will use the following general lemma, which is an easy consequence of the polar decomposition and encodes the additional information that our statement is formulated in a tracial setting. Lemma 3.10. Let x be an element of any tracial W ∗-probability space (M, τ ) over some complex Hilbert space H. Let pker(x) and pker(x∗) de- note the orthogonal projections onto ker(x) and ker(x∗), respectively. The projections pker(x) and pker(x∗) belong both to M and satisfy τ (pker(x)) = τ (pker(x∗)). Thus, in particular, if ker(x) is non-zero, then also ker(x∗) is a non- zero subspace of H. Proof. We consider the polar decomposition x = v(x∗x)1/2 = (xx∗)1/2v of x, where v ∈ M is a partial isometry mapping ran(x∗) to ran(x), such that v∗v = pran(x∗) and vv∗ = pran(x). Hence, it follows that 1 − v∗v = pran(x∗)⊥ = pker(x) from which we may deduce by traciality of τ that indeed and 1 − vv∗ = pran(x)⊥ = pker(x∗), τ (pker(x)) = τ (1 − v∗v) = τ (1 − vv∗) = τ (pker(x∗)). This concludes the proof. (cid:3) 16 T. MAI, R. SPEICHER, AND M. WEBER Combining Lemma 3.10 with Proposition 3.6 will provide us with the desired reduction argument. Before giving the precise statement, let us introduce some notation. If p ∈ M is any projection, we define a linear mapping ∆p,j : ChX1, . . . , Xni → ChX1, . . . , Xni for j = 1, . . . , n by ∆p,jP := (τ ⊗ id)((p ⊗ 1)∂jP ) for any P ∈ ChX1, . . . , Xni. Corollary 3.11. Let P ∈ ChX1, . . . , Xni and w = w∗ ∈ vN(X1, . . . , Xn) be given, such that P w = 0 holds true. If w 6= 0, then there exists a projection 0 6= p ∈ vN(X1, . . . , Xn) such that (∆p,jP )w = 0. Proof. Since P w = 0 and w 6= 0, we see that {0} 6= ran(w) ⊆ ker(P ), such that we also must have ker(P ∗) 6= {0} according to Lemma 3.10. The projection p := pker(P ∗) ∈ vN(X1, . . . , Xn) thus satisfies p 6= 0 and P ∗p = 0. Proposition 3.6 tells us that (p ⊗ 1)(∂jP )(1 ⊗ w) = 0 for j = 1, . . . , n holds true. Hence, we get that (∆p,jP )w = (τ ⊗ id)((p ⊗ 1)∂jP )w = (τ ⊗ id)((p ⊗ 1)(∂jP )(1 ⊗ w)) = 0, which concludes the proof. (cid:3) Now, we are prepared to finish the proof of Theorem 3.1. It suffices to show that, if P ∈ ChX1, . . . , Xni and w = w∗ ∈ vN(X1, . . . , Xn) with w 6= 0 are given such that P w = 0, then P = 0 follows. For seeing this, write P = a0 + d Xk=1 n Xi1,...,ik=1 ai1,...,ikXi1 . . . Xik. Assume that the total degree d of P satisfies d ≥ 1. We choose any summand of highest degree ai1,...,idXi1 . . . Xid of P , which is non-zero. Iterating Corollary 3.11, we see that there are non-zero projections p1, . . . , pd ∈ vN(X1, . . . , Xn) such that (∆pd,id . . . ∆p1,i1P )w = 0. But since we can easily check that ∆pd,id . . . ∆p1,i1P = τ (pd) . . . τ (p1)ai1,...,id, this leads us to ai1,...,id = 0, which contradicts our assumption. Thus, P must be constant, and since w 6= 0, we end up with P = 0. This concludes the proof of Theorem 3.1. ABSENCE OF ALGEBRAIC RELATIONS AND OF ZERO DIVISORS 17 References [Dab10a] Y. Dabrowski, A free arXiv:1008.4742v3 (2010). stochastic partial differential equation, [Dab10b] , A note about proving non-Γ under a finite non-microstates free Fisher information assumption, J. Funct. Anal. 258 (2010), no. 11, 3662 -- 3674. [Shl00] D. Shlyakhtenko, Free entropy with respect to a completely positive map, [SS13] Am. J. Math. 122 (2000), no. 1, 45 -- 81. D. Shlyakhtenko and P. Skoufranis, Freely independent random variables with non-atomic distributions, arXiv:1305.1920 (2013). [Voi93] D. Voiculescu, The analogues of entropy and of Fisher's information mea- sure in free probability theory, I, Commun. Math. Phys. 155 (1993), no. 1, 71 -- 92. [Voi94] [Voi96] [Voi97] [Voi98] [Voi99] , The analogues of entropy and of Fisher's information measure in free probability theory, II, Invent. Math. 118 (1994), no. 3, 411 -- 440. , The analogues of entropy and of Fisher's information measure in free probability theory, III: The absence of Cartan subalgebras, Geom. Funct. Anal. 6 (1996), no. 1, 172 -- 199. , The analogues of entropy and Fisher's information measure in free probability theory, IV: Maxiumum entropy and freeness, Fields Inst. Commun. 12 (1997), 293 -- 302. , The analogues of entropy and of Fisher's information measure in free probability theory, V: Noncommutative Hilbert transforms, Invent. Math. 132 (1998), no. 1, 189 -- 227. , The analogues of entropy and of Fisher's information measure in free probability theory, VI: Liberation and mutual free information, Adv. Math. 146 (1999), no. 2, 101 -- 166, art. no. aima.1998.1819. [Voi02] , Free entropy, Bull. Lond. Math. Soc. 34 (2002), no. 3, 257 -- 278. Universitat des Saarlandes, FR 6.1−Mathematik, 66123 Saarbrucken, Germany E-mail address: [email protected] Universitat des Saarlandes, FR 6.1−Mathematik, 66123 Saarbrucken, Germany E-mail address: [email protected] Universitat des Saarlandes, FR 6.1−Mathematik, 66123 Saarbrucken, Germany E-mail address: [email protected]
1505.04751	1	1505	2015-05-18T18:32:05	Finite digraphs and KMS states	[ "math.OA" ]	The paper contains a description of the KMS states and ground states of a generalized gauge action on the C*-algebra of a finite graph.	math.OA	math	FINITE DIGRAPHS AND KMS STATES JOHANNES CHRISTENSEN AND KLAUS THOMSEN 1. Introduction In a recent paper by an Huef, Laca, Raeburn and Sims, [aHLRS], the authors describe an algorithm by which it is possible to determine all the KMS states of the gauge action on the C ∗-algebra of a finite graph. Their results cover also the gauge action on the Toeplitz extension of the algebra and extend the result of Enomoto, Fujii and Watatani, [EFW], which deals with strongly connected graphs. Almost simultaneously with this work, Carlsen and Larsen obtained an abstract description of the KMS states for some of the generalized gauge actions on the C ∗-algebra of a finite graph as well as its Toeplitz extension. Their work build on and extend methods and results obtained by Exel and Laca in [EL] and bring our knowledge about the KMS states of the actions they consider to the point where the work on the gauge action begins in [aHLRS]. It is the purpose of the present paper to take the steps from the abstract to the concrete which were taken by an Huef, Laca, Raeburn and Sims, but now for all the generalized gauge actions. The point of departure for our work are results of the second author from [Th3] from which it follows that the relevant results of Carlsen and Larsen from [CL] remain valid for all generalized gauge actions, provided attention is restricted to the KMS states that are gauge invariant; a condition which is automatically satisfied for the actions considered by Carlsen and Larsen. What we do first is to develop the approach from [aHLRS] to make it applicable to generalized gauge actions. In this way we obtain a description of the gauge invariant KMS states for all generalized gauge actions. The main input for this is a generalization of the Perron-Frobenius theory for positive matrices which was obtained by Victory, [Vi]. See also [Ta]. The theory handles arbitrary finite non-negative matrices and can also be used to simplify some of the steps in [aHLRS]. We give here a new proof of the relevant results from [Vi] and [Ta] by using ideas from [aHLRS]. In order to identify the KMS states that are not gauge invariant we use results by Neshveyev, [N], in a form presented in [Th1]. By combining the result with our study of the gauge invariant KMS states we obtain in Theorem 5.2 our main result which describes the β-KMS states for all β ∈ R\{0} and for an arbitrary generalized gauge action on the C ∗-algebra of a finite graph. As with the gauge action, [aHLRS], it is a sub-collection of the components and the sinks in the graph that parametrize the extremal KMS states, although in general some of the components, corresponding to a loop without exits, may contribute a family of extremal KMS states parametrized by a circle. Which components and sinks play a role depends very much on the action, as we illustrate by examples. It is intrinsic to our approach that the case β = 0, where the KMS states are the trace states, must be handled separately as we do in Section 5.1. For completeness Version: August 8, 2018. 1 2 JOHANNES CHRISTENSEN AND KLAUS THOMSEN we describe also in a final section the ground states for the same actions. While there are no ground states for the gauge action unless the graph has sinks, this is no longer the case for generalized gauge actions and even for strongly connected graphs their structure can be very rich. Acknowledgement. The authors thank Astrid an Huef and Iain Raeburn for dis- cussions on the subject of this paper. 2. Preparations Let G be a finite directed graph with vertex set V and edge set E. The maps r, s : E → V associate to an edge e ∈ E its source vertex s(e) ∈ V and range vertex r(e) ∈ V . Thus the set of edges emitted from a vertex v is the set s−1(v) while r−1(v) is the set of edges terminating at v. A sink in G is a vertex v that does not emit an edge, i.e. s−1(v) = ∅. Formulated in terms of generators and relations the C ∗-algebra C ∗(G) of G is the universal C ∗-algebra generated by a set Se, e ∈ E, of partial isometries and a set Pv, v ∈ V , of mutually orthogonal projections such that 1) S∗ e Se = Pr(e) ∀e ∈ E, and 2) Pv = Xe∈s−1(v) SeS∗ e for every vertex v ∈ V which is not a sink. (2.1) A finite path µ in G is an element µ = e1e2 · · · en ∈ En for some n ∈ N such that r(ei) = s(ei+1), i = 1, 2, · · · , n − 1. For such a path we set Sµ = Se1Se2 · · · Sen−1Sen. The number µ = n is the length of the path. We consider a vertex v as a path ν of length 0, and set Sν = Pv in this case. Let Pf (G) denote the set of finite paths in G. Then A = Span {SµS∗ ν : µ, ν ∈ Pf (G)} (2.2) is a dense ∗-subalgebra of C ∗(G). Let F : E → R be a function. The universal property of C ∗(G) guarantees the existence of a one-parameter group αF αF t (Pv) = Pv ∀v ∈ V, t , t ∈ R, on C ∗(G) such that and αF t (Se) = eiF (e)tSe ∀e ∈ E. For β ∈ R a β-KMS state for αF is a state ϕ on C ∗(G) such that ϕ(ab) = ϕ(cid:0)bαF iβ(a)(cid:1) for all a, b ∈ A, cf. Definition 5.3.1 in [BR]. When F is constant 1 the automorphism group {α1 t } is the so-called gauge action and we study first the gauge invariant KMS states for αF , i.e the KMS states ϕ for αF with the additional property that ϕ ◦ α1 t = ϕ for all t ∈ R. For this purpose we use the following description of the gauge invariant KMS states. It was obtained by Carlsen and Larsen in [CL] when F is strictly positive (in which case all KMS states for αF are gauge-invariant). The general case follows from Theorem 2.8 in [Th3]. Let B be a non-negative matrix over V with the property that Bvw > 0 iff there is an edge in G from v to w. A vector ψ ∈ [0, ∞)V is almost harmonic for B (or almost B-harmonic) when Bvwψw = ψv (2.3) Xw∈V FINITE DIGRAPHS AND KMS STATES 3 for every vertex v ∈ V which is not a sink, and normalized when Pv∈V ψv = 1. When the identity (2.3) holds for all v ∈ V we say that ψ is harmonic for B (or B-harmonic). Thus an almost B-harmonic vector ψ is B-harmonic if and only if ψs = 0 for every sink s ∈ V . For β ∈ R, consider the matrix A(β) = (A(β)vw) over V defined such that A(β)vw = Xe∈vEw e−βF (e), where vEw = s−1(v) ∩ r−1(w). For a finite path µ = e1e2 · · · en in G, set F (µ) = F (e1) + F (e2) + · · · + F (en). Lemma 2.1. ([CL], [Th3]) For every normalized A(β)-almost harmonic vector ψ there is a unique gauge invariant β-KMS state ωψ for αF such that ωψ (SµS∗ ν ) = δµ,νe−βF (µ)ψr(µ) (2.4) for every pair µ, ν of finite paths in G. Furthermore, every gauge invariant β-KMS state for αF arises from a normalized A(β)-almost harmonic vector in this way. By Lemma 2.1 the study of the gauge invariant KMS states becomes a study of normalized almost harmonic vectors for the family A(β), β ∈ R, of non-negative matrices over V . 3. Almost harmonic vectors for a non-negative matrix Let B be a non-negative matrix over V with the property that Bvw > 0 iff there is an edge in G from v to w. We seek to obtain a description of the B-almost harmonic vectors. We shall need the following well-known lemma, cf. e.g. 6.43 in [W]. Lemma 3.1. (Riesz decomposition.) Let ψ = (ψv)v∈V ∈ [0, ∞[V be a non-negative vector such that for all v ∈ V . It follows that there are unique non-negative vectors h, k ∈ [0, ∞[V such that h is B-harmonic and Bn vwkw (3.1) for all v ∈ V . The vector k is given by Bvwψw ≤ ψv Xw∈V ∞Xn=0 ψv = hv +Xw∈V kv = ψv −Xw∈V n→∞Xw∈V ∞Xn=0 hv = lim Bn vs < ∞ Bvwψw, Bn vwψw. while We say that a sink s ∈ V is B-summable when 4 JOHANNES CHRISTENSEN AND KLAUS THOMSEN for all v ∈ V . For such a sink we define a vector φs ∈ [0, ∞)V by Lemma 3.2. φs in an extremal normalized B-almost harmonic vector. Proof. The only assertion which may not be straightforward to verify is that φs is extremal in the set of normalized B-almost harmonic vectors. To show this, consider a B-almost harmonic vector ϕ with the property that ϕ ≤ φs. Since Bm vwϕw ≤ Bm vwφs → 0 (3.2) as m → ∞, it follows that the harmonic part from the Riesz decomposition of ϕ is zero. Thus n=0 Bn vs φs v = P∞ Pw∈VP∞ . n=0 Bn ws n=m Bn vs n=0 Bn ws w ≤ P∞ Pw∈VP∞ ∞Xn=0 ϕv =Xw∈V vwkw Bn where kv = ϕv −Pw∈V Bvwϕw. Note that kv = 0 when v is not a sink since ϕ is B-almost harmonic, and that ks′ = ϕs′ for every sink s′. Note also that φs s′ = 0 for every sink s′ in G other than s. Since ϕ ≤ φs it follows that the same is true for ϕ. Hence Bn vsϕs = tφs v, ϕv = ∞Xn=0 t = ϕsXw∈V Bn ws. ∞Xn=0 (cid:3) where By combining Lemma 3.1 and Lemma 3.2 we obtain the following Proposition 3.3. Let ψ be a normalized B-almost harmonic vector. There are a unique (possibly empty) set S of summable sinks in G, unique positive numbers ts ∈]0, 1], s ∈ S, and a unique B-harmonic vector h such that ψ = h +Xs∈S tsφs. We turn to a study of the B-harmonic vectors. For any pair of subsets E, D ⊆ V we let BE,D denote the E × D-matrix obtained by restricting B to E × D, and we set BE = BE,E for any subset E ⊆ V . Write v w between two vertexes v, w when there is a finite path µ = e1 · · · en in G such that s(e1) = v and r(en) = w, and v ∼ w when v w and w v. Then ∼ is an equivalence relation since we consider a vertex v as a finite path (of length 0) from v to v. A component C in G is an equivalence class in V / ∼ such that BC 6= 0. For any collection F of vertexes in G we define the closure of F to be the set of vertexes that 'talk' to an element of F , i.e. v ∈ F if and only if there is a vertex w ∈ F such that v w. In contrast the hereditary closure of a set F consists of the vertexes w ∈ V such that v w for some v ∈ F . The hereditary closure will be denoted by bF . component C in G is B-harmonic when In the following we denote the spectral radius of a finite matrix A by ρ(A). A FINITE DIGRAPHS AND KMS STATES 5 a) ρ(cid:0)BC(cid:1) = 1 and b) ρ(cid:16)BC\C(cid:17) < 1 if C\C 6= ∅. This definition, as well as the proof of the following lemma, is inspired by Theorem 4.3 in [aHLRS]. Lemma 3.4. Let C be a B-harmonic component in G. There is a unique normalized B-harmonic vector φC such that BCφCC = φCC and φC v 6= 0 ⇔ v ∈ C. there is a strictly positive vector xC ∈ [0, ∞)C such that BCxC = xC. Since Proof. Existence: Since ρ(cid:0)BC(cid:1) = 1 it follows from Perron-Frobenius theory that ρ(cid:16)BC\C(cid:17) < 1, the matrix 1C\C − BC\C is invertible and we set φC =(cid:16)1C\C − BC\C(cid:17)−1 BC\C,CxC + xC, which is a strictly positive vector in [0, ∞)C. For any pair of vertexes v, w ∈ C\C, Using this and that no vertex in C talks to a vertex in C\C, we find that BC\C,CxC + BC\C,CxC + BCxC and hence n lim sup n ≤ ρ(cid:16)BC\C(cid:17) < 1, v,w(cid:1) 1 (cid:0)Bn ∞Xn=0(cid:16)BC\C(cid:17)n (cid:16)1C\C − BC\C(cid:17)−1 BCφC = BC\C(cid:16)1C\C − BC\C(cid:17)−1 = . = BC\C BC\C,CxC + BC\C,CxC + xC ∞Xn=0(cid:16)BC\C(cid:17)n = = ∞Xn=1(cid:16)BC\C(cid:17)n ∞Xn=0(cid:16)BC\C(cid:17)n = φC. BC\C,CxC + BC\C,CxC + xC (3.3) BC\C,CxC + xC Set φC v = 0 when v /∈ C and normalize the resulting vector in [0, ∞)V . It follows from (3.3) that φC is B-harmonic. Since φCC is multiple of xC by construction, it follows that BCφCC = φCC. Uniqueness: If ψ is a normalized B-harmonic vector such that BCψC = ψC and ψv 6= 0 ⇔ v ∈ C, it follows from Perron-Frobenius theory that there is a λ > 0 v ∀v ∈ C. Then ψ − λφC is vector supported in C\C such that such that ψv = λφC BC\C(ψ − λφC) = ψ − λφC. Since ρ(cid:16)BC\C(cid:17) < 1, it follows first that ψ = λφC and then that ψ = φC because both vectors are normalized. (cid:3) The following theorem is equivalent to the Frobenius-Victory theorem stated as Theorem 2.7 in [Ta]. 6 JOHANNES CHRISTENSEN AND KLAUS THOMSEN Theorem 3.5. Let ψ ∈ [0, 1]V be a normalized B-harmonic vector. There is a unique collection C of B-harmonic components in G and positive numbers tC ∈ ]0, 1], C ∈ C, such that tCφC. (3.4) ψ =XC∈C Proof. Set c(cid:13) = {v ∈ V : ψv > 0}. Let v ∈ c(cid:13). Since Bn that vvψv ≤ ψv for all n, it follows Hence lim sup n (Bn vv) 1 n ≤ 1. ρ (B c(cid:13)) = sup v∈ c(cid:13) lim sup n (Bn vv) 1 n ≤ 1. On the other hand, the fact that B c(cid:13)ψ c(cid:13) = ψ c(cid:13) implies that ρ (B c(cid:13)) ≥ 1, and we conclude that ρ (B c(cid:13)) = 1. (3.5) Since ρ (B c(cid:13)) = sup C ρ(cid:0)BC(cid:1) , where we take the supremum over the components of G contained in c(cid:13), the collection C′ of components C from G such that C ⊆ c(cid:13) and ρ(BC) = 1 is not empty. Order the elements of C′ such that C ≤ C ′ when the elements in C talk to the elements of C ′. Let C be the minimal elements of C′ with respect to this order. Let D ∈ C. We claim that D is a B-harmonic component, i.e. we assert that ρ(cid:16)BD\D(cid:17) < 1. Since D ⊆ c(cid:13) it follows from (3.5) that ρ(cid:16)BD\D(cid:17) ≤ 1. If ρ(cid:16)BD\D(cid:17) = 1, there must be one of G's components, say D′, contained in D\D such that ρ(cid:0)BD′(cid:1) = 1. But then D′ ∈ C′, D′ 6= D and D′ ≤ D, contradicting the minimality of D. Hence D is B-harmonic as claimed, and we conclude that C consists of B-harmonic components. Let D ∈ C. Then BDψD ≤ ψD so it follows from Perron-Frobenius theory that there is tD ≥ 0 such that ψD = tDψDD. Since ψD and ψDD are strictly positive, tD is positive too. Set η = ψ −XD∈C tDψD. We claim that η = 0. To show this, set K = SD∈C D, and note that ηK = 0. Let H be the hereditary closure of K, i.e. H = bK. Consider a D ∈ C. When v ∈ (H\K) ∩ D, there is a path from (some element of) D′ ⊆ K to v and a path from v to (some element of) D. Note that D′ 6= D since otherwise v would have to be an element of D ⊆ K. But D′ 6= D is impossible since D is minimal for the order on C′. Hence (H\K) ∩ D = ∅, showing that ψDH\K = 0. It follows that ηH\K = ψH\K, and hence that ηH ≥ 0. Let w ∈ H. There is an l ∈ N and v ∈ K such that Bl vwηw ≥ 0, implying that ηw = 0. Hence ηH = 0. Now note that vw 6= 0. Since Blη = η we find that 0 = ηv =Pu∈V Bl vuηu ≥ Bl (3.6) ρ(cid:0)B c(cid:13)\H(cid:1) < 1 FINITE DIGRAPHS AND KMS STATES 7 since all components D in c(cid:13) with ρ(cid:0)BD(cid:1) = 1 are contained in H. Since Bvwηw = ηv (cid:0)B c(cid:13)\Hη(cid:1)v = Xw∈ c(cid:13)\H Bvwηw =Xw∈V for all v ∈ c(cid:13)\H, it follows from (3.6) that η c(cid:13)\H = 0. Thus η = 0 as claimed and (3.4) follows. To prove the uniqueness part of the statement let D be a collection of B-harmonic components in G and sC, C ∈ D, positive numbers such that ψ =XC∈D sCφC. It follows that C ⊆ C ′ and that either C ′ = C or C ⊆ C ′\C ′. However, ρ(BC) = 1 Then c(cid:13) =SC∈C C =SC∈D C, so when C ∈ D there is a C ′ ∈ C such that C∩C ′ 6= ∅. while ρ(cid:16)BC′\C′(cid:17) < 1, and it follows therefore that C = C ′. In this way we conclude that D = C. Since the preceding argument shows that C ∩ C ′ = ∅ when C and C ′ are distinct elements from C, we find that and hence that sC = tC for all C ∈ C. (cid:3) sCφCC = ψC = tCφCC, Corollary 3.6. The normalized B-harmonic vectors constitute a finite dimensional simplex whose set of extreme points is (cid:8)φC : C a B-harmonic component in G(cid:9) . Combining Theorem 3.5 with Proposition 3.3 we obtain the following Corollary 3.7. The set of normalized B-almost harmonic vectors constitute a finite dimensional simplex whose set of extreme points is (cid:8)φC : C a B-harmonic component in G(cid:9) ∪ {φs : s a B-summable sink in G} . 4. Gauge invariant KMS states It follows from Lemma 2.1 and Corollary 3.7 that the gauge invariant β-KMS states for αF are determined by the A(β)-harmonic components and the A(β)- summable sinks. In this section we complete the description of the gauge invariant KMS states for β 6= 0 by finding the A(β)-harmonic components and the A(β)- summable sinks for each β ∈ R\{0}. 1 4.1. A(β)-harmonic components. A loop in G is a finite path µ = e1e2 · · · en (of positive length, i.e. n ≥ 1) such that s(e1) = r(en). If a component C only contains a single loop, we call it circular. Lemma 4.1. Let C ⊆ V be a component. The function is log-convex and continuous. R ∋ β 7→ ρ(cid:0)A(β)C(cid:1) 1We could have handled the case β = 0 here also, but it does simplify things a little when β 6= 0, and we will have to consider the β = 0 case separately for other reasons anyway. 8 JOHANNES CHRISTENSEN AND KLAUS THOMSEN vertex on this loop. It follows that log ρ(cid:0)A(β)C(cid:1) ≥ 1 Proof. Since C is a component there is a loop in C, of length p, say. Let v be a vv, showing that the logarithm of the function we consider takes finite values for all β. Its continuity follows therefore from its log-convexity which is established as follows. Let v ∈ C and β, β′ ∈ R, t ∈ [0, 1]. For each n ∈ N let vEnv denote the set of paths of length n from v back to itself. Then p log(cid:0)A(β)C(cid:1)p (cid:0)A(tβ + (1 − t)β′)C(cid:1)n Then Holders inequality shows that vv = Xµ∈vEnv (cid:0)A(tβ + (1 − t)β′)C(cid:1)n ρ(cid:0)A(tβ + (1 − t)β′)C(cid:1) = lim sup e−(tβ+(1−t)β′ )F (µ) = Xµ∈vEnv(cid:0)e−βF (µ)(cid:1)t(cid:16)e−β′F (µ)(cid:17)1−t vv ≤(cid:0)(cid:0)A(β)C(cid:1)n vv(cid:1)t(cid:0)(cid:0)A(β′)C(cid:1)n vv(cid:1)1−t (cid:0)(cid:0)A(tβ + (1 − t)β′)C(cid:1)n vv(cid:1) 1 . n n It follows that is dominated by the product which is what we needed to prove. ρ(cid:0)A(β)C(cid:1)t ρ(cid:0)A(β′)C(cid:1)1−t , . (cid:3) Lemma 4.2. Let C be a component in G which is not circular. ii) If F (µ) < 0 for all i) If F (µ) > 0 for all loops µ in C, there is a unique β0 ∈ R such that loops µ in C, there is a unique β0 ∈ R such that ρ(A(β0)C) = 1. This β0 is positive and ρ(cid:0)A(β)C(cid:1) < 1 if and only if β > β0. ρ(A(β0)C) = 1. This β0 is negative and ρ(cid:0)A(β)C(cid:1) < 1 if and only if β < β0. µ1, µ2 in C such that F (µ1) < 0 < F (µ2), it follows that ρ(cid:0)A(β)C(cid:1) > 1 for if F (µ) = 0 for some loop in C or there are loops all β ∈ R. iii) In all other cases, i.e. Proof. Some of the following arguments have appeared in [Th3]. i): We claim that β 7→ ρ(A(β)C) is strictly decreasing. To see this, set a = min {F (µ) : µ is a loop in C of length µ ≤ #C} . Consider β′ < β and a loop µ in C of length n. Then µ = µ1µ2 · · · µm, where each µi is a loop in C of length ≤ #C, and e−β′F (µ)eβF (µ) =Yj e(β−β′)F (µj ) ≥ em(β−β′)a ≥ e n #C (β−β′)a. Summing over all loops of length n starting and ending at the same vertex v in C, it follows first that and then that ρ(cid:0)A(β′)C(cid:1) = lim sup n n vv ≥ e (cid:0)A(β′)C(cid:1)n #C (β−β′)a(cid:0)A(β)C(cid:1)n n ≥ ρ(cid:0)A(β)C(cid:1) e (cid:0)(cid:0)A(β′)C(cid:1)n vv(cid:1) 1 vv , 1 #C (β−β′)a > ρ(cid:0)A(β)C(cid:1) . This proves the claim. Note that A(0)C is the adjacency matrix of the subgraph H of G whose vertex set is C. This is a finite strongly connected graph and it is well-known, and easy to show, that ρ(A(0)C) > 1 because H by assumption consists of more than a single loop. In view of Lemma 4.1 it suffices now to show that FINITE DIGRAPHS AND KMS STATES 9 limβ→∞ ρ(A(β)C) = 0. To this end note that any path in H of length ≥ #C must visit at least one vertex twice. It follows that for any path µ ∈ Pf (H) of length n with r(µ) = s(µ) there is a finite collection {ν1, ν2, · · · , νN } ⊆ {µ ∈ Pf (H) : 1 ≤ µ ≤ #C, s(µ) = r(µ)} such that N ≥ n #C and Let β > 0 and v ∈ C. Then F (νj) ≥ Na ≥ na #C . e−βF (µ) ≤ A(0)n vve− βan #C , F (µ) = NXj=1 vv = Xµ∈vEnv ρ(cid:0)A(β)C(cid:1) = lim sup A(β)n n Hence n ≤ ρ(cid:0)A(0)C(cid:1) e− βa (cid:0)(cid:0)A(β)C(cid:1)n vv(cid:1) 1 #C . Since a > 0, it follows that limβ→∞ ρ(A(β)C) = 0. The proof of ii) is analogous to that of i). iii): Assume first that F (µ) = 0 for some loop in C. Since we assume that C is not circular, there is a path ν such that ν = mµ for some m ∈ N, s(ν) = r(ν) = s(µ) and ν is not the composition of m copies of µ. It follows that, with v = s(µ), ≥ (e−βmF (µ) + e−βF (ν))n = (1 + e−βF (ν))n (cid:0)A(β)C(cid:1)nmµ vv ≥(cid:16)(cid:0)A(β)C(cid:1)mµ vv (cid:17)n for all n ∈ N, showing that ρ(cid:0)A(β)C(cid:1) ≥(cid:0)1 + e−βF (ν)(cid:1) 1 mµ > 1 for all β ∈ R. Assume then that there are loops µ1, µ2 in C such that F (µ1) < 0 < F (µ2). We may assume that µ1 and µ2 start at the same vertex v, if necessary after a modification of µ1 or µ2. Then vv for all n ∈ N, proving that (cid:0)A(β)C(cid:1)nµ1µ2 ρ(cid:0)A(β)C(cid:1) ≥ max(cid:26)e−β ≥ max(cid:8)e−βnµ2F (µ1), e−βnµ1F (µ2)(cid:9) µ2 (cid:27) > 1 F (µ1) µ1 , e−β F (µ2) for all β 6= 0. This completes the proof because ρ(cid:0)A(0)C(cid:1) > 1 since C is not circular. (cid:3) Lemma 4.3. Let C be a circular component consisting of the vertexes in the loop µ. Then for all β ∈ R. Proof. Left to the reader. ρ(cid:0)A(β)C(cid:1) = e−β F (µ) µ (cid:3) Let C be a component. It follows from Lemma 4.2 and Lemma 4.3 that when F (µ) > 0 for every loop µ in C, or F (µ) < 0 for every loop in C, there is a unique number βC ∈ R such that ρ(cid:0)A(βC)C(cid:1) = 1. 10 JOHANNES CHRISTENSEN AND KLAUS THOMSEN Definition 4.4. A non-circular component C in G is a KMS component of positive type when i) F (µ) > 0 for every loop µ in C, and ii) βC′ < βC for every component C ′ in C\C, if any. Similarly, a non-circular component C in G is a KMS component of negative type when i) F (µ) < 0 for every loop µ in C, and ii) βC < βC′ for every component C ′ in C\C, if any. Lemma 4.5. i) Let β > 0. A non-circular component C is A(β)-harmonic if and only if C is a KMS component of positive type and βC = β. ii) Let β < 0. A non-circular component C is A(β)-harmonic if and only if C is a KMS component of negative type and βC = β. Proof. The proofs of the two cases are identical and we consider here only case i): By In view of Lemma 4.2 the first condition is equivalent to F (µ) being strictly positive definition, C is A(β)-harmonic if and only if ρ(cid:0)A(β)C(cid:1) = 1 and ρ(cid:16)A(β)C\C(cid:17) < 1. for every loop µ in C and that βC = β. Note that ρ(cid:16)A(βC)C\C(cid:17) = 0 when C\C is non-empty, but does not contain any components, while ρ(cid:16)A(βC)C\C(cid:17) = maxnρ(cid:16)A(βC)C′(cid:17) : C ′ a component in C\Co otherwise. In view of i) in Lemma 4.2 and Lemma 4.3 this shows that the second condition, Assume then that C is a KMS component of negative type. If there are no compo- nents in C\C, we set IC = R. Otherwise, set IC = ]−∞, βC[, where βC = max(cid:8)βC′ : C ′ a component in C\C(cid:9) . βC = min(cid:8)βC′ : C ′ a component in C\C(cid:9) . ρ(cid:16)A(βC)C\C(cid:17) < 1, holds if and only if F (µ) > 0 for every loop µ in C\C and βC′ < βC for every component in C\C. (cid:3) We consider then the circular components. Definition 4.6. A circular component C in G is a KMS component of positive type when i) F (ν) = 0 for the loop ν in C, ii) F (µ) > 0 for all loops µ in C \ C, if any. Similarly, a circular component C in G is a KMS component of negative type when i) F (ν) = 0 for the loop ν in C, and ii) F (µ) < 0 for all loops µ in C \ C, if any. Unlike non-circular components, a circular component C can be a KMS component of both positive and negative type. This occurs when there are no loops in C\C. Let C be a circular component. Assume that C is a KMS component of positive type. If there are no components in C\C, it follows ρ(cid:16)A(β)C\C(cid:17) = 0 for all β ∈ R and we set IC = R in this case. Otherwise, set IC = ]βC, ∞[, where FINITE DIGRAPHS AND KMS STATES 11 In analogy with Lemma 4.5 we have the following. Lemma 4.7. i) Let β > 0. A circular component C is A(β)-harmonic if and only if C is a KMS component of positive type and β ∈ IC. ii) Let β < 0. A circular component C is A(β)-harmonic if and only if C is a KMS component of negative type and β ∈ IC. Proof. Basically the same as for Lemma 4.5. (cid:3) 4.2. A(β)-summable sinks. Definition 4.8. A sink s in G is a KMS sink of positive type when F (µ) > 0 for every loop µ in {s}, if any, and a KMS sink of negative type when F (µ) < 0 for every loop µ in {s}, if any. When there are no loops in {s} we set Is = R. When s is a KMS sink of positive type with components in {s}, we set Is = ]βs, ∞[ where βs = maxnβC′ : C ′ a component in {s}o . Similarly, when s is a KMS sink of negative type with components in {s}, we set Is = ]−∞, βs[ where βs = minnβC′ : C ′ a component in {s}o . i) Let β > 0. A sink s in G is A(β)-summable if and only if s is Lemma 4.9. a KMS sink of positive type and β ∈ Is. ii) Let β < 0. A sink s in G is A(β)-summable if and only if s is a KMS sink of negative type and β ∈ Is. Proof. Left to the reader. (cid:3) 4.3. The gauge invariant β-KMS states, β 6= 0. For β ∈ R\{0}, let C(β) be the set of non-circular KMS components C such that βC = β, and Z(β) the set of circular KMS components D such that β ∈ ID. Let S(β) be the set of KMS sinks s with β ∈ Is. We can then summarise our findings with regard to the gauge invariant KMS states as follows. Theorem 4.10. Let β ∈ R\{0}. For every gauge invariant β-KMS state ϕ for αF there are unique functions f : C(β) → [0, 1], g : Z(β) → [0, 1] and h : S(β) → [0, 1] such thatPC f (C) +PD g(D) +Ps h(s) = 1 and ϕ(SµS∗ ν ) = δµ,νe−βF (µ)φr(µ) for all finite paths µ, ν, where φ ∈ [0, ∞)V is the vector φv = XC∈C(β) f (C)φC v + XD∈Z(β) g(D)φD v + Xs∈S(β) h(s)φs v. 12 JOHANNES CHRISTENSEN AND KLAUS THOMSEN 5. Including the KMS states that are not gauge invariant To handle KMS states that are not gauge invariant we draw on the results of Neshveyev, [N]. For this it is necessary to introduce the groupoid picture of C ∗(G). Originally graph C ∗-algebras were introduced using groupoids, [KPRR], but only for row-finite graphs without sinks. For general graphs the realization as a groupoid C ∗-algebra was obtained by A. Paterson in [Pa]. To describe the groupoid for a general graph, possibly infinite but countable, let Pf (G) and P (G) denote the set of finite and infinite paths in G, respectively. The range and source maps, r and s on edges, extend in the natural way to Pf (G); the source map also to P (G). A vertex v ∈ V will be considered as a finite path of length 0 and we set r(v) = s(v) = v when v is considered as an element of Pf (G). Let V∞ be the set of vertexes v that are either sinks, or infinite emitters in the sense that s−1(v) is infinite. The unit space ΩG of G is the union ΩG = P (G) ∪ Q(G), where Q(G) = {p ∈ Pf (G) : r(p) ∈ V∞} is the set of finite paths that terminate at a vertex in V∞. In particular, V∞ ⊆ Q(G) because vertexes are considered to be finite paths of length 0. For any p ∈ Pf (G), let p denote the length of p. When p ≥ 1, set Z(p) = {q ∈ ΩG : q ≥ p, qi = pi, i = 1, 2, · · · , p} , and Z(v) = {q ∈ ΩG : s(q) = v} when v ∈ V . When ν ∈ Pf (G) and F is a finite subset of Pf (G), set The sets ZF (ν) form a basis of compact and open subsets for a locally compact 2 When µ ∈ Pf (G) and x ∈ ΩG, we can define the Hausdorff topology on ΩG. concatenation µx ∈ ΩG in the obvious way when r(µ) = s(x). The groupoid G consists of the elements in ΩG × Z × ΩG of the form (µx, µ − µ′, µ′x), for some x ∈ ΩG and some µ, µ′ ∈ Pf (G). The product in G is defined by (µx, µ − µ′, µ′x)(νy, ν − ν′, ν′y) = (µx, µ + ν − µ′ − ν′, ν′y), when µ′x = νy, and the involution by (µx, µ − µ′, µ′x)−1 = (µ′x, µ′ − µ, µx). To describe the topology on G, let ZF (µ) and ZF ′(µ′) be two sets of the form (5.1) with r(µ) = r(µ′). The topology we shall consider has as a basis the sets of the form {(µx, µ − µ′, µ′x) : µx ∈ ZF (µ), µ′x ∈ ZF ′(µ′)} . (5.2) With this topology G becomes an ´etale second countable locally compact Hausdorff groupoid and we can consider the reduced C ∗-algebra C ∗ r (G) as in [Re]. As shown by Paterson in [Pa] there is an isomorphism C ∗(G) → C ∗ r (G) which sends Se to 1e, where 1e is the characteristic function of the compact and open set {(ex, 1, r(e)x) : x ∈ ΩG} ⊆ G, 2Since we here deal with finite graphs where there are no infinite emitters, the topology has as an alternative basis the sets Z(ν), corresponding to ZF (ν) with F = ∅. )ZF (ν) = Z(ν)\ [µ∈F Z(µ)! . (5.1) FINITE DIGRAPHS AND KMS STATES 13 and Pv to 1v, where 1v is the characteristic function of the compact and open set {(vx, 0, vx) : x ∈ ΩG} ⊆ G. In the following we use the identification C ∗(G) = C ∗ r (G) and identify ΩG with the unit space of G via the embedding ΩG ∋ x 7→ (x, 0, x). In this way we get a canonical embedding C(ΩG) ⊆ C ∗(G) and there is a conditional expectation P : C ∗(G) → C(ΩG) defined such that P (f )(x) = f (x, 0, x) when f ∈ Cc(G), cf. [Re]. This conditional expectation can be used to characterise the gauge invariant KMS states because it follows from Theorem 2.2 in [Th2] that a KMS state for αF is gauge invariant if and only if it factorises through P . To describe the automorphism group αF in the groupoid picture we define a continuous homomorphism cF : G → R by cF (ux, u − u′, u′x) = F (u) − F (u′). The automorphism group αF on C ∗ r (G) is then defined such that t (f )(γ) = eitcF (γ)f (γ) αF when f ∈ Cc(G), cf. [Re]. Thanks to this picture of C ∗(G) and αF , and because we consider finite graphs in this paper, we can draw on the results of Neshveyev, [N], to obtain a decomposition of the KMS states into those that are gauge invariant and those that are not. Since the groupoid G has the additional properties required in Section 2 of [Th1] we can use the description obtained in Theorem 2.4 of [Th1] when β 6= 0. Of the β-KMS states considered in Theorem 2.4 in [Th1], it is only those of the form ωϕ O which may not factor through P . Here O is an orbit in ΩG under the canonical action of the groupoid G on its unit space, and O must be consistent and β-summable for ωϕ O to be defined. Furthermore, the formula for ωϕ O shows that it is only if the points in O have non-trivial isotropy group in G that ωϕ O does not factor through P . Note that the isotropy group Gx x ⊆ G of an element x ∈ ΩG is trivial unless x is an infinite path in G which is pre-period. Its orbit under G is then the orbit of an infinite periodic path. We may therefore assume that there is a loop δ in G such that x = δ∞ ∈ P (G). Then Gx x = {(x, kp, x) : k ∈ Z} , where p is the period of δ∞. We may assume that p = δ and find then that cF (x, kp, x) = kF (δ). It follows that the G-orbit Gx is consistent in the sense used in [Th1] if and only if F (δ) = 0. If the component of G containing δ contains a second loop, there will be another loop δ′ in G starting and ending at the same vertex as δ. Then xn = δnδ′δ∞, n ∈ N, are distinct elements in Gx, and when we use the notation from [Th1], we have that This shows that lx(xn) = e−F (δ′). lx(z)β = ∞ Xz∈Gx 14 JOHANNES CHRISTENSEN AND KLAUS THOMSEN for all β ∈ R, and we conclude therefore that Gx is not β-summable for any β ∈ R. It follows that the only G-orbits of elements with non-trivial isotropy groups which can be both consistent and β-summable in the sense of [Th1], are the G-orbits of a periodic infinite path lying in a circular component consisting of a loop δ with F (δ) = 0. On the other hand, for such an infinite path x the corresponding G-orbit will be β-summable if and only if e−βF (µ) < ∞, (5.3) Xµ∈E∗ δ s(x) where E∗ δ s(x) denotes the set of finite paths µ in G that terminate at s(x) ∈ V and do not contain δ. Note that (5.3) will hold if and only if C is a circular KMS component with β ∈ IC. In this case the β-KMS state ωϕ O is defined for every state ϕ on C ∗ (Gx x ), but it will only be extremal when ϕ is a pure state. By using the identification C ∗(Gx x ) = C(T) this means that the extremal β-KMS states occurring in Theorem 2.4 in [Th1] that are not gauge invariant arise from a number λ ∈ T, considered as a pure state on C(T), and a component C of zero type with β ∈ IC. We will denote this extremal β-KMS state by ωλ C. The formula for this state, as it was given in [Th1], becomes λke−βF (µ)f (µx, kp, µx) (5.4) ωλ C(f ) = Xν∈E∗ δ s(x) Xk∈Z Xµ∈E∗ δ s(x) −1 e−βF (ν) C(a) =ZT ωµ when f ∈ Cc(G). A general state ϕ on C(T) is given by integration against a Borel probability measure µ on T and the corresponding β-KMS state ωϕ O from [Th1], which we in the present setting will denote by ωµ C, is then given as an integral ωλ C(a) dµ(λ). (5.5) The conclusions we need here can then be summarised in the following way. Lemma 5.1. Let β ∈ R\{0}. For every β-KMS state ϕ for αF there is a Borel probability measure ν on ΩG, Borel probability measures µD, D ∈ Z(β), on T and numbers t and tD, D ∈ Z(β), in [0, 1] such that t +PD∈Z(β) tD = 1 and ϕ(a) = tZΩG P (a) dν + XD∈Z(β) tDωµD D (a). (5.6) The numbers t and tD are uniquely determined by ϕ, as are the Borel probability measures µD with tD > 0. The measure ν in Lemma 5.1 have certain properties which reflect that ϕ is a KMS state, and they can be found in [Th1], but what matters here is only that a 7→ZΩG P (a) dν is β-KMS state which is gauge invariant. It is therefore a convex combination of the states ϕC, ϕs, ϕD given by the formula (2.4) when the vector ψ occurring there is substituted by the A(β)-almost harmonic vectors φC, C ∈ C(β), φs, s ∈ S(β), and φD, D ∈ Z(β), respectively. Note that the state ϕD corresponding to a component D ∈ Z(β) is the same as the state ωm D from (5.5) when m is the normalized Lebesgue FINITE DIGRAPHS AND KMS STATES 15 measure on T. We can therefore now use Theorem 2.4 in [Th1] and combine Lemma 5.1 with Theorem 4.10 to obtain the following description of the β-KMS states when β 6= 0. Theorem 5.2. For β ∈ R\{0}, • let C(β) be the set of non-circular KMS components C in G with βC = β, • let S(β) be the set of KMS sinks s in G with β ∈ Is, and • let Z(β) be the set of circular KMS components D with β ∈ ID. For every β-KMS state ϕ for αF there are numbers αC ∈ [0, 1], C ∈ C(β), αs ∈ [0, 1], s ∈ S(β), and αD ∈ [0, 1], D ∈ Z(β), and Borel probability measures µD, D ∈ Z(β), on T, such thatPC αC +Ps αs +PD αD = 1, and αsϕs + XD∈Z(β) αCϕC + Xs∈S(β) ϕ = XC∈C(β) αDωµD D . The numbers αC, αs, αD are uniquely determined by ϕ, as are the Borel probability measures µD for the components D ∈ Z(β) with αD > 0. 5.1. Trace states. We need a different approach when β = 0. Since the 0-KMS states are the trace states of C ∗(G) we must determine these. Let Z(0) denote the set of circular components C in G with the property that C\C does not contain any components, and similarly S(0) the set of sinks s in G such that {s}\{s} does not contain a component. For every C ∈ Z(0) the set V \C is hereditary and saturated, and there is a surjective ∗-homomorphism πC : C ∗(G) → C ∗(C), where C is considered as a directed graph with vertex set C ⊆ V when s ∈ S(0) there is also a surjective ∗-homorphism πs : C ∗(G) → C ∗({s}), where {s} is considered as a directed graph with vertex set {s} ⊆ V and the edge set and the edge set (cid:8)e ∈ E : s(e), r(e) ∈ C(cid:9), cf. Theorem 4.1 in [BPRS]. Similarly, ne ∈ E : s(e), r(e) ∈ {s}o. When s ∈ S(0) we let ns be the number of paths in G terminating at s. When C ∈ Z(0) we choose a vertex vC ∈ C and set nC = # {µ ∈ Pf (G) : r(µ) = vC, s(µi) 6= vC, for i ≤ µ} , where the condition that s(µi) 6= vC is negligible when µ = 0. Theorem 5.3. For every s ∈ S(0), C ∗({s}) ≃ Mns(C), and for every C ∈ Z(0), C ∗(C) ≃ MnC (C(T)) , C ∈ Z(0). For every trace state ω on C ∗(G) there are unique numbers αs ∈ [0, 1] and αC ∈ [0, 1], and trace states ωs on C ∗({s}) and ωC on C ∗(C), s ∈ S(0), C ∈ Z(0), such that and αC = 1 Xs∈S(0) αs + XC∈Z(0) αsωs ◦ πs + XC∈Z(0) ω = Xs∈S(0) αCωC ◦ πC. 16 JOHANNES CHRISTENSEN AND KLAUS THOMSEN For the proof of Theorem 5.3 set Then N is hereditary and saturated, and the set {Pv : v ∈ N} generates an ideal IN in C ∗(G) such that C ∗(G)/IN ≃ C ∗( G) where G is the graph with vertex set N = V \ [C∈Z(0) V = [C∈Z(0) C ∪ [s∈S(0) {s} . C ∪ [s∈S(0) {s} and edge set E = {e ∈ E : r(e) /∈ N}, cf. Theorem 4.1 in [BPRS]. Lemma 5.4. Let ω be a trace state on C ∗(G). Then ω(IN ) = 0. Proof. It suffices to show that ω(Pv) = 0 when v ∈ N. To this end consider a loop µ in G with vertexes v1, . . . , vn, v1. The Cuntz-Krieger relations (2.1) imply: ω(Pv1) = ω(cid:0) Xe∈s−1(v1) = Xe∈s−1(v2) SeS∗ e(cid:1) = Xe∈s−1(v1) ω(S∗ e Se) = Xe∈s−1(v1) ω(Pr(e)) ≥ ω(Pv3) = · · · ≥ ω(Pvn) = Xe∈s−1(vn) ω(Pr(e)) ≥ ω(Pv2) ω(Pr(e)) ≥ ω(Pv1). Hence we must have equality everywhere, which implies that ω(Pr(e)) = 0 if e ∈ s−1(vi) for some i, but e /∈ µ. It follows from this that ω(Pw) = 0 when w ∈ [C∈C bC \ [C′∈Z(0) C ′ where C is the set of components. Hence if s is a sink in G it follows that ω(Ps) = 0 unless s ∈ S(0). Consider a vertex v ∈ N. If v is sink, ω(Pv) = 0 and we are done. Otherwise, if ω(Pv) > 0, the Cuntz-Krieger relations (2.1) implies that there is an edge e1 ∈ s−1(v) such that ω(Pr(e1)) > 0. Then r(e1) can not be a sink and we can find an edge e2 such that s(e2) = r(e1) and ω(Pr(e2)) > 0. We can continue this construction of edges ei indefinitely so there are i < i′ such that s(ei) = r(ei′), and the path eiei+1 · · · ei′ is contained in a component C. Since ω(Pr(ei)) > 0 this component must be circular and without components in C\C, which contradicts that v ∈ N. It follows that ω(Pv) = 0. (cid:3) For each C ∈ Z(0), fix a vertex vC ∈ C, and set vs = s for s ∈ S(0). For all v ∈ V and a ∈ Z(0) ∪ S(0), we define: N a v = {µ ∈ Pf ( G) s(µ) = v , r(µ) = va , s(µi) 6= va for i ≤ µ} where the condition that s(µi) 6= va is negligible when µ = 0. We define N a = v for a ∈ Z(0) ∪ S(0). Sv∈ V N a Lemma 5.5. C ∗( G) ≃(cid:16) Ms∈S(0) M#N s(C)(cid:17) ⊕(cid:16) MC∈Z(0) M#N C (C(T))(cid:17) FINITE DIGRAPHS AND KMS STATES 17 Proof. For a ∈ Z(0) ∪ S(0), let eα,β, α, β ∈ N a be the standard matrix units in MN a(C) ≃ M#N a(C). For v ∈ V , set Pv = Xa∈Z(0)∪S(0) Xα∈N a v eα,α. Then Pv, v ∈ V , are mutually orthogonal projections. For each f ∈ E such that s(f ) /∈ {va : a ∈ Z(0) ∪ S(0)}, set Sf = Xa∈Z(0)∪S(0) Xα∈N a r(f ) ef α,α. If s(f ) ∈ {va : a ∈ Z(0) ∪ S(0)}, then s(f ) = vC for some C ∈ Z(0), and we let µC denote the unique shortest path in G with s(µC) = r(f ) and r(cid:0)µC(cid:1) = vC. We define an element Sf ∈ C (T, MN C (C)) such that Sf (z) = zevC ,µC . It is straightforward to verify that Pv, v ∈ V , and Sf , f ∈ E, is a Cuntz-Krieger family, i.e. they satisfy (2.1) relative to G. Since Pv, Sf ∈(cid:16) Ms∈S(0) M#N s(C)(cid:17) ⊕(cid:16) MC∈Z(0) M#N C (C(T))(cid:17) for all v ∈ V and all f ∈ E, the universal property of C ∗( G) gives us a canonical ∗-homomorphism C ∗( G) → (cid:16) Ms∈S(0) M#N s(C)(cid:17) ⊕(cid:16) MC∈Z(0) M#N C (C(T))(cid:17). To show that this is an isomorphism, note first that it is surjective because the target algebra is generated as a C ∗-algebra by Pv, v ∈ V , and Sf , f ∈ E. For the injectivity we shall appeal to the gauge-invariant uniqueness theorem, Theorem 2.1 in [BPRS]. For a a ∈ S(0) ∪ Z(0), define for each ω ∈ T the unitary: U a ω = Xα∈N a ωαeα,α For s ∈ S(0) we define an automorphism ψs and for C ∈ Z(0) we define an automorphism on M#N C (C(T)) by ψC U C ω . It is straightforward to check that: ω on M#N s(C) by ψs ω f (ω#Cz)U C ω(A) = U s ωAU s ω, ω (f )(z) = T ∋ ω → ψω := (Ms∈S ψs ω) ⊕ ( MC∈Z(0) ψC ω ) is an action, and that we for f ∈ E and v ∈ V have: ψω( Sf ) = ω Sf ψω( Pv) = Pv for all ω ∈ T. It follows therefore from Theorem 2.1 in [BPRS] that the homomor- phism under consideration is injective. (cid:3) 18 JOHANNES CHRISTENSEN AND KLAUS THOMSEN Proof of Theorem 5.3: Consider C ∈ Z(0) and let C ∗(G) → M#N C (C(T)) be the surjective ∗-homomorphism obtained by composing the quotient map C ∗(G) → C ∗( G) with the projection C ∗( G) → M#N C (C(T)) obtained from Lemma 5.5. The kernel of this ∗-homomorphism is the same as the kernel of πC : C ∗(G) → C ∗(C), namely the ideal generated by (cid:8)Pv : v /∈ C(cid:9) . In the same way we see that C ∗({s}) ≃ It follows that C ∗(C) ≃ M#N C (C(T)). M#N s(C) when s ∈ S(0). The statements regarding a trace state ω follow from Lemma 5.4 and Lemma 5.5. (cid:3) 6. Ground states To describe the ground states we use again the groupoid picture described in Section 5 in order to adapt the approach from Section 5 in [Th4] to the present setting. The fixed point algebra of αF is the C ∗-algebra of the open sub-groupoid F = {(µx, µ − µ′, µ′x) : x ∈ ΩG, F (µ) = F (µ′)} of G. The conditional expectation Q : C ∗(G) → C ∗ r (F ) extending the restriction map Cc(G) → Cc(F ) can be described as a limit: Q(a) = lim R→∞ cf. the proof of Theorem 2.2 in [Th3]. 1 RZ R 0 αF t (a) dt, (6.1) When x ∈ ΩG, z ∈ Pf (G), write z ⊆ x when 1 ≤ z and x[1,z] = z or z = 0 and z = s(x). An element x ∈ ΩG has minimal F -weight when the following holds: z, z′ ∈ Pf (G), z ⊆ x, r(z′) = r(z) ⇒ F (z′) ≥ F (z) . We denote the set of elements in ΩG with minimal F -weight by Min(F, G). Then Min(F, G) is closed in ΩG and F -invariant in the sense that (x, k, y) ∈ F , x ∈ Min(F, G) ⇒ y ∈ Min(F, G). It follows that the reduction F Min(F,G) of F to Min(F, G), defined by F Min(F,G) = {(µx, µ − µ′, µ′x) : x ∈ ΩG, F (µ) = F (µ′), µx ∈ Min(F, G)} , is a locally compact ´etale groupoid. Furthermore, there is a surjective ∗-homomorphism R : C ∗ r (F ) → C ∗ r(cid:0)F Min(F,G)(cid:1) extending the restriction map Cc(F ) → Cc(cid:0)F Min(F,G)(cid:1). Now the proof of Theorem 5.3 in [Th4] can be repeated almost ad verbatim to yield the following. Theorem 6.1. The map ω 7→ ω ◦ R ◦ Q is an affine homeomorphism from the state space of C ∗ r(cid:0)F Min(F,G)(cid:1) onto the ground states of αF . The structure of the C ∗-algebra C ∗ When F is constant zero, it is equal to C ∗(G), and when F is strictly positive it is isomorphic to Cn, where n is the number of sinks in G. If G consists of three edges, ei, and a vertex v with r(ei) = s(ei) = v, i = 1, 2, 3, and if F (e1) = F (e2) = 0 while r(cid:0)F Min(F,G)(cid:1) varies a lot with the choice of F . FINITE DIGRAPHS AND KMS STATES 19 F (e3) = 1, we find that C ∗(G) is the Cuntz-algebra O3 while C ∗ copy of O2. Which of the ground states are weak* limits, for β → ∞, of β-KMS states, can be decided by combining Theorem 6.1 with Theorem 5.2. It follows, for instance, that they all are when F = 1, while none of them are in the last mentioned example. r(cid:0)F Min(F,G)(cid:1) is a 7. An example Consider the following graph G. The two sinks are s1 and s2 and there are four components labelled C1 through C4. In order to define various functions on the edge set we have labelled four edges a, b, c and d. a b d c s1 s2 C1 C2 C3 C4 Consider first the gauge action where F (e) = 1 for all edges e. The two sinks are both KMS sinks in this case; with intervals Is1 = R and Is2 = (cid:3) log 2 the components it is only C2 and C4 that are KMS components, both of positive type and with βC2 = βC4 = log 2 2 . There are three extremal β-KMS states when β = log 2 2 , coming from s1, and two when β > log 2 2 , coming from s1 and s2. This 'KMS spectrum' away from 0 can be described by the following figure. 2 , coming from s1, C2 and C4, one when β < log 2 2 , ∞(cid:2). Of s1 s2 C1 C2 C3 C4 R log(2)/2 KMS spectrum (β 6= 0) for the gauge action on C ∗(G). To define a different generalized gauge action, let E be the set of edges in G, and set F1(e) = 1 when e ∈ E\{a, b, c} while F1(a) = F1(b) = −2 and F1(c) = 0. If we describe the KMS-spectrum for the action αF1 by a diagram as was done for the gauge action, the picture becomes the following. The red line describes the contribution from the circular KMS component C3 and hence each point on it represents a family of extremal KMS states parametrized by a circle. 20 JOHANNES CHRISTENSEN AND KLAUS THOMSEN s1 s2 C1 C2 C3 C4 R − log(2) log(2)/2 KMS spectrum (β 6= 0) for the generalized gauge action αF1 on C ∗(G). Finally we consider F2 defined such that F2(e) = 1 when e ∈ E\{a, d}, F2(a) = −1 2. For the generalized gauge action αF2 we find the following KMS and F2(d) = − 3 spectrum. − log(4) s1 s2 C1 C2 C3 C4 R KMS spectrum (β 6= 0) for the generalized gauge action αF2 on C ∗(G). The structure of the ground states vary also for the three actions. For the gauge action there are two extremal ground states coming from the sinks, while for the actions αF1 and αF2 there are infinitely many. Concerning αF1 the sinks still con- tribute two, but the infinite path c∞ has minimal F1-weight and contributes a family of extremal ground states naturally parametrized by a circle. The sink s1 is the only sink which gives rise to an extremal ground state for the action αF2, but now the loop of period 2 beginning with the edge a is an element of Min(F2, G) and gives rise to a family of extremal ground states naturally parametrized by a circle. The 0-KMS states are of course the same for all three actions. They are the trace states on the algebra, and by using Theorem 5.3 we see that they can be identified with the trace states on M2(C) ⊕ M3(C(T)), where the sink s1 is responsible for the first summand and the component C1 for the second. References [BPRS] [BR] [CL] [EFW] [EL] T. Bates, D. Pask, I. Raeburn and W. Szymanski, The C ∗-algebra of Row-Finite Graphs, New York Jour. of Math. 6 (2000), 307-324. O. Bratteli and D.W. Robinson, Operator Algebras and Quantum Statistical Mechan- ics I + II, Texts and Monographs in Physics, Springer Verlag, New York, Heidelberg, Berlin, 1979 and 1981. T.M. Carlsen and N. Larsen, Partial actions and KMS states on relative graph C ∗- algebras, arXiv:1311.0912. M. Enomoto, M. Fujii and Y. Watatani, KMS states for gauge action on OA, Math. Japon. 29 (1984), 607-619. R. Exel and M. Laca, Partial dynamical systems and the KMS condition, Comm. Math. Phys. 232 (2003), 223-277. FINITE DIGRAPHS AND KMS STATES 21 [aHLRS] [KPRR] [N] [Pa] [Re] [Ta] [Th1] [Th2] [Th3] [Th4] [Vi] [W] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on the C ∗-algebras of reducible graphs, Ergodic Th. & Dynam. Syst., to appear. A. Kumjian, D. Pask, I. Raeburn and J. Renault, Graphs, Groupoids, and Cuntz- Krieger algebras, J. Func. Analysis 144 (1997), 505-541. S. Neshveyev, KMS states on the C ∗-algebras of non-principal groupoids, J. Operator Theory 70 (2013), 513-530. A.L.T. Paterson, Graph inverse semigroups, groupoids and their C ∗-algebras, J. Op- erator Theory 48 (2002), 645-662. J. Renault, A Groupoid Approach to C ∗-algebras, LNM 793, Springer Verlag, Berlin, Heidelberg, New York, 1980. B. Tam, The Cone-Theoretic Approach to the Spectral Theory of Positive Linear Operators: The Finite-Dimensional Case, Taiwan. J. Math. 5, 2001, 207-277. K. Thomsen, KM S-states and conformal measures, Comm. Math. Phys. 316 (2012), 615-640. DOI: 10.1007/s00220-012-1591-z K. Thomsen, KMS weights on groupoid and graph C ∗-algebras, J. Func. Analysis 266 (2014), 2959-2988. K. Thomsen, KMS weights on graph C ∗-algebras, arXiv:1409.3702 K. Thomsen, KMS weights on graph C ∗-algebras II. Factor types and ground states, arXiv:1412.6762 H. D. Victory, Jr., On nonnegative solutions of matrix equations, SIAM J. Alg. Disc. Meth. 6, 1985, 406-412. W. Woess, Denumerable Markov Chains, EMS Textbooks in Mathematics, 2009. E-mail address: [email protected]; [email protected] Department of Mathematics, Aarhus University, Ny Munkegade, 8000 Aarhus C, Denmark
1403.0927	3	1403	2015-02-10T19:17:28	Distance between unitary orbits of normal elements in simple C*-algebras of real rank zero	[ "math.OA" ]	Let $x, y$ be two normal elements in a unital simple C-algebra $A.$ We introduce a function $D_c(x, y)$ and show that in a unital simple AF-algebra there is a constant $1>C>0$ such that $$ C\cdot D_c(x, y)\le {\rm dist}({\cal U}(x),{\cal U}(y))\le D_c(x,y), $$ where ${\cal U}(x)$ and ${\cal U}(y)$ are the closures of the unitary orbits of $x$ and of $y,$ respectively. We also generalize this to unital simple C-algebras with real rank zero, stable rank one and weakly unperforated $K_0$-group. More complicated estimates are given in the presence of non-trivial $K_1$-information.	math.OA	math	Distance between unitary orbits of normal elements in simple C∗-algebras of real rank zero Shanwen Hu and Huaxin Lin∗ Abstract Let x, y be two normal elements in a unital simple C ∗-algebra A. We introduce a function Dc(x, y) and show that in a unital simple AF-algebra there is a constant 1 > C > 0 such that C · Dc(x, y) ≤ dist(U(x),U(y)) ≤ Dc(x, y), where U(x) and U(y) are the closures of the unitary orbits of x and of y, respectively. We also generalize this to unital simple C ∗-algebras with real rank zero, stable rank one and weakly unperforated K0-group. More complicated estimates are given in the presence of non-trivial K1-information. 1 Introduction Let H be a Hilbert space and let B(H) be the C ∗-algebra of all bounded operators. The study of normal operators in B(H) has a long history. It has been an interesting and important problem to determine when two normal operators are unitarily equivalent in a subalgebra A of B(H). Any detailed account of history will inevitably involve an enormous amount of literature. We will choose to limit ourselves to the immediate concerns of this paper. We study the distance between unitary orbits of normal elements. Some of the pioneer works on this subject have been made by Ken Davidson in [6], [5] and [7]. More recent work on the subject can be found in [29] and [30]. Let A be a unital C ∗-algebra and let x ∈ A be a normal element. The unitary orbit of x is defined to be the set {u∗xu : u ∈ U (A)}, where U (A) is the unitary group of A. Denote by U (x) the closure of the unitary orbit of x. Suppose that y ∈ A is another normal element. Denote by X and Y the spectrum of x and of y, respectively. Let ϕX , ϕY : C(X ∪ Y ) → A be the homomorphisms defined by ϕX (f ) = f (x) and ϕY (f ) = f (y) for all f ∈ C(X ∪ Y ). Suppose that A has a unique tracial state τ. Denote by µτ ◦ϕX and µτ ◦ϕY the two probability measures on X ∪ Y defined by the positive linear functionals τ ◦ ϕX and τ ◦ ϕY , respectively. For each open subset O ⊂ C and r > 0, denote by Or = {ξ ∈ C : dist(ξ, O) < r}. Define and define rO = inf{r : µτ ◦ϕX (O) ≤ µτ ◦ϕY (Or)} DT (x, y) = sup{rO : O open subset of C}. For finite dimensional Hilbert spaces, when A = Mn, the n × n matrix algebra over C, an application of the Marriage Theorem shows that ∗corresponding author: [email protected] dist(U (x),U (y)) ≤ DT (x, y) (e 1.1) 1 (see [12] and [6], for example). We first realize that (e 1.1) also holds for the case that A is a UHF-algebra. Apart from the application of the Marriage Theorem, the proof is also based on an important fact that normal elements in A can be approximated by normal elements with finite spectrum ([15]). If we allow A to be a general unital simple AF-algebra with a unique tracial state, the above bound no long works because of the presence of possible infinitesimal elements in K0(A). Even without infinitesimal elements in K0(A), in the case that A has infinitely many extremal tracial states, the usage of the Marriage Theorem has to mix appropriately with the Riesz interpolation property. With the help of the Cuntz semigroups, which are more appropriate tools to compare positive elements, we are able to establish a modified upper bound formula for the distance between unitary orbits of two normal elements in a general unital AF-algebra (see 3.7 below). The fact that normal elements in an AF-algebra can be approximated by normal elements with finite spectrum plays an essential role in the proof. This follows from the result that a pair of almost commuting self-adjoint matrices is close to a pair of commuting self-adjoint matrices (see [16] and [11]). In [17], it was shown that, in a unital separable simple C ∗-algebra of real rank zero, stable rank one and with weakly unperforated K0(A), a normal element x can be approximated by those normal elements with finite spectrum if λ − x ∈ Inv0(A) (the connected component of invertible elements of A containing the identity) for all λ 6∈ sp(x). In this case we also prove that the same upper bound works for distance between unitary orbits of two normal elements in A which have vanishing K1 information (see 3.6). The distance between unitary orbits of normal elements in unital purely infinite simple C ∗- algebras were recently studied in [29]. In that case, one could get a precise formula for the distance at least for the case that K0(A) = 0 when no K1-information involved. However, the distance is basically given by dH(sp(x), sp(y)), the Hausdorff distance between the spectra. It is the presence of the trace in the finite C ∗-algebras that makes our upper bound more complicated and sophisticated. But it is exactly this phenomenon that is exciting. However, it is more desirable, in this current study, to include the cases that C ∗-algebras have non-trivial K1-groups and normal elements have non-trivial K1-information. To study the unitary orbits of normal elements, one has to know when two normal elements are approximately unitarily equivalent. When A is a unital purely infinite simple C ∗-algebra, or A is a unital separable simple C ∗-algebra with finite tracial rank, we know exactly when two normal elements are approximately unitarily equivalent (see [4], [21] and [23]). These works actually deal with the problem of unitary orbits of homomorphisms from C(X), the C ∗-algebra of continuous functions on a compact metric space X, to a unital purely infinite simple C ∗- algebra or a unital separable simple C ∗-algebra with finite tracial rank. These studies are closely related to the Elliott program of classification of amenable C ∗-algebras. To consider the unitary orbits of normal elements in a unital separable simple C ∗-algebra A with real rank zero, stable rank one and weakly unperforated K0(A), we first present a theorem that two normal elements x, y ∈ A are approximately unitarily equivalent if and only if sp(x) = sp(y), (ϕx)∗i = (ϕy)∗i, i = 0, 1 and τ ◦ ϕx = τ ◦ ϕy for all τ ∈ T (A) (the tracial state space of A), where ϕx, ϕy : C(sp(x)) → A are defined by ϕx(f ) = f (x) and by ϕy(f ) = f (y) for all f ∈ C(sp(x)), respectively. This is a generalization of the similar result in [21] (which only works for unital simple C ∗-algebra with tracial rank zero). Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A) and let x, y ∈ A be two normal elements with sp(x) = X and sp(y) = Y, respectively. We first consider the case that (λ − x)−1(λ − y) ∈ Inv0(A) for all λ 6∈ X ∪ Y. With the help of a Mayer-Vietoris Theorem, we are able to present a reasonable upper bound for the distance between unitary orbits of x and y (see 6.7 below) in the same sprit of 3.6 mentioned above. However, there are normal elements with sp(x) = sp(y) which induce exactly the same 2 map on the Cuntz semigroup, but, for any given λ 6∈ X ∪ Y, (λ − x)−1(λ − y) 6∈ Inv0(A). In this case the same upper bound mentioned above is zero. Nevertheless, as first found by Davidson ([5]), dist(U (x),U (y)) ≥ sup{dist(λ, sp(x)) + dist(λ, sp(y))}, (e 1.2) where the supremum is taken among those λ 6∈ sp(x)∪sp(y) such that (λ−x)−1(λ−y) 6∈ Inv0(A). Based on 6.7, we also present an upper bound for distance between unitary orbits of normal elements in A, which is a combination of the upper bound in 6.7 together with the sprit of the lower bound given by Davidson (see 7.3 below). Then, of course, there is the issue of the lower bound. It was shown by Davidson ([6]) that there is a universal constant 1 > C > 0 such that dist(U (x),U (y)) ≥ C · DT (x, y) (e 1.3) in the case that A = Mn (or A = B(H) for infinite dimensional Hilbert space H with some modification). The constant is computed at least 1/3. It was shown even in the case that n ≥ 3, the constant C cannot be made equal to 1 ([13]). We show that a similar lower bound (with the same constant C) holds for unital AF-algebras and, more generally, for unital separable simple C ∗-algebras of tracial rank zero. A different lower bound dc(x, y) is also presented. There are cases that dc(x, y) = Dc(x, y) with dist(U (x),U (y)) = dc(x, y) and cases that dc(x, y) < Dc(x, y) with dist(U (x),U (y)) = Dc(x, y). Briefly, the paper is organized as follows: Section 2 serves as a preliminary of the paper. Some metrics associated with the measure distribution and the Cuntz semigroups are introduced. In Section 3, we present an upper bound for the distance between unitary orbits of normal elements in unital AF-algebras and in unital separable simple C ∗-algebras A with real rank zero, stable rank one and weakly unperforated K0(A), when K1-information of the relevant normal elements vanish. In Section 4, we show, with the metric introduced in Section 2, that normal elements can always be approximated by normal elements with finite spectrum in a unital simple C ∗-algebra with stable rank one, real rank zero and weakly unperforated K0(A). Another function De c (x, y) which is a modification of Dc(x, y) is also introduced. In Section 5, we show exactly when two normal elements in A are approximately unitarily equivalent. In Section 6, we present an upper bound for the distance between unitary orbits of normal elements x, y ∈ A under the condition that λ − x and λ − y give the same K1 element (but not necessarily zero) when λ 6∈ sp(x)∪ sp(y). In Section 7, we present a general upper bound for the distance between unitary orbits of normal elements in A (without any assumption on K1-information). In Section 8, we discuss the lower bound of the distance between unitary orbits of normal elements. Acknowledgements The most research of this work was done when both authors were at the Research Center for Operator Algebras in the East China Normal University and both were partially supported by the Center. The second named author also acknowledges the support from a grant of NSF. The authors would also like to thank the referee for carefully checking and for numerous suggestions. 2 Preliminaries Definition 2.1. Let A be a unital C ∗-algebra. Denote by U (A) the unitary group of A. Let x ∈ A be a normal element. Define U (x) to be the closure of {u∗xu : u ∈ U (A)}. 3 Definition 2.2. Fix a compact metric space Ω. Let r > 0. For each subset S ⊂ Ω, define Sr = {t ∈ Ω : dist(t, S) < r}, S−r = {t ∈ Ω : dist(t, Sc) > r} and, denote by ¯S the closure of S. Definition 2.3. Let A be a unital C ∗-algebra. Denote by T (A) the tracial state space of A. Let Aff(T (A)) be the space of all real affine continuous functions on T (A). Denote by ρA : K0(A) → Aff(T (A)) the order preserving homomorphism defined by ρA([p])(τ ) = (τ ⊗ T r)(p) for all projections in Mn(A), where T r is the standard trace on Mn, n = 1, 2, .... Definition 2.4. Let A be unital C ∗-algebra. Denote by Inv0(A) the connected component of the set of invertible elements which contains the identity of A. Let x, y ∈ A and let λ 6∈ sp(x)∪sp(y). Then λ − x and λ − y are invertible. Denote by [λ − x] the corresponding element in K1(A). So [λ − x] = [λ − y] means that they represent the same element in K1(A). In the case that A is of stable rank one, [λ − x] = [λ − y] is equivalent to (λ − x)−1(λ − y) ∈ Inv0(A). Definition 2.5. Let A be a C ∗-algebra and let a, b ∈ A+ be two positive elements. We write a . b if there is a sequence of elements {xn} ⊂ A such that x∗ nbxn → a as n → ∞. If a . b and b . a, then we write [a] = [b] and say that a and b are equivalent in Cuntz semi-group. If p, q ∈ A are two projections, then p . q means that there is a partial isometry w ∈ A such that w∗w = p and ww∗ ≤ q. Definition 2.6. Let ǫ > 0. Denote by fǫ the continuous function on [0,∞) such that 0 ≤ fǫ ≤ 1; f (t) = 1 if ξ ∈ [ǫ,∞) and f (t) = 0 if ξ ∈ [0, ǫ/2] and f (t) is linear in (ǫ/2, ǫ). Let b ∈ A+ defined dτ (b) = lim ǫ→0 τ (fǫ(b)) for τ ∈ T (A). A is said to have strict comparison for positive elements, if dτ (a) < dτ (b) for all τ ∈ T (A) implies that a . b. In this paper, we mainly study those C ∗-algebras A which have real rank zero, stable rank one and weak unperforated K0(A). Such C ∗-algebras always have strict comparison for positive elements. When A has tracial rank zero (see 3.6.2 of [18]), we write T R(A) = 0. If A is a unital simple C ∗-algebra with T R(A) = 0, then A has real rank zero, stable rank one and weakly unperforated K0(A). Definition 2.7. Let Ω be a compact metric space and let O ⊂ Ω be an open subset. Throughout this paper, fO denotes a positive function with 0 ≤ fO ≤ 1 whose support is exactly O, i.e., fO(t) > 0 for all t ∈ O and fO(t) = 0 for all t 6∈ O. If x, y ∈ A are two normal elements with X = sp(x) and Y = sp(y), we let Ω = X ∪ Y. Let ϕX , ψY : C(Ω) → A be unital homomorphisms defined by ϕX (f ) = f (x) and ψY (f ) = f (y) for all f ∈ C(Ω). Definition 2.8. Let W (A) be the Cuntz semi-group which are equivalence classes of positive elements in M∞(A). Definition 2.9. Let A be a unital C ∗-algebra and Ω be a compact metric space. Denote by Hom1(C(Ω), A) the set of all unital homomorphisms κ from C(Ω) into A. 4 Definition 2.10. Let O ⊂ Ω be an open subset. Given κ ∈ Hom1(C(Ω), A), [κ(fO)] does not depend on the choice of fO. If κ1, κ2 ∈ Hom1(C(Ω), A), define Dc(κ1, κ2) = sup{inf{d > 0 : κ1(fO) . κ2(fOd)} : O ⊂ Ω, open}. (e 2.4) Remark 2.11. The definition of Dc(·,·) is not symmetric as a priori. However, when A is a unital simple C ∗-algebra with stale rank one, then the definition in (e 2.4) is in fact symmetric. Moreover, in general, W (C(Ω)) is not determined by open subsets of Ω. In the definition it would required that κ1([fO]) . κ2([fOd]), for all f ∈ M∞(C(Ω))+ whose supports in O and all fOd ∈ M∞(C(Ω))+ with supports in Od. We will not study these in full generality. Definition 2.12. Let A be a unital C ∗-algebra and let Ω be a compact metric space. For κ1, κ2 ∈ Hom1(C(Ω), A) we write κ1 ∼ κ2 if [κ1(fO)] = [κ2(fO)] in W (A) (e 2.5) for all open subsets O ⊂ Ω. It is easy to see that "∼" is an equivalence relation. Put Hc,1(C(Ω), A) = Hom1(C(Ω), A)+/ ∼ . It follows from [26] that if κ1 ∼ κ2 then they induce the same semi-group homomorphisms from W (C(Ω)) into W (A) if the covering dimension of Ω is at most two and the second co Homology subsets H 2(X) = {0} for each compact subsets X. This is particular true when Ω is a compact subset of the plane which is the primary concern of this research. Let ϕ ∈ Hom1(C(Ω), A). Then kerϕ = {f ∈ C(Ω) : fX = 0} for some compact subset X ⊂ Ω. The compact subset X is called the spectrum of ϕ. Sometimes, we denote ϕ ∈ Hom1(C(Ω), A) with spectrum X by ϕX . Definition 2.13. Let A be a unital C ∗-algebra and let X and Y be two compact subsets of a compact metric space Ω. Suppose that ϕ : C(X) → A and ϕ : C(Y ) → A are two unital monomorphisms. We define ϕX : C(Ω) → A by ϕ ◦ πX and ϕY : C(Y ) → A by ϕ ◦ πY , where πX and πY are the projections onto X and Y , respectively. Let a, d ∈ A+ with a, d ≤ 1. In the following as well as in the proof of 2.15 below, we will write a << d. if there are b, c ∈ A+ with 0 ≤ b, c ≤ 1 such that ab = a, b . c and dc = c. (e 2.6) The following follows immediately from [27] (see also [25], and 4.2 and 4.3 of [28]). Lemma 2.14. Suppose that A is a unital C ∗-algebra with stable rank one, 0 ≤ a, d ≤ 1 are elements in A such that a << d. (e 2.7) Then 1 − d . 1 − a. Proof. There are b, c ∈ A such that 0 ≤ b, c ≤ 1, ab = a, b . c and cd = c. Since ab = a, for any 1/2 > ǫ > 0, afǫ(b) = a. Since b . c, by Proposition 2.4 of [27], there is a unitary u ∈ A such that u∗fǫ(b)u ∈ Her(c). It follows that u∗au = d1/2u∗aud1/2 ≤ d. Therefore 1 − d ≤ 1 − u∗au = u∗(1 − a)u. It follows that 1 − d . 1 − a. 5 (e 2.8) (e 2.9) Proposition 2.15. Let A be a unital simple C ∗-algebra with stable rank one and let Ω be a com- pact metric space. Then (Hc,1(C(Ω), A), Dc) is a metric space. That is: for any ϕX , ϕY , ϕZ ∈ Hc,1(C(Ω), A), Dc(ϕX , ϕY ) = 0 ⇐⇒ ϕX ∼ ϕY , Dc(ϕX , ϕY ) = Dc(ϕY , ϕX ), Dc(ϕX , ϕZ ) ≤ Dc(ϕX , ϕY ) + Dc(ϕY , ϕZ ). (e 2.10) (e 2.11) (e 2.12) Proof. Let d = Dc(ϕX , ϕY ). We will show that Dc(ϕY , ϕX ) = d. Suppose O ⊂ Ω is an open subset. For any d > ǫ > 0, let F = {t : dist(t, O) ≥ d + ǫ} and K = Fd+ǫ. Define f, g ∈ C(Ω) as the following: f (t) = 0 if t 6∈ Fǫ/8, 0 < f (t) < 1 if t ∈ Fǫ/8 \ F and f (t) = 1 if t ∈ F and g(t) = 0 if t 6∈ K, 0 < g(t) < 1 if t ∈ K \ Fd+ǫ/2 and g(t) = 1, if t ∈ Fd+ǫ/2. (e 2.13) (e 2.14) Since Fǫ/8 ⊂ F ǫ/8 ⊂ Fǫ/4 ⊂ (Fǫ/4)d+ǫ/16 ⊂ Fd+5ǫ/16 ⊂ Fd+5ǫ/16 ⊂ Fd+ǫ/2 ⊂ Fd+ǫ/2 ⊂ K (e 2.15) (e 2.16) and ϕX (fFǫ/4) . ϕY (f(Fǫ/4)d+ǫ/16), we have By 2.14, Note that ϕX (f ) << ϕY (g). 1 − ϕY (g) . 1 − ϕX (f ). O ⊂ {t : dist(t, F ) ≥ d + ǫ} = {t : dist(t, F ) < d + ǫ}c ⊂ K c. It follows that, for any t ∈ O, g(t) = 0, which implies fO ≤ 1 − g. Hence ϕY (fO) . 1 − ϕY (g). (e 2.17) (e 2.18) (e 2.19) (e 2.20) On the other hand, if f (t) 6= 1 or t 6∈ F , then dist(t, O) < d + ǫ, or t ∈ Od+ǫ. Therefore Supp(1 − f ) ⊂ Od+ǫ, so we may assume that 1 − f ≤ fOd+ǫ by choosing a right representation of fOd+ǫ. It follows that Combining (e 2.20),(e 2.18) and (e 2.22), we obtain 1 − ϕX(f ) . ϕX(fOd+ǫ). 6 (e 2.21) (e 2.22) ϕY (fO) . ϕX(fOd+ǫ) for all ǫ > 0 and for all open subsets O ⊂ Ω. This implies that Dc(ϕY , ϕX ) ≤ d. (e 2.23) (e 2.24) By symmetry, this proves that (e 2.11) holds. Suppose that Dc(ϕX , ϕY ) = 0. For any non-empty open subset O, any r > 0, recall that O−r = {t : dist(t, Oc) > r}. Then there is δ > 0 such that for any r ∈ (0, δ), O−r 6= ∅. It is easy to show, for any S ⊂ Ω, (S−r)r ⊂ S. For any ǫ ∈ (0, δ), then ϕX (fO−ǫ) . ϕY (f(O−ǫ)ǫ) . ϕY (fO). This shows that, for any σ ∈ (0, δ), fσ(ϕX (fO)) . ϕY (fO). (e 2.25) It follows from 2.4 of [27] that ϕX (fO) . ϕY (fO). Similarly, by (e 2.11), we have Dc(ϕY , ϕX ) = 0, so ϕY (fO) . ϕX (fO), so [ϕX (O)] = [ϕY (O)] and (e 2.10) holds. Finally, suppose that Dc(ϕX , ϕY ) = d1, Dc(ϕY , ϕZ ) = d2, Dc(ϕX , ϕZ ) = d3. Then for any open O and any ǫ > 0, ϕX (fO) . ϕY (fOd1+ǫ) . ϕZ (fOd1+d2+2ǫ). Therefore and (e 2.12) holds. d3 ≤ d1 + d2 Proposition 2.16. Let A be a simple unital C ∗-algebra with stable rank one, and let Ω be a compact metric space. Then, for any finite subset of projections P ⊂ C(Ω), there exists a δ > 0 satisfying the following: if ϕ, ψ : C(Ω) → A are two unital homomorphisms such that then Moreover, Dc(ϕ, ψ) < δ, [ϕ(p)] = [ψ(p)] in W (A) for all p ∈ P. [ϕ(p)] = [ψ(p)] in K0(A) for all p ∈ P. 7 (e 2.26) (e 2.27) (e 2.28) Proof. Without loss of generality, we may assume that P consists of mutually orthogonal non- zero projections. There are mutually disjoint clopen subsets {Ei : i = 1, 2..., m} such that P = {pi = χEi, i = 1, 2, ..., m}. Let d = min 1≤i≤m{dist(Ei, X\Ei)}. Now choose 0 < δ < d. Note that, for any d > r > 0, (Ei)r = Ei, i = 1, 2, ..., m. If Dc(ϕ, ψ) < δ, then for any i and d > r > δ, which implies By symmetry, we have ϕ(fEi) . ψ(f(Ei)r ) = ψ(fEi) [ϕ(pi)] . [ψ(pi)]. [ψ(pi)] . [ϕ(pi)]. Thus we get [ϕ(pi)] = [ψ(pi)], in W (A) i = 1, 2, ..., m. Lemma 2.17. Let A be a unital simple C ∗-algebra of (stable rank one) and let Ω be a compact metric space. For any ǫ > 0, there exists δ > 0 and a finite subset F ⊂ C(Ω) satisfying the following: Suppose that ϕ, ψ, ρ : C(Ω) → A are three unital homomorphisms such that then kϕ(f ) − ψ(f )k < δ for all f ∈ F, Dc(ϕ, ρ) − Dc(ψ, ρ) < ǫ (e 2.29) (e 2.30) Proof. Let d = Dc(ϕ, ρ) ≥ 0. Let ǫ > 0 be given. any open subset G ⊂ Ω, there is an integer i, Since Ω is compact, there are only finitely many open subsets {O1, O2, ..., On} such that, for G ⊂ Oi ⊂ Gǫ. (e 2.31) Let gi = f(Oi)ǫ, i = 1, 2, ..., n. It follows from [27] that there is δ > 0 satisfying the following: If h1, h2 : C(Ω) → A are two unital homomorphisms such that kh1(gi) − h2(gi)k < δ, i = 1, 2, ..., n, fǫ(h1(gi)) . h2(gi) (e 2.32) (e 2.33) then i = 1, 2, ..., n. Choose this δ. Let F = {gi : i = 1, 2, ..., n}. Let G ⊂ Ω be an open subset. There is an integer i such that If G ⊂ Oi ⊂ Gǫ. kϕ(gi) − ψ(gi)k < δ, i = 1, 2, ..., n, 8 (e 2.34) (e 2.35) then fǫ(ψ(gi)) . ϕ(gi). (e 2.36) (e 2.37) Since the support of fǫ(gi) contains Oi, G ⊂ Oi and Oi ⊂ Gǫ, we obtain that ψ(fG) . ψ(fǫ(gi)) . ϕ(gi). ϕ(fG2ǫ) . ρ(fGd+3ǫ), (e 2.38) Since this holds for all open sets G ⊂ Ω, we conclude that Dc(ψ, ρ) ≤ d + 3ǫ = Dc(ϕ, ρ) + 3ǫ. By symmetry, Lemma follows. Dc(ϕ, ρ) ≤ Dc(ψ, ρ) + 3ǫ. (e 2.39) (e 2.40) Definition 2.18. Let A be a unital C ∗-algebra and let Ω be a compact metric space. Fix κ ∈ Hc,1(C(Ω), A) and ǫ > 0. Let X be the spectrum of κ. We say κ has a finite ǫ-approximation in Hc,1(C(Ω), A), if there is a finite subset F ⊂ X and ϕF ∈ Hom1(C(Ω), A)+ with the spectrum F such that Dc(κ, ϕF ) < ǫ. (e 2.41) Note that [ϕF (fO)] can be represented by a projection p ∈ A. Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and with weakly unperforated K0(A). We show that (see 4.12 below), if κ ∈ Hc,1(C(Ω), A) is induced by a homomorphism h : C(Ω) → A, then κ has a finite ǫ-approximation for any ǫ > 0. Warning: a homomorphism ϕ has a finite ǫ-approximation in Hc,1(C(Ω), A) does not imply that it is close to a homomorphisms with finite spectrum. Definition 2.19. Let A be a unital C ∗-algebra with T (A) 6= ∅. Let Ω be a compact metric space, and X, Y ⊂ Ω be two compact subsets. Let ϕX : C(X) → A and ψY : C(Y ) → A be two unital monomorphisms. Let πX : C(Ω) → C(X) and πY : C(Ω) → C(Y ) be the quotient maps. Define ϕX = ϕ ◦ πX and ψY = ψ ◦ πY . For each open subset O ⊂ Ω, define rO(ϕX , ψY ) = inf{r > 0 : dτ (ϕX (fO)) ≤ dτ (ψY (fOr )) for all τ ∈ T (A)} and r+ O(ϕX , ψY ) = inf{r > 0 : dτ (ϕX (fO)) < dτ (ψY (fOr )) for all τ ∈ T (A)}. (e 2.42) (e 2.43) Define Put DT (ϕX , ψY ) = sup{rO(ϕX , ψY ) : O open}. (e 2.44) a(ϕX , ψY ) = sup{dist(ζ, X) : ζ ∈ Y } and b(ϕX , ψY ) = sup{dist(ξ, Y ) : ξ ∈ X}. Define DT (ϕX , ψY ) = max{a(ϕX , ψY ), sup{r+ O(ϕX , ψY ) : O open and O ∩ X 6= X}}. (e 2.45) 9 Note that, if X ⊂ O, then dτ (ϕX (fO)) = 1 for all τ ∈ T (A). Therefore DT (ϕX , ψY ) ≥ a(ϕX , ψY ). (e 2.46) Since X is compact, there is ξ ∈ X such that b(ϕX , ψY ) = dist(ξ, Y ). If ǫ > 0 and O(ξ, ǫ) is the open ball with center at ξ and radius ǫ, then rO(ξ,ǫ)(ϕX , ψY ) ≥ b(ϕX , ψY ) − ǫ. (e 2.47) It follows that DT (ϕX , ψY ) ≥ max{a(ϕX , ψY ), b(ϕX , ψY )} = dH (X, Y ), where dH(X, Y ) is the Hausdorff distance between X and Y . Lemma 2.20. Let A be a unital simple C ∗-algebra with T (A) 6= ∅ and let O ⊂ Ω be an open set with O ∩ X 6= X. If r > DT (ϕX , ψY ), then inf{dτ (ψY (fOr )) − dτ (ϕX (fO)) : τ ∈ T (A)} > 0. (e 2.48) Proof. Put d = DT (ϕX , ψY ) and η = (1/4)(r − d) > 0. Let f1 ∈ C(Ω) with 0 ≤ f1 ≤ 1 such that f1(t) = 1 if t ∈ O and f1(t) = 0 if t 6∈ Oη. Let f2 ∈ C(Ω) with 0 ≤ f2 ≤ 1 such that f2(t) = 1 if t ∈ Od+2η and f2(t) = 0 if t 6∈ Or. Then dτ (ψY (fOr )) ≥ τ (ψY (f2)) ≥ dτ (ψY (fOd+2η)) > dτ (ϕX (fOη )) ≥ τ (ϕX (f1)) ≥ dτ (ϕX (fO)) for all τ ∈ T (A). Since T (A) is compact, we conclude that inf{τ (ψY (f2)) − τ (ϕX (f1)) : τ ∈ T (A)} > 0. Therefore inf{dτ (ψY (fOr )) − dτ (ϕX (fO)) : τ ∈ T (A)} ≥ inf{τ (ψY (f2)) − τ (ϕX (f1)) : τ ∈ T (A)} > 0. (e 2.49) (e 2.50) (e 2.51) (e 2.52) (e 2.53) Proposition 2.21. Let A be a unital C ∗-algebra with T (A) 6= ∅. If A has the strict comparison for positive elements. Let ϕX , ϕY , ϕZ be three unital homomorphisms from C(Ω) into A. Then DT (ϕX , ϕY ) ≤ Dc(ϕX , ϕY ) ≤ DT (ϕX , ϕY ), DT (ϕX , ϕY ) = DT (ϕY , ϕX ), DT (ϕX , ϕY ) = DT (ϕY , ϕX ), DT (ϕX , ϕZ ) ≤ DT (ϕX , ϕY ) + DT (ϕY , ϕZ ), DT (ϕX , ϕZ ) ≤ DT (ϕX , ϕY ) + DT (ϕY , ϕZ ). If X or Y is connected, then DT (ϕX , ϕY ) = Dc(ϕX , ϕY ) = DT (ϕX , ϕY ). Proof. If Dc(ϕX , ϕY ) < d, then for any open O, So for any τ ∈ T (A), ϕX (fO) . ϕY (fOd). dτ (ϕX (fO)) ≤ dτ (ϕY (fOd)). 10 (e 2.54) (e 2.55) (e 2.56) (e 2.57) (e 2.58) (e 2.59) This implies the first inequality of (e 2.54). If DT (ϕX , ϕY ) < d, then for any open subset O with O ∩ X 6= X, by 2.20, inf{dτ (ϕY (fOd)) − dτ (ϕX (fO)) : τ ∈ T (A)} > 0. (e 2.60) Since A has strict comparison for positive elements, we have ϕX (fO) . ϕY (fOd). If O ∩ X = X, then X ⊂ O. Since d > a(ϕX , ψY ), we have that Y ⊂ Xd ⊂ Od. It follows ϕX (fO) . ϕY (fOd). Thus we get the second inequality of (e 2.54). To show (e 2.55), let DT (ϕX , ϕY ) ≤ d. It suffices to show that DT (ϕY , ϕX ) ≤ d. Suppose O ⊂ Ω is an open subset. For any ǫ > 0, let F = {t : dist(t, O) ≥ d + ǫ} and K = Fd+ǫ. Define f, g ∈ C(Ω) with f (t) = 0 if t 6∈ Fǫ/2, 0 < f (t) < 1, if t ∈ Fǫ/2 \ Fǫ/4, f (t) = 1 if t ∈ Fǫ/4 and g(t) = 0 if t 6∈ K,0 < g(t) < 1, if t ∈ K \ Fd+ǫ/2, g(t) = 1 if t ∈ Fd+ǫ/2. So, if 1 − f (t) 6= 0, then t 6∈ F. Hence t ∈ Od+ǫ. Therefore, by choosing a right fOd+ǫ, we may assume 1 − f ≤ fOd+ǫ. Then (Fǫ/2)d+ǫ/2 ⊂ Fd+ǫ = K. (e 2.61) By the definition, dτ (ϕX (f )) = dτ (ϕX (fFǫ/2)) ≤ dτ (ϕY (f(Fǫ/2)d+ǫ/2)) ≤ dτ (ϕY (fK)) = dτ (ϕY (g)). (e 2.62) Thus Since fO ≤ 1 − g, we have Hence Thus we have 1 − dτ (ϕY (g)) ≤ 1 − dτ (ϕX (f )). ϕY (fO) . 1 − ϕY (g). 1 − dτ (ϕX (f )) ≤ dτ (ϕX (fOd+ǫ)). dτ (ϕY (fO)) ≤ dτ (ϕX (fOd+ǫ)) (e 2.63) (e 2.64) (e 2.65) (e 2.66) for all ǫ > 0 and for all open subsets O ⊂ Ω. Thus we get DT (y, x) ≤ d. To show (e 2.56), let d > DT (ϕX , ϕY ) and let O ⊂ Ω. If O ∩ Y 6= Y, we have dτ (ϕX (fFǫ/2)) < dτ (ϕY (f(Fǫ/2)d+ǫ/2)). (e 2.67) We will follow the proof of (e 2.55). By (e 2.67), instead of " ≤ " we will have " < " in (e 2.62). It follows, as in the proof of (e 2.55), sup{r+ O(ϕY , ϕX ) : O open and O ∩ Y 6= Y } ≤ d. 11 Since a(ϕY , ϕX ) ≤ DT (ϕY , ϕX ) = DT (ϕX , ϕY ) < d, then DT (ϕY , ϕX ) = max{a(ϕY , ϕX ), sup{r+ O(ϕY , ϕX ) : O open and O ∩ Y 6= Y }} ≤ d. (e 2.68) We get (e 2.56). Now we turn to (e 2.57). c4 = DT (ϕX , ϕY ), then for any ǫ > 0, any open set O, If c1 = DT (ϕX , ϕY ), c2 = DT (ϕY , ϕZ ), c3 = DT (ϕX , ϕY ) and and If dτ (ϕX (fO)) ≤ dτ (ϕY (fOc1+ǫ)) ≤ dτ (ϕZ (fOc1+c2+2ǫ)) dτ (ϕX (fO)) < dτ (ϕY (fOc3+ǫ)) < dτ (ϕZ (fOc3+c4+2ǫ)) a(ϕX , ϕY ) = d1, a(ϕY , ϕZ ) = d2, then Z ⊂ Yd2 ⊂ (Xd1 )d2 ⊂ Xd1+d2, so a(ϕX , ϕZ ) = inf{r > 0 : Z ⊂ Xr} ≤ d1 + d2 = a(ϕX , ϕY ) + a(ϕY , ϕZ ). From these, we obtain (e 2.57) and (e 2.58). To show (e 2.59), assume X is connected and DT (ϕX , ϕY ) = d. It is suffices to show DT (ϕX , ϕY ) ≤ d. For any open O with O ∩ X 6= X, since X is connected, there is δ > 0 such that for any 0 < ǫ < δ, O ∩ X 6= Oǫ/2 ∩ X. So, since A is simple, for any τ ∈ T (A), τ (ϕX (fO)) < τ (ϕX (fOǫ/2)) ≤ τ (ϕY (fOd+ǫ)). (e 2.69) On the other hand, since a(ϕX , ϕY ) ≤ DT (ϕX , ϕY ), by definition, DT (ϕX , ϕY ) ≤ d. This ends the proof of (e 2.59). Note there exists C ∗-algebra A, DT (ϕ, ψ) = Dc(ϕ, ψ), even when neither X nor Y are connected. Proposition 2.22. Let A be a unital simple C ∗-algebra with stable rank one, kerρA(K0(A)) = {0} and with strict comparison for positive elements. Suppose that A has a unique tracial state. Then DT (ϕ, ψ) = Dc(ϕ, ψ). Proof. Let ϕ, ψ : C(Ω) → A be two unital homomorphisms with spectrum X and Y, respectively, and let τ be the unique tracial state of A. By (e 2.54) of 2.21, it suffices to show that DT (ϕ, ψ) ≥ Dc(ϕ, ψ). Let d = DT (ϕ, ψ) and d1 > d. For any open subset O ⊂ Ω, If inequality holds in (e 2.70), then by the strict comparison, dτ (ϕ(fO)) ≤ dτ (ψ(fOd1 )). Otherwise, suppose that equality holds in (e 2.70). If, for every 1/2 > ǫ > 0, ϕ(fO) . ψ(fOd1 ). (e 2.70) (e 2.71) dτ (fǫ(ϕ(fO))) = dτ (ϕ(fǫ(fO))) < dτ (ϕ(fO)) = dτ (ψ(fOd1 )), (e 2.72) by the strict comparison again, ϕ(fǫ(fO)) . ψ(fOd1 ) for all 1/2 > ǫ > 0. (e 2.73) 12 It follows that ϕ(fO) . ψ(fOd1 ). Otherwise there is an 1/2 > ǫ > 0 such that Since A is simple, we conclude that, for δ > 0, dτ (fǫ(ϕ(fO))) = dτ (ϕ(fO)). O ∩ X = O−δ∩X = {ξ ∈ X : dist(x, Oc) > δ}, (e 2.74) (e 2.75) which implies that O∩ X is a clopen set relative to X. Let q = ϕ(fO) which is then a projection in this case. We also have, for any d < d2 < d1, dτ (ϕ(fO)) = dτ (ϕZ (fOd2 )) = dτ (ϕY (fOd1 )). (e 2.76) The same argument above shows that Od2 ∩ Y = Od1 ∩ Y. It follows that p = ψ(fOd1 projection. Since kerρA(K0(A)) = {0}, and τ (p) = τ (q), ϕ(fO) = q ∼ p = ψ(fOd1 ). ) is a (e 2.77) It follows that Dc(ϕ, ψ) ≤ d1 for all d1 > d. This completes the proof. Definition 2.23. For ϕ, ψ ∈ Hom(C(Ω), A), let δT (ϕ, ψ) = inf{r > 0 : dτ (ϕ(fO)) ≤ dτ (ψ(fOr )), for all O = O(λ, d) , τ ∈ T (A)}, δc(x, y) = inf{r > 0 : ϕ(fO).ψ(fOr ), for all O = O(λ, d) ,} and dT (ϕ, ψ) = max{δT (ϕ, ψ), δT (ψ, ϕ)}, dc(ϕ, ψ) = max{δc(ϕ, ψ), δc(ψ, ϕ)}. For any normal elements x, y ∈ A, if X = sp(x), Y = sp(y), Ω = X ∪ Y and ϕX , ϕY are defined by ϕX (f ) = f (x), ϕY (f ) = f (y) for any f ∈ C(Ω), define dT (x, y) = dT (ϕX , ϕY ), dc(x, y) = dc(ϕX , ϕY ). Proposition 2.24. Let Ω be a compact metric space and A be a unital simple C ∗-algebra. Suppose that ϕX , ϕY , ϕZ : C(Ω) → A are unital homomorphisms with spectrum X, Y, Z ⊂ Ω, respectively. Then dT (ϕX , ϕY ) = dT (ϕY , ϕX ), dc(ϕX , ϕY ) = dc(ϕY , ϕX ) dT (ϕX , ϕY ) ≥ dH(X, Y ), dT (ϕX , ϕY ) ≤ dc(ϕX , ϕY ) ≤ Dc(ϕX , ϕY ), dT (ϕX , ϕY ) ≤ dT (ϕX , ϕZ ) + dT (ϕZ , ϕY ), dc(ϕX , ϕY ) ≤ dc(ϕX , ϕZ ) + dc(ϕZ , ϕY ). (e 2.78) (e 2.79) (e 2.80) (e 2.81) (e 2.82) Proof. The identities in (e 2.78) follows from the definition. The inequality in (e 2.79) also follows from the definition immediately since A is assumed to be simple. If dc(ϕX , ψY ) = r, then for any ξ ∈ Ω, any d > 0, any ǫ > 0, ϕX (fO(ξ,d)) . ϕY (fO(ξ,d+r+ǫ)). It follows that, for any τ ∈ T (A), dτ (ϕX (fO(ξ,d))) ≤ dτ (ϕY (fO(ξ,d+r+ǫ))). 13 This implies dT (ϕX , ψY ) ≤ dc(ϕX , ψY ). It is obvious that dc(ϕX , ψY ) ≤ Dc(ϕX , ψY ). So (e 2.80) holds. The proofs of (e 2.81) and (e 2.82) are similar, we show (e 2.81) only. If dT (ϕX , ϕZ ) < d1, dT (ϕZ , ϕY ) < d2, then for any ξ ∈ C, any d > 0, any ǫ > 0, dT (ϕX (fO(ξ,d))) ≤ dT (ρZ (fO(ξ,d+d1+ǫ))) ≤ dT (ψY (fO(ξ,d+d1+d2+2ǫ))), therefore dT (ϕX , ψY ) ≤ d1 + d2, this implies (e 2.81) holds. Remark 2.25. There are examples such that dT (x, y) = Dc(x, y). Let n ≥ 4 be an integer. Let X = {ekπi/n : 0 ≤ k ≤ n} and Y = rX = {rekπi/n : 0 ≤ k ≤ n} for some 0 < r < 1. Let A be any unital simple C ∗-algebra with T (A) 6= ∅ which has n mutually orthogonal non-zero projections {e1, e2, ..., en} such that Pn k=1 ei = 1. Define x = Pn k=1 re(k−1)πi/nei. Then one computes that k=1 e(k−1)πi/nei and y = Pn DT (x, y) = 1 − r = dT (x, y) = dc(x, y) = Dc(x, y). (e 2.83) On the other hand, of course, there are also examples that dT (x, y) < DT (x, y). Let {e1, e2, e3} be mutually orthogonal and equivalent projections with e1 + e2 + e3 = 1, Then dT (x, y) = dc(x, y) = 1 < √2 = DT (x, y) = Dc(x, y). x = −e1 + e3, y = −ie1 + ie3. In the case that A has stable rank one, strict comparison and kerρA = {0}, then DT (·,·) is a distance on Hc,1(C(Ω), A). In general, because of the definition of Hc,1(C(Ω), A), DT , dT , dc are not a distance on Hc,1(C(Ω), A). 3 Distance between unitary orbits of normal elements with triv- ial K1 Let Zk be the direct sum of n copies of the abelian group Z. Put Zk + = {(n1, n2, ..., nk) : nj ≥ 0 : j = 1, 2, ..., k}. (e 3.84) It is well known that (Zk, Zk +) is an unperforated (partially) ordered group with the Riesz interpolation property. Let R ⊂ {1, 2, ..., m} × {1, 2, ..., n} be a subset and let A ⊂ {1, 2, ..., m}. Define RA ⊂ {1, 2, ..., n} to be the subset of those j′s such that (i, j) ∈ R, for some i ∈ A. Lemma 3.1. If {ai}m satisfying: for any A ⊂ {1, ..., m}, j=1 bj, and R ⊂ {1, ..., m} × {1, ..., n} i=1 ai = Pn + with Pm i=1,{bi}n j=1 ⊂ Zk X i∈A ai ≤ X j∈RA bj, (e 3.85) then there are {cij} ⊂ Zk + such that n X j=1 cij = ai, m X i=1 cij = bj, for all i, j (e 3.86) and cij = 0 unless (i, j) ∈ R. (e 3.87) 14 Proof. Write ai = (ai(1), ai(2), ..., ai(k)) and bj = (bj(1), bj (2), ..., bj (k)), (e 3.88) i = 1, 2, ..., m and j = 1, 2, ..., n. It follows from Lemma 1.2 of [12] that, for each s (s = 1, 2, ..., k), there are ci,j(s) ∈ Z+ such that n X j=1 cij(s) = ai(s), m X i=1 cij(s) = bj(s), for all i, j. (e 3.89) and Define Note that We also have cij(s) = 0 unless (i, j) ∈ R. (e 3.90) cij = (cij(1), cij (2), ..., cij (k)), i = 1, 2, ..., m and j = 1, 2, ..., n (e 3.91) cij = 0 unless (i, j) ∈ R. m X i=1 cij = ai and n X j=1 cij = bj. (e 3.92) (e 3.93) Lemma 3.2. Let (G, G+) be a countable torsion free unperforated partially ordered abelian group with the Riesz interpolation property. If {ai}m j=1 bj, and R ⊂ {1, ..., m} × {1, ..., n} satisfying: for any A ⊂ {1, ..., m}, j=1 ⊂ G+ with Pm i=1,{bi}n i=1 ai = Pn X i∈A ai ≤ X j∈RA bj, (e 3.94) then there are {cij} ⊂ G+ such that n X j=1 cij = ai, m X i=1 cij = bj, for all i, j (e 3.95) and cij = 0 unless (i, j) ∈ R. (e 3.96) Proof. Let G be a countable unperforated ordered Riesz group. It follows from [8] that G = limn→∞(Gn, hn) with Gn is order isomorphic to Zr(n) (with positive cone Zr(n) + ), where hn : Gn → Gn+1. Denote by hn,n+k = hn+k ◦ hn+(k−1) ◦ ··· ◦ hn : Gn → Gn+k, k = 1, 2, ..., n = 1, 2, ..., and denote by hn,∞ : Gn → G the homomorphism induced by the inductive limit system. Moreover, G+ = ∪∞ n=1hn,∞((Gn)+). For each A⊂{1, 2, ..., m}, denote by gA = X j∈RA bj − X i∈A ai 15 (e 3.97) Note that there are no more than 2m many A′s. There exists n1 > 0 such that there are ai,k, bj,k ∈ (Gk)+ and gA,k ∈ (Gk)+ for all i, j and A ⊂ {1, 2, ..., m} (e 3.98) (e 3.99) such that, for k1 > k ≥ n1, hk,k1(ai,k) = ai,k1, hk,k1(bj,k) = bj,k1, hk,k1(gA,k) = gA,k1, hk,∞(ai,k) = ai, hk,∞(bj,k) = bj and hk,∞(g′ A) = gA. Note, since each Gk is isomorphic to Zr(n), there is an integer n2 > n1 such that m X i=1 hn1,n2(ai,n1) − n X j=1 hn1,n2(bj,n1) = 0 and gA,n2 = X j∈RA hn1,n2(bj,n1) − X i∈A hn1,n2(ai,n1). (e 3.100) (e 3.101) Thus, we obtain, by applying 3.1, ci,j,n2 ∈ (Gn2 )+, (i, j) ∈ R such that n X j=1 ci,j,n2 = ai,n2 and m X i=1 ci,j,n2 = bj,n2. (e 3.102) Moreover, Define ci,j = hn2,∞(ci,j,n2). Then, ci,j ≥ 0 and by (e 3.102) and (e 3.103), ci,j,n2 = 0 unless (i, j) ∈ R. n X j=1 ci,j = ai, m X i=1 ci,j = bj and ci,j = 0 unless (i, j) ∈ R. (e 3.103) (e 3.104) (e 3.105) Lemma 3.3. Let (G, G+) be a countable weakly unperforated partially ordered abelian group with the Riesz interpolation property. If {ai}m i=1,{bi}n j=1 bj, R ⊂ {1, ..., m} × {1, ..., n} satisfying: for any A ⊂ {1, ..., m}, bj, j=1 ⊂ G+ with Pm i=1 ai = Pn (e 3.106) X i∈A ai ≤ X j∈RA then there are {cij} ⊂ G+ such that n X j=1 cij = ai, m X i=1 and cij = bj, for all i, j (e 3.107) cij = 0 unless (i, j) ∈ R. (e 3.108) Proof. It follows from [9] that one may write 0 → T → G → G0 → 0, where G0 is a countable unperforated ordered group with the Riesz interpolation property and T is a countable abelian torsion group. Moreover, g ∈ G+ \ {0} if and only if d(g) ∈ (G0)+, 16 where d : G → G0 is the quotient map. Furthermore, there exists a sequence of abelian groups Sn and Tn such that Sn is order isomorphic to Zr(n) and Tn = Z/k(1, n)Z ⊕ Z/k(2, n)Z ⊕ ··· ⊕ Z/k(t(n), n)Z such that G = limn→∞(Sn ⊕ Tn, ın), where ın : Sn ⊕ Tn → Sn+1 ⊕ Tn+1. Denote by ın,∞ : Sn ⊕ Tn → G. Note (Sn ⊕ Tn)+ = {(s, f ) : s ∈ (Sn)+ \ {0}} ∪ {(0, 0)}, n = 1, 2, ..., and G+ = lim∞(Sn ⊕ Tn)+. Let π′ n : Sn ⊕ Tn → Tn be the projection maps. Let ı′ n ◦ ınSn. Let Fn = Zt(n) and πn : Fn → Tn be the quotient map. Define Hn = Sn ⊕ Fn. Since Fn is free, there is a homomorphism jn : Fn → Fn+1 such that n : Sn → Sn+1 be defined by ı′ n : Sn ⊕ Tn → Sn and let π′′ n = π′ commutes. Since Sn is free, there is h′ Fn jn−→ Fn+1 yπn+1 ın−→ Tn+1 yπn Tn n : Sn → Fn+1 such that ın−→ Sn+1 ⊕ Tn+1 Sn yh′ n Fn+1 πn+1−→ yπ′′ n+1 Tn+1 commutes. Define hn : Hn → Hn+1 by n ⊕ h′ hnSn = ı′ n and hnFn = jn, n = 1, 2, .... Define (Hn)+ = {(s, f ) : s ∈ (Sn)+ \ {0}} ∪ {(0, 0)}. Let H = limn→∞(Hn, hn) and let F = limn→∞(Fn, hnFn). Define H+ = ∪∞ n=1hn,∞(Hn)+, where hn,∞ : Hn → H is the homomorphism induced by the inductive limit system. Define d1 : H → H/F. Then it is clear that H/F is order isomorphic to G0. Moreover, if h ∈ H, then h ∈ H+ if and only if d1(h) ∈ (G0)+. Therefore H is also a torsion free weakly unperforated ordered group with Riesz interpolation property. Define qn : Hn → Sn⊕ Tn by qnSn = idSn and qnFn = πn, n = 1, 2, .... One has the following commutative diagram: hn−→ Hn yqn Sn ⊕ Tn Hn+1 yqn+1 ın−→ Sn+1 ⊕ Tn+1 It induces a quotient map Π : H → G. It is an order preserving map. Π(a′ j) = bj, i = 1, 2, ..., m and j = 1, 2, ..., n. Then a′ i, b′ Now let ai, bj ∈ G+, i = 1, 2, ..., m and j = 1, 2, ..., n, as described. Let a′ i) = ai and Π(b′ i, b′ j ∈ H+ and j ∈ S such that X i∈A i ≤ X a′ j∈RA b′ j in H+. Since H satisfies the assumption in 3.2, there are c′ ij ∈ H+ such that m X j=1 ij = a′ c′ i, n X i=1 c′ ij = b′ j and c′ ij = 0 unless (i, j) ∈ R. 17 (e 3.109) (e 3.110) Put cij = Π(c′ ij ), i = 1, 2, ..., m and j = 1, 2, ..., n. Then m X j=1 cij = ai, n X i=1 cij = bj and cij = 0 unless (i, j) ∈ R. (e 3.111) (e 3.112) Lemma 3.4. Let A be a unital separable simple C ∗-algebra with stable rank one and weakly unperforated K0(A) which has the Riesz interpolation property and let Ω be a compact metric space. Suppose that ϕX (f ) = Pm j=1 f (ζj)qj for all f ∈ C(Ω), where {p1, p2, ..., pm} and {q1, q2, ..., qn} are two sets of mutually orthogonal non-zero projections in A such that Pm j=1 qj = 1A and ξi, ζj ∈ Ω. Let d > 0. Then Dc(ϕX , ϕY )≤ d if and only if, for any ǫ > 0, there are projections pi,j, qi,j ∈ A such that i=1 f (ξi)pi and ϕY (f ) = Pn i=1 pi = Pn pi = n X j=1 pi,j, qj = n X i=1 qi,j, [pi,j] = [qi,j] in K0(A) and max{dist(ξi, ζj) : qi,j 6= 0} < d + ǫ. Proof. Suppose d = Dc(ϕX , ϕY ). Let ǫ > 0. Put R = {(i, j) : dist(ξi, ζj) ≤ d + ǫ}. For any A ⊂ {1, ..., m}, put OA = {ξi : i ∈ A} and ORA = {ζj : j ∈ RA}. Then )] = X pii = [ϕX (fOA)] ≤ [ϕY (f(OA)d+ǫ)] = [ϕY (fORA [pi] = hX X j∈RA i∈A i∈A (e 3.113) (e 3.114) (e 3.115) [qj]. It follows from 3.3, there projection rij such that [pi] = n X j=1 [rij], [qj] = m [rij], i = 1, 2, ..., m; j = 1, 2, ..., n, X i=1 where rij = 0 unless (i, j) ∈ R. Then there are {pij} and {qij} with [pij] = [qij] = [rij], satisfying pi = Then The converse is obvious. n pij, qj = X j=1 m X i=1 qij, i = 1, 2, ..., m; j = 1, 2, ..., n. max{dist(ξi, ζj) : qij 6= 0} ≤ d + ǫ. Lemma 3.5. Let A be a unital separable simple C ∗-algebra with stable rank one and weakly unperforated K0(A) which has the Riesz interpolation property and let x, y ∈ A be two normal elements with finite spectrum. Then, dist(U (x),U (y)) ≤ Dc(x, y). (e 3.116) 18 Proof. Let ǫ > 0. Put d = Dc(x, y)+ǫ. We assume x = Pm j=1 are mutually orthogonal projections with Pm and {qj}n 3.4 that there are {pij} and {qij} with [pij] = [qij] = [rij], satisfying i=1 λipi, y = Pn i=1 pi and Pn j=1 µjqj, where {pi}m i=1 j=1 qj = 1. It follows from n pi = pij, qj = X j=1 m X i=1 qij, i = 1, 2, ..., m; j = 1, 2, ..., n. Let u ∈ U (A) with u∗piju = qij, i = 1, 2, ..., m; j = 1, 2, ..., n. ku∗xu − yk = kX i,j (λi − µj)qijk ≤ max{λi − µj : qij 6= 0} ≤ d. Theorem 3.6. Let A be a unital simple separable C ∗-algebra with real rank zero, stable rank one and with weakly unperforated K0(A). Suppose that x and y are two normal elements in A with [λ − x] = 0 and [µ − y] = 0 in K1(A) for all λ 6∈ sp(x) and for all µ 6∈ sp(y). Then dist(U (x),U (y)) ≤ Dc(x, y). (e 3.117) (e 3.118) Proof. Let ǫ > 0. The assumption (e 3.117) implies that λ − x ∈ Inv0(A) for all λ 6∈ sp(x). It follows from [17] that, for any δ > 0, there is a normal element x1 ∈ A with finite spectrum in sp(x) such that It follows from 2.17 that, for sufficiently small δ, we may assume that kx − x1k < min{δ, ǫ/8}. Dc(x, x1) < ǫ/8. (e 3.119) (e 3.120) Exactly the same argument shows that there is normal element with finite spectrum in sp(y) such that It follows that ky − y1k < ǫ/8 and Dc(y, y1) < ǫ/8. Dc(x1, y1) < Dc(x, y) + ǫ/4. (e 3.121) (e 3.122) Since x1 and y1 both have finite spectrum, it follows from 3.5 that there exists a unitary u ∈ A such that ku∗x1u − y1k < Dc(x1, y1)+ǫ/8 < Dc(x, y) + ǫ/4 + ǫ/8 = Dc(x, y) + 3ǫ/8. (e 3.123) It follows that for all ǫ > 0. dist(U (x),U (y)) < ǫ/8 + ku∗x1u − y1k + ǫ/8 < Dc(x, y) + 5ǫ/8 (e 3.124) (e 3.125) 19 Theorem 3.7. Let A be a unital separable AF-algebra and let x, y ∈ A be two normal elements. Then dist(U (x),U (y)) ≤ Dc(x, y). Proof. Fix two normal elements x, y ∈ A. Let ǫ > 0. Let ǫ > η > 0. For any δ > 0 with δ < η/4, there exists a finite dimensional C ∗-subalgebra B ⊂ A with 1B = 1A and two elements x′, y′ ∈ B such that kx − x′k < δ and ky − y′k < δ. (e 3.126) It follows from [16] that, for some sufficiently small δ, there are normal elements x1, y1 ∈ B such that kx′ − x1k < η/4 and ky′ − y1k < η/4. kx − x1k < η/4 + δ and ky − y1k < η/4 + δ. (e 3.127) (e 3.128) Thus So, by 2.17, Dc(x, x1) < ǫ/4, Dc(y, y1) < ǫ/4 and Dc(x1, x2) < Dc(x, y) + ǫ/2. (e 3.129) Now x1 and y1 have finite spectra. We then complete the proof as in 3.6. 4 Maps with finite dimensional ranges Definition 4.1. Let Ω be a compact metric space, X, Y ⊂ Ω be two compact subsets. Let A be a unital C ∗-algebra and let ϕX : C(Ω) → A and ϕY : C(Ω) → A be two homomorphisms with spectrum X and Y, respectively. Let {hn : C(Ω) → A} be a sequence of homomorphisms with fi- nite dimensional ranges, i.e, hn(f ) = Pk(n) i=1 f (ξ(i, n))pn,i for all f ∈ C(Ω), where ξ(i, n) ∈ X ∩ Y and where {pn,1, pn,2, ..., pn,k(n)} is a sequence of finite subsets of mutually orthogonal projec- tions. We assume that, for any ǫ > 0, there is N ≥ 1 such that {ξ(1, n), ξ(2, n), ..., ξ(k(n), n)} is ǫ-dense in X ∩ Y. Denote by en = Pk(n) i=1 pn,i, n = 1, 2, .... Suppose that lim n→∞ sup{τ (en) : τ ∈ T (A)} = 0. Let {un} ⊂ U (A) be a sequence of unitaries, let qn = u∗ nenun, let ϕX,n : C(Ω) → (1 − en)A(1 − en) and let ϕY,n : C(Ω) → (1 − qn)A(1 − qn) be two unital homomorphisms with spectrum in X and Y, respectively. Suppose that (e 4.130) lim n→∞ Dc(ϕX , hn + ϕX,n) = 0 and lim n→∞ Dc(ϕY , Ad un ◦ hn + ϕY,n) = 0. (e 4.131) Define ϕ′ by Dc(ϕX,n, ϕ′ Y,n(f ) = unϕY,n(f )u∗ Now we defined n for all f ∈ C(Ω). Then ϕ′ Y,n : C(Ω) → (1 − en)A(1 − en). Denote Y,n) the distance defined in 2.10 for A = (1 − en)A(1 − en). De c (ϕX , ϕY ) = inf{lim inf n→∞ Dc(ϕX,n, ϕ′ Y,n)}, where the infimum is taken among all possible such non-zero hn, ϕX,n, un and ϕY,n. It is important to note that De c(ϕX , Ad u ◦ ϕX ) = 0 20 (e 4.132) (e 4.133) for any compact subset X ⊂ Ω and any unitary u ∈ A. Note since n ◦ (Ad un ◦ hn + ϕY,n) = hn + Ad u∗ Adu∗ n ◦ ϕY,n, when we consider De by ϕY,n and Ad u∗ n ◦ ϕY,n respectively, and write c(ϕX , ϕY ) for a pair ϕX and ϕY , we may always replace ϕY by Ad u∗ n ◦ ϕY lim n→∞ Dc(ϕX , hn + ϕX,n) = 0 and lim n→∞ Dc(ϕY , hn + ϕY,n) = 0 (e 4.134) This will be used throughout the paper without further notice. If no such non-zero maps {hn} exists, we define De c (ϕX , ϕY ) = Dc(ϕX , ϕY ). In particular, if X ∩ Y = ∅, De c (ϕX , ϕY ) = Dc(ϕX , ϕY ). When X ∩ Y 6= ∅ and A is a unital separable simple C ∗-algebra of real rank zero, stable rank one and weakly unperforated K0(A), non-zero {hn} can always be found (see 4.14 below). In general, From the definition, by 2.15, We also note that Dc(ϕX , ϕY ) ≤ De c (ϕX , ϕY ). De c(ϕX , ϕY ) = De c(ϕY , ϕX ). De c (ϕX , ϕY ) = inf{lim sup n→∞ Dc(ϕX,n, ϕ′ Y,n)}. To see this, take a subsequence {nk} such that lim k→∞ Dc(ϕX,nk , ϕ′ Y,nk ) = lim sup n→∞ Dc(ϕX,nk , ϕ′ Y,nk ). Then (e 4.135) (e 4.136) (e 4.137) lim inf k→∞ Thus, by the definition of De such that Dc(ϕX,nk , ϕ′ Y,nk ) = lim k→∞ Dc(ϕX,nk , ϕ′ Y,nk ). c (·, ·), (e 4.137) holds. Furthermore, there exists a subsequence {nk} lim k→∞ Dc(ϕX,nk , ϕY,nk ) = De c (ϕX , ϕY ). To see this, for each k, there exists such sequence {hn,k}, ϕX,n,k and ϕY,n,k such that lim inf n→∞ Dc(ϕX,n,k, ϕY,n,k) ≤ De c(ϕX , ϕY ) + 1/k. Choose {nk} such that Then Dc(ϕX,nk,k, ϕY,nk,k) ≤ De c (ϕX , ϕY ) + 1/k., lim k→∞ Dc(ϕX,nk,k, ϕY,nk,k) = De c(ϕX , ϕY ). Remark 4.2. Let A be a unital simple C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A). Suppose that ϕX (f ) = Pm i=m+1 f (ζi)pi and ϕY (f ) = Pm i=m+1 f (ηi)qi for all f ∈ C(Ω), where {p1, p2, ...., pn1} and {q1, q2, ..., qn2} are two sets of mutually orthogonal non-zero projections such that Pn1 i=1 qi, i=1 pi = 1A = Pn2 i=1 f (ξi)qi + Pn2 i=1 f (ξi)pi +Pn1 21 {ξ1, ξ2, ..., ξm} = X ∩ Y and ζi ∈ X \ Y and ηi ∈ Y \ X. Let ei,n ≤ pi and e′ projections such that [e1,n] = [ei,n] = [e′ i,n] for all i and n, and i,n ≤ qi be non-zero m lim n→∞ Then sup{τ ( X i=1 ei,n) : τ ∈ T (A)} = 0. (e 4.138) ϕX (f ) = ϕY (f ) = m X i=1 m X i=1 m f (ξi)ei,n + ( X i=1 m f (ξi)e′ i,n + ( X i=1 f (ξi)(pi − ei,n) + f (ξi)(qi − e′ i,n) + n1 X i=m+1 n2 X i=m+1 f (ζi)pi) and (e 4.139) f (ηi)qi) (e 4.140) for all f ∈ C(Ω). It makes sense that one insists that ei,n pairs with e′ pairs according to the distance Dc defined in 2.10. After all, ei,n and e′ same point ξi ∈ X ∩ Y. Proposition 4.3. Let Ω be a compact metric space and X ⊂ Ω. Let ∆ : (0, 1) → (0, 1) be an increasing function (with limr→0 ∆(r) = 0). For any ǫ > 0, let r0 = ǫ/16. There is a finite subset of mutually orthogonal projections {f1, f2, ..., fn} ⊂ C(Xǫ/64), a finite subset H ⊂ C(Xǫ/64)+ and σ > 0 satisfying the following: Suppose that A is a unital simple C ∗-algebra with stable rank one and with strict comparison for positive elements and suppose that ϕ, ψ : C(Xǫ/64) → A are two homomorphisms such that i,n and the rest of them i,n corresponds to the ϕ∗0([fi]) = ψ∗0([fi]), i = 1, 2, ..., n, τ ◦ ϕ(g) − τ ◦ ψ(g) < σ for all g ∈ H and µτ ◦ϕ(O) ≥ ∆(r) for all open balls O ∈ Xǫ/64 with radius r > r0 and for all τ ∈ T (A), then Moreover, σ can be chosen to be Dc(ϕ, ψ) < ǫ. σ = (1/4)∆(ǫ/16). (e 4.141) (e 4.142) (e 4.143) (e 4.144) (e 4.145) Proof. Let ǫ > 0. Let Ω1 = Xǫ/64. Since Ω1 is compact, there are ξ1, ξ2, ..., ξm1 ∈ Ω1 such that ∪m1 i=1O(ξi, ǫ/16) ⊃ Ω1. Let G1, G2, ..., GK be all possible finite union of those O(ξi, ǫ/16)'s. If O ⊂ Ω1 is any open subset, there is Gj such that O ⊂ Gj and dH (O, Gj) ≤ ǫ/8. (e 4.146) Without loss of generality, we may assume that (G1)5ǫ/64, (G2)5ǫ/64, ..., (GK0 )5ǫ/64 are clopen sets and (GK0+1)5ǫ/64, (GK0+2)5ǫ/64, ..., (GK )5ǫ/64 are not closed. Therefore, for j > K0, there is ζ ′ j ∈ Ω1 such that ζ ′ j, Gj ) = 5ǫ/64. This implies there exists ζj ∈ X such that 4ǫ/64 < dist(ζj, Gj) < 6ǫ/64. There is η > 0 such that ζj 6∈ (Gj)η+ǫ/16, K0 < j ≤ K. j 6∈ (Gj)5ǫ/64 and dist(ζ ′ Note that O(ζj, ǫ/16) ⊂ (Gj)10ǫ/64 \ (Gj )η. (e 4.147) Therefore 22 inf{dτ (ϕ(f(Gj )10ǫ/64) − dτ (ϕ(fGj ))) : τ ∈ T (A)} ≥ ∆(ǫ/16). (e 4.148) Let S1, S2, ..., Sn be clopen subsets of Ω1 such that Sj = (Gj )5ǫ/64, j = 1, 2, ..., K0, and n ≥ K0. Let fj be the characteristic function of Sj j = 1, 2, ..., n. Let F ⊂ C(Ω1)+ be the finite subset which contains gj, j = 1, 2, ..., K, where 0 ≤ gj ≤ 1, gj(t) = 1 on (Gj )10ǫ/64 and gj(t) = 0 if t 6∈ (Gj)ǫ/4. Choose 0 < σ = (1/4)∆(ǫ/16). Now suppose that ψ : C(Ω1) → A is a unital homomorphisms which satisfies the assumption Let O ⊂ Ω1 be an open subset. Then we may assume that (e 4.146) holds. In particular, (e 4.141), (e 4.142) and (e 4.143). If 1 ≤ j ≤ K0, by the assumption (e 4.141), [ϕ(f(Gj )5ǫ/64 )] = [ψ(f(Gj )5ǫ/64 )], j = 1, 2, ..., K0. (Gj)ǫ/4 ⊂ Oǫ. (e 4.149) (e 4.150) Therefore ϕ(fO) . ϕ(f(Gj )5ǫ/64 ) . ψ(f(Gj )5ǫ/64 ) . ψ(f(Gj )ǫ/4) . ψ(fOǫ). (e 4.151) If K0 < j ≤ K, dτ (ψ(fOǫ)) ≥ τ (ψ(gj )) > τ (ϕ(gj )) − σ ≥ dτ (ϕ(g(Gj )10ǫ/64)) − σ > dτ (ϕ(gGj )) ≥ dτ (ϕ(fO)) (e 4.152) (e 4.153) (e 4.154) for all τ ∈ T (A). Combining (e 4.151) and (e 4.152), since A has the strict comparison for positive elements, we obtain Dc(ϕ, ψ) < ǫ. (e 4.155) Corollary 4.4. Let ǫ > 0. Let X be a compact subset of the plane. Suppose that X = ⊔n i=1Si, where each Si is an ǫ/64-connected component, i = 1, 2, ..., n, suppose that A is a unital simple C ∗-algebra of stable rank one, real rank zero and weakly unperforated K0(A). Suppose that {λ1, λ2, ..., λm0}, {µ1, µ2, ..., µm1} and {ζ1, ζ2, ..., ζm2} are ǫ/16-dense in X, and suppose that {e0,1, e0,2, ..., e0,m0}, {e1,1, e1,2, ..., e1,m1} and {e2,1, e2,2, ..., e2,m2} are mutually orthogonal non- zero projections in A such that P = m1 X j=1 e1,j = m2 X j=1 e2,j, e0,j ∈ (1 − P )A(1 − P ), j = 1, 2, ..., m0, 8τ (P ) < τ (e0,j) for all τ ∈ T (A), j = 1, 2, ..., m0 and ϕ∗0([χSi]) = ψ∗0([χSi]), i = 1, 2, ..., n, j=1 f (µj)e1,j and ψ(f ) = Pm0 j=1 f (λj)e0,j + Pm1 where ϕ(f ) = Pm0 Then there is a unitary u ∈ A such that ku∗ϕ(z)u − ψ(z)k < ǫ, where z ∈ C(X) is the identity function on X. 23 (e 4.156) (e 4.157) (e 4.158) (e 4.159) (e 4.160) j=1 f (λj)e0,j + Pm2 j=1 f (ζj)e2,j. Proof. The proof is virtually contained in that of 4.3. We can keep all notations and proof of 4.3 up to the definition of σ. We define Since A is unital and simple, σ > 0. By the proof of 4.3, for any O, there is Gj such that σ = min{inf{τ (e0,j) : τ ∈ T (A)} : j = 1, 2, ..., m0}. O ⊂ Gj ⊂ (Gj)ǫ/8 ⊂ (Gj)5ǫ/64 = Sj. If Oǫ/16 = Sj, then In general, ψ(fO) . ψ(χSj ) . ϕ(χSj ) . ϕ(fOǫ/16). (e 4.161) dτ (ψ(fO)) < dτ (ϕ(fO)) + (1/4)σ (e 4.162) for all τ ∈ T (A). If Oǫ/16 6= Sj, there is ξ 6∈ Oǫ/16 such that dist(ξ, Oǫ/16) < ǫ/64. There is λj such that Then Then, by (e 4.162), dist(ξ, λj) < ǫ/16. λj 6∈ O and λj ∈ Oǫ/8. dτ (ψ(fO)) ≤ dτ (ϕ(fO)) + (1/4)σ < dτ (ϕ(fO)) + dτ (ϕ(O(λj , ǫ/16))) ≤ dτ (ϕ(fOǫ)). It follows that ψ(fO) . ϕ(fOǫ). This holds for any open set O ⊂ X. Therefore Dc(ψ(z), ϕ(z)) < ǫ. The corollary then follows from 3.5. (e 4.163) (e 4.164) (e 4.165) (e 4.166) (e 4.167) (e 4.168) The following follows from Proposition 11.1 of [23] immediately. Proposition 4.5. Let A be a unital simple C ∗-algebra with T (A) 6= ∅ and let X be a compact metric space. Suppose that ϕ : C(X) → A is a unital monomorphism. Then there is an increasing function ∆ : (0, 1) → (0, 1) with limr→0 ∆(r) = 0 such that µτ ◦ϕ(O) ≥ ∆(r) for all τ ∈ T (A) (e 4.169) where O is an open ball of X with radius r ∈ (0, 1). 24 Corollary 4.6. Let A be a unital simple C ∗-algebra with stable rank one and with strict com- parison for positive elements, let Ω be a compact metric space, let X ⊂ Ω be a compact subset and let ϕX → A be a unital homomorphism with spectrum X. For any ǫ > 0, there is δ > 0, a finite set of clopen subsets S1, S2, ..., Sk in Xǫ/64 and a finite subset F ⊂ C(Xǫ/64)+ such that for any unital homomorphism ϕY : C(Xǫ/64) → A with the property that τ (ϕX (f )) − τ (ϕY (f )) < δ for all f ∈ F and [ϕX (χSi)] = [ϕY (χSi)], i = 1, 2, ..., k, then Dc(ϕX , ϕY ) < ǫ. (e 4.170) (e 4.171) (e 4.172) Proposition 4.7. Let Ω be a compact metric space, let X ⊂ Ω be a compact subset, let A be a unital simple C ∗-algebra of stable rank one and with strict comparison for positive elements and let ϕX : C(Ω) → A be a unital homomorphism with spectrum X. Suppose that {hn : C(Ω) → A} is a sequence of unital homomorphisms such that Then lim n→∞ Dc(ϕX , hn) = 0. lim n→∞ De c (ϕX , hn) = 0. (e 4.173) (e 4.174) Proof. Fix ϕX . Let ǫ > 0. Without loss of generality, we may assume that Ω = Xǫ/64. Let S1, S2, ..., Sk be a finite set of clopen subsets of Ω, F ⊂ C(Ω)+ be a finite subset and δ > 0 be required by 4.6 for ϕX and ǫ. Let Yn be the spectrum of hn. By (e 4.173), we assume that Yn ⊂ Xǫ/64 for all n, without loss of generality. Furthermore, we may further assume that X∩Yn 6= ∅ for all n. Suppose that {en} ⊂ A is a sequence of projections, {ϕ0,n : C(Ω) → enAen} is a sequence of unital homomorphisms with finite spectrum that is ǫn-dense in X ∩ Yn and with limn→∞ ǫn = 0, and two sequences of unital homomorphisms {ϕ1,n, ϕ2,n : C(Ω) → (1− en)A(1− en)} such that the spectrum of ϕ1,n is in X, the spectrum of ϕ2,n is in X ∩ Yn, n = 1, 2, ... and (e 4.175) lim n→∞ sup{τ (en) : τ ∈ T (A)} = 0, Dc(ϕX , ϕ0,n + ϕ1,n) = 0 and lim n→∞ lim n→∞ Dc(hn, ϕ0,n + ϕ2,n) = 0. (e 4.176) (e 4.177) By 2.16, we may assume that, for all n ≥ 1, [ϕX (χSi)] = [ϕ0,n(χSi)] + [ϕ1,n(χSi)] = [ϕ0,n(χSi)] + [ϕ2,n(χSi)], (e 4.178) j = 1, 2, ..., k and n = 1, 2, .... Since A has stable rank one, this implies that [ϕ1,n(χSi)] = [ϕ2,n(χSi)], j = 1, 2, ..., k and n = 1, 2, .... (e 4.179) By (e 4.175), there exists N ≥ 1 such that, for all n ≥ N, It follows from (e 4.176), (e 4.177) and (e 4.173) that sup{τ (en) : τ ∈ T (A)} < δ. (e 4.180) sup{τ (ϕ1,n(f )) − τ (ϕ2,n(f )) : τ ∈ T (A)} < δ for all f ∈ F. (e 4.181) 25 It follows from 4.6 that It follows that Dc(ϕ1,n, ϕ2,n) < ǫ for all n ≥ N. lim sup n→∞ De c (ϕX , hn) = 0. (e 4.182) (e 4.183) In fact, we proved the following: Corollary 4.8. With the same assumption above, for a fixed unital homomorphism ϕ : C(Ω) → A, and for any ǫ > 0, there exists δ > 0 such that if ψ : C(Ω) → A is another unital homomor- phism such that Dc(ϕ, ψ) < δ, then De c (ϕ, ψ) < ǫ. Since for any ϕ, Dc(ϕ, ϕ) = 0, we have De c (ϕ, ϕ) = 0. We also have the following: (e 4.184) (e 4.185) Proposition 4.9. Let A be a unital simple C ∗-algebra with strict comparison for positive ele- ments. Then De c(ϕX , ϕY ) ≤ DT (ϕX , ϕY ). (e 4.186) Proof. By (e 2.55) in 2.21 and (e 4.135), we may assume X ∩ Y 6= ∅ and let Ω = X ∪ Y . Suppose that {en} ⊂ A is a sequence of projections, hn : C(Ω) → enAen is a sequence of unital homomorphisms with spectrum being ǫn-dense in X ∩ Y and with limn→∞ ǫn = 0, and two sequences of unital homomorphisms ϕXn, ϕYn : C(Ω) → (1 − en)A(1 − en) such that Xn ⊂ X, n = 1, 2, ... and lim n→∞ lim n→∞ sup{τ (en) : τ ∈ T (A)} = 0, Dc(ϕX , ϕXn + hn) = 0 and lim n→∞ Dc(ϕY , ϕYn + hn) = 0 (e 4.187) (e 4.188) (e 4.189) (see the lines around (e 4.134)). For any fixed ǫ > 0, let r0 = DT (ϕX , ϕY ). We will show De c (ϕX , ϕY ) < r0 + 3ǫ. It is suffices to show there exists integer N such that for all n > N , Dc(ϕXn , ϕYn) < r0 + 3ǫ. (e 4.190) There are finitely many open subsets {Ok : k = 1, 2, ..., K} of Ω such that for any open set O ⊂ Ω, there exists Ok such that O ⊂ Ok ⊂ Oǫ. (e 4.191) We may assume that Ok ∩ X 6= X, k = 1, 2, ..., J; Ok ∩ X = X, k = J + 1, J + 2, ..., K. 26 Define δ = inf{dτ (ϕY (f(Ok)r0+ǫ)) − dτ (ϕX (fOk )) : τ ∈ T (A), k = 1, 2, ..., J}. (e 4.192) By 2.20, δ > 0. Let N be large such that for any n > N , sup{τ (en) : τ ∈ T (A)} < δ 2 (e 4.193) and Dc(ϕX , ϕXn + hn) < ǫ, Dc(ϕY , ϕYn + hn) < ǫ. (e 4.194) Since DT (ϕX , ϕXn + hn) ≤ Dc(ϕX , ϕXn + hn) < ǫ, for any open subset O with Oǫ ∩ X 6= X, we have dτ (ϕXn (fO)) ≤ dτ (ϕXn(fO)) + dτ (hn(fO)) ≤ dτ (ϕX (fOǫ)). (e 4.195) Suppose Oǫ ⊂ Ok ⊂ (Oǫ)ǫ, then k ≤ J and by (e 4.192), dτ (ϕX (fOǫ)) ≤ dτ (ϕX (fOk)) < dτ (ϕY (f(Ok)r0+ǫ)) − δ 2 . (e 4.196) Using the fact that Dc(ϕX , ϕXn + hn) < ǫ, we also have dτ (ϕY (f(Ok)r0+ǫ)) ≤ dτ (ϕYn(f(Ok)r0+2ǫ)) + dτ (hn(f(Ok)r0+2ǫ)) (e 4.197) and dτ (ϕYn (f(Ok)r0+2ǫ)) ≤ dτ (ϕYn (fOr0+3ǫ)) (e 4.198) Further since dτ (hn(f(Ok)r0+2ǫ)) ≤ τ (en) < δ we obtain 2 , combing (e 4.195),(e 4.196),(e 4.197)and (e 4.198), dτ (ϕXn(fO)) < dτ (ϕYn (fOr0+3ǫ)). (e 4.199) The strict comparison for positive elements implies [ϕXn (fO)] ≤ [ϕYn(fOr0+3ǫ)]. By the exactly the same argument, [ϕYn(fO)] ≤ [ϕXn(fOr0+3ǫ)]. Since DT (ϕX , ϕY ) ≥ dH(X, Y ), for any open set O with Oǫ∩X = X, we have Or0+3ǫ∩Y = Y , this implies [ϕXn (fO)] ≤ [ϕYn(fOr0+3ǫ)]. Therefore Dc(ϕXn, ϕYn ) ≤ r0 + 3ǫ holds for all n > N . That is the end of the proof. Corollary 4.10. Let A be a unital simple C ∗-algebra with strict comparison for positive ele- ments and with T (A) 6= ∅. Suppose that X, or Y are connected compact subsets. Then Dc(ϕX , ϕY ) = De c (ϕX , ϕY ). (e 4.200) Proof. This follows from 4.9 and the last part of 2.21. 27 Proposition 4.11. Let A be a unital simple C ∗-algebra with stable rank one, with the strict comparison for positive elements. Suppose that ϕX , ϕY : C(X ∪ Y ) → A are two unital homo- morphisms with spectrum X and Y, respectively. Let {ξ1, ξ2, ..., ξk} ⊂ X ∩ Y. Suppose that there is a sequence of finite subsets of mutually orthogonal non-zero projections {e1,n, e2,n, ..., ek,n} of A such that lim n→∞ Dc(ϕX,n, ϕX ) = 0, lim n→∞ Dc(ϕY,n, ϕY ) = 0, k and lim n→∞ sup{ X i=1 τ (ei,n) : τ ∈ T (A)} = 0, (e 4.201) (e 4.202) where ϕX,n(f ) = Pk i=1 f (ξi)ei,n + ψY,n(f ) for all f ∈ C(X ∪ Y ), and ψX,n, ψY,n : C(X ∪ Y ) → (1 − pn)A(1 − pn) are a unital homomorphisms with the spectrum in X and Y, respectively, where pn = Pk i=1 ei,n (please see lines around (e 4.134)). i=1 f (ξi)ei,n + ψX,n(f ), ϕY,n(f ) = Pk Then lim inf n→∞Dc(ψX,n, ψY,n) ≤ De c (ϕX , ϕY ) and let Ω = X ∪ Y. Let {ξ(n) Proof. Let d = De subsets of X ∩ Y which are ǫn-dense with limn→∞ ǫn = 0, let {e(n) of mutually orthogonal non-zero projections in A with c(ϕX , ϕY ). , ξ(n) 1 2 (e 4.203) , ..., ξ(n) 1 , e(n) k(n)} be a sequence of finite 2 , ..., e(n) k(n)} be a sequence k(n) lim n→∞ sup{ X i=1 τ (e(n) i ) : τ ∈ T (A)} = 0, (e 4.204) and {un} be be a sequence of unitary such that lim n→∞ lim n→∞ Dc(ϕX , hX,n,0 + hX,n,1) = 0, lim n→∞ Dc(hX,n,1, hY,n,1) = d = De c(ϕX , ϕY ), Dc(ϕY , hY,n,0 + hY,n,1)) = 0 and (e 4.205) (e 4.206) i=1 f (ξ(n) where hX,n,0(f ) = hY,n,0(f ) = Pk(n) for all f ∈ C(Ω) and hX,n,1, hY,n,1 : C(Ω) → (1−En)A(1−En) are unital homomorphisms with spectrum in X and Y, respectively, and where En = Pk(n) , n = 1, 2, .... Without loss of generality, we may also assume that k(n) ≥ k and ξ(n) i = ξi, i = 1, 2, ..., k. By (e 4.202), since A has strict comparison, by passing to a subsequence of ϕi,n (ψX,n and ψY,n), if necessary, we may further assume that i=1 e(n) )e(n) i i i ej,n ≤ e(n) j , j = 1, 2, ..., k and n = 1, 2, .... (e 4.207) Define h′ X,n,0(f ) = k X i=1 k f (ξ(n) i )ei,n, k(n) h′′ X,n,0(f ) = X,n,1 = h′′ h′ i )(e(n) f (ξ(n) X X,n,0 + hX,n,1 and h′ i=1 X Y,n,1 = h′′ i − ei,n) + i=k+1 X,n,0 + hY,n,1. Then, by assumption, lim n→∞ Dc(h′ X,n,0 + ψX,n, h′ X,n,0 + h′ X,n,1) = 0 28 (e 4.208) f (ξ(n) i )e(n) i for all f ∈ C(Ω) (e 4.209) (e 4.210) (e 4.211) and lim n→∞ Dc(h′ X,n,0 + ψY,n, h′ X,n,0 + hY,n,1) = 0. (e 4.212) By the proof of 4.7, (e 4.211) and (e 4.212) imply that lim n→∞ Dc(ψX,n, h′ X,n,1) = 0 and lim n→∞ Dc(ψY,n, h′ Y,n,1) = 0. (e 4.213) It follows that lim inf n→∞ Dc(ψX,n, ψY,n) ≤ lim n→∞ Dc(h′ lim sup n→∞ Dc(ψX,n, h′ X,n,1) + X,n,1, h′ Y,n,1) + lim n→∞ Dc(ψY,n, h′ Y,n,1) = lim sup n→∞ ≤ lim sup n→∞ Dc(h′ X,n,1, h′ Y,n,1) Dc(hX,n,1, hY,n,1) = d. (e 4.214) (e 4.215) (e 4.216) (e 4.217) Lemma 4.12. Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A). Let ϕX : C(X) → A be a unital homomorphism. Then, for any ǫ > 0, any σ > 0, any η > 0 and any finite η-dense subset {ξ1, ξ2, ..., ξm} ⊂ X, there is a projection e ∈ A with τ (e) < σ for all τ ∈ T (A), a unital homomorphism ϕ0 : C(X) → eAe with spectrum {ξ1, ξ2, ..., ξm} and a unital homomorphism ϕ1 : C(X) → (1 − e)A(1 − e) with finite spectrum such that Dc(ϕX , ϕ0 + ϕ1) < ǫ. (e 4.218) Proof. Since A is simple and has real rank zero and stable rank one with weakly unperfo- rated K0(A), K0(A) has Riesz interpolation property by a theorem of Zhang ([33]). Moreover ρA(K0(A)) is dense in Aff(T (A)). By [10], there exists a unital simple AH-algebra of no dimen- sion growth B of real rank zero (therefore T R(B) = 0 -- see [20]) such that (K0(B), K0(B)+, [1B]) = (K0(A), K0(A)+, [1A]) and K1(B) = {0}. (e 4.219) (e 4.220) It follows from [19] that there exists a unital homomorphism h : B → A such that h∗0 gives the identity in (e 4.219). It follows from [22] that there exist unital monomorphisms ψ′ X : C(X) → B such that (h ◦ ψ′ X )∗0 = (ϕX )∗0 and τ ◦ h ◦ ψ′ X = τ ◦ ϕX for all τ ∈ T (A). Define ψX = h ◦ ψ′ X . Then (ψX )∗0 = (ϕX )∗0 and τ ◦ ψX = τ ◦ ϕX for all τ ∈ T (A). These, in particular, by 4.6, imply that Dc(ϕX , ψX ) = 0. (e 4.221) (e 4.222) (e 4.223) So, without loss of generality, we may assume now that A = B. In particular, B has tracial rank zero. 29 Let ǫ > 0. Let δ > 0 be a positive number, S1, S2, ..., Sk be a finite set of mutually disjoint clopen subsets of X and let F ⊂ C(X)+ be a finite subset required by 4.6 for ǫ > 0 and ϕX . We may assume that X = ⊔m i=1Si and 1C(X) ∈ F. By Lemma 4.3 of [22], there is a projection p 6= 1A, a unital homomorphism h : C(X) → pAp with finite spectrum such that for all τ ∈ T (A), and τ ◦ h(f ) − τ ◦ ϕX (f ) < δ/2 for all f ∈ F and τ ◦ h(χSi) < τ ◦ ϕX(χSi), i = 1, 2, ..., k, h(f ) = m X i=1 f (ξi)ei + h1(f ) for all f ∈ C(X), (e 4.224) (e 4.225) (e 4.226) where {e1, e2, ..., em} ⊂ pAp is a set of mutually orthogonal non-zero projections and h1 : C(X) → (p − Pm i=1 ei) is a unital homomorphism with finite spectrum in X. Note that (e 4.224) implies that i=1 ei)A(p − Pm τ (1 − p) < δ/2 for all τ ∈ T (A). (e 4.227) By (e 4.225), there are mutually orthogonal projections q1, q2, ..., qk ∈ (1 − p)A(1 − p) such i=1 χSi = 1C(X) and ϕX is unital, j=1 f (ζj)qj + h(f ) for all f ∈ C(X), that [ϕX (χSi)] = [qi] + [h(χSi)], i = 1, 2, ..., k. Since Pk p + Pk i=1 qi = 1A. Define ψX : C(X) → A by ψX(f ) = Pk where ζj ∈ Sj is a point, j = 1, 2, ..., k. We compute that, [ψX (χSi)] = [ϕX (χSi)], i = 1, 2, ..., k. Moreover, by (e 4.227) and (e 4.224), τ (ϕX (f )) − τ (ψX (f )) < δ for all τ ∈ T (A). It follows from 4.6 that Dc(ϕX , ψX ) < ǫ. (e 4.228) (e 4.229) (e 4.230) Since A is simple and has (SP), we can find non-zero projections e′ Put e = Pm i. Define ϕ0(f ) = Pm i=1 f (ξi)ei for all f ∈ C(X) and defined i=1 e′ i ≤ ei such that Pm i=1 τ (e′ i) < σ. ϕ1(f ) = k X j=1 f (ζj)qj + m X i=1 f (ξi)(ei − e′ i) + h1(f ) (e 4.231) for all f ∈ C(X). Note that ϕ0 + ϕ1 = ψX. Lemma follows. Corollary 4.13. Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and with weakly unperforated K0(A) and let X be a compact metric space. Suppose that ϕX : C(X) → A is a unital homomorphism. Then, there exists a sequence of unital homomorphisms ϕn : C(X) → A with finite dimensional range such that lim n→∞ De c (ϕX , ϕn) = 0. (e 4.232) Remark 4.14. In the case that A has real rank zero, stable rank one and weakly unperforated K0(A), 4.12 shows that, in the definition of De c(ϕX , ϕY ), if X ∩ Y 6= ∅, we can also assume that the sequence of non-zero {hn} exists. 30 Proposition 4.15. Let Ω be a compact metric space, let A be a unital simple C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A), and let ϕX , ϕY , ϕZ : C(Ω) → A be unital homomorphisms with spectrum X, Y and Z, respectively. If, in addition, X ∩ Y ⊂ Z (e 4.233) c (ϕX , ϕZ ) + De c (ϕZ , ϕY ). De c (ϕX , ϕY ) ≤ De Proof. If X ∩ Y = ∅, then De c (ϕX , ϕY ) = Dc(ϕX , ϕY ) ≤ Dc(ϕX , ϕZ ) + Dc(ϕZ , ϕY ) ≤ De c (ϕX , ϕZ ) + De c (ϕZ , ϕX ). By the definition and from 4.14 above, we have nonzero sequences of projections {e(n, i)} of So, we assume that X ∩ Y 6= ∅. A, unital homomorphisms h(n, i) : C(Ω) → e(n, i)Ae(n, i) and unital homomorphisms ϕ(n, i), ϕ(Z, n, i), : C(Ω) → (1 − e(n, i))A(1 − e(n, i)) such that sup τ (e(n, i)) = 0 τ ∈T (A) Dc(ϕX , h(n, 1) + ϕ(n, 1)) = 0, lim n→∞ Dc(ϕY , h(n, 2) + ϕ(n, 2)) = 0; (e 4.235) (e 4.234) Dc(ϕZ , h(n, i) + ϕ(Z, n, i)) = 0; lim n→∞ lim n→∞ lim n→∞ De c (ϕX , ϕZ ) = lim n→∞ c (ϕY , ϕZ ) = lim n→∞ De Dc(ϕ(n, 1), ϕ(Z, n, 1)); Dc(ϕ(n, 2), ϕ(Z, n, 2)) and lim n→∞ Dc(h(n, 1) + ϕ(Z, n, 1), h(n, 2) + ϕ(Z, n, 2)) = 0, (e 4.236) (e 4.237) (e 4.238) (e 4.239) the spectrum of h(n, 1) is X ′ n, that of ϕ(n, 1) is in X, that of ϕ(n, 2) is in Y, ϕ(Z, n, i) is in Z, where X ′ n is a finite subset of Z ∩ Y which are ǫn-dense in X ∩ Z and in Y ∩ Z with limn→∞ ǫn = 0. Since X ∩ Y ⊂ Z, we may assume, without loss of generality, that X ′ n is a finite subset of X ∩ Z and Y ′ n and the spectrum of h(n, 2) is Y ′ n ∩ Y ′ n is ǫn-dense in X ∩ Y. We write h(n, i)(f ) = r(n,i) X j=1 f (ζ(n, j, i))q(n, j, i) for all f ∈ C(Ω), (e 4.240) where {ζ(n, 1, 1), ζ(n, 2, 1), ..., ζ(n, r(n, 1), 1)} = X ′ n, {ζ(n, 1, 2), ζ(n, 2, 2), ..., ζ(n, r(n, 2), 2)} = Y ′ n and {q(n, 1, i)), q(n, 2, i), ..., q(n, r(n, i), i)} is a set of mutually orthogonal non-zero projec- tions. We may further assume that ζ(n, j, 1) = ζ(n, j, 2), j = 1, 2, ..., k(n) ≤ r(n, 1), r(n, 2), (e 4.241) where {ζ(n, 1, 1), ζ(n, 2, 1), ..., ζ(n, k(n), 1)} is ǫn-dense in X ∩ Y. Let Xn be the spectrum of ϕ(n, 1) and Yn be the spectrum of ϕ(n, 2), n = 1, 2, .... Without loss of generality, we may assume that X ′ n ⊂ Yn, n = 1, 2, .... Note that, without changing the sums h(n, i) + ϕ(n, i), h(n, i) + ϕ(Z, n, i) and (e 4.234) -- (e 4.238), one can choose smaller q(n, j, i), j = 1, 2, ..., r(n, i), i = 1, 2 and n = 1, 2, .... We may assume that, since A is simple and has (SP), we may assume that r(n, 1) = r(n, 2) and n ⊂ Xn and Y ′ [q(n, j, 1)] = [q(n, j′, 2)], j, j′ = 1, 2, ...., k(n), n = 1, 2, .... (e 4.242) To simplify the notation, we may further assume that q(n, j, 1) = q(n, j, 2), j = 1, 2, ..., k(n), n = 1, 2, .... (e 4.243) 31 Put ϕ(n, i)′(f ) = ϕ(Z, n)(f ) = ϕ(Z, n, 2)′(f ) = r(n,i) X j=k(n)+1 r(n,1) X j=k(n)+1 r(n,1) X j=k(n)+1 f (ζ(n, j, i))q(n, j, i) + ϕ(n, i)(f ), (e 4.244) f (ζ(n, j, i))q(n, j, 1) + ϕ(Z, n, 1)(f ) and (e 4.245) f (ζ(n, j, 2))q(n, j, 2) + ϕ(Z, n, 2)(f ) (e 4.246) for all f ∈ C(Ω). It follows that lim sup n→∞ Dc(ϕ(n, 1)′, ϕ(Z, n)) ≤ lim sup n→∞ Dc(ϕ(n, 1), ϕ(Z, n, 1)) = De c (ϕX , ϕZ ). (e 4.247) By (e 4.239), (e 4.243), and the proof of 4.7, lim n→∞ Dc(ϕ(Z, n), ϕ(Z, n, 2)′) = 0. It follows that lim sup n→∞ Dc(ϕ(n, 2)′, ϕ(Z, n)) n→∞ ≤ lim sup ≤ lim sup n→∞ Dc(ϕ(n, 2)′, ϕ(Z, n, 2)′) + lim n→∞ c(ϕY , ϕZ ). Dc(ϕ(n, 2), ϕ(Z, n, 2)) = De Dc(ϕ(Z, n, 2)′, ϕ(Z, n)) However, by (e 4.247) and (e 4.249), n→∞ De c (ϕX , ϕY ) ≤ lim sup ≤ lim sup ≤ De n→∞ c (ϕX , ϕZ ) + De c(ϕZ , ϕY ). Dc(ϕ(n, 1)′, ϕ(n, 2)′) (Dc(ϕ(n, 1)′, ϕ(Z, n)) + Dc(ϕ(n, 2)′, ϕ(Z, n))) (e 4.248) (e 4.249) (e 4.250) (e 4.251) (e 4.252) (e 4.253) (e 4.254) Definition 4.16. Let A be a unital C ∗-algebra and let x, y ∈ A be two normal elements with sp(x) = X and sp(y) = Y, respectively. Define ϕX , ϕY : C(X ∪ Y ) → A to be unital homomorphisms defined by ϕX(f ) = f (x) and ϕY (f ) = f (y) for all f ∈ C(X ∪ Y ). We will use the notation De c (x, y) for De c (ϕX , ϕY ). 5 Approximate unitary equivalence The purpose of this section is to present 5.6 and 5.5. In the case that A is a unital simple C ∗-algebra with T R(A) = 0, much more general results were presented in [21]. However, in the spirit of [17], the exact condition for when two normal elements are approximately unitarily equivalent can be obtained in unital simple C ∗-algebra A with real rank zero, stable rank one and with weakly unperforated K0(A). We are also interested in 5.7. The following is proved in [17]. 32 Theorem 5.1. Let ǫ > 0. For any unital simple C ∗-algebra A of real rank zero with (IR) and any normal element x ∈ A with kxk ≤ 1 such that λ − x ∈ Inv0(A) (e 5.255) for all λ ∈ C with dist(λ, sp(x)) ≥ ǫ/8, there is a normal element with finite spectrum x0 ∈ A such that Proof. This was exactly proved in the proof of the Theorem of [17]. Note that the set kx − x0k < ǫ. (e 5.256) in that proof is chosen for r = ǫ/8. F1 = {λ ∈ C : dist(λ, sp(x)) < r} Lemma 5.2. Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A), let X be a compact subset of the plane and let ∆ : (0, 1) → (0, 1) be an increasing function such that limr→0 ∆(r) = 0. Then, for any ǫ > 0, there exists d > 0 with d < ǫ/128, there exists a finite subset {f1, f2, ..., fn} ⊂ C(Xd/2) of mutually orthogonal projections with Pn i=1 fi = 1C(Xd/2), a finite subset H ⊂ C(Xd/2)+ satisfying the following: if h : C(Xd/2) → A is a homomorphism such that µτ ◦h(O) ≥ ∆(r) (e 5.257) for any open balls O of X with radius r ≥ ǫ/32, then for any homomorphism ϕ : C(Xǫ/128) → A such that ϕ∗0([fi]) = h∗0([fi]), i = 1, 2, ..., n, λ − h(z), λ − ϕ(z) ∈ Inv0(A) if dist(λ, X) ≥ d, and τ ◦ h(g) − τ ◦ ϕ(g) < (1/4)∆(ǫ/32) for all g ∈ H, then there exists a unitary u ∈ A such that ku∗h(z)u − ϕ(z)k < ǫ, (e 5.258) (e 5.259) (e 5.260) (e 5.261) where z ∈ C(Xd/2) is the identity function on Xd/2. Proof. Let ǫ > 0. We choose δ > 0 which is required by 2.17 for ǫ/8 (with Xǫ/16 = Ω). Let d = min{δ/8, ǫ/221}. (e 5.262) 1, f ′ n} ⊂ C(Xǫ/128) be a subset of projections be as required by 4.3 for ∆, Xd/2 (in , i = 1, 2, ..., n. Now assume A, h and ϕ 2, ..., f ′ Let {f ′ place of X) and ǫ/2 (instead of ǫ). Define fi = f ′ be as stated above.. By applying 4.3, one has iXd/2 Dc(h, ϕ) < ǫ/2. (e 5.263) It follows from 5.1 that, if (e 5.259) holds, there are normal elements x0 and y0 with finite spectrum such that kh(z) − x0k < min{ǫ/16, δ} and kϕ(z) − y0k < min{ǫ/16, , δ}. (e 5.264) 33 By 2.17, we have that Therefore Dc(h(z), x0) < ǫ/8 and Dc(ϕ(z), y0) < ǫ/8 (e 5.265) Dc(x0, y0) < 3ǫ/4. (e 5.266) Since A is simple and has real rank zero and stable rank one with weakly unperforated K0(A), K0(A) has Riesz interpolation property by a theorem of Zhang [33], by 3.5, there exists a unitary u ∈ A such that Combining this with (e 5.264), we conclude that ku∗x0u − y0k < 3ǫ/4. ku∗h(z)u − ϕ(z)k < ǫ. (e 5.267) (e 5.268) Lemma 5.3. Let A be a unital infinite dimensional separable simple C ∗-algebra with real rank zero, stable rank one and with weakly unperforated K0(A) and X ⊂ C be a compact subset. Let p1, p2, ..., pn ∈ C(X) be n mutually orthogonal projections with Pn i=1 pi = 1C(X) such that {[p1, ], [p2], ..., [pn]} generates a subgroup G of K0(C(X)). Suppose that κ0 : G → K0(A) is an order preserving homomorphism with κ0([1C(X)]) = [1A] and with κ0([pi]) > 0, i = 1, 2, ..., n, and κ1 : K1(C(X)) → K1(A) is a homomorphism. Then there is a unital monomorphism ϕ : C(X) → A such that ϕ∗0G = κ0 and ϕ∗1 = κ1. (e 5.269) Proof. Since A is simple and has real rank zero and stable rank one with weakly unperforated K0(A), as in the proof of 5.2, K0(A) has Riesz interpolation property and ρA(K0(A)) is dense in Aff(T (A)). It follows from [10] that there is a unital simple AH-algebra B with no dimension growth and real rank zero such that (K0(A), K0(A)+, [1A], K1(A)) = (K0(B), K0(B)+, [1B], K1(B)). (e 5.270) It follows from 4.6 of [19] that there exists a unital embedding ı : B → A which carries the above identification. We will also use the fact that ρB(K0(B)) = ρA(K0(A)) is dense in Aff(T (A)). Therefore it suffices to show that the lemma holds for A = B. There are mutually orthogonal nonzero projections e1, e2, ..., en ∈ B such that κ0([pi]) = ei, i = 1, 2, ..., n. Let Xi be the clopen subset of X corresponding to the projection pi, i = 1, 2, ..., n. Each eiBei is an infinite dimensional unital simple C ∗-algebra with T R(eiBei) = 0. Therefore there is a monomorphism ψi : C(Xi) → eiBei, i = 1, 2, ..., n. Define a unital monomorphism ψ : C(X) → B by ψ(f ) = Pn i=1 ψi(f pi) for all f ∈ C(X). Note that (e 5.271) We also have that ψ∗0 and ψ♯ : C(X)s.a. → Aff(T (B)) are compatible and ψ♯ is strictly positive. It follows from Corollary 5.3 of [22] that there is a unital monomorphism ϕ : C(X) → B such that κ0 = ψ∗0G. Lemma follows. ϕ∗0 = ψ∗0 and ϕ∗1 = κ1. 34 (e 5.272) (e 5.273) Lemma 5.4. Let A be a unital simple C ∗-algebra of real rank zero, stable rank one with weakly unperforated K0(A), let X ⊂ C be a compact subset of the plane and let ∆ : (0, 1) → (0, 1) be a non-decreasing function such that limt→0 ∆(t) = 0. For any 1 > r0 > 0, any ǫ > 0, any η > 0, any η1 > 0 with η1 < r0/4, any η2 > 0 and any finite subset G ⊂ C(Xη1 )+, where Xη1 = {λ ∈ C : dist(λ, X) < η1}, there is a finite subset H ⊂ C(X)s.a. satisfying the following: If x, y ∈ A are normal elements with sp(x), sp(y) ⊂ X such that τ ◦ ϕ(g) − τ ◦ ψ(g) < η/2 for all g ∈ H and µτ ◦ϕ(O) ≥ ∆(r) for all τ ∈ T (A) (e 5.274) (e 5.275) for all open balls O of X with radius r ≥ r0, where ϕ, ψ : C(X) → A are defined by ϕ(f ) = f (x) and ψ(f ) = f (y) for all f ∈ C(X), respectively, then there exists {λ1, λ2, ..., λn} ⊂ X which is r0- dense in X, non-zero mutually orthogonal projections {e1, e2, ..., en} ⊂ A, two normal elements x0, y0 ∈ eAe, where e = 1 − Pn i=1 ei and a unitary u ∈ A such that n n X i=1 X i=1 λiei + x0)k < ǫ/2, ku∗yu − ( kx − ( τ ◦ ϕ0(g) − τ ◦ ψ0(g) < η for all g ∈ G and for all τ ∈ T (A), sp(x0), sp(y0) ⊂ Xη1, τ ( λiei + y0)k < ǫ/2 ei) < η2 for all τ ∈ T (A) and n X i=1 µτ ◦ϕ0(O) ≥ (1/2)∆(r/6) (e 5.276) (e 5.277) (e 5.278) (e 5.279) (e 5.280) for all open balls O ⊂ Xη1 with radius r ≥ 3r0 and for all τ ∈ T (A), where ϕ0, ψ0 : C(Xη1 ) → A is defined by n n ϕ0(f ) = X i=1 f (λi)ei + f (x0) and ψ0(f ) = f (λi)ei + f (y0) X i=1 for all f ∈ C(Xη1). Proof. To simplify the notation, we may assume that X is a subset of the unit disk. Note that, since A has real rank zero and stable rank one, so does pAp for any projection p ∈ A. It follows that pAp has (IR) (see [11]). Let 1 > ǫ > 0 be given. Let ǫ1 > 0 be such that ǫ/4 > ǫ1 > 0. By [11], there exists δ1 > 0 such that, for any C ∗-algebra D with (IR), any element z ∈ D with kzk ≤ 2 and kz∗z − zz∗k < δ1, then there is a normal element z0 ∈ D such that kz0 − zk < ǫ1. (e 5.281) (e 5.282) Let η, η1, η2 > 0 be given and a finite subset G ⊂ C(Xη1)+ be given. Denote by ϕ′ : C(Xη1 ) → A the homomorphism defined by ϕ′(f ) = f (x) for all f ∈ C(Xη1 ). Since η1 < r0/4, every open ball of Xη1 of radius r > r0 contains an open balls of X of radius r/2. It follows that for all open balls O ⊂ Xη1 with radius r > r0 and for all τ ∈ T (A). µτ ◦ϕ′(O) ≥ ∆(r/2) (e 5.283) 35 We will applying Lemma 2 of [15]. Note that since A has real rank zero, non-zero projections pk described in that lemma exists. Thus, we obtain non-zero mutually orthogonal projections p1, p2, ..., pn ∈ A and p′ n ∈ A such that 2, ..., p′ 1, p′ n kx − ( X i=1 λipi + pxp)k < min{δ1/16, ǫ1/16} and n ky − ( X i=1 λip′ i + p′yp′)k < min{δ1/16, ǫ1/16}, (e 5.284) (e 5.285) where p = Pn i=1 pi, p′ = Pn i=1 p′ i. Since sp(x) and sp(y) are r0-dense by (e 5.275), we may assume that {λ1, λ2, ..., λn} is r0- dense. Since A is simple and has real rank zero, there are possibly smaller non-zero projections ei ≤ pi such that ei . p′ i, i = 1, 2, ..., n. In other words, since A has stable rank one, there is a unitary u ∈ A such that n τ ( X i=1 ei) < η2 for all τ ∈ T (A), λiei + n X i=1 λi(pi − ei) + pxp)k < min{δ1/16, ǫ1/16} and n (e 5.286) (e 5.287) n kx − ( X i=1 where y′ Then 0 = u∗y0u + Pn i=1 λi(u∗p′ X ku∗yu − ( iu − ei). Put x′ i=1 λiei + y′ 0 = Pn 0)k < min{δ1/16, ǫ1/16}, i=1 λi(pi − ei) + pxp. Let e = 1 − Pn (e 5.288) i=1 ei. 0)∗(x′ k(x′ 0) − (x′ 0)(x0)∗k < δ1 and k(y′ 0)∗(y′ 0) − (y′ 0)(y′ 0)∗k < δ1. (e 5.289) By the choice of δ1, by applying [11], there exist normal elements x0, y0 ∈ eAe such that kx0 − x′ 0k < ǫ1 and ky0 − y′ 0k < ǫ1. (e 5.290) It follows that n n kx − ( X i=1 λiei + x0)k < ǫ1 and ku∗yu − ( X i=1 λiei + y0)k < ǫ1. (e 5.291) Define ϕ0, ψ0 : C(Xη1 ) → A by X ϕ0(f ) = n i=1 f (λi)ei + f (x0) and ψ0(f ) = n X i=1 f (λi)ei + f (y0) for all f ∈ C(Xη1). Now, we will choose ǫ1 sufficiently small to begin with. First, by applying Lemma 3.4 of [24], we will choose H sufficiently large and σ sufficiently small (independent of A and normal elements given) so that µτ ◦ϕ0(O) ≥ (1/2)∆(r/6) (e 5.292) for all open balls O of Xη1 of radius r ≥ 3r0, if (e 5.277) holds. In particular, we choose H ⊃ G and σ < η/2. Since G is finite and given, with sufficiently smaller ǫ1, we also have, by (e 5.291), and by assumption (e 5.277), τ ◦ ϕ0(g) − τ ◦ ψ0(g) < η for all g ∈ G (e 5.293) and for all τ ∈ T (A). 36 Theorem 5.5. Let A be a unital separable simple C ∗-algebra of real rank zero, stable rank one with weakly unperforated K0(A), let X ⊂ C be a compact subset of the plane and let ∆ : (0, 1) → (0, 1) be a non-decreasing function such that limt→0 ∆(t) = 0. For any ǫ > 0 there is a finite subset of unitaries V ⊂ C(X), a finite subset of projections {p1, p2, ..., pn} ⊂ C(X), a finite subset H ⊂ C(X)s.a., σ > 0 and r0 > 0 satisfying the following: If x, y ∈ A are normal elements with sp(x), sp(y) ⊂ X such that ϕ∗0([pi]) = ψ∗0([pi]) ϕ∗1V = ψ∗1V , τ ◦ ϕ(g) − τ ◦ ψ(g) < σ for all g ∈ H and µτ ◦ϕ(O) ≥ ∆(r) for all τ ∈ T (A) (e 5.294) (e 5.295) (e 5.296) (e 5.297) for all open balls O of X with radius r ≥ r0, where ϕ, ψ : C(X) → A are defined by ϕ(f ) = f (x) and ψ(f ) = f (y) for all f ∈ C(X), respectively. then there exists a unitary u ∈ A such that ku∗xu − yk < ǫ. (e 5.298) Proof. To simplify notations, we may assume that X is a compact subset of the unit disk. Let ∆1(r) = (1/64)∆(r/12) for all r ∈ (0, 1). Let ǫ > 0. Choose F = {z}, where z is the identity function on the unit disk. Let ǫ0 = ǫ/64. Choose r0 = ǫ0/16. Let d > 0 with d < ǫ0/220 be required by 5.2 for ǫ0 (in place of ǫ) and ∆1 (in place of ∆). Put Y = Xd/2. Let {f1, f2, ..., fn} ⊂ C(Y ) be a finite subset of mutually orthogonal projections and let H1 ⊂ C(Y ) (in place of H) be required by 5.2 for ǫ0 (in place of ǫ) and for ∆1 (in place of ∆). We may assume that 1C(Y ) = Pn i=1 fi and fi corresponding to r0/214-connected components. Note that Y is homeomorphic to a finite CW complex in the plane. Let {v1, v2, ..., vn1} ⊂ Y be a set of unitaries which generates K1(C(Y )). Choose ǫ1 > 0 satisfying the following: if x′, y′ be two normal elements in a unital C ∗-algebra B with sp(x′), sp(y′) ⊂ Y and kx′ − y′k < ǫ1, then (ϕ′)∗0([fi]) = (ψ′)∗0([fi]), i = 1, 2, ..., n and (ϕ′)∗1 = (ψ′)∗1, (e 5.299) (e 5.300) where ϕ′, ψ′ : C(Y ) → B are defined by ϕ′(f ) = f (x′) and ψ′(f ) = f (y′) for all f ∈ C(Y ). Let ǫ2 = min{ǫ1/4, ǫ0/2}. Let η = (1/210)(∆1(ǫ0/64)), η1 = min{d/2, r0/64} and η2 = η. Let H ⊂ C(X)s.a be a finite subset be required by 5.4 for r0, ǫ2 (in place of ǫ), η, η1, η2, H1 (in place of G) and ∆. Let pi = fiX , i = 1, 2, ..., n and let uj = vjX , j = 1, 2, ..., n1. Put V = {u1, u2, ..., un1}. Let σ = η/2. Now suppose that x, y are two normal elements in A satisfying the assumption for the above V, {p1, p2, ..., pn}, H, σ and r0. By 5.4, there exists {λ1, λ2, ..., λm} ⊂ X which is r0-dense, there are non-zero mutually orthogonal projections {e1, e2, ..., em} ⊂ A, two normal elements x0, y0 ∈ eAe, where e = 1 − 37 Pm i=1 ei and a unitary w ∈ A such that m m i=1 X λiei + x0)k < ǫ2/2, kw∗yw − ( kx − ( τ ◦ ϕ0(g) − τ ◦ ψ0(g) < η for all g ∈ H1 and for all τ ∈ T (A), sp(x0), sp(y0) ⊂ Xη1 , τ ( λiei + y0)k < ǫ2/2, ei) < η2 for all τ ∈ T (A) and X i=1 m X i=1 µτ ◦ϕ0(O) ≥ (1/2)∆(r/6) (e 5.301) (e 5.302) (e 5.303) (e 5.304) (e 5.305) for all open balls O ⊂ Xη1 with radius r ≥ 3r0 and for all τ ∈ T (A), where ϕ0, ψ0 : C(Xη1 ) → A is defined by ϕ0(f ) = m X i=1 f (λi)ei + f (x0) and ψ0(f ) = m X i=1 f (λi)ei + f (y0) for all f ∈ C(Xη1). mutually orthogonal projections ei,0, ei,1, ei,2 such that Since A is a unital simple C ∗-algebra with real rank zero, there are, for each i, non-zero n ei = ei,0 + ei,1 + ei,2 and 9τ ( X i=1 (ei,1 + ei,2)) < τ (ej) (e 5.306) for all τ ∈ T (A) and j = 1, 2, ..., m. Define ϕ′ 0(f ) = m X i=1 f (λi)ei, ϕ0,0(f ) = f (λi)ei,0, m X i=1 m ϕ0,1(f ) = m X i=1 f (λi)ei,1 and ϕ0,2 = f (λi)ei,2 X i=1 (e 5.307) (e 5.308) for all f ∈ C(Y ). Put P1 = Pm i=1 ei,1 and P2 = Pn i=1 ei,2. We have τ (P1 + P2) < η2/8 for all τ ∈ T (A). (e 5.309) It follows from 5.3 that there are unital monomorphisms H1 : C(Y ) → P1AP1 and H2 : C(Y ) → P2AP2 such that (H1)∗0([fi]) = (ϕ0,1)∗0([fi]), (H2)∗0([fi]) = (ϕ0,2)∗0([fi]), i = 1, 2, ..., n and (H1)∗0([vj]) = −(ϕx)∗1([vj ]) and (H2)∗1([vj]) = (ϕx)∗1([vj]), j = 1, 2, ..., n1, (e 5.310) (e 5.311) where ϕx : C(Y ) → A is defined by ϕx(f ) = f (x) for all f ∈ C(Y ). Let x1 = H1(z) + H2(z), where z is the identity function on Y. Then, by 5.1, there are {µ1, µ2, ...., µm1} which is r0/212- dense and mutually orthogonal non-zero projections {e1,3, e2,3, ..., em1,3} in (P1 + P2)A(P1 + P2) such that kx1 − m1 X j=1 µjej,3k < r0/212. 38 (e 5.312) By 4.4, there is a unitary v ∈ (1 − e)A(1 − e) such that kv∗ϕ′ 0(z)v − (x1 + ϕ0,0(z))k < r0/211. Define ϕ′ x, ψ′ y : C(Y ) → (1 − P2)A(1 − P2) by ϕ′ x(f ) = H1(f ) + f (x0) + ϕ00(f ) and ψ′ y(f ) = H1(f ) + f (y0) + ϕ00(f ) for all f ∈ C(Y ). We have, for all λ ∈ C, x(z), λ − ψ′ λ − ϕ′ y(z) ∈ Inv0((1 − P2)A(1 − P2)) (e 5.313) (e 5.314) (e 5.315) if λ 6∈ Xd/2. By the choice of ǫ1, (e 5.301) and (e 5.310), and the fact that A has stable rank one, we check that (ϕ′ x)∗0([fi]) = (ψ′ y)∗0([fi]), i = 1, 2, ..., n. (e 5.316) For each open ball O ⊂ Xd/2 with radius r > r0, we estimate that, by (e 5.304) and (e 5.305), (e 5.317) µτ ◦ϕ′ x(O) > µτ ◦ϕ0(O) − τ (P1) − τ (P2) (e 5.318) (e 5.319) (e 5.320) (e 5.321) (e 5.322) (e 5.323) ≥ (1/2)∆(r/6) − η2 ≥ (1/2)∆(r/6) − 2−10∆1(ǫ0/64) > 1/4∆(r/6) ≥ ∆1(r) for all r ≥ 3r0 and for all τ ∈ T (A). It follows that µt◦ϕ′ x(O) > ∆1(r) for all t ∈ T ((1 − P2)A(1 − P2)) and for all open balls O with radius r > 3r0. For f ∈ H1, by (e 5.304) and (e 5.302) τ ◦ ϕ′ x(f ) − τ ◦ ψ′ y(f ) < τ ◦ ϕ0(f ) − τ ◦ ψ0(f ) + 2τ (1 − e) < η + 2η2 ≤ (3/210)(∆1(ǫ0/64)) for all τ ∈ T (A). It follows that x(f ) − t ◦ ψ′ for all t ∈ T ((1 − P2)A(1 − P2)). t ◦ ϕ′ y(f ) < (3/210)(∆1(ǫ0/64)) 1 − η < (1/4)∆(ǫ0/64) (e 5.324) It follows from 5.2 that there is a unitary w0 ∈ (1 − P2)A(1 − P2) such that 0ϕ′ kw∗ y(z)w0 − ϕ′ x(z)k < ǫ0. (e 5.325) Put w1 = v + e and w2 = w0 + (1− P2) and u = w1w∗ and (e 5.313) again, 2w∗ 1. Then, by (e 5.301), (e 5.313), (e 5.307) x ≈ ǫ2/2 ϕ0(z) ≈r0/211 w1(x1 + ϕ00(z) + x0)w∗ w1(H2(z) + H1(z) + ϕ00(z) + x0)w∗ 1 1 = 2ϕ′ ≈ ǫ0 w1(H2(z) + w∗ w1w∗ = ≈ r0/211 w1w∗ ≈ ǫ2/2 uyu∗. 1(ϕ′ y(z)w2)w∗ 1 2(H2(z) + H1(z) + ϕ00(z) + y0)w2w∗ 1 2(w∗ 0(z) + y0)w1w2w∗ 1 (e 5.326) (e 5.327) (e 5.328) (e 5.329) (e 5.330) (e 5.331) But ǫ2/2 + r0/211 + ǫ0 + r0/211 + ǫ2/2 < ǫ. 39 Theorem 5.6. Let A be a unital separable simple C ∗-algebra of real rank zero, stable rank one and weakly unperforated K0(A) and let x ∈ A be a normal element with sp(x) = X. For any ǫ > 0 there is a finite subset V ⊂ K1(C(sp(x))), a finite subset P ⊂ K0(C(sp(x))), a finite subset H ⊂ C(sp(x))s.a., σ > 0 satisfying the following: If y ∈ A is normal element with sp(y) ⊂ X such that ϕ∗0P = ψ∗0P ϕ∗1V = ψ∗1V and τ ◦ ϕ(g) − τ ◦ ψ(g) < σ for all g ∈ H for all τ ∈ T (A), then there exists a unitary u ∈ A such that ku∗xu − yk < ǫ. (e 5.332) (e 5.333) (e 5.334) (e 5.335) Theorem 5.7. Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and with weakly unperforated K0(A). Let x, y ∈ A be two normal elements. Suppose that Dc(x, y) = 0 and [λ − x] = [λ − y] in K1(A) for all λ 6∈ X ∪ Y. Then dist(U (x),U (y)) = 0 (e 5.336) Proof. Since A is a unital simple C ∗-algebra, the assumption of Dc(x, y) = 0 implies that sp(x) = sp(y) = X. Let ϕ, ψ : C(X) → A be the unital monomorphisms induced by x and y, respectively. The assumption implies that ϕ∗1 = ψ∗1. It follows from 2.16 that we also have ϕ∗0 = ψ∗0. Thus the theorem follows from 5.6. 6 Distance between unitary orbits for normal elements with non-zero K1 Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A). 5.7 provides a clue how to described an upper bound for the distance between unitary orbits for normal elements in A. If two normal elements x, y ∈ A have the same spectrum and induce the same homomorphism from K1(C(sp(x))) to K1(A), then an upper bound for the distance between their unitary orbits can be similarly described. When they have different spectrum and with non-trivial K1 information, however, things are very different. This section deals with the case that (λ − x)−1(λ − y) ∈ Inv0(A) for all λ 6∈ sp(x) ∪ sp(y). Note that the assumption allows the case that λ − x 6∈ Inv0(A) for λ ∈ Y \ X and λ − y 6∈ Inv0(A) for λ ∈ X \ Y. This can be done partly because we are able to borrow a Mayer-Vietoris Theorem. Definition 6.1. Let A be a unital simple C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A) and let Ω be a compact metric space. Let F1 and F2 be two finite subsets of Ω. Suppose that κF1, κF2 ∈ Hc,1(C(Ω), A)+ are two elements represented by two homomorphisms whose spectra are F1 and F2, respectively. Suppose also that κF1(fO) and κF2(fO) are projections for all open subsets O ⊂ Ω. Suppose that F1 = {x1, x2, ..., xm} and F2 = {y1, y2, ..., yn}. Suppose that Dc(κF1, κF2) = r. (e 6.337) Then, as proved earlier in 3.4, for any ǫ > 0, there are ai,j ∈ W (A), 1 ≤ i ≤ n and 1 ≤ j ≤ m such that n X j=1 ai,j = κF1([f{xi}]), m X i=1 ai,j = κF2([f{yj}]) and xi − yj ≤ r+ǫ, if ai,j 6= 0 40 (e 6.338) (e 6.339) By a paring of κF1 and κF2 we mean a subset R(κF1, κF2) ⊂ {1, 2, ..., m} × {1, 2, ..., n} of those pairs of (i, j) such that (e 6.338) and (e 6.339) hold. i=1Si and Y = ⊔m2 Definition 6.2. Given a pair of κ1 and κ2 with spectra X and Y, we say that the pair has a hub at X ∩ Y , if X = ⊔m1 k=1Gk, where {S1, S2, ..., Sm1} is a set of mutually disjoint clopen subsets of X and {G1, G2, ..., Gm2} is a set of mutually disjoint clopen subsets of Y, there exists ǫ0 > 0 such that, for any 0 < ǫ < ǫ0, there are finite ǫ-approximations κF1 of κ1 and κF2 of κ2 satisfying the following: There is a pairing R(κF1, κF2) such that, for each pair (t, k) with St ∩ Gk 6= ∅, there is a pair (i, j) ∈ R(κF1, κF2) such that xi, yj ∈ St ∩ Gk. when X = Y are connected. More examples will be presented in 6.10. Obvious examples that any pairs (κ1, κ2) have a hub at X ∩ Y are when X ∩ Y = ∅, or Let x, y ∈ A be two normal elements with X = sp(x) and Y = sp(y). Denote by ϕX : C(X) → A and ϕY : C(Y ) → A the homomorphisms induced by x and y. We say the pair (x, y) has a hub at X ∩ Y, if the pair (ϕX , ϕY ) has a hub at X ∩ Y. Lemma 6.3. Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and with weakly unperforated K0(A), and let x, y ∈ A be two normal elements with X = sp(x) and Y = sp(y). For any ǫ > 0, any finite subset GX ⊂ C(X)s.a. and any finite subset GY ⊂ C(Y )s.a., there exist mutually orthogonal projections {e1, e2, ..., en} ⊂ A with Pn i=1 ei = 1A, λ1, λ2, ..., λn ∈ sp(x) and µ1, µ2, ..., µn ∈ sp(y) such that max{τ ◦ g(x) − τ ◦ g(x1) : g ∈ GX} < ǫ/2, max{τ ◦ g(y) − τ ◦ g(y1) : g ∈ GY } < ǫ/2 for all τ ∈ T (A), c (y, y1) < ǫ/2, c (x, y) + ǫ and kx1 − y1k < Dc(x, y) + ǫ, e(x, x1) < ǫ/2, Dc(y, y1) ≤ De Dc(x, x1) ≤ Dc De c(x1, y1) < De where and x1 = n X i=1 λiei, y1 = n X i=1 µiei 1≤i≤nλi − µi ≤ Dc(x, y) + ǫ max (e 6.340) (e 6.341) (e 6.342) (e 6.343) (e 6.344) (e 6.345) Moreover, if X ∩ Y 6= ∅, for any σ > 0 and η > 0, we may require that m0 m0 x1 = X i=1 λie(i, 0) + x1,1 and y1 = X i=1 λie(i, 0) + y1,1, (e 6.346) τ (e(i, 0)) < σ for all τ ∈ T (A), Dc(x1,1, y1,1) ≤ De c(x, y) + ǫ (e 6.347) m0 X i=1 where {e(1, 0), e(2, 0), ..., e(m0 , 0)} is a set of mutually orthogonal projections, {λ1, λ2, ..., λm0} is η-dense in X ∩ Y, x1,1, y1,1 ∈ (1 − Pm0 i=1 e(i, 0)) are normal elements with finite spectrum in X and Y, respectively, i=1 e(i, 0))A(1 − Pm0 i=1Fj and Y = ⊔m2 In the above, if X = ⊔m1 k=1Gk, where F1, F2, ..., Fm1 are η/2-connected components of X and G1, G2, ..., Gm2 are η/2-connected components of Y, we may assume that {λi} is η-dense in X and {µi} is η-dense in Y. In particular, we may require that {λi}∩ Fj 6= ∅, {µi} ∩ Gk 6= ∅. Moreover, if Fj ∩ Gk 6= ∅, we may further assume that there exist λi(j), µi(k) ∈ Fj ∩ Gk such that λi(j) − µi(k) < η. Furthermore, if the pair (x, y) has a hub at X ∩ Y, then, we may require that, for each pair (j, k) with Fj ∩ Gk 6= ∅, there are λi, µi ∈ Fj ∩ Gk. 41 Proof. The main part of this lemma follows from 4.13. In fact that the existence of x1 and y2 satisfy everything up to (e 6.342) follows immediately from 4.13. By the assumption, Dc(x1, y1) ≤ Dc(x1, x) + Dc(x, y) + Dc(y, y1) < Dc(x, y) + ǫ. Note that sp(x1) ∩ Y, Y ∩ X ⊂ X. It follows from 4.15 that c(x, y). c (x1, x) + De De c (x1, y) ≤ De Similarly Therefore De c (x1, y1) ≤ De c (x1, y) + De c(y, y1). De c(x1, y1) ≤ De c (x, y) + De c(x1, x) + De c (y, y1). (e 6.348) (e 6.349) (e 6.350) replaced by u∗y1u, we may assume (e 6.343) holds. Applying 3.5, there exists a unitary u, such that kx1 − u∗y1uk < Dc(x, y) + ǫ. Let y1 be Further by applying the proof of 3.5, (e 6.344) and (e 6.345) hold. The second part of the statement with (e 6.346) and (e 6.347) follows from the definition of De c (·,·) and 4.11. The third and fourth parts of the statement follow from (e 6.346) and the fact that the finite η-approximations of ψX and ψY can be made for arbitrarily small η. Suppose, in addition, that the pair (x, y) has a hub at X ∩ Y. For each pair (j, k) with Fj ∩ Gk 6= ∅, we may assume that, by choosing sufficiently better finite approximation, without loss of generality, that, there is λi′ ∈ Fj ∩ Gk and there is µi′′ ∈ Fj ∩ Gk. By the assumption that the pair (x, y) has a hub at X ∩ Y and its definition, we may further assume, there are λi, µi ∈ Fj ∩ Gk. Corollary 6.4. Let A be a unital separable simple C ∗-algebra of real rank zero, stable rank one and weakly unperforated K0(A) and let x, y ∈ A be normal elements. Then Dc(x, y) ≤ De Proof. We will prove the second inequality. c(x, y) ≤ 2Dc(x, y). (e 6.351) Put X = sp(x) and Y = sp(y). By 6.3, there are two sequences of normal elements xk, yk ∈ A with lim k→∞ xk = De c (xk, x) = 0, lim k→∞ De c (yk, y) = 0, (e 6.352) m(k) X i=1 λ(k, i)p(k, i) and yk = m(k) X i=1 µ(k, i)p(k, i), (e 6.353) where λ(k, i) ∈ X and µ(k, i) ∈ Y, lim k→∞ max {1≤i≤m(k)} λ(k, i) − µ(k, i) ≤ Dc(x, y) (e 6.354) and {p(k, 1), p(k, 2), ..., p(k, m(k))} is a sequence of mutually orthogonal non-zero projections in A with Pm(k) i=1 p(k, i) = 1A, k = 1, 2, .... Note that lim k→∞ Dc(xk, yk) = Dc(x, y). (e 6.355) 42 We may write xk = m(k) X i=1 λ(k, i)p′(k, i) + m(k) X i=1 λ(k, i)p′′(k, i) and yk = m(k) X i=1 µ(k, i)p′(k, i) + m(k) X i=1 µ(k, i)p′′(k, i), (e 6.356) (e 6.357) where p′(k, i), p′′(k, i) 6= 0 and p′(k, i) + p′′(k, i) = p(k, i), i = 1, 2, ..., m(k), k = 1, 2, .... To simplify the notation, we may write that xk = 2m(k) X i=1 λ(k, i)p(k, i) and yk = 2m(k) X i=1 µ(k, i)p(k, i), (e 6.358) λ(k, m(k) + i) = λ(k, i) and µ(k, m(k) + i) = µ(k, i), i = 1, 2, ..., m(k), (e 6.359) where p(k, i) 6= 0 for all i and k. Without loss of generality, we may assume that Sk = {λ(k, 1), λ(k, 2), ..., λ(k, r(k))} ⊂ X ∩ Y and λ(k, j) 6∈ Y, r(k) < j ≤ 2m(k). Let Tk = {µ(k, α(1)), µ(k, α(2)), ..., µ(k, α(g(k)))} ⊂ X ∩ Y and µ(k, j) 6∈ X if j 6= α(i) (1 ≤ i ≤ g(k)). We may also assume that Sk and Tk both are ǫk-dense in X ∩ Y and limk→∞ ǫk = 0. A standard argument allows us to assume, without loss of generality, that λ(k, i) = µ(k, α(i)), i = 1, 2, ..., f (k), where f (k) ≤ min{r(k), g(k)} and Wk = {λ(k, i) : 1 ≤ i ≤ f (k)} is δk-dense in X ∩ Y and limk→∞ δk = 0. By (e 6.358) and (e 6.359), we may assume that α(i) > f (k), i = 1, 2, ..., f (k). In particular, {α(i) : 1 ≤ i ≤ f (k)} ∩ {1, 2, ..., f (k)} = ∅. (e 6.360) Since A is simple and has (SP), there is a sequence of finite sets of non-zero projections e(k, i) ≤ p(k, i) such that p(k, i) − e(k, i) 6= 0, [e(k, i)] = [e(k, 1)], m(k) i = 1, 2, ..., f (k), (e 6.361) and X i=1 τ (e(k, i)) < 1/k for all τ ∈ T (A), k = 1, 2, .... Let uk ∈ U (A) be a sequence of unitaries such that k = e(k, i), i = 1, 2, ..., f (k), ke(k, i)uk = e(k, α(i)), uke(k, α(i))u∗ u∗ u∗ k(p(k, i) − e(k, i))uk = (p(k, i) − e(k, i)), uk(p(k, α(i)) − e(k, α(i)))u∗ i = 1, 2, ..., f (k) and u∗ kp(k, j)uk = p(k, j), if j 6∈ {i : 1 ≤ i ≤ f (k)}. k = (p(k, α(i)) − e(k, α(i))), (e 6.362) (e 6.363) (e 6.364) (e 6.365) (e 6.366) (e 6.367) (e 6.368) Define x0,k = x1,k = f (k) X i=1 f (k) X i=1 λ(k, i)e(k, i), λ(k, i)(p(k, i) − e(k, i)) + 2m(k) X i=f (k)+1 λ(k, i)pk,i (e 6.369) 43 y0,k = y1,k = f (k) X i=1 f (k) X i=1 µ(k, α(i))e(k, α(i)) (e 6.370) µ(k, i)(p(k, α(i)) − e(k, α(i))) + 2m(k) X i=f (k)+1 µ(k, i)p(k, i). (e 6.371) Note that uky0,ku∗ k = f (k) X i=1 µ(k, α(i))uke(k, α(i))u∗ k = x0,k. Note now that λ(k, i) = µ(k, α(i)), i = 1, 2, ..., f (k), Wk is δk-dense in X ∩ Y and f (k) lim k→∞ sup{τ ( X i=1 e(k, i)) : τ ∈ T (A)} = 0 (e 6.372) Moreover we still have that lim k→∞ Dc(x, x0,k + x1,k) = 0 and lim k→∞ Dc(y, y0,k + y1,k) = 0. (e 6.373) Therefore Note also, by (e 6.360) De c(x, y) ≤ lim inf k→∞ Dc(x1,k, y1,k). (e 6.374) x1,k = f (k) X i=1 λ(k, i)(p(k, i) − e(k, i)) + f (k) X i=1 λ(k, α(i))(p(k, α(i)) − e(k, α(i))) (e 6.375) + f (k) X i=1 λ(k, α(i))e(k, α(i)) + X j6∈{i,α(i):1≤i≤f (k)} λ(k, j)p(k, j) and (e 6.376) y1,k = = f (k) X i=1 f (k) X i=1 µ(k, α(i))(p(k, α(i)) − e(k, α(i))) + X j6∈{α(i):1≤i≤f (k)} µ(k, j)p(k, j) (e 6.377) µ(k, i)(p(k, i) − e(k, i)) + f (k) X i=1 µ(k, α(i))(p(k, α(i)) − e(k, α(i))) (e 6.378) + f (k) X i=1 µ(k, i)e(k, i) + X j6∈{i,α(i):1≤i≤f (k)} µ(k, j)p(k, j). (e 6.379) Since, for i = 1, 2, ..., f (k), [ek,i] = [ek,α(i)], λ(k, i) = µ(k, α(i)) and k→∞ λ(k, α(i)) − µ(k, i) lim sup ≤ lim sup = lim sup k→∞ (λ(k, α(i)) − µ(k, α(i)) + µ(k, α(i)) − µ(k, i)) k→∞ λ(k, α(i)) − µ(k, α(i)) + lim sup k→∞ λ(k, i) − µ(k, i) ≤ Dc(x, y) + Dc(x, y). 44 (e 6.380) (e 6.381) (e 6.382) (e 6.383) It follows that (by comparing the last expresses of x1,k and y1,k above and using (e 6.354)) Therefore, by (e 6.374), Dc(x1,k, y1,k) ≤ 2Dc(x, y). De c(x, y) ≤ 2Dc(x, y). (e 6.384) (e 6.385) The following is a useful observation for the proof of the main results in this section. Lemma 6.5. Let A be a unital C ∗-algebra and let X and Y be two compact subsets of the plane. Suppose that x, y ∈ A are two normal elements with sp(x) = X and sp(y) = Y and suppose that ϕX : C(X) → A and ϕY : C(Y ) → A are induced unital monomorphisms by x and by y, respectively. Suppose also that [λ − x] = [λ − y] in K1(A) for all λ 6∈ X ∪ Y. Then (ϕX ◦ ı1)∗1 = 0, (e 6.386) where I = {f ∈ C(X) : fX∩Y = 0} and ı1 : I → C(X) is the embedding. Consequently, if X ∩ Y = ∅, then I = C(X) and (ϕX )∗1 = 0. In this case (e 6.386) means that Proof. Let πX : C(X ∪ Y ) → C(X) and πY : C(X ∪ Y ) → C(Y ) be quotient maps. Define ψ1 = ϕX ◦ πX and ψ2 = ϕY ◦ πY . The assumption implies that (ψ1)∗1 = (ψ2)∗1. (e 6.387) Put Note that J ∼= C0(X ∪ Y \ Y ). But X ∪ Y \ Y = X \ X ∩ Y. Therefore there is a natural isomorphism h : I → J. Let ı2 : J → C(X ∪ Y ) be the embedding. Then J = {f ∈ C(X ∪ Y ) : fY = 0}. Thus But Therefore (e 6.388) (e 6.389) (e 6.390) (e 6.391) (e 6.392) πX ◦ ı2 ◦ h = ı1. (ϕX )∗1 ◦ (ı1)∗1 = (ϕX )∗1 ◦ (πX)∗1 ◦ (ı2 ◦ h)∗1 = (ψ2)∗1 ◦ (ı2 ◦ h)∗1. ψ2 ◦ ı2 = 0. (ϕX ◦ ı1)∗1 = 0. 45 Lemma 6.6. Let A be a unital infinite dimensional simple C ∗-algebra with real rank zero, stable rank one and with unperforated K0(A), let x ∈ A be a normal element and let ϕX : C(X) → A be the unital monomorphism induced by x, where X = sp(x). Suppose that y ∈ A is another normal element such that sp(y) = Y and [λ − x] = [λ − y] in K1(A) for all λ 6∈ X ∪ Y and suppose that X ∩ Y 6= ∅. Suppose also that functions f1, f2, ..., fn ∈ C(X ∩ Y ) are n mutually orthogonal projections with 1C(X∩Y ) = Pn i=1 fi. Then, for any non-zero projection e ∈ A and any n mutually orthogonal nonzero projections such that Pn i=1 ei = e, there is a normal element x1 ∈ eAe with sp(x1) = for any normal elements x0, y0 ∈ (1 − e)A(1 − e) with finite X ∩ Y satisfying the following: spectrum in X and Y, respectively, fi(x1) = ei, i = 1, 2, ..., n, (ψ1)∗1 = (ϕX )∗1 and (ψ2)∗1 = (ϕY )∗1, (e 6.393) (e 6.394) (e 6.395) where ψ1 : C(X) → A and ψ2 : C(Y ) → A are defined by ψ1(f ) = f (x0 + x1) for all f ∈ C(X) and ψ2(f ) = f (y0 + x1) for all f ∈ C(Y ). Proof. Let πX : C(X ∪ Y ) → C(X), πY : C(X ∪ Y ) → C(Y ), πX X∩Y : C(X) → C(X ∩ Y ) and πY X∩Y : C(Y ) → C(X ∩ Y ) be quotient maps. By the Universal Coefficient Theorem, there are κ1 ∈ KK(C(X), A) such that κ1K1(C(X)) = −(ϕX )∗1 and κ1K0(C(X)) = 0 and κ2 ∈ KK(C(Y ), A) such that κ2K1(C(Y )) = (ϕY )∗1 and κ2K0(C(Y )) = 0. Consider the pull back: πX−→ C(X ∪ Y ) yπY C(Y ) C(X) yπX X∩Y πY X∩Y−→ C(X ∩ Y ). (e 6.396) By a Mayer-Vietoris Theorem (see, for example, 21.5.1 of [1]), one has the following six-term exact sequence: (−[ϕX X∩Y ],[ϕY X∩Y ]) −→ KK(C(X), A) ⊕ KK(C(Y ), A) [ϕX ]+[ϕY ] −→ KK(C(X ∪ Y ), A) (−[ϕX X∩Y ],[ϕY X∩Y ]) ←−   y KK 1(C(X ∩ Y ), A). KK(C(X ∩ Y ), A) x   KK 1(C(X ∪ Y ), A) [ϕX ]+[ϕY ] ←− KK 1(C(X), A) ⊕ KK 1(C(Y ), A) By the assumption and the proof of 6.5, (ϕX )∗1 ◦ (πX)∗1 = (ϕX ◦ πX)∗1 = (ϕY ◦ πY )∗1 = (ϕY )∗1 ◦ (πY )∗1. It follows that ([ϕX X∩Y ] + [ϕY X∩Y ])(κ1, κ2) = 0. (e 6.397) The exactness of the Mayer-Vietoris sequence above shows that there is κ3 ∈ KK(C(X ∩ Y ), A) such that or (−[ϕX ], [ϕY ])(κ3) = (κ1, κ2), − [ϕX X∩Y ](κ3) = κ1 and [ϕY X∩Y ](κ3) = κ2. 46 (e 6.398) (e 6.399) Let κ3K1(C(X∩Y )) = λ be as an element in Hom(K1(C(X ∩ Y ), K1(A)). Then (e 6.399) implies that λ ◦ (ϕX X∩Y )∗1 = (ϕX )∗1 and λ ◦ (ϕY X∩Y )∗1 = (ϕY )∗1 (e 6.400) It follows from 5.3 that there is a unital monomorphism ψ′ 1 : C(X ∩ Y ) → eAe such that (ψ′ 1)∗1 = λ and ψ′ 1(fi) = ei, i = 1, 2, ..., n. (e 6.401) Let zX∩Y ∈ C(X ∩ Y ) be the identity function on X ∩ Y. Choose x1 ∈ eAe such that x1 = 1(zX∩Y ). Choose any normal element x0 ∈ B ⊂ eAe, where B is a finite dimensional C ∗- ψ′ subalgebra of eAe with 1B = e such that sp(x0) ⊂ X. Define ψ′′ : C(X) → (1 − e)A(1 − e) by ψ′′(f ) = f (x1) for all f ∈ C(X) and define ψ1 : C(X) → A by ψ1(f ) = f (x0 + x1). Then, since x0 ∈ B, ψ′′(f ) ∈ B for all f ∈ B. It follows that (ψ′′)∗1 = 0. Therefore, by (e 6.400), (e 6.402) X∩Y )∗1 = (ϕX )∗1. (ψ1)∗1 = λ ◦ (πX Define ψ2 : C(Y ) → A by ϕ2(g) = g(x1 + y0) for all g ∈ C(Y ) and for any normal element y0 ∈ B with sp(y0) ⊂ Y. We also have (ψ2)∗1 = λ ◦ (πY X∩Y )∗1 = (ϕY )∗1. (e 6.403) Let A be a unital simple C ∗-algebra with T (A) 6= ∅. Let ϕX : C(X) → A be a unital monomorphism. Denote by ϕ♯ X : C(X)s.a. → Aff(T (A)) the unital affine continuous map induced by ϕ. If s : C(X) → C is a state of C(X), denote by µs the probability Borel measure induced by s. Theorem 6.7. Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and with weakly unperforated K0(A). Let x, y ∈ A be two normal elements with sp(x) = X and sp(y) = Y. Denote Z = X ∪ Y. Suppose that [λ − x] = [λ − y] in K1(A) for all λ 6∈ Z. (1) Then dist(U (x),U (y)) ≤ De c (x, y). (2) Moreover, if the pair (x, y) has a hub at X ∩ Y, then dist(U (x),U (y)) ≤ Dc(x, y). (e 6.404) (e 6.405) Proof. Denote by ϕX : C(X) → A the unital monomorphism defined by ϕX(f ) = f (x) for all f ∈ C(X) and ϕY : C(Y ) → A defined by ϕY (f ) = f (y) for all f ∈ C(Y ). Let ǫ > 0. Let V1 ⊂ K1(C(X)) (in place of V) be a finite subset, P1 ⊂ K0(C(X)) (in place of P) be a finite subset, H1 ⊂ C(X)s.a. (in place of H) be a finite subset and let σ1 > 0 (in place of σ) be required by 5.6 for ǫ/16 and x. Let V2 ⊂ K1(C(Y )) (in place of V) be a finite subset, P2 ⊂ K0(C(Y )) (in place of P) be a finite subset, H2 ⊂ C(Y )s.a. (in place of H) be a finite subset and let σ2 > 0 (in place of σ) be required by 5.6 for ǫ/16 and y. Without loss of generality, we may assume that H1 and H2 are in the unit balls of C(X) and C(Y ), respectively. Moreover, we may assume that P1 = {f1, f2, ..., fm1} and P2 = {g1, g2, ..., gm2}, (e 6.406) 47 where fi ∈ C(X) and gi ∈ C(X) which are projections such that 1C(X) = m1 X i=1 fi and 1C(Y ) = m2 X i=1 gi. (e 6.407) Denote by F1, F2, ..., Fm1 the clopen sets of X corresponding to projections f1, f2, ..., fm1 , and by G1, G2, ..., Gm2 the clopen subsets of Y corresponding to projections g1, g2, ..., gm2 . Let X∩Y = ⊔I j=1Sj, where S1, S2, ..., SI are distinct ǫ/8-connected components of X ∩ Y. In Let η > 0 be such that η < ǫ. By applying 6.3, there are mutually orthogonal projections particular, dist(SI , Sj) ≥ ǫ/8. If X ∩ Y = ∅, this notation simply means I = 0 (see 6.5). e1, e2, ..., en ∈ A with Pn i=1 ei = 1A, λ1, λ2, ..., λn ∈ X and µ1, µ2, ..., µn ∈ Y such that max{τ (g(x)) − τ (g(x1)) : g ∈ H1} < σ1/2 for all τ ∈ T (A), max{τ (g(y)) − τ (g(y1)) : g ∈ H2} < σ2/2 for all τ ∈ T (A), Dc(x, x1) ≤ De Dc(y, y1) ≤ De De c (x1, y1) < De 1≤i≤nλi − µi < Dc(x, y) + η/4, max c(x, x1) < η/4, c (y, y1) < η/4, c (x, y) + η/4 and kx1 − y1k < Dc(x, y) + η/4 where x1 = n X i=1 λiei and y1 = n X i=1 µiei. By the proof of 6.3 (by choosing even smaller δ) and by 2.16, we may assume that [fi(x1)] = [fi(x)], i = 1, 2, ..., m1 and [gj(y1)] = [gj(y)] j = 1, 2, ..., m2 (e 6.408) (e 6.409) (e 6.410) (e 6.411) (e 6.412) (e 6.413) (e 6.414) (e 6.415) (e 6.416) We first consider case (1). Since the case that X ∩ Y = ∅ will be dealt with in case (2), we will assume that X ∩ Y 6= ∅. By the second part of 6.3, we may also assume that, x1 = I X i=1 λie(0) i + x′ 2, y2 = I X i=1 λie(0) i + y′ 2, Dc(x′ 2, y′ 2) < De c (x, y) + ǫ/4, τ ( I X i=1 e(0) i ) < min{σ1/2, σ2/2} (e 6.417) (e 6.418) 2 , ..., e(0) for all τ ∈ T (A), where {e(0) 2, y′ λi ∈ Si, i = 1, 2, ..., I, x′ and Y, respectively, and where p0 = PI 1 , e(0) I } is a set of mutually orthogonal non-zero projections, 2 ∈ (1 − p0)A(1 − p0) are normal elements with finite spectrum in X i=1 e(0) Let hj = χSj ∈ C(X ∩ Y ), j = 1, 2, ..., I. By applying 6.6, there is a normal element x0 ∈ p0Ap0 with sp(x0) = X ∩ Y such that . i (ψ1)∗1V1 = (ϕX )∗1V1, (ψ2)∗1V2 = (ϕY )∗1V2, hj(x0) = e(0) j , j = 1, 2, ..., I, (e 6.419) (e 6.420) (e 6.421) where ψ1 : C(X) → p0Ap0 is defined by ψ1(f ) = f (x0) for all f ∈ C(X) and ψ2 : C(Y ) → p0Ap0 is defined by ψ2(f ) = f (x0) for all f ∈ C(Y ). Now consider x3 = x0 + x′ 2. Note 2 and y3 = x0 + y′ 48 that sp(x3) ⊂ X and sp(y3) ⊂ Y. Define ψ3 : C(X) → A by ψ3(f ) = f (x3) for all f ∈ C(X) and ψ4 : C(Y ) → A by ψ4(f ) = f (y3) for all f ∈ C(Y ). Since x′ 2 have finite spectra, by (e 6.419) and (e 6.420), we have 2 and y′ (ϕX )∗1V1 = (ψ3)∗1V1 and (ϕY )∗1V2 = (ψ4)∗1V2. For each i, if Fi ∩ Y = ∅, i.e., Fi ∩ Gk = ∅ for all k, we compute that 2)] = [fi(x1)] = [fi(x)] [ψ3(fi)] = [fi(x3)] = [fi(x′ (e 6.422) (e 6.423) i = 1, 2, ..., I. If Fi ∩ Y 6= ∅, let Hi be the subset of {j : j = 1, 2, ..., I} such that hj ≤ fi. We then have [ψ3(fi)] = X j∈Hi [e(0) j ] + [fi(x′ 2)] = [fi(x1)] = [fi(x)]. (e 6.424) Similarly, if Gi ∩ X = ∅, [ψ4(gi)] = [gi(y3)] = [gi(y′ 2)] = [gi(y1)] = [gi(y)]. (e 6.425) If Gi ∩ X 6= ∅, let H ′ j∈H ′ i i be the subset of {j : j = 1, 2, ..., I} such that hj ≤ gi. We then have [ψ4(gi)] = X 2)] = [gi(y1)] = [gi(y)]. ] + [gi(y′ [e(0) j (e 6.426) In other words, (ψ3)∗0P1 = (ϕX )∗0P1 and (ψ4)∗0P2 = (ϕY )∗0P2 . (e 6.427) By applying 5.6, using (e 6.422), (e 6.427), (e 6.408), (e 6.409) and (e 6.418), we obtain a unitary u1, u2 ∈ A such that ku∗ 1xu1 − x3k < ǫ/16 and ku∗ 2yu2 − y3k < ǫ/16. By (e 6.418) and by 3.5 there is a unitary u3 ∈ (1 − p0)A(1 − p0) such that 3x′ ku∗ 2u3 − y′ 2k < De c (x, y) + ǫ/16. Put u4 = p0 + u3. Then ku∗ 3x′ 4x3u4 − y3k = k(x0 + u∗ 2u3) − (x0 + y′ 3x′ 2)k = ku∗ 2u3 − y′ 2k < De c (x, y) + ǫ/16. Therefore dist(U (x),U (y)) < De c (x, y) + 5ǫ/8. (e 6.428) (e 6.429) (e 6.430) (e 6.431) (e 6.432) This proves the case (1). Now we turn to case (2). If X ∩ Y = ∅, by the assumption that [λ − x] and [λ − y] are the same in K1(A) for all λ 6∈ X ∪ Y, λ − x ∈ Inv0(A) for all λ 6∈ X and λ − y ∈ Inv0(A) for all λ 6∈ Y. Thus this special case has been proved in 3.6. Thus we will then assume again X ∩ Y 6= ∅. Some of the argument above will be repeated. In this case, we may assume that, if Fj ∩ Gk 6= ∅, there are at least one i such that λi, µi ∈ Fj ∩ Gk. 49 Note that, in this case, Fj ∩ Gk is a non-empty clopen subset of X ∩ Y. In fact X ∩ Y is a disjoint union of those Fj ∩ Gk. Call them T1, T2, ..., Tk. Then k ≤ n. We may assume that {r(1), r(2), ..., r(k)} ⊂ {1, 2, ..., n} such that λr(j), µr(j) ∈ Tj, j = 1, 2, ..., k. Since A is simple and infinite dimensional, one can find a non-zero projection e(0) r(j) ≤ er(j) such that [e(0) r(j)] = [e(0) r(1)] in K0(A), j = 1, 2, ..., rk, and k τ ( X j=1 e(0) r(j)) < min{σ1/2, σ2/2} for all τ ∈ T (A). (e 6.433) j=1 e(0) Let p0 = Pk i 6∈ {r(1), r(2), ..., r(k)}. Put r(j), p = 1A − p0, pi = ei − e(0) i , if i ∈ {r(1), r(2), ..., r(k)} and pi = ei, if x2 = n X i=1 λipi and y2 = n X i=1 µipi. Note, by (e 6.413), that kx2 − y2k ≤ max 1≤i≤nλi − µi < Dc(x, y) + η/4. (e 6.434) (e 6.435) Let h′ j = χTj , j = 1, 2, ..., k. By applying 6.6, there is a normal element x0 ∈ p0Ap0 with sp(x0) = X ∩ Y such that (ψ1)∗1V1 = (ϕX )∗1V1, (ψ2)∗1V2 = (ϕY )∗1V2, i(x0) = e(0) h′ r(i), i = 1, 2, ..., k, (e 6.436) (e 6.437) (e 6.438) where ψ1 : C(X) → p0Ap0 is defined by ψ1(f ) = f (x0) for all f ∈ C(X) and ψ2 : C(Y ) → p0Ap0 is defined by ψ2(f ) = f (x0) for all f ∈ C(Y ). Now consider x3 = x0 + x2 and y3 = x0 + y2. Note that sp(x3) ⊂ X and sp(y3) ⊂ Y. Define ψ3 : C(X) → A by ψ3(f ) = f (x3) for all f ∈ C(X) and ψ4 : C(Y ) → A by ψ4(f ) = f (y3) for all f ∈ C(Y ). Since x2 and y2 have finite spectra, by (e 6.419), we have (ϕX )∗1V1 = (ψ3)∗1V1 and (ϕY )∗1V2 = (ψ4)∗1V2 For each i, if Fi ∩ Y = ∅, i.e., i 6∈ {r(1), r(2), ..., r(k)}, we compute that [ψ3(fi)] = [ψ1(fi)] = [ϕX (fi)]. (e 6.439) (e 6.440) If Fi ∩ Y 6= ∅, we also have [er(j) − e(0) r(j)] + X λj ∈Fi,j6=r(j) [ei]) (e 6.441) (e 6.442) (e 6.443) [ψ3(fi)] = X h′ j≤fi = [ X λi∈Fi [e(0) r(j)] + ( X h′ j ≤fi ej] = [ϕX (fi)], j = 1, 2, ..., k. In other words, (ψ3)∗0P1 = (ϕX )∗0P1. 50 By applying 5.6, using (e 6.443), (e 6.439) and (e 6.408) and (e 6.433), we obtain a unitary such that On the other hand, for each j, ψ2(gj) = Ph′ It follows that ku∗xu − x3k < ǫ/16. h′ i(x0). i≤gj [ψ2(gj)] = X µr(i)∈Gj [e(0) r(i)]. Therefore, [ψ4(gj )] = [(ψ2)(gj )] + [gj(y2)] [e(0) r(i)] + X = X µr(i)∈Gj µr(j)∈Gj [er(i) − e(0) r(i)] + (e 6.444) (e 6.445) (e 6.446) [es] (e 6.447) X µs∈Gj ,s6∈{r(i):1≤i≤r(k)} = X µs∈Gj [es] = [ϕY (gj)]. It follows that (ψ4)∗0P2 = (ϕY )∗0P2. (e 6.448) (e 6.449) It follows from (e 6.449), (e 6.439), (e 6.409), (e 6.433) and 5.6 that there is a unitary v ∈ A such that kv∗yv − y3k < ǫ/16. (e 6.450) We also have (using (e 6.435)) kx3 − y3k = k(x0 + x2) − (x0 + y2)k = kx2 − y2k < Dc(x, y) + ǫ/4. (e 6.451) It follows that ku∗xu − v∗yvk < ǫ/16 + Dc(x, y) + ǫ/4 + ǫ/16 < Dc(x, y) + ǫ/2. (e 6.452) Therefore dist(U (x),U (y)) < Dc(x, y) + ǫ/2 (e 6.453) for all ǫ > 0. The theorem follows. Remark 6.8. It is probably helpful to be reminded that and if both X and Y are connected, De c(x, y) = Dc(x, y). De c (x, y) ≤ min{DT (x, y), 2Dc(x, y)}. Corollary 6.9. Let A be a unital separable simple infinite dimensional C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A) and let x ∈ A be a normal element with sp(x) = X. Then, for any ǫ > 0, any σ > 0 and any finite subset {ξ1, ξ2, ..., ξk} ⊂ sp(x), there 51 is a set of mutually orthogonal non-zero projections {e1, e2, ..., ek} of A and a normal element x0 ∈ (1 − p)A(1 − p) with sp(x0) = X such that kx − (x0 + k X i=1 ξiei)k < ǫ, ei) < σ for all τ ∈ T (A), (ϕ1)∗1 = (ϕ2)∗1, (e 6.454) (e 6.455) (e 6.456) k τ ( X i=1 i=1 f (ξi)ei for all f ∈ C(X). i=1 ei, ϕ1, ϕ2 : C(X) → A is defined by ϕ1(f ) = f (x) and ϕ2(f ) = f (x0) + where p = Pk Pk Proof. This is merely a refinement of that of 5.4. The issue is that we now insist that sp(x0) = X. The proof is contained in the proof of the case (1) in the proof of 6.7 by letting X = Y . With the notation in the proof of the case (1) in 6.7, we have x3 = x0 + x′ 2 has finite spectrum but sp(x′ 2) can be ǫ/16-dense in X. Note that sp(x0) = X. We may assume, without loss of generality, that x′ 2, where {p1, p2, ..., pk} is a set of mutually orthogonal non-zero projections and where x′′ i=1 pi) with sp(x′′ i ≤ pi, i = 1, 2, ..., k, such that 2) ⊂ X. Since A is simple and has the property (SP), there are non-zero projections e′ 2 is a normal element in (1 − Pk i=1 pi)A(1 − Pk i=1 ξkpi + x′′ 2, where x′ 2 = Pk k X i=1 τ (e′ i) < σ for all τ ∈ T (A). We still have, as in (e 6.428), a unitary u1 ∈ A such that 1xu1 − x3k < ǫ/16 ku∗ Then we have kx − (u1(x0 + x′′ 2)u∗ 1 + k X i=1 ξiu1e′ iu∗ 1)k < ǫ/16. (e 6.457) (e 6.458) (e 6.459) Choose the new x0 to be u1(x0 + x′′ 2)u∗ 1 and ei to be u1e′ iu∗ 1. Corollary 6.10. Let A be a unital separable simple C ∗-algebra of real rank zero, stable rank one and weakly unperforated K0(A), let x, y ∈ A be two normal elements with sp(x) = X and sp(y) = Y. Then the pair (x, y) has hub at X ∩ Y, if one of the following holds: (1) X = Y is connected; (2) X ∩ Y is connected and it contains an open ball with radius Dc(x, y); (3) for every connected component S of X, either S = X ∩ Y or dist(ξ, X ∩ Y ) > Dc(x, y) (4) X ∩ Y = ∅. Consequently, for all ξ ∈ S; if one of the above conditions holds. dist(U (x),U (y)) ≤ Dc(x, y), 52 Proof. It is clear that (1) and (4) follow from the definition immediately. It is also clear that (3) holds since X ∩ Y must be connected and no point in X ∩ Y can pair with any point outside X ∩ Y with a distance no more than Dc(x, y). i=1Fi and Y = ⊔m2 j=1Gj, where {F1, F2, ..., Fm1} and {G1, G2, ..., Gm2} are of mutually disjoint clopen sets. Since X ∩ Y is connected, we may assume that X ∩ Y = F1 ∩ G1, and Fi ∩ Gj = ∅, if (i, j) 6= (1, 1). To see (2), let X = ⊔m1 Let d = min{dist(G1, Gj ) : j = 2, 3, ..., m2} > 0. Choose ǫ0 = d/16. Let 0 < ǫ < ǫ0. Suppose that F = {λ1, λ2, ..., λK} ⊂ X and G = {µ1, µ2, ..., µL} ⊂ Y are finite subsets, x1 = PK j=1 µjpj, where {e1, e2, ..., eK} and {p1, p2, ..., pL} are two sets of mutually orthogonal projections in A such that PK i=1 ei = 1A = PL i=1 λkei and y1 = PL j=1 pj, and such that Dc(x1, x) < ǫ and Dc(y1, y) < ǫ. (e 6.460) Therefore Dc(x, y) − 2ǫ < Dc(x1, y1) < Dc(x, y) + 2ǫ. (e 6.461) Then there is λi ∈ F1 ∩ G1 = X ∩ Y such that dist(λi, Gj) > (Dc(x, y) − ǫ) + d > Dc(x1, y1) + d − 3ǫ > Dc(x, y). for all j 6= 1. In other words there is j such that µj ∈ F1 ∩ G1 and (i, j) ∈ Rx1,y2 (see 6.1). Therefore the pair (x, y) has a hub at X ∩ Y. Corollary 6.11. Let A be a unital separable simple C ∗-algebra of real rank zero, stable rank one and weakly unperforated K0(A), let x, y ∈ A be two normal elements with sp(x) = X and sp(y) = Y. If X or Y is connected and if [λ − x] = [λ − y] in K1(A) for all λ 6∈ X ∪ Y , then Proof. If X or Y is connected, by 4.10, Dc(x, y) = De c(x, y). The corollary follows from 6.7. dist(U (x),U (y)) ≤ Dc(x, y). 7 Distance between unitary orbits of normal elements with dif- ferent K1 maps In this section we will show that, without the condition that [λ − x] = [λ − y] in K1(A) for λ 6∈ X ∪ Y in the statement of 6.7, Dc(x, y) alone may have little to do with the distance of the unitary orbits of x and y as 7.2 shows. However, 7.3 provides us some description of the upper as well as lower bound for the distance between unitary orbits of normal elements. Let A be a unital C ∗-algebra and let x, y ∈ A be two normal elements. Let X = sp(x) and Y = sp(y). Denote by dH(X, Y ) the Hausdorff distance between the subset X and Y. Define ρ(x, y) = max{dH (X, Y ), ρ1(x, y)}, (e 7.462) where ρ1(x, y) = sup{dist(λ, X) + dist(λ, Y ) : λ 6∈ X ∪ Y, (λ − x)(λ − y)−1 6∈ Inv0(A)}. (e 7.463) 53 Let ρx(x, y) = sup{dist(λ, X) : λ 6∈ X ∪ Y, (λ − x)−1(λ − y) 6∈ Inv0(A)} and (e 7.464) ρy(x, y) = sup{dist(λ, Y ) : λ 6∈ X ∪ Y, (λ − x)−1(λ − y) 6∈ Inv0(A)}. (e 7.465) The following is a result of Ken Davidson ([5]). The proof is exact the same as that in [5]. Proposition 7.1. Let A be a unital C ∗-algebra and let x, y ∈ A be two normal elements. Then (e 7.466) dist(U (x),U (y)) ≥ ρ(x, y). Theorem 7.2. Let A be a unital, infinite dimensional, separable simple C ∗-algebra with real rank zero, stable rank one, weakly unperforated K0(A) and K1(A) 6= {0}. Then that [u1] 6= [u2] in K1(A), (1) for any unitary u1 ∈ A with sp(u1) = T (the unit circle), there is a unitary u2 ∈ A such Dc(u1, u2) = 0 and dist(U (u1),U (u2)) = 2; (e 7.467) (2) For any compact subset X ⊂ C such that C \ X is not connected and for any normal element x ∈ A with sp(x) = X, there exists a normal element y ∈ A such that dist(U (x),U (y)) ≥ 2 sup{dist(λ, sp(x)) : λ in bounded components of C \ sp(x)} and Dc(x, y) = 0. (e 7.468) (e 7.469) (e 7.470) Proof. It is clear that (1) follows from (2). So we will prove (2). As in the beginning of the proof of 5.3, there is a unital simple AH-algebra B with slow dimension growth and with real rank zero such that (K0(B), K0(B)+, [1B], K1(B)) = (K0(A), K0(A)+, [1A], K1(A)) and since ρB(K0(B)) and ρA(K0(A)) are dense in Aff(T (B)) and Aff(T (A)), respectively, the above also gives an affine homeomorphism from Aff(T (B)) to Aff(T (A)) which is compatible with the above identification. We will use this fact in the proof of (2). Let x ∈ A be a normal element with sp(x) = X. Put d = sup{dist(λ, sp(x)) : λ in bounded components of C \ sp(x)}. Let S be the union of all bounded components of C \ sp(x). Then sup{λ : λ ∈ S} ≤ kxk. In particular, d ≤ kxk. There is λ0 ∈ S such that dist(λ0, sp(x)) = d0 > 0. So d ≥ d0. The set (e 7.471) S1 = {ξ ∈ S : kxk ≥ dist(ξ, sp(x)) ≥ d0} is compact. It follows that there is λ ∈ S1 such that dist(λ, sp(x)) = d. (e 7.472) Let ϕ1 : C(X) → A be the unital monomorphism defined by ϕ(f ) = f (x) for all f ∈ C(X). By [22], since K1(A) 6= {0}, there exists a unital monomorphism ψ : C(X) → B ⊂ A such that ψ∗0 = ϕ∗0, τ ◦ ψ = τ ◦ ϕ for all τ ∈ T (B) = T (A) and [λ − ψ(z)] 6= [λ − x] in K1(A), where z : X → X is the identity function. 54 Let y ∈ U (ψ(z)). It follows from 4.6 and 2.21 that Dc(x, y) = 0. Since (λ − x)−1(λ − ψ(z)) 6∈ Inv0(A), by 7.1, dist(U (x),U (y)) ≥ 2dist(λ, sp(x)) = 2d. (e 7.473) (e 7.474) Please note that the above 7.2 does not follow from the following theorem. Theorem 7.3. Let A be a unital separable simple C ∗-algebra with real rank zero, stable rank one and weakly unperforated K0(A) and let x, y ∈ A be two normal elements. Then ρ(x, y) ≤ dist(U (x),U (y)) ≤ min{D1, D2}, (e 7.475) where D1 = max{DT (x, y), max{ρx(x, y), ρy(x, y)}} + min{ρx(x, y), ρy(x, y)}, D2 = De c (x, y) + 2 min{ρx(x, y), ρy(x, y)}. (e 7.476) (e 7.477) Proof. Let X = sp(x) and Y = sp(y). Let d = DT (x, y), d/2 > ǫ > 0 and let S = {λ ∈ C : λ 6∈ X ∪ Y, (λ − x)−1(λ − y) 6∈ Inv0(A)}. (e 7.478) Note that the closure ¯S of S is compact. Let ξ1, ξ2, ..., ξL′ be a finite subset of X ∪ Y ∪ ¯S such that it is ǫ/32-dense in X ∪ Y ∪ ¯S. Let N1, N2, ..., NL be all possible finite unions of O(ξi, ǫ/32)′s such that Nj ∩ Y 6= Y for j = 1, 2, ..., L. For each i, let ηi = inf{dτ (f(Ni)d+ǫ/32 (x)) − dτ (fNi(y)) : τ ∈ T (A)}. It follows from 2.20 that ηi > 0, i = 1, 2, ..., L. Choose 0 < η < min{ǫ/4, min{ηi : 1 ≤ i ≤ L}/8}. (e 7.479) (e 7.480) gi(t) = 1 if t ∈ (Ni)d+ǫ/8, g(t) = 0 if t 6∈ (Ni)d+ǫ/4, i = 1, 2, ..., L. Let δ > 0 with δ < min{ǫ/210, η/16}. Let gi ∈ C(X ∪ Y ∪ ¯S) be such that 0 ≤ gi(t) ≤ 1, There are distinct points ζ1, ζ2, ..., ζK ∈ ¯S such that i=1 O(ζi, δ/4) ⊃ ¯S. ∪K Let S1, S2, ..., SK be compact subsets of ¯S such that (e 7.481) ζi ∈ Si and diam(Si) < δ/2, i = 1, 2, ..., K. Since sp(x) is compact, there are λ1, λ2, ..., λK ∈ sp(x) such that dist(λi, ζi) = dist(sp(x), ζi), (e 7.482) (e 7.483) i = 1, 2, ..., K. Let H = {z, gi : 1 ≤ i ≤ L}, where z represents the identity function on X ∪ Y ∪ ¯S. 55 By 6.9 and 4.11, there are nonzero mutually orthogonal projections {e1, e2, ..., eK , eK+1, ..., ek}, a unitary w in A and normal elements x0 ∈ (1 − Q1)A(1 − Q1) with sp(x0) = X and y0 ∈ (1 − Q2)A(1 − Q2) with sp(y0) = Y satisfy the following: kf (x) − (f (x0) + k X i=1 f (λi)ei)k < δ/4 for all f ∈ H, k f (λi)ei)wk < δ/4 for all f ∈ H X kf (y) − w∗(f (y0) + [λ − x] = [λ − x1] for all λ 6∈ X and τ ( ei) < η/2, i=K+1 k Dc(x0 + K X i=1 λiei, y0) < De c (x, y) + δ/4 X i=1 (e 7.484) (e 7.485) (e 7.486) (e 7.487) (e 7.488) i=K+1 ei, x1 = x0 + Pk for all τ ∈ T (A), where Q1 = Pk i=1 λiei. and {λK+1, λK+2, ..., λk} is δ/4-dense in X ∩ Y. As in the proof of 5.3, there are normal elements hi ∈ eiAei such that sp(hi) = Si, λ − hi ∈ Inv0(eiAei) for all λ 6∈ Si, i = 1, 2, ..., K. i=1 ei and Q2 = Pk Define x2 = x0 + K X i=1 hi + k X i=K+1 λiei It follows that kx − x2k ≤ kx − x1k + kx1 − x2k X X K K λiei − < δ/4 + k ≤ δ/4 + max{kλiei − hik : 1 ≤ i ≤ K} < δ/4 + ρx(x, y). hik i=1 i=1 (e 7.489) (e 7.490) (e 7.491) (e 7.492) (e 7.493) Let Z = X ∪ ¯S. Define ψ : C(Ω) → A by ψ(f ) = f (x2) for all f ∈ C(Ω) and define ψY : C(Ω) → A by ψY (g) = g(y) for g ∈ C(Ω). Let λ 6∈ Z ∪ Y = X ∪ Y ∪ ¯S. By the assumption and (e 7.485), Since δ < η/16, by (e 7.484), [λ − x2] = [λ − y] for all λ 6∈ Z ∪ Y. τ (fi(x1)) > τ (fi(x)) − η/16 for all τ ∈ T (A), i = 1, 2, ..., L. Let d1 = dH (sp(x2), Y ) = max{dH (X, Y ), ρy(x, y)}. (e 7.494) (e 7.495) (e 7.496) Let O ⊂ Y ∪ X ∪ ¯B be an open subset with O ∩ Y 6= Y. If Oǫ/2 ∩ Y = Y, since A is simple, dτ (ψY (fO)) < dτ (ψY (fOǫ/2)) for all τ ∈ T (A). (e 7.497) 56 But we also have that Od1+ǫ ∩ Z⊃ sp(x2). Then ψ(fOd1+ǫ) = 1A. It follows that dτ (ψY (fO)) < dτ (ψY (fOǫ/2)) = dτ (ψ(fOd1+ǫ)). (e 7.498) If Oǫ/2 ∩ Y 6= Y, let Oǫ/32 ∩{ξ1, ξ2, ..., ξL′} = ξk1, ξk2, ..., ξkl. Then O ⊂ ∪l It follows that there is j such that j=1O(ξkj , ǫ/32) ⊂ Oǫ/16. By (e 7.487), (e 7.495), (e 7.499) and (e 7.479), we have O ⊂ Nj ⊂ (Nj)d+ǫ/16 ⊂ Od+ǫ/8. dτ (ψ(fOd+ǫ)) − dτ (ψY (fO)) > dτ (fOd+ǫ(x0)) − dτ (ψY (fNj )) − η/2 ≥ τ (fOd+ǫ(x1)) − dτ (ψY (fNj )) − η/2 − η/2 ≥ τ (gj (x1)) − dτ (ψY (fNj )) − δ/4 − η > τ (gj (x)) − dτ (ψY (fNj )) − η/16 − δ/4 − η ≥ dτ (f(Nj )d+ǫ/16(x)) − dτ (ψY (fNj )) − 17η/16 − δ/4 ≥ ηj − 17η/16 − η/64 > 0 for all τ ∈ T (A). By (e 7.500)-(e 7.505) and (e 7.496) DT (x2, y) ≤ max{DT (x, y) + ǫ, ρy(x, y)}. It follows from this, (e 7.494), 4.9 and 6.7 that dist(U (x2),U (y)) ≤ max{DT (x, y) + ǫ, ρy(x, y)}. Combining this with (e 7.490) and (e 7.493), we have (δ < ǫ/210) dist(U (x),U (y)) ≤ max{DT (x, y) + ǫ, ρy(x, y)} + ρx(x, y) + ǫ/212 for all ǫ > 0. Therefore dist(U (x),U (y)) ≤ max{DT (x, y), ρy(x, y)} + ρx(x, y). Since we may switch the position of x and y, we conclude that dist(U (x),U (y)) ≤ max{DT (x, y), max{ρx(x, y), ρy(x, y)}} + min{ρx(x, y), ρy(x, y)}. On the hand, we have De c (x2, y) ≤ Dc(x0 + K X i=1 hi, y0) ≤ ρx(x, y) + Dc(x0 + λiei, y0) ≤ ρx(x, y) + De c (x, y) + δ/4. It follows from 6.7 that dist(U (x2),U (y)) ≤ ρx(x, y) + De c(x, y) + δ/4. By (e 7.493), dist(U (x),U (y)) ≤ De c (x, y) + 2ρx(x, y) + δ/4. Since we can exchange x with y in the above proof, finally, we conclude that dist(U (x),U (y)) ≤ De c (x, y) + 2 min{ρx(x, y), ρy(x, y)}. (e 7.499) (e 7.500) (e 7.501) (e 7.502) (e 7.503) (e 7.504) (e 7.505) (e 7.506) (e 7.507) (e 7.508) (e 7.509) (e 7.510) (e 7.511) (e 7.512) (e 7.513) (e 7.514) (e 7.515) 57 In some special case below exact formula for distance can be stated. Corollary 7.4. Let A be a unital separable simple C ∗-algebra of real rank zero, stable rank one and weakly unperforated K0(A) and let x, y ∈ A be two normal elements. If Dc(x, y) = 0, then (e 7.516) dist(U (x),U (y)) = ρ1(x, y). If X = Y and X is connected, then dist(U (x),U (y)) ≤ max{Dc(x, y), (1/2)ρ1(x, y)} + (1/2)ρ1(x, y). (e 7.517) Proof. In this case sp(x) = sp(y) and dH (X, Y ) = 0. Therefore ρx(x, y) = ρy(x, y) and ρ1(x, y) = ρx(x, y) + ρy(x, y) = 2ρx(x, y). In case that X is connected, by 2.21, Dc(x, y) = DT (x, y). Thus the corollary follows from 7.3. 8 Lower bound Last section gives both upper bound and lower bound for the distance between unitary orbits of normal elements. However, the lower bound are all given by the bounded components of C \ X ∪ Y which give different K1-information of the corresponding normal elements. In this section, we will discuss the lower bound for distance between unitary orbits of normal elements who have the same K1-information outside of X ∪ Y. Theorem 8.1. There exists a constant C > 0 satisfying the following: Let A be a unital separable AF-algebra and let x, y ∈ A be two normal elements. Then C · Dc(x, y) ≤ dist(U (x),U (y)) ≤ Dc(x, y). (e 8.518) Proof. Let C = c−1 be in the statement of Theorem 4.2 of [6]. Without loss of generality, we may assume that kxk,kyk ≤ 1. It follows from 3.6 that it suffices to show that We will show that dist(U (x),U (y)) ≥ C · Dc(x, y). kx − yk ≥ C · Dc(x, y), (e 8.519) (e 8.520) Put d = Dc(x, y). Let ǫ > 0. It follows from [17] that there are λ1, λ2, ..., λn ∈ sp(x), µ1, µ2, ..., µm ∈ sp(y), two sets of mutually orthogonal non-zero projections {p1, p2, ..., pn} and {q1, q2, ..., qm} in A such that kx − n X i=1 λipik < ǫ/16 and ky − m X j=1 µjqjk < ǫ/16. (e 8.521) Put x1 = Pn i=1 λipi and y1 = Pm may also assume that j=1 µjqj. Without loss of generality, by the virtue of 2.17, we Since Dc(·,·) is a metric, Dc(x, x1) < ǫ/16 and Dc(y, y1) < ǫ/16. (e 8.522) Dc(x1, y1) ≥ Dc(x, y) − ǫ/8. (e 8.523) 58 Let ǫ > δ > 0 be given. Since A is an AF-algebra, there is a finite dimensional C ∗-algebra B ⊂ A such that there are mutually orthogonal projections {p′ m} in B such that n} and {q′ 2, ..., p′ 1, q′ 2, ..., q′ 1, p′ i = 1, 2, ..., n and j = 1, 2, ..., m. Put x2 = Pn virtue of 2.17 and choosing sufficiently small δ, (e 8.524) j=1 µjq′ j. Therefore, by the kpi − p′ ik < δ/16n and kqj − q′ i=1 λip′ jk < δ/16m, i and y2 = Pm kx1 − x2k < ǫ/16, ky1 − y2k < ǫ/16 and Dc(x1, x2) < ǫ/16 and Dc(y1, y2) < ǫ/16. It follows from (e 8.523) and (e 8.525) that This has to hold in B too. By Theorem 4.2 of [6], Dc(x2, y2) ≥ d − ǫ/4. dist(x2, y2) ≥ C(d − ǫ/4). It follows that for any ǫ > 0. kx − yk ≥ C(d − ǫ/4) − ǫ/4 = C · d − Cǫ/4 − ǫ/4 (e 8.525) (e 8.526) (e 8.527) (e 8.528) (e 8.529) So we get (e 8.520). For any unitary u, by (e 8.520), we have ku∗xu − yk ≥ C · Dc(u∗xu, y) = C · Dc(x, y). That implies (e 8.519) holds. Lemma 8.2. Let A be a unital simple C ∗-algebra with T R(A) = 0, let x, y ∈ A be two normal elements. Let η > 0. Suppose that τ (f (x)) > τ (g(y)) (e 8.530) for some positive functions f ∈ C(Xη) and g ∈ C(Yη) and for some τ ∈ T (A). Then, for any ǫ > 0, there is a projection p ∈ A, and a finite dimensional C ∗-subalgebra B ⊂ A with 1B = p, and normal elements x0, y0 ∈ (1 − p)A(1 − p), x1, y1 ∈ B such that sp(x1) ⊂ Xη, sp(y1) ⊂ Yη with kx − (x0 + x1)k < ǫ, ky − (y0 + y1)k < ǫ t0(f (x1)) > t0(g(y1)) (e 8.531) (e 8.532) for some t0 ∈ T (B). Proof. Let τ (f (x)) > τ (g(y)) for some τ ∈ T (A) and let d = τ (f (x)) − τ (g(y)) > 0. Fix a separable C ∗-subalgebra C ⊂ A such that x, y ∈ C. Since A has tracial rank zero, there exists a sequence of projections {pn} and a sequence of finite dimensional C ∗-subalgebras {Bn} such that 1Bn = pn, n = 1, 2, ...., such that n→∞kpnc − cpnk = 0 for all c ∈ C; lim lim dist(pncpn, Bn) = 0 and n→∞ max{t(1 − pn) : t ∈ T (A)} = 0. lim n→∞ 59 (e 8.533) (e 8.534) (e 8.535) It follows from [16] and [11] that there exists normal elements x(0) x(1) n , y(1) n , y(0) n ∈ Bn such that n ∈ (1 − pn)A(1 − pn), n→∞kx − (x(0) lim n + x(1) n )k = 0 and lim n→∞ky − (y(0) n + y(1) (e 8.536) n )k = 0. n ), sp(y(1) Therefore, we may assume that sp(x(0) that n ), sp(x(1) n ) ⊂ Xη and sp(y(0) n ) ⊂ Yη. It follows n→∞kf (x) − (f (x(0) lim n ) + f (x(1) n ))k = 0 and lim n→∞kg(y) − (g(y(0) n ) + g(y(1) n ))k = 0. (e 8.537) By (e 8.535), we may assume that It follows, for all sufficiently large n, that t(1 − pn) < d/4 for all t ∈ T (A). τ (f (x(1) n )) > τ (f (x)) − d/4 − d/4 > τ (g(y(1) n )) (e 8.538) (e 8.539) Note Bn is a finite direct sum of simple C ∗-algebras. If, for all tracial states t ∈ T (Bn), n )) ≤ t(g(y(1) then, by [3], there is a sequence {zk,n} ⊂ Bn such that t(f (x(1) n )), ∞ X k=1 k,nzk,n = f (x(1) z∗ n ) and ∞ X k=1 zk,nzk,n ≤ g(y(1) n ). Since Bn ⊂ A, this would imply that τ (f (x(1) n )) ≤ τ (g(y(1) n )) for all τ ∈ T (A) (e 8.540) (e 8.541) (e 8.542) which contradicts with (e 8.539). Theorem 8.3. There is a constant C > 0 satisfying the following: Let A be a unital separable simple C ∗-algebra with T R(A) = 0 and let x, y ∈ A be two normal elements. Then dist(U (x),U (y)) ≥ C · DT (x, y). If [λ − x] = [λ − y] in K1(A) for all λ 6∈ sp(x) ∪ sp(y), then De c (x, y) ≥ dist(U (x),U (y)) ≥ C · DT (x, y). (e 8.543) (e 8.544) Proof. Note that the second part of the theorem follows from the first part and 6.7. So we will only prove the first part of the theorem. Let C be the constant c−1 in the statement of Theorem 4.2 of [6]. Let r = DT (x, y) and let 1/2 > ǫ > 0. Suppose that K > 0 such that kxk, kyk ≤ K and D is a closed ball with the center at the origin and radius larger than K. Denote by ϕX , ϕY : C(D) → A the unital homomorphisms defined by ϕX(f ) = f (x) and ϕY (f ) = f (y) for all f ∈ C(D). We will show that kx − yk ≥ Cr. 60 (e 8.545) There is an open subset O of D such that dτ (ϕX (fO)) > dτ (ϕY (fOs1 )), for some τ ∈ T (A), where s1 = r − ǫ/8. It follows that dτ (ϕX (fO)) > τ (ϕY (g)), (e 8.546) (e 8.547) where g ∈ C(D) such that 0 ≤ g ≤ 1, g(ξ) = 1 if dist(ξ, O) < r − ǫ/2 and g(ξ) = 0 if dist(ξ, O) ≥ r − ǫ/4. There is δ > 0 such that τ (fδ(ϕX (fO))) > τ (ϕY (g)). (e 8.548) By 8.2, there is a projection p ∈ A, a finite dimensional C ∗-algebra B ⊂ A with 1B = p, normal elements x0, y0 ∈ (1 − p)A(1 − p), x1, y1 ∈ B with sp(x0), sp(y0), sp(x1), sp(y1) ⊂ D such that (e 8.549) kx − (x0 + x1)k < ǫ/16, ky − (y0 + y1)k < ǫ/16 and t0(fδ(fO(x1))) > t0(g(y1)) for some t0 ∈ B. Therefore dt0 (fO(x1)) ≥ t0(fδ(fO(x1))) > dt0(fOr−ǫ/2(y1)). It follows that, in B, It follows from Theorem 2.4 of [6] that Dc(x1, y1) ≥ r − ǫ/2. It follows from (e 8.549) that kx1 − y1k ≥ C(r − ǫ/2). kx − yk ≥ k(x0 + x1) − (y0 + y1)k − kx − (x0 + x1)k − k(y0 + y1) − yk ≥ kx1 − y1k − ǫ/8 ≥ Cr − Cǫ/2 − ǫ/8. Hence kx − yk ≥ C · DT (x, y). (e 8.550) (e 8.551) (e 8.552) (e 8.553) (e 8.554) (e 8.555) (e 8.556) Lemma 8.4. Let A be a unital simple C ∗-algebra with T R(A) = 0. Suppose that p, q ∈ A are two non-zero projections such that for some τ ∈ T (A). Then τ (p) > τ (q) k(1 − q)pk = 1 61 (e 8.557) (e 8.558) Proof. We first apply 8.2. Let x = p, y = q and f = g be identity function on [0, 1]. Then, for any ǫ > 0, there are a non-zero projection e ∈ A and a finite dimensional C ∗-algebra B ⊂ A with 1B = e, non-zero projections p1, q1 ∈ B and p0, q0 ∈ (1 − e)A(1 − e) such that kp − (p0 + p1)k < ǫ/4, kq − (q0 + q1)k < ǫ/4 and t0(p1) > t0(q1) (e 8.559) for some t0 ∈ T (B). Since B is finite dimensional, we write B = Mn1 ⊕ Mn2 ⊕ ··· ⊕ Mnk . Accordingly, we may write p1 = (p1,1, ..., p1,nk ), q1 = (q1,1, ..., q1,nk ), (e 8.560) (e 8.561) where p1,i, q1,i,∈ Mni, i = 1, 2, ..., nk. The last condition in (e 8.559) implies, for some i, rankp1,i > rankq1,i. (e 8.562) Let Mni act on Hi (dimHi = ni). Then, by counting the rank, (1Bi − q1,i)Hi ∩ p1,iHi 6= {0}. Therefore It follows that k(1Bi − q1,i)p1,ik = 1. k(e − q1)p1k = 1 (e 8.563) (e 8.564) Let ξ ∈ (1Bi − q1,i)Hi ∩ p1,iHi be a unit vector. Define a projection e0 ∈ B(Hi) = Mni by e0(x) = hx, ξiξ for all x ∈ Hi. Then e0 ∈ Mni ⊂ A is a non-zero projection. Moreover, Therefore e0 ≤ e − q1 and e0 ≤ p1. k(1 − q)pk ≥ k1 − (q0 + q1)(p0 + p1)k − ǫ/2 ≥ ke(1 − (q0 + q1)(p0 + p1))ek − ǫ/2 = k(e − q1)p1k − ǫ/2 = 1 − ǫ/2. It follows that (e 8.558) holds. (e 8.565) (e 8.566) (e 8.567) (e 8.568) Theorem 8.5. Let A be a separable simple C ∗-algebra with T R(A) = 0 and let x, y ∈ A be two normal elements. Then If A is a finite dimensional C ∗-algebra then dist(U (x),U (y)) ≥ dT (x, y). dist(U (x),U (y)) ≥ dc(x, y). (e 8.569) (e 8.570) Proof. Let 0 < d < dT (x, y). Note that in a finite dimensional C ∗-algebra dc(x, y) = dT (x, y). Let ǫ > 0. We assume that ǫ < dT (x,y)−d . We also assume that kxk,kyk > 2ǫ. Denote X = sp(x) and Y = sp(y). Choose any pair of x′ ∈ U (x) and y′ ∈ U (y). Since sp(x′) = X, sp(y′) = Y, dc(x′, y′) = dc(x, y), and dT (x′, y′) = dT (x, y), to simplify the notation, without loss of generality, it suffices to show that kx − yk ≥ dT (x, y). 4 62 By the assumption, there is an open disc O = O(λ, η) such that dτ (fO(x)) > dτ (fOd+ǫ(y)) (e 8.571) for some τ ∈ T (A) (including the case that A is finite dimensional). projection of y corresponding to Od in A∗∗. Let e1 be the spectrum projection of x corresponding to open set Oη+ǫ and e2 be the spectrum Denote by z = y(1 − e2) = (1 − e2)y. Then sp(1−e2)A∗∗(1−e2)(z) = Y \ Od ∩ Y (e 8.572) (as an element in (1 − e2)A∗∗(1 − e2)). The inequality (e 8.571) implies that Y 6= Od ∩ Y. In particular, We also note that dist(λ, sp(1−e2)A∗∗(1−e2)(z)) ≥ d. kxe1 − λe1k < η + ǫ. Therefore It follows that One has k(1 − e2)(y − λ)e1k ≤ k(1 − e2)(y − x)e1k + k(1 − e2)(x − λ)e1k < k(1 − e2)(y − x)e1k + η + ǫ. k(1 − e2)(y − x)e1k > k(1 − e2)(y − λ)e1k − η − ǫ. (1 − e2)(y − λ)e1 = (y − λ)(1 − e2)e1 = (1 − e2)(y − λ)(1 − e2)e1 = (z − λ)(1 − e2)e1. (e 8.573) (e 8.574) (e 8.575) (e 8.576) (e 8.577) (e 8.578) (e 8.579) (e 8.580) Let z1 be the inverse of z − λ in (1 − e2)A∗∗(1 − e2). Then k(1 − e2)e1k ≤ kz1(z − λ)(1 − e2)e1k ≤ kz1kk(z − λ)(1 − e2)e1k. (e 8.581) It follows that k(z − λ)(1 − e2)e1k ≥ k(1 − e2)e1k kz1k = dist(λ, sp(1−e2)A∗∗(1−e2)(z))k(1 − e2)e1k ≥ (η + d)k(1 − e2)e1k. By (e 8.577) and (e 8.578), one concludes that ky − xk ≥ k(1 − e2)(y − x)e1k > k(1 − e2)(y − λ)e1k − η − ǫ ≥ k(z − λ)(1 − e2)e1k − η − ǫ ≥ (η + d)k(1 − e2)e1k − η − ǫ. If A is finite dimensional, then e1, e2 ∈ A. By (e 8.571), ranke1 > ranke2. 63 (e 8.582) (e 8.583) (e 8.584) (e 8.585) (e 8.586) (e 8.587) (e 8.588) As in the proof of 8.4, this implies that k(1 − e2)e1k = 1. From this and from (e 8.585) to (e 8.587), It follows that kx − yk ≥ d. kx − yk ≥ dc(x, y). (e 8.589) (e 8.590) Let e0 be the spectral projection of x corresponding to the closed set {ξ ∈ C : dist(ξ, λ) ≤ η} and e3 be the spectral projection of y corresponding to the open subset Od+ǫ in A∗∗. Note that e0 is a closed projection and e3 is an open projection. If A is a simple infinite dimensional C ∗-algebra with T R(A) = 0, by [2], there are projections p1, q1 ∈ A such that By (e 8.571), It follows from 8.4 that By (e 8.585), e0 ≤ q1 ≤ e1 and e2 ≤ p1 ≤ e3. τ (q1) > τ (p1). k(1 − p1)q1k = 1. kx − yk ≥ (η + d)k(1 − e2)e1k − η − ǫ ≥ (η + d)k(1 − p1)(1 − e2)e1q1k − η − ǫ = (η + d)k(1 − p1)q1k − η − ǫ ≥ d−ǫ. (e 8.591) (e 8.592) (e 8.593) (e 8.594) (e 8.595) (e 8.596) (e 8.597) The theorem follows. Remark 8.6. Suppose that x, y are normal elements in a separable simple C ∗-algebra A with T R(A) = 0, and sp(x) is a subset of a straight line L1 and sp(y) is a subset of another straight line. If L1 and L2 are parallel, by applying 8.5, one can show that dc(x, y) = Dc(x, y). Hence by 8.5 and 3.6, dist(U (x),U (y)) = Dc(x, y) = dc(x, y). If L1 and L2 are perpendicular, one can also show dist(U (x),U (y)) = Dc(x, y). However, in 2.25 of section 2, there are x, y in finite dimesional C ∗-algebra, with x = x∗, y = iy∗ and dc(x, y) = 1 < √2 = Dc(x, y), So we get an example such that dc(x, y) < dist(U (x),U (y)) = Dc(x, y). Corollary 8.7. Let A be a unital AF-algebra and let x, y ∈ A be two normal elements. Suppose that Dc(x, y) = dc(x, y). Then dist(U (x),U (y)) = Dc(x, y). (e 8.598) 64 Corollary 8.8. Let A be a unital simple separable C ∗-algebra with T R(A) = 0 and let x, y ∈ A be two normal elements with sp(x) = X and sp(y) = Y. Then max{C · DT (x, y), dT (x, y), ρ1(x, y)} ≤ dist(U (x),U (y)) ≤ min{D1, D2}, (e 8.599) where D1 = max{DT (x, y), max{ρx(x, y), ρy(x, y)}} + min{ρx(x, y), ρy(x, y)} and D2 = max{De c (x, y) + min{ρx(x, y), ρy(x, y)}} + min{ρx(x, y), ρy(x, y)}. (e 8.600) Suppose that for all λ 6∈ X ∪ Y. Then [λ − x] = [λ − y] in K1(A) max{C · DT (x, y), dT (x, y)} ≤ dist(U (x),U (y)) ≤ De c(x, y). (e 8.601) (e 8.602) References [1] B. Blackadar, K-theory for operator algebras, Mathematical Sciences Research Institute Publications, 5. Springer-Verlag, New York, 1986. viii+338 pp. [2] L. G. Brown, Interpolation by projections in C ∗ -- algebras of real rank zero, J. Operator Theory 26 (1991), 383 -- 387. [3] J. Cuntz and G. K. Pedersen, Equivalence and traces on C ∗-algebras, J. Funct. Anal. 33 (1979), 135 -- 164. [4] M. Dadarlat, Approximately unitarily equivalent morphisms and inductive limit C ∗-algebras, K-Theory 9 (1995), no. 2, 117 -- 137. [5] K. Davidson, The distance between unitary orbits of normal elements in the Calkin algebra, Proc. Roy. Soc. Edinburgh Sect. A 99 (1984), 35 -- 43 [6] K. Davidson, The distance between unitary orbits of normal operators, Acta Sci. Math. (Szeged) 50 (1986), 213 -- 223. [7] K. Davidson, Estimating the distance between unitary orbits, J. Operator Theory 20 (1988), 21 -- 40. [8] E. Effros, D. Handelman and C. L. Shen, Dimension groups and their affine representations, Amer. J. Math. 102, (1980), 385 -- 407. [9] G. A. Elliott, Dimension groups with torsion, Internat. J. Math. 1 (1990), 361 -- 380. [10] G. A. Elliott and G. Gong, On the classification of C ∗-algebras of real rank zero, II Ann. of Math. 144 (1996), 497 -- 610. [11] P. Friis and M. Rørdam, Almost commuting self-adjoint matrices, a short proof of Huaxin Lin's theorem, J. reine angew. Math. 479 (1996), 121 -- 131 [12] F. Hiai and Y. Nakamura, Distance between unitary orbits in von Neumann algebras, Pacific J. Math. 138 (1989), 259 -- 294. 65 [13] J. Holbrook, Spectral variation of normal matrices, Lin. Alg. and its Appl., 174 (1992), 131 -- 144. [14] S. Hu, H. Lin and Y. Xue, Limits of homomorphisms with finite-dimensional range, Internat. J. Math. 16 (2005), 807 -- 821. [15] H. Lin, Approximation by normal elements with finite spectra in simple AF-algebras, J. Operator Theory 31 (1994), 83 -- 98. [16] H. Lin, Almost commuting selfadjoint matrices and applications, Operator algebras and their applications (Waterloo, ON, 1994/1995), 193 -- 233, Fields Inst. Commun., 13, Amer. Math. Soc., Providence, RI, 1997. [17] H. Lin, C ∗-algebras with weak (FN), J. Funct. Anal. 150 (1997), 65 -- 74. [18] H. Lin, An introduction to the classification of amenable C ∗-algebras. World Scientific Pub- lishing Co., Inc., River Edge, NJ, 2001. xii+320 pp. ISBN: 981-02-4680-3. [19] H. Lin, Embedding an AH-algebra into a simple C ∗-algebra with prescribed KK-data, K- Theory 24 (2001), 135 -- 156. [20] H. Lin, Simple AH-algebras of real rank zero, Proc. Amer. Math. Soc. 131 (2003), 3813 -- 3819. [21] H. Lin, Classification of homomorphisms and dynamical systems, Trans. Amer. Math. Soc. 359, (2007), 859 -- 895. [22] H. Lin, The range of approximate unitary equivalence classes of homomorphisms from AH- algebras, Math. Z. 263 (2009), 903 -- 922. [23] H. Lin, Approximate unitary equivalence in simple C ∗-algebras of tracial rank one, Trans. Amer. Math. Soc. 364 (2012), 2021 -- 2086. [24] H. Lin, Homomorphisms from AH-algebras, prepirnt, arXiv:1102.4631. [25] L. Robert, Classification of inductive limits of 1-dimensional NCCW complexes, Adv. Math. 231 (2012), 2802 -- 2836. [26] L. Robert, The Cuntz semigroup of some spaces of dimension at most two, C. R. Math. Acad. Sci. Soc. R. Can. 35 (2013), 22 -- 32. [27] M. Rørdam, On the structure of simple C ∗-algebras tensored with a UHF-algebra. II, J. Funct. Anal. 107 (1992), 255 -- 269. [28] M. Rørdam and W. Winter, The Jiang-Su algebra revisited, J. reine angew. Math. 642 (2010), 129-155. [29] P. Skoufranis, Closed unitary and similarity orbits of normal operators in purely infinite C ∗-algebras, J. Funct. Anal. 265 (2013), 474 -- 506. [30] D. Sherman, Unitary orbits of normal operators in von Neumann algebras, J. Reine Angew. Math. 605 (2007), 95 -- 132. [31] J. Villadsen, The range of the Elliott invariant of the simple AH-algebras with slow dimen- sion growth, K-Theory 15, (1998), 1 -- 12. 66 [32] V. Sunder and K. Thomsem, Unitary orbits of selfadjoints in some C ∗-algebras, Houston J. Math. Vol 18 (1992),127 -- 137. [33] S. Zhang, A Riesz decomposition property and ideal structure of multiplier algebras, J. Operator Theory 24 (1990), 209 -- 325 67
1909.04710	1	1909	2019-09-10T19:18:05	Graded C-algebras and twisted groupoid C-algebras	[ "math.OA" ]	Let $C^$-algebra that is acted upon by a compact abelian group. We show that if the fixed-point algebra of the action contains a Cartan subalgebra $D$ satisfying an appropriate regularity condition, then $A$ is the reduced $C^$-algebra of a groupoid twist. We further show that the embedding $D \hookrightarrow A$ is uniquely determined by the twist. These results generalize Renault's results on Cartan subalgebras of $C^*$-algebras.	math.OA	math	GRADED C ∗-ALGEBRAS AND TWISTED GROUPOID C ∗-ALGEBRAS JONATHAN H. BROWN, ADAM H. FULLER, DAVID R. PITTS, AND SARAH A. REZNIKOFF Abstract. Let C ∗-algebra that is acted upon by a compact abelian group. We show that if the fixed-point algebra of the action contains a Cartan subalgebra D satisfying an appropriate regularity condition, then A is the reduced C ∗-algebra of a groupoid twist. We further show that the embedding D ֒→ A is uniquely determined by the twist. These results generalize Renault's results on Cartan subalgebras of C ∗-algebras. 1. Introduction Abelian operator algebras are well understood: abelian C ∗-algebras are all isomorphic to spaces of continuous functions on a locally compact Hausdorff space; abelian von Neumann algebras are all isomorphic to L∞-spaces. The study of non-abelian operator algebras is often aided by the presence of appropriate abelian subalgebras. This idea was exemplified by Feldman and Moore's characterization of von Neumann algebras containing Cartan subalgebras in 1977 [17]. Cartan embeddings arise naturally in many examples, including finite dimensional von Neumann algebras and von Neumann algebras constructed from free actions of discrete groups on abelian von Neumann algebras. Feldman and Moore [17] gave a complete classification of Cartan subalgebras in terms of measured equivalence relations. To transfer Feldman and Moore's theory to the topological setting, Renault [32] defined Cartan subalgebras for C ∗-algebras. Definition 1.1. [32, Definition 5.1] Let A be a C ∗-algebra. A maximal abelian C ∗-algebra D ⊆ A is a Cartan subalgebra of A if (1) there exists a faithful conditional expectation E : A → D; (2) D contains an approximate unit for A; (3) the set of normalizers of D, i.e. the n ∈ A such that nDn∗ ∈ D and n∗Dn ∈ D, generate A as a C ∗-algebra. When D is a Cartan subalgebra of A, we call (A, D) a Cartan pair. Renault [32], building on work by Kumjian [18], showed that there is a one-to-one correspon- dence between Cartan pairs of separable C ∗-algebras and C ∗-algebras of second countable twisted groupoids; that is, between Cartan pairs and the reduced C ∗-algebra generated by an extension of groupoids T × G(0) → Σ → G. In Renault's result, G must be topologically principal: Renault refers to G as the Weyl groupoid of the Cartan pair. It is reasonable to seek a larger class of inclusions D ⊆ A with D abelian that can be used to construct twists. This idea has recently successfully been pursued by several authors, and larger classes of inclu- sions have been shown to arise as C ∗-algebras of twists. In particular, motivated by shift spaces and the work by Matsumoto and Matui [23, 25, 24], Brownlowe, Carlsen, and Whittaker [7] were able to construct a Weyl type groupoid from a general graph C ∗-algebra and its canonical diagonal and use this construction to show that diagonal-preserving isomorphisms of these inclusions come 1 precisely from isomorphisms of Weyl type groupoids. This led to work proving similar results for Leavitt path algebras [6] and Steinberg algebras [2]. The paper [2] in particular inspired this work. Steinberg algebras are algebraic analogues of groupoid C ∗-algebras [36, 11]. In [2], the authors consider Steinberg algebras associated to groupoids G equipped with a homomorphism c : G → Γ where Γ is an abelian group and c−1(0) is topologically principal. The Steinberg algebra is then naturally graded by Γ; the authors use this grading to reconstruct G. It is well known that algebras graded by an abelian group Γ correspond in the C ∗-algebraic theory to C ∗-algebras endowed with a Γ action (for example see [38], [29]). In this paper, we construct groupoids from inclusions of an abelian C ∗-algebra D into a C ∗- algebra A endowed with the action of a compact abelian group. In particular, the aim of our work is to generalize Renault's characterization of Cartan pairs by reduced C ∗-algebras of twisted groupoids. Our results apply to examples appearing naturally in the study of higher-rank graph and twisted higher-rank graph C ∗-algebras (See Example 7.2 below). We start with a C ∗-algebra A and a discrete abelian countable group Γ such that the dual group Γ acts continuously on A by automorphisms. Let AΓ be the points in A fixed by the action of Γ: this is a subalgebra of A called the fixed point algebra. Assume AΓ contains a Cartan subalgebra D. If in addition the normalizers of D in A densely span A we call (A, D) a Γ-Cartan pair. We note that the normalizers of D in AΓ are homogeneous of degree 0. In particular, if the action by Γ is trivial, then (A, D) is a Cartan pair. If (A, D) is a Γ-Cartan pair, then following Kumjian's construction, we show how to create a twisted groupoid (Σ; G) that is graded by Γ. This yields the following commutative diagram (1.2) T × G(0) ι / Σ q / G ❃❃❃❃❃❃❃❃ cΣ cG Γ where cΣ and cG are homomorphisms. We prove in Theorem 4.36 that there is a natural isomor- phism between (A, D) and the reduced crossed product (C ∗ r (Σ; G), C0(G(0))). Next, if Σ → G is a twist satisfying the commutative diagram (1.2), we show that the inclusion C0(G(0)) ֒→ C ∗ r (Σ; G) satisfies our hypotheses, so Theorem 4.36 allows us to construct a new twist from this inclusion. The natural question is: does our construction in Theorem 4.36 recover Σ → G? We answer this affirmatively in Theorem 6.4. This second question is the main focus of [7], [6] [2], and [10] in the case that the twist is trivial. The paper [10] by Carlsen, Ruiz, Sims and Tomforde is similar in scope to our present work. While [10] is also concerned with translating the results of [2] to a C ∗-algebraic framework, their work avoids twists altogether, instead focusing on showing rigidity results along the lines of our Theorem 6.4. The results of [10] apply to C ∗-algebras already known to arise from groupoids, however it does contain some remarkable innovations which allows the authors to address C ∗- algebras endowed with co-actions of a possibly nonabelian group. Furthermore, Carlsen, Ruiz, Sims and Tomforde relax the requirement that the abelian subalgebra D must be Cartan in the fixed point algebra. Whether or not a C ∗-algebra satisfies the Universal Coefficient Theorem (UCT) remains the main stumbling block in the classification program for simple nuclear C ∗-algebras. Indeed, Tikuisis, White and Winter [37] have shown that all separable, unital, simple, nuclear C ∗-algebras with finite nuclear dimension that satisfying the UCT are classifiable. Recent results of Barlak and Li [4] show that if A is a nuclear C ∗-algebra containing a Cartan subalgebra, then A satisfies the UCT. We discuss in Example 7.3 how their results also apply to our setting. 2 / / This paper is organized as follows. We begin with preliminaries on twists (Section 2). In Section 3 we define Γ-Cartan pairs and review the relationship between topological grading and strong group actions. In Section 4 we prove our main theorem, which shows that a large class of C ∗-algebras are isomorphic to the reduced C ∗-algebra of a twist. In Section 5 we then provide a few basic results concerning a natural Γ-Cartan pair that arises in the presence of a twist. In Section 6 we prove our rigidity result, Theorem 6.4, which shows that if the inclusion in the previous section comes from a twist then our construction recovers the twist. Section 7 gives some examples to which our theorems apply. Notably, in Example 7.2 we show that the twisted higher-rank graph C ∗-algebras introduced in [21] and [22] give examples of Γ- Cartan pairs. Moreover, the groupoid description of twisted higher-rank graph C ∗-algebras given in [22] yields groupoids isomorphic to ours. Finally, in an appendix, we describe how we can obtain the results of Section 4 by using a coaction of a non-abelian group (instead of an action of an abelian group); note that in this case the grading on the C ∗-algebra is by the group itself, rather than by its dual. The authors thank John Quigg for pointing out this alternative construction. Acknowledgments. Whilst conducting this research JHB and AHF made use of meeting space made available by the Columbus Metropolitan Library. We would like to thank CML for their important work in the community. AHF would like to thank Christopher Schafhauser for patiently answering his questions on [4]. This work was partially supported by grants from the Simons Foundation (#316952 to David Pitts and #360563 to Sarah Reznikoff) and by the American Institute of Mathematics SQuaREs Program. 2.1. ´Etale groupoids. A groupoid G is a small category in which every morphism has an inverse. The unit space G(0) of G is the set of identity morphisms. The maps s, r : G → G(0), given by s(γ) = γ−1γ and r(γ) = γγ−1, are the source and range maps. For S, T ⊆ G we denote 2. Preliminaries ST := {γη : γ ∈ S, η ∈ T, r(η) = s(γ)}. If either S or T is the singleton set {γ} we remove the set brackets from the notation and write Sγ or γT . A topological groupoid is a groupoid G endowed with a topology such that inversion and compo- sition are continuous. An open set B ⊆ G is a bisection if r(B) and s(B) are open and rB and sB are homeomorphisms onto their images. The groupoid G is ´etale if there is a basis for the topology on G consisting of bisections. When G is ´etale then G(0) is open and closed in G. For x ∈ G(0), the isotropy group at x is xGx := {γ ∈ G : r(γ) = s(γ) = x} and the isotropy subgroupoid is the set G′ = {γ ∈ G : r(γ) = s(γ)}. A topological groupoid G is topologically principal if {x ∈ G(0) : xGx = {x}} is dense in G(0); it is effective if the interior of G′ is G(0). If G is second countable these notions coincide [5, Lemma 3.1], but in the general (not necessarily second countable) case, effective is the more useful notion. Unless explicitly stated otherwise, for the remainder of this paper, we make the following as- sumptions. Standing Assumptions on Groupoids. Throughout, all groupoids are: (1) locally compact and (2) Hausdorff. 3 2.2. Twists. The main focus of this paper is on twists and their C ∗-algebras. We provide a brief account of the necessary background here. Much of this background can also be found in [32]. We also encourage the reader to consult the recent expository article by Sims [33]. We now expand on a few details that are particularly relevant to our context. A twist is the analog of a central extension of a discrete group by the circle T. Here is the formal definition. Definition 2.1 (see [33, Definition 5.1.1]). Let Σ and G be topological groupoids with G ´etale, and let T × G(0) be the product groupoid. That is, (z1, x1)(z2, x2) is defined if and only if x1 = x2, in which case the product is given by (z1, x1)(z2, x2) = (z1z2, x1); inversion is (z, x)−1 = (z−1, x), and the topology is the product topology. The unit space of T × G(0) is {1} × G(0). The pair (Σ, G) is a twist if there is an exact sequence T × G(0) ι→ Σ q → G where (1) ι and q are continuous groupoid homomorphisms with ι one-to-one and q onto; (2) ι{1}×G(0) and qΣ(0) are homeomorphisms onto Σ(0) and G(0), respectively (identify Σ(0) and G(0) using q); (3) q−1(G(0)) = ι(T × G(0)); (4) for every γ ∈ Σ and z ∈ T, ι(z, r(γ))γ = γι(z, s(γ)); and (5) for every g ∈ G there is an open bisection U with g ∈ U and a continuous function φU : U → Σ such that q ◦ φU = id U and the map T × U ∋ (z × h) 7→ ι(z, r(h)) φU (h) is a homeomorphism of T × U onto q−1(U ). (Conditions (1 -- 3) say the sequence is an extension, (4) says the extension is central, and (5) says G is ´etale and the extension is locally trivial.) A twist is often denoted simply by Σ → G. For z ∈ T and γ ∈ Σ we will write z · γ := ι(z, r(γ))γ and γ · z := γι(z, s(γ)) for the action of T on Σ arising from the embedding of T × G(0) into Σ. Notice that this action of T on Σ is free. Also, for γ ∈ Σ, we will often denote q(γ) by γ; indeed, we use the name γ for an arbitrary element of G. Remark 2.2. By [40, Exercise 9K(3)] the map q : Σ → G is a quotient map. The C ∗-algebra of the twist is constructed from the completion of an appropriate function algebra Cc(Σ; G). This algebra can be constructed in two different ways and both will be used in this note. First description of Cc(Σ; G): Sections of a line bundle. The first way to construct Cc(Σ; G) is by considering sections of a complex line bundle L over G. Define L to be the quotient of C × Σ by the equivalence relation on C × Σ given by (λ, γ) ∼ (λ1, γ1) if and only if there exists z ∈ T such that (λ1, γ1) = (zλ, z · γ). We sometimes write L = (C × Σ)/T. Use [λ, γ] to denote the equivalence class of (λ, γ). Observe that for any z ∈ T, (2.3) [λ, z · γ] = [zλ, γ]. With the quotient topology, L is Hausdorff. The (continuous) surjection P : L → G is given by P : [λ, γ] 7→ γ. For γ ∈ G and γ0 ∈ q−1( γ), the map C ∋ λ 7→ [λ, γ0] ∈ P −1( γ) is a homeomorphism, so L is a complex line bundle over G. In general, there is no canonical choice of γ0. However, when γ ∈ G(0), Σ(0) ∩ q−1( γ) is a singleton set, so there is a canonical choice: take γ0 to be the unique element 4 of Σ(0) ∩ q−1( γ). Thus, recalling that Σ(0) and G(0) have been previously identified (using qΣ(0)), when x ∈ G(0) = Σ(0), we sometimes identify P −1(x) with C via the map λ 7→ [λ, x] = λ · [1, x]. Finally, there is a continuous map : L → [0, ∞) given by ([λ, γ]) := λ. When f : G → L is a section and γ ∈ G, we will sometimes write f ( γ) instead of (f ( γ)). Since Σ is locally trivial, L is locally trivial as well. Indeed, given ℓ ∈ L, let B be an open bisection of G containing P (ℓ). Let φB : B → Σ be a continuous function satisfying the conditions of Definition 2.1(5). Then for every element ℓ1 ∈ P −1(B), there exist unique λ ∈ C and γ ∈ B so that ℓ1 = [λ, φB( γ)]. It follows that the map [λ, φB( γ)] 7→ (λ, γ) is a homeomorphism of P −1(B) onto C × B, so L is locally trivial. There is a partially defined multiplication on L, given by [λ, γ][λ′, γ′] = [λλ′, γγ′], whenever γ and γ′ are composable in Σ. When [λ, γ], [λ′, γ′] ∈ L satisfy γ = γ′, let (2.4) where z is the unique element of T so that γ′ = z · γ. There is also an involution on L given by [λ, γ] + [λ′, γ′] := [λ + zλ′, γ], (2.5) [λ, γ] = [λ, γ−1]. We use the symbol Cc(Σ; G) to denote the set of "compactly supported" continuous sections of L, that is, (2.6) Cc(Σ; G) := {f : G → L f is continuous, P ◦ f = id G, and ◦ f has compact support}. Notation 2.7. For f ∈ Cc(Σ; G), we denote the support of ◦ f by supp(f ); we denote its open support by supp′(f ). Further, let C(Σ; G) and C0(Σ; G) be, respectively, the set of continuous sections and continuous sections vanishing at infinity of the bundle L. We endow Cc(Σ; G) with a ∗-algebra structure where addition is pointwise (using (2.4)), multi- plication is given by convolution: (2.8) and the involution is from (2.5): f ∗ g( γ) = Xη1 η2= γ f ( η1)g( η2) = Xr( η)=r( γ) f ( η)g( η−1 γ), f ∗( γ) = f ( γ−1). Note that if f, g are supported on bisections B1, B2 and ηi ∈ Bi then f ∗ g( η1 η2) = f ( η1)g( η2). We can identify C0(G(0)) with a subalgebra of continuous sections of the line bundle L by C0(G(0)) → C0(Σ; G) by φ 7→ γ 7→([φ( γ), ι(1, γ)] 0 γ ∈ G(0) otherwise! . Note that this identification takes pointwise multiplication on C0(G(0)) to the convolution on Cc(Σ; G). Second description of Cc(Σ; G): Covariant functions. A function f on Σ is covariant if for every z ∈ T and γ ∈ Σ, f (z · γ) = z f (γ). The second way to describe Cc(Σ; G) is as the set of compactly supported continuous covariant functions on Σ, that is, (2.9) Cc(Σ; G) := {f ∈ Cc(Σ) : ∀γ ∈ Σ ∀z ∈ T f (z · γ) = zf (γ)}. 5 Addition is pointwise, the involution is f ∗(γ) = f (γ−1), and the convolution multiplication is given by (2.10) f ∗ g(γ) = Xη∈G r( η)=r( γ) f (η)g(η−1γ), where for each η with r( η) = r( γ), only one representative η of η is chosen. It is easy to verify that this is well-defined. Equivalence of the descriptions. To proceed, we need to be more explicit on how these two descrip- tions of Cc(Σ; G) are the same. Take f ∈ Cc(Σ) such that f (z · γ) = zf (γ) for all γ ∈ Σ and z ∈ T. Let f be the section of the line bundle given by f ( γ) = [f (γ), γ]. Note that by the definition of the line bundle, this is well-defined. On the other hand, consider a compactly supported continuous section f : G → L. For γ ∈ Σ, the fact that P ◦ f = id G yields P(cid:16)[1, γ]−1 f ( γ)(cid:17) = s( γ). Hence there exists λγ ∈ C such that [1, γ]−1 f ( γ) = λγ · [1, s(γ)], that is, Define f : Σ → C by f ( γ) = λγ · [1, γ] = [λγ, γ]. f (γ) = λγ. Then f is continuous and compactly supported since f is and satisfies (2.11) f (z · γ) = zf (γ). We have thus described a linear isomorphism between the spaces defining Cc(Σ; G) given in (2.6) and (2.9). It is a routine matter to show this linear map is a ∗-algebra isomorphism, so that the two descriptions coincide. Notice that γ ∈ supp(f ) if and only if γ ∈ supp( f ). Remark 2.12. Technically, the support of a function f : Σ → C satisfying the covariance condition (2.11) is a subset of Σ, but (2.11) allows us to regard both supp(f ) and supp′(f ) as subsets of G. We shall do this. Thus the notions of support are the same whether f is viewed as a covariant function or as a section of the line bundle. To define the reduced groupoid C ∗-algebra, we need to define regular representations. For x ∈ G(0), let Hx = ℓ2(Σx, Gx) be the set of square summable sections of the line bundle LGx; that is, Hx = {χ : Gx → P −1(Gx) for γ ∈ Gx, P (χ( γ)) = γ, and ◦ χ ∈ ℓ2(Gx)}. Given χ1, χ2 ∈ Hx and γ ∈ Gx, P(cid:16)χ2( γ)χ1( γ)(cid:17) = x ∈ G(0), so that we obtain a unique λ γ ∈ C so that χ2( γ)χ1( γ)) = λ γ · [1, x]. We may therefore define an inner product on Hx: hχ1, χ2i is the unique element of C such that (2.13) χ2( γ)χ1( γ) = hχ1, χ2i · [1, x] = [hχ1, χ2i , x]. Xγ∈Gx The regular representation of Cc(Σ; G) on Hx is then defined as follows. For f ∈ Cc(Σ, G) and χ ∈ Hx, (πx(f )χ) ( γ) = Xη∈G, ζ∈Gx η ζ= γ f ( η)χ( ζ) = Xη∈G r( η)=r( γ) 6 f ( η)χ( η−1 γ) ( γ ∈ Gx). The reduced C ∗-algebra of (Σ; G), denoted C ∗ r (Σ; G), is the completion of Cc(Σ; G) under the norm kf kr = supx∈G(0) kπx(f )k. Remark 2.14. Viewing Cc(Σ, G) as the space of compactly supported sections of the line bundle affords us an alternative way to describe the regular representations, as follows. Given x ∈ G(0), define a linear functional ǫx on Cc(Σ, G) by defining ǫx(f ) to be the unique scalar such that f (x) = [ǫx(f ), x]. Note that for f ∈ Cc(Σ, G), ǫx(f ∗f ) = Xs( γ)=x (f ( γ))2 ≥ 0, number of open bisections U1, . . . , Un for G such that supp(g) ⊆ Sn so ǫx is positive. Morever, if also g ∈ Cc(Σ, G), then [16, Proposition 3.10] shows there exist a finite j=1 Uj, from which it follows that (εx(f ∗g∗gf ))1/2 ≤ n kgk∞ εx(f ∗f )1/2. Thus the GNS construction may be applied to εx to produce a representation (πεx, Hεx ) of Cc(Σ; G). Letting Lεx be the left kernel of εx, the map Cc(Σ; G)/Lεx ∋ g + Lεx 7→ gGx is isometric and so determines an isometry W : Hεx → Hx. As G is ´etale, for γ ∈ Gx, there is an open bisection U for G with γ ∈ U , and hence we may find g ∈ Cc(Σ; G) supported in U with g( γ) 6= 0. Thus, if h ∈ ℓ2(Σx, Gx) has finite support, there exists f ∈ Cc(Σ, G) with f Gx = h. This implies that W is onto, and a calculation shows that W πεx = πxW . This shows πx and πεx are unitarily equivalent representations of Cc(Σ; G). Of course, the same applies when Cc(Σ; G) is viewed as compactly supported continuous covariant functions on Σ: in this case εx(f ) = f (x). For x ∈ G(0), it will be useful to have a fixed orthonormal basis for Hx. For η ∈ Gx, we select δ η ∈ Hx such that ( ◦ δ η)( γ) =(1 if γ = η 0 otherwise. and insist in particular that δx(x) = [1, x]. Then {δ η : η ∈ Gx} is an orthonormal basis for Hx. δ η( η) ∈ L. By choosing (and fixing) η ∈ q−1( η) there exists a unique λ η ∈ T such that In the sequel, we will have occasion to consider the element It is sometimes useful to informally regardDπx(f )δ η, δ ζE as a product of elements of L, and we now give a formula which provides this description. For f ∈ Cc(Σ; G), x ∈ G(0) and η, ζ ∈ Gx, the definition of πx(f ) and the inner product on Hx yield δ η( η) = [λ η, η]. hDπx(f )δ η, δ ζE , xi = δ ζ( ζ)f ( ζ η−1)δ η( η). In particular,hDπx(f )δx, δ ζE , xi = δ ζ( ζ)f ( ζ). Therefore, f ( ζ) =Dπx(f )δx, δ ζE · δ ζ( ζ) and (cid:12)(cid:12)(cid:12)Dπx(f )δx, δ ζE(cid:12)(cid:12)(cid:12) = (f ( ζ)). (2.15) (2.16) (2.17) Example 2.18. Suppose that σ is a normalized continuous 2-cocycle on the ´etale groupoid G. This is a continuous function from the set of composable pairs G(2) into T such that σ(γ, s(γ)) = 1 = σ(r(γ), γ) and for all composable triples, (γ1, γ2, γ3), σ(γ2, γ3)σ(γ1γ2, γ3)σ(γ1, γ2γ3)σ(γ1, γ2) = 1. 7 Define Σ := T×σG, where T×σG is the Cartesian product of T and G with the product topology and multiplication defined by (z1, γ1)(z2, γ2) = (z1z2σ(γ1, γ2), γ1γ2). In this case, L may be identified with C × G by φ : [λ, (z, γ)] 7→ (λz, γ) and we identify sections of L with functions on G by f = p1 ◦ φ ◦ f where f ∈ Cc(G; Σ) where p1 is the projection onto the first factor. Now for compactly supported sections f, g of L, f ∗ g( γ) =X f ( η)g( η−1 γ) =X( f ( η), η)(g( η−1 γ), η−1 γ) =X[ f ( η), (1, η)][g( η−1 γ), (1, η−1 γ)] =X[ f ( η)g( η−1 γ), (σ( η, η−1 γ), γ)] =X[ f ( η)g( η−1 γ) σ( η, η−1 γ), (1, γ)] =(cid:16)X f ( η)g( η−1 γ) σ( η, η−1 γ), γ)(cid:17) . This last sum is the convolution formula for f , g in Cc(Σ; G) used by Renault in [30]. In particular, if σ is trivial then we get the usual convolution formula for ´etale groupoid C ∗-algebras. We will use the following proposition to find useful subalgebras of the twisted groupoid C ∗- algebra. Lemma 2.19. Let T × G(0) ΣH := q−1(H). Then ι → Σ q → G be a twist and H be an open subgroupoid of G. Define T × H (0) ι H(0) → ΣH qΣH→ H is a twist. Moreover the map κ : Cc(ΣH ; H) ֒→ Cc(Σ; G) defined by extending functions by zero extends to an inclusion of C ∗ r (ΣH ; H) into C ∗ r (Σ; G). Proof. That T × H (0) γ ∈ ΣH if and only if γ ∈ H. H(0) → ΣH ι qΣH→ H is a twist comes from the facts that q(ι(λ, x)) = x and View elements of Cc(Σ; G) and Cc(ΣH; H) as sections of line bundles. By definition, the respec- tive line bundles are LΣH = (C × ΣH)/T and LΣ = (C × Σ)/T. Therefore, LΣH = LΣH. Since H and ΣH are open, we may define κ : Cc(ΣH; H) ֒→ Cc(Σ; G) by extending functions by zero. For each x ∈ X, let εx be defined as in Remark 2.14 and let εH r (ΣH; H). Let (πx, Hx) and (πH r (Σ; G) and C ∗ r (Σ; G) and LH x := εx ◦ κ. Then εx and εH x x , HH x ) be their associated GNS r (ΣH; H) be the left kernels of εx and εH x x (h∗ ∗ h) = εx(κ(h)∗ ∗ κ(h)), so the map on Cc(ΣH ; H) defined x → Hx. A calculation shows that for x ⊆ C ∗ extend to states on C ∗ representations and let Lx ⊆ C ∗ respectively. For h ∈ Cc(ΣH; H), εH by (h + LH h ∈ Cc(ΣH; H), WxπH x ) 7→ (κ(h) + Lx) extends to an isometry Wx : HH x (h) = πx(κ(h))Wx so that WxπH x (h)W ∗ x = πx(κ(h))(WxW ∗ x ). Thus, for h ∈ Cc(ΣH; H), r (ΣH ;H) = sup (2.20) khkC ∗ x∈X(cid:13)(cid:13)πH x (h)(cid:13)(cid:13) = sup x∈X kWxπx(κ(h))W ∗ x k ≤ sup x kπx(κ(h))k = kκ(h)kC ∗ r (Σ;G) . Let B = κ(Cc(ΣH; H)), so B is a C ∗-subalgebra of C ∗ r (Σ; G). By (2.20), the map κ(h) 7→ h extends to a ∗-epimorphism Θ : B ։ C ∗ r (ΣH ; H). Now let ∆ : C ∗ r (Σ; G) → C0(X) be the faithful conditional expectation determined by Cc(Σ; G) ∋ f 7→ f X; likewise let ∆H : C ∗ r (ΣH; H) → C0(X) be determined by C0(ΣH; H) ∋ h 7→ hX . For h ∈ C0(ΣH; H), ∆(κ(h)) = ∆H(h). Therefore, for b ∈ B, ∆(b) = ∆H(η(b)). So for b ∈ B, Θ(b∗b) = 0 implies b = 0 by the faithfulness of ∆. It follows that Θ is a ∗-isomorphism of B onto 8 r (ΣH; H). Therefore, Θ−1 is a ∗-isomorphism of C ∗ C ∗ we needed to show. r (ΣH; H) onto κ(C0(ΣH ; H)), which is what (cid:3) The following proposition allows us to view elements of C ∗ r (Σ; G) as functions in C0(Σ; G). This proposition was originally proved in the case of Example 2.18 above by Renault in [30, Proposition II.4.2]. Renault uses it without proof in the full generality of twists in [32]. As we know of no proof of [30, Proposition II.4.2] for twists, we provide a proof here at the level of generality we will require. Note that C0(Σ, G) can be made into a Banach space with kf k = sup γ∈G (f ( γ)). Proposition 2.21. Let (Σ; G) be a twist with G ´etale. Then the inclusion map j : Cc(Σ; G) → C0(Σ; G) extends to a norm-decreasing injective linear map of C ∗ r (Σ; G) into C0(Σ; G). Moreover, the algebraic operations of adjoint and convolution on Cc(Σ; G) extend to corresponding operations on j(C ∗ r (Σ; G) and γ ∈ G, r (Σ; G)), that is, for every a, b ∈ C ∗ j(a∗)( γ) = j(a)( γ−1) and (2.22) j(ab)( γ) = Xr( η)=r( γ) j(a)( η) j(b)( η−1 γ). Proof. The algebra Cc(Σ; G) may be regarded as a subalgebra of C ∗ r (Σ; G) or as its image under j in C0(Σ; G). First we show that for f ∈ Cc(Σ; G) we have kf kr ≥ kf k∞. To see this, for γ ∈ G consider δs( γ). We have (2.23) kf kr ≥ kπs( γ)(f )k ≥ kπs( γ)(f )δs( γ)k = hπs( γ)(f )(δs( γ)), πs( γ)(f )(δs( γ))i1/2 =s Xs( η)=s( γ) f (η)2 ≥ f ( γ). Thus j extends to a norm decreasing linear map j : C ∗ r (Σ; G) → C0(Σ; G). We turn to showing that j is injective. Since j is norm-decreasing, the equalities in (2.17) extend to every element of C ∗ r (Σ; G). Therefore, for any γ ∈ Gx, and a ∈ C ∗ r (Σ; G), kπx(a)δ γ k2 = Xu∈Gx hπx(a)δ γ , δ ui 2 = Xu∈Gx πx(a)δ γ( u)2 = πx(a)δ γ( γ)2 = j(a)( γ γ−1)2. So if j(a) = 0, then πx(a) = 0 for every x ∈ G(0). Thus a = 0, so j is injective. To verify the first equality in (2.22), observe that it holds for a ∈ Cc(Σ; G). For general a ∈ r (Σ; G), observe that for any f ∈ Cc(Σ; G), the fact that j is contractive yields C ∗ (j(a∗( η)) − j(a)( η)) ≤ (j(a∗ − f ∗)( η)) + (j(f − a)( η−1)) ≤ 2 ka − f kr . As the right-most term in this inequality can be made as small as desired by choosing f appropri- ately, we obtain the first equality. Before establishing the second, for a ∈ C ∗ r (Σ; G) and x ∈ G(0), define kak2,x := kπx(a)δxk . Then max{kak2,x , ka∗k2,x} ≤ kakr and and, using the first equality in (2.22), kak2 ka∗k2 2,x = Xη∈Gx 2,x = Xη∈xG hπx(a)δx, δ ηi 2 = Xη∈Gx j(a)( η)2 j(a)( η)2. 9 To establish the second equality in (2.22), first note it holds when a, b ∈ Cc(Σ; G). Now let r (Σ; G) be arbitrary. Suppose (fi), (gi) are nets in Cc(Σ; G) such that kfi − akr → 0 and a, b ∈ C ∗ kgi − bkr → 0. Then  Xr( η)=r( γ) =  Xr( η)=r( γ) ≤ kf ∗ j(fi)( η) j(gi)( η−1 γ) − Xr( η)=r( γ) j(fi)( η) j(gi − b)( η−1 γ) + Xr( η)=r( γ) i k2,r( γ) kgi − bk2,s( γ) + kf ∗ i − a∗k2,r( γ) kbk2,s( γ) j(a)( η) j(b)( η−1 γ) j(fi − a)( η) j(b)( η−1 γ) ≤ kfikr kgi − gkr + kfi − akr kbkr , from which it follows that lim i→∞ Xr( η)=r( γ) j(fi)( η) j(gi)( η−1 γ) = Xr( η)=r( γ) j(a)( η) j(b)( η−1 γ). Therefore, for every γ ∈ G, j(ab)( γ) =(cid:10)πs( γ)(ab)δs( γ), δ γ(cid:11) δ γ( γ) = lim j(figi)( γ) fi( η)gi( η−1 γ) = Xr( η)=r( γ) = lim Xr( η)=r( γ) a( η)b( η−1 γ), as desired. (cid:3) Definition 2.24. Let G be an ´etale groupoid and Γ a discrete abelian group. A twist graded by Γ is a twist T × G(0) ֒→ Σ ։ G over G together with continuous groupoid homomorphisms cΣ : Σ → Γ and cG : G → Γ such that the diagram, (2.25) T × G(0) / Σ / G ❃❃❃❃❃❃❃❃ cΣ cG Γ commutes. We will sometimes abbreviate (2.25) and simply say Σ → G is a Γ-graded twist. For ω ∈ Γ and t ∈ Γ we denote the natural pairing ω(t) by hω, ti. We will use additive notation for the group Γ and multiplicative notation for the group Γ. We now show that the grading maps cΣ and cG induce an action of Γ on C ∗ r (Σ; G). This fact is well known to experts but we include a proof for completeness. Lemma 2.26. Suppose Σ → G is a Γ-graded twist. There exists a continuous action of Γ on C ∗ r (Σ; G) characterized by (ω · f )( γ) = hω, cG( γ)if ( γ) where ω ∈ Γ and f ∈ Cc(Σ; G). 10 / / Proof. First we check that the action is multiplicative. For this we compute (ω · f ) ∗ (ω · g)( γ) = Xr( η)=r( γ) = Xr( η)=r( γ) = hω, c( γ)i Xr( η)=r( γ) (ω · f )( η)(ω · g)( η−1 γ) hω, c( η)if ( η)hω, c( η−1 γ)ig( η−1 γ) f ( η)g( η−1 γ) = (ω · (f ∗ g))( γ). Now let L be the line bundle over G associated to Σ and for x ∈ G(0) let Lx := LGx. Consider the regular representation πx of C ∗ r (Σ; G) associated to x ∈ G(0). For χ ∈ Hx define χω ∈ Hx by χω( γ) := hω, c( γ)iχ( γ). Then kχωk2 = kχk2, so the mapping χ 7→ χω is a unitary Wω ∈ B(Hx). So for f ∈ Cc(Σ; G), hω, c( η)if ( η)χ( η−1 γ) πx(ω · f )χ( γ) = Xr( η)=r( γ) = Xr( η)=r( γ) = hω, c( γ)i Xr( η)=r( γ) hω, c( γ)ihω, c( γ)ihω, c( η−1)if ( η)χ( η−1 γ) f ( η)χω( η−1 γ) = hω, c( γ)iπx(f )χω( γ). This then implies that kπx(ω · f )χk = kπx(f )χωk. So now kπx(ω · f )k = sup kχk=1 kπx(ω · f )χk = sup kχk=1 kπx(f )χωk = sup kχk=1 kπx(f )χk = kπx(f )k and since this holds for all x we get kω · f kr = kf kr as desired. Now suppose that we have nets ωi → ω and ai → a ∈ C ∗ r (G; Σ). Consider ωi · ai − ω · a = ωi · ai − ωi · a + ωi · a − ω · a. Since kω · akr = kakr, to show ωi · ai → ω · a it suffices to show ωi · a → ω · a. For i sufficiently large we can assume kakr sup hωi − ω · c( η)i < ǫ. Now kπx(ωi · a − ω · a)χk = k Xr( η)=r( γ) hωiω−1, c( η)ia( η)χ( η−1 γ)k = khωiω−1, c( γ)i Xr( η)=r( γ) ≤ hωiω−1, c( γ)ikakr < ǫ. a( η)χωiω−1( η−1 γ)k Since this holds for all x ∈ G(0) we get the result. (cid:3) Remark 2.27. When elements of Cc(Σ; G) are viewed as in (2.9), the action of Γ on C ∗ characterized by r (Σ; G) is where ω ∈ Γ and f ∈ Cc(Σ) is covariant. 11 (ω · f )(γ) = hω, cΣ(γ)i f (γ), 3. Γ-Cartan pairs and abelian group actions In this section we define the main objects of our study, Γ-Cartan pairs, and explore the rela- tionship between Γ-Cartan pairs and strongly continuous actions of compact abelian groups on C ∗-algebras. We first give some preliminary results on topologically graded C ∗-algebras. Definition 3.1. A C ∗-algebra A is topologically graded by a (discrete abelian) group Γ if there exist a family of linearly independent closed linear subspaces {At}t∈Γ of A such that t = A−t, • AtAs ⊆ At+s, • A∗ • A is densely spanned by {At}t∈Γ; and • there is a faithful conditional expectation from A onto A0. Definition 3.2. Let A be a C ∗-algebra topologically graded by a group Γ. We call an element a ∈ A homogeneous if a ∈ At for some t. Let D ⊆ A0 be an abelian subalgebra. We denote the set of normalizers of D in A by N (A, D) or simply N . Also, n is a homogeneous normalizer if it is both a normalizer and homogeneous: that is, n is a normalizer and n ∈ At for some t ∈ Γ. We denote the set of homogeneous normalizers by Nh(A, D) or simply Nh. Notice that for n ∈ Nh and d ∈ D we have nd, dn ∈ Nh. The term topologically graded was introduced by Exel [14]; see also [15]. An action of a compact abelian group on a C ∗-algebra produces a topological grading, which we now describe in some detail. Let Γ be a discrete abelian group and A a C ∗-algebra. As is customary, we say Γ acts strongly on A if there is a strongly continuous group of automorphisms on the C ∗-algebra A indexed by Γ. That is, there is a map Γ × A → A, written (ω, a) 7→ ω · a such that: (1) for every ω, a 7→ ω · a is an automorphism βω of A; (2) the map ω 7→ βω is a homomorphism of Γ into Aut(A); and (3) for each a ∈ A, the map ω 7→ ω · a is norm continuous. Let AΓ be the fixed point algebra under this action. For t in Γ and a ∈ A define (3.3) and let Φt(a) :=ZΓ (ω · a)hω−1, tidω, At = Φt(A) be the range of Φt. Then for each t ∈ Γ, Φt is a completely contractive and idempotent linear map. The following simple fact is worth noting. Lemma 3.4. The map Φ0 : A → AΓ = A0 is a faithful conditional expectation. Sketch of Proof. That Φ0 is a conditional expectation is clear, so it remains to show Φ0 is faithful. If Φ0(a∗a) = 0, then for every state ρ on A,RΓ ρ(ω · (a∗a)) dω = 0. Thus ρ(ω · (a∗a)) = 0 for every state ρ and every ω ∈ Γ. Taking ω to be the unit element gives ρ(a∗a) = 0 for every state, so a∗a = 0. (cid:3) We now characterize the homogeneous elements of A. The following lemma is a generalization of [1, Lemma 5.2.10]. where it is proved for Γ = Z. Lemma 3.5. Suppose Γ acts strongly on A. The following statements hold for all t ∈ Γ, a, b ∈ A. (1) a ∈ At iff for every σ ∈ Γ, ω · a = hω, tia. (2) a ∈ At iff a∗ ∈ A−t. (3) If a ∈ At, b ∈ As then ab ∈ At+s. 12 (4) If a ∈ At and s ∈ Γ, then Φs(a) =(a if s = t; otherwise. 0 Proof. Let a ∈ At and σ ∈ Γ. Then Conversely if σ · a = hσ, tia for every σ ∈ Γ, then σ · a = σ · Φt(a) =ZΓ = hσ, tiΦt(a) = hσ, tia. ((σω) · a)(cid:10)ω−1, t(cid:11) dω =ZΓ (ω · a)hω−1, tidω =ZΓ Φt(a) =ZΓ (ω · a)(cid:10)ω−1, tσ(cid:11) dω ahω, tihω−1, tidω = a. Items (2) and (3) follow immediately since σ · (a∗) = (σ · a)∗ = hσ, tia∗ = hσ, −tia∗ and σ · (ab) = (σ · a)(σ · b) = hσ, ti ahσ, sib = hσ, t + siab. Lastly for (4), Φs(a) =ZΓ (ω · a)hω−1, sidω = aZΓ hω, tihω−1, sidω = δs,ta. (cid:3) The following lemma and its corollary show the homogeneous spaces {At}t∈Γ together densely span in A. We thank Ruy Exel for showing us the simple proof. Lemma 3.6. Suppose the compact abelian group Γ acts strongly on the C ∗-algebra A, and a ∈ A. Then a ∈ span{Φt(a) : t ∈ Γ}. Proof. Let B := span{Φt(a) : t ∈ Γ}. Suppose ρ is a bounded linear functional on A which annihilates B. Define ga : Γ → C by ga(ω) = ρ(ω · a). Compute the Fourier transform of ga: for t ∈ Γ, ga(t) =ZΓ = ρ(cid:18)ZΓ ga(ω)hω, ti dβ (ω · a)hω, ti dω(cid:19) = ρ(Φt(a)) = 0. Since the Fourier transform is one-to-one, ga = 0. Taking ω = 1, we get ρ(a) = 0. As this does not depend on the choice of ρ, by the Hahn-Banach theorem, a ∈ B (cid:3) As an immediate corollary we get that {At}t∈Γ have dense span in A. Corollary 3.7. Suppose the compact abelian group Γ acts on the C ∗-algebra A. For t ∈ Γ, let At := {a ∈ A : β · a = hβ, ti a for every β ∈ Γ}. Then A = span{At : t ∈ Γ}. Remark 3.8. Lemmas 3.5 and 3.6 show that if Γ acts strongly on A, then A is topologically graded by Γ. In particular, when Σ → G is a Γ-graded twist, Lemma 2.26 shows that C ∗ r (Σ; G) is topologically graded by Γ. In [29, Theorem 3] the converse to Lemma 3.5 is proved: it is shown that if A is topologically graded by Γ, then there is a strongly continuous action of Γ on A such that a ∈ At if and only if a =ZΓ (ω · a)hω−1, ti dω. We now observe that the proof of Lemma 3.6 can be used to show that if span N (A, D) = A then span Nh(A, D) = A. Here are the details. 13 Proposition 3.9. Suppose Γ acts on A and that D is a MASA in A0. If n ∈ N , then for every t ∈ Γ, Φt(n) ∈ Nh and n ∈ span{Φt(n) : t ∈ Γ}. Proof. Fix n ∈ N . By Lemma 3.6 it suffices to show Φt(n) ∈ Nh. Let d ∈ D. Then Φt(n)∗dΦt(n) ∈ A0. For e ∈ D, and ω ∈ Γ, ω · e = e. So Φt(n)∗dΦt(nn∗n)e = Φt(n)∗dZΓ ω · n(ω · (n∗ne))hω−1, tidω = Φt(n)∗dZΓ = Φt(n)∗dnen∗Φt(n) = Φt(n)∗nen∗dΦt(n) =ZΓ = n∗neΦt(n)∗dΦt(n) = en∗nΦt(n)∗dΦt(n) = eΦt(nn∗n)∗dΦt(n). ω · (nen∗n)hω−1, tidω ω · (n∗nen∗)hω, tidωdΦt(n) This relation holds if we replace n∗n by a polynomial in n∗n and by taking limits we see that it holds if we replace n∗n by (n∗n)1/k for any k ∈ N. Since limk n(n∗n)1/k = n, we find that Φt(n)∗dΦt(n) commutes with every element of D. Since D is a MASA in A0, Φt(n)∗dΦt(n) ∈ D. A similar argument shows that Φt(n)dΦt(n)∗ ∈ D. So Φt(n) ∈ Nh. (cid:3) We now define a main object of study. Definition 3.10. Let A be C ∗-algebra topologically graded by a discrete abelian group Γ and D an abelian C ∗-subalgebra of A0. We say the pair (A, D) is Γ-Cartan if (1) D is Cartan in A0, (2) N (A, D) spans a dense subset of A. The following observations are simple but important. In particular, for Γ-Cartan pairs we may focus on homogeneous normalizers in place of more general normalizers. Lemma 3.11. Suppose (A, D) is a Γ-Cartan pair. The following statements hold. (1) The span of the homogeneous normalizers, Nh(A, D), is dense in A. (2) If (ei) is an approximate unit for A0, then (ei) is an approximate unit for A. (3) For any n ∈ N (A, D), n∗n and nn∗ belong to D. (4) Any approximate unit for D is an approximate unit for A. Proof. As noted in Remark 3.8, a topological grading arises from an action of a compact abelian group. By Proposition 3.9, Nh(A, D) spans a dense subset of A. Now suppose (ei) is an (not necessarily countable) approximate unit for A0. Let n ∈ Nh. Then nn∗ and n∗n belong to A0. Since (ei) is an approximate unit for A0, (3.12) (ein − n)(ein − n)∗ = einn∗ei − nn∗ei − einn∗ + nn∗ → 0, whence ein → n. Similarly, nei → n. Hence for any a ∈ span Nh, eia → a and aei → a. Since span Nh is dense in A, (ei) is an approximate unit for A. Since (A0, D) is a Cartan pair, D contains an approximate unit (ei) for A0. By part (2), (ei) is also an approximate unit for A. Then for any n ∈ N (A, D), D ∋ n∗ein → n∗n, so n∗n ∈ D. Likewise, nn∗ ∈ D. Finally, if (ei) is an approximate unit for D and n ∈ N , (3.12) together with the fact that nn∗ ∈ D, gives ein → n; likewise nei → n. As before, span N = A implies (ei) is an approximate unit for A. (cid:3) 4. Twists from Γ-Cartan pairs Throughout this section, we consider a fixed Γ-Cartan pair (A, D). The purpose of this section r (Σ; G) and D ∼= C0(G(0)). is to define a twist D × T → Σ → G from the pair (A, D) so that A ∼= C ∗ 14 This task is completed in Theorem 4.36. Our methods follow those found in Kumjian [18] and Renault [32], and also use techniques from Pitts [26]. Renault and Kumjian construct a twist from the Weyl groupoid associated to a Cartan pair by first considering its groupoid G of germs and then using the multiplicative structure of the normalizers to construct the twist as an extension Σ of G by T × G(0). Finally, they recognize Σ as a family of linear functionals on A. To a certain extent, we follow the Kumjian-Renault approach. We will define Σ and G in two ways. We first construct sets Σ and G using the Weyl groupoid (the topologies and groupoid operations come later). After doing so, we identify Σ as a family of linear functionals and G as as a family of (non-linear) functions on A. The product on Σ and G is obtained by translating the product on A to Σ utilizing the first approach, and the second approach makes defining the topologies on Σ and G straightforward. Viewing Σ and G as functions highlights the parallel between the Gelfand theory for commutative C ∗-algebras and relationship of the twist and the pair (A, D) more transparent. To begin, we fix some notation. Write X := D. We generally identify D with C0(X); thus for x ∈ X and d ∈ D, we write d(x) instead of d(x). Let E denote the faithful conditional expectation E : A0 → D. By [29] there is a corresponding strong action of Γ on A. We denote by Φt the completely contractive map Φt : A → At as defined in Equation (3.3). Set By Lemma 3.4, ∆ is a faithful conditional expectation of A onto D. For n ∈ N , Lemma 3.11 gives n∗n, nn∗ ∈ D; let ∆ := E ◦ Φ0. dom(n) := {x ∈ D : n∗n(x) > 0} and ran(n) := {x ∈ D : nn∗(x) > 0}. By the definition of normalizer, ndn∗ ∈ D for all d ∈ D. So Nh acts on D by conjugation. As D is abelian, this induces a partial action α on the spectrum. The following result of Kumjian gives a precise description of this action. Proposition 4.1. [18, Proposition 1.6] Let n ∈ N . Then there exists a unique partial homeomor- phism αn : dom(n) → ran(n) such that for each d ∈ D and x ∈ dom(n), When the action is clear from the context, we will sometimes write, (n∗dn)(x) = d(αn(x)) (n∗n)(x). n.x := αn(x). By [32, Lemma 4.10] (or [18, Corollary 1.7]), for n, m ∈ N and d ∈ D we have αn ◦ αm = αmn, αn∗ = α−1 n , and αd = idsupp′(d) . The collection {αn : n ∈ N } is an inverse semigroup, sometimes called the Weyl semigroup of the inclusion (A, D). Dual to the Weyl semigroup is a collection of partial automorphisms {θn : n ∈ N } of D. Given n ∈ N , nn∗D and n∗nD are ideals of D whose Gelfand spaces may be identified with ran(n) and dom(n) respectively. By [27, Lemma 2.1], the map nn∗D ∋ d 7→ n∗dn ∈ n∗Dn extends uniquely to a ∗-isomorphism θn : nn∗D → n∗nD such that for every d ∈ nn∗D, (4.2) dn = nθn(d) and for every x ∈ dom(n), (4.3) θn(d)(x) = d(αn(x)). 15 Lemma 4.4. Suppose n ∈ Nh(A, D) and x ∈ X such that ∆(n)(x) 6= 0. Then x is in the interior of the set of fixed points of αn and there exists h ∈ D such that h(x) = 1 and nh = hn ∈ D. Proof. First note n ∈ A0 because 0 6= ∆(n)(x) = E(Φ0(n))(x), and thus Φ0(n) 6= 0. Furthermore, x ∈ dom(n) and [26, Lemma 2.5] gives αn(x) = x. We claim that x is actually in the interior of the set of fixed points of αn. If not, then there exists a net (xi) in dom(n) such that αn(xi) 6= xi and xi → x. Then ∆(n)(xi) → ∆(n)(x) 6= 0. However, by [26, Lemma 2.5] again, ∆(n)(xi) = 0 for all i, a contradiction. Now let F be the interior of the set of fixed points of αn and J := {d ∈ D : supp d ⊆ F }. For S ⊆ D let S⊥ = {a ∈ D : ax = 0 for all x ∈ S}. Note that J ⊥⊥ is the fixed point ideal K0 for n (see [26, Definition 2.13]). Then by [26, Lemma 2.15] there exists h ∈ D with h(x) = 1 and nh = hn ∈ D′. But n ∈ A0 and h ∈ D, so nh ∈ A0 ∩ D′ = D because D is maximal abelian in A0. Thus nh = hn ∈ D. This completes the proof. (cid:3) The following is an interesting structural fact about the relationship between ∆ and the action of Nh on D, which is used when defining the inverse operation on Σ. Proposition 4.5. For any n ∈ Nh and a ∈ A, n∗∆(a)n = ∆(n∗an). Proof. To begin, we claim that for m ∈ Nh, (4.6) n∗∆(nm)n = n∗n∆(mn). Since the terms on both sides of (4.6) belong to D, it suffices to show that for every x ∈ X (4.7) As n = limk→∞ n(n∗n)1/k, both sides of (4.7) vanish if (n∗n)(x) = 0. Thus to obtain (4.7) it suffices to prove that for x ∈ dom(n) (n∗∆(nm)n)(x) = (n∗n∆(mn))(x). (4.8) ∆(nm)(αn(x)) = ∆(mn)(x), and this is what we shall do. Suppose first that ∆(nm)(αn(x)) 6= 0. Lemma 4.4 shows there exists k ∈ D with k(αn(x)) = 1 and nmk = knm ∈ D. Then ∆(nm)(αn(x)) = (k∆(nm))(αn(x)) = ∆(knm)(αn(x)) = (knm)(αn(x)) = (n∗(knm)n)(x) (n∗n)(x) = ∆((n∗kn)mn)(x) (n∗n)(x) = (n∗kn)(x) (n∗n)(x) ∆(mn)(x) = k(αn(x))∆(mn)(x) = ∆(mn)(x). Next, suppose ∆(mn)(x) 6= 0 and put y = αn(x). We do a similar calculation. Another application of Lemma 4.4 produces h ∈ D with h(x) = 1 and mnh = hmn ∈ D. As x = αn∗(y), ∆(mn)(x) = (mnh)(αn∗ (y)) = (n(mnh)n∗)(y) (nn∗)(y) = ∆(nm(nhn∗))(y) (nn∗)(y) = ∆(nm)(y) (nhn∗)(y) (nn∗)(y) = ∆(nm)(αn(x)). = ∆(nm)(αn(x))h(αn∗ (y)) = ∆(nm)(αn(x))h(x) We have shown that ∆(mn)(x) 6= 0 if and only if ∆(nm)(αn(x)) 6= 0, and, when this occurs, ∆(mn)(x) = ∆(nm)(αn(x)). Thus (4.8) holds, completing the proof of the claim. 16 By varying m and using the facts that n∗n ∈ D (Lemma 3.11(3)) and span Nh = A, (4.6) implies that for every a ∈ A, n∗∆(na)n = (n∗n)∆(an) = ∆(n∗(na)n). Therefore, for every a ∈ nA, n∗∆(a)n = ∆(n∗an). Given k ∈ N, there exists a sequence of polynomials {pj} each of which vanish at the origin such that (nn∗)1/k = limj pj(nn∗). Thus, for a ∈ A and k ∈ N, (nn∗)1/ka(nn∗)1/k ∈ nA. Hence for each a ∈ A, n∗∆(a)n = lim k n∗(nn∗)1/k∆(a)(nn∗)1/kn = lim k ∆(n∗(nn∗)1/ka(nn∗)1/kn) = ∆(n∗an). n∗(cid:16)∆((nn∗)1/ka(nn∗)1/k)(cid:17) n k This completes the proof. = lim (cid:3) 4.1. Local equivalence relations from homogeneous normalizers. Let G := {(n, x) ∈ Nh × X : n∗n(x) 6= 0}. We now define two equivalence relations on G arising as germs of the subsemigroup of the Weyl semigroup arising from homogeneous normalizers. While we shall define the groupoids Σ and G in the twist Σ → G as functions on A, the equivalence relations below will enable us to define the multiplicative structure on Σ and G. Definition 4.9. For (n, x), (n′, x′) ∈ G. Consider (1) x = x′, (2Σ) there exist d, d′ ∈ Cc(X) such that d(x) > 0, d′(x) > 0 and nd = n′d′, (2G) there exist d, d′ ∈ Cc(X) such that d(x) 6= 0, d′(x) 6= 0 and nd = n′d′. By (4.2) and (4.3), the latter two conditions may equivalently be replaced with the following conditions. (2Σ (2G ′) There exist d, d′ ∈ Cc(X) such that d(αn(x))) > 0, d′(αn(x))) > 0 and dn = d′n′. ′) There exist b, b′ ∈ Cc(X) such that d(αn(x))) 6= 0, d′(αn(x))) 6= 0 and dn = d′n′. Note that in conditions (2Σ) and (2G), we may assume that d, d′ ∈ n∗nD ∩ n′∗n′D; likewise we may assume d, d′ ∈ nn∗D ∩ n′n′∗D in conditions (2Σ ′) and (2G ′). Define ∼Σ as the relation given by (1) and (2Σ) and ∼G as the relation given by (1) and (2G). We omit the proof that these are equivalence relations. We denote the equivalence classes by [n, x]Σ, [n, x]G respectively. We shall omit the subscript when the proof does not depend on which relation is used. Following Renault [32], define ΣA,D,Γ := G/ ∼Σ and GA,D,Γ := G/ ∼G . We omit the A, D, Γ from the notation and write Σ and G respectively when the inclusion and grading are clear from context. Essentially, G is a modification of the groupoid of germs of the α action and Σ is a twist on this. The following is a useful observation. Lemma 4.10. For i = 1, 2 suppose (ni, xi) ∈ G and [n1, x1] = [n2, x2]. Then n∗ 1n2 ∈ A0. Proof. We do this only for ∼G, leaving the obvious modifications for ∼Σ to the reader. By definition of ∼G, x1 = x2 =: x and there exist d1, d2 ∈ D with di(x) 6= 0 and n1d1 = n2d2. Then nidi ∈ Nh 1n1d1 is a non-zero element of D. Since ni ∈ Nh, there exists t ∈ Γ such that and d∗ n∗ 1n2 ∈ At. But At is a D-bimodule, so d∗ It is useful to have an alternative description of the equivalence relations ∼Σ and ∼G before 1n2d2 ∈ A0 ∩ At, whence t = 0. 1n2d2 = d∗ 1n∗ 1n∗ 1n∗ (cid:3) continuing. 17 Definition 4.11. For (n, x), (n′, x′) ∈ G, consider the properties (i) x = x′, (iiΣ) ∆(n∗n′)(x) > 0, and (iiG) ∆(n∗n′)(x) 6= 0. Define ≈Σ to be the relation on G given by (i) and (iiΣ), and define ≈G be the relation on G given by (i) and (iiG). Proposition 4.12. The relations ≈Σ and ∼Σ are the same. Likewise, the relations ≈G and ∼G are the same. Proof. We prove ≈Σ and ∼Σ are the same. The proof for ≈G and ∼G is similar. Suppose (n, x) ≈Σ (n′, x′). Then x = x′ and ∆(n∗n′)(x) > 0. Since n, n′ are homogeneous normalizers, so is n∗n′. Lemma 4.4 implies there is an h ∈ D such that hn∗n′ = n∗n′h ∈ D and h(x) > 0. Now consider the equalities, n(n∗n′h)((n′h)∗n′h) = nn∗[((n′h)(n′h)∗)n′h] = n′h(nn∗ ◦ α(n′h)∗)(n′h)∗(n′h). Take d = (n∗n′h)((n′h)∗n′h) and d′ = h(nn∗ ◦ α(n′h)∗ )(n′h)∗(n′h), so that nd = n′d′. Note d(x) = n∗n′(x)h(x)((n′h)∗n′h)(x) > 0 and d′(x) = h(x)(nn∗ ◦ α(n′h)∗)(x)((n′h)∗(n′h))(x) > 0. Thus (n, x) ∼Σ (n′, x). The converse follows immediately from the definitions. (cid:3) 4.2. Viewing Σ as linear functionals. Our next goal is to show that Σ may be identified as a family of linear functionals on A and G as a family of functions on A. This highlights the role of the inclusion (A, D) in producing Σ and G and will allow us to easily define Hausdorff topologies on Σ and G. In addition, for a ∈ A we will define a : Σ → C by a([n, x]) = [n, x](a). The main result of this section shows that the map a 7→ a is an isomorphism A ∋ a 7→ a ∈ C ∗ r (Σ; G) which in a natural sense extends the the Gelfand transform. We write A# for the Banach space dual of A. For f ∈ A#, let f ∗ ∈ A# be defined by A ∋ a 7→ f (a∗) and let f be the function on A defined by f (a) = f (a). For a non-empty subset K ⊆ A#, write K := {f : f ∈ K}. Equip K with the relative weak-∗ topology and K with the quotient topology arising from the surjective map, K ∋ f 7→ f . Then K and K are Hausdorff. Put S := {x ◦ ∆ : x ∈ X}, so S consists of all states of the form A ∋ a 7→ ∆(a)(x). Then S is a family of state extensions of pure states on D to all of A. We make the following observations. Observations 4.13. (1) With the relative weak-∗ topology (i.e. the σ(A#, A)-topology) on S, the restriction map, S ∋ ψ 7→ ψD is a homeomorphism of S onto X. (2) Lemma 4.4 implies that if ψ ∈ S, then for every n ∈ Nh, (4.14) ψ(n)2 ∈ {0, ψ(n∗n)}. This condition is a variant of the notion of compatible state introduced in [27], the difference being that (4.14) is required to hold only for elements of Nh rather than all of N as in [27]. The compatibility condition (4.14) implies (using [27, Proposition 4.4(iii)] and the Cauchy- Schwartz inequality) that in the GNS representation (πψ, Hψ) associated to ψ, the set of vectors V := {n + Lψ : n ∈ Nh} has the property that any two vectors in V are either orthogonal or parallel; here Lψ is the left kernel of ψ, Lψ := {a ∈ A : ψ(a∗a) = 0}. Notice also that span V is dense in Hψ. 18 For any (n, x) ∈ G, define an element of A# by (4.15) ψ(n,x)(a) := ∆(n∗a)(x) n(x) . Simple calculations show that(cid:13)(cid:13)ψ(n,x)(cid:13)(cid:13) = 1 and for d1, d2 ∈ D and a ∈ A, ψ(n,x)(d1ad2) = d1(αn(x))ψ(n,x)(a)d2(x). In other words, in the language of [12, Section 2], ψ(n,x) is a norm-one eigenfunctional with source s(ψ(n,x)) = x and range r(ψ(n,x)) = αn(x). Furthermore, observe that (n, x) ∈ G ⇔ (n∗, αn(x)) ∈ G and a calculation using Proposition 4.5 shows that for (n, x) ∈ G, (4.16) ψ∗ (n,x) = ψ(n∗,αn(x)). For later use, notice that for d ∈ D with d(x) > 0 and z ∈ T, (4.17) Let (4.18) ψ(nd,x) = ψ(n,x) and ψ(zn,x) = zψ(n,x). E := {ψ(n,x) : (n, x) ∈ G}. Since a state ψ on a C ∗-algebra B is uniquely determined by ψ, it also makes sense to define source and range maps on E by s(ψ(n,x)) = x and r(ψ(n,x)) = αn(x). Then the source and range maps carry E and E onto X. Given ψ ∈ E, write ψ = ψ(n,x) ∈ E, and choose m ∈ Nh such that ψ(m) 6= 0. Notice that for any a ∈ A, we have (4.19) ∆(a)(x) = ψ(ma) ψ(m) and ∆(a)(m.x) = ψ(am) ψ(m) . (Indeed, since ψ(m) 6= 0, Lemma 4.4 gives n ∼G m, and a computation gives (4.19).) Setting (4.20) s(ψ) := ψ(ma) ψ(m) and r(ψ) = ψ(am) ψ(m) , then s(ψ) = s(ψ) ◦ ∆ and r = r(ψ) ◦ ∆. Thus s(ψ) and r(ψ) are the (necessarily unique) elements of S satisfying s(ψ)D = s(ψ) and r(ψ)D = r(ψ). Also, notice that S ⊆ E, for if ψ = x ◦ ∆ ∈ S, then ψ = ψ(d,x) for any d ∈ D with d(x) > 0. Also, it follows easily (using Lemma 4.4) that (4.21) S = {ψ(n,x) : ∆(n)(x) > 0} = {ψ(d,x) : d ∈ D and d(x) > 0}. We list a few additional properties of E and E. Lemma 4.22. The following statements hold. (1) The map ψ(n,x) 7→ [n, x]Σ is a well-defined bijection of E onto Σ. (2) If g ∈ E and m ∈ Nh satisfies g(m) > 0, then g = ψ(m,s(g)). (3) E ∪ {0} is weak-∗ compact; in particular E is locally compact. Furthermore, s, r are contin- uous mappings of E onto X. (4) The map ψ(n,x) 7→ [n, x]G is a well-defined bijection of E onto G. (5) If φ ∈ E and m ∈ Nh satisfies φ(m) 6= 0, then φ = ψ(m,s(φ)). (6) E is locally compact and s, r are continuous mappings of E onto X. 19 Proof. We prove statements (1), (2) and (3), leaving the others to the reader. To establish the first, it suffices to show ψ(n1,x1) = ψ(n2,x2) if and only if [n1, x1]Σ = [n2, x2]Σ and this is what we do. Suppose ψ(n1,x1) = ψ(n2,x2). Applying the source map gives x1 = x2; write x := x1 = x2. Now 0 < ψ(n2,x)(n2) = ψ(n1,x)(n2) = ∆(n∗ 1n2)(x) n1(x) , which implies ∆(n∗ 1n2)(x) > 0. Proposition 4.12 now gives [n1, x1]Σ = [n2, x2]Σ. Conversely, if [n1, x1]Σ = [n2, x2]Σ, then x1 = x2 =: x and there exists d1, d2 ∈ D with d1(x) > 0 and d2(x) > 0 such that n1d1 = n2d2. Then ψ(n1,x1) = ψ(n1,x) = ψ(n1d1,x) = ψ(n2d2,x) = ψ(n2,x2). This gives statement (1). For statement (2), write φ = ψ(n,x) and apply Proposition 4.12 and part (1). Turning now to statement (3), s(ψ(n,x)) = x◦∆ and r(ψ(n,x)) = αn(x)◦∆, so the maps s, r : E → S are surjective. They are continuous by (4.20), so s, r : E → X are also continuous surjections. Next suppose that ψ(ni,xi) is a net in E converging weak-∗ to φ ∈ A#; write ψi := ψ(ni,xi). If φ = 0, there is nothing to do. So suppose φ 6= 0. By [12, Proposition 2.3], φ is an eigenfunctional. Thus if x := s(φ), continuity of s yields xi → x. Since span Nh is dense in A, there exists n ∈ Nh such that φ(n) > 0. Since n(n∗n)1/k → n as k → ∞, we have 0 < φ(n) = limk φ(n(n∗n)1/k) = φ(n) limk(n∗n)1/k(x). Thus n∗n(x) 6= 0 and so (n, x) ∈ G. Since ψi → φ, ψi(n) is eventually non-zero, so we may as well assume that ψi(n) 6= 0 for every λ. Proposition 4.12 implies (ni, xi) ∼G (n, xi). Hence there exists zi ∈ T such that ψi = ψ(zin,xi) = zi · ψ(n,xi). Therefore, 0 < φ(n) = lim zi ∆(n∗n)(xi) n(xi) = lim zi n(xi). As n(xi) → n(x), we conclude zi → 1. It follows that φ = lim ψi = lim zi · ψ(n,xi) = ψ(n,x), so φ ∈ E. Thus E ∪ {0} is a closed subset of the unit ball of A#, and hence is compact. (cid:3) Notation 4.23. We use the bijections of Lemma 4.22 to identify Σ with E (respectively G with E) and will use E and Σ interchangeably (resp. G and E) depending upon what is convenient for the context. Thus for a ∈ A, we will often write [n, x]Σ(a) and [n, x]G(a) instead of ψ(n,x)(a) and ψ(n,x)(a). Then Σ and G become Hausdorff topological spaces of functions on A. When convenient, we will also identify S with X via the restriction mapping from 4.13(1). 4.3. The twist associated to a Γ-Cartan pair. We are now prepared to place groupoid struc- tures on G and Σ. This is done exactly as in [26, Definition 8.10 and Theorem 8.12]; for convenience, we provide sketches of the proofs using the present notation. Lemma 4.24. Σ and G are Hausdorff topological groupoids under the following operations: • Multiplication: [m, αn(x)][n, x] = [mn, x]; • Inversion: [n, x]−1 = [n∗, αn(x)]. The map x 7→ [d, x] for d ∈ D with d(x) > 0 identifies X with the unit space of Σ and G. Furthermore, under this identification r([n, x]) = αn(x) and s([n, x]) = x. Proof. We sketch the proof for Σ. The proof for G is left to the reader (details may be found in [26]). That inversion is well-defined and continuous follows from (4.15), (4.16), and Lemma 4.22. Also, it is clear that inversion is involutive. Next we show multiplication is well-defined. Suppose [m1, y]Σ = [m2, y]Σ and [n1, x]Σ = [n2, x]Σ. Using the bijection in Lemma 4.22 we can identify ψ and φ with [m1, y]Σ = [m2, y]Σ and [n1, x]Σ = 20 [n2, x]Σ respectively. We have y = αni(x). By the definition of ∼Σ we can assume m2 = m1d and n2 = n1d′ where d(y) > 0 and d′(x) > 0. So to show that multiplication is well defined it suffices to show that ψ(m1n1,x) = ψ(m1dn1d′,x). But this follows since m1dn1d′ = m1n1θn1(d)d′ and we know from equation (4.17) that ψ(νb,x) = ψ(ν,x) for all ν ∈ Nh, x ∈ dom(ν) and b ∈ D with b(x) > 0. Multiplication is associative since multiplication in the C ∗-algebra is. Suppose [m, x], [n, y] ∈ Σ are such that the composition [m, x][n, y] is defined. Then x = αn(y). We must show that [m, x][n, y][n∗, αn(y)] = [m, x] and [m∗, αm(y)][m, x][n, y] = [n, y]. But these equalities follow from Lemma 4.22(2) because [mnn∗, αn(y)](m) > 0 and [m∗mn, y](n) > 0. This completes the proof that Σ is a groupoid when equipped with the indicated operations. Since Σ(0) = {[m, x]−1 Σ [m, x]Σ : (m, x) ∈ G} we obtain Σ(0) = {[d, x]Σ ∈ Σ : d ∈ D and d(x) > 0} = S. It follows that the map X ∋ x 7→ [d, x]Σ where d ∈ D is chosen so that d(x) > 0, is a bijection of X onto Σ(0). Similarly, the map X ∋ x 7→ [d, x]G where d ∈ D satisfies d(x) > 0 (or merely satisfies d(x) 6= 0)) is a bijection of X onto G(0). For (n, x) ∈ G, r([n, x]) = [nn∗, αn(x)] and s([n, x]) = [n∗n, x]). This gives the desired identifi- cation of the range and source maps. We have already observed that inversion is continuous and we now verify that multiplication is continuous. Let E(2) be the set of composable pairs, that is, the collection (ψ, φ) ∈ E × E with s(ψ) = r(φ). Suppose (φi)i∈I and (ψi)i∈I are nets in E converging to φ, ψ ∈ E respectively, and such that (φi, ψi) ∈ E(2) for all λ. Since s and r are continuous, we find that s(φ) = limi s(φi) = limi r(ψi) = r(ψ), so (φ, ψ) ∈ E(2). Let n, m ∈ Nh be such that φ(n) > 0 and ψ(m) > 0. There exists i0, so that i ≥ i0 implies φi(n) and ψi(m) are non-zero. For each i ≥ i0, there exist scalars λi, λ′ i ∈ T such that φi = λi[n, s(φi)] and ψi = λ′ i[m, s(ψi)]. Since lim i φi(n) = φ(n) = lim i lim ψi(n) = ψ(n) = lim i [n, s(ψi)](n), we conclude that lim λi = 1 = lim λ′ [n, s(φi)](n) and i i. So for any a ∈ A, (φψ)(a) = s(ψ)((nm)∗a) (s(ψ)((nm)∗(nm)))1/2 = lim i s(ψi)((nm)∗a) (s(ψi)((nm)∗(nm)))1/2 = lim i [n, s(φi)][m, s(ψi)] = lim i (φiψi)(a), giving continuity of multiplication. Define q : Σ → G and ι : T × G(0) → Σ by (cid:3) q([n, x]Σ) := [n, x]G and ι(λ, [d, x]G) = [λd, x]Σ. Then q and ι are continuous groupoid homomorphisms with q surjective and ι injective. Moreover, q−1(G(0)) = {[d, x]Σ : d ∈ D and d(x) 6= 0} = ι(T × G(0)). Furthermore, for (n, x) ∈ G, and λ ∈ T, ι(λ, [nn∗, αn(x)]G) [n, x]Σ = [λn, x]Σ = [n, x]Σ ι(λ, [n∗n, x]G). We thus have a central extension of groupoids, T × G(0) q → G. ι ֒→ Σ 21 (4.26) [d, x]Σ 7→(cid:18) d(x) d(x) , x(cid:19) . Also, for λ ∈ T and (n, x) ∈ G, (4.25) λ · [n, x]Σ = [λn, x]Σ. As G(0) may be identified with X, we usually identify ι(T × G(0)) with T × X by Under this identification, the extension of groupoids above becomes T × X ֒→ Σ q → G. Remark 4.27. We have already seen an action of T on Σ: λ · [n, x]Σ = [λn, x]Σ. When elements of Σ are identified with their corresponding elements of E via the map in Lemma 4.22, there is another action of T on Σ, namely scalar multiplication of linear functionals. These actions differ: if scalar multiplication of linear functionals is denoted by juxtaposition, then For n ∈ Nh, let λ[n, x]Σ = λ · [n, x]Σ. Z(n) := {[n, x]G : x ∈ dom(n)}. Lemma 4.28. For each n ∈ Nh, Z(n) is an open bisection for G and {Z(n) : n ∈ Nh} is a base for the topology on G. Moreover, q−1(Z(n)) is homeomorphic to T × Z(n). In particular, G is an ´etale groupoid and the bundle Σ → G is locally trivial. Proof. The relevant definitions and an application of Lemma 4.12 yield q−1(Z(n)) = {[m, y]Σ ∈ Σ : [m, y](n) 6= 0}, which is an open subset of Σ. Thus Z(n) is an open subset of G. We claim rZ(n) and sZ(n) are homeomorphisms of Z(n) onto ran(n) and dom(n) respectively. As r is the composition of the source map with the inversion map, it suffices to show this for s only. First note that sZ(n) : Z(n) → dom(n) is a bijection by definition. By Lemma 4.22(6), s, r : G → X are continuous. Next we show (sZ(n))−1 is a continuous function from dom n to Z(n). If xi ∈ dom n is a net and xi → x ∈ dom n, then ψ(n,xi) → ψ(n,x) (by definition of the weak-* topology), so [n, xi]G → [n, x]G by Lemma 4.22(4). Thus the claim holds, and Z(n) is therefore an open bisection. Let U ⊆ G be open and choose [n, x]G ∈ U . Then V := U ∩ Z(n) is an open bisection, so s(V ) is an open subset of X containing x. Let d ∈ D be such that supp d ⊆ s(V ) and d(x) = 1. Since dom(nd) = dom(n) ∩ supp′(d) ⊆ s(V ), we find [n, x]G ∈ Z(nd) = s−1(dom(nd)) ⊆ V ⊆ U. Thus, {Z(n) : n ∈ Nh} is a base for the topology on G. As {Z(n) : n ∈ Nh} covers G, G is ´etale. Consider the map τ : T × dom(n) → q−1(Z(n)) defined by (z, x) 7→ [zn, x]Σ. This map is a homeomorphism, and as sZ(n) : Z(n) → dom(n) is a homeomorphism, we see Σ → G is locally trivial. (cid:3) The following summarizes our discussion so far. Proposition 4.29. Σ and G are locally compact Hausdorff topological groupoids, G is ´etale, and T × X ֒→ Σ q ։ G is a twist. Define a map gr : Nh → Γ by by taking n ∈ At to t. This induces maps cΣ : Σ → Γ and cG : G → Γ given by (4.30) cΣ([n, x]Σ) = gr(n) and cG([n, x]G) = gr(n). 22 Notice that the definition of the topologies and the groupoid multiplications imply that cΣ and cG are continuous homomorphisms. We therefore have produced the graded twist, T × G(0) / Σ ❄❄❄❄❄❄❄❄ cΣ / G cG Γ. 4.4. Every Γ-Cartan pair is a twisted groupoid C ∗-algebra. For a ∈ A, define a function a : Σ → C by [n, x]Σ 7→ , that is, a([n, x]Σ) = ψ(n,x) (a). ∆(n∗a)(x) (n∗n)1/2(x) By construction, a is a continuous function on Σ, and for z ∈ T, a(z · [n, x]Σ) = za([n, x]Σ), so a may be regarded as a continuous section of the line bundle over the twist Σ → G. Thus we may regard the open support of a as a subset of G, as in Remark 2.12. Lemma 4.31. Suppose n ∈ Nh, then the open support of n, supp′(n), is the set Z(n) Proof. Consider n for n ∈ Nh. Then n[m, y] = ∆(m∗n)(y) (m∗m)1/2(y) . This is zero unless ∆(m∗n) 6= 0. Now Proposition 4.12 gives [m, y]G = [n, y]G, that is [m, y]G ∈ Z(n). (cid:3) Lemma 4.32. The map Ψ : A → C(Σ; G) given by a 7→ a is linear and injective. Proof. This map is linear since ∆ is. Injectivity will follow since span Nh is dense in A. Indeed, suppose a ≡ 0. Then for all n ∈ Nh, ∆(n∗a)(y) = 0 for all y ∈ Dom(n). Thus ∆(n∗a) = 0 for all n ∈ Nh. By assumption a ∈ span(Nh); take a net νi ∈ span(Nh) such that νi → a. By linearity, Thus by continuity ∆(a∗a) = 0. Since ∆ is faithful, a = 0. (cid:3) ∀i ∆(ν∗ i a) = 0. Now let Nh,c := {n ∈ Nh : supp n is compact} and Ac := span Nh,c (no closure). Note that Ac is a ∗-algebra and by Lemma 4.31, for a ∈ Ac, a ∈ Cc(Σ; G). Lemma 4.33. Let (A, D) be a Γ-Cartan pair. Then (1) Nh,c is dense in Nh. (2) Ac is dense in A. (3) Ψ : a 7→ a sends Ac bijectively onto Cc(Σ; G) and Dc = D ∩ Ac bijectively onto Cc(X). (4) Ψ is a ∗-algebra homomorphism. Proof. For (1), let (ei) be an approximate unit for D with ei ∈ Cc(X) for every i. By Lemma 3.11(4), (ei) is also an approximate unit for A. Thus n = lim nei. So to prove Nh,c is dense it suffices to show that nd ∈ Nh,c for all d ∈ Cc(X). Given d ∈ Cc(X), let d1 ∈ Cc(X) be such that supp′(d1) ⊇ supp(d). By Lemma 4.31, supp′(dnd1) = Z(nd1). Recalling that sZ(nd1) is a homeomorphism of 23 / / Z(nd1) onto dom(nd1), we see that Z(nd) is compact because Z(nd) ⊆ Z(nd1) and dom(nd) = supp′(n∗n) ∩ supp′(d) has compact closure in dom(nd1). Now (2) follows immediately from (1). Lemma 4.32 shows that Ψ is injective and Ψ(Dc) = Cc(S) ≃ Cc(X), so to obtain (3), we must show Ψ(Ac) = Cc(Σ, G). Now Cc(Σ; G) is the span of sections of the line bundle supported on sets of the form Z(n), as the Z(n) form a basis for G and we can use a partition of unity argument. Thus it suffices to show that for n ∈ Nh, every f ∈ Cc(Σ; G) with support in Z(n) is in the image of Ψ. To proceed note the following. i. The line bundle is trivial over Z(n): this is true because q−1(Z(n)) = {z · [n, x]Σ : x ∈ dom(n), z ∈ T} and the map T × Z(n) ∋ (z, [n, x]G) 7→ [zn, x]Σ is a homeomorphism of T × Z(n) onto q−1(Z(n)). ii. The source map of G sends Z(n) homeomorphically to {x : n∗n(x) 6= 0} because Z(n) is an open bisection. Now let f be a section of the line bundle supported on Z(n). By the first item above we can (n∗n)1/2 . We view f as a function. By item (ii), f = d ◦ (sZ(n)) for some d ∈ D. Now take a = nd show a = f . Indeed, a[m, x] = ∆(dn∗m)(x) (n∗n)1/2(x)(m∗m)1/2(x) , which is 0 unless the germ of m is the same as n. So we can assume that n = m. Hence the above becomes Thus a = f and the claim holds. Part (3) now follows. n∗n(x) d(x)n∗n(x) = d(x) = f ([n, x]) m, n ∈ Nh,c. Using (2.10) we compute: It remains to show (4). By linearity it is enough to check thatdmn = m ∗ n and cm∗ = ( m)∗ for m ∗ n([ν, x]Σ) = m([v, y]Σ) n([w, x]Σ) X[v,y]Σ[w,x]Σ=[ν,x]Σ (again, for each factorization [v, y]G[w, x]G = [ν, x]G only one factorization [v, y]Σ[w, x]Σ = [ν, x]Σ is chosen). But m([v, y]Σ) n([w, x]Σ) = 0 unless [v, y]G = [m, y]G and [w, x]G = [n, x]G. When this occurs, there exist zv, zw ∈ T so that [v, y]Σ = [zvm, y]Σ and [w, x]Σ = [zwn, x]Σ. As [v, y]Σ[w, x]Σ = [ν, x]Σ, we have y = αn(x) and [ν, x]Σ = [zvzwmn, x]Σ. So m ∗ n([ν, x]Σ) = m([zvm, y]Σ)n([zwn, x]Σ) as desired. = zvzwp(m∗m)(y)(n∗n)(x) = zvzw(cid:18) (n∗m∗mn)(x) =dmn([zvzwmn, x]Σ) =dmn([ν, x]Σ), To see cm∗ = ( m)∗, use Lemma 4.22 and (4.16): cm∗([n, x]Σ) = ψ(n,x)(m∗) = ψ∗ Σ ) = ( m)∗([n, x]Σ). (n,x)(m) = ψ(n∗,αn(x))(m) = m([n, x]−1 (n∗n)(x) (n∗n)(x)(cid:19)1/2 (cid:3) Lemma 4.34. When Cc(Σ; G) is equipped with the reduced norm, ΨAc : Ac → Cc(Σ; G) is an isometric ∗-isomorphism. 24 Proof. Lemma 4.33 gives ΨAc is a ∗-isomorphism of Ac onto Cc(Σ; G). Fix x ∈ X. Using Remark 2.14, we may regard πx as the GNS representation of Cc(Σ; G) arising from the functional εx. On the other hand, the state ρx := x◦∆ determines the GNS representation (πρx, Hρx ) of A. Let Lx ⊆ C ∗ r (Σ; G) and Lρx ⊆ A be the left kernels of εx and ρx respectively. We claim that for n, m ∈ Nh,c, (4.35) ρx(mn) = εx( m ∗ n). To see this, choose d ∈ D with d(x) = 1, so that [d, x]Σ ∈ Σ(0). For [n, x]Σ with [n, x]−1 [d, x]Σ, a computation similar to that used in the proof of Lemma 4.33(4) gives Σ [n, x]Σ = εx( m ∗ n) = m ∗ n([d, x]Σ) = m([n∗, αn(x)])n([n, x]Σ) = ∆(nm)(αn(x)) p(nn∗)(αn(x)) (n∗n)(x)1/2 = ∆(nm)(αn(x)) (4.8) = ∆(mn)(x) = ρx(mn). For a ∈ Ac, (4.35) gives Thus for a ∈ Ac, the map a+Lρx 7→ a+Lx extends to an isometry Wx : Hρx → Hx. Lemma 4.33(3) implies that Wx is onto. For m, n ∈ Nh, ρx(a∗a) = εx(ba∗ ∗ a). Wxπρx(m)(n + Lρx) = Wx(mn + Lρx) = m ∗ n + Lx = πx( m)Wx(n + Lρx). It follows that Wxπρx(m)W ∗ x = πx( m). Hence for a ∈ Ac, Wxπρx(a)W ∗ x = πx(Ψ(a)). Finally, for a ∈ Ac, kΨ(a)kC ∗ r (Σ;G) = sup x kπx(Ψ(a))k = sup x kπρx(a)k = kakA , with the last equality following from the fact that ∆ is faithful (as in the proof of Lemma 4.32). (cid:3) We now come to the main result of this section. Theorem 4.36. Let (A, D) be a Γ-Cartan pair. Then there exists a graded twist T × G(0) / Σ / G ❃❃❃❃❃❃❃❃ cΣ cG Γ and a Γ-covariant ∗-isomorphism Ψ : A → C ∗ r (Σ; G) such that Ψ(D) = C0(G(0)). Proof. Lemmas 4.33 and 4.34 show that Ψ determines a ∗-isomorphism of A onto C ∗ r (Σ; G) and the construction of the graded twist shows Ψ is Γ-covariant. It remains to show that Ψ(D) = C0(X). For d ∈ D and [n, x]Σ ∈ Σ, d([n, x]Σ) = ∆(n∗)(x) n(x) 25 d(x). / / Changing perspective to viewing d as a section of the line bundle instead of as a covariant function and recalling that [n, x]G ∈ G(0) if and only if ∆(n)(x) 6= 0, we get n(x) d(x), [n, x]Σ(cid:21) d([n, x]G) =(cid:20) ∆(n∗)(x) (0 hd(x), [ ∆(n∗)(x) (2.3), (4.25) = n(x) n, x]Σi if [n, x]G /∈ G(0) if [n, x]G ∈ G(0). As [ ∆(n∗)(x) represent the same element of the line bundle L. Thus n(x) n, x]Σ ∈ Σ(0), under the identification of X with Σ(0),hd(x), [ ∆(n∗)(x) n(x) n, x]Σi and [d(x), x] d([n, x]G) =(0 [d(x), x]] if [n, x]G /∈ G(0) if [n, x]G ∈ G(0), showing that Ψ(d) ∈ C0(X). On the other hand, if f ∈ Cc(Σ; G) vanishes off G(0), define d ∈ D as follows. For x ∈ X, choose n ∈ Nh so that (n∗n)(x) 6= 0; then let d(x) be the unique scalar satisfying f ([n∗n, x]G) = [d(x), x] ∈ L. Then d(x) = f . It follows that C0(X) = Ψ(D) and the proof is complete. (cid:3) 5. Γ-Cartan Pairs from Γ-Graded Twists In the previous section, we associated a graded twist (Σ, G, Γ) to a Γ-Cartan pair (A, D) and showed that (A, D) can be recovered from (Σ, G, Γ). The purpose of this section is to produce a Γ-Cartan pair from any suitable twist graded by the abelian group Γ. Throughout this section we assume the following: Assumptions 5.1. We fix a Γ-graded twist (5.2) T × G(0) / Σ q / G ❃❃❃❃❃❃❃❃ cΣ cG Γ, with G ´etale (and Hausdorff) where the diagram commutes and (1) Γ is a discrete abelian group; (2) cG and cΣ are (continuous) groupoid homomorphisms; and (3) c−1 G (0) is effective. The homomorphisms cG and cΣ are often called cocycles in the literature as they are elements of the first groupoid cohomology group with coefficients. We will persist in referring to cΣ and cG as cocycles here. For notational convenience, let The commutativity of (5.2) yields P := c−1 Σ (0) and R := c−1 G (0). P = q−1(R) and G(0) = R(0). Also, the continuity of cG ensures R is a clopen subgroupoid of G. Thus we obtain the twist (5.3) Since R is ´etale and effective, it follows from Renault's work in [32, Section 4] that C0(G(0)) is a r (P; R). (Renault makes the assumption that R(0) is second countable, but that Cartan MASA in C ∗ T × G(0) ֒→ P qP ։ R. 26 / / assumption is not required to show that (C ∗ r (P; R), C0(R(0))) is a Cartan pair. A close inspection of [32] shows that he uses R effective instead of R topologically principal, but these notions coincide when R(0) is second countable.) By Lemma 2.19, the inclusion Cc(P; R) ֒→ Cc(Σ; G) given by extension by zero extends to a ∗-monomorphism Proposition 5.4. The image of C ∗ C ∗ r (P; R)) = C ∗ r (Σ; G), that is, i(C ∗ r (Σ; G). r (P; R) ֒→ C ∗ i : C ∗ r (P; R) under i is the fixed point algebra of the action of Γ on r (Σ; G)Γ. Proof. First notice that i(C ∗ r (P; R)) ⊆ C ∗ r (Σ; G)Γ: indeed, if f ∈ Cc(P; R) and γ ∈ G, ω · i(f )( γ) = hω, cG( γ)ii(f )( γ) =(0 hω, 0ii(f )( γ) = i(f )( γ) if γ /∈ R if γ ∈ R = i(f )( γ). We now turn to showing C ∗ r (Σ; G)Γ ⊆ i(C ∗ r (P; R)). First suppose a ∈ C ∗ r (Σ; G)Γ ∩ Cc(Σ; G). Then for γ ∈ G, a( γ) =ZΓ (ω · a)( γ) dω (2.26) = ZΓ hω, cG( γ)i a( γ) dω = a( γ)ZΓ hω, cG( γ)i dω, which vanishes unless γ ∈ R. Thus a ∈ i(Cc(P, R)). For general a ∈ C ∗ Also note that Φ0(Cc(Σ; G)) ⊆ Cc(Σ; G), so Φ0(fi) ∈ i(Cc(P; R)). It follows that a ∈ i(C ∗ r (Γ; G)Γ, choose a net fi ∈ Cc(Σ; G) so that fi → a. Then Φ0(fi) → Φ0(a) = a. r (P; R)). (cid:3) Here is the main result of this section. Proposition 5.5. The pair (C ∗ r (Σ; G), C0(G(0))) is a Γ-Cartan pair. Proof. It is well known that C0(G(0)) is an abelian subalgebra of C ∗ proximate unit for C ∗ have already observed that C0(G(0)) is a Cartan MASA in C ∗ C0(G(0)) is a Cartan MASA in C ∗ (C ∗ r (Σ; G) that contains an ap- r (Σ; G). We r (P; R), so Lemma 5.4 shows that r (Σ; G). Thus (cid:3) r (Σ; G), C0(G(0))) is Γ-Cartan. We close this section with a result describing the supports of homogeneous normalizers. This is r (Σ; G) [31, Lemma 3.2]. Lemma 2.26 gives an action of Γ on C ∗ r (Σ; G)Γ. Since G is ´etale, span N is dense in C ∗ necessary for the proof of Lemma 6.2 below. Lemma 5.6. Let a ∈ C ∗ normalizer if and only if Sa is a bisection in c−1(t) for some t in Γ. Proof. An element a ∈ C ∗ r (Σ; G) is homogeneous of degree t if and only if r (Σ; G) and Sa be the open support of a. Then a is a homogeneous ω · ahω, tidω a =ZΓ ⇔ a(γ) =ZΓ ⇔ t = c(γ) for all γ ∈ Sa. hω, c(γ)i)a(γ)hω, tidω = a(γ)ZΓ hω, c(γ)ihω, tidω Thus a ∈ At if and only if Sa ⊆ c−1(t). By the argument in [32, Proposition 4.7], if a is a normalizer, then S−1 a is a homogeneous normalizer, Sa ⊂ c−1(t) for some t, whence S−1 effective and S−1 a Sa is open, we obtain S−1 a Sa ⊆ G(0). This implies that Sa is a bisection. a Sa ⊆ Iso(G). Thus, when a Sa ⊆ Iso(G) ∩ R. As R is (cid:3) 27 6. Analysis of C ∗ r (Σ; G) Throughout this section we fix a Γ-graded twist (Σ, G, Γ) satisfying Assumptions 5.1. Let (A, D) be the Γ-Cartan pair constructed from (Σ, G, Γ) in Proposition 5.5. An application of Theorem 4.36 to (A, D) yields another Γ-graded twist (Σ1, G1, Γ) also satisfying Assumptions 5.1. Our goal is to show that (Σ1, G1, Γ) and (Σ, G, Γ) are isomorphic in the sense that there are topological groupoid isomorphisms ΥΣ : Σ → Σ1 and ΥG : G → G1 such that the diagram, (6.1) T × G(0) ι Σ id×Υ G(0) ΥΣ cΣ1 T × G(0) 1 ι1 / Σ1 cΣ ❄❄❄❄❄❄❄❄ @ q Γ q1 G ΥG cG ⑦⑦⑦⑦⑦⑦⑦⑦ _❃❃❃❃❃❃❃❃ cG1 G1 commutes. Throughout, we will use the notation established in Section 4 for (A, D): thus X = G(0), D = C(X), ∆ = E ◦ Φ0, etc. Further, notice that for a ∈ Cc(Σ; G), ∆(a) is nothing more than aX . Thus, for every a ∈ A, ∆(a)(x) = εx(a). Lastly, recall from Section 4 that E = {ψ(n,x) : (n, x) ∈ G} is a family of linear functionals on A which becomes a topological groupoid when equipped with the weak-∗ topology, product ψ(m,αn(x))ψ(n,x) = ψ(mn,x), and inverse ψ−1 (n,x) = ψ(n∗,αn(x)). From Section 4 we have Σ1 = E and G1 = E. To begin, for γ ∈ Σ, consider the linear functional εγ on A determined by Proposition 2.21, described as follows. For a ∈ A, there is a unique scalar εγ(a) such that j(a)( γ) ∈ L is represented by (εγ(a), γ) ∈ T × Σ, that is, Alternatively, if a is viewed as a covariant function on Σ, j(a)( γ) = [εγ(a), γ]. Note that εγ is a norm-one linear functional on A. εγ(a) = a(γ). Lemma 6.2. The map ΥΣ : Σ → A# given by γ 7→ εγ is a homeomorphism of Σ onto E. Further- more, ΥΣ is an isomorphism of topological groupoids. Proof. Fix γ ∈ Σ, put x := s(γ) and choose n ∈ Nh such that εγ(n) > 0. By Lemma 5.6, n is supported on a homogeneous bisection, whence γ is the unique element of supp n whose εγ = ψ(n,x), it suffices to show εγ(m) = ψ(n,x)(m) for every m ∈ Nh. Choosing m ∈ Nh, we have source is x. Thus (n∗n)(x) = Pσ1σ2=x n(σ2)n(σ2) = n(γ)n(γ) > 0, so (n, x) ∈ G. To show ∆(n∗m)(x) = (n∗m)(x) =P σ∈Gx n(σ)m(σ). As the terms in this sum are zero unless σ ∈ supp n, and supp n ∩ Gx = { γ}, we have ψ(n,x)(m) = ∆(n∗m)(x) n(x) = n(γ)m(γ) n(γ) = εγ(m) because n(γ) > 0. Thus εγ = ψ(n,x), as desired. Now suppose (n, x) ∈ G. Since n is supported on an open bisection by Lemma 5.6, there is a unique element of supp n whose source is x. Therefore, there is a unique element γ ∈ Σ satisfying s(γ) = x and εγ(n) > 0. The argument of the previous paragraph shows ψ(n,x) = εγ. We have thus shown that ΥΣ(Σ) = E. Notice that our work also shows that ΥΣ is bijective. 28 / / / / / / / @ _ Recall that G is ´etale, Cc(Σ; G) is dense in A, and elements of E are norm one linear functionals on A. So if (γi) is a net in Σ and γ ∈ Σ, γi → γ ⇔ for every a ∈ Cc(Σ; G), εγi(a) → εγ(a) ⇔ for every a ∈ A, εγi(a) → εγ(a) ⇔ (εγi) converges weak-∗ to εγ. Thus ΥΣ is a homeomorphism. We now observe that ΥΣ preserves the groupoid operations. First, suppose γ ∈ Γ and n ∈ Nh is such that εγ(n) > 0. Then εγ = ψ(n,s(γ)). For d ∈ D, we have εγ(dn) = d(r(γ))n(γ). On the other hand, ψ(n,s(γ))(dn) = d(αn(s(γ))p(n∗n)(s(γ)) = d(αn(s(γ)))n(γ). But as this holds for every d ∈ D and n(γ) > 0, we conclude that (6.3) r(γ) = αn(s(γ)). Suppose the product of γ1, γ2 ∈ Σ is defined. For i = 1, 2, choose ni ∈ Nh so that εγi (ni) > 0. As ni are supported in open bisections of G (Lemma 5.6), εγ1γ2(n1n2) = (n1n2)(γ1γ2) = n1(γ1)n2(γ2) = εγ1(n1)εγ2 (n2) > 0. We therefore obtain ΥΣ(γ1γ2) = ψ(n1n2,s(γ2)) (6.3) = ψ(n1,r(γ2))ψ(n2,s(γ2)) = ΥΣ(γ1)ΥΣ(γ2). For γ ∈ Σ and n ∈ Nh such that εγ(n) > 0, we have εγ−1(n∗) = n∗(γ−1) = n(γ) > 0, so εγ−1 = ψ(n∗,r(γ)). As r(ψ(n,s(γ))) = αn(s(γ)) = r(γ) = αn(s(γ)), we obtain ΥΣ(γ−1) = (ΥΣ(γ))−1. Finally, suppose z ∈ T and γ ∈ Σ. Choose n ∈ N so that εγ(n) > 0. Then εz·γ(zn) = (zn)(z·γ) = z(zn)(γ) = n(γ) > 0. Thus, εz·γ = ψ(zn,s(γ)). So by Remark 4.27, ΥΣ(z · γ) = z · ΥΣ(γ). (cid:3) Writing Σ1 := E and G1 := E, we thus have defined the two left vertical arrows in (6.1). It follows that if γ ∈ G, then for σ1, σ2 ∈ q−1( γ), q2(ΥΣ(σ1)) = q2(ΥΣ(σ2)). Therefore the map ΥG : G → G1 given by ΥG( γ) := q1(ΥΣ(γ)) is a well-defined isomorphism of groupoids. That ΥG is a homeomorphism follows from the fact that Σ → G and Σ1 → G1 are locally trivial and ΥΣ is a homeomorphism (or use the fact that q, q1 are quotient maps and ΥΣ is a homeomorphism). Now suppose γ ∈ G and cG( γ) = t. Then for γ ∈ q−1( γ), cΣ(γ) = t. Choosing n ∈ Nh with εγ(n) > 0, we obtain supp n ⊆ c−1(t). So by (4.30) we obtain cΣ1(ψ(n,s(γ))) = cG1(ψ(n,s(γ))) = t. It follows that (6.1) commutes. Thus we have proved the following theorem. Theorem 6.4. Let Σ → G be a Γ-graded twist satisfying Assumptions 5.1. Let Σ1 → G1 be the twist constructed from (C ∗ r (Σ; G), C0(G(0))). For each γ ∈ Σ, choose a homogeneous normalizer n ∈ C ∗ r (Σ, G) such that n(γ) > 0. Then the map ΥΣ : Σ → Σ1 given by γ 7→ [n, s(γ)]Σ1 descends to a well-defined isomorphism of twists such that the diagram (6.1) commutes. Corollary 6.5. Suppose Σ → G and Σ′ → G′ are Γ-graded twists satisfying Assumptions 5.1. Suppose further that Ξ : C ∗ is an isomorphism of C ∗-algebras such that r (Σ; G) → C ∗ r (Σ′; G′) 29 (1) Ξ is equivariant for the induced Γ actions; and (2) ΞC0(G(0)) : C0(G(0)) → C0(G′(0)) is an isomorphism. Then there exists groupoid isomorphisms υΣ, υG such that the following diagram commutes. T × G(0) ι Σ id×vG G(0) υΣ T × G′(0) ι′ / Σ′ cΣ ❄❄❄❄❄❄❄❄ ?⑧⑧⑧⑧⑧⑧⑧⑧ c′ Σ q Γ q′ G υG cG ⑧⑧⑧⑧⑧⑧⑧⑧ _❄❄❄❄❄❄❄❄ c′ G / G′ Proof. By Theorem 6.4, there are isomorphisms ΥΣ : Σ → Σ1, ΥG : G → G1, ΥΣ′ : Σ′ → Σ′ 1, r (Σ; G)Γ isomorphically onto ΥG′ : G′ → G′ r (Σ′; G′)Γ. Thus by construction, Σ1 ∼= Σ′ C ∗ 1. The result then follows from composition of isomorphisms. (cid:3) 1. Since Ξ is an equivariant isomorphism it takes C ∗ 1, G1 ∼= G′ 7. Examples Example 7.1. Let G be a finite discrete abelian group. Take A = C ∗(G). Then C ∗(G) = span{δg : g ∈ G}. As pointed out in [10], if G = H, then C ∗(G) ∼= C ∗(H). So it is surprising that we would be able to recover G using Theorem 6.4. However, as this example illustrates, the induced action of Γ required in Theorem 6.4 plays a crucial role. Suppose c : G → Γ is a homomorphism of G into a discrete abelian group, with c−1(0) topologi- cally principal. Then c−1(0) = {0}, so c is injective and c(G) is isomorphic to G as a subgroup of Γ. So A0 = Cδ0 and we consider the inclusion D = A0 ⊆ A. Notice that for ω ∈ G, ω ·δg(h) = hω, c(h)i δg(h) so that ω ·δg = hω, c(g)iδg . Thus, by Lemma 3.5, δg ∈ Ac(g) and furthermore δg /∈ At for t 6= c(g). Since C ∗(G) = span{δg : g ∈ G} we have At =(Cδg 0 t = c(g) otherwise. Thus the homogeneous normalizers of A0 are all of the form λδg for some λ ∈ C. Take X = {∗} = A0. Now [λδg, ∗]Σ = [λ′δg′, ∗]Σ ⇔ g = g′ and λλ′ > 0, [λδg, ∗]G = [λ′δg′, ∗]G ⇔ g = g′. So here ΥG : g 7→ [δg, ∗]G and ΥΣ(z, g) 7→ [zδg, ∗]Σ where ΥΣ : T × G → ΣW are the desired isomorphisms from Theorem 6.4. Example 7.2. Let Λ be a k-graph. That is, Λ is a small category endowed with a functor, the degree map, d : Λ → Nk, that satisfies the following unique factorization property: if λ ∈ Λ and d(λ) = m + n there exists unique µ, ν ∈ Λ such that d(µ) = m, d(ν) = n and λ = µν. We assume Λ has no sources in that for all objects v and all m ∈ Nk there exists µ with r(µ) = v and d(µ) = m. In [22], Kumjian, Pask, and Sims introduce categorical cohomology on a k-graph Λ. In particular, they define a 2-cocycle with coefficients in T to be a function φ : Λ ∗ Λ → T such that φ(λ1, λ2) + φ(λ1λ2, λ3) = φ(λ2, λ3) + φ(λ1, λ2λ3) 30 / / / / / ? / _ where Λ ∗ Λ := {(µ, ν) : s(µ) = r(ν)} and λi are defined so that all of the compositions above make sense. Denote the set of these 2-cocycles by Z2(Λ, T). They prove in [22, Theorem 4.15], that there is an isomorphism from the second cubical cohomology group they defined in [21] to this categorical cohomology group. They define in [22, Definition 5.2] the twisted k-graph C ∗-algebra by φ ∈ Z2(Λ, T) to be the universal C ∗-algebra C ∗(Λ, φ) generated by elements tµ, µ ∈ Λ of a C ∗-algebra satisfying the following. (1) The tv for v ∈ d−1(0) are mutually orthogonal projections, (2) tµtν = φ(µ, ν)tµν whenever s(µ) = r(ν), (3) t∗ λtλ = ts(λ), and (4) for all v ∈ d−1(0) and n ∈ Nk, tv = Xr(λ)=v d(λ)=n tλt∗ λ. By the universal property of C ∗(Λ, φ), d : Λ → Nk induces an action of Tk on C ∗(Λ, φ) charac- terized by z · tµt∗ ν = zd(µ)−d(ν)tµt∗ ν. Let and let C = span{tµt∗ ν : d(µ) = d(ν)}, D = span{tµt∗ µ}. By [22, Lemma 7.4] C is the fixed point algebra C ∗(Λ, φ)Tk , for this action. Moreover, as elements of the generating set {tµt∗ ν : d(µ) = d(ν)} are all normalizers for D, to show D is Cartan in C it suffices to show D is maximal abelian and that there is a conditional expectation from C onto D. The conditional expectation P from C onto D is given by, for µ, ν ∈ Λ with d(µ) = d(ν) The C ∗-algebra C is an AF-algebra. This is shown in [22, Proposition 7.6]. We recap and reframe some of those details to show D is Cartan in C. For each n ∈ Nk let P (tµt∗ ν) = δµ,ν tµt∗ ν. and let Cn = span{tµt∗ ν : µ, ν ∈ Λn}, Dn = span{tµt∗ µ : µ ∈ Λn}. When m ≤ n we embed Cm in Cn using condition (4) in the definition of C ∗(Λ, φ) above: µ, ν ∈ Λm then if tµt∗ tµtλt∗ λt∗ ν ∈ Cn. For each v ∈ Λ0 and n ∈ Nk denote by K(Λnv) the compact operators on the Hilbert space ℓ2(Λnv). Using a matrix unit argument, it is observed in [22, Proposition 7.6(1)] that Note that this embedding also gives Dm ⊆ Dn. We have then that and ν = Xλ∈Λn−ms(µ) C = [n∈Nk D = [n∈Nk Cn, Dn. Cn = Mv∈Λ0 31 K(Λnv). Pn(a) = Xµ∈Λn (tµt∗ µ)a(tµt∗ µ), As Dn is formed by the self-adjoint matrix units, Dn is a maximal abelian subalgebra of Cn. Further there is a faithful conditional expectation Pn on Cn. We can describe this conditional expectation by where the series convergences in the strong operator topology. Using this formula we can extend Pn to all of B(H). A simple calculation shows that Pn(B(H)) = D′ n, i.e., Pn is a conditional expectation onto D′ n. We further note that the embeddings of Cm into Cn for m ≤ n, give PnCm = Pm. The conditional [28, expectation P : C → D can then be described as the direct limit of the maps {Pn} (see e.g. Proposition A.8]). To show that D is maximal abelian in C we use an argument similar to that found in [35, Chapter 1]. Suppose a ∈ D′ ∩ C. Since a ∈ C there is a net (an) with an ∈ Cn such that lim n Further, since a ∈ D′, we have that a ∈ D′ n for each n ∈ Nk, and thus Pn(a) = a for each n ∈ Nk. Hence kan − ak = 0. kPn(an) − ak = kPn(an − a)k ≤ kan − ak. And thus the net (Pn(an))n converges to a. Since Pn(an) ∈ Dn it follows that a ∈ D, and therefore D is maximal abelian in C. such that C ∗(Λ, φ) ∼= C ∗ [sµs∗ Thus (C ∗(Λ, φ), D) is a Zk-Cartan pair. Hence by Theorem 4.36 there exists a twist ΣW → GW r (ΣW ; GW ). Notice that here ΣW and GW consist of elements of the form ν, x] with µ, ν ∈ Λ and x ∈ Λ∞. In [22], the authors construct a groupoid GΛ and a continuous cocycle ς such that C ∗(Λ, φ) ∼= r (GΛ, ς) [22, Theorem 6.7]. By Theorem 6.4, ΣW ∼= T ×ς GΛ and GW ∼= GΛ, that is Theorem 4.36 C ∗ recovers the construction in [22]. We provide some details of the isomorphisms of twists given above, but to proceed we need to provide a few details of the construction in [22]. The groupoid construction in [22] is standard and goes back to [19, 30, 20]. We say Λ∞ := {x : N × N → Λ, x is a degree preserving functor} and σp : Λ∞ → Λ∞ by σp(x)(m, n) = x(m + p, n + p). GΛ := {(x, ℓ − m, y) ∈ Λ∞ × Zk × Λ∞ : ℓ, m ∈ Nk, σℓx = σmy} with the topology on GΛ given by basic open sets Z(µ, ν) := {(µx, d(µ) − d(ν), νx) : x ∈ Λ∞, r(x) = s(µ) = s(ν)}. It turns out that under our hypotheses each Z(µ, ν) is compact and open and so there exists a subset P ⊆ Λ × Λ such that {Z(µ, ν) : (µ, ν) ∈ P} is a partition of GΛ. Thus for each γ ∈ GΛ there exists (µγ, νγ) ∈ P such that γ ∈ Z(µγ, νγ). Now by [22, Lemma 6.3] for γ, η ∈ GΛ composable we can find y ∈ Λ∞, α, β, ζ ∈ Λ such that γ = (µγαy,d(µγ) − d(νγ), νγ αy), η = (µηβy, d(µη) − d(νη), νηβy), γη = (µγηζy, d(µγη) − d(νγη), νγηζy) and ςφ(γ, η) := (φ(µγ, α) − φ(νγ, α) + (φ(µη, β) − φ(νη, β)) − (φ(µγη, ζ) − φ(νγη, ζ)) is a well-defined continuous groupoid cocycle (see [30] for the definition). We can then define and then the twist is ΣΛ,φ = T ×ς GΛ. T × G(0) Λ → ΣΛ,φ → GΛ. 32 Now the isomorphisms given in Theorem 6.4 are given by ΥΣ :(z, (µγ x, d(µγ) − d(νγ), νγx)) 7→ [sµγ s∗ ΥG :(µγx, d(µγ) − d(νγ), νγx) 7→ [sµγ s∗ νγ , νγx]G. νγ , νγx]Σ Example 7.3. A main result in [4] is that any separable, unital, nuclear C ∗-algebra which contains a Cartan subalgebra satisfies the UCT. In fact, more is shown. It is shown in [4, Theorem 3.1] that if Σ → G is a twist, where G is an ´etale Hausdorff locally compact second countable groupoid where the reduced C ∗-algebra C ∗ r (Σ; G) satisfies the UCT. Hence we have the following corollary to Theorem 4.36 and [4, Theorem 3.1]. r (Σ; G) is nuclear, then C ∗ Corollary 7.4. Let A be a separable and nuclear C ∗-algebra. If A contains an abelian subalgebra D such that (A, D) is a Γ-Cartan pair for a some discrete abelian group Γ, then A satisfies the UCT. Appendix A. Nonabelian groups In the previous sections we assumed we had gradings by discrete abelian groups and actions by the dual group, as this case is familiar and doesn't involve the introduction of coactions. However all of the results of the paper can be extended to gradings by nonabelian groups, by replacing actions with coactions: in this short appendix we outline how to extend our results to the nonabelian case for those readers already familiar with coactions. For those readers interested in more information on coactions we recommend [13, Appendix A]. Throughout this appendix, all tensor products of C ∗-algebras are spatial tensor products. (A.1) Uses of Commutativity. The alert reader will no doubt have noticed that we used the commutativity of Γ in a few key places: (1) to define a Γ grading on a C ∗-algebra A when an action of Γ on A is given; (2) to define maps Φt : A → At, including the faithful conditional expectation Φ0 onto A0 the fixed point algebra (Lemma 3.5); (3) to define an action of Γ on C ∗ (4) to show the fixed point algebra of the action above contains C0(G(0)) as a Cartan subalgebra r (Σ; G) where c : G; Σ 7→ Γ (Lemma 2.26); and (Proposition 5.4). Now suppose that Γ is a not necessarily abelian discrete group whose identity we denote by e. For s, t ∈ Γ, we will use δs to denote the indicator function of the set {s} and δs,t for the Kronecker δ (so δs,t = 1 if s = t and 0 if s 6= t). Let Λ : C ∗ r (Γ) → B(ℓ2(Γ)) be the left regular representation of Γ. Also, the map δs 7→ δs ⊗ δs ∈ C ∗ r (Γ) ⊗ C ∗ r (Γ) extends to a ∗-homomorphism νΓ : C ∗ r (Γ) → C ∗ r (Γ) ⊗ C ∗ r (Γ). Let ν : A → M (A ⊗ C ∗ r (Γ)) be a (reduced) coaction. This means that ν is a non-degenerate ∗-homomorphism of A into M (A ⊗ C ∗ r (Γ)) such that (i) ν(A)(I ⊗ C ∗ (ii) (ν ⊗ idΓ) ◦ ν = (idA ⊗ νΓ) ◦ ν (these maps belong to B(A, M (A ⊗ C ∗ r (Γ)) ⊆ A ⊗ C ∗ r (Γ); and r (Γ) ⊗ C ∗ r (Γ))). Notice that since Γ is discrete, I ⊗ δe is the identity of M (A ⊗ C ∗ that actually, r (Γ)). Thus condition (i) implies ν : A → A ⊗ C ∗ r (Γ). Furthermore, the fact that Γ is discrete implies that ν is non-degenerate in the sense that ν(A)(I ⊗ C ∗ r (Γ), see [3] or [13, Remark A.22(3)]. r (Γ)) is dense in A ⊗ C ∗ For t ∈ Γ, there is a slice map St : A ⊗ C ∗ r (Γ) → A characterized by a ⊗ b 7→ ahΛ(b)δe, δti, 33 (see [39] or [13, §A.4]). Further, it follows from the second statement of [13, Lemma A.30] that for x ∈ A ⊗ C ∗ r (Γ), x = 0 ⇔ for all t ∈ Γ, St(x) = 0. Thus, every element x ∈ A ⊗ C ∗ r (Γ) has a uniquely determined "Fourier series", Define continuous maps Φt : A → A by Then for every a ∈ A, (A.2) x ∼Xt∈Γ St(x) ⊗ δt. Φt := St ◦ ν. ν(a) ∼Xt∈Γ Φt(a) ⊗ δt. While we will not need this fact here, the ∗-homomorphism property of ν and the series representa- tion (A.2) implies that the "coefficient maps" {Φt}t∈Γ behave much as Fourier coefficients do under convolution multiplication and adjoints: for every t ∈ Γ and a, b ∈ A, Φt(ab) =Xs∈Γ Φts−1(a)Φs(b) and Φt(a∗) = Φt−1(a)∗. What we do require is that condition (ii) in the definition of coaction given above implies that for every s, t ∈ Γ and a ∈ A, Φs(Φt(a)) =(0 if s 6= t, Φt(a) when s = t. Define At := Φt(A). Note that as Se arises from the (faithful) trace on C ∗ expectation. r (Γ), Φe : A → Ae is a faithful conditional We will call an element a ∈ A homogeneous if a ∈ At for some t ∈ Γ. This gives us the Γ-grading and an analog of Lemma 3.5, which addresses the first two points of Paragraph A.1. We now address the third and fourth items of Paragraph A.1. Assume we have a twist with a cocycle as in Section 5 but with Γ not necessarily abelian: (A.3) T × G(0) / Σ ❄❄❄❄❄❄❄❄ cΣ / G cG Γ. The proof of [10, Lemma 6.1] goes through without change to show there exists a coaction ν : C ∗ r (Σ; G) → C ∗ r (Σ; G) ⊗ C ∗ r (Γ) characterized by ν(f ) = f ⊗ δt where f ∈ Cc(G) and supp(f ) ⊆ c−1 G (t). Note that for f ∈ Cc(G) and supp(f ) ⊆ c−1 G (t), Φs(f ) = Ss(f ⊗ δt) = f hΛ(δt)δe, δsi = f hδt, δsi = f δs,t. Now for f ∈ Cc(G) then f =Pt∈Γ f c−1 G (t) so that the above computation show that Φs(f ) = f c−1 G (s). 34 / / By continuity of Φs and the map j : C ∗ G (e) and P = c−1 in Section 5, let R = c−1 if a ∈ C ∗ Φe(fi) = fiR ∈ Cc(R; P), so that a ∈ C ∗(R; P), giving Proposition 5.4. Theorem A.4. Let A be a C ∗-algebra and let D be an abelian C ∗-algebra of A such that: r (Σ; G) → C0(Σ; G) we get that j(Φs(a)) = j(a)c−1 G (s). As Σ (e). Then Lemma 2.19 goes through without change and r (Σ; G) with fi → a, by the above we have Φe(fi) → a and r (Σ; G)e and a net fi ∈ C ∗ • there is a coaction ν of a discrete group Γ on A; • D is Cartan in the algebra Ae; and • span Nh(A, D) = A. Let Ac be the algebraic span of Nh. Then there exists a Γ-graded twist T × X ֒→ Σ → G and a ∗-isomorphism Ψ : Ac → C ∗ r (Σ; G) which induces an isomorphism A → C ∗(Σ; G) taking D to C0(G(0)). Proof. The arguments of Section 4 go through without change. (cid:3) Remark A.5. Compare the conditions of Theorem A.4 to the definition of Γ-Cartan (3.10), noting that by Lemma 3.11 (1) the conditions on the normalizers coincide when Γ is abelian. While we have not checked details, we expect that if Γ is a (not necessarily abelian) discrete group, and the hypotheses of Theorem A.4 are weakened so that the condition span Nh(A, D) = A is replaced with span N (A, D) = A, then it is still true that Nh(A, D) = A. If this is the case, a Γ-Cartan pair could then be defined to be an inclusion of C ∗-algebras D ⊆ A satisfying the weakened hypotheses of Theorem A.4. All the results of this paper would be valid for this notion of Γ-Cartan pairs. Likewise, the arguments of Sections 5 and 6 yield the following result for (possibly non-commutative) discrete groups Γ. Theorem A.6. Let Σ → G be a locally trivial twist, with a cocycle c into a discrete group Γ. Then (C ∗ r (Σ; G), C0(G(0))) is a Γ-Cartan pair. Let Σ1 → G1 be the Γ-graded twist constructed from (C ∗ r (Σ, G) such that n(γ) > 0. Then the map r (Σ; G), C0(G(0))). For each γ ∈ Σ, choose a homogeneous normalizer n ∈ C ∗ descends to a well-defined isomorphism of twists such that the following diagram commutes. ΥΣ : Σ → Σ1 given by γ 7→ [n, s(γ)]Σ1 T × X ΥΣ ι ;✇✇✇✇✇✇✇✇✇ #●●●●●●●●● ι1 Σ Σ1 cΣ ❅❅❅❅❅❅❅❅ ?⑦⑦⑦⑦⑦⑦⑦⑦ cΣ1 G ΥG q Γ cG ~⑥⑥⑥⑥⑥⑥⑥⑥ `❅❅❅❅❅❅❅❅ cG1 q1 / G1 References 1. G. Abrams, P. Ara, and M. Siles Molina, Leavitt Path Algebras, Lecture Notes in Mathematics, vol. 2192, Springer, London, 2017. MR 3729290 2. P. Ara, J. Bosa, R. Hazrat, and A. Sims, Reconstruction of graded groupoids from graded Steinberg algebras, preprint 2016. (arXiv:1601.02872v1 [math.RA]). 3. S. Baaj and G. Skandalis, C*-alg´ebres de Hopf et th´eorie de Kasparov ´equivariante, K-Theory 2 (1989), 683 -- 721 4. S. Barlak and X. Li, Cartan subalgebras and the UCT problem. Adv. Math. 316 (2017), 748 -- 769. 5. J.Brown, L. Clark, C. Farthing, and A. Sims, Simplicity of algebras associated to ´etale groupoids, Semigroup Forum 88 (2014), 433 -- 452. 35 / / ~ ; # ? / ` 6. J. Brown, L. Clark, and A. an Huef, Diagonal-preserving ring ∗-isomorphisms of Leavitt path algebras, preprint 2015. (arXiv:1510.05309v3 [math.RA]). 7. N. Brownlowe, T. Carlsen, and M. Whittaker, Graph algebras and orbit equivalence, Ergodic Theory Dynam. Systems 37 (2017), 389 -- 417. 8. T.Carlsen, ∗-isomorphism of Leavitt path algebras over Z, preprint 2016. (arXiv:1601.00777v2 [math.RA]). 9. T. Carlsen, E. Ruiz, and A. Sims, Equivalence and stable isomorphism of groupoids, and diagonal-preserving stable isomorphisms of graph C ∗-algebras and Leavitt path algebras, Proc. Amer. Math. Soc. 145 (2017), 1581 -- 1592. 10. T. Carlsen, E. Ruiz, A. Sims, and M. Tomforde, Reconstruction of groupoids and C ∗-rigidity of dynamical systems, preprint 2017, (arXiv:1711.01052v1 [math.OA]). 11. L.O. Clark, C. Farthing, A. Sims and M. Tomforde, A groupoid generalisation of Leavitt path algebras, Semigroup Forum, 89 (2014), 501 -- 517. 12. A. Donsig and D. Pitts, Coordinates Systems and Bounded Isomorphisms, J. Operator Theory. 59(2) (2008), 359 -- 416. 13. S. Echterhoff, S. Kaliszewski, J. Quigg, and I. Raeburn, A Categorical Approach to Imprimitivity Theorems for C ∗-Dynamical Systems, preprint 2005, (arXiv:0205322v2). 14. R. Exel, Amenability for Fell Bundles, J. Reine Angew. Math., 492 (1997), 41 -- 73. MR 1488064. 15. Partial dynamical systems, Fell bundles and applications, Mathematical Surveys and Mono-graphs, vol. 224, American Mathematical Society, Providence, RI, 2017. MR 3699795 16. R. Exel, Inverse semigroups and combinatorial C ∗-algebras, Bull. Braz. Math. Soc. (N.S.) 39 (2008), no. 2, 191 -- 313. MR 2419901 (2009b:46115) 17. J. Feldman and C. C. Moore, Ergodic equivalence relations, cohomology and von Neumann algebras I, II, Trans. Amer. Math. Soc. 234 (1977), 289 -- 359. 18. A. Kumjian, On C ∗-diagonals, Can. J. Math., 38 (1986), 969 -- 1008. 19. A. Kumjian and D. Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20. 20. A. Kumjian, D. Pask, I. Raeburn, and J. Renault, Graphs, Groupoids, and Cuntz-Krieger Algebras, J. Funct. Anal. 144 (1997), 505 -- 541. 21. A. Kumjian, D. Pask, and A. Sims, Homology for higher-rank graphs and twisted C ∗-algebras, J. Funct. Anal. 263 (2012), 1539 -- 1574. 22. 23. K. Matsumoto and H. Matui, Continuous orbit equivalence of topological Markov shifts and Cuntz-Krieger alge- , On twisted higher-rank graph C ∗-algebras, Trans. Amer. Math. Soc. 367 (2015), 5177 -- 5216. bras, Kyoto J. Math. 54 (2014) 863 -- 877. 24. K. Matsumoto, Classification of Cuntz-Krieger algebras by orbit equivalence of topological Markov shifts, Proc. Amer. Math. Soc. 141 (2013), 2329 -- 2342. 25. H. Matui, Homology and topological full groups of ´etale groupoids on totally disconnected spaces, Proc. Lond. Math. Soc. (3) 104 (2012), 27 -- 56. , Structure for regular inclusions. I, J. Operator Theory. 78(2) (2017), 357 -- 416. 26. D. Pitts, Structure for regular inclusions, preprint 2012, (arXiv:1202.6413v2 [math.OA]). 27. 28. I. Raeburn, Graph algebras. CBMS Regional Conference Series in Mathematics, 103. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2005. 29. 30. J. Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, vol. 793, Springer-Verlag, New , On graded C ∗-algebras, Bull. Aust. Math. Soc. 97 (2018), no. 1, 127 -- 132. MR 3744874 York, 1980. , Repr´esemtation des produits crois´es d'alg`ebres de groupoıdes, J. Operator Theory 18 (1987), 67 -- 97. , Cartan subalgebras in C ∗-algebras, Irish Math. Soc. Bulletin 61 (2008), 29 -- 63. 31. 32. 33. A. Sims, Hausdorff ´etale groupoids and their C ∗-algebras, arXiv:1710.10897v2, October 2018. 34. J. Spielberg. Groupoids and C ∗-algebras for categories of paths, Trans. Amer. Math. Soc. 366, 5771 -- 5819. 35. S¸. Stratila and D. Voiculescu, Representations of AF-algebras and of the group U (∞). Lecture Notes in Mathe- matics, Vol. 486. Springer-Verlag, Berlin-New York, 1975. 36. B. Steinberg, A groupoid approach to inverse semigroup algebras, Adv. Math. 223 (2010), 689 -- 727. 37. A. Tikuisis, S. White, W. Winter, Quasidiagonality of nuclear C∗-algebras. Ann. of Math. (2) 185 (2017), no. 1, 229 -- 284. 38. M. Tomforde, Uniqueness theorems and ideal structure for Leavitt path algebras J. Algebra 318 (2007), 270 -- 299. 39. J. Tomiyama, Applications of Fubini type theorem to the tensor products of C ∗-algebras, Tohoku Math. Journ. 19 (1967), 213 -- 226. 40. S. Willard, General topology, Addison-Wesley Publishing Co., Reading, Mass.-London-Don Mills, Ont., 1970. MR 0264581 (41 #9173) 36 (J.H. Brown) Department of Mathematics, University of Dayton, 300 College Park Dayton, OH 45469-2316 U.S.A. E-mail address: [email protected] (A.H. Fuller) Department of Mathematics, Ohio University, Athens, OH 45701 U.S.A. E-mail address: [email protected] (D.R. Pitts) Department of Mathematics, University of Nebraska-Lincoln, Lincoln, NE 68588-0130 U.S.A. E-mail address: [email protected] (S. A. Reznikoff) Department of Mathematics, Kansas State University, 138 Cardwell Hall, Manhat- tan, KS, U.S.A. E-mail address: [email protected] 37
1610.02600	1	1610	2016-10-09T00:20:43	Relative Morita equivalence of Cuntz--Krieger algebras and flow equivalence of topological Markov shifts	[ "math.OA", "math.DS" ]	In this paper, we will introduce notions of relative version of imprimitivity bimodules and relative version of strong Morita equivalence for pairs of $C^$-algebras $(\mathcal{A}, \mathcal{D})$ such that $\mathcal{D}$ is a $C^$-subalgebra of $\mathcal{A}$ with certain conditions. We will then prove that two pairs $(\mathcal{A}_1, \mathcal{D}_1)$ and $(\mathcal{A}_2, \mathcal{D}_2)$ are relatively Morita equivalent if and only if their relative stabilizations are isomorphic. In particularly, for two pairs $(\mathcal{O}_A, \mathcal{D}_A)$ and $(\mathcal{O}_B, \mathcal{D}_B)$ of Cuntz--Krieger algebras with their canonical masas, they are relatively Morita equivalent if and only if their underlying two-sided topological Markov shifts $(\bar{X}_A,\bar{\sigma}_A)$ and $(\bar{X}_B,\bar{\sigma}_B)$ are flow equivalent. We also introduce a relative version of the Picard group ${\operatorname{Pic}}(\mathcal{A}, \mathcal{D})$ for the pair $(\mathcal{A}, \mathcal{D})$ of $C^*$-algebras and study them for the Cuntz--Krieger pair $(\mathcal{O}_A, \mathcal{D}_A)$.	math.OA	math	Relative Morita equivalence of Cuntz -- Krieger algebras and flow equivalence of topological Markov shifts Kengo Matsumoto Department of Mathematics Joetsu University of Education Joetsu, 943-8512, Japan May 13, 2018 Abstract In this paper, we will introduce notions of relative version of imprimitivity bimod- ules and relative version of strong Morita equivalence for pairs of C ∗-algebras (A,D) such that D is a C ∗-subalgebra of A with certain conditions. We will then prove that two pairs (A1,D1) and (A2,D2) are relatively Morita equivalent if and only if their relative stabilizations are isomorphic. In particularly, for two pairs (OA,DA) and (OB,DB) of Cuntz -- Krieger algebras with their canonical masas, they are relatively Morita equivalent if and only if their underlying two-sided topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are flow equivalent. We also introduce a relative version of the Picard group Pic(A,D) for the pair (A,D) of C ∗-algebras and study them for the Cuntz -- Krieger pair (OA,DA). Contents 1. Introduction 2. Relative σ-unital C ∗-algebras. 3. Relative imprimitivity bimodules and relative Morita equivalence. 4. Isomorphism of relative stabilizations. 5. Relative full corners. 6. Relative Morita equivalence in Cuntz -- Krieger pairs. 7. Corner isomorphisms in Cuntz -- Krieger pairs. 8. Relative Picard groups. 9. Relative Picard groups of Cuntz -- Krieger pairs. 10. Appendix: Picard groups of Cuntz -- Krieger algebras. 1 Introduction In [28], M. Rieffel introduced the notion of imprimitivity bimodule for C ∗-algebras as a Hilbert C ∗-bimodule satisfying certain conditions from a viewpoint of representation theory of groups, so that he defined the notion of strong Morita equivalence in C ∗-algebras. Let A and B be C ∗-algebras. An A -- B-bimodule X means a Hilbert C ∗-bimodule with 1 i and with a right B- a left A-module structure and an A-valued inner product Ah iB satisfying some comparability module structure and a B-valued inner product h It is said to be full if the ideals spanned by conditions (see [24], [28], [10], [25], etc.). {Ahx yi x, y ∈ X} and {hx yiB x, y ∈ X} are dense in A and in B, respectively. If a full A -- B-bimodule X further satisfies the condition Ahx yiz = xhy ziB for x, y, z ∈ X, it is called an A -- B-imprimitivity bimodule. Two C ∗-algebras A and B are said to be strong Morita equivalent if there exists an A -- B-imprimitivity bimodule, which means that A and B have same representation theory. Brwon -- Green -- Rieffel in [3] have shown that two σ- unital C ∗-algebras A and B are strong Morita equivalent if and only if they are stably isomorphic, that is A ⊗ K is isomorphic to B ⊗ K, where K denotes the C ∗-algebra of compact operators on a separable infinite dimensional Hilbert space. In this paper, we will study Morita equivalence of C ∗-algebras from a view point of symbolic dynamical systems. For an irreducible non-permutation matrix A = [A(i, j)]N with entries in {0, 1}, two-sided topological Markov shift ( ¯XA, ¯σA) are defined as a topo- logical dynamical system on the shift space ¯XA consisting of two-sided sequences (xn)n∈Z of xn ∈ {1, . . . , N} such that A(xn, xn+1) = 1 for all n ∈ Z with the shift homeomorphism ¯σA((xn)n∈Z) = (xn+1)n∈Z on the compact Hausdorff space ¯XA. J. Cuntz and W. Krieger introduced a C ∗-algebra OA associated to the matrix A ([7]). The C ∗-algebra is called the Cuntz -- Krieger algebra, which is a universal unique C ∗-algebra generated by partial isometries S1, . . . , SN subject to the relations: i,j=1 SjS∗ j = 1, S∗ i Si = N Xj=1 N Xj=1 A(i, j)Sj S∗ j , i = 1, . . . , N. (1.1) in ··· S∗ Since the stable isomorphism class of OA does not have complete informations about the underlying dynamical system ( ¯XA, ¯σA), we need some extra structure to OA to study ( ¯XA, ¯σA). In this paper, we consider the pair (OA,DA) where DA is the C ∗-subalgebra of OA generated by the projections of the form: Si1 ··· SinS∗ i1, i1, . . . , in = 1, . . . , N . We call the pair (OA,DA) the Cuntz -- Krieger pair. As in [17], the isomorphism class of the pair (OA,DA) is a complete invariant of the continuous orbit equivalence class of the underlying one-sided topological Markov shift (XA, σA). As one of the remarkable rela- tionships between symbolic dynamics and Cuntz -- Krieger algebras, Cuntz -- Krieger showed in [7] that if topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are flow equivalent, then there exists an isomorphism Φ : OA ⊗ K −→ OB ⊗ K such that Φ(DA ⊗ C) = DB ⊗ C, where C denotes the maximal commutative C ∗-subalgebra of K consisting of the diagonal elements. Recently H. Matui and the author have proved that the converse implication also holds, so that ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are flow equivalent if and only if there exists an isomorphism Φ : OA ⊗ K −→ OB ⊗ K such that Φ(DA ⊗ C) = DB ⊗ C ([21]). We call the pair (OA ⊗ K,DA ⊗ C) the stabilized Cuntz -- Krieger pair or the relative stabilization of (OA,DA), so that the isomorphism class of the relative stabilization of (OA,DA) is a complete invariant for the flow equivalence class of the underlying two-sided topological Markov shift ( ¯XA, ¯σA). In this paper, we will introduce notions of relative version of imprimitivity bimodules and of relative version of strong Morita equivalence for pairs of C ∗-algebras (A,D) such 2 that D is a C ∗-subalgebra of A for which D has an orthogonal countable approximate unit for A. Such a pair is said to be relative σ-unital. If D contains the unit of A, the pair is relative σ-unital. Relative version of strong Morita equivalence is called the relative Morita equivalence. We will first show the following theorem for relative σ-unital pair (A,D) of C ∗-algebras: Theorem 1.1 (Lemma 3.9, Theorem 4.7 and Theorem 5.5). Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. Then the following assertions are mutually equiva- lent: (1) (A1,D1) and (A2,D2) are relatively Morita equivalent. (2) (A1 ⊗ K,D1 ⊗ C) and (A2 ⊗ K,D2 ⊗ C) are relatively Morita equivalent. (3) There exists an isomorphism Φ : A1 ⊗ K −→ A2 ⊗ K of C ∗-algebras such that Φ(D1 ⊗ C) = D2 ⊗ C. (4) (A1,D1) and (A2,D2) are complementary relative full corners. We will second apply the above theorem to the Cuntz -- Krieger pair (OA,DA) and clarify relationships between relative Morita equivalence and flow equivalence of underlying topological dynamical systems. Theorem 1.2 (Theorem 6.3 and Theorem 7.4, cf. [21, Corollary 3.8]). Let A, B be irreducible non-permutation matrices with entries in {0, 1}. Let (OA,DA), (OB,DB) be the associated Cuntz -- Krieger pairs. Then the following assertions are mutually equivalent: (1) (OA,DA) and (OB,DB) are relatively Morita equivalent. (2) (OA ⊗ K,DA ⊗ C) and (OB ⊗ K,DB ⊗ C) are relatively Morita equivalent. (3) There exists an isomorphism Φ : OA ⊗ K −→ OB ⊗ K of C ∗-algebras such that Φ(DA ⊗ C) = DB ⊗ C. (4) (OA,DA) and (OB,DB) are corner isomorphic. (5) The two-sided topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are flow equivalent. By using J. Franks' s Theorem [8] (cf. [1], [23]), the last assertion (5) is equivalent to the following (6): (6) The groups ZN /(id − A)ZN and ZM /(id − B)ZM are isomorphic and det(id − A) = det(id − B), where N is the size of the matrix A and M is that of B. Hence we know that the group ZN /(id − A)ZN with the value det(id − A) is a complete invariant of the relative Morita equivalence class of the Cuntz -- Krieger pair (OA,DA). In [3], Brown -- Green -- Rieffel introduced the notion of the Picard group Pic(A) for a C ∗- algebra to study equivalence classes of imprimitivity bimodules of C ∗-algebras. Natural isomorphism classes [X] of imprimitivity bimodules X over A form a group under the relative tensor product [X] · [Y ] = [X ⊗A Y ]. The group is called the Picard group for the C ∗-algebra A and is written Pic(A), that are considered as a sort of generalizations of 3 automorphism group Aut(A) of A. We will introduce relative version of the Picard group Pic(A,D) as the group of (A,D) -- (A,D)-relative imprimitivity bimodules and study their structure for the Cuntz -- Krieger pairs (OA,DA). Let Aut◦(OA,DA) = {α ∈ Aut(OA) α(DA) = DA, α∗ = id on K0(OA)}. Its quotient group Aut◦(OA,DA)/ Int(OA,DA) by Int(OA,DA) is denoted by Out◦(OA,DA). Let Aut1(ZN /(id − At)ZN ) be a subgroup of Aut(ZN /(id − At)ZN ) defined by Aut1(ZN /(id − At)ZN ) = {ξ ∈ Aut(ZN /(id − At)ZN ) ξ([1]) = [1]} where [1] ∈ ZN /(id − At)ZN denotes the class of the vector (1, . . . , 1) in ZN . It is well- known that there exists an isomorphism ǫA : K0(OA) −→ ZN /(id − At)ZN such that ǫ([1OA]) = [1] ([6]). We will obtain the following structure theorem for Pic(OA,DA). Theorem 1.3 (Theorem 9.8 and Theorem 9.9). Let A be an irreducible non-permutation matrix. Then there exist short exact sequences: 1 −→ Out◦(OA,DA) ¯Ψ −→ Pic(OA,DA) K∗−→ Aut(ZN /(id − At)ZN ) −→ 1, 1 −→ Out(OA,DA) ¯Ψ −→ Pic(OA,DA) K∗−→ Aut(ZN /(id − At)ZN )/Aut1(ZN /(id − At)ZN ) −→ 1. In Appendix of the paper, we refer to the ordinary Picard groups Pic(OA) for Cuntz -- Krieger algebras OA, and especially the ordinary Picard groups Pic(ON ) for Cuntz algebras ON (Theorem 10.4 and Corollary 10.5). 2 Relative σ-unital C∗-algebras For a C ∗-algebra A, we denote by M (A) its multiplier C ∗-algebra (cf. [32]). The locally convex topology on M (A) generated by the seminorms x −→ kxak, x −→ kaxk for a ∈ A is called the strict topology. Throughout the paper, we denote by {eij}i,j∈N the matrix units on the separable infinite dimensional Hilbert space ℓ2(N). The C ∗-algebra generated by them is denoted by K which is the C ∗-algebra of all compact operators on ℓ2(N). The C ∗-subalgebra of K generated by diagonal projections {ei,i}i∈N is denoted by C. A C ∗-algebra is said to be σ-unital if it has a countable approximate unit. We will first introduce a notion of relative version of a σ-unital C ∗-algebra. Definition 2.1. A pair (A,D) of C ∗-algebras A,D is called relative σ-unital if it satisfies the following conditions: (1) D is a C ∗-subalgebra of A. (2) D contains a countable approximate unit for A. (3) There exists a sequence an ∈ A, n = 1, 2, . . . of elements such that (a) a∗ (b) P∞ (c) anda∗ ndan, anda∗ n=1 a∗ n ∈ D for all d ∈ D and n = 1, 2, . . . . nan = 1 in the strict topology of M (A). m = 0 for all d ∈ D and n, m ∈ N with n 6= m. 4 (i) a∗ nan, ana∗ n ∈ D for all n = 1, 2, . . . . k=1 a∗ n=1 a∗ We call the sequence {an}n∈N satisfying the three conditions (a), (b), (c) a relative kak belongs to D and the sequence {bn}n∈N is a countable approximate approximate unit for the pair (A,D). Remark 2.2. By the above condition (2), we know that M (D) is a C ∗-subalgebra of M (A) in natural way (cf. [32, p 46, 2G]). Lemma 2.3. Assume that (A,D) is a relative σ-unital pair of C ∗-algebras. Let {an}n∈N be a relative approximate unit for (A,D). Then we have (ii) bn =Pn unit for A. Proof. (i) Take and fix k ∈ N. Since P∞ kak ≤ 1 so that kakk ≤ 1. As D has an approximate unit for A, for any ǫ > 0, there exists d ∈ D kak − a∗ such that kak − dakk < ǫ, so that ka∗ kdak ∈ D ensures us that a∗ (ii) Since bn = Pn k=1 a∗ kak converges to 1 in the strict topology of M (A), {bn}n∈N is an approximate unit for A. Lemma 2.4. Let D be a C ∗-subalgebra of A. Then (A,D) is relative σ-unital if and only if there exists a sequence dn ∈ D, n = 1, 2, . . . such that (a) dn ≥ 0, n = 1, 2, . . . . (b) P∞ (c) dndd∗ Proof. Suppose that (A,D) is relative σ-unital. Take a relative approximate unit {an}n∈N in A. Put dn = a∗ nan. By the preceding lemma, dn belongs to D and satisfies the desired properties. Conversely, suppose that there exists a sequence dn in D satisfying the above three conditions. Put an = √dn, which becomes a relative approximate unit for (A,D). nan = 1 in M (A), we have 0 ≤ a∗ kdakk < ǫ. The condition a∗ kak belongs to D. Similarly we know that aka∗ m = 0 for all d ∈ D and n, m ∈ N with n 6= m. n=1 dn = 1 in the strict topology of M (A). k belongs to D. We call the sequence {dn}n∈N in D satisfying the conditions (a), (b), (c) in Lemma 2.4 an orthogonal approximate unit for (A,D). Example 2.5. 1. If a C ∗-subalgebra D of A contains the unit of A, the pair (A,D) is relative σ-unital by putting d1 = 1 and dn = 0 for n = 2, 3, . . . . 2. Let A = K and D = C. Then the pair (A,D) is relative σ-unital by putting dn = en,n, n ∈ N where {en,m}n,m∈N is the matrix units of K. More generally we know the following proposition. Proposition 2.6. If (A,D) is relative σ-unital, so is (A ⊗ K,D ⊗ C). Proof. Take an orthogonal approximate unit {dn}n∈N in D for the pair (A,D). Put d(n,m) = dn ⊗ em,m for n, m = 1, 2, . . . . It is straightforward to see that the sequence d(n,m), n, m = 1, 2, . . . becomes an orthogonal approximate unit for the pair (A ⊗ K,D ⊗ C). We call the pair (A ⊗ K,D ⊗ C) the relative stabilization for (A,D). Corollary 2.7. If a C ∗-subalgebra D of A contains the unit of A, both the pairs (A,D) and (A ⊗ K,D ⊗ C) are relative σ-unital. 5 3 Relative imprimitivity bimodules and relative Morita equiv- alence Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. Definition 3.1. Let X be an A1 -- A2-Hilbert C ∗-bimodule. Put XD = {x ∈ X A1hxd2 xi ∈ D1 for all d2 ∈ D2, hx d1xiA2 ∈ D2 for all d1 ∈ D1}. The A1 -- A2-Hilbert C ∗-bimodule X is called an (A1,D1) -- (A2,D2)-relative imprimitivity bimodule if it satisfies the following conditions: (1) X is an A1 -- A2-imprimitivity bimodule. (2) There exists a sequence xn ∈ XD, n = 1, 2, . . . such that n=1hxn xniA2 = 1 in the strict topology of M (A2). (a) P∞ (b) A1hxnd2 xmi = 0 for all d2 ∈ D2 and n, m ∈ N with n 6= m. (3) There exists a sequence yn ∈ XD, n = 1, 2, . . . such that n=1 A1hyn yni = 1 in the strict topology of M (A1). (a) P∞ (b) hyn d1ymiA2 = 0 for all d1 ∈ D1 and n, m ∈ N with n 6= m. Remark 3.2. (1) Since X is an A1 -- A2-imprimitivity bimodule, norms on X defined 2k for by their inner products coincide each other, that is, kA1hx xik x ∈ X (cf. [25, Proposition 3.1]). We denote the norm by kxk. 1 1 2 = khx xiA2k (2) The above elements xn, yn ∈ XD in Definition 3.1 satisfy the inequalities A1hxn xni ≤ 1, hyn yniA2 ≤ 1 (3.1) because of the inequality A1hxn xni ≤ kA1hxn xnik = khxn xniA2k ≤ k ∞ Xn=1 hxn xniA2k = 1 and of a similar inequality for hyn yniA2 . (3) Both the left action of A1 and the right action of A2 on X are non-degenerate, that is, A1X = X = XA2. More strongly we see that D1X = X = XD2. In fact, for d1 ∈ D1 and x ∈ X, the following inequalities hold kx − d1xk2 = kA1hx − d1x x − d1xik = kA1hx xi − d1A1hx xi − A1hx xid∗ 1 + d1A1hx xid∗ 1k ≤ kA1hx xi − d1A1hx xik + kA1hx xi − d1A1hx xikkd∗ 1k. As D1 has a countable approximate unit for A1, we have a sequence d1(n) in D1 such that limn→∞ kx − d1(n)xk = 0 so that D1X = X. 6 Lemma 3.3. For x ∈ XD we have (i) A1hx xi ∈ D1. (ii) hx xiA2 ∈ D2. Proof. (i) Let x ∈ XD. For d2 ∈ D2, we have hx − xd2 x − xd2iA2 = hx xiA2 − hx xiA2 d2 − d∗ 2hx xiA2 + d∗ 2hx xiA2d2. (3.2) Now D2 contains an approximate unit for A2, the equality (3.2) shows that for any ǫ > 0 there exists an element d2 ∈ D2 such that khx − xd2 x − xd2iA2k < ǫ. Since X is an A1 -- A2-imprimitivity bimodule, we see that kA1hx − xd2 x − xd2ik < ǫ by [25, Lemma 3.11]. By the Cauchy -- Schwartz inequality (cf. [25, Lemma 2.5]), we have kA1hx − xd2 xik2 =kA1hx − xd2 xi∗ A1hx − xd2 xik ≤kA1hx − xd2 x − xd2ikkA1hx xik <ǫkA1hx xik. Hence we have kA1hx xi − A1hxd2 xik2 = kA1hx − xd2 xik2 < ǫkA1hx xik. (3.3) As A1hxd2 xi belongs to D1, we conclude that A1hx xi belongs to D1. (ii) is similarly shown to (i). Lemma 3.4. n=1 A1hz xnixn for z ∈ X which converges in the norm of X, and n=1 ynhyn ziA2 for z ∈ X which converges in the norm of X, and (i) We have z =P∞ A1hxn xmi = 0 for n, m ∈ N with n 6= m. (ii) We have z =P∞ hyn ymi = 0 for n, m ∈ N with n 6= m. Proof. (i) As X = XD2, for z ∈ X and ǫ > 0 there exists d2 ∈ D2 such that kz − zd2k < ǫ. SinceP∞ n=1hxn xniA2 = 1 in the strict topology of M (A2), we may find K ∈ N such that kPK Xn=1 kz − Xn=1 n=1 d2hxn xniA2 − d2k < ǫ. Therefore we have A1hz xnixnk K K zd2hxn xniA2 − zhxn xniA2k K Xn=1 hxn xniA2k K Xn=1 K =kz − ≤kz − zd2k + kzd2 − zhxn xniA2k Xn=1 zd2hxn xniA2k + k Xn=1 ≤kz − zd2k + kzkkd2 − =(2 + kzk)ǫ, K K Xn=1 d2hxn xniA2k + k(zd2 − z) 7 so that P∞ As in the proof of Lemma 3.3, for n, m ∈ N with n 6= m, there exists d2(k) ∈ D2 such n=1 A1hz xnixn converges to z in the norm of X. that n→∞kA1hxn xmi − A1hxnd2 xmik2 = 0. lim Since A1hxnd2 xmi = 0, we have A1hxn xmi = 0. The sequences {xn}n∈N,{yn}n∈N ⊂ XD satisfying the conditions (2), (3) in Definition 3.1 are called a relative left basis, a relative right basis, respectively. The pair ({xn},{yn}) is called a relative basis for X. We arrive at our definition of relative version of strong Morita equivalence. (A2,D2). Definition 3.5. Two relative σ-unital pairs of C ∗-algebras (A1,D1) and (A2,D2) are said to be relatively Morita equivalent if there exists an (A1,D1) -- (A2,D2)-relative imprimitivity bimodule. In this case we write (A1,D1) ∼RME Lemma 3.6. Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. If there exists an isomorphism θ : A1 −→ A2 of C ∗-algebras such that θ(D1) = D2, then we have (A2,D2). In particular, for a relative σ-unital pair (A,D) of C ∗-algebras, (A1,D1) ∼RME we have (A,D) ∼RME Proof. Let an ∈ A1, n ∈ N be a relative approximate unit for (A1,D1). Put Xθ = A1 as vector space having module structure and inner products given by (A,D). a1 · x · a2 := a1xθ−1(a2) A1hx yi = xy∗, for a1 ∈ A1, a2 ∈ A2, x ∈ Xθ, for x, y ∈ Xθ. hx yiA2 = θ(x∗y) Put xn = an, n ∈ N. We have for d1 ∈ D1, d2 ∈ D2 A1hxnd2 xni = anθ−1(d2)a∗ n ∈ D1, hxn d1xniA2 = θ(a∗ nd1an) ∈ D2 so that xn ∈ (Xθ)D. We also have Xn=1 hxn xniA2 = ∞ ∞ Xn=1 θ(a∗ nan) = 1, and A1hxnd2 xmi = anθ−1(d2)a∗ m = 0 for all d2 ∈ D2 and n, m ∈ N with n 6= m. Similarly by putting yn = a∗ A1hynd2 yni = a∗ n we have nθ−1(d2)an ∈ D1, hyn d1yniA2 = θ(and1a∗ n) ∈ D2 so that yn ∈ (Xθ)D. We also have Xn=1 ∞ A1hyn yni = a∗ nan = 1, ∞ Xn=1 8 and hyn d1ymi = θ(and1a∗ m) = 0 for all d1 ∈ D1 and n, m ∈ N with n 6= m. Hence ({xn},{yn}) is a relative basis for Xθ so that Xθ becomes an (A1,D1) -- (A2,D2)- relative imprimitivity bimodule to show (A1,D1) ∼RME (A2,D2). We will next show that the relation ∼RME pairs of C ∗-algebras. is an equivalence relation in relative σ-unital Lemma 3.7. Suppose that X12 is an (A1,D1) -- (A2,D2)-relative imprimitivity bimodule and X23 is an (A2,D2) -- (A3,D3)-relative imprimitivity bimodule. Then the relative tensor product X12⊗A2 X23 of bimodules is an (A1,D1) -- (A3,D3)-relative imprimitivity bimodule. Proof. Take relative bases ({xn},{yn}) for X12 and ({zn},{wn}) for X23. We will show that the pair ({xn ⊗ zm}n,m,{yn ⊗ wm}n,m) becomes a relative basis for X12 ⊗A2 X23. For d3 ∈ D3, d1 ∈ D1, we have A1h(xn ⊗ zm)d3 xn ⊗ zmi =A1hxn ⊗ (zmd3) xn ⊗ zmi = A1hxnA2hzmd3 zmi xni, hxn ⊗ zm d1(xn ⊗ zm)iA3 =hxn ⊗ zm (d1xn) ⊗ zmiA3 = hzm hxn d1xniA2 zmiA3. As A2hzmd3 zmi ∈ D2, we have A1hxnA2hzmd3 zmi xni ∈ D1 so that A1h(xn ⊗ zm)d3 xn ⊗ zmi ∈ D1. Similarly we know that A1hxnA2hzmd3 zmi xni ∈ D3. We also have ∞ Xn,m=1 hxn ⊗ zm xn ⊗ zmiA3 = = = ∞ ∞ Xn,m=1 Xm=1 Xm=1 ∞ hzm hxn xniA2zmiA3 ∞ Xn=1 hxn xniA2)zmiA3 hzm ( hzm zmiA3 = 1. For d3 ∈ D3, we have A1h(xn ⊗ zm)d3 xl ⊗ zki = A1hxn ⊗ (zmd3) xl ⊗ zki = A1hxnA2hzmd3 zki xli. If m 6= k, then A2hzmd3 zki = 0. If n 6= l, then A1hxnA2hzmd3 zki xli = 0 because A2hzmd3 zki ∈ D2. Hence if (n, m) 6= (l, k), we have A1h(xn ⊗ zm)d3 xl ⊗ zki = 0 and know that the sequence {xn ⊗ zm}n,m is a relative left basis for X12 ⊗A2 X23. By a similar argument, we know that {yn ⊗ wm}n,m is a relative right basis for X12 ⊗A2 X23, so that ({xn ⊗ zm}n,m,{yn ⊗ wm}n,m) is a relative basis for X12 ⊗A2 X23. Therefore we have Proposition 3.8. Relative Morita equivalence ∼RME σ-unital pairs of C ∗-algebras. is an equivalence relation in relative 9 Proof. The refrexisive law follows from Lemma 3.6. We will show the symmetric law. Suppose that (A1,D1) ∼RME (A2,D2) via relative imprimitivity bimodule X12. Then its conjugate module ¯X12 denoted by X21 becomes an (A2,D2) -- (A1,D1)-relative imprimitiv- ity bimodule (see [28, Definition 6.17], cf. [10, p. 3443]), so that (A1,D2) ∼RME (A1,D1). The transitive law follows from Lemma 3.7. Lemma 3.9. Let (A,D) be a relative σ-unital pair of C ∗-algebras. Then we have (A,D) ∼RME (A ⊗ K,D ⊗ C). Proof. Let an ∈ A, n ∈ N be a relative approximate unit for (A,D). Let {en,m}n,m∈N be the matrix units of K. Define X = A ⊗ e1,1K. By identifying A with A ⊗ Ce1,1, X has a natural structure of A -- A⊗K-imprimitivity bimodule. Put xn,m = an⊗e1,m ∈ X, n, m ∈ N. For d1 ∈ D and d2 = d ⊗ f ∈ D ⊗ C, we have Ahxn,md2 xn,mi =(an ⊗ e1,m)(d ⊗ f )(an ⊗ e1,m)∗ = anda∗ n ⊗ e1,mf e∗ 1,m ∈ D ⊗ Ce1,1, hxn,m d1xn,miA⊗K =(an ⊗ e1,m)∗(d ⊗ e1,1)(an ⊗ e1,m) = a∗ ndan ⊗ em,1e1,1e1,m ∈ D ⊗ C, so that xn,m belongs to XD under the identification between D with D ⊗ Ce1,1. We also have ∞ ∞ hxn,m xn,miA⊗K = Xn,m=1 in M (A ⊗ K). For d2 = d ⊗ f ∈ D ⊗ C, we have Xn,m=1 (an ⊗ e1,m)∗(an ⊗ e1,m) = ∞ Xn,m=1 a∗ nan ⊗ e∗ 1,me1,m = 1 ⊗ 1 Ahxn,md2 xk,li = (an ⊗ e1,m)(d ⊗ f )(ak ⊗ e1,l)∗ = anda∗ k ⊗ e1,mf e∗ 1,l. k = 0. If m 6= l, we have e1,mf e∗ 1,l = 0. Hence if (n, m) 6= (k, l), we If n 6= k, we have anda∗ have Ahxn,md2 xk,li = 0. Put yn = a∗ n ⊗ e1,1. Then for d1 ∈ D and d2 = d ⊗ f ∈ D ⊗ C, we have Ahynd2 yni =(a∗ hyn d1yniA⊗K =(a∗ n ⊗ e1,1)(d ⊗ f )(a∗ n ⊗ e1,1)∗(d ⊗ e1,1)(a∗ n ⊗ e1,1)∗ = a∗ ndan ⊗ e1,1f e∗ 1,1 ∈ D ⊗ Ce1,1, n ⊗ e1,1) = anda∗ n ⊗ e1,1 ∈ D ⊗ C, so that yn belongs to XD. We also have ∞ Xn=1 Ahyn yni = ∞ Xn=1 a∗ nan ⊗ e1,1 = 1 ⊗ e1,1, and hyn d1ymi = (a∗ n ⊗ e1,1)∗(d ⊗ e1,1)(a∗ m ⊗ e1,1) = anda∗ m ⊗ e1,1 = 0 for n 6= m. Therefore X becomes an (A,D) -- (A ⊗ K,D ⊗ C)-relative imprimitivity bimodule, so that (A,D) ∼RME (A ⊗ K,D ⊗ C). 10 Example 3.10. For m, k ∈ N, let A1 = Mm(C),D1 = diag(Mm(C)) = Cm, and A2 = Mk(C),D2 = diag(Mk(C)) = Ck. Then we have (A1,D1) ∼RME (A2,D2). We will present an (A1,D1) -- (A2,D2)-relative imprimitivity bimodule in the followung way. Let A0,D0 be Mm+k(C), diag(Mm+k(C)), respectively. Let p1, p2 be the projections in D0 defined by m k m k p1 = ( We then have A1 = p1A0p1, z } { z } { 0,··· , 0), 1,··· , 1, D1 = D0p1 p2 = ( z } { 0,··· , 0, and A2 = p2A0p2, z } { 1,··· , 1). D1 = D0p2. Put X = p1A0p2 with natural A1 -- A2-bimodule structure and inner products such taht Ahx yi = xy∗, hx yiA2 = x∗y for x, y ∈ X. It is not difficult to see that X becomes an (A1,D1) -- (A2,D2)-relative imprimitivity bi- module so that (A1,D1) ∼RME (A2,D2). 4 Isomorphism of relative stabilizations In this section, we devote to proving the following theorem, which is a relative version of a part of Brown -- Green -- Rieffel Theorem [3, Theorem 1.2]. Theorem 4.1. Suppose (A1,D1) ∼RME A1 ⊗ K −→ A2 ⊗ K of C ∗-algebras such that Φ(D1 ⊗ C) = D2 ⊗ C. (A2,D2). Then there exists an isomorphism Φ : Suppose that X is an (A1,D1) -- (A2,D2)-relative imprimitivity bimodule. Let ¯X be [10, p.3443]). The corresponding the conjugate bimodule of X ([28, Definition 6.17], cf. element in ¯X to y ∈ X is denoted by ¯y. It is straightforward to see that ¯X is (A2,D2) -- (A1,D1)-relative imprimitivity bimodule. We define the relative linking pair (A0,D0) by setting ¯y A0 = {(cid:20)a1 D0 = {(cid:20)d1 0 x a2(cid:21) a1 ∈ A1, a2 ∈ A2, x ∈ X, ¯y ∈ ¯X}, d2(cid:21) d1 ∈ D1, d2 ∈ D2}. 0 (4.1) (4.2) As in [3, p.350], the products between two elements of A0 is defined by hy ziA2 + a2b2(cid:21) . ¯w b2(cid:21) :=(cid:20)a1b1 + A1hx wi a2(cid:21)(cid:20)b1 (cid:20)a1 ¯yb1 + a2 ¯w a1z + xb2 x ¯y z Let X ⊕ A2 be the Hilbert C ∗-right module over A2 with the natural right action of A2 and A2-valued right inner product defined by h(cid:20) x a2(cid:21) (cid:20) y b2(cid:21)iA2 := hx yiA2 + a2b2. 11 The algebra A0 acts on X ⊕ A2 by a2(cid:21)(cid:20) z x (cid:20)a1 ¯y b2(cid:21) =(cid:20) a1z + xb2 hy ziA2 + a2b2(cid:21) . As seen in [25, Lemma 3.20], A0 itself is a C ∗-subalgebra of all bounded adjointable operators on the Hilbert C ∗-right module X ⊕ A2. We set P2 =(cid:20)0 0 0 1(cid:21) . P1 =(cid:20)1 0 0 0(cid:21) , (4.3) They satisfy P1 + P2 = 1 and P1A0P1 = A1, D0P1 = D1 and P2A0P2 = A2, D0P2 = D2. (4.4) To prove Theorem 4.1, we provide several lemmas. 0 n=1 U ∗ m = 0 for n 6= m. 0(cid:21) ∈ A0, n ∈ N. The sequence Un satisfies the following conditions: nUn which converges in the strict topology of M (A0). Lemma 4.2. Let ({xn},{yn}) ⊂ XD be a relative bases for X. (i) Put Un =(cid:20)0 xn (a) P2 =P∞ (b) UnU ∗ (c) UnD0U ∗ (d) U ∗ (ii) Put Tn =(cid:20) 0 (a) P1 =P∞ (b) TnT ∗ (c) TnD0T ∗ (d) T ∗ n ≤ P1 and UnU ∗ n ⊂ D0P1 = D1. nD0Un ⊂ D0P2 = D2. ¯yn 0(cid:21) ∈ A0, n ∈ N. The sequence Tn satisfies the following conditions: n Tn which converges in the strict topology of M (A0). n ≤ P2 and TnT ∗ n ⊂ D0P2 = D2. nD0Tn ⊂ D0P1 = D1. m = 0 for n 6= m. n=1 T ∗ 0 Proof. (i) For d1 ∈ D1, d2 ∈ D2, we have U ∗ n(cid:20)d1 0 0 d2(cid:21) Un =(cid:20)0 0 hxn d1xniA2(cid:21) . 0 (4.5) Since xn ∈ XD and d1 ∈ D1, we have hxn d1xniA2 ∈ D2, so that U ∗ shows (d). Since we have nD0Un ⊂ D0P2, which U ∗ nUn =(cid:20)0 0 hxn xniA2(cid:21) 0 n=1hxn xniA2 = 1 implies P∞ the equality P∞ for d1 ∈ D1, d2 ∈ D2, we have Un(cid:20)d1 d2(cid:21) U ∗ 0 0 n =(cid:20)A1hxnd2 xni 0 0(cid:21) . 0 n=1 U ∗ nUn = P2 which shows (a). And also (4.6) (4.7) 12 Since xn ∈ XD and d2 ∈ D2, we have A1hxnd2 xni ∈ D1 so that UnD0U ∗ shows (c). Since we have n ⊂ D0P1, which UnU ∗ m =(cid:20)A1hxn xmi 0 0(cid:21) . 0 (4.8) n ≤ P1 and A1hxn xmi = 0 for n, m ∈ N with 0(cid:21) = Un + Tn. It then follows that T ∗ n Tn = P2 + P1 = 1. the inequality A1hxn xni ≤ 1 implies UnU ∗ n 6= m. which shows (b). Lemma 4.3. The pair (A0,D0) is relative σ-unital. Proof. Keep the above notations. Put an =(cid:20) 0 xn Xn=1 Xn=1 Xn=1 a∗ nan = nUn + U ∗ ¯yn ∞ ∞ ∞ For d1 ∈ D1, d2 ∈ D2, we have d2(cid:21) an = U ∗ n(cid:20)d1 n(cid:20)d1 a∗ 0 0 0 0 0 d2(cid:21) Tn d2(cid:21) Un + T ∗ n(cid:20)d1 hxn d1xniA2(cid:21) ∈ D1 ⊕ D2 = D0. 0 0 0 =(cid:20)h¯yn d2 ¯yniA1 d2(cid:21) a∗ 0 Similarly we have an(cid:20)d1 UndU ∗ approximate unit for (A0,D0) to show (A0,D0) is relative σ-unital. m = (Un+Tn)d(Um+Tm)∗ = m = 0 for d = d1 + d2 ∈ D1 ⊕ D2 and n 6= m. Hence {an} is a relative n ∈ D1⊕D2. We also have anda∗ m + TndT ∗ 0 j=1 jk,jsjk,j = fj, sjk,js∗ Let us decompose the set N of natural numbers into disjoint infinite subsets N = ∪∞ Nj, and decompose Nj for each j once again into disjoint infinite sets Nj = ∪∞ Njk. Let {ei,j}i,j∈N be the set of matrix units which generate the algebra K = K(ℓ2(N)). Put ei,i and fjk = Pi∈Njk the projections fj = Pi∈Nj ei,i. Take a partial isometry sjk,j such that s∗ jk,j. We set for n = 1, 2, . . . , Xk=1 Uk ⊗ snk,n, Xl=1 jk,j = fjk and put sj,jk = s∗ wn = P1 ⊗ sn0,n + un, zn = P2 ⊗ sn0,n + tn. Tl ⊗ snl,n, un = tn = k=0 ∞ ∞ Then we have Lemma 4.4 (cf. [20, Lmma 3.3]). For each n ∈ N, we have (i) wn is a partial isometry in M (A0 ⊗ K) such that nwn = 1 ⊗ fn. n ≤ P1 ⊗ fn. (a) w∗ (b) wnw∗ (c) wn(D0 ⊗ C)w∗ n ⊂ D1 ⊗ C. 13 (d) w∗ n(D0 ⊗ C)wn ⊂ D2 ⊗ C. (ii) zn is a partial isometry in M (A0 ⊗ K) such that nzn = 1 ⊗ fn. n ≤ P2 ⊗ fn. (a) z∗ (b) znz∗ (c) zn(D0 ⊗ C)z∗ (d) z∗ n ⊂ D2 ⊗ C. n(D0 ⊗ C)zn ⊂ D1 ⊗ C. Proof. (i) Since u∗ nun = P2 ⊗ fn, we have w∗ nwn = P1 ⊗ fn + u∗ nun = P1 ⊗ fn + P2 ⊗ fn = 1 ⊗ fn. As un(P2 ⊗ sn,n0) = (P2 ⊗ sn,n0)u∗ n = 0, we have wnw∗ n = P1 ⊗ fn0 + unu∗ n = P1 ⊗ fn0 + ∞ Xk=1 UkU ∗ k ⊗ fnk. Since fn0, fnk ≤ fn, we have (ii) is similarly shown to (i). wnw∗ n ≤ P1 ⊗ fn. We will construct and study the unitary V1 in M (A0⊗K) such that Ad(V1) : A0⊗K −→ Let fn,m be a partial isometry satisfying f ∗ n,m = fn. The following n,mfn,m = fm, fn,mf ∗ A1 ⊗ K and Ad(V1)(D0 ⊗ C) = D1 ⊗ C lemma is straightforward. Lemma 4.5 (cf. [20, Lemma 3.4]). We put v1 = w1 = P1 ⊗ s10,1 + u1, v2n = (P1 ⊗ fn − v2n−1v∗ v2n−1 = wn(1 ⊗ fn − v∗ 2n−2v2n−2) 2n−1)(P1 ⊗ fn,n+1) for 1 ≤ n ∈ N, for 2 ≤ n ∈ N. Then we have for n ∈ N (a) v∗ 2n−2v2n−2 + v∗ 2n−1v2n−1 = 1 ⊗ fn. 2n−1 + v2nv∗ 2n = P1 ⊗ fn. (b) v2n−1v∗ (c) vn(D0 ⊗ C)v∗ (d) v∗ n ⊂ D1 ⊗ C. n(D1 ⊗ C)vn ⊂ D0 ⊗ C. By the above lemma, one may see that the summationP∞ n=1 vn converges in M (A0⊗K) to certain partial isometry written V1 in the strict topology of M (A0 ⊗ K). Similarly we obtain a partial isometry V2 in M (A0⊗K) constructed from the preceding sequences tn, zn of partial isometries. As a consequence, we obtain the following proposition. Proposition 4.6. Assume that (A1,D1) ∼RME pair defined in (4.1), (4.2). (A2D2). Let (A0,D0) be the relative linking 14 (i) There exists an isometry V1 in M (A0 ⊗ K) such that (a) V ∗ 1 V1 = 1 ⊗ 1. (b) V1V ∗ 1 = P1 ⊗ 1. (c) V1(D0 ⊗ C)V ∗ (d) V ∗ 1 = D1 ⊗ C. 1 (D1 ⊗ C)V1 = D0 ⊗ C. (ii) There exists an isometry V2 in M (A0 ⊗ K) such that (a) V ∗ 2 V2 = 1 ⊗ 1. (b) V2V ∗ 2 = P2 ⊗ 1. (c) V2(D0 ⊗ C)V ∗ (d) V ∗ 2 = D2 ⊗ C. 2 (D2 ⊗ C)V2 = D0 ⊗ C. Therefore we reach the following theorem Theorem 4.7. Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. Then (A1,D1) ∼RME (A2,D2) if and only if there exists an isomorphism Φ : A1 ⊗ K −→ A2 ⊗ K of C ∗-algebras such that Φ(D1 ⊗ C) = D2 ⊗ C. Proof. Suppose (A1,D1) ∼RME sition 4.6. Put Φ = Ad(V2V ∗ of C ∗-algebras such that Φ(D1 ⊗ C) = D2 ⊗ C. (A2,D2). Take isometries V1, V2 in M (A0 ⊗ K) as in Propo- 1 ) which gives rise to an isomorphism Φ : A1 ⊗ K −→ A2 ⊗ K Converse implication comes from Lemma 3.6 and Lemma 3.9. 5 Relative full corners It is well-known that two C ∗-algebras are strong Morita equivalent if and only if they are complementary full corners of some C ∗-algebra ([3, Theorem 1.1]). In this section, we will study a relative version of this fact. Definition 5.1. For a relative σ-unital pair (A,D) of C ∗-algebra, a projection P ∈ M (D) is said to be relative full for (A,D) if it satisfies the following conditions (1) P d = dP for all d ∈ D. (2) There exists an sequence an ∈ A, n = 1, 2, . . . such that (a) a∗ ndan ∈ D, anda∗ n=1 a∗ n ∈ DP for all d ∈ D and n = 1, 2, . . . . nP an = 1 − P in the strict topology of M (A). (c) anda∗ (b) P∞ We call the sequence {an}n∈N satisfying the three conditions (a), (b), (c) a relative full m = 0 for all d ∈ D and n, m ∈ N with n 6= m. sequence for P . 15 Remark 5.2. By the above condition (b), we know that (b′) a∗ ndP an ∈ D(1 − P ) for all d ∈ D, because we have (a∗ ndP an)∗a∗ ndP an = a∗ nP d∗ana∗ ndP an ≤ kd∗ana∗ ndka∗ nP an ≤ 1 − P. Definition 5.3. Relative σ-unital pairs (A1,D1) and (A2,D2) of C ∗-algebras are said to be complementary relative full corners if there exists a relative σ-unital pair (A0,D0) of C ∗-algebras such that there exist relative full projections P1, P2 ∈ M (D0) such that and i = 1, 2. P1 + P2 = 1 PiA0Pi = Ai, D0Pi = Di, (5.1) Proposition 5.4. Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. If they are complementary relative full corners, then we have (A1,D1) ∼RME Proof. Let (A0,D0) and Pi ∈ M (D0), i = 1, 2 be a relative σ-unital pair of C ∗-algebras and projections, respectively, satisfying Definition 5.3. Let {an} and {bn} be relative full sequences for the projections P1, P2, respectively. We set X = P1A0P2 and define two sequences by xn = P1anP2 and yn = P1b∗ nP2. For d ∈ D0, put di = dPi, i = 1, 2. It then follows that (A2,D2). A1hxnd2 xni = P1anP2d2P2a∗ hxn d1xniA2 = P2a∗ nP1 ∈ D0P1 = D1, nP1d1P1anP2 ∈ D0P2 = D2. Hence xn belongs to XD. We also have ∞ Xn=1 hxn xniA2 = ∞ Xn=1 P2a∗ nP1anP2 = P2, and A1hxnd2 xmi = P1anP2dP2a∗ mP1 = 0 for n 6= m. Hence {xn} is a relative left basis for X. Similarly we have A1hynd2 yni = P1b∗ hyn d1yniA2 = P2bnP1d1P1b∗ nP2d2P2bnP1 ∈ D0P1 = D1, nP2 ∈ D0P2 = D2. Hence yn belongs to XD. We also have ∞ Xn=1 A1hyn yni = ∞ Xn=1 P1b∗ nP2bnP1 = P1, and hyn d1ymiA2 = P2bnP1dP1b∗ mP2 = 0 for n 6= m. Hence {yn} is a relative right basis for X. Therefore X is an (A1,D1) -- (A2,D2)-relative imprimitivity bimodule, so that we have (A1,D1) ∼RME (A2,D2). 16 We obtain the following theorem. Theorem 5.5. Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. Then (A2,D2) if and only if (A1,D1) and (A2,D2) are complementary relative (A1,D1) ∼RME full corners. Proof. The if part has been proved in Proposition 5.4. To show the only if part, suppose (A1,D1) ∼RME (A2,D2). Take (A0,D0) the linking pair defined in (4.1), (4.2). Let P1, P2 be the projections in M (D0) defined by (4.3). Take the sequence Un, Tn as in Lemma 4.2. The proof of Lemma 4.2 shows us that the sequences an := Un and bn := Tn are relative full sequences for P1 and P2, respectively, so that P1 and P2 are relative full projections for (A0,D0). Since P1 + P2 = 1, the equalities (4.4) show that (A1,D1) and (A2,D2) are complementary relative full corners. 6 Relative Morita equivalence in Cuntz -- Krieger pairs In this section, we will study relative Morita equivalence particularly in Cuntz -- Krieger algebras from a viewpoint of symbolic dynamical systems. For a nonnegative matrix A = [A(i, j)]N i,j=1, the associated directed graph GA = (VA, EA) consists of the vertex set VA = {vA N} of N -vertices and the edge set EA = {a1, . . . , aNA} where there are A(i, j) edges from vA j . For ai ∈ EA, denote by t(ai), s(ai) the terminal vertex of i ai and the source vertex of ai, respectively. The graph GA has the NA × NA transition matrix AG = [AG(ai, aj)]NA i,j=1 of edges defined by 1 , . . . , vA to vA AG(ai, aj) =(1 0 if t(ai) = s(aj), otherwise (6.1) for ai, aj ∈ EA. The Cuntz -- Krieger algebra OA for the matrix A is defined as the Cuntz -- Krieger algebra OAG for the matrix AG which is the universal C ∗-algebra generated by partial isometries Sai indexed by edges ai, i = 1, . . . , NA subject to the relations: Saj S∗ aj = 1, S∗ aiSai = NA Xj=1 NA Xj=1 AG(ai, aj)Saj S∗ aj for i = 1, . . . , NA. (6.2) The subalgebra DA is defined as the algebra DAG. The pair (OA,DA) is called the Cuntz -- Krieger pair for the matrix A. Since 1 ∈ DA ⊂ OA, the pair (OA,DA) is relative σ-unital. As in [17], the isomorphism class of the pair (OA,DA) is exactly corresponding to the continuous orbit equivalence class of the underlying one-sided topological Markov shift (XA, σA). Its complete classification result has been obtained in [21, Theorem 3.6]. Let A, B be irreducible square matrices with entries in nonnegative integers. In [34], R. F. Williams proved that the two-sided topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are topologically conjugate if and only if the matrices A, B are strong shift equivalent. Two nonnegative matrices A, B are said to be elementary equivalent if there exist nonnegative rectangular matrices C, D such that A = CD, B = DC. If there exists a finite sequence of nonnegative matrices A0, A1, . . . , An such that A = A0, B = An and Ai is elementary equivalent to Ai+1 for i = 1, 2, . . . , n − 1, then A and B are said to be strong shift 17 equivalent ([34]). Hence topological conjugacy of two-sided topological Markov shifts is generated by a finite sequence of elementary equivalence of underlying matrices. Let us denote by Bk( ¯XA) the set of admissible words with length k of the topological Markov shift ( ¯XA, ¯σA). Put B∗( ¯XA) = ∪∞ k=0Bk( ¯XA). In this section we will first show the following proposition. Proposition 6.1. Suppose that two nonnegative square matrices A and B are elementary (OB,DB). equivalent such that A = CD and B = DC. Then we have (OA,DA) ∼RME Proof. Suppose that the size of A is N and that of B is M so that C is an N × M matrix and D is an M × N matrix, respectively. We set the square matrix Z =(cid:20) 0 C D 0(cid:21) as block matrix, and we see Z 2 =(cid:20)CD 0 0 DC(cid:21) =(cid:20)A 0 0 B(cid:21) . Let us consider the Cuntz -- Krieger algebra OZ for the matrix Z. Since EZ = EC ∪ ED, we may write the canonical generating partial isometries of OZ as Sc, Sd, c ∈ EC, d ∈ ED so that Pc∈EC d = 1 and SdS∗ ScS∗ Z(c, d)SdS∗ d, S∗ dSd = Xc∈EC Z(d, c)ScS∗ c c +Pd∈ED c Sc = Xd∈ED S∗ for c ∈ EC , d ∈ ED. Let us denote by Sa, a ∈ EA (resp. Sb, b ∈ EB) the canonical generating partial isometries of OA (resp. OB) satisfying the relations (1.1). As Z 2 = (cid:20)A 0 0 B(cid:21) , we have a bijective correspondence ϕA,CD from EA to a subset of EC × ED (resp. ϕB,DC from EB to a subset of ED × EC ) such that ScSd 6= 0 (resp. SdSc 6= 0) if and only if ϕA,CD(a) = cd (resp. ϕB,DC(b) = dc) for some a ∈ EA (resp. b ∈ EB), we may identify cd (resp. dc) with a (resp. b) through the map ϕA,CD (resp. ϕB,DC ). We may then write Scd = Sa (resp. Sdc = Sb) where Scd denotes ScSd (resp. Sdc denotes SdSc). Define the d. Both of them belong to projections in OZ by PA =Pc∈EC DZ and satisfy PA + PB = 1. It has been shown in [16] (cf. [20]) that SdS∗ ScS∗ PAOZ PA = OA, PBOZ PB = OB, DZ PB = DB. (6.3) c and PB =Pd∈ED DZPA = DA, We put X = PAOZ PB which has a natural structure of OA − OB imprimitivity bimodule under the identification PAOZ PA = OA, PBOZ PB = OB. We will prove that X becomes (OA,DA) -- (OB,DB)-relative imprimitivity bimodule. Put EC = {c1, . . . , cNC} and ED = {d1, . . . , dND} for the matrices C and D respectively. For k = 1, . . . , ND, take c(k) ∈ EC such that c(k)dk ∈ B2(XZ ) so that we have S∗ c(k)Sc(k) ≥ Sdk S∗ dk . Similarly for l = 1, . . . , NC , take d(l) ∈ ED such that d(l)cl ∈ B2(XZ ) so that we have S∗ d(l)Sd(l) ≥ SclS∗ cl. 18 We set xk = Sc(k)Sdk S∗ dk yl = Sd(l)SclS∗ cl for k = 1, . . . , ND, for l = 1, . . . , ND. For d1 ∈ DA, d2 ∈ DB, we have OAhxkd2 xki =Sc(k)Sdk S∗ hxk d1xkiOB =Sdk S∗ dk d2Sdk S∗ dk S∗ dk Sc(k)d1Sc(k)Sdk S∗ c(k) ∈ DA, dk ∈ DB so that xk belongs to XD and similarly yl belongs to XD. We also have ND Xk=1 hxk xkiOB = = = For n 6= m, we have ND ND Xk=1 Xk=1 Xk=1 ND (Sc(k)Sdk S∗ dk )∗(Sc(k)Sdk S∗ dk ) Sdk S∗ dk S∗ c(k)Sc(k)Sdk S∗ dk Sdk S∗ dk = PB dnd2SdmS∗ OAhxnd2 xmi = Sc(n)SdnS∗ l=1 OAhyl yli = PA and hyn d1ymiOB = 0 for n 6= m, so that X Similarly we have PNC becomes (OA,DA) -- (OB,DB)-relative imprimitivity bimodule. shifts is generated by strong shift equivalences and expansions A → A defined bellow. In [23], Parry -- Sullivan proved that the flow equivalence relation of topological Markov c(m) = 0. dmS∗ For an N × N matrix A = [A(i, j)]N i,j=1 with entries in {0, 1}, put A = , (6.4)   0 0 A(1, 1) 1 0 A(2, 1) ... 0 A(N, 1) ... ··· A(1, N ) ··· ··· A(2, N ) 0 ... ··· A(N, N )   which is called the expansion of A at the vertex 1. The expansion of A at other vertices are similarly defined. (O A,D A). Proposition 6.2. (OA,DA) ∼RME Proof. Let {0, 1, . . . , N} be the set of symbols for the topological Markov shifts ( ¯X A, ¯σ A) defined by the matrix A. Let us denote by S0, S1, . . . , SN the canonical generating partial isometries of the Cuntz -- Krieger algebra O A satisfying PN Si = PN j for i = 0, 1, . . . , N . Put P =PN i . The identities A(i, j) Sj S∗ j = 1, S∗ Sj S∗ Si S∗ j=0 j=0 i=1 i P + S0 S∗ 0 = P + S∗ 1P S1 = 1 S∗ 1 P S1 = S∗ 1 S1 = S0 S∗ 0 , (6.5) 19 hold, so that we have PO AP = OA, (6.6) We put X = PO A which has a natural structure of OA−O A imprimitivity bimodule under the identification PO AP = OA,D AP = DA. We will prove that X becomes (OA,DA) -- (O A,D A)-relative imprimitivity bimodule. We set x1 = P, x2 = P S1 and y1 = P . For d1 ∈ DA, d2 ∈ D A, we have D AP = DA. OAhx1d2 x1i =P d2P ∈ DA, OAhx2d2 x2i =P S1d2 S∗ hd1x1 x1iO A hd1x2 x2iO A 1 P ∈ DA, =P d1P ∈ DA ⊂ D A, = S∗ 1P d1P S1 ∈ D A, so that x1, x2, y1 ∈ XD. We also have 2 Xk=1 hxk xkiO A OAhx1d2 x2i =P d2(P S1)∗ = P d2 S∗ OAhx2d2 x1i =P S1d2P ∗ = 0. =P ∗P + (P S1)∗(P S1) = P + S∗ 1P S1 = 1, 1P = 0, Hence X becomes (OA,DA) -- (O A,D A)-relative imprimitivity bimodule. We have thus obtained the following theorem. Theorem 6.3. If two-sided topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are flow equivalent, then the Cuntz -- Krieger pairs (OA,DA) and (OB,DB) are relatively Morita equivalent. 7 Corner isomorphisms in Cuntz -- Krieger pairs Let A, B, Z be square irreducible non-permutation matrices with entries in nonnegative integers. Definition 7.1. Two Cuntz -- Krieger pairs (OA,DA) and (OZ ,DZ ) are said to be elemen- tary corner isomorphic if there exists a projection P ∈ DZ such that POZ P = OA, DZ P = DA. (7.1) Two Cuntz -- Krieger pairs (OA,DA) and (OB,DB) are said to be corner isomorphic if there exists a finite chain of Cuntz -- Krieger pairs (OZi,DZi), i = 0, 1, . . . , n such that Z0 = A, Zn = B, and either (OZi,DZi) and (OZi+1,DZi+1) or (OZi+1,DZi+1) and (OZi,DZi) are elementary corner isomorphic for all i = 0, 1, . . . , n. That is, the equivalence relation generated by elementary corner isomorphisms in Cuntz -- Krieger pairs is the corner iso- morphism. We will prove the following theorem. 20 Theorem 7.2. If two Cuntz -- Krieger pairs (OA,DA) and (OB,DB) are corner isomorphic, then there exists an isomorphism Φ : OA ⊗ K −→ OB ⊗ K of C ∗-algebras such that Φ(DA ⊗ C) = DB ⊗ C. Proof. We use the notation Z of matrix instead of B, so that we suppose that (OA,DA) and (OZ ,DZ ) are elementary corner isomorphic by a projection P ∈ DZ satisfying (7.1). Although by showing that X = POZ is an (OA,DA) -- (OZ ,DZ )-relative imprimitivity bimodule, we know that (OA,DA) and (OZ ,DZ ) are relatively Morita equivalent, and hence there exists an isomorphism Φ : OA ⊗ K −→ OB ⊗ K of C ∗-algebras such that Φ(DA ⊗ C) = DB ⊗ C, we will directly construct such a isomorphism Φ in the following way. We may assume that the projection Q = 1 − P is not zero. Let S1, . . . , SNZ be the canonical generating partial isometries of the Cuntz -- Krieger algebra OZ . As Q ∈ DZ , one may find a finite family of admissible words µ(k) ∈ B∗(XZ ), k = 1, . . . , N1 such that µ(1) = ··· = µ(N1) and Q = PN1 µ(k). Since Z is irreducible, we may find admissible words ν(k) ∈ B∗(XZ ) for each µ(k) such that ν(1) = ··· = ν(N1) and k=1 Sµ(k)S∗ Sν(k)Sµ(k) 6= 0, As ν(k)µ(k) is an admissible word in XZ , we know S∗ P ≥ Sν(k)S∗ ν(k), k = 1, . . . , N1. ν(k)Sν(k) ≥ Sµ(k)S∗ µ(k). Put Uk = Sν(k)Sµ(k)S∗ µ(k), k = 1, . . . , N1. Then we have U ∗ k Uk = N1 Xk=1 N1 Xk=1 and Sµ(k)S∗ µ(k)S∗ ν(k)Sν(k)Sµ(k)S∗ µ(k) = UkU ∗ k ≤ Sν(k)S∗ ν(k) ≤ P. The sequence further satisfies the following l = 0 UkU ∗ UkDZ U ∗ As in the proof of Theorem 4.1, by setting for k 6= l, k ⊂ DZ P = DA, Sµ(k)S∗ µ(k) = Q, N1 Xk=1 U ∗ kDZ Uk ⊂ DZ Q ⊂ DZ . un = N1 Xk=1 Uk ⊗ snk,n, wn = P ⊗ sn0,n + un, n = 1, 2, . . . . we have a sequence wn, n ∈ N of partial isometries in M (OZ ⊗ K) such that (1) w∗ nwn = 1 ⊗ fn. (2) wnw∗ n ≤ P ⊗ fn. (3) wn(DZ ⊗ C)w∗ n ⊂ DZ P ⊗ C. 21 (4) w∗ n(DZ ⊗ C)wn ⊂ DZ Q ⊗ C. Let fn,m be a partial isometry satisfying f ∗ n,mfn,m = fm, fn,mf ∗ n,m = fn. We put v1 = w1 = P ⊗ s10,1 + u1, v2n = (P ⊗ fn − v2n−1v∗ v2n−1 = wn(1 ⊗ fn − v∗ 2n−2v2n−2) 2n−1)(p ⊗ fn,n+1) for 1 ≤ n ∈ N, for 2 ≤ n ∈ N. By the same way as Lemma 4.5, we have for n ∈ N (1) v∗ 2n−2v2n−2 + v∗ 2n−1v2n−1 = 1 ⊗ fn. Hence the summationP∞ VA in the strict topology of M (OZ ⊗ K), so that (1) V ∗ n=1 vn converges in M (OZ⊗K) to certain partial isometry written 2n = P ⊗ fn. 2n−1 + v2nv∗ (2) v2n−1v∗ (3) vn(DZ ⊗ C)v∗ (4) v∗ n ⊂ DZ P ⊗ C. n(DZ P ⊗ C)vn ⊂ DZ ⊗ C. AVA = 1 ⊗ 1. (2) VAV ∗ A = P ⊗ 1. (3) VA(DZ ⊗ C)V ∗ (4) V ∗ A = DZ P ⊗ C. A(DZ P ⊗ C)VA = DZ ⊗ C. Putting ΦA = Ad(VA) : OZ ⊗K −→ OA ⊗K which is an isomorphism between OZ ⊗K and OA ⊗ K such that Φ(DZ ⊗ C) = DA ⊗ C. Therefore we have the following theorem. Theorem 7.3. The Cuntz-Krieger pairs (OA,DA) and (OB,DB) are corner isomorphic if and only if there exists an isomorphism Φ : OA ⊗ K −→ OB ⊗ K of C ∗-algebras such that Φ(DA ⊗ C) = DB ⊗ C. Proof. The only if part follows from Theorem 7.2. We will show the if part. Suppose that there exists an isomorphism Φ : OA ⊗ K −→ OB ⊗ K of C ∗-algebras such that Φ(DA ⊗ C) = DB ⊗ C. By [21], the two-sided topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are flow equivalent, so that the two matrices are connected by a finite chain of strong shift equivalences and symbol expansions. In the proofs of Proposition 6.1 and Proposition 6.2, we know that (OA,DA) and (OB,DB) are corner isomorphic. Therefore we may summarize our discussions in the following way. Theorem 7.4. Let A, B be irreducible non-permutation matrices with entries in {0, 1}. Let OA,OB be the associated Cuntz -- Krieger algebras. Then the following assertions are mutually equivalent: (1) (OA,DA) ∼RME (OB,DB). 22 (2) (OA ⊗ K,DA ⊗ C) ∼RME (3) There exists an isomorphism Φ : OA ⊗ K −→ OB ⊗ K of C ∗-algebras such that (OB ⊗ K,DB ⊗ C). Φ(DA ⊗ C) = DB ⊗ C. (4) (OA,DA) and (OB,DB) are corner isomorphic. (5) The two-sided topological Markov shifts ( ¯XA, ¯σA) and ( ¯XB, ¯σB) are flow equivalent. Proof. (1) ⇐⇒ (2) comes from Lemma 3.9. (1) ⇐⇒ (3) comes from Theorem 4.7. (3) ⇐⇒ (4) comes from Theorem 7.3. (5) =⇒ (1) comes from Theorem 6.3. (3) =⇒ (5) comes from [21, Corollary 3.8]. We note that the implication (5) =⇒ (3) is seen in [7], and that the equivalence between (3) and (5) is seen in [21, Corollary 3.8]. 8 Relative Picard groups Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. Let X, Y be (A1,D1) -- (A2,D2)-relative imprimitivity bimodule. Then X and Y are said to be equivalent if there exists an isomorphism ϕ : X −→ Y of A1 -- A2-imprimitivity bimodule such that hϕ(x1) ϕ(x2)i = hx1 x2i for x1, x2 ∈ X for both left and right inner products. As ϕ : X −→ Y preserves the bimodule structures and inner products of X and Y , we know ϕ(XD) = YD. We denote by [X] the equivalence class of relative imprimitivity bimodule. For a relative σ-unital pair (A,D) of C ∗-algebras, we define a relative version of Picard group as follows. Definition 8.1. The relative Picard group Pic(A,D) for (A,D) is defined by the group of equivalence classes [X] of (A,D) -- (A,D)-relative imprimitivity bimodule by the product [X] · [Y ] := [X ⊗A Y ]. We remark that the identity element of the group Pic(A,D) is the class of the identity (A,D) -- (A,D)-relative imprimitivity bimodule X = A defined by the module structure and the inner products: a · x · b = axb, Ahx yi := xy∗, hx yiA := x∗y for a, b, x, y ∈ A. (8.1) Since (A,D) is relative σ-unital, the above X becomes an (A,D) -- (A,D)-relative imprim- itivity bimodule as seen in Lemma 3.6. Lemma 8.2. If (A1,D1) ∼RME (A2,D2), we have Pic(A1,D1) = Pic(A2,D2). Hence we have Pic(A,D) = Pic(A⊗K,D ⊗C) for every relative σ-unital pair (A,D) of C ∗-algebras. 23 Proof. Let X be (A1,D1) -- (A2,D2)-relative imprimitivity bimodule, and ¯X its conjugate module, which is (A2,D2) -- (A1,D1)-relative imprimitivity bimodule. It is easy to see that the correspondence [Y ] ∈ Pic(A1,D1) −→ [ ¯X ⊗A1 Y ⊗A1 X] ∈ Pic(A2,D2) yields an isomorphism as groups, because [ ¯X ⊗A1 X] is the unit of the group Pic(A2,D2) and [X ⊗A2 ¯X] is the unit of the group Pic(A1,D1). If θ : A1 −→ A2 is an isomorphism of C ∗-algebras such that θ(D1) = D2, then we write θ : (A1,D1) −→ (A2,D2) and call an isomorphism of relative σ-unital pairs of C ∗-algebras. As in Lemma 3.6, any isomorphism θ : (A1,D1) −→ (A2,D2) gives rise to a (A1,D1) -- (A2,D2)-relative imprimitivity bimodule Xθ. The following lemma is straightforward. Lemma 8.3. Let θ12 : (A1,D1) −→ (A2,D2) and θ23 : (A2,D2) −→ (A3,D3) be isomor- phisms of relative σ-unital pairs of C ∗-algebras. Then we have [Xθ12 ⊗A2 Xθ23 ] = [Xθ23◦θ12 ]. Therefore we have a contravariant functor from the category of relative σ-unital C ∗- algebras with isomorphisms θ : (A1,D1) −→ (A2,D2) as morphisms into the category of relative σ-unital C ∗-algebras with equivalence classes of relative imprimitivity bimodules. Let Aut(A,D) be the group of automorphisms θ on A such that θ(D) = D, that is, Aut(A,D) := {θ ∈ Aut(A) θ(D) = D} We denote by U (A,D) the group of unitaries u ∈ M (A) satisfying uDu∗ = D. We denote by Ad(u) the automorphism of (A,D) defined by Ad(u)(a) = uau∗ for a ∈ A. Let us denote by Int(A,D) the subgroup of Aut(A,D) consisting of such automorphisms of (A,D). Hence Int(A,D) is a normal subgroup of Aut(A,D). By the preceding lemma, we have an anti-homomorphism θ ∈ Aut(A,D) −→ [Xθ] ∈ Pic(A,D). The following proposition and its corollary are achieved by a similar manner to Brown -- Green -- Rieffel's argument [3, Proposition 3.1] and [3, Corollary 3.2]. Proposition 8.4 (cf. Aut(A,D) into Pic(A,D) is exactly Int(A,D). That is, we have an exact sequence: [3, Proposition 3.1]). The kernel of the anti-homomorphism from 1 −→ Int(A,D) −→ Aut(A,D) −→ Pic(A,D). Corollary 8.5 (cf. [3, Corollary 3.2]). Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. Let α, β : (A1,D1) −→ (A2,D2) be isomorphisms. If Xα and Xβ are equivalent, then there exists a unitary u ∈ U (A,D) such that β = Ad(u) ◦ α. The following lemma is also a relative version of [3, Lemma 3.3]. 24 Lemma 8.6 (cf. [3, Lemma 3.3]). Let X be an (A1,D1) -- (A2,D2)-relative imprimitivity bimodule. Let (A0,D0) be the linking pair of X defined by (4.1) and (4.2). Then X is equivalent to Xθ for some isomorphism θ : (A1,D1) −→ (A2,D2) if and only if there exists a partial isometry v ∈ M (A0) such that v∗v =(cid:20)1 0 0 0(cid:21) , vv∗ =(cid:20)0 0 0 1(cid:21) (8.2) and vD0v∗ = D0vv∗, v∗D0v = D0v∗v. (8.3) In this case, θ is defined by θ(a) = vav∗, a ∈ A1. Remark 8.7. Under the equality (8.2), the second equality of (8.3) follows from the first equality of (8.3). Because the first one of (8.3) ensures us the equality (8.4) By (8.2), v∗v commutes with any elements of D0 so that (8.4) goes to the second equality of (8.3). v∗vD0v∗v = v∗D0vv∗v. Proof of Lemma 8.6. Although the proof basically follows the proof of [3, Lemma 3.3], we give it for the sake of completeness. Suppose that X is equivalent to Xθ for some isomorphism θ : (A1,D1) −→ (A2,D2). By this isomorphism, the linking algebra A0 of X is identified with that of Xθ. Hence Xθ = A1 and A0 = {(cid:20)a1 ¯y x a2(cid:21) a1 ∈ A1, a2 ∈ A2, x ∈ Xθ, ¯y ∈ ¯Xθ}. v(cid:20) z c2(cid:21) =(cid:20) 0 θ(z)(cid:21) , We define operators v, v∗ on Xθ ⊕ A2 by v∗(cid:20) z c2(cid:21) =(cid:20)θ−1(c2) (cid:21) where Xθ = A1 so that θ(z) ∈ A2. Put a2(cid:21) v, x a2(cid:21)) =(cid:20)a1 Rv((cid:20)a1 Lv((cid:20)a1 ¯y x ¯y ¯y 0 It is straightforward to see that for z ∈ Xθ, c2 ∈ A2 x a2(cid:21)) = v(cid:20)a1 ¯y x a2(cid:21) . Rv((cid:20)a1 ¯y x a2(cid:21))(cid:20)a′ 1 x′ ¯y′ a′ 2(cid:21) =(cid:20)a1 ¯y x a2(cid:21) Lv((cid:20)a′ 1 x′ ¯y′ a′ 2(cid:21)). Hence the pair (Lv, Rv) defines an element of M (A0) as a double centralizer of A0. Simi- larly (Lv∗ , Rv∗ ) defines an element of M (A0) such that (Lv, Rv)∗ = (Lv∗ , Rv∗ ), so that we may write (Lv, Rv) = v. It then follows that v∗v(cid:20) z vv∗(cid:20) z c2(cid:21) =(cid:20)z 0(cid:21) c2(cid:21) c2(cid:21) =(cid:20) 0 v∗v =(cid:20)1 0 0 0(cid:21) , vv∗ =(cid:20)0 0 0 1(cid:21) . and hence and hence 25 It is direct to see that v(cid:20)a1 0 0(cid:21) v∗ =(cid:20)0 0 θ(a1)(cid:21) . 0 0 This means that θ(a1) = va1v∗ for a1 ∈ A1 under the identification between a1 and(cid:20)a1 0 0(cid:21) for a1 ∈ A1. Since θ : A1 −→ A2 satisfies θ(D1) = D2 and D1 = D0v∗v, D2 = D0vv∗, we have 0 vD0v∗ = vD1v∗ = θ(D1) = D2 = D0vv∗ and v∗D0v = D0v∗v. Conversely suppose that a partial isometry v ∈ M (A0) satisfies the equalities (8.2) and (8.3). It is easy to see that there exists an element θ(a) in A2 for each a ∈ A1 such that v(cid:20)a 0 0 0(cid:21) v∗ =(cid:20)0 0 θ(a)(cid:21) 0 and the correspondence a ∈ A1 −→ θ(a) ∈ A2 gives rise to an isomorphism of C ∗-algebras. The conditions (8.2) and (8.3) implies that vD0v∗ = D0vv∗ = D2 and v∗D0v = D0v∗v = D1 so that we have vD1v∗ = vv∗D0vv∗ = D2. This implies that θ(D1) = D2. We will next show that X is equivalent to Xθ. We identify A1 with its image in A0 and then we will define a map η : X −→ A1(= Xθ) by Since for x ∈ X. η(x) :=(cid:20)0 x 0 0(cid:21) v 0 0(cid:21)(cid:20)0 x 0 0(cid:21) vv∗v =(cid:20)0 x v∗vη(x)v∗v =(cid:20)1 0 0 0(cid:21) v = η(x), we see that η(x) ∈ A1. By a routine calculation, we know that η is a bimodule homomor- phism from X to Xθ which preserves both inner products, and hence η gives rise to an isomorphism between X and Xθ. The following theorem is also a relative version of a Brown-Green-Rieffel' s theorem We will give its proof for the sake of completeness. Theorem 8.8 (cf. [3, Theorem 3.4]). Let (A1,D1) and (A2,D2) be relative σ-unital pairs of C ∗-algebras. Let X be an (A1 ⊗ K,D1 ⊗ C) -- (A2 ⊗ K,D2 ⊗ C)-relative imprimitivity bimodule. Then there exists an isomorphism θ : A1 ⊗K −→ A2 ⊗K satisfying θ(D1 ⊗C) = D2⊗C such that X is equivalent to Xθ. Furthermore θ is unique up to left multiplication by an element of Int(A2 ⊗ K,D2 ⊗ C), that is if X is equivalent to Xϕ for some isomorphism ϕ : (A1⊗K,D1⊗C) −→ (A2⊗K,D2⊗C), then there exists a unitary u ∈ U (A2⊗K,D2⊗C) such that ϕ = Ad(u) ◦ θ. Proof. The uniqueness follows immediately from Corollary 8.5. Now let X be an (A1⊗K,D1⊗C) -- (A2⊗K,D2⊗C)-relative imprimitivity bimodule. We put ¯Ai = Ai ⊗ K, ¯Di = Di ⊗ K for i = 1, 2. Let ( ¯A0, ¯D0) be the linking pair for X defined from ¯Ai, ¯Di, i = 1, 2 and X by (4.1) and (4.2). By the assumption that ( ¯A1, ¯D1) ∼RME 26 ( ¯A2, ¯D2) with Theorem 4.7, Proposition 4.6 tells us that there exist vi ∈ M ( ¯A⊗K), i = 1, 2 such that in M ( ¯A0 ⊗ K), v∗ i vi = 1 ⊗ 1 1 = P1 ⊗ 1 where P1 =(cid:20)1 0 0 0(cid:21) v1v∗ 2 = P2 ⊗ 1 where P2 =(cid:20)0 0 0 1(cid:21) v2v∗ i = 1, 2, in M ( ¯A0) in M ( ¯A0) and Put a partial isometry w = v2v∗ i = ¯Di ⊗ C, vi( ¯D0 ⊗ C)v∗ i ( ¯Di ⊗ C)vi = ¯D0 ⊗ C, v∗ 1 ∈ M ( ¯A ⊗ K) so that we have 0(cid:21) , w∗w = P1 ⊗ 1 =(cid:20)1 ⊗ 1 0 0 ww∗ = P2 ⊗ 1 =(cid:20)0 0 1 ⊗ 1(cid:21) 0 i = 1, 2. in M ( ¯A0 ⊗ K). and w( ¯D1 ⊗ C)w∗ = ¯D2 ⊗ C, w∗( ¯D2 ⊗ C)w = ¯D1 ⊗ C. Let p ∈ C be the rank one projection p = e1,1, so that ¯A1⊗ p⊗K is a corner of ¯A1⊗K⊗K. Hence by [2, Lemma 2.5] there exists a partial isometry ¯v1 ∈ M ( ¯A1 ⊗ K ⊗ K) such that ¯v∗ 1 ¯v1 = 1 ⊗ 1 ⊗ 1, ¯v1¯v∗ 1 = 1 ⊗ p ⊗ 1. By the construction of ¯v1, we see that ¯v1( ¯D1 ⊗ C ⊗ C)¯v∗ 1 = ¯D1 ⊗ p ⊗ C, 1( ¯D1 ⊗ p ⊗ C)¯v1 = ¯D1 ⊗ C ⊗ C. ¯v∗ We can identify ¯A1 and ¯D1 with ¯A1 ⊗ 1⊗ K and ¯D1 ⊗ 1 ⊗ C, respectively, so that we have ¯v1 ∈ M ( ¯A1 ⊗ K) and ¯v∗ 1 ¯v1 = 1 ⊗ 1, 1 = ¯D1 ⊗ p, ¯v1( ¯D1 ⊗ C)¯v∗ ¯v1¯v∗ 1 = 1 ⊗ p, 1( ¯D1 ⊗ p)¯v1 = ¯D1 ⊗ C. ¯v∗ Similarly we have ¯v2 ∈ M ( ¯A2 ⊗ K) and ¯v∗ 2 ¯v2 = 1 ⊗ 1, 2 = ¯D2 ⊗ p, ¯v2( ¯D2 ⊗ C)¯v∗ ¯v2¯v∗ 2 = 1 ⊗ p, 2( ¯D2 ⊗ p)¯v2 = ¯D2 ⊗ C. ¯v∗ Define ¯v ∈ M ( ¯A0 ⊗ K) by 0 ¯v =(cid:20)0 0 ¯v2(cid:21) w(cid:20)¯v∗ 1 0 0 0(cid:21) in M ( ¯A0 ⊗ K). 27 We then have and 0 0 1 ⊗ 1(cid:21) w(cid:20)¯v∗ 1 0 0 0(cid:21) 0 ¯v∗¯v =(cid:20)¯v1 0 0(cid:21) w∗(cid:20)0 =(cid:20)¯v1 0 0(cid:21) w∗w(cid:20)¯v∗ =(cid:20)1 ⊗ p 0 0(cid:21) 0 0 1 0 0 0(cid:21) 0 0 ¯v∗ 2(cid:21) 0 0 ¯v¯v∗ =(cid:20)0 0 ¯v2(cid:21) w(cid:20)1 ⊗ 1 0 0(cid:21) w∗(cid:20)0 =(cid:20)0 0 ¯v2(cid:21) ww∗(cid:20)0 2(cid:21) =(cid:20)0 0 1 ⊗ p(cid:21) . 0 0 ¯v∗ 0 0 We will next show that ¯v( ¯D0⊗p)¯v∗ = ( ¯D0⊗p)¯v¯v∗. For(cid:20)d1 0 we have ¯v(cid:20)d1 ⊗ p 0 0 d2 ⊗ p(cid:21) ¯v∗ =(cid:20)0 0 ¯v2(cid:21) w(cid:20)¯v∗ 0 1(d1 ⊗ p)¯v1 0 0 0 d2(cid:21) ∈ ¯D0 with di ∈ ¯Di, i = 1, 2, 0(cid:21) w∗(cid:20)0 0 0 ¯v∗ 2(cid:21) . 1(d1 ⊗ p)¯v1 ∈ ¯D1 ⊗ C, we have w(cid:20)¯v∗ 1(d1 ⊗ p)¯v1 0 0 0(cid:21) w∗ ∈ w( ¯D1 ⊗ C)w∗ = ¯D2 ⊗ C so Since ¯v∗ that ¯v(cid:20)d1 ⊗ p 0 0 d2 ⊗ p(cid:21) ¯v∗ ∈ ¯v2( ¯D2 ⊗ C)¯v2 = ¯D2 ⊗ p = ( ¯D0 ⊗ p)¯v¯v∗. Therefore we have ¯v( ¯D0 ⊗ p)¯v∗ ⊂ ( ¯D0 ⊗ p)¯v¯v∗ and similarly ¯v∗( ¯D0 ⊗ p)¯v ⊂ ( ¯D0 ⊗ p)¯v∗¯v so that we have By the equalities and ¯v( ¯D0 ⊗ p)¯v∗ = ( ¯D0 ⊗ p)¯v¯v∗ ¯v∗¯v =(cid:20)1 ⊗ p 0 0(cid:21) , 0 ¯v∗( ¯D0 ⊗ p)¯v = ( ¯D0 ⊗ p)¯v∗¯v. ¯v¯v∗ =(cid:20)0 0 0 1 ⊗ p(cid:21) , we know that ¯v commutes with 1⊗p so that we can regard ¯v as an element of M ( ¯A0⊗p) = M ( ¯A0). Thus we obtain a partial isometry ¯v in M ( ¯A0) such that 0 1 ¯A2(cid:21) , ¯v∗¯v =(cid:20)1 ¯A1 ¯v¯v∗ =(cid:20)0 0(cid:21) , 0 0 0 and ¯v ¯D0¯v∗ = ¯D0¯v¯v∗ and ¯v∗ ¯D0¯v = ¯D0¯v∗¯v. Therefore by Lemma 8.6, we conclude that X is equivalent to Xθ for some isomorphism θ : ( ¯A1, ¯D1) −→ ( ¯A2, ¯D2). 28 Recall that the subgroups Aut(A ⊗ K,D ⊗ C) and Int(A ⊗ K,D ⊗ C) of automorphism group Aut(A ⊗ K) are defined by Aut(A ⊗ K,D ⊗ C) = {β ∈ Aut(A ⊗ K) β(D ⊗ C) = D ⊗ C}, Int(A ⊗ K,D ⊗ C) = {β ∈ Int(A ⊗ K) β(D ⊗ C) = D ⊗ C}. Corollary 8.9. Let (A,D) be a relative σ-unital pair of C ∗-algebras. For any relative imprimitivity bimodule [X] ∈ Pic(A⊗K,D⊗C), there exists an automorphism θ ∈ Aut(A⊗ K,D ⊗ C) such that [X] = [Xθ]. Thus we have a exact sequence 1 −→ Int(A ⊗ K,D ⊗ C) −→ Aut(A ⊗ K,D ⊗ C) −→ Pic(A ⊗ K,D ⊗ C) −→ 1. Let us denote by Out(A ⊗ K,D ⊗ C) the quotient group Aut(A ⊗ K,D ⊗ C)/ Int(A ⊗ K,D ⊗ C). We then have Corollary 8.10. Let (A,D) be a relative σ-unital pair of C ∗-algebras. We have Pic(A,D) = Out(A ⊗ K,D ⊗ C). Proof. By Lemma 8.2, we see that Pic(A,D) = Pic(A ⊗ K,D ⊗ C) so that we have the desired equality by the preceding corollary. 9 Relative Picard groups of Cuntz -- Krieger pairs In this section, we will study the relative Picard group Pic(A,D) for the Cuntz -- Krieger pairs (OA,DA). By [13, Lemma 1.1], for a unitary u ∈ M (OA ⊗ K), the automorphism Ad(u) acts trivially on K0(OA ⊗ K). We will first show the following proposition which is a relative version of [13, Lemma 3.13] (Lemma 10.1 in Appendix). Proposition 9.1. Let β ∈ Aut(OA ⊗ K) satisfy β(DA ⊗ C) = DA ⊗ C and β∗ = id on K0(OA). Then there exists a unitary u ∈ M (OA ⊗K) and an automorphism α ∈ Aut(OA) such that β = Ad(u) ◦ (α ⊗ id) and u(DA ⊗ C)u∗ = DA ⊗ C, α∗ = id on K0(OA), α(DA) = DA. To show the above proposition, we provide several lemmas. Lemma 9.2. Let β ∈ Aut(OA ⊗ K) satisfy β(DA ⊗ C) = DA ⊗ C and β∗ = id on K0(OA). Then for each k ∈ N, there exists a partial isometry wk ∈ OA ⊗ K such that wk(DA ⊗ C)w∗ w∗ kwk = 1 ⊗ ekk, k ⊂ DA ⊗ C, wkw∗ k = β(1 ⊗ ekk), w∗ k(DA ⊗ C)wk ⊂ DA ⊗ C. Proof. Let us denote by Ns(OA ⊗ K,DA ⊗ C) the normalizer semigroup {v ∈ OA ⊗ K v is a partial isometry; v(DA ⊗ C)v∗ ⊂ DA ⊗ C, v∗(DA ⊗ C)v ⊂ DA ⊗ C} of partial isometries in OA ⊗K. Denote by K0(OA⊗K,DA ⊗C) the Murray-von-Neumann equivalence classes of projections in DA ⊗ C by partial isometries in Ns(OA ⊗ K,DA ⊗ C). (9.1) (9.2) 29 It has been proved in [18] that there exists a natural isomorphism between K0(OA) and K0(OA⊗K,DA⊗C). Since [β(1⊗ekk)] = β∗([1⊗ekk]) = [1⊗ekk], we have β(1⊗ekk) ∼ 1⊗ekk in K0(OA⊗K,DA⊗C). We may find a partial isometry wk ∈ OA⊗K satisfying the desired conditions. Lemma 9.3. Let β ∈ Aut(OA ⊗ K) satisfy β(DA ⊗ C) = DA ⊗ C and β∗ = id on K0(OA). Then there exists a unitary w ∈ M (OA ⊗ K) such that (Ad(w∗) ◦ β)(1 ⊗ ekk) = 1 ⊗ ekk, (Ad(w∗) ◦ β)(OA ⊗ ekk) = OA ⊗ ekk, Ad(w∗) ◦ β(DA ⊗ ekk) = DA ⊗ ekk, for all k ∈ N. Proof. Take wk ∈ OA ⊗ K be a partial isometry for each k ∈ N satisfying the conditions of Lemma 9.2. k=1 wk converges in the strict topology of M (OA ⊗K). By (9.1) and (9.2), we have w∗w = ww∗ = 1 and w(DA ⊗C)w∗ = w∗(DA ⊗ C)w = DA ⊗ C. We then see that w(1 ⊗ ekk)w∗ = ww∗ It is easy to see that the summation P∞ k = β(1 ⊗ ekk) kwkw∗ = wkw∗ so that (Ad(w∗) ◦ β)(1 ⊗ ekk) = 1 ⊗ ekk. For x ∈ OA, we have (Ad(w∗) ◦ β)(x ⊗ ekk) =(Ad(w∗) ◦ β)((1 ⊗ ekk)(x ⊗ ekk)(1 ⊗ ekk)) =(1 ⊗ ekk)(Ad(w∗) ◦ β)((x ⊗ ekk))(1 ⊗ ekk) so that (Ad(w∗) ◦ β)(OA ⊗ ekk) = OA ⊗ ekk. As β(DA ⊗ C) = DA ⊗ C and w∗(DA ⊗ C)w = DA ⊗ C, we have Ad(w∗) ◦ β(DA ⊗ ekk) = DA ⊗ ekk. Proof of Proposition 9.1. Suppose that β ∈ Aut(OA ⊗K) satisfies β(DA ⊗C) = DA ⊗C and β∗ = id on K0(OA). Take a unitary w ∈ M (OA ⊗ K) satisfying the conditions of Lemma 9.3. Put βw = Ad(w∗)◦β ∈ Aut(OA⊗K). Since (Ad(w∗)◦β)(OA⊗ekk) = OA⊗ekk, we may find an automorphism αk ∈ Aut(OA) for k ∈ N such that for x ∈ OA. αk(x) ⊗ ekk = βw(x ⊗ ekk) By replacing β with βw, we may assume that βw(x ⊗ ekk) = αk(x) ⊗ ekk. For j, k ∈ N, we have β(x ⊗ ejk) = β((1 ⊗ ejk)(x ⊗ ejk)) = β(1 ⊗ ejk) · (αk(x) ⊗ ejk). By putting x = 1, we see that β(1 ⊗ ejk) = (1 ⊗ ejj)β(1 ⊗ ejk)(1 ⊗ ekk) so that there exists wjk ∈ OA such that w∗ jk = wkj and β(1 ⊗ ejk) = wjk ⊗ ejk. Since jkwjk ⊗ ekk = β(1 ⊗ ejk)∗β(1 ⊗ ejk) = β(1 ⊗ ekk) = 1 ⊗ ekk w∗ so that w∗ jkwjk = 1 and similarly wjkw∗ jk = 1. We also have for a ∈ DA w∗ jkawjk ⊗ ejj = β((1 ⊗ ejk)(a ⊗ ekk)(1 ⊗ ekj)) = β(a ⊗ ejj) = αj(a) ⊗ ejj 30 so that wjkDAw∗ jk = DA. Since β(x ⊗ ejk) = β(1 ⊗ ejk) · (αk(x) ⊗ ekk) = wjkαk(x) ⊗ ejk and similarly β(x ⊗ ejk) = αj(x)wjk ⊗ ejk, we see wjkαk(x) ⊗ ejk = αj(x)wjk ⊗ ejk and jkαj(x)wjk for x ∈ OA. Put u =P∞ hence αk(x) = w∗ k=1 w1k ⊗ ekk which is easily proved to be a unitary in M (OA ⊗ K). It then follows that β(x ⊗ ejk) =β((1 ⊗ ej1)(x ⊗ e11)(1 ⊗ e1k)) =(wj1 ⊗ ej1)(α1(x) ⊗ e11)(w1k ⊗ e1k) =(wj1α1(x)w1k) ⊗ ejk =u∗(α1(x) ⊗ ejk)u for x ∈ OA so that β = Ad(u∗) ◦ (α1 ⊗ id). Since w1kDAw∗ 1k = DA, we have u(DA ⊗ C)u∗ = DA ⊗ C. By [13, Lemma 1.1], we know that Ad(u)∗ = id on K0(OA) so that α0∗ = (β−1)∗ = id on K0(OA). We thus have the following theorem. Theorem 9.4. Let β ∈ Aut(OA ⊗ K). Then β satsifies the following condition β(DA ⊗ C) = DA ⊗ C and β∗ = id on K0(OA) (9.3) if and only if there exists an automorphism α ∈ Aut(OA) and a unitary u ∈ M (OA ⊗ K) such that β = Ad(u) ◦ (α ⊗ id) and u(DA ⊗ C)u∗ = DA ⊗ C, α∗ = id on K0(OA), α(DA) = DA. (9.4) (9.5) The following proposition shows that the expression β in the form (9.4) and (9.5) is unique up to inner automorphisms on OA invariant globally DA. Proposition 9.5. Supose that β ∈ Aut(OA ⊗ K,DA ⊗ C) is of the form β = Ad(u) ◦ (α ⊗ id) = Ad(u′) ◦ (α′ ⊗ id) for some automorphisms α, α′ ∈ Aut(OA,DA) and unitaries u, u′ ∈ M (OA ⊗ K) satisfying both the conditions (9.4) and (9.5). Then there exists a unitary V ∈ OA such that u = u′(V ⊗ 1), α = Ad(V ∗) ◦ α′ (9.6) Proof. For x ⊗ K ∈ OA ⊗ K, we have u(α(x) ⊗ K)u∗ = u′(α′(x) ⊗ K)u′∗. Put v = u′∗u ∈ M (OA ⊗ K) which is a unitary satisfying v(α(x) ⊗ K) = (α′(x) ⊗ K)v. We in particularly see that v(1 ⊗ ejk) = (1 ⊗ ejk)v for all j, k ∈ Z+. Define V ∈ OA by setting V ⊗ e11 = (1 ⊗ e11)v(1 ⊗ e11). As v commutes with 1 ⊗ e11, we know that V is a unitary in OA. We then have and V DAV ∗ = DA. u′∗u(1 ⊗ ekk) = v(1 ⊗ ek1)(1 ⊗ e1k) = (1 ⊗ ek1)v(1 ⊗ e1k) = (1 ⊗ ek1)(V ⊗ e11)(1 ⊗ e1k) = (V ⊗ 1)(1 ⊗ ek1)(1 ⊗ e11)(1 ⊗ e1k) = (V ⊗ 1)(1 ⊗ ekk) 31 for all k ∈ Z+. Hence we have u′∗u = V ⊗ 1. As we have for x ∈ OA α(x) ⊗ e11 = v∗(α′(x) ⊗ e11)v = v∗(1 ⊗ e11)(α′(x) ⊗ e11)(1 ⊗ e11)v = (V ∗ ⊗ e11)(α′(x) ⊗ e11)(V ⊗ e11) = V ∗α′(x)V ⊗ e11 so that α(x) = V ∗α′(x)V for x ∈ OA. As α(DA) = α′(DA) = DA, we have V DAV ∗ = DA. Corollary 9.6. Let β ∈ Aut(OA ⊗ K). Let us denote by 1A the unit of the C ∗-algebra OA. Then β satsifies the following condition β(DA ⊗ C) = DA ⊗ C and β∗([1A ⊗ e11]) = [1A ⊗ e11] on K0(OA ⊗ K) if and only if there exists an automorphism α ∈ Aut(OA) and a unitary u ∈ M (OA ⊗ K) such that β = Ad(u) ◦ (α ⊗ id) and u(DA ⊗ C)u∗ = DA ⊗ C, α∗ = β∗ on K0(OA), α(DA) = DA. Proof. The if part is clear. It suffices to show the only if part. Suppose that β ∈ Aut(OA⊗ K) satsifies the following conditions and β∗([1A ⊗ e11]) = [1A ⊗ e11] on K0(OA ⊗ K). β(DA ⊗ C) = DA ⊗ C Since β ∈ Aut(OA ⊗ K) satisfies β∗([1A ⊗ e11]) = [1A ⊗ e11], By [30], there exists an automorphism α◦ of OA such that α◦∗ = β∗ on K0(OA). Hence by using [18, Proposition 5.1], we may find an automorphism α1 of OA such that α1(DA) = DA and α1∗ = α◦∗ on K0(OA). Put β1 := β ◦ (α−1 1 ⊗ id) ∈ Aut(OA ⊗ K). We have β1(DA ⊗ C) = DA ⊗ C and β1∗ = β∗ ◦ α−1 1∗ = id on K0(OA). By Theorem 9.4, one may take an automorphism α2 ∈ Aut(OA) and a unitary u ∈ M (OA ⊗ K) such that β1 = Ad(u) ◦ (α2 ⊗ id) and u(DA ⊗ C)u∗ = DA ⊗ C, α2∗ = id on K0(OA), α2(DA) = DA. Put α := α2 ◦ α1 ∈ Aut(OA). We then have β = Ad(u) ◦ (α ⊗ id), α∗ = β∗ on K0(OA), α(DA) = DA. Let Aut◦(OA,DA) = {α ∈ Aut(OA) α(DA) = DA, α∗ = id on K0(OA)}. Since Int(OA,DA) is a subgroup of Aut◦(OA,DA), we may consider the quotient group Aut◦(OA,DA)/ Int(OA,DA) which we denote by Out◦(OA,DA). Thanks to Theorem 9.4 32 and Corollary 9.6, we know the following theorem on the relative Picard group Pic(OA,DA), which are relative versions of the results shown in Appendix after this section. Let Ψ : Aut(OA,DA) −→ Aut(OA ⊗ K,DA ⊗ C) be the homomorphism defined by Ψ(α) = α⊗ id. Since Ψ(Int(OA,DA)) ⊂ Int(OA ⊗K,DA ⊗C), it induces a homomorphism from Out(OA,DA) to Out(OA ⊗ K,DA ⊗ C) written ¯Ψ. The following is a corollary of Proposition 9.5. Corollary 9.7. The homomorphism ¯Ψ : Out(OA,DA) −→ Out(OA ⊗K,DA ⊗C) is injec- tive. Proof. Suppose that α ∈ Aut(OA,DA) satisfies α ⊗ id = Ad(u′) for some u′ ∈ U (OA ⊗ K,DA ⊗ C). Put α′ = id and u = 1 in the statement of Proposition 9.5 to have a unitary V ∈ U (OA,DA) such that u′ = V ⊗ 1 and α = Ad(V ). By [13, Lemma 1.1], we may define a homomorphism K∗ : Out(OA ⊗ K,DA ⊗ C) −→ Aut(K0(OA ⊗ K)) by setting K∗([α]) = α∗ for [α] ∈ Out(OA ⊗ K,DA ⊗ C). Theorem 9.8. Let A be an irreducible non-permutation matrix. Then the following short exact sequence holds: 1 −→ Out◦(OA,DA) ¯Ψ −→ Out(OA ⊗ K,DA ⊗ C) K∗−→ Aut(K0(OA ⊗ K)) −→ 1. (9.7) Hence there exists a short exact sequence: 1 −→ Out◦(OA,DA) ¯Ψ −→ Pic(OA,DA) K∗−→ Aut(ZN /(id − At)ZN −→ 1. (9.8) Proof. We will show the exactness of (9.7). The injectivity of the homomorphism ¯Ψ : Out◦(OA,DA) −→ Out(OA ⊗ K,DA ⊗ C) follows from Corollary 9.7. The inclusion re- lation ¯Ψ(Out◦(OA,DA)) ⊂ Ker(K∗) is clear. Conversely for any [β] ∈ Ker(K∗), we know β ∈ Aut(OA ⊗ K) satisfy β∗ = id on K0(OA). By Theorem 9.4, there exist a unitary u ∈ M (OA ⊗ K) and an automorphism α ∈ Aut(OA,DA) such that β = Ad(u) ◦ (α ⊗ id) and α∗ = id on K0(OA). Hence we have [β] = [α ⊗ id] = ¯Ψ([α]) and [α] ∈ Aut◦(OA,DA)/ Int(OA,DA), so that we have ¯Ψ(Out◦(OA,DA)) = Ker(K∗) For any ξ ∈ Aut(K0(OA ⊗ K)), ξ gives rise to an automorphism of the abelian group ZN /(id − At)ZN . The group ZN /(id − At)ZN is isomorphic to the Bowen -- Franks group BF (A) = ZN /(id − A)ZN of the matrix A. By Huang's theorem [9, Theorem 2.15] and its proof, any automorphism of the Bowen -- Franks group BF (A) comes from an flow equivalence of the underlying topological Markov shifts ( ¯XA, ¯σA). It implies that there exists an automorphism ψ ∈ Aut(OA ⊗ K) such that ψ(DA ⊗ C) = DA ⊗ C and ψ∗ = ξ on K0(OA). Hence ψ belongs to Aut(OA ⊗ K,DA ⊗ C) such that K∗(ψ) = ξ. Consequently the sequence (9.7) is exact. Let Aut1(ZN /(id − At)ZN ) be a subgroup of Aut(ZN /(id − At)ZN ) defined by Aut1(ZN /(id − At)ZN ) = {ξ ∈ Aut(ZN /(id − At)ZN ) ξ([1]) = [1]} where [1] ∈ ZN /(id − At)ZN denotes the class of the vector (1, . . . , 1) in ZN . 33 Theorem 9.9. Let A be an irreducible non-permutation matrix. Then there exists a short exact sequence: 1 −→ Out(OA,DA) K∗−→ Aut(ZN /(id − At)ZN )/Aut1(ZN /(id − At)ZN ) −→ 1. ¯Ψ−→ Pic(OA,DA) Proof. It suffices to show the exactness at the middle. The inclusion relation ¯Ψ(Out(OA)) ⊂ Ker(K∗) is clear. Conversely, by Rørdam's result [30] again, for any ξ ∈ Aut(K0(OA ⊗K)) with ξ([1]) = [1], there exists an automorphism α◦ of OA such that α◦∗ = ξ on K0(OA). By [18, Proposition 5.1], we may find an automorphism α1 of OA such that α1(DA) = DA and α1∗ = α◦∗ on K0(OA). Hence α1 ∈ Aut(OA,DA) such that ¯Ψ([α1]) = ξ so that ¯Ψ(Out(OA)) = Ker(K∗), and the sequence is exact. 10 Appendix: Picard groups of Cuntz -- Krieger algebras In this appendix, we will refer to the Picard groups of Cuntz -- Krieger algebras and es- pecially Cuntz algebras. As examples of the Picard groups for some interesting class of C ∗-algebras, K. Kodaka has studied the Picard groups for irrational rotation C ∗-algebras Pic(Aθ) to show that Pic(Aθ) is isomorphic to Out(Aθ) if θ is not quadratic, and a semidi- rect product Out(Aθ) ⋊ Z if θ is quadratic ([12], [13]). He also studied the Picard group of certain Cuntz algebras in [13]. He proved that Pic(ON ) = Out(ON ) for N = 2, 3. He also showed that there exists a short exact sequence: 1 −→ Out(ON ) ¯Ψ −→ Pic(ON ) K∗−→ Aut(Z/(1 − N )Z) −→ 1 (10.1) for N = 4, 6. Since Aut(Z/(1 − N )Z) is trivial for N = 2, 3, the Kodaka's results say that the exact sequence (10.1) holds for N = 2, 3, 4, 6. We will show that the above exact sequence holds for all 1 < N ∈ N (Theorem 10.4). As a corollary we know that the Picard group Pic(ON ) of the Cuntz algebra ON is a semidirect product Out(ON ) ⋊ Z/(N − 2)Z if N − 1 is a prime number. We first refer to the Picard groups of Cuntz -- Krieger algebras. Let u ∈ M (A) be a unitary in the multiplier C ∗-algebra M (A) of a C ∗-algebra A. The automorphism Ad(u) on A acts trivially on its K-group K0(A) by [13, Lemma 1.1]. Lemma 10.1 (Kodaka [13, Lemma 1.3]). Let β ∈ Aut(OA⊗K) satisfy β∗ = id on K0(OA). Then there exists a unitary u ∈ M (OA ⊗K) and an automorphism α ∈ Aut(OA) such that β = Ad(u) ◦ (α ⊗ id) and α∗ = id on K0(OA). For a C ∗-algebra A, we put Aut◦(A) = {α ∈ Aut(A) α∗ = id on K0(A)} which is a subgroup of Aut(A). Since Ad(u)∗ = id on K0(A) for a unitary u ∈ M (A), we see that Int(A) is a subgroup of Aut◦(A). The quotient group Aut◦(A)/ Int(A) is denoted by Out◦(A). 34 Let A be a unital C ∗-algebra. Let Ψ : Aut(A) −→ Aut(A ⊗ K) be the homomorphism defined by Ψ(α) = α ⊗ id for α ∈ Aut(A). It induces a homomorphism ¯Ψ : Out(A) −→ Out(A⊗K). If ¯Ψ([α]) = id for some α ∈ Aut(A), we have Ψ(α) = Ad(W ) for some unitary W ∈ M (A ⊗ K). Hence we see that α(x) ⊗ K = W (x ⊗ K)W ∗ for all x ∈ A, K ∈ K. (10.2) Since 1 ⊗ e11 = α(1) ⊗ e11 = W (1 ⊗ e11)W ∗. the unitary W commutes 1⊗ e11 so that there exists a unitary w ∈ A such that w ⊗ e11 = (1 ⊗ e11)W (1 ⊗ e11). We then have α(x) ⊗ e11 = (1 ⊗ e11)W (x ⊗ e11)(1 ⊗ e11) = wxw∗ ⊗ e11 for all x ∈ A. Hence α = Ad(w) ∈ Int(A). This means that the map ¯Ψ : Out(A) −→ Out(A ⊗ K) is injective. Any automorphism β ∈ Aut(A⊗K) induces an automorphism β∗ of K0(A⊗K), which we denote by K∗(β) ∈ Aut(A ⊗ K). By [3, Theorem 1.2] with [3, Corollary 3.5], we know Pic(A) = Pic(A ⊗ K) = Out(A ⊗ K). Proposition 10.2. Let A be an irreducible non-permutation matrix. Then the following short exact sequence holds: 1 −→ Out◦(A) ¯Ψ −→ Out(OA ⊗ K) K∗−→ Aut(K0(OA ⊗ K)) −→ 1. Hence there exists a short exact sequence: 1 −→ Out◦(OA) ¯Ψ −→ Pic(OA) K∗−→ Aut(ZN /(id − At)ZN ) −→ 1. (10.3) (10.4) Proof. We will show the exactness of (10.3). We have already known that the injectivity of ¯Ψ : Out◦(OA) −→ Out(OA ⊗ K). By definition of the group Aut◦(OA), the inclusion relation ¯Ψ(Out◦(OA)) ⊂ Ker(K∗) is clear. Conversely for any [β] ∈ Ker(K∗), we know that β ∈ Aut(OA ⊗ K) satisfy β∗ = id on K0(OA). By Lemma 10.1, there exists a unitary u ∈ M (OA ⊗ K) and an automorphism α ∈ Aut(OA) such that β = Ad(u) ◦ (α ⊗ id) and α∗ = id on K0(OA). Hence we have [β] = [α ⊗ id] = ¯Ψ([α]) and [α] ∈ Aut◦(OA)/ Int(OA). Therefore we have ¯Ψ(Out◦(OA)) = Ker(K∗) By Rørdam's result [30], for any ξ ∈ Aut(K0(OA ⊗ K)), there exists an automorphism β of OA ⊗ K such that β∗ = ξ. Therefore the map K∗ is surjective to prove the exactness of the sequence (10.3). Let Aut1(ZN /(id − At)ZN ) be a subgroup of Aut(ZN /(id − At)ZN ) defined by Aut1(ZN /(id − At)ZN ) = {ξ ∈ Aut(ZN /(id − At)ZN ) ξ([1]) = [1]} where [1] ∈ ZN /(id − At)ZN denotes the class of the vector (1, . . . , 1) in ZN . 35 Proposition 10.3. Let A be an irreducible non-permutation matrix. Then there exists a short exact sequence: ¯Ψ −→ Pic(OA) 1 −→ Out(OA) K∗−→ Aut(ZN /(id − At)ZN )/Aut1(ZN /(id − At)ZN ) −→ 1. Proof. It suffices to show the exactness at the middle. The inclusion relation ¯Ψ(Out◦(OA)) ⊂ Ker(K∗) is clear. Conversely, by Rørdam's result [30] again, for any ξ ∈ Aut(K0(OA ⊗K)) with ξ([1]) = [1], there exists an automorphism β of OA such that β∗ = ξ. The sequence is exact. We will finally mention about the Picard groups of Cuntz algebras. By using Propo- sition 10.2, we know the following theorem. For N = 2, 3, 4, 6, Kodaka has already shown in [12, Corollary 15, Remark 17]. Theorem 10.4. For each 1 < N ∈ N, there exists a short exact sequence: K∗−→ Aut(Z/(1 − N )Z) −→ 1 (10.5) Proof. Since K0(ON ) = Z/(1 − N )Z by [5] and the unit 1 of the C ∗-algebra ON corre- sponds to the generator [1] of the cyclic group Z/(1 − N )Z, the fact α(1) = 1 for any automorphism α ∈ Aut(ON ) ensures us that α∗ = id on K0(ON ). Hence we see that Aut◦(ON ) = Aut(ON ) and hence Out◦(ON ) = Out(ON ). Therefore the exact sequence (10.3) goes to (10.5). ¯Ψ−→ Pic(ON ) 1 −→ Out(ON ) As a corollary, we have Corollary 10.5. Suppose that N − 1 is a prime number. Then the Picard group Pic(ON ) of the Cuntz algebra ON is a semidirect product Out(ON ) ⋊ Z/(N − 2)Z of the outer automorphism group by the cyclic group Z/(N − 2)Z: Pic(ON ) = Out(ON ) ⋊ Z/(N − 2)Z. Proof. As N − 1 is a prime number, an automorphism η of the cyclic group Z/(1 − N )Z is determined by η(1) which can take its value in {1, 2, . . . , N − 2}, so that we have Aut(Z/(1 − N )Z) is isomorphic to Z/(N − 2)Z. Since N is not prime, by [12, Theorem 16], for any k ∈ N with 1 ≤ k ≤ N−1, there exists βk ∈ Aut(ON ⊗K) such that (βk)∗ = k·id on K0(ON ). Hence the correspondence k ∈ {1, 2, . . . , N − 1} −→ [βk] ∈ Pic(ON ) gives rise to a cross section for the exact sequence (10.5). Therefore the exact sequence (10.5) splits and yields a decomposition of Pic(ON ) into a semidirect product Out(ON ) ⋊ Z/(N − 2)Z. Remark 10.6. After the first draft of the paper was completed, the following paper has appeared in arXiv. Kazunori Kodaka, Tamotsu Teruya: The strong Morita equivalence for inclusions of C ∗ -- algebras and conditional expectations for equivalence bimodules, arXiv:1609.08263. In the above paper, Morita equivalence for pairs of C ∗-algebras is defined. However, their definition of Morita equivalence is different from ours. Acknowledgments: This work was supported by JSPS KAKENHI Grant Number 15K04896. 36 References [1] R. Bowen and J. Franks, Homology for zero-dimensional nonwandering sets, Ann. Math. 106(1977), pp. 73 -- 92. [2] L. G. Brown, Stable isomorphism of hereditary subalgebras of C ∗-algebras, Pacific J. Math. 71(1977), pp. 335 -- 348. [3] L. G. Brown, P. Green and M. A. Rieffel, Stable isomorphism and strong Morita equivalence of C ∗-algebras, Pacific J. Math. 71(1977), pp. 349 -- 363. [4] D. Crocker, A. Kumjian, I. Raeburn and D. P. Williams, An equivariant Brauer group and actions of groups of C ∗-algebras, J. Funct. Anal. 146(1997), pp. 151 -- 184. [5] J. Cuntz, K-theory for certain C ∗-algebras, Ann. Math. 117(1981), pp. 181 -- 197. [6] J. Cuntz, A class of C ∗-algebras and topological Markov chains II: reducible chains and the Ext-functor for C ∗-algebras, Invent. Math. 63(1980), pp. 25 -- 40. [7] J. Cuntz and W. Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56(1980), pp. 251 -- 268. [8] J. Franks, Flow equivalence of subshifts of finite type, Ergodic Theory Dynam. Systems 4(1984), pp. 53 -- 66. [9] D. Huang, Flow equivalence of reducible shifts of finite type, Ergodic Theory Dynam. Systems 14(1994), pp. 695 -- 720. [10] T. Kajiwara and Y. Watatani, Jones index theory by Hilbert C ∗-modules and K-theory, Trans. Amer. Math. Soc. 352(2000), pp. 3429 -- 3472. [11] B. P. Kitchens, Symbolic dynamics, Springer-Verlag, Berlin, Heidelberg and New York (1998). [12] K. Kodaka, Picard groups of irrational rotation C ∗-algebras, J. London Math. Soc. 56(1997), pp. 179 -- 188. [13] K. Kodaka, Full projections, equivalence bimodules and automorphisms of stable algebras of unital C ∗-algebras, J. Operator Theory 37(1997), pp. 357 -- 369. [14] , A. Kumjian, On C ∗-diagonals, Canad. J. Math. 38(1986), pp. 969 -- 1008. [15] D. Lind and B. Marcus, An introduction to symbolic dynamics and coding, Cambridge University Press, Cambridge (1995). [16] K. Matsumoto, Strong shift equivalence of symbolic dynamical systems and Morita equivalence of C ∗-algebras, Ergodic Theory Dynam. Systems 24(2004), pp. 199 -- 215. [17] K. Matsumoto, Orbit equivalence of topological Markov shifts and Cuntz -- Krieger algebras, Pacific J. Math. 246(2010), 199 -- 225. 37 [18] K. Matsumoto, Classification of Cuntz -- Krieger algebras by orbit equivalence of topological Markov shifts, Proc. Amer. Math. Soc. 141(2013), pp. 2329 -- 2342. [19] K. Matsumoto, Continuous orbit equivalence, flow equivalence of Markov shifts and circle actions on Cuntz -- Krieger algebras, preprint, arXiv:1501.06965v4, to appear in Math. Z. [20] K. Matsumoto, Topological conjugacy of topological Markov shifts and Cuntz -- Krieger algebras, preprint, arXiv:1604.02763. [21] K. Matsumoto and H. Matui, Continuous orbit equivalence of topological Markov shifts and Cuntz -- Krieger algebras, Kyoto J. Math.54(2014), pp. 863 -- 878. [22] P. S. Muhly, D. Pask and M. Tomforde, Strong shift equivalence of C ∗- correspondences, Israel J. Math. 167 (2008), pp. 315 -- 346. [23] W. Parry and D. Sullivan, A topological invariant for flows on one-dimensional spaces, Topology 14(1975), pp. 297 -- 299. [24] W. L. Paschke, Inner product modules over B∗-algebras, Trans. Amer. Math. Soc. 182(1973), pp. 443 -- 468. [25] I. Raeburn and D. P. Williams, Morita equivalence and continuous-trace C ∗- algebras, Mathematical Surveys and Monographs, vol(60) Amer. Math. Soc. (1998). [26] J. Renault, Cartan subalgebras in C ∗-algebras, Irish Math. Soc. Bull. 61(2008), pp. 29 -- 63. [27] J. Renault, Examples of masas in C ∗-algebras, Operator structures and dynamical systems, pp. 259 -- 265, Contemp. Math., 503, Amer. Math. Soc., Providence, RI, 2009. [28] M. A. Rieffel, Induced representations of C ∗-algebras, Adv. in Math. 13(1974), pp. 176 -- 257. [29] M. A. Rieffel, Morita equivalence for C ∗-algebras and W ∗-algebras, J. Pure Appl. Algebra 5(1974), pp. 51 -- 96. [30] M. Rørdam, Classification of Cuntz-Krieger algebras, K-theory 9(1995), pp. 31 -- 58. [31] M. Tomforde, Strong shift equivalence in the C ∗-algebraic setting: graphs and C ∗- correspondences, Operator theory, Operator Algebras, and Applications, 221 -- 230, Contemp. Math., 414, Amer. Math. Soc., Providence, RI, 2006. [32] N. E. Wegge-Olsen, K-Theory and C ∗-algebras, Oxford University Press, Oxford (1993). [33] Y. Watatani, Index for C ∗-algebras, 424(1990), Memoirs of Amer. Math. Soc. [34] R. F. Williams, Classification of subshifts of finite type, Ann. Math. 98(1973), pp. 120 -- 153. erratum, Ann. Math. 99(1974), pp. 380 -- 381. 38
1501.00135	6	1501	2015-11-16T00:15:07	Classification of finite simple amenable ${\cal Z}$-stable $C^*$-algebras	[ "math.OA" ]	We present a classification theorem for a class of unital simple separable amenable ${\cal Z}$-stable $C^$-algebras by the Elliott invariant. This class of simple $C^$-algebras exhausts all possible Elliott invariant for unital stably finite simple separable amenable ${\cal Z}$-stable $C^$-algebras. Moreover, it contains all unital simple separable amenable $C^$-alegbras which satisfy the UCT and have finite rational tracial rank.	math.OA	math	A classification of finite simple amenable Z-stable C-algebras Guihua Gong, Huaxin Lin and Zhuang Niu Abstract We present a classification theorem for a class of unital simple separable amenable Z- stable C ∗-algebras by the Elliott invariant. This class of simple C ∗-algebras exhausts all possible values of the Elliott invariant for unital stably finite simple separable amenable Z- stable C ∗-algebras. Moreover, it contains all unital simple separable amenable C ∗-algebras which satisfy the UCT and have finite rational tracial rank. 1 Introduction The concept of a C∗-algebra exists harmoniously in many areas of mathematics. The origi- nal definition of C∗-algebras axiomatized the norm closed self-adjoint subalgebras of B(H), the algebra of all bounded linear operators on a Hilbert space H. Thus C∗-algebras are operator algebras. The study of C∗-algebras may also be viewed as the study of a non-commutative ana- logue of topology. This is because every unital commutative C∗-algebra is isomorphic to C(X), the algebra of continuous functions on a compact Hausdorff space X (by the Gelfand transform). If we take our space X and equip it with a group action via homeomorphisms, we enter the realm of topological dynamical systems, where remarkable progress has been made by considering the transformation C∗-algebra C(X) ⋊ G arising from the crossed product construction. Analo- gously, we can consider the study of general crossed products as the study of non-commutative topological dynamical systems. Then there is also naturally appeared the group C∗-algebra of G, which is fundamental in the study of abstract harmonic analysis. C∗-algebra theory is also closely related to the non-commutative geometry. Since von Neumann algebras are C∗-algebras, there are deep and inseparable interactions between C∗-algebra theory and von Neumann al- gebra theory. The list goes on. Therefore, naturally, it would be of great interest to classify C∗-algebras. We have already seen some breathtaking advancements. Early classification theorems include the work of Glimm in the 1950s who classified the Uni- formly Hyperfinite algebras (UHF-algebras) by supernatural numbers. We then saw Dixmier’s and Brattelli’s work on matroid C∗-algebras and AF-algebras respectively. G. A. Elliott gave a classification of AF-algebras by dimension groups in 1976 ([23]) using what is now known as Elliott’s intertwining argument. By 1989, Elliott had began his classification program by classifying simple AT-algebras of real rank zero by scaled ordered K-theory. Since then there has been rapid progress in the program to classify separable amenable simple C∗-algebras now known as the Elliott program. Elliott-Gong ([26]) and Elliott-Gong-Li ([27]) (together with a reduction theorem by Gong ([35])) classified the unital simple AH-algebras with no dimension growth by the Elliott invariant (see 2.4 below). The Elliott’s intertwining argument gave a framework to carry out further classification and focused one’s efforts on studying the maps between algebras and invariants of certain building blocks (for AF algebras these building blocks would be finite dimensional algebras and for AH algebras these building blocks would be certain homogeneous C∗-algebras). In particular, one wants to know when maps between invariants are induced by maps between building blocks (sometimes referred to as an existence theorem) and to know when maps between the building 1 blocks are approximately unitarily equivalent (often called an uniqueness theorem). To classify C∗-algebras without assuming some particular inductive limit structure, one would like to es- tablish abstract existence theorems and uniqueness theorems. These efforts became the engine for these rapid developments ([56], [52], [61] and [18] for example). Both existence theorems and uniqueness theorems used KL-theory ([93]) and the total K-theory (K(A)) developed by Dadarlat and Loring ([17]). These existence and uniqueness theorems provide not only the technical tools for the classification program but also provide the foundation to understand the morphisms in the category of C∗-algebras. The rapid developments mentioned above includes the Kirchberg-Phillips ([46], [47], [87]) classification of purely infinite simple separable amenable C∗-algebras which satisfy the UCT by their K-theory. There is also the classification of unital simple amenable C-algebras in the UCT class which have finite tracial rank ([52], [59] , [63] and [73]). On the other hand, it had been suggested in [19] and [2] that unital simple AH-algebras without a dimension growth condition might behave differently. It was Villadsen ([101] and [102]) who showed that unital simple AH-algebras may have perforated K0-groups and may have stable rank of any positive integer values. M. Rørdam exhibited a unital separable simple C-algebra which is finite but not stably finite ([94]). It was Toms ([100]) who showed that there are unital simple AH-algebras of stable rank one with the same Elliott invariant that are not isomorphic. Before that, Jiang-Su ([44]) constructed a unital simple ASH-algebra Z of stable rank one which has the same Elliott invariant as that of C. In particular, Z has no non-trivial projections. If A is a unital separable amenable simple C∗-algebra which belongs to a classifiable class, then one should expect A⊗Z ∼= A, since they have the same K-theory. Such C∗-algebras are called Z-stable C∗-algebras. The existence of non-elementary simple C∗-algebras which were not Z stable was first proved by Gong-Jiang-Su (see [36]). Toms’ counterexample is in particular not Z-stable. Thus Z-stability should be added to into the hypotheses if one uses the conventional Elliott invariant. (The classes already classified up to that time can easily be shown to be Z-stable.) The next development in this direction came from a new approach of W. Winter who made use of the Z-stability assumption in a remarkably innovative way ([105]). His idea was to view A ⊗ Z as a (stationary) inductive limit of algebras of paths in A ⊗ Q with endpoints in A ⊗ Mp and A ⊗ Mq (where Q is the UHF-algebra with its K0(Q) = Q, p and q are coprime supernatural numbers and Mp and Mq their associated UHF algebras). Suppose that the endpoint algebras are classifiable. Winter showed that if somehow there is a continuous path of isomorphisms from one endpoint to the other, then the class of all A can also be classified. Winter’s procedure provided a new framework to carry out classification, however to actually execute the continuation from endpoint to endpoint alluded to above, one has to establish new types of uniqueness and existence theorems. In other words, just like before, in Elliott’s intertwining argument, the new procedure ultimately, but not surprisingly, depends on some existence and uniqueness theorems. However, this time we need the uniqueness and existence theorems with respect to asymptotic unitary equivalence of the maps involved rather approximate unitary equivalence, which is significantly more demanding. For example, in the case of the existence theorem, we need to construct maps which lift a prescribed KK-element rather than a KL-element, which was once thought to be out of reach for general stably finite algebras since the KK-functor does not preserve inductive limits. It was an unexpected usage of the Basic Homotopy Lemma that made this possible. Moreover, the existence theorem also needs to respect the prescribed rotation related map. The existence theorems are very different from ones in the early study. Inevitably, the uniqueness theorem also becomes more complicated (again the Basic Homotopy Lemma plays the key role). Once we overcame these new hurdles, the class of all Z-stable C∗-algebras A whose tensor products with all UHF-algebras of infinite 2 type are of tracial rank zero were shown to be classified by the Elliott invariant in [105], [72] and [75]. This class was then expanded to the class B of unital simple separable amenable Z-stable C∗-algebras which satisfy the UCT and have finite tracial rank after tensoring some It is shown in [69] that C∗-algebras in the class B can be infinite dimensional UHF-algebra. classified up to isomorphism by the Elliott invariant. These constitute a significant expansion of the class of C∗-algebras which can be classified because there are a great amount of unital simple C∗-algebras which do not have finite tracial rank whose tensored products with a UHF- algebra U do. For example the Jiang-Su algebra Z is projectionless, but Z ⊗ U ∼= U for any infinite dimensional UHF-algebra U. In fact, the class B exhausts all those Elliott invariant with weakly unperforated simple rational Riesz groups as K0-groups under the restriction that the maps from tracial state spaces to state spaces of K0 map the extremal points to extremal points ([76]). The class B not only contains all unital simple separable amenable C∗-algebras with finite tracial rank in the UCT class and the Jiang-Su algebra but also contains many other simple C∗-algebras. More importantly it unifies the previously classified classes such as the so-called dimension drop algebras as well as those dimension drop circle algebras which were known not to be AH-algebras ([68]). However the restriction on the parings of tracial state spaces and state spaces of K0-groups prevents the class B from including those inductive limits of so-called “point-line” algebras, which we called the Elliott-Thomsen building blocks and leads us to the main goal of our paper. The goal of this article is to give a classification of a certain class of unital simple separable amenable C∗-algebras satisfying the UCT by the Elliott invariant. This class is significant because it exhausts all possible values of the Elliott invariant for all unital separable simple amenable stably finite Z-stable C∗-algebras and it also contains the class B properly. We introduce a class of separable simple C∗-algebras which will be called simple C∗-algebras of generalized tracial rank at most one. The definition follows the same spirit as that of tracial rank one (or zero). But, instead of using only finite direct sums of matrix algebras of continuous functions on a one-dimensional finite CW complex, we use certain C-subalgebras of interval algebras. These were first introduced into the Elliott program by Elliott and Thomsen ([32]), sometimes also called one dimensional non-commutative CW complexes (NCCW), as a model to approximate C∗-algebras (tracially). This class will be denoted by B1 (see 9.1 below). Denote by N1 the family of unital separable amenable simple C∗-algebras A which satisfy the UCT such that A⊗ Q ∈ B1, and by N0 the subclass of those C∗-algebras A such that A⊗ Q ∈ B0, where Q is the UHF-algebra with K0(Q) = Q. The main theorem of this article has two parts. The first part states (see 29.8) that if A and B are two unital separable simple Z-stable C∗-algebras in N1. Then A ∼= B if and only if Ell(A) ∼= Ell(B) (see the definition 2.4 below). As a consequence, two unital C∗-algebras in B1 which satisfy the UCT are isomorphic if and only if they have the same Elliott invariant. The second part of the main theorem states that, given any countable weakly unperforated simple ordered group G0 with order unit e, any countable abelian group G1, and any metrizable Choquet simple T and any surjective continuous affine map s : T → Se(G0) (the state space of G0), there exists a unital C∗-algebra A ∈ N1 such that Ell(A) = (K0(A), K0(A)+, [1A], K1(A), T (A), rA) = (G0, (G0)+, e, G1, T, s). The article is organized as follows. Section 2 serves as preliminaries for the article which contains a number of conventions. In Section 3, we study the class of Elliott-Thomsen building blocks, denoted by C, a class of unital C-subalgebras of finite direct sums of interval algebras. We also use C0 for the subclass of C∗-algebras in C with trivial K1. The Elliott-Thomsen building blocks are also called point-line algebras, or one dimensional non-commutative CW complexes as studied in [21] and [22]. Section 4 and 5 discuss the uniqueness theorem for maps from C∗- algebras in C to finite dimensional C∗-algebras. Section 8 presents a uniqueness theorem for maps 3 from a C-algebra in C to another C-algebra in C. This is done by using a homotopy lemma established in Section 6 and existence theorems established in Section 7 to bridge uniqueness In Section 9, the classes B1 and B0 are theorems in Section 4 and 5 to ones in Section 8. introduced. Properties of C∗-algebras in class B1 are discussed in Section 9, 10 and 11. C∗- algebras in B1 are simple and are of generalized tracial rank one (or zero). These unital simple C∗-algebras may also be characterized as tracially approximated by subhomogeneous C∗-algebras with one dimensional spectra. We show, for example, in Section 10, they are Z-stable. Section 12 is dedicated to the main uniqueness theorem for C∗-algebras in B0 used in the isomorphism theorem in Section 21. Sections 13 and 14 are devoted to the range theorem. It includes one of the main results: We show in Section 13 that, given any six-tuples of possible Elliott invariant for unital separable simple Z-stable C∗-algebras, there is a unital simple Z-stable C∗-algebra in N1 whose Elliott invariant is exactly as the given one. Section 14 gives a similar construction for a set of restricted Elliott invariants. In this case, simple C∗-algebras constructed are inductive limits of C∗-algebras which are finite direct sums of some homogeneous C∗-algebras and C∗- algebras in C0. This subclass plays an important role in this article. It should be pointed out, however, that there are unital simple Z-stable C∗-algebras in N1 which can not be written as inductive limits of finite direct sums of homogeneous C∗-algebras and C∗-algebras in C0 (or in C1). Section 15 to 19 could all be described as parts of existence theorems. These deal with the issues of existence for maps from C∗-algebras in C to finite dimensional C∗-algebras and then to C∗-algebras in C which match the prescribed K0-maps and tracial information. Ordered structure and combined simplex information of these C∗-algebras become complicated. We also need to consider maps from homogeneous C∗-algebras to C∗-algebras in C. The mixture with higher dimensional CW complexes does not ease the difficulties. However, in Section 18, we show that, at least under certain restrictions, any given compatible triple which consists of a strictly positive KL-element, a map on the tracial state space and a homomorphism on a quotient of the unitary group, it is possible to construct a homomorphism between C∗-algebras in B0 which matches the triple. Variations of this are also discussed. In Section 19, we show that N1 = N0. In Section 21, we show that any unital simple C∗-algebras with the form A ⊗ U with A ∈ B0 is isomorphic to a unital C∗- algebras constructed in Section 14, and two such C∗-algebras are isomorphic if (and only if) they have the same Elliott invariant. This isomorphism theorem is special but is also the foundation of our main isomorphism theorem in Section 29. In the next seven sections, we show that Winter’s strategy may be carried out in the general case. This requires a much more sensitive uniqueness and existence theorem. The uniqueness theorem now requires an asymptotic unitary equivalence theorem which is proved in Section 27. To do this, we first need another Basic Homotopy Lemma. Sections 22, 24 and 25 establish the needed homotopy lemma while Section 22 serves as the existence theorem for the homotopy lemma. Sections 26 and 28 are for the rotation maps and existence theorem. In Section 29, we show that two Z-stable C∗-algebras in N0 are isomorphic if and only if they have the same Elliott invariant. Since we have shown in section 19 that N0 = N1. This completes the proof of the main isomorphism theorem for all Z-stable C∗-algebras in N1. Acknowledgements: A large part of this article were written during the summers of 2012, 2013 and 2014 when all three authors visited The Research Center for Operator Algebras in East China Normal University. They were partially supported by the center. Both the first named author and the second named author have been supported partially by NSF grants. The work of the third named author has been partially supported by a NSERC Discovery Grant, a Start-Up Grant from the University of Wyoming, and a Simons Foundation Collaboration Grant. Authors would like to thank Michael Yuan Sun for his efforts to read part of the earlier version of this research and for his comments. In Section 20, we continue to study the existence theorem. 4 2 Preliminaries Definition 2.1. Let A be a unital C∗-algebra. Denote by U (A) the unitary group of A, and denote by U0(A) the normal subgroup of U (A) consisting of those unitaries which are in the connected component of U (A) containing 1A. Denote by DU (A) the commutator subgroup of U0(A) and CU (A) the closure of DU (A) in U (A). Definition 2.2. Let A be a unital C∗-algebra and let T (A) denote the simplex of tracial states of A. Let τ ∈ T (A). We say that τ is faithful if τ (a) > 0 for all a ∈ A+ \ {0}. Denote by Tf (A) the set of all faithful tracial states. Denote by Aff(T (A)) the space of all real continuous affine functions on T (A) and denote by LAff b(T (A)) the set of all bounded lower-semi-continuous real affine functions on T (A). Suppose that T (A) 6= ∅. There is an affine map raff : As.a. → Aff(T (A)) defined by raff (a)(τ ) = a(τ ) = τ (a) for all τ ∈ T (A) s.a. the space raff (As.a.) and Aq For each integer n ≥ 1 and a ∈ Mn(A), write τ (a) = (τ ⊗ Tr)(a), where Tr is the (non- and for all a ∈ As.a.. Denote by Aq normalized) standard trace on Mn. Definition 2.3. Let A be a unital stably finite C∗-algebra with T (A) 6= ∅. Denote by ρA : K0(A) → Aff(T (A)) the order preserving homomorphism defined by ρA([p]) = τ (p) for any projection p ∈ Mn(A), n = 1, 2, ... (see the above convention). A map s : K0(A) → R is said to be a state if s is an order preserving homomorphism such that s([1A]) = 1. The set of states on K0(A) is denoted by S[1A](K0(A)). + = raff (A+). Denote by rA : T (A) → S[1A](K0(A)) the map defined by rA(τ )([p]) = τ (p) for all projections p ∈ Mn(A) (for all integer n) and for all τ ∈ T (A). Definition 2.4. Let A be a unital simple C∗-algebra. The Elliott invariant of A, denote by Ell(A), is the following six tuple Ell(A) = (K0(A), K0(A)+, [1A], K1(A), T (A), rA). Suppose that B is another unital simple C∗-algebra. We write Ell(A) ∼= Ell(B), if there is an order isomorphism κ0 : K0(A) → K0(B) such that κ0([1A]) = [1B], an isomorphism κ1 : K1(A) → K1(B) and an affine homeomorphism κρ : T (B) → T (A) such that rA(κρ(t))(x) = rB(t)(κ0(x)) for all x ∈ K0(A) and for all t ∈ T (B). ε if t ∈ [ε/2, ε]. Note that0 ≤ f ≤ 1 and fεfε/2 = fε. Definition 2.5. Let X be a compact metric space, x ∈ X be a point and let r > 0. Denote by B(x, r) = {y ∈ X : dist(x, y) < r}. Let ε > 0. Define fε ∈ C0((0,∞)) to be a function with fε(t) = 0 if t ∈ [0, ε/2], fε(t) = 1 if t ∈ [ε,∞) and fε(t) = 2 t−ε/2 Definition 2.6. Let A be a C∗-algebra. Let a, b ∈ Mn(A)+. Following Cuntz ([12]), we write a . b if there exists a sequence (xn) ⊂ Mn(A) such that limn→∞ x∗nbxn = a. If a . b and b . a, then we write a ∼ b. The relation “∼” is an equivalence relation. Denote by W (A) the Cuntz semi-group of the equivalence classes of positive elements in ∪∞m=1Mm(A) with orthogonal addition (i.e., [a + b] = [a ⊕ b]). If p, q ∈ Mn(A) are projections, then p . q if and only if p is Murray -von Neumann equivalent to a subprojection of q. In particular, when A is stably finite, then p ∼ q if and only if they are von Neumann equivalent. 5 K(A) = Mi=0,1 Ki(A) ⊕ Mi=0,1Mk≥2 Ki(A, Z/kZ). (e 2.1) Denote by QT (A) the set of normalized quasi-traces on A. For a ∈ A+ and τ ∈ QT (A), define dτ (a) = lim ε→0 τ (fε(a)). Suppose that QT (A) 6= ∅. We say A has strict comparison for positive elements, if for any a, b ∈ Mn(A) (for all integer n ≥ 1) dτ (a) < dτ (b) for τ ∈ QT (Mn(A)) implies a . b. Definition 2.7. Let A be a C∗-algebra. Denote by A1 the unit ball of A. Aq,1 the intersection of A+ ∩ A1 in Aq +. Definition 2.8. Let A be a unital C-algebra and let u ∈ U (A). We write Ad u for the au- tomorphism a 7→ u∗au for all a ∈ A. Suppose B ⊆ A is a unital C-subalgebra. Denote by Inn(B, A) the set of all those monomorphisms ϕ : B → A such that there exists a sequence of unitaries {un} ⊂ A so that ϕ(b) = limn→∞ u∗nbun for all b ∈ B. Definition 2.9. Denote by N the class of separable amenable C∗-algebras which satisfy the Universal Coefficient Theorem (UCT). Denote by Z the Jiang-Su algebra ([44]). Note that Z has a unique trace and Ki(Z) = Ki(C) + is the image of (i = 0, 1). A C-algebra A is said to be Z-stable if A ∼= A ⊗ Z. Definition 2.10. Let A be a unital C-algebra. Recall that, following Dadarlat and Loring ([17]), one defines There is a commutative C∗-algebra Ck such that one may identify Ki(A⊗Ck) with Ki(A, Z/kZ). Let A be a unital separable amenable C-algebra, and let B be a σ-unital C-algebra. Following Rørdam ([93]), KL(A, B) is the quotient of KK(A, B) by those elements represented by limits of trivial extensions (see [61]). In the case that A satisfies the UCT, Rørdam defines KL(A, B) = KK(A, B)/P, where P is the subgroup corresponding to the pure extensions of the K∗(A) by K∗(B). In [17], Dadarlat and Loring proved that KL(A, B) = HomΛ(K(A), K(B)). (e 2.2) Now suppose that A is stably finite. Denote by KK(A, B)++ the set of those elements κ ∈ KK(A, B) such that κ(K0(A)\{0}) ⊂ K0(B)+\{0}. Suppose further that both A and B are unital. Denote by KKe(A, B)++ the subset of those κ ∈ KK(A, B)++ such that κ([1A]) = [1B]. Denote by KLe(A, B)++ the image of KKe(A, B)++ in KL(A, B). Definition 2.11. Let A and B be two C∗-algebras and let ϕ : A → B be a map. We will sometime, without warning, continue to use ϕ for the induced map ϕ⊗idMn : A⊗Mn → B⊗Mn. Also ϕ⊗ 1Mn : A → B ⊗ Mn is used for the amplification which maps a to a⊗ 1Mn, the diagonal element with a repeated n times. Throughout the paper, if ϕ is a homomorphism, we will use ϕ∗i : Ki(A) → Ki(B), i = 0, 1, for the induced homomorphism. We will use [ϕ] for the element in KL(A, B) (or KK(A, B) if there is no confusion) which is induced by ϕ. Suppose that T (A) 6= ∅ and T (B) 6= ∅. Then ϕ induces an affine map ϕT : T (B) → T (A) defined by ϕT (τ )(a) = τ (ϕ(a)) for all τ ∈ T (B) and a ∈ As.a.. Denote by ϕ♯ : Aff(T (A)) → Aff(T (B)) the affine continuous map defined by ϕ♯(f )(τ ) = f (ϕT (τ )) for all f ∈ Aff(T (A)) and τ ∈ T (B). Definition 2.12. Let A be a unital separable amenable C∗-algebra and let x ∈ A. Suppose that kxx∗− 1k < 1 and kx∗x− 1k < 1. Then xx−1 is a unitary. Let us use hxi to denote xx−1. Let F ⊂ A be a finite subset and ε > 0 be a positive number. We say a map L : A → B is F-ε-multiplicative if kL(xy) − L(x)L(y)k < ε for all x, y ∈ F. 6 Let P ⊂ K(A) be a finite subset. There is ε > 0 and a finite subset F satisfying the following: for any unital C∗-algebra B and any unital F-ε-multiplicative contractive completely positive linear map L : A → B, L induces a homomorphism [L] defined on G(P), where G(P) is the subgroup generated by P, to K(B) such that kL(p) − qk < 1 and khL(u)i − vk < 1, (e 2.3) where q ∈ Mn(B) (for some n ≥ 1) is a projection such that [q] = [L]([p]) in K0(B) and v ∈ Mn(B) is a unitary such that [v] = [L]([u]) for all [p] ∈ P ∩ K0(A) and [u] ∈ P ∩ K1(A). This also applies to P ∩ Ki(A, Z/kZ) with a necessary modification, by replacing L, by L⊗ idCk , where Ck is the commutative C-algebra described in 2.10, for example. Such a triple (ε,F,P) is called a KL-triple for A. Suppose that Ki(A) is finitely generated. Then, by [17], there is n0 ≥ 1 such that ev- k=2 Ki(A, Z/kZ). Therefore, for some large P, if (ε,F,P) is a KL-triple for A, then , where Gn0 = Li=0,1 Ki(A) ⊕ ery element κ ∈ HomΛ(K(A), K(B)) is determined by κGn0 Li=0,1Ln0 [L] defines an element in KL(A, B) = KK(A, B). In this case, we say (ε,F) is a KK-pair. Definition 2.13. Let A be a unital C∗-algebra. Consider the tensor product A ⊗ C(T). By the Kunneth Theorem, the tensor product induces two canonical injective homomorphisms: β(0) : K0(A) → K1(A ⊗ C(T)) and β(1) : K1(A) → K0(A ⊗ C(T)). (e 2.4) In this way, one may write Ki(A ⊗ C(T)) = Ki(A) ⊕ β(i−1)(Ki−1(A)), i = 0, 1. For each i ≥ 2, one also obtains the following injective homomorphisms: β(i) k : Ki(A, Z/kZ) → Ki−1(A ⊗ C(T), Z/kZ), i = 0, 1. Thus one may write (e 2.5) (e 2.6) Ki(A ⊗ C(T), Z/kZ) = Ki(A, Z/kZ) ⊕ β(i−1) k (Ki−1(A, Z/kZ)), i = 0, 1. (e 2.7) If x ∈ K(A), we use β(x) for β(i)(x) if x ∈ Ki(A) and β(i) k (x) if x ∈ Ki(A, Z/kZ). So one has an injective homomorphism and writes β : K(A) → K(A ⊗ C(T)) K(A ⊗ C(T)) = K(A) ⊕ β(K(A)). (e 2.8) (e 2.9) Let h : A ⊗ C(T) → B be a unital homomorphism. Then h induces a homomorphism h∗i,k : Ki(A ⊗ C(T), Z/kZ) → Ki(B, Z/kZ), k = 0, 2, 3, ... and i = 0, 1. Suppose that ϕ : A → B is a unital homomorphism and v ∈ U (B) is a unitary such that ϕ(a)v = vϕ(a) for all a ∈ A. Then ϕ and v induce a unital homomorphism h : A ⊗ C(T) → B by h(a ⊗ z) = ϕ(a)v for all a ∈ A, where z ∈ C(T) is the identity function on the unit circle T. We use Bott(ϕ, v) for all homomorphisms h∗i−1,k ◦ β(i) k and we write Bott(ϕ, v) = 0 (e 2.10) if h∗i,k ◦ β(i) K1(B). We also use botti(ϕ, v) for h∗i−1 ◦ β(i), i = 0, 1. k = 0 for all k and i. In particular, since A is unital, (e 2.10) implies that [v] = 0 in 7 Suppose that A is a unital separable amenable C∗-algebra, ϕ : A → B is a homomorphism and v ∈ B is a unitary. For any ε > 0 and any finite subset F ⊂ A, there is δ > 0 and a finite subset G ⊂ A such that if k[ϕ(g), v]k < δ for all f ∈ G, (e 2.11) then, by 2.8 of [66], there exists a unital F-ε-multiplicative contractive completely positive linear map L : A ⊗ C(T) → B such that kL(f ) − ϕ(f )k < ε for all f ∈ F and kL(1 ⊗ z) − vk < ε, (e 2.12) where z ∈ C(T) is the standard unitary generator of C(T). Therefore, for each finite subset Q ⊂ K(A ⊗ C(T)), there is δ > 0 and a finite subset G such that, when (e 2.11) holds, [L]Q is well defined. Let P ⊂ K(A) be a finite subset. There are δP > 0 and a finite subset FP if (e 2.11) holds for δP (in place of δ) and FP (in place of F), then satisfying the following: [L]β(P) is well defined. In this case, we will write Bott(ϕ, v)P (x) = [L]β(P)(x) for all x ∈ P. In particular, when [L]β(P) = 0, we will write Bott(ϕ, v)P = 0. (e 2.13) (e 2.14) When K∗(A) is finitely generated, KL(A, B) = HomΛ(K(A), K(B)) is determined by a finitely generated subgroup of K(A) (see [17]). Let P be a finite subset which generates this subgroup. Then, in this case, instead of (e 2.15), we may write that Bott(ϕ, v) = 0. In general, if P ⊂ K0(A), we will write bott0(ϕ, v)P = Bott(ϕ, v)P and if P ⊂ K1(A), we will write (e 2.15) (e 2.16) bott1(ϕ, v)P = Bott(ϕ, v)P . (e 2.17) Definition 2.14. Let A be a unital C∗-algebra. Each element u ∈ U0(A) can be written as u = eih1eih2 ··· eihk for h1, h2, ..., hk ∈ As,a. We write that cer(u) ≤ k if u = eih1eih2 ··· eihk for some selfadjoint elements h1, h2, ..., hk. We write that cer(u) = k if cer(u) ≤ k and u is not a norm limit of unitaries {un} with cer(un) ≤ k − 1. We write cer(u) = k + ε if cer(u) 6= k and there exists a sequence of unitaries {un} ⊂ A such that un ∈ U0(A) with cer(un) ≤ k. Define exponential length of u by cel(u) = inf(cid:8)length of (u(t))0≤t≤1 u(t) ∈ U0(A), u0 = u, u1 = 1(cid:9). Obviously if u = eih1eih2 ··· eihk , then (see [90]) cel(u) ≤ kh1k + kh2k + ··· + khkk . Definition 2.15. Suppose that A is a unital C∗-algebra with T (A) 6= ∅. Recall that CU (A) is the closure of commutator subgroup of U0(A). Let u ∈ U (A). We use ¯u for the image in U (A)/CU (A). It was proved in [98] that there is a splitting short exact sequence: 0 → Aff(T (A))/ρA(K0(A)) → ∞[n=1 U (Mn(A))/CU (Mn(A)) → K1(A) → 0. (e 2.18) 8 Let Jc be a fixed splitting map. Then one may write ∞[n=1 U (Mn(A))/CU (Mn(A)) = Aff(T (A))/ρA(K0(A)) ⊕ Jc(K1(A)). (e 2.19) If A has stable rank k, then K1(A) = U (Mk(A))/U0(Mk(A)). Note It follows from Theorem 3.10 of [38] that csr(C(T, A)) ≤ tsr(A) + 1 = k + 1. U0(Mn(A))/CU (Mn(A)) = U0(Mk(A))/CU (Mk(A)). (e 2.20) ∞[n=1 Then one has the following splitting short exact sequence 0 → Aff(T (A))/ρA(K0(A)) → U (Mk(A))/CU (Mk(A)) → U (Mk(A))/U0(Mk(A)) → 0 (e 2.21) and one may write U (Mk(A))/CU (Mk(A)) = Aff(T (A))/ρA(K0(A)) ⊕ Jc(K1(A)) = Aff(T (A))/ρA(K0(A)) ⊕ Jc(U (Mk(A))/U0(Mk(A))). (e 2.22) (e 2.23) For each piecewise smooth and continuous path {u(t) : t ∈ [0, 1]} ⊂ U (Mk(A)), define DA({u(t)})(τ ) = τ ( du(t) dt u∗(t))dt, τ ∈ T (A). 0 1 2πiZ 1 For each {u(t)}, the map DA({u}) is a real continuous affine function on T (A). Let DA : U0(Mk(A))/CU (Mk(A)) → Aff(T (A))/ρA(K0(A)) denote the de la Harpe and Skandalis ([15]) determinant given by DA(¯u) = DA({u}) + ρA(K0(A)), u ∈ U0(Mk(A)), where {u(t) : t ∈ [0, 1]} ⊂ Mk(A) is a piecewise smooth and continuous path of unitaries with u(0) = 1 and u(1) = u. It is known that the de la Harp and Skandalis determinant is independent of the choice of representatives for ¯u and the choice of path {u(t)}. Define kDA(¯u)k = inf{kDA({v})k : v(0) = 1, v(1) = v and ¯v = ¯u}, (e 2.24) where kDA({v})k = supτ∈T (A) kDA({v})(τ )k. We will fix a metric on U (Mk(A))/CU (Mk(A)). Suppose that u, v ∈ U (Mk(A)). Define dist(¯u, ¯v) =(cid:26) 2, if uv∗ 6∈ U0(Mk(A)), kDA(uv∗)k, otherwise. This is a metric. Note that, if u, v ∈ U0(Mk(A)), then dist(uv, 1k) ≤ dist(¯u, 1k) + dist(¯v, 1k). 9 (e 2.25) (e 2.26) Definition 2.16. Let A be a unital separable amenable C∗-algebra. For any finite subset U ⊂ U (A), there exists δ > 0 and a finite subset G ⊂ A satisfying the following: If B is another unital C∗-algebra and if L : A → B is an F-ε-multiplicative contractive completely positive linear map, then hL(u)i is a well defined element in U (B)/CU (B) for all u ∈ U . We will write L‡(¯u) = hL(u)i. Let G(U ) be the subgroup generated by U . We may assume that L‡ is a well- defined homomorphism on G(U ) so that L‡(u) = hL(u)i for all u ∈ U . In what follows, whenever we write L‡, we mean that ε is small enough and F is large enough so that L‡ is well defined (see Appendix in [73]). Moreover, for an integer k ≥ 1, we will also use L‡ for the map on U (Mk(A))/CU (Mk(A)) induced by L ⊗ idMk . In particular, when L is a unital homomorphism, the map L‡ is well defined on U (A)/CU (A). Definition 2.17. Let C and B be unital C∗-algebras and let ϕ1, ϕ2 : C → B be two monomor- phisms. Define Mϕ1,ϕ2 = {(f, c) : C([0, 1], B) ⊕ C : f (0) = ϕ1(c) and f (1) = ϕ2(c)}. (e 2.27) Denote by πt : Mϕ1,ϕ2 → B the point evaluation at t ∈ [0, 1]. One has the following short exact sequence: 0 → SB ı→ Mϕ1,ϕ2 πe→ C → 0, where ı : SB → Mϕ,ψ is the embedding and πe is the quotient map from Mϕ,ψ to C. Denote by π0, π1 : Mϕ,ψ → C by the point evaluations at 0 and 1, respectively. Since both ϕ1 and ϕ2 are injective, one may identify πe by the point-evaluation at 0 for convenience. Suppose that [ϕ] = [ψ] in KL(C, B). Then Mϕ,ψ corresponds a trivial element in KL(A, B). In particular, the corresponding extensions 0 → Ki(B) ı∗→ Ki(Mϕ,ψ) πe→ Ki(C) → 0 (i = 0, 1) are pure. Definition 2.18. Suppose that T (B) 6= ∅. Let u ∈ Ml(Mϕ,ψ) (for some integer l ≥ 1) be a unitary which is a piecewise smooth continuous function on [0, 1]. Then DB({u(t)})(τ ) = 1 2πiZ 1 0 τ ( du(t) dt u∗(t))dt for all τ ∈ T (B). (see 2.2 for the extension of τ on Ml(B)) as defined in 2.15. Suppose that τ ◦ ϕ = τ ◦ ψ for all τ ∈ T (B). Then there exists a homomorphism Rϕ,ψ : K1(Mϕ,ψ) → Aff(T (B)) defined by Rϕ,ψ([u])(τ ) = DB({u(t)})(τ ) as above which is independent of the choice of the piecewise smooth paths u in [u]. We have the following commutative diagram: K0(B) ρB ց ı∗−→ Aff(T (B)) K1(Mϕ,ψ) ւ Rϕ,ψ Suppose, in addition, that [ϕ1] = [ϕ2] in KK(C, B). Then the following exact sequence splits: 0 → K(SB) → K(Mϕ1,ϕ2) [πe] ⇋ θ K(C) → 0. (e 2.28) We may assume that [π0]◦[θ] = [ϕ1] and [π1]◦[θ] = [ϕ2]. In particular, one may write K1(Mϕ,ψ) = K0(B) ⊕ K1(C). Then we obtain a homomorphism Rϕ,ψ ◦ θK1(C) : K1(C) → Aff(T (B)). 10 We say the rotation map vanishes if there exists such a map θ such above that Rϕ,ψ◦θK1(C) = 0. Denote by R0 the set of those elements λ ∈ Hom(K1(C), Aff(T (B))) for which there is a ho- momorphism h : K1(C) → K0(B) such that λ = ρB◦h. It is a subgroup of Hom(K1(C), Aff(T (B))). One has a well-defined element Rϕ,ψ ∈ Hom(K1(C), Aff(T (B)))/R0 (which is independent of the choice of θ). In this case, there exists a homomorphism θ′1 : K1(C) → K1(Mϕ,ψ) such that (πe)∗1 ◦ θ′1 = idK1(C) and Rϕ,ψ ◦ θ′1 ∈ R0 if and only if there is Θ ∈ HomΛ(K(C), K(Mϕ,ψ)) such that [πe] ◦ Θ = [idC] in KK(C, B) and Rϕ,ψ ◦ ΘK1(C) = 0. In other words, Rϕ,ψ = 0 if and only if there is Θ described above such that Rϕ,ψ ◦ ΘK1(C) = 0. When Rϕ,ψ = 0, one has that θ(K1(C)) ⊂ kerRϕ,ψ for some θ so that (e 2.28) holds. In this case θ also gives the following: kerRϕ,ψ = kerρB ⊕ K1(C). Definition 2.19. Let C be a C∗-algebra, let a, b ∈ C be two elements and let ε > 0. We write a ≈ε b if ka − bk < ε. Suppose that A is another C∗-algebra, L1, L2 : C → A are two maps and F ⊂ C is a subset. We write L1 ≈ε L2 on F, (e 2.29) if kL1(c) − L2(c)k < ε for all c ∈ F. Definition 2.20. Let A and B be C∗-algebras, and assume that B is unital. Let H ⊆ A+ \{0} be a finite subset, let T : A+ \ {0} → R+ \ {0} and let N : A+ \ {0} → N be two maps. Then a map L : A → B is said to be T × N -H-full if for any h ∈ H, if there are b1, b2, ..., bN (h) ∈ B such that kbik ≤ T (h) and b∗i L(h)bi = 1B. N (h)Xi=1 3 The Elliott-Thomsen building blocks To generalize the class of C∗-algebras of tracial rank at most one, we naturally consider all subhomogeneous C∗-algebras with one dimensional spectrum which, in particular, include circle algebras as well as dimension drop algebras. We begin, however, with the following special form: Definition 3.1 (See [32] and [25]). Let F1 and F2 be two finite dimensional C∗-algebras. Sup- pose that there are two unital homomorphisms ϕ0, ϕ1 : F1 → F2. Denote the mapping torus Mϕ1,ϕ2 by A = A(F1, F2, ϕ0, ϕ1) = {(f, g) ∈ C([0, 1], F2) ⊕ F1 : f (0) = ϕ0(g) and f (1) = ϕ1(g)}. These C∗-algebras have been introduced into the Elliott program by Elliott and Thomsen ([32]), and in [25], Elliott used this class of C-algebras and some other building blocks with 2- dimensional spectra to realize any weakly unperforated ordered group as the K0-group of a simple ASH C-algebra. Denote by C the class of all unital C∗-algebras of the form A = A(F1, F2, ϕ0, ϕ1) and all finite dimensional C∗-algebras. These C∗-algebras will be called Elliott-Thomsen building blocks. A unital C∗-algebra C ∈ C is said to be minimal if it is not a direct sum of more than one copy of C∗-algebras in C. If A ∈ C is minimal and is not finite dimensional, in what follows, we may assume that kerϕ0 ∩ kerϕ1 = {0}. In general, if A ∈ C, and kerϕ0 ∩ kerϕ1 6= {0}. Then, we 11 1 For t ∈ (0, 1), define πt can write A = A1⊕ (kerϕ0∩ kerϕ1), where A1 = A(F ′1, F2, ϕ′0, ϕ′1), F1 = F ′1⊕ (kerϕ0∩ kerϕ1) and for i = 1, 2. (Note that A1 = A(F ′1, F2, ϕ′0, ϕ′1) satisfies the condition kerϕ′0 ∩ kerϕ′1 = ϕ′i = ϕiF ′ {0}, and that kerϕ0 ∩ kerϕ1 is a finite dimensional C∗-algebra.) : A → F2 by πt((f, g)) = f (t) for all (f, g) ∈ A. For t = 0, define π0 : A → ϕ0(F1) ⊂ F2 by π0((f, g)) = ϕ0(g) for all (f, g) ∈ A. For t = 1, define π1 : A → ϕ1(F1) ⊂ F2 by π1((f, g)) = ϕ1(g)) for all (f, g) ∈ A. In what follows, we will call πt a point-evaluation of A at t. There is a canonical map πe : A → F1 defined by πe(f, g) = g for all pair (f, g) ∈ A. It is a surjective map. The notation πe will be used for this map throughout this paper. If A ∈ C, then A is the pull-back of A /❴❴❴❴❴❴ C([0, 1], F2) πe F1 (ϕ0,ϕ1) (π0,π1) / F2 ⊕ F2 (e 3.1) Every such pull-back is an algebra in C. Infinite dimensional C-algebras in C are also called one-dimensional non-commutative finite CW complexes (NCCW) (see [21] and [22]). Denote by C0 the sub-class of those C∗-algebras A in C such that K1(A) = {0}. It follows from Theorem 6.22 of [21] that C∗-algebras in C are semiprojective, an important feature that we will use later without warning. Lemma 3.2. Let f ∈ C([0, 1], Mk(C)) and let a0, a1 ∈ Mk(C) be invertible elements with ka0 − f (0)k < ε and ka1 − f (1)k < ε. Then there exists an invertible element g ∈ C([0, 1], Mk(C)) such that g(0) = a0, g(1) = a1 and kf (t) − g(t)k < ε for all t ∈ [0, 1]. Proof. Let S⊂Mk(C) be the set consisting of all singular matrices. Then Mk(C) is a 2k2- dimensional differential manifold (diffeomorphic to R2k2 ) with S being finite union of closed submanifold of codimension at least two. Since each continuous map between two differential manifolds (perhaps with boundary) can be approximated arbitrarily well by smooth maps, we can find f1 ∈ C∞([0, 1], Mk(C)) with f1(0) = a0 and f1(1) = a1 and kf1(t) − f (t)k < ε′ < ε. Apply the relative version of the transversality theorem—the corollary on page 73 of [41] and its proof (see page 70 and 68 of [41]), for example, with Z = S, Y = Mk(C), X = [0, 1] with ∂X = {0, 1}, one can find g ∈ C∞([0, 1], Mk (C)) with g(0) = f1(0), g(1) = f1(1), g(t) /∈ S and kg(t) − f1(t)k < ε − ε′. Proposition 3.3. If A ∈ C, then A has stable rank one. Proof. Let (f, a) be in A with f ∈ C([0, 1], F2) and a ∈ F1 with f (0) = ϕ0(a) and f (1) = ϕ1(a). For any ε > 0, since F1 is a finite dimensional C∗-algebra, there is an invertible element b ∈ F1 such that kb − ak < ε. Since ϕ0, ϕ1 are unital, ϕ0(b) and ϕ1(b) are invertible, kϕ0(b) − f (0)k < ε and kϕ1(b) − f (1)k < ε. By Lemma 3.2, there exists an invertible element g ∈ C([0, 1], F2) (applied to each direct sum- mand of F2) such that g(0) = ϕ0(b), g(1) = ϕ1(b) and This is what as desired. kg − fk < ε. 12 / ✤ ✤ ✤ / 3.4. Let F1 = MR(1) ⊕ MR(2) ⊕ ··· ⊕ MR(l) and M2 = Mr(1) ⊕ Mr(2) ⊕ ··· ⊕ Mr(k). A = (F1, F2, ϕ0, ϕ1). Denote by m the greatest common factor of(cid:8)R(1), R(2),··· , R(l)(cid:9). Then each r(j) is also a multiple of m. Let eF1 = M R(1) (C) and eF2 = (C) ⊕ ··· ⊕ M R(l) (C) ⊕ ··· ⊕ M r(k) (C) ⊕ M R(2) (C) ⊕ M r(2) M r(1) m m m m m m (C). Let eϕ0,eϕ1 : eF1 → eF2 be maps such that eϕ0∗,eϕ1∗ : K0(eF1) = Zl −→ Zk satisfying eϕ0∗ = (aij)k×l and eϕ1∗ = (bij)k×l. That is, the maps which are represented by the same matrices represent ϕ0∗ and ϕ1∗. By 3.14, A(F1, F2, ϕ0, ϕ1) ∼= Mm(A( F1, F2, ϕ0, ϕ1)). Proposition 3.5. Let A = A(F1, F2, ϕ0, ϕ1). Then K1(A) = Zl/Im(ϕ0∗0 − ϕ1∗0) and with positive cone being K0(A)∩Zl K0(A) ∼= = ϕ1∗ v1 v2 ... vl v1 v2 ... vl v1 v2 ... vl    ∈ Zl , ϕ0∗     + = +, and scale  ∈ Zl, where Zl R1 R2 ... Rl (e 3.2) ⊂Zl . v1 v2 ... vl   ; vi ≥ 0 Moreover, the map πe : A → F1 induces the natural order embedding (πe)∗0 : K0(A) → K0(F1) = Zl, in particular, kerρA = {0}. Furthermore, if K1(A) = {0}, then Zl/K0(A) ∼= K1(C0((0, 1), F2)) is torsion free, and in this case, [πe]Ki(A,Z/kZ) is injective for all k ≥ 2 and i = 0, 1. Proof. Most of these assertions are known. We sketch the proof here. Consider the short exact sequence 0 / C0((0, 1), F2) / A πe / / F1 / 0. We obtain 0 / K0(A) πe∗ / / K0(F1) / K0(F2) / K1(A) / 0, (e 3.3) where the map K0(F1) → K0(F2) is given by ϕ0∗0 − ϕ1∗0. In particular, (πe)∗0 is injective. If p ∈ πe(Mm(A)) is a projection, then (ϕ0)∗0(p) = (ϕ1)∗0(p). Therefore ϕ0(p) and ϕ1(p) have the same rank. It follows that there is a projection q ∈ Mm(A) such that πe(q) = p. This implies that (πe)∗0 is an order embedding. This also implies that kerρA = {0}. Other descriptions of Ki(A) (i = 0, 1) also follow. The quotient group K0(F1)/(πe)∗0(K0(A)) is always torsion free, since K0(F1)/(πe)∗0(K0(A)) is a subgroup of K0(F2). It happens that K0(F1)/(πe)∗0(K0(A)) = K0(F2) holds if and only if K1(A) = {0}. In the case that K1(A) = {0}, one also computes that K0(A, Z/kZ) may be identified with K0(A)/kK0(A) and K1(A, Z/kZ) = {0} for all k ≥ 2. To see that [πe]K0(A,Z/kZ) is injective for k ≥ 2, let ¯x ∈ K0(A, Z/kZ) = K0(A)/kK0(A) and let x ∈ K0(A) be such that its image in K0(A)/kK0(A) is ¯x. If [πe](¯x) = 0, then (πe)∗0(x) ∈ kK0(F1). Let y ∈ K0(F1) be such that ky = (πe)∗0(x). Then k¯y = (πe)∗0(x) = 0 in K0(F1)/(πe)∗0(K0(A)). This implies that K0(F2) has torsion. A contradiction. Therefore [πe]K0(A/Z/kZ) is injective for k ≥ 2. Then the rest of proposition also follows from (e 3.3). 13 / / / / / / / g(t) = ϕ0(h)(1 − 2t), ϕ1(h)(2t − 1), 0 ≤ t ≤ 1 2 , 2 < t ≤ 1. 1 3.6. Let us describe how to identify a unitary u ∈ A with [u] ∈ K1(A) = Zk/(ϕ1∗ − ϕ0∗)(Zl). Let u = (f, a) ∈ A, where a = (a1, a2, ..., al) ∈ Ll j=1 MR(j)(C) = F1. Note that every unitary in MN (C) can be written as eih for some self-adjoint element h ∈ MN (C)). Write h = (h1, h2, ..., hl) ∈Ll j=1 M{j}(C) = F1 such that a = eih. Let (g, h) ∈ A be defined by Obviously, (g, h) is well defined and inside A. Let v = (f1, 1F1), where f1(t) = f (t)e−ig(t). Let U (s) = (f (t)e−ig(t)s, eih(1−s)) for s ∈ [0, 1]. Then {U (s) : s ∈ [0, 1]} is a continuous path of unitaries in A with U (0) = u and U (1) = (f1, 1F1). Therefore uv∗ ∈ U0(A). In particular, [u] = [v] ∈ K1(A), and v = (v1, v2, ..., vk) with vj(0) = vj(1) = 1. Write f = (a1, a2, ..., ak) and f e−igs = (b1(s), b2(s), ..., bk(s)), where aj, bj(s) ∈ C([0, 1], Mr(j)) for all s ∈ [0, 1]. Suppose that det(am(t)) = 1 for all t ∈ [0, 1], m = 1, 2, ..., k, and det(eihj ) = 1 for j = 1, 2, ..., l. Define Bm(s) = det(bm(s))∗, B(s) = (B1(s), B2(s), ..., Bk(s)) and d(s) = (det(eih1(1−s)), det(eih2(1−s)), ..., det(eihk(1−s))). Define W (s) = (Bj(s)U (s), d(s)∗eih(1−s)). Then {W (s) : s ∈ [0, 1]} is a continuous path of unitaries in A such that W (0) = u and U (1) = (B(1)f1, 1). Note that det(Bm(1)bm(1)) = 1 for 1 ≤ m ≤ k. In other words, if det(u(ψ)) = 1 for all irreducible representations ψ, then we may also assume that det(v(ψ)) = 1 for ψ. Note that v ∈ C0((0, 1), F2e), the unitilization of the ideal C0((0, 1), F2)⊂A. Hence u defines an element (s1, s2, ..., sk) ∈ Zk = K1(C0((0, 1), F2)), where sj is winding number of the map t ∈ [0, 1] −→ det(vj(t)) ∈ T ⊆ C. In particular, if det(u(ψ)) = 1 for all irreducible representations ψ, then u ∈ U0(A). Such a k-tuple gives an element [(s1, s2, ..., sk)] ∈ Zk/(ϕ1∗ − ϕ0∗)(Zl) . Lemma 3.7. Let A = A(F1, F2, ϕ0, ϕ1) be as in Definition 3.1. A unitary u ∈ U (A) is in CU (A) if and only if for each irreducible representation ψ of A, one has that det(ψ(u)) = 1. Proof. Obviously, the condition is necessary. One only has to show that the condition is also sufficient. From the discussion 3.6 above, if u ∈ U (A) with det(ψ(u)) = 1 for all irreducible representations ψ, then u ∈ U0(A). Write F1 = MR(1) ⊕ RR(2) ⊕ ··· ⊕ MR(l) and F2 = Mr(1) ⊕ Mr(2) ⊕ ··· ⊕ Mr(k). Then, since u ∈ U0(A), we may write that u =Qm n=1 exp(i2πhn) for some hn ∈ As.a., n = 1, 2, ..., m. We may write hn = (hnI , hnq) ∈ A with ϕi(hnq) = hnI (i), i = 0, 1. For any irreducible representation ψ of A,Pn j=1 Trψψ(hj ) ∈ Z, where Trψ is the standard (unnormalized) trace on ψ(A) ∼= Mn(ψ) for some integer n(ψ). Put Hj(t) = Trj(Pm n=1 πj(hn(t))) and H(t) = (H1, H2, ..., Hk) for t ∈ [0, 1], where Trj is the standard trace on Mr(j) and πj : F2 → Mr(j) is the projection map, j = 1, 2, ..., k. Put a(ψ) =Pn j=1 Trψ(hj ) for those ψ corresponding to irreducible representations of F1. Since Pn j=1 Trψψ(hj ) ∈ Z, there is a projection p ∈ MN (F1) such that Trψψ(p) = a(ψ) for some N ≥ 1. Then, Tr′j(π′j ◦ ϕ0(p)) = Hj(0) = Hj(1) = Tr′j(ϕ1(p)), where Tr′j is the standard trace on MN (Mr(j)) and π′j : MN (F2) → MN (Mr(j)) is the projection map. There is a projection Pj ∈ MN (C([0, 1], Mr(j))) such that Pj(0) = πj ◦ ϕ0(p) and Pj(1) = πj ◦ ϕ1(p). Put 14 P ∈ MN (C([0, 1], F1)) such that π′j(P ) = Pj and e = (P, p). Then e ∈ MN (A). Consider the continuous path u(t) =Qm n=1 exp(i2πhnt) for t ∈ [0, 1]. Then u(0) = 1, u(1) = u and du(t) τ ( dt u∗(t)) = i2π τ (hn) for all τ ∈ T (A). But, for all a ∈ A and τ ∈ T (A), αstrs(a) + trj(πn(a))dµj (t), mXn=1 kXj=1Z(0,1) τ (a) = lXs=1 s=1 αs +Pk 2πiZ 1 τ ( 1 0 where µj is a Borel measure on (0, 1), trs is the tracial state on MR(s) and trj is a tracial state on Mr(j), αi ≥ 0 andPl j=1 kµjk = 1. We compute that du(t) u(t)∗)dt = τ (e) for all τ ∈ T (A). dt In other words, DA({u(t)}) ∈ ρA(K0(A)) (see 2.15). It follows from [98] and the fact that A has stable rank one that u ∈ CU (A) (see also [38]). The following is known (see [96], [97], and [86]). ku − u1k < ε, Lemma 3.8. Let u be a unitary in C([0, 1], Mn). Then, for any ε > 0, there exist continuous functions hj ∈ C([0, 1])s.a. such that where u1 = exp(iπH), H =Pn j=1 hjpj and {p1, p2, ..., pn} is a set of mutually orthogonal rank one projections in C([0, 1], Mn), and exp(iπhj (t)) 6= exp(iπhk(t)) if j 6= k for all t ∈ (0, 1), u1(0) = u(0) and u1(1) = u(1). Furthermore, if det(u(t)) = 1 for all t ∈ [0, 1], then we may also assume that det(u1(t)) = 1 for all t ∈ [0, 1]. Remark 3.9. It follows from the above (also see Theorem 1-2 [96] and [97]) that u, v ∈ U (Mk(C[0, 1])) are approximately unitarily equivalent (that is, for any ε, there is w ∈ U (Mk(C[0, 1])) such that ku − wvw∗k < ε) if and only if for each t ∈ [0, 1], u(t) and v(t) have the same set of eigenvalues counting multiplicities (see also Lemma 3.1 of [71]). j=1 Mr(j), there are real valued functions hj Lemma 3.10. Let A = A(F1, F2, ϕ0, ϕ1) be as in Definition 3.1. For any unitary (f, a) ∈ U (A), ε > 0, there is a unitary (g, a) ∈ U (A) such that kg − fk < ε and, for each block Mr(j)⊂F2 = Lk : [0, 1] → R such that gj = Pn j=1 hjpj and {p1, p2, ..., pn} is a set of mutually orthogonal rank one projections in C([0, 1], Mr(j)) and exp(iπhj(t)) 6= exp(iπhk(t)) if j 6= k for all t ∈ (0, 1). Moreover, if (f, a) ∈ CU (A), one can choose (g, a) ∈ CU (A). Proof. By 3.8, for each unitary f j ∈ C([0, 1], Mr(j)), one can approximates f j by gj 1 to within ε such that gj(0) = f j(0), gj(1) = f j(1) and for each t in the open interval (0, 1), gj (t) has distinct eigenvalues. If (f, a) ∈ U (A), then (g, a) ∈ U (A) too. Combining 3.8 with 3.7, the last statement also follows. 2, ..., hj kj 1, hj Remark 3.11. In (3.10), one may assume that hj small t ∈ (0, δ), one may assume 1(0), hj 2(0), ..., hj kj (0) ∈ [0, 1). For an arbitrarily max{hj i (t); 1 ≤ i ≤ kj} − min{hj i (t); 1 ≤ i ≤ kj} < 1 (e 3.4) 15 i1 i2 i2 (t) 6= hj (t) 6= e2πihj and hj i1 e2πihj Hence, one may assume that max{hj (t) for i1 6= i2. From the choice of gj, we know that for any t ∈ (0, 1), (t). That is, hj (t) 6∈ Z. This implies that (e 3.4) holds for all t ∈ (0, 1). i1 (t)− hj i2 i (1); 1 ≤ i ≤ kj} − min{hj i (1); 1 ≤ i ≤ kj} ≤ 1. (e 3.5) Lemma 3.12. For any u ∈ CU (A), one has that cer(u) ≤ 2 + ε and cel(u) ≤ 4π. Moreover, there exists a continuous path of unitaries {u(t) : t ∈ [0, 1]} ⊂ CU (A) with length at most 4π such that u(0) = 1A and u(1) = u. Proof. Case (i): the case that u = (f, a) with a = 1. As in 3.10, up to approximation within arbitrarily small pre-given number ε > 0, u is unitarily equivalent to v = (g, a) ∈ CU (A) with gj(t) = diag(cid:0)e2πihj 1(t), e2πihj 2(t), ..., e 2πihj kj (t)(cid:1) with distinct eigenvalues for each t ∈ (0, 1). (Note that since f (0) = f (1) = 1 = g(0) = g(1), the unitary to intertwining the approximation of u and v can be chosen to be 1 at t = 0, 1, and therefore, the unitary is in A = A(F1, F2, ϕ0, ϕ1).) Furthermore, one can assume that 1(0) = hj hj 1(0) = ··· = hj kj (0) = 0. Since det(gj(t)) = 1 for all t ∈ [0, 1], one has that hj 1(t) + hj 2(t) + ··· + hj kj (t) ∈ Z. By the continuity of each hj s(t), we know that (e 3.6) hj s(t) = 0. s (cid:8)hj Furthermore, by hj Otherwise, min kjXs=1 s(1) ∈ Z (since gj(1) = 1), we know that hj s (cid:8)hj s(1)(cid:9) ≥ 1 which implies that s(1)(cid:9) ≤ −1 and max s(1)(cid:9) ≥ 2, s (cid:8)hj s(1)(cid:9) − min s (cid:8)hj and this contradicts to Remark 3.11. That is, one has proved that h = (cid:0)(h1, h2,··· , hk), 0(cid:1), where hj(t) = diag(cid:0)hj (t)(cid:1) is an element in A = A(F1, F2, ϕ0, ϕ1) with h(0) = 2(t),··· , hj h(1) = 0. As g = e2πih, we have cer(u) ≤ 1 + ε. We also have tr(h(t)) = 0 for all t. s(t)(cid:9) ≤ 1 (see (e 3.5) in Remark 3.11) s (cid:8)hj s(t)(cid:9) − min s (cid:8)hj s(1) = 0 for all s ∈ {1, 2,··· , kj}. It follows from (e 3.6) above and max 1(t), hj max that kj hj s(t) ⊂ (−1, 1), t ∈ [0, 1], s = 1, 2,··· , kj . Hence k2πhk ≤ 2π which implies cel(u) ≤ 2π. Moreover let u(s) = exp(is2h). Then u(0) = 1 and u(1/2) = u. Since tr(s2h(t)) = 2s · tr(h(t)) = 0 for all t, one has that u(s) ∈ CU (A) for all s ∈ [0, 1/2]. 1 . So aj = exp(2πihj ) for hj ∈ F j Case (ii): The general case. Since a = (a1, a2,··· , al) with det(aj) = 1 for aj ∈ F j 1 with tr(hj) = 0 and khjk < 1. Define H ∈ A(F1, F2, ϕ0, ϕ1) by H(t) = ϕ0(h1, h2, ..., hl) · (1 − 2t), ϕ1(h1, h2, ..., hl) · (2t − 1), if 0 ≤ t ≤ 1 2 , if 1 2 < t ≤ 1. 16 (Note H( 1 2 ) = 0, H(0) = ϕ0(h1, h2,··· , hl), H(1) = ϕ1(h1, h2,··· , hl) and therefore H ∈ A(F1, F2, ϕ0, ϕ1). Moreover tr(H(t)) = 0 for all t. Then u′ = u· exp(−2πiH) ∈ A(F1, F2, ϕ0, ϕ1) with u′(0) = u′(1) = 1. By Case (i), cer(u′) ≤ 1 + ε and cel(u′) ≤ 2π, we have cer(u) ≤ 2 + ε and cel(u) ≤ 2π + 2πkHk ≤ 4π. Furthermore, we note that exp(−2πsH) ∈ CU (A) for all s as in Case (i). 3.13. Let F1 = MR1(C) ⊕ MR2(C) ⊕ ··· ⊕ MRl(C), let F2 = Mr1(C) ⊕ Mr2(C) ⊕ ··· ⊕ Mrk (C) and let ϕ0, ϕ1 : F1 → F2 be unital homomorphisms, where Rj and ri are positive integers. Then ϕ0 and ϕ1 induce homomorphisms ϕ0∗, ϕ1∗ : K0(F1) = Zl −→ K0(F2) = Zk by matrices (aij)k×l and (bij)k×l, respectively, where ri =Pl Proposition 3.14. For fixed finite dimensional C-algebras F1, F2, the C-algebra A = A(F1, F2, ϕ0, ϕ1) is completely determined by ϕ0∗, ϕ1∗ : Zl −→ Zk (up to isomorphisms). Proof. Let B = A(F1, F2, ϕ′0, ϕ′1) with ϕ′0∗ = ϕ0∗, ϕ′1∗ = ϕ1∗. It is well known that there exist two unitaries u0, u1 ∈ F2 such that u0ϕ0(a)u∗0 = ϕ′0(a), j=1 aijRj for i = 1, 2, ..., k. a ∈ F1 and u1ϕ1(a)u∗1 = ϕ′1(a), a ∈ F1. Since U (F2) is path connected, there is a unitary path u : u(1) = u1. Define ϕ : A → B by ϕ(f, a) = (g, c), [0, 1] → U (F2) with u(0) = u0 and where g(t) = u(t)f (t)u(t)∗. Then a straightforward calculation shows that the map ϕ is a -isomorphism. Theorem 3.15. Let A = A(F1, F2, ϕ1, ϕ2) be in C. Then K0(A)+ is finitely generated by its min- imal elements; in other words, there is an integer m ≥ 1 and finitely many minimal projections of Mm(A) such that these minimal projections generate the positive cone K0(A)+. Proof. We first show that K0(A)+ \ {0} has only finitely many minimal elements. Suppose otherwise that {qn} is an infinite set of minimal elements of K0(A)+ \ {0}. Write qn = (m(1, n), m(2, n), ..., m(l, n)) ∈ Zl +, where m(i, n) are non-negative integers, i = 1, 2, ..., l and n = 1, 2, .... If there is an integer M ≥ 1 such that m(i, n) ≤ M for all i and n, then {qn} is a finite set. So we may assume that {m(i, n)} is unbounded for some 1 ≤ i ≤ l. There is a subsequence of {nk} such that limk→∞ m(i, nk) = ∞. To simplify the notation, without loss of generality, we may assume that limn→∞ m(i, n) = ∞. We may assume that, for some j, {m(j, n)} is bounded. Otherwise, by passing to a subsequence, we may assume that limn→∞ m(i, n) = ∞ for all i ∈ {1, 2, ..., l}. Therefore limn→∞ m(i, n) − m(i, 1) = ∞. It follows that, for some n ≥ 1, m(i, n) > m(i, 1) for all i ∈ {1, 2, ..., l}. Therefore qn ≥ q1 which contradicts the fact that qn is minimal. By passing to a subsequence, we may write {1, 2, ..., l} = N ⊔ B such that limn→∞ m(i, n) = ∞ if i ∈ N and {m(i, n)} is bounded if i ∈ B. Therefore {m(j, n)} has only finitely many different values if j ∈ B. Thus, by passing to a subsequence again, we may assume that m(j, n) = m(j, 1) if j ∈ B. Therefore, for some n > 1, m(i, n) > m(i, 1) for all n if i ∈ N and m(j, n) = m(j, 1) for all n if j ∈ B. It follows that qn ≥ q1. This is impossible since qn is minimal. This shows that K0(A)+ has only finitely many minimal elements. To show that K0(A)+ is generated by these minimal elements, fix an element q ∈ K0(A)+ \ {0}. It suffices to show that q is a finite sum of minimal elements in K0(A)+. If q is not minimal, consider the set of all elements in K0(A)+\{0} which are smaller than q. This set is finite. Choose one which is minimal among them, say p1. Then p1 is minimal element in K0(A)+\{0}, otherwise 17 there is one smaller than p1. Since q is not minimal, q 6= p1. Consider q − p1 ∈ K0(A)+ \ {0}. If q − p1 is minimal, then q = p1 + (q − p1). Otherwise, we repeat the same argument to obtain a minimal element p2 ≤ q − p1. If q − p1 − p2 is minimal, then q = p1 + p2 + (q − p1 − p2). Otherwise we repeat the same argument. This process is finite. Therefore q is a finite sum of minimal elements in K0(A)+ \ {0}. Theorem 3.16. The exponential rank of A = A(F1, F2, ϕ0, ϕ1) is at most 3 + ε. Proof. For each unitary u ∈ U0(A), by 3.6, one can write u = veih, where v = (g, h) with v(0) = v(1) = 1 ∈ F2. So we only need to prove the exponential rank of v is at most 2+ ε. Consider v as an element in C0((0, 1), F2)e) which defines an element (s1, s2, ..., sk) ∈ Zk = K1(C0((0, 1), F2)). Since [v] = 0 in K1(A), there are (m1, m2, ..., ml) ∈ Zl such that (s1, s2, ..., sk) = ((ϕ1)∗0 − (ϕ0)∗0)((m1, m2, ..., ml)). Note that (ϕ0)∗0((R1, R2, ..., Rl)) = (ϕ1)∗0((R1, R2, ..., Rl)) = (r1, r2, ..., rk) = [1F2] ∈ K0(F2). Increasing (m1, m2,··· , ml) by adding a positive multiple of(cid:0)R1, R2, ..., Rl(cid:1), we can assume mj ≥ 0 for all j ∈ {1, 2, ..., l}. Let a = (m1P1, m2P2, ..., mlPl), where Let h be defined by 1 Pj = h(t) = 0 . . . 0  ∈ M{j}(C) ⊂ F1 . ϕ0(a)(1 − 2t), ϕ1(a)(2t − 1), 0 ≤ t ≤ 1 2 , 2 < t ≤ 1. 1 Then (h, a) defines a selfadjoint element in A. One also has e2πih ∈ C0((0, 1), F2e), since e2πih(0) = e2πih(1) = 1. Furthermore, e2πih defines (ϕ1∗ − ϕ0∗)((m1, m2, ..., ml)) ∈ Zk as an element in K1(C0((0, 1), F2)) = Zk. Let w = ve−2πih. Then w satisfies w(0) = w(1) = 1 and w ∈ C0((0, 1), F2e) defines (0, 0, ..., 0) ∈ K1(C0((0, 1), F2)) up to an approximation within a sufficiently small ε, one can assume that w = (w1, w2,··· , wk) such that for all j = 1, 2,··· , k, 1(t), e2πihj 2(t), ..., e 2πihj kj (1) wj(t) = diag(cid:0)e2πihj (2) the numbers e2πihj (3) hj 1(0) = hj 2(0) = ··· = hj kj Since wj(1) = 1, one has that hj (0) = 0. i (1) ∈ Z. (t)(cid:1); (t) 1(t), e2πihj 2(t), ..., e 2πihj kj are distinct for all t ∈ (0, 1); On the other hand, the unitary w defines 2(t) + ··· + hj hj 1(t) + hj kj (t) ∈ Z ∼= K1(C0((0, 1), F j 2 )) which is zero by the property of w. From (e 3.5), one has hj i1 (1) − hj i2 (1) ≤ 1. This implies hj 1(1) = hj 2(1) = ··· = hj kj (1) = 0. Hence h =(cid:0)(h1, h2, ..., hk), 0(cid:1) defines a selfadjoint element in A and w = e2πih. 18 3.17. Let A = A(F1, F2, ϕ0, ϕ1) ∈ C, where F1 = Mr1 ⊕ Mr2 ⊕···⊕ Mrl and F2 = MR1 ⊕ MR2 ⊕ ··· ⊕ MRk . Let us calculate the Cuntz semigroup of A. It is well known the extreme points of T(A) are canonically one-to-one corresponding to the irreducible representations of A, which are given by kaj=1 (0, 1)j ∪ {ρ1, ρ2, ..., ρl} = Irr(A), where (0, 1)j is the same open interval (0, 1). We use subscript j to indicate the j-th copy. The affine space Aff(T(A)) can be identified with the subset of consisting of (f1, f2, ..., fk, g1, g2, ..., gl) satisfying the condition l copies {z } kMj=1 C([0, 1]j , R) ⊕ (R ⊕ R ⊕ ··· ⊕ R) fi(0) = 1 Ri lXj=1 aijgj · rj and fi(1) = 1 Ri lXj=1 bijgj · rj, where (aij)k×l = ϕ0∗ and (bij)k×l = ϕ1∗ as in 3.13. π ∈ Irr(A) For any selfadjoint h ∈ (A ⊗K)+, one can define a map Dh : Irr(A) → Z+ ∪ {∞} as for any Dh(π) = lim n→∞ Tr(π ⊗ idK(h where Tr is the unnormalized trace. Then we write Dh = (D1 where Di h = Dh(0,1)i , which satisfies the following conditions h is lower semi-continuous on each (0, 1)j , (1) Dj 1 n )) , h, D2 h, ..., Dk h, Dh(ρ1), Dh(ρ2), ..., Dh(ρl)), h(t) ≥Pj aijDh(ρj) and lim inf t→1 Di (2) lim inf t→0 Di It is straight forward to verify that the image of the map h ∈ (A ⊗ K)+ → Dh is the subset of Map(Irr(A), Z+ ∪ {∞}) consisting elements satisfying the above two conditions. Note that Dh(π) = rank(π ⊗ idK(h)) for each π ∈ Irr(A) and h ∈ (A ⊗ K)+. The following result is well known to experts (for example, see [11]). h(t) ≥Pj bijDh(ρl). Theorem 3.18. Let A = A(F1, F2, ϕ0, ϕ1) ∈ C and let n ≥ 1 be an integer. (a) The following are equivalent: (1) h ∈ Mn(A)+ is Cuntz equivalent to a projection; (2) 0 is an isolate point in the spectrum of h; (3) Dj h continuous on each (0, 1)j , lim inf t→0 Di h. (b) For h1, h2 ∈ Mn(A)+, h1 . h2 if and only if Dh1(π) ≤ Dh2(π) for each π ∈ Irr(A). In particular, A has strictly comparison for positive elements. h and lim inf t→1 Di h(t) =Pj aijrj h(t) =Pj bijrj Proof. For part (b), obviously, h1.h2 implies Dh1(ϕ) ≤ Dh2(ϕ) for each ϕ ∈ Irr(A). Conversely assume that h1 = (f, a) and h2 = (g, b) satisfy that Dh1(ϕ) ≤ Dh2(ϕ) for each ϕ ∈ Irr(A). First, there are strictly positive functions s1, s2 ∈ C0((0, 1]) such that s1(a) and s2(b) are projections in F1. Note that si(hi) are Cuntz equivalent to hi (i = 1, 2). By replacing hi by si(hi), without loss of generality we may assume that a and b are projections. 19 Fix 1/4 > ε > 0. There exists δ1 > 0 such that kf (t) − ϕ0(a)k < ε/64 for all t ∈ [0, 2δ1] andkf (t) − ϕ1(a)k < ε/16 for all t ∈ [1 − 2δ1, 1].(e 3.7) It follows that fε/8(f (t)) is a projection in [0, 2δ1] and [1 − 2δ1, 1]. Put h0 = fε/8(h1). Note that h0 = (fε/8(f ), a). Then Di (π) is constant in (0, 2δ1] ⊂ (0, 1)i and [1 − 2δ1, 1] ⊂ (0, 1)i, h0 i = 1, 2, ..., k. Choose δ2 > 0 such that Di h2(t) ≥ dT ri(ϕ0(b)) for all t ∈ (0, 2δ2]i and Dh2(t) ≥ dT r(ϕ1(b)) for all t ∈ [2δ2, 1)i,(e 3.8) where T ri is the standard trace on MRi, i = 1, 2, ..., k. Choose δ = min{δ1, δ2}. Since a . b in F1, there is a unitary ue ∈ F1 such that u∗eaue = q ≤ b, where q ≤ b is a sub-projection. Let u0 = ϕ0(ue) and u1 = ϕ1(ue) ∈ F2. Then one can find a unitary u ∈ A such that u(0) = u0 and u(1) = u1. Without loss of generality, by replacing h0 by u∗h0u, we may assume that a = q. Let β : [0, 1] → [0, 1] be a continuous function which is 1 on the boundary and 0 on [δ, 1− δ]. Let h3 = (f2, b− a) with f2(t) = β(t)ϕ0(b− a) for t ∈ [0, δ], f2(t) = β(t)ϕ1(b− a) for t ∈ [1− δ, δ], and f2(t) = 0 for t ∈ [δ, 1 − δ]. Define h′1 = h0 + h3. Note that h′1 has the form (f′, b) for some f′ ∈ C([0, 1], F2). Let πi : F2 → MRi be the projection map. Then h0 ≤ h′1, Di h′ 1 (π) = Di h0(π) for all π ∈ (δ, 1 − δ) ⊂ (0, 1)i rank(h′1(π)) ≤ rank(h0(t)) + rank(h3(t)) ≤ rank(πi(ϕ0(a))) + rank(πi(ϕ0(b − a))) h2(π) for all π ∈ (0, δ) ⊂ (0, 1)i and rank(h′1(π)) ≤ rank(h0(t)) + rank(h3(t)) ≤ rank(πi(ϕ1(a))) + rank(πi(ϕ1(b − a))) h2(π) for all π ∈ (1 − δ, 1) ⊂ (0, 1)i. = rank(πi(ϕ1(b))) ≤ Di = rank(πi(ϕ0(b))) ≤ Di It follows that Dh′ 1 (π) ≤ Dh2(π) for all π ∈ Irr(A). (e 3.9) (e 3.10) (e 3.11) (e 3.12) (e 3.13) (e 3.14) It follows from (e 3.14) and Theorem 1.1 of [91] that h′1 . h2 in C([0, 1], F2). Since C([0, 1], F2) has stable rank one, there is a unitary w ∈ C([0, 1], F2) such that w∗h′1w = h4 ∈ h2C([0, 1], F2)h2. Since h′1 = (f′, b), h2 = (g, b) and b is a projection, h4(0) = ϕ0(b) and h4 = ϕ1(b). In other words, h4 ∈ A. In particular, h4 ∈ h2Ah2 and h4 . h2. Note that w∗ϕi(b)w = ϕi(b), i = 0, 1. Using a continuous path of unitaries which commutes ϕi(b) (i = 0, 1) and connects to the identity, it is easy to find a sequence of unitaries un ∈ C([0, 1], F2) such that un(0) = un(1) = 1F2 such that (e 3.15) u∗nh′1un = h4. lim n→∞ Then un ∈ A. We also have that unh4u∗n → h′1. Thus h′1 . h4. It follows that fε(h1) . h0 . h′1 . h4 . h2 (e 3.16) for all ε > 0. This implies that h1 . h2 and part (b) follows. To prove part (a), we note that (1) and (2) are obviously equivalent and both implies (3). That of (3) implies (1) follows from the computation of K0(A) in 3.5 and part (b). Lemma 3.19. Let C ∈ C, and let p ∈ C be a projection. Then pCp ∈ C. Moreover, if p is full and C ∈ C0, then pCp ∈ C0. Proof. We may assume that C is not of finite dimensional. Write C = C(F1, F2, ϕ0, ϕ1). Denote by pe = πe(p), where πe : C → F1 is the map defined in 3.1. 20 For each t ∈ [0, 1], write πt(p) = p(t) and p ∈ C([0, 1], F2) such that πt(p) = p(t) for all t ∈ [0, 1]. Then ϕ0(pe) = p(0), ϕ1(pe) = p(1), and pCp = {(f, g) ∈ C : f (t) ∈ p(t)F2p(t), and g ∈ p1F1p1}. (e 3.17) Put p0 = p(0). There is a unitary W ∈ C([0, 1], F2) such that W ∗ pW = p0. Define Φ : pC([0, 1], F2)p → C([0, 1], p0F2p0) by Φ(f ) = W ∗f W for all f ∈ pC([0, 1], F2)p. Put F ′1 = p1F1p1 and F ′2 = p0F2p0. Define ψ0 = Ad W (0) ◦ ϕ0F ′ and ψ1 = Ad W (1) ◦ ϕ1F ′ . Put 1 1 C1 = {(f, g) ∈ C([0, 1], F ′2) ⊕ F ′1 : f (0) = ψ0(g) and f (1) = ψ1(g)}, and note that C1 ∈ C. Define Ψ : pCp → C1 by Ψ((f, g)) = (Φ(f ), g) for all f ∈ pC([0, 1], F2)p and g ∈ F ′1. (e 3.18) It is ready to verify that Ψ is an isomorphism. If p is full and C ∈ C0, then, by a result of Brown ([6]), the hereditary C∗-subalgebra pCp is stably isomorphic to C, and hence K1(pCp) = K1(C) = {0}; that is, pCp ∈ C0. Class C and C0 are not closed under quotient. However, we have the following: Lemma 3.20. Any quotient of a C-algebra in C (or in C0) can be locally approximated by C-algebras in C (or in C0). More precisely, let A ∈ C (or A ∈ C0), let B be a quotient of A, let F ⊂ B be a finite set and let ε > 0, there exists a unital C∗-subalgebra B0 ⊂ B with B0 ∈ C (or B0 ∈ C0) such that dist(x, B0) < ε for all x ∈ F. Proof. Let A ∈ C. We may consider only those A’s which are not finite dimensional. Let I be an ideal of A. Write A = A(E, F, ϕ0, ϕ1), where E = E1 ⊕ ··· ⊕ El, where Ei ∼= Mki, and F = F1 ⊕ ··· ⊕ Fs with Fj ∼= Mmj Let J = {f ∈ C([0, 1], F ) : f (0) = f (1) = 0} ⊂ A. As before, we may write [0, 1]j for the the spectrum of the j-th summand of C([0, 1], Fj ), whenever it is convenient. Put ϕi,j = πj ◦ ϕi : E → Fj, where πj : F → Fj is the quotient map, i = 0, 1. Then A/I may be written (with a re-indexing) as C( Ij, Fj), f (0j ) = ϕ0,j(a), if 0j ∈ Ij, f (1j) = ϕ1,j(a), if 1j ∈ Ij}, {(f, a) : a ∈ E, f ∈ M1≤j≤s′ where s′ ≤ s, l ≤ l′, E =Ll′ i=1 Ei and ϕi,j = ϕi,j E (l′ ≤ l) and Ij ⊂ [0, 1]j is a compact subset. It follows from [34] that there is a sequence of Xn,j which is a finite disjoint union of closed intervals (includes points) such that Ij is an inverse limit of Xn,j and each map sn,j : Ij → Xn,j is surjective. Moreover Xn,j can be identified with disjoint union of closed subintervals of [0, 1]. With this identification, we may further assume that sn,j(0) = 0, if 0 ∈ Ij and sn,j(1) = 1, if 1 ∈ Ij. Indeed, if Ij contains an interval [0, d′n,j], we may assume that one of subintervals of Xn,j is an interval with the form [0, dn,j] and sn,j(0) = 0. Otherwise, there exist two sequences δn,j,2 > δn,j,1 > 0 with limn→∞ δn,j,2 = 0 such that Ij ∩ (δn,j,1, δn,j,2) = ∅. By redefining sn,j so that sn,j( Ij ∩ [0, δn,j,1]) = {0} and keeping sn,j Ij∩[δn,j,2,1], we can assume that sn,j(0) = 0. This can also be done at 1. Write C( Ij, Fj ) = ∪∞n=1(C(Xn,j, Fj)). Let sn :Ls′ j=1 C(Xn,j, Fj ) → Ls′ j=1 C( Ij, Fj ) be the map induced by sn,j. Put Cj = C(Xn,j, Fj ). Then, for all sufficiently large n, for each f ∈ F, there is g ∈Ls′ kf Ij − s∗n,j(gXn,j )k < ε/4, g(0j ) = f (0j), if 0j ∈ Ij and g(1j ) = f (1j), if 1j ∈ Ij. (e 3.19) j=1 Cj such that 21 Note that Cj is a unital C∗-subalgebra of C( Ij, Fj), j = 1, 2, ..., s. Define C = {(f, a) : f ∈ s′Mj=1 Cj, f (sn,j(0j)) = ϕ0,j(a), if 0j ∈ Ij and f (sn,j(1j)) = ϕ1,j(a)}. Then g in (e 3.19) is in C and F ⊂ε C. Moreover, C ∈ C and C is isomorphic to a quotient of A. This proves the lemma in the case that A ∈ C. Now suppose that A ∈ C0. Since C is isomorphic to a quotient of A, it suffices to show that, for any ideal I ⊂ A, if K1(A) = {0}, then K1(A/I) = {0}. To this end, we consider the following six-term exact sequence: K0(I) / K0(A) / K0(A/I) δ1 K1(A/I) K1(A) K1(I). By 3.3, A and A/I has stable rank one, it follows from Proposition 4 of [78] that δ1 = 0. Since K1(A) = {0}, it follows that K1(A/I) = {0}. Lemma follows. As the end of this section, we would like to return to the beginning of this section by stating the following proposition which will not be used. Proposition 3.21 (Theorem 2.15 of [31]). Let A be a unital C∗-algebra which is a subhomoge- neous C∗-algebra with one dimensional spectrum. Then, for any finite subset F ⊂ A and any ε > 0, there exists unital C∗-subalgebra B of A which is in C such that Proof. We use the fact that A is an inductive limit of C∗-algebras in C ([31]). Therefore, there is C∗-algebra C ∈ C and a unital homomorphism ϕ : C → A such that dist(x, B) < ε for all x ∈ F. Then we apply Lemma 3.20. dist(x, ϕ(C)) < ε/2 for all x ∈ F. 4 Maps to finite dimensional C∗-algebras i=1 ai · zi and b =Pn Lemma 4.1. Let z1, z2, ..., zn be positive integers which may not be distinct. There is a positive integer T = n · max{zizj : 1 ≤ i, j ≤ n} such that for any two nonnegative integer linear i=1 a′i · zi i=1 bi · zi, there are two combinations a′ =Pn combinations a =Pn and b′ =Pn i=1 b′i · zi with a′ = b′, 0 ≤ a′i ≤ ai, 0 ≤ b′i ≤ bi, and min{a − a′, b − b′} ≤ T . Consequently, if δ > 0 and a − b < δ, we also have that max{a − a′, b − b′} < δ + T. Proof. To prove the first part, let T = n · maxi,j{zizj}. It is enough to prove that if a, b > T, then there are nonzero 0 < a′ =Pn i=1 b′i · zi with 0 ≤ a′i ≤ ai, 0 ≤ b′i ≤ bi. But if both a > T and b > T , then there are two (not necessary distinct) index i, j, with ai ≥ zj and bj ≥ zi. Then choose a(1)′ = b′jzj with a′i = zj and b′j = zi. If , b − b(1)′} ≤ T. Then we are done. If not, we repeat this on a − a(1) and b − b(1) min{a − a(1)′ and obtain a(2) ≤ a − a(1) and b(2) ≤ b − b(1) such that a(2) = b(2). Put a(2)′ + a(2) = Pi a(2) and b(2)′ i zi and =Pi b(2) , b− b(2)′} ≤ T, then b(2)′ we are done. Otherwise, we continue. An inductive argument shows the first part of lemma follows. To see the second part, assume that a − a′ ≤ T. Then b − b′ < a − b + T. and we can also have a(2)′ = b(2)′ i ≤ b′i for all i. If min{a− a(2)′ i=1 a′i · zi = b′ =Pn = a′izi and b(1)′ = a(1)′ + b(2). Note we have a(2)′ = b(1)′ i zi with 0 ≤ a(2) i ≤ ai and 0 ≤ b(2) 22 / / O O o o o o Theorem 4.2. (see 2.10 of [74], Theorem 4.6 of [62] and 2.15 of [50]) Let X be a connected compact metric space, and let C = C(X). Let F ⊆ C be a finite set, and let ǫ > 0 be a constant. There is a finite set H1 ⊆ C + such that for any σ1 > 0 there is a finite subset H2 ⊆ C and σ2 > 0 such that for any unital homomorphisms ϕ, ψ : C → Mn for a matrix algebra Mn satisfying (1) ϕ(h) > σ1 and ψ(h) > σ1 for any h ∈ H1, and (2) tr ◦ ϕ(h) − tr ◦ ψ(h) < σ2 for any h ∈ H2, then there is a unitary u ∈ Mn such that kϕ(f ) − u∗ψ(f )uk < ǫ for any f ∈ F. We would also like to state another version of the above theorem. Theorem 4.3. Let A = C(X), where X is a compact metric space and let ∆ : Aq,1 be an order preserving map. + \{0} → (0, 1) For any ε > 0, any finite set F ⊂ A, there exist a finite set P ⊂ K(A), a finite set H1 ⊂ A1 + \ {0}, a finite set H2 ⊂ As.a. and δ > 0 satisfying the following: If ϕ1, ϕ2 : A → Mn (for some integer n ≥ 1) are two unital homomorphisms such that [ϕ1]P = [ϕ2]P , τ ◦ ϕ1(h) ≥ ∆(h) for all h ∈ H1 and τ ◦ ϕ1(g) − τ ◦ ϕ2(g) < δ for all g ∈ H2, then, there exist a unitary u ∈ Mn such that kAd u ◦ ϕ1(f ) − ϕ2(f )k < ε for all f ∈ F. (e 4.1) Remark 4.4. Let X be a compact metric space and let A = C(X). Suppose that ϕ1, ϕ2 : A → Mn are two homomorphisms. Then (ϕi)∗1 = 0, i = 1, 2. Let P ⊂ K(A) be a finite subset and let G be the subgroup generated by P. There exists a finite CW complex Y and a unital homomorphism h : C(Y ) → C(X) such that G ⊂ [h](K(C(Y ))). Write K0(C(Y )) = Zk ⊕ kerρC(Y ), where Zk is generated by mutually orthogonal projections {p1, p2, ..., pk} which correspond to k different path connected components Y1, Y2, ..., Yk of Y. Fix ξi ∈ Yi, Let Ci = C0(Yi \ {ξi}), i = 1, 2, ..., k. Since Yi is path connected, by considering the point-evaluation at ξi, it is easy to see that, for any ϕ : C(Y ) → Mn, [ϕ]K(Ci) = 0. Let ¯G = [h](K(C(Y ))). Suppose that τ ◦ ϕ1(pi) = τ ◦ ϕ2(pi), i = 1, 2, ..., k. Then, from the above, one computes that [ϕ1]P = [ϕ2]P . (e 4.2) We will use this fact in the next proof. Lemma 4.5. Let X be a compact metric space, let F be a finite dimensional C∗-algebra and let A = P C(X, F )P, where P ∈ C(X, F ) is a projection. Let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. For any ε > 0, any finite subset F ⊂ A and any σ > 0, there exists a finite subset H1 ⊂ A1 + \ {0}, a finite subset H2 ⊂ As.a. and δ > 0 satisfying the following: If ϕ1, ϕ2 : A → Mn (for some integer n ≥ 1) are two unital homomorphisms such that τ ◦ ϕ1(h) ≥ ∆(h) for all h ∈ H1, and τ ◦ ϕ1(g) − τ ◦ ϕ2(g) < δ for all g ∈ H2, 23 then, there exist a projection p ∈ Mn, a unital homomorphism H : A → pMnp, unital homo- morphisms h1, h2 : A → (1 − p)Mn(1 − p) and a unitary u ∈ Mn such that kAd u ◦ ϕ1(f ) − (h1(f ) + H(f ))k < ε, kϕ2(f ) − (h2(f ) + H(f )k < ε for all f ∈ F, and τ (1 − p) < σ, where τ is the tracial state of Mn. Proof. We first prove the case that A = C(X). Let ∆1 = (1/2)∆. Let P ⊂ K(A) be a finite set, H′1 ⊂ A1 + \ {0} (in place of H1) be a finite set, H′2 ⊂ As.a. (in place of H2) be a finite set and δ1 > 0 (in place of δ) required by 4.3 for ε/2 (in place of ε), F and ∆1. + \ {0}. So, in what follows, H′2 ⊂ A1 Without loss of generality, we may assume that 1A ∈ F, 1A ∈ H′1 ⊂ H′2 and H′2 ⊂ A1 + \ {0}. Put σ0 = min{∆1(g) : g ∈ H′2}. (e 4.3) Let G be the subgroup generated by P and let ¯G be defined in 4.4. Let P0 be a set of generators of ¯G∩ K0(A). Without loss of generality, we may assume that P0 = {p1, p2, ..., pk1}∪ {z1, z2, ..., zk2}, where pi ∈ C(X) are projections (corresponding to clopen subsets) and zj ∈ kerρA(K0(A)). Without loss of generality, we may assume that {pi : 1 ≤ i ≤ k1} is a set of mutually orthogonal projections such that 1A =Pk1 Let H1 = H′1 ∪ {pi : 1 ≤ i ≤ k1} and H2 = H′2 ∪ H1. Let σ1 = min{∆(g) : g ∈ H2}. Choose Suppose now that ϕ1, ϕ2 : A → Mn are two unital homomorphisms described in the lemma k=1 f (xk,j)qk,j for all f ∈ C(X), where {qk,j : 1 ≤ k ≤ n} (j = 1, 2) δ = min{σ0 · σ/4k1, σ0 · δ1/4k1, σ1/16k1}. for the above H1, H2 and ∆. is a set of mutually orthogonal rank one projections and xk,j ∈ X. We have We may write ϕj(f ) =Pn i=1 pi. τ ◦ ϕ1(pi) − τ ◦ ϕ2(pi) < δ, i = 1, 2, ..., k1, (e 4.4) where τ is the tracial state on Mn. Therefore, there exists a projection P0,j ∈ Mn such that τ (P0,j) < k1δ < σ0 · σ, j = 1, 2, (e 4.5) rank(P0,1) = rank(P0,2), unital homomorphisms ϕ1,0 : A → P0,1MnP0,1, ϕ2,0 : A → P0,2MnP0,2, ϕ1,1 : A → (1 − P0,1)Mn(1 − P0,1) and ϕ1,2 : A → (1 − P0,2)Mn(1 − P0,2) such that ϕ1 = ϕ1,0 ⊕ ϕ1,1, ϕ2 = ϕ2,0 ⊕ ϕ2,1, τ ◦ ϕ1,1(pi) = τ ◦ ϕ1,2(pi), i = 1, 2, ..., k1. (e 4.6) (e 4.7) By replacing ϕ1 by Ad v◦ϕ1, simplifying the notation, without loss of generality, we may assume that P0,1 = P0,2. It follows from (e 4.7) that (see 4.4) By (e 4.5) and the choice of σ0, we also have [ϕ1,1]P = [ϕ2,1]P . τ ◦ ϕ1,1(g) ≥ ∆1(g) for all g ∈ H′1 and τ ◦ ϕ1,1(g) − τ ◦ ϕ1,2(g) < σ0 · δ1 for all g ∈ H′2. 24 (e 4.8) (e 4.9) (e 4.10) Therefore t ◦ ϕ1,1(g) ≥ ∆1(g) for all g ∈ H′1 and t ◦ ϕ1,1(g) − t ◦ ϕ1,2(g) < δ1 for all g ∈ H′2, (e 4.11) (e 4.12) where t is the tracial state on (1 − P1,0)Mn(1 − P1,0). By applying 4.3, there exists a unitary v1 ∈ (1 − P1,0)Mn(1 − P1,0) such that kAd v1 ◦ ϕ1,1(f ) − ϕ2,1(f )k < ε/16 for all f ∈ F. (e 4.13) Put H = ϕ2,1 and p = P1,0. The lemma for the case that A = C(X) follows. For the case that A = Mr(C(X)), let e1 ∈ Mr be a rank one projection. Put B = e1(Mr(C(X)))e1 ∼= C(X). Consider ψ1 = ϕ1B and ψ2 = ϕ2B. Since ϕ1 and ϕ2 are unital, rank(ψ1(1B)) = rank(ψ2(1B)). By replacing ψ1 by Ad v ◦ ψ1 for some unitary v ∈ Mn, we may assume that ψ1(1B) = ψ2(1B). Then what has been proved could be applied to ψ1 and ψ2. The case of A = Mn(C(X)) then follows. For the case A = C(X, F ), let F = Mr1 ⊕ Mr2 ⊕ ··· ⊕ Mrk . Denote by Ei the projection of A corresponding to the identity of Mri, i = 1, 2, ..., k. The same argument used the above shows that we can find a projection Qj ∈ Mn (j = 1, 2), unital homomorphisms ψj,0 : A → QjMnQj and unital homomorphisms ψj,1 : A → (1 − Qj)Mn(1 − Qj) such that rank(Q1) = rank(Q2), τ (Qj) < δ, τ (ψj,1(Ei)) = τ (ψj,2(Ei)), i = 1, 2 and ϕj = ψj,0 ⊕ ψj,1, (e 4.14) (e 4.15) j = 1, 2. By replacing ϕ1 by Ad v ◦ ϕ1 for a suitable unitary v ∈ Mn, we may assume Q1 = Q2 and ψ1,1(Ei) = ψ2,1(Ei), i = 1, 2, ..., k. Thus this case has been reduced to the case that A = Mr(C(X)). Therefore the case that A = C(X, F ) also follows from the above proof. Let us consider the case that A = P C(X, F )P. Note that {rank(P (x)) : x ∈ X} is a finite set of positive integers. Therefore the set Y = {x ∈ X : rank(P (x)) > 0} is compact and open. Then we may write A = P C(Y, F )P. Thus we may assume that P (x) > 0 for all x ∈ X. This implies easily that P is a full projection of C(Y, F ). Then, by [6], A⊗K ∼= C(Y, F )⊗K. It follows that there is an integer m ≥ 1 and a projection e ∈ Mm(A) such that eMm(A)e ∼= C(X, F ). By extending ϕ1 and ϕ2 to maps from eMm(A)e, one easily sees that the proof reduces to the case that A = C(X, F ). Corollary 4.6. Let X be a compact metric space, let F be a finite dimensional C∗-algebra and let A = P C(X, F )P, where P ∈ C(X, F ) is a projection. Let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. Let 1 > α > 1/2. + \ {0} and any integer K ≥ 1. There is an integer N ≥ 1, a finite subset H1 ⊂ A1 + \ {0}, a finite subset H2 ⊂ As.a., δ > 0 satisfying the following: If ϕ1, ϕ2 : A → Mn (for any integer n ≥ N ) are two unital homomorphisms such that For any ε > 0, any finite subset F ⊂ A, any finite subset H0 ⊂ A1 τ ◦ ϕ1(h) ≥ ∆(h) for all h ∈ H1, and τ ◦ ϕ1(g) − τ ◦ ϕ2(g) < δ for all g ∈ H2, then, there exist mutually orthogonal nonzero projections e0, e1, e2, ..., eK ∈ Mn such that e1, e2, ..., eK i=1 ei = 1Mn, and there are unital homomorphisms h1, h2 : are equivalent, e0 . e1 and e0 +PK 25 A → e0Mne0, ψ : A → e1Mne1 and a unitary u ∈ Mn such that kAd u ◦ ϕ1(f ) − (h1(f ) + diag( ψ(f ), ψ(f ), ..., ψ(f )))k < ε, kϕ2(f ) − (h2(f ) + diag( ∆(g) and τ ◦ ψ(g) ≥ α where τ is the tracial state of Mn. K z ψ(f ), ψ(f ), ..., ψ(f )))k < ε for all f ∈ F, for all g ∈ H0, z K } K } { { Proof. By applying 4.5, it is easy to see that it suffices to prove the following statement: Let X, F, P A and α be as in the corollary. Let ε > 0, let F ⊂ A be a finite subset, let H0 ⊂ A1 N ≥ 1, a finite subset H1 ⊂ A1 some n ≥ N ) is a unital homomorphism such that + \{0} and let K ≥ 1. There is an integer + \ {0} satisfying the following: Suppose that H : A → Mn (for τ ◦ H(g) ≥ ∆(g) for all g ∈ H0. (e 4.16) Then there are mutually orthogonal projections e0, e1, e2, ..., eK ∈ Mn, a unital homomorphism ϕ : A → e0Mne0 and a unital homomorphism ψ : A → e1Mne1 such that kH(f ) − (ϕ(f ) ⊕ diag( ψ(f ), ψ(f ), ..., ψ(f )))k < ε for all f ∈ F, τ ◦ ψ(g) ≥ α∆(g)/K for all g ∈ H0. z K } { (e 4.17) (e 4.18) We make one further reduction: Using the argument at the end of the proof of 4.5, it suffices to prove the above statement for A = C(X). Put σ0 = ((1 − α)/4) min{∆(g) : g ∈ H0} > 0. Let ε1 = min{ε/16, σ0} and let F1 = F ∪ H0. Choose d0 > 0 such that f (x) − f (x′) < ε1 for all f ∈ F1, (e 4.19) (e 4.20) provided that x, x′ ∈ X and dist(x, x′) < d0. Choose ξ1, ξ2, ..., ξm ∈ X such that ∪m dist(x, ξ) < r}. There is d1 > 0 such that d1 < d0/2 and B(ξj, d1) ∩ B(ξi, d1) = ∅, j=1B(ξj, d0/2) ⊃ X, where B(ξ, r) = {x ∈ X : (e 4.21) if i 6= j. There is, for each j, a function hj ∈ C(X) with 0 ≤ hj ≤ 1, hj(x) = 1 if x ∈ B(ξj, d1/2) and hj(x) = 0 if x 6∈ B(ξj, d1). Define H1 = H0 ∪ {hj : 1 ≤ j ≤ m} and put σ1 = min{∆(g) : g ∈ H1}. (e 4.22) Choose an integer N0 ≥ 1 such that 1/N0 < σ1 · (1 − α)/4 and N = 4m(N0 + 1)2(K + 1)2. (e 4.16). Let Y1 = B(ξ1, d0/2) \ ∪m B(ξj, d0/2) \ (∪j−1 B(ξj, d1) ⊂ Yj. We write that Now let H : C(X) → Mn be a unital homomorphism with n ≥ N satisfying the assumption i=3B(ξi, d1), Yj = i=j+1B(ξi, d1)), j = 1, 2, ..., m. Note that Yj ∩ Yi = ∅ if i 6= j and i=2B(ξi, d1), Y2 = B(ξ2, d0/2) \ (Y1 ∪ ∪m i=1 Yi ∪ ∪m H(f ) = nXi=1 f (xi)pi = mXj=1 (Xxi∈Yj f (xi)pi) for all f ∈ C(X), (e 4.23) 26 where {p1, p2, ..., pn} is a set of mutually orthogonal rank one projections in Mn, {x1, x2, ..., xn} ⊂ X. Let Rj be the cardinality of {xi : xi ∈ Yj}. Then, by (e 4.16), Rj ≥ N τ ◦ H(hj ) ≥ N ∆( hj) ≥ (N0 + 1)2Kσ1 ≥ (N0 + 1)K 2, j = 1, 2, ..., m. (e 4.24) Write Rj = SjK+rj, where Sj ≥ N0Km and 0 ≤ rj < K, j = 1, 2, ..., m. Choose xj,1, xj,2, ..., xj,rj ⊂ {xi ∈ Yj} and denote Zj = {xj,1, xj,2, ..., xj,rj}, j = 1, 2, ..., m. Therefore we may write H(f ) = mXj=1 ( Xxi∈Yj\Zj f (xi)pi) + mXj=1 ( rjXi=1 f (xj,i)pj,i) (e 4.25) for f ∈ C(X). Note that the cardinality of {xi ∈ Yj \ Zj} is KSj, j = 1, 2, ..., m. Define ( KXk=1 mXj=1 mXj=1 pi =PK Ψ(f ) = f (ξj)Pj = f (ξj)Qj,k) for all f ∈ C(X), (e 4.26) j=1 Qj,k, k = 1, 2, ..., K. Note that k=1 Qj,k and rank Qj,k = Sj, j = 1, 2, ..., m. Put e0 =Pm i=1(Prj i=1 pj,i), where Pj =Pxi∈Yj\Zj ek =Pm rank(e0) = rj < mK and rank(ek) = Sj mXj=1 z z K } K } { { (e 4.27) (e 4.28) (e 4.29) (e 4.30) (e 4.31) It follows that e0 . e1 and ei is equivalent to e1. Moreover, we may write Sj ≥ N0mK > mK, j = 1, 2, .., K. Ψ(f ) = diag( ψ(f ), ψ(f ), ..., ψ(f )) for all f ∈ A, j=1 f (ξj)Qj,1 for all f ∈ A. We also estimate that kH(f ) − (ϕ(f ) ⊕ diag( ψ(f ), ψ(f ), ..., ψ(f )))k < ε1 for all f ∈ F1. where ψ(f ) =Pm We also compute that τ ◦ ψ(g) ≥ (1/K)(∆(g) − ε1 − mK N0Km ) ≥ α ∆(g) K for all g ∈ H0. Remark 4.7. If we also assume that X has infinitely many points, then Lemma 4.6 holds without mentioning the integer N. This can be seen by taking larger H1 which will force the integer n large. Definition 4.8. Denote by ¯D′0 the class of all finite dimensional C∗-algebras. For k ≥ 1, denote by ¯D′k the class of all C∗-algebras with the form: A = {(f, a) ∈ P C(X, F )P ⊕ B : fZ = Γ(a)}, where X is a compact metric space, F is a finite dimensional C∗-algebra, P ∈ C(X, F ) is a projection, Z ⊂ X is a nonempty proper subset of X, B ∈ ¯D′k−1 and Γ : B → P C(Z, F )P is 27 a unital homomorphism, where we assume that there is dX,Z > 0 (we will denote it by dX if there is no danger of confusion) such that, for any 0 < d ≤ dX , there exists sd ∗ : X d → Z such that sd ∗(x) = x for all x ∈ Z and lim d→0kfZ ◦ sd ∗ − fX dk = 0 for all f ∈ C(X, F ), (e 4.32) where X d = {x ∈ X : dist(x, Z) < d}. We also assume that, for any 0 < d < dX /2 and for any d > δ > 0, there is a homeomorphism r : X \ X d−δ → X \ X d such that dist(r(x), x) < δ for all x ∈ X \ X d−δ. (e 4.33) For the convenience, we will also assume, in addition, that F = Ms(1)L Ms(2) ···L Ms(k1), P C(X, F )P = PjC(X, Ms(j))Pj , k1Mj=1 where each Pj has a constant rank r(j) > 0. Let Am be a unital C∗-algebra in ¯D′m. For 0 ≤ k < m, let Ak ∈ ¯D′k such that Ak+1 = {(f, a) ∈ Pk+1C(Xk+1, Fk+1)Pk+1 ⊕ Ak : fZk+1 = Γk+1(a)}, where Fk+1 is a finite dimensional C∗-algebra, Pk+1 ∈ C(Xk+1, Fk+1) is a projection, Γk+1 : Ak → Pk+1C(Zk+1, Fk+1)Pk+1 is a unital homomorphism, k = 0, 1, 2, ..., m − 1. Denote by ∂k+1 : Pk+1C(Xk+1, Fk+1)Pk+1 → Qk+1C(Zk+1, Fk+1)Qk+1 the map defined by f 7→ fZk+1 (Qk+1 = Pk+1Zk+1). We use π(k+1) : Ak+1 → Ak for the quotient map and λk+1 : Ak+1 → Pk+1C(Xk+1, Fk+1)Pk+1 for the map defined by (f, a) 7→ f. For each k, one has the following commutative diagram: e /❴❴❴❴❴❴❴ λk Ak PkC(Xk, Fk)Pk π(k+1) e Ak−1 Γk ∂k / QkC(Zk, Fk)Qk, (e 4.34) In general, suppose that A = Am ∈ ¯D′m is constructed in the following sequence A0 = F0, A1 = A0⊕Q1C(Z1,F1)Q1P1C(X1, F1)P1, A2 = A1⊕Q2C(Z2,F2)Q2P2C(X2, F2)P2,··· , Am = Am−1 ⊕QmC(Zm,Fm)Qm PmC(Xm, Fm)Pm, where Qi = PiZi, i = 1, 2, ..., m. With π(k+1) A = Am → Ak and homomorphism Λk : A = Am → PkC(Xk, Fk)Pk as follows: e and λk above we can define quotient map Πk : Πk = π(k+1) e ◦ π(k+2) e ◦ ··· ◦ π(m−1) e ◦ π(m) e and Λk = λk ◦ Πk. Combining all Λk we get the inclusion homomorphism Λ : A → mMk=0 PkC(Xk, Fk)Pk with X0 being single point set. For each k ≥ 1, we may write Pk(C(Xk, Fk))Pk = Pk,1C(Xk, Ms(k,1))Pk,1⊕Pk,2C(Xk, Ms(k,2))Pk,2⊕··· Pk,tk C(Xk, Ms(k,tk))Ptk , 28 / ✤ ✤ ✤ / where Pk,j ∈ C(Xk, Ms(k,j)) is a projection of rank r(k, j) at each x ∈ Xk. For each x ∈ Xk and j ≤ tk and f ∈ A, denoted by π(x,j)(f ) ∈ Mr(k,j), the evaluation of jth component of Λk(f ) at point x. Then for each pair (x, j), π(x,j) is a finite dimensional representation of A, and furthermore if we assume x ∈ Xk \ Zk, then π(x,j) is an irreducible representation. For each integer m ≥ 0, denote by ¯Dm the class of C∗-algebras which are finite direct sums of those C∗-algebras in ¯D′m. If in the above definition of ¯D′k, in addition, we assume that X is path connected, Z has finitely many path connected components and Xi \ Zi is path connected, then we will use D′k for the resulting class of C∗-algebras and we use Dk for the class of C∗-algebras which are finite direct sums of those C∗-algebras in D′m. Note that Dk ⊂ ¯Dk and C ⊂ D1. Remark 4.9. Let A ∈ ¯Dk (or A ∈ Dk). It is easy to check that C(T)⊗A ∈ ¯Dk+1 (or A ∈ Dk+1). First of all it is easy to check that if F0 is a finite dimensional C∗ algebra, then C(T)⊗ F0 ∈ ¯D1 by putting F1 = F0 and X1 = T with Z1 = {1} ⊂ T and Γ1 : F0 → C(Z1, F1) ∼= F0 to be the identity map. And if a pair of spaces (Xk, Zk) satisfies the condition described as pair (X, Z) in the above definition, in particular, the existence of the retraction sd ∗ and homeomorphism r as in e 4.32 and e 4.33, then the pair (Xk × T, Zk × T) also satisfies the same condition. Remark 4.10. Let A ∈ Dk be a unital C∗-algebra. There is an integer l ≥ 1 such that (K0(A))/ ker(ρA) ⊂ Zl as an order group, where Zl is equipped with the usual order. This has been proved for k = 0, 1. Now assume that k > 1 and A ∈ Dk. Then the above statement holds for any unital C∗-algebra in Dd with 0 ≤ d < k. Consider the short exact sequence 0 → I →j A → C →πe 0, (e 4.35) where C ∈ Dd for some d < k, I = P C0(X \ Z, F ))P and j : I → A is the embedding. Note ρA ◦ j∗0 = 0, since X is path connected and Z ⊂ X is non empty. It follows that ρA factors through (πe)∗0. Hence the map K0(A)/ ker(ρA) → K0(C)/ ker(ρC) is injective and consequently the statement holds for A. In particular for A ∈ Dk as notation in 4.8, we have K0(A)/ ker(ρA) ⊂ (K0(A0))/ ker(ρA0), K0(A)/ ker(ρA) ∼= (Π0)∗0(K0(A)) ⊂ K0(A0) ∼= Zl.(e 4.36) Lemma 4.11. Let A ∈ ¯Dk be a unital C∗-algebra. Let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. Let 1 > α > 1/2. + \ {0} be a finite subset and K ≥ 1 be an + \ {0}, a finite subset integer. There exist an integer N ≥ 1, δ > 0, a finite subset H1 ⊂ A1 H2 ⊂ As.a. satisfying the following: Let ε > 0, F ⊂ A be a finite subset, H0 ⊂ A1 If ϕ1, ϕ2 : A → Mn(for some integer n ≥ N ) are two unital homomorphisms such that τ ◦ ϕ1(g) ≥ ∆(g) for all g ∈ H1 and τ ◦ ϕ1(g) − τ ◦ ϕ2(g) < δ for all g ∈ H2, (e 4.37) (e 4.38) Mn such that e1, e2, ..., eK are mutually equivalent, e0 . e1, and e0 +PK where τ is the tracial state on Mn, then there exist mutually orthogonal projections e0, e1, e2, ..., eK ∈ i=1 ei = 1Mn, unital ho- momorphisms h1, h2 : A → e0Mne0, a unital homomorphism ψ : A → e1Mne1 and a unitary u ∈ Mn such that kAd u ◦ ϕ1(f ) − (h1(f ) ⊕ diag( ψ(f ), ψ(f ), ..., ψ(f )))k < ε, kϕ2(f ) − (h2(f ) ⊕ diag( ∆(g) and τ ◦ ψ(g) ≥ α where τ is the tracial state on Mn. K z ψ(f ), ψ(f ), ..., ψ(f )))k < ε for all f ∈ F for all g ∈ H0, { K } { z K } 29 (e 4.39) (e 4.40) (e 4.41) Proof. We will use induction on integer k ≥ 0. The case k = 0 follows from 4.6. Now assume that the lemma holds for integers 0 ≤ k ≤ m. We assume that A ∈ ¯D′m+1. We will retain the notation for A as an algebra in ¯D′m+1 in In particular, Xm+1 = X, Zm+1 = Z, Y = X \ Z, X 0 = Z = X \ Y, the later part of 4.8. I = P C0(Y, F )P ⊂ A. We will write A = {(f, b) ∈ P C(X, F )P ⊕ B : fX 0 = Γ(b)}, (e 4.42) where B ∈ ¯D′m is a unital C∗-algebra and will be identified with A/I. We also keep the notation λ : A → P C(X, F )P in the pull-back in the last part of 4.8. We will use fS for λ(f )S for f ∈ A and S ⊂ X in the proof when there is no confusion. Let dX > 0 be given in 4.8. Denote by πI : A → A/I the quotient map. We may write that P C(X, F )P = k2Mj=1 PjC(X, Ms(k,j))Pj , (e 4.43) where Pj ∈ C(X, Ms(k,j)) is a projection of rank r(j) at each x ∈ X. We may assume that A/I has irreducible representations with rank l1, l2, ..., lk1 . Choose T = (k1k2) · maxi,j{zizj : zi, zj ∈ {l1, l2, ..., lk1 , r(1), r(2), ..., r(k2)}}. 1/N00 < (1−β) 64 . Choose β =p1 − (1 − α)/8 =p(7 + α)/8. Note that 1 > β2 > α. Fix N00 ≥ 4 such that Fix ε > 0, a finite subset F ⊂ A and a finite subset H0 ⊂ A1 + \ {0}. We may assume that 1A ∈ H0 ⊂ F. Without loss of generality, we may also assume that F ⊂ As.a. and kfk ≤ 1 for all f ∈ F. Write I = {f ∈ P C(X, F )P : fX 0 = 0}. Put There is d > 0 such that δ00 = min{∆(g)/2 : g ∈ H0} > 0. (e 4.44) kπx,j(f ) − πx′,j(f )k < min{ε, δ00}/256KN00 for all f ∈ F, (e 4.45) provided that dist(x, x′) < d for any pair x, x′ ∈ X (here we identify πx,j(f ) with πx,j(Λ(f )) —see the 4.8). Put ε0 = min{ε, δ00}/16KN00. We also assume that, for any x ∈ X d = {x ∈ X : dist(x, X 0) < d} choosing smaller d if necessary, kπx,j ◦ sd ◦ (Λ(f )Z ) − πx,j(f )k < ε0/16, where sd : QC(Z, F )Q → PX dC(X d, F )PX d is induced by sd sd also induces a map (still denoted by sd) (e 4.46) ∗ : X d → Z (see 4.8). Note that sd : B → B ⊕QC(Z,F )Q PX dC(X d, F )PX d, where Q = PZ . To simplify the notation, we may assume that d < dX /2. For any b > 0, as in 4.8, we will continue to use X b for {x ∈ X : dist(x, X 0) < b}. Let Y0 = X \ X d/2. Put B0,0 = P C(Y0, F )P. Let FI,0 = {fY0 : f ∈ F} and let H0,I,0 = {hY0 : h ∈ H0}. Let f0,0 ∈ C0(Y )+ be such that 0 ≤ f0,0 ≤ 1, f0,0(x) = 1 if x ∈ X \ X d, f0,0(x) = 0 if x /∈ Y0 Let ∆I,0 : (B0,0)q,1 and f0,0(x) > 0 if dist(x, X 0) > d/2. + \ {0} → (0, 1) be defined by ∆I,0(g) = β∆(bg′) for all g ∈ (B0,0)1 30 + \ {0}, (e 4.47) Let H1,I,0 ⊂ (B0,0)1 where g′ = f0,0· P · g which is viewed as an element in I 1 +. Note that if g 6= 0, then f0,0· P · g 6= 0. So ∆I,0 : (B0,0)q,1 + → (0, 1) is an order preserving map. Let N I ≥ 1 be an integer (in place of N ) as required by 4.6 for B0,0 (in place of A), ∆I,0 (in place of ∆), ε0/16 (in place of ε), FI,0 (in place of F) and H0,I,0 (in place of H0) and 2K (in place of K). + \ {0} be a finite subset (in place of H1), H2,I,0 ⊂ (B0,0)s.a. (in place of H2) and δ1 > 0 (in place of δ) be as required by 4.6 for ε0/16 (in place of ε), FI,0 (in place of F), 2K (in place of K) and H0,I,0 associated with B0,0 (in place of A) and ∆I,0 (in place of ∆) and β (in place of α). Without loss of generality, we may assume that kgk ≤ 1 for all g ∈ H2,I,0. Let Fπ = πI (F). Let g′0 ∈ C(X)+ with 0 ≤ g′0 ≤ 1 such that g′0(x) = 0 if dist(x, X 0) < d/256 and g′0(x) = 1 if dist(x, Y0) ≤ d/16. Define g0 = 1A−g′0·P. Since g′0 ∈ I, we view g0 as an element in A. Note that (1− g′0)P · sd(g) may be viewed as an element in P (C(X, F )P as (1− g′0)(x) = 0 for x /∈ X d. (Recall that sd : QC(Z, F )Q → P C(X d, F )P is also regarded as map sd : B = A/I → B ⊕QC(Z,F )Q PX dC(X d, F )PX d .) Hence g0 · sd(g) may be identified with ((1 − g′0)P · sd(g), g) ∈ A. Define ∆π(g) = β∆( \g0 · sd(g)) for all g ∈ (A/I)1 +. (e 4.48) + \ {0} → (0, 1) is an order preserving map. We will late used the fact that g0(x) = 0 if dist(x, Y0) ≤ d/16. ∆π : (A/I)q,1 Note if g is nonzero, so is sd(g). Since g0X 0 = 1, we have that g0 · sd(g) 6= 0. It follows that Put H0,π = πI (H0). Let N π ≥ 1 be the integer associated with A/I(= B), ∆π, ε0/16, Fπ and H0,π (as required by the inductive assumption that the lemma holds for integer m). Let H1,π ⊂ (A/I)q,1 + be a finite subset (in place of H1), H2,π ⊂ A/Is.a. be a finite subset (in place of H2) and let δ2 > 0 (in place of δ) be as required by the inductive assumption that this lemma for the case that k = m for ε0/16 (in place of ε), Fπ (in place of F), H0,π (in place of H0) and 2K associated with A/I (in place of A), ∆π (in place of ∆) and β (in place of α). Without loss of generality, we may assume that khk ≤ 1 for all h ∈ H2,π. Set δ000 = min{δ1, δ2, ε0}. There is an integer N0 ≥ 256 such that 1/N0 < ∆( \f0,0 · P ) · δ000 2 · min{∆I,0(g) : g ∈ H1,I,0} · min{∆π(g) : g ∈ H1,π}/64KN00.(e 4.49) Define Yk to be the closure of {y ∈ Y : dist(y, Y0) < kd 64N 2 0 }, k = 1, 2, ..., 4N 2 0 . Let FI,k = {fYk : f ∈ F} and let H0,I,k = {hYk : h ∈ H0}. Put B0,k = PYk C(Yk, F )PYk . Let f0,k ∈ C0(Y )+ be such that 0 ≤ f0,k ≤ 1, f0,k(x) = 1 if x ∈ Yk−1, f0,k(x) = 0 if x 6∈ Yk and f0,k(x) > 0 if dist(x, Y0) < kd 64N 2 0 , k = 1, 2, ..., 4N 2 0 . Let rk : Yk → Y0 be a homeomorphism such that dist(rk(x), x) < d/4 for all x ∈ Yk, k = 1, 2, ..., 4N 2 0 (see 4.8). Let ∆I,k : (B0,k)q,1 + \ {0} → (0, 1) be defined by (e 4.50) (e 4.51) ∆I,k(g) = β∆(bg′) for all g ∈ (B0,k)1 k which is viewed as an element in I 1 +, +. Note that if g 6= 0, then where g′ = (f0,k · P · g) ◦ r−1 f0,k · 1A · g 6= 0. So ∆I,k : (B0,k)q,1 + \ {0} → (0, 1) is an order preserving map. Let F′I,k = {f ◦ rk : f ∈ FI,0} and H′0,I,k = {g ◦ rk : g ∈ H0,I,0}, k = 1, 2, ..., 4N 2 0 . Any unital homomorphism Φ : B0,k → C (for any unital C∗-algebra C) induces a unital homomorphism Ψ : B0,0 → C by Ψ(f ) = Φ(f ◦ rk) for all f ∈ B0,0. Note also that f 7→ f ◦ rk is an isomorphism from B0,0 onto B0,k. 31 Then, for ε0/16 (in place of ε), F′I,k (in place of F), K and H′0,I,k associated with B0,k (in place of A), ∆I,k (in place of ∆) and β(in place of α), to apply 4.6, one can choose H1,l,k (in place of H1) to be H1,I,0 ◦ rk, H2,I,k (in place of H2) to be H2,I,0 ◦ rk and δ1 (in place of δ). We also note that kf − fY0 ◦ rkk < min{ε, δ00}/64KN00 for all f ∈ FI,k. (e 4.52) Define σ0 = min{min{∆I,0(g) : g ∈ H1,I,0}, min{∆π(g) : g ∈ H1,π}}. Choose an integer N ≥ N π + N I such that T N < σ0 · min{δ1/64, δ2/64, ε0/64K}/N00. (e 4.53) Put H1 = ∪4N 2 0 k=0{f0,k · 1A · g ◦ rk : g ∈ H1,I,0} ∪ {(g0 · 1A · sd(g), g) : g ∈ H1,π} ∪ GI ∪ Gπ, (e 4.54) where GI is a finite subset in {g· P : g ∈ C0(Y ) : gY \Y0 = 0} which contains at least N mutually orthogonal non-zero positive scalar elements and Gπ is a finite subset of J which contains at least N mutually orthogonal non-zero scalar elements, where J = {(f · P, b) ∈ A : f ∈ C(X) : fX\X d/256 = 0} such that 0 ≤ f ≤ 1 for all f ∈ GI ∪ Gπ. With the convention that r0 : Y0 → Y0 is the identity map, put H′2 = ∪4N 2 0 k=0{f0,k · g ◦ rk : g ∈ H2,I,0} ∪ {f0,k, 0 ≤ k ≤ 4N 2 0}. (e 4.55) Define g′0,k ∈ C(X)+ so that 0 ≤ g′0,k ≤ 1, g′0,k(x) = 0 if x 6∈ Yk, g′0,k(x) = 1 if x ∈ Yk−1 and g′0,k(x) > 0 if x ∈ Yk \ Yk−1, k = 1, 2, .... Put g0.k = 1A − g′0,k · P. Note, since g′0,k · P ∈ I, g0,k ∈ A. Define k=1{(g0,k · sd(g), g) ∈ A : g ∈ H2,π} ∪ F. H′′2 = ∪4N 2 0 Put Let H2 = H′2 ∪ H′′2. δ = σ0 · min{δ1/64, δ2/64, ε0/64K} 4KN1N00 . (e 4.56) (e 4.57) Now let ϕ1, ϕ2 : A → Mn (for some integer n ≥ N ) be two unital homomorphisms such that they satisfy the assumption for the above H1, H2 and δ. Consider two finite Borel measures on Y defined by ZY f µi = τ ◦ ϕi(f · P ) for all f ∈ C0(Y ), i = 1, 2, (e 4.58) where τ is the tracial state on Mn. Note that {Yk \ Yk−1 : k = 1, 2, ...., 4N 2 0 } is a family of 4N 2 least 2N0 of k′s such that 0 disjoint Borel sets. There are at µ1(Yk \ Yk−1) < 1/N0. 32 (e 4.59) It is clear then there is at least one of them satisfies µi(Yk \ Yk−1) < 1/N0, i = 1, 2. Fix one of such k. We may write that ϕ1 = Σ1 π ⊕ Σ1 b ⊕ Σ1 s ⊕ Σ1 I and ϕ2 = Σ2 π ⊕ Σ2 b ⊕ Σ2 s ⊕ Σ2 I , (e 4.60) (e 4.61) I and Σ2 I are finite direct sums of the form πx,j for x ∈ Yk−1, Σ1 where Σ1 direct sums of the from πx,j for x ∈ Yk \ Yk−1, Σ1 πx,j for x ∈ Y \ Yk and Σ1 representations of A/I (note that these πx,j or ¯πx,i can be repeated). s are finite b are finite direct sum of the form π are finite direct sum of the form ¯πx,i given by irreducible b and Σ2 s and Σ2 π and Σ2 Define ψ1,0 , ψ2,0 I I : B0,k−1 → Mn by ψi,0 I (f ) = Σi I (f ) for all f ∈ B0,k−1, i = 1, 2. By the choice of H2, we estimate that (e 4.62) (e 4.63) (e 4.64) (e 4.65) (e 4.66) I (1B0,k−1 ) I (1B0,k−1 ) − τ ◦ ψ2,0 I (f0,k · P ) − τ ◦ ϕ1(f0,k · P ) + τ ◦ ϕ1(f0,k · P ) − τ ◦ ϕ2(f0,k · P ) τ ◦ ψ1,0 ≤ τ ◦ Σ1 + τ ◦ ϕ2(f0,k · P ) − Σ2 ≤ δ + ∆(cf00 · P )δ1 min{∆I,0(g) : g ∈ H1,I,0}/32KN00. I(f0,k · P ) < 1/N0 + δ + 1/N0 It follows from 4.1 that there are two mutually equivalent projections p1,0 and p2,0 ∈ Mn such that pi,0 commutes with ψi,0 I (1B0,k−1 ) = pi,0. i = 1, 2, and I (f ) for all f ∈ B0,k−1 and pi,0ψi,0 I (1B0,k−1 ) − τ (pi,0) < δ + ∆( \f00 · P )δ1 min{∆I,0(g) : g ∈ H1,I,0}/32KN00 + T /n,(e 4.67) 0 ≤ τ ◦ ψi,0 i = 1, 2. Since Y0 ⊂ Yk−1, supp(f00)= Y0 ⊂ Yk−1. Therefore, using (e 4.49), τ ◦ ψ1,0 I (1B0,k−1 ) ≥ τ ◦ ψ1,0 I (f00 · P ) ≥ ∆( \f00 · P ) > max{σ0, 8/(N0δ1)}. Hence τ (p2,0) > max{σ0, 8/(N0δ1)}/2. (e 4.68) (e 4.69) Put qi,0 = ψi,0 Define ψ1 ψ2 I : B0,k−1 → p2,0Mnp2,0 by ψ2 (e 4.67) among other items) I (1B0,k−1 ) − pi,0, i = 1, 2. There is a unitary U1 ∈ Mn such that U∗1 p1,0U1 = p2,0. I (f )U1 for all f ∈ B0,k−1 and define I (f ) for all f ∈ B0,k−1. We compute that (using I : B0,k−1 → p2,0Mnp2,0 by ψ1 I (f ) = U∗1 p1,0ψ1,0 I (f ) = p2,0ψ2,0 τ ◦ ψ1 I (g ◦ r−1 k−1) ≥ ∆(\f0,k−1 · 1A · g) − min{∆(g : g ∈ H1,I,0}/2N00 > β∆(\f0,k−1 · 1A · g) = ∆I,0(g ◦ r−1 k−1) for all g ∈ H1,I,0. Therefore t ◦ ψ1 I (g) ≥ ∆I,k(g) for all g ∈ H1,I,k, 33 (e 4.70) (e 4.71) (e 4.72) where t is the tracial state on p2,0Mnp2,0. We also estimate that t ◦ ψ1 I (g) − t ◦ ψ2 I (g) = (1/τ (τ ◦ ψ1 I (g) − τ ◦ ψ2 I (g) I (g) I (g) − τ ◦ ψ1,0 I (g) − τ (ϕ1(f0,k · 1A · g)) ≤ (1/τ (p1,0))τ ◦ ψ1 + (1/τ (p1,0))τ ◦ ψ1,0 + (1/τ (p2,0))τ (ϕ1(f0,k · 1A · g)) − τ (ϕ2(f0,k · 1A · g)) + (1/τ (p2,0))τ (ϕ2(f0,k · 1A · g) − τ ◦ ψ2,0 + τ ◦ ψ2,0 < (1/τ (p2,0))(τ (q1,0) + 1/N0 + δ + 1/N0 + τ (q2,0)) < δ1 I (g) − τ ◦ ψ2 I (g) I (g) (e 4.79) for all g ∈ H2,I,k (the last step uses (e 4.69)). Note also that p2,0 has rank at least N I . It follows (by applying 4.6) that there are mutually orthogonal projections eI 2K ∈ p2,0Mnp2,0 such that eI 1, two unital homomorphisms ψ1,I,0, ψ2,I,0 : B0,k−1 → eI 1 and a unitary u1 ∈ p2,0Mnp2,0, such that 0 . eI 0, a unital homomorphism ψI : B0,k−1 → eI 0 +P2K j are equivalent to eI i = p2,0, eI 1 and eI 2, ..., el 0MneI 1MneI i=1 eI 0, eI 1, eI (e 4.73) (e 4.74) (e 4.75) (e 4.76) (e 4.77) (e 4.78) (e 4.80) (e 4.81) (e 4.82) (e 4.83) (e 4.84) (e 4.85) (e 4.86) (e 4.87) (e 4.88) (e 4.89) (e 4.90) (e 4.91) kAd u1 ◦ ψ1 and kψ2 kAd u1 ◦ ψ1 and kψ2 z z z z 2K 2K } } 2K 2K } } { { { { I (f ) − (ψ1,I,0(f ) ⊕ diag( ψI (f ), ψI (f ), ..., ψI (f )))k < ε0/16 I (f ) − (ψ2,I,0(f ) ⊕ diag( ψI (f ), ψI (f ), ..., ψI (f )))k < ε0/16 for all f ∈ F′I,k−1. By (e 4.52), the above implies that I (f ) − (ψ1,I,0(f ) ⊕ diag( ψI (f ), ψI (f ), ..., ψI (f )))k < ε0/8 I (f ) − (ψ2,I,0(f ) ⊕ diag( ψI (f ), ψI (f ), ..., ψI (f )))k < ε0/8 for all f ∈ FI,k−1. For each x ∈ X \ Yk such that πx,j appeared in Σ1 b, by (e 4.45), kπx,j(f ) − πx,j ◦ sd ◦ πI (f ))k < ε0/16 for all f ∈ F. b, or Σ2 Define Σπ,b,i = Σi b ◦ sd : A/I → Mn, i = 1, 2. Define Φ1 : A/I → (1 − p2,0)Mn(1 − p2,0) by Φ1(f ) = Ad U1 ◦ (Σ1 π ⊕ Σπ,b,1)(f ) for all f ∈ A/I. Define Φ2 : A/I → (1 − p2,0)Mn(1 − p2,0) by Φ2(f ) = (Σ1 π ⊕ Σπ,b,2)(f ) for all f ∈ A/I. Note that Φ1(1A/I ) = Σ1 π(g0,k) ⊕ Σ1 b(g0,k) and Φ2(1A/I ) = Σ1 π(g0,k) ⊕ Σ2 b(g0,k). We estimate that τ ◦ Φ1(1A/I ) − τ ◦ Φ2(1A/I ) ≤ τ ◦ Φ1(1A/I ) − τ ◦ ϕ1(g0,k) +τ ◦ ϕ1(g0,k) − τ ◦ ϕ2(g0,k) + τ ◦ ϕ2(g0,k) − τ ◦ Φ2(g0,k) < 1/N0 + δ + 1/N0 < δ + ∆( \f00 · P )δ2 min{∆π(g) : g ∈ H1,π}/32N00. . 34 It follows from 4.1 that there are two mutually equivalent projections p1,1 and p2,1 ∈ (1 − p2,0)Mn(1 − p2,0) such that pi,1 commutes with Φi(f ) for all f ∈ A/I and pi,1Φi(1A/I ) = pi,1. i = 1, 2, and 0 ≤ τ ◦ Φi(1A/I ) − τ (pi,1) < δ + ∆(cf00)δ2 min{∆π(g) : g ∈ H1,π}/(32N00) + T /n (e 4.92) i = 1, 2. Since g0(x) = 0 if dist(x, Y0) ≤ d/16, we have τ ◦ Φ1(1A/I ) > ∆( \g0 · sd(1)) > ∆π(1) ≥ max{σ0, 8/(N0δ2)}, and τ (p2,1) ≥ max(σ0,8/(N0δ2)) U∗2 p1,1U2 = p2,1. Define Φ1 and define Φ2 2 . Put qi,1 = Φi(1A/I ) − pi,1, i = 1, 2. There is a unitary U2 ∈ (1 − p2,0)Mn(1 − p2,0) such that π(f ) = U∗2 p1,1Φ1(f )U2 for all f ∈ B0,k−1 π : A/I → p2,1Mnp2,1 by Φ1 π : A/I → p2,1Mnp2,1 by Φ2 π(f ) = p2,1Φ2(f ) for all f ∈ A/I. We compute that τ ◦ Φ1 π(g) ≥ ∆( \g0 · 1A · sd(g)) − σ0/N00 > β∆( \g0 · 1A · sd(g)) = ∆π(g) for all g ∈ H1,π. Therefore t1 ◦ Φ1 π(g) ≥ ∆π(g) for all g ∈ H1,π, where t1 is the tracial state on p2,1Mnp2,1. We also estimate (similar to the estimate of (e 4.79)) that (e 4.93) (e 4.94) (e 4.95) t1 ◦ Φ1 π(g) − t1 ◦ Φ2 π(g) − τ ◦ Φ2 π(g) π(g) − τ ◦ Φ1(g) π(g) = (1/τ (p2,1))τ ◦ Φ1 ≤ (1/τ (p2,1))τ ◦ Φ1 + (1/τ (p2,1))τ ◦ Φ1(g) − τ ◦ ϕ1(g0,k · 1A · sd(g)) (e 4.98) + (1/τ (p2,1))τ ◦ ϕ1(g0,k · 1A · sd(g)) − τ ◦ ϕ2(g0,k · 1A · sd(g)) (e 4.99) + (1/τ (p2,1))τ ◦ ϕ2(g0,k · 1A · sd(g)) − τ ◦ Φ2(g) (e 4.100) + (1/τ (p2,1))τ ◦ Φ2(g) − τ ◦ Φ2 π(g) < (1/τ (p2,1))(1/N0 + δ + 1/N0) < δ2 (e 4.101) (e 4.102) (e 4.97) (e 4.96) for all g ∈ H2,π. It follows from 4.6 that there are mutually orthogonal projections eπ p2,1Mnp2,1 such that eπ A/I → eπ such that 2 , ..., eπ 1 , two unital homomorphisms ψ1,π,0, ψ2,π,0 : 1 and a unitary u2 ∈ p2,1Mnp2,1, j are equivalent to eπ 1 and eπ 0 Mneπ 1 Mneπ 0 . eπ 0 , eπ 1 , eπ 2K ∈ kψi π(f ) − (ψi,π,0 ◦ πI (f ) ⊕ diag( ψπ(πI (f )), ψπ(πI (f )), ..., ψπ(πI (f ))))k < ε0/8 (e 4.105) { kAd u2 ◦ Φ1 and kΦ2 for all f ∈ Fπ. Let ψ1 π : A → p2,1Mnp2,1 by ψ2 ψ2 0 , a unital homomorphism ψπ : A/I → eπ } } π(f ) − (ψ1,π,0(f ) ⊕ diag( 2K π(f ) = p2,1(Σ2 π(f ) − (ψ2,π,0(f ) ⊕ diag( π : A → p2,1Mnp2,1 by ψ1 π ⊕ Σ2 z 2K } 35 2K ψπ(f ), ψπ(f ), ..., ψπ(f )))k < ε0/16 z z ψπ(f ), ψπ(f ), ..., ψπ(f )))k < ε0/16 π ⊕ Σ1 π(f ) = Ad u2 ◦ Ad U2(p2,1(Σ1 b)(f ) for all f ∈ A. Then, by (e 4.84), { { (e 4.103) (e 4.104) b)(f )) and define for all f ∈ F, i = 1, 2. 2i−1 ⊕ eI Put ei = eI 2i ⊕ eπ 2i−1 ⊕ eπ 2i, i = 1, 2, ..., K. Define ψ : A → e1Mne1 by ψ(f ) = diag(ψI (fYk−1), ψI (fYk−1), ψπ ◦ πI (f ), ψπ ◦ πI (f )) for all f ∈ A. By (e 4.59), (e 4.67) and (e 4.92), τ (qi,0) + τ (qi,1) + τ (Σi s(1A)) < 1/64K + 1/N0 + 1/64K < 1/16K. (e 4.106) We have, for f ∈ A, ϕ2(f ) = ψ2 Put e0 = eI 0 ⊕ eπ 0 + q2,1 + Σ2 π(f ) ⊕ q2,1(Σ2 π + Σ2 b)(f ) ⊕ Σ2 s(1A) + q2,0. Then 0) + τ (eπ τ (e0) < τ (eI s(f ) ⊕ ψ2 I (fYk−1) ⊕ q2,0ψ2,0 I (fYk−1). (e 4.107) (e 4.108) In other words, e0 . e1. Moreover e1 is equivalent to each ei, i = 1, 2, ..., K. Define h2 : A → e0Mne0 by, for each f ∈ A, 0 ) + 1/16K ≤ τ (e1). h2(f ) = ψ2,I,0(fYk−1) ⊕ ψ2,π,0(πI (f )) ⊕ q2,1(Σ2 π + Σ2 b)(f ) ⊕ Σ2 s(f ) ⊕ q2,0ψ2,0 I (fYk−1). It follows from (e 4.83), (e 4.105) and above that kϕ2(f ) − (h2(f ) ⊕ diag( ψ(f ), ψ(f ), ..., ψ(f ))k < ε0/8 for all f ∈ F. (e 4.109) Similarly, there exists a unitary U ∈ Mn and a unital homomorphism h2 : A →e0Mne0 such that z K } { kAd U ◦ ϕ1(f ) − (h1(f ) ⊕ diag( ψ(f ), ψ(f ), ..., ψ(f ))k < ε0/8 for all f ∈ F. (e 4.110) Since we also assume that H0 ⊂ F in the above proof, it is easy to check, by the choice of ε0 and β, that z K } { τ ◦ ψ(g) ≥ α ∆((g)) K for all g ∈ H0. (e 4.111) Remark 4.12. If we assume that A is infinite dimensional, then Lemma 4.11 still holds without the assumption about the integer N. This could be easily seen by taking a larger H1. The following is known and is taken from Theorem 3.9 of [61] Theorem 4.13. Let A be a unital separable amenable C∗-algebra which satisfies the UCT and let B be a unital C∗-algebra. Suppose that h1, h2 : A → B are two homomorphisms such that [h1] = [h2] in KL(A, B). Suppose that h0 : A → B is a unital full monomorphism. Then, for any ε > 0 and any finite subset F ⊂ A, there exits an integer n ≥ 1 and a unitary W ∈ Mn+1(B) such that kW ∗diag(h1(a), h0(a), ..., h0(a))W − diag(h2(a), h0(a), ..., h0(a))k < ε for all a ∈ F, and W ∗pW = q, where p = diag(h1(1A), h0(1A), ..., h0(1A)) and q = diag(h2(1A), h0(1A), ..., h0(1A)). In particular, if h1(1A) = h2(1A), W ∈ U (pMn+1(B)p). 36 Proof. This is a slight variation of Theorem 3.9 of [61]. If h1 and h2 are both unital, then it is exactly the same as Theorem 3.9 of [61]. So suppose that h1 is not unital. Let A′ = C ⊕ A. Choose p0 = 1B − h1(1A) and p1 = diag(p0, 1B). Put B′ = p1M2(B)p1. Define h′1 : A′ → B′ by h′1(λ ⊕ a) = λ · diag(p0, p0) ⊕ h1(a) for all λ ∈ C and a ∈ A, and define h′2 : A′ → B′ by h′2(λ ⊕ a) = λ · diag(p0, 1B − h2(1A)) ⊕ h2(a) for all λ ∈ C and a ∈ A. Then h′1 and h′2 are unital and [h′1] = [h′2] in KL(A′, B′). Define h′0 : A′ → B′ by h′0(λ ⊕ a) = λ · p0 ⊕ h0(a) for all λ ∈ C and a ∈ A. Note that h′0 is full in B′. So, Theorem 3.9 of [61] applies. It follows that there is an integer n ≥ 1 and a unitary W ′ ∈ Mn+1(B′) such that k(W ′)∗diag(h′1(a), h′0(a), ..., h′0(a))W ′ − diag(h′2(a), h′0(a), ..., h′0(a))k < min{1/2, ε/2} for all a ∈ F ∪ {1A}. In particular, k(W ′)∗pW ′ − qk < min{1/2, ε/2}. There is a unitary W1 ∈ Mn+1(B′) such that kW1 − 1Mn+1(B′)k < ε/2 and W ∗1 (W ′)∗pW ′W1 = q. Put W = W ′W1. Then (e 4.112) (e 4.113) kW ∗diag(h1(a), h0(a), ..., h0(a))W − diag(h2(a), h0(a), ..., h0(a))k < ε (e 4.114) for all a ∈ F. Lemma follows. Lemma 4.14. (cf. 5.3 of [56], Theorem 3.1 of [37], [16], 5.9 of [61] and Theorem 7.1 of [67]) Let A be a unital separable amenable C∗-algebra which satisfies the UCT and let ∆ : Aq,1 + \{0} → (0, 1) be an order preserving map. For any ε > 0 and any finite subset F ⊂ A, there exists δ > 0, a finite subset G ⊂ A, a finite subset P ⊂ K(A), a finite subset H ⊂ A1 + \ {0} and an integer K ≥ 1 satisfying the following: For any two unital G-δ-multiplicative contractive completely positive linear maps ϕ1, ϕ2 : A → Mn (for some integer n) and any unital G-δ-multiplicative contractive completely positive linear map ψ : A → Mm with m ≥ n such that τ ◦ ψ(g) ≥ ∆(g) for all g ∈ H and [ϕ1]P = [ϕ2]P , there exists a unitary U ∈ MKm+n such that kAd U ◦ (ϕ1 ⊕ Ψ)(f ) − (ϕ2 ⊕ Ψ)(f )k < ε for all f ∈ A, (e 4.115) (e 4.116) where z K } { Ψ(f ) = diag( ψ(f ), ψ(f ), ..., ψ(f )) for all f ∈ A. Proof. This follows from 4.13. Fix ∆ as given. Suppose the lemma is false. Then there exist ε0 > 0 and a finite sub- set F0 ⊂ A, an increasing sequence of finite subsets {Pn} of K(A) whose union is K(A), an increasing sequence of finite subsets {Hn} ⊂ A1 + and if a ∈ Hn and f1/2(a) 6= 0, then f1/2(a) ∈ Hn+1, three sequences of increasing integers {R(n)} and {r(n)}, {s(n)} (with s(n) ≥ r(n)), two sequences of contractive completely positive linear maps ϕ1,n, ϕ2,n, : A → Mr(n) with the properties that + \ {0} whose union is dense in A1 [ϕ1,n]Pn = [ϕ2,n]Pn and lim n→∞kϕi,n(ab) − ϕi,n(a)ϕi,n(b)k = 0 for all a, b ∈ A, i = 1, 2, (e 4.117) (e 4.118) 37 a sequence of unital contractive completely positive linear maps ψn : A → Ms(n) with the properties that τn ◦ ψn(g) ≥ ∆(g) for all g ∈ Hn and n→∞kψn(ab) − ψn(a)ψn(b)k = 0 for all a, b ∈ A lim (e 4.119) (e 4.120) such that inf{sup{kAd Un ◦ (ϕ1,n(f ) ⊕ ψR(n) n (f )) − (ϕ2,n(f ) ⊕ ψR(n) n (f )k : f ∈ F}} ≥ ε0, (e 4.121) where τn is the normalized trace on Mr(n), ψ(R(n))(f ) = diag( f ∈ A, and the infimum is taken among all unitaries Un ∈ Mr(n). Note that, by (e 4.119), since {Hn} is increasing, for any g ∈ Hn ⊂ A1 ψn(f ), ψn(f ), ..., ψn(f )) for all z R(n) } { +, we compute that (e 4.122) where pm is the spectral projection of ψm(g) corresponding to the subset {λ > ∆(g)/2} for all m ≥ n. It follows that (for all sufficiently large m) there are element xg,i,m ∈ Ms(m) with kxg,i,mk ≤ 1/∆(g), i = 1, 2, ..., N (g) such that τm(pm) ≥ ∆(g)/2, x∗g,i,mψm(g)xg,i,m = 1s(m), (e 4.123) where 1 ≤ N (g) ≤ 1/∆(g) + 1. Put Xg,i = {xg,i,m}, i = 1, 2, ..., N (g). Then Xg,i ∈Q∞n=1 Mr(n). Let Q({Mr(n)}) = Q∞n=1 Mr(n)/L∞n=1 Mr(n), Q({Ms(n)}) = Q∞n=1 Ms(n)/L∞n=1 Ms(n), and let Π1 :Q∞n=1 Mr(n) →Q∞n=1 Mr(n)/L∞n=1 Mr(n), Π2 :Q∞n=1 Ms(n) →Q∞n=1 Ms(n)/L∞n=1 Ms(n) be quotient maps. Denote by Φi : A → Q({Mr(n)}) the homomorphisms Π1 ◦ {ϕi,n} and denote by ¯ψ : A → Q({Ms(n)}) the homomorphism Π2 ◦ {ψn}. For each g ∈ ∪∞n=1Hn, Π2(Xi,g)∗ ¯ψ(g)Π2(Xi,g) = 1Q({Ms(n)}). (e 4.124) N (g)Xi=1 N (g)Xi=1 Note that if g ∈ ∪n=1Hn and f1/2(g) 6= 0, then g1/2(g) ∈ ∪n=1Hn, and ∪n=1Hn is dense in A1 +. real rank zero. One computes that This implies that ¯ψ is full. Note that bothQ∞n=1 Mr(n) and Q({Mr(n)} have stable rank one and (e 4.125) By applying Theorem 4.13 (Theorem 3.9 of [61]), there exists an integer K ≥ 1 and a unitary U ∈ P MK+1(Q({Ms(n)})P, where P = diag(1Q({Mr(n)}, 1MK (Q({Ms(n)}))), such that [Φ1] = [Φ2] in KL(A, Q({Mr(n)}). kAd U ◦ (Φ1(f ) ⊕ diag( for all f ∈ F0. It follows that there are unitaries K z z ¯ψ(f ), ..., ¯ψ(f ))k < ε0/2 ¯ψ(f ), ..., ¯ψ(f )) − (Φ2(f ) ⊕ diag( } } { { K (e 4.126) such that, for all large n, {Un} ∈ Mr(n)+Ks(n) ∞Yn=1 kAd Un ◦ ϕ1,n(f ) ⊕ diag( for all f ∈ F0. This leads to a contradiction with (e 4.121) when we choose n with R(n) ≥ K. ψn(f ), ..., ψn(f )) − ϕ2,n(f ) ⊕ diag( ψn(f ), ..., ψn(f ))k < ε0/2 (e 4.127) z K } { z K } { 38 Remark 4.15. This lemma holds in a much more general setting and variations of it has appeared. We state this version here for our immediate purpose (see 12.1 and part (1) of 12.2 for more comments). It should be noted that, in the following statement the integer L and Ψ depend not only on ε, F, G, but also depend on B as well as ϕ1 and ϕ2. Lemma 4.16. Let C be a unital amenable separable residually finite dimensional C∗-algebra which satisfies the UCT. For any ε > 0, any finite subset F ⊂ C, there exists a finite subset G ⊂ C, δ > 0, a finite subset P ⊂ K(C) satisfying the following: For any unital G-δ-multiplicative contractive completely positive linear maps ϕ1, ϕ2 : C → A (for any unital C∗-algebra A) such that [ϕ1]P = [ϕ2]P , (e 4.128) there exist an integer L ≥ 1, a unital homomorphism Ψ : C → ML(C) ⊂ ML(A), and a unitary U ∈ U (ML+1(A)) such that kAd U ◦ diag(ϕ1(f ), Ψ(f )) − diag(ϕ2(f ), Ψ(f )k < ε (e 4.129) for all f ∈ F. Proof. The proof is almost the same as that of Theorem 9.2 of [65]. Suppose that the lemma is false. We then obtain positive number ε0 > 0 and a finite F0 ⊂ C, a sequence of finite subsets Pn ⊂ K(C) with Pn ⊂ Pn+1 and ∪nPn+1 = K(C), a sequence of unital C∗-algebras {An}, a sequence of unital contractive completely positive linear maps {L(1) n } (from C to An) n } and {L(2) such that n (a)L(i) n (b)k = 0 for all a, b ∈ C, n (ab) − L(i) n→∞kL(i) lim [L(1) n ]Pn = [L(2) n ]Pn, inf{sup{ku∗ndiag(L(1) n (a), Ψn(a))un − diag(L(2) n (a), Ψn(a)k : a ∈ F0} ≥ ε0, (e 4.130) (e 4.131) (e 4.132) where the infimum is taken among all integers k > 1, all possible unital homomorphisms Ψn : C → Mk(C) and all possible unitaries U ∈ Mk+1(An). We may assume that 1C ∈ F. Define Bn = An ⊗ K, B = Q∞n=1 Bn and Q1 = B/L∞n=1 Bn. Let π : B → Q1 be the quotient map. Define ϕj : C → B by ϕj(a) = {L(j) n (a)} and define ¯ϕj = π◦ ϕj, j = 0, 1. Note that ¯ϕj : C → Q1 is a homomorphism as in the proof of 9.2 of [65], we have [ ¯ϕ1] = [ ¯ϕ2] KL(C, Q). Fix an irreducible representation ϕ′0 : C → Mr. Denote by pn the unit of the unitization Bn of 0 (c) = ϕ′0(c)⊗1 B : C → Mr( Bn) = Mr⊗ Bn by ϕ(n) Bn, n = 1, 2, .... Define a homomorphism ϕ(n) 0 for all c ∈ C. Put eA = {1An}, P = {1Mr( Bn)} + eA. Put also Q2 = π(P )Mr+1( Q1)π(P ) and define ¯ϕj = ¯ϕ′j ⊕ π ◦ {ϕ(n) 0 }, j = 1, 2. Then [ ¯ϕ1] = [ ¯ϕ2] in KK(C, Q2). (e 4.133) The point to add π ◦ {ϕ(n) 0 } is that, now, ¯ϕ1 and ¯ϕ2 are unital. It follows from Theorem 4.3 of [16] that there exists an integer K > 0, a unitary u ∈ M1+K(Q1) and a unital homomorphism 39 ψ : C → MK(C) ⊂ MK (Q2) (by identifying MK (C) with the unital subalgebra of MK(Q2)) such that Ad u ◦ diag( ¯ϕ1, ψ) ≈ε0/4 diag( ¯ϕ2, ψ) on F0. There exists a unitary V = {Vn} ∈ M1+K(P Mr+1( B)P ) such that π(V ) = u. It follows (by identifying MK (C) with MK (C) ⊗ 1Q2) that for all sufficiently large n, Ad Vn ◦ diag(L(n) 1 ⊕ ϕ(n) n , ψ) on F0. Denote by, for each integer k ≥ 1, en,k,0 = diag( 1An, 1An , ..., 1An ) ∈ An ⊗ K = Bn, 2 ⊕ ϕ(n) { k n , ψ) ≈ε0/3 diag(L(n) } { z K z e′n,k = diag(1An, en,k,0, en,k,0, ..., en,k,0) ∈ P M1+r(Bn)P and e′′n,k = diag( z e′n,k, e′n,k, ..., e′n,k) ∈ MK (P M1+r(Bn)P ). } r } { (e 4.134) (e 4.135) It should be noted that e′′n,k commutes with ψ and e′n,k commutes with ϕ(0) n . Put en,k = e′n,k⊕e′′n,k in M1+K (P M1+r(Bn)P ). Then {en,k} forms an approximate identity for M1+K (P M1+r(Bn)P ). Note that Vn ∈ M1+K(P Mr+1( B)P ). It is easy to check that lim k→∞k[Vn, en,k]k = 0. (e 4.136) It follows that there exists a unitary Un,k ∈ en,kM1+K(P M1+r(Bn)P )en,k for each n and k such that lim k→∞ken,kVnen,k − Un,kk = 0. (e 4.137) For each k, there is N (k) = rk + K(rk + 1) such that MN (k)(An) = ((e′n,k − 1An) ⊕ e′′n,k)M1+K (P Mr+1(Bn)P )((e′n,k − 1An) ⊕ e′′n,k). (e 4.138) Moreover en,kM1+K (P M1+r(Bn)P )en,k = MN (k)+1(An). Define Ψn(c) = (e′n,k−1An)ϕ(n) 1An) ⊕ e′′n,kψ(c)en,k for c ∈ C. Then, for all large k and large n, 0 (c)(e′n,k− Ad Un ◦ diag(L(n) 1 , Ψn) ≈ε0/2 diag(L(n) 2 , Ψn) on F0. (e 4.139) This gives a contradiction. Theorem 4.17. Let A ∈ ¯Ds be a unital subhomogeneous C∗-algebra and let ∆ : Aq,1 (0, 1) be an order preserving map. + \ {0} → For any ε > 0 and finite subset F ⊂ A, there exists δ > 0, a finite subset P ⊂ K(A), a finite If ϕ1, ϕ2 : A → Mn are two unital homomorphisms such that + \ {0} and a finite subset H2 ⊂ As.a. satisfying the following: subset H1 ⊂ A1 [ϕ1]P = [ϕ2]P , τ ◦ ϕ1(g) ≥ ∆(g) for all g ∈ H1 and τ ◦ ϕ1(h) − τ ◦ ϕ2(h) < δ for all h ∈ H2, then there exists a unitary u ∈ Mn such that kAd u ◦ ϕ1(f ) − ϕ2(f )k < ε for all f ∈ F. 40 (e 4.140) (e 4.141) (e 4.142) (e 4.143) Proof. If A has finite dimensional, the lemma is known. So, in what follows, we will assume that A is infinite dimensional. Define ∆0 : Aq,1 + \ {0} → (0, 1) by ∆0 = (3/4)∆. Fix ε > 0 and a finite subset F ⊂ A. Let + \ {0} (in place of H) be a finite subset and an integer P ⊂ K(A) be a finite subset, H0 ⊂ Aq,1 K ≥ 1 be as required by 4.14 for ε/2 (in place of ε), F and ∆0. Choose ε0 > 0 and a finite subset G ⊂ A such that ε0 < ε and [Φ′1]P = [Φ′2]P for any pair of unital homomorphisms from A, provided that kΦ′1(g) − Φ′2(g)k < ε0 for all g ∈ G. (e 4.144) (e 4.145) Let α = 3/4. Let N ≥ 1 be an integer, δ1 > 0 (in place of δ), H1 ⊂ A1 We may assume that F ⊂ G and ε0 < ε/2. + \ {0} be a finite subset and H2 ⊂ As.a. be a finite subset as required by 4.11 for ε0/2 (in place ε), G (in place of F), H0, K and ∆0 (in place of ∆). By choosing larger H1, since A has infinite dimensional, we may assume that H1 contains at least N many mutually orthogonal non-zero positive elements. Now suppose that ϕ1, ϕ2 are two unital homomorphisms satisfying the assumption for the above P, H1 and H2. The assumption (e 4.141) implies that n ≥ N. By applying 4.11, we obtain a unitary u1 ∈ Mn, mutually orthogonal non-zero projections e0, e1, e2, ..., eK ∈ Mn i=0 ei = 1Mn, e0 . e1, e1 are equivalent to ei, i = 1, 2, ..., K, unital homomorphisms Φ1, Φ2 : A → e0Mne0 and a unital homomorphism ψ : A → e1Mne1 such that with PK kAd u1 ◦ ϕ1(f ) − (Φ1(f ) ⊕ Ψ(f ))k < ε0/2 for all f ∈ G, kϕ2(f ) − (Φ2(f ) ⊕ Ψ(f ))k < ε0/2 for all f ∈ G and τ ◦ ψ(g) ≥ (3/4)∆( g)/K for all g ∈ H0, { } z ψ(a), ψ(a), ..., ψ(a)) for all a ∈ A and τ is the tracial state on Mn. Since [ϕ1]P = [ϕ2]P , by the choice of ε0 and G, we compute that where Ψ(a) = diag( K Moreover, [Φ1]P = [Φ2]P . t ◦ ψ(g) ≥ (3/4)∆(g) for all g ∈ H0, (e 4.146) (e 4.147) (e 4.148) (e 4.149) (e 4.150) if t is the tracial state of e1Mne1. By 4.14, there is a unitary u2 ∈ Mn such that kAd u2 ◦ (Φ1 ⊕ Ψ)(f ) − (Φ1 ⊕ Ψ(f ))k < ε/2 for all f ∈ F. (e 4.151) Put U = u2u1. Then, by (e 4.146), (e 4.147) and (e 4.151), kAd U ◦ ϕ1(f ) − ϕ2(f )k < ε0/2 + ε/2 + ε0/2 < ε for all f ∈ F. (e 4.152) Lemma 4.18. Let A ∈ Ds be a unital C∗-algebra and let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. Let P0 ⊂ K0(A) be a finite subset. Then there exists an integer N (P0)≥ 1 and a finite subset H ⊂ A1 +\{0} satisfying the following: For any unital homomorphism ϕ : A → Mk 41 (for some k ≥ 1) and any unital homomorphism ψ : A → MR for some integer R ≥ N (P0)k such that τ ◦ ψ(g) ≥ ∆(g) for all g ∈ H, there exists a unital homomorphism h0 : A → MR−k such that [ϕ ⊕ h0]P0 = [ψ]P0 . (e 4.153) (e 4.154) Proof. Let G0 be a subgroup of K0(A) generated by P0. We may also assume, without loss of generality, that P0 = {[p1], [p2], ..., [pm1 ]} ∪ {z1, z2, ..., zm2}, where p1, p2, ..., pm1 ∈ Ml(A) are projections and zj ∈ kerρA, j = 1, 2, ..., m2. We prove the lemma by induction. Assume first A = P C(X, F )P , where X is path connected. This, of course, includes the case that X is a single point. There is d > 0 such that kπx,j ◦ pi − πx′,j ◦ pik < 1/2, i = 1, 2, ..., m1, (e 4.155) provided that dist(x, x′) < d, where πx,j is identified with πx,j ⊗ idMl. Since X is compact, we may assume that {x1, x2, ..., xm3} is a d/2-dense set. Write PxiF Pxi = Mr(i,1) ⊕ Mr(i,2) ⊕ ··· ⊕ Mr(i,k(xi)), i = 1, 2, ..., m3. There are hi,j ∈ C(X) with 0 ≤ hi,j ≤ 1, hi,j(xi) = 1Mr(i,j) and hi,jhi′,j′ = 0 if (i, j) 6= (i′, j′). Put gi,j = hi,j · P ∈ A, j = 1, 2, ..., k(xi), i = 1, 2, ..., m3, Let Moreover we assume that hi,j(x) = 0 if dist(x, xi) ≥ d. σ0 = min{∆( hi,j) : 1 ≤ j ≤ k(xi), 1 ≤ i ≤ m3} and N (P0) ≥ 2/σ0. (e 4.156) Put H = {hi,j : 1 ≤ j ≤ k(xi), 1 ≤ i ≤ m3}. Now suppose that ϕ : A → Mk and ψ : A → MR with R ≥ N (P0)k and τ ◦ ψ(g) ≥ ∆(g) for all g ∈ H. (e 4.157) Write ϕ = Lm3 i,j Πyi,j, where Πyi,j(i) is Ti,j copies of πyi,j. Note k − Ti > 0 for all i. Since R ≥ N (P0)k, (e 4.157) implies that ψ may be viewed as direct sum of at least ∆(chj,i) · (2k/σ0) > 2k (e 4.158) copies of πx,j with dist(x, xi) < d, i = 1, 2, ..., m3. Rewrite ψ = Σ1 ⊕ Σ2, where Σ1 contains exactly Ti,j copies of πx,j with dist(x, xi) < d for each i and j. Then rank Σ1(pi) = rankϕ(pi), i = 1, 2, ..., m1. (e 4.159) Put h0 = Σ2. Note for any unital homomorphism h : A → Mn, [h(z)] = 0 for all z ∈ kerρA. This proves the case that A = P C(X, F )P as above, in particular, the case that A ∈ D0. Now assume the lemma holds for any C∗-algebra A ∈ Dm. Let A be a C-algebra in D′m+1. We assume that A ⊂ P C(X, F )P ⊕ B is a unital C∗- subalgebra and I = {f ∈ P C(X, F )P : fX 0 = 0}, where X 0 = X \ Y and Y is an open subset of X and B ∈ D′m and A/I ∼= B. We assume that, if dist(x, x′) < 2d, kπx,j(pi) − πx′,j(pi)k < 1/2 and kπx,j ◦ s ◦ πI (pi) − πx,j(pi)k < 1/2, (e 4.160) where s : A/I → Ad = {fX d : f ∈ A} is an injective homomorphism given by 4.8. We also assume that 2d < dY . Define ∆π : (A/I)q,1 + \ {0} → (0, 1) by ∆π(g) = ∆( \g0 · P · s(g)) for all g ∈ (A/I)1 + \ {0}, (e 4.161) 42 where g0 ∈ C(X d)+ with 0 ≤ g ≤ 1, g0(x) = 1 if x ∈ X 0, g0(x) > 0 if dist(x, X 0) < d/2 and g0(x) = 0 if dist(x, X d) ≥ d/2}. Note that g0 · s(g) > 0 if g ∈ (A/I)+ \ {0}. Therefore ∆π is indeed an order preserving map from (A/I)q,1 + \ {0} into (0, 1). Note that A/I ∈ D′m. By the inductive assumption, there is an integer Nπ(P0) ≥ 1, a finite + \{0} satisfying the following: if ϕ′ : A/I → Mk′ is a unital homomorphism subset Hπ ⊂ (A/I)1 and ψ′ : A/I → MR′ is a unital homomorphism for some R′ > Nπ(P0)k′ such that t ◦ ψ′(g) ≥ ∆π(g) for all g ∈ Hπ, where t is the tracial state of NR′, then there exists a unital homomorphism hπ : A/I → MR′−k′ such that (ϕ′ ⊕ hπ)∗0 ¯P0 = (ψπ)∗0 ¯P0, where ¯P0 = {(πI )∗0(p) : p ∈ P0}. Let r : Y d/2 → Y d be a homeomorphism such that dist(r(x), x) < d for all x ∈ Y d/2. Let C = {fY d : f ∈ I}. Define ∆I : C q,1 + \ {0} → (0, 1) by ∆I(g) = ∆( \f0 · g ◦ r) for all g ∈ C q,1 + \ {0}, (e 4.162) where f0 ∈ C0(Y )+ with 0 ≤ f0 ≤ 1, f0(x) = 1 if x ∈ Y d, f0(x) = 0 if dist(x, X 0) ≤ d/2 and f0(x) > 0 if dist(x, X 0) > d/2}. Note that C = PY dC(Yd, F )PY d. By what has been proved, there is an integer NI (P0) ≥ 1 and a finite subset HI ⊂ C 1 is a unital homomorphism and ψ′′ : C → MR′′ (for some R′′ ≥ NI (P0)k′′) is another unital homomorphism such that + \ {0} satisfying the following: if ϕ′′ : C → Mk′′ where t is the tracial state on MR′′, then there exists a unital homomorphism h′′ : C → MR′′−k′′ such that t ◦ ψ′′(g) ≥ ∆I (g) for all g ∈ HI , (ϕ′′ ⊕ h′′)∗0P0 = (ψ′′∗0)P0 . (e 4.163) Put σ = min{min{∆π(g) : g ∈ Hπ}, min{∆I (g) : g ∈ Hπ}} > 0. (e 4.164) Let N = (Nπ(P0) + NI (P0))/σ and let H = {g0 ◦ s(g) : g ∈ Hπ} ∪ {f0 · g ◦ r}. (e 4.165) Now suppose that ϕ and ψ satisfy the assumption for N = N (P0) and H as above. We may write ϕ = Σϕ,π⊕Σϕ,I, where Σϕ,π corresponds to a finite direct sum of irreducible representations of A which factors through A/I and Σϕ,I corresponds to the finite direct sum of irreducible representations of I. We also write ψ = Σψ,π ⊕ Σψ,b ⊕ Σψ,I ′, (e 4.166) where Σψ,π corresponds to the finite direct sum of irreducible representations of A which factors through A/I, Σψ,b which corresponds to the finite direct sum of irreducible representations which factors through point-evaluations at x ∈ Y with dist(x, X 0) < d/2 and Σψ,I ′ corresponds the rest of irreducible representations (which can be factors through point-evaluations at x ∈ Y with dist(x, X 0) ≥ d/2). 43 Put qπ = (Σψ,π ⊕ Σψ,b)(1A) and k′ = rankΣϕ,π(1A). Define ψπ : A/I → Mrank(qπ) by ψπ(a) = (Σψ,π ⊕ Σψ,b) ◦ πI ◦ s(a) for all a ∈ A/I. Then t ◦ ψπ(g) ≥ τ ◦ ψπ(g) ≥ ∆( \g0 · s(πI (g))) = ∆π(g) for all g ∈ Hπ. (e 4.167) where t is the tracial state on Mrank(qπ ). Note that Therefore τ ◦ ψ(g0) ≥ ∆( g0) = ∆π(d1A/I ), rank(qπ) ≥ R∆π(1A/I ) ≥ Nπ(P0)k′. (e 4.168) (e 4.169) By the inductive assumption, there is a unital homomorphism hπ : A/I → Mrankqπ−k′ such that (e 4.170) (Σϕ,π ⊕ hπ)∗0P0 = (ψπ)∗0P0. Put qI = Σψ,I ′(1A) and k′′ = rank(Σϕ,I(1A)). Define ψI : C → MrankqI by ψI (a) = Σψ,I ′(a ◦ r) for all a ∈ C. (e 4.171) Then t ◦ ψI (g) ≥ τ ◦ Σψ,I ′(a ◦ r) ≥ ψ(f0 · a ◦ r) ≥ ∆( \f0 · g ◦ r) = ∆I (g) for all g ∈ HI , (e 4.172) where t is the tracial state on Mrank(qI ). Note that Therefore τ ◦ ψ(f0) ≥ ∆( f0) = ∆I( 1A). rank(qI ) ≥ R∆I(1A) ≥ NI(P0)k′′. (e 4.173) (e 4.174) There is 0 < d1 < d < dY such that all irreducible representations appeared in Σϕ,I factor through point-evaluations at x with dist(x, X d) ≥ d1. Put r′ : Y d1 → Y d. Define ϕI : C → Mk′′ by ϕI (f ) = Σϕ,I(f ◦ r′). By the inductive assumption, there is a unital homomorphism hI : C → Mrank(qI )−k′′ such that (Σϕ,I ⊕ hI )∗0P0 = (ψI )∗0P0. (e 4.175) Define h : A → MR−k by h(a) = hπ(πI (a)) ⊕ hI (aY d) ⊕ Σψ,b(a) for all a ∈ A. Then, for each i, (e 4.176) rankϕ(pi) + rankh(pi) = rank(Σϕ,π(pi)) + rank(Σϕ,I (pi)) + rank(Σψ,b(pi)) + rankhπ(pi) + rankhI (pi) = rankψπ(pi) + rank(Σψ,b(pi)) + rankψI (pi) = rankψ(pi), i = 1, 2, ..., m1. Since (ϕ)∗0(zj) = h∗0(zj) = ψ∗0(zj), j = 1, 2, ..., m2, we conclude that This completes the induction process. (ϕ ⊕ h)∗0P0 = ψ∗0P0. 44 (e 4.177) (e 4.178) (e 4.179) (e 4.180) Lemma 4.19. Let A ∈ Ds be a unital C∗-algebra and let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. For any ε > 0 and any finite subset F, there exist a finite subset H ⊂ A1 + \ {0} and an integer L ≥ 1 satisfying the following: For any unital homomorphism ϕ : A → Mk and any unital homomorphism ψ : A → MR for some R ≥ Lk such that tr ◦ ψ(h) ≥ ∆(h) for all h ∈ H, there exist a unital homomorphism ϕ0 : A → MR−k and a unitary u ∈ MR such that kAd u ◦ diag(ϕ(f ), ϕ0(f )) − ψ(f )k < ε (e 4.181) (e 4.182) for all f ∈ F. Proof. Let δ > 0, P ⊂ K(A) be a finite subset, H1 ⊂ A1 + \ {0} be a finite subset, H2 ⊂ As.a. be a finite subset and N0 be an integer as required by 4.17 for ε/4 (in place of ε), F, (1/2)∆ and A. Without loss of generality, we may assume that H2 ⊂ A1 + \ {0}. Let σ0 = min{min{∆(g) : g ∈ H1}, min{∆(g) : g ∈ H2}} Let G be a subgroup of K(A) generated by P. Put P0 = P ∩ K0(A). We may also assume, without loss of generality, that P0 = {[p1], [p2], ..., [pm1 ]} ∪ {z1, z2, ..., zm2}, where p1, p2, ..., pm1 are projections in Ml(A) and zj ∈ kerρA, j = 1, 2, ..., m2. Let j ≥ 1 be an integer such that K0(A, Z/j′Z) ∩ G = ∅ for all j′ ≥ j. Put J = j!. Let N (P0) ≥ 1 be an integer and H3 ⊂ A1 + \ {0} be a finite subset as required by 4.18 for P0. Let ps = (a(s) i,j )l×l, s = 1, 2, ..., m1, and choose ε0 > 0 and a finite subset F0 such that [ψ′]P = [ψ′′]P , (e 4.183) provided that kψ′(a) − ψ′′(a)k < ε0 for all a ∈ F0. Put F2 = F ∪ F1 ∪ H2 and put ε1 = min{ε/16, ε1/2}. Let K > 8((N (P0) + 1)(J + 1)/δσ) be an integer. Let H0 = H1 ∪ H3. Let N1 ≥ 1 (in place of N ) be an integer, δ1 > 0 (in place of δ), H4 ⊂ A1 + \ {0} (in place of H1) be a finite subset and H5 ⊂ As.a. (in place of H2) be a finite subset as required by 4.11 for ε1 (in place of ε) F2 (in place of F), H0 and K. Let L = K(K + 1) and let H = H4 ∪ H0 as well as α = 15/16. Suppose that ϕ and ψ satisfy the assumption for the above L and H. Then, by applying 4.11, there are mutually orthogonal projections e0, e1, e2, ..., eK ∈ MR such that e0 . e1, ei is equivalent to e1, i = 1, 2, ..., K, a unital homomorphism ψ0 : A → e0MRe0 and a unital homomorphism ψ1 : A → e1MRe1 such that kψ(a) − (ψ0(a) ⊕ diag( z K } { ψ1(a), ψ1(a), ..., ψ1(a)))k < ε1 for all a ∈ F2 and τ ◦ ψ1(g) ≥ (15/16) for all g ∈ H0. ∆(g) K Put Ψ = ψ0 ⊕ diag( ψ1(a), ψ1(a), ..., ψ1(a))). We compute that z K } { [Ψ]P = [ψ]P . Let R0 = rank(e1). Then R0 = Rτ ◦ ψ1(1A) ≥ Lk(15/16) ∆(c1A) K ≥ k(K + 1)(15/16)∆(c1A) ≥ k(15/16)8N (P0)(J + 1)/δ. 45 (e 4.184) (e 4.185) (e 4.186) (e 4.187) (e 4.188) Moreover, t ◦ ψ1(g) ≥ (15/16)∆(g) for all g ∈ H3, (e 4.189) where t is the tracial state on MR1. It follows from 4.18 that there exists a unital homomorphism h0 : A → MR0−k such that (ϕ ⊕ h0)∗0P0 = (ψ1)P0. Put Then h1 = h0 ⊕ diag( ϕ ⊕ h0, ϕ ⊕ h0, ..., ϕ ⊕ h0) and ψ2 = diag( ψ1, ψ1, ..., ψ1) z J } { z J−1 } { [ϕ ⊕ h1]P = [ψ2]P . Put Ψ′ = diag( K−J } z ψ1, ψ1, ..., ψ1). Let ϕ0 = h1 ⊕ ψ0 ⊕ Ψ′. Then { [ϕ ⊕ ϕ0]P = [ψ0 ⊕ ψ2 ⊕ Ψ′]P = [ψ]P (e 4.190) (e 4.191) (e 4.192) (e 4.193) Since J/(K − J) < δ and by (e 4.181), by applying 4.17, there is a unitary u ∈ MR such that kAd u ◦ (ϕ(f ) ⊕ ϕ0(f )) − ψ(f )k < ε for all f ∈ F. (e 4.194) 5 Almost multiplicative maps to finite dimensional C∗-algebras Note that in the following statement, n is given and (G, δ) depends on n. Lemma 5.1. Let n ≥ 1 be an integer and let A be a unital separable C∗-algebra. For any ε > 0 and any finite subset F ⊂ A, there exist δ > 0 and a finite subset G ⊂ A such that, for any unital G-δ-multiplicative contractive completely positive linear map ϕ : A → Mn, there exists a unital homomorphism ψ : A → Mn such that kϕ(a) − ψ(a)k < ε for all a ∈ F. Proof. Suppose that the lemma is not true for certain finite set F⊂A and ε0> 0. Let {Gk}∞k=1 be a sequence of finite subsets of A with Gk ⊂ Gk+1 and ∪kGk = A and let {δk} be a monotone decreasing sequence of positive numbers with δk → 0. Since the lemma is assumed not to be true, there are unital Gk-δk-multiplicative contractive completely positive linear maps ϕk : A → Mn such that inf{maxa∈Fkϕk(a) − ψ(a)k : ψ : A → Mn homomorphisms} ≥ ε0. (e 5.1) For each pair (i, j) with 1 ≤ i, j ≤ n, let li,j : Mn → C be the map defined by taking matrix a ∈ Mn to the entry of ith row and jth column of a. Let ϕi,j k = li,j ◦ ϕk : A → C. Note that the unit ball of the dual space of A (as Banach space) is weak∗-compact. Since A is separable, there is a subsequence (instead of subnet) of {ϕk} (still denote by ϕk) such that {ϕi,j k } is weak∗ convergent for all i, j. In other words, {ϕk} convergent pointwise. Let ψ0 be the the limit. Then ψ0 is a homomorphism and for k large enough, we have This is a contradiction to (e 5.1) above. kϕk(a) − ψ0(a)k < ε0, for all a ∈ F. 46 Lemma 5.2. (cf. Lemma 4.5 of [58]) Let A be a unital C∗-algebra arising from a locally trivial continuous field of Mn over a compact metric space X. Let T be a finite subset of tracial states on A. For any finite subset F ⊂ A and for any ε > 0 and any σ > 0, there is an ideal J ⊂ A such that kτJk < σ for all τ ∈ T, a finite dimensional C∗-subalgebra C ⊂ A/J and a unital homomorphism π0 from A/J such that dist(π(x), C) < ε for all x ∈ F and π0(A/J) = π0(C) ∼= C, (e 5.2) j=1 πxj . where π : A → A/J is the quotient map. Proof. This follows from Lemma 4.5 of [58]. Choose xi ∈ Fi j = 1, 2, ..., k. One can then choose π0 =Lk Lemma 5.3. (cf. Lemma 4.7 of [58]) Let A be a unital separable subhomogeneous C-algebra. Let T ⊂ T (A) be a finite subset. For any finite subset F ⊂ A, ε > 0 and σ > 0, there is an ideal J ⊂ A such that kτJk < σ for all τ ∈ T, a finite dimensional C∗-subalgebra C ⊂ A/J and a unital homomorphism π0 from A/J such that dist(π(x), C) < ε for all x ∈ F and π0(A/J) = π0(C) ∼= C, (e 5.3) where π : A → A/J is the quotient map. Proof. The proof is in fact contained in that of Lemma 4.7 of [58]. Each time Lemma 4.5 of [58] applied, one can apply 5.2 instead. The complete proof is omitted here. Lemma 5.4. Let A be a unital subhomogeneous C-algebra. Let ε > 0, let F ⊂ A be a finite subset and let σ0 > 0. There exist δ > 0 and a finite subset G ⊂ A satisfying the following: Suppose that ϕ : A → Mn (for some integer n ≥ 1) is a G-δ-multiplicative contractive completely positive linear map. Then, there exists a projection p ∈ Mn and a unital homomorphism ϕ0 : A → pMnp such that kϕ(a) − [(1 − p)ϕ(a)(1 − p) + ϕ0(a)]k < ε for all a ∈ F and kpϕ(a) − ϕ(a)pk < ε for all a ∈ F, tr(1 − p) < σ0, (e 5.4) (e 5.5) (e 5.6) where tr is the normalized trace on Mn. Proof. We assume that the lemma is false. Then there exists ε0 > 0, a finite subset F0, a positive number σ0 > 0, an increasing sequence of finite subsets Gn ⊂ A such that Gn ⊂ Gn+1 and such sequence of integers {m(n)} and a sequence of unital Gn-δn-multiplicative contractive completely positive linear maps ϕn : A → Mm(n) satisfying the following: that ∪n=1Gn is dense in A, a sequence of decreasing positive numbers {δn} withP∞n=1 δn < ∞, a inf{max{kϕn(a) − [(1 − p)ϕn(a)(1 − p) + ϕ0(a)k : a ∈ F0}} ≥ ε0 (e 5.7) where infimum is take among all projections p ∈ Mm(n) with trn(1 − p) < σ0, where trn is the normalized trace on Mm(n) and all possible homomorphisms ϕ0 : A → pMm(n)p. By the virtue of 5.1, one may also assume that m(n) → ∞ as n → ∞. Note that {trn ◦ ϕn} is a sequence of (not necessary tracial) states of A. Let t0 be a weak limit of {trn ◦ ϕn}. Since A is separable, there is a subsequence (instead of subnet) of {trn ◦ ϕn} converging to t0. Without loss of generality, we may assume that trn ◦ ϕn converges to t0. By the Gn-δn-multiplicativity of ϕn, we know that t0 is a tracial state on A. 47 ∞Mn=1 ({Mm(n)}) = {{an} : an ∈ Mm(n) and lim n→∞kank = 0}. Consider the idealL∞n=1({Mm(n)}), where Denote by Q the quotient Q∞n=1({Mm(n)}/L∞n=1({Mm(n)}). Let πω :Q∞n=1({Mm(n)}) → Q be the quotient map. Let A0 = {πω({ϕn(f )}) : f ∈ A} which is a subalgebra of Q. Then Ψ is a unital homomorphism from A to Q∞n=1(Mm(n))/L({Mm(n)}) with Ψ(A) = πω(A0). If a ∈ A has zero image in πω(A0), that is, ϕn(a) → 0, then t0(a) = limn→∞ trn(ϕn(a)) = 0. So we may view t0 as a state on πω(A0) = Ψ(A). It follows from Lemma 5.3 that there is an ideal I ⊂ Ψ(A) and a finite dimensional C∗- subalgebra B ⊂ Ψ(A)/I and a unital homomorphism π00 : Ψ(A)/I → B such that dist(πI ◦ Ψ(f ), B) < ε0/16 for all f ∈ F0, k(t0)Ik < σ0/2 and π00B = idB. (e 5.8) (e 5.9) Note that π00 can be regarded as map from A to B with kerπ00 ⊃ I. There is, for each f ∈ F0, an element bf ∈ B such that kπI ◦ Ψ(f ) − bfk < ε0/16. (e 5.10) Put C′ = B + I and I0 = Ψ−1(I) and C1 = Ψ−1(C′). For each f ∈ F0, there exists af ∈ C1 ⊂ A such that kf − afk < ε0/16 and πI ◦ Ψ(af ) = bf . (e 5.11) Let a ∈ (I0)+ be a strictly positive element and let J = Ψ(a)QΨ(a) be the hereditary C∗- subalgebra of Q generated by Ψ(a). Put C2 = Ψ(C1) + J. Then J is an ideal of C2. Denote by πJ : C2 → B the quotient map. Since Q and J have real rank zero and C2/J has finite dimension, by Lemma 5.2 of [51], C2 has real rank zero. It follows that 0 → J → C2 → B → 0 is a quasidiagonal extension. As in Lemma 4.9 of [58], there is a projection P ∈ J and a unital homomorphism ψ0 : B → (1 − P )C2(1 − P ) such that kP Ψ(af ) − Ψ(af )Pk < ε0/8 and kΨ(af ) − [P Ψ(af )P + ψ0 ◦ πJ ◦ Ψ(af )]k < ε0/8 (e 5.12) for all f ∈ F0. Let H : A → ψ0(B) be defined by H = ψ0 ◦ π00 ◦ πI ◦ Ψ. One estimates that kP Ψ(f ) − Ψ(f )Pk < ε0/2 and kΨ(f ) − [P Ψ(f )P + H(f )]k < ε0/2 (e 5.13) (e 5.14) H(A) → Q∞n=1({Mm(n)}) such that π ◦ H1 ◦ H = H. One may write H1 = {hn}, where each for all f ∈ F0. Note that dimH(A) < ∞, and that H(A)⊂Q. There is a homomorphism H1 : hn : H(A) → Mm(n) is a (not necessary unital) homomorphism, n = 1, 2, .... There is also a sequence of projections qn ∈ Mm(n) such that π({qn}) = P. Let pn = 1 − qn, n = 1, 2, .... Then, for sufficiently large n, by (e 5.13) and (e 5.14), k(1 − pn)ϕn(f ) − ϕn(f )(1 − pn)k < ε0 and kϕn(f ) − [(1 − pn)ϕn(f )(1 − pn) + hn ◦ H(f )]k < ε0 (e 5.15) (e 5.16) 48 for all f ∈ F0. Moreover, since P ∈ J, for any η > 0, there is b ∈ I0 with 0 ≤ b ≤ 1 such that However, by (e 5.9), kΨ(b)P − Pk < η. 0 < t0(Ψ(b)) < σ0/2 for all b ∈ I0 with 0 ≤ n ≤ 1. (e 5.17) By choosing sufficiently small η, for all sufficiently large n, This contradicts with (e 5.7). trn(1 − pn) < σ0. Corollary 5.5. Let A be a unital subhomogeneous C-algebra. Let η > 0, let E ⊂ A be a finite subset and let η0 > 0. There exist δ > 0 and a finite subset G ⊂ A satisfying the following: Suppose that ϕ, ψ : A → Mn (for some integer n ≥ 1) are two G-δ-multiplicative contractive completely positive linear maps. Then, there exist projections p, q ∈ Mn with rank(p) = rank(q) and unital homomorphisms ϕ0 : A → pMnp and ψ0 : A → qMnq such that a ∈ E, kpϕ(a) − ϕ(a)pk < η, kϕ(a) − [(1 − p)ϕ(a)(1 − p) + ϕ0(a)]k < η, kqψ(a) − ψ(a)qk < η, kψ(a) − [(1 − q)ψ(a)(1 − q) + ψ0(a)]k < η, a ∈ E where tr is the normalized trace on Mn. and tr(1 − p) = tr(1 − q) < η0, For convenience of future use, we used η, η0 and E to replace ε, σ0 and F in 5.4. Proof. By 5.4, we can get such decomposition for ϕ and ψ separately, then the only missing part is that rank(p) = rank(q). Let {z1, z2, ..., zm} be the set of ranks of irreducible representations of A and let T be the number given by 4.1 corresponding to {z1, z2, ..., zm}. We apply 5.4 to η0/2 instead of σ0 (and, η and E in places of ε and F). By 5.1, we can assume the size n of matrix Mn is so large that T n < η0/2. By 4.1, we can take sub-representations out of ϕ0 and ψ0 (one of them has size at most T ) so that the remainder of ϕ0 and ψ0 have same size—that is for rank(new p) = rank(new q), and tr(1 − (new p)) = tr(1 − (new q)) < η0 Lemma 5.6. Let A ∈ Ds be an infinite dimensional unital C∗-algebra, let ε > 0 and let F ⊂ A be a finite subset. let ε0 > 0 and let G0 ⊂ A be a finite subset. Let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. + \ {0} is a finite subset, ε1 > 0 is a positive number and K ≥ 1 is an integer. There exists δ > 0, σ > 0 and a finite subset G ⊂ A and a finite subset H2 ⊂ A1 + \ {0} satisfying the following: Suppose that L1, L2 : A → Mn (for some integer n ≥ 1) are unital G-δ-multiplicative contractive completely positive linear maps Suppose that H1 ⊂ A1 2 + T n < η0. tr ◦ L1(h) ≥ ∆(h) and tr ◦ L2(h) ≥ ∆(h) for all h ∈ H2, and tr ◦ L1(h) − tr ◦ L2(h) < σ for all h ∈ H2. (e 5.18) (e 5.19) Then there exist mutually orthogonal projections e0, e1, e2, ..., eK ∈ Mn such that e1, e2, ..., eK i=1 ei = 1, and there exist unital are pairwise equivalent, e0 . e1, tr(e0) < ε1 and e0 +PK 49 G0-ε0-multiplicative contractive completely positive linear maps ψ1, ψ2 : A → e0Mke0, a unital homomorphism ψ : A → e1Mke1, and unitary u ∈ Mn such that ψ(f ), ψ(f ), ..., ψ(f ))k < ε and K } { kL1(f ) − diag(ψ1(f ), z kuL2(f )u∗ − diag(ψ2(f ), z K } ψ(f ), ψ(f ), ..., ψ(f ))k < ε { (e 5.20) (e 5.21) (e 5.22) for all f ∈ F, where tr is the tracial state of Mn. Moreover, for all g ∈ H1. tr(ψ(g)) ≥ ∆(g) 3K Proof. First note that the following statement is evident. For any C∗-algebra A, a finite subset G0 ⊂ A and ε0 > 0, there are finite subset G′ ⊂ A with δ′ > 0, F′ ⊂ A and with ε′ > 0 satisfying If L : A → B is a unital G′-δ′-multiplicative contractive completely the following condition. positive linear map , p0, p1 ∈ B are projections with p0 + p1 = 1B, and L′0 : A → p0Bp0, L′1 : A → p1Bp1 are linear maps with kL(f ) − diag(L′0(f ), L′1(f ))k < ε′ for all f ∈ F′, then both L′0 and L′1 are ε0-G0-multiplicative. Therefore if ε is sufficiently small and F is sufficiently large relative to (ε0,G0), then (e 5.20) and (e 5.20) imply ψ1 and ψ2 are G0-ε0- multiplicative. Put ε1 = min{ε/16, ε′/16, 1 64K min{∆(h), h ∈ H1}}. (e 5.23) + \ {0}. Let ∆1 = (3/4)∆. Let δ1 > 0 (in place of δ), H1,0 ⊂ A1 + \{0} (in place of H1) be a finite subset, H2,0 ⊂ As.a. (in place of H2) be a finite subset as required by 4.11 (see also its remark 4.12) for ε1 (in place of ε), F ∪ F′ (in place of F), 2K (in place of K), ∆1 and A as well as α = (3/4). We may assume that H0,2 ⊂ A1 Let η0 = min{δ1/16, ε1/16, min{τ (h) : h ∈ H1,0 ∪ H2,0}. Let δ2 > 0 (in place of δ) and let G1 ⊂ A (in place of G) be the finite subset as required by 5.5 for η = η0 · min{ε1, δ1/4}, η0 and E = F ∪ H1,0 ∪ H2,0 ∪ H1. Let δ = η0 · min{δ2/2, δ1/2,}, σ = min{η0/2, η1/2}, let G = G′ ∪ G1 ∪ F ∪ F′ ∪ E and let H2 = H1,0 ∪ H2,0 ∪ H1. Now suppose that L1 and L2 satisfy the assumption of the lemma with respect to the above δ, σ and G and H2. It follows from 5.5 that there exists a projection p ∈ Mn, two unital homomorphisms ϕ1, ϕ2 : A → pMnp and a unitary u1 ∈ Mn such that kAd u1 ◦ L1(a) − ((1 − p)u∗1L1(a)u1(1 − p) + ϕ1(a))k < η, kL2(a) − ((1 − p)L2(a)(1 − p) + ϕ2(a)k < η for all a ∈ E and where τ is the tracial state on Mn. We compute that τ (1 − p) < η0, τ ◦ ϕ1(g) ≥ ∆(g) − η − η0 ≥ (3/4)∆(g) for all g ∈ H1,0 and τ ◦ ϕ1(g) − τ ◦ ϕ2(g) < 2η + 2η0 + δ < η0δ1 for all g ∈ H2,0. 50 (e 5.24) (e 5.25) (e 5.26) (e 5.27) (e 5.28) It follows from 4.11 (and its remark (4.12)) that there exist mutually orthogonal projections q0, q1, ..., q2K ∈ pMnp such that q0 . q1 and qi is equivalent to q1 for all i = 1, 2, ..., 2K, two unital homomorphisms ϕ1,0, ϕ2,0 : A → q0Mnq0, a unital homomorphism ψ′ : A → e1Mne1 and a unitary u2 ∈ pMnp such that 2K kAd u2 ◦ ϕ1(a) − (ϕ1,0 ⊕ diag( and kϕ2(a) − (ϕ2,0 ⊕ diag( z ψ′(a), ψ′(a), ..., ψ′(a)))k < ε1 z ψ′(a), ψ′(a), ..., ψ′(a)))k < ε1 { { τ ◦ ψ(g) ≥ (3/4)2∆(g)2K for all g ∈ H1. } } 2K for all a ∈ F ∪ F′. Moreover, (e 5.29) (e 5.30) (e 5.31) Let u = ((1 − p) + u2)u1, e0 = (1 − p) ⊕ q0, ei = q2i−1 ⊕ q2i, i = 1, 2, ..., K, let ψ1(a) = (1− p)u∗1L1(a)u1(1− p)⊕ ψ1,0, ψ2(a) = (1− p)L2(a)(1− p)⊕ ψ2,0 and ψ(a) = diag(ψ′(a)⊕ ψ′(a)) for a ∈ A. Then kAd u ◦ L1(f ) − (ψ1(f ) ⊕ diag( ψ(f ), ψ(f ), ..., ψ(f )))k < ε and kL2(f ) − (ψ2(f ) ⊕ diag( ψ(f ), ψ(f ), ..., ψ(f )))k < ε z z K K } } { { (e 5.32) (e 5.33) for all f ∈ F, τ ◦ ψ(g) ≥ for all g ∈ H1. ∆(g) 3K Moreover ψ1 and ψ1 are G0-ε0-multiplicative. Corollary 5.7. Let A0 ∈ Ds be a unital C∗-algebra, let ε > 0 and let F ⊂ A be a finite subset. Let ∆ : (A0)q,1 + \ {0} is a finite subset, σ > 0 is positive number and n ≥ 1 is + \ {0} satisfying the following: Suppose that an integer. There exists a finite subset H2 ⊂ (A0)1 ϕ : A = A0 ⊗ C(T) → Mk (for some integer k ≥ 1) is a unital homomorphism and + \ {0} → (0, 1) be an order preserving map. Suppose that H1 ⊂ (A0)1 (e 5.34) tr ◦ ϕ(h ⊗ 1) ≥ ∆(h) for all h ∈ H2. (e 5.35) Then there exist mutually orthogonal projections e0, e1, e2, ..., en ∈ Mk such that e1, e2, ..., en are i=0 ei = 1, and there exists unital homomorphisms ψ0 : A = A0 ⊗ C(T) → e0Mke0 and ψ : A = A0 ⊗ C(T) → e1Mke1 such that one may write that equivalent and Pn kϕ(f ) − diag(ψ0(f ), and tr(e0) < σ z ψ(f ), ψ(f ), ..., ψ(f ))k < ε n } { for all f ∈ F, where tr is the tracial state on Mk. Moreover, tr(ψ(g ⊗ 1)) ≥ The following is taken from 9.4 of [63]. ∆(g) 2n for all g ∈ H1. 51 (e 5.36) (e 5.37) (e 5.38) Lemma 5.8. Let A be a unital separable C∗-algebra. For any ε > 0, any finite subset H ⊂ As.a., there exists a finite subset G ⊂ A and δ > 0 satisfying the following: Suppose that ϕ : A → B (for some unital C∗-algebra B) is a G-δ-multiplicative contractive completely positive linear map and t ∈ T (B) is a tracial state of B. Then, there exists a tracial state τ ∈ T (A) such that t ◦ ϕ(h) − τ (h) < ε for all h ∈ H. (e 5.39) Theorem 5.9. Let A ∈ Ds be a unital C∗-algebra. Let ∆ : Aq,1 preserving map. Let ε > 0 and let F ⊂ A be a finite subset. There exists a finite subset H1 ⊂ A1 + \ {0} → (0, 1) be an order + \ {0}, a finite subset G ⊂ A, δ > 0, a finite subset P ⊂ K(A), a finite subset H2 ⊂ As.a. and σ > 0 satisfying the following: Suppose that L1, L2 : A → Mk (for some integer k ≥ 1) are two unital G-δ-multiplicative contractive completely positive linear maps such that [L1]P = [L2]P , tr ◦ L1(h) ≥ ∆(h) for all h ∈ H1 and tr ◦ L1(h) − tr ◦ L2(h) < σ for all h ∈ H2, then there exists a unitary u ∈ Mk such that kAd u ◦ L1(f ) − L2(f ) < ε for all f ∈ F. (e 5.40) (e 5.41) (e 5.42) (e 5.43) Proof. The proof is exactly the same as that of 4.17. As in the proof of 4.17, we will use 4.14. However, here we will use 5.6 instead of 4.11. 6 Homotopy Lemma in finite dimensional C∗-algebras Lemma 6.1. Let A be a unital separable C∗-algebra and let ϕ : A → Mk (for some integer k ≥ 1) be a unital linear map. Suppose that u ∈ Mk is a unitary such that uϕ(a) = ϕ(a)u for all a ∈ A. Then there exists a continuous path of unitaries {ut : t ∈ [0, 1]} ⊂ Mk such that u0 = u, u1 = 1, utϕ(a) = ϕ(a)ut for all a ∈ A and for all t ∈ [0, 1]. Moreover, length({ut}) ≤ π. (e 6.1) (e 6.2) (e 6.3) Proof. There is d > 0 such that spectrum of u has a gap containing an arc with length at least d. There is a continuous function h from sp(u) to [−π, π] such that Therefore exp(ih(u)) = u. ϕ(a)h(u) = h(u)ϕ(a) for all a ∈ A. (e 6.4) (e 6.5) Note that h(u) ∈ (Mk)s.a. and kh(u)k ≤ π. Define ut = exp(i(1 − t)h(u)) (t ∈ [0, 1]). Then u0 = u and u1 = 1. Also for all a ∈ A and t ∈ [0, 1]. Moreover, one has utϕ(a) = ϕ(a)ut as desired. lenghth({ut}) ≤ π, 52 Lemma 6.2. Let A ∈ Ds be a unital C-algebra, let H ⊂ (A ⊗ C(T))s.a. be a finite subset, let 1 > σ > 0 be a positive number and let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. Let ε > 0, G0 ⊂ A ⊗ C(T) be a finite subset, P0,P1 ⊂ K(A) be finite subsets and P = P0 ∪ β(P1) ⊂ K(A ⊗ C(T)). There exist δ > 0, a finite subset G ⊂ A ⊗ C(T) and a finite subset H1 ⊂ (A ⊗ C(T))1 + \ {0} satisfying the following: Suppose that L : A ⊗ C(T) → Mk (for some integer n ≥ 1) is a G-δ-multiplicative contractive completely positive linear map such that (e 6.6) tr ◦ L(h) ≥ ∆(h) for all h ∈ H1, and [L]β(P1) = 0. (e 6.7) Then there exists a unital G0-δ-multiplicative contractive completely positive linear map ψ : A ⊗ C(T) → Mk such that u = ψ(1 ⊗ z) is a unitary, uψ(a ⊗ 1) = ψ(a ⊗ 1)u for all a ∈ A [L]P = [ψ]P and tr ◦ L(h) − tr ◦ ψ(h) < σ for all h ∈ H. (e 6.8) (e 6.9) (e 6.10) Proof. Let H and σ0, ε and G0 are given. Without loss of generality, we may assume that H ⊂ G0 which is in the unit ball of A and σ < ε/4. We may also assume that G0 = {g ⊗ f : g ∈ G0A and f ∈ G1T}, where G0A ⊂ A and G1T ⊂ C(T) are finite subsets. To simplify matter further, we may assume, without loss of generality, that G1T = {1C(T), z}, where z ∈ C(T) is the standard unitary generator. We may assume that G0A is sufficiently large and ε is sufficiently small such that [L1]P is well defined for any unital G0-ε-multiplicative contractive completely positive linear map from A ⊗ C(T) and for any unital G0A-ε-multiplicative contractive completely positive linear map L2 from A⊗ C(T) such that [L1]P0 = [L2]P0 (e 6.11) L1 ≈ε L2 on G0A. (e 6.12) Let n be an integer such that 1/n < σ/2. Note that A ⊗ C(T) ∈ Ds. Let δ > 0, G ⊂ A ⊗ C(T) and H1 ⊂ A ⊗ C(T)1 + \ {0} (in place of H2) be finite subsets required by 5.6 for A ⊗ C(T) (in place of A), ε/2 (in place of ε), G0 (in place of F), H (in place of H1) and ∆. Now suppose that L : A ⊗ C(T) → Mk satisfies the assumption for the above δ, G and H1. It follows from 5.6 that there is a projection e0 ∈ Mk and a G0-ε/2-multiplicative contractive completely positive linear mapψ0 : A⊗ C(T) → e0Mke0 and a unital homomorphism ψ1 : A ⊗ C(T) → (1 − e0)Mk(1 − e0) such that tr(e0) < 1/n < σ, kL(a) − ψ0(a) ⊕ ψ1(a)k < ε for all a ∈ G0. (e 6.13) (e 6.14) Define ψ : A⊗ C(T) → Mk by ψ(a) = ψ0(a)⊕ ψ1(a) for all a ∈ A and ψ(1⊗ z) = e0 ⊕ ψ1(1⊗ z). Put u = ψ(1 ⊗ z). One verifies that this ψ and u satisfy all requirements. 53 Lemma 6.3. Let A ∈ Ds be a unital C∗-algebra and let ∆ : (A ⊗ C(T))q,1 + \ {0} → (0, 1) be an order preserving map. Let ε > 0 and let F ⊂ A be a finite subset. There exists a finite subset H1 ⊂ A1 +\{0}, a finite subset G ⊂ A, δ > 0 and a finite subset P ⊂ K(A) such that, if L : A ⊗ C(T) → Mk (for some integer k ≥ 1) is G′-δ-multiplicative contractive completely positive linear map, where G′ = {g⊗ f : g ∈ G, f = {1, z, z∗}} and u ∈ Mk is a unitary such that + \{0}, a finite subset H2 ⊂ C(T)1 kL(1 ⊗ z) − uk < δ, [L]β(P) = 0 and tr ◦ L(h1 ⊗ h2) ≥ ∆( \h1 ⊗ h2) (e 6.15) (e 6.16) (e 6.17) for all h1 ∈ H1 and h2 ∈ H2, then there exists a continuous path of unitaries {ut : t ∈ [0, 1]} ⊂ Mk with u0 = u and u1 = 1 such that kL(f ⊗ 1)ut − utL(f ⊗ 1)k < ε for all f ∈ F and t ∈ [0, 1]. Moreover, length({ut}) ≤ π + ε. (e 6.18) (e 6.19) Proof. Let ∆1 = (1/2)∆, F0 = {f ⊗ 1 : 1⊗ z : f ∈ F} and let B = A⊗ C(T). Then B ∈ ¯Ds. Let H′ ⊂ B1 +\{0} (in place of H1), H0 ⊂ As.a. (in place of H2) be a finite subset, G1 ⊂ A⊗ C(T) (in place of G), δ1 > 0 (in place of δ), P′ ⊂ K(B) (in place of P) be the finite sets and constants as required by 5.9 (for B instead of A) for ε/16 (in place of ε), F0 (in place of F) and ∆. Without loss of generality, we may assume that H0 ⊂ B1 + \{0} and further, to simplify notation, we may assume that H0 = H′. We may assume that there are finite sets H′1 ⊂ A1 +\{0}, and G′1 ⊂ A such that +\{0}, H′2 ⊂ C(T)1 H′ = {h1 ⊗ h2 : h1 ∈ H′1 and h2 ∈ H′2} and G1 = {g⊗f : g ∈ G′1 and f ∈ {1, z, z∗}}. We may also assume that 1A ∈ H′1 and 1C(T) ∈ H′2. Without loss of generality, one may assume that P′ = P0 ⊔ P1, (e 6.20) where P0 ⊂ K(A) and P1 ⊂ β(K(A)) are finite subsets. Let P ⊂ K(A) be a finite subset such that β(P) = P1. Let σ = min{∆1(h) : h ∈ H′}. (e 6.21) There is δ2 > 0 (in place of δ) with δ2 < ε/16, a finite subset G2 ⊂ A ⊗ C(T)(in place of G) + \{0} (in place of H1) required by 6.2 for σ, ∆, H′(in place and a finite subset H3 ⊂ (A⊗ C(T))1 of H), min{ε/16, δ1/2} (in place of ε), G1 (in place of G0), P0 and P (in place of P1). We may also assume that for a finite set G′2 ⊂ A. We may further assume that G2 = {g ⊗ f : g ∈ G′2 and f ∈ {1, z, z∗}} H3 = {h1 ⊗ h2 : h1 ∈ H4 and h2 ∈ H5} for finite sets H4 ⊂ A1 H1 = H′1 ∪ H4 and H2 = H′2 ∪ H5. +\{0} and H5 ⊂ C(T)1 +\{0}. Let G = F∪G′1∪G′2, δ = min{δ1/2, δ2/2, ε/16}, 54 Now suppose that one has a linear map L : A⊗ C(T) → Mk and a unitary u ∈ Mk satisfy the assumption with the above H1, H2, G, P, δ and σ. It follows from 6.2 that there is a unital G1- min{ε/16, δ1/2}-multiplicative contractive completely positive linear map ψ : A ⊗ C(T) → Mk such that w = ψ(1 ⊗ z) is a unitary, wψ(g ⊗ 1) = ψ(g ⊗ 1)w for all g ∈ A, [ψ]P ′ = [L]P ′, tr ◦ L(g) − tr ◦ ψ(g) < σ for all g ∈ H3. (e 6.22) (e 6.23) (e 6.24) It follows that (e 6.25) for all h ∈ H′. Combining (e 6.22), (e 6.16), (e 6.17), (e 6.24) and (e 6.18), by applying 5.9, one obtains a unitary U ∈ Mk such that tr ◦ ψ(h) ≥ tr ◦ L(h) − σ ≥ ∆1(h) kAd U ◦ ψ(f ) − L(f )k < ε/16 for all f ∈ F0. Let w1 = Ad U ◦ ϕ(1 ⊗ z). Then ku − wk ≤ ku − L(1 ⊗ z)k + kL(1 ⊗ z) − Ad U ◦ ψ(1 ⊗ z)k < δ + ε/16 < ε/8. Then there is a continuous path of unitaries {ut ∈ [0, 1/2]} ⊂ Mk such that (e 6.26) (e 6.27) (e 6.28) (e 6.29) kut − uk < ε/8, kut − wk < ε/8, u0 = u, u1/2 = w and length({ut : t ∈ [0, 1/2]}) < επ/8. (e 6.30) It follows from 6.1 that there exists a continuous path of unitaries {ut : t ∈ [1/2, 1]} ⊂ Mk such that u1/2 = w, u1 = 1 and ut(Ad U ◦ ϕ(f ⊗ 1)) = (Ad U ◦ ϕ(f ⊗ 1))ut for all t ∈ [1/2, 1] and f ∈ A ⊗ 1. Moreover, It follows that Furthermore, length({ut : t ∈ [1/2, 1]}) ≤ π. length({ut : t ∈ [0, 1]}) ≤ π + επ/6. kutL(f ⊗ 1) − L(f ⊗ 1)utk < ε for all f ∈ F (e 6.31) (e 6.32) (e 6.33) (e 6.34) and t ∈ [0, 1]. Lemma 6.4. (Lemma 2.8 of [66] ) Let A be a unital separable amenable C∗-algebra. Let ε > 0, let F0 ⊂ A be a finite subset and let F ⊂ A⊗ C(T) be a finite subset. There exists a finite subset G ⊂ A and δ > 0 satisfying the following: For any G-δ-multiplicative contractive completely positive linear map ϕ : A → B (for some unital C∗-algebra B) and any unitary u ∈ B such that (e 6.35) there exists a unital F-ε-multiplicative contractive completely positive linear map L : A⊗C(T) → B such that kϕ(g)u − uϕ(g)k < δ for all g ∈ G, kϕ(f ) − L(f ⊗ 1)k < ε and kL(1 ⊗ z) − uk < ε (e 6.36) for all f ∈ F0, where z ∈ C(T) is the identity function on the unit circle. 55 + \ {0} and let H2 ⊂ C(T)1 Lemma 6.5. Let A ∈ ¯Ds be a unital C∗-algebra, Let ε > 0 and let F ⊂ A be a finite subset. Let H1 ⊂ A1 + \ {0} be finite subsets. For any order preserving map ∆ : Aq,1 + \{0} → (0, 1), there exists a finite subset G ⊂ A, a finite subset H′1 ⊂ A+\{0} and δ > 0 satisfy the following: for any unital G-δ-multiplicative contractive completely positive linear map ϕ : A → Mk (for some integer k ≥ 1) and any unitary u ∈ Mk such that kuϕ(g) − ϕ(g)uk < δ for all g ∈ G and tr ◦ ϕ(h) ≥ ∆(h) for all h ∈ H′1, there exists a continuous path of unitaries {ut : t ∈ [0, 1]} ⊂ Mk such that u0 = u, u1 = w, kutϕ(f ) − ϕ(f )utk < ε for all f ∈ G and t ∈ [0, 1], (e 6.37) (e 6.38) (e 6.39) (e 6.40) tr ◦ L(h1 ⊗ h2) ≥ ∆(ch1)τm(h2)/4 for all h1 ∈ H1 and h2 ∈ H2, where L : A ⊗ C(T) → Mk is a contractive completely positive linear map such that kL(f ⊗ 1) − ϕ(f )k < ε for all f ∈ F, and kL(1 ⊗ z) − wk < ε, (e 6.41) and τm is the tracial state on C(T) induced by the Lebesgue measure on the circle. Moreover, Proof. There exists an integer n ≥ 1 such that length({ut}) ≤ π + ε. (1/n) nXj=1 f (eθ+j2πi/n) ≥ (63/64)τm(f ) (e 6.42) (e 6.43) for all f ∈ H2 and for any θ ∈ [−π, π]. We may also assume that 16π/n < ε. Let σ1 = (1/210) inf{t(h) : h ∈ H1} · inf{τm(g) : g ∈ H2}. Let F′ = {f ⊗ 1, f ⊗ z : f ∈ F ∪H1}. Let δ1 > 0 (in place of δ) and G1 ⊂ A ⊗ C(T) (in place of G) be a finite subset as required by 5.4 for ε/32 (in place of ε), F′ (in place of F) and σ1/16 (in place of σ0). Without loss of generality, one may assume that, for a finite set G2 ⊂ A, G1 = {g ⊗ 1, 1 ⊗ z : g ∈ G2}. Let H′1 ⊂ A+ \{0} (in place of H2) be a finite subset as required by 5.7 for min{ε/32, σ1/16} (in place of ε), F ∪H1 (in place of F), H1 (in place of H), (190/258)∆ (in place of ∆) and σ1/16 (in place of σ) and integer n. Put H′ = {h1 ⊗ h2, h1 ⊗ 1, 1 ⊗ h2 : h1 ∈ H1 and h2 ∈ H2}. Let G3 = G2 ∪H1 ∪H′1. To simplify the notation, without loss of generality, one may assume that G3 and F′ are all in the unit ball of A ⊗ C(T). Let δ2 = min{ε/64, δ1/2, σ1/16}. Let G4 ⊂ A be a finite subset (in place of G) and let δ3 (in place of δ) be positive as required by 6.4 for G3 (in place of F0), F′ (in place of F), and δ2 (in place of ε). Let G = G4 ∪ G3 and δ = min{δ1/4, δ2/2, δ3/2}. Now let ϕ : A → Mk be a unital G-δ- multiplicative contractive completely positive linear map , and let u ∈ Mk be a unitary such that (e 6.37) and (e 6.38) hold for the above δ, σ, G and H′1. 56 It follows from 6.4 that there exists a G3-δ2-multiplicative contractive completely positive linear map L1 : A ⊗ C(T) → Mk such that kL1(g ⊗ 1) − ϕ(g)k < δ2 for all g ∈ G2 and kL1(1 ⊗ z) − uk < δ2. (e 6.44) We then have that tr ◦ L1(h ⊗ 1) ≥ tr ◦ ϕ(h) − δ2 ≥ ∆(h) − σ1/16 ≥ (191/256)∆(h) (e 6.45) (e 6.46) for all h ∈ H1. It follows from 5.4 that there exist a projection p ∈ Mk and a unital homomor- phism ψ : A ⊗ C(T) → pMkp such that kpL1(f ) − L1(f )pk < min{ε/32, σ1/16} for all f ∈ F′, (e 6.47) kL1(f ) − ((1 − p)L1(f )(1 − p) + ψ(f ))k < min{ε/32, σ1/16} for all f ∈ F′ (e 6.48) and tr(1 − p) < σ1/16. (e 6.49) Note that pMkp ∼= Mm for some m ≤ k. It follows from (e 6.46), (e 6.38), (e 6.48) and (e 6.49) that tr ◦ ψ(h) ≥ (191/256)∆(h) − σ1/16 − σ1/16 ≥ (190/256)∆(h) for all h ∈ H1. (e 6.50) By 5.7, there are mutually orthogonal projections e0, e1, e2, ..., en ∈ pMkp such that e1, e2, ..., en are equivalent and there are unital homomorphisms ψ0 : A ⊗ C(T) → e0Mke0 and ψ1 : A ⊗ C(T) → e1Mke1 such that kψ(f ) − diag(ψ0(f ), z Let w′0 = ψ1(1 ⊗ z). One may write n } { ψ1(f ), ..., ψ1(f ))k < min{ε/32, σ1/6} for all f ∈ F1 and tr(e0) < σ1/16. (e 6.51) (e 6.52) w′0 = diag(exp(ia1), exp(ia2), ..., exp(ian)), where aj ∈ ejMkej is a selfadjoint element with kajk ≤ π. Choose a continuous path of unitaries {w′t,j : t ∈ [0, 1]} ⊂ ejMkej such that w′0,j = exp(iaj), w′1,j = exp(i(2πj/n)), and length({w′t,j}) ≤ π + ε/4. (e 6.53) Moreover, one can choose such w′t,j in the commutant of ψ1(A). There is a unitary w′′0 ∈ (1 − p)Mk(1 − p) such that kw′′0 − (1 − p)L1(1 ⊗ z)(1 − p)k < ε/16. u′0 = w′′0 ⊕ ψ0(1 ⊗ z) ⊕ w′0. Put Then u′0 is a unitary and ku − u′0k ≤ ku − L1(1 ⊗ z)k + kL1(1 ⊗ z) − u′0k ≤ δ2 + ε/16 < ε/8. 57 (e 6.54) (e 6.55) (e 6.56) (e 6.57) One obtains a continuous path of unitaries {wt ∈ [0, 1]} ⊂ Mk such that w0 = u, w1 = w′′0 ⊕ ψ0(1 ⊗ z) ⊕ diag(w′1,1, w′1,2, ..., w′1,n) kwtϕ(f ) − ϕ(f )wtk < ε for all f ∈ F, and length({wt}) ≤ π + ε. Define L : A ⊗ C(T) → Mk by L(a ⊗ f ) = (1 − p)L1(a ⊗ f )(1 − p) ⊕ (diag(ψ0(a), for all a ∈ A and f ∈ C(T). It follows that z ψ1(a), ..., ψ1(a))f (w1)) n } { (e 6.58) (e 6.59) (e 6.60) kL(f ⊗ 1) − ϕ(f )k < ε for all f ∈ F and kL(1 ⊗ z) − w1k < ε. (e 6.61) One also has that (note that w′1,j is scalar) tr ◦ L(h1 ⊗ h2) ≥ tr(ψ0(h1) + tr(diag( ψ1(a), ..., ψ1(a))f (w1)) z n } { nXj=1 ≥ tr ◦ ψ(h1) nXj=1 ≥ (190/256)∆(ch1)(63/64)τm(h2) ≥ ∆(ch1) · τm(h2)/4 h2(ei2πj/n) ≥ (190/256)∆(ch1) h2(ei2πj/n) (e 6.62) (e 6.63) (e 6.64) for all h1 ∈ H1 and h2 ∈ H2. Definition 6.6. Let A be a unital C∗-algebra with T (A) 6= ∅ and let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. Suppose that τm : C(T) → C is the tracial state given by the normalized Lebesgue measure. Define ∆1 : (A ⊗ C(T))q,1 ∆1(h) = sup{ Lemma 6.7. Let A ∈ Ds be a unital C∗-algebra. Let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. For any ε > 0 and any finite subset F ⊂ A, there exists a finite subset H ⊂ A1 +\{0}, δ > 0, a finite subset G ⊂ A and a finite subset P ⊂ K(A) satisfying the following: For any unital G-δ-multiplicative contractive completely positive linear map ϕ : A → Mk (for some integer k ≥ 1) and any unitary v ∈ Mk such that : h ≥ \h1 ⊗ h2 and h1 ∈ A+ \ {0}, h2 ∈ C(T)+ \ {0}}. + \ {0} → (0, 1) by ∆(h1)τm(h2) (e 6.65) 4 tr ◦ ϕ(h) ≥ ∆(h) for all h ∈ H, kϕ(g)v − vϕ(g)k < δ for all g ∈ G and Bott(ϕ, v)P = {0}, then there exists a continuous path of unitaries {ut : t ∈ [0, 1]} ⊂ Mk such that u0 = v, u1 = 1, and kϕ(f )ut − utϕ(f )k < ε for all t ∈ [0, 1] and f ∈ F. Moreover, length({ut}) ≤ 2π + ε. (e 6.66) (e 6.67) (e 6.68) (e 6.69) (e 6.70) Proof. Let ∆1 be as in 6.6. Let H1 ⊂ A1 + \ {0} be finite subsets, G1 ⊂ A (in place of G) be a finite subset, δ1 > 0 (in place of δ) and P ⊂ K(A) be a finite subset required by 6.3 for ε/4 (in place of ε), F and ∆1. + \ {0} and H2 ⊂ C(T)1 58 Let G2 ⊂ A (in place of G) be a finite subset, H′1 ⊂ A+ \ {0} be a finite subset, δ2 > 0 (in place of δ) be as required by 6.5 for min{ε/16, δ1/2} (in place of ε), G1 ∪ F (in place of F) and H1 and H2. Let G = G2 ∪ G1 ⊂ F and let δ = min{δ2, ε/16}. Let H = H1. Now suppose that ϕ : A → Mk is a unital G-δ-multiplicative contractive completely positive linear map and u ∈ Mk is a unitary which satisfy the assumption for the above H, δ, G and P. By applying 6.5, one obtains a continuous path of unitaries {ut : t ∈ [0, 1/2]} ⊂ Mk such that u0 = u, u1 = w, kutϕ(g) − ϕ(g)utk < min{δ1, ε/4} (e 6.71) for all g ∈ G1 ∪ F and t ∈ [0, 1/2]. Moreover, there is a unital contractive completely positive linear map L : A ⊗ C(T) → Mk such that kL(g ⊗ 1) − ϕ(g)k < min{δ1, ε/4} for all g ∈ G1 ∪ F, kL(1 ⊗ z) − wk < min{δ1, ε/4} and tr ◦ L(h1 ⊗ h2) ≥ ∆(h1)τm(h2)/4 for all h1 ∈ H1 and h2 ∈ H2. Furthermore, length({ut : t ∈ [0, 1/2]}) ≤ π + ε/4. Note that [L]β(P) = Bott(ϕ, w)P = Bott(ϕ, u)P = 0. (e 6.72) (e 6.73) (e 6.74) (e 6.75) (e 6.76) By (e 6.72), (e 6.73), (e 6.76) and (e 6.74), applying 6.3, one obtains a continuous path of unitaries {ut ∈ [1/2, 1]} ⊂ Mk such that u1/2 = w, u1 = 1, kutϕ(f ) − ϕ(f )utk < ε/4 for all f ∈ F and length({ut : t ∈ [1/2, 1]}) ≤ π + ε/4. (e 6.77) (e 6.78) Therefore {ut : t ∈ [0, 1]} ⊂ Mk is a continuous path of unitaries in Mk with u0 = u and u1 = 1 such that kutϕ(f ) − ϕ(f )utk < ε for all f ∈ F and length({ut : t ∈ [0, 1]}) ≤ 2π + ε. (e 6.79) (e 6.80) 7 An Existence Theorem for Bott maps Lemma 7.1. Let A be a unital amenable residually finite dimensional C-algebra which satisfies the UCT, let G = Zr ⊕ Tor(G) ⊂ K0(A) be a finitely generated subgroup with [1A] ∈ G and let J0, J1 ≥ 0 be integers. For any δ > 0, any finite subset G ⊂ A and any finite subset P ⊂ K(A) with P ∩ K0(A) ⊂ G, there exist integers N0, N1, ..., Nk and unital homomorphisms hj : A → MNj , j = 1, 2, ..., k, satisfying the following: For any κ ∈ HomΛ(K(A), K(K)), with κ([1A]) = J1 and J0 = max{κ(gi) : gi = ( i−1 z } { 0, ..., 0, 1, 0, ..., 0) ∈ Zr : 1 ≤ i ≤ r}, 59 (e 7.1) there exists a G-δ-multiplicative contractive completely positive linear map Φ : A → MN0+κ([1A]) such that [Φ]P = (κ + [h1] + [h2] + ··· + [hk])P . (e 7.2) (Note that N0 =Pk i=1 Ni). Proof. It follows from 6.1.11 of [55] (see also [52] and [18]) that, for each such κ, there is a unital G-δ-multiplicative contractive completely positive linear map Lκ : A → Mn(κ) (for integer n(κ) ≥ 1) such that [Lκ]P = (κ + [hκ])P , (e 7.3) where hκ : A → MNκ is a unital homomorphism. There are only finitely many different κP so that (e 7.1) holds, say κ1, κ2, ..., κk. Set hi = hκi, i = 1, 2, ..., k. Let Ni = Nκi, i = 1, 2, ...k. Note that Ni = J1 + n(κi), if κ([1A]) = J1, and Ni = −J1 + n(κi), if κ([1A]) = −J1. Define If κ = κi, define Φ : A → MN0+κ([1A]) by Ni. N0 = kXi=1 Φ = Lκi +Xj6=i hj. The lemma follows. Lemma 7.2. Let A be a unital C∗-algebra as in 7.1 and let [1A] ∈ G = Zr ⊕ Tor(G) ⊂ K0(A) be a finitely generated subgroup. There exists Λi ≥ 0, i = 1, 2, ..., r, satisfying the following: For any δ > 0, any finite subset G ⊂ A and any finite subset P ⊂ K(A) with P ∩ K0(A) ⊂ G, there exist integers N (δ,G,P, i) ≥ 1, i = 1, 2, ..., r, satisfying the following: i−1 z } { 0, ..., 0, 1, 0, ..., 0) ∈ Zr, there exists a unital G-δ-multiplicative contractive completely positive linear map L : A → MN1 and a homomorphism h : A → MN1 such that Let κ ∈ HomΛ(K(A), K(K)) and Si = κ(gi), where gi = ( [L]P = (κ + [h])P , (e 7.4) i=1(N (δ,G,P, i) ± Λi) · Si. : G → Z be a homomorphism defined by ψ+ where N1 =Pr Proof. Let ψ+ i (gj ) = 0, if j 6= i, and i ψ+ i Tor(G) = 0, and let ψ−i (gi) = −1 and ψ−i (gj) = 0, if j 6= i, and ψ−i Tor(G) = 0, i = 1, 2, ..., r. Note that ψ−i = −ψ+ i and κ−i = ψ−i , i = 1, 2, ..., r. Let κ+ Let N0(i) ≥ 1 (in place of N0) be required by 7.1 for δ, G, J0 = 1 and J1 = Mi. Define N (δ,G,P, i) = N0(i), i = 1, 2, ..., r. i , where Si = κ(gi), i = 1, 2, ..., r. By applying 7.1, one obtains G-δ-multiplicative contractive completely positive linear maps L±i i , κ−i ∈ HomΛ(K(A), K(K)) be such that κ+ Let κ ∈ HomΛ(K(A), K(K)). Then κG = Pr i ([1A]) and a homomorphism h±i i , i = 1, 2, ..., r. Let Λi = ψ+ i ([1A]), i = 1, 2, ..., r. i (gi) = 1, ψ+ i G = ψ+ : A → MN0(i) such that : A → MN0(i)+κ± i=1 Siψ+ [L±i ]P = (κ±i + [h±i ])P , i = 1, 2, ..., r. (e 7.5) 60 i=1 L±,Si i Define L =Pr , where L±,Si : A → MSiN0(i) is defined by { for all a ∈ A. One checks that L : A → MN1, where N1 = Pr i ([1A]) if Si > 0, or Λ′i = −ψ+ Λ′i = ψ+ completely positive linear map and L±i (a), ..., L±i (a)) L±,Si(a) = diag( } z Si i=1 Si(Λ′i + N (δ,G,P, i)) with +([1A]) if Si < 0, is a unital G-δ-multiplicative contractive for some homomorphism h : A → MN1. [L]P = (κ + [h])P i h±,0 i i=1 h±,Si Remark 7.3. Note that in the above proof, h may be written asPr . Let us agree that is zero. Let S ≥ 1 be a fixe integer. Let us assume that max{Si} ≤ S as a restriction on κ. Choose new N1 to bePr i=1(N (δ,G,P, i)±Λi)·Sr. Let h0 =P0≤si≤SPr : A → MN1 Then h0 is independent of κ. Choose Φ = L⊕Psi6Si h±,si . Then Φ : A → MN1 and [Φ]P = (κ+[h0])P . In other words, h0 can be chosen before κ is given as long as κ has the restriction as mentioned above (as in 7.1). Lemma 7.4. Let A ∈ Ds be a unital C∗-algebra and let P ⊂ K(A) be a finite subset. Suppose that G ⊂ K(A) is the group generated by P, and G1 = G ∩ K1(A) = Z r ⊕ Tor(K1(A)). Let F ⊂ A, let ε > 0, and let ∆ : Aq,1 + \{0}, and an integer N ≥ 1 satisfying the following: Let κ ∈ KK(A ⊗ C(T), C) and put There exist δ > 0, a finite subset G ⊂ A, a finite subset H ⊂ A1 + \ {0} → (0, 1) be an order preserving map. i=1 h±,si i i K = max{κ(β(gi)) : 1 ≤ i ≤ r}, (e 7.6) i−1 z } { 0, ..., 0, 1, 0, ..., 0) ∈ Zr. Then for any unital G-δ-multiplicative contractive completely where gi = ( positive linear map ϕ : A → MR such that R ≥ N (K + 1) and tr ◦ ϕ(h) ≥ ∆(h) for all h ∈ H, (e 7.7) (e 7.8) (e 7.9) there exists a unitary u ∈ MR such that k[ϕ(f ), u]k < ε for all f ∈ F and Bott(ϕ, u)P = κ ◦ βP . Proof. To simplify notation, without loss of generality, we may assume that F is a subset of the unit ball. Let ∆1 = (1/8)∆ and ∆2 = (1/16)∆. Let ε0 > 0 and let G0 ⊂ A be a finite subset satisfy the following: If ϕ′ : A → B (for any unital C∗-algebra B) is a unital G0-ε0-multiplicative contractive completely positive linear map and u′ ∈ B is a unitary such that kϕ′(g)u′ − u′ϕ′(g)k < 4ε0 for all g ∈ G0, (e 7.10) then Bott(ϕ′, u′)P is well defined. Moreover, if ϕ′ : A → B is another unital G0-ε0-multiplicative contractive completely positive linear map , then Bott(ϕ′, u′)P = Bott(ϕ′′, u′′)P , (e 7.11) 61 provided that kϕ′(g) − ϕ′′(g)k < 4ε0 and ku′ − u′′k < 4ε0 for all g ∈ G0. (e 7.12) We may assume that 1A ∈ G0. Let G′0 = {g ⊗ f : g ∈ G0} and f = {1C(T), z, z∗}, where z is the identity function on the unit circle T. We also assume that if Ψ′ : A ⊗ C(T) → C for a unital C∗-algebra C is a G′0-ε0-multiplicative contractive completely positive linear map, then there exists a unitary u′ ∈ C such that kΨ′(1 ⊗ z) − u′k < 4ε0. (e 7.13) Without loss of generality, we may assume that G0 is in the unital ball of A. Let ε1 = min{ε/64, ε0/512} and F1 = F ∪ G0. Let H0 ⊂ A1 + \ {0} (in place of H) be a finite subset and L ≥ 1 be an integer as required by 4.19 for ε1 (in place of ε) and F1 (in place of F) as well as ∆2 (in place of ∆). Let H1 ⊂ A1 + \ {0}, G1 ⊂ A (in place of G), δ1 > 0 (in place of δ), P1 ⊂ K(A) (in place of P), H2 ⊂ As.a., and 1 > σ > 0 be required by 5.9 for ε1 (in place of ε), F1 (in place of F) and ∆1. We may assume that [1A] ∈ P2, H2 is in the unit ball of A and H0 ⊂ H1. Without loss of generality, we may assume that δ1 < ε1/16, σ < ε1/16, and F1 ⊂ G1. Put P2 = P ∪ P1 and H′1 = H1 ∪ H0. Suppose that A has irreducible representations of rank r1, r2, ..., rk. Fix an irreducible rep- resentation π0 : A → Mr1. Let N (p) ≥ 1 (in place of N (P0)) and H′0 ⊂ A1 + \ {0} (in place of H) be a finite subset required by 4.18 for {1A} (in place of P0) and (1/16)∆. Let H′1 = H1 ∪ H′0. Let G0 = G ∩ K0(A) and write G0 = Zs1 ⊕ Zs2 ⊕ Tor(G0), where Zs2 ⊕ Tor(G0) ⊂ kerρA. j−1 z } { 0, ..., 0, 1, 0, ..., 0) ∈ Zs1 ⊕ Zs2, j = 1, 2, ..., s2. Note that A ⊗ C(T) ∈ Ds and A ⊗ C(T) Let xj = ( has irreducible representations of rank r1, r2, ..., rk. Let ¯r = max{(π0)∗0(xj) : 0 ≤ j ≤ s1 + s2}. Let P3 ⊂ K(A ⊗ C(T)) be a finite subset set containing P2, {β(gj) : 1 ≤ j ≤ r} and a finite subset which generates β(Tor(G1)). Choose δ2 > 0 and finite subset G = {g ⊗ f : g ∈ G2, f ∈ {1, z, z∗}} in A ⊗ C(T), where G2 ⊂ A is a finite subset such that, for any unital G-δ2-multiplicative contractive completely positive linear map Φ′ : A ⊗ C(T) → C (for any unital C∗-algebra C with T (C) 6= ∅), [Φ′]P3 is well defined and [Φ′]Tor(G0)⊕β(Tor(G1) = 0. (e 7.14) We may assume G2 ⊃ G1 ∪ F1. K(A) ⊕ β(K(A)). Consider the subgroup of K0(A ⊗ C(T)) given by Let σ1 = min{∆2(h) : h ∈ H′1}. Note K0(A⊗C(T)) = K0(A)⊕β(K1(A)) and K(A⊗C(T)) = Zs1 ⊕ Zs2 ⊕ Zr ⊕ Tor(K0(A) ⊕ β(Tor(K1(A)). Let δ3 = min{δ1, δ2}. Let N (δ3,G,P3, i) and Λi, i = 1, 2, ..., s1 + s2 + r, be required by 7.2 n1 − 1 62 for A ⊗ C(T). Choose an integer n1 ≥ N (p) such that N (δ3,G,P3, i) + 1 + Λi)N (p) (Ps1+s2+r i=1 < min{σ/16, σ1/2}. (e 7.15) Choose n > n1 such that n1 + 2 n < min{σ/16, σ1/2, 1/(L + 1)}. (e 7.16) Let ε2 > 0 and let F2 ⊂ A be a finite subset such that [Ψ]P2 is well defined. Let ε3 = min{ε2/2, ε1} and F3 = F1 ∪ F2. Denote by δ4 > 0 (in place of δ), G3 ⊂ A (in place of G), H3 ⊂ A1 + \ {0} (in place of H2) the finite sets as required by 5.6 for ε3 (in place of ε), F3 ∪ H′1 (in place of F), δ3/2 (in place of ε0), G2 (in place of G0), ∆, H′1 (in place of H), min{σ/16, σ1/2} (in place of σ) and n2 (in place of K) required by 5.6 (with L1 = L2). Set G = F3 ∪ G1 ∪ G2 ∪ G3, set δ = min{ε3/16, δ4, δ3/16}, and set G5 = {g ⊗ f : g ∈ G4, f ∈ N (δ3,G0,P3, i) + Λi + 1) and define {1, z, z∗}}. Let H = H′1 ∪ H3. Define N0 = (n + 1)N (p)(Ps1+s2+r N = N0 + N0¯r. Fix any κ ∈ KK(A ⊗ C(T), C) with i=1 K = max{κ(β(gi) : 1 ≤ j ≤ r}. Let R > N (K + 1). Suppose that ϕ : A → MR is a unital G-δ-multiplicative contractive completely positive linear map such that tr ◦ ϕ(h) ≥ ∆(h) for all h ∈ H. (e 7.17) e1, e2, ..., en2 are equivalent, tr(e0) < min{σ/64, σ1/4} and e0 +Pn2 Then, by 5.6, there exist mutually orthogonal projections e0, e1, e2, ..., en2 ∈ MR such that i=1 ei = 1MR, and there exists a unital G2-δ3/2-multiplicative contractive completely positive linear map ψ0 : A → e0MRe0 and a unital homomorphism ψ : A → e1MRe1 such that kϕ(f ) − (ψ0(f ) ⊕ n2 z } ψ(f ), ψ(f ), ..., ψ(f ))k < ε3 for all f ∈ F3 and tr ◦ ψ(h) ≥ ∆(h)/3n2 for all h ∈ H1. { (e 7.18) (e 7.19) Let α ∈ HomΛ(K(A ⊗ C(T)), K(Mr)) be defined as follows: αK(A) = [π0] and αβ(K(A)) = κβ(K(A)). Let max{κ ◦ β(gi) : i = 1, 2, ..., r,π0(xj) : 1 ≤ j ≤ s1 + s2} ≤ max{K, ¯r}. Applying 7.2, we obtain a unital G-δ3-multiplicative contractive completely positive linear N (δ3,G0,P3, j) + Λi) max{K, ¯r}, and a map Ψ : A ⊗ C(T) → MN ′ homomorphism H0 : A ⊗ C(T) → H0(1A)MN ′ , where N′1 ≤ N1 =Ps1+s2+r H0(1A) such that such that j=1 1 1 [Ψ]P3 = (α + [H0])P3 . In particular, since [1A] ∈ P2 ⊂ P3, rankΨ(1A) = r1 + rank(H0). Note that N′1 + N (p) R N1 + N (p) N (K + 1) ≤ < 1/(n + 1). (e 7.20) (e 7.21) Let R1 = rank e1. Then R1 ≥ R/(n + 1). So, from (e 7.21), one has that R1 ≥ N1 + N (p). In other words, R1 − N′1 ≥ N (p). Note that t ◦ ψ(g) ≥ (1/3)∆(g) ≥ ∆2(g) for all g ∈ H′0, 63 1 Define L1 : A → E1MRE1 by L1(a) = π0(a) ⊕ H0A(a) ⊕ h0(a ⊗ 1) ⊕ ( where t is the tracial state on MR1. By applying 4.18 to the case that π0 ⊕ H0 (in place ϕ), ψ, where ψ is an amplification of ψ with ψ repeated n times, and P0 = {[1A]}, we obtain a unital . Define ψ′0 : A ⊗ C(T) → e0MRe0 by ψ′0(a ⊗ f ) = homomorphism h0 : A ⊗ C(T) → MnR1−N ′ ψ0(a) · f (1) · e0 for all a ∈ A and f ∈ C(T), where 1 ∈ T. Define ψ′ : A ⊗ C(T) → e1MRe1 by ψ′(a ⊗ f ) = ψ(a) · f (1) · e0 for all a ∈ A and f ∈ C(T). Put E1 = diag(e1, e2, ..., enn1 ). n(n1−1) } } z z [L1]P1 = [L2]P1, tr ◦ L1(h) ≥ ∆1(h), tr ◦ L2(h) ≥ ∆1(h) for all h ∈ H1, and tr ◦ L1(g) − tr ◦ L2(g) < σ for all g ∈ H2. a ∈ A, and define L2 : A → E1MRE1 by L2(a) = Ψ(a ⊗ 1) ⊕ h0(a ⊗ 1) ⊕ ( a ∈ A. Note that ψ(f ), ..., ψ(f )) for ψ(f ), ..., ψ(f )) for n(n1−1) { { It follows from 5.9 that there exists a unitary w1 ∈ E1MRE1 such that kad w1 ◦ L2(a) − L1(a)k < ε1 for all a ∈ F1. Define E2 = (e1 + e2 + ··· + en2) and define Φ : A → E2MRE2 by Φ(f )(a) = diag( ψ(a), ψ(a), ..., ψ(a)) for all a ∈ A. z n2 } { Then (e 7.22) (e 7.23) (e 7.24) (e 7.25) (e 7.26) (e 7.29) (e 7.30) (e 7.31) (e 7.32) (e 7.33) tr ◦ Φ(h) ≥ ∆2(h) for all h ∈ H0. (e 7.27) n1+2 > L + 1. By applying 4.19, we obtain a unitary w2 ∈ E2MRE2 n By (e 7.16), one has that and a unital homomorphism H1 : A → (E2 − E1)MR(E2 − E1) such that kad w2 ◦ diag(L1(a), H1(a)) − Φ(a)k < ε1 for all a ∈ F1. (e 7.28) Put w = (e0 ⊕ w1 ⊕ (E2 − E1))(e0 ⊕ w2) ∈ MR. Define H′1 : A ⊗ C(T) → (E2 − E1)MR(E2 − E1) by H′1(a ⊗ f ) = H1(a) · f (1) · (E2 − E1) for all a ∈ A and f ∈ C(T). Define Ψ1 : A → MR by n1−1 Ψ1(f ) = ψ′0(f ) ⊕ Ψ(f ) ⊕ h0 ⊕ It follows from (e 7.25), (e 7.28) and (e 7.18) that z ψ′(f ), ..., ψ′(f )) ⊕ H′1(f ) for all f ∈ A ⊗ C(T). } { kϕ(a) − w∗Ψ1(a ⊗ 1)wk < ε1 + ε1 + ε3 for all a ∈ F. Now pick a unitary v ∈ MR such that Put u = w∗vw. Then, we estimate that kΨ1(1 ⊗ z) − vk < 4ε1. Moreover, by (e 7.25),(e 7.20) and (e 7.11), one has k[ϕ(a), u]k < min{ε, ε0) for all a ∈ F1. Bott(ϕ, u)P = κ ◦ βP . 64 8 A Uniqueness Theorem for C∗-algebras in Ds The main goal of this section is to prove Theorem 8.4. Definition 8.1. Let A be a unital C∗-algebra and let C ∈ C, where C = C(F1, F2, ϕ0, ϕ1) is as in 3.1. Suppose that L : A → C is a contractive completely positive linear map. Define Le = πe ◦ L. Then Le : A → F1 is a contractive completely positive linear map such that ϕ0 ◦ Le = π0 ◦ L and ϕ1 ◦ Le = π1 ◦ L. (e 8.1) Moreover, if δ > 0 and G ⊂ A and L is G-δ-multiplicative, then Le is also G-δ-multiplicative. Lemma 8.2. Let A be a unital C∗-algebra and let C ∈ C, where C = C(F1, F2, ϕ0, ϕ1) is as in 3.1. Let L1, L2 : A → C be two unital completely positive linear maps, let ε > 0 and let F ⊂ A be a subset. Suppose that there are unitaries w0 ∈ π0(C) ⊂ F2 and w1 ∈ π1(C) ⊂ F2 such that (e 8.2) kw∗0π0 ◦ L1(a)w0 − π0 ◦ L2(a)k < ε and kw∗1π1 ◦ L1(a)w1 − π1 ◦ L2(a)k < ε for all a ∈ F. (e 8.3) (e 8.4) (e 8.5) Then there exists a unitary u ∈ F1 such that kϕ0(u)∗π0 ◦ L1(a)ϕ0(u) − π0 ◦ L2(a)k < ε, and kϕ1(u)∗π1 ◦ L1(a)ϕ1(u) − π1 ◦ L2(a)k < ε for all a ∈ F. We may assume that there are k(0) and k(1) such that ϕ0Mni = 0 if i > k(0), and ϕ1Mni Proof. Write F1 = Mn1 ⊕ Mn2 ⊕ ··· ⊕ Mnk and F2 = Mr1 ⊕ Mr2 ⊕ ··· ⊕ Mrl. We may assume that, kerϕ0 ∩ kerϕ1 = {0}. (See 3.1.) is injective, i = 1, 2, ..., k(0) with k(0) ≤ k, ϕ0Mni is injective, i = k(1), k(1) + 1, ..., k with k(1) ≤ k, ϕ1Mni j=k(1) Mnj . Note that k(1) ≤ k(0) + 1, ϕ0F1,0 and ϕ1F1,1 are injective. Note ϕ0(F1,0) = ϕ0(F1) = π0(C) and ϕ1(F1,1) = ϕ1(F1) = π1(C). Let ψ0 = (ϕ0F1,0)−1 and ψ1 = (ϕ1F1,1)−1. For each fixed a ∈ A, since Li(a) ∈ C (i = 0, 1), there are elements = 0, if i < k(1). Write F1,0 = Lk(0) i=1 Mni and F1,1 = Lk ga,i = ga,i,1 ⊕ ga,i,2 ⊕ ··· ⊕ ga,i,k(0) ⊕ ··· ⊕ ga,i,k ∈ F1, such that ϕ0(ga,i) = π0◦Li(a) and ϕ1(ga,i) = π1◦Li(a), i = 1, 2, where ga,i,j ∈ Mnj , j = 1, 2, ..., k and i = 1, 2. Note that such ga,i is unique since kerϕ0∩ kerϕ1 = {0}. Since w0 ∈ π0(C) = ϕ0(F1), there is a unitary u0 = u0,1 ⊕ u0,2 ⊕ ··· ⊕ u0,k(0) ⊕ ··· ⊕ u0,k such that ϕ0(u0) = w0. Note that the first k(0) components of u0 is uniquely determined by w0 (since ϕ0 is injective on this part) and the components after k(0)’th component can be chosen arbitrarily (since ϕ0 = 0 on this part). Similarly there exists u1 = u1,1 ⊕ u1,2 ⊕ ··· ⊕ u1,k(1) ⊕ ··· ⊕ u1,k such that ϕ1(u1) = w1 Now by e 8.2 and e 8.3, we have kϕ0(u0)∗ϕ0(ga,1)ϕ0(u0) − ϕ0(ga,2)k < ε and kϕ1(u1)∗ϕ1(ga,1)ϕ1(u1) − ϕ1(ga,2))k < ε for all a ∈ F. (e 8.6) (e 8.7) 65 Since ϕ0 is injective on F i use k(1) ≤ k(0) + 1), we have 1 for i ≤ k(0) and ϕ1 is injective on F i 1 for i > k(0) (note that we k(u0,i)∗(ga,1,i)u0,i − (ga,2,i)k < ε k(u1,i)∗(ga,1,i)u1,i − (ga,2,i)k < ε for all i > k(0), for all i ≤ k(0), and for all a ∈ F. (e 8.8) (e 8.9) Let u = u0,1 ⊕ ··· ⊕ u0,k(0) ⊕ u1,k(0)+1 ⊕ ··· ⊕ u1,k ∈ F1—that is, for the first k(0)’s components of u, we use u0’s corresponding components, and for the last k − k(0) components of u, we use u1’s. From e 8.8 and e 8.8. we have ku∗ga,1u − ga,2k < ε for all a ∈ F. Apply ϕ0 and ϕ1 to the above inequality, we get e 8.4 and e 8.5 as desired. Lemma 8.3. Let A be a unital C∗-algebra and let C ∈ C, where C = C(F1, F2, ϕ0, ϕ1) is as in 3.1. Suppose L, Ψ : A → C are two G-δ-multiplicative contractive completely positive linear maps for some 1 > δ > 0 and a subset G ⊂ A. Suppose that g ∈ U (A), 1/2 > γ > 0 and there is v ∈ CU (C) such that khL(g∗)ihΨ(g)i − vk < γ. Then, there is ve ∈ CU (F1) such that khLe(g∗)ihΨe(g)i − vek < γ, π0(ve) = v and ϕ1(ve) = π1(v). (e 8.10) (e 8.11) Proof. The element ve = πe(v) satisfy the condition of the lemma. Theorem 8.4. Let A ∈ Ds be a unital C∗-algebra with finitely generated Ki(A) (i = 0, 1). Let F ⊂ A, let ε > 0 be a positive number and let ∆ : Aq,1 + \{0} → (0, 1) be an order preserving map. There exists a finite subset H1 ⊂ A1 + \ {0}, γ1 > 0, γ2 > 0, δ > 0, a finite subset G ⊂ A and a finite subset P ⊂ K(A), a finite subset H2 ⊂ A, a finite subset U ⊂ Jc(K1(A)) (see 2.15) for which [U ] ⊂ P, and N ∈ N satisfying the following: For any unital G-δ-multiplicative contractive completely positive linear maps ϕ, ψ : A → C for some C ∈ C such that [ϕ]P = [ψ]P , τ (ϕ(a)) ≥ ∆(a), τ (ψ(a)) ≥ ∆(a), for all τ ∈ T (C) a ∈ H1, τ ◦ ϕ(a) − τ ◦ ψ(a) < γ1 for all a ∈ H2, and there exists a unitary W ∈ C ⊗ MN such that dist(ϕ‡(u), ψ‡(u)) < γ2 for all u ∈ U , kW (ϕ(f ) ⊗ 1MN )W ∗ − (ψ(f ) ⊗ 1MN )k < ε, for all f ∈ F. (e 8.12) (e 8.13) (e 8.14) (e 8.15) (e 8.16) Proof. Since K∗(A) is finitely generated, there is n0 such that κ ∈ HomΛ(K(A), K(C)) is deter- mined by its restriction to K(A, Z/nZ), n = 0, ..., n0. Set N = n0!. Let ∆1 be defined in 6.6 for the given ∆. Let H′1 ⊂ A+ \ {0} (in place of H1) for ε/32 (in place of ε) and F required by 6.7. Let δ1 > 0 (in place of δ), G1 ⊂ A (in place of G) be a finite subset and let P0 ⊂ K(A) (in place of P) be a finite subset required by 6.7 for ε/32 (in place of ε), F and ∆1. We may assume that δ1 < ε/32 and (2δ1,G1) is a KK-pair (see the end of 2.12). 66 τ (bott1(u, v)) = (τ (log(u∗vuv∗)) 1 2π√−1 Moreover, we may assume that δ1 is sufficiently small that if kuv − vuk < 3δ1, then the Exel formula holds for any pair of unitaries u and v in any unital C∗-algebra C with tracial rank zero and any τ ∈ T (C) (see Theorem 3.6 of [65]). Moreover if kv1 − v2k < 3δ1, then bott1(u, v1) = bott1(u, v2). Let g1, g2, ..., gk(A) ∈ U (Mm(A)(A)) (m(A) ≥ 1 is an integer) be a finite subset such that { ¯g1, ¯g2, ..., ¯gk(A)} ⊂ Jc(K1(A)) and such that {[g1], [g2], ..., [gk(A)]} forms a set of generators for K1(A). Let U = { ¯g1, ¯g2, ..., ¯gk(A)} ⊂ Jc(K1(A)) be a finite subset. Let U0 ⊂ A be a finite subset such that {g1, g2, ..., gk(A)} ⊆ {(ai,j) : ai,j ∈ U0}. Let δu = min{1/256m(A)2, δ1/16m(A)2}, Gu = F ∪ G1 ∪ U0 and let Pu = P0. Let δ2 > 0 (in place of δ), G2 ⊂ A (in place of G), H′2 ⊂ A+ \ {0} (in place of H), N1 ≥ 1 (in place of N ) be the finite subsets and the constants as required by 7.4 for the data δu (in place of ε), Gu (in place of F), Pu (in place of P) and ∆ and with ¯gj (in place of gj), j = 1, 2, ..., k(A) (with k(A) = r). Let d = min{∆(h) : h ∈ H′2}. Let δ3 > 0 and let G3 ⊂ A ⊗ C(T) be finite subset satisfying the following: For any G3-δ3-multiplicative contractive completely positive linear map L′ : A ⊗ C(T) → C′ (for any unital C∗-algebra C′ with T (C′) 6= ∅), τ ([L](β(¯gj)) < d/8, j = 1, 2, ..., k(A). (e 8.17) Without loss of generality, we may assume that G3 = {g ⊗ z : g ∈ G′3 and z ∈ {1, z, z∗}}, Let H′3 ⊆ A1 where G′3 ⊂ A is a finite subset (by choosing a smaller δ3 and large G′3). Let ε′′1 = min{d/27m(A)2, δu/2, δ2/2m(A)2, δ3/2m(A)2} and let ¯ε1 > 0 (in place of δ) and G4 ⊂ A (in place of G) be a finite subset as required by 6.4 for ε′′1 (in place of ε) and Gu ∪ G′3. Put ε1 = min{ε′1, ε′′1, ¯ε1}. Let G5 = Gu ∪ G′3 ∪ G4. + \{0} (in place of H1), δ4 > 0 (in place of δ), G6 ⊂ A (in place of G), H′4 ⊂ As.a. (in place of H2), P1 ⊂ K(A) (in place of P) and σ4 > 0 (in place of σ2) be the finite subsets and constants as required by Theorem 5.9 with respect to ε1/4 (in place ε) and G5 (in place of F) and ∆. + \ {0} and δ5 > 0 and a finite subset G7 ⊂ A such that, for any Mm and unital G7-δ5-multiplicative contractive completely positive linear map L′ : A → Mm, if tr ◦ L′(h) > 0 for all h ∈ H′5, then m ≥ N2((8/d) + 1). Choose N2 ≥ N1 such that (k(A) + 1)/N2 < d/8. Choose H′5 ⊂ A1 Put δ = min{ε1/16, δ4/4m(A)2, δ5/4m(A)2}, G = G5 ∪ G6 ∪ G7, and P = Pu ∪ P1. Put H1 = H′1 ∪ H′2 ∪ H′3 ∪ H′4 ∪ H′5 and let H2 = H′4. Let γ1 = σ4 and let 0 < γ2 < min{d/16m(A)2, δu/9m(A)2, 1/256m(A)2}. Now suppose that C ∈ C and ϕ, ψ : A → C are two unital G-δ-multiplicative contractive completely positive linear maps satisfying the condition of the theorem for the given ∆, H1, δ, G, P, H2, γ1, γ2 and U . Let 0 = t0 < t1 < ··· < tn = 1 67 be a partition of [0, 1] so that kπt ◦ ϕ(g) − πt′ ◦ ϕ(g)k < ε1/16 and kπt ◦ ψ(g) − πt′ ◦ ψ(g)k < ε1/16 (e 8.18) for all g ∈ G, provided t, t′ ∈ [ti−1, ti], i = 1, 2, ..., n. ··· ⊕ MnF (2). By the choice of H′5, one has that We write C = A(F1, F2, h0, h1), F1 = Mm1 ⊕ Mm2 ⊕ ··· ⊕ MmF (1) and F2 = Mn1 ⊕ Mn2 ⊕ nj ≥ N2(8/d + 1) and ms ≥ N2(8/d + 1), 1 ≤ j ≤ F (2), 1 ≤ s ≤ F (1). (e 8.19) Applying Theorem 5.9, one obtains a unitary wi ∈ F2, if 0 < i < n, w0 ∈ h0(F1), if i = 0, and w1 ∈ h1(F1), if i = 1, such that kwiπti ◦ ϕ(g)w∗i − πti ◦ ψ(g)k < ε1/16 for all g ∈ G5. (e 8.20) and h1(we) = wn. It follows from 8.2 that we may assume that there is a unitary we ∈ F1 such that h0(we) = w0 By (e 8.15), there is a unitary ωj ∈ Mm(A)(C) such that ωj ∈ CU (Mm(A)(C)) and kh(ϕ ⊗ idMm(A)(g∗j )ih(ψ ⊗ idMm(A))(gj )i − ωjk < γ2, j = 1, 2, ..., k(A). (e 8.21) (note that we now have wi as well as ωi in the proof.) Write ωj = e(j)Yl=1 exp(√−1a(l) j ) for some selfadjoint element a(l) j ∈ Mm(A)(C), l = 1, 2, ..., e(j), j = 1, 2, ..., k(A). Write a(l) j = (a(l,1) j , a(l,2) j , ..., a (l,nF (2)) j ) and ωj = (ωj,1, ωj,2, ..., ωj,F (2)) in C([0, 1], F2) = C([0, 1], Mn1 ) ⊕ ··· ⊕ C([0, 1], MnF (2) ), where ωj,s = exp(√−1a(l,s) j 1, 2, ..., F (2). Then ), s = ns(ts ⊗ Trm(A))(a(l,s) j 2π (t)) ∈ Z, t ∈ [0, 1], e(j)Xl=1 where ts is the normalized trace on Mns, s = 1, 2, ..., F (2). In particular, e(j)Xl=1 ns(t ⊗ Trm(A))(a(l,s) j (t)) = e(j)Xl=1 ns(t ⊗ Trm(A))(a(l,s) j (t′)) for all t, t′′ ∈ [0, 1]. (e 8.22) Let Wi = wi ⊗ idMm(A), i = 0, 1, ...., n and We = we ⊗ idMm(A)(F1). Then kπi(hϕ ⊗ idMm(A))(g∗j )i)Wi(πi(hϕ ⊗ idMm(A))(gj )i)W ∗i − ωj(ti)k < 3m(A)2ε1 + 2γ2 < 1/32. We also have (with ϕe = πe ◦ ϕ) khϕe ⊗ idMm(A))(g∗j )iWe(hϕe ⊗ idMm(A))(gj)i)W ∗e − πe(ωj)k < 3m(A)2ε1 + 2γ2 < 1/32. (e 8.23) (e 8.24) (e 8.25) It follows from (e 8.23) that there exists selfadjoint elements bi,j ∈ Mm(A)(F2) such that exp(√−1bi,j) = ωj(ti)∗(πihϕ ⊗ idMm(A))(g∗j )i)Wi(πi(hϕ ⊗ idMm(A))(gj )i)W ∗i , (e 8.26) 68 and be,j ∈ Mm(A)(F1) such that exp(√−1be,j) = πe(ωj)∗(πehϕ ⊗ idMm(A))(g∗j )i)We(πe(hϕ ⊗ idMm(A))(gj )i)W ∗e , (e 8.27) and Write kbi,jk < 2 arcsin(3m(A)2ε1/4 + 2γ2), j = 1, 2, ..., k(A), i = 0, 1, ..., n, e. (e 8.28) bi,j = (b(1) i,j , b(2) i,j , ..., bF (2) i,j ) ∈ F2 and be,j = (b(1) e,j , b(2) e,j , ..., b(F (1)) e,j ) ∈ F1. We have that Note that h0(be,j) = b0,j and h1(be,j) = bn,j. (e 8.29) (πi(hϕ ⊗ idMm(A)(g∗j )i))Wi(πi(hϕ ⊗ idMm(A))(gj)i)W ∗i = πi(ωj) exp(√−1bi,j), (e 8.30) j = 1, 2, ..., k(A) and i = 0, 1, ..., n, e. Then, ns 2π (ts ⊗ TrMm(A))(b(s) i,j ) ∈ Z, (e 8.31) where ts is the normalized trace on Mns, s = 1, 2, ..., F (2), j = 1, 2, ..., k(A), and i = 0, 1, ..., n. We also have ms 2π (ts ⊗ TrMm(A))(b(s) e,j ) ∈ Z, (e 8.32) where ts is the normalized trace on Mms, s = 1, 2, ..., F (1), j = 1, 2, ..., k(A). Put λ(s) i,j = ns 2π (ts ⊗ TrMm(A))(b(s) i,j ) ∈ Z, where ts is the normalized trace on Mns, s = 1, 2, ..., n, j = 1, 2, ..., k(A) and i = 0, 1, 2, ..., n. Put λ(s) e,j = ms 2π (ts ⊗ TrMm(A))(b(s) e,j ) ∈ Z where ts is the normalized trace on Mms, s = 1, 2, ..., F (1) and j = 1, 2, ..., k(A). Denote λi,j = (λ(1) i,j , λ(2) i,j , ..., λ(F (2) i,j ) ∈ ZF (2), and λe,j = (λ(1) e,j , λ(2) e,j , ..., λ(F (1) e,j ) ∈ ZF (1). We have, by (e 8.28), λ(s) i,j ns < d/4, s = 1, 2, ..., F (2), and λ(s) e,j ms < d/4, s = 1, 2, ..., F (1), (e 8.33) (e 8.34) j = 1, 2, ..., k(A), i = 0, 1, 2, ..., n. Define α(0,1) define α(0,1) e i : K1(A) → ZF (2) by mapping [gj ] to λi,j, j = 1, 2, ..., k(A), i = 0, 1, 2, ..., n, and : K1(A) → ZF (1) by mapping [gj ] to λe,j, j = 1, 2, ..., k(A). We write K0(A⊗C(T)) = 69 K0(A) ⊕ β(K1(A))) (see 2.13 for the definition of β). Define αi : K∗(A ⊗ C(T)) → K∗(F2) as follows: On K0(A ⊗ C(T)), define αiK0(A) = [πi ◦ ϕ]K0(A), αiβ(K1(A)) = αi ◦ βK1(A) = α(0,1) i and on K1(A ⊗ C(T)), define αiK1(A⊗C(T)) = 0, i = 0, 1, 2, ..., n. Also define αe ∈ Hom(K∗(A ⊗ C(T)), K∗(F1)), by αeK0(A) = [πe ◦ ϕ]K0(A), αeβ(K1(A)) = αi ◦ βK1(A) = α(0,1) e on K0(A ⊗ C(T)) and (αe)K1(A⊗C(T)) = 0. Note that (h0)∗ ◦ αe = α0 and (h1)∗ ◦ αe = αn. (e 8.35) (e 8.36) (e 8.37) (e 8.38) Since A⊗ C(T) satisfies the UCT, the map αe can be lifted to an element of KK(A⊗ C(T), F1) which is still denoted by αe. Then define α0 = αe × [h0] and αn = αe × [h1] (e 8.39) in KK(A ⊗ C(T), F2). For i = 1, ..., n − 1, also pick a lifting of αi in KK(A ⊗ C(T), F2), and still denote it by αi. We estimate that k(w∗i wi+1)πti ◦ ϕ(g) − πti ◦ ϕ(g)(w∗i wi+1)k < ε1/4 for all g ∈ G5, i = 0, 1, ..., n − 1. Let Λi,i+1 : C(T) ⊗ A → F2 be a unital contractive completely positive linear map given by the pair w∗i wi+1 and πti ◦ ϕ (by 6.4, see 2.8 of [66]). Denote Vi,j = hπti ◦ ϕ ⊗ idMm(A)(gj)i, j = 1, 2, ..., k(A) and i = 0, 1, 2, ..., n − 1. Write Vi,j = (Vi,j,1, Vi,j,2, ..., Vi,j,F (2)) ∈ F2, j = 1, 2, ..., k(A), i = 0, 1, 2, ..., n. Similarly, write We have Wi = (Wi,1, Wi,2, ..., Wi,F (2)) ∈ F2, i = 0, 1, 2, ..., n. kWiV ∗i,jW ∗i Vi,jV ∗i,jWi+1Vi,jW ∗i+1 − 1k < 1/16 kWiV ∗i,jW ∗i Vi,jV ∗i+1,jWi+1Vi+1,jW ∗i+1 − 1k < 1/16 (e 8.40) (e 8.41) and there is a continuous path Z(t) of unitaries such that Z(0) = Vi,j and Z(1) = Vi+1,j. Since kVi,j − Vi+1,jk < δ1/12, j = 1, 2, ..., k(A), we may assume that kZ(t) − Z(1)k < δ1/6 for all t ∈ [0, 1]. We also write Z(t) = (Z1(t), Z2(t), ..., ZF (2)(t)) ∈ F2 and t ∈ [0, 1]. We obtain a continuous path WiV ∗i,jW ∗i Vi,jZ(t)∗Wi+1Z(t)W ∗i+1 70 which is in CU (Mnm(A)) for all t ∈ [0, 1] and kWiV ∗i,jW ∗i Vi,jZ(t)∗Wi+1Z(t)W ∗i+1 − 1k < 1/8 for all t ∈ [0, 1]. It follows that (1/2π√−1)(ts ⊗ TrMm(A))[log(Wi,sV ∗i,j,sW ∗i,sVi,j,sZs(t)∗Wi+1,sZs(t)W ∗i+1,s)] is a constant integer, where ts is the normalized trace on Mns. In particular, (1/2π√−1)(ts ⊗ TrMm(A))(log(Wi,sV ∗i,j,sW ∗i,sWi+1,sVi,j,sW ∗i+1,s)) = (1/2π√−1)(ts ⊗ TrMm(A))(log(Wi,sV ∗i,j,sW ∗i,sVi,jV ∗i+1,j,sWi+1Vi,j,sW ∗i+1,s)). (e 8.43) (e 8.42) One also has WiV ∗i,jW ∗i Vi,jV ∗i+1,jWi+1Vi+1,jW ∗i+1 = (ωj(ti) exp(√−1bi,j))∗ωj(ti) exp(√−1bi+1,j) = exp(−√−1bi,j)ωj(ti)∗ωj(ti+1) exp(√−1bi+1,j). Note that, by (e 8.21) and (e 8.18), for t ∈ [ti, ti+1], kωj(ti)∗ωj(t) − 1k < 3(3ε′1 + 2γ2) < 3/32, j = 1, 2, ..., k(A), i = 0, 1, ..., n − 1. By Lemma 3.5 of [74], It follows that (by the Exel formula (see [43]), using (e 8.43), (e 8.46) and (e 8.48)) (ts ⊗ Trm(A))(log(ωj,s(ti)∗ωj,s(ti+1))) = 0. (t ⊗ Trm(A))(bott1(Vi,j, W ∗i Wi+1)) (e 8.44) (e 8.45) (e 8.46) (e 8.47) (e 8.48) (e 8.49) (e 8.55) (e 8.56) (e 8.57) = ( = ( = ( = ( = ( 1 1 1 2π√−1 2π√−1 2π√−1 2π√−1 2π√−1 1 1 (e 8.50) )(t ⊗ Trm(A))(log(V ∗i,jW ∗i Wi+1Vi,jW ∗i+1Wi)) )(t ⊗ Trm(A))(log(WiV ∗i,jW ∗i Wi+1Vi,jW ∗i+1)) )(t ⊗ Trm(A))(log(WiV ∗i,jW ∗i Vi,jV ∗i+1,jWi+1Vi+1,jW ∗i+1)) )(t ⊗ Trm(A))(log(exp(−√−1bi,j)ωj(ti)∗ωj(ti+1) exp(√−1bi+1,j)) (e 8.53) )[(t ⊗ Trk(n))(−√−1bi,j) + (t ⊗ Trk(n))(log(ωj(ti)∗ωj(ti+1)) (e 8.54) (e 8.51) (e 8.52) +(t ⊗ Trk(n))(√−1bi,j)] = 1 2π (t ⊗ Trk(n))(−bi,j + bi+1,j) for all t ∈ T (F2). In other words, bott1(Vi,j, W ∗i Wi+1)) = −λi,j + λi+1,j j = 1, 2, ..., m(A), i = 0, 1, ..., n − 1. Consider α0, ..., αn ∈ KK(A ⊗ C(T), F2) and αe ∈ KK(A ⊗ C(T), F1). Note that αi(gj) = λi,j , 71 and by (e 8.33), one has ms, nj ≥ N2(8/d + 1). Applying 7.4 (using (e 8.34), among other items), there are unitaries zi ∈ F2, i = 1, 2, ..., n−1, and ze ∈ F1 such that k[zi, πti ◦ ϕ(g)]k < δu for all g ∈ Gu Bott(zi, πti ◦ ϕ) = αi and Bott(ze, πe ◦ ϕ) = αe. (e 8.58) (e 8.59) Put One verifies (by (e 8.39)) that z0 = h0(ze) and zn = h1(ze). Bott(z0, πt0 ◦ ϕ) = α0 and Bott(zn, πtn ◦ ϕ) = αn. (e 8.60) Let Ui,i+1 = zi(wi)∗wi+1(zi+1)∗, i = 0, 1, 2, ..., n − 1. Then k[Ui,i+1, πti ◦ ϕ(g)]k < min{δ1, δ2}, g ∈ Gu, i = 0, 1, 2, ..., n − 1. (e 8.61) Moreover, for i = 0, 1, 2, ..., n − 1, bott1(Ui,i+1, πti ◦ ϕ) = bott1(zi, πti ◦ ϕ)) + bott1((w∗i wi+1, πti ◦ ϕ)) +bott1((z∗i+1, πti ◦ ϕ) = (λi,j) + (−λi,j + λi+1,j) + (−λi+1,j) = 0. k=1 K∗(A ⊗ C(T), Z/kZ), one has N x = 0. Therefore Bott((Ui,i+1, ..., Ui,i+1 Note that for any x ∈L∗=0,1Ln0 } ts ◦ πt ◦ ϕ(h) ≥ ∆(h) for all h ∈ H′1, i = 0, 1, 2, ..., n − 1. Note that, by the assumption (e 8.13), ), (πti ◦ ϕ, ..., πti ◦ ϕ {z N {z } N ))P = N Bott(Ui,i+1, πti ◦ ϕ)P = 0, (e 8.62) (e 8.63) By applying 6.7, using (e 8.63), (e 8.61) and (e 8.62), there exists a continuous path of uni- where ts is the normalized trace on Mns, 1 ≤ s ≤ F (2). taries, { Ui,i+1(t) : t ∈ [ti, ti+1]} ⊂ F2 ⊗ MN (C) such that Ui,i+1(ti) = idF2⊗MN (C), Ui,i+1(ti+1) = (ziw∗i wi+1z∗i+1) ⊗ 1MN (C), (e 8.64) and k Ui,i+1(t)(πti ◦ ϕ(f ), ..., πti ◦ ϕ(f ) } {z N ) Ui,i+1(t)∗ − (πti ◦ ϕ(f ), ..., πti ◦ ϕ(f ) } {z N for all f ∈ F and for all t ∈ [ti, ti+1]. Define W ∈ C ⊗ MN by )k < ε/32 W (t) = (wiz∗i ⊗ 1MN ) Ui,i+1(t) for all t ∈ [ti, ti+1], i = 0, 1, ..., n − 1. Note that W (ti) = wiz∗i ⊗ 1MN , i = 0, 1, ..., n. Note also that W (0) = w0z∗0 ⊗ 1MN = h0(wez∗e ) ⊗ 1MN 72 (e 8.65) (e 8.66) and So W ∈ C ⊗ MN . One then checks that, by (e 8.18), (e 8.65) , (e 8.58) and (e 8.20), W (1) = wnz∗n ⊗ 1MN = h1(wez∗e ) ⊗ 1MN . kW (t)((πt ◦ ϕ)(f ) ⊗ 1MN )W (t)∗ − (πt ◦ ψ)(f ) ⊗ 1MNk < kW (t)((πt ◦ ϕ)(f ) ⊗ 1MN )W (t)∗ − W (t)((πti ◦ ϕ)(f ) ⊗ 1MN )W ∗(t)k +kW (t)(πti ◦ ϕ)(f )W (t)∗ − W (ti)(πti ◦ ϕ)(f )W (ti)∗k +kW (ti)((πti ◦ ϕ)(f ) ⊗ 1MN )W (ti)∗ − (wi(πti ◦ ϕ)(f )w∗i ) ⊗ 1MNk +kwi(πti ◦ ϕ)(f )w∗i − πti ◦ ψ(f )k +kπti ◦ ψ(f ) − πt ◦ ϕ(f )k < ε1/16 + ε/32 + δu + ε1/16 + ε1/16 < ε (e 8.67) (e 8.68) (e 8.69) (e 8.70) (e 8.71) (e 8.72) (e 8.73) for all f ∈ F and for t ∈ [ti, ti+1]. Remark 8.5. With a minor modification, the proof also works without assuming that K∗(A) is finitely generated. In Theorem 8.4, the multiplicity N only depends on K(A) as K(A) is finitely generated. However, if K∗(A) is not finitely generated, it also depends on F and ǫ. Moreover, if K∗(A) is torsion free, or if K1(C) = 0, then the multiplicity N can be chosen to be 1. Corollary 8.6. Theorem 8.4 holds if A is replaced by Mm(A) for any integer m ≥ 1. 9 C∗-algebras in B1 Definition 9.1. Let A be a unital infinite dimensional simple C∗-algebra. We say A ∈ B1 if the following hold: Let ε > 0, let a ∈ A+ \ {0} and let F ⊂ A be a finite subset. There exists a nonzero projection p ∈ A and a C∗-subalgebra C ∈ C with 1C = p such that kxp − pxk < ε for all x ∈ F, dist(pxp, C) < ε for all x ∈ F and 1 − p . a. (e 9.1) (e 9.2) (e 9.3) In the above, if C can always be chosen in C0, that is, K0(C) = {0}, then we say that A ∈ B0. Definition 9.2. Let A be a unital simple C∗-algebra. We say A has the generalized tracial rank at most one, if the following hold: Let ε > 0, let a ∈ A+ \ {0} and let F ⊂ A be a finite subset. There exists a nonzero projection p ∈ A and a C∗-subalgebra C which is a subhomogeneous C∗-algebra with one dimensional spectrum, or C is finite dimensional C∗-algebra with 1C = p such that kxp − pxk < ε for all x ∈ F, dist(pxp, C) < ε for all x ∈ F and 1 − p . a. (e 9.4) (e 9.5) (e 9.6) In this case, we write gT R(A) ≤ 1. Remark 9.3. It follows from 3.21 that gT R(A) ≤ 1 if and only if A ∈ B1. Let D be a class of unital C∗-algebras. 73 Definition 9.4. Let A be a unital simple C∗-algebra. We say A is tracially approximately in D, denoted by A ∈ TAD, if the following hold: p ∈ A and a C∗-subalgebra C ∈ D with 1C = p such that For any ε > 0, any a ∈ A+\{0} and any finite subset F ⊂ A, there exist a nonzero projection kxp − pxk < ε for all x ∈ F, dist(pxp, C) < ε for all x ∈ F, and 1 − p . a. (e 9.7) (e 9.8) (e 9.9) (see Definition 2.2 of [30]). Note that B0 = TAC0 and B1 = TAC. If in the above definition, only (e 9.7) and (e 9.8) hold, then we say A has the property (LD). The following proposition was first appear in an unpublished paper of the second named author distributed in 1998. Proposition 9.5. Let A be a unital simple C∗-algebra which has the property (LD). Then, for any ε > 0 and any finite subset F ⊂ A, there exists a projection p ∈ A and a C∗-subalgebra C ∈ D with 1C = p such that k[x, p]k < ε for all f ∈ F, dist(pxp, C) < ε, and, kpxpk ≥ kxk − ε for all x ∈ F. (e 9.10) (e 9.11) (e 9.12) Proof. Fix ε > 0 and a finite subset F ⊂ A. It is clear that, without loss of generality, we may assume that x ∈ A+ and kxk = 1 for all x ∈ F. Let f ∈ C(([0, 1])+ be such that f (t) = 0 if t ∈ [0, 1 − ε/2], f (t) = 1 if t ∈ [1 − ε/4, 1] and f (t) is linear in (1 − ε/2, 1 − ε/4). For each such x ∈ F, there exist y1(x), y2(x), ..., yk(x),x ∈ A such that yi(x)∗f (x)yi(x) = 1A. (e 9.13) k(x)Xi=1 Put 1 σ(x) = and σ = min{σ(x) : x ∈ F}. Choose δ > 0 and a large finite subset F1 ⊃ F such that, for any projection q, k(x)(max{1,kyi(x)k2} implies that, for all x ∈ F, kqy − yqk < δ for all y ∈ F1 (e 9.14) qyi(x)∗qf (x)qyi(x)q − qk < σ/16 and kf (qxq) − qf (x)qk < σ/16. (e 9.15) k k(x)Xi=1 Now, since A has property (LD), there is a projection p ∈ A and a C∗-subalgebra C ∈ D with 1C = p such that kpy − ypk < min{ε/2, δ} and dist(pyp, C) < ε for all y ∈ F1. (e 9.16) Therefore, by the choice of δ and F1, k k(x)Xi=1 pyi(x)pf (x)pyi(x)pk ≥ 1 − σ/16 and kf (pxp) − pf (x)pk < σ/16 for all x ∈ F. 74 (e 9.17) (e 9.18) Therefore It follows that Therefore kpf (x)pk ≥ 15σ/16 for all x ∈ F. kf (pxp)k ≥ 14σ/16. kpxpk > 1 − ε for all x ∈ F. (e 9.19) (e 9.20) Theorem 9.6. Let A be a unital separable simple C∗-algebra in B1 (or in B0). Then either A is an inductive limit of unital C∗-algebras An ∈ C (or An ∈ C0) with A =S∞n=1 An, or A has the property (SP). Proof. This follows from Definition 9.1 immediately. Let F1,F2, ...,Fn, ... be a sequence of increasing finite subsets of the unit ball of A whose union is dense in the unit ball. If A does not have property (SP), then there is a non-zero positive element a ∈ A such that aAa 6= A and aAa has no non-zero projection. Then, for each n ≥ 1, there is a projection 1A − pn . a and a C∗-subalgebra Cn ∈ C (or C0) such that 1Cn = pn and kpnx − xpnk < 1/2n and dist(pnxpn, Cn) < 1/2n for all x ∈ Fn. (e 9.21) Since aAa does not have any non-zero projection, one has 1A − pn = 0. In other words, 1A = pn and dist(x, Cn) < 1/2n for all x ∈ Fn, n = 1, 2, .... (e 9.22) It follows thatS∞n=1 Cn = A. Since each C∗-algebra Cn is semiprojective (see 3.1, also Theorem 6.22 of [21]), A is in fact an inductive limit of C∗-algebras in C0 (or C0). Theorem 9.7. Let A ∈ B1. Then A has stable rank one. Proof. This follows from (3.3) and 4.3 of [30] (or a similar result in [33]). Lemma 9.8. Let D be a family of unital separable C∗-algebras which are residually finite dimensional. Any unital separable simple C-algebra with property LD can be embedded in Q Mr(n)/L Mr(n) for some sequence of integers {r(n)}. Proof. Let A be a unital separable simple C-algebra with property (LD). Let F1 ⊆ F2 ⊆ ··· ⊆ Fi ⊆ ··· be an increasing sequence of finite subsets of A with the union being dense in A. Since A has property (LD), for each n, there is a projection pn ∈ A and Cn ⊆ A with 1Cn = pn and Cn ∈ D such that kpnf − f pnk < 1/2n+2, kpnf pnk ≥ kfk − 1/2n+2 and pnf pn ∈1/2n+2 Cn for all f ∈ Fn.(e 9.23) For each a ∈ Fn, there exists c(a) ∈ Cn such that kpnapn − c(a)k < 1/2n+1. There are unital homomorphisms π′n : Cn → Bn, where Bn is a finite dimensional C∗-subalgebra such that kπ′n(c(a))k = kc(a))k ≥ kak − 1/2n+1 for all ∈ Fn, n = 1, 2, .... (e 9.24) There is an integer r(n) ≥ 1 such that Bn is unitally embedded into Mr(n). Denote by πn : Cn → Mr(n) the composition of π′n and the embedding. Note Cn ⊂ pnApn. Then there is contractive completely positive linear map Φ′n : pnApn → Mr(n) such that Φ′nCn = πn. 75 (e 9.25) Define Φn : A → Mr(n) by Φn(a) = Φ′n(pnapn) for all a ∈ A. It is a unital contractive completely positive linear map. Moreover, kΦn(pnapn) − Φn(c(a))k < 1/2n+1 for all a ∈ Fn, n = 1, 2, ...,∞. Combining with (e 9.23), we obtain that kΦn(f )k ≥ kfk − 1/2n for all f ∈ Fn, n = 1, 2, .... Define Φ : A →Q∞n=1 Mr(n) by Φ(a) = {π′n ◦ Φn(a)} for all a ∈ A. Let ∞Yn=1 ∞Mn=1 Π : Mr(n) → Mr(n)/ Mr(n) ∞Yn=1 (e 9.26) (e 9.27) be the quotient map. Put Ψ = Π◦Φ. One easily checks that Ψ is in fact a unital homomorphism. Moreover, by (e 9.24), Ψ is not zero. Since A is simple, it is a unital monomorphism. Lemma 9.9. Let D be the family of unital separable residually finite dimensional C∗-algebras and let A be a unital simple separable C∗-algebra which has property (LD) and property (SP). Then A satisfies the following Popa condition: Let ε > 0 and let F ⊂ A be a finite subset. There exists a finite dimensional C∗-subalgebra F ⊂ A with P = 1F such that k[P, x]k < ε, P xP ∈ε F and kP xPk ≥ kxk − ε (e 9.28) for all x ∈ F. In particular, if A ∈ B1, then A satisfies the Popa condition. Proof. We may assume that F ⊂ A1 and 0 < ε < 1/2. Without loss of generality, we may assume that Since A has property (LD), there is a projection p ∈ A and a C∗-subalgebra D ⊆ A with D ∈ D and p = 1D such that d = min{kxk : x ∈ F} > 0. kpx − xpk < dǫ/16, pxp ∈dǫ/16 D and kpxpk ≥ (1 − ε/16)kxk (e 9.29) for all x ∈ F (see 9.5). Let F′ ⊂ D be a finite subset such that, for each x ∈ F, there exists x′ ∈ F′ such that kpxp − x′k < dε/16. Since D ∈ D, there is a unital surjective homomorphism π : D → D/kerπ such that F1 := D/kerπ is a finite dimensional C∗-algebra and kπ(x′)k ≥ (1 − ε/16)kx′k for all x′ ∈ F′. (e 9.30) Let B = kerπAkerπ. B is a hereditary C∗-subalgebra of A. Let C be the closure of D + B. Note that 1C = 1D = p. As in the proof of 5.2 of [51], B is an ideal of C and C/B ∼= D/kerπ = F1. The lemma then follows from Lemma 2.1 of [81]. In fact, since pAp has property (SP), by Lemma 2.1 of [81], there a projection P ∈ pAp and a monomorphism h : F1 → P AP such that h(1F ) = P, kP x′ − x′Pk < ε/16 and kh ◦ π(x′) − P x′Pk < ε · d/16 (e 9.31) (e 9.32) for all x′ ∈ F′. 76 Put F = h(F1). Then, one estimates that, for all x ∈ F, kP x − xPk ≤ kP px − P x′k + kP x′ + x′Pk + kx′P − xpPk < ε/16 + ε/16 + ε/16 < ε, P xP ≈ε/16 P x′P ∈ε/16 F1 and kP xPk = kP pxpPk ≥ kP x′Pk − dε/16 ≥ kh ◦ π(x′)k − dε/8 = kπ(x′)k − dε/8 ≥ kx′k − dε/16 − dε/8 ≥ kpxpk − dε/4 ≥ (1 − ε/16)kxk − dε/4 ≥ kxk − ε. (e 9.33) (e 9.34) (e 9.35) (e 9.36) (e 9.37) (e 9.38) Theorem 9.10. Let A ∈ B1 (or A ∈ B0). Then, for any projection p ∈ A, one has that pAp ∈ B1 (or pAp ∈ B0). Proof. Let 1/4 > ε > 0, let a ∈ (pAp)+ \ {0} and let F ⊂ pAp be a finite subset. Since A is unital and simple, there are x1, x2, ..., xm ∈ A such that x∗i pxi = 1A. mXi=1 (e 9.39) Put F1 = {p, x1, x2, ..., xm, x∗1, x∗2, ..., x∗m} ∪ F. Let K = m2 max{kxk : x ∈ F1}. Since A ∈ B1, there is a projection e ∈ A and a unital C∗-subalgebra C1 ∈ C (or C1 ∈ C0) with 1C1 = e such that kxe − exk < ε/64(K + 1) for all x ∈ F1, dist(exe, C1) < ε/64(K + 1) for all x ∈ F1 and 1 − e . a. Since p ∈ F1, there is a projection q ∈ C1 such that It follows that kepe − qk < ε/64(K + 1). kpep − qk < ε/32(K + 1). Moreover, there are y1, y2, ..., ym ∈ C1 such that k mXi=1 y∗i qyi − ek < ε. (e 9.40) (e 9.41) (e 9.42) (e 9.43) (e 9.44) It follows that q is full in C1. It follows from 3.19 that qC1q ∈ C(or qC1q ∈ C0). There is a unitary u ∈ A such that ku − 1k < ε/16(K + 1) and u∗qu ≤ p. Put q1 = u∗qu and C = u∗qC1qu. Then C ∈ C (or C ∈ C0) and 1C = q1. We also have kepe − q1k < ε/64(K + 1) + ε/8(K + 1) = 9ε/64(K + 1). If x ∈ F, then kq1x − xq1k ≤ 2k(q1 − epe)xk + kepex − xepek < 18ε/64 + ε/16(K + 1) < ε for all x ∈ F. 77 (e 9.45) (e 9.46) (e 9.47) Similarly, we estimate that We also have dist(qxq, C) < ε for all x ∈ F. (e 9.48) k(p − q1) − (p − pep)k = kq1 − pepk < ε/32(K + 1) + ε/8(K + 1) = 5ε/32(K + 1). (e 9.49) Put η = 5ε/32(K + 1) < 1/16. Let fη/2(t) ∈ C0((0,∞) be as in 2.5. Then, by 2.2 of [92], p − q1 = fη(p − q1) . p − pep . 1 − e . a. (e 9.50) This shows that pAp ∈ B1. Theorem 9.11. Let D be a class of until C∗-algebras which is closed under tensor products with a finite dimensional C-algebra and which satisfies the strict comparison property for positive elements (see 2.6). Let A be a unital simple C-algebra in the class TAD, then A has strict comparison for positive elements. In particular, if A ∈ B1, then A has strictly comparison for positive elements and K0(A) is weakly unperforated. Proof. By a result of Rødram (see, for example, Corollary 4.6 of [95]), to show that A has strict comparison for positive elements, it is enough to show that W (A) is almost unperforated, i.e., for any positive elements a, b in a matrix algebra over A, if (n + 1)[a] ≤ n[b] for some n ∈ N, then [a] ≤ [b]. Let a, b be such positive elements. Since any matrix algebra over A is still in TAD, let us assume that a, b ∈ A. First we consider the case that A does not have (SP) property. In this case, by the proof of 9.6, A = ∪∞n=1An, where An ∈ D ({An} be not be increasing). Without loss of generality, we may assume that 0 ≤ a, b ≤ 1. Let ε > 0. It follows from an argument of Rordam (see Lemma 5.6 of [85]) that there exists an integer m ≥ 1, a′, b′ ∈ Am such that n+1 diag( fε/2(a′), fε/2(a′), ..., fε/2(a′)) . diag( z ka′ − ak < ε/2, kb′ − bk < ε/2, b′ . b and } b′, b′, ..., b′) in Am. } { { n z Since Am has strict comparison (see part (b) of 3.18), one has It follows, using 2.1 of [92], that fε/2(a′) . b′ in Am. fε(a) . fε/2(fε/2((a)) . fε/2(a′) . b′ . b (e 9.51) (e 9.52) (e 9.53) (e 9.54) for every ε > 0. It follows that a . b. Now we assume that A has (SP). Let 1/4 > ε > 0. We may further assume that kbk = 1. Since A has (SP) and simple, there are mutually orthogonal and mutually equivalent non-zero projections e1, e2, ..., en+1 ∈ f3/4(b)Af3/4(b). Put E = e1 + e2 +··· + en+1. By 2.4 of [92], we also have that (n + 1)[fε/2(a)] ≤ n[fδ(b)] (e 9.55) 78 for some ε > δ > 0. Put η = min{ε/4, δ/4, 1/8}. It follows from the definition 9.1 that there is a C∗-subalgebra C = pAp ⊕ S with S ∈ D and a′, b′, E′, e′i ∈ C (i = 1, 2, ..., n + 1) such that 0 ≤ a′, b′ ≤ 1 and E′, e′i are projections in C, a − a′ < η, b′ . fδ(b),kf1/2(b′)E′ − E′k < η, n+1Xi=1 e′i, kei − e′ik < η and kE − E′k < η < 1, E′ = (e 9.56) (e 9.57) and n+1 } { z diag( b′, b′, ..., b′) and (n + 1)[E′] < [b′] fε/2(a′), fε/2(a′), ..., fε/2(a′)) . diag( } (see Lemma 5.6 of [85]). Moreover, the projection p can be chosen so that p . e1. From (e 9.56), there is a projection e′′i , E′′ ∈ f1/2(b′)Cf1/2(b′) (i = 1, 2, ..., n + 1) such that kE′ − E′′k < 2η, ke′′i − e′ik < 2η, i = 1, 2, ..., n + 1 and E′′ = Pn+1 i=1 e′′i (we also assume that e′′1, e′′2, ..., e′′n+1 are mutually orthogonal). Note that e′i and e′′i are equivalent. Choose a function g ∈ C0((0, 1])+ with g ≤ 1 such that g(b′)f1/2(b′) = f1/2(b′) and [g(b′)] = [b′] in W (C). In particular, g(b′)E′′ = E′′. in C (e 9.58) n z { Write a′ = a′0 ⊕ a′1, g(b′) = b′0 ⊕ b′1, e′′i = ei,0 ⊕ ei,1 and E′′ = E′0 ⊕ E′1 with a′0, b′0, E′0, ei,0 ∈ pAp and a′1, b′1, ei,1, E′1 ∈ S, i = 1, 2, ..., n+1. Note that E′1b′1 = E′1b′1 = E′1. This, in particular, implies that τ (b′1) ≥ (n + 1)τ (e1,1) for all τ ∈ T (S). (e 9.59) It follows from (e 9.58) that dτ (fε/2(a′1)) ≤ n n + 1 dτ (b′1), for all τ ∈ T(S). Since (b′1 − e1,1)e1,1 = 0 and b′1 = (b′1 − e1,1) + e1,1, for all τ ∈ T (S), dτ ((b′1 − e1,1)) = dτ (b′1) − τ (e1,1) > dτ (b′1) − 1 n + 1 dτ (b′1) ≥ dτ (fε/2(a′1))). Since S has the strict comparison, one has and therefore fε/2(a′1)+ . (b′1 − e′1), fε(a) . fε/2(a′) . p ⊕ fε/2(a′1) . p ⊕ (b′1 − e1,1) . e1 ⊕ (b′1 − e1,1) . e1 ⊕ (b′1 − e1,1) + (b′0 − e1,0) ∼ e′′1 ⊕ (g(b′) − e′′1) ∼ g(b′) ∼ b′ . b. (e 9.60) (e 9.61) (e 9.62) Since ǫ is arbitrary, one has that a . b. Hence one always has that a . b, and therefore W (A) is almost unperforated. Let F,G ⊆ C be finite subsets, let ǫ > 0 and δ > 0 be positive. Let H ⊆ C 1 Lemma 9.12. Let D be a class of unital amenable C-algebras, let A be a separable unital C-algebra which is T AD and let C be a unital ( amenable) C-algebra. + be a finite subset, and let T : C+ \ {0} → R+ \ {0} and N : C+ \ {0} → N be maps. Let ∆ : C q,1 + \ {0} → (0, 1) be an order preserving map. Let H1 ⊆ C 1 + \ {0}, H2 ⊆ Cs.a. and U ⊆ U (C)/CU (C) be finite subsets. Let σ1 > 0 and σ2 > 0 be constants. Let ϕ, ψ : C → A be two unital G-δ-multiplicative linear maps such that 79 (1) ϕ and ψ are T × N -H-full (see the definition 2.20), (2) τ ◦ ϕ(c) > ∆(c) and τ ◦ ψ(c) > ∆(c) for any c ∈ H1, (3) τ ◦ ϕ(c) − τ ◦ ψ(c) < σ1 for any τ ∈ T(A) and any c ∈ H2, (4) dist(ϕ‡(u), ψ‡(u)) < σ2 for any u ∈ U . Then, for any finite subset F′ ⊆ A and ǫ′ > 0, there exists a C-subalgebra D ⊆ A with (a) kpa − apk < ǫ′, (b) pap ∈ǫ′ D, (c) τ (1 − p) < ǫ′, for any τ ∈ T (A). There are also (completely positive) linear map j0 : A → (1 − p)A(1 − p) and a unital D ∈ D such that if p = 1D, then, for any a ∈ F′, contractive completely positive linear map j1 : A → D such that j0(a) = (1 − p)a(1 − p) for all a ∈ A and kj1(a) − papk < 3ǫ′ for all a ∈ F. Moreover, define ϕ0 = j0 ◦ ϕ, ψ0 = j0 ◦ ψ, ϕ1 = j1 ◦ ϕ and ψ1 = j1 ◦ ψ. With a sufficiently large F′ and small enough ǫ′, one has that ϕ0, ψ0, ϕ1 and ψ1 are G-2δ- multiplicative and (5) kϕ(c) − (ϕ0(c) ⊕ ϕ1(c))k < ǫ and kψ(c) − (ψ0(c) ⊕ ψ1(c))k < ǫ, for any c ∈ F, (6) ϕ0, ψ0 and ϕ1, ψ1 are 2T × N -H-full, (7) τ ◦ ϕ1(c) > ∆(c)/2 and τ ◦ ψ1(c) > ∆(c)/2 for any c ∈ H1, (8) τ ◦ ϕ1(c) − τ ◦ ψ1(c) < 2σ1 for any τ ∈ T(D) and any c ∈ H2, (9) dist(ϕ‡i (u), ψ‡i (u)) < 2σ2 for any u ∈ U , i = 0, 1. Proof. Without loss of generality, one may assume that each element of F, G, H2, or F′ has norm at most one and 1A ∈ F′. Since ϕ and ψ are T × N -H-full, for each h ∈ H, there are a1,h, ..., aN (h),h and b1,h, ..., bN (h),h in A with kai,hk,kbi,hk ≤ T (h) such that a∗i,hϕ(h)ai,h = 1A and N (h)Xi=1 N (h)Xi=1 b∗i,hψ(h)bi,h = 1A. (e 9.63) Put d0 = min{∆(h) : h ∈ H1}. By (3), there are, for each c ∈ H2, x1,c, x2,c, ..., xt(c),c ∈ A such that k t(c)Xi=1 x∗i,cxi,c − ϕ(c)k < σ1 and k t(c)Xi=1 xi,cx∗i,c − ψ(c)k < σ1. (e 9.64) Let t(H2) = max{kxi,ck : 1 ≤ i ≤ t(c) : c ∈ H2}. 80 For the given finite subset F′ ⊆ A and given ǫ′ > 0, since A can be tracially approximated by the C∗-algebras in the class D, there exists a C-subalgebra D ⊆ A with D ∈ D such that if p = 1D, then, for any a ∈ F′, (i) kpa − apk < ǫ′, (ii) pap ∈ǫ′ D, (iii) τ (1 − p) < ǫ′, for any τ ∈ T (A). To make F′ large and ε′ small, we may assume that F′ also contains F, G, H, ϕ(G ∪ F), ψ(G ∪F), ϕ(H), ψ(H), H1, H2, xi,c, x∗i,c, i = 1, 2, ..., t(c) and c ∈ H2, as well as ai,h, a∗i,h, bi,h, b∗i,h, i = 1, 2, ..., N (h) and h ∈ H, and ε′ < min{min{1/64(T (h) + 1)(N (h) + 1) : h ∈ H}, ε, δ, d0, σ1, σ2}/64(t(H2) + 1)2. For each a ∈ F′, choose da ∈ D such that kpap − dak < ǫ′ (choose d1A = 1D). Consider the finite subset {dadb : a, b ∈ F′} ⊆ D. Since D is an amenable C∗-subalgebra of pAp, there is unital completely positive linear map L : pAp → D such that Define j1 : A → D by j1(a) = L(pap). Then, for any a ∈ F′, one has kL(dadb) − dadbk < ǫ′, a, b ∈ F′. kj1(a) − papk = kL(pap) − papk = kL(da) − dak + 2ǫ′ = 2ε′. (e 9.65) Note that j0 and j1 are F′-7ǫ′ -multiplicative, and ka − j0(a) ⊕ j1(a)k < 4ǫ′ for all a ∈ F′. Define ϕ0 = j0 ◦ ϕ, ψ0 = j0 ◦ ψ, ϕ1 = j1 ◦ ϕ and ψ1 = j1 ◦ ψ. Then (by the choices of F′ and ε′), maps ϕ0, ψ0, ϕ1, and ϕ1 are G-2δ-multiplicative, and for any c ∈ F, kϕ(c) − (ϕ0(c) ⊕ ϕ1(c))k < ǫ and kψ(c) − (ψ0(c) ⊕ ψ1(c))k < ǫ. (e 9.66) Apply j1 on both sides of both equations in (e 9.63). One obtains two invertible elements eh :=PN (h) i=1 j1(a∗i )ϕ1(h)j1(ai) and fh :=PN (h) i=1 j1(b∗i )ψ1(h)j1(bi) such that ke− 1 h k − 1 < 1 and kf− 1 2 2 h k − 1 < 1. Note that N (h)Xi=1 h j1(a∗i )ϕ1(h)j1(ai)e− 1 e− 1 2 2 h = 1D, h j1(b∗i )ψ1(h)j1(bi)f− 1 f− 1 2 2 h = 1D, N (h)Xi=1 and kj1(bi)f− 1 kj1(ai)e− 1 2 h k < 2T (h) 2 h k < 2T (h). Therefore, ϕ1 and ψ1 are 2T × N -H-full, and this proves (6). The same calculation also shows that ϕ0 and ψ0 are 2T × N -H-full. Note that we have shown (5) and (5) hold. Since ε′ < d0/16, it is also easy to check that (7) holds. To see (8), one notes that k t(c)Xi=1 d∗xi,cdxi,d − ϕ1(c)k < 2σ1 and k t(c)Xi=1 81 dxi,cd∗xi,d − ψ1(c)k < 2σ1 (e 9.67) for all c ∈ H2. Then (8) also holds. Let us show that (9) holds with sufficiently large F′ and sufficiently small ǫ′. Choose unitaries u1, u2, ..., un ∈ C such that U = {u1, u2, ..., un}. Pick unitaries w1, w2, ..., wn ∈ A such that each wi is a commutator and dist(hϕ(ui)ihψ(u∗i )i , wi) < σ2. Choose F′ sufficiently large and ǫ′ sufficiently small such that there are commutators w′1, w2, ..., w′n ∈ CU (D) and commutators w′′1 , w′′2 , ..., w′′n ∈ (1 − p)A(1 − p) satisfying kj1(wi) − w′ik < σ2/2 and kj0(wi) − w′′i k < σ2/2, 1 ≤ i ≤ n, (see Appendix of [73]) and khϕk(ui)ihψk(u∗i )i − jk(hϕ(ui)ihψ(u∗i )i)k < σ2/2, 1 ≤ i ≤ n and k = 0, 1. Then khϕk(ui)ihψk(u∗i )i − w′ik ≤ khϕk(ui)ihψk(u∗i )i − jk(wi)k + kjk(wi) − w′ik ≤ khϕk(ui)ihψk(u∗i )i − jk(hϕ(ui)ihψ(u∗i )i)k + kjk(hϕ(ui)ihψ(u∗i )i) − jk(wi)k + σ2/2 ≤ 2σ2, k = 0, 1. (e 9.68) (e 9.69) (e 9.70) (e 9.71) (e 9.72) This proves (9). 10 Z-stability Lemma 10.1. Let A ∈ B1 (or B0) be a unital infinite dimensional simple C∗-algebra. Then, for any ε > 0, any a ∈ A+ \ {0}, any finite subset F ⊂ A and any integer N ≥ 1, there exists a projection p ∈ A and a C∗-subalgebra C ∈ C (or C0) with 1C = p satisfies the following: (1) dim(π(C)) ≥ N 2 for every irreducible representation π of C; (2) kpx − xpk < ε for all x ∈ F; (3) dist(pxp, C) < ε for all x ∈ F and (4) 1 − p . a. Proof. Since A is an infinite dimensional simple C∗-algebra, there are N +1 mutually orthogonal non-zero positive elements a1, a2, ..., aN +1 in A. Since A is simple, there are xi,j ∈ A, j = 1, 2, ..., k(i), i = 1, 2, ..., N + 1, such that x∗i,jaixi,j = 1A. k(i)Xj=1 Let K = (N + 1) max{kxi,jk + 1 : 1 ≤ j ≤ k(i), 1 ≤ i ≤ N + 1}. Let a0 ∈ A+ \ {0} be such that a0 . ai for all 1 ≤ i ≤ N + 1. Since a0Aa0 is also an infinite dimensional simple C∗-algebra, one obtains a01, a02 ∈ a0Aa0 which are mutually orthogonal and nonzero. One then obtains a non-zero element a ∈ a01Aa01 such that a . a02. Let F = {ai : 1 ≤ i ≤ N + 1} ∪ {xi,j : 1 ≤ j ≤ k(i), 1 ≤ i ≤ N + 1} ∪ {a}. 82 Now since A ∈ B1, there is a projection p ∈ A and C ∈ C with 1C = p such that (1) kxp − pxk < min{1/2, ε}/2K for all x ∈ F, (2) dist(pxp, C) < min{1/2, ε}/2K for all x ∈ F, and (3) 1 − p . a. Thus, with a standard computation, we obtain mutually orthogonal non-zero positive ele- ments b1, b2, ..., bN +1 ∈ C and yi,j,∈ C (1 ≤ j ≤ k(i)), i = 1, 2, ..., N + 1, such that k k(i)Xj=1 y∗i,jbiyi,j − pk < min{1/2, ε/2}. For each i, we find another element zi ∈ C such that z∗i yi,jbiyi,jzi = p. k(i)Xj=1 (e 10.1) (e 10.2) Let π be an irreducible representation of C. Then by (e 10.2), π(z∗i yi,j)π(bi)π(yi,jzi) = π(p). (e 10.3) k(i)Xj=1 Therefore π(b1), π(b2), ..., π(bN +1) are mutually orthogonal non-zero positive elements in π(A). Then (e 10.3) implies that π(C) ∼= Mn with n ≥ N + 1. This proves the lemma. Corollary 10.2. Let A ∈ B1 be a unital simple C∗-algebra. Then, for any ε > 0 and f ∈ Aff(T (A))++, there exists a C∗-subalgebra C ∈ C0 in A, an element c ∈ C+ such that dimπ(C) ≥ (4/ε)2 for each irreducible representation π of C, 0 < τ (f ) − τ (c) < ε/2 for all τ ∈ T (A). (e 10.4) (e 10.5) The following is known. Lemma 10.3. Let C = Mn([0, 1]) and g ∈ LAff b(T (C))+. Then there exists a ∈ C+ with 0 ≤ a ≤ 1 such that 0 ≤ g(t) − dt(a) ≤ 1/n for all t ∈ [0, 1], where dt(a) = limk→∞ a1/k(t) for all t ∈ [0, 1]. Proof. We will use the proof of Lemma 5.2 of [9]. For each 0 ≤ i ≤ n − 1, define Xi = {t ∈ [0, 1] : g(t) > i/n}. Since g is lower semi-continuous, Xi is open in [0, 1]. There is a continuous function gi ∈ C([0, 1])+ with 0 ≤ gi ≤ 1 such that {t ∈ [0, 1] : gi(t) 6= 0} = Xi, i = 0, 1, ..., n − 1. Let e1, e2, ..., en be n mutually orthogonal rank one projections in C = Mn(C([0, 1]). Define a = n−1Xi=1 giei ∈ C. 83 (e 10.6) Then 0 ≤ a ≤ 1. Put Yi = {t ∈ [0, 1] : (i+1)/n ≥ gi(t) > i/n} = Xi\Sj>i Xj, i = 0, 1, 2, ..., n−1. These are mutually disjoint sets. Note that [0, 1] = ([0, 1] \ X0) ∪ Yi. n−1[i=0 If x ∈ ([0, 1] \ X0) ∪ Y0, dt(a) = 0. So 0 ≤ g(t) − dt(a)(t) ≤ 1/n for all such t. If t ∈ Yj, dt(a) = j/n. Then It follows that 0 ≤ g(t) − dt(a) ≤ 1/n for all t ∈ Yj. 0 ≤ g(t) − dt(a) ≤ 1/n for all t ∈ [0, 1]. (e 10.7) (e 10.8) (e 10.9) Lemma 10.4. Let F1 and F2 be two finite dimensional C∗-algebras such that each simple sum- mand of F1 and F2 has rank at least k, where k ≥ 1 is an integer. Let ϕ0, ϕ1 : F1 → F2 be unital homomorphisms. Let C = A(ϕ0, ϕ1, F1, F2). Then, for any f ∈ LAff b(T (C))+ with 0 ≤ f ≤ 1, there exists a positive element a ∈ M2(C) such that max τ∈T (C)dτ (a) − f (τ ) ≤ 2/k. Proof. Let I = {g ∈ C : g(0) = g(1) = 0}. Note that C/I is a finite dimensional C∗-algebra. Let T = {τ ∈ T (C) : kerτ ⊃ I}. Then T may be identified with T (C/I). Let f ∈ LAff b(T (C))+ with 0 ≤ f ≤ 1. It is easy to see that there exists b ∈ (C/I)+ for some integer m1 ≥ 1 such that 0 ≤ f (τ ) − dτ (b) for all τ ∈ T and max{f (τ ) − dτ (b) : τ ∈ T} ≤ 1/k, (e 10.10) m < f (τq(s)) ≤ r+1 and furthermore, if f (τ ) > 0, then f (τ )−dτ (b) > 0. As matter of fact, we can choose b to be rank r projection in ith block MR(s) of F1 and r m , where τq(s) is the normalized trace of MR(s) regarded as an element in T (C/I) (since m ≥ k). For such a choice, we have dτ (b) = τ (b) for all τ . Note that b = (bq(1), bq(2), ..., bq(l)) ∈ C/I = F1 = MR(1) ⊕ MR(2) ⊕ ··· ⊕ MR(l). Let b0 = ϕ0(b) ∈ F2 and b1 = ϕ1(b) ∈ F2. Write F2 = Mr1 ⊕ Mr2 ⊕ ··· ⊕ Mrm. Write b0 = b0,1 ⊕ b0,2 ⊕ ··· ⊕ b0,rm and b1 = b1,1 ⊕ b1,2 ⊕ ··· ⊕ b1,rm, where b0,j, b1,j ∈ Mrj , j = 1, 2, ..., m. Let τt,j = trj ◦ Ψj ◦ πt, where trj is the normalized trace on Mrj , Ψj : F2 → Mrj is the quotient map and πt : A → F2 is the evaluation at t ∈ (0, 1). Since f is lower semi-continuous on T (C), lim inf t→0 f (τt,j) ≥ trj(b0,j) and lim inf t→1 f (τt,j) ≥ trj(b1,j). s=1 αs = 1. Therefore, if Note that trj(b0,j) = Pl trj(b0,j) > 0, then s=1 αstrq(s)(bq(s)) for some αs ≥ 0 with Pl lXs=1 αstrq(s)(bq(s)) = trj(b0,j). lXs=1 αstrq(s)) > f ( 84 Hence, if trj(b0,j) > 0 (or trj(b1,j) > 0), then lim inf t→0 f (τt,j) > trj(b0,j) (or lim inf t→1 f (τt,j) > trj(b1,j)). Therefore, there exists 1/8 > δ > 0, such that f (τt,j) ≥ trj(b0,j) for all t ∈ (0, 2δ) and f (τt,j) ≥ trj(b1,j) for all t ∈ (1 − 2δ, 1), j = 1, 2, ..., l. Let δ − t δ )b0 if t ∈ [0, δ) c(t) = ( c(t) = 0 if t ∈ [δ, 1 − δ] and c(t) = ( t − 1 + δ δ )b1 for all t ∈ (1 − δ, 1] Note that c ∈ A. Define gj(0) = 0 gj(t) = f (τt,j) − trj(b0,j) for all t ∈ (0, δ] gj(t) = f (τt,j) for all t ∈ (δ, 1 − δ) gj(t) = f (τt,j) − trj(b1,j) for all t ∈ [1 − δ, 1) and gj(1) = 0. (e 10.11) (e 10.12) (e 10.13) (e 10.14) (e 10.15) (e 10.16) (e 10.17) (e 10.18) (e 10.19) (e 10.20) One verifies that gj is lower semi-continuous on [0, 1]. It follows 10.3 that there exists a1 ∈ C([0, 1], F2)+ such that 0 ≤ gj(t) − dtrt,j (a1) ≤ 1/rj ≤ 1/k for all t ∈ [0, 1]. (e 10.21) Note that a1(0) = 0 and a1(1) = 0. Therefore a1 ∈ I ⊂ C. Now let a = c ⊕ a1 ∈ M2(C). Note that dτ (a) = dτ (c) + dτ (a1) = dτ (b) if t ∈ T, dtrt,j (a) = dtrt,j (c) + dtrt,j (a1) = dt,j(b0) + dtrt,j (a1) for all t ∈ (0, δ), dtrt,j (a) = dtrt,j (a1) for all t ∈ [δ, 1 − δ] and dtrt,j (a) = dtrt,j (c) + dtrt,j (a1) = dt,j(b1) + dtrt,j (a1) for all t ∈ (1 − δ, 1). Then, combining with (e 10.10), (e 10.13), (e 10.14), (e 10.15), (e 10.16), (e 10.17), (e 10.18), (e 10.19) and (e 10.21), we have 0 ≤ f (τ ) − dτ (a) ≤ 2/k for all τ ∈ T and τ = trt,j, j = 1, 2, ..., l, t ∈ (0, 1). (e 10.22) Since T ∪ {trt,j : 1 ≤ j ≤ l, and t ∈ (0, 1)} contains all extremal points of T (C), we conclude that 0 ≤ f (τ ) − dτ (a) ≤ 2/k for all τ ∈ T (C). (e 10.23) Theorem 10.5. Let A ∈ B1 be a unital simple C∗-algebra. Then the map W (A) → V (A) ⊔ LAff b(A)++ is surjective. 85 Proof. The proof follows the same lines of Theorem 5.2 of [9]. It suffices to show that the map a 7→ dτ (a) is surjective from W (A) onto LAff b(T (A)). Let f ∈ LAff b(A)+ with f (τ ) > 0 for all τ ∈ T (A). We may assume that f (τ ) ≤ 1 for all τ ∈ T (A). As in the proof of 5.2 of [9], it suffices to find a sequence of ai ∈ M2(A)+ such that ai . ai+1, [an] 6= [an+1] (in W (A)) and Using the semi-continuity of f, we find a sequence fn ∈ Aff(T (A))++ such that lim n→∞ dτ (an) = f (τ ) for all τ ∈ T (A). fn(τ ) < fn+1(τ ) for all τ ∈ T (A), n = 1, 2, .... fn(τ ) = f (τ ) for all τ ∈ T (A). lim n→∞ (e 10.24) (e 10.25) Since fn+1 − fn is continuous and strictly positive on the compact set T, there is εn > 0 such that (fn − fn+1)(τ ) > εn for all τ ∈ T (A), n = 1, 2, .... It follows from 10.2, for each n, there is a C∗-subalgebra Cn of A with Cn ∈ C and an element bn ∈ (Cn)+ such that (e 10.26) (e 10.27) (e 10.28) dimπ(Cn) ≥ (16/εn)2 for each irreducible representation π of Cn, 0 < τ (fn) − τ (bn) < εn/4 for all τ ∈ T (A). Applying 10.4, one obtains an element an ∈ M2(Cn)+ such that 0 < t(bn) − dt(an) < εn/4 for all t ∈ T (Cn). It follows that 0 < τ (fn) − dτ (an) < εn/2 for all τ ∈ T (A). (e 10.29) One then checks that limn→∞ dτ (an) = f (τ ) for all τ ∈ T (A). Moreover, dτ (an) < dτ (an+1) for all τ ∈ T (A), n = 1, 2, .... It follows from 9.11 that an . an+1, [an] 6= [an+1], n = 1, 2, .... This ends the proof. Theorem 10.6. Let A ∈ B1 be a unital simple C∗-algebra. Then W (A) has 0-almost divisible property. Proof. Let a ∈ Mn(A)+ \ {0} and k ≥ 1 be an integer. We need to show that there exists an element x ∈ Mm′(A)+ for some m′ ≥ 1 such that (e 10.30) in W (A). It follows from 10.5 that, since kdτ (a)/(k2 + 1) ∈ LAff b(T (A)), there is x ∈ M2n(A)+ such that k[x] ≤ [a] ≤ (k + 1)[x] Then, dτ (x) = kdτ (a)/(k2 + 1) for all τ ∈ T (A). It follows from 9.11 that kdτ (x) < dτ (a) < (k + 1)dτ (x) for all τ ∈ T (A). k[x] ≤ [a] ≤ (k + 1)[x]. (e 10.31) (e 10.32) (e 10.33) Theorem 10.7. Let A ∈ B1 be a unital separable simple amenable C∗-algebra. Then A⊗Z ∼= A. Proof. Since A ∈ B1, A has finite weak tracial nuclear dimension (see 8.1 of [73]). By 9.11, A has the strict comparison property for positive elements. Note, by 9.10, every unital hereditary C∗-subalgebra of A is in B1. Thus, by 10.6, its Cuntz semigroup also has 0-almost divisibility. It follows from 8.3 of [73] that A is Z-stable. 86 11 The unitary groups Theorem 11.1. (cf. Theorem 6.5 of [63]) Let K ∈ N be an integer and let B be a class of unital C∗-algebras which has the property that cer(B) ≤ K for all B ∈ B. Let A be a unital simple C∗-algebra which is tracially in B and let u ∈ U0(A). Then, for any ε > 0, there exists a unitary u1, u2 ∈ A such that u1 has exponential length no more than 2π, u2 has exponential rank K and Moreover, cer(A) ≤ K + 2 + ε. Proof. The proof is exactly the same as that of Theorem 6.5 of [63]. ku − u1u2k < ε. Corollary 11.2. Any C-algebra in the class B1 has exponential rank at most 5 + ǫ. Proof. By Theorem 3.16, C∗-algebras in C0 has exponential rank at most 3 + ǫ. Therefore, by Theorem 11.1, any C∗-algebra in B1 has exponential rank at most 5 + ǫ. Theorem 11.3. Let L > 0 be a positive number and let B be a class of unital C∗-algebras such that cel(v) ≤ L for every unitary v in their closure of commutator subgroups. Let A be a unital simple C∗-algebra which is tracially in B and let u ∈ CU (A). Then u ∈ U0(A) and cel(u) ≤ 3π + L. Proof. Let 1 > ε > 0. There are v1, v2, ...., vk ∈ U (A) such that (e 11.1) and vi = aibia∗i b∗i , where ai, bi ∈ U (A). Let N be an integer in Lemma 6.4 of [63] (for L = 8π +ε). Since A is tracially in B, there is a projection p ∈ A a unital C∗-subalgebra in B with 1B = p such that ku − v1v2 ··· vkk < ε/16 kai − (a′i ⊕ a′′i )k < ε/32k, kbi − (b′i ⊕ b′′i )k < ε/32k, i = 1, 2, ..., k kYi=1 (a′ib′i(a′i)∗(b′i)∗ ⊕ a′′i b′′i (a′′i )∗(b′′i )∗k < ε/8, ku − (e 11.2) (e 11.3) where a′i, b′i ∈ U ((1 − p)A(1 − p)), a′′i , b′′i ∈ U0(B) and 6N [1 − p] ≤ [p]. Put w = kYi=1 a′ib′i(a′i)∗(b′i)∗ and z = kYi=1 a′′i b′′i (a′′i )(b′′i )∗. (e 11.4) Then z ∈ CU (B). Therefore celB(z) ≤ L in B ⊂ pAp. It is standard to show that a′ib′i(a′i)∗(bi)∗ ⊕ (1 − p) ⊕ (1 − p) is in U0(M4((1 − p)A(−1p))) and it has exponential length no more than 4(2π) + 2ε/16k. This implies cel(w ⊕ (1 − p) ⊕ (1 − p)) ≤ 8πk + ε/4 in U (M4((1 − p)A(1 − p))). It follows from Lemma 6.4 of [63] that cel(w ⊕ p) ≤ 2π + ε/4. (e 11.5) It follows that cel((w ⊕ p)((1 − p) ⊕ z)) < 2π + ε/4 + L + ε/16. Therefore cel(u) ≤ 2π + L + ε. 87 Corollary 11.4. Let A ∈ B1, and let u ∈ CU (A). Then u ∈ U0 and cel(u) ≤ 7π. Proof. This follows from Lemma 3.12 and Theorem 11.3. Lemma 11.5. Let A be a unital C∗-algebra, let U be an infinite dimensional UHF-algebra and let B = A ⊗ U. Then U0(B)/CU (B) is torsion free and divisible. Proof. Since B = A ⊗ U , and U is a UHF-algebra with infinite type, for any projection p in a matrix algebra over B, there are projections e1, e2, ..., em ∈ B for some m such that p = e1 ⊕ e2 ⊕ ··· ⊕ em. Therefore ρB(K0(B)) is spanned by the image of the projections in B, and by Theorem 3.2 of [98], the group U0(B)/CU (B) is isomorphic to Aff(T (B))/ρB(K0(B)). Since Aff(T (B)) is divisible (it is a vector space), so is its quotient. We will show that Aff(T (B))/ρB(K0(B)) is torsion free. Suppose that a ∈ Aff(T (B)) so that na ∈ ρB(K0(B)) for some integer n > 1. Let ε > 0. There exists a pair of projections p, q ∈ Mn(B) such that na(τ ) − τ (p) − τ (q) < ε/2 for all τ ∈ T (B), regarded as unnormalized trace on Mn(B). Since B = A ⊗ U, there are mutually orthogonal projections p1, p2, ..., pn, pn+1, q1, q2, ..., qn, qn+1 such that p1 + p2 + ··· + pn + pn+1 = p and q1 + q2 + ··· + qn + qn+1 = q [p1] = [pj], [q1] = [qj], j = 1, 2, ..., n and τ (pn+1), τ (qn+1) < ε/2 for all τ ∈ T (B). (e 11.6) It follows that a(τ ) − (τ (p1) − τ (q1)) < ε for all τ ∈ T (B). This implies that a ∈ ρB(K0(B)). Therefore Aff(T (B))/ρB(K0(B)) is torsion free and the lemma follows. Theorem 11.6. Let A be a unital C∗-algebra such that there is an integer K > 0 such that cel(u) ≤ K for all u ∈ CU (A). Suppose that U0(A)/CU (A) is torsion free and suppose that u, v ∈ U (A) such that u∗v ∈ U0(A). Suppose also that there is k ∈ N such that cel((uk)∗vk) < L for some L > 0. Then Proof. It follows from [90] that, for any ε > 0, there are a1, a2, ..., aN ∈ As.a. such that cel(u∗v) ≤ K + L/k. (e 11.7) (e 11.8) (uk)∗vk = exp(√−1aj) and NYj=1 NXj=1 kajk ≤ L + ε/2. Choose w =QN to be torsion free, it follows that j=1 exp(−√−1aj/k). Then (u∗vw)k ∈ CU (A). Since U0(A)/CU (A) is assumed u∗vw ∈ CU (A). (e 11.9) Thus, cel(u∗vw) ≤ K. Note that cel(w) ≤ L/k + ε/2k, It follows that cel(u∗v) ≤ K + L/k + ε/2k. 88 Corollary 11.7. Let A be a unital simple C∗-algebra in B1, B = A ⊗ U, where U is a UHF- algebra of infinite type. Then (1) U0(B)/CU (B) is torsion free and divisible; and (2) if u, v ∈ U (B) with cel((u∗)kvk) ≤ L for some integer k > 0, then cel(u∗v) ≤ 7π + L/k. Proof. The lemma follows from Lemma 11.5, Corollary 11.4, and Theorem 11.6. Corollary 11.8. Let An be a sequence of unital separable simple C∗-algebras in B1 and let Bn = An ⊗ U. Then the kernel of the map K1(Yn Bn) → bYn K1(Bn) → 0 (e 11.10) is a divisible and torsion free. Proof. By Corollary 11.2, the exponential rank of each Bn is bounded by 6. Since each Bn has stable rank one, by (2) of Proposition 2.1 of [37], the kernel of the map in (e 11.10) is divisible. Cor 11.8 implies that {An} has exponential length divisible rank 7π + L/k (see Definition 2.1 of [37]). It follows Lemma 2.2 of [37] that the kernel is also torsion free. Lemma 11.9. Let K ≥ 1 be an integer. Let A be a unital simple C∗-algebra in B1. Let e ∈ A be a projection and let u ∈ U0(eAe). Suppose that w = u + (1− e) and suppose η ∈ (0, 2]. Suppose also that [1 − e] ≤ K[e] in K0(A) and dist( ¯w, ¯1) ≤ η. (e 11.11) Then, if η < 2, one has celeAe(u) < ( Kπ 2 + 1/16)η + 6π and dist(¯u, ¯e) < (K + 1/8)η, and if η = 2, one has celeAe(u) < Kπ 2 cel(w) + 1/16 + 6π. Proof. We assume that (e 11.11) holds. Note that η ≤ 2. Put L = cel(w). We first consider the case that η < 2. There is a projection e′ ∈ M2(A) such that [(1 − e) + e′] = K[e]. To simplify notation, by 9.10 and by replacing A by (1A + e′)M2(A)(1A + e′) and w by w + e′, without loss of generality, we may now assume that [1 − e] = K[e] and dist( ¯w, ¯1) < η. (e 11.12) η 32K(K+1)π > ε > 0 with ε + η < 2, since T R(A) ≤ 1, there exists a projection p ∈ A For any There is R1 > 1 such that max{L/R1, 2/R1, ηπ/R1} < min{η/64, 1/16π}. and a C∗-subalgebra D ∈ C with 1D = p such that (1) k[p, x]k < ε for x ∈ {u, w, e, (1 − e)}, 89 (2) pwp, pup, pep, p(1 − e)p ∈ε D, (3) there is a projection q ∈ D and a unitary z1 ∈ qDq such that kq−pepk < ε, kz1−quqk < ε, kz1 ⊕ (p − q) − pwpk < ε and kz1 ⊕ (p − q) − c1k < ε + η; (4) there is a projection q0 ∈ (1 − p)A(1 − p) and a unitary z0 ∈ q0Aq0 such that kq0−(1−p)e(1−p)k < ε, kz0−(1−p)u(1−p)k < ε, kz0⊕(1−p−q0)−(1−p)w(1−p)k < ε, kz0 ⊕ (1 − p − q0) − c0k < ε + η, (5) [p − q] = K[q] in K0(D), [(1 − p) − q0] = K[q0] in K0(A); (6) 2(K + 1)R1[1 − p] < [p] in K0(A); (7) cel(1−p)A(1−p)(z0 ⊕ (1 − p − q0)) ≤ L + ε, where c1 ∈ CU (D) and c0 ∈ CU ((1 − p)A(1 − p)). 2 arcsin( ε+η 2 ) such that (by (3) above) By Lemma 3.7, one has that detD(c1) = 1. Since ε + η < 2, there is h ∈ Ds.a with khk ≤ It follows that or It follows that (z1 ⊕ (p − q)) exp(ih) = c1. detD((z1 ⊕ (p − q)) exp(ih)) = 1, DD(z1 ⊕ (p − q) exp(ih))(t) = 0 for all t ∈ T (D). DD(z1 ⊕ (p − q))(t) ≤ 2 arcsin( ε + η 2 ) for all τ ∈ T (D). By (5) above, one obtains that DqDq(z1)(t) ≤ K2 arcsin( ε + η 2 ) for all t ∈ T (qDq). If 2K arcsin( ε+η 2 ) ≥ π, then 2K( ε+η 2 ) π 2 ≥ π. It follows that K(ε + η) ≥ 2 ≥ dist(z1, q). (e 11.13) (e 11.14) (e 11.15) (e 11.16) (e 11.17) (e 11.18) Since those unitaries in D with det(u) = 1 (for all points) are in CU (D) (see Lemma 3.7), from (e 11.17), one computes that, when 2K arcsin( ε+η 2 ) < π, dist(z1, q) < 2 sin(K arcsin( ε + η 2 )) ≤ K(ε + η). By combining both (e 11.18) and (e 11.19), one obtains that dist(z1, q) ≤ K(ε + η) ≤ Kη + η 32(K + 1)π . By (e 11.17), it follows from Lemma 3.12 that (e 11.19) (e 11.20) celqDq(z1) ≤ 2K arcsin ε + η 2 + 4π ≤ K(ε + η) π 2 + 4π ≤ (K π 2 + 1 64(K + 1) )η + 4π. (e 11.21) 90 By (5) and (6) above, (K + 1)[q] = [p − q] + [q] = [p] > 2(K + 1)R1[1 − p]. Since K0(A) is weakly unperforated, one has 2R1[1 − p] < [q]. (e 11.22) There is a unitary v ∈ A such that v∗(1 − p − q0)v ≤ q. Put v1 = q0 ⊕ (1 − p − q0)v. Then (e 11.23) v∗1(z0 ⊕ (1 − p − q0))v1 = z0 ⊕ v∗(1 − p − q0)v. Note that k(z0 ⊕ v∗(1 − p − q0)v)v∗1c∗0v1 − q0 ⊕ v∗(1 − p − q0)vk < ε + η. (e 11.24) Moreover, by (7) above, It follows from (e 11.22) and Lemma 6.4 of [63] that cel(z0 ⊕ v∗(1 − p − q0)v) ≤ L + ε, cel(q0+q)A(q0+q)(z0 ⊕ q) ≤ 2π + (L + ε)/R1. Therefore, combining (e 11.21), (e 11.25) (e 11.26) cel(q0+q)A(q0+q)(z0 + z1) ≤ 2π + (L + ε)/R1 + (K π 2 + 1 64(K + 1) )η + 6π. (e 11.27) By (e 11.25), (e 11.22) and Lemma 3.1 of [67], in U0((q0 + q)A(q0 + q))/CU ((q0 + q)A(q0 + q)), dist(z0 + q, q0 + q) < (L + ε) R1 . (e 11.28) Therefore, by (e 11.19) and (e 11.28), dist(z0 ⊕ z1, q0 + q) < (L + ε) R1 + Kη + η 32(K + 1)π < (K + 1/16)η. (e 11.29) We note that It follows that ke − (q0 + q)k < 2ε and ku − (z0 + z1)k < 2ε. (e 11.30) dist(¯u, ¯e) < 4ε + (K + 1/16)η < (K + 1/8)η. (e 11.31) Similarly, by (e 11.27), celeAe(u) ≤ 4επ + 2π + (L + ε)/R1 + (K π 2 + 1 64(K + 1) )η + 4π (e 11.32) < (K π 2 + 1/16)η + 6π. (e 11.33) This proves the case that η < 2. 91 Now suppose that η = 2. Define R = [cel(w) + 1]. Note that cel(w) R < 1. There is a projection e′ ∈ MR+1(A) such that It follows from Lemma 3.1 of [67] that [(1 − e) + e′] = (K + RK)[e]. dist(w ⊕ e′, 1A + e′) < cel(w) R + 1 . (e 11.34) Put K1 = K(R + 1). To simplify notation, without loss of generality, we may now assume that [1 − e] = K1[e] and dist( ¯w, ¯1) < cel(w) R + 1 . It follows from the first part of the lemma that celeAe(u) < ( K1π 2 + 1 16 ) cel(w) R + 1 + 6π ≤ Kπcel(w) 2 + 1 16 + 6π. (e 11.35) (e 11.36) Theorem 11.10. Let A be a unital simple C∗-algebra of stable rank one and let e ∈ A be a non- zero projection. Then the map u 7→ u+(1−e) induces an isomorphism j from U (eAe)/CU (eAe) onto U (A)/CU (A). Proof. This was originally proved here following 11.9 (with additional assumption that A ∈ B1). However, by 3.3, A has stable rank one. Thus this follows from Theorem 4.6 of [38]. Corollary 11.11. Let A be a unital simple C∗-algebra of stable rank one. Then the map z m } { 1, 1, ..., 1) from A to Mn(A) induces an isomorphism from U (A)/CU (A) onto j : a → diag(a, U (Mn(A))/CU (Mn(A)) for any integer n ≥ 1. Proof. This follows from 11.10 but also follows from 3.11 of [38]. 12 A Uniqueness Theorem for C∗-algebras in B0 The following follows from Theorem 7.1 of [67]. Theorem 12.1. (cf. Theorem Theorem 7.1 of [67]) Let A be a unital separable amenable C∗- algebra which satisfies the UCT, let T × N : A+ \ {0} → R+ \ {0} × N be a map and let L : U (M∞(A)) → R+ be another map. For any ε > 0 and any finite subset F ⊂ A, there exists δ > 0, a finite subset G ⊂ A, a finite subset H ⊂ A+ \ {0}, a finite subset U ⊂ ∪m=1U (Mm(A)), a finite subset P ⊂ K(A) and an integer n > 0 satisfying the following: for any unital separable simple C∗-algebra C in B1, if ϕ, ψ, σ : A → B = C ⊗ U, where U is a UHF-algebra of infinite type, are three G-δ-multiplicative contractive completely positive linear maps and σ is unital and T × N -H-full (see 2.20) with the properties that [ϕ]P = [ψ]P and cel(hψ(u)i∗hϕ(v)i) ≤ L(v) (e 12.1) for all v ∈ U and σ is unital, there exists a unitary u ∈ Mn+1(B) such that ku∗diag(ϕ(a), σ(a))u − diag(ψ(a), σ(a))k < ε for all a ∈ F, where σ(a) = diag( σ(a), σ(a), ..., σ(a)) for all a ∈ A. z n } 92 { Proof. Note that B has stable rank one, cer(Mm(B)) ≤ 6 (see 11.2), exponential length divisible rank E(L, k) = 7π + L/k ((2) of 11.7) and K0(B) are weakly unperforated (in particular, it has K0-divisible rank T (L, k) = L + 1). The class of unital C∗-algebras have just mentioned properties will be denoted by C. We sketch the proof below. Suppose that the theorem is false. Then there exists ε0 > 0 and a finite subset F ⊂ A such that there are a sequence of positive numbers {δn} with δn ց 0, an increasing sequence {Gn} ⊂ A of finite subsets such that ∪nGn is dense in A, an increasing sequence {Pn} ⊂ K(A) of finite subsets with ∪n=1Pn = K(A), an increasing sequence of finite subsets {Un} ⊂ U (M∞(A)) such that ∪n=1Un∩U (Mm(A)) is dense in U (Mm(A)) for each integer m ≥ 1, an increasing sequence of finite subsets {Hn} ⊂ A1 +\{0} such that if a ∈ Hn and f1/2(a) 6= 0, then f1/2(a) ∈ Hn+1 and ∪n=1Hn is dense in A1 +, a sequences of integers {k(n)} with limn→∞ k(n) = ∞, a sequence of unital C∗-algebras Bn ∈ C, two sequences of Gn-δn-multiplicative contractive completely positive linear maps ϕn, ψn : A → Bn, such that (e 12.2) [ϕn]Pn = [ψn]Pn and cel(hϕn(u)ihψn(u∗)i) ≤ L(u) for all u ∈ Un and a sequence of unital Gn-δn-multiplicative contractive completely positive linear map σn : A → Bn which is also F -Hn-full satisfying inf{supkv∗ndiag(ϕn(a), Sn(a))vn − diag(ψn(a), Sn(a))k : a ∈ Fk} ≥ ε0, (e 12.3) where the infimum is taken among all unitaries vn ∈ Mk(n)l(n)+1(Bn) and where Sn(a) = diag( σn(a), σn(a), ..., σn(a)) for all a ∈ A. Let C0 =L∞n=1 Bn, C =Q∞n=1 Bn, Q(C) = C/C0 and π : C → Q(C) is the quotient map. Define Φ, Ψ, S : A → C by Φ(a) = {ϕn(a)}, Ψ(a) = {ψn(a)} and S(a) = {σn(a)} all a ∈ A. Note that π ◦ Φ, π ◦ Ψ and π ◦ S are homomorphisms. As in the proof of 7.1 of [67], since Bn ∈ C, one computes that [π ◦ Φ] = [π ◦ Ψ] in KL(A, Q(C)). (e 12.4) One can also check that π ◦ S is a full homomorphism. It follows from 4.13 that there exists an integer K ≥ 1 and a unitary U ∈ MK+1(Q(C)) such that kU∗diag(π ◦ Φ(a), Σ(a))U − diag(π ◦ Ψ(a))k < ε0/4 for all a ∈ F. (e 12.5) It follows that there exists a V = {vn} ∈ C and an integer N ≥ 1 such that, for any n ≥ N, k(n) ≥ K such that kv∗ndiag(ϕn(a), σn(a))vn − diag(ψn(a), σn(a))k < ε0/2 (e 12.6) for all a ∈ F, where σn(a) = diag( σn(a), σn(a), ..., σn(a)) for all a ∈ A. This contradicts with (e 12.3). Remark 12.2. (1) The assumption that B = C ⊗ U where C ∈ B1 and U is an infinite UHF- algebra is just for our purpose in this paper. The same theorem holds for a much more general setting. Let R, r : N → N, T : N2 → N, E : R+ × N → R+ be fixed maps. Let CR,r,T,E be the class of unital C∗-algebras which have Ki-r-cancellation, Ki-divisible rank T and exponential 93 z z k(n) } K } { { length divisible length E, and cer(Mm(B)) ≤ R for all B ∈ CR,r,T,E. Then Theorem 12.1 holds for B ∈ CR,r,T,E just as 7.1 of [67]. One should also note that the proof is not new. It is just the same as the combination of those of Theorem 5.9 of [61] and Theorem 4.8 of [56]. One should also note that ϕ and ψ are not assumed to be unital. Thus Theorem 4.14 can also viewed as a special case of Theorem 12.1. (2) Suppose that there exists an integer n0 ≥ 1 such that U (Mn0(A))/U0(Mn0(A)) → U (Mn0+k(A))/U0(Mn0+k(A)) is an isomorphism for all k ≥ 1. Then L may be replaced by a map from U (Mn0(A)) to R+, and U can be chosen in U (Mn0(A)). Moreover, the condition that cel(hϕ(u)ihψ(u)∗) ≤ L(u) may, in practice, be replaced by a stronger condition that, for all u ∈ U , dist(ϕ‡(u), ψ‡(u)) < L (e 12.7) where U ⊂ ∪∞m=1U (Mm(A))/CU (Mm(A)) is a finite subset and where L < 2 is a given constant. To see this, let U be a finite subset of U (Mm(A)) for some large m whose image in the group U (Mm(A))/CU (Mm(A)) is U . Then (e 12.7) implies that khϕ(u)ihψ(u∗)i − vk < 2 (e 12.8) for some v ∈ CU (Mm(A)), provided that δ is sufficiently small and G is sufficiently large. Since cel(v) ≤ 7π, we conclude that (e 12.9) for all u ∈ U ,and take L : U∞(A) → R+ to be constant 8π + 1. Furthermore, we may assume cel(hϕ(u)ihψ(u∗)i) ≤ π + 7π + 1 U ⊂ U (Mn0(A))/CU (Mn0 (C)), if K1(C) = U (Mn0(C))/U0(Mn0(C)). Lemma 12.3. Let A be a unital separable simple C∗-algebra with T (A) 6= ∅. There exists a map ∆0 : Aq,1 + \ {0}, there exits δ > 0 and a finite subset G ⊂ A such that, for any unital C∗-algebra B with T (B) 6= ∅ and any unital G-δ-multiplicative contractive completely positive linear map ϕ : A → B, one has + \ {0} → (0, 1) satisfying the following: For any finite subset H ⊂ A1 (e 12.10) (e 12.11) (e 12.12) τ ◦ ϕ(h) ≥ ∆0(h)/2 for all h ∈ H for all τ ∈ T (B). Moreover, one may assume that ∆0(c1A) = 3/4. ∆0(h) = min{3/4, inf{τ (h) : τ ∈ T (A)}}. Proof. Define, for each h ∈ A1 + \ {0}, Let H ⊂ A1 + \ {0} be a finite subset. Define d = min{∆0(h)/4 : h ∈ H} > 0. Let δ > 0 and let G ⊂ A be a finite subset as required by 5.8 for ε = d and F = H. Suppose that ϕ : A → B is a unital G-δ-multiplicative contractive completely positive linear map. Then, by 5.8, for each t ∈ T (B), there exists τ ∈ T (A) such that t ◦ ϕ(h) − τ (h) < d for all h ∈ H. (e 12.13) It follows that t ◦ ϕ(h) > τ (h) − d for all h ∈ H. Thus t ◦ ϕ(h) > ∆0(h) − d > ∆0(h)/2 for all h ∈ H and for all t ∈ T (B). (e 12.14) 94 Lemma 12.4. Let C be a unital C-algebra, and let ∆ : C q,1 + \{0} → (0, 1) be an order preserving map. There exists a map T × N : C+ \ {0} → R+ \ {0} × N satisfying the following: For any finite subset H ⊂ C 1 + \ {0} and any unital C-algebra A with the strict comparison of positive elements, if ϕ : C → A is a unital contractive completely positive linear map satisfying τ ◦ ϕ(h) ≥ ∆(h) for all h ∈ H for all τ ∈ T(A), (e 12.15) then ϕ is (T × N )-H-full. Proof. For each δ ∈ (0, 1), let gδ : [0, 1] → [0, +∞) be the continuous function defined by gδ(t) =( 0 if t ∈ [0, δ 4 ], otherwise, fδ/2(t) t where fδ/2 is as defined in 2.5. Note that gδ(t)t = fδ/2(t) for all t ∈ [0, 1]. (e 12.16) Let h ∈ C 1 + \ {0}. Then define T (h) = k(g∆(bh)) 1 2k= 2 δ and N (h) = ⌈ 2 ∆(bh)⌉. Indeed, let H ⊆ C 1 Then the function T × N satisfies the lemma. + \ {0} be a finite subset. Let A be a unital C∗-algebra with the strict comparison for positive elements, and let ϕ : C → A be a unital positive linear map satisfying (e 12.17) Put ϕ(h)− = (ϕ(h) − ∆(h) τ ◦ ϕ(h) ≥ ∆(h) for all h ∈ H for all τ ∈ T(A). 2 )+. Then, by (e 12.17), one has that, since 0 ≤ h ≤ 1, dτ (ϕ(h)−) = dτ ((ϕ(h) − ∆(h) 2 )+) ≥ τ ((ϕ(h) − ∆(h) 2 )+) ≥ ∆(h) 2 for all τ ∈ T(A). The above also shows that ϕ(h)− 6= 0. Since A has the strict comparison for positive elements, one has K h(ϕ(h)−i > h1Ai , where K = ⌈ 2 ∆(h)⌉ and where hxi denotes the class of x in W (A). Therefore there is a partial isometry v = (vij)K×K ∈ MK(A) such that vv∗ = 1A and v∗v ∈ (ϕ(h)− ⊗ 1K )MK (A)(ϕ(h)− ⊗ 1K). Note that c(fδ/2(ϕ(h))⊗1K ) = c(fδ/2(ϕ(h))⊗1K ) = c for all c ∈ (ϕ(h)− ⊗ 1K )MK (A)(ϕ(h)− ⊗ 1K ), where δ = ∆(h), and therefore Consider the upper-left corner of MK (A), one has that v(fδ/2(ϕ(h)) ⊗ 1K)v∗ = vv∗ = 1A. v1,ifδ/2(ϕ(h))v∗1,i = 1A, KXi=1 and therefore, by (e 12.16), one has v1,i(g∆( f )(ϕ(h))) 1 2 ϕ(h)(g∆( f )(ϕ(h))) 1 2 v∗1,i = 1A. KXi=1 95 Since v is a partial isometry, one has that kvi,jk ≤ 1, i, j = 1, ..., K, and therefore kv1,i(g∆( f )(ϕ(f ))) 1 2k ≤ k(g∆( f )(ϕ(f ))) 1 2k ≤ k(g∆( f )) 1 2k = T (f ). Hence the map ϕ is T × N -H-full, as desired. Theorem 12.5. Let C be a unital C∗-algebra in Ds (see 4.8). Let F ⊂ C be a finite subset, let ε > 0 be a positive number and let ∆ : C q,1 + \ {0} → (0, 1) be an order preserving map. There exists a finite subset H1 ⊂ C 1 + \ {0}, there exists γ1 > 0, 1 > γ2 > 0, δ > 0, a finite subset G ⊂ C and a finite subset P ⊂ K(C), a finite subset H2 ⊂ Cs.a. and a finite subset U ⊂ ∪∞m=1U (Mm(C))/CU (Mm(C)) for which [U ] ⊂ P satisfying the following: For any unital G-δ-multiplicative contractive completely positive linear maps ϕ, ψ : C → A, where A = A1 ⊗ U for some A1 ∈ B1 and a UHF-algebra U of infinite type satisfying [ϕ]P = [ψ]P , τ (ϕ(a)) ≥ ∆(a), τ (ψ(a)) ≥ ∆(a) for all τ ∈ T (A) and for all a ∈ H1, τ ◦ ϕ(a) − τ ◦ ψ(a) < γ1 for all a ∈ H2 and dist(ϕ‡(u), ψ‡(u)) < γ2 for all u ∈ U , there exists a unitary W ∈ A such that kW ∗ϕ(f )W − ψ(f )k < ε for all f ∈ F. (e 12.18) (e 12.19) (e 12.20) (e 12.21) (e 12.22) Let H1,1 ⊆ C 1 Proof. Let T ′ × N : C+ \ {0} → R+{0} × N be the map of Lemma 12.4 with respect to C and ∆/4. Let T = 2T ′. Define L = 1. Let δ0 > 0 (in place of δ), G0 ⊆ C (in place of G), H0 ⊆ C+ \ {0} (in place of H), U0 ⊆ U (Mn0(C))/CU (Mn0 (C)) (in place of U ), P0 ⊆ K(C) (in place of P) be finite subsets and n1 (in place of n) be an integer as required by Theorem 12.1 with respect to C (in place of A), T × N , L, F and ǫ/2 —-(see (2) of 12.2). + \ {0} (in place of H1), H1,2 ⊆ A (in place of H2), γ1,1 > 0 (in place of γ1), γ1,2 > 0 (in place of γ2), δ1 > 0 (in place of δ), G1 ⊆ C (in place of G), P1 ⊆ K(C) (in place of P), U1 ⊆ Jc(K1(C)) (in place of U ) and n2 (in place of N ) be the finite subsets and constants of Theorem 8.4 with respect to C (in place of A), ∆/4, F and ǫ/4. Put G = G0 ∪ G1, δ = min{δ0/4, δ1/4}, P = P0 ∪ P1, H1 = H1,1, H2 = H1,2, U = U0 ∪ U1, γ1 = γ1,1/2, γ2 = γ1,2/2. One asserts these are desired finite subsets and constants (for F and ǫ). We may assume that γ2 < 1/4. In fact, let A = A1 ⊗ U , where A ∈ B1 and U is a UHF-algebra of infinite type. Let ϕ, ψ : C → A be G-δ-multiplicative maps satisfying (e 12.18) to (e 12.21) for the above chosen G, H1, P, H2, U , γ1 and γ2. Since A = A1 ⊗ U , A ∼= A ⊗ U. Moreover, j ◦ ı : A → A is approximately inner, where ı : A → A ⊗ U is defined by a 7→ a ⊗ 1U and j : A ⊗ U → A is an isomorphism. Thus, we may assume that A = A1⊗ U ⊗ U = A2⊗ U, where A2 = A1⊗ U. Moreover, without loss of generality, we may assume that the images of ϕ and ψ are in A2. Since A2 ∈ B1, for a finite subset G′′ ⊆ A2 and δ′ > 0, there is a projection p ∈ A2, a C-subalgebra D ∈ C1 with p = 1D such that (1) kpg − gpk < δ′ for any g ∈ G′′, (2) pgp ∈δ′ D, 96 (3) τ (1 − p) < min{δ′, γ1/4, 1/8m} for any τ ∈ T(A). Define j0 : A2 → (1 − p)A2(1 − p) by j0(a) = (1 − p)a(1 − p) for all a ∈ A2. For any ε′′ > 0 and any finite subset F′′ ⊂ A2, there is also a unital contractive completely positive linear map j1 : A → D such that kj1(a) − papk < ε′′ for all a ∈ F′′, provided that G′′ is sufficiently large and δ′ is sufficiently small. Therefore, in particular, we may assume that kϕ(c) − (j0 ◦ ϕ(c) ⊕ j1 ◦ ϕ(c))k < ε/16 and kψ(c) − (j0 ◦ ψ(c) ⊕ j1 ◦ ψ(c))k < ε/16 (e 12.23) (e 12.24) for all c ∈ F. Choose an integer m ≥ 2(n1 + 1)n2 and mutually orthogonal and mutually equivalent pro- i=1 ei = 1U . Define ϕ′i, ψ′i : C → A ⊗ U by ϕ′i(c) = ϕ(c) ⊗ ei and jections e1, e2..., em ∈ U withPm ψ′i(c) = ψ(c) ⊗ ei for all c ∈ C, i = 1, 2, ..., m. Note that [ϕ′1]P = [ϕ′i]P = [ψ′1]P = [ψ′i]P , (e 12.25) i = 1, 2, ..., m. Note also that ϕ′i, ψ′i : C → eiAei are G-δ-multiplicative. Write m = kn2 + r, where k ≥ n1 + 1 and r < n2 are integers. Define ϕ, ψ : C → (1 − p)A2(1 − p) ⊕Lm i=kn2+1 A2 ⊗ ei by ϕ(c) = j0 ◦ ϕ(c) ⊕ ψ(c) = j0 ◦ ψ(c) ⊕ mXi=kn2+1 mXi=kn2+1 j1 ◦ ϕ(c) ⊗ ei and j1 ◦ ψ(c) ⊗ ei (e 12.26) (e 12.27) for all c ∈ C. With sufficiently large G′′ and small δ′, we may assume that ϕ and ψ are G-2δ- multiplicative and, by (e 12.25), Moreover, by 9.12, we may further assume that [ ϕ]P = [ ψ]P . dist( ϕ‡(¯v), ψ‡(¯v)) < γ2 ≤ L (e 12.28) (e 12.29) i , ψ1 for all ¯v ∈ U . Define ϕ1 i (c) = j1 ◦ ϕ(c) ⊗ ei and ψ1 By 9.12 and by choosing even larger G′′ and smaller δ′, we may assume that i (h)) ≥ ∆(h)/2 for all h ∈ H1 i (h) ≥ ∆(h)/2 and τ (ψ1 i : C → D ⊗ ei by ϕ1 τ ◦ ϕ1 i (c) = j1 ◦ ψ(c) ⊗ ei. and for all τ ∈ T (pAp ⊗ ei), t ◦ ϕ1 i (c) − t ◦ ψ1 i (c) < γ1,1 for all c ∈ H2 and dist((ϕ1 i )‡(¯v), (ψ1 i )‡(¯v)) < γ1,2 for all ¯v ∈ U . By applying 12.4, ϕ1 are G-2δ-multiplicative contractive completely positive linear maps and i are T × N -H1-full. Moreover, we may also assume that ϕ1 i and ψ1 i and ψ1 i [ϕ1 i ]P = [ψ1 i ]P . 97 (e 12.33) (e 12.30) (e 12.31) (e 12.32) Define Φ, Ψ : C →Lkn2 i=1 D ⊗ ei by Φ(c) = ϕ1 i (c) and Ψ(c) = kn2Mi=1 ψ1 i (c) kn2Mi=1 (e 12.34) for all c ∈ C. By (e 12.33), (e 12.30), (e 12.31), (e 12.32) and by 8.4, there exits a unitary W1 ∈ (e 12.35) (e 12.36) (Pkn2 i=1 p ⊗ ei)(A2 ⊗ U )(Pkn2 i=1 p ⊗ ei) such that kW ∗1 Φ(c)W1 − Ψ(c)k < ε/4 for all c ∈ F. Note that τ (1 − p) + mXkn2+1 τ (ei) < (1/m) + (r/m) ≤ n2/m for all τ ∈ T (A). Note also that k ≥ n1. By (e 12.28), (e 12.29), since ψ1 multiplicative, by applying 12.1, there exists a unitary W2 ∈ A such that i is T × N -H1- kW ∗2 ( ϕ(c) ⊕ Ψ(c))W1 − ( ψ(c) ⊕ Ψ(c))k < ε/2 for all c ∈ F. (e 12.37) Set Then we compute that W = (diag(1 − p, ekn2+1, ekn2+2, ..., em) ⊕ W1)W2. kW ∗( ϕ(c) ⊕ Φ(c))W − ( ψ(c) ⊕ Ψ(c))k < ε/2 + ε/4 for all c ∈ F. By (e 12.23), we have kW ∗ϕ(c)W − ψ(c)k < ε for all c ∈ F, as desired. (e 12.38) (e 12.39) Remark 12.6. First we note that, given (e 12.20) and with an arrangement at the beginning of the proof, we only need one inequality in (e 12.19). Note also that the condition that U ⊂ ∪∞m=1U (Mm(C))/CU (Mm(C)) can be replaced by U ⊂ Jc(K1(C)). To see this, we note that we may write, by 2.15, ∞[m=1 U (Mm(C))/CU (Mm(C)) = Aff(T (C))/ρC (K0(C)) ⊕ Jc(K1(C)). (e 12.40) So, without loss of generality, we may assume that U = U0∪U1, where U0 ⊂ Aff(T (C))/ρC (K0(C)) and U1 ⊂ Jc(K1(C)). Using the de La Harpe and Skandalis determinant (see again 2.15), with sufficiently small δ, γ1 and sufficiently large G and H1, the condition (e 12.20) implies that dist(ϕ‡(u), ψ‡(u)) < γ2 for all u ∈ U0. (e 12.41) This particularly implies that, when K1(C) = 0, U and γ2 are not needed in the statement of 12.5. When K1(C) is torsion, ϕ‡Jc(K1(C)) = 0, since Aff(T (A))/ρA(ρA(K0(A)) is torsion free (see 11.5). In general, we can further assume that U ⊂ Jc(K1(C)) generates a free group. Let G(U ) be the subgroup generated by U . Write G(U ) = G1 ⊕ Gt, where G0 is free and Gt is torsion. Let 98 g1, g2, ..., gk1 be the generators of G1 and f1, f2, ..., fk2 be the generators of Gt. As in the proof, we may assume that [x] ∈ P for all x ∈ U′, where U′ is a finite subset of unitaries such that [U′] = U . By choosing even larger P and U , we may assume that gi, fj ∈ U . There is an integer m(j) ≥ 1 such that m(j)fj = 0. It follows that m(j)ϕ‡(fj) = m(j)ψ‡(fj) = 0. With sufficiently small δ, γ1 and sufficiently large G, condition (e 12.18) implies that π ◦ ψ‡(fj) = π ◦ ϕ‡(fj), where π :S∞m=1 U (Mm(A))/CU (Mm(A)) → K1(A) is the quotient map. Therefore Note m(j)Jc ◦ π ◦ ϕ‡(fj) = 0. It follows that Jc ◦ π ◦ ψ‡(fj) = Jc ◦ π ◦ ϕ‡(fj). (e 12.42) m(j)(ϕ‡(fj) − Jc ◦ π ◦ ϕ‡(fj)) = 0. However, ϕ‡(fj) − Jc ◦ π ◦ ψ‡(fj) ∈ Aff(T (A))/ρA(K0(A)). Since A1 ∈ B1, by 11.7, Aff(T (A))/ρA(K0(A)) is torsion free. Therefore Similarly, ψ‡(fj) − Jc ◦ π ◦ ψ‡(fj) = 0. Thus, by (e 12.42), ϕ‡(fj) − Jc ◦ π ◦ ϕ‡(fj) = 0 ϕ‡(fj) = ψ‡(fj). (e 12.43) In other words, with sufficiently small δ, γ1 and sufficiently large G, (e 12.18) implies (e 12.43). So Gt can be dropped. Therefore in the statement in 12.5, we may assume that U generates a free subgroup of Jc(K1(C)). Furthermore, if C has stable rank k, then Jc(K1(C)) may be replaced by Jc(U (Mk(C))/U0(Mk(C))), In the case that C has stable rank one, then U may be assumed to be a sub- see (e 2.23). set of Jc(U (C)/U0(C)). In the case that C = C′ ⊗ C(T) for some C′ with stable rank one, then stable rank of C is no more than 2. Therefore, in this case, U may be assumed to be in Jc(U (M2(C))/U0(M2(C)), or U (M2(C))/CU (M2(C)). The introduction of ∆ and condition (e 12.19) are for convenience which can be replaced by original fullness condition. ∆ can be replaced by a map T × N : C+ \{0} and condition (e 12.19) is replaced by ϕ and ψ are T × N -H1-full as indicated in the proof. Corollary 12.7. Let ε > 0 be a positive number and let ∆ : C(T)q,1 + \ {0} → (0, 1) be an order preserving map. There exists a finite subset H1 ⊂ C(T)1 + \ {0}, there exists γ1 > 0, 1 > γ2 > 0 and any finite subset H2 ⊂ C(T)s.a. satisfying the following: For any two unitaries u1 and u2 in a unital separable simple C∗-algebra A ∈ B1 such that [u1] = [u2] ∈ K1(A) and τ (f (u1)), τ (f (u2)) ≥ ∆( f ) (e 12.44) for all τ ∈ T (C) and for all f ∈ H1, τ (g(u1)) − τ (g)(u2) < γ1 for all g ∈ H2 and dist( ¯u1, ¯u2) < γ2, (e 12.45) there exists a unitary W ∈ C such that kW ∗u1W − u2k < ε. 99 (e 12.46) Lemma 12.8. Let A be a C∗-algebra and X be a compact metric space. Suppose that y ∈ (A ⊗ C(X))+ \ {0}. Then exists a(y) ∈ A+ \ {0}, f (y) ∈ C(X)+ \ {0} and ry ∈ A ⊗ C(X) such that ka(y)k ≤ kyk, kf (y)k ≤ 1, kryk ≤ kak and r∗yyry = a(y) ⊗ f (y). Proof. Identify A ⊗ C(X) with C(X, A). Let x0 ∈ X such that ky(x0)k = kyk. There is δ > 0 such that ky(x)− y(x0)k < kyk/16 for x ∈ B(x0, 2δ). Let Y = B(x0, δ). Let z(x) = fkyk/4(y(x0)) for all x ∈ Y. Note z 6= 0. By 2.2 of [92], there exist r ∈ C(Y, A) with krk ≤ kyk such that r∗(yY )r = z. Choose g ∈ C(X)+ \ {0} such that 0 ≤ g ≤ 1, g(x) = 0 if dist(x, x0) ≥ δ and g(x) = 1 if dist(x, x0) ≤ δ/2. Then one may view rg1/2, zg ∈ C(X, A). Put ry = rg1/2, a(y) = fkyk/4(y(x0)) and f (y) = g. Then r∗yyry = zg = a(y) ⊗ f. Theorem 12.9. Let A1 ∈ B1 be a unital simple C∗-algebra which satisfies the UCT, A = A1 ⊗ C(X), where X be a point or X = T. For any ε > 0, any finite subset F ⊂ A and any order preserving map ∆ : C(X)1 + \ {0} → (0, 1), there exists δ > 0, a finite subset G ⊂ A, σ1, σ2 > 0, a finite subset P ⊂ K(A), a finite subset H1 ⊂ C(X)1 + \ {0}, a finite subset U ⊂ U (M2(A))/CU (M2(A)) (—see 12.10) and a finite subset H2 ∈ As.a satisfying the following: Let B′ ∈ B1, let B = B′ ⊗ U for some UHF-algebra U of infinite type and let ϕ, ψ : A → B be two unital G-δ-multiplicative contractive completely positive linear maps such that [ϕ]P = [ψ]P , τ ◦ ϕ(1 ⊗ h) ≥ ∆(h) for all h ∈ H1 and τ ∈ T (B), τ ◦ ϕ(a) − τ ◦ ψ(a) < σ1 for all a ∈ H2 and dist(ϕ‡(¯u), ψ‡(¯u)) < σ2 for all ¯u ∈ U . Then there exists a unitary u ∈ U (B) such that kAd u ◦ ϕ(f ) − ψ(f )k < ε for all f ∈ F. (e 12.47) (e 12.48) (e 12.49) (e 12.50) (e 12.51) Proof. Let ε > 0 and let F ⊂ A be a finite subset. Without loss of generality, we may assume that F = {a ⊗ f : a ∈ F1 and f ∈ F2}, Let L = 1. Let ∆ : C(X)1 where F1 ⊂ A is a finite subset and F2 ⊂ C(X) is also a finite subset. We further assume that F1 and F2 are in the unit ball of A and C(X), respectively. + \ {0} → (0, 1) be an order preserving map. Let T ′ × N′ : C(X)+ \ {0} → R+ \ {0} × N be a map given by 12.4 with respect to 3∆/16. Since A1 is a unital separable simple C∗-algebra, the identity map on A1 is T ′′ × N′′-full for some T ′′ × N′′ : (A1)+ \ {0} → R+ \ {0} × N. Define a map T × N : A+ \ {0} → R+ \ {0} × N as follows: For any y ∈ A+ \ {0}, by 12.8, there exist a(y) ∈ (A1)+ \ {0}, f (y) ∈ C(X)+ \ {0} and ry ∈ A ⊗ C(X) with kryk ≤ kyk such that r∗yyry = a(y) ⊗ f (y), ka(y)k ≤ kyk and kf (y)k ≤ 1. There are xa(y),1, xa(y),2, ..., xa(y),N ′′ (a(y)) ∈ A1 with max{kxa(y),ik : 1 ≤ i ≤ N′′(a(y))} = T ′′(a(y)) such that Then define N ′′(a(y))Xi=1 x∗a(y),ia(y)xa(y),i = 1A1. (T × N )(y) = (1 + max{T ′′(a(y)), T ′(f (y))} · max{1,kyk}, N′′(a(y)) · N′(f (y))). (e 12.52) 100 Let ε/16 > δ1 > 0 (in place of δ), G1 ⊂ A (in place of G), H0 ⊂ A+ \ {0} (in place of H), U1 ⊂ U (M2(A))/CU (M2(A)) (in place of U —see (2) of 12.2), P1 ⊂ K(A) (in place of P) and n ≥ 1 be the finite subsets and constants as required by 12.1 for A, L, ε/16 (in place of ε), F and T × N. Without loss of generality, we may assume that δ1 < ε and G1 = {a ⊗ g : a ∈ G′1 and g ∈ G′′1}, (e 12.53) where G′1 ⊂ A1 and G′′1 ⊂ C(X) are finite subsets. We may further assume that F1 ⊂ G′1 and F′2 ⊂ G′′1 , and both are in the unit ball. In particular, F ⊂ G1. We may also assume that H0 = {a ⊗ f : a ∈ H′0 and f ∈ H′′0}, (e 12.54) where H′0 ⊂ (A1)+ \ {0} and H′′0 ⊂ C(X)+ \ {0} be finite subsets. For convenience, we may further assume that, for the above integer n ≥ 1, 1/n < inf{∆(h) : h ∈ H′′0}/16. (e 12.55) Let U1 = { ¯v1, ¯v2, ..., ¯vK}, where v1, v2, ..., vK ∈ U (M2(A)). Put U0 = {v1, v2, ..., vK}. Choose a finite subset G′u ⊂ A such that vj ∈ {(ai,j)1≤i,j≤2 : ai,j ∈ G′u} for all vj ∈ U0. (e 12.56) Choose δ′1 > 0 and a sufficiently large finite subset Gv ⊂ A1 which satisfying the following: If p ∈ A1 is a projection such that kpx − xpk < δ′1 for all x ∈ Gv, then there are unitaries wj ∈ (diag(p, p) ⊗ 1C(X))M2(A)(diag(p, p) ⊗ 1C(X)) such that kdiag(p, p)vj diag(p, p) − wjk < δ1/16n for all vj ∈ U0 and j = 1, 2, ..., K. (e 12.57) Let G′2 = F1 ∪ G′1 ∪ H′0 ∪ Gu ∪ {a(y), xa(y),j , x∗a(y),j : y ∈ H′0}, and M1 = 64(max{kxa(y),jk : y ∈ H′0} + 1) · max{N (y) : y ∈ H′0}. (e 12.58) Put δ′′1 = min{δ′1, δ1}/(64(n+1)M1). Since A1 ∈ B1, there exists mutually orthogonal projections p′0, p′1 ∈ A1, a C∗-subalgebra C ∈ C and 1C = p1, unital δ′′1 /16-G′2-multiplicative contractive completely positive linear maps ı′00 : A1 → p′0A1p′0 and ı′01 : A1 → C such that n+1 diag( z p′0, p′0, ..., p′0) . p′1 and kx − ı′00(x) ⊕ ı′01(x)k < δ′′1 } { (e 12.59) for all x ∈ G′1 ∪ G′2, where ı′00(a) = p0ap0 for all a ∈ A1. Define p0 = p′0 ⊗ 1C(X), p1 = p′1 ⊗ 1C(X). Without loss of generality, we may assume that p′0 6= 0. Since A1 is simple, there is an integer N0 > 1 such that This also implies that N0[p′0] ≥ [p′1] in W (A1). N0[p0] ≥ [p1]. 101 (e 12.60) (e 12.61) Define ı00 : A → p0Ap0 by ı00(a ⊗ f ) = ı′00(a) ⊗ f and ı01 : A → C ⊗ C(X) by ı01(a ⊗ f ) = ı′01(a) ⊗ f for all a ∈ A1 and f ∈ C(X). Define L0 : A → A by L0(a) = ı00(a) ⊕ ı01(a) for all a ∈ A. For each vj ∈ U0, there exists a unitary wj ∈ M2(p0Ap0) such that kdiag(p0, p0)vjdiag(p0, p0) − wjk < δ1/16n, j = 1, 2, .... (e 12.62) Note that C1 = C ⊗ C(X)⊂ A. ı♯ 0 : C q → Aq 12.3 and define 1 be defined by ı♯ Let ı0 : C → A be the natural embedding as C is a unital C∗-subalgebra of p1A1p1. Let 1 \{0} → (0, 1) be the map given by 0(c) = c for c ∈ C. Let ∆0 : Aq,1 0( h1))∆(h2)/4 : h ≥ h1 ⊗ h2, h1 ∈ C \ {0} and h2 ∈ C(X)+ \ {0}} ∆1(h) = sup{∆0(ı♯ for all h ∈ (C1)+ \ {0}. Let G′3 = ı′0,1(G′1 ∪G′2). Let G3 = {a⊗ f : a ∈ G′3 and f ∈ G′′1}. Let H3 ⊂ (C1)+ \{0} (in place of H1), γ′1 > 0 (in place of γ1), γ′2 > 0 (in place of γ2), δ2 > 0 (in place of δ), G4 ⊂ C1 (in place of G), P2 ⊂ K(C1) (in place of P), H′4 ⊂ (C1)s.a. (in place of H2), U2 ⊂ U (M2(C1))/U0(M2(C1)) (in place of U —see 12.6) be the finite subsets and constants as required by 12.5 for δ1/16 (in place of ε) and G3 (in place of F), ∆1/2 (in place of ∆) and for C1 (in place of A). Let U2 ⊂ U (M2(C1)) be a finite subset which has an one-to-one correspondence to its image in U (M2(C1))/CU (M2(C1)) which is exactly U2. We also assume that {[u] : u ∈ U2} ⊂ P2. Without loss of generality, we may assume that H3 = {h1 ⊗ h2 : h1 ∈ H′3 and h2 ∈ H′′3}, where H′3 ⊂ C+ \ {0} and H′′3 ⊂ C(X)+ \ {0} are finite subsets, and G4 = {a ⊗ f : a ∈ G′4 and f ∈ G′′4}, (e 12.63) (e 12.64) where G′4 ⊂ C and G′′4 ⊂ C(X) are finite subsets. 12.3 for ∆0 and H′0 ∪ H′3. Let δ3 > 0 (in place of δ) and let G5 ⊂ A1 (in place of G) be the finite subset as required by Set δ = min{1/16, ε/16, δ1 , δ2, δ3} 128N0(n + 1)(T (H4) + N (H4)) , (e 12.65) and set G6 = {G′2 ∪ L0(G′2) ∪ ı′01(G′2) ∪ n−1[j=1 {Ad uj ◦ ı′01(G′2)} ∪ G′4 ∪ G5 and G = {a ⊗ f : a ∈ G6 and f ∈ G′′1 ∪ H′′0 ∪ G′′4} ∪ {pj : 0 ≤ j ≤ 1} ∪ {vj, wj : 1 ≤ j ≤ K}. To simplify notation, without loss of generality, we may assume that G ⊂ A1. Let P = P1 ∪{pj : 0 ≤ j ≤ 1} ∪ [ı](P2), where ı : C1 → A is the embedding. Let H1 = H′′0 ∪ H′′3. Let U′2 = {diag(1−p1, 1−p1)+w : w ∈ U2} and let U′′0 = {wj +diag(p1, p1) : 1 ≤ j ≤ K}. Let U = {¯v : v ∈ U1 ∪ U′2} and let H2 = H′′4. Let σ1 = min{ 1 16nN0}. 4n , Now we assume that B is as in the statement, ϕ, ψ : A → B are two unital G-δ-multiplicative contractive completely positive linear maps satisfying the assumption for the above defined δ, G, P, H1, U , H2, σ1 and σ2. 16nN0} and σ2 = min{ 16nN0 γ′ 2 γ′ 1 1 , 102 Note that B′ is in B1 and B = B′ ⊗ U. We may also write B = B1 ⊗ U, where B1 = B′ ⊗ U, since U is strongly self absorbing. Without loss of generality, by the fact that U is strongly self absorbing, we may assume that the image of both ϕ and ψ are in B1. By replacing ψ by Ad u0 for some unitary u0 ∈ U (B1) if necessary, we may assume that ψ ◦ ı′00(1A) = ϕ ◦ ı′00(1A) = q. (e 12.66) There is an integer m ≥ n and mutually orthogonal and mutually unitarily equivalent projections j=1 ei = 1U . Define ϕ′0, ψ′0 : A → qB1q ⊗ 1U (see (e 12.66)) by e1, ..., em ∈ U such thatPm for all a ∈ A. Define Φ′, Ψ : A → (1 − q)B1(1 − q) ⊗ 1U by ϕ′0(a) = ϕ ◦ ı00(a) ⊗ 1U and ψ′0(a) = ψ ◦ ı00(a) ⊗ 1U Φ′(a) = ϕ ◦ ı01(a) ⊗ 1U and Ψ′(a) = ϕ ◦ ı01(a) ⊗ 1U for all a ∈ A. Define Φ, Ψ : C1 → (1 − q)B1(1 − q) by Φ = ϕ ◦ ı and Ψ = ψ ◦ ı. Define ψ′i : A → (1 − q)B1(1 − q) ⊗ ei by ψ′i(a) = Ψ′(a)ei for all a ∈ A. (e 12.67) (e 12.68) (e 12.69) (e 12.70) Note, by the choice of δ and G, Φ and Ψ are G-δ-multiplicative. By (the proof of) 12.3, τ (Φ(h)) ≥ ∆1(h)/2 for all h ∈ H3. By assumptions, for all τ ∈ T (B1), τ (Φ(c)) − τ (Ψ(c)) < σ1 for all c ∈ H′′4. Therefore, for all t ∈ T ((1 − q)B1(1 − q)), t(Φ(c)) − t(Ψ(c)) < γ′1 for all c ∈ H′′4. Since [ı](P2) ⊂ P, by assumptions, one also has [Φ]P2 = [Ψ]P2 . One also computes that (as elements in M2((1 − q)B1)(1 − q))) dist(Φ‡(¯v)Ψ‡(¯v∗)) < γ′2 for all v ∈ U2. (e 12.71) (e 12.72) (e 12.73) (e 12.74) (e 12.75) By the choices of δ, G4, γ′1, γ′2, P2, H3, H′′4 and U2, and by applying 12.5, there exists u1 ∈ (1 − q)B1(1 − q) such that ku∗1Φ(c)u1 − Ψ(c)k < δ1/16 for all c ∈ G3. Thus, by (e 12.59), ku∗1Φ′(a)u1 − Ψ′(a)k < δ1/16 + δ′′ for all a ∈ G1. We check that, by assumption (e 12.48) and (e 12.55), (e 12.76) (e 12.77) τ (Ψ′(h)) ≥ 15∆(h)/16 for all h ∈ H′′0 and for all τ ∈ T (B1). (e 12.78) 103 By 12.4, it follows that Ψ′ is T ′ × N′-H′′0-full. Note that k N ′′(a(y))Xi=1 Ψ′(x∗a(y),i)Ψ′(a(y))Ψ′(xa(y),i) − (1 − q)k < 4N′′(a(y))T ′′(a(y))δ (e 12.79) for all a ∈ H′0. We conclude that Ψ′ is T × N -H0-full. i = 1, 2, ..., m. It is easy to see that It follows that ψ′i is T × N -H0-full, [ψ′0]P1 = [ϕ′0]P1. (e 12.80) Assumptions also imply that dist(ϕ‡(wj + diag(p1, p1)), ψ‡(wj + diag(p1, p1))) < σ2 (e 12.81) j = 1, 2, ..., K. By e 12.61, applying (11.9) and then (e 12.62), we compute that dist((ϕ′0)‡(vj), (ψ′0)‡(vj)) < γ′2, j = 1, 2, ..., K. (e 12.82) It follows from 12.1 and its remark 12.2 that there exists a unitary u2 ∈ B such that ku∗2(ϕ′(a) ⊕ ψ′1(a) ⊕ ··· ⊕ ψ′n(a))u2 − (ψ′0(a) ⊕ ψ′1(a) ⊕ ··· ⊕ ψ′n(a))k < ε/16 (e 12.83) for all a ∈ F. In other words, ku∗2(ϕ′0(a) ⊕ Ψ′(a))u2 − ψ′0(a) ⊕ Ψ′(a)k < ε/16 for all a ∈ F. (e 12.84) Thus, by (e 12.77), ku∗2(ϕ′0(a) ⊕ u∗1Φ′(a)u1)u2 − ψ0(a) ⊕ Ψ(a)k < ε/16 + δ1/16 + δ′′ for all a ∈ F. (e 12.85) Let u = (q + u1)u2 ∈ U (B). Then, by (e 12.59), one has ku∗ϕ(a)u − ψ(a)k < ε for all a ∈ F, (e 12.86) as desired. Remark 12.10. As in remark 12.6, the condition that U ⊂ U (M2(A))/CU (M2(A)) can be replaced by U ⊂ Jc(U (M2(A))/U0(M2(A))). Moreover, if X is a point, or equivalently, A ∈ B1, we may take U ⊂ Jc(U (A)/U0(A)), since A has stable rank one. Furthermore, in this case , we do not need ∆ and (e 12.48). Let G be the subgroup generated by the finite subset U ⊂ Jc(K1(A)) and G = G1 ⊕ T or(G1), where G1 is free. Since Aff(T (B))/ρB(K0(B)) is torsion free (by 11.5), we may further assume that G is free. In the case that A =S∞n=1 An, in the theorem above, one may choose U to be in An for some sufficiently large n. 13 The range of invariant Notation 13.1. Let A be a unital subhomogeneous C-algebra, that is, the maximal dimension of irreducible representations of A is finite. Let us use RF (A) to denote the set of equivalence classes of all (not necessarily irreducible) finite dimensional representations. In this section, by {y} = {x1, x2, ..., xk}, we mean that the representation corresponding to y is direct sum of the representations x1, x2, ..., xk. If some of x repeats k times, then we use x∼k to denote this. That 104 2 1 , z∼k2 , ..., z∼km is, {y} may be written as {z∼k1 we do not insist that each zi should be irreducible. m }, where for each j, zj = xi for some i. Note that Let us use Sp(A) to denote the set of equivalence classes of all irreducible representations of A, which may be viewed as a subset of RF (A). Since A is of type I, the set Sp(A) has an one-to-one correspondence to the set of primitive ideals of A. Let X ⊂ Sp(A) be a closed subset, quotient algebra A/IX . If ϕ : B → A is a homomorphism, then we will use ϕX : B → AX to denote the composition π ◦ ϕ, where π : A → AX is the quotient map. As usual, if B1 is a subset of B, we will also use ϕB1 to denote the restriction of ϕ on B1. These two notation will not be confused, since it will be clear from content which notation we refer to. then X corresponds to the idea IX =Tψ∈X ker ψ. In this section, let us use AX to denote the If ϕ : A → B is a homomorphism, then we write Sp(ϕ) = {x ∈ Sp(A) : ker ϕ ⊆ ker x}. Notation 13.2. In this section we will use the concept of sets with multiplicity. There- fore X1 = {x, x, x, y} is different from X2 = {x, y}. We will also use x∼k for a simplified } (see 1.1.7 of [35]). For example, {x∼2, y∼3} = {x, x, y, y, y}. Let notation for {x, x, ..., x } n } and Y = {x∼j1 X = {x∼i1 , ..., x∼in zero which means the element xk does not appear in the set X (or Y )). k = 1, 2, ..., n, then we say that X ⊂ Y (see 3.21 of [35]). We define n } (some of the ik’s (or jk’s) may be If ik ≤ jk for all , ..., x∼jn {z , x∼j2 , x∼i2 k 1 2 1 2 X ∪ Y = {x∼max(i1,j1) 1 , x∼max(i2,j2) 2 , ..., x∼max(in,jn) n }. 1 2 1 2 n }. , x∼j2 , x∼ki2 , ..., x∼kin ,··· , x∼jn By X∼k we mean the set {x∼ki1 If ϕ : A → Mm(C) is a homomorphism, let us use SP (ϕ) to denote the corresponding equivalent class of ϕ in RF (A). Any finite subset of RF (A) also defines an element in RF (A) which is the equivalent class of the direct sum of all corresponding representations in the set with correct multiplicities. If both X and Y are finite subsets of RF (A) with multiplicities, we say X ⊂ Y if the representation corresponding to X is equivalent to a sub-representation of that corresponding to Y . That is, if we rewrite X and Y as X = {x∼i1 n } and Y = {x∼j1 n }, with xi being irreducible representation, then ik ≤ jk for each k ∈ {1, 2,··· , n}. It is clear that for two homomorphisms ϕ1 : A → Mm1(C) and ϕ2 : A → Mm2(C), we have SP (ϕ1) ⊂ SP (ϕ2) if and only if ϕ1 is equivalent to sub-representation of ϕ2. Strictly speaking, an element in RF (A) is regarded as a set with multiplicity whose elements are in Sp(A). But when we write X = {z∼k1 m }, we do not insist that zi itself in Sp(A). It may be a list of several elements in Sp(A)—that is, we do not insists zi to be irreducible (but as we know, it can always be decomposed into irreducible ones). So in this notation, we do not differentiate {x} and x, both give same element in RF (A) and same set with multiplicity whose elements in Sp(A). Comparing with notation in 13.1, if ϕ : A → Mm(C) is a homomorphism and if SP (ϕ) = {x∼k1 }, with x1, x2, ..., xi being irreducible representations, then Sp(ϕ) = {x1, x2, ..., xi} ⊂ Sp(A). So Sp(ϕ) is an ordinary set which is a subset of Sp(A), while SP (ϕ) is a set with multiplicity, whose elements are also elements in Sp(A). ,··· , z∼km ,··· , x∼in , ..., x∼ki , x∼k2 , z∼k2 , x∼i2 1 2 1 2 1 2 i Notation 13.3. Let us recall some notations from 4.8. Suppose that A = Am ∈ Dm constructed in the following sequence A0 = F0, A1 = A0 ⊕Q1C(Z1,F1)Q1 P1C(X1, F1)P1, A2 = A1 ⊕Q2C(Z2,F2)Q2 P2C(X2, F2)P2, ··· (e 13.1) is as in 4.8. Let Λ : A →Lm where x ∈ Xk and j the positive integer represented j-th block of Ak =Llk k=0 PkC(Xk, Fk)Pk, be the inclusion homomorphism as 4.8. Let π(x,j), j=1 Pk,jC(Xk, Fk)Pk,j, 105 be the finite dimensional representation of A as in 4.8. According to 13.1, one has that π(x,j) ∈ RF (A). Suppose all Xk are metric spaces. Let ϕ : A → M•(C) be a homomorphism. We say that SP (ϕ) is δ-dense in Sp(A) if for each irreducible representation θ of A, there are two points x, y,∈ Xk and j (same k and j) such that dist(x, y) < δ, and such that θ ⊂ π(y,j) and π(x,j) ⊂ SP (ϕ). This will be used in this section and next. If Xk = [0, 1], then we use the ordinary metric of interval [0, 1]. n,∞(I). Consequently, the C∗-algebra of the given inductive system is simple. 13.4. Let A = lim−→(An, ϕn,m) be an inductive limit with all An ∈ Dl(n) for some l(n). The limit C∗ algebra is simple if the inductive system satisfy the following condition: for any n and δ > 0, there is an integer m > n such that for any y ∈ Sp(Am), SP (ϕn,my) is δ-dense— consequently, for any m′ > m and any y ∈ Sp(Am′), SP (ϕn,m′y) is δ-dense. The proof of this claim is standard. Indeed, for any nonzero element f ∈ An, it must be nonzero on a open ball of radius δ (for some δ > 0) of a certain Xk (a space appears in the construction of An). Then, by the given condition, there is an m, such that for any m′ > m and any irreducible representation θ of Am′, one has θ(ϕn,m′(f )) 6= 0. It follows that the ideal I generated by ϕn,m′(f )) in Am′ equals Am′—otherwise, any irreducible representation θ of Am′/I 6= 0 (which is also a irreducible representation of Am′ ) satisfies θ(ϕn,m′(f )) = 0, and this contradicts with the fact that θ(ϕn,m′(f )) 6= 0 for all irreducible representation θ. It is well known that an ideal I of the limit C-algebra lim(An, ϕn,m) is always the limit of ideals lim(In, ϕn,mIn), where In = ϕ−1 Definition 13.5. Denote by N0 the class of those unital simple C∗-algebras A in N for which A ⊗ U ∈ N ∩ B0, for any UHF-algebra U of infinite type (see 2.9 for definition of class N ). Denote by N1 the class of those unital C∗-algebras A in N for which A ⊗ U ∈ N ∩ B1, for any UHF-algebra of infinite type. In section 19, we will show that N1 = N0. Also denoted by N Z0 (respectively N Z1 ) the class of all Z-stable C∗-algebras in N0 (respec- tively N1). 13.6. Let (G, G+, u) be a scaled ordered abelian group (G, G+) with order unit u ∈ G+ \ {0}. The scale is given by {g ∈ G+ : g ≤ u}. Sometimes we will also call u the scale of the group. Let S(G) := Su(G) be the state space of G. Suppose that ((G, G+, u), K, ∆, r) is a weakly unperforated Elliott invariant—that is, (G, G+, u) is a simple scaled ordered countable group, K is a countable abelian group, ∆ is a metrizable Choquet simplex, and r : ∆ → S(G) is a surjective affine map such that for any x ∈ G, x ∈ G+ \ {0} if and only if r(τ )(x) > 0 for all τ ∈ ∆. (e 13.2) The above condition (e 13.2) is also called weakly unperforated for the simple ordered group. (Note that this condition is equivalent to that x ∈ G+ \ {0} if and only if for any f ∈ S(G), f (x) > 0. The latter condition does not mention Choquet simplex ∆.) In this paper, we only consider the Elliott invariant for stably finite simple C∗-algebras and therefore ∆ is not empty. The above weakly unperforated condition is also equivalent to that x > 0 if nx > 0 for some positive integer n. In this section, we will show that for any weakly unperforated Elliott invariant ((G, G+, u), K, ∆, r), there is a unital simple C-algebra A in the class N Z0 such that ((K0(A), K0(A)+, [1A]), K1(A), T (A), rA) ∼= ((G, G+, u), K, ∆, r). In [25], Elliott constructed an inductive limit A with Ell(A) = ((G, G+, u), K, ∆, r). In this section, we will use modified building blocks to construct a simple C∗-algebra A such that Ell(A) is as described, and, in addition, as a direct consequence of the construction, one has that A ∈ N Z0 . 106 13.7. Our construction will be a modification of the Elliott construction mentioned above. As matter of fact, for the case that K = {0} and G torsion free, our construction uses the same building blocks in C0, as in [25]. We will repeat a part of the construction of Elliott for this case. There are two steps in Elliott construction: Step 1. Construct an inductive limit with inductive limit of ideals C1 −→ C2 −→ ··· −→ C I1 −→ I2 −→ ··· −→ I such that the non-simple limit C has the described Elliott invariant and the quotient C/I is a simple AF algebra. For the case of K = {0} and G torsion free, we will use the notation Cn and C for the construction, and reserve An and A for the general case. Step 2. Modify the above inductive limit to make C (or A in the general case) simple without changing the Elliott invariant of C (or A). For reader’s convenience, we will repeat Step 1 of Elliott construction with minimum modi- fication. For Step 2, we will use a slightly different way to modify the inductive limit which will be more suitable for our purpose—that is, to construct an inductive limit A ∈ N Z0 with possible nontrivial K1 and nontrivial Tor(K0(A)). 13.8. Let ρ : G → Aff(∆) be the dual map of r : ∆ → S(G). That is, for every g ∈ G, τ ∈ ∆, ρ(g)(τ ) = r(τ )(g) ∈ R. Since G is weakly unperforated, one has that g ∈ G+ \ {0} if and only if ρ(g)(τ ) > 0 for all τ ∈ ∆. Note that Aff(∆) is an ordered vector space with the strict order, i.e., f ∈ Aff(∆)+ \{0} if and only if f (τ ) > 0 for all τ ∈ ∆. Since ∆ is a metrizable compact convex set, there is a countable dense subgroup G1 ⊂ Aff(∆). Put H = G ⊕ G1 and define ρ : H → Aff(∆) by ρ((g, f ))(τ ) = ρ(g)(τ ) + f (τ ) for all (g, f ) ∈ G ⊕ G1 and τ ∈ ∆. Define H+ \ {0} to be the set of elements (g, f ) ∈ G ⊕ G1 with ρ((g, f ))(τ ) > 0 for all τ ∈ ∆. The order unit (or scale) u ∈ G+ could be regarded as (u, 0) ∈ G ⊕ G1 = H as the order unit of H+ (still denote it by u). Then (H, H+, u) is a simple ordered group. Using the fact that ρ(H) is dense and therefore has the Riesz interpolation property, it is straight forward to prove that (H, H+, u) is a Riesz group. We give a brief proof of this fact as below. Let a1, a2, b1, b2 ∈ H with ai < bj for i, j ∈ {1, 2}, then for each τ ∈ ∆, max{ρ(a1)(τ ), ρ(a2)(τ )} < min{ρ(b1)(τ ), ρ(b2)(τ )}. Let d = minτ∈∆ (min{ρ(b1)(τ ), ρ(b2)(τ )} − max{ρ(a1)(τ ), ρ(a2)(τ )}) > 0. Since ρ(H) is dense in ∆, there is a c ∈ H such that max{ρ(a1)(τ ), ρ(a2)(τ )} < ρ(c)(τ ) < min{ρ(b1)(τ ), ρ(b2)(τ )} for all τ ∈ ∆. Consequently, ai < c < bj for i, j ∈ {1, 2}. As a direct summand of H, the subgroup G is relatively divisible subgroup of H, i.e., if g ∈ G, m ∈ N \ {0}, and h ∈ H such that g = mh, then there is g′ ∈ G such that g = mg′. Note that G $ H. Now we assume, until 13.21, that G is torsion free and K = 0. Then H is also torsion free. Therefore H is a dimension group. 13.9. In 13.8 we can choose the dense subgroup G1 ⊂ Aff(∆) to contain at least three elements x, y, z ∈ Aff(∆) such that x, y and z are Q-linearly independent. With this choice, when we write H as inductive limit of finite direct sums of copies of ordered group (Z, )+) as in Theorem 2.2 of [20] , we can assume all Hn have at lease three copies of Z. H1 −→ H2 −→ ··· 107 Note that the homomorphism γn,n+1 : Hn = Zpn −→ Hn+1 = Zpn+1 is given by a pn+1 × pn matrix c = (cij) of nonnegative integers, where i = 1, 2, ..., pn+1; j = 1, 2, ..., pn and cij ∈)+ := {0, 1, 2, ...}. For M > 0, if all cij ≥ M , then we will say γn,n+1 is at least M -large or has multiplicity at least M . Note that since H is a simple group, passing to subsequences, we may assume that at each step γn,n+1 is at least Mn-large for arbitrary choice of Mn depending on our construction up to step n. 13.10. As in 13.8 and 13.9, we have G ⊆ H with G+ = H+ ∩ G and both G and H share the same order unit u ∈ G ⊆ H. As in 13.9, write H as inductive limit of Hn—finite direct sum of ordered groups (Z, Z+). Let Gn = γ−1 n,∞(γn,∞(Hn) ∩ G), where γn,∞ : Hn → H is induced map to the limit. We may assume u ∈ Gn ⊆ Hn for each n and (Gn)+ = (Hn)+ ∩ Gn. Since G is a relatively divisible subgroup of H and H is torsion free, the quotient Hn/Gn is torsion free group and therefor a direct sum of copies of Z, denoted by Zln. Then we have the following commutative diagram γ12G1 γ12 G1 H1 G2 H2 H1/G1 γ12 / / H2/G2 ··· ··· / ··· / G / H / H/G Let Hn =(cid:0)Zpn, Zpn realized as the K0 group of Fn =Lpn + , un(cid:1), where un = ([n, 1], [n, 2], ..., [n, pn]) ∈ (Z+ \ {0})pn. Then Hn can be i=1 M[n,i](C)—that is, (K0(Fn), K0(Fn)+, 1Fn) = (Hn, (Hn)+, un). If there are infinitely many n such that the inclusions Gn → Hn are isomorphisms, then, by passing to subsequence, we have that Gn → Hn is also an isomorphism which contradict with G $ H in13.8. Hence, without lose of generality, we can assume that for all n, Gn $ Hn, and therefore Hn/Gn 6= 0. Hn/Gn being identified as a map (still denoted by) To construct the C-algebra with K0 being (Gn, (Gn)+, un), we consider the map π : Hn → π : Zpn −→ Zln. as in [25]. We emphasis that ln > 0 for all n. Such a map can be realized as difference of two maps b0, b1 : Zpn −→ Zln corresponding to two ln × pn matrices of strictly positive integers b0 = (b0,ij) and b1 = (b1,ij). That is, π t1 t2 ... tp  =(cid:0)b1 − b0(cid:1)  t1 t2 ... tp  ∈ Zln, for any  t1 t2 ... tp  ∈ Zpn. 108 / / _ / / _ / _ / / / / / / / Note that un = ([n, 1], [n, 2],··· , [n, pn]) ∈ Gn and hence π(un) = 0. Consequently, b1 [n, 1] [n, 2] ... [n, pn]  = b0 pnXj=1 {n, 1} {n, 2}... {n, pn} ,  [n, 1] [n, 2] ... [n, pn] ,  pnXj=1 {n, i} , b1,ij[n, j] = b0,ij[n, j]. i.e., denote that Let En = ⊕ln i=1M{n,i}(C). We can choose any two homomorphisms β0, β1 : Fn → En such that (β0)∗0 = b0 and (β1)∗0 = b1. Then define which is A(Fn, En, β0, β1) as in the definition of 3.1. Using the following six-term exact sequence Cn =n(f, a) ∈ C([0, 1], En) ⊕ Fn; f (0) = β0(a), f (1) = β1(a)o, K0(C0(cid:0)(0, 1), En(cid:1) / K0(Cn) / K0(Fn) K1(Fn) K1(Cn) K1(C0(cid:0)(0, 1), En(cid:1), and the fact that π is surjective, we have K1(Cn) = 0 and (cid:16)K0(Cn), K0(Cn)+, 1Cn(cid:17) =(cid:16)Gn, (Gn)+, un(cid:17). Note that in the above diagram the map K0(Fn) = Zpn → K1(C0(cid:0)(0, 1), En(cid:1)) = Zln is given by (b1 − b0) ∈ Mln×pn(Z), which is surjective, as quotient map Hn(cid:0) = Zpn(cid:1) → Hn/Gn(cid:0) = Zln(cid:1). Zpn) −→ Hn/Gn(= Zln). For example, if (mij) ∈ Mln×pn(cid:0)Z+ \ {0}(cid:1) is any ln × pn matrix of As observed in [25], in the construction of Cn, we have the freedom to choose the pair of the K0 map (β0)∗0 = b0 and (β1)∗0 = b1 as long as the difference is the same map π : Hn(= positive integer, then we can replace b0,ij by b0,ij + mij and, at the same time, replace b1,ij by b1,ij + mij. That is, we can assume that each entry of b0 (and of b1) is larger than any fixed integer M which depends on Cn−1 and ψn−1,n : Fn−1 → Fn. Also, we can make all the entries of one column (say, the third column) of both b0 and b1 much larger than all the entries of another column (say, the second), by choosing mi3 >> mj2 for all i, j. 13.11. For later use, we will also deal with the case that Gn = Z• ⊕ G′n and Hn = Z• ⊕ H′n, with the inclusion map being identity for the first • copies of Z. In this case, the quotient map Hn(= Zpn) → Hn/Gn(= Zln) given by the matrix b1 − b0 maps the first • copies of Z to zero. For this case, it will be much more convenient to assume that the first •-columns of both matrices b0 and b1 are zero and each entry of the last (pn − •)-column of them are larger than any previously given integer M . Now we have that the entries of the matrices b0, b1 are strictly positive integers except the ones in the first •-columns which are zero. Consider the following diagram 109 / / O O o o o o Zp0 1 ⊕ G′1 γ12G1/ Zp0 2 ⊕ G′2 Zp0 1 ⊕ H′1 γ12 / Zp0 2 ⊕ H′2 H1/G1 γ12 / H2/G2 ··· ··· / ··· / G / H / H/G. n ⊕ G′n and Hn = Zp0 That is Gn = Zp0 p0 n = pn − p0 n copies of Z. Write p1 strictly positive integers for last p1 since ln > 0 (see 13.10), we have p1 n ⊕ H′n, with the inclusion map being identity for the first n. Then we assume that the entries of the matrices b0, b1 are n-columns and are zeros for the first p0 n-columns. Note that n > 0. n ⊕ G′n and Hn = Zp0 n column of matrices b0 and b1 equal to each others). However, for the case that p0 The inductive limits H = lim−→(Hn, γn,n+1) and G = lim−→(Gn, γn,n+1Gn) (with Gn ⊂ Hn) constructed in 13.10 are in fact special cases of the current construction when we assume that n = 0. One notices that, for the case Gn = Zp0 n ⊕ H′n, with the inclusion p0 map being identity for the first p0 n copies of Z, one can still use the construction of 13.10 to make all the entries of b0 and b1 (not only the entries of last p1 n) to be strictly positive (of course with first p0 n 6= 0 for all n, we will get better decomposition as we will discuss in the next section. In fact, for the case that p0 n = 0 for all n. n+1 M[n,i](C), then the If we write Fn (in 13.10)) asLp0 i=1 M[n,i](C). Moreover, the algebra Cn = n(f, a) ∈ maps β0 and β1 are zero on the part Lp0 C([0, 1], En) ⊕ Fn; f (0) = β0(a), f (1) = β1(a)o in 13.10 can be written asLp0 i=1 M[n,i](C) ⊕ C′, where C′ =n(f, a) ∈ C([0, 1], En)⊕F ′n; f (0) = β0(a), f (1) = β1(a)o, which is A(F ′n, En, β0F ′ i=1 M[n,i](C) ⊕ F ′n, where F ′n =Lpn n 6= 0 for all n, the construction is much simpler than the case that p0 as in the notation of 3.1. n n i=p0 n n, β1F ′ n) 13.12. Let us emphasis that once (Hn, (Hn)+, un) =(cid:0)Zpn, (Z+)pn, ([n, 1], [n, 2], ..., [n, pn])(cid:1), j=1 b1,ij[n, j] =Ppn are fixed, then the algebras Fn =Lpn {n, i} ,Ppn i=1 M[n,i](C), En =Lln are determined up to isomorphism. b0, b1 : Zpn → Zln, i=1 M{n,i}(C), where j=1 b0,ij[n, j], and Cn = A(Fn, En, β0, β1), with K0βi = bi (i = 0, 1), To construct the inductive limit, we not only need to construct Cn’s (later on An’s for the general case) but also need to construct ϕn,n+1 which realizes the corresponding K-theory map— that is (ϕn,n+1)∗0 = γn,n+1Gn. In additional, we need to make the limit algebras to have the desired tracial state space. In order to do all these, we need some extra conditions for b0, b1 for Cn+1 (or for An+1) depending on Cn (or An). We will divide the construction into several steps with gradually stronger conditions for b0, b1 (for Cn+1)—of course depending on Cn and the map γn,n+1 : Hn → Hn+1, to guarantee the construction can go through. Let Gn ⊂ Hn = Zpn, Gn+1 ⊂ Hn+1 = Zpn+1, and γn,n+1 : Hn → Hn+1, with γn,n+1(Gn) ⊂ Gn+1 be as in 13.9 and 13.10 (also see 13.11). Then γn,n+1 induces a map γn,n+1 : Hn/Gn(= Zln) → Hn+1/Gn+1 = (Zln+1). Let γn,n+1 : Hn (= Zpn) → Hn+1 (= Zpn+1) be given by a matrix c = (cij) ∈ Mpn+1×pn (Z+\{0}) and γn,n+1 : Zln → Zln+1 (as a map from Hn/Gn → Hn+1/Gn+1) be given by matrix d = (dij). Let πn+1 : Hn+1(= Zpn+1) → Hn+1/Gn+1(= Zln+1) be the quotient map. 110 / _ / / _ / _ / / / / / / / and reserve b0, b1 for Cn. Of course πn+1 = b′1 − b′0. Let us use b′0, b′1 : Zpn+1 → Zln+1 to denote the maps required for the construction of Cn+1, We will prove that if b′0, b′1 satisfy: (♦) b0,ji, b1,ji > lnXk=1(cid:0)djk + 2(cid:1) max(b0,ki, b1,ki) (e 13.3) for all i ∈ {1, 2, ..., pn} and for all j ∈ {1, 2, ..., ln+1}, where b0 = b′0 · c = (b0,ji) and b1 = b′1 · c = (b1,ji), then one can construct the homomorphism ϕn,n+1 : Cn → Cn+1 to realized the desired K-theory map (see 13.13 below). If b′0, b′1 satisfy stronger condition: (♦♦) b0,ji, b1,ji > 22n lnXk=1 (djk + 2){n, k}! (e 13.4) Ppn i=1 b1,ki[n, i] =Ppn for all i ∈ {1, 2, ..., pn} and for all j ∈ {1, 2, ..., ln+1}, then we can prove the limit algebra constructed has the desired tracial state space (see 13.15—13.19). (It follows from that {n, k} , If we want to make the limit algebra to be simple, then we need to modify the connecting homomorphisms ϕn,n+1 (but not the algebras Cn and Cn+1), and we need even stronger condition for b′0 = (b′0,ij) and b′1 = (b′1,ij) below: i=1 b0,ki[n, i], we have that (♦♦) is stronger than (♦).) b0,il > 22n lnXk=1 and b1,il > 22n lnXk=1 (dik + 2) · {n, k} + ((n − 1) (dik + 2) · {n, k} + ((n − 1) b0,kl) · b′0,i1! lnXk=1 b0,kl) · b′1,i1! , lnXk=1 (e 13.5) (e 13.6) for each l ∈ {1, 2, ..., pn} and i ∈ {1, 2, ..., ln+1}. But we will not do the modification for this special case since it only covers the case that the algebras have trivial K1 group and torsion free K0 group. Later on, we will deal with general case involving general K1 group and K0-group; the conditions will be (♦♦♦)1 and (♦♦♦)2 specified in 13.29 below. Let us emphasis that when we modify inductive limit system to make it simple, we never change the algebras Cn (or An in the general case), what will be changed are the connecting homomorphisms. Let A and B be two C∗-algebras, ϕ : A → B be a homomorphism and π ∈ RF (B). For the rest of this section, we will use ϕπ for the composition π ◦ ϕ, in particular, in the following statement and its proof. This notation is consistent with 13.1. In the following lemma we will give the construction of ϕn,n+1 if b′0 and b′1 satisfy the condition (♦) in 13.12—of course, the condition depends on the previous step. So this lemma provides the n + 1-step of the construction. Again, we first have G, then obtains H, Hn and Gn as constructed in 13.8 and 13.9. Lemma 13.13. Let (Hn, (Hn)+, un) =(cid:0)Zpn, (Z+)pn, ([n, 1], [n, 2], ..., [n, pn ])(cid:1), M[n,i](C), b0, b1 : Zpn → Zln, En = M{n,i}(C), β0, β1 : Fn → En lnMi=1 Fn = pnMi=1 (e 13.7) 111 with (β0)∗0 = b0, (β1)∗0 = b1, and Cn = A(Fn, En, β0, β1) with K0(Cn) = Gn be as in 13.10 or more generally as in 13.11. Let (Hn+1, H + n+1, un+1) =(cid:0)Zpn+1, (Z+)pn+1, ([n + 1, 1], [n + 1, 2], ..., [n + 1, pn+1])(cid:1), let γn,n+1 : Hn → Hn+1 be an ordered homomorphism with ϕn,n+1(un) = un+1 (as in 13.10 or 13.11 ), let Gn+1⊂Hn+1 be a subgroup containing un+1 (as in 13.9) and satisfying γn,n+1(Gn)⊂Gn+1. Let πn+1 : Hn+1(= Zpn+1) → Hn+1/Gn+1(= Zln+1) denote the quotient map. Suppose that the matrices b′0 = (b′0,ij), b′1 = (b′1,ij) : Zpn+1 → Zln+1 satisfy b′1 − b′0 = πn+1 and satisfy the condition (♦) in 13.12. (As a convention, we assume that the entries of first n+1 columns of the matrices b′0 and b′1 are zeros and the entries of the last p1 p0 columns are strictly positive. Note that p0 n+1 might be zero as in the special case 13.10.) n+1 = pn+1 − p0 n+1 Put Fn+1 =Lpn+1 i=1 M[n+1,i](C) and En+1 =Lln+1 i=1 M{n+1,i}(C) with {n + 1, i} = b′1,ij[n + 1, j] = pn+1Xj=1 b′0,ij[n + 1, j], pn+1Xj=1 and pick unital homomorphisms β′0, β′1 : Fn+1 → En+1 with Set (β′0)∗0 = b′0 and (β′1)∗0 = b′1. Cn+1 = A(Fn+1, En+1, β′0, β′1). Then there is a homomorphism ϕn,n+1 : Cn → Cn+1 satisfying the following conditions: (1) K0(Cn+1) = Gn+1 as a scaled ordered group (as already verified in 13.10). (2) (ϕn,n+1)∗0 : K0(Cn) = Gn → K0(Cn+1) = Gn+1 satisfies (ϕn,n+1)∗0 = γn,n+1Gn. (3) ϕn,n+1(C0(cid:0)(0, 1), En(cid:1))⊂C0(cid:0)(0, 1), En+1(cid:1). (4) Let ϕn,n+1 : Fn → Fn+1 be the quotient map induced by ϕn,n+1 (note from (3), we know that this quotient map exists), then ( ϕn,n+1)∗0 = γn,n+1 : K0(Fn) = Hn → K0(Fn+1) = Hn+1. (5) For each y ∈ Sp(Cn+1), Sp(Fn) ⊂ Sp(ϕn,n+1y). (6) For each j0 ∈ {1, 2, ..., ln+1}, i0 ∈ {1, 2, ..., ln}, one of the following holds: tains t ∈ (0, 1)i0⊂Sp(Cn); or (i) for each t ∈ (0, 1)j0⊂Sp(In+1) =Sln+1 (ii) for each t ∈ (0, 1)j0⊂Sp(In+1) =Sln+1 tains 1 − t ∈ (0, 1)i0⊂Sp(Cn). j=1 (0, 1)j⊂Sp(Cn+1), Sp(ϕn,n+1t) ∩ (0, 1)i0 con- j=1 (0, 1)j⊂Sp(Cn+1), Sp(ϕn,n+1t) ∩ (0, 1)i0 con- Consequently (as ln+1 > 0, that is En+1 6= 0 or there is at least one interval in Sp(Cn+1)), the following is true: if X⊂Sp(Cn+1) is δ-dense, thenSx∈X Sp(ϕn,n+1x) is δ-dense in Sp(Cn) (see 13.3). (In general, ϕn,n+1 is injective.) 112 Remark 13.14. Let In = C0(cid:0)(0, 1), En(cid:1) and In+1 = C0(cid:0)(0, 1), En+1(cid:1). If ϕn,n+1 : Cn → Cn+1 is as described in 13.13, then we have the following exact sequences: 0 0 / K0(Cn) K0(Cn/In) / K1(In) γn,n+1Gn γn,n+1 γn,n+1 / K0(Cn+1) / K0(Cn+1/In+1) / K1(In+1) / 0 / 0 where K0(Cn) and K0(Cn+1) are identified with Gn and Gn+1, K0(Cn/In)(= K0(Fn)) and K0(Cn+1/In+1)(= K0(Fn+1)) are identified with Hn and Hn+1, K1(In) is identified with Hn/Gn, and K1(In+1) is identified with Hn+1/Gn+1. Moreover, γn,n+1 is induced by γn,n+1 : Hn → Hn+1. Consider the matrix c = (cij) ∈ Mpn+1×pn Z+ \ {0}) which is induced by the map γn,n+1 : Hn (= Zpn) → Hn+1 (= Zpn+1) and consider the matrix d = (dij) which is induced by the map γn,n+1 : Zln → Zln+1 (as a map from Hn/Gn → Hn+1/Gn+1). Note that πn : Hn(= Zpn) → Hn/Gn(= Zln) is given by b1 − b0 ∈ Mln×pn(Z). Here, in the situation of 13.11, we assume the first p0 n columns are strictly positive. Let πn+1 : Hn+1(= Zpn+1) → Hn+1/Gn+1(= Zln+1) be the quotient map. Write Gn+1 = Zp0 n+1 columns being zero and the last p1 n+1 columns being strictly positive, such that πn+1 = b′1 − b′0 and n+1 ⊕ H′n+1, then we can choose both b′0 and b′1, with the first p0 n columns of both b0 and b1 are zero and last p1 n+1 ⊕ G′n+1, Hn+1 = Zp0 n+1 = pn+1 − p0 (♦) b0,ji, b1,ji > lnXk=1(cid:0)djk + 2(cid:1) max(b0,ki, b1,ki) (e 13.8) for all i ∈ {1, 2, ..., pn} and for all j ∈ {1, 2, ..., ln+1}, where b0 = b′0 · c = (b0,ji) and b1 = b′1 · c = (b1,ji). Note that ln+1 > 0 and p1 n+1 > 0. So we can make (♦) hold by only increasing the last p1 n+1 columns of the the matrices b′0 and b′1—that is the first p0 n+1 column of the matrices are still kept to be zero, since all the entries in c are strictly positive. Note that, even though the first n+1 column of b′0 and b′1 (as ln+1 × pn+1 matrices) are zero, but all entries of b0 and b1 (as p0 ln+1 × pn matrix) has been made strictly positive. Again note that the case of 13.10 is a special case of 13.11 for p0 Proof of 13.13. Suppose that b′0, b′1 satisfy b′1 − b′0 = πn+1 and (e 13.3) (and the first p0 columns to be zero). Now let En+1, β′0, β′1 : Fn+1 → En+1, and Cn+1 = A(Fn+1, En+1, β′0, β′1) be as constructed in 13.10. We will define ϕn,n+1 : Cn → Cn+1 to satisfy Conditions (2)–(6) of 13.13 (Condition (1) is a property for Cn+1 which is verified in 13.10). n+1 = 0, so one does not need to deal with this case separately. n+1 As usual, let us use F i There exists a unital homomorphism ϕn,n+1 : Fn → Fn+1 such that n) to denote the i-th block of Fn (or En). n (or Ei ( ϕn,n+1)∗0 = γn,n+1 : K0(Fn) = Hn → K0(Fn+1) = Hn+1, where γn,n+1 is defined in 13.9. Note that Sp(Cn+1) =`ln+1 j=1 (0, 1)jS Sp(Fn+1) (see §3). Write Sp(Fn+1) = (θ′1, θ′2, ..., θ′pn+1 ) and Sp(Fn) = (θ1, θ2, ..., θpn ). To define ϕn,n+1 : Cn → Cn+1, we need to specify ϕn,n+1y = ey ◦ ϕn,n+1 below, where for each y ∈ Sp(Cn+1), ϕn,n+1y : Cn ey / Cn+1 / Cn+1y , and ey is the point-evaluation at y. 113 / / / / / / / / / / / To actually construct ϕn,n+1, we first construct a homomorphism ψ : C([0, 1], En) → C([0, 1], En+1). This can be done by define, for each j, the map ψj : C([0, 1], En) → C([0, 1], Ej n+1) for each j = 1, 2, ..., ln+1. Let (f1, f2, ..., fln) ∈ C([0, 1], En). For any k ∈ {1, 2, ..., ln}, if djk > 0, then let Fk(t) = diag(cid:0) fk(t), fk(t), ..., fk(t) } {z djk (cid:1) ∈ C(cid:16)[0, 1], Mdjk·{n,k}(C)(cid:17) ; if djk < 0, then let (cid:1) ∈ C(cid:16)[0, 1], Mdjk·{n,k}(C)(cid:17) ; if djk = 0, then let djk {z Fk(t) = diag(cid:0) fk(1 − t), fk(1 − t), ..., fk(1 − t) } Fk(t) = diag(cid:0)fk(t), fk(1 − t)(cid:1) ∈ C(cid:0)[0, 1], M2·{n,k}(C)(cid:1) . ψj(f1, f2, ..., fln )(t) = diag(F1(t), F2(t), ..., Fln (t)) ∈ C(cid:16)[0, 1], M(Pln With the above notation, define k=1 d′ k·{n,k})(C)(cid:17) , (e 13.9) where d′k = djk 2 if djk 6= 0, if djk = 0. Note that {n + 1, j} = pn+1Xl=1 b′0,jl[n + 1, l] = b′0,jlcli[n, i] = b0,ji[n, i]. (e 13.10) pn+1Xl=1 pnXi=1 pnXi=1 Recall that b0 = b′0 · c =(cid:0)b0,ji(cid:1). From (e 13.3), e 13.10, d′k ≤ dkj + 2 and {n, k} =P b0,ki[n, i], we have {n + 1, j} = pnXi=1 b0,ji[n, i] > pnXi=1(cid:0) lnXk=1(cid:0)djk + 2(cid:1)b0,ki(cid:1)[n, i] ≥ k·{n,k})(C)(cid:17) can be regarded as a corner of the C- n+1) = C(cid:0)[0, 1], M{n+1,j}(C)(cid:1), and consequently, ψj can be regarded as map- n+1). Put all ψj together we get ψ : C([0, 1], En) → C([0, 1], En+1). d′k{n, k}. lnXk=1 k=1 d′ Hence the C-algebra C(cid:16)[0, 1], M(Pln algebra C([0, 1], Ej ping into C([0, 1], Ej Define ψ0, ψ1 : Cn → En+1 to be ψ0(f ) = ψ(f )(0) and ψ1(f ) = ψ(f )(1) for any f ∈ Cn⊂C([0, 1], En). Since ψ0(C0((0, 1), En)) = 0 and ψ1(C0((0, 1), En)) = 0, this de- fines maps α0, α1 : Fn → En+1. Note that for each j ∈ {1, 2, ..., ln+1}, the maps αj 1 : Fn → En+1 → Ej n+1 have spectra 0, αj SP (αj 0) =nθ∼i1 1 , θ∼i2 2 , ..., θ∼ipn pn o SP (αj 1) =nθ∼i′ 1 1 , θ∼i′ 2 2 , ..., θ∼i′ pn pn o and 114 b0,kl = b1,kl = 0 and consequently, il = i′l = 0. Put Note that i′l − il = Pln α0 = β′0 ◦ ϕn,n+1 : Fn Then for each j ∈ {1, 2, ..., ln+1}, the maps αj 0, αj / En+1 / Fn+1 β′ 0 / ϕn,n+1 SP (αj where , θ∼¯i2 2 , ..., θ∼¯ipn pn o and b′0,jkckl = b0,jl and 1 0) =nθ∼¯i1 pn+1Xk=1 ¯il = SP (αj and α1 = β′1 ◦ ϕn,n+1 : Fn 1 : Fn → En+1 → Ej 1) =nθ∼¯i′ pn+1Xk=1 ¯i′l = 2 1 1 b′1,jkckl = b1,jl. n+1 have spectra , θ∼¯i′ , ..., θ∼¯i′ 2 pn pn o , n in the case 13.11, then ϕn,n+1 / Fn+1 β′ 1 / / En+1 . respectively (see 13.2 for the notation used here), where and (b0,kl + b1,kl) djkb1,kl + Xdjk>0 djkb1,kl + Xdjk<0 il = Xdjk<0 i′l = Xdjk>0 k=1 djk(b1,kl − b0,kl), and note that if l ≤ p0 djkb0,kl + Xdjk=0 djkb0,kl + Xdjk=0 (b0,kl + b1,kl). From (e 13.3), we have that ¯il > il and ¯i′l > i′l. Furthermore ¯i′l − ¯il = Ppn+1 k=1 (b′1,jk − b′0,jk)ckl. Since (b′1 − b′0)c = d(b1 − b0), we have that ¯i′l − ¯il = i′l − il, and hence ¯i′l − i′l = ¯il − il , rl > 0. Note that these numbers are defined for the homomorphisms αj 1 : Fn → Ej n+1. So strictly speaking, rl should be denoted as rj 0, αj 0, αj 1, αj l > 0. Define a unital homomorphism Φ : Cn → C([0, 1], En+1) = ⊕ln+1 Φj(f1, f2,··· , fln, a1, a2, ..., apn) = diag(cid:0)ψj(f1, f2,··· , fln), a∼rj Again, define the maps Φ0, Φ1 : Cn → En+1 by 1 1 j=1 C([0, 1], Ej n+1) by , a∼rj 2 2 , ..., a∼rj pn pn (cid:1). Φ0(F ) = Φ(F )(0) and Φ1(F ) = Φ(F )(1), for F = (f1, f2, ..., fln , a1, a2, ..., apn) ∈ Cn. These two maps induce two quotient maps α0, α1, : Fn → En+1. 0 = αj 0)∗0 = (αj 0 and Ad U1◦ αj 0)∗0 and (αj 1 = αj 1 resp.) does. That is, (αj 0 (αj From our calculation, for each j ∈ {1, 2, ..., ln+1}, the map αj 1 resp.) has same spectrum as αj 0 (αj 1)∗0 = (αj 1)∗0. There are unitaries U0, U1 ∈ En+1 such that Ad U0◦ αj 1. Choose a unitary path U ∈ C([0, 1], En+1) such that U (0) = U0 and U (1) = U1. Finally, set ϕn,n+1 : Cn → C([0, 1], En+1) to be defined as ϕn,n+1 = Ad U ◦ Φ. From the construction, we have that ψ(C0(cid:0)(0, 1), En(cid:1))⊂C0(cid:0)(0, 1), En+1(cid:1) and consequently, Φ(C0(cid:0)(0, 1), En(cid:1))⊂C0(cid:0)(0, 1), En+1(cid:1). We conclude that ϕn,n+1 (In)⊂In+1. Since AdU (0) ◦ αj 0 = β′0 ϕn,n+1 and Ad U (1) ◦ αj 1 = αj 1 = β′1 ϕn,n+1 we have that ϕn,n+1 (Cn) ⊂ Cn+1 and furthermore, the quotient map from Cn/In → Cn+1/In+1 induced by ϕn,n+1 is the same as ϕn,n+1 (see definition of αj 1). Note that Condition (2)–(4) hold evidently. Condition (6) follows from (e 13.9) and the the definition of Fk(t). If x ∈ Sp(Fn+1) ⊂ Sp(Cn+1), then Sp(Fn) ⊂ Sp(ϕn,n+1x)(= Sp( ϕn,n+1x)), by the fact that all entries of c are strictly positive. If x ∈ (0, 1)j = Sp(C0((0, 1), Ej n+1), then each θi, as the only element in Sp(F i i > 0 times in Sp(ϕn,n+1x). Consequently, we also have Sp(Fn) ⊂ Sp(ϕn,n+1x). Hence Condition (5) holds. This finishes the proof. n)(⊂ Sp(Fn)), appears rj 0 and αj 0 = αj 115 / / 13.15. Let ϕ : Cn → Cn+1 be as in the proof above. We will calculate the map ϕ♯ n,n+1 : Aff(T (Cn)) → Aff(T (Cn+1)). Recall from 3.17 that Aff(T (Cn))⊂ lnMi=1 C([0, 1]i, R) ⊕ Rpn consisting of (f1, f2, ..., fln , h1, h2, ..., hpn ) which satisfies the conditions 1 (∗) fi(0) = {n, i}X b0,ijhj · [n, j] {n, i}X b1,ijhj · [n, j], i=1 C([0, 1]i, R)⊕Rpn+1 consisting of (f′1, f′2, ..., f′ln+1 fi(1) = (∗∗) 1 (e 13.11) (e 13.12) , h′1, h′2, ..., h′pn+1 ) which and Aff(T (Cn+1))⊂Lln+1 satisfies f′i(0) = Let Then 1 {n + 1, i}X b′0,ijh′j · [n + 1, j] and f′i(1) = 1 {n + 1, i}X b′1,ijh′j · [n + 1, j]. ϕ♯ n,n+1(f1, f2, ..., fln , h1, h2, ..., hpn ) = (f′1, f′2, ..., f′ln+1, h′1, h′2, ..., h′pn+1 ). h′i = 1 [n + 1, i] pnXj=1 cijhj[n, j]. Recall that (cij)pn+1×pn is the matrix corresponding to ( ϕn,n+1)∗0 = γn,n+1 for ϕn,n+1 : Fn → Fn+1, and note that since ϕn,n+1 is unital, one has cij[n, j] = [n + 1, j]. pnXj=1 Also note that 1 {n + 1, i}Xdik>0 f′i(t) = where ri l = = pn+1Xk=1 pn+1Xk=1 dikfk(t){n, k} + Xdik<0 + Xdik=0 b′0,ikckl −Xdik<0 b′1,ikckl −Xdik>0 dikb1,kl + Xdik>0 dikb1,kl + Xdik<0 n+1 = (pn+1 − p0 116 dikfk(1 − t){n, k}+ (fk(t) + fk(1 − t)){n, k} + ri pnXl=1 , (e 13.13) lhl[n, l] (b0,kl + b1,kl) (b0,kl + b1,kl) . (e 13.14) dikb0,kl + Xdik=0 dikb0,kl + Xdik=0 It follows from the last paragraph of 13.10 and 13.11 that, when we define Cn+1, we can always n+1) columns of the matrices b′0 = (b′0,ik) and increase the entries of the last p1 b′1 = (b′1,ik) by adding an arbitrarily (but same for b′0 and b′1) matrix (mik)ln+1×(p1 each mik > 0 sufficiently large, to the last p1 strengthen the requirement (♦) (see (e 13.3)) to , with n+1 columns of the matrices. In particular we can n+1 (♦♦) b0,il = pn+1Xk=1 b′0,ikckl and b1,il = pn+1Xk=1 b′1,ikckl! > 22n lnXk=1 (dik + 2){n, k}!(e 13.15) for all i∈{1, ..., ln+1}. This condition and e 13.14 (and note that b0,kl ≤ {n, k}, b1,kl ≤ {n, k} for any k ≤ ln) imply 22n − 1 22n b0,il and equivalently ri l ≥ n,n+1 : Aff(T (Cn)) → Aff(T (Cn+1)), and ϕ♯ n : Aff(T (Cn)) → Aff(T (Fn)) and π♯ 0 ≤ b0,il − ri Recall that ϕ♯ n,n+1 : Aff(T (Fn)) → Aff(T (Fn+1)) are the maps induced by corresponding homomorphisms (i.e., by the homomorphisms ϕn,n+1 and ϕn,n+1; and same for π♯ n+1 : Aff(T (Cn+1)) → Aff(T (Fn+1)), which are the maps induced by quotient maps πn : Cn → Fn and πn+1 : Cn+1 → Fn+1, respectively. Since ϕn,n+1 ◦ πn = πn+1 ◦ ϕn,n+1, we have (e 13.16) b0,il. l < 1 22n π♯ n+1 ◦ ϕ♯ n,n+1 = ϕ♯ n,n+1 ◦ π♯ n : Aff(T (Cn+1)) → Aff(T (Fn+1)). 0, β♯ n ◦ Γn = idAff(T (Fn)) as below. 13.16. For each n, we will define a map Γn : Aff(T (Fn)) → Aff(T (Cn)) which is a right inverse n : Aff(T (Cn)) → Aff(T (Fn))—that is, π♯ of π♯ Recall that Cn = A(Fn, En, β0, β1) with unital homomorphisms β0, β1 : Fn → En whose K-theory maps satisfy (β0)∗0 = b0 = (b0,ij) and (β1)∗0 = b1 = (b1,ij). For (h1, h2, ..., hpn ) ∈ Rpn, the linear maps β♯ 1 : Aff(T (Fn)) = Rpn → Aff(T (En)) = Rln are given by the matrices {n,i} (cid:17) and(cid:16) b1,ij [n,j] {n,i} (cid:17) , where i ∈ {1, 2, ..., ln}, j ∈ {1, 2, ..., pn}—that is, (cid:16) b0,ij [n,j]  j=1 b0,1j[n, j] · hj j=1 b0,2j[n, j] · hj j=1 b0,lnj[n, j] · hj h1 h2 ... hpn  =  β♯ 0 1 . (For β♯ 1, one replaces b0,ij by b1,ij in the construction above). For h = (h1, h2, ..., hpn ), set 1 1 ... {n,1}Ppn {n,2}Ppn {n,ln}Ppn 1(h) + (1 − t) · β♯ Γ′n(h)(t) = t · β♯ which is an element in C([0, 1], Rln ) =Lln 0(h), i=1 C([0, 1]i, R). Finally, define the map Γn : Aff(T (Fn)) = Rpn → Aff(T (Cn))⊂ C([0, 1]j , R) ⊕ Rpn by lnMj=1 C([0, 1]j , R) ⊕ Rpn. Γn(h) = (Γ′n(h), h) ∈ lnMj=1 Note that Γn(h) ∈ Aff(T (Cn)) (see (e 13.11) and (e 13.12) in 13.15). Evidently, π♯ idAff(T (Fn)). n ◦ Γn = 117 (e 13.17) Lemma 13.17. If condition (♦♦) (see (e 13.15)) holds, then for any f ∈ Aff(T (Cn)) with kfk ≤ 1, and f′ = ϕ♯ n,n+1(f ) ∈ Aff(T (Cn+1)), we have kΓn+1 ◦ π♯ n+1(f′) − f′k < 2 22n . Proof. Write f = (f1, f2, ..., fln , h1, h2, ..., hpn ) ∈ Aff(T (Cn)) and f′ = (f′1, f′2, ..., f′ln+1 , h′1, h′2, ..., h′pn+1 ) ∈ Aff(T (Cn+1)). (e 13.18) Since π♯ n+1 ◦ Γn+1 = idAff(T (Fn+1)), one has Γn+1 ◦ π♯ n+1(f′) := g′ := (g′1, g′2, ..., g′ln+1 , h′1, h′2, ..., h′pn+1 ); Note that the evaluations of f′ at zero, (f′1(0), f′2(0), ..., f′ln+1 that is, f′ and g′ have the same boundary value (h′1, h′2,··· , h′pn+1). one, (f′1(0), f′2(0), ..., f′ln+1 13.15, we also have (0)), and the evaluation at at (0)) are completely determined by h′1, h′2, ..., h′pn+1 . From (e 13.13) of f′i(t)−f′i(0) = 1 {n + 1, i}Xdik>0 And from (e 13.15), one has dikfk(t){n, k} + Xdik<0 + Xdik=0 dikfk(1 − t){n, k} + (fk(t) + fk(1 − t)){n, k} − f′i(t) − f′i(0) < 2 22n . 2 22n . Similarly, we have f′i(t) − f′i(1) < But, by the definition of Γn+1, we have Combining with g′i(0) = f′i(0) and g′i(1) = f′i(1), we have g′i(t) = tg′i(1) + (1 − t)g′i(0). as desired. g′i(t) − f′i(t) < 2 22n for all i, 118 l )hl[n, l] . pnXl=1 (b0,il − ri (fk(t) + fk(1 − t)){n, k}(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xdik>0 dikfk(t){n, k} + Xdik<0 dikfk(1 − t){n, k} + Xdik=0 ≤ lnXk=1 (dik + 2){n, k} ≤ 1 22n b0,il < 1 22n · {n + 1, i}. Note that b0,il < {n + 1, i}, sincePpn (e 13.16), we have l=1 b0,il[n, l] = {n + 1, i} and [n, l]≥ 1. Combining this with Remark 13.18. In the proof of Lemma 13.17, the key point is that for any f′ = ϕ♯ Aff(T (Cn+1)), we have that all components of f′, as Rl to a constant function. component f′j = ϕj n,n+1(f ) ∈ n+1-valued functions on (0, 1), are close In the definition of ϕn,n+1 of the proof of 13.13, we know that each n,n+1(f1, f2, ..., fln , a1, a2, ..., apn ) is written as diag(cid:0)ψj(f1, f2, ..., fln ), a∼rj up to unitary equivalence. The major part of diag(cid:0)a∼rj 1 1 1 1 , a∼rj 2 2 , ..., a∼rj pn pn (cid:1) , a∼rj 2 2 , ..., a∼rj pn pn (cid:1) is constant (of course up to unitary equivalence)—that is, each ai appeared above is a fixed matrix as in Fn and does not depend on t ∈ (0, 1)j ⊂ SpC0((0, 1), En+1) as a function. In fact, Condition (e 13.15) implies that this part occupies more than 22n−1 of the whole size {n + 1, j} of M{n+1,j}(C[0, 1]) = 22n C([0, 1], Ej n+1). In the rest of this section, we will use this argument several times. That is, some condition similar to (♦♦) (see (e 13.15)) will imply that the map ϕ# n,n+1 has the property that for any f , the image f′ = ϕ# n,n+1(f ) is close to constant (on certain connected subsets of spectrum), and therefore it “almost only” depends on its value at finitely many points (in the above case, it depend on its values on the spectrum of Fn). Then pass to the inductive limit and use the approximately intertwining argument, one can prove that the tracial state space of the limit algebra is as desired—in the case above, lim(Af f T Cn, ϕ# n,n+1) (see 13.19 below). n,n+1) = lim(Af f T Fn, ϕ# Theorem 13.19. If Cn and ϕn,n+1 : Cn → Cn+1 are as in 13.13 and 13.14 with condition (♦) (see (e 13.3)) being replaced by the stronger condition (♦♦) (see (e 13.15)), then the inductive limit C = lim(Cn, ϕn) has the Elliott invariants (K0(C), K0(C)+, 1C, T (C), rC ) = (G, G+, ∆, r) (here we assume G is torsion free and K1(C) = {0}). Proof. From the construction, we have the following infinite commutative diagram: I1 C1 I2 / C2 I3 C3 / ··· I / ··· C C1/I1 / C2/I2 / C3/I3 / ··· C/I , where Cn/In = Fn. Also from the construction, we have the following diagram: 0 0 0 K0(C1) = G1 K0(C2) = G2 K0(C3) = G3 K0(C1/I1) = H1 / K0(C2/I2) = H2 K0(C3/I3) = H3 K1(I1) / K1(I2) / K1(I3) 0 / 0 / 0 119 / ··· / ··· / ··· / ··· / ··· . / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / So K0(C) = lim(Gn, γn,n+1Gn ) = G. The inductive limits also induce the following diagram: Aff(T (C1)) Aff(T (C2)) Aff(T (C3)) ··· Aff(T (C)))) π# 1 π# 2 π# 3 Aff(T (F1)) / Aff(T (F2)) / Aff(T (F3)) / ··· Aff(T ((C/I))) . By the construction of Γn and 13.17 we have the following diagram approximately intertwining: Aff(T (C1)) ϕ# 1,2 π# 1 Γ1 Aff(T (F1)) eϕ# 1,2 / Aff(T (C2)) π# 2 Γ2 Aff(T (F2)) ϕ# 2,3 ϕ# 2,3 Aff(T (C3)) π# 3 Γ3 Aff(T (F3)) ··· Aff(T (C)) ··· Aff(T (C/I)) . Hence by (2.2)–(2.4) of [24], it induces unital ordered isomorphism between unital ordered real Banach spaces Aff(T (C)) and Aff(T ((C/I))) = Aff(∆). This isomorphism π# ∞ and its inverse Γ∞ are continuous with respect to the weak ∗ topology of Aff(T (C)) and Aff(T ((C/I)))—that is, for any net lα ∈ Aff(T (C)) and l ∈ Aff(T (C)), if lα(τ ) → l(τ ) for all τ ∈ T (C), then ∞(lα))(τ1) → (π# (π# ∞(l))(τ1) for all τ1 ∈ T (C/I). Furthermore, T (C) (or T (C/I)) is the set of all positive linear maps τ from Aff(T (C)) (or Aff(T (C/I))) to R with τ (1) = 1, which is continuous with respect to weak ∗ topology on Aff(T C) (or Aff(T (C/I))). Consequently T (C) = T (C/I) = ∆, as they can be recovered from Aff(T (C)) = Aff(T (C/I)) = Aff(∆) and as they are the sets of all positive unital linear weak ∗ continuous map from the same set Aff(T (C)) = Aff(T (C/I)) = Aff(∆) to R. Furthermore, the homomorphisms πi : Ci → Ci/Ii, i = 1, 2, ..., induce the inclusion map ι : K0(C) = G → K0(C/I) = H and the isomorphism Aff(T (C)) → Aff(T (C/I)) = Aff(∆) by the above two approximately intertwining diagram, and therefore it is compatible with the map rC and r—that is π# ∞(rC (x)) = r(K0(π)(x)) for any x ∈ K0(C). (Here, we use the facts that r : G → Aff(∆) factors through H and K0(π)(x) ∈ G ⊂ H for x ∈ K0(C).) 13.20. From the construction, the algebra C in 13.19 has an ideal I = lim(In, ϕn,mIn) and therefore is not a simple C∗-algebra. However, one can modify the homomorphism ϕn,n+1 to a map ψn,n+1 to have the following properties: (1) ψn,n+1 ∼h ϕn,n+1 . (2) kψ♯ n,n+1 − ϕ♯ (3) Sp(ψn,n+1θ′ n,n+1k ≤ 1 2n . ) is 1 1 n -dense in Sp(Cn), where θ′1 is the first point in Sp(Fn+1) = {θ′1, θ′2,··· , θ′pn+1}⊂Sp(Cn+1) (see 13.3 for concept of δ-dense). (4) Like ϕn,n+1, for any x ∈ Sp(Cn+1), we have Sp(Fn) ⊂ Sp(ψn,n+1x). (5) Again like ϕn,n+1, the homomorphism ψn,n+1 is injective and furthermore, if X⊂Sp(Cn+1) is δ-dense in Sp(Cn+1), then ∪x∈XSp(ψn,n+1x) is δ-dense in Sp(Cn). 120 / / / / / / / / / / / / / / / / U U / / U U / / U U (See the end of 13.1, 13.2 and 13.3 for the notation used here.) For each n and x ∈ Sp(Cn+2), combining (3) (for ψn,n+1) and (4) (for ψn+1,n+2), we have that Sp(ψn,n+2x) = [ y∈Sp(ψn+1,n+2x) Sp(ψn,n+1y) ⊃ Sp(ψn,n+1θ′ 1 ) (e 13.19) is 1/n-dense in Sp(Cn) Combining with (5), for any n and m ≥ n + 2, and any x ∈ Sp(Cm), Sp(ψn,mx) is m−2 dense in Sp(Cn). Consequently, the new inductive limit B = lim(Cn, ψn,n+1 ) is simple by 13.4. By (1) and (2), 1 (K0(C), K0(C)+, 1C, T (C), rC ) ∼= (K0(B), K0(B)+, 1B, T (B), rB ). Since we will construct the algebra C ∈ N Z0 with general weakly unperforated K0 group (without the torsion free condition) and general K1 group, we will not give detailed construction of the above special case. The following theorem is in §3 of [26] (see Lemma 3.21 and Corollary 3.22 there). Proposition 13.21. Let X and Y be path connected finite CW complexes of dimension at most three, with base point x0 ∈ X and y0 ∈ Y , such that the cohomology groups H 3(X) and H 3(Y ) are finite. Let α0 : K0(C(X)) → K0(C(Y )) be a homomorphism satisfying that α0 is at least 12-large and that α0(K0(C(X))+ \{0})⊂K0(C(Y ))+\{0}, and let α1 : K1(C(X)) → K1(C(Y )) be any homomorphism. Let P ∈ M∞(C(X)) be any non-zero projection and Q ∈ M∞(C(Y )) be a projection with α0([P ]) = [Q] (such projections always exist.) Then there exists a unital homomorphism ϕ : P M∞(C(X))P → QM∞(C(Y ))Q such that ϕ∗0 = α0 and ϕ∗1 = α1, and such that ϕ(P M∞(C0(X \ {x0}))P )⊂QM∞(C0(Y \ {y0}))Q That is, if f ∈ P M∞(C(X))P satisfies f (x0) = 0, then ϕ(f )(y0) = 0. Remark 13.22. In the proof of the above proposition in [26], one reduced the general case to the case that P is rank one trivial projection (P M∞(C(X))P = C(X)). If one further assumes that α0 is at least 13-large and Y 6= {pt}, then the homomorphism ϕ in the proposition can be chosen to be injective. To prove this we choose a surjective homotopy trivial continuous map g : Y → X, which induces an injective homomorphism g∗ : C(X) → C(Y ). Then apply the theorem to the new map (still denoted by α0) α0 := α0 − (g∗)∗0 : K0(C(X)) → K0(C(Y )) which is at least 12-large and α1 to obtain ϕ1 : C(X) → QM∞C(Y )Q. Then ϕ = diag(ϕ1, g∗) : C(X) → (Q ⊕ 1)M∞C(Y )(Q ⊕ 1) has the desired property. The following is perhaps known. Lemma 13.23. Let 0 → E → H → H/E → 0 be a short exact sequence of countable abelian groups with H/E torsion free. And let γ′ 1,2 γ′ 2,3 / H3 γ′ 3,4 / H2 H1 / ··· / H/E be an inductive system with limit H/E such that each Hi is finitely generated free abelian group. Then there are an increasing sequence of finitely generated subgroups E1 ⊂ E2 ⊂ ··· ⊂ En ⊂ ··· ⊂ E with E =S∞i=1 Ei, and an inductive system γ2,3 γ1,2 E1 ⊕ H1 / E2 ⊕ H2 / E3 ⊕ H3 γ3,4 / ··· / H 121 / / / / / / / / with limit H with the following properties: (i) γn,n+1(En) ⊂ En+1 and γn,n+1En is the inclusion from En to En+1. (ii) Let πn+1 : En+1 ⊕ Hn+1 → Hn+1 be the canonical projection, then πn+1 ◦ γn,n+1Hn = (Here, we do not assume γ′n,n+1 to be injective.) γ′n,n+1. Proof. Let E = {ei}∞i=1. We will construct the system (En ⊕ Hn, γn,n+1), inductively. Let us assume we already have En ⊂ E with {e1, e2,··· , en} ⊂ En and the map γn,∞ : En ⊕ Hn → H such that γn,∞En is the inclusion and π ◦ γn,∞Hn = γ′n,∞, where π : H → H/E is the quotient map. Note that γ′n+1,∞(Hn+1) is a finitely generated free abelian subgroup of H/E, one has a lifting map γn+1,∞ : Hn+1 → H such that π ◦ γn+1,∞ = γ′n+1,∞. For each h ∈ Hn, we have γ(h) := γn,∞(h) − γn+1,∞(γ′n,n+1(h)) ∈ E. Let En+1 ⊂ E be the finitely generated subgroup generated by En ∪ {en+1} ∪ γ(Hn) and extend the map γn+1,∞ on En+1 ⊕ Hn+1 by defining it to be inclusion on En+1. And finally let γn,n+1 : En ⊕ Hn → En+1 ⊕ Hn+1 be defined by γn,n+1(e, h) = (e + γ(h), γ′n,n+1(h)) ∈ En+1 ⊕ Hn+1 for each (e, h) ∈ En ⊕ Hn. Evidently, γn,∞ = γn+1,∞ ◦ γn,n+1. 13.24. Now let ((G, G+, u), K, ∆, r) be the one given in 13.6 in general, i.e., G may have torsion and K may not be zero. Let G1 ⊂ Aff(∆) be a dense subgroup with at least three Q-linearly independent elements and H = G ⊕ G1 be as in 13.8. The order unit u ∈ G+ could be regarded as (u, 0) ∈ G ⊕ G1 = H as the order unit of H+. Note that H has torsion Tor(G) = Tor(H) here. Then we have splitting short exact sequence 0 −→ G −→ H −→ H/G(= G′) −→ 0 with H/Tor(H) dimension group. By applying 13.23 to the short exact sequence 0 → Tor(H) → H → H/Tor(H) → 0, we can write H as inductive limit of finitely generated abelian groups γ1,2 / H2 H1 γ2,3 / H3 γ3,4 / ··· / H, j=1H j n with H 1 where Hn = ⊕pn cone should be given by (Hn)+\{0} = (Zpn cone with n = Z⊕ Tor(Hn) and H i + \{0, ..., 0 {z } pn (Hn)+ =(cid:0)(Z+ \ {0} ⊕ Tor(Hn) ∪ {0, 0}(cid:1) ⊕ Zpn−1 + n = Z for all i ≥ 2. Presumingly the positive })⊕ Tor(Hn). But we change it to a smaller + \ {0, ..., 0 with the order unit un =(cid:0)([n, 1], τn), [n, 2], ..., [n, pn](cid:1), where τn ∈ Tor(Hn) and [n, i] are positive integers. Evidently, by simplicity of ordered group H, this modification will not change the positive cone of the limit—in fact, since all the entries of the map c : Zpn → Zpn+1 are strictly positive, any element in (Zpn }) ⊕ Tor(Hn) will be sent into (Z+ \ {0} ⊕ Tor(Hn+1) ⊕ {z } (Z+ \ {0})pn+1−1 which is a subset of H + n+1, by γn,n+1. The map γn,n+1 : Hn → Hn+1 could be represented by the pn+1 × pn matrix of homomorphism c = (cij), where cij : H j n+1. If i > 1, j ≥ 1, then each cij = cij is a positive integer which defines a map Z(= H j n+1) by sending m to cijm (for j > 1) or a map Z ⊕ Tor(Hn) → Z by sending (m, t) ∈ Z ⊕ Tor(Hn) to cijm (note that the i-th component of γn,n+1(Tor(Hn)) is 0 if i > 1). If i = 1, then cij = c1j + Tj, where Tj : H j n+1 and c1j is a positive integer. Since γn,n+1 satisfies γn,n+1(Tor(Hn))⊂Tor(Hn+1), it induces the map n → Tor(Hn+1)⊂H 1 n) → Z(= H j n → H i pn γ′n,n+1 : Hn/Tor(Hn) = Zpn −→ Hn+1/Tor(Hn+1) = Zpn+1. 122 / / / / Then γ′n,n+1 is given by the matrix c = (cij ) of all entries positive integers. By passing to a sub- sequence, we can require cij to be larger than any previously given number which only depends on the construction up to nth step—we will specify how large cij should be in 13.29 below (see (♦♦♦)1 there) n,∞(G) with (Gn)+ = (Hn)+ ∩ Gn. Then the unit un ∈ (Hn)+ is also From Tor(G) = Tor(H) we have Tor(Gn) = Tor(Hn). Furthermore we have the following 13.25. Let Gn = Hn ∩ γ−1 the unit for (Gn)+, since u ∈ G ⊂ H. commutative diagram: 0 G1 H1 γ12G1 γ12 0 G2 H2 H1/G1 γ12 / H2/G2 0 0 ··· ··· ··· 0 / G / H / H/G 0 where γn,n+1 is induced by γn,n+1. Note that the inductive limit of quotient groups H1/G1 → H2/G2 → ··· → H/G has no apparent order structure. Also write the group K (in 13.24) as inductive limit K1 χ12 / / K2 χ23 / / K3 χ34 / / ··· / K , where each Kn is finitely generated. 13.26. Recall from [26] that the finite CW complexes T2,k (or T3,k) is defined to be a 2- dimensional connected finite CW complex with H 2(T2,k) = Z/k and H 1(T2,k) = 0 (or 3- dimensional finite CW complex with H 3(T3,k) = Z/k and H 1(T3,k) = 0 = H 2(T3,k)). (In [26] the spaces are denoted by TII,k and TIII,k.) For each n, there is a space X′n which is of form X′n = S1 ∨ S1 ∨ ··· ∨ S1 ∨ T2,k1 ∨ T2,k2 ∨ ··· ∨ T2,ki ∨ T3,m1 ∨ T3,m2 ∨ ··· ∨ T3,mj n = Z ⊕ Tor(Hn) and K1(C(X′1)) = Kn. Let xn be the base point of X′n with K0(C(X′1)) = H 1 which is a common point of all spaces S1, T2,k, T3,k appeared above in the wedge ∨. And there is a projection Pn ∈ M∞(C(Xn)) such that [Pn] = ([n, 1], τn) ∈ K0(C(Xn)) = Z ⊕ Tor(K0(C(Xn))) where ([n, 1], τn) is the first component of unit un ∈ Hn. We may assume that Pn(xn) = 1M[n,1]. Note that rank(Pn) = [n, 1]. Assume Pn(xn) = 1M[n,1](C), where M[n,1](C) is identified with upper left corner of M∞(C). Define Xn = [0, 1] ∨ X′n with 1 ∈ [0, 1] identified with the base point xn ∈ X′n. We name 0 ∈ [0, 1], by the symbol θ1. So [0, 1] is identified with [θ1, θ1 + 1]. Under this identification, we have Xn = [θ1, θ1 + 1] ∨ X′n. Pn ∈ M∞(C(X′n)) can be extended to 123 / / / / / / / / / / / / / / / a projection, still called Pn ∈ M∞(C(Xn)) by Pn(θ1 + t) = 1[n,1](C) for each t ∈ (0, 1). Now we will also call θ1 the base point of Xn. The old base point of X′n is θ1 + 1. Let Fn = PnM∞(C(Xn))Pn ⊕ pnMi=2 Then Jn is an ideal of Fn with Fn = Fn/Jn = ⊕pn given by π(g, a2, a3, ..., apn ) = (g(θ1), a2, a3, ..., apn ) where g ∈ PnM∞(C(Xn))Pn. Let us denote the spectral of Fn = ⊕pn Jn = {f ∈ PnM∞(C(Xn))Pn : f (θ1) = 0}. M[n,i](C) and let i=1M[n,i](C). i=1 n by θ1, θ2, ..., θpn. Then the map π : Fn → Fn is F i (e 13.20) (e 13.21) 13.27. The map Hn/Tor(Hn)(cid:0) = Zpn(cid:1) −→ Hn/Gn (cid:0) = Zln(cid:1) (induced by the quotient map Hn → Hn/Gn; here we use the fact that Tor(Hn) = Tor(Gn)) can be realized as a difference of two maps b0, b1 : Zpn → Zln corresponding to two ln × pn matrices of strictly positive integer entries b0 = (b0,ij), b1 = (b1,ij). Exactly as in 13.10 (in which we did the special case of torsion free K0 and trivial K1), we can define {n, i} := pnXj=1 b0,ij[n, j] = b1,ij[n, j] . pnXj=1 Fn → En be any homomorphisms such that (e 13.22) Let En = Lln i=1 M{n,j}(C) and let β0, β1 : (β0)∗0 = b0 and (β1)∗0 = b1. Now we define: An =(cid:8)(f, g) ∈ C([0, 1], En) ⊕ Fn; f (0) = β0(π(g)), f (1) = β1(π(g))(cid:9), i=2 M[n,i](C) as defined in (e 13.20) and recall that from where Fn = PnM∞(C(Xn))Pn ⊕Lpn 13.26, the map π : Fn → Fn is given by π(g, a2, a3, ..., apn) = (g(θ1), a2, a3, ..., apn ), where g ∈ PnM∞(C(Xn))Pn. Similar to the construction of Cn (see 13.12), the algebra An depends only on Fn, b0, b1 : K0( Fn)(= K0(Fn)/Tor(K0(Fn))) = Zpn → Zln. That is, once Fn, b0, b1 are specified, the construction of the algebra An is considered to be done. Note that from (e 13.22), β0, β1 are unital homomorphisms and therefore An is a unital algebra. Note that this algebra, in general, is NOT a direct sum of a homogeneous algebra and an algebra in C0. In fact, An ∈ D2. Later we will deal with a nicer special case so that An is a direct sum of a homogeneous C∗-algebra and a C∗-algebra in C0. 13.28. Let In = {(f, g) ∈ An g = 0}. Then An/In = Fn, and we denote the quotient map An → An/In = Fn by π2. Let J′n = {(f, g) ∈ An g ∈ Jn ⊂ Fn} = π−1 2 (Jn). Denote the quotient algebra An/J′n by ¯An and the quotient map An → An/J′n = ¯An by π1. Note that ¯An = A( Fn, En, β0, β1) =(cid:8)(f, g) ∈ C([0, 1], En)⊕ Fn; f (0) = β0((g)) f (1) = β1(π(g))(cid:9).(e 13.23) 124 The homomorphism ϕn,n+1 : An → An+1 to be constructed should satisfy the conditions ϕn,n+1(In) ⊂ In+1 and ϕn,n+1(J′n) ⊂ J′n+1, and therefore it induces two homomorphism ψn,n+1 : An/In = Fn → An+1/In+1 = Fn+1 and ¯ϕn,n+1 : An/J′n = ¯An → An+1/J′n+1 = ¯An+1. Let ¯In , π1(In). Conversely, if two homomorphisms ψn,n+1 : Fn → Fn+1 and ¯ϕn,n+1 : ¯An → ¯An+1 satisfy the following conditions: (a) ψn,n+1(Jn) ⊂ Jn+1 and ¯ϕn,n+1( ¯In) ⊂ ¯In+1, and (b) the homomorphism Fn/Jn = Fn → Fn+1/Jn+1 = Fn+1 induced by ψn,n+1 and the homo- morphism ¯An/ ¯In = Fn → ¯An+1/ ¯In+1 = Fn+1 induced by ¯ϕn,n+1 are the same, = ¯ϕn,n+1Sp Fn then, there is a unique homomorphism ϕn,n+1 : An → An+1 satisfying ϕn,n+1(In) ⊂ In+1, ϕn,n+1(J′n) ⊂ J′n+1, and it induces the homomorphisms ψn,n+1 and ¯ϕn,n+1. Therefore to con- struct ϕn,n+1, we only needs to construct ψn,n+1 and ¯ϕn,n+1 separately with the same restriction map ψn,n+1Sp Fn Note that ¯An is as same as Cn in 13.10–13.19 with Fn and ¯In = π1(In) in place of Fn and In in 13.10–13.19. Therefore the construction of ¯ϕn,n+1 can be carried out as in 13.10–13.19, with the map Fn → Fn+1 being given by the matrix c as in 13.24 above—of course, we need to assume that the corresponding maps b0, b1 in this case (see 13.27) satisfy (♦♦). So, in what follows, we will focus on the construction of ψn,n+1. But before the construction, we will specify the conditions to be used in the future. : Fn → Fn+1 13.29. The construction of An+1 and ϕn,n+1 will be done by induction. Suppose that, we already have the part of the inductive sequence: ϕ1,2 / A2 ϕ2,3 / A3 A1 ϕ3,4 ϕn−1,n / An, / ··· satisfies the following conditions: for each i = 1, 2,··· n − 1, (a) ϕi,i+1(Ii) ⊂ Ii+1 and ϕi,i+1(J′i) ⊂ J′i+1 and therefore induces two homomorphisms ψi,i+1 : Fi → Fi+1 and ¯ϕi,i+1 : ¯Ai → ¯Ai+1; (b) all the homomorphism ϕi,i+1, ψi,i+1 and ¯ϕi,i+1 are injective, in particular ψi,i+1Ji : Ji → Ji+1 is injective; (c) the induced map ¯ϕi,i+1 : ¯Ai → ¯Ai+1 satisfies all the conditions (1)–(6) of 13.13 with i in place of n, with ¯Ai in place of Cn (and of course naturally with ¯Ai+1 in place of Cn+1), and with Gi/Tor(Gi) (or Gi+1/Tor(Gi+1)) and Hi/Tor(Hi) (or Hi+1/Tor(Hi+1)) in place of Gn (or Gn+1) and Hn (or Hn+1), respectively; (d) the matrices b′0 and b′1 for each Ai+1 satisfy the condition (♦♦) in (e 13.15) with i + 1 in place of n + 1 (this condition will be strengthened below). Now we need to construct An+1 and late on the homomorphism ϕn,n+1. Choose a finite set Y ⊂Xn \ {θ1} (where θ1 is the base point of Xn) satisfying that for n -dense in Xi. This can be done since the corresponding i+1) is injective for each i. Let us denote t ∈ (0, 1)j ⊂ : Ji(⊂ F 1 n)(cid:1) by t(j) to distinguish spectrum from different direct summands of C([0, 1], En). each i < n, Sy∈Y Sp(ϕi,ny ) is 1 Sp(cid:0)C([0, 1], Ej i ) → Ji+1(⊂ F 1 map ψi,i+1Ji Let T ⊂ Sp(An) be defined by k n T =(cid:26)( )(j); j = 1, 2,··· , ln; k = 1, 2,··· , n − 1(cid:27) . 125 / / / / (♦♦♦)2 b0,il =Ppn+1 b1,il =Ppn+1 k=1 b′0,ik · ckl > 22n(cid:16)Pln k=1 b′1,ik · ckl > 22n(cid:16)Pln k=1(dik + 2) · {n, k} k=1(dik + 2) · {n, k} +(L1 + (n − 1)Pln +(L1 + (n − 1)Pln k=1 b0,kl) · b′0,i2(cid:17) and k=1 b0,kl) · b′1,i2(cid:17) ,  Let Y = {y1, y2, ..., yL1} ⊂ Xn and let L = ln · (n − 1) + L1 = #(T ∪ Y ). Set An =(cid:8)(f, g) ∈ C([0, 1], En)⊕ Fn; f (0) = β0(π(g)), f (1) = β1(π(g))(cid:9), with (β0)∗0 = b0 = First we need that the matrix c = (cij) given by the map γ′n,n+1 : Hn/Tor(Hn) = Zpn −→ (b0,ij) and (β1)∗0 = b1 = (b1,ij). Put M = max{b0,ij; i = 1, 2,··· , pn; j = 1, 2,··· , ln}. Hn+1/Tor(Hn+1) = Zpn+1 satisfies cij > 13 · 22n · M L (♦♦♦)1 This can be done by increasing n + 1 to large enough m, and renaming Hm as Hn+1 and γ′n,m as γ′n,n+1 (see the end of 13.24). Let An+1 = (cid:8)(f, g) ∈ C([0, 1], En+1) ⊕ Fn+1 : f (0) = β′0(π(g)), f (1) = β′1(π(g))(cid:9) be con- structed as in 13.27. In this construction, we will choose b′0 = (b′0,ij) and b′1 = (b′1,ij) (corre- sponding to β′0 and β′1) to satisfy the strengthened condition (stronger than (♦♦)) for all i, j. for each l ∈ {1, 2, ..., pn} and i ∈ {1, 2, ..., ln+1}, where b0 = b′0·c = (b0,ij) and b1 = b′1·c = (b1,ij). To make the above inequalities hold, we only need to increase the entries of the third columns of b′0 and b′1 by adding same big positive number to make it much larger than the second column of the matrices (see the end of 13.10). With the above choice of An+1 satisfying the conditions (♦♦♦)1 and (♦♦♦)2, we will begin the construction of ϕn,n+1 by constructing ψn,n+1 first. 13.30. Recall that K0(Fn) = (Hn, un), K0(Fn+1) = (Hn+1, un+1) and the map c = (cij) : Hn → Hn+1 is as in 13.24. Assume that cij > 13 for any ij (which is a consequence of (♦♦♦)1). We shall define the unital homomorphism ψn,n+1 : Fn → Fn+1 to satisfy the following conditions: (1) (ψn,n+1)∗0 = γn,n+1 : K0(Fn)(= Hn) −→ K0(Fn+1)(= Hn+1) and (ψn,n+1)∗1 = χn,n+1 : K1(Fn)(= Kn) −→ K1(Fn+1)(= Kn+1); (2) ψn,n+1 (Jn) ⊂ Jn+1, and the map ψn,n+1 : Fn → Fn+1, induced by ψn,n+1 satisfies ( ψn,n+1)∗0 = c = (cij) : K0( Fn)(= Zpn) → ( Fn+1)∗0(= Zpn+1). Since a homomorphism from a finite dimensional C-algebra to another is completely deter- mined by its K-theory map up to unitary equivalence, the map ψn,n+1 : Fn → Fn+1 is completely determined by the condition ( ψn,n+1 )∗0 = c = (cij) up to unitary equivalence and such a map is necessarily unital (since K-theory map c keeps the order unit). n,n+1 is ψi,j n,n+1. n,n+1(Jn) = 0. n+1 for j ≥ 2) n /Jn and is completely determined by ψn,n+1. Note that, for any i and n+1, and, consequently, ψi,j n+1 = 0 for any j ≥ 2, one has that ψ1,j n → F j n+1 = F j n+1 for j ≥ 2 (note that F j If i ≥ 2 and j ≥ 2, then F i n+1 = F j n and F j n = F i Since ψn,n+1(Jn) ⊂ Jn+1 and Jn+1 ∩ F j n,n+1 : F 1 Therefore each component ψ1,j factors through F 1 any j ≥ 2, ψi,j project to the first summand. Therefore one only needs to define ψ−,1 n = F 1 n,n+1 is completely determined by ψi,j n+1, the map from Fn to Fn+1, then n,n+1 : Fn → F 1 n,n+1. 126 First one can find mutually orthogonal projections Q1, Q2, ..., Qpn such that Q1 + Q2 + ··· + Qpn = Pn+1 n n,n+1([1F i and γi,1 is rank(Qi)/rank(1F i ψi,1 n → QiF 1 n,n+1 : F i χn,n+1 : K1(F 1 Condition (2). n ]) = [Qi] ∈ K0(Pn+1M∞(C(Xn+1))Pn+1). Since cij > 13 for all i, j—that ) = c1,i > 13, by 13.21 (and 13.22), there are unital homomorphisms n+1) and n) = 0 for i ≥ 2) and satisfy n+1Qi which realize the K-theory map γi,1 n+1)(= Kn+1) (note that K1(F i n )(= Kn) → K1(F 1 n) → K0(F 1 n,n+1 : K0(F i 13.31. homomorphism ψi,1 the simple form described below. n,n+1 : F i n → QiF 1 For the convenience of later construction, we need the projection Qi and we need the n+1Qi not only satisfies the conditions in 13.30, but also has n+1 and denoted by ψi : To simplify the notation, we denote ψ−,1 n → QiF 1 F i n+1Qi the corresponding partial map. Note that rank(Qi) = c1i[n, i]. We may require that the projection Qi (for any i ≥ 2, the case i = 1 will be discussed later) satisfy the following condition: n → F 1 i=1 F i n,n+1 by ψ : Lpn (a) Qi[θ′ 1+1] is of the form 1,θ′ diag(0c11[n,1], 0c12[n,2], ..., 0c1 i−1[n,i−1], 1c1i[n,i], 0c1 i+1[n,i+1], ..., 0c1pn [n,pn]), 1,θ′ 1+1] is identified with 1[n+1,1] ∈ M[n+1,1](C[θ′1, θ′1 + 1]) ⊂ M∞(C[θ′1, θ′1 + 1]). when Pn+1[θ′ (In the above, recall that [θ′1, θ′1 + 1] is identified with the interval [0, 1] in Xn+1 = [0, 1] ∨ X′n+1, and θ′1 is the first element in Sp( Fn+1) = {θ′1, θ′2, ..., θ′pn+1}. Here we reserve {θ1, θ2, ..., θpn} for Sp( Fn). ) Fix i ≥ 2 (the case of i = 1 will be discussed later). Let {eij} be the matrix unit of n = M[n,i](C) and let q = ψ1(e11). It is well known that QiM∞(C(Xn+1))Qi can be identified F i with qM∞(C(Xn+1))q ⊗ M[n,i](C) so that the homomorphism ψi is given by ψi(cid:0)(aij)(cid:1) = q ⊗ (aij) ∈ qM∞(C(Xn+1))q ⊗ M[n,i](C). (e 13.24) Note that rank(q) = c1i as rank(Qi) = c1i[n, i]. Denote c1i − 1 by d. We can write q = q1 + q2 + ··· + qd + p, where q1, q2,··· , qd are mutually equivalent trivial rank 1 projections and p is a (possible nontrivial) rank 1 projection. Under the identification Qi = q ⊗ 1M[n,i], we denote that qj = qj ⊗ 1M[n,i] and p = p ⊗ 1M[n,i]. From the definition of ψi (see (e 13.24) above), we know that ψi(F i n) commutes with q1, q2, ..., qd, and p. We can further require the homomorphisms ψi (i ≥ 2) satisfy q2[θ′ QiM∞(C[θ′1, θ′1 + 1])Qi identified with Mci1[n,i](C[θ′1, θ′1 + 1]). That is 1+1], 1+1] are diagonal matrix with 1[n,i] in the correct place when the above projections q1, q2,··· , qd, and p for ψi can be chosen such that q1[θ′ 1+1], and p[θ′ 1+1] ,..., qd[θ′ (b) 1,θ′ 1,θ′ 1,θ′ 1,θ′ qj = diag(0[n,i], ..., 0[n,i] , 1[n,i], 0[n,i], ..., 0[n,i]) and p = diag(0[n,i], ..., 0[n,i] , 1[n,i]). j−1 {z } d {z } Lemma 13.32. Let i ≥ 2 and ψi : F i n+1Qi be as in 13.31 above. In particular, the projection Qi satisfies Condition (a) and the homomorphism ψi satisfies Condition (b) above. Suppose m ≤ d = c1i − 1. Let Λ : QiF 1 n+1Qi) be the amplification defined by Λ(a) = a ⊗ 1m. There is a projection Ri ∈ Mm(QiF 1 n+1Qi) satisfying the following conditions: (ii) Ri(θ′1) = Qi(θ′1) ⊗(cid:18) 1m−1 0 {z m−1 quently, rank(Ri) = ci1(m − 1)[n, i] = (d + 1)(m − 1)[n, i]. n → QiF 1 n+1Qi → Mm(QiF 1 0 (cid:19) = diag(Qi(θ′1), ..., Qi(θ′1) , 0)) ∈ Mm(F 1 } (i) Ri commutes with Λ(ψi(F i n+1θ′ ). Conse- n)). 0 1 127 Let π : Mm(QiF 1 Then π takes RiMm(QiF 1 (see (ii) above). n+1Qi) → Mm(Qi(θ′1) F 1 n+1Qi)Ri onto Mm−1(Qi(θ′1) F 1 n+1Qi(θ′1)) be the map induced by π : F 1 n+1Qi(θ′1)) ⊂ Mm(Qi(θ′1) F 1 n+1 → F 1 n+1. n+1Qi(θ′1)) (iii) There is an inclusion unital homomorphism ι : Mm−1(Qi(θ′1) F 1 n+1Qi(θ′1)) ֒→ RiMm(QiF 1 1)) and such that Ri(Λ(ψi(F i n+1Qi)Ri such that π ◦ ι = idMm−1(Qi(θ′ Proof. In the proof of this lemma, i ≥ 2 is fixed. So the notation q, q1, q2, ..., qd, p, r and Λ1 are for this fixed i and only kept the meaning in this proof (later on, we will use them for the proof of similar lemma for the case i = 1 in which we will work on the corner Mm(Q1F 1 n)))Ri ⊂ Image(ι). n+1Qi(θ′ 1) F 1 The homomorphism Λ◦ψi : F i n+1Qi) can be regarded as Λ1⊗id[n,i], Note that q = q1 + q2 +··· + qd + p with {qi} being mutually equivalent rank one projections. n+1q ⊗ n+1q) is the unital homomorphism given by Λ1(c) = c · (q ⊗ 1m). n+1q) = qF 1 1+1] is also a trivial rank one projection. Let r ∈ Mm(qF 1 where Λ1 : C → Mm(qF 1 Furthermore, p[θ′ Mm(C) be defined as below. n = M[n,i](C) → Mm(QiF 1 1,θ′ n+1Q1)). 0 r(θ′1) = q(θ′1) ⊗(cid:18) 1m−1 0 0 (cid:19) = (q1(θ′1) + q2(θ′1) + ··· + qd(θ′1) + p(θ′1)) ⊗(cid:18) 1m−1 0 0 (cid:19) , r(θ′1 + 1) = (q1(θ′1 + 1) + q2(θ′1 + 1) + ··· + qm−1(θ′1 + 1)) ⊗ 1m + 0 (cid:19) , + (qm(θ′1 + 1) + ··· + qd(θ′1 + 1)) ⊗(cid:18) 1m−1 0 0 0 and for x ∈ X′n+1 ⊂ Xn+1, r(x) = (q1(x) + q2(x) + ··· + qm−1(x)) ⊗ 1m + (qm(x) + ··· + qd(x)) ⊗(cid:18) 1m−1 0 0 (cid:19) . 0 In the above, between θ′1 and θ′1 + 1, r(t) can be defined to be any continuous path connecting the projections r(θ′1) and r(θ′1 + 1), both of rank (d + 1)(m− 1) = (m− 1)m + (d− m + 1)(m− 1). (Note that all qi(t) and p(t) are constant on [θ′1, θ′1 + 1].) n+1Qi) with Mm(qF i Let Ri = r⊗1[n,i], under the identification of Mm(QiF 1 n+1q)⊗1[n,i]. Since n+1q) sends C to the center of Mm(qF i Λ1 : C → Mm(qF i n+1q), we have that r commutes with n)) as Λ ◦ ψi = Λ1 ⊗ id[n,i]. That Λ1(C) and consequently, Ri = r ⊗ 1[n,i] commutes with Λ(ψi(F i is, condition (i) holds. Condition (ii) follows from definition of r(θ′1) and Ri(θ′1) = r(θ′1) ⊗ 1[n,i]. n+1q) = Mm(qM∞(C(Xn+1))q) is a trivial projection of rank (d + 1)(m − 1) and r(θ′1) = q(θ′1) ⊗ 1m−1. One identifies Note that r ∈ Mm(qF 1 r(θ′1)Mm(q(θ′1) F 1 n+1q(θ′1))r(θ′1) = Mm−1(q(θ′1) F 1 n+1q(θ′1)) ∼= M(d+1)(m−1)(C). ij, 1 ≤ i, j ≤ (d + 1)(m − 1) be the matrix units for M(d+1)(m−1). Since r is a trivial n+1q)r, 1 ≤ i, j ≤ (d + 1)(m − 1) with rij(θ′1) = r0 n+1q)r ∼= M(d+1)(m−1)(C(Xn+1)). Here by Let r0 projection, one can construct rij ∈ rMm(qF 1 serving as matrix units for M(d+1)(m−1) ⊂ rMm(qF 1 matrix units, we mean rijrkl = δjkril and r =P(d+1)(m−1) ι1 : Mm−1(q(θ′1) F 1 ij) = rij. Finally define ι = ι1 ⊗ id[n,i], using the identification Ri = r ⊗ 1[n,i] and n+1q(θ′1))r(θ′1)) ֒→ rMm(qF 1 n+1q(θ′1))(∼= r(θ′1)Mm(q(θ′1) F 1 rii. We can define n+1q)r i=1 ij by ι1(r0 Qi = q ⊗ 1[n,i]. Then (iii) follows. 128 13.33. Now we continue the discussion of 13.31 for homomorphism ψ1 : F 1 n+1Q1. We know that rank(Q1) = c11[n, 1], where [n, 1] = rank(Pn) for F 1 n = PnM∞(C(Xn))Pn. Note that c11 > 13. Denote d = c11 − 13. We can assume that the projection Q1 and the homomorphism ψ1 : F 1 (a) Q1 = 1d[n,i] ⊕ Q := Q′ ⊕ Q ∈ M∞(C(Xn+1)) such that Q[θ′ 1,θ′ n+1Q1 satisfy the following conditions 1+1] = 113[n,i] (but in the n → Q1F 1 n → Q1F 1 lower right corner of Q1[θ′ 1,θ′ 1+1]); (b) ψ1 : F 1 n → Q1F 1 n → Md[n,i](C(Xn+1)) = Q′M∞(C(Xn+1))Q′ and ψ2 : F 1 ψ1 : F 1 described below: n+1Q1 is decomposed into two parts ψ1 = ψ1 ⊕ ψ2 with n → QM∞(C(Xn+1)) Q (b1) the unital homomorphism ψ1 : F 1 defined by ψ1(f ) = diag(f (θ1), f (θ1), ..., f (θ1) n → Md[n,i](C(Xn+1)) = Q′M∞(C(Xn+1))Q′ is {z n → QM∞(C(Xn+1)) Q is a homomorphism satisfy- ing (ψ2)∗0 = c11 − d = c11 − d + T1 (where T1 : H 1 is as in 13.24) and (ψ2)∗1 = χn,n+1 : K1(Fn)(= Kn) → K1(Fn+1)(= Kn+1) (such ψ2 exists because of 13.21); n )) → Tor(Hn+1) ⊂ H 1 ) as a constant function on Xn+1; n(= K0(F 1 } (b2) the unital homomorphism ψ2 : F 1 n+1 d (b3) furthermore, ψ2 is injective (see 13.22) and the definition of ψ2[θ′ 1,θ′ 1+1] is given as below: for t ∈ [0, 1 2 ], ψ2(f )(θ′1 + t) = diag(f (θ1), f (θ1), ..., f (θ1) ) and for t ∈ [ 1 2 , 1], ψ2(f )(θ′1 + t) = diag(f (θ1 + 2t), f (θ1 + 2t), ..., f (θ1 + 2t) ). 13 {z {z 13 } } Here, f (θ1 + s) ∈ Pn(θ1 + s)M∞(C)Pn(θ1 + s) is regarded as an [n, 1] × [n, 1] matrix for each s ∈ [0, 1] by using the fact Pn[θ1,θ1+1] = 1[n,1]. Let us remark that in the decomposition ψ1 = ψ1 ⊕ ψ2 : F 1 n+1, the first part ψ1 : F 1 F 1 M[n,1](C), and the restriction ψ1[θ′ n → Md[n,i](C(Xn+1))(= Q′M∞(C(Xn+1))Q′ factors through F 1 n → Q1M∞(C(Xn+1))Q1 ⊂ n = 2 ] also factors through F 1 n , as 1+ 1 1,θ′ ψ1(f )(x) = diag(f (θ1), ..., f (θ1) ) for any x ∈ [θ′1, θ′1 + 1 2 ]. d+13 {z } n → Q1F 1 n+1Q1 satisfy Conditions (a) and (b) n+1Q1 → n+1Q1) be amplification defined by Λ(a) = a⊗1m. There is a projection R1 ∈ Mm(Q1F 1 Lemma 13.34. Suppose that Q1 and ψ1 : F 1 (including (b1),(b2) and (b3)). Suppose that 13m ≤ d = c11 − 13. Let Λ : Q1F 1 Mm(Q1F 1 satisfying the following conditions: (i) R1 commutes with Λ(ψ1(F 1 n )). n+1Q1) (ii) R1(θ′1) = Q1(θ′1)⊗(cid:18) 1m−1 0 Let π : Mm(Q1F 1 m−1 quently, rank(R1) = c11(m − 1)[n, 1] = (d + 13)(m − 1)[n, 1]. takes R1Mm(Q1F 1 above). n+1Q1)R1 onto Mm−1(Q1(θ′1) F 1 n+1Q1) → Mm(Q1(θ′1) F 1 0 (cid:19) = diag(Q1(θ′1),··· , Q1(θ′1) } n+1Q1(θ′1)) be induced by π : F 1 n+1Q1(θ′1)) ⊂ Mm(Q1(θ′1) F 1 {z 0 , 0)) ∈ Mm(F 1 n+1θ′ 1 ). Conse- n+1 → F 1 n+1, then π n+1Q1(θ′1)) (see (ii) 129 (iii) There is an inclusion unital homomorphism ι : Mm−1(Q1(θ′1) F 1 such that π ◦ ι = idMm−1(Q1(θ′ 1) F 1 n+1Q1(θ′ n+1Q1(θ′1)) ֒→ R1Mm(Q1F 1 1)) and such that R1(Λ(ψ1(F 1 n+1Q1)R1 n )))R1 ⊂ Image(ι). The notation Λ, d, m in the above lemma, and q, q1, q2, ...qd, p, r and Λ1 in the proof below, are also used in Lemma 13.32 and its proof for the case i ≥ 2 (comparing with i = 1 here). Since they are used for the same purpose, we choose the same notation. Proof. The map ψ1 : F 1 n π / F 1 n / Md[n,1](C(Xn+1)) = Q′M∞(C(Xn+1))Q′ (where Q′ = 1d[n,1]) can be written as (Λ1⊗ id[n,1])◦π, where Λ1 : C → Md(C(Xn+1)) is the map sending c ∈ C to c · 1d. We write Λ1(1) := q′ = q1 + q2 + ··· + qd, with each qi a trivial constant projection of rank 1. Here q′ is constant subprojection of Q′ with Q′ = q′ ⊗ 1[n,1]. Consider the map ψ2 := ψ2[θ′ 2 ] : 1,θ′ F 1 n → Q1F 1 n+1Q1[θ′ 2 ]. As pointed out in 13.33, ψ2 has the factorization 2 ] and ψ1 := ψ1[θ′ 1,θ′ 2 ] = (ψ1 + ψ2)[θ′ 2 ] : F 1 1+ 1 n → QF 1 Q[θ′ 1,θ′ 1+ 1 1+ 1 1+ 1 1+ 1 1,θ′ 1,θ′ n+1 F 1 n π / / F 1 n / M13[n,1](C[θ′1, θ′1 + 1 2 ]). Hence ψ1 has the factorization F 1 n π / / F 1 n / M(d+13)[n,1](C[θ′1, θ′1 + 1 2 ]). The map ψ1 can be written as (Λ2 ⊗ id[n,1])◦ π, where Λ2 : C → Md+13(C[θ′1, θ′1 + 1 2 ]) is the map defined by sending c ∈ C to c · 1d+13. We write Λ2(1) := q = q1 + q2 + ··· + qd + p with each qi being the restriction of qi appeared in the definition of Λ1(1) on [θ′1, θ′1 + 1 2 ], and p is rank 13 trivial projection. Here q is a constant projection on [θ′1, θ′1 + 1 2 ] = q ⊗ 1[n,1]. 2 ] and Q1[θ′ 1,θ′ Let r ∈ Mm(qF 1 r(θ′1) = q(θ′1) ⊗(cid:18) 1m−1 0 for t ∈ [ 1 0 (cid:19) = (q1(θ′1) + q2(θ′1) + ··· + qd(θ′1) + p(θ′1)) ⊗(cid:18) 1m−1 0 0 (cid:19) ; n+1q ⊗ Mm(C) be defined as below. n+1q) = qF 1 2 , 1], 1+ 1 0 0 r(θ′1 + t) =(cid:16)q1(θ′1 + t) + q2(θ′1 + t) + ··· + q13(m−1)(θ′1 + t)(cid:17) ⊗ 1m+ +(cid:16)q13(m−1)+1(θ′1 + t) + ··· + qd(θ′1 + t)(cid:17) ⊗(cid:18) 1m−1 0 0 (cid:19) ; 0 and for x ∈ X′n+1 ⊂ Xn+1, r(x) = (q1(x) + q2(x) + ··· + q13(m−1)(x)) ⊗ 1m + (q13(m−1)+1(x) + ··· + qd(x)) ⊗(cid:18) 1m−1 0 0 (cid:19) . 0 In the above, between θ′1 and θ′1 + 1 2 , r(t) can be defined to be any continuous path connecting the projections r(θ′1) and r(θ′1+ 1 2 ), both of rank (d+13)(m−1) = 13(m−1)m+(d−13(m−1))(m−1). (Note that all qi(x) are constant on x ∈ Xn+1 = [θ′1, θ′1 + 1] ∨ X′n+1 and p(t) is constant for t ∈ [θ′1, θ′1 + 1].) Note that for x ∈ [θ′1 + 1 2 ) which is 2 , θ′1 + 1]∨ X′n+1, r(x) has same form as r(θ′1 + 1 130 / / / / constant sub-projection of constant projection q′ ⊗ 1m. We are going to define R1 to be r⊗ 1[n,i] under certain identification. Note that the projection Q1 is identified with q ⊗ 1[n,1] only on interval [θ′1, θ′1+1] so the definition of R1 will be divided into two parts. For the part on [θ′1, θ′1+ 1 2 ], we use the identification of Q1 with q ⊗ 1[n,1], and for the part that x ∈ [θ′1 + 1 2 , θ′1 + 1] ∨ X′n+1, we use the identification of Q′ = 1d[n,1] with q′ ⊗ 1m (of course, we use the fact that r is sub- projection of q′ on this part). This is the only difference between the proof of this lemma and that of Lemma 13.32. The definition of ι : Mm−1(Q1(θ′1) F 1 n+1Q1)R1 and verification that ι and R satisfy the conditions are exact as same as the proof of 13.32, with (d + 1)(m − 1) replaced by (d + 13)(m − 1). n+1Q1(θ′1)) ֒→ R1Mm(Q1F 1 Combining 13.32 and 13.34, we have the following theorem which is used to conclude the algebra A (will be constructed later) satisfies that A ⊗ U is in B0. Theorem 13.35. Suppose that 1 < m ≤ min{ c11−13 F 1 n+1 be the composition 13 , c12 − 1, c13 − 1, ..., c1pn − 1}. Let ψ : Fn → ψn,n+1 / Fn+1 Fn π1 / / F 1 n+1. (The map π1 is the quotient map to the first block.) Let Λ : F 1 fication by Λ(a) = a ⊗ 1m. There is a projection R ∈ Mm(F 1 n+1) = F 1 is an inclusion unital homomorphism ι : Mm−1( F 1 satisfying the following conditions: (i) R commutes with Λ(ψ(Fn)). (ii) R(θ′1) = 1 F 1 n+1 ⊗(cid:18) 1m−1 0 0 (cid:19) . Consequently, the map π : F 1 0 n+1 → Mm(F 1 n+1) = F 1 n+1 ⊗ Mm−1(C) ֒→ RMm(F 1 n+1) be the ampli- n+1 ⊗ Mm(C) and there n+1)R, n+1 → F 1 n+1, takes RMm(F 1 n+1)R onto Mm−1( F 1 n+1) . (iii) π ◦ ι = idMm−1( F 1 n+1), and R(Λ(ψ(Fn)))R ⊂ Image(ι). i=1 Ri ∈ Mm(F 1 n+1Q1) is given in Lemma 13.34 n+1), where R1 ∈ Mm(Q1F 1 n+1Qi) (for i ≥ 2) are given in Lemma 13.32. Then the theorem follows. Proof. Choose R =Lpn and Ri ∈ Mm(QiF 1 13.36. Set F = lim−→(Fn, ψn,n+1). Since ψn,n+1(Jn) ⊂ Jn+1, this procedure also gives an inductive limit of quotient algebras F = lim( Fn, ψn,n+1), where Fn = Fn/Jn. Evidently, F is an AF algebra with K0( F ) = H/Tor(H). Theorem 13.37. If the matrix c = (cij) of γn,n+1 : Hn/Tor(Hn) → Hn+1/Tor(Hn+1) satisfies cij > 13 · 22n for each i, j (which is true by (♦♦♦)1 in 13.29), then tracial state space T (F ) of F is T ( F ) = ∆—that is, the Elliott invariant of F is ((K0(F ), K0(F )+, [1F ]), K1(F ), T (F ), rF ) ∼= ((H, H+, u), K, ∆, r). Proof. Note that ψn,n+1 satisfies (1) in 13.30, and consequently, ((K0(F ), K0(F )+, [1F ]), K1(F )) ∼= ((H, H+, u), K, ). Let πn : Fn → Fn be the quotient map. Then π♯ n : Aff(T (Fn)) = C(Xn, R) ⊕ Rpn−1 → Aff(T ( Fn)) = Rpn is given by πn(g, h2, h3, ..., hpn ) = (g(θ1), h2, h3, ..., hpn ). Define Γn : Aff(T ( Fn)) = Rpn → Aff(T (Fn)) = C(Xn, R)⊕Rpn−1 to be the right inverse of π♯ n given by Γn(h1, h2, h3, ..., hpn ) = (g, h2, h3, ..., hpn ) with g being constant function g(x) = h1 for all x ∈ Xn. Then with the con- dition cij > 13 · 22n, we can prove the following claim: Claim: For any f ∈ Aff(T (Fn)) with kfk ≤ 1 and f′ = ψ♯ n+1(f′) − f′k < n,n+1(f ) ∈ Aff(T (Fn+1)), we have 2 22n . kΓn+1 ◦ π♯ 131 / Proof of claim: Write f = (g, h2, ..., hpn ) and f′ = (g′, h′2, ..., h′pn+1 ). Then Γn+1 ◦ π♯ n+1(f′) = (g′′, h′2, ..., h′pn+1 ) with g′′(x) = g′(θ′1) for all x ∈ Xn+1. Recall that ψi,1 n,n+1 is denoted by Q as in n → QF 1 n+1Q′ and ψ2 : F 1 ψi : F i 13.31 and 13.33. Note that n+1Qi and ψ1 = ψ1 + ψ2 with ψ1 : F 1 n → QiF 1 n → Q′F 1 n+1 rank(Qi) rank(Pn+1) = Hence j=1 cj,1 ci,1Ppn Ppn g′ = c1,1 − 13 j=1 cj,1 , rank(Q′) rank(Pn+1) = c1,1 − 13 j=1 cj,1 Ppn ψ# 1 (g) + ψ# 2 (g) + 13 j=1 cj,1 Ppn Also from the construction in 13.31 and 13.33, we know that ψ# constant. So we have g′(x) − g′(θ′1) ≤ and the claim follows. < 2 22n , 2 × 13 j=1 cj,1 Ppn , and rank( Q) rank(Pn+1) = . 13 j=1 cj,1 Ppn pnXi=2 ci,1Ppn (ψi)#(hi). j=1 cj,1 1 (g) and (ψi)#(hi) (i ≥ 2) are The proof of the theorem is then as same as that of 13.19 with 13.17 replaced by the above claim. Namely, one can prove the following approximately intertwining diagram replacing the diagram there. Aff(T (F1)) π# 1 Γ1 Aff(T ( F1)) ψ# 1,2 ψ# 1,2 / Aff(T (F2)) π# 2 Γ2 Aff(T ( F2)) ψ# 2,3 ψ# 2,3 Aff(T (F3)) π# 3 Γ3 Aff(T ( F3)) ··· Aff(T (F )) ··· Aff(T ( F )) , 13.38. where π# is induced by the quotient map πi : Fi → Fi. Exactly the same as the end of the proof i of Lemma 13.19, one can get the isomorphism between ((K0(F ), K0(F )+, [1F ]), K1(F ), T (F ), rF ) and ((H, H+, u), K, ∆, r) including the compatibility (note that the quotient map Hi = K0(Fi) → Hi/Tor(Hi) = K0( Fi) is also induced by quotient map πi : Fi → Fi). Let An =n(f, g) ∈ C([0, 1], En) ⊕ Fn; f (0) = β0π(g), f (1) = β1π(g)o be as in 13.27, where β0, β1 : Fn → En are two unital homomorphisms, and π is as in the end of 13.26. Let An+1 =n(f, g) ∈ C([0, 1], En+1)⊕ Fn+1; f (0) = β′0π(g), f (1) = β′1π(g)o be defined with unital homomorphisms β′0, β′1 : Fn+1 → En+1. Note that An+1 satisfies the conditions (♦♦♦)1 and (♦♦♦)2 in 13.29. Recall that Jn ⊂ F 1 n=1, resp.) is the ideal of elements vanishing on θ1 ∈ Sp( F 1 n ) ⊂ Sp(F 1 Let In (or In+1), J′n (or J′n+1), Fn = An/In (or Fn+1 = An+1/In+1), and ¯An = An/J′n (or n (and Jn+1 ⊂ F 1 n+1) ⊂ Sp(F 1 n+1) respectively). n ) (and θ′1 ∈ Sp( F 1 ¯An+1 = An+1/J′n+1) be as in 13.28. We still use the matrix b0 = (b0,ij) (or b1 = (b1,ij) ) to denote (β0)∗0 (or (β1)∗0) as before. Also b′0 = (b′0,ij) = (β′0)∗0, b′1 = (b′1,ij) = (β′1)∗0 as in 13.14. We still use c = (cij)pn+1×pn to denote the map Hn/Tor(Hn) → Hn+1/Tor(Hn+1) and d = (dij)ln+1×ln to denote the map Hn/Gn → Hn+1/Gn+1. It follows that, as in 13.13 and 13.14, if b′0 and b′1 satisfy the condition (♦) (see e 13.3) which is weaker than (♦♦♦)2, then one can use ψn,n+1 : Fn → Fn+1 to define ¯ϕn,n+1 : ¯An → ¯An+1 (exactly as Cn → Cn+1 there) which satisfies ¯ϕn,n+1(In) ⊂ In+1. From the construction, we know that ( ¯ϕn,n+1 )∗,0 = γ′n,n+1Gn/Tor(Gn) and γ′n,n+1Gn/Tor(Gn) : K0( ¯An) = Gn/(Tor(Gn)) → K0( ¯An+1) = Gn+1/(Tor(Gn+1)), (e 13.25) 132 / / / / / / / T T / / T T / / T T where γ′n,n+1 : Hn/(Tor(Hn)) → Hn+1/(Tor(Hn+1)) is induce by γn,n+1 as in 13.24. Let π1 : An → An/Jn = ¯An (or An+1 → ¯An+1) and π2 : An → An/In = Fn (or An+1 → Fn+1) be the quotient maps as in 13.28. Then we can combine the above definition of ¯ϕn,n+1 : ¯An → ¯An+1 and ψn,n+1 : Fn → Fn+1 to define ϕn,n+1 : An → An+1, as below. Let f ∈ An and x ∈ Sp(An+1). If x ∈ Sp( ¯An+1), then (ϕn,n+1 (f ))(x) = ¯ϕn,n+1(π1(f )(x)); (e 13.26) and if x ∈ Sp(Fn+1), then (ϕn,n+1 (f ))(x) = ψn,n+1(π2(f )(x)). (e 13.27) Note that for x ∈ Sp(Fn+1) ∩ Sp( ¯An+1) = Sp( Fn+1), ¯ϕn,n+1(π1(f )(x)) = ψn,n+1(π2(f )(x)) = ψn,n+1(π1(π2(f ))), where π1(π2(f )) = π2(π1(f )) ∈ Fn. By (e 13.25), (e 13.26), (e 13.27) above and (1) of 13.30, one has In this way we define an inductive limit (ϕn,n+1 )∗,0 = γn,n+1Gn and (ϕn,n+1)∗,1 = χn,n+1. A1 −→ A2 −→ A3 −→ ··· −→ An −→ ··· −→ A with(cid:0)K0(A), K0(A)+, 1A(cid:1) = (G, G+, u) and K1(A) = K. Similar to 13.15 (see (∗) and (∗∗) there) Aff(T (An))⊂ lnMi=1 C([0, 1]i, R) ⊕ C(Xn, R) ⊕ Rpn−1 consists of (f1, f2, ..., fln , g, h2, ..., hpn ) which satisfies the conditions 1 1 fi(0) = {n, i}(cid:0)b0,i1 g(θ1)[n, 1] + {n, i}(cid:0)b1,i1 g(θ1)[n, 1] + pnXj=2 pnXj=2 For h = (h1, h2, ..., hpn ) ∈ Aff T ( Fn), let Γ′n(h)(t) = t · β♯ element C([0, 1], Rln ) =Lln i=1 C([0, 1]i, R). And let fi(1) = and b0,ijhj · [n, j](cid:1) b1,ijhj · [n, j](cid:1), 1(h) + (1 − t) · β♯ Γn : Aff(T ( Fn)) = Rpn → Aff(T (An)) ⊂ be defined by lnMj=1 C([0, 1]j , R) ⊕ C(Xn, R) ⊕ Rpn−1 (∗) (∗∗) 0(h) which gives an Γn(h1, h2, ..., hpn ) = (Γ′n(h1, h2, ..., hpn ), g, h2, ..., hpn ) ∈ where g ∈ C(Xn, R) is the constant function g(x) = h1. lnMj=1 C([0, 1]j , R) ⊕ C(Xn, R) ⊕ Rpn−1, 133 If we further assume at each step that the K-group homomorphisms b′0 = (β′0)∗0 and b′1 = (β′1)∗0 satisfy (e 13.15), and assume that c = (cij ) satisfies cij > 13· 22n, then by using 13.17 and the claim in 13.37, we will the following claim: Claim: For any f ∈ Aff(T (An)) with kfk ≤ 1 and f′ = ϕ♯ n+1(f′) − f′k < n,n+1(f ) ∈ Aff(T (An+1)), we have 4 22n , kΓn+1 ◦ π♯ n+1 : Aff(T (An+1)) → Aff(T ( Fn+1)) is induced by canonical quotient map πn+1 : where π♯ An+1 → Fn+1. Proof of the claim: For any n ∈ Z+, write n : Aff(T ( Fn)) Γn = Γ1 n ◦ Γ2 Γ2 n / / Aff(T (Fn)) Γ1 n / / Aff(T (An)) with Γ2 n : Aff(T ( Fn)) = Rpn → Aff(T (Fn)) = C(X, R) ⊕ Rpn−1 defined by Γ2 n(h1, h2, ..., hpn ) = (g, h2, ..., hpn ), where g is the constant function g(x) = h1, and with Γ1 Aff(T (An)) defined by n : Aff(T (Fn)) = C(X, R) ⊕ Rpn−1 → Γ1 n(g, h2, ..., hpn ) = (Γ′n(g(θ1), h2, ..., hpn ), g, h2, ..., hpn ). n = (π2 Also write π# n)# ◦ (π1 n)#, where π1 n : An → Fn and π2 For any f ∈ Aff(T (An)) with kfk ≤ 1, write f1 = (π1 n,n+1(f ) and f′1 = ψ♯ n,n+1(f1). By the condition cij > 13· 22n and the claim in the proof of Theorem 13.37, we have (e 13.28) n : Fn → Fn are quotient maps. n)#(f ), f′ = ϕ♯ kΓ2 n+1 ◦ (π2 n+1)♯(f′1) − f′1k < 2 22n . By condition (♦♦) (which is weaker than (♦♦♦)2), applying Lemma 13.17 to ¯An → ¯An+1 as Cn → Cn+1 (note that the definition of Γ′n is the same as the definition of Γ′n in 13.16 and ¯ϕn,n+1 : ¯An → ¯An+1 is the same as ϕn,n+1 : Cn → Cn+1), we have n+1)♯(f′) − f′(cid:17)Sp( ¯An+1)k < n+1)♯(f′) − f′(cid:1)Sp(Fn+1) = 0. So we have n+1)♯(f′) − f′k < Note that(cid:0)Γ1 k(cid:16)Γ1 n+1 ◦ (π1 n+1 ◦ (π1 n+1 ◦ (π1 2 22n . 2 22n . (e 13.29) (e 13.30) kΓ1 Consequently (applying (e 13.28)(e 13.30)), we get n+1 ◦ (π2 n,n+1 ◦ (π1 n+1)♯ ◦ (π1 n)♯(f ) − f′k n+1)♯(f′) − f′k n+1 ◦ (π2 n+1 ◦ Γ2 n+1)♯ ◦ ψ# n+1(f′) − f′k = kΓ1 n+1 ◦ Γ2 kΓn+1 ◦ π♯ = kΓ1 n+1 ◦ ϕn,n+1 = ψn,n+1 ◦ π1 (since π1 n) n+1 ◦ (π2 n+1 ◦ Γ2 2 n+1(f′1) − f′k + 22n n+1 ◦ ψ# n,n+1 ◦ (π1 n)#(f ) − f′k + 2 n+1)#(f′) − f′k + n+1 ◦ (π1 22n < = kΓ1 < kΓ1 = kΓ1 = kΓ1 n+1)♯ ◦ ψ# n,n+1(f1) − f′k = kΓ1 (by (e 13.28)) n+1 ◦ Γ2 n+1 ◦ (π2 n+1)♯(f′1) − f′k n+1)# ◦ ϕ# n,n+1(f ) − f′k + 2 22n 2 n+1 ◦ (π1 22n = kΓ1 2 2 22n . 22n + 134 So the claim is proved. Use the claim, one can prove the following approximately intertwining diagram: Aff(T (A1)) π# 1 Γ1 Aff(T ( F1)) ϕ# 1,2 ψ# 1,2 / Aff(T (A2)) π# 2 Γ2 Aff(T ( F2)) ϕ# 2,3 ψ# 2,3 Aff(T (A3)) π# 3 Γ3 Aff(T ( F3)) ··· Aff(T (F )) ··· Aff(T ( F )) , where π# i are induced by the quotient map πi : Ai → Fi. With the diagram above and noting that Aff(T ( F )) = Aff(∆), exactly the same argument as the proof of 13.19 with the above approximately intertwining diagram replacing the diagram there, we can prove (cid:0)(K0(A), K0(A)+, 1A), K1(A), T (A), rA(cid:1) ∼=(cid:0)(G, G+, u), K, ∆, r(cid:1). 13.39. The algebra A in 13.38 is not simple, we need to modify the homomorphism ϕn,n+1 to make the limit algebra simple. Let us emphasize that every homomorphism ϕ : An → An+1 is completely determined by ϕx = πx ◦ ϕ for each x ∈ Sp(An+1), where the map πx : An+1 → An+1x is the corresponding irreducible representation. Note that from the definition of ϕ : An → An+1 and from the assumption that cij > 13 for each entry of c = (cij ), we know that for any x ∈ Sp(An+1), Sp(ϕn,n+1x) ⊃ Sp( Fn) = (θ1, θ2, ..., θpn). (e 13.31) (See 13.1 and 13.2 for notations. Also note that each irreducible representation of Fn can be identified with one of Fn.) To see the above is true, one notes that for x ∈ Sp( ¯An+1), the homomorphism ϕn,n+1x is defined to be ¯ϕn,n+1x, and in turn, ¯ϕn,n+1 is defined in the proof of 13.13 (it was called ϕn,n+1 there) which satisfies condition (5)—the same condition as above. For x ∈ Sp(Fn+1), we have ϕn,n+1x = ψn,n+1x. From the definition of ψn,n+1 (see 13.30,13.31 and 13.33), we know that, if x ∈ Sp(F i n+1) (for i ≥ 2), then SP (ψn,n+1x) contains exactly cij copies of θj; and if x ∈ Sp(F 1 n+1), then SP (ψn,n+1x) contains exactly c1j copies of θj (for j ≥ 2) and at least c11 − 13 copies of θ1. Hence the condition also holds for this case. To make the limit algebra simple, we need to make the set Sp(ϕn,mx) sufficiently dense in Sp(An), for any x ∈ Sp(Am), provided m large enough. Write Sp(An) = ∪ln i=2 M[n,i](C)). Let 0 < d < 1/2 and let Z ⊂ Sp(An). Recall from 13.3 that, we say Z is d-dense in Sp(An) if the following holds: Z ∩ (0, 1)j is d-dense in (0, 1)j with the usual metric, Z ∩ Xn is d-dense in Xn with a given metric of Xn and if ξ ∈ Sn is an isolated point of Sp(An), then ξ ∈ Y. j=1(0, 1)j ∪ Xn ∪ Sn, where Sn = Sp(Lpn n,n+1k ≤ 1 2n . Now we will change ϕn,n+1 to a map ξn,n+1 : An → An+1 satisfying: (i) ξn,n+1 is homotopic equivalent to ϕn,n+1. (ii) kϕ♯ (iii) for any y ∈ Sp(An+1), if x ∈ Sp(An) satisfies x ∈ SP (ϕn,n+1y), then x ∈ SP (ξn,n+1y) (we avoid to say SP (ξn,n+1y) ⊃ SP (ϕn,n+1y), since multiplicity of the irreducible rep- resentation x ∈ Sp(An) in SP (ξn,n+1y) may not be larger or equal to its multiplicity in SP (ϕn,n+1y), when both SP (ξn,n+1y) and SP (ϕn,n+1y) are regarded as sets with multi- plicities), consequently, ξn,n+1 also satisfies that for any y ∈ Sp(An+1), n,n+1 − ξ♯ SP (ξn,n+1y) ⊃ Sp( Fn) = (θ1, θ2, ..., θpn), (e 13.32) since ϕn,n+1 has the same property. 135 / / / / / / / T T / / T T / / T T (iv) For any i ≤ n, Sp(ξi,n+1θ′ 2 ) is 1 n - dense in Sp(Ai), n+1)⊂Sp( Fn+1) = {θ′1, θ′2, ..., θ′pn+1} is the second point (note that θ′1 is the base n ), and we do not want to modify this one). where θ′2 ∈ Sp(F 2 point of [θ′1, θ′1 + 1] ∨ X′n+1 = Xn+1 = Sp(F 1 (v) For any x ∈ Sp(Fn+1) satisfying x 6= θ′2, In particular we have, ϕn,n+1x = ξn,n+1x. (v’) ϕn,n+1Sp(F 1 n+1) = ξn,n+1Sp(F 1 n+1), or equivalently, for any x ∈ Xn+1 = Sp(F 1 n+1), ϕn,n+1x = ξn,n+1x. (e 13.33) (e 13.34) Note that, in (iv) we do not need that Sp(ξi,n+1x) to be sufficiently dense in Sp(Ai) for This property is important for us to apply 13.35 to prove that the limit algebra A has the property that A ⊗ U ∈ B0 for any UHF-algebra U . all x ∈ Sp(An+1), but only need Sp(ξi,n+1θ′ ) to be sufficiently dense. Then combined with condition (iii) for ξn+1,n+2, we will have Sp(ξi,n+2x) to be sufficiently dense in Sp(Ai) for all x ∈ Sp(An+2), since 2 Sp(ξi,n+2x) = Sp(ξi,n+1y) ⊃ Sp(ξi,n+1θ′ 2 ). (e 13.35) [y∈Sp(ξn+1,n+2x) Namely, it follows from (iii) (for n + 1 in place of n) and (iv) that for any i ≤ n and any x ∈ Sp(An+2), the set Sp(ξi,n+2x) is 1 13.40. Suppose that we have constructed n - dense in Sp(Ai). ξ1,2 / A2 A1 ξ2,3 ξn−1,n / An, / ··· such that for all i ≤ n − 1, ξi,i+1 satisfy Conditions (i)–(v) (with i in place of n) in 13.39. We will construct the map ξn,n+1 : An → An+1. Let An =(cid:8)(f, g) ∈ C([0, 1], En) ⊕ Fn; f (0) = β0(π(g)), f (1) = β1(π(g))(cid:9) with Fn = PnM∞(C(Xn))Pn ⊕ M[n,i](C), (e 13.36) pnMi=2 where rank(Pn) = [n, 1]. By Condition (iii) of 13.39, applying to each j ∈ {i, i + 1, ..., n − 1} (in place of n), we have that for each x ∈ Sp(Ai), if x ∈ SP (ϕi,nz ) for some z ∈ Sp(An), then x ∈ SP (ξi,nz ). Recall n -dense in Xi. Combine this two facts, we have that n -dense in Xi. that the finite set Y ⊂ Xn from 13.29 satisfies that for each i < n,Sy∈Y SP (ϕi,ny ) is 1 (vi) ∪y∈Y SP (ξi,ny ) is also 1 Also recall that, we denote t ∈ (0, 1)j⊂Sp(cid:0)C([0, 1], Ej different direct summand of C([0, 1], En), and we denote 0 ∈ [0, 1]j by 0j and 1 ∈ [0, 1]j by 1j. Note that 0j and 1j do not correspond to irreducible representations. In fact, 0j corresponds to the direct sum of irreducible representations for the set n)(cid:1) by t(j) to distinguish spectrum from nθ∼b0,j1 1 , θ∼b0,j2 2 , ..., θ∼b0,jpn n o 136 / / / and 1j corresponds to the set Again recall from 13.29, T⊂Sp(An) is defined by 2 1 n o . , θ∼b1,j2 , ..., θ∼b1,jpn nθ∼b1,j1 )(j); j = 1, 2, ..., ln; k = 1, 2, ..., n − 1(cid:27) . T =(cid:26)( k n We need to modify ϕn,n+1 to ξn,n+1 such that If this is done, then combining with (vi) above, we know that (iv) of 13.39 holds. SP(cid:0)ξn,n+1θ′ 2(cid:1)⊃T ∪ Y. (e 13.37) Recall that in conditions (♦♦♦)1 and (♦♦♦)2, L = ln · (n − 1) + L1 = #(T ∪ Y ) and M = max{b0,ij; i = 1, 2, ..., pn; j = 1, 2, ..., ln}, where we write Y = {y1, y2, ..., yL1} ⊂ Xn. 13.41. We only defined for each x ∈ Sp(Fn+1) with x 6= θ′2, ξn,n+1x = ϕn,n+1x. Now we define ξn,n+1θ′ . To simplify the notation, write ξn,n+1 := ξ. Note that 2 We use y1, y2, ..., yL1 to replace L1 copies of θ1 in {θ∼c21 1 SP (ϕn,n+1θ′ 2 ) =(cid:8)θ∼c21 1 , θ∼c22 2 n , ..., θ∼c2pn (cid:9) . }, so {θ∼c21 1 } becomes {θ∼(c21−L1) 1 , y1, y2, ..., yL1}. 2 we use f (yi) to replace one copy of f (θ1). Note that That is, in the definition of ξn,n+1(f )θ′ f (yi) = Pn(yi)M∞(C)Pn(yi) can be identified with M[n,1](C) = F 1 n . For late use we also choose a path yi(s) (0 ≤ s ≤ 1) from θ1 to yi. That is, yi(0) = θ1, yi(1) = yi, and fix identi- fication of Pn(yi(s))M∞(C)Pn(yi(s)) with M[n,1](C) = F 1 n = Pn(θ1)M∞(C)Pn(θ1)—such an identification could be chosen to be continuously depending on s. (Here, we only use the fact that any projection (or vector bundle) over the interval is trivial to make such identification. Since the projection Pn itself may not be trivial, it is possible that the paths for different yi and yj (i 6= j) may intersect at yi(s1) = yj(s2) and we may use different identification of Pn(yi(s1))M∞(C)Pn(yi(s2)) = Pn(yj(s2))M∞(C)Pn(yj(s2)) with M[n,1](C) for i and j.) When we talk about f (yi(s)) later, we will considered it to be an element in M[n,1](C) (rather than in Pn(yi(s))M∞(C)Pn(yi(s))). Also we use ( k n )(j) ∈ (0, 1)j to replace In summary, we have that SP (ξθ′ 2 1 2 1 2 n n , θ∼a2 , θ∼b0,j2 , ..., θ∼apn , ..., θ∼b0,jpn 0j =nθ∼b0,j1 ) =(cid:8)θ∼a1 a1 = c21 − L1 −(cid:16)Pln a2 = c22 −(cid:16)Pln apn = c2pn −(cid:16)Pln o . (cid:9) ∪ T ∪ Y with j=1 b0,j1(cid:17) (n − 1) j=1 b0,j2(cid:17) (n − 1) j=1 b0,jpn(cid:17) (n − 1) . ... 137 From (♦♦♦)1 of 13.40, we know that 22n − 1 22n c2i, ai ≥ (e 13.38) ) remain the same. The map 2 is not unique, but is unique up to unitary equivalence. Note that ϕn,n+1θ′ and as many as ai points (counting multiplicity) of SP (ϕn,n+1θ′ ξθ′ is homotopic to ξθ′ Ω(s) : An → F 2 (s · k n )(j) ∈ (0, 1)j . simplified notation for ξn,n+1. It is obvious that Ω(0) = ϕn,n+1θ′ n+1 as the definition of ξθ′ : An → F 2 n+1. Namely, for each s ∈ [0, 1], one can define homomorphism n )(j) ∈ (0, 1)j by . Again recall ξ is , by replacing yi by yi(s), and ( k , and Ω(1) = ξθ′ : An → F 2 n+1 2 2 2 2 2 2 n+1 = F 2 n+1 ≥ 2, then, for i ≤ p0 If An+1 is in the case of 13.11, with p0 n+1) is a closed and open subset of Sp(An+1) since F i n+1 is separate from C([0, 1], En+1) in the definition. On the n+1 has a single spectrum θ′2. That is, θ′2 is an isolate point in Sp(An+1). other hand, F 2 Then we can define ξx = ϕn,n+1x for any x ∈ Sp(An+1) \ {θ′2} to finish the construction of ξ = ξn,n+1 (this case is much easier). Also, for this case, we do not need the estimation (♦♦♦)2 (only (♦♦♦)1 and the old (♦♦) (see (e 13.15)) will be enough). 13.42. In 13.41, we have defined the part ξ (restricted to Fn+1 and also to Fn+1); that is n+1, Sp(F i ξ′ = π ◦ ξ : An / An+1 π / / Fn+1 , where π is the quotient map modulo the ideal In+1 + J′n+1 (see 13.28 for notation of In and J′n). Recall that ξ is the simplified notation for ξn,n+1. Now we define ξ[0,1]j for each [0, 1]j ⊂ Sp(C[0, 1], Ej n+1). We already know the definition of ξ0j = πj ◦ β′0 ◦ ξ′ : An → Ej n+1 and ξ1j = πj ◦ β′1 ◦ ξ′ : An → Ej n+1, where πj : En+1 → Ej for the corresponding points in Sp(C([0, 1], Ej n+1. Here we use 0j, 1j ∈ [0, 1]j⊂Sp(C([0, 1], Ej to obtain the definition of ξ[0,1]j Now we need to connect ξ0j . Let ψj : C([0, 1], En) → n+1) be as defined in the proof of 13.13 (and 13.14). Let Γ : An → C([0, 1], En) be C([0, 1], Ej the natural inclusion map. As in the proof of 13.13, for the original map ϕn,n+1, we have n+1)), and reserve 0j, 1j and ξ1j n)). SP (ϕn,n+10j SP (ϕn,n+11j ) =(cid:26)SP (ψj ◦ Γ0j ) =(cid:26)SP (ψj ◦ Γ1j ), θ∼rj 1 1 , θ∼rj 2 2 , ..., θ∼rj ) θ∼rj 1 1 , θ∼rj 2 2 , ..., θ∼rj where rj l = pn+1Xk=1 Note that SP (ξ0j b′0,jkckl −Xdjk<0 nθ∼b0,i1 ) and SP (ξ1j 1 2 , θ∼b0,i2 djkb1,kl + Xdjk>0 djkb0,kl + Xdjk=0 ) are obtained by replacing some subsets , ..., θ∼b0,ipn pn o = 0i ∈ Sp(C([0, 1], Ei n)) 138 pn pn pn (cid:27) and pn (cid:27) , (b0,kl + b1,kl) . (e 13.39) / ) by ( k 0,j1 , (θ′2)∼b′ 0,j2 , ..., (θ′pn+1 )∼b′ ) and SP (ϕn,n+11j in SP (ϕn,n+10j n )(i) and replacing some of θ1 each by one of yk (k = 1, 2, ..., L1) (see 13.1 and 13.2 for notation). We would like to give precise calculation for the numbers of θ1 or 0i to be replaced. Recall that 0j = {(θ′1)∼b′ if i 6= 2. So the set SP (ξ0j From the part of ξ already defined, recall that ξθ′ ) ) by doing the following replacement: replace each group of b′0,j2 is obtained from SP (ϕn,n+10j copies of θ1 by same number copies of yi for each i = 1, 2, ..., L1(totally L1 · b′0,j2 copies of θ1 have been replaced by some yi’s), and replace each group of b′0,j2 copies of 0i for i = 1, 2, ..., ln, by same number of copies of ( k n )(i) for each k = 1, 2, ..., n − 1 (totally (n − 1) · b′0,j2 copies of 0i =nθ∼b0,i1 o have been replaced by some ( k 0,jpn+1}, and 1j = {(θ′1)∼b′ = ϕn,n+1θ′ 1,j2 , ..., (θ′pn+1 )∼b′ n )(i)’s). Consequently 1,j1 , (θ′2)∼b′ , ..., θ∼b0,ipn , θ∼b0,i2 pn 1 2 i i 1,jpn+1}. SP (ξ0j where ) =nSP (ψj ◦ Γ0j ), θ∼s1 , θ∼s2 , ..., θ∼spn 1 2 pn o ∪ (T ∪ Y )∼b′ i=1 b0,i1) · b′0,j2 0,j2 , s1 = rj s2 = rj spn = rj 1 − (L1 + (n − 1)Pln 2 − ((n − 1)Pln pn − ((n − 1)Pln i=1 b0,i2) · b′0,j2 i=1 b0,ipn) · b′0,j2 . Exactly as in the above argument, we have SP (ξ1j ) =nSP (ψj ◦ Γ1j ), θ∼t1 1 , θ∼t2 2 where , ..., θ∼tpn pn o ∪ (T ∪ Y )∼b′ i=1 b0,i1) · b′1,j2 1,j2 , i=1 b0,i2) · b′1,j2, ... ... t1 = rj t2 = rj tpn = rj 1 − (L1 + (n − 1)Pln 2 − ((n − 1)Pln pn − ((n − 1)Pln i=1 b0,ipn) · b′1,j2 . pn (cid:9) ∪ (T ∪ Y )∼b′ n θ∼t1 , θ∼t2 1 2 , ..., θ∼tpn 0,j2 of the set SP (ξ0j pn o ∪ (T ∪ Y )∼b′ 1,j2 ) and ) already be paired by the homomorphism ψj defined before, we only need to pair ) now. Of course the part of SP (ψj ◦ Γ0j ) and SP (ξ1j ) with the part We can pare two sets SP (ξ0j SP (ψj ◦ Γ1j the part(cid:8) θ∼s1 , ..., θ∼spn , θ∼s2 1 2 ). ) with all ti copies of θi in SP (ξ1j If b′1,j2 ≥ b′0,j2 (the other case b′0,j2 > b′1,j2 ) with some b′0,j2 (≤ b′1,j2) copies of x in SP (ξ1j of the set SP (ξ1j can be done similarly), then si ≥ ti. First we do the obvious pairing. That is, we can pair ti copies of θi from ); and we can also pair, for each x ∈ Y ∪ T , all b′0,j2 SP (ξ0j copies of x in SP (ξ0j ). After this procedure of pairing, what left in SP (ξ1j {θ∼(s1−t1) there are b′1,j2 − b′0,j2 copies of y left in SP (ξ1j same number of copies of θ1 in SP (ξ0j for each ( k ) is }. And we will pair these parts as below. For each y ∈ Y , ), and we pair all those copies (for each y) with )—it will totally cost L1 · (b′1,j2 − b′0,j2) copies of θ1; and ), we pair all those 0,j2), and what left in SP (ξ0j ) is the set (Y ∪ T )∼(b′ ,··· , θ∼(spn−tpn ) n )(i) ∈ T , there are also b′1,j2 − b′0,j2 copies of ( k n )(i) left in SP (ξ1j , θ∼(s2−t2) 1,j2−b′ pn 1 2 139 copies (for each ( k totally cost n )(i)) with same number of copies of 0i =nθ∼b0,i1 1 , θ∼b0,i2 2 , ..., θ∼b0,ipn pn o—it will i=1 b0,i1) · (b′1,j2 − b′0,j2) i=1 b0,i2) · (b′1,j2 − b′0,j2) i=1 b0,ipn) · (b′1,j2 − b′0,j2) copies of θ1, copies of θ2, ... copies of θpn. Hence, putting all them together, totally we will use exactly ((n − 1)Pln ((n − 1)Pln ((n − 1)Pln lnXi=1 ((n − 1) (L1 + (n − 1) copies of θ1 and (e 13.40) b0,i1) · (b′1,j2 − b′0,j2) = s1 − t1 lnXi=1 b0,ik) · (b′1,j2 − b′0,j2) = sk − tk copies of θk, (e 13.41) for each k ∈ {2, 3, ..., pn}. That is, the set (Y ∪ T )∼(b′ 1,j2−b′ 0,j2) is paired exactly with , θ∼(s2−t2) 2 , ..., θ∼(spn−tpn ) pn }. Then, we can write 1 {θ∼(s1−t1) ) =nSP (ψj ◦ Γ0j SP (ξ0j ) =nSP (ψj ◦ Γ1j and a pairing Θ between Ξ0 and Ξ1 such that each pair σ ∈ Θ is of the following form: σ = (x ∈ Ξ0, x ∈ Ξ1) with x ∈ Y ∪ T ∪ {θ1, θ2, ...θpn}, σ = (cid:0)0i ∈ Ξ0, ( k (θ1 ∈ Ξ0, yk ∈ Ξ1). For any σ ∈ Θ, define ψσ : An → C([0, 1], Ej ψσ(f ) as: ), Ξ1o , n )(i) ∈ Ξ1(cid:1), or σ = ), Ξ0o and SP (ξ1j n) by sending f ∈ An to ψσ(f )(t) = f (x), f ((t k n )(i)), f (yk(t)), if σ = (x ∈ Ξ0, x ∈ Ξ1) if σ =(cid:0)0i ∈ Ξ0, ( k n )(i) ∈ Ξ1(cid:1) if σ = (θ1 ∈ Ξ0, yk ∈ Ξ1) (case 1), (case 2), (case 3).  (Of course, if b′0,j2 > b′1,j2, then the case (2) above will be changed to σ =(cid:0)( k n )(i) ∈ Ξ0, 0i ∈ Ξ1(cid:1) and the function ψσ(f )(t) should be defined by ψσ(f )(t) = f (((1 − t) k n )(i)); and the case (3) above will be changed to σ = (yk ∈ Ξ0, θ1 ∈ Ξ1) and the function ψσ(f )(t) should be defined by ψσ(f )(t) = f (yk(1 − t)).) Finally let ξj(f ) = diag(cid:0)ψj(f ), ψσ1(f ), ψσ2 (f ), ..., ψσ• (f )(cid:1) where {σ1, σ2, ..., σ•} = Θ. Then this ξj has the expected spectrum at 0j and 1j. After conjugating a unitary, we can get ξj(f )(0j) = πj ◦ β′0 ◦ ξ′(f ), and ξj(f )(1j) = πj ◦ β′1 ◦ ξ′(f ). (e 13.42) Combining this ξj together with the previous defined ξ into Fn+1, we get the definition of ξ : An → An+1. That is for x ∈ Sp(C0((0, 1), Ej n)) ⊂ Sp( ¯An+1), define ξx to be ξjx; and for x ∈ Sp(Fn+1), define ξx to be as previously defined ξx on this part. The condition e 13.42 says these two parts of definition compatible on the their boundary Sp( Fn+1) = Sp( ¯An+1)∩Sp(Fn+1). Notice that ψσ (part of ξ) is homotopic to the constant map, ψ′σ(f )(t) = f (0i) for case 2 and x = ( k n )(i) ∈ T in case 1, or 140 ψ′σ(f )(t) = f (θ1) for case 3 and x ∈ Y in case 1, which just matches the definition of the corresponding part of ϕn,n+1 (see the penultimate para- graph of 13.41 also). That is, ξn,n+1 is homotopic to ϕn,n+1, which is (i) of 13.39. From (∗) above, the definition of sl and tl, and condition (♦♦♦)2 of 13.29, we have sl > 22n − 1 22n rj l and tl > 22n − 1 22n rj l . This implies that, for each x ∈ Sp(An+1), SP (ϕn,n+1x) and SP (ξx) differ by a fraction at most 1 22n . Thus we have kϕ♯ n,n+1 − ξ♯k < 1 22n , which is (ii) of 13.39. Evidently, (iii) and (v) of 13.39 follows from the construction. Note that we have already made (e 13.37) hold. Therefore (iv) follows, as we mentioned immediately after the equation (e 13.37). Therefore our inductive construction is completed. 13.43. Let B = lim(An, ξn,n+1). Recall that A = lim(An, ϕn,n+1 ). Using (i) and (ii), exactly as the same as the proofs of Theorems 13.19 and 13.37, we have that(cid:0)(K0(B), K0(B)+, 1B), K1(B), T B, rB(cid:1) is isomorphic to(cid:0)(K0(A), K0(A)+, 1A), K1(A), T (A), rA(cid:1) (including the compatibility). Hence (cid:0)(K0(B), K0(B)+, 1B), K1(B), T (B), rB(cid:1) ∼= (cid:0)(G, G+, u), K, ∆, r(cid:1). On the other hand, we have that, for each n, and m > n + 1, Sp(ξn,n+1x ) is m−2 dense in Sp(An) for any x ∈ Sp(Am) (see the end of 13.39). This condition implies that the limit algebra B is simple (see 13.4). Notice that in the definition of ξn,n+1 we have 1 n+1) = ϕn,n+1Sp(F 1 ξn,n+1Sp(F 1 Then 13.35 implies the following corollary. Corollary 13.44. For any m > 0 and any Ai, there is an integer n ≥ i and a projection R ∈ Mm(An+1) such that (1) R commutes with Λξn,n+1(An), where Λ : An+1 → Mm(An+1) is the amplify map sending n+1). a to an m × m diagonal matrix: Λ(a) = diag(a,··· , a); (2) Recall ¯An+1 = A(cid:16) Fn+1, En+1, β′0, β′1(cid:17), where β′0, β′1 : Fn+1 → En+1 is as in the definition of An+1 (see (e 13.23). Then there is an injective homomorphism such that RΛ(ξn,n+1(An))R⊂ι(Mm−1( ¯An+1)). ι : Mm−1( ¯An+1) −→ RMm(An+1)R Proof. We have already defined the part of R in Mm(F 1 n+1) with property described in 13.35 (the definition is given by combining 13.32 and 13.34). Note that this R works for the homomorphism ξn,n+1 in place of ϕn,n+1 because ξn,n+1Sp(F 1 R(θ′1) = 1 F 1 n+1) = ϕn,n+1Sp(F 1 n+1). In particular, We extend the definition of R as follows. For each x ∈ Sp(An+1) \ Sp(F 1 n+1), define 0 n+1 ⊗(cid:18) 1m−1 0 0 (cid:19) . R(x) = 1An+1x ⊗(cid:18) 1m−1 0 0 (cid:19) . 0 141 One can use the map ι combining with the identity map id : Mm−1( ¯An+1) → Mm−1(An+1/J′n+1) to get the corollary. (Note that ¯An+1 = An+1/J′n+1). Corollary 13.45. Let B be as constructed above. Then B ⊗ U ∈ B0 for every UHF-algebra U of infinite dimension. Proof. In the above corollary, we know that ¯An+1 ∈ C0 and therefore Mm−1( ¯An+1) and ι(Mm−1( ¯An+1)) are in C0. Also, for all normalized traces τ ∈ T (Mm(An+1)), we have τ (1 − R) = 1/m. Fix an integer k ≥ 1 and any finite subset F ⊂ B ⊗ Mk, let F1 ⊂ B be a finite subset such that {(fij)k×k : fij ∈ F1} ⊃ F. Now, evidently, the inductive limit algebra B = lim(An, ξn,m) have the following property: For any finite set F1 ⊂ B, ε > 0, δ > 0, and any m > 1/δ, there is a unital C∗-subalgebra C ⊂ Mm(B) which is in C0 such that (i) k[1C, diag{f, ..., f }]k < ε/k2, for all f ∈ F1, (ii) dist(1C(diag{f, ..., f })1C , C) < ε/k2, for all f ∈ F1, and {z }m {z }m (iii) τ (1 − 1C) = 1/m < δ for all τ ∈ T (Mm(B)). Thus the above (i), (ii), (iii) hold by replacing B by Mk(B), F1 by F ⊂ Mk(B) and ε/k2 by ε. Now B ⊗ U can be written as lim(B ⊗ Mkn, ιn,m) with k1k2k3 ··· and kn+1/kn → ∞, and ιn,n+1 is the amplification by sending f ∈ B ⊗ Mkn to diag(f, ..., f ) ∈ B ⊗ Mkn+1, where f repeated kn+1/kn times. To show B ⊗ U ∈ B0, let F ⊂ B ⊗ U be a finite subset and let a ∈ (B ⊗ U )+ \ {0}. There is an integer m0 > 0 such that τ (a) > 1/m0 for all τ ∈ B ⊗ U . Without loss of generality, we may assume that F ⊂ B ⊗ Mkn with kn+1 > m0. Then by what has been proved above for B ⊗ Mkn with m = kn+1/kn (and note that ιn,n+1 is the amplification), there is a unital C∗ subalgebra C ⊂ B ⊗ Mkn+1 with C ∈ C0 such that k[1C, ιn,n+1(f )]k < ε, for all f ∈ F, such that dist(1C(ιn,n+1(f ))1C , C) < ε for all f ∈ F, and such that τ (1 − 1C) = 1/m < δ for all τ ∈ T (Mkn+1(B)). Then ιn+1,∞(C) is the desired subalgebra. (Note that 1 − 1C . a follows from strict comparison property of B ⊗ U .) It follows that B ⊗ U ∈ B0. Theorem 13.46. For any simple weakly unperforated Elliott invariant (cid:0)(G, G+, u), K, ∆, r(cid:1), there is an unital simple algebra A ∈ N Z0 which is an inductive limit of (An, ϕn,m) with An described in the end of 13.27, with ϕn,m injective, such that kn (cid:0)(K0(A), K0(A)+, 1A), K1(A), T (A), rA(cid:1) ∼=(cid:0)(G, G+, u), K, ∆, r(cid:1). Proof. By 13.45, A ∈ N0. Since A is a unital simple inductive limit of subhomogeneous C- algebras with no dimension growth. By Corollary 6.5 of [103], A is Z-stable. Corollary 13.47. For any simple weakly unperforated Elliott invariant (cid:0)(G, G+, u), K, ∆, r(cid:1) with K = 0 and G torsion free,there is an unital Z-stable simple algebra which is an inductive limit of (An, ϕn,m) with An in C0 described in 3.1, with ϕn,m injective, and such that (cid:0)(K0(A), K0(A)+, 1A), K1(A), T (A), rA(cid:1) ∼=(cid:0)(G, G+, u), 0, ∆, r(cid:1). Proof. In the construction of An, just let all the spaces Xn involved to be the space of a single point. 142 14 Models for C∗-algebras in A0 with (SP) property 14.1. For technical reasons, in the construction of our model algebras, it is important for us to be able to decompose An into direct sum of two parts: the homogeneous part which stores the information of Inf K0(A) and K1(A) and the part of algebra in C0 which stores information of K0(A)/ Inf K0(A), T (A) and paring between them. But this can not be done in general for the algebras in N0 (see 13.5), since some algebras in N0 has minimal projections (or even the unit itself is a minimal projection). But we will prove this can be done if the Elliott invariant satisfies an extra condition (SP) described below. Note that if A ∈ N0 then the Elliott invariant of A ⊗ Mp has the condition (SP) even though the Elliott invariant of A itself may not satisfy the condition. Let(cid:0)(G, G+, u), K, ∆, r(cid:1) be a weakly unperforated Elliott invariant as 13.6. We say that it has the (SP) property if for any real number s > 0, there is g ∈ G+ \ {0} such that τ (g) < s for any state τ on G, or equivalently, r(τ )(g) < s for any τ ∈ ∆. In this case, we will prove that the algebra in 13.46 can be chosen to be in class B0 (rather than in the larger class N0 = {A: A ⊗ Mp ∈ B0}). Roughly speaking, for each An, we will separate the part of homogeneous algebra which will store all the information of infinitesimal part of K0 and K1, and it will be in the corner PnAnPn with Pn small compare to 1An in the limit algebra. In fact, the construction of this case is much easier, since the homogeneous blocks can be separate out from the part of C0—we will first write the group inclusion Gn ֒→ Hn as in 13.11. 14.2. Let ((G, G+, u), K, ∆, r) be the one given in 13.6 or 13.24. As in 13.8, let ρ : G → Aff ∆ be dual to the map r. Denote by the kernel of the map ρ by Inf(G)—the infinitesimal part of G, that is Inf(G) = {g ∈ G : ρ(g)(τ ) = 0, ∀τ ∈ ∆}. Let G1 ⊆ Aff ∆ be a countable dense subgroup which is Q linearly independent with ρ(G)—that is, if g ∈ ρ(G) ⊗ Q and g1 ∈ G1 ⊗ Q satisfy g + g1 = 0, then both g and g1 are zero. Note that such G1 exist, since as Q vector space, dimension of ρ(G) ⊗ Q is countable, but dimension of Aff ∆ is uncountable. Again as in 13.8, let H = G ⊕ G1 with H+ \ {0} to be the set of (g, f ) ∈ G ⊕ G1 with ρ(g)(τ ) + f (τ ) > 0 for all τ ∈ ∆. The scale u ∈ G+ is regarded as (u, 0) ∈ G ⊕ G1 = H as the scale of H+. Since ρ(u)(τ ) > 0, it follows that u is an order unit for H. Since G1 is Q linearly independent with ρ(G), we know Inf(G) = Inf(H)—that is, when we embed G into H, it does not create more elements in the infinitesimal group. Evidently Tor(G) = Tor(H) ⊂ Inf(G). Let G′ = G/ Inf(G) and H′ = H/ Inf(H), then we have the following diagram: 0 0 / Inf(G) / G G′ 0 / Inf(H) / H / H′ / 0. Let G′+ (or H′+) and u′ be the image of G+ (or H+) and u under the quotient map from G to G′ (or from H to H′). Then (G′, G′+, u′) is weakly unperforated group without infinitesimal elements. Note that G and H share same unit u, and therefore G′ and H′ share same unit u′. Since r(τ )Inf(G) = 0 for any τ ∈ ∆, the map r : ∆ → Su(G) induces a map r′ : ∆ → Su′(G′). Hence ((G′, G′+, u′),{0}, ∆, r′) is a weakly unperforated Elliott Invariant with trivial K1 group and no infinitesimal elements in the K0-group. 143 / / / / _ / / _ / / / / 14.3. With the same argument as that of 13.10, we have the following diagram of inductive limits: G′1 ι1 H′1 α′ 12 / G′2 α′ 23 ι2 γ′ 12 / / H′2 γ′ 23 ··· / ··· / G′_ ι / H′, where each H′n is a direct sum of finite copies of ordered groups Z, αn,n+1 = γn,n+1G′ H′n/G′n is a free abelian group. n, and By 13.23, we can construct an increasing sequence of finitely generated subgroups Inf1 ⊂ Inf2 ⊂ Inf3 ⊂ ··· ⊂ Inf(G), with Inf(G) =S∞i=1 Infn and the inductive limit γ1,2 γ2,3 Inf 1 ⊕H′1 / Inf2 ⊕H′2 / Inf 3 ⊕H′3 γ3,4 / ··· / H Put Hn := Infn ⊕H′n and Gn = Infn ⊕G′n. Since G′n is a subgroup of H′n, the group Gn is also a subgroup of Hn. Define αn,n+1 : Gn → Gn+1 by αn,n+1 = γn,n+1Gn, which is compatible with α′n,n+1 in the sense that (ii) of 13.23 holds. Hence we get the following diagram of inductive limits: G1 α12 / G2 α23 / ··· ι1 ι2 H1 γ12 / / H2 γ23 / / ··· / G ι / H, with αn,n+1(Infn) ⊂ Infn+1 and αn,n+1(Inf n) is the inclusion map. Note that all notations discussed so far in this section will be used for the rest of this section. Lemma 14.4. Let (G, G+, u) = lim((Gn, (Gn)+, un), αn,m) and (H, H+, u) = lim((Hn, (Hn)+, un), γn,m) For any n with Gn be as above. Again suppose that(cid:0)(G, G+, u), K, ∆, r(cid:1) has (SP) property. ιn֒→ Hn, any positive integer L, and for any D = Zk (for k an arbitrary positive integer), there are m and positive maps (κ, id) : Hn → D ⊕ Hn, (κ′, id) : Gn → D ⊕ Gn, η : D ⊕ Hn → Hm, and η′ : D ⊕ Gn → Gm such that the following diagram commutes: αn,m Gn $■■■■■■■■■ (κ′,id) Gm :ttttttttt η′ ιn (id,ιn) ιm D ⊕ Gn D ⊕ Hn (κ,id) :✉✉✉✉✉✉✉✉✉ η $❏❏❏❏❏❏❏❏❏ γn,m / Hm, Hn and such that the following are true: (1) For any positive element x ∈ H′+ consequently, each component of κ(un) = κ′(un) in Zk is strictly positive. n , each component of κ(x) in Zk is strictly positive and 144 / _ / / _ / / / / / / / / _ / _ / _ / / / $ : _ $ : / (2) For any τ ∈ ∆, r(τ )((αm,∞ ◦ η′)(1D))(= r(τ )((γm,∞ ◦ η)(1D))) < 1/L. (3) Each component of the map η : D ⊕ H′n = Zk ⊕ Zpn → H′m = Zpm is L-large—that is all entries in the (k + pn) × pm matrix corresponding to the map are larger than L. Proof. We will use the following fact several times: the positive cone of G′n (and of H′n) is finitely generated (note that even though Gn and Hn are finitely generated, their positive cone may not be finitely generated). For Hn, pick an arbitrary positive nonzero homomorphism λ : Hn → Z n , λ(x) > 0. Denote by λ′ = λ ◦ ιn : Gn → Z. It follows from so that for any nonzero x ∈ H′+ positivity that such map λ satisfies λ(Infn) = 0. Since G has the (SP) property, there is p′ ∈ G+ \ {0} such that for any a ∈ (G′n)+ (not use (Gn)+ here, but we regard it as subset of (Gn)+) the element αn,∞(a) − k · λ′(a) · p′ is positive and for any a ∈ H′+ the element n (not use (Hn)+ here, but again we regard it as subset of (Hn)+), γn,∞(a) − k · λ(a) · p′ is positive, where the map αn,∞ and γn,∞ are the homomorphisms from Gn to G and from Hn to H respectively. Moreover, one may require that r(τ )(λ(un) · p′) < 1/2kL for all τ ∈ ∆. (e 14.1) Since (G′n)+ and H′+ n are finitely generated, there is an integer m and p ∈ G+ m such that αm,∞(p) = p′, αn,m(a) − k · λ′(a) · p ∈ G+ γn,m(a) − k · λ(a) · p ∈ H + m for all a ∈ (G′n)+ and m for all a ∈ H′+ n . Then define αn : Gn → Gm and γn : Hn → Hm by αn : Gn ∋ a 7→ αn,m(a) − k · λ′(a) · p ∈ Gm and γn : Hn ∋ a 7→ γn,m(a) − k · λ(a) · p ∈ Hm. (e 14.2) By the choice of p, the maps αn and γn are positive. (Note that αn = αn,m and γn = γn,m on Infn.) A direct calculation shows the following diagram commutes: αn,m Gn $■■■■■■■■■ (κ′,id) Gm :ttttttttt η′ ιn (id,ιn) ιm D ⊕ Gn D ⊕ Hn Hn where D = Zk, (κ,id) :✉✉✉✉✉✉✉✉✉ η $❏❏❏❏❏❏❏❏❏ γn,m / Hm, κ′(a) = (λ′(a), ..., λ′(a) ) ∈ D and κ(a) = (λ(a), ..., λ(a) ) ∈ D, k {z } 145 k {z } / / $ : _ $ : / η′((m1, ..., mk, g)) = (m1 + ··· + mk)p + αn(g) and η1((m1, ..., mk, g)) = (m1 + ··· + mk)p + γn(g). The order of D ⊕ Gn and D ⊕ Hn are the standard order on direct sums, i.e., (a, b) ≥ 0 if and only if a ≥ 0 and b ≥ 0. Since the maps αn and γn are positive, the maps η′ and η are positive. Condition (1) follows from the construction; Condition (2) follows from (e 14.1), and Condition (3) follows from the simplicity of H, if one passes to further stage (choose larger m). 14.5. Write K (the K1 part of invariant) as the union of increasing sequence of finitely generated abelian subgroups: K1 ⊂ K2 ⊂ K3 ⊂ ··· ⊂ K with K =S∞i=1 Ki. For a finitely generated abelian group G, we use rank G to denote the minimum number of possible generated set of G—that is, G can be written as a direct sum of rank(G) cyclic groups (e.g., Z or Z/mZ, m ∈ N). Let dn = max{2, 1 + rank(Inf n) + rank(Kn)}. Apply 14.4 with k = dn (and suitable choice of L = Ln > 13 · 2n to be determined later (e 14.5)) for each n and pass to subsequence, we can obtain the following diagram of inductive limits: Zd1 ⊕ G1 α1,2 Zd2 ⊕ G2 α2,3 / ··· (id,ι1) (id,ι2) Zd1 ⊕ H1 / Zd2 ⊕ H2 γ2,3 / ··· γ1,2 / G ι / H. Consider the ordered group Zdn ⊕ Gn = Zdn ⊕ Infn ⊕G′n with the scale (κ′(un), un), where κ′ is as in 14.4; and let us still denote it by (Gn, un). Similarly, consider Zdn⊕Hn = Zdn⊕Infn ⊕H′n with the scale (κ(un), un), where κ is also as in 14.4; and let us still denote it by (Hn, un). We will also use αn,n+1 and γn,n+1 for αn,n+1 and γn,n+1 in the above diagram. Let G′′n = Zdn⊕Infn, then with the new notation, we have Gn = G′′n ⊕ G′n and Hn = G′′n ⊕ H′n. Now we have the following diagram G′′1 ⊕ G′1 α1,2 G′′2 ⊕ G′2 α2,3 / ··· (id,ι1) (id,ι2) / G ι / H. G′′1 ⊕ H′1 γ1,2 / G′′2 ⊕ H′2 γ2,3 / ··· The positive cones of G′n and H′n are as before. Write G′′n = Ldn i=1(G′′n)i, with (G′′n)i = Z for i ≤ 1 + rank(Kn) and (G′′n)i = Z ⊕ (a cyclic group) for 1 + rank(Kn) < i ≤ dn, and the direct sum of those cyclic groups is Infn. Here the positive cone of (G′′n)i is given by the strict positivity of first coordinate for nonzero positive elements. And an element in G′′n is positive if each of its components in (G′′n)i is positive. The order on the groups G′′n ⊕ G′n and G′′n ⊕ H′n are the standard order on direct sums, i.e., (a, b) ≥ 0 if and only if a ≥ 0 and b ≥ 0. Since each entry of γn,n+1 : Zdn ⊕ H′n → Zdn+1 ⊕ H′n+1 is strictly positive, the infinitesimal part can be put in any given summand without affect the order structure of the limit. Let un = (u′′n, u′n) ∈ G′′n ⊕ G′n ⊂ H′n ⊕ H′′n be the unit of Gn and of Hn. Note that Inf(G) ⊂ ∞[n=1 (Inf n) ⊂ ∞[n=1 αn,∞(G′′n). (e 14.3) Definition 14.6. A C∗-algebra is said to be in the class H if it is the direct sum of the algebras of the form P (C(X) ⊗ Mn)P , where X = {pt}, [0, 1], S1, S2, T2,k and T3,k. 146 / / _ _ / / _ / / / / / _ _ / / _ / / / 14.7. For each n as in 13.10 applied to G′n ⊂ H′n, one can find finite dimensional C∗-algebras Fn and En, unital homomorphisms β0, β1 : Fn → En, and the C-algebra Cn = A(Fn, En, β0, β1) := {(f, a) ∈ C([0, 1], En) ⊕ Fn; f (0) = β0(a), f (1) = β1(a)} such that (K0(Fn), K0(Fn)+, [1Fn]) = (H′n, H′+ n , u′n), (K0(Cn), K0(Cn)+, 1Cn) = (G′n, (Gn)+, u′n), K1(Cn) = {0}, and furthermore K0(Cn) is identified with ker((β1)∗0 − (β0)∗0) = {x ∈ K0(Fn); ((β0)∗0 − (β1)∗0)(x) = 0 ∈ K0(En)}. We can also find a unital C∗ algebra Bn ∈ H such that (K0(Bn), K0(Bn)+, 1Bn) = (G′′n, G′′+ n , u′′n) (e 14.4) and K1(Bn) = Kn. More precisely, we have that Bn = Ldn K1(Bi In particular, the algebra B1 at least one block B2 replace the single point spectrum by interval [0, 1]. n) = (G′′n)i and n) is either a cyclic group for the case 2 ≤ i ≤ 1 + rank(Kn) or zero for the other cases. n can be chosen to be a matrix algebra over C. And we assume, for n, the spectrum is not a single point (note that dn ≥ 2), otherwise, we will Now, we can extend the maps β0 and β1 to β0, β1 : Bn ⊕ Fn → En, by defining them to be Consider An = Bn ⊕ Cn. Then the C-algebra An can be written as n, with K0(Bi zero on Bn. i=1 Bi An = {(f, a) ∈ C([0, 1], En) ⊕ (Bn ⊕ Fn); f (0) = β0(a), f (1) = β1(a)}. For each block Bi n, choose a base point xn,i ∈ Sp(Bi n). Let In = C0(cid:0)(0, 1), En(cid:1) be the ideal of Cn as before. And let Jn be the ideal of Bn consisting of functions vanishing on all base points {xn,i}dn i=1. Applying Proposition 13.21, and exactly as in 13.13 and 13.30, we will have non-simple inductive limit A′ = lim(Bn ⊕ Cn, ϕn,m) with (ϕn,n+1)∗0 = αn,n+1, K1(ϕn,n+1) is the inclusion from Kn to Kn+1, Ell(A) ∼= ((G, G+, u), K, ∆, r), and with ϕn,n+1(In) ⊂ In+1, ϕn,n+1(Jn) ⊂ Jn+1. Furthermore the map π ◦ ϕn,n+1Bn is injective for projection π : Bn+1 ⊕ Cn+1 → Bn+1, since at least one block of Bn+1 is not a single point. (We do not need 13.31 to 13.35 and 13.44 because the homogeneous algebra Bn is separate from Cn ∈ C0. It is also true that Bn occupies relatively smaller space compare to those occupied by Cn in the limit algebra—namely, by (2) of 14.4 and the choice of L = Ln ≥ 13 · 2n in 14.5, one has 1 τ (αn,∞(1Bn)) τ (αn,∞(1Cn)) < , 13 · 2n − 1 for all τ .) We need to modify ϕn,n+1 to ψn,n+1 to make the algebra simple. However, it is much easier than what we did in 13.39 to 13.43. We only need to modify the partial map from An to B1 n+1, the first block of An+1 = Bn+1 ⊕ Cn+1, and keep other part of the map to be same—that is, we only need to make SP (ψn,n+1xn+1,1) to be dense enough in Sp(An). In order to do this, one chooses a finite set X ⊂ Sp(An) dense enough to play the role of Y ∪ T (as in 13.29) and let Ln in 14.5 satisfies Ln > 13 · 2n · (#(X)) · (max{size(Bi n), size(F i n), size(Ei n)}). (e 14.5) 147 (Note in 13.39 to 13.43, we modify the set SP (ϕn,n+1θ′ ), which will force us to change the definition of the map for the point in Sp(In+1). But now, the maps β0 and β1 in the construction of An+1 are zero on Bn+1 and {xn+1,1} is an isolated point in Sp(Bn+1) (and is also isolated in Sp(An+1)), so the modification of SP (ϕn,n+1xn+1,1) will not affect other points; see the end of 13.41 also.) Let us emphasis that, in our original construction of An+1, we can assume the Ln involved in the construction satisfying the above condition. Therefore the algebra An+1 will not be changed, when we modify the connecting maps to make the limit algebra simple. We get the following theorem: 2 Theorem 14.8. Let ((G, G+, u), K, ∆, r) be a six-tuple of the following objects: (G, G+, u) is a weakly unperforated simple order-unit group, K is a countable abelian group, ∆ is a separable Choquet simplex and r : ∆ → Su(G) is surjective affine map, where Su(G) the compact convex set of the states on (G, G+, u). Assume that (G, G+, u) has the (SP) property in the sense that for any real number s > 0, there is g ∈ G+ \ {0} such that τ (g) < s for any state τ on G. with injective ψi,i+1, where Ai = Bi ⊕ Ci, Bi ∈ H, Ci ∈ C0 with K1(Ci) = {0} such that (1) limi→∞ sup{τ (ψi,∞(1Bi )) : τ ∈ T (A)} = 0, Then there is a unital simple C-algebra A ∈ B0 which can be written as A = limn→∞(Ai, ψi,i+1) (2) ker ρA ⊂S∞i=1[ψi,∞]0(kerρBi), and (3) Ell(A) ∼= ((G, G+, u), K, ∆, r). Moreover, the inductive system (Ai, ψi) can be chosen so that ψi,i+1 = ψ(0) i,i+1 : Ai → A(0) i+1 and A(1) ψ(0) 1 A(0) i+1 injective. i,i+1 ⊕ ψ(1) i,i+1 with i+1 of Ai+1 with i+1 is a nonzero finite dimensional C-algebra, and [ψi,∞]1 is i+1 for C-subalgebras A(0) = 1Ai+1 such that A(0) i+1 and ψ(1) i,i+1 : Ai → A(1) + 1 A(0) i+1 Proof. Condition (2) follows from (e 14.3) and (e 14.4). Now the only item that has not been proved is the assertion that A ∈ B0. By (1) above and the fact that Ci ∈ C0, it remains to show that A has strict comparison for projections. This actually follows from our construction immediately. Note, by 3.18, Ci has strictly comparison. Moreover, dimensions of the underline spaces do not increase. A standard argument shows that the inductive limit has strict comparison for projections ( or for positive elements). Another way to see this is that since A is a unital simple inductive limit of subhomogeneous C-algebras with no dimension growth, it then follows from Corollary 6.5 of [103] that A is Z-stable. Hence A has strict comparison for positive elements. Remark 14.9. Note that A(0) i+1 can be chosen to be the first block B1 i+1, so we have lim i→∞ τ (ψi+1,∞(1 A(0) i+1 )) = 0 uniformly for τ ∈ T (A). Remark 14.10. Let ξn,n+1 be the partial map of ψn,n+1 : Bn → Bn+1, and let ξn,m : Bn → Bm be the corresponding composition ξm−1,m ◦ ξm−2,m−1 ◦ ··· ◦ ξn,n+1. Let en = ξ1,n(1B1). Then, from the construction, we know that the algebra B = lim−→(enBnen, ξn,m) is simple, as we know that SP (ξn,n+1xn,1) is dense enough in Sp(Bn). Note that the simplicity of B does not follow from simplicity of A itself, since it is not a corner of A. The following is an analog of Theorem 1.5 of [57]. 148 Corollary 14.11. Let A1 be a simple separable C-algebra in B1, and let A = A1 ⊗ U for an infinite dimensional UHF-algebra U . There exists an inductive limit algebra B as constructed in Theorem 14.8 such that A and B have the same Elliott invariant. Moreover, the C-algebra B satisfies the following properties: Let G0 be a finitely generated subgroup of K0(B) with decomposition G0 = G00 ⊕ G01, where G00 vanishes under all states of K0(A). Suppose P ⊂ K(B) is a finite subset which generates a subgroup G such that G0 ⊂ G ∩ K0(B). Then, for any ǫ > 0, any finite subset F ⊂ B, any 1 > r > 0, and any positive integer K, there is an F-ǫ-multiplicative map L : B → B such that: (1) [L]P is well defined. (2) [L] induces the identity maps on the infinitesimal part of G ∩ K0(B), G ∩ K1(B), G ∩ K0(B, Z/kZ) and G ∩ K1(B, Z/kZ) for k = 1, 2, ..., and i = 0, 1. (3) ρB◦[L](g) ≤ rρB(g) for all g ∈ G∩K0(B), where ρ is the canonical positive homomorphism from K0(A) to Aff(S(K0(A), K0(A)+, [1A]0)). (4) For any positive element g ∈ G01, we have g − [L](g) = Kf for some f ∈ K0(B)+. Proof. Without loss of generality, by replacing A1 by A1 ⊗ U, we may assume that Ell(A1) has (SP) property. Consider Ell(A1), which satisfies the condition of Theorem 14.8, and therefore by the first part of Theorem 14.8, there is a inductive system B1 = lim−→(Ti ⊕ Si, ψi,i+1) such that (i) Ti ∈ H and Si ∈ C0 with K1(Si) = {0}, (ii) lim τ (ϕi,∞(1Ti )) = 0 uniformly on τ ∈ T (B1), (iii) ker(ρB1 ) =S∞i=1(ψi,∞)∗0(ker(ρTi)), and (iv) Ell(B1) = Ell(A1). Put B = B1 ⊗ U . Then Ell(A) = Ell(B). Let P ∈ K(B) be a finite subset, and let G be the subgroup generated by P which contains G0. Then there is a positive integer M′ such that G ∩ K∗(B, Z/kZ) = {0} if k > M′. Put M = M′!. Then M g = 0 for any g ∈ G ∩ K∗(B, Z/kZ), k = 1, 2, ... . Let ε > 0, F ⊆ B and 0 < r < 1 be given. Choose a finite subset G ⊆ B and 0 <ǫ′ < ǫ such that F ⊆ G and for any G-ǫ′-multiplicative map L : B → B, the map [L]P is well defined, and [L] induce a homomorphism on G. By choosing a sufficiently large i0, we may assume that [ψi0,∞](K(Ti0 ⊕ Si0)) ⊃ G. In partic- ). Let G′ ⊂ K(Ti0 ⊕ Si0) ular, we may assume that, by (iii) above, G∩ kerρB1 ⊂ (ψi0,∞)∗0(kerρTi0 be such that [ψi0,∞](G′) = G. Since B = B1 ⊗ U, we may write that U = lim−→(Mm(n), ın,n+1), where m(n)m(n + 1) and ın,n+1 : Mm(n) → Mm(n+1) is defined by a 7→ a ⊗ 1m(n+1). One may assume that for each f ∈ G, there exists i > i0 such that for some f0 ∈ T ′i , fi ∈ S′i, and m > 2M K/r, where m = m(i + 1)m(i + 2)··· m(n), T ′i = ψ′i,∞(Ti ⊗ Mm(i)) and S′i = ψ′i,∞(Si ⊗ Mm(i)) and where ψi,∞ = ψi,∞ ⊗ ıi,∞. Moreover, one may assume that τ (1T ′ Choose a large n so that m = M0 + l with M0 divisible by KM and 0 ≤ l < KM . Then ) < r/2 for all τ ∈ T (A1). i define the map L : (T ′i ⊕ S′i) ⊗ Mm → (T ′i ⊕ S′i) ⊗ Mm 149 f = f0 ⊕ f1 . . . f0 ⊕ f1  ∈ (T ′i ⊕ S′i) ⊗ Mm (e 14.6) to be where El = diag(1S′ i L((fi,j ⊕ gi,j)m×m) = (fi,j)m×m ⊕ El(gi,j)m×mEl, , 1S′ , 0, 0, ..., 0 , ..., 1S′ ), which is a contractive completely positive linear map from (T ′I ⊕ S′i) ⊗ Mm to B. We then extend L to a completely positive linear map B → (1B − El)B(1B − El). Also define i l {z i {z }M0 } R : (T ′i ⊕ S′i) ⊗ Mm → T ′i ⊕ S′i to be R((fi,j ⊕ gi,j)i,j) = g1,1, and extend it to a contractive completely positive linear map B → B, where T ′i ⊕ S′i is regarded as a corner of (T ′i ⊕ S′i) ⊗ Mm ⊆ B. Then L and R are G-ǫ′-multiplicative. Hence [L]P is well defined. Moreover, τ (L(1A)) < τ (1T1 ) + l m < r 2 + M K 2M K/r = r for all τ ∈ T (A). Note that for any f in the form of (e 14.6), one has f = L(f ) + R(f ), where R(f ) may be written as Hence for any g ∈ G, R(f ) = diag{0, 0, ..., 0 }. , (0 ⊕ g1,1), ..., (0 ⊕ g1,1) l {z } M0 {z } g = [L](g) + M0[R](g). Then, if g ∈ (G0,1)+ ⊆ (G0)+, one has, g − [L](g) = M0[R](g) = K(( M0 K )[R](g)). And if g ∈ G ∩ Ki(B, Z/kZ) (i = 0, 1), one also has g − [L](g) = M0[R](g). Since M g = 0 and MM0, one has g − [L](g) = 0. Since L is identity on ψ′i,∞(Ti ⊗ Mm(i)) and i > i0, by (iii), L is identity map on G ∩ ker ρB. Since K1(Si) = 0 for all i, L induces the identity map on G ∩ K1(B). It follows that L is the desired map. Related to the above we have the following decomposition: Proposition 14.12. Let A1 be a separable amenable C-algebra in B1 (or B0) and let A = A1 ⊗ U for some infinite dimensional UHF-algebra U . Let G ⊆ A, P ⊆ K(A) be finite subsets, P0 ⊂ A ⊗ K be a finite subset of projections, and let ǫ > 0, 0 < r0 < 1 and M ∈ N be arbitrary. Then there is a projection p ∈ A, a C∗-subalgebra B ∈ C (or in C0) with p = 1B and G-ǫ- multiplicative unital contractive completely positive linear maps L1 : A → (1 − p)A(1 − p) and L2 : A → B such that (1) kL1(x) + L2(x) − xk < ε for all x ∈ G; 150 (2) [Li]P is well defined, i = 1, 2; (3) [L1]P + [ı ◦ L2]P = [id]P ; (4) τ ◦ [L1](g) ≤ r0τ (g) for all g ∈ P0 and τ ∈ T (A); (5) For any x ∈ P, there exists y ∈ K(B) such that x − [L1](x) = [ı ◦ L2](x) = M [ı](y) and, (6) for any d ∈ P0, there exist positive element f ∈ K0(B)+ such that d − [L1](d) = [ı ◦ L2](d) = M [ı](f ), where ı : B → A is the embedding. Moreover, we can require that 1 − p 6= 0. Proof. Since A is in B1 (or B0), there is a sequence of projections pn ∈ A and a sequence of C∗-subalgebra Bn ∈ B1 (B0) with 1Bn = pn such that lim n→∞k(1 − pn)a(1 − pn) + pnapn − ak = 0, lim dist(pnapn, Bn) = 0 and n→∞ max{τ (1 − pn) : τ ∈ T (A)} = 0 lim n→∞ (e 14.7) (e 14.8) (e 14.9) for all a ∈ A. Since each Bn is amenable, one obtains easily a sequence of unital contractive completely positive linear maps Ψn : A → Bn such that lim n→∞kpnapn − Ψn(a)k = 0 for all a ∈ A. In particular, lim n→∞kΨn(ab) − Ψn(a)Ψn(b)k = 0 for all a, b ∈ A. (e 14.10) (e 14.11) Let j : A → A⊗ U be defined by j(a) = a⊗ 1U . There is a unital homomorphism s : A⊗ U → A and a sequence of unitaries un ∈ A ⊗ U such that lim n→∞ka − Ad un ◦ s ◦ j(a)k = 0 for all a ∈ A. (e 14.12) (e 14.13) (e 14.14) There are nonzero projection e′n ∈ U and en ∈ U such that z t(en) = 0 and 1 − en = diag( lim n→∞ M } e′n, e′n, ..., e′n), { where t ∈ T (U ) is the unique tracial state on U. Choose N ≥ 1 such that 0 < t(en) < r0/2 and max{τ (1 − pn) : τ ∈ T (A)} < r0/2. Define Φn : A → (1 − pn)A(1 − pn) by Φn(a) = (1 − pn)a(1 − pn) for all a ∈ A. Define Φ′n(a) = Φn(a) ⊕ Ad un ◦ s(a ⊗ en) and Ψ′n(a) = Ad un ◦ s(Ψ(a) ⊗ (1 − en)) for all n ≥ N. Note that u∗ns(Bn ⊗ (1 − en))un ∈ C1 (or in C0). It is then easy to verify that, if we choose a large n, the maps L1 = Φ′n and L2 = Ψ′n meet the requirements. 151 15 Positive maps from K0-group of C-algebras in the class C. This section contains some technical lemmas about positive homomorphisms from K0(C) for some C ∈ C. Lemma 15.1. (Compare 2.8 of [53]) Let G ⊂ Zl (for some l > 1) be a subgroup. There is an integer M > 0 satisfying the following condition: Let 1 > σ1, σ2 > 0 be any given numbers. There is an integer R > 0 such that: if a set of l positive numbers αi ∈ R+ (i = 1, 2,··· l) satisfy αi ≥ σ1 for all i and satisfy lXi=1 αimi ∈ Z for all (m1, m2, ..., ml) ∈ G, (e 15.1) then for any integer K ≥ R, there exist a set of rational positive numbers βi ∈ 1 1, 2,··· l) such that KM Z+ (i = lXi=1 αi − βi < σ2, i = 1, 2, ..., l and ϕG = ϕG, (e 15.2) i=1 αini and ϕ((n1, n2, ..., nl)) =Pl where ϕ((n1, n2, ..., nl)) =Pl Zl. Proof. Denote by ej ∈ Zl the element having 1 in the j-th coordinate and 0 elsewhere. First we consider the case that Zl/G is finite. In this case there is an integer M ≥ 1 such that M ej ∈ G for all j = 1, 2, ..., l. It follows that ϕ(M ej) ∈ ϕ(G)⊂ Z, j = 1, 2, ..., l. Hence αj = ϕ(ej) ∈ 1 M Z+. We choose βj = αj, j = 1, 2, ..., l, and ϕ = ϕ. The lemma follows—that is, for any σ1, σ2, we can choose R = 1. i=1 βini for all (n1, n2, ..., nl) ∈ Now we assume that Zl/G is not finite. Regard Zl as a subset of Ql and set H0 to be the vector subspace of Ql spanned by elements in G. The assumption that Zl/G is not finite implies that H0 has dimension p < l. Moreover G ∼= Zp. Let g1, g2, ..., gp ∈ G be free generators of G. View them as elements in H0 ⊂ Ql and write gi = (gi,1, gi,2,··· , gi,l), i = 1, 2, ..., p. (e 15.3) Define L : Qp → Ql to be L = (fi,j)l×p, where fi,j = gj,i, i = 1, 2, ..., l and j = 1, 2, ..., p. Then L∗ = (gi,j)p×l. We also view L∗ : Ql → Qp. Define T = L∗L : Qp → Qp which is invertible. Note that T = T ∗ and (T −1)∗ = T −1. Note also that the matrix representation (ai,j)l×p of L ◦ T −1 is an l × p matrix with entries in Q. There is an integer M1 ≥ 1 such that ai,j ∈ 1 M1 Z, i = 1, 2, ..., p and j = 1, 2, ..., l. Let H00 = kerL∗. It has dimension l − p > 0. Let P : Ql → H00 be an orthogonal projection which is a Q-linear map. Represent P as a l × l matrix. Then its entries are in Q. There is an integer M2 ≥ 1 such that all entries are in 1 It is important to note that M1 and M2 depend on G only and are independent of {αj : 1 ≤ j ≤ l}. Let M = M1M2. Suppose that σ1, σ2 ∈ (0, 1) are two positive numbers and the numbers αi ≥ σ1 (i = 1, 2,··· , l) satisfy the condition (e 15.1). j=1 αjgi,j ∈ Z, i = 1, 2, ..., p. Put The condition (e 15.1) is equivalent to that b = (b1, b2, ..., bp)T and α = (α1, α2, ..., αl)T . Then b = L∗α. M2 Z. We will use the fact that L∗ = L∗(1 − P ). bi: = Pl 152 If we write L(T ∗)−1b = c = c1 c2 ... cl  , (e 15.4) then, since b ∈ Zp, one has that ci ∈ 1 Let K ≥ R be any integer. Note that M1 Z. Choose an integer R ≥ 1 such that 1/R < σ1σ2/(4l2). L∗c = L∗L(T ∗)−1b = L∗LT −1b = L∗α. (e 15.5) Thus α − c ∈ kerL∗ as a subspace of Rl. For the space Rl, we use k · k1 and k · k2 to denote the l1 and l2 norm on it. Then we have kvk2 ≤ kvk1 ≤ lkvk2 for all v ∈ Rl. Since H00 is dense in the real subspace of kerL∗, there exists ξ ∈ H00 such that kα − c − ξk2≤ kα − c − ξk1 < σ1σ2/4l. Pick η ∈ Ql such that ξ = P η. Then there is η0 ∈ Ql such that Kη0 ∈ Zl and Since P has norm 1 with respect to l2 norm, kη0 − ηk2 ≤kη0 − ηk1 < σ1σ2/2l. kα − c − P η0k1≤ lkα − c − P η0k2 ≤ l(kα − c − ξk2 + kP (η0 − η)k2) < σ1σ2. Put β = c + P η0 = (β1, β2, ..., βl)T . Note that M2K(P η0) ∈ Zl , and that M1c ∈ Zl. We have KM β ∈ Zl, and L∗β = L∗c = L∗α and kα − βk1 < σ1σ2. Moreover, since αi ≥ σ1, βi > 0, i = 1, 2, ..., l. (e 15.6) (e 15.7) (e 15.8) (e 15.9) (e 15.10) Since P η0 ∈ H00, one has that L∗β = L∗(1 − P )β = L∗(1 − P )c = L∗c = L∗α = b. Define ϕ : Ql → Q by for all x ∈ Ql. Note L∗ei = gi, where ei is the element in Zp with the i-th coordinate being 1 and 0 elsewhere. So ϕ(x) = hx, βi (e 15.11) ϕ(gi) = hLei, αi = hei, L∗αi = hei, L∗βi = hLei, βi = hgi, βi = ϕ(gi), (e 15.12) i = 1, 2, ..., p. It follows that ϕ(g) = ϕ(g) for all g ∈ G. Hence ϕG = ϕG. Note that ϕ(Zl) ⊂ KM Z, since βi ∈ 1 KM Z+, i = 1, 2, ..., l. 1 153 M = M1 which only depends on G and l (by replacing β by c in the proof). If we do not need to approximate {αi : 1 ≤ i ≤ l}, then R can be chosen to be 1, with From the proof 15.1, since L and (T ∗)−1 depends only on g1, g2, ..., gp, we have the following: Lemma 15.2. Let G ⊂ Zl be an ordered subgroup with order unit e, and let g1, g2, ..., gp (p ≤ l) be a set of free generators of G. For any ε > 0, there exists δ > 0 satisfying the following: if ϕ : G → R is a homomorphism such that then, there is β = (β1, β2, ..., βl) ∈ Rl with βi < ε, i = 1, 2, ..., l, such that ϕ(gi) < δ, i = 1, 2, ..., p, ϕ(g) = ψ(g) for all g ∈ G, where ψ : Zl → R defined by ψ((m1, m2, ..., ml)) =Pl Corollary 15.3. Let G ⊂ Zl be an order subgroup. Then, there exists an integer M ≥ 1 satisfying the following: for any positive map κ : G → Zn (for any integer n ≥ 1) with every element in κ(G) divisible by M , there is R0 ≥ 1 such that, for any integer K ≥ R0, there is a positive homomorphism κ : Zl → Zn such that κG = Kκ. Proof. We first prove the case that n = 1. i=1 βimi for all mi ∈ Z. Let S ⊂ {1, 2, ..., l} be a subset and denote by Z(S) the subset Z(S) = {(m1, m2, ..., ml) : mi = 0 if i 6∈ S}. every element in κ(G) is divisible by M . Let ΠS : Zl → Z(S) be the projection and G(S) = ΠS(G). Let M (S) be the integer (in place of M ) as in 15.1 associated with G(S) ⊂ Z(S). Put M =QS⊂{1,2,...,l} M (S). Now assume that κ : G → Z is a positive homomorphism with multiplicity M —that is By applying 2.8 of [53], we obtain a positive homomorphism β : Zl → R such that βG = κ. Define fi = β(ei), where ei is the element in Zl with 1 at the i-th coordinate and 0 elsewhere, i = 1, 2, ..., l. Then fi ≥ 0. Choose S such that fi > 0 if i ∈ S and fi = 0 if i 6∈ S. Evidently if ξ1, ξ2 ∈ Zl satisfy ΠS(ξ1) = ΠS(ξ2), then β(ξ1) = β(ξ2), and if we further assume ξ1, ξ2 ∈ G then κ(ξ1) = κ(ξ2). Hence the maps β and κ induce the maps β′ : Z(S) → R and κ′ : G(S) → Z such that β = β′ ◦ ΠS and κ = κ′ ◦ ΠS. In addition, we have β′G(S) = κ′. . Apply 15.1 to αi = fi/M > σ1 for i ∈ S and to G(S) ⊂ Z(S), we obtain the number R(κ) (depending on σ1 and σ2 and therefore depending on κ) as in the lemma. For any K ≥ R(κ), it follows from the lemma there are βi ∈ 1 KM )+ (for i ∈ S), such M κ′, where κ′ : Z (S) → Q is defined by κ′({mi}i∈S ) = Pi∈S βimi. Evidently that κ′G(S) = 1 κ = KM (κ′ ◦ ΠS)Zl → Z is as desired for this case. This prove the case n = 1. In general, let si : Zn → Z be the projection to the i-th summand, i = 1, 2, ..., n. Apply the case n = 1 to each of the the maps κi := si ◦ κ (for i = 1, 2,··· n) to obtain R(κi). And let R0 = maxi R(κi). For any K ≥ R0, by what has been proved, we obtain κi : Zl → Z such that (e 15.13) κiG = Ksi ◦ κG, i = 1, 2, .... Let σ1 = 2σ2 = min{fi: i∈S} 2M Define κ : Zl → Zn by κ(z) = (κ1(z), κ2(z), ..., κl(z)). The lemma follows. 154 G ϕ / G ϕ / ··· / lim−→ G. Lemma 15.4 (Lemma 3.2 of [30]). Let G = K0(S), where S ∈ C and let r : G → Z be a strictly positive homomorphism. Then, for any order unit u ∈ G+, there exists a natural number m such that if the map θ : G → G is defined by g 7→ r(g)u, then there exists an integer m ≥ 1 such that the positive homomorphism id + mθ : G → G factors through Ln Proof. Let u be an order unit of G, and define the map ϕ : G → G by ϕ(g) = g + r(g)u; that is, ϕ = id +θ. Define Gn = G and ϕn : G → G by ϕn(g) = ϕ(g) for all g and n. Consider the inductive limit i=1 Z positively for some n. Then the ordered group lim−→ G has the Riesz decomposition property. In fact, let a, b, c ∈ lim−→ G+ such that Without loss of generality, one may assume that a 6= b + c. b and ϕn,∞(c′) = c, and furthermore We may assume that there are a′, b′, c′ ∈ G for the n-th G such that ϕn,∞(a′) = a, ϕn,∞(b′) = a ≤ b + c. a′ < b′ + c′. (e 15.14) A straightforward calculation shows that for each k, there is m(k) ∈ N such that ϕn,n+k(a′) = a′+m(k)r(a′)u, ϕn,n+k(b′) = b′+m(k)r(b′)u, and ϕn,n+k(c′) = c+m(k)r(c′)u. Moreover, the sequence (m(k)) is strictly increasing. Since r is strictly positive, combing with (e 15.14), we have that r(a′) < r(b′) + r(c′) (in Z). There are l(a′)i ∈ Z+ such that l(a′)1 + l(a′)2 = r(a′), l(a′)1 ≤ r(b′) and l(a′)2 ≤ r(c′). Without loss of generality, we may assume d = r(b′) − l(a′)1 > 0 (otherwise we let d = r(c′) − l(a′)2). Since u is an order unit, there is m1 ∈ Z+ such that m1du > a′. Choose k ≥ 1 such that m(k) > m1. Let a1 = a′ + m(k)l(a′)1u and a2 = m(k)l(a′)2u. Then a2=m(k)l(a′)2u ≤ m(k)r(c′)u ≤ c′ + m(k)r(c′)u = ϕn,n+k(c′). Moreover, a1 = a′ + m(k)l(a′)1u ≤ m(k)du + m(k)l(a′)1u ≤ b′ + m(k)r(b′)u = ϕn,n+k(b′). Note These imply that ϕn,n+k(a′) = a1 + a2 ≤ ϕn,n+k(b′) + ϕn,n+k(c′). a=ϕn+k,∞(a1) + ϕn+k,∞(a2) ≤ b + c, ϕn+k,∞(a1) ≤ b and ϕn+k,∞(a2) ≤ c. (e 15.15) (e 15.16) This implies that the limit group lim−→ G has the Riesz decomposition property. Since G is unperforated, so is lim−→ G. It then follows from the Effros-Handelman-Shen Theorem ([20]) that the ordered group lim−→ G is a dimension group. Therefore, for a sufficiently large k ∈ N, the map ϕk must factor through the ordered groupLn Z positively for some n. Since ϕk has the desired form id +m(k)θ, the lemma follows. 155 / / / Lemma 15.5. Let (G, G+, u) be an group with order unit u such that the positive cone G+ is generated by finitely many positive elements which are smaller than u. Let λ : G → K0(A) be an order preserving map such that λ(u) = [1A] and λ(G+ \ {0}) ⊂ K0(A)+ \ {0}, where A ∈ B1 (A ∈ B0). Let a ∈ K0(A)+ \ {0} with a ≤ [1A]. Let P ⊂ G+ \ {0} be a finite subset. Suppose that there exists an integer N ≥ 1 such that N λ(x) > [1A] for all x ∈ P. There are two positive homomorphisms λ0, λ1 : G → K0(A) and a C∗-subalgebra S′ ⊂ A with S′ ∈ C (S′ ∈ C0) satisfying the following: λ = λ0 + λ1, λ1 = ı∗0 ◦ γ, 0 ≤λ0(u) < a and γ(g) > 0 (e 15.17) for all g ∈ G+ \ {0}, where γ : G → K0(S′) with γ([u]) = [1S′] and where ı : S′ → A is the embedding. Moreover, N γ(x) ≥ γ(u) for all x ∈ P. Furthermore, if A = A1 ⊗ U , where U is an infinite dimensional UHF-algebra and A1 ∈ B1 (or B0), then, for any integer K ≥ 1, we can require that S′ = S ⊗ MK for some S ∈ C (or C0) and γ has multiplicity K. Proof. Let {g1, g2, ..., gm} ⊂ G+ be the set of generators of G+ with gi < u. Since A has stable rank one, it is easy to check that there are projections q1, q2, ..., qm ∈ A such that λ(gi) = [qi], i = 1, 2, ..., m. For convenience, to simplify notation, without loss of generality, we may assume that P = {g1, g2, ..., gm}. Define By the assumption, there are vi ∈ MN (A) such that Qi = diag( qi, qi, ..., qi), i = 1, 2, ..., m. z N } { v∗i vi = 1A and viv∗i ≤ Qi, i = 1, 2, ..., m. (e 15.18) Since A ∈ B1 (or B0), there exists a sequence of projections {pn} of A, a sequence of C∗- subalgebra Sn ∈ C (C0) with pn = 1Sn and a sequence of unital contractive completely positive linear maps Ln : A → Sn such that lim n→∞ka − ((1 − pn)a(1 − pn) + pnapn)k = 0, and lim n→∞ sup τ∈T (A){τ (1 − pn)} = 0. lim n→∞kLn(a) − pnapnk = 0 (e 15.19) (e 15.20) It is also standard to find, for each i, a projection e′i,n ∈ (1 − pn)A(1 − pn), a projection ei,n ∈ MN (Sn) and partial isometries wi,n ∈ MN (Sn) such that (e 15.21) (e 15.22) n→∞k(1 − pn)qi(1 − pn) − e′i,nk = 0, lim w∗i,nwi,n = pn, wi,nw∗i,n ≤ ei,n, n→∞k(Ln⊗idMN )(vi) − wi,nk = 0 and lim lim n→∞k(Ln ⊗ idMN )(Qi) − ei,nk = 0.(e 15.23) Let Ψn : A → (1 − pn)A(1 − pn) be defined by Ψn(a) = (1 − pn)a(1 − pn) for all a ∈ A. We will use [Ψn]◦λ for λ0 and [Ln◦λ] for γ for some large n. The fact that λ0 and γ are homomorphisms follows from Lemma 7.1 of [60]. To see that λ0 is positive, we use (e 15.21) and the fact that G+ is finitely generated. It follows from (e 15.22) and (e 15.23) that N γ(x) ≥ γ(u) for all x ∈ P. Since we assume that the positive cone of G+ is generated by P, this also shows that γ(x) > 0 for all x ∈ G+ \ {0}. By (e 15.20), we can choose large n so that 0 ≤λ0(u) < a. It should be noted when A does not have (SP), one can choose λ = λ1, and λ0 = 0. If A = A1 ⊗ U, then, without loss of generality, we may assume that pn ∈ A1 for all n. Choose a sequence of nonzero projections en ∈ U such that t(1 − en) = r(n)/K, where t is the unique tracial state on U and r(n) are positive rational numbers such that limn→∞ t(en) = 0. Thus Sn ⊗ (1 − en) ⊂ Bn where Bn ∼= Sn ⊗ MK and pn ⊗ (1 − en) = 1Bn . We check that the lemma follows if we replace Ψn by Ψ′n, where Ψ′n(a) = (1 − pn)a ⊗ 1U + pna ⊗ en 156 Lemma 15.6. (see Lemma 3.6 of [30] or Lemma 2.8 of [83]) Let G = K0(S), where S ∈ C. Let H = K0(A) for A = A1 ⊗ U, where A1 ∈ B1 (or B0) and U is an infinite dimensional UHF- algebra. Let M1 ≥ 1 be a given integer and d ∈ K0(A)+ \ {0}. Then for any strictly positive homomorphism θ : G → H with multiplicity M1, and any integers M2 ≥ 1 and K ≥ 1 such that Kθ(x) > [1A] for all x ∈ G+ \ {0}, one has a decomposition θ = θ1 + θ2, where θ1 and θ2 are positive homomorphisms from G to H such that the following diagrams commute: G ❆❆❆❆❆❆❆❆ ϕ1 θ1 G1 H >⑤⑤⑤⑤⑤⑤⑤⑤ ψ1 G ❆❆❆❆❆❆❆❆ ϕ2 θ2 G2 , H >⑤⑤⑤⑤⑤⑤⑤⑤ ψ2 where θ1([1S ]) ≤ d, G1 ∼= Ln Z for some natural number n and G2 = K0(S′) for some C∗- subalgebra S′ of A which is in the class C (or in C0), ϕ1, ψ1 are positive homomorphisms and ψ2 = ı∗0, where ı : S′ → A is the embedding. Moreover, ϕ1 has the multiplicity M1, ϕ2 has the multiplicity M1M2, 2Kϕ2(x) > ϕ2([1S ]) > 0 for all x ∈ G+ \ {0}. Proof. Let u = [1S], and let m be as in Lemma 15.4. Suppose that S = A(F1, F2, ψ0, ψ1) with F1 = MR1⊕MR1⊕···⊕MRl. It is easy to find a strictly positive homomorphism η0 : K0(F1) → Z. Define r : G → Z by r(g) = η0 ◦ (πe)∗0. By replacing S with Mr(S) and A by Mr(A) for some integer r ≥ 1, without loss of generality we may assume that S has a finite subset of projections P = {p1, p2, ..., pl} such that every projection q ∈ S is equivalent to one of projections in P and {[pi] : 1 ≤ i ≤ l} generates K0(S)+ (see 3.15). Let σ0 = min{ρA(d)(τ ) : τ ∈ T (A)}. Note that since A is simple, one has that σ0 > 0. Let σ1 = inf{τ (θ([p]) : p ∈ P, τ ∈ T (A)} > 0. Since A = A1 ⊗ U, A has the (SP) property, there is a projection f0 ∈ A+ \ {0} such that 0 < τ (f0) < min{σ0, σ1}/8N r(u) for all τ ∈ T (A). (e 15.24) Since A = A1 ⊗ U, we may choose f0 so that f0 = M1h for some nonzero h ∈ K0(A)+. Put θ′0 : G → K0(A) by θ′0(g) = r(g)h for all g ∈ G. And let θ′ = M1θ′0. Then 2θ′(x) < θ(x) for all x ∈ G+ \ {0}. Since θ has multiplicity M1, one has that θ(g) − θ′(g) is divisible by M1 for any g ∈ G. By the choice of σ0, one checks that θ − θ′ is strictly positive. Moreover, 2KρA((θ(x) − θ′(x))(τ ) > 2KρA(θ(x))(τ ) − KρA(θ(x))(τ )=KρA(θ(x))(τ ) (e 15.25) (e 15.26) Applying 15.5, one obtains a C∗-subalgebra S′ ∈ A, a homomorphism θ1 : G → K0(A) and ≥ ρA([1A])(τ ) for all τ ∈ T (A). strictly positive homomorphism ϕ2 : G → K0(S′) such that θ − θ′ = θ1 + ı∗0 ◦ ϕ2, τ (θ1(u)) < τ (h) mM1M2 , τ ∈ T (A), 2Kϕ2(x) > ϕ2([1S ]), ϕ2([1S ]) = [1S′], where m is from 15.4 and ϕ2 has multiplicity M1M2, and where ı : S′ → A is the embedding. Put θ2 = ı∗0 ◦ ϕ2, and ψ2 = ı∗0. 157 (e 15.27) (e 15.28) / / / / > > Since θ(g) − θ′(g) is divisible by M1 and any element in θ2(G) is divisible by M1, one has that any elements in θ1(G) is divisible by M1. Therefore, the map θ1 can be decomposed further as M1θ′1, and one has that θ − θ′ = M1θ′1 + θ2. Therefore, there is a decomposition θ = θ′ + M1θ′1 + θ2 = M1θ′0 + M1θ′1 + θ2. Put Then, θ1 = M1(θ′0 + θ′1). We then show that θ1 has the desired factorization property. For θ′0+θ′1, one has the following ρA(θ1([1S]))(τ ) < τ (d)/2 for all τ ∈ T (A). farther decomposition: for any g ∈ G, θ′0(g) + θ′1(g) = r(g)h + θ′1(g) = r(g)(h − mθ′1(u)) + r(g)mθ′1(u) + θ′1(g) = r(g)(h − mθ′1(u)) + θ′1(mr(g)u) + θ′1(g) = r(g)(h − mθ′1(u)) + θ′1(mr(g)u + g). By (e 15.28), h− mθ′1(u)) > 0. By Lemma 15.4, g → mr(g)u + g factors thoughLn Z positively for some n. Therefore, the map M1(θ′0 + θ′1) factors though L(1+n)M1 Z positively. So there are positive homomorphisms ϕ1 : G →L(1+n)M1 Z and ψ1 :L(1+n)M1 Z. → K0(A) such that θ1 = ψ1 ◦ ϕ1 and ϕ1 has multiplicity of M1. 16 Existence Theorems for affine maps on tracial state spaces for building blocks Lemma 16.1. Let A be a unital separable C∗-algebra with T (A) 6= ∅ and let H ⊂ A be a finite subset. Then, for any σ > 0, there exists an integer N > 0 and a finite subset E ⊂ ∂e(T (A)) satisfying the following: For any τ ∈ T (A) and any k ≥ N , there are {t1, t2, ..., tk} ⊂ E such that τ (h) − 1 k kXi=1 ti(h) < σ for all h ∈ H. (e 16.1) (If τ is a (possibly unnormalized) trace on A with kτk ≤ 1, then there are {t1, t2, ..., tk′} with k′ ≤ k such that Suppose that A is a subhomogeneous C-algebras. Then, there are {π1, π2, ..., πk} ⊆ Irr(A) such that τ (h) − 1 k (tr1 ◦ π1(h) + tr2 ◦ π2(h) + ··· + trk ◦ πk(h)) < σ for all h ∈ H, (e 16.2) where πj ∈ E and trj is the canonical trace of πj(A). Moreover, if, for each l, Al has no isolated points, then {π1, π2, ..., πk} can be required to consist of distinct points, where Al is the set of all irreducible representations of A with rank exactly l. Remark: Note that in (e 16.2), πi may not be distinct. But the subset E of irreducible representations can be chosen independent of τ (but dependent of σ and H). 158 τ (h) − ti(h) < σ for all h ∈ H.) 1 k k′Xi=1 Proof. Without loss of generality, one may assume that kfk ≤ 1 for all f ∈ H. Note the tracial state space T (A) is weak - compact. Therefore, there are τ1, τ2, ..., τm ∈ T (A) such that, for any τ ∈ T (A), there is j ∈ {1, 2, ..., m} such that τ (f ) − τj(f ) < σ/4 for all f ∈ H. (e 16.3) Note that the set of extreme points of T (A) is the set of those tracial states induced by irreducible representations of A. By the Krein-Milman Theorem, there are t′1, t′2, ..., t′n ∈ ∂e(T (A)) and nonnegative numbers {αi,j} such that τj(f ) − nXi=1 αi,jt′i(f ) < σ/8 and nXi=1 αi,j = 1. (e 16.4) Put E = {t′1, t′2, ..., t′n}. Choose N > 32mn/σ. Let τ be possibly unnormalized trace on A with 0 < τ (1) ≤ 1. Suppose that j is so chosen that kτ (f )/τ (1) − τj(f )k < σ/4 for all f ∈ H as in e 16.3. Then, for any k ≥ N, there exist positive rational numbers ri,j and positive integers pi,j (1 ≤ i ≤ n and 1 ≤ j ≤ m) such that, nXi=1 pi,j k nXi=1 ri,j ≤ 1, or nXi=1 ri,j = 1 if τ (1) = 1 and τ (1)αi,j − ri,j < σ 8n , 1 ≤ i ≤ n, pi,j ≤ k or nXi=1 pi,j = k if τ (1) = 1. ri,j = Note that Then, by (e 16.6), τ (f ) − pi,j k ( nXi=1 )t′i(f ) < σ/4 + σ/8 = 3σ/8 for all f ∈ H. (e 16.5) (e 16.6) (e 16.7) It is then clear that (e 16.1) holds by repeating each t′i pi,j times. Irr(A). It follows that (e 16.2) holds. Now suppose that A is subhomogeneous. Then every t′i has the form tri◦πi, where {π1, π2, ..., πn} ⊂ There exists δ > 0 such that for any irreducible representation x, y ∈ Al with dist(x, y) < δ, we have f (x) − f (y) < σ/64k for all f ∈ H. If Al has no isolated points, for each i, choose πi,j distinct points in a neighborhood O(π′i, δ) of π′i (in Al) with diameter less than δ. Let {π1,j, π2,j, ..., πk,j} be the resulting set of k elements (see (e 16.7)). Then, one has τ (f ) − 1 k (f (π1,j) + f (π2,j) + ··· + f (πk,j)) < σ for all f ∈ H, (e 16.8) as desired. The following is well known. 159 Lemma 16.2. Let C =Lk i=1 C(Xi) ⊗ Mr(i), where each Xi is connected compact metric space. Let H ⊆ C be a finite subset and let σ > 0. Then there is an integer N ≥ 1 satisfying the following: for any positive homomorphism κ : K0(C) → K0(Ms) = Z with κ([1Mr(i)]) ≥ N and any τ ∈ T (C) such that ρC(x)(τ ) = tr(κ(x)) for all x ∈ K0(C), where tr is the tracial state on Ms, there is a homomorphism ϕ : A → Ms such that ϕ∗0 = κ and tr ◦ ϕ(h) − τ (h) < σ for all h ∈ H. Lemma 16.3. Let C = C(T) ⊗ F1, where F1 = MR(1) ⊕ MR(2) ⊕ ··· ⊕ MR(l), or C = F1. Let H ⊆ C be a finite subset, and let ǫ > 0. There is δ > 0 satisfying the following: For any Ms, any order-unit map κ : K0(C) → K0(Ms) and any tracial state τ ∈ T (C) such that ρMs(κ(p))(tr) − τ (p) < δ for all projections p in C, where tr is the tracial state on Ms, there is a tracial state τ ∈ T (C) such that and tr(κ([p])) = τ (p), τ (h) − τ (h) < ǫ for all h ∈ H. Proof. We may assume that H is in the unit ball of C. Let δ = ε/l. We may write that τ = j=1 αj = 1. Let βj = j=1 βjτj. Then tr(κ(p)) = τ (p) for all projections p ∈ C; Pl j=1 αjτj, where τj is a tracial state on C(T) ⊗ Mr(j), αj ∈ R+ and Pk tr(κ([1MR(j) ]), j = 1, 2..., l. Put τ =Pl and for any h ∈ H, τ (h) − τ (h) ≤ as desired. lXj=1 βj − αj < ε, e e e ◦ π(2) : Am → Am−1, Πk = π(k+1) : Am → A0 be the quotient map. In the following statement, we keep the notation in 4.8 for C∗-algebra A = Am ∈ Dm. In : Am → Ak and particular, Πm−1 = π(m) Π0 = π(1) e ◦ ··· ◦ π(m) Lemma 16.4. Let A ∈ Dm be a unital C∗-algebra for some m ≥ 0, and let ∆ : Aq,1 + \{0} → (0, 1) be an order preserving map. Let H ⊂ A be a finite subset, let P ⊂ Mm′(A) be a finite subset of projections (for some integer m′ ≥ 1) and let ε > 0. Then there are ∆0 : (A0)q,1 + \ {0} → (0, 1), a finite subset H1 ⊆ A1 + \ {0} and a positive integer K such that for any τ ∈ T (A) satisfying ◦ ··· ◦ π(m) e ◦ π(k+2) e e τ (h) > ∆(h) for all h ∈ H1 and any positive homomorphism κ : K0(A) → Z with s = κ([1A])∈ N such that ρA(x)(τ ) = (1/s)(κ(x)) for all x ∈ K0(A), (e 16.9) (e 16.10) there are rational numbers r1, r2, .., rk ≥ 1, finite dimensional irreducible representations π1, π2, ..., πk of A, πk+1, πk+2, ..., πk+l of A0 and a trace τ0 on A0 such that, for all h ∈ H, ritri ◦ πi(h) + τ0 ◦ Π0(h)) < ε/2, τ0 ◦ Π0(h)) − ritriπi ◦ Π0(h) < ε/2,(e 16.11) τ (h) − ( kXi=1 τ (p) = kXi=1 ritri ◦ πi(p) + τ0 ◦ Π0(p) = kXi=1 160 ritri ◦ πi(p) + ritriπi ◦ Π0(p) (e 16.12) k+lXi=k+1 k+lXi=k+1 for all p ∈ P; τ0 ◦ Π0(h) ≥ ∆0(\Π0(h)) for all h ∈ H1; (e 16.13) and moreover, Kri ∈ Z for i = 1, 2,··· , k, and sKri ∈ Z, for i = k + 1, k + 2,··· k + l. Proof. Without loss of generality, we may assume that A ∈ D′m. Write Am and keep the notation in 4.8 for Am. So Xm = X. Put Ym = Xm \ Zm, and Yk = Xk \ Zk, k = 0, 1, ..., m. Let π(k) : Ak → Ak−1, λk : Ak → PkC(Xk, Fk)Pk and Γk+1 : Ak → Pk+1C(Zk+1, Fk+1)Pk+1 be as e defined in 4.8. Let Πk : Am → Ak be defined by π(k+1) for k = 0, 1, 2, ..., m − 1. Let Ik = QkC0(Yk, Fk)Qk ⊂ Ak, k = 1, 2, ..., m, where Qk = PkYk , as in 4.8 . Note that Ik = kerπ(k) e . Note also that we assume that Pk(x) 6= 0 for all x ∈ Xk, k = 1, 2, ..., m. By replacing A by MN (A), without loss of generality, we may assume that K0(A) is generated by {p1, p2, ..., pc}, where pi ∈ A are projections, i = 1, 2, ..., c. Moreover, we may assume that, without loss of generality, P = {p1, p2, ..., pc} and P ⊂ H. To simplify notation, we may further j=1 Mr(0,j) and assume that ej ∈ Mr(0,j) is a rank one j = {x ∈ Xj : dist(x, Zj) < d}. Choose δ0 > 0 such that there is assume that p1 = 1A. Write A0 = Ll(0) Recall for any d > 0, X d ◦ π(k+2) ··· π(m) projection. e e e continuous map s(j,δ0) : X δ0 j → Zj satisfying ∗ (x) = x for all x ∈ Zj, dist(s(j,δ0) s(j,δ0) ∗ kΓj(Πj−1(h))(s(k,δ0) ∗ (x), x) ≤ δ0 for all x ∈ X δ0 (y)) − λj(Πj(h))(y)k < ε/2m+1 for all h ∈ H ∗ j (e 16.14) (e 16.15) and y ∈ X δ0 j = {y ∈ Xj : dist(y, Zj) < δ0}, j = 1, 2, ..., m. Write PkC(Xk, Fk)Pk = l(k)Mj=1 Pk,j(C(Xk, Ms(k,j))Pk,j, where Pk,j ∈ C(Xk, Ms(k,j)) is a projection of rank r(k, j) at each x ∈ Xk, (k = 0, 1, 2, ..., m) and r(i) = r(0, i), i = 1, 2, ..., l. Note that (K0(Ak))/ ker(ρAk ) ⊂ ρAk−1(K0(Ak−1))/ ker(ρAk−1) ⊂ Zl (see 4.10). Let Sk,i : PkC(Xk, Fk)Pk → Pk,iC(Xk, Ms(k,j))Pk,i be the quotient map, i = 1, 2, ..., l(k), k = 0, 1, 2, ..., m. Choose χk ∈ (Ik)+ such that kχkk ≤ 1, χk(x) is a scalar multiple of λk(1Ak )(x) for all x ∈ Yk, χk(x) = λk(1Ak )(x) if dist(x, Zk) > δ0 and χk(x) = 0 if dist(x, Zk) < δ0/2 and for all x ∈ Xk, k = 1, 2, ..., m. 6= ∅, define an element bk ∈ C0(Yk) such that bk(x) = 0 if x ∈ Zk, bk(x) > 0 for all x ∈ Yk ∩ X δ0 k . and Π0(hj ) = ej and h(k) Note that Yk ∩ X δ0 Let bk,i = Sk,i(bk · Qk)∈ Pk,iC0(Yk, Ms(k,i))Pk,i, i = 1, 2, ..., l(k). Let hj ∈ A+ with khjk ≤ 1 k j ∈ A+ with kh(k) j ) = (1 − χk)Πk(hj) = (1 − χk)1/2Πk(hj )(1 − χk)1/2, j k ≤ 1 such that Πk(h(k) j = 1, 2, ..., l(0) and k = 0, 1, 2, ..., m−1. Let H0 = {hj : 1 ≤ j ≤ l(0)}. For each b ∈ Ak−1= Ak/Ik define Lk(b) = (1 − χk)sk,δ0(b) = (1 − χk)1/2sk,δ0(b)(1 − χk)1/2∈ Ak, for k = 1, 2, ..., m, where sk,δ0 : Ak/Ik → C(X δ0 for all a ∈ Ak/Ik = Ak−1 and x ∈ X δ0 positive linear map. k , Fk) defined by sk,δ0(a)(x) = Γk(a)(sk,δ0 (x)) k . Note that Lk : Ak−1 → Ak is a contractive completely ∗ 161 For each h ∈ Ak−1, define h′(k) = Lm ◦ Lm−1 ◦ ··· ◦ Lk(h)∈ Am, k = 1, 2, ..., m. Define ∆k : (Ak−1)(q,1) + \ {0} → (0, 1) by k = 1, 2, ..., m − 1. Evidently for any h ∈ A+ k−1, k−2 (k ≥ 2) we have ∆k(h) = ∆(dh′(k)) for all h ∈ A+ ∆k−1(h) = ∆k( \Lk−1(h)). Put (e 16.16) H1 = {1A} ∪ H ∪ H0 ∪ ∪m k=1{h′(k) : h ∈ H ∪ H0} ∪ ∪m k=1{b′k,i : 1 ≤ i ≤ l(k)}, where b′k,i = Lm ◦ Lm−1 ◦ ··· ◦ Lk+1(bk,i). Put σ1 = min{∆(h) : h ∈ H1}/2 and σ2 = σ1 · ε · δ0 64m · (Pm k=1 l(k) + l) . (e 16.17) Let K′ ≥ 1 be an integer (in place of the product M R there) in 15.1 associated with G = (Π0)∗0(K0(A)) = (K0(A))/ ker(ρA) ⊂ Zl (see (e 4.36)), σ1 and σ2. Let {d1, d2, ..., dm} be the set of ranks of all possible irreducible representations of A. Let K′′ = (Qm Let Hk = λk ◦ Πk(H), k = 1, 2, ..., m − 1. Note that Yk \ X δ0 k = {x ∈ Xk : dist(x, Zk) ≥ δ0} is compact. Let Kk ≥ 1 be integers (in place of N ) in 16.1 for σ2/2m (in place of σ) and Qk(C(Yk \ X δ0 k , Fk)Qk (in place of A), and HkYk\X j=1 dj). δ0 k (in place of H ⊂ A). 2·σ2 σ2 k=1 l(k)+l) , let K = (K m 1 64·l(Pm Choose an integer K0 ≥ 1 such that 1/K0 < letK = KK′K′′. For each τ ∈ T (A), we may write k=1 Kk) and 0 )(Qm τ (f ) = l(m)Xi=1ZYm am,itrm,i(f (t))dµj ) + tm−1 ◦ π(m) e (f ) for all f ∈ A, (e 16.18) where µi is a Borel probability measure on Ym, trm,i is the normalized trace on Mr(m,i) (here we regard that f (t) ∈ Mr(m,i) by identifying Pm,i(t)Ms(m,i)Pm,i(t) ∼= Mr(m,i)), tm−1 is an unnor- malized trace on Am−1, am,i ∈ R+ andPi am,i + ktm−1k = 1. We claim that for each k∈ {1, 2,··· , m + 1}, ( ( 2m τ (p) = τ (h) − ( (m − k + 1)ε d(s, i, j)trs,i(Λs(h)(ys,i,j))) + tk−1 ◦ Πk−1(h)) ≤ l(s)Xi=1Xj mXs=k l(s)Xi=1Xj mXs=k KmKm−1 ··· Kk · d(s, i, j) ∈ Z+,Ps,i,j d(s, i, j) + ktkk = 1 and ys,i,j ∈ Ys\X δ0 d(s, i, j)trs,i(Λs(h)(ys,i,j))) + tk−1 ◦ Πk−1(p) for all p ∈ P, and where K m−k+1 s−1, trs,i is the tracial state on Mr(s,i) and tk−1 is a unnormalized trace in (Ak−1), s = k, k + 1, ..., m. m+1 is a sum of empty set (so any such sum is zero). Hence the above holds for k = m + 1. We will prove the claim by reverse induction. Here the convention is, when k = m + 1, τm = τ, Πm = idA andPm tk−1(h) ≥ ∆k(h) − ((m − k + 1)σ2)/2m for all h ∈ Πk−1(H1)⊂ (Ak−1)+, 0 for all h ∈ H, 162 Suppose that the above holds for k ≤ m + 1. One may write i=1 tk−1(a) = trk−1,i(λk−1(a))dµk−1,i + t′k−2 ◦ Πk−2(a) for all a ∈ Ak−1,(e 16.19) ak−1,iZYk−1 ak−1,i + kt′k−2k = ktk−1k, trk−1,i is the tracial state on Mr(k−1,i), t′k−2 is a trace on Ak−2 and µk−1,i is a probability Borel measure on Xk−1. Here we also identify trk−1,i with trk−1,i ◦ πy ◦ Sk−1,i and where πy : Pk,iC(Xk−1, Ms(k−1,i))Pk,i → Mr(k−1,i) is the point evaluation at y ∈ Yk−1. k−1, j = 1, 2, ..., m(k − 1, i) ≤ Kk−1 (for convenient we will denote yk−1,i,j by ti,j) such that l(k−1)Xi=1 wherePl(k−1) It follows from 16.1 that there are yk−1,i,j ∈ Yk−1 \ X δ0 ak−1,iZYk−1\X d(k − 1, i, j)trk−1,i(f (ti,j)) < trk−1,i(f )dµk−1,i(t) − 2m (e 16.20) ak−1,iσ2 2 δ0 k−1 m(k−1,i)Xj=1 for all f ∈ H1, where K0Kk−1 · d(k − 1, i, j) ∈ Z, i = 1, 2, ..., l(k − 1). Put η(f ) = ak−1,iZYk−1\X δ0 k−1 trk−1,i(f )dµk−1,i(t) − m(k−1,i)Xj=1 d(k − 1, i, j)trk−1,i(f (ti,j)). Then η(f ) < ak−1,iσ2 2 2m for all f ∈ H1. (e 16.21) For each i, define ρi : A → C by ρi(f ) = ak−1,iZYk−1 trk−1,i(λk−1(f ))dµk−1,i − for all f ∈ A. Put αi = ρi(1A) and βi = µk−1,i(Yk−1∩X δ0 Yk−1 ∩ X δ0 k−1, we have d(k − 1, i, j)trk−1,i(f (ti,j))(e 16.22) m(k−1,i)Xj=1 k−1), i = 1, 2, ..., k(l). Since supp(bbk−1,i) ⊂ βi > tk−1(\bk−1,i) > ∆k(\bk−1,i) > 2σ2. It follows from (e 16.20) and 1A ∈ H, that ak−1,iβi − αi < ak−1,iσ2 2/2m, or ak−1,i − αi βi < ak−1,iσ2/2m (e 16.23) (e 16.24) and consequently, αi > ak−1,iβi − ak−1,iσ2/2m > 0. Define Ti : A → C by Ti(f ) = αi βiZYk−1∩X δ0 k−1 trk−1,i(sk−1,δ0 ◦ Πk−2(f ))dµk−1,i for all f ∈ A. It follows from (e 16.15) that Ti(f ) − αi βi ZYk−1∩X trk−1,i(Πk−1(f ))dµk−1,i < δ0 k−1 εαi 2m+1 (e 16.25) 163 for each f ∈ H. Also, (e 16.21) implies trk−1,i(Πk−1(f ))dµk−1,i αi ρi(f ) − βi ZYk−1∩X = η(f ) + (cid:18)ak−1,i − < ak−1,iσ2/2m + ak−1,iβi − αi < ak−1,iσ2 βi(cid:19)ZYk−1∩X δ0 k−1 αi δ0 k−1 2/2m−1 trk−1,i(Πk−1(f ))dµk−1,i for each f ∈ H. Note that αi < ak−1,i and σ2 < ε/64. We have ak−1,iε Ti(f ) − ρi(f ) < 2m for f ∈ H. (e 16.26) (e 16.27) (e 16.28) (e 16.29) It follows from (e 16.20) that ak−1,iZYk−1∩X δ0 k−1 trk−1,i(f )dµi − Ti(f )= ρi(f ) − Ti(f ) + η(f ) < ak−1,iε (e 16.30) for all f ∈ H. Since each projection Πk−1(p) is constant on each open set Yk−1 and since Γk−1 ◦ Πk−2(p) = λk−1(Πk−1(p))Zk−1 , one checks that Ti(p) = αi βi ZYk−1∩X = αi · trk−1,i(p) = ρi(p) trk−1,i(sk−1,δ0 ◦ Πk−1(pj))dµk−1,i δ0 k−1 (e 16.31) (e 16.32) for all p ∈ P. Define t′′k−1(f ) = t′k−2(Πk−2(f )) + l(k−1)Xi=1 Ti(f ) + l(k−1)Xi=1 m(k−1,i)Xj=1 d(k − 1, i, j)trk−1,i(f (ti,j)) for all f ∈ A and define tk−2(a) = t′k−2(a) + l(k−1)Xi=1 αi βi Z Yk−1∩X δ0 k−1 for all a ∈ Ak−2. We estimate that trk−1,i ◦ sk−1,δ0(Πk−2(a))dµk−1,i (e 16.33) t′′k−1(h) − tk−1(h)= l(k−1)Xi=1 Ti(f ) − ρi(f ) < ( l(k−1)Xi=1 ak−1,i)ε/2m ≤ ε/2m (e 16.34) for all h ∈ H. Therefore τ (h) − mXs=k−1 ( l(s)Xi=1 m(s,i)Xj=1 d(s, i, j)trs,i(λs(h)(ys,i,j)) + tk−2 ◦ Πk−2(h) ≤ (m − k + 2)ε/2m(e 16.35) for all h ∈ H and τ (p) = mXs=k−1 (Xi,j d(s, i, j)trs,i(λs(p)(ys,i,j)) + tk−2 ◦ Πk−2(p) (e 16.36) 164 tk−1(Lk−1(Πk−2(h))) = for all p ∈ P. Note that ys,i,j ∈ Ys \ X δ0 and using (e 16.24), compute that, for all h ∈ H1, ak−1,iZYk−1 l(k−1)Xi=1 ak−1,iZYk−1∩X l(k−1)Xi=1 βiZYk−1∩X δ0 l(k−1)Xi=1 l(k−1)Xi=1 (ak−1,i − δ0 k−1 αi = = + s and (1 − χs)(ys,i,j) = 0. We then, by (e 16.19),(e 16.33) trk−1,i(Lk−1(Πk−2(h))dµk−1,i + t′k−2 ◦ Πk−2(h)(e 16.37) trk−1,i(Lk−1(Πk−2(h))dµk−1,i + t′k−2 ◦ Πk−2(h)(e 16.38) trk−1,i(Lk−1(Πk−2(h))dµk−1,i + t′k−2 ◦ Πk−2(h)(e 16.39) k−1 l(k−1)Xi=1 α βi )ZYk−1∩X βiZYk−1∩X αi trk−1,i(Lk−1(Πk−2(h))dµk−1,i(e 16.40) δ0 k−1 trk−1,i(sk−1,δ0(Πk−2(h))dµk−1,i(e 16.41) ≤ +t′k−2 ◦ Πk−2(h) + σ2/2m = tk−2(Πk−2(h)) + σ2/2m.(e 16.42) δ0 k−1 It follows that (using (e 16.16) , for all h ∈ H1, tk−2 ◦ Πk−2(h) ≥ tk−1(Lk−1(Πk−2(h))) − σ2/2m ≥ ∆k( \(Lk−1(Πk−2(h)))) − (m − k + 2)σ2/2m = ∆k−1( \Πk−2(h)) − (m − k + 2)σ2/2m. (e 16.43) (e 16.44) (e 16.45) Note that from our construction, there are hj ∈ H1 ⊂ A1 This completes the induction (stop at k − 2 = 0) and proves the claim. Then define ∆0 = ∆1/2. The inequality (e 16.12) and first half of both inequality (e 16.11) and equation (e 16.12) follows from the claim for k = 1. Of course we denote Λs(h)(ys,i,j) by πi(h) for certain irreducible representation πi, and denote d(s, i, j) by ri. Furthermore we have Kri ∈)+, for i = 1, 2,··· , k. Note so far we only have definition of ri for first half of (e 16.11). + with Π0(hj) = ej, j = 1, 2, ..., l . Write τ0 = Pl j=1 α′jtj ◦ Sj, where tj is a tracial state on Mr(0,j), Sj : A0 → Mr(0,j) is the quotient map. Then from (e 16.13) and ej ∈ Π0(H1), we know α′j ≥ σ1. From the first half of (e 16.12), Kri ∈)+ , for i = 1, 2,··· k, and (e 16.10), we have that ritri ◦ πi(p)(cid:1) ∈ Z, s KK′′τ0 ◦ Π0(p) = s KK′′(cid:0)τ (p) − for all p ∈ P. kXi=1 Let rk+j = 1 Let ¯αj = s KK′′α′j, then from the above, we have for any (m1, m2,··· , ml) ∈ (Π0)∗0K0(A) ⊂ K0(A0) = Zl, we havePl j=1 ¯αjmj ∈ Z. Apply (15.1), we have ¯βj (1 ≤ j ≤ l) such thatK′ ¯βj ∈ Z +,Pj k ¯βj − ¯αjk < σ2, andPl j=1 ¯αjmj for all (m1, m2,··· , ml) ∈ (Π0)∗0K0(A). Lemma 16.5. Let A ∈ Dm be a unital C∗-algebra. Let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. Let H ⊆ A be a finite subset and let σ > 0. Then there are finite subset H1 ⊆ A1 ¯βj and πk+j = Sj for j = 1, 2,··· , l. The lemma now follows easily. + \ {0} and a positive integer K such that for any τ ∈ T (A) satisfying ¯βjmj =Pl s KK ′′ j=1 τ (h) > ∆(h) for all h ∈ H1 165 (e 16.46) and any positive homomorphism κ : K0(A) → K0(Ms) with s = κ([1A]) such that ρA(x)(τ ) = (1/s)(κ(x)) for all x ∈ K0(A), (e 16.47) there is a unital homomorphism ϕ : A → MsK such that ϕ∗0 = Kκ and tr′ ◦ ϕ(h) − τ (h) < σ for all h ∈ H, where tr′ is the tracial state on MsK. Proof. Lemma follows easily from 16.4 . Let ∆0 : Π0(A) = A0 : (A0)q,1 j=1 Mr(0,j) and ej ∈ Mr(0,j) is a rank one projection. Without loss of generality, we may assume that A ∈ D′m, and we may also assume that projections in A generates K0(A), by replacing A by MN (A) for some integer N ≥ 1. Let P ⊂ A be a finite subset of projections such that {ρA([p]) : p ∈ P} generates ρA(K0(A))+. Write A0 =Ll(0) possible irreducible representations of A. Let K2 = (Qm + \ {0} and K1 be the integer K in the statement of 16.4 for σ/4 (in place of ε), H and ∆. Let {d1, d2, ..., dm} be the set of ranks of all j=1 dj) and let K = K1K2. For given κ and τ as described in the statement, rational numbers r1, r2, .., rk ≥ 1, finite dimensional irreducible representations π1, π2, ..., πk of A, irreducible representations of A0 such that + \ {0} → (0, 1), H1 ⊂ A1 τ (h) − ( ritri ◦ πi(h) + ritriπi ◦ Π0(h) < σ/4, τ (p) = ritri ◦ πi(p) + ritriπi ◦ Π0(p) for all p ∈ P, kXi=1 kXi=1 k+lXi=k+1 k+lXi=k+1 (e 16.48) (e 16.49) and sK1ri ∈ Z. trace on MR(i). Then we have Define r′i = ri/R(i) and Tri = R(i)tri, where R(i) is rank of πi. So Tri is an unnormalized τ (h) − ( τ (p) = r′iTriπi ◦ Π0(h) < σ/4, kXi=1 kXi=1 r′iTri ◦ πi(p) + k+lXi=k+1 r′iTri ◦ πi(h) + k+lXi=k+1 r′iTriπi ◦ Π0(p) for all p ∈ P, (e 16.50) (e 16.51) and sK1K2r′i ∈ Z. Define T : A → C by T (a) =Pk Note that T ∈ T (A). Therefore i=1 r′iTri ◦ πi(a) +Pk+l i=k+1 r′iTri ◦ πi ◦ Π0(a) for all a ∈ A. (1/s) ◦ κ(p) = T (p) for all p ∈ K0(A) and τ (h) − T (h) < σ/4 for all h ∈ H. (e 16.52) (e 16.53) Put m0 =Pk+l i=1 m(i)R(i), where Ri is rank of πi and m(i) = sK1K2r′i. Define ϕ : A → Mm0 i=1 ¯πi(a)LLk+1 by Ψ(a) =Lk i=k+1 ¯πi(Π0(a)) for all a ∈ A, where ¯πi is m(i) copies of πi. Let κϕ : K0(A) → Z be the map induced by ϕ. Then using (e 16.52), we have, κϕ([1A]) = K1K2κ([1A])). Therefore m0 = sK1K2 = sK, tr′(p) = T (p) for all projections p ∈ P, where tr′ is the normalized trace on MsK and tr′(h) = T (h) for all h ∈ A. This completes the proof. 166 Lemma 16.6. Let A ∈ Dm be a unital C∗-algebra (m ≥ 1). Let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. Let H ⊆ A be a finite subset and let ǫ > 0. There exist a finite subset H1 ⊆ A1 + \ {0} and a finite subset of projections P ⊂ Mn(A) (for some n ≥ 1), and there is δ > 0 such that if a tracial state τ ∈ T (A) satisfies τ (h) > ∆(h) for all h ∈ H1 and any order-unit map κ : K0(A) → K0(Ms) satisfying tr(κ([p])) − τ (p) < δ (e 16.54) for all projections p ∈ P, where tr is the canonical tracial state on Ms, there is a tracial state τ ∈ T (A) such that tr(κ(p)) = τ (p) for all p ∈ K0(A) and τ (h) − τ (h) < ǫ for all h ∈ H. Proof. Without loss of generality, we may assume that projections in A generates K0(A), by replacing A by MN (A) for some integer N ≥ 1. Let P ⊂ A be a finite subset of projections such that {ρA([p]) : p ∈ P} generates ρA(K0(A))+. Let H1 = H∪P and let σ0 = min{∆(h) : h ∈ H1}. Let δ = min{ε · σ0/128, 1/16}. Now suppose that τ and κ satisfies the assumption for the above mentioned H1, P and δ. Recall τ may be viewed as an order preserving map from K0(A) to R. Define η = (1 − ε/3)(κ/s − τ ) : K0(A) → R. Let dA : K0(A) → K0(A)/kerρA be a quotient map and let γ : K0(A)/kerρA → R given such that γ ◦ dA = (ε/3s)κ + η. Note that, as in 4.10 , (e 16.55) So one may view that γ is a homomorphism from (Π0)∗0(K0(A)) to R. For each p ∈ P, from the assumption (e 16.54), one computes that K0(A)/kerρA ∼= (Π0)∗0(K0(A)) ⊂ K0(A0). Therefore η([p]) < (1 − ε/3)δ. γ((Π0)∗0([p])) = (ε/3)(κ([p])/s) + η([p]) > (ε/3)(∆(p) − δ) − (1 − ε/3)δ (e 16.56) (e 16.57) (e 16.58) (e 16.59) ≥ (ε/3)(1 − 1/128)σ0 − (1 − ε/3)εσ0/128 )] > 0 = εσ0[( 1 ε ) − ( 1 128 − 3 · 128 1 3 − 3 · 128 for all p ∈ P. In other words, γ is positive. By applying 2.8 of [53], there is a positive ho- momorphism γ1 : K0(A0) → R such that γ1 ◦ (Π0)∗0 = γ. It is well-known that there is a (non-normalized) trace T0 on A0 such that γ([p]) = T0(p) for all projections p ∈ A. Consider the trace τ′ = (1 − ε/3)τ + T0 ◦ Π0 on A. Then, for all projection p ∈ A, τ′(p) = (1 − ε/3)τ (p) + T0 ◦ Π0(p) = (1 − ε/3)τ (p) + (ε/3s)κ([p]) + η([p]) (e 16.60) = (1 − ε/3)τ (p) + (ε/3s)κ([p]) + (1 − ε/3)(κ([p])/s − τ (p)) = (1/s)κ([p]).(e 16.61) Since (1/s)κ([1A]) = 1, τ′ ∈ T (A). We also compute that, by (e 16.56) Therefore, we also have T0 ◦ Π0(1A) = γ ◦ ρA([1A]) < ε/2 τ′(h) − τ (h) < ε for all h ∈ H. (e 16.62) (e 16.63) 167 Lemma 16.7. Let A ∈ Dm be a unital C∗-algebra. Let ∆ : Aq,1 + \ {0} → (0, 1) be an order preserving map. Let H ⊆ A be a finite subset and let σ > 0. Then there are a finite subset H1 ⊆ A1 + \ {0}, δ > 0, a finite subset P ⊂ K0(A) and a positive integer K such that for any τ ∈ T (A) satisfying τ (h) > ∆(h) for all h ∈ H1 and any positive homomorphism κ : K0(A) → K0(Ms) with s = κ([1A]) such that ρA(x)(τ ) − (1/s)(κ(x)) < δ for all x ∈ P, there is a unital homomorphism ϕ : A → MsK such that ϕ∗0 = Kκ and where tr′ is the tracial state on MsK. tr′ ◦ ϕ(h) − τ (h) < σ for all h ∈ H, Proof. Note that there is an integer n ≥ 1 such that projections in Mn(A) generate K0(A). Therefore this lemma is a corollary of 16.5 and 16.6. Remark 16.8. Since K0(A) is finite generated and (Π0)∗0(K0(A)+) ⊂ Zl + is also finitely gener- ated (though K0(A)+ itself may not be finitely generated), in what follows, we can choose finite subset P to be a set of generators of K0(A) as abelian group, and {(Π0)∗0([p])}p∈P generate (Π0)∗0(K0(A)+) as abelian semigroup. In particular, we can choose n such that P ⊂ Mn(A). Lemma 16.9. Let C ∈ Dm be a unital C-algebra with finitely generated Ki(C) (i = 0, 1). Let ∆ : C q,1 + \ {0} → (0, 1) be an order preserving map. Let F,H ⊆ C be finite subsets, and let ǫ > 0, σ > 0. Then there are a finite subset H1 ⊆ C 1 +\{0}, δ > 0, a finite subset P ⊂ K0(C) and a positive integer K such that for any continuous affine map γ : T (C([0, 1])) → T (C) satisfying γ(τ )(h) > ∆(h) for all h ∈ H1 and for all τ ∈ T (C([0, 1])) and any positive homomorphism κ : K0(C) → K0(Ms(C([0, 1]))) with κ([1C ]) = s such that ρA(x)(γ(τ )) − (1/s)τ (κ(x)) < δ for all τ ∈ T (C([0, 1])) for all x ∈ P, there is an F-ǫ-multiplicative completely positive linear map ϕ : C → MsK(C([0, 1])) such that ϕ0 = Kκ and τ ◦ ϕ(h) − γ′(τ )(h) < σ for all h ∈ H, where γ′ : T (MsK(C([0, 1]))) → T (C) is induced by γ. Furthermore ϕ0 = π0 ◦ ϕ and ϕ1 = π1 ◦ ϕ are true homomorphisms. In the case that C ∈ C, the map ϕ can be chosen to be a homomorphism. Proof. Since any C∗-algebras in C are semi-projective, the second part of the statement follows directly from the first part of the statement. Thus, let us only show the first part of the statement. Without loss of generality, one may assume that F ⊆ H. To simplify notation, without loss of generality, by replacing C by Mr(C) for some r ≥ 1, we may assume that the set of projections in C generates ρA(K0(C)). Since the K-theory of C is finitely generated, there is M ∈ N such that M p = 0 for all p ∈ Tor(Ki(C)), i = 0, 1. Let H1,1 ⊂ C 1 + \ {0} (in place of H1) and σ1 > 0 (in place of δ) be the finite subsets and the positive constant of Theorem 4.17 with respect to C (in the place of A), min{σ, ǫ} (in the 168 place of ǫ) and H (in the place of F), and ∆/2. (We will not need the finite set P in Theorem 4.17, since K0(C) is finitely generated and when we apply Theorem 4.17, we will require both map induce same KL maps.) Let H1,2 ⊆ C (in the place of H1) be a finite subset, δ > 0 be a positive number, P ⊂ K0(C) be a finite subset and K′ be an integer required by Lemma 16.7 with respect to C, ∆/2 (in place of ∆), H ∪ H1,1 (in the place of H) and min{σ/16, σ1/8,{∆(h)/4 : h ∈ H1,1}} (in the place of σ). We may assume that P is the set of projections in C, as mentioned in 16.8. Put H1 = H1,1 ∪ H1,2 and K = M K′. Then, let γ : T (C([0, 1])) → T (C) be a continuous affine map with γ(τ )(h) > ∆(h) for all h ∈ H1, and let κ : K0(C) → K0(Ms(C([0, 1]))) with κ([1C ]) = s such that ργ(τ )(x) − (1/s)τ (κ(x)) < δ for all x ∈ P for all τ ∈ T (C([0, 1])). Since γ is continuous, there is a partition 0 = x0 < x1 < ··· < xn = 1 such that for any 0 ≤ i ≤ n − 1, and any x ∈ [xi, xi+1], one has γ(τx)(h) − γ(τxi)(h) < min{σ/8, σ1/4} for all h ∈ H1, where τx ∈ T (Ms(C([0, 1]))) is the extremal trace which concentrates at x. For any 0 ≤ i ≤ n, consider the trace τi = γ(τxi) ∈ T (C). It is clear that τi(x) − tr(κ(x)) < δ for all x ∈ P and τi(h) > ∆(h) for all h ∈ H1,2. By Lemma 16.7, there exists a unital homomorphism ϕ′i : C → MsK ′(C) such that (e 16.64) as we identify K0(C([0, 1], Ms)) with Z and [ϕ′i]0 = Kκ tr ◦ ϕ′i(h) − τxi(h) < min{σ/16, ∆(h)/4, σ1/8; h ∈ H1,1} for all h ∈ H ∪ H1,1. (e 16.65) In particular, by (e 16.65), one has that, for any 0 ≤ i ≤ n − 1, tr ◦ ϕ′i(h) − tr ◦ ϕ′i+1(h) < σ1 for all h ∈ H1,1. Note that γ(τxi)(h) > ∆(h) for any h ∈ H1,1 by the assumption. (e 16.65) that, for any 0 ≤ i ≤ n, It then also follows from Define the amplification ϕ′′i as tr ◦ ϕ′i(h) > ∆(h)/2 for all h ∈ H1,1. ϕ′′i := ϕ′i ⊗ 1MM (C) : C → MsK(C)= MM (sK ′)(C). One has that [ϕ′′i ] = [ϕ′′i+1] in KL(C, MsK ). It then follows from Theorem 4.17 that there is a unitary u1 ∈ Ms(C) such that kϕ′′0(h) − Adu1 ◦ ϕ′′1(h)k < min{σ, ǫ} for all h ∈ H. 169 Consider the maps Adu1 ◦ ϕ′′1 and ϕ′′2. Applying Theorem 4.17 again, one obtains a unitary u2 ∈ MsK(C) such that kAdu1 ◦ ϕ′′1(h) − Adu2 ◦ ϕ′′2(h)k < min{σ, ǫ} for all h ∈ H. Repeat this argument for all i = 1, ..., n, one obtains unitaries ui ∈ MsK(C) such that kAdui ◦ ϕ′′i (h) − Adui+1 ◦ ϕ′′i+1(h)k < min{σ, ǫ}, for all h ∈ H. Then define ϕ0 = ϕ′′0 and ϕi = Adui ◦ ϕ′′i , and one has kϕi(h) − ϕi+1(h)k < min{σ, ǫ} for all h ∈ H. (e 16.66) Define the linear map ϕ : C → MsK([0, 1]) by ϕ(f )(t) = t − xi xi+1 − xi ϕi(f ) + xi+1 − t xi+1 − xi ϕi+1(f ), if t ∈ [xi, xi+1]. Since each ϕi is a homomorphism, by (e 16.66), the map ϕ is H-ǫ-multiplicative; in particular, it is F-ǫ-multiplicative. It is clear that ϕ∗0 = Kκ. On the other hand, for any x ∈ [xi, xi+1] for some i = 1, ..., n − 1, one has that for any h ∈ H, tr(ϕi(f )) + γ(τx)(h) − τx ◦ ϕ(h) x − xi xi+1 − xi x − xi xi+1 − xi = γ(τx)(h) − ( < γ(τx)(h) − ( γ(τxi)(h) + < γ(τx)(h) − γ(τxi+1)(h)) + 3σ/8 < σ/2, xi+1 − x xi+1 − xi xi+1 − x xi+1 − xi tr(ϕi+1(f ))) γ(τxi+1)(h)) + σ/4 (by (e 16.65)) (by (e 16.64)) (by (e 16.64)). Hence for any h ∈ H, γ(τx)(h) − τx ◦ ϕ(h) < γ(τx)(h) − τx ◦ L(h) + σ/2 < σ, and therefore γ(τ )(h) − τ ◦ ϕ(h) < σ for any τ ∈ T (MsK(C([0, 1]))). Thus the map ϕ satisfies the statement of the lemma. Note that the restriction of ϕ to the boundaries are ϕ0 and ϕn which are homomorphisms. Theorem 16.10. Let C ∈ Dm be a unital C∗-algebra with finite generated Ki(C) (i = 0, 1). Let ∆ : C q,1 + \ {0} → (0, 1) be an order preserving map. Let F,H ⊆ C be finite subsets, and let 1 > σ, ǫ > 0. There exist a finite subset H1 ⊆ C 1 + \ {0}, δ > 0, a finite subset P ⊂ K0(C)and a positive integer K such that for any continuous affine map γ : T (D) → T (C) satisfying γ(τ )(h) > ∆(h) for all h ∈ H1 for all τ ∈ T (D), where D is a C-algebra in C, any positive homomorphism κ : K0(C) → K0(D) with κ([1C ]) = s[1D] for some integer s ≥ 1 satisfying ρC(x)(γ(τ )) − (1/s)τ (κ(x)) < δ for all τ ∈ T (D) 170 and for all x ∈ P, there is a F-ǫ-multiplicative positive linear map ϕ : C → MsK(D) such that ϕ∗0 = Kκ and (1/(sK))τ ◦ ϕ(h) − γ(τ )(h) < σ for all h ∈ H and τ ∈ T (D). In the case that C ∈ C, the map ϕ can be chosen to be a homomorphism. Proof. As in the proof of 16.9, since C∗-algebras in C are semi-projective, we will only prove the first part of the statement. Without loss of generality, one may also assume that F ⊆ H. By replacing C by Mm(C) for some integer m ≥ 1, we may find a finite subset P of projections in C as in Remark 16.8. Without loss of generality, one may also assume that P ⊆ H. Since the K-group of C is finitely generated (as abelian groups), there is M ∈ N such that M x = 0 for all x ∈ Tor(Ki(C)), i = 0, 1. Let H1,1 ⊂ C 1 +\{0} (in place of H1) and σ1 > 0 (in place of δ) be a positive number required by Theorem 4.17 with respect to C (in the place of A), min{σ/4, ǫ/2} (in the place of ǫ), H (in the place of F) and ∆. Let H1,2 ⊆ C (in place of H1) be a finite subset, let σ2 (in place of δ) be a positive number, and K1 (in place of K) be an integer required by Lemma 16.7 with respect to H ∪ H1,1 and 2 min{σ/16, σ1/4, min{∆(h)/2 : h ∈ H1,1}} (in the place of σ) and ∆. Note we can choose P above — see 16.8. Let H1,3 (in place of H1), σ3 > 0 (in place of δ) and K2 (in place of K) be finite sub- set and constants required by Lemma 16.9 with respect to C, H ∪ H1,1 (in the place of H), min{σ/16, σ1/4, min{∆(h)/2 : h ∈ H1,1}} (in the place of σ), ε/4 (in place of ε), H (in place of F) and ∆ (with the same P above). Put H1 = H1,1 ∪ H1,2 ∪ H1,3∪P, δ = min{σ1/2, σ2, 1/4} and K = M K1K2. Let D = A(F1, F2, ψ0, ψ1) = {(f, a) ∈ C([0, 1], F2) ⊕ F1 : f (0) = ψ0(a) and f (1) = ψ1(a)} 1 be any C∗-algebra in C, and let γ : T (D) → T (C) be a given continuous affine map satisfying γ(τ )(h) > ∆(h) for all h ∈ H1 for all τ ∈ T (D). Let κ : K0(C) → K0(Ms(D)) be any positive map with s[1D] = κ([1C ]) satisfying ρC(x)(γ(τ )) − (1/s)τ (κ(x)) < δ for all τ ∈ T (D) and for all x ∈ P. Write C([0, 1], F2) = I1 ⊕ I2 ⊕ ··· ⊕ Ik with Ii = C([0, 1], Mri ), i = 1, ..., k. Note that γ induces a continuous affine map γi : T (Ii) → T (C) by γi : T (Ii) → T (C) defined by γi(τ ) = γ(τ ◦ πi) for each 1 ≤ i ≤ k, where πi is the restriction map D → Ii defined by (f, a) → f[0,1]j . It is clear that for any 1 ≤ i ≤ k, one has that γi(τ )(h) > ∆(h) for all h ∈ H1,3 and for all τ ∈ T (Ii) and ρC(x)(γi(τ )) − τ ((πi)∗0 ◦ κ(x)) < δ ≤ σ3 for all τ ∈ T (Ms(Ii)) (e 16.67) (e 16.68) and for all x ∈ P and for any 1 ≤ i ≤ k. Also write F1 = MR1 ⊕ ··· ⊕ MRl and denote by π′j : D → MRj the corresponding evaluation of D. Since γ(τ )(h) > ∆(h) for all h ∈ H1,2 and for all τ ∈ T (D), and ρC(x)(γ(τ )) − (1/s)τ (κ(x)) < δ for all τ ∈ T (D) (e 16.69) 171 and for all x ∈ P, one has that, for each j, γ ◦ (π′j)∗(tr)(h) > ∆(h) for all h ∈ H1,2 and ρC (x)(γ ◦ (π′j)∗(tr′)) − tr([π′j] ◦ κ(x)) < δ ≤ σ2, (e 16.70) where tr is the tracial state on MsRj and tr′ is the tracial state on MRj , for all x ∈ K0(C) and where γ ◦ (π′j)∗(tr) = γ(tr ◦ πj). It follows from Lemma 16.7 that there is a homomorphism ϕ′j : C → MRj ⊗ MsK1K2 such that (ϕ′j)∗0 = (π′j)∗0 ◦ K1K2κ and (e 16.71) tr ◦ ϕ′j(h) − (γ ◦ (π′j)∗)(tr′)(h) < min{σ/16, σ1/4, min{∆(h)/2 : h ∈ H1,1}}(e 16.72) for all h ∈ H∪H1,1, where tr is the tracial state on MRj ⊗ MsK and where tr′ is the tracial state on MRj . Denote by ϕ′ = lMj=1 ϕ′j : C → F1 ⊗ MsK1K2(C). Applying Lemma 16.9 to (e 16.67) and (e 16.68), one has that, for any 1 ≤ i ≤ k, there is an H-ε/4-multiplicative contractive completely positive linear map ϕi : C → Ii⊗ MsK1K2 such that (ϕi)∗0 = (πi)∗0 ◦ K1K2κ and (1/sK1K2)τ ◦ ϕi(h)− ((γ ◦ (πi)∗)(τ ))(h) < min{σ/16, σ/4, min{∆(h)/2 : h ∈ H1,1}} (e 16.73) for all h ∈ H ∪ H1,1 ∪ G1, where τ ∈ T (Ii). Furthermore, as in the conclusion of Lemma 16.9, the restrictions of ϕi to both boundaries are homomorphisms. For each 1 ≤ i ≤ k, denote by πi,0 and πi,1 the evaluations of Ii ⊗ Ms at the point 0 and 1 (e 16.74) respectively. Then one has ψ0,i ◦ πe = πi,0 ◦ πi. It follows that lXj=1 (π′j)∗0) ◦ K1K2κ (ψ0,i ◦ ϕ′)∗0 = (ψ0,i)∗0 ◦ ( = (ψ0,i)∗0 ◦ (πe)∗0 ◦ K1K2κ = (πi,0 ◦ πi)∗0 ◦ K1K2κ = (πi,0)∗0 ◦ (ϕi)∗0. Moreover, note that by (e 16.72), and by (e 16.73), tr ◦ (ψ0,i ◦ ϕ′)(h) ≥ ∆(h)/2 for all h ∈ H1,1, tr ◦ (πi,0 ◦ ϕi)(h) ≥ ∆(h)/2 for all h ∈ H1,1. It also follows from (e 16.72) and (e 16.73) that tr ◦ (ψ0,i ◦ ϕ′)(h) − tr ◦ (πi,0 ◦ ϕi)(h) < σ1/2 for all h ∈ H1,1. Consider amplifications ϕ′i : = ϕi ⊗ 1MM (C) : C → Ii ⊗ MsK and ϕ′′ : = ϕ′ ⊗ 1MM (C) : C → F1 ⊗ MsK . 172 (e 16.75) (e 16.76) (e 16.77) (e 16.78) (e 16.79) (e 16.80) (e 16.81) Then, one has [ψ0,i ◦ ϕ′′] = [(πi,0)∗0 ◦ (ϕ′i)] in KL(C, MrisK). Therefore, by Theorem 4.17, there is a unitary ui,0 ∈ Mri ⊗ MsK such that kAdui,0 ◦ πi,0 ◦ ϕ′i(f ) − ψ0,i ◦ ϕ′′(f )k < min{σ/4, ǫ/2} for all f ∈ H. Exactly the same argument shows that there is a unitary ui,1 ∈ Mri ⊗ MsK such that kAdui,1 ◦ πi,1 ◦ ϕi′(f ) − ψ1,i ◦ ϕ′′(f )k < min{σ/4, ǫ/2} for all f ∈ H. Choose two paths of unitaries {ui,0(t) : t ∈ [0, 1/2]} ⊂Mri ⊗ MsK such that ui,0(0) = ui,0 and ui,0(1/2) = 1Mri⊗MsK , and {ui,1(t) : t ∈ [1/2, 1]} ⊂ Mri ⊗ MsK such that ui,1(1/2) = 1Mri⊗MsK and ui,1(1) = ui,1 Put ui(t) = ui,0(t) if t ∈ [0, 1/2) and ui(t) = ui,1(t) if t ∈ [1/2, 1]. Define ϕi : C → Ii ⊗ MsK by πt ◦ ϕi = Ad ui(t) ◦ πt ◦ ϕ′i, where πt : Ii ⊗ MsK → Mri ⊗ MsK is the point-evaluation at t ∈ [0, 1]. One has that, for each i, kπi,0 ◦ ϕi(f ) − ψ0,i ◦ ϕ′′(f )k < min{σ/4, ǫ/2} and kπi,1 ◦ ϕi(f ) − ψ1,i ◦ ϕ′′(f )k < min{σ/4, ǫ/2} for all f ∈ H. (e 16.82) For each 1 ≤ i ≤ k, let ǫi < 1/2 be a positive number such that k ϕi(f )(t) − ψ0,i ◦ ϕ′′(f )k < min{σ/4, ǫ/2} for all f ∈ H for all t ∈ [0, ǫi] and k ϕi(f )(t) − ψ1,i ◦ ϕ′′(f )k < min{σ/4, ǫ/2} for all f ∈ H for all t ∈ [1 − ǫi, 1].(e 16.83) (ψ0,i ◦ ϕ′′) + t ǫi ϕi(f )(ǫi), ǫi (ψ1,i ◦ ϕ′′) + 1−t ǫi ϕi(f )(ǫi), if t ∈ [0, ǫi], if t ∈ [ǫi, 1 − ǫi] , if t ∈ [1 − ǫi, 1]. Define Φi : C → Ii ⊗ MsK to be (ǫi−t) ϕi(f )(t), (t−1+ǫi) ǫi Φi(t) = The map Φi is not necessarily a homomorphism, but it is H-ǫ-multiplicative; in particular, it is F-ǫ-multiplicative. Moreover, it satisfies the relations πi,0 ◦ Φi(f ) = ψ0,i ◦ ϕ′′(f ) and πi,1 ◦ Φi(f ) = ψ1,i ◦ ϕ′′(f ) for all f ∈ H, i = 1, ..., k.(e 16.84) Define Φ′(f ) : C → C([0, 1], F2) ⊗ MsK by πi,t ◦ Φ′ = Φi, where πi,t : C([0, 1], F2) ⊗ MsK → Mri ⊗ MsK defined by the point evaluation at t ∈ [0, 1] (on the i-th summand) and define Φ′′ : C → F1 by Φ′′(f ) = ϕ′(f ) for all f ∈ C. Define ϕ(f ) = (Φ′(f ), Φ′′(f )). It follows from (e 16.84) that ϕ is F-ǫ-multiplicative positive linear map from C to D ⊗ MsK. It follows from (e 16.71) that [πe ◦ ϕ(p)] = [ϕ′(p)] = (πe)∗0 ◦ Kκ([p]) for all p ∈ P. Since (πe)∗0 : K0(D) → Zl is injective, one has It follows from (e 16.73) and (e 16.72) that one calculates that ϕ∗0 = Kκ. (1/sK)τ ◦ ϕ(h) − γ(τ )(h) < σ for all h ∈ H and for all τ ∈ T (D). 173 (e 16.85) (e 16.86) Lemma 16.11. Let C ∈ C. For any ε > 0 and any finite subset H ⊂ Cs.a., there exists a finite subset of extremal traces T ⊆ T (C) and a continuous affine map λ : T (C) → ▽, where ▽ is the convex hull of of T such that λ(τ )(h) − τ (h) < ε, h ∈ H, τ ∈ T (C). (e 16.87) Proof. Without loss of generality, we may assume that H is in the unit ball of C. Write C = C(F1, F2, ψ0, ψ1), where F1 = MR1 ⊕ MR2 ⊕ ··· ⊕ MRl and F2 = Mr1 ⊕ Mr2 ⊕ ··· ⊕ Mrk . Let πe,i : C → MRi be the surjective homomorphism defined by the composition of πe and the projection from F1 onto MRi, and πI,j : C → C([0, 1], Mrj ) the restriction which may also be viewed as the restriction of the projection from C([0, 1], F2) to C([0, 1], Mri ). Denote by πt ◦ πI,j the composition of πI,j and the point-evaluation at t ∈ [0, 1]. There is δ > 0 such that, for any h ∈ H, kπI,j(h)(t) − πI,j(h)(t′)k < ε/16 for all h ∈ H (e 16.88) and t − t′ < δ, t, t′ ∈ [0, 1]. Let g1, g2, ..., gn be a partition of unity over interval [δ, 1 − δ] with respect to an open cover 6= 0 implies that s − s′ ≤ 1. with order 2 such that each supp(gi) has diameter < δ and gsgs′ Let ts ∈ supp(gs)∩[δ, 1 − δ] be a point. We may assume that ts < ts+1, We may further choose t1 = δ and tn = 1 − δ and assume that g1(δ) = 1 and gn(1 − δ) = 1 by choosing an appropriate open cover of order 2. Extend gs to [0, 1] by defining gs(t) = 0 if t ∈ [0, δ) ∪ (1 − δ, 1] for s = 2, 3, ..., n − 1 and g1(t) = g1(δ)(t/δ) for t ∈ [0, δ) and gn(t) = gn(1 − δ)(1 − t)/δ for t ∈ (1 − δ, 1]. (e 16.89) Define g0 = 1−Pn s=1 gs. Then g0(t) = 0 for all t ∈ [δ, 1−δ]. In what follows, we view gs as gs·idC as in the center of C. In particular, g0 is identified with (g0, 1F1) so that g0(0) = ψ0(1F1) and g0(1) = ψ1(1F1). Let gs,j = πI,j(gs), s = 1, 2, ..., n, j = 1, 2, ..., k. Let pi ∈ F1 be the projection corresponding to the summand MRi. Choose di ∈ C([0, 1], F2) such that di(t) = ψ0(pi) for t ∈ [0, δ] and di(t) = ψ1(pi) for t ∈ [1− δ, 1] and 0 ≤ di(t) ≤ 1 for t ∈ (δ, 1− δ). Note that di ∈ C. i=1 g0di = g0. Without loss of generality, we may assume that {ts : 1 ≤ s ≤ n} is a set of distinct points. Denote by tri the tracial state on MRi and tr′j the tracial state on Mrj , i = 1, 2, ..., l and j = 1, 2, ..., k. Let Moreover,Pl T = {tri ◦ πe,i : 1 ≤ i ≤ l} ∪ n[s=1 {tr′j ◦ πts ◦ πI,j : 1 ≤ j ≤ k}. (e 16.90) Let ▽ be the convex hull of T . Define λ : T (C) → ▽ by λ(τ )(f ) = kXj=1 nXs=1 τ (gs,j)tr′j ◦ (πI,j(f )(ts)) + lXi=1 τ (g0di)tri ◦ πe,i(f ), (e 16.91) where gs,j = gs · 1rj ∈ C0((0, 1), Mrj ) ⊂ C, for all f ∈ C. It is clear that λ is a continuous affine map. Note that if h ∈ C, λ(trj ◦ πe,j)(h) = trj ◦ πe,j(g0di)tri ◦ πe,i(h) = trj ◦ πe,j(g0dj)trj ◦ πe,j(h) = trj ◦ πe,j(h). lXi=1 174 (e 16.92) (e 16.93) (e 16.94) If τ (f ) = tr′j ◦ (πI,j(f )(t)) with t ∈ (δ, 1 − δ), then if h ∈ H, tr′j ◦ (πI,j(hgs)(t))) τ (h) = ≈2ε/16 = = nXs=1 tr′j ◦ (πI,j(h)(t)) = ( nXs=1 tr′j ◦ πI,j(h(ts)gs(t)) nXs=1 kXi=1 nXs=1 τ (gs,j)tr′j ◦ πI,j(h(ts)) lXi=1 τ (gs,i)tr′i ◦ πI,i(h(ts)) + gs,j(t)tr′j ◦ πI,j(h(ts)) = nXs=1 0 · tri ◦ πe,i(h) = λ(τ )(h), (e 16.95) (e 16.96) (e 16.97) (e 16.98) where we notice that τ (gs,i) = 0 if i 6= j. If τ has the form τ (f ) = tr′j ◦ (πI,j(f )(t)) for some fixed t ∈ (0, δ), then for h ∈ H with h = (h0, h1), where h0 ∈ C([0, 1], F2) and h1 ∈ F1 such that ψ0(h1) = h0(0) = h(0) and ψ1(h1) = h0(1) = h(1), τ (h) = tr′j ◦ (πI,j(h(t))) = tr′j ◦ (πI,j(h(t)g1(t))) + tr′j ◦ (πI,j(h(t)g0(t))) tr′j ◦ (πI,j(h(δ)g1(t)) + tr′j ◦ πI,j(h(0)g0(t))) ≈ε/8 = g1(t)tr′j ◦ πI,j(h(δ)) + g0(t)tr′j ◦ πI,j(ψ0(h1)) = τ (g1,j)tr′j ◦ πI,j(h(t1)) + g0(t) lXi=1 tr′j ◦ πI,j(ψ0(h1pi)) lXi=1 = τ (g1,j)tr′j ◦ πI,j(h(t1)) + g0(t)( lXi=1 tr′j ◦ (πI,j(di(t)))tri(πe,i(h)) tr′j ◦ (πI,j(g0(t)di(t)))tri(πe,i(h)) = τ (g1,j)tr′j ◦ πI,j(h(t1)) + = τ (g1,j)tr′j ◦ πI,j(h(t1)) + g0(t) lXi=1 lXi=1 = τ (g1,j)tr′j ◦ πI,j(h(t1)) + = λ(τ )(h). τ (g0di)tri(πe,i(h)) tr′j ◦ πI,j(ψ0(pi))tri ◦ πe,i(h1)) The same argument as above shows that, if then It follows that τ (f ) = tr′j ◦ (πI,j(f )(t)), t ∈ (1 − δ, 1), τ (h) ≈ε/8 λ(τ )(h) for all h ∈ H. and for all extreme points of τ ∈ T (C). By the Choquet Theorem, for each τ ∈ T (C), there exist a Borel probability measure µτ on the extreme points ∂eT (C) of T (C) such that τ (h) − λ(τ )(h) < ε/8 for all h ∈ H τ (f ) =Z∂eT (C) f (t)dµτ for all f ∈ Aff(T (C)). 175 Therefore, for each h ∈ H, τ (h) =Z∂e(T (C) h(t)dµτ ≈ε/8Z∂eT (C) as desired. h(λ(t))dµτ for all τ ∈ T (C), Lemma 16.12. Let C be a unital stably finite C-algebra, and let A ∈ B1 (or B0). Let α : T (A) → T (C) be a continuous affine map. (1) For any finite subset H ⊆ Aff(T (C)), any σ > 0, there is a C-subalgebra D ⊆ A and a continuous affine map γ : T (D) → T (C) such that D ∈ C (or C0,) and h(γ(ı(τ ))) − h(α(τ )) < σ for all τ ∈ T (A) for all h ∈ H, where ı : T (A) ∋ τ → 1 τ (p) τD ∈ T (D), and p = 1D. (2) If there are a finite subset H1 ⊆ C + and σ1 > 0 such that α(τ )(g) > σ1 for all g ∈ H1 for all τ ∈ T (A), the affine map γ can be chosen so that γ(τ )(g) > σ1 for all g ∈ H1 for any τ ∈ T (D). (3) If the positive cone of K0(C) is generated by a finite subset P of projections and there is an order-unit map κ : K0(C) → K0(A) which is compatible to α, then, for any δ > 0, the C-subalgebra D and γ can be chosen so that there are positive homomorphisms κ0 : K0(C) → K0((1 − p)A(1 − p)) and κ1 : K0(C) → K0(D) such that κ1 is strictly positive, κ = κ0 + ı ◦ κ1, where ı : D → A is the embedding,, and γ(τ )(p) − τ (κ1([p]) < δ for all p ∈ P and τ ∈ T (D). (e 16.99) (4) Moreover, if A ∼= A⊗ U for some infinite dimensional UHF-algebra, for any given positive integer K, the C-algebra D can be chosen so that D = MK(D1) for some D1 ∈ C (D1 ∈ C0) and κ1 = Kκ′1, where κ′1 : K0(C) → K0(D1) is a strictly positive homomorphism. Furthermore, κ0 can also be chosen to be strictly positive. Proof. Write H = {h1, h2, ..., hm}. We may assume that khik ≤ 1, i = 1, 2, ..., m. Choose f1, f2, ..., fm ∈ A such that that τ (fi) = hi(α(τ )) for all τ ∈ T (A) and kfik ≤ 2, i = 1, 2, ..., m (see 9.2 of [63]). Put F = {1A, f1, f2, ..., fm}. Let δ > 0 and let G1 (in place of G) be a finite subset required by Lemma 9.4 of [63] for A, σ/16 (in place of ε) and F. Let σ1 = min{σ/16, δ/16}. We may assume that G1 ⊃ F. Put G = {g, gh : g, h ∈ G1}. Since A ∈ B1 (or B0) there is a D ∈ C (or C0), and h′ ∈ (1 − p)A(1 − p) and h′′ ∈ D with p = 1D such that kh − (h′ + h′′)k < σ1/16, h ∈ G1 and τ (1 − p) < σ1/2, τ ∈ T (A). (e 16.100) Moreover, since D is amenable, without loss of generality, we may further assume that there is a unital contractive completely positive linear map L : A → D such that L(h) = h′′ and L 176 is G-σ1/2-multiplicative. By the choice of δ and G, it follows from Lemma 9.4 of [63] that, for each τ ∈ T (D), there is γ′(τ ) ∈ T (A) such that τ (L(h)) − γ′(τ )(h) < σ/16 for all h ∈ F. (e 16.101) Applying 16.11, one obtains t1, t2, ..., tn ∈ ∂eT (D) and a continuous affine map λ : T (D) → ∆ such that (e 16.102) τ (f ) − λ(τ )(f ) < σ1/16 for all τ ∈ T (D) and f ∈ F, where ∆ is the convex hull of {t1, t2, ..., tn}. Define λ1 : ∆ → T (A) by λ1(ti) = γ′(ti), i = 1, 2, ..., m. (e 16.103) Define γ = α ◦ λ1 ◦ λ. Then hj(γ(π(τ ))) = hj(α ◦ λ1 ◦ λ(ı(τ ))) λ1 ◦ λ(π(τ ))(fj ) = ≈σ/16 λ(ı(τ ))(f′′j ) ≈σ/16 ≈σ/8 ı(τ )(f′′j ) τ (gj) = hj(α(τ )), and this proves (1). Note that it follows from the construction that γ(τ ) ∈ α(T (A)), and hence (2) also holds. With 15.5, (3) and (4) follows straightforwardly except the ”Furthermore” part. To see that, we note that we may choose D ⊂ A ⊗ 1U . Choose a projection e ∈ U such that 0 < t0(e) < δ0 < δ − max{γ(τ )(p) − τ (κ1([p]) : p ∈ P and τ ∈ T (D)}, where t0 is the unique tracial state of U. We then replace κ1 by κ2 : K0(A) → K0(D2), where D2 = D ⊗ (1 − e) and κ2([p]) = κ1([p]) ⊗ [1 − e]. Define κ3([p]) = κ1([p]) ⊗ [e]. Then let κ4 : K0(C) → K0((1 − p + (1 ⊗ e))A(1 − p + (1 ⊗ e)) be defined by κ4 = κ0 + [ı] ◦ κ3, where ı : D ⊗ e → A ⊗ U ∼= A is the embedding. We then replace κ0 by κ4. Note that, now, κ4 is strictly positive. 17 Maps from homogeneous C∗-algebras to C∗-algebras in C. Lemma 17.1. Let X be a connected finite CW-complex and let C = C(X). Let H ⊆ C be a finite subset, and let σ > 0. There exists a finite subset H1,1 ⊆ C + satisfying the following: for any σ1,1 > 0, there is a finite subset H1,2 ⊆ C + satisfying the following: for any σ1,2 > 0, there is a positive integer M such that for any D ∈ C with the dimension of any irreducible representation of D at least M , for any continuous affine map γ : T (D) → T (C) satisfying and γ(τ )(h) > σ1,1 for all h ∈ H1,1 for all τ ∈ T (D), γ(τ )(h) > σ1,2 for all h ∈ H1,2 for all τ ∈ T (D), there is a homomorphism ϕ : C → D such that τ ◦ ϕ(h) − γ(τ )(h) < σ for all h ∈ H. Moreover, if D ∈ C0, then there is a point-evaluation Ψ : C → D such that [ϕ] = [Ψ]. 177 Proof. Without loss of generality, one may assume that every element of H has norm at most 1. Let η > 0 such that for any f ∈ H and any x, x′ ∈ X with d(x, x′) < η, one has f (x) − f (x′) < σ/4. Since X compact, one can choose a finite subset H1,1 ⊆ C + such that for any open ball 24 , there is an nonzero element h ∈ H1,1 with supp(h) ⊂ Oη/24. We assume Oη/24 ⊂ X, of radius η that khk ≤ 1 for all h ∈ H1,1. Consequently, if there is σ1,1 > 0 such that then τ (h) > σ1,1 for all h ∈ H1,1, µτ (Oη/24) > σ1,1 for any open ball Oη/24 with radius η/24, where µτ is the probability measure induced by τ. Fixed σ1,1 > 0. Let δ and G ⊆ C(X) (in the place of G) be the constant and finite subset of Lemma 6.2 of [70] with respect to σ/2 (in the place of ǫ), H (in the place of F), and σ1,1/η (in the place of σ). Let H1,2 ⊆ C(X) (in the place of H1) be the finite subset of Theorem 4.2 with respect to δ (in the place of ǫ) and G (in the place of F). Let σ1,2 > 0. Then let H2 ⊆ C(X) (in the place of H2) and σ2 be the finite subset and positive constant of Theorem 4.2 with respect to σ1,2 and H1,2 (in the place of σ1 and H1). Let M (in place of N ) be the constant of Theorem 2.1 of [50] with respect to H2∪H1,2∪H1,1 (in the place of F ) and min{σ/4, σ2/4, σ1,2/2, σ1,1/2} (in the place of ǫ). Let D = D(F1, F2, ψ0, ψ1) be a C∗-algebra in C with dimensions of the irreducible rep- resentations at least M . Let γ : T (D) → T (C) be a map satisfying the lemma. Write C([0, 1], F2) = I1 ⊕ ··· ⊕ Ir with Ii = C([0, 1], Mri ), i = 1, ..., k. Then γ induces a continu- ous map γi : T (Ii) → T (C) by γi(τ ) = γ(τ ◦ πi), where πi is the restriction map D → Ii. It is then clear that γi(τ )(h) > σ1,2 for all h ∈ H1,2, τ ∈ T (Ii). Also write F1 = MR1 ⊕···⊕ MRl, and denote by π′j : D → Mlj the corresponding evaluation of D. By Theorem 2.1 of [50], for each 1 ≤ i ≤ k, there is a homomorphism ϕi : C(X) → Ii such that τ ◦ ϕi(h) − γi(τ )(h) < min{σ/4, σ2/4, σ1,2/2, σ1,1/2} for all h ∈ H2 ∪ H1,2 ∪ H1,1; (e 17.1) and for any j, there is also a homomorphism ϕ′j : C(X) → MRj such that tr ◦ ϕ′j(h) − γ ◦ (π′j)∗(tr)(h) < min{σ/4, σ2/4, σ1,2/2, σ1,1/2} for all h ∈ H2 ∪ H1,2 ∪ H1,1. (e 17.2) Denote by ϕ′ =Lj ϕ′j and by πt : Ii → Mri the point-evaluation at t ∈ [0, 1]. It follows that tr ◦ (ψ0,i ◦ ϕ′)(h) ≥ σ1,2/2 tr ◦ (ψ0,i ◦ ϕ′) − tr ◦ (π0 ◦ ϕi) ≤ σ2/2 for all h ∈ H2 and and tr ◦ (π0 ◦ ϕi)(h) ≥ σ1,2/2 for all h ∈ H1,2. (e 17.3) By Theorem 4.2, there is a unitary ui,0 ∈ Mri such that Adui,0 ◦ π0 ◦ ϕi(f ) − ψ0,i ◦ ϕ′(f ) < δ for all f ∈ G. Exactly the same argument shows that there is a unitary ui,1 ∈ Mri such that kAdui,1 ◦ π1 ◦ ϕi(f ) − ψ1,i ◦ ϕ′(f )k < δ for all f ∈ H. 178 Choose two paths of unitaries {ui,0(t) : t ∈ [0, 1/2]} ⊂ Mri such that ui,0(0) = ui,0 and and ui,1(1) = ui,1 ui,0(1/2) = 1Mri Put ui(t) = ui,0(t) if t ∈ [0, 1/2) and ui(t) = ui,1(t) if t ∈ [1/2, 1]. Define ϕi : C → Ii by and {ui,1(t) : t ∈ [1/2, 1]} ⊂ Mri such that ui,1(1/2) = 1Mri Then πt ◦ ϕi = Ad ui(t) ◦ πt ◦ ϕi. kπ0 ◦ ϕi(f ) − ψ0,i ◦ ϕ′(f )k < δ and kπ1 ◦ ϕi(f ) − ψ1,i ◦ ϕ′(f )k < δ (e 17.4) for all f ∈ G, i = 1, ..., k. Note that it also follows from (e 17.1) and (e 17.2) that tr ◦ (ψ0,i ◦ ϕ′)(h) ≥ σ1,1/2 and tr ◦ (πi,0 ◦ ϕi)(h) ≥ σ1,1/2 for all h ∈ H1,1/2. Hence µτ◦(ψ0,i◦ϕ′)(Oη/24) ≥ σ1,1 and µτ◦(π0◦ ϕi)(Oη/24) ≥ σ1,1. Thus, by Lemma 6.2 of [70], for each 1 ≤ i ≤ k, there are two unital homomorphisms Φ0,i, Φ′0,i : C(X) → C([0, 1], Mri ) such that π0 ◦ Φ0,i = ψ0,i ◦ ϕ′, πt ◦ Φ0,i(f ) − ψ0,i ◦ ϕ′(f ) < σ/2, π0 ◦ Φ′0,i = π0 ◦ ϕi, πt ◦ Φ′0,i(f ) − π0 ◦ ϕi < σ/2 for all f ∈ H and t ∈ [0, 1], and there is a unitary wi,0 ∈ Mri (in the place of u) such that π1 ◦ Φ0,i = Adwi,0 ◦ π1 ◦ Φ′0,i. The same argument shows that, for each 1 ≤ i ≤ k, there are two unital homomorphisms Φ1,i, Φ′1,i : C(X) → C([0, 1], Mri ) such that π1 ◦ Φ1,i = ψ1,i ◦ ϕ′, πt ◦ Φ1,i(f ) − ψ1,i ◦ ϕ′(f ) < σ/2, π1 ◦ Φ′1,i = π1 ◦ ϕi, πt ◦ Φ′1,i(f ) − π0 ◦ ϕi < σ/2 for all f ∈ H and t ∈ [0, 1], and there is a unitary wi,1 ∈ Mri (in the place of u) such that π0 ◦ Φ1,i = Adwi,1 ◦ π0 ◦ Φ′1,i. wi,0(0) = wi,0, wi,0(1) = 1Mri Choose two continuous paths {wi,0(t) : t ∈ [0, 1]}, {wi,1(t) : t ∈ [0, 1]} in Mri such that For each 1 ≤ i ≤ k, by the continuity of γi, there is 1 > ǫi > 0 such that and wi,1(1) = 1Mri and wi,1(0) = wi,1. γi(τx)(h) − γi(τy)(h) < σ/4 for all h ∈ H, provided that x − y < ǫi, where τx and τy are the extremal trace of Ii concentrated on x and y respectively. ǫi − 1)) ◦ π1 ◦ Φ′0,i, Define the map ϕi : C → Ii by ǫi ◦ Φ0,i, π 3t Ad(wi,0( 3t ǫi ◦ Φ′0,i, π3− 3t 1/2−ǫi ◦ ϕi, π t−ǫi 1−2πi /3−1/2 ◦ Φ′1,i, π 1−2εi /3−t Ad(wi,1( (1−εi/3)−t π t−1+εi/3 πt ◦ ϕi = ǫi/3 ◦ Φ1,i, ǫi/3  )) ◦ π0 ◦ Φ′1,i, 179 t ∈ [0, ǫi/3), t ∈ [ǫi/3, 2ǫi/3), t ∈ [2ǫi/3, ǫi), t ∈ [ǫi, 1/2), t ∈ [1/2, 1 − 2εi/3], t ∈ [1 − 2ǫi/3, 1 − ǫi/3], t ∈ [1 − ǫi/3, 1]. Then, π0 ◦ ϕi = ψ0,i ◦ ϕ′ and π1 ◦ ϕi = ψ1,i ◦ ϕ′. One can also estimate, by the choice of εi and the definition of ϕi, that τt ◦ ϕi(h) − γi(τt)(h) < σ for all t ∈ [0, 1], where τt is the extremal tracial state of Ii concentrated on t ∈ [0, 1]. Define Φ : C(X) → C([0, 1], F2) by Φ(f ) =Lk ϕi(f ) for all f ∈ C(X). Define ϕ : C(X) → C([0, 1], F2) ⊕ F2 by (Φ(f ), ϕ′(f )). By (e 17.5), ϕ is a homomorphism from C(X) to D. By (e 17.6) and (e 17.2), one has that i=1 (e 17.5) (e 17.6) τ ◦ ϕ(g) − γ(τ )(h) < σ for all h ∈ H We may write πe ◦ ϕ(f ) = Pm and for all τ ∈ T (D), as desired. To see the last part of the lemma, one assumes that D ∈ C0. Consider πe◦ ϕ : C → F1 (where D = D(F1, F2, ψ0, ψ1) as above). Since πe◦ϕ has finite dimensional range, it is a point-evaluation. i=1 f (xi)pi for all f ∈ C(X), where {x1, x2, ..., xm} ⊂ X and {p1, p2, ..., pm} ⊂ F1 is a set of mutually orthogonal projections. Let I = {f ∈ C(X) : f (x1) = 0} and ı : I → C(X) be the embedding. It follows that [πe ◦ ϕ ◦ ı] = 0. By 3.5, [πe] is injective on each Ki(A) and each Ki(A, Z/kZ) (k ≥ 2, i = 0, 1). Hence [ϕ ◦ ı] = 0. Choose Ψ(f ) = f (x) · 1A. Since X is connected, [ϕ] = [Ψ] (see the end of the Remark 4.4). Corollary 17.2. Let X be a connected finite CW-complex, and put C = C(X). Let ∆ : C q,1 + \ {0} → (0, 1) be an order preserving map. Let H ⊆ C be a finite subset and let σ > 0. Then there exists a finite subset H1 ⊆ C 1 +\{0} and a positive integer M such that for any D ∈ C with the dimension of any irreducible representation of D at least M, for any continuous affine map γ : T (D) → T (C) satisfying γ(τ )(h) > ∆(h) for all h ∈ H1 and for all τ ∈ T (D), there is a homomorphism ϕ : C → D such that Proof. Let H1,1 be the subset of Lemma 17.1 with respect to H and σ. Then put τ ◦ ϕ(h) − γ(τ )(h) < σ for all h ∈ H. σ1,1 = min{∆(h) : h ∈ H1,1}. Let H1,2 be the finite subset of Lemma 17.1 with respect to σ1,1, and then put σ1,2 = min{∆(h) : h ∈ H1,2}. Let M be the positive integer of Lemma 17.1 with respect to σ1,2. Then it follows from Lemma 17.1 that the finite subset and the positive integer M satisfy the statement of the corollary. H1 := H1,1 ∪ H1,2 Theorem 17.3. Let X be a connected finite CW complex, and let A ∈ B0 be a unital separable simple C-algebra. Suppose that γ : T (A) → Tf (C(X)) is a continuous affine map. Then, for any σ > 0, any finite subset H ⊂ C(X)s.a., there exists a unital homomorphism h : C(X) → A such that [h] = [Ψ] and τ ◦ h(f ) − γ(τ )(f ) < σ for all f ∈ H, (e 17.7) where [Ψ] is a point-evaluation. 180 Proof. Without loss of generality, one may assume that every element of H has norm at most one. Let H1,1 be the finite subset of Lemma 17.1 with respect to H (in place of H), σ/4 (in place of σ), and C (in place of C). Since γ(T (A)) ⊆ Tf (C(X)), there is σ1,1 > 0 such that γ(τ )(h) > σ1,1 for all h ∈ H1,1 for all τ ∈ T (A). Let H1,2 ⊆ C+ be the finite subset of Lemma 17.1 with respect to σ1,1. Again, since γ(T (A)) ⊆ Tf (C(X)), there is σ1,2 > 0 such that γ(τ )(h) > σ1,2 for all h ∈ H1,2 for all τ ∈ T (A). Let M be the constant of Lemma 17.1 with respect to σ1,2. Note that A ∈ B0. By (1) and (2) of Lemma 16.12, there is a C-subalgebra D ∈ A with D ∈ C0, a continuous affine map γ′ : T (D) → T (C) such that 1 γ′( τD)(f ) − γ(τ )(f ) < σ/4 for all τ ∈ T (A) for all f ∈ H, (e 17.8) τ (p) where p = 1D, τ (1 − p) < σ/(4 + σ), γ′(τ )(h) > σ1,1 for all τ ∈ T (D) for all h ∈ H1,1 and γ′(τ )(h) > σ1,2 for all τ ∈ T (D) for all h ∈ H1,2. (e 17.9) (e 17.10) Moreover, since A is simple, one may assume that the dimension of any irreducible repre- sentation of D is at least M (see 10.1). Thus, by (e 17.9) and (e 17.10), one applies Lemma 17.1 to D, C, and γ′ (in the place of γ) to obtain a homomorphism ϕ : C → D such that (e 17.11) Moreover, we may assume that [ϕ] = [Φ0] for some point evaluation Φ0 : C → D, since we τ ◦ ϕ(f ) − γ′(τ )(f ) < σ/4 for all f ∈ H for all τ ∈ T (D). assume that D ∈ C0. Pick a point x ∈ X, and define h : C → A by f 7→ f (x)(1 − p) ⊕ ϕ(f ) for all f ∈ C. For any f ∈ H, one has τ ◦ h(f ) − γ(τ )(f ) ≤ τ ◦ ϕ(f ) − γ(τ )(f ) + σ/4 1 τD)(f ) + σ/2 τ (p) < τ ◦ ϕ(f ) − γ′( 1 < τ ◦ ϕ(f ) − τ ◦ ϕ(f ) + 3σ/4 < σ. Define Φ : C → A by Φ(f ) = f (x)(1 − p) ⊕ Φ0(f ) for all f ∈ C. Then [h] = [Φ]. Corollary 17.4. The statement of Theorem 17.3 holds for C = P Mm(C(X))P , where X is a connected finite CW-complex and P is a projection in Mm(C(X)). τ (p) 18 KK-attainability of the building blocks Definition 18.1. (9.1 of [63]) Let D be a class of unital C∗-algebras. A C∗-algebra C is said to be KK-attainable with respective to D if for any A ∈ D and any α ∈ KK(C, A)++, there exists a sequence of completely positive linear maps Ln : C → A such that lim n→∞kLn(ab) − Ln(a)Ln(b)k = 0 for all a, b ∈ C and [Ln] = α. (e 18.2) In what follows, we will use Bu0 for the class of those C∗-algebras of the form A ⊗ U, where (e 18.1) A ∈ B0 and U is any UHF-algebras of infinite type. 181 Theorem 18.2. (Theorem 5.9 of [52]) Let A be a separable C-algebra satisfying UCT and let B be a amenable separable C-algebra. Assume that A is the closure of an increasing sequence {An} of residually finite dimensional C∗-subalgebras. Then for any α ∈ KL(A, B), there exist two sequences of completely positive contractions ϕ(i) n : A → B ⊗ K (i = 1, 2) satisfying the following: n (ab) − ϕ(i) (1) kϕ(i) (2) for any n, the images of ϕ(2) n (a)ϕ(i) n (b)k → 0 as n → ∞, and for any finite subset P ⊂ K(A), [ϕ(i) n ]P are well defined for sufficiently large n, n are contained in a finite dimensional sub-C-algebra of B⊗K (3) for each finite subset P ⊂ K(A), there exists m > 0 such that n ]P = α + [ϕ(2) [ϕ(1) n ]P for all n ≥ m, n is a homomorphism on An. (4) for each n, we may assume that ϕ(2) Lemma 18.3. Let C = A(F1, F2, ϕ0, ϕ1) ∈ C and let A ∈ C be another C∗-algebra. Let κ : K0(C) → K0(A) be an order preserving homomorphism such that, for any nonzero element p ∈ K0(C)+, there exists an integer N ≥ 1 such that N κ(p) > [1A]. Then there is σ > 0 satisfying the following: For any τ ∈ T (A), there exists t ∈ T (C) such that t(h) ≥ σZ[0,1] T(ı(h(t)))dµ(t) for all h ∈ C+ and τ (κ(p)) τ (κ([1C ]) = t(p) for all p ∈ K0(C)+, (e 18.3) (e 18.4) where ı : C → C([0, 1], F2) ⊕ F1 is the natural embedding and T(b) =Pk i=1 tri(b) for all b ∈ F2 and where tri is the normalized tracial state on the i-th simple summand of F2, and µ is the Lebesgue measure on [0, 1]. Proof. Write A = A(F ′1, F ′2, ϕ′0, ϕ′1) ={(g, c) ∈ C([0, 1], F ′2) ⊕ F ′1 : g(0) = ϕ0′(c) and g(1) = ϕ1′(c), c ∈ F ′1}. Suppose that F ′1 has l′ . Let πi : K0(F ′1) → Z be the i-th projection. View πi ◦ κ as a positive homomorphism from K0(C) → R. Since simple summands so that K0(F ′1) = Zl′ one has that N κ(x) > [1A] for all x ∈ K0(C)+ \ {0}, (πi ◦ κ)(x) > 0 for all x ∈ K0(C)+ \ {0}. (e 18.5) It follows from [40] that there are positive homomorphisms Li : K0(C([0, 1], F2)⊕ F1) → R such that Li ◦ ı∗0 = πi ◦ κ and Li(ej ) = αi,j>0, j = 1, 2, ...k + l, i = 1, 2, ..., l′, where ej = ψj ◦ ı([1C ]) and ψj is the projection from C([0, 1], F2) ⊕ F1 to the j-th summand of C([0, 1], F2), 1 ≤ j ≤ k, and ψj is the projection from C([0, 1], F2)⊕ F1 to (j − k)-th simple summand of F1, k + 1 ≤ j ≤ k + l. Moreover, by (e 18.5), for each i,Pj αi,j > 0. Note that k+lXj=1 rankψj(p) rankψj(1C ) αi,j = rankψj(p) rankψj(1C ) k+lXj=1 Li(ej) = Li ◦ ı∗0([p]) = πi ◦ κ([p]) (e 18.6) for any projection p ∈ Mm(C), where m ≥ 1 is an integer. Choose min{αi,j : 1 ≤ j ≤ k + l, 1 ≤ i ≤ l′} σ = . kPi,j αi,j 182 1 tτ (h) ≥ σZ[0,1] ) ) ) kXj=1 kXj=1 kXj=1 1 Pj αi,j Pj αi,j Pj αi,j Pj αi,j 1 1 ti(p) = ( = ( = ( = ( αi,jZ[0,1] tr′j(p) tr′j(1C ) dµ(t) + αi,j rank(ψj(p)) rank(ψj(1C )) + Li(ei) rankψj(p) rankψj(1C ) + αi,j αi,j tr′j(p) rankψj(p) tr′j(1C ) lXj=k+1 rankψj(1C )) rankψj(1C ) rankψj(p) Li(ei) lXj=k+1 lXj=k+1 )(πi ◦ κ(p)) = πi ◦ κ(p)/πi ◦ κ([1C ]) For any τ ∈ T (A), by 2.8 of [53], there is τ′ ∈ T (F ′1) such that τ′ ◦ πe(x) = τ (x) for all x ∈ K0(A). i=1 λi,τ = 1 and each tri is the tracial state of the Write τ′ =Pl′ i-th simple summand of F ′1. For each i, define i=1 λi,τ tri, where 0 ≤ λi ≤ 1,Pl′ αi,jZ[0,1] ti(s) = ( j=1 αi,j )( 1 kXj=1 tr′j(s) tr′j(1C ) dµ(t) + Pk+l for all s ∈ C, where each tr′j is the tracial state on the j-th simple summand of F2 for 1 ≤ j ≤ k, and each tr′j is the tracial state on the j-th simple summand of F1 for k + 1 ≤ j ≤ l. In general, if τ ∈ T (A), define tτ =Pl′ i=1 λi,τ ti. It is straightforward to verify that αi,j lXj=k+1 tr′j(s) tr′j(1C ) ) T(ı(h))dµ(t) for all h ∈ C+. Moreover, for each i, by (e 18.6), for all projection p ∈ Mm(C). This implies that tτ (x) = τ (κ(x))/τ (κ([1C ])) for all x ∈ K0(C) and for all τ ∈ T (A). Proposition 18.4. Let S ∈ C and N ≥ 1. There exists an integer K ≥ 1 satisfying the following: For any positive homomorphism κ : K0(S) → K0(A) which satisfies κ([1S ]) ≤ [1A] and N κ([p]) > [1A] for any p ∈ K0(S)+ \ {0}, where A ∈ C, there exists a homomorphism ϕ : S → MK(A) such that ϕ∗0 = Kκ. If we further assume κ([1S ]) = [1A], then ϕ can be chosen to be unital. Proof. Write S = A(F1, F2, ϕ0, ϕ1) ={(f, g) : (f, g) ∈ C([0, 1], F2) ⊕ F1 : f (0) = ϕ0(g), f (1) = ϕ1(g)}. Denote by ı : S → C([0, 1], F2) ⊕ F1 the embedding. To simplify the notation, without loss of generality, by replacing S by Mr(S) for some integer r ≥ 1, we may assume that projections in S generates K0(S). Note that, by the assumption, κ([p]) associated with a full projection of MN (A). Without loss of generality, by applying 3.19, we may assume that that [1A] = κ([1S ]). 183 Let σ > 0 be given by 18.3 (associated with integer N ). Define ∆ : Sq,1 + \ {0} → (0, 1) by ∆(h) = σZ[0,1] T(h(t))dµ(t) for all h ∈ S+, where T(c) = Pk i=1 tri(c) for all c ∈ C([0, 1], F2), tri is the normalized tracial state on the i-th simple summand of F2 (so we assume that F2 has k simple summands), and where µ is the Lebesgue measure on [0, 1]. Let H1, δ > 0, P and K be the finite subset and constants of Theorem 16.10 with respect to S, ∆, an arbitrarily chosen H and an arbitrarily chosen 1 > σ1 > 0 (in place of σ). We may assume that P ⊂ S is a finite subset of projections such that every projection q ∈ S is equivalent to one of projections in P. Let Q ⊂ Mr′(A) (for some r′ ≥ 1 be a finite subset of projections such that κ(P) can be represented by projections in Q. It follows from 16.11 that there is a finite subset T of extreme points of T (A) and there exists a continuous affine map γ′ : T (A) → ▽ such that γ′(τ )(p) − τ (p) < δ/2 for all p ∈ Q, (e 18.7) where ▽ is the convex hull of T . exact. Therefore, by [5] and [42], for each s ∈ ▽, there is tracial state ts ∈ T (S) such that Note that any C∗-algebra in the class C is of type I, it is amenable and in particular it is rS(ts)(x) = rA(s)(κ(x)) for all x ∈ K0(S), (e 18.8) where rS : T (S) → S[1S](K0(S)) and rA : T (A) → S[1A](K0(A)) are the induced maps from the tracial state spaces to the state space of K0, respectively. It follows from 18.3 that we may choose ts such that ts(h) ≥ ∆(h) for all h ∈ S+. (e 18.9) For each s ∈ T , define λ(s) = ts which satisfies (e 18.8) and (e 18.9). This extends to a continuous affine map λ : ▽ → T (S). Put γ = λ ◦ γ′. Then, for any τ ∈ T (S), γ(τ )(h) ≥ ∆(h) for all h ∈ H1 and γ(τ )(q) − τ (κ([q])) = λ(γ′(τ ))(q) − γ′(τ )(κ([q])) + γ′(τ )(κ([q])) − τ (κ([q])) = γ′(κ([q])) − κ([q]) < δ/2 (e 18.10) (e 18.11) (e 18.12) for all projections q ∈ S. One then applies 16.10 to obtain a unital homomorphism ϕ : S → MK (A) such that [ϕ] = Kκ. Lemma 18.5. Let C ∈ C. Then there is M > 0 satisfying the following: Let A1 ∈ B1 and +(C)) → let A = A1 ⊗ U for some infinite dimensional UHF-algebra and let κ : (K0(C), K0 +(A)) be a strictly positive homomorphism with multiplicity M . Then there exists a (K0(A), K0 homomorphism ϕ : C → Mm(A) (for some integer m ≥ 1) such that ϕ∗0 = κ and ϕ∗1 = 0. Proof. Write C = A(F1, F2, ϕ0, ϕ1). Denote by M the constant of 15.3 for G = K0(C) ⊂ K0(F1) = Zl. Let κ : K0(C) → K0(A) be a unital positive homomorphism satisfying the condition of the lemma. Since K0(A) is finitely generated, simple and κ is strictly positive, there is N such that for any nonzero positive element x ∈ K0(C)+, one has that N κ(x) > 2[1A]. Let K be the natural number of Proposition 18.4 with respect to C and N . 184 We may also assume that Mr(C) contains a set of minimal projections such that every minimal element of K0(C)+ \ {0} is represented by a minimal projection from the set. By Lemma 15.6, for any positive map κ with multiplicity M , one has κ = κ1 + κ2 and there are positive homomorphisms λ1 : K0(C) → Zn, γ1 : Zn → K0(A), λ2 : K0(C) → K0(C′) such that λ1 has multiplicity M, λ2 has multiplicity M K, κ1 = γ1 ◦ λ1, κ2 = ı∗0 ◦ λ2 and C′ ⊂ A is a C∗-subalgebra with C′ ∈ C, where ı : C′ → A is the embedding. Moreover, λ2([1C ]) = [1C′] and N λ2(x) > λ([1C ]) > 0 (e 18.13) for any x ∈ K0(C)+ \ {0}. Let R0 be as in 15.3 associated with K0(C) = G ⊂ Zl and λ1 : K0(C) → Zn , which has multiplicity M . Let λ1([1C ]) = (r1, r2, ..., rn), where ri ∈ Z+, i = 1, 2, ..., n. Put F3 = Mr1 ⊕ Mr2 ⊕ ··· ⊕ Mrn. Since A has stable rank one, there is a homomorphism ψ0 : F3 → A such that (ψ0)∗0 = γ1. Write U = limn→∞(MR(n), hn), where hn : MR(n) → MR(n+1) is a unital embedding. Choose R(n) ≥ R0. Consider the unital homomorphism jF3 : F3 → F3 ⊗ MR(n) defined by jF3(a) = a⊗ 1MR(n) for all a ∈ F3, and consider the unital homomorphism ψ0⊗ hn,∞ : F ⊗ MR(n) → A⊗ U defined by ψ0 ⊗ hn,∞(a⊗ b) = ψ0(a)⊗ hn,∞(b) for all a ∈ F3 and b ∈ MR(n). We check, for any projection p ∈ F3, ((ψ0 ⊗ hn,∞) ◦ jF3)∗0([p]) = [ψ0(p) ⊗ 1U ] = (ψ0)∗0([p]) ∈ K0(A). (e 18.14) It follows that (ψ0 ⊗ hn,∞)∗0 ◦ (jF3)∗0 = (ψ0)∗0. (e 18.15) Now the map (jF3)∗0 ◦ λ1 : K0(C) → K0(F3) = Zn has multiplicity M R(n). Applying 15.3, we obtain a positive homomorphism λ′1 : K0(F1) = Zl → K0(F3) such that (λ′1)∗0 ◦ (πe)∗0 = (jF3)∗0 ◦ λ1. The above can be summarized by the following commutative diagram: Zl ↑(πe)∗0 K0(C) λ′ 1→ K0(F3) ⊗ MR(n) λ1−→ ↑(jF3 )∗0 K0(F3) (ψ0⊗hn,∞)∗0 −→ γ1=(ψ0)∗0−→ K0(A ⊗ U ) k K0(A). We obtain a homomorphism h0 : F1 → F3 ⊗ MR(n) such that (h0)∗0 = λ′1. Define h1 = h0 ◦ πe : C → F3 ⊗ MR(n) and h2 = (ψ0 ⊗ ψn,∞) ◦ h1 : C → A ⊗ U. Then, by the commutative diagram above, (e 18.16) Since λ2 has multiplicity K, there exists λ′2 : K0(C) → K0(C′) such that Kλ′2 = λ2. Since (h2)∗0 = κ1. K0(C′) is weakly unperforated, λ′2 is positive. Moreover, by (e 18.13), KN λ′2(x) > Kλ′2([1C ]) = λ2([1C ]) = [1C′] > 0. (e 18.17) Since K0(C′) is weakly unperforated, we have N λ′2(x) > λ′2([1C ]) > 0 for all x ∈ K0(C)+ \ {0}. (e 18.18) There is a projection e ∈ Mk(C′) for some integer k ≥ 1 such that λ′2([1C ]) = [e]. Define C′′ = eMk(C′)e. By (e 18.17), e is full in C′. In fact K[e] = [1C′′]. In other words, MK (C′′) ∼= C′. By 3.19, C′′ ∈ C. By applying 18.4, we obtain a unital homomorphism h′2 : C → C′′ such that (ϕ′2)∗0 = Kλ′2 = λ2. Put h2 = ı ◦ λ2. Note that [1A] = κ([1C ]) = κ1([1C ]) + κ2([1C ]), by conjugating a unitary, without loss of generality, we may assume that h1(1C ) + h2(1C ) = 1A. Then it is easy to check that ϕ : C → A defined by ϕ(c) = h1(c) + h2(c) for all c ∈ C meets the requirements. 185 Lemma 18.6. (cf. Lemma 9.8 of [63]) Let A be a unital C∗-algebra and let B1 be a unital separable simple C∗-algebra in B0 and let B = B1 ⊗ U for some UHF-algebra of infinite type and C ∈ C0 be a C∗-subalgebra of B. Let G ⊂ K(A) be a finitely generated subgroup. Suppose that there exists an F-δ-multiplicative contractive completely positive linear map ψ : A → C ⊂ B such that [ψ]G is well defined. Then, for any ε > 0, there exists C∗-subalgebra C1 ∼= C of B and an F-δ-multiplicative contractive completely positive linear map L : A → C1 ⊂ B such that (e 18.19) [L]G∩K0(A,Z/kZ) = [ψ]G∩K0(A,Z/kZ) and τ (1C1) < ε for all τ ∈ T (B) and for all k ≥ 1 so that G ∩ K0(A, Z/kZ) 6= {0}, where L and ψ are viewed as maps to B. Furthermore, if [ψ]G∩K0(A) is positive then so is [L]G∩K0(A). Proof. We use the fact that U ⊗ U ∼= U. So we may write, without loss of generality, that C ⊂ B1 ⊗ U ⊗ 1. Without loss of generality, we may assume that ψ(1A) = p is a projection. Let 1 > ε > 0. Suppose that G ∩ K0(A, Z/kZ) = {0} for all k ≥ K for some integer K ≥ 1. Find a projection e0 ∈ U such that τ0(e0) < ε and 1U = e0 +Pm i=1 pi, where m = 2lK! and 1/l < ε and p1, p2, ..., pm are mutually orthogonal and mutually equivalent projections in U. Choose C1 = C⊗e0. Then C1 ∼= C. Let ϕ : C → C1 be the isomorphism defined by ϕ(c) = c ⊗ e0 for all c ∈ C. Put L = ϕ ◦ ψ. Note that K1(C) = K1(C1) = {0}. Both [L] and [ψ] maps K0(A, Z/kZ) to K(B)/kK0(B) and factor through K0(C, Z/kZ). It follows that [L]G∩K0(A.Z/kZ) = [ψ]G∩K0(A,Z/kZ). In case that [ψ]G∩K0(A) is positive, from the definition of L, it is clear that [L]G∩K0(A) is also positive. Theorem 18.7. Let C and A be unital stably finite C∗-algebras and let α ∈ KKe(C, A)++. (i) If C ∈ H, or C ∈ C, and A1 is a unital simple C-algebra in B0 and A = A1 ⊗ U for some UHF-algebra U of infinite type, then there exists a sequence of completely positive linear maps Ln : C → A such that lim n→∞kLn(ab) − Ln(a)Ln(b)k = 0 for all a, b ∈ C and [Ln] = α; (e 18.20) (ii) If C ∈ C0 and A1 ∈ B1, the above also holds; (iii) If C = Mn(C(S2)) for some integer n ≥ 1, A = A ⊗ U and A1 ∈ B1, then there is a unital homomorphism h : C → A such that [h] = α; (iv) If C ∈ H with torsion K1(C), C 6= Mn(C(S2)), and A = A1 ⊗ Q, where A1 is unital and A has stable rank one, then there exists a unital homomorphism h : C → A such that [h] = α; (v) If C = Mn(C(T)) for some integer n ≥ 1, then for any unital C∗-algebra A with stable rank one, there is a unital homomorphism h : C → A such that [h] = α. Proof. Let us first consider (iii). This is a special case of Lemma 2.19 of [83]. Let us provide a proof here. In this case one has that K0(C) = Z ⊕ kerρC ∼= Z ⊕ Z is free and K1(C) = {0}. Write A = A1 ⊗ U, where K0(U ) = D ⊂ Q is identified with a dense subgroup of Q and 1U = 1. Let α0 = αK0(C). Then α0([1C ]) = [1A] and α0(x) ∈ kerρA for all x ∈ kerρC. Let ξ ∈ kerρC = Z be a generator and α0(ξ) = ζ ∈ kerρA. Let B0 be a the unital simple AF-algebra with where (K0(B), K0(B)+, [1B0 ]) = (D ⊕ Z, (D ⊕ Z)+, (1, 0)), (D ⊕ Z)+ = {(d, m) : d > 0, m ∈ Z} ∪ {(0, 0)}. 186 It follows from [29] that there is a unital homomorphism h0 : C → B such that h∗0(ξ) = (0, 1). There is a positive order-unit preserving homomorphism λ : D → K0(A) (given by the embedding a → 1A ⊗ a from U → A1 ⊗ U ). Define a homomorphism κ0 : K0(B) → K0(A) by κ0(r) = λ(r) for all r ∈ D and κ0((0, 1)) = ζ. Since A has stable rank one, it is known and easy to find a unital homomorphism ϕ : B → A such that (e 18.21) ϕ∗0 = κ0. Define L = ϕ ◦ h0. Then, [L] = α. This proves (iii). For (iv), we note that Ki(A) is torsion free and divisible. Write C = P Mn(C(X))P, where X is a connected finite CW complex and P ∈ Mn(C(X)) is a projection. Note that X 6= S2. In this case K0(C) = Z ⊕ Tor(K0(C)), K1(C) = {0}, or K0(C) = Z and K1(C) is finite. Suppose that P has rank r ≥ 1. Choose x0 ∈ X. Let πx0 : C → Mr be defined by πx0(f ) = f (x0) for all f ∈ C. Suppose that e = (1, 0) ∈ Z ⊕ Tor(K0(C)) or e = 1 ∈ Z. Choose a projection p ∈ A such that [p] = α0(e) (this is possible since A has stable rank one). There is a unital homomorphism h0 : Mr → A such that h0(e1,1) = p, where e1,1 ∈ Mr is a rank one projection. Define h : C → A by h = h0 ◦ πx0. One verifies that [h] = α. If C ∈ H and C 6= S2, the statement follows from the same argument as that of Lemma 9.9 of [63] by replacing Lemma 9.8 of [63] by 18.6 above (and replace F by C and F1 by C1) in the proof of Lemma 9.9 of [63]. Now we prove (i) and (ii). Assume that C ∈ C. By considering Ad w ◦ LnC for suitable unitary w (in Mr(A)), we may replace C by Mr(C) for some r ≥ 1 so that K0(C)+ is generated by minimal projections {p1, p2, ..., pd} ⊂ C (see 3.15). Since A is simple and α(C+ \ {0}) ⊂ K0(A)+ \ {0}, there exists an integer N ≥ 1 such that N α([p]) > 2[1A] for all [p] ∈ K0(C)+ \ {0}. (e 18.22) Let M ≥ 1 (in place of K) be the integer given by 18.4 associated with N and C. Since C has a separating family of finite-dimensional representations, by Theorem 18.2, there exist two sequences of completely positive contractions ϕ(i) n : C → A⊗K (i = 0, 1) satisfying the following: n (ab) − ϕ(i) (a) kϕ(i) (b) for any n, ϕ(1) finite subset P ⊂ K(C), the map [ϕ(i) (c) for each finite subset P ⊂ K(C), there exists m > 0 such that n ]P for all n > m, n (a)ϕ(i) n is a homomorphism with finite dimensional range and, consequently, for any n ]P are well defined for all sufficiently large n, [ϕ(0) n ]P = α + [ϕ(1) n (b)k → 0, for all a, b ∈ C, as n → ∞, Since C is semiprojective and the positive cone of the K0-group is finitely generated, there are homomorphisms ϕ0 and ϕ1 from C → A ⊗ K such that [ϕ0] = α + [ϕ1]. Without lose of generality, let us assume that ϕ0 and ϕ1 are -homomorphisms from C to Mr(A) for some r. Note that Mr(A) ∈ B0 (or in B1, when C ∈ C0). Since Ki(C) is finitely generated (i = 0, 1), there exists n0 ≥ 1 such that every element κ ∈ KL(C, A) is determined by κ on Ki(C) and Ki(C, Z/nZ) for 2 ≤ n ≤ n0, i = 1, 2 (see Corollary 2.11 of [17]). Let P ⊂ K(C) be a finite subset which generates Mi=0,1 (Ki(C) ⊕ M2≤n≤n0 Ki(C, Z/nZ)). Choose K = n0!. Let G be a finite subset of Mr(A) which contains {ϕ0(pi), ϕ1(pi); i = 1, ..., d}. We may assume that {[pi], i = 1, 2,··· d} ⊂ P. 187 Let T = max{τ (ϕ0(pi)) + KM τ (ϕ1(pi)) : 1 ≤ i ≤ d; τ ∈ T (A)}. Choose r0 > 0 such that N T r0 < 1/2. (e 18.23) Let Q = [ϕ0](P) ∪ [ϕ1](P) ∪ α(P). Let 1 > ε > 0. By Lemma 14.12, for ε and r0 above, there is a non-zero projection e ∈ Mr(A), a C∗-subalgebra B ∈ C0 (or B ∈ C for case (ii)) with e = 1B, G-ε-multiplicative contractive completely positive linear maps L1 : Mr(A) → (1−e)Mr(A)(1−e) and L2 : Mr(A) → B with the following properties: (1) kL1(a) + L2(a) − ak < ε/2 for all a ∈ G; (2) [Li]Q is well defined, i = 1, 2; (3) [L1]Q + [ı ◦ L2]Q = [id]Q; (4) τ ◦ [L1](g) ≤ r0τ (g) for all g ∈ P0 and τ ∈ T (A); (5) For any x ∈ Q, there exists y ∈ K(B) such that x − [L1](x) = [ı ◦ L2](x) = KM [ı](y) and, (6) There exist positive elements {fi} ⊂ K0(B)+ such that for i = 1, ..., n, α([pi]) − [L1](α([pi]) = [ı ◦ L2](α([pi])) = KM ı∗0(fi). where ı : B → A is the embedding. By (5), since K = n0!, [ı ◦ L2 ◦ ϕ0]Ki(C,Z/nZ)∩P = [ı ◦ L2 ◦ ϕ1]Ki(C,Z/nZ)∪P = 0, i = 0, 1, (e 18.24) and n = 1, 2, ..., n0. It follows that [L1 ◦ ϕ0]Ki(C,Z/nZ)∩P = [ϕ0]Ki(C,Z/nZ)∩P [L1 ◦ ϕ1]Ki(C,Z/nZ)∩P = [ϕ1]Ki(C,Z/nZ)∩P , and (e 18.25) (e 18.26) i = 0, 1 and n = 1, 2, ..., n0. Furthermore, consequently for the case B ∈ C0 we have K1(B) = 0, and It follows that [ı ◦ L2]K1(C)∩P = 0. [L1 ◦ ϕ0]K1(C)∩P = [ϕ0]K1(C)∩P , [L1 ◦ ϕ1]K1(C)∩P = [ϕ1]K1(C)∩P . (e 18.27) (e 18.28) (e 18.29) In the second case when we assume that C ∈ C0 and A1 ∈ B1, then K1(C) = 0. Therefore (e 18.28) above also holds. Denote by Ψ := ϕ0 ⊕LKM−1 ϕ1. One then has [L1 ◦ Ψ]Ki(C, Z/nZ)∩P = [L1 ◦ ϕ0]Ki(C, Z/nZ)∩P + (KM − 1)[L1 ◦ ϕ1]Ki(C, Z/nZ)∩P = [L1 ◦ ϕ0]Ki(C, Z/nZ)∩P − [L1 ◦ ϕ1]Ki(C, Z/nZ)∩P = [ϕ0]Ki(C, Z/nZ)∩P − [ϕ1]Ki(C, Z/nZ)∩P = αKi(C, Z/nZ)∩P , 188 where i = 0, 1, n = 1, 2, ..., n0! By (e 18.22), (4) and (e 18.23), N (τ (α([pi]) − [L1 ◦ Ψ]([pi])) ≥ 2 − N r0T ≥ 3/2 for all τ ∈ T (A). (e 18.30) Since the strict order on K0(A) is determined by traces, one has that α([pi])−[L1◦Ψ]([pi]) > [1A]. Moreover, one also has α([pi]) − [L1 ◦ Ψ]([pi]) = α([pi]) − ([L1 ◦ α]([pi]) + KM [L1 ◦ ϕ1]([pi])) = (α([pi]) − [L1 ◦ α]([pi])) − KM [L1 ◦ ϕ1]([pi]) = KM (ı∗0(fj) − [L1] ◦ [ϕ1]([pi])) = KM f′j, where f′j = fj − [L1] ◦ [ϕ1]([pi]). Note that f′j ∈ K0(A)+ \ {0}, j = 1, 2, ..., d. Define a homomorphism β : K0(C) → K0(A) by β([pj]) = M f′j, j = 1, 2, ..., d. Since the K0(C)+ is generated by [p1], [p2], ..., [pd], β(K0(C)+ \ {0}) ⊂ K0(A)+ \{0}. Since β has multiplicity M , By the choice of M and by Lemma 18.5, there exists a -homomorphism h : C → Mr(A) (may not be unital) such that h∗0 = β and h∗1 = 0. Consider the map ϕ′ := L1 ◦ Ψ ⊕ (LK i=1 h) : C → A ⊗ K, one has that [ϕ′]K0(C)∩P = [L1 ◦ Ψ]K0(C)∩P + Kβ = αK0(C)∩P . It is clear that and therefore K[h]Ki(C, Z/nZ)∩P = 0, i = 0, 1, n = 1, 2, ..., n0 [ϕ′]Ki(C, Z/nZ)∩P = [L1 ◦ Ψ]Ki(C, Z/nZ)∩P = αKi(C, Z/nZ)∩P , where i = 0, 1, n = 1, 2, ..., n0 . We also have that [ϕ′]K1(C)∩P = [L1 ◦ Ψ]K1(C)∩P = αK1(C)∩P . Therefore Since [ϕ′(1C )] = [1A] and A has stable rank one, there is a unitary u in a matrix algebra of [ϕ′]P = αP . A such that the map ϕ = ad(u) ◦ ϕ′ satisfies ϕ(1C ) = 1A, as desired. Case (v) is standard and is well known. Corollary 18.8. Any C-algebra A of Theorem 14.8 is KK-attainable with respective to Bu0. Proof. Note that A is an inductive limit of C-algebras which are finite direct sums of C∗- algebras in H and C0. Since KK-attainability passes to inductive limits, by Theorem 18.7, A is KK-attainable with respective to Bu0. Corollary 18.9. Let C ∈ C, let A ∈ Bu0 and α : KK(C, A)++ be such that α([1C ]) = [p] for some projection p ∈ A and α is strictly positive. Then there exists a homomorphism ϕ : C → A such that ϕ∗0 = α. Proof. This is a special case of Theorem 18.7 since C is semiprojective. 189 Corollary 18.10. Let A ∈ Bu0. Then there exists a unital simple C-algebra B1 = lim−→(Cn, ϕn), for B = B1 ⊗ U , we where each Cn is in C0, and a UHF algebra U of infinite type such that have (K0(B), K0(B)+, [1B], T (B), rB) = (ρA(K0(A)), ρA(K0(A)+), ρA([1A]), T (A), rA). Moreover, for each n, there is a unital homomorphism hn : Cn ⊗ U → A such that ρA ◦ (hn)∗0 = (ϕn,∞ ⊗ idU )∗0. (e 18.31) Proof. Consider the tuple (ρA(K0(A)), ρA(K0(A)+), ρA([1A]), T (A), rA). Since A ∼= A⊗ U1, for a UHF algebra U1 of infinite type, it has the (SP) property (see [4]), and therefore the ordered group (K0(A), K0(A)+, [1A]) has the (SP) property in the sense of Theorem 14.8; that is, for any positive real number 0 < s < 1, there is g ∈ K0(A)+ such that τ (g) < s for any τ ∈ T (A). Then it is clear that the scaled ordered group (ρA(K0(A)), (ρA(K0(A)+), ρA(1A)) also has the (SP) property in the sense of Theorem 14.8. Therefore, by Theorem 14.8, there is a simple unital C-algebra B1 = lim−→(Cn, ϕn), where each Cn ∈ C0 such that (K0(B1), K0(B1)+, [1B1 ], T (B1), rB1 ) ∼= (ρA(K0(A)), ρA(K0(A)+), ρA(1A), T (A), rA). Let U = U1 and B = B1 ⊗ U . Recall that A ⊗ U = A. Therefore (K0(B), K0(B)+, [1B], T (B), rB) ∼= (ρA(K0(A)), ρA(K0(A)+), ρA(1A), T (A), rA). Clearly, B has the inductive limit decomposition B = lim−→(Cn ⊗ U, ϕn ⊗ idU ). For each n, consider the positive homomorphism (ϕn,∞)∗0 : K0(Cn) → K0(B1) ∼= ρA(K0(A)). Since K0(Cn) is torsion free, K0(Cn)+ is finitely generated, and the strict order on the projections of A is determined by traces, there is a positive homomorphism κn : K0(Cn) → K0(A) such that ρA ◦ κn = (ϕn,∞)∗0 and κ([1Cn ]0) = [1A]0. By Corollary 18.9, there is a unital homomorphism h′n : Cn → A such that (h′n)∗0 = κn. It is clear that hn := h′n ⊗ idU satisfies the desired condition. Lemma 18.11. Let C ∈ C. Let σ > 0 and let H ⊂ Cs.a. be any finite subset. Let A ∈ Bu0. Then for any κ ∈ KLe(C, A)++ and any continuous affine map γ : T (A) → Tf (C) which is compatible to κ, there is a unital homomorphism ϕ : C → A such that [ϕ]∗ = κ and τ ◦ ϕ(h) − γ(τ )(h) < σ for all h ∈ H. Moreover, the above also holds if C ∈ C0 and A ∈ Bu1. Proof. Without loss of generality, one may assume that every element of H has norm at most one. Let κ and γ be given. Define ∆ : C q,1 + \ {0} → (0, 1) by ∆(h) = inf{γ(τ )(h)/2 : τ ∈ T (A)}. Let H1 ⊆ C +, δ, and K be the finite subset, positive constant and the positive integer of 16.10 with respect to C, ∆, H and σ/8 (in place of σ). Let P ⊂ K0(C)+ be a finite subset which generates K0(C)+. 190 By Lemma 16.12, there is a C∗-subalgebra D ⊆ A with D ∈ C and with 1D = p ∈ A, a continuous affine map γ′ : T (D) → T (C) such that γ′( 1 τ (p) τ )(f ) − γ(τ )(f ) < σ/8 for all τ ∈ T (A) for all f ∈ H, (e 18.32) where p = 1′D, τ (1 − p) < σ/(8 + σ), γ′(τ )(h) > ∆(h) for all τ ∈ T (D) for all h ∈ H1. (e 18.33) Denote by ı : D → pAp the embedding. Moreover, by 16.12, there are positive homomorphisms κ0,0 : K0(C) → K0((1− p)A(1− p)) and κ1,0 : K0(C) → K0(D) such that κ1,0 is strictly positive and has the multiplicity K, κK0(A) = κ0,0 + ı∗0 ◦ κ1,0 and γ′(τ )(p) − τ (κ1,0(p)) < δ, p ∈ P, τ ∈ T (D). Since we assume that A ⊗ U ∼= A, by the last part of 16.12, we may assume that κ0,0 is also strictly positive. Therefore, by Lemma 16.10 (for suitable modified projection p which still satisfies (e 18.32) and (e 18.33)), there is a homomorphism ϕ1 : C → D ⊗ U ⊂ A ⊗ U such that (ϕ1)∗0 = κ1,0 and τ ◦ ϕ1(h) − γ′(τ )(h) < σ/4 for all h ∈ H. Since A is simple, Ki((1 − p)A(1 − p)) = Ki(A), i = 0, 1. Let κ0 = κ − [ı ◦ ϕ1] ∈ KL(C, A) = KL(C, (1 − p)A(1− p)). Then κ0K0(C) = κ0,0. Since κ0K0(C) = κ0.0, it is strictly positive. Note that (1−p)A(1−p)⊗U ∼= (1−p)A(1−p). Therefore, by Theorem 18.7, since C is semi-projective, there is a homomorphism ϕ0 : C → (1− p)A(1− p) such that [ϕ0] = κ0. Note that, this holds for both the case that A ∈ Bu0 and the case that C ∈ C0 and A ∈ Bu1. Consider the homomorphism ϕ := ϕ0 ⊕ ı ◦ ϕ1 : C → (1 − p)A(1 − p) ⊕ D ⊆ A. One has that [ϕ] = κ and, for all h ∈ H, τ ◦ ϕ(h) − γ(τ )(h) ≤ τ ◦ ϕ1(h) − γ(τ )(h) + σ/4 1 < τ ◦ ϕ1(h) − γ′( 1 < τ ◦ ϕ1(h) − τ (p) τD)(f ) + σ/2 τ (p) τ ◦ ϕ1(h) + 3σ/4 < σ as desired. It turns out that the KK-attainability implies the following existence theorem. Proposition 18.12. Let A ∈ B0, and assume that A is KK-attainable with respective to Bu0. Then for any B ∈ Bu0, any α ∈ KL++(A, B), and any affine continuous map γ : T (B) → T (A) which is compatible to α, there is a sequence of completely positive linear maps Ln : A → B such that lim n→∞Ln(ab) − Ln(a)Ln(b) = 0 for all a, b ∈ A, [Ln] = α and lim n→∞ sup{τ ◦ Ln(f ) − γ(τ )(f ) : τ ∈ T (A)} = 0 for all f ∈ A. Proof. The proof is the same as that of Proposition 9.7 of [63]. Instead of using Lemma 9.6 of [63], one uses Lemma 18.11. 191 Remark 18.13. The condition that A = A ⊗ U (A ∈ B1, or A ∈ B0) in the statements in Section 18 can be easily eased to the condition that A ∈ B1 (or B0) such that A is tracially approximately divisible. As a consequence, with almost no additional efforts, one can also ease the the same condition for B. Since we eventually do not need to assume A ∼= A⊗ U, to shorten the length of this article, we conveniently use the stronger assumption. 19 The class N1 Let A be a unital C∗-algebra such that A⊗Q ∈ B1. In this section, we will show that A⊗Q ∈ B0. Note that this is proved without assuming A ⊗ Q is nuclear. However, it implies that N1 = N0. If we assume that A ⊗ Q has finite nuclear dimension, using 18.11 and a characterization of of T AS by Winter, a much shorter proof of 19.2could easily given here. Lemma 19.1. Let A ∈ B1 such that A ∼= A ⊗ Q. Then the following holds: For any ε > 0, any two non-zero mutually orthogonal elements a1, a2 ∈ A+ and any finite subset F ⊂ A, there exists a projection q ∈ A and a C∗-subalgebra C1 ∈ C with 1C1 = q such that (1) k[x, q]k < ε/16, x ∈ F, (2) pxp ∈ε/16 C1, x ∈ F, and (3) 1 − q . a1. Suppose ∆ : (C1)q,1 + \ {0} → (0, 1) is an order preserving map such that + \ {0}. τ (c) ≥ ∆(c) for all τ ∈ T (A) and c ∈ (C1)1 ( By 12.3 such ∆ always exists.) Suppose also that H ⊂ (C1)+ \ {0} and F1 ⊂ C1 are finite subsets. Then, there exists another projection p ∈ A and a C∗-subalgebra C2 ∈ C with p = 1C2 such that p ≤ q, and a unital homomorphism H : C1 → C2 such that (4) k[x, p]k < ε/16, x ∈ F, (5) kH(y) − pypk < ε/16, y ∈ F1, and (6) 1 − p . a1 + a2. Moreover K1(C1) = Zm ⊕ G0 such that H∗1(G0) = {0}, and H∗1Zm and (j ◦ H)∗1Zm are both injective, where j : C2 → A is the embedding (m could be zero, in this case, G0 = K1(C1)). Furthermore, we may assume that τ (j ◦ H(c)) ≥ 3∆(c)/4 for all c ∈ H for all τ ∈ T (A). Proof. Since A ∈ B1, there exists a projection q ∈ A and a C∗-subalgebra C1 ∈ C with 1C1 = q such that (a) k[x, q]k < ε/16, x ∈ F, (b) pxp ∈ε/16 C1, x ∈ F, and (c) 1 − q . a1. There are two non-zero mutually orthogonal elements a′2 and a3 ∈ a2Aa2. Note that A ∼= A ⊗ Q. Therefore K1(A) is torsion free. Denote by j : C1 → qAq the embedding. Since K1(C1) is finitely generated, we may write K1(C1) = G1 ⊕ G0, where G1 ∼= Zm1, j∗1G1 is injective and j∗1G0 = 0. Define σ = min{∆(h)/16 : h ∈ H} > 0. 192 s=1 exp(ihs,k), hs,k ∈ (qAq)s.a, s = 1, 2, ..., l(k), k = 1, 2, ..., l. y ∈ F1 such that kpxp − yk < ε/16. hs,k, exp(ihs,k), s = 1, 2, ..., l(k), k = 1, 2, ..., l. Choose an element a′′2 ∈ a′2Aa′2 such that dτ (a′′2) < σ for all τ ∈ T (A) , where recall dτ (a) = limn→∞ τ (a Suppose that G0 is generated by [v1], [v2], ..., [vl], where vi ∈ U (C1) (note C1 has stable rank one). Then, we may write vk = Ql(k) (since j∗1([vk]) = 0 inK1(A)) where Let F1 be a finite subset of C1 which also has the following property: if x ∈ F, there is Let F2 be a finite subset of qAq to be determined which at least contains F ∪ F1 ∪ H and Let 0 < δ < min{ε/64, σ/4} to be determined. Since qAq ∈ B1, one obtains a non -zero (d) k[x, q1]k < δ, x ∈ F2, (e) q1xq1 ∈δ C2, x ∈ F2, (f) 1 − q1 . a2′′. With sufficiently small δ and large F2, using the semi-projectivity of C1, we obtain unital projection q1 ∈ qAq and a C∗-subalgebra C2⊂qAq such that 1 n ). homomorphisms h1 : C1 → C2 such that kh1(a) − q1aq1k < min{ε/16, σ} for all a ∈ F1 ∪ H. One computes that τ (j ◦ h1(c)) ≥ 7∆(c)/8 for all c ∈ H, (e 19.1) (e 19.2) where we also use j for the embedding from C2 into qAq. Note that, when F2 is large enough, (h1)∗1(G0) = {0}. We may write K1(C1) = G2⊕G2,0⊕G2,0,0, where G2 ∼= Zm2 with m2 ≤ m1, G2 is a subgroup of G1, G2,0,0 ⊃ G0, (h1)∗1(G2,0,0) = {0}, (j ◦ h1)∗1G2,0 = 0 and (j ◦ h1)∗1G2 is injective. Here we use the fact that K1(A) is torsion free. If G2,0 = {0}, we are done. Otherwise m2 < m1. We also note that (j ◦ h1)∗1(G2,0 ⊕ G2,0,0) = {0}. We will repeat the process to construct h2, and consider h2◦h1. Then we may write K1(C1) = G3⊕G3,0⊕G3,0,0 with G3 ∼= Zm3 (m3 ≤ m2) G3 ⊂ G2, G3,0,0 ⊃ G2,0⊕G2,0,0, (j◦h2◦h1)∗1(G3,0) = {0} and (j ◦ h2 ◦ h1)∗1G3 is injective. Again, if G3,0 = {0}, we are done (choose H = h2 ◦ h1). Otherwise, m3 < m2 < m1. We continue this process. Since G1 is a commutative noetherian ring, this process stops at a finite stage. This proves the lemma. Theorem 19.2. Let A1 be a unital separable amenable C∗-algebra such that A1⊗ Q ∈ B1. Then A1 ⊗ Q ∈ B0. Proof. Let A = A1 ⊗ Q. Suppose that A ∈ B1. Let ε > 0, let a ∈ A+ \ {0} and let F ⊂ A. Since A has property (SP), we obtain three non-zero and mutually orthogonal projections e0, e1, e2 ∈ aAa. There exists a projection q1 ∈ A and a C∗-subalgebra C1 ∈ C with 1C1 = q1 such that k[x, q1]k < ε/16 and q1xq1 ∈ε/16 C1 for all x ∈ F, and 1 − q1 . e0. (e 19.3) (e 19.4) Let F1 ⊂ C1 be a finite subset such that, for any x ∈ F, there is y ∈ F1 such that kq1xq1 − yk < ε/16. For each h ∈ (C1)+ \ {0}, define ∆(h) = (1/2) inf{τ (h) : τ ∈ T (A)}. Then ∆ : (C1)q,1 + \ {0} be a finite subset, γ1, γ2 > 0, δ > 0, G ⊂ C1 be a finite subset, P ⊂ K(C1) be a finite subset, H2 ⊂ (C1)s.a. be a + \ {0} → (0, 1) preserves the order. Let H1 ⊂ (C1)1 193 finite subset and let U ⊂ J (1) C = C1, for ε/16 (in place of ε) and F1 (in place of F) and ∆/2 (in place of ∆). such that q2 ≤ q1, and a unital homomorphism H : C1 → C2 such that By 19.1, there exists another projection q2 ∈ A and a C∗-subalgebra C2 ∈ C with q2 = 1C2 (U (C1)/U0(C1)) be a finite subset required by 12.5 and 12.6 for c k[x, q2]k < ε/16 for all x ∈ F, kH(y) − q2yq2k < ε/16 for all y ∈ F1, τ (j ◦ H(c)) ≥ 3∆(c)/4 for all c ∈ H and 1 − q2 . e0 + e1. (e 19.5) (e 19.6) (e 19.7) (e 19.8) Moreover, we may write K1(C1) = Zm ⊕ G00, where H∗1(G00) = {0}, H∗1Zm and (j ◦ H)∗1Zm are injective, where j : C2 → A is the embedding. Let A2 = q2Aq2 and denote by j1 : C2 → A2 the embedding. By 14.8, there exists a unital simple C∗-algebra B ∼= B ⊗ Q such that B = limn→∞(Bn, ın) such that each Bn = Bn,0 ⊕ Bn,1 with Bn,0 ∈ H and Bn,1 ∈ C0, ın is injective, lim n→∞ max{τ (1Bn,0 ) : τ ∈ T (B)} = 0 and Ell(B) = Ell(A2). (e 19.9) We may assume that U = U1 ∪ U0, such that π(U1) generates Zm, and π(U0) ⊂ G0, where π : U (C1)/CU (C1) → K1(C1) is the quotient map. Here J (1) : K1(C1) → U (C1)/CU (C1) is a fixed splitting map defined in 2.15. Suppose that ¯v1, ¯v2, ..., ¯vm form a set of free generators for J (1) c (Zm). Without loss of generality, we may assume that U1 = {¯v1, ¯v2, ..., ¯vm}. Put c γ3 = min{(∆/2)(h) : h ∈ H1}. Note H‡(U0) ⊂ U0(C2)/CU (C2). Choose a finite subset H3 ⊂ (C2)s.a. and σ > 0 which have for any two unital homomorphisms h1, h2 : C2 → D (for any unital the following property: C∗-algebra D of stable rank one), if then τ ◦ h1(g) − τ ◦ h2(g) < σ for all g ∈ H3, dist(h‡1(¯v), h‡2(¯v)) < γ2/8 (e 19.10) (e 19.11) for all ¯v ∈ H‡(U0) ⊂ U0(C2)/CU (C2). Without loss of generality, we may assume that khk ≤ 1 for all h ∈ H1 ∪ H2 ∪ H3. Let κ : Ell(A2) → Ell(B) be the above identification. So κ ◦ [j2] ∈ KKe(C2, B)++. It follows from 18.11 that there exists a unital homomorphism ϕ : C2 → B such that [ϕ] = κ ◦ [j2] and τ (ϕ(h)) − γ(τ )(h) < min{γ1, γ2, γ3, σ}/8 (e 19.12) (e 19.13) for all h ∈ H(H1) ∪ H(H2) ∪ H3, where γ : T (B) → Tf (C1) is induced by κ and the embedding j1. In particular, ϕ∗1 is injective on H∗1(Zm). Since C2 is semi-projective, without loss of generality, we may assume that ϕ(C2) ⊂ B1. We may also assume that B1 = B1,0 ⊕ B1,1 and τ (1B1,0 ) < min{τ (κ([e2]))/4, γ1/8, γ2/8, γ3/8, σ/8} for all τ ∈ T (B). (e 19.14) Furthermore, identifying Ell(A2) and Ell(B), we have (ı1,∞ ◦ ϕ)∗1 = ϕ∗1, which is is injective on H∗1(Zm). Consequently one has that (ı1,∞)∗1 is injective on ϕ∗1(H∗1(Zm)). 194 Let G1 = H‡(J (1) c (Zm)) ⊂ U (C2)/CU (C2) and G0 = H‡(J (1) (G00)). Note, by the con- struction, G0 ⊂ U0(C2)/CU (C2). Since κK1(C) is an isomorphism and (ı1,∞)∗1 is injective on ϕ∗1(H∗1(Zm)), ϕ‡G1 is injective Let H4 = P (ϕ(H1 ∪ H2 ∪ H3)), where P : B1 → B1,1 is the projection. Let e3∈A be a projection such that [e3] = κ−1([ı1,∞)(1B1,1 )]). It follows from the second part of 18.11 that there is a unital monomorphism ψ1 : B1,1 → e3Ae3 such that c [ψ1] = κ−1 ◦ [(ı1,∞)B1,1 ] and (e 19.15) τ ◦ ψ1(g) − γ′(τ )(ı1,∞(g)) < min{τ (([e3]))/4, γ1/8, γ2/8, γ3/8, σ/8} for all g ∈ H4 (e 19.16) for all τ ∈ T (A), where γ′ : T (A) → Tf (B1,1) induced by κ−1 and ı1,∞. Write B1,0 = B1,0,1⊕B1,0,2, where B1,0,1 is a finite direct sum of circle algebras and K1(B1,0,2) is finite. Since B ∼= B ⊗ Q, we may assume that (ı1,∞)∗1K1(B1,0.2) = {0}. Since A2 ∼= A2 ⊗ Q, by part (iii) and (iv) of 18.7, there exists a unital homomorphism ψ2 : B1,0,2 → e4A2e4 such that [ψ2] = κ−1 ◦ [(ı1,∞)B1,0 ], where e4 is a projection orthogonal to e2 and [e4] = κ−1 ◦ [ı1,∞](1B1,0,2 ). As in the proof of 18.7, (ψ2)∗1 = 0. Let ψ3 : B1,1 ⊕ B1,0,2 → (e3 + e4)A2(e3 + e4) by ψ3 = ψ1 ⊕ ψ2. Let P1 : B1 → B1,0,1. Then, since (ı1,∞)∗1K1(B1,0,2) = {0}, (P1)∗1ϕ∗1(H∗1(Zm)) is injective. Also P ‡1 is injective on ϕ‡◦H‡(J (1) (Zm))⊂ U (B1,0,1)/CU (B1,0,1). Then G′1 ∼= Zm. It follows from 18.7 that there is a unital homomorphism ψ′4 : B1,0,1 → (1 − e3 − e4)A2(1 − e3 − e4) such that [ψ′4] = κ−1 ◦ [ı1,∞B1,0,1 ]. Let (Zm)). Put G′1 = P ‡1◦ϕ‡◦H‡(J (1) c c zi = P ‡1 ◦ ϕ‡ ◦ H‡(¯vi) and ξi = ψ‡3 ◦ (1B1 − P1)‡ ◦ ϕ‡ ◦ H‡(¯vi), (e 19.17) i = 1, 2, ..., m. It should be noted that, since (ψ1)∗1 = (ψ2)∗1 = 0, ξi ∈ U0(A2)/CU (A2), i = 1, 2, ..., m. Moreover, since (ψ′4)∗1 ◦ (P1)∗1 ◦ ϕ∗1(x) = j∗1(x) for all x ∈ Zm ⊂ K1(C1), π(¯vi)π((ψ′4)‡(zi))−1 = 0 in K1(A). Define a homomorphism λ : G′1 → U (A2)/CU (A2) by λ(zi) = ¯vi((ψ′4)‡(zi)ξi)−1, i = 1, 2, ..., m. (e 19.18) (e 19.19) (e 19.20) Note that λ(zi) ∈ U0(A)/CU (A), i = 1, 2, ..., m. By 11.5, U0(A)/CU (A) is divisible. There exists a homomorphism ¯λ : U (B1,0,1)/CU (B1,0,1) → U (A2)/CU (A2) such that ¯λG′ = λ. Define a homomorphism λ1 : U (B1,0,1)/CU (B1,0,1) → U (A2)/CU (A2) by λ1(x) = (ψ′4)‡(x)¯λ(x) for all x ∈ U (B1,0,1)/CU (B1,0,1). By 11.10, the homomorphism 1 U ((1 − e3 − e4)A(1 − e3 − e4))/CU ((1 − e3 − e4)A2(1 − e3 − e4)) → U (A2)/CU (A2) is an isomorphism. Since B1,0,1 is a circle algebra, one easily obtains a unital homomorphism such that ψ4 : B1,0,1 → (1 − e3 − e4)A2(1 − e3 − e4) [ψ4] = [ψ′4] = κ−1 ◦ [ı1,∞B1,0,1 ] and ψ‡4 = λ1. (e 19.21) 195 Define ψ : B1 → A2 by ψ = ψ3 ⊕ ψ4. Then, we have [ψ] = κ−1 ◦ [ı1,∞]. Since [ı1,∞ ◦ ϕ] = κ ◦ [j], we compute that [ψ ◦ ϕ ◦ H] = κ−1 ◦ [ı1,∞] ◦ [ϕ ◦ H] = [j ◦ H] in KK(C1, A). Put C′1 = H(C1) and ψ′ = ψ ◦ ϕC′ 1 . Then ψ′ : C′1 → A2 is a monomorphism. We have [ψ′] = [jC′ 1 ]. (e 19.22) (e 19.23) (e 19.24) By (e 19.13), (e 19.14) and (e 19.16), τ ◦ ψ ◦ ϕ(g) − τ (g) < min{σ, γ1, γ2, γ3, σ} for all g ∈ H(H1) ∪ H(H2) ∪ H3. (e 19.25) In particular, τ ◦ ψ′(g) − τ (j(g)) < min{γ1, γ2, γ4} for all g ∈ H(H1) ∪ H(H2). We then compute that By (e 19.20) and the definition of λ1, we have τ ◦ ψ′(h) ≥ ∆(h)/2 for all h ∈ H1. (ψ′)‡(¯vi) = ¯vi, i = 1, 2, ..., m. By the choice of H3 and σ, we also have dist((ψ′)‡(¯v), (j ◦ H)‡(¯v)) < σ2 for all ¯v ∈ U0. (e 19.26) (e 19.27) (e 19.28) (e 19.29) It follows from 12.5 and 12.6 that there is a unitary U ∈ q2Aq2 such that (e 19.30) Let C3 = Ad U ◦ ψ(B1,1) and p = 1C3 . and p = Ad U ◦ ψ(1B1,1 ). For any y ∈ F1, ϕ(y)1B1,1 = 1B1,1, ϕ(y). Therefore Ad U ◦ ψ′(y)p = pAd U ◦ ψ ◦ ϕ(y) = pAd U ◦ ψ′(y). Hence, kAd U ◦ ψ′(x) − xk < ε/16 for all x ∈ F1. kpy − ypk ≤ kpy − pAd U ◦ ψ′(y)k + kpAd U ◦ ψ′(y) − ypk < ε/8 + ε/8 = ε/4 (e 19.31) for all y ∈ F1. Thus, combining with (e 19.3) and (e 19.5), kpx − xpk = kpq2x − xq2pk < 2ε/16 + kpq2xq2 − q2xq2pk < ε for all x ∈ F. (e 19.32) Let x ∈ F. Choose y ∈ F1 such that kq2xq2 − q2yq2k < ε/16. Then, by (e 19.30), kpxp − pAd U ◦ ψ′(y)pk ≤ kpxp − pq2xq2pk +kpq2xq2p − pypk + kpyp − pAd U ◦ ψ′(y)pk < ε/16 + ε/16 = ε/8. (e 19.33) (e 19.34) (e 19.35) (e 19.36) (e 19.37) However, pAd U ◦ ψ′(y)p = Ad U ◦ (ψ(ϕ(q2)ϕ(y)ϕ(q2)) ∈ C3 for all y ∈ F1. Therefore We then estimate that pxp ∈ε C3. [1 − p] ≤ [1 − q2] + [ψ(1B1,0 )] ≤ [e0 ⊕ e1 ⊕ e2] ≤ [a]. Since B1,1 ∈ C0, by applying 3.20, A ∈ B0. Corollary 19.3. If A is a unital separable amenable simple C∗-algebra such that A ⊗ Q ∈ B1, then, for any infinite dimensional UHF-algebra U, A ⊗ U ∈ B0. Proof. It follows from 19.2 that A ⊗ Q ∈ B0. Then, by 3.20 and by [79], A ⊗ U ∈ B0 for every UHF-algebra of infinite type. 196 20 KK-attainability of the C-algebras in B0 In the following, let us show an existence theorem for the maps from an algebra in the class B0 to a C-algebra in Theorem 14.8. The procedure is similar to that of Section 2 of [59], and roughly, we will construct a map factors through the C-subalgebras (in C0) of the given C-algebra in B0, and also require this map to carry the given KL-element. But since positive cone of the K0-groups of a C-algebra in C0 in general is not free, extra work has to be done to take care of this issue. 20.1. Let us proceed as that of Section 2 of [59]. Let A ∈ B0, and assume that A has the (SP) property. By Lemma 9.8, the C-algebra A can be embedded as a C-subalgebra of Q Mnk /L Mnk for some (nk), and therefore A is MF in the sense of Blackadar and Kirchberg (Theorem 3.2.2 of [3]). Since A is assumed to be amenable, by Theorem 5.2.2 of [3], the C- algebra A is strong NF, and hence, by Proposition 6.1.6 of [3], there is an increasing family of RFD sub-C-algebras {An} such that their union is dense in A. Let {x1, x2, ..., xn, ...} be a dense sequence of elements in the unit ball of A. Let P0 ⊆ M∞(A) be a finite subset of projections. We assume that x1 and P0 ⊆ M∞(A1). Consider a finite subset F0 ⊂ A1 ⊂ A with {1, x1} ⊆ F0, δ0 > 0, and a homomorphism h0 from A1 to a finite-dimensional C-algebra F0 which is non-zero on F0. Since A is assumed to have the (SP) property, by Lemma 2.1 of [81], there is a non-zero homomorphism h′ : F0 → A such that (1) e0x − xe0 < δ0/256 and (2) h′ ◦ h0(x) − e0xe0 < δ0/256 for all x ∈ F0, where e0 = h′(1). Since F0 has finite dimension, it follows from Arveson’s Extension Theorem that the ho- momorphism h0 : A1 → F0 can be extended to a contractive completely positive linear map from A to F0, and let us still denote it by h0. Put H = h′ ◦ h0 : A → A. Note that e0 = H(1). Since the hereditary C-subalgebra (1 − e0)A(1 − e0) is in the class B0 again, there is a projection q′1 ≤ 1 − e0 and a C- subalgebra S′1 ∈ C0 with 1S′ = q′1 such that 1 (3) q′1x − xq′1 < δ0/256 for any x ∈ (1 − e0)F0(1 − e0), (4) dist(q′1xq′1, S′1) < δ0/256 for any x ∈ (1 − e0)F0(1 − e0), and (5) τ (1 − e0 − q′1) < 1/16 for any tracial state τ on A. Put q1 = q′1 + e0 and S1 = S′1 ⊕ h′(F0). One has (6) q1x − xq1 < δ0/64 for any x ∈ {ab : a, b ∈ F0}, (7) dist(q1xq1, S1) < δ0/64 for any x ∈ {ab : a, b ∈ F0}, and (8) τ (1 − q1) = τ (1 − q′1 − e0) < 1/16 for any tracial state τ on A. Let F′0 ⊂ S1 be a finite subset such that dist(q1yq1,F′0) < δ0/16 for all y ∈ {ab : a, b ∈ F0}. Let G1 be a finite generating set of S1 which is in the unit ball. Since S1 is amenable, there is a contractive completely positive linear map L′0 : q1Aq1 → S1 such that (e 20.1) Set L0(a) = L′0(q1aq1) for any a ∈ A. Then L0 is a completely positive contraction from A to S1 such that kL′0(s) − sk < δ0/256 for all s ∈ G1 ∪ F′0. kL0(s) − sk < δ0/256 for all s ∈ G1 ∪ F′0. (e 20.2) 197 We estimate that L0 is {x1}-δ0/16-multiplicative. Since S1 is semi-projective, there is δ′1 > 0 and a finite subset G′1 ⊃ G1 in the unit ball such that for any G′1-δ′1-multiplicative contractive completely positive linear map L from S1 to any C-algebra, there is a homo- morphism h from S1 to that C-algebra such that L(a) − h(a) < δ0/2 for any a ∈ {L0(x1)} ∪ G1. Set F1 = {x2} ∪ F0 ∪ G′1. Set δ1 = min{δ′1, δ0/256}. Also there is a projection q2 ∈ A and a C-subalgebra S2 ∈ C0 with 1S2 = q2 such that (9) q2x − xq2 < δ1/16 for any x ∈ {ab : a, b ∈ F1}, (10) dist(q2xq2, S2) < δ1/16 for any x ∈ {ab : a, b ∈ F1}, and (11) τ (1 − q2) ≤ 1/32 for any tracial state τ on A. Let F′1 ⊂ S2 be a finite subset such that dist(q2yq2,F′1) < δ1/16 for all y ∈ {ab : a, b ∈ F1}. Let G2 be a finite generating set of S2 which is in the unit ball. Therefore there exists a completely positive contraction L′1 : q2Aq2 → S2 such that kL′1(a) − ak < δ1/16 for all a ∈ G2∪F′1. Define L1(a) = L′1(q2aq2) for all a ∈ A. It is a completely positive contraction from A to S2 such that kL1(a) − ak < δ1/16 for all a ∈ G2∪F′1, We estimate, from (9), the choice of F′1 and the above inequality, that L1 is F1-δ1/2- multiplicative. Therefore, there is a homomorphism h1 : S1 → S2 such that L1(a) − h1(a) < δ0/2 for any a ∈ {L0(x1)} ∪ G1. Since S2 is semi-projective, there is δ′2 > 0 and a finite subset G′2 ⊃ G2 in the unit ball such that for any G′2-δ′2-multiplicative contractive completely positive linear map L from S2 to any C-algebra, there is a homomorphism h′ from S2 to that C-algebra such that kL(a) − h′(a)k < δ1/2, a ∈ {L1(x1), L1(x2)} ∪ G2. Put F2 = {x3} ∪ F1 ∪ F′1 ∪ G′1 ∪ G′2. Set δ2 = min{δ′2, δ1/256}. Also there is a projection q3 ∈ A and a C-subalgebra S3 ∈ C0 with 1S3 = q3 such that (12) q3x − xq3 < δ2/16 for any x ∈ {ab : a, b ∈ F2}, (13) dist(q3xq3, S3) < δ2/16 for any x ∈ {ab : a, b ∈ F2}, and (14) τ (1 − q2) ≤ 1/64 for any tracial state τ on A. Let F′2 ⊂ S3 be a finite subset such that dist(q3yq3,F′2) < δ3/16 for all y ∈ {ab : a, b ∈ F2}. Let G3 be a finite generating set of S3 which is in the unit ball. There then exists a completely positive contraction L′2 : q3Aq3 → S3 such that kL′2(a) − ak < δ2/16 for all a ∈ G3 ∪ F′2. 198 Define L2(a) = L′2(q3aq3) for all a ∈ A. It is a completely positive contraction from A to S2 such that kL2(a) − ak < δ2/16 for all a ∈ G3 ∪ F′2. As estimated above, L2 is F2-δ2/2-multiplicative. Since S3 is semi-projective, there is δ′3 > 0 and a finite subset G′3 ⊃ G3 in the unit ball such that for any G′3-δ′3-multiplicative contractive completely positive linear map L from S3 to a C-algebra, there is a homo- morphism h′ from S3 to that C-algebra such that kL(a) − h′(a)k < δ2/2, Put F3 = {x4} ∪ F2 ∪ F′2 ∪ G′2 ∪ G′3. On the other hand, there is a homomorphism and h2 : S2 → S3 such that a ∈ {L2(x1), L2(x2), L2(x3)} ∪ G3. kL2(a) − h2(a)k < δ1, a ∈ {L1(x1), L1(x2)} ∪ G2. union in the unit ball of A, a decreasing sequence of positive numbers {δn} withP∞n=1 δn < Repeating this construction, one obtains a sequence of finite subsets F0,F1, ..., with dense ∞, a sequence of projections {qn} ⊂ A, a sequence of C-subalgebras Sn ∈ C0 with 1Sn = qn with finite generating subsets Gn ⊂ Sn, and a sequence of homomorphisms hn : Sn → Sn+1 and a sequence of Fn-δn/2-multiplicative contractive completely positive linear maps Ln : A → Sn+1 such that (15) qnx − xqn < δn−1/16 for all x ∈ Fn. (16) dist(qnxiqn, Sn) < δn−1/16, i = 1, ..., n. (17) τ (1 − qn) < 1/2n+1 for all tracial states τ on A. (18) Gn ⊂ Fn+1, where Gn is a finite set of generators for Sn. (19) Ln+1(a)− hn+1(a) < δn for all a ∈ Ln(Fn)∪Gn. and kLn(a)− ak < δn for all a ∈ Gn+1. 20.2. Let Ψn : A → (1−qn+1)A(1−qn+1) denote the cut-down map sending a to (1−qn+1)a(1− qn+1), and let Jn : A → A denote the map sending a to Ψn(a) ⊕ Ln(a). Note that Ψn and Jn are Fn-δn/2-multiplicative. Set Jm,n = Jn−1 ◦ ··· ◦ Jm and hm,n = hn−1 ◦ ··· ◦ hm : Sm → Sn. Note that Jm,n is Fm-δm-multiplicative. We also use Ln, Ψn, Jn, Jm,n, hm, and hm,n for their extensions on a matrix algebra over A. Using the same argument as that of Lemma 2.7 of [59], one has the following lemma. Lemma 20.3 (Lemma 2.7 of [59]). Let P ⊂ Mk(A) be a finite set of projections. Assume that F1 is sufficiently large and δ0 is sufficiently small such that [Ln ◦ J1,n]P and [Ln ◦ J1,n]G0 are well defined, where G0 is the subgroup generated by P. Then lim n→∞ sup τ∈T (A)τ ([ιn+1 ◦ Ln ◦ J1,n]([p])) − τ ([p]) = 0 for any p ∈ P. Furthermore, for any projection p ∈ P and k ≥ 1, we have τ (hk,k+n+1 ◦ [Lk−1]([p])) − τ (hk,k+n ◦ [Lk−1]([p]) < (1/2)n+k for all τ ∈ T (A) and lim n→∞ τ (hk,k+n ◦ [Lk−1]([p]) ≥ (1 − for all p ∈ P and τ ∈ T (A). 199 nXi=1 1/2i+k)τ ([Lk−1]([p])) > 0 Remark 20.4. Since A is stably finite and assumed to be amenable, therefore exact, any positive state of K0(A) is the restriction of a tracial state of A ([5] and [42]). Thus, the lemma above still holds if one replaces the trace τ by any positive state τ0 on K0(A). 20.5. Fix a finite subset P of projections of Mr(A) (for some r ≥ 1) and an integer N ≥ 1 such that [LN +i]P , [JN +i]P and [ΨN +i]P are all well defined. Keep notation in 20.2. Then, on P, [LN +1 ◦ JN ] = [LN +1 ◦ LN ] ⊕ [LN +1 ◦ ΨN ] = [hN +1 ◦ LN ] ⊕ [LN +1 ◦ ΨN ], and [LN +2 ◦ JN,N +2] = [LN +2 ◦ LN +1 ◦ JN ] ⊕ [LN +2 ◦ ΨN +1 ◦ JN ] = [LN +2 ◦ LN +1 ◦ LN ] ⊕ [LN +2 ◦ LN +1 ◦ ΨN ] ⊕[LN +2 ◦ ΨN +1 ◦ JN ] = [hN +1,N +3] ◦ [LN ] ⊕ [LN +2 ◦ LN +1 ◦ ΨN ] ⊕ [LN +2 ◦ ΨN +1 ◦ JN ]. Moreover, on P, [LN +n ◦ JN,N +n−1] = [hN +1,N +n+1] ◦ [LN ] ⊕ [LN +n ◦ ΨN +n−1 ◦ JN,N +n−1] ⊕[LN +n ◦ LN +n−1 ◦ ΨN +n−2 ◦ JN,N +n−2] ⊕[LN +n ◦ LN +n−1 ◦ LN +n−2 ◦ ΨN +n−3 ◦ JN,N +n−3] ⊕··· ⊕[LN +n ◦ LN +n−1 ◦ ··· ◦ LN +2 ◦ ΨN +1 ◦ JN ] ⊕[LN +n ◦ LN +n−1 ◦ ··· ◦ LN +1 ◦ ΨN ]. N +2 = LN +2 ◦ ΨN +1, ..., ψN N +n = LN +n ◦ ΨN +n−1, Set ψN n = 1, 2, .... N = LN , ψN N +1 = LN +1 ◦ ΨN , ψN 20.6. For each Sn, since the abelian group K0(Sn) is finitely generated and torsion free, there is a set of free generators {en ln} ⊆ K0(Sn). By Theorem 3.15, the positive cone of the K0(Sn) is finitely generated; denote a set of generators by {sn 0 (Sn). Then there is an rn × ln integer-valued matrix R′n such that ~rn = R′n~en, rn} ⊆ K + 2 , ..., sn 2 , ..., en 1 , sn 1 , en 1 , sn ln )T. In particular, for any ordered group H, 2 , ..., en where ~rn = (sn and any elements h1, h2, ..., hln ∈ H, the map en 7→ hi, i = 1, ..., ln induces an abelian-group homomorphism ϕ : K0(Sn) to H, and the map ϕ is positive (or strictly positive) if and only if rn)T and ~en = (en 2 , ..., sn 1 , en i R′n ~h ∈ H rn + (or R′n ~h ∈ (H+ \ {0})rn ), where ~h = (h1, h2, ..., hln )T ∈ H ln. Moreover, for each en i )+, (en (en i )− ∈ K0(Sn)+ and fix this decomposition. Define a rn × 2ln matrix i , write it as en i = (en i )+ − (en i )− for (e 20.3) (e 20.4) (e 20.5) (e 20.6) (e 20.7) (e 20.8) (e 20.9) (e 20.10) (e 20.11) Rn = R′n Then one has .   0 ··· 1 −1 0 1 −1 ··· 0 ... ... . . . ··· 0 0 0 ... 0 ... 0 0 0 0 0 ... ... 1 −1 ~rn = Rn~en,±, 200 where ~en,± = ((en ments h1,+, h1,−, ..., hln,+, hln,− ∈ H, the map en (or strictly positive) homomorphism if and only if 1 )−, ..., (en 1 )+, (en )+, (en ln )−)T. Hence, for any ordered group H, and any ele- 7→ (hi,+ − hi,−), i = 1, ..., ln induces a positive ln i Rn~h± ∈ H rn + (or Rn~h± ∈ (H+ \ {0})rn ) , where ~h± = (h1,+, h1,−, ..., hln,+, hln,−)T ∈ H ln. of Sn, there are integers mn one has Since {en 2 , ..., en 1 , en ln} is a set of generators of K0(Sn), for any projection p in a matrix algebra ln ([p]) such that for any homomorphism τ : K0(Sn) → R, 1 ([p]), ..., mn τ ([p]) = h ~mn(p), τ (~en)i = lnXi=1 mn i ([p])τ (en i ) = lnXi=1 mn i ([p])τ ((en i )+) − mn i ([p])τ ((en i )−), where ~mn([p]) = (mn 1 ([p]), ..., mn ln ([p]))T and ~en = (en 1 , en 2 , ..., en ln ). For each p ∈ Mm(A), for some integer m ≥ 1, denote by [ψk,k+j(p)] an element in K0(Sk+j+1) associated with ψk,k+j(p). Let ın : Sn → A be the imbedding. Denote by ln)). (ın)∗0 : ~en 7→ (((ın)∗0(en 2 ), ..., (ın)∗0(en 1 ), (ın)∗0(en Then, by Lemma 20.3 and Remark 20.4, one has the following lemma. Lemma 20.7. With the notion same as the above, for any p ∈ P0, for each fixed k, one has that τ (p) = lim n→∞ mk+j+1 i ([ψk k+j(p)])τ ((ık+n ◦ hk+j,k+n)∗0(ek+j+1 i ))) (ψk k+j(p)])τ ((ık+n ◦ hk+j,k+n)∗0(ek+j+1 i )−)) uniformly on S(K0(A)). Moreover, ρA ◦ (ın)∗0 ◦ hk+j,k+n(ek+j element in Aff(S(K0(A))) as n → ∞ uniformly. Proof. We first compute that, if j > 1, i,± ) converges to a strictly positive mk+j+1 i ([ψk k+j]([p]))τ ((ık+n ◦ hk+j,k+n)∗0(ek+j+1 i = τ ([Lk+n ◦ ··· ◦ Lk+j+1 ◦ Lk+j ◦ Ψk+j−1]([p])) )) (e 20.12) (e 20.13) and, if j = 1, mk+1 i ([ψk k+1]([p]))τ ((ık+n ◦ hk+1,k+n)∗0(ek+1 i = τ ([Lk+n ◦ ··· ◦ Lk+j+1 ◦ Lk]([p])). )) (e 20.14) (e 20.15) ( lk+j+1Xi=1 nXj=1 − mk+j+1 i lk+j+1Xi=1 lk+1Xi=1 201 Thus (see 20.5) mk+j i ([ψk k+j(p)])τ ((ın ◦ hk+j,k+n)∗0(ek+j i ))) = τ ([Lk+n ◦ ··· ◦ Lk+j+1 ◦ Lk]([p])) τ ([Lk+n ◦ ··· ◦ Lk+j+1 ◦ Lk+j ◦ Ψk+j−1]([p])) ( + lk+jXi=1 nXj=1 nXj=2 nXj=2 + = τ ([Lk+n ◦ ··· ◦ Lk+j+1 ◦ Lk]([p])) τ ([Lk+n ◦ ··· ◦ Lk+j+1 ◦ Lk+j ◦ Ψk+j−1 ◦ Jk,k+j−1]([p])) = τ ([Lk+n ◦ Jk,n+k−1]([p])). (e 20.16) (e 20.17) (e 20.18) (e 20.19) (e 20.20) (e 20.21) Thus the first part of the lemma then follows from 20.3. The second part also follows. One then has the following Corollary 20.8. Let P be a finite subset of projections in a matrix algebra over A, let G0 be the subgroup of K0(A) generated by P and let k ≥ 1 be an integer. Denote by ρ : G0 → ΠZ the map defined by [p] 7→ (mk 1(q0),−mk 1(q0), mk 2(q0),−mk 2(q0),··· mk lk (q0),−mk 1 mk+1 mk+2 1 1 (q1),−mk+1 (q2),−mk+2 1 (q1), mk+1 (q2), mk+2 2 2 lk (q0), (q1), ..., mk+1 lk+1 (q2), ..., mk+2 lk+2 2 (q1),−mk+1 (q2),−mk+2 2 lk+1 (q1),−mk+1 (q2),−mk+2 lk+2 (q1), (q2),··· ),(e 20.22) where qi = [ψk k+i(p)], i = 0, 1, 2, .... If ρ(g) = 0, then τ (g) = 0 for any trace over A. By the definition of the map ρ and H = h′ ◦ h0 : A → F0 → A, using the same argument as that of Lemma 2.12 of [59], one has the following lemma. Lemma 20.9. Let P be a finite subset of projections in Mk(A1) ⊆ Mk(A). Then there is a finite subset F1 ⊂ A1 and δ0 > 0 such that if the above construction starts with F1 and δ0, then ker ρ ⊂ ker[H] and ker ρ ⊂ ker[h0]. The K0-part of the existence theorem will almost factor through the map ρ, and this lemma will help us to handle the elements of K0(A) which vanish under ρ. Moreover, to get a such K0-homomorphism, one also needs to find a copy of the generating set of the positive cone of K0(S) inside the image of ρ as an ordered group for certain algebra S ∈ C0. In order to do so, one needs the following technical lemma, which is essentially Lemma 3.4 of [59]. Lemma 20.10 (Lemma 3.4 of [59]). Let S be a compact convex set, and let Aff(S) be the affine continuous functions on S. Let D be a dense ordered subgroup of Aff(S), and let G be an ordered group with the strict order determined by a surjective homomorphism ρ : G → D. Let {xij}1≤i≤r,1≤j<∞ be an r × ∞ matrix having rank r and with xij ∈ Z for each i, j. Let j ∈ G such that ρ(g(n) g(n) j } is a sequence of positive elements in D such that a(n) j → aj(> 0) uniformly on S as n → ∞. , where {a(n) ) = a(n) j j 202 Further suppose that there is a sequence of integers s(n) satisfying the following condition: j j )s(n)×1 be the part of (g(n) Letfvn = (g(n) Denote by yn = ρ(r)(fyn). Then there exists z = (zj)r×1 such that yn → z on S uniformly. fyn = (xij)r×s(n)fvn With the above condition, there exist δ > 0 and a positive integer K > 0 satisfying the )1≤j<∞ and let For some sufficiently large n, if M is a positive integer, and if z′ ∈ ( 1 K 3 G)r (i.e., there is z′′ ∈ Gr such that K 3 z′′ = z′) satisfies z− M z′ < δ, where z′ = (z′1, z′2,··· , z′r) with z′j = ρ(z′j) then there is a u = (cj )s(n)×1 ∈ Gs(n) following: + such that (xij)r×s(n) u = z′. Moreover, if each s(n) can be written as s(n) =Pn for each k, Rk is a rk × lk matrix with entries in Z so that ¯Rn = diag(R1, R2, ..., Rn) k=1 lk, where lk are positive integers, and satisfies ¯Rn¯gn > 0 (e 20.23) as an element in GPn that k=1 rk , where ¯gn = (gn 1 , gn 2 , ..., gn s(n))T , n = 1, 2, ...,, then we may choose u ¯Rn u > 0. (e 20.24) j j ) a(n) s(k−1)+1 Proof. The proof is repeating the argument of Lemma 3.4 of [59], and we need to show that if (e 20.23) holds, then u = (cj)s(n)×1 can be chosen to make (e 20.24) hold. a(n) M )T is strictly positive. Since a(n) s(k) M , ..., By (e 20.23), any entry of Rk( ρ(g(n) M < δ′k/2 and kaj/M − a(n) ) − j ∈ G+, then any entry of Rk(f (k) j → aj and the strict order on G is determined by ρ, there exists an integer Lk ≥ 1 and δ′k > 0 such that if n ≥ Lk and ρ(f (k) for some f (k) Lk+1 > Lk, k = 1, 2, ..... One may assume that {δ′k} is decreasing. Without loss go generality, let us assume that (xij)r×r has rank r. Set A′n = (xij)r×s(n). Then there is an invertible matrix B′ ∈ Mr(Q) such that B′A′n = Cn, where Cn = (Ir, D′n) for some r × (s(n) − r) matrix D′n. Moreover, there is an integer K such that all entries of KB′ and K(B′)−1 are integers. Evidently, KD′n is also a matrix with integer entries. ) and, for all n ≥ Ln, Since D is dense in Aff(S), there are ξn ∈ Gs(n) such that ξn = ( d(n) s(k))T is positive. We may assume that j = s(k − 1) + 1, ..., s(k) s(k−1)+1, ..., f (k) j /Mk < δ′k/4, j K 3ρ( d(n) j ) − a(n) j M < min{ 1 s(n)2 · 2n , δ′n/4}, j = 1, 2, ...., s(n). (e 20.25) Let wn = K 3ξn, and let ρ( d(n) j ) = d(n) j . Then K 3d(n) j → K 3dj = aj M > 0 203 uniformly on S. Set y′n = A′n wn = K 3A′nξn, and set y′n = ρ(s(n))(y′n), where ρ(m)Gm → Dm is the nature map induced by ρ : G → D. Then one has y′n → z/M uniformly on S. We then has Cn wn = K 3Cnξn = K 3B′A′nξn = B′ y′n and where v′n = (K 3 d(n) 1 , ..., K 3 d(n) r )T and Dn = (0, D′n). Ir v′n = B′ y′n − Dn wn, Since d(n) j → dj > 0 uniformly on S, there is N1 > 0 such that d(n) j ≥ inf{ dj 2 (τ ) : τ ∈ S} > 0 for all n ≥ N1 and j = 1, 2, ..., r. Let kr be a natural number such that r < s(kr), and set 0 < ǫ < min{δ′kr /8, inf{dj(τ ) : τ ∈ S}/32K 4 : j = 1, 2, ..., r}. There is N2 such that if n ≥ N2, then B′yn − B′z∞ < ǫ/4; and there is N3 > 0 such that B′(y′n) − B′(z/M )∞ < ǫ/4, n ≥ N3. There is δ > 0 depending only on B′ such that if z − M z′ < δ, one has B′y′n − B′z′∞ < ǫ/2, n ≥ N3. Let z′ and z′′ be described in the lemma. Set N = max{N1, N2, N3}. Since both KB′ and KDn are matrix over Z, u′ = KB′z′′ − KDnξn ∈ Gr. Let B′z′ − Dn wn = K 3B′z′′ − K 3Dnξn = K 2u′, (e 20.26) where u′ = (c1, c2, ..., cr) ∈ Gr and n ≥ N. Set u′′ = K 2u′, and let One may write and one has ρ(r)(u′) = (c1, c2, ..., cr) ∈ Dr. Iru′′ = B′z′ − Dn wn, ρ(r)(u′′) − ρ(r)(v′n)∞ = ρ(r)(B′ z′ − Dn wn) − ρ(r)(B′ y′n − Dn wn)∞ < ǫ. Set We calculate that u = (K 2c1, ..., K 2cr, K 3 d(n) r+1, ..., K 3 d(n) s(n))T . B′A′n u = (Ir, D′n)u = u′′ + (0, D′n)u = B′z′ − Dn wn + (0, D′n)u = B′z′. Consequently, A′n u = z′, since B′ is invertible. 204 In order to show that ¯Rn u > 0, it suffices to verify that K 2ρ(cj) − ρ(g(Lkr ) ) j M < δ′kr /2, j = 1, ..., r (e 20.27) and, if n ≥ Ln, K 3ρ( d(n) j ) − ρ(g(Ln) ) j M < δ′n/2, j = r + 1, ..., s(n). (e 20.28) Equation (e 20.28) follows from (e 20.25) directly. Let us verify (e 20.27). By (e 20.26), one has K 2ρ(u′) − ρ(g′) = B′ρ(z′) − Dnρ( w)n − ρ(g′) ≤ B′y′n − Dnρ( w)n − ρ(g′) + ǫ/2 ≤ B′ρ(A′n wn) − Dnρ( w)n − ρ(g′) + δ′kr /16 = ρ((B′A′n − Dn)( wn) − g′) + δ′kr /16 = ρ((Cn − Dn)( wn) − g′) + δ′kr /16 = K 3ρ(( )r×1) − ρ(g′) + δ′kr /16 ≤ δ′kr /2, d(n) j where g′ = (g(Lk) Equation (e 20.27). Thus, u is the desired solution. /M, ..., g(Lkr ) /M, g(Lkr ) 1 2 r /M ), and ρr : Gr → Dr is denoted by ρ. This verifies Definition 20.11. A unital stably finite C-algebra A is said to have the K0-density property if the image ρ(K0(A)) is dense in Aff(S[1](K0(A)), where S[1](K0(A)) is the convex set of the states of K0(A), i.e., the convex set of all positive homomorphisms r : K0(A) → R satisfying r([1]) = 1. Remark 20.12. By Corollary 7.9 of [39], the linear space spanned by ρ(K0(A)) is always dense in Aff(S[1](K0(A)). Therefore, the unital stably finite C-algebra in the form of A⊗U for a UHF- algebra U always has the K0-density property. Moreover, any unital stably finite C-algebra A which is tracially approximately divisible has the K0-density property. Remark 20.13. Not all C-algebras in B1 with the (SP) property satisfy the density property. The following is one example: Consider G = {(a, b) ∈ Q ⊕ Q : 2a − b ∈ Z} and G+ \ {0} = (Q+ \ {0} ⊕ Q+ \ {0}) ∩ G (e 20.29) and 1 = (1, 1) ∈ G+ as unit. Then (G, G+, 1) is a weakly unperforated rational Riesz simple ordered group (but not a Riesz group—see [76]). Evidently G has (SP) property; but the image of G is not dense in Aff(S[1](G) = R ⊕ R, as (1/2, 1/2) ∈ R ⊕ R is not in the closure of the image of G. We leave the details to the readers. Proposition 20.14. Let A ∈ B0 satisfy the density property and let B1 be an inductive limit C-algebra in Theorem 14.8 such that (K0(A), K0(A)+, [1A], K1(A)) ∼= (K0(B), K0(B)+, [1B], K1(B)), where B = B1⊗ U for a UHF-algebra U of infinite type. Let α ∈ KL(A, B) be an element which implements the isomorphism above. Then, for any P ⊂ P (A), there is a sequence of completely positive linear maps Ln : A → B such that lim n→∞Ln(ab) − Ln(a)Ln(b) = 0 for all a, b ∈ A and [Ln]P = αP as n → ∞. 205 Proof. By Lemma 9.8, A is the closure of an increasing union of RFD C∗-subalgebras {An}. We may assume P ⊂ K(A1). Let G = G(P) be the subgroup generated by P and let P0 ⊂ P be such that P0 generate G ∩ K0(A). Write P0 = {p1, ..., pl}, where p1, ..., pl are projections in a matrix algebra over A. Let G0 be the group generated by P0. Let F1 be a finite subset of A1 and let δ0 > 0 be such that any F1-δ0-multiplicative linear map, the map [L]P is well defined. Moreover, one requires that F1 and δ0 satisfy Lemma 20.9. Let k0 be an integer such that G(P) ∩ Ki(A, Z/kZ) = {0} for any k ≥ k0, i = 0, 1. By Theorem 18.2, there are two F1-δ0/2 multiplicative contractive completely positive linear maps Φ0, Φ1 from A to B ⊗ K such that [Φ0]P = αP + [Φ1]P and the image of Φ1 is in a finite dimensional C∗-subalgebra. Moreover, we may also assume that Φ1 is a homomorphism when it is restricted on A1, and the image is a finite-dimensional C-algebra. With Φ1 in the role of h0, we can proceed with the construction as described at the beginning of this section. We will keep the same notation. Consider the map ρ : G(P) ∩ K0(A) → l∞(Z) defined in Corollary 20.8. The linear span of {ρ(p1), ..., ρ(pl)} over Q will have finite rank, say r. So, we may assume that {ρ(p1), ..., ρ(pr)} are linearly independent and the Q-linear span of them give us the whole subspace. Therefore, there is an integer M such that for any g ∈ ρ(G0), the element M g is in the subgroup generated by {ρ(p1), ..., ρ(pr)}. Let xij = (ρ(pi))j, and zi = ρB(α([pi]) ∈ D, where D = ρB(K0(B)) in Aff(S[1](K0(B)). Since A is assumed to have the density property, so is B. Therefore the image D is a dense subgroup of Aff(S[1](K0(B))). , ek+j i j = ρB(g(n) j α([ın ◦ hk+j,n(ek+j i,+ )]) = g(n) i,− )]) = g(n) s(j−1)+2i, Let {Sj} be the sequence of C∗-subalgebras in C0 in the construction at the beginning of this section. Fix k ≥ 1. Let ek+j i,± ∈ K0(Sk+j), i = 1, 2, ..., lk+j and Rk+j be rk+j × 2lk+j matrix as described in 20.6. Let s(j) =Pj i=1 2lk+i, j = 1, 2, .... Put s(j−1)+2i−1, α([ın ◦ hk+j,N0+n(ek+j ), j = 1, 2, ..., s(n) =Pn i = 1, 2, ..., lk, and a(n) D+\{0}. It follows from Lemma 20.3 that limn→∞ a(n) s(j−1)+2i−1 = as(j−1)+2i−1 = ρB(α(g(n) 0 and limn→∞ a(n) 20.7,Pn j=1 xija(n) j → zi uniformly. Furthermore, by 20.6, ¯Rn¯gn > 0, ¯gn = (gn z M K 3(k0+1)!−1 Φ1 ⊕ ··· ⊕ Φ1). Since Φ1 factors through a finite-dimensional C-algebra, it is zero when restricted to K1(A) ∩ G and K1(A, Z/kZ) ∩ G for 2 ≤ k ≤ k0. Moreover, the map Φ1 ⊕ ··· ⊕ Φ1) vanishes on K0(A, Z/kZ) for 2 ≤ k ≤ k0. Therefore we have ( j=1 lk+j, n = 1, 2, .... Note that a(n) j ∈ s(j−1)+2i = as(j−1)+2i = ρB(α(g(n) s(j−1)+2i)) > s(j−1)+2i−1)) > 0 uniformly. Moreover, by So, Lemma 20.10 applies. Fix K and δ obtained from Lemma 20.10. Let Ψ := Φ0 ⊕ ( M K 3(k0+1)! 1 , gn 2 , ..., gn s(n))T . (e 20.30) } { z } { [Ψ]K1(A)∩G = αK1(A)∩G, [Ψ]K1(A,Z/kT)∩G = αK1(A,Z/kZ)∩G and [Ψ]K0(A,Z/kZ)∩G = αK0(A,Z/kZ)∩G. We may assume Ψ(1A) is a projection in Mr(B) for some integer r. We may also assume that there exist projections {p′1, ..., p′l} in B ⊗ K which are sufficiently close to {Ψ(p1), ..., Ψ(pl)} respectively, so that [p′i] = [Ψ(pi)]. Note that B ∈ B0, and hence the strict order on the projections of B is determined by traces. Thus there is a projection q′i ≤ p′i such that [q′i] = M K 3(k0 +1)![Φ1(pi)]. Set e′i = p′i−q′i, and let P1 = Ψ(P)∪Φ1(P)∪{p′i, q′i, e′i; i = 1, ..., l}. Denote by G1 the group generated by P1. Recall that G0 = G(P) ∩ K0(A), and 206 decompose it as G00 ⊕ G01, where G00 is the infinitesimal part of G0. Fix this decomposition and denote by {d1, ..., dt} the positive elements which generate G01. Applying Corollary 14.11 to Mr(B) with any finite subset G, any ǫ > 0 and any 0 < r0 < δ < 1, one has a G-ǫ-multiplicative map L : Mr(B) → Mr(B) with the following properties: (1) [L]P1 and [L]G1 are well defined; (2) [L] induces the identity maps on the infinitesimal part of G1∩K0(B), G1 ∩ K1(B), G1 ∩ K0(B, Z/kZ) and G1 ∩ K1(B, Z/kZ) for the k with G1 ∩ Ki(B, Z/kZ) 6= {0}, i = 0, 1; (3) τ ◦ [L](g) ≤ r0τ (g) for all g ∈ G1 ∩ K0(B) and τ ∈ T (B); (4) There exist positive elements {fi} ⊂ K0(B)+ such that for i = 1, ..., t, α(di) − [L](α(di)) = M K 3(k0 + 1)!fi. Using the compactness of T (B) and the strict comparison for positive elements for B, the positive number r0 can be chosen sufficiently small such that τ ◦ [L] ◦ [Ψ]([pi]) < δ/2 for all τ ∈ T (B), and α([pi]) − [L ◦ Ψ]([pi]) > 0, i = 1, 2, ..., l. j=1 mjdj +sj, where mj ∈ Z and sj ∈ G00. Note, by (2) above, (α−[L]◦α)(si) = 0. Then we have Let [pi] =Pt α([pi]) − [L ◦ Ψ]([pi]) = α([pi]) − ([L ◦ α]([pi]) + M K 3(k0 + 1)![L ◦ Φ1]([pi])) = (α(P mjdj) − [L ◦ α](P mjdj)) − M K 3(k0 + 1)![L ◦ Φ1]([pi]) = M K 3(k0 + 1)!(P mjfj − [L] ◦ [Φ1]([pi])) = M K 3(k0 + 1)!f′i, where f′i =P mjfj−[L]◦[Φ1]([pi]), for i = 1, 2, ..., l. Define β([pi]) = K 3(k0 +1)!f′i , i = 1, 2, ..., l. Let us now construct a map h′ : A → B. It will be constructed by factoring through the K0-group of some C∗-algebra in the class C0 in the construction given at the beginning of this section. Let zi′ = β([pi]), and z′i = ρB( zi′) ∈ Aff(S[1](K0(B))). Then we have: M z′ − z∞ = maxi{ρB(α([pi]) − [L ◦ Ψ]([pi])) − ρ(α([pi]))} = maxi{supτ∈T (B){τ ◦ [L] ◦ [Ψ]([pi])} ≤ δ/2, where z = (z1, z2, ..., zr) and z′ = (z′1, z′2, ..., z′r). By Lemma 20.10, for sufficiently large n, one obtains u = (u1, u2, ..., us(n)) ∈ K0(B)s(n) such that + (e 20.31) (e 20.32) X xijuj = z′i. More importantly, It follows from 20.6 that the maps ¯Rn u > 0. ek+j i 7→ (us(j−1)+2i−1 − us(j−1)+2i), 1 ≤ j ≤ n, 1 ≤ i ≤ lk+j defines strictly positive homomorphism, κ(k+j) from K0(Sk+j) to K0(B) which defines a strictly positive homomorphism from K0(D) to K0(B), where D = Sk+1 ⊕ ··· ⊕ Sk+n. Since B ∈ Bu0, by Corollary 18.9, there is a homomorphism h′ : D → Mm(B) for some large m such that ∗0Sk = κ(k) h′ 0 . By (e 20.31), one has, keeping the notation in the construction at the beginning of this section, 0 ∗0([ψk h′ k+1(pi)], [ψk k+2(pi)], ..., [ψk n+k(pi)]) = β([pi]), i = 1, ..., r. 207 Now, define h′′ : A → D →Mm(B) by h′′ = h′ ◦ (ψk k+1 ⊕ ψk k+2⊕··· ⊕ ψk k+n). Then h′′ is F-δ-multiplicative. For any x ∈ ker ρ, by Lemma 20.9, x ∈ ker ρB ◦ α ∩ ker[H] and x ∈ ker[h0] = ker[Φ1]. Therefore, we have [Φ1](x) = 0 and [Ψ](x) = α(x). Note that α(x) also vanishes under any state of (K0(B), K0 +(B)), we have [L] ◦ α(x) = α(x). So, we get α(x) − [L ◦ Ψ](x) = 0. Therefore α − [L ◦ Ψ]kerρ = 0. Therefore we may view α − [L ◦ Ψ] as a homomorphism from ρ(G0). Since M g is in the subgroup generated by ρ([p1]), ..., ρ([pr]) for any g ∈ ρ(G0). Recall that (α − [L ◦ Ψ])([pi]) = M β([pi]), i = 1, 2, ..., r. Set h to be M copies of h′′. The map h is F-δ-multiplicative, and i = 1, ..., r. [h]([pi]) = α([pi]) − [L] ◦ [Ψ]([pi]) (e 20.33) Note that [h] also has the multiplicity M K 3(k0 + 1)!, and D ∈ C0 with trivial K1 groups. One can concludes that h induces zero map on G ∩ K1(A), G ∩ K1(A, Z/kZ) and G ∩ K1(A, Z/kZ) for k ≤ k0. Therefore, we have Set L1 = (L ◦ Ψ) ⊕ h. It is F-δ multiplicative and [h]P = αP − [L] ◦ [Ψ]P . [L1]P = [h]P + [L] ◦ [Ψ]P = αP . We may assume L1(1A) = 1B by taking a conjugation with a partial isometry. Then L1 is an F-δ-multiplicative map from A to B, and [L1]P = αP . Corollary 20.15. Let A be a amenable C-algebra in the class B1, and assume that A satisfies the K0-density property and the UCT. Then A is KK-attainable with respect to Bu0. Proof. Let C be any C-algebra in Bu0, and let α ∈ KL++(A, C). We may write that C = C1⊗U for some C1 ∈ B0 and for some UHF-algebra of infinite type. By Theorem 14.8, there is a C- algebra B which is an inductive limit of C-algebras in the class C0 together with homogeneous C∗-algebras in the class H such that (K0(A), K0(A)+, [1A]0, K1(A)) ∼= (K0(B), K0(B)+, [1B]0, K1(B)). Since A satisfies the UCT, there is an invertible β ∈ KL++(A, B) such that β carries the isomorphism of K-theories of A and B. Applying Proposition 20.14 to β and applying Corollary 18.8 to α ◦ β−1, one has the desired conclusion. Theorem 20.16. Let A ∈ B1 be a amenable C∗-algebra satisfying the K0-density property and the UCT, and let B ∈ Bu0. Then for any α ∈ KL++(A, B), and any γ : T (B) → T (A) which is compatible to α, there is a sequence of completely positive linear maps Ln : A → B such that lim n→∞Ln(ab) − Ln(a)Ln(b) = 0 for all a, b ∈ A [Ln] = α and lim n→∞ sup τ∈T (B)τ ◦ Ln(f ) − γ(τ )(f ) = 0 for all f ∈ A. Proof. It follows from Corollary 20.15 and Proposition 18.12 directly. 208 21 The isomorphism theorem z { s } Definition 21.1. Put C′ = P Mn(C(X))P , where T ⊔ ··· ⊔ T⊔Y for some connected finite CW-complex Y with torsion K1-group (no restriction on K0(C(Y ))) and dimension no more than 3, and P is a projection in Mn(C(X)) with rank r ≥ 6. Then K1(P Mn(C(X))P ) = Tor(K1(C′)) ⊕ G1 for some torsion free group G1 ∼= Zs. Let D′ = Ls i=1 Mr(C(T)). Let Π′i : P Mn(C(X))P → Ei (Ei = Mr(C(T))) be defined by Π′i(f )(x) = fTi, where Ti is the i-th circle, for all f ∈ C′, i = 1, 2, ..., s. Define Π′ : P Mn(C(X))P → D′ by Π′(f ) = (Π′1(f ), Π′2(f ), ..., Π′s(f )) for all f ∈ C′. We have that K1(D′) ∼= G1. Denote by C a finite direct sum of the C-algebras of the form C′ above, matrix algebras and C∗-algebras in C0 (with trivial K1-group). Denote by D the direct sum of D′ corresponding to the C-algebras in the form C′. In other words, C = D ⊕ C0, where C0 is a direct sum of C∗-algebras in C0 and those with the form P Mn(C(Y ))P, where Y is a connected finite CW complex with K1(C(Y )) is a finite abelian group. Then, one has that U(C)/CU(C) ∼= U0(C)/CU(C) ⊕ K1(D) ⊕ Tor(K1(C)). Here we identify K1(D)⊕Tor(K1(C)) with a subgroup of U (C)/CU (C). Denote by π0, π1, π2 the projection maps from U(C)/CU(C) to each component according to the decomposition above. Define P ′ : C → P Mn(C(X))P to be the projection and Π = Π′ ◦ P ′. We will frequently refer to the above notation later in this section. As in [63], we have the following lemmas to control the maps from U (C)/CU (C) in the approximate intertwining argument in the proof of 21.9. The proofs are the repetition of the corresponding arguments in [63]. Lemma 21.2 (See Lemma 7.2 of [63]). Let C be as above, let U ⊂ U(C) be a finite subset, and let F be the group generated by U . Suppose that G is a subgroup of U(C)/CU(C) which contains F , the image of F in U (C)/CU (C), also contains π1(U(C)/CU(C)), and π2(U(C)/CU(C)). Suppose that the composition map γ : F → U(D)/CU(D) → U(D)/U0(D) is injective—that is, if x, y ∈ F and x 6= y, then [x] 6= [y] in U (D)/U0(D). Let B be a unital C-algebra and Λ : G → U(B)/CU(B) be a homomorphism such that Λ(G∩ (U0(C)/CU(C))) ⊂ U0(B)/CU(B). Let θ : π2(U(C)/CU(C)) → U(B)/CU(B) be defined by θ(g) = Λπ2(U(C)/CU(C))(g−1) for any g ∈ π2(U(C)/CU(C)). Then there is a homomorphism β : U(D)/CU(D) → U(B)/CU(B) with β(U0(D)/CU(D)) ⊂ U0(B)/CU(B), and such that β ◦ Π‡ ◦ π1( ¯w) = Λ( ¯w)(θ ◦ π2( ¯w)) for all w ∈ F. If furthermore B ∼= B1 ⊗ U for a unital C-algebra B1 ∈ B0 and a UHF-algebra U , and Λ(G) ⊂ U0(B)/CU(B), then β ◦ Π‡ ◦ (π1) ¯F = Λ ¯F . The above may be summarized by the following commutative diagram: inclusion / G Λ+θ◦π2 U (B)/CU (B) ¯F π1 π1( ¯F ) Π‡ 6♠♠♠♠♠♠♠♠♠♠♠♠♠ β U (D)/CU (D) 209 / 6 Proof. The proof is exactly the same as that of Lemma 7.2 of [63]. Let κ1 : U (D)/CU (D) → K1(D) ⊂ U (C)/CU (C) be the quotient map. Let η : π1(U (C)/CU (C)) → K1(D) be the map defined by η = κ1 ◦ Π‡π1(U (C)/CU (C)). Note that η is an isomorphism—- as we regard both π1(U (C)/CU (C)) and K1(D) as the same subgroup of U (C)/CU (C), this η is the identity map. Since γ is injective and γ(F ) is free, we conclude that κ1 ◦ Π‡ ◦ π1 is also injective on F . Since U0(C)/CU (C) is divisible (6.6 of [63]), there is a homomorphism λ : K1(D) → U0(C)/CU (C) such that λκ1◦Π‡◦π1(F ) = π0 ◦ ((κ1 ◦ Π‡ ◦ π1)F )−1, where ((κ1◦Π‡◦π1)F )−1 : η◦π1(F ) → F is the inverse map of the injective map (κ1◦Π‡◦π1)F . This could be viewed as the following commutative diagram: K1(D) 8qqqqqqqqqq λ '❖❖❖❖❖❖❖❖❖❖❖ (η ◦ π1)( ¯F ) π0◦(η◦π1)−1 U0(C)/CU (C). Define Then, for any w ∈ F , β = Λ((η−1 ◦ κ1) ⊕ (λ ◦ κ1)). β(Π‡ ◦ π1(w)) = Λ(η−1(κ1 ◦ Π‡(π1(w))) ⊕ λ ◦ κ1(Π‡(π1(w)))) = Λ(π1(w) ⊕ π0(w)). Define θ : π2(U (C)/CU (C)) → U (B)/CU (B) by Then θ(x) = Λ(x−1) for all x ∈ π2(U (C)/CU (C)). β(Π‡(π1(w))) = Λ(w)θ(π2(w)) for all w ∈ F. For the second part of the statement, one assume that Λ(G) ⊆ U0(B)/CU (B). Then Λ(π2(U (C)/CU (C))) is a torsion subgroup of U0(B)/CU (B). But U0(B)/CU(B) is torsion free by Lemma 11.5, and hence θ = 0. Lemma 21.3 (See Lemma 7.3 of [63]). Let B ∈ B1 be a separable simple C-algebra, and let C be as above. Let U ⊂ U(B) be a finite subset, and let F be the subgroup generated by U such that κ1( ¯F ) is free, where κ1 : U(B)/CU(B) → K1(B) is the quotient map. Suppose that α : K1(C) → K1(B) is an injective homomorphism and L : ¯F → U(C)/CU(C) is an injective homomorphism with L( ¯F ∩ U0(B)/CU(B)) ⊂ U0(C)/CU(C) such that π1 ◦ L is one-to-one and α ◦ κ′1 ◦ L(g) = κ1(g) for all g ∈ ¯F , where κ′1 : U(C)/CU(C) → K1(C) is the quotient map. Then there exists a homomorphism β : U(C)/CU(C) → U(B)/CU(B) with β(U0(C)/CU(C)) ⊂ U0(B)/CU(B) such that β ◦ L(f ) = f for all f ∈ ¯F . Proof. The proof is exactly the same as that of Lemma 7.3 of [63]. Let G be the preimage of α◦κ′1(U (C)/CU (C)) under κ1. So we have the short exact sequence 0 → U0(B)/CU (B) → G → α ◦ κ′1(U (C)/CU (C)) → 0. 210 ' / / 8 Since U0(B)/CU (B) is divisible, there is an injective homomorphism γ : α ◦ κ′1(U (C)/CU (C)) → G such that κ1 ◦ γ(g) = g for any g ∈ α ◦ κ′1(U (C)/CU (C)). Since α ◦ κ′1 ◦ L(f ) = κ1(f ) for any f ∈ F , we have F ⊆ G. Moreover, note that (γ ◦ α ◦ κ′1 ◦ L(f ))−1f ∈ U0(B)/CU (B) for all f ∈ F . Define ψ : L(F ) → U0(B)/CU (B) by ψ(x) = γ ◦ α ◦ κ′1(x−1)L−1(x) for x ∈ L(F ). Since U0(B)/CU (B) is divisible, there is a homomorphism ψ : U (C)/CU (C) → U0(B)/CU (B) such that ψL(F ) = ψ. Now define β(x) = γ ◦ α ◦ κ′1(x) ψ(x). Hence β(L(f )) = f for f ∈ F . Lemma 21.4. Let A be a unital separable C∗-algebra such that {ρA([p]) : p ∈ A projections} is dense in real linear span of {RρA([p]) : p ∈ A projections}. Then, for any finite dimensional C∗-subalgebra B ⊂ A, U (B) ⊂ CU (A). Proof. Let u ∈ U (B). Since B has finite dimensional, u = exp(ih) for some h ∈ Bs.a.. We may write h =Pn i=1 λipi, where λi ∈ R and {p1, p2, ..., pn} is a set of mutually orthogonal projections. By the assumption and applying Proposition 3.6 of [38], one has u ∈ CU (A). Lemma 21.5 (See Lemma 7.4 of [63]). Let B ∼= A ⊗ U , where A ∈ B0 and U is an infinite dimensional UHF-algebra. Let C = C1 ⊕ C2, where C1 = P Mn(C(X))P with X be as defined in 21.1, C2 ∈ C0. Let F be a group generated by a finite subset U ⊂ U(C) such that (π1) ¯F is one- to-one. Let G be a subgroup containing ¯F , π1(U(C)/CU(C)) and π2(U(C)/CU(C)). Suppose that α : U(C)/CU(C) → U(B)/CU(B) is a homomorphism such that α(U0(C)/CU(C)) ⊂ U0(B)/CU(B). Then, for any ǫ > 0, there are δ > 0 and finite subset G ⊆ C satisfying the following: if ϕ = ϕ0 ⊕ ϕ1 : C → B (by such decomposition, we mean there is a projection e0 such that ϕ0 : C → e0Be0 and ϕ1 : C → (1B − e0)B(1B − e0)) is a G-δ-multiplicative completely positive linear contraction such that (1) ϕ0 maps identity of each summand of C to a projection, (2) G is sufficiently large and δ is sufficiently small depending only on F and C (such that ϕ‡ is well defined on a subgroup of U(C)/CU(C) containing all of ¯F , π0( ¯F ), π1(U(C)/CU(C)), and π2(U(C)/CU(C))), (3) ϕ0 is homotopic to a homomorphism with finite dimensional image, [ϕ0]K0(C) is well defined and [ϕ]K1(C) = α∗, where α∗ : K1(C) → K1(B) is induced map, (4) τ (ϕ0(1C )) < δ for all τ ∈ T (B) (assume e0 = ϕ0(1C )), then there is a homomorphism Φ : C → e0Be0 such that [ϕ0]K0(C) and (i) ΦC1 is homotopic to a homomorphism with finite dimensional image and (Φ)∗0 = (ii) α( ¯w)−1(Φ ⊕ ϕ1)‡( ¯w) = ¯gw where gw ∈ U0(B) and cel(gw) < ǫ for any w ∈ U . 211 Proof. The argument is exactly the same as that of Lemma 7.4 of [63] (see also [27] and [80]) . Since the source algebra and target algebra in this lemma are different from the ones in 7.4 of [63], we will repeat some of the argument here. We will retain the notations in 21.1. By 21.2, there are homomorphisms β1, β2 : U (D)/CU (D) → U (B)/CU (B) with βi(U0(D)/CU (D)) ⊂ U0(B)/CU (B) (i = 1, 2) and homomorphisms such that θ1, θ2 : π2(U (C)/CU (C)) → U (B)/CU (B) β1 ◦ Π‡(π1( ¯w)) = α( ¯w)θ1(π2( ¯w)) and β2 ◦ Π‡(π1( ¯w)) = ϕ‡( ¯w∗)θ2(π2( ¯w)) (e 21.1) for all ¯w ∈ ¯F . Moreover, θ1(g) = α(g−1) and θ2(g) = ϕ‡(g) for all g ∈ π2( ¯F ). Since ϕ0 is homotopic to a homomorphism with finite dimensional range, using condition (3), we compute that θ1(g)θ2(g) ∈ U0(B)/CU (B) for all g ∈ π2( ¯F ). (e 21.2) Since π2(U (C)/CU (C)) is torsion free and U0(B)/CU (B) is torsion free (see 11.5), we conclude that θ1(g)θ2(g) = 1 for all g ∈ π2( ¯F ). (e 21.3) To simplify notation, without loss of generality, we may write C = D ⊕ C0 as in 21.1. To keep the same notation as in the proof of 7.4 of [63], we also write D = C (1) and C0 = ⊕l1 j=2C (j), where each C (j) is in C0 and is minimal (cannot be written as finite direct sum of more than one copies of C∗-algebras in C0), or a C∗-algebra of the form P Mn(C(Y ))P with connected spectrum Y. Note K1(C (j)) = {0} for all j ≥ 2. We then proceed the proof of 7.4 of [63] and construct Φ1 exactly the same way as in the proof of 7.4 of [63]. We will keep the notation used in the proof of 7.4 of [63]. We will again use the fact, by Corollary 11.7, the group U0(B)/CU(B) is torsion free, and by Theorem 11.10, the map U (eBe)/CU (eBe) → U (B)/CU (B) is an isomorphism, where e ∈ B is a nonzero projection. By the assumption, ϕ0C0 is homotopy to ϕ00 : C0 → ϕ0(1C )Bϕ0(1C ) such that B0 = ϕ00(C0) is a finite dimensional C∗-subalgebra in (e0 − E1)B(e0 − E1) with 1B0 = e0 − E1, where E1 ≤ e0 is a non-zero projection such that [E1] = (ϕ0)∗0([1D]). Let Φ2 : C → B0 be defined by Φ2(f, g) = ϕ00(g) for all f ∈ D and g ∈ C0. It is important to note (as the main difference from this lemma and that of 7.4 of [63]) that, with the assumption, by 21.4, for any w ∈ U (C), Φ2(w) ∈ CU (B). In other words, the general case could be reduced to the case that C = D. The rest of the proof is exactly the same as that of 7.4 of [63]. Lemma 21.6 (See Lemma 7.5 of [63]). Let B ∼= A ⊗ U , where A ∈ B1 and U an infinite dimensional UHF-algebra. Let U ⊂ U(B) be a finite subset and F be the subgroup generated by U such that κ1( ¯F ) is free, where κ1 : U(B)/CU(B) → K1(B) is the quotient map. Let C be as above and let ϕ : C → B be a homomorphism such that (ϕ)∗1 is one-to-one. Suppose that j, L : ¯F → U(C)/CU(C) are two injective homomorphisms with j(F ∩ U0(B)), L(F ∩ U0(B)) ⊂ U0(C)/CU(C) such that κ1 ◦ ϕ‡ ◦ L = κ1 ◦ ϕ‡ ◦ j = κ1 ¯F , and they are one-to-one. Then, for any ǫ > 0, there exists δ > 0 such that if ϕ can be decomposed as ϕ = ϕ0 ⊕ ϕ1 : C → B, where ϕ0 and ϕ1 are homomorphisms satisfying the following: 212 (1) τ (ϕ0(1C )) < δ for all τ ∈ T (B) and (2) ϕ0 is homotopic to a homomorphism with finite dimensional image, then there is a homomorphism ψ : C → e0Be0 (e0 = ϕ0(1C )) such that (4) [ψ] = [ϕ0] in KL(C, B) and (5) (ϕ‡ ◦ j( ¯w))−1(ψ ⊕ ϕ1)‡(L( ¯w)) = ¯gw where gw ∈ U0(B) and cel(gw) < ǫ for any w ∈ U . Proof. The proof is the same as that of Lemma 7.5 of [63]. Note that instead of using Lemma 7.4 of [63], one uses Lemma 21.5. Remark 21.7. Roles of Lemma 21.5 and Lemma 21.6 played in the proof of the following isomorphism theorem (21.9) are the same as those in the proof of Theorem 10.4 in [63]. The following statement is well known. For the reader’s convenience, we include a proof. Lemma 21.8. Let (An, ϕn,n+1) be a unital inductive sequence of separable C-algebras, and denote by A = lim−→ An. Assume that A is amenable. Let F ⊆ A be a finite subset, and let ǫ > 0. Then there is an integer m ≥ 1 and a unital completely positive linear map Ψ : A → Am such that kϕm,∞ ◦ Ψ(f ) − fk < ǫ for all f ∈ F. sequences (x1, x2, ..., xn, ...) satisfying that there is N with xn+1 = ϕn(xn), n = N, N + 1, ... . Since A is amenable, by the Choi-Effros lifting theorem, there is a unital completely positive Proof. Regard A as the C-subalgebra ofQ An/L An generated by the equivalence classes of the linear map Φ : A →Q An such that π ◦ Φ = idA, where π is the quotient map. In particular, this implies that (e 21.4) lim k→∞kπk ◦ Φ(a) − akk = 0, if a = (a1, a2, ..., ak, ...) ∈ A. Write F = {f1, f2, ..., fl}, and for each fi, fix a representative In particular Then, for each fi, one has fi = (fi,1, fi,2, ..., fi,k, ...). lim k→∞ ϕk,∞(fi,k) = fi. (e 21.5) lim sup k→∞ kϕk,∞ ◦ πk ◦ Φ(fi) − fik k→∞ kϕk,∞ ◦ πk ◦ Φ(fi) − ϕk,∞(fi,k)k k→∞ kπk ◦ Φ(fi) − fi,kk = 0 (by (e 21.4)). = lim sup ≤ lim sup (by (e 21.5)) There then exists m ∈ N such that Thus, the unital completely positive linear map kϕm,∞ ◦ πm ◦ Φ(fi) − fik < ǫ, 1 ≤ i ≤ l. satisfies the lemma. Ψ := πm ◦ Φ 213 Theorem 21.9. Let A1 ∈ B0 be a unital separable simple C-algebra satisfying the UCT, and denote by A = A1 ⊗ U for a UHF-algebra U of infinite type. Let C be a C-algebra in Theorem 14.8. If Ell(C) ∼= Ell(A), then there is an isomorphism ϕ : C → A which carries the identification of Ell(C) ∼= Ell(A). Moreover, if there is a homomorphism Γ : Ell(C) → Ell(A), then there is a -homomorphism ϕ : C → A such that ϕ induces Γ. Proof. We only prove the first part of the statement. The second part can be proved in a similar way (one only has to do one-sided intertwining arguments in this case). Ell(C) ∼= Ell(A). We will repeatedly apply Theorem 12.9. Let δ(1) c,1 , σ(1) Let α ∈ KL(C, A) with α−1 ∈ KL(A, C) and γ : T (A) → T (C) be given by the isomorphism Assume that C = lim−→(Cn, ιn) be as in 14.8. Let G1 ⊆ G2 ⊆ ··· ⊆ C and F1 ⊆ F2 ⊆ ··· ⊆ A be increasing sequences of finite subsets with dense union. Let ε1 > ε2 > ··· > 0 be a decreasing sequence of positive numbers with finite sum. c ⊆ C (in place of G), σ(1) c ⊆ U (C) (in place of U ) and H(1) c ⊆ Cs.a (in place of H2) be as required by 12.9 for C (in the place of A with X being a point), ε1 (in the place of ε), G1 (in the place of F). Note, in this case, X is a point, by Remark 12.10, we do not introduce the map ∆ and H1. the image of U (Cn) for some large n ≥ 1. Moreover, as 12.10, we let U (1) U (M2(C))). As in the remark of 12.10, with sufficiently large n ≥ 1, we may assume that U (1) c,2 > 0 (in place of σ1 and σ2 respectively), P (1) is in c ⊂ U (C) (instead in c ⊆ K(C) (in place of P), U (1) c > 0 (in place of δ), G(1) c Denote by Fc ⊆ U (C) the subgroup generated by U (1) . Write Fc = (Fc)0⊕ Tor(Fc) according to the decomposition described in 21.1, where (Fc)0 is torsion free. Without loss of generality (by choosing smaller σ(1) c,2 ), one may assume that c c = U (1) U (1) c,0 ⊔ U (1) c,1 where U (1) that U (1) c,0 · U (1) that for each u ∈ U (1) c,0 generates (Fc)0 and U (1) c ⊂ U (1) c,1 generates Tor(Fc)—namely, we can choose U (1) by U (1) c,1 ; then by choosing smaller σ(1) c,2 one can replace U (1) c,1 , one has that uk ∈ CU (C), where k is the order of u. c c,0 and U (1) c,0 ⊔ U (1) c,1 so c,1 . Note By Theorem 20.16, there is a G(1) c -δ(1) c -multiplicative map L1 : C → A such that c and = αP (1) [L1]P (1) τ ◦ L1(f ) − γ(τ )(f ) < σ(1) c c,1 /3 for all f ∈ H(1) c for all τ ∈ T (A). (e 21.6) (e 21.7) Without loss of generality, one may assume that L‡1 is well defined and injective on (Fc)0. Moreover, if k is the order of u, one may also assume that dist(L1(uk), CU (A)) < σ(1) c,2 /2 for all u ∈ U (1) c,2 . a,1, σ(1) To apply Theorem 12.9 second time, let δ(1) a,2 > 0 (in the place of σ1 and σ2 respectively), P (1) a ⊆ C (in the place of G), σ(1) a ⊆ K(A) (in the place of P), U (1) a ⊆ U (A) (in the place of U ) and H(1) a ⊆ As.a (in the place of H2) be as required by Theorem 12.9 for A (in the place of A with X being a point), ε1 (in the place of ε), and F1 (in the place of F). Again, note that X is a point, in the application of Theorem 12.9. a > 0 (in the place of δ), G(1) 214 is finite, we can write Fa = (Fa)0 ⊕ Tor(Fa), where (Fa)0 is torsion free. Fix this decomposition. Without loss of generality (by choosing smaller σ(1) Denote by Fa ⊆ U (A) the subgroup generated by U (1) a,2), one may assume that a . Since U (1) a a = U (1) U (1) a,0 ⊔ U (1) a,1 generates Tor(Fa). a,1 a,0 generates (Fa)0 and U (1) where U (1) a , we can assume P (1) a ⊃ κ1,A((Fa)0)(in K1(A)), where κ1,A : U (A)/CU (A) → K1(A) is the quotient map. Note that for each u ∈ U (1) δ′ < δ(1) such that, with σ(1) a,1 , one has that uk ∈ CU (A), where k is the order of u. By Theorem 20.16 and amenability of C, there are a finite subset G′ ⊃ G(1) a , a positive number a , a sufficiently large integer n ≥ 1 and there is a G′-δ′-multiplicative map Φ′1 : A → Cn By enlarging P (1) (e 21.8) a,c,1 = min{σ(1) a,1, σ(1) c,1}/3, = α−1P (1) [ın ◦ Φ′1] τ ◦ ın ◦ Φ′1(f ) − γ−1(τ )(f ) < σ(1) a ∪[L1](P (1) P (1) ) c and c ) a ∪[L1](P (1) a,c,1 for all f ∈ H(1) c ) and τ ∈ T (C).(e 21.9) Moreover, one may assume that Φ′1 ◦ L1 is G(1) -multiplicative, (Φ′1)‡ is defined and injective on (Fa)0, and (Φ′1◦L1)‡ is well defined and injective on (Fc)0. Furthermore, one may also assume that (e 21.10) a ∪ L1(H(1) -δ(1) dist(ın ◦ Φ′1(uk), CU (C)) < σ(1) a,2/2 for all u ∈ U (1) a,1 , c c where k is the order of u, and dist((ın ◦ Φ′1 ◦ L1)(vk′ ), CU (A)) < σ(1) c,2 for all v ∈ U (1) c,1 , where k′ is the order of v. It then follows from (e 21.6) and (e 21.8) that and it follows from (e 21.7) and (e 21.9) that [ın ◦ Φ′1 ◦ L1]P (1) c = [id]P (1) c ; (e 21.11) (e 21.12) τ ◦ ın ◦ Φ′1 ◦ L1(f ) − τ (f ) < 2σ(1) c,1 /3 for all f ∈ H(1) c for all τ ∈ T (C). (e 21.13) Recall that (Fc)0 ⊆ U (C)/CU (C) is the subgroup generated by U (1) c,0 . Since we have assumed c,0 is in the image of U (Cn)/CU (Cn), there is an injective homomorphism j : (Fc)0 → (e 21.14) that U (1) U (Cn)/CU (Cn) such that . ι‡n ◦ j = id(Fc)0 Moreover, by (e 21.12), κ1,C ◦ ι‡n ◦ (Φ′1 ◦ L1)‡(Fc)0 U (C)/CU (C) → K1(C) be the quotient map. σc,2 (in place of ε), ιn (in place of ϕ), j and (Φ′1 ◦ L1)‡(Fc)0 of C, one has a decomposition ιn = ι(0) n such that n ⊕ ι(1) Let δ be the constant of Lemma 21.6 with respect to Cn (in place of C), C (in place of B), (in place of L). By the construction = κ1,A ◦ ι‡n ◦ j = κ1,C(Fc)0 , where κ1,C : (1) τ (ι(0) c,1 /3} for all τ ∈ T (C), and n (1Cn )) < min{δ, σ(1) n has finite dimensional range. Then, by Lemma 21.6, there is a homomorphism h : Cn → e0Ce0, where e0 = ι(0) n (1Cn), such that (2) ι(0) 215 n ] in KL(Cn, C), and (3) [h] = [ι(0) (4) for each u ∈ U (1) c,0 , one has that (ι‡n ◦ j(u))−1(h ⊕ ι(1) n )‡((Φ′1 ◦ L1)‡(u)) = gu (e 21.15) for some gu ∈ U0(C) with cel(gu) < σ(1) c,2 . Define Φ1 = (h ⊕ ι(1) n ) ◦ Φ′1. By (e 21.8) and (3), one has [Φ1] a ∪[L1](P (1) P (1) c ) = α−1P (1) a ∪[L1](P (1) c ) (e 21.16) Note that Φ1 is still G(1) Φ′1 replaced by Φ1. That is a -δ(1) a -multiplicative, and hence (e 21.10) and (e 21.11) still hold with dist(Φ1(uk), CU (C)) < σ(1) a,2/2 for all u ∈ U (1) a,1 , where k is the order of u, and dist((Φ1 ◦ L1)(vk′ ), CU (A)) < σ(1) c,2 for all v ∈ U (1) c,1 , where k′ is the order of v. By (e 21.12) and (e 21.13), one has τ ◦ Φ1 ◦ L1(f ) − τ (f ) < σ(1) [Φ1 ◦ L1]P (1) c,1 for all f ∈ H(1) and = [id]P (1) for all τ ∈ T (C). c c c Moreover, for any u ∈ U (1) c,0 , one has (by (e 21.14) and (e 21.15)) (Φ1 ◦ L1)‡(u) = (ι‡n ◦ j(u)) · gu = u · gu ≈σ(1) c,2 u. (e 21.17) (e 21.18) (e 21.19) (e 21.20) (e 21.21) Let u ∈ U (1) c,1 with order k. By (e 21.18), there is a self-adjoint element b ∈ C with b < σ(1) c,2 such that (where, we notice that (u∗)k ∈ CU (C)) and hence (u∗)k(Φ1 ◦ L1)(uk) exp(2πib) ∈ CU (C), ((u∗)(Φ1 ◦ L1)(u) exp(2πib/k))k ∈ CU (C). Note that (u∗)(Φ1 ◦ L1)(u) exp(2πib/k) ∈ U0(C) and U0(C)/CU (C) is torsion free (Corollary 11.7). One has that (u∗)(Φ1 ◦ L1)(u) exp(2πib/k) ∈ CU (C). In particular, this implies that dist((Φ1 ◦ L1)‡(¯u), ¯u) < σ(1) c,2 /k for all ¯u ∈ U (1) c,1 (e 21.22) Together with (e 21.21), one has that dist((Φ1 ◦ L1)‡(¯u), ¯u) < σ(1) c,2 for all u ∈ U (1) c . (e 21.23) 216 Therefore, by (e 21.19), (e 21.20), and (e 21.23), applying Theorem 12.9, one obtains a unitary U1 such that U∗(Φ1 ◦ L1(f ))U − f < ε1 for all f ∈ G1. By replacing Φ1 by Ad(U ) ◦ Φ1, without loss of generality, one may assume that In other words, one has the following diagram: Φ1 ◦ L1(f ) − f < ε1 for all f ∈ G1. C id / C ?⑦⑦⑦⑦⑦⑦⑦⑦ Φ1 L1 A c,1 , σ(2) We will continue to apply Theorem 12.9. Let δ(2) c,2 > 0 (in the place of σ1 and σ2 respectively), P (2) which is approximately commutative on the subset G1 within ε1. c ⊆ C (in the place of G), σ(2) c ⊆ K(C) (in the place of P), U (2) c ⊆ Cs.a (in the place of H2) be as required by Theorem 12.9 for C (in the place of A with X being a point), ε2 (in the place of ε), and G2 (in the place of F). . Since U (2) is finite, we can write )0 is torsion free. Fix this decomposition. Without loss c ⊆ U (C) the subgroup generated by U (2) c ⊆ U (C) (in the place of U ) and H(2) c > 0 (in the place of δ), G(2) F (2) ), where (F (2) c = (F (2) c of generality (by choosing smaller σ(2) c,2 ), one may assume that )0 ⊕ Tor(F (2) Denote by F (2) c c c c = U (2) U (2) c,0 ⊔ U (2) c,1 generates Tor(F (2) c,1 c c )0 and U (2) c,0 generates (F (2) where U (2) that uk ∈ CU (C), where k is the order of u. G0-δ0-multiplicative contractive completely positive linear maps L′′1, L′′2 : C → A, if kL′′1(c) − L′′2(c)k < δ0 for all c ∈ G0, then There are a finite subset G0 ⊂ C and a positive number δ0 > 0 such that, for any two ). Note that, for each u ∈ U (2) c,1 , one has c c and = [L′′2]P (2) [L′′1]P (2) τ ◦ L′′1(h) − τ ◦ L′′2(h) < min{σ(2) a ) for all τ ∈ T (A). Note, by Lemma 21.8, for any finite subset G′′ ⊂ C and any δ′′ > 0, there exists a large m a,1/3}/2 for all h ∈ H(2) c ∪ Φ1(H(1) c,1 /3, σ(1) and a unital contractive completely positive linear map L0,2 : C → Cm such that kım ◦ L0,2(g) − gk < δ′′ for all g ∈ G′′. (e 21.24) Let κ1,Cm : U (Cm)/CU (Cm) → K1(Cm) and κ1,A : U (A)/CU (A) → K1(A) be the quotient maps, respectively. We may assume that, with sufficiently large G′′ and sufficiently small δ′′, (L0,2 ◦ Φ1)‡ is defined and injective on (Fa)0, and moreover, and by (e 21.16), for any g ∈ (Fa)0 (note that P (1) κ1,Cm ◦ (L0,2 ◦ Φ1)‡(g) = [L0,2 ◦ Φ1](κ1,A(g)), g ∈ (Fa)0, a ⊃ κ1,A((Fa)0)(in K1(A))), α ◦ [ım] ◦ [L0,2 ◦ Φ1](κ1,A(g)) = α ◦ [ım ◦ L0,2] ◦ [Φ1](κ1,A(g)) = α ◦ [Φ1](κ1,A(g)) = κ1,A(g). 217 / ? Hence, α ◦ [ιm] ◦ κ1,Cm ◦ (L0,2 ◦ Φ1)‡(g) = α ◦ [ιm] ◦ [L0,2 ◦ Φ1](κ1,A(g)) = κ1,A(g) for all g ∈ (Fa)0. It follows from Lemma 21.3 that there is a homomorphism β : U (Cm)/CU (Cm) → U (A)/CU (A) with β(U0(Cm)/CU (Cm)) ⊆ U0(A)/CU (A) such that β ◦ (L0,2 ◦ Φ1)‡(f ) = f for all f ∈ (Fa)0. By Theorem 20.16, there is a G′′-δ′′-multiplicative map L′2 : C → A such that (e 21.25) (e 21.26) a ) (e 21.27) c = αP (2) [L′2]P (2) τ ◦ L′2(f ) − γ(τ )(f ) < min{σ(2) and c c,1 /3, σ(1) a,1/3}/2 for all f ∈ H(2) c ∪ Φ1(H(1) and for all τ ∈ T (A). We may choose that G′′ ⊃ G0 ∪ G(2) c and δ′′ < min{δ0, δ(2) c }. Define L′′2 : C → A by L′′2(c) = L′2 ◦ ım ◦ L0,2. Then, by (e 21.24), since L′2 is contractive, kL′′2(c) − L′2(c)k < δ0 for all c ∈ G0. It follows that c = [L′2]P (2) [L′′2]P (2) τ ◦ L′′2(f ) − γ(τ )(f ) < min{σ(2) = αP (2) and c,1 /3, σ(1) c c a,1/3} for all f ∈ H(2) c ∪ Φ1(H(1) a ) and for all τ ∈ T (A). G(1) a -δ(1) By choosing G′′ sufficiently large and δ′′ sufficiently small, one may assume that L′′2 ◦ Φ1 is a -multiplicative, and dist((L′′2 ◦ Φ1)(uk), CU (A)) < σ(1) a,2 for all u ∈ U (1) a,1 , where k is the order of u (see (e 21.18)). Moreover, by the construction of C (see 14.8), one may assume that L′2 ◦ ιm = h0 ⊕ ι(1) m , where h0, ι(1) m : Cm → A are G′-η′-multiplicative for a sufficiently large G′ and suffi- ciently small η′ (only depends on Cm and (Φ1)‡((Fa)0)) so that (L2 ◦ ιm)‡ is defined on a sub- group of U (Cm)/CU (Cm) containing (Φ1)‡((Fa)0), π0((Φ1)‡((Fa)0)), π1(U (Cm)/CU (Cm)) and π2(U (Cm)/CU (Cm)). Moreover, since every finite dimensional C∗-algebra is semi-projective and since L′2 is chosen after Cm is chosen, we may assume that the map h0 is a homomorphism and has finite dimensional range, and τ (h0(1Cm )) < min{δ′, σa,1/3} for any τ ∈ T (A), where δ′ is the constant (in place of δ) of Lemma 21.5 with respect to σa,2 (in place of ε). Then, by Lemma 21.5, there is is a homomorphism ψ0 : Cm → e′0Ae′0, where e′0 = ψ0(1Cm ), (i) ψ0 is homotopically trivial, and [ψ0]0 = [h0]0, and (ii) for any u ∈ U (1) a,0 , one has such that β(Φ‡1(u))−1(ψ0 ⊕ ι(1) m )‡(Φ‡1(u)) = gu (e 21.28) for some gu ∈ U0(A) with cel(gu) < σa,2. 218 Define L2 = (ψ0 ⊕ ι1 one then has m) ◦ L0,2 : Cm → A. Then, for any u ∈ Ua, by (e 21.28) and (e 21.25), (L2 ◦ Φ1)‡(u) = β(Φ‡1(u)) · gu = u · gu ≈σa,2 u for all u ∈ U (1) a,0 . Moreover, it is also clear that τ ◦ L2 ◦ Φ1(f ) − τ (f ) < σ(1) a [L2 ◦ Φ1]P (1) = [id]P (1) a,1 for all f ∈ H(1) a a and for all τ ∈ T (A). (e 21.29) (e 21.30) (e 21.31) c c -δ(2) has order k), Note that L2 is still G(2) -multiplicative. One then has that for arbitrary u ∈ U (1) dist((L2 ◦ Φ1)(uk), CU (A)) < σ(1) a,2. Therefore, there is a self-adjoint element h ∈ A with h < σ(1) a,2 such that a,1 (with u and hence Note that (u∗)k(L2 ◦ Φ1)(uk) exp(2πih) ∈ CU (A), ((u∗)(L2 ◦ Φ1)(u) exp(2πih/k))k ∈ CU (A). (u∗)(L2 ◦ Φ1)(u) exp(2πih/k) ∈ U0(A) and U0(A)/CU (A) is torsion free (Corollary 11.7). On has that (u∗)(L2 ◦ Φ1)(u) exp(2πih/k) ∈ CU (A). In particular, this implies that Together with (e 21.29), one has that dist((L2 ◦ Φ1)‡(u), u) < σ(1) a,2. dist((L2 ◦ Φ1)‡(u), u) < σ(1) a,2 for all u ∈ U (1) a . (e 21.32) Then, applying Theorem 12.9 with (e 21.30), (e 21.31) and (e 21.32), one obtains a unitary W ∈ A such that W ∗(L2 ◦ Φ1(f ))W − f < ε1 for all f ∈ F1. Redefine L2 to be Ad(W ) ◦ L2, and one has L2 ◦ Φ1(f ) − f < ε1 for all f ∈ F1. That is, one has the following diagram C L2 C id / ?⑦⑦⑦⑦⑦⑦⑦⑦ Φ1 L1 A / A id with the upper triangle approximately commutes on G1 up to ε1 and the lower triangle approx- imately commutes on F1 up to ε1. Since δ(2) c have been chosen and embedded into the construction of L2, the construction can continue. and H(2) c,2 , P (2) c,1 , σ(2) , G(2) , σ(2) c c c 219 / ? / By repeating this argument, one obtains the following approximate intertwining diagram L1 A where ?⑧⑧⑧⑧⑧⑧⑧⑧ Φ1 / A id C id / C / C id / ?⑧⑧⑧⑧⑧⑧⑧⑧ Φ2 L2 / A id / C id / ?⑧⑧⑧⑧⑧⑧⑧⑧ Φ3 L3 / A id L4 / ··· =④④④④④④④④④ ··· , Φn ◦ Ln(g) − g < εn for all g ∈ Gn and Ln+1 ◦ Φn(f ) − f < εn for all f ∈ Fn. By the choices of Gn, Fn and the fact that P∞n=1 εn < ∞, the standard Elliott approximate intertwining argument applies, which shows that A ∼= B, as desired. This proves the first part of the proof. The proof of the second part is basically the same but simpler since we only need to have a one-sided approximate intertwining. In particular, we do not need to construct Φ1. Thus, once L1 is constructed, we can go to construct L2. We can first construct L′2 as above right after (e 21.25). It is important that we do not need to assume that L‡1 is injective on Fc0, since we will apply 21.5 but not 21.6. Theorem 21.10. Let A1, B1 ∈ B0 be two unital separable simple C∗-algebras which satisfying the UCT. Let A = A1 ⊗ U and B = B1 ⊗ U, where U is a UHF-algebra of infinite type. Suppose that Ell(A) = Ell(B). Then there exists an isomorphism ϕ : A → B which carries the identification of Ell(A) = Ell(B). Proof. By Theorem 14.8, there is a C-algebra C as in 14.8, such that Ell(C) ∼= Ell(A) ∼= Ell(B). By the first part of Proposition 21.9, one has that C ∼= A and C ∼= B. In particular, A ∼= B. Corollary 21.11. Let A and B be as in 21.10. Suppose that there is a homomorphism Γ from Ell(A) to Ell(B) such that Γ([1A]) = [1B]. Then there exists a unital homomorphism ϕ : A → B such that ϕ induces Γ. Proof. By Theorem 21.10, one may assume that A is a unital C-algebra described in 14.8. Then the Corollary follows from the second part of Theorem 21.9. 22 More existence theorems Lemma 22.1. Let X be a finite CW complex, C = P Mk(C(X))P and let A1 ∈ B0 be a unital simple C∗-algebra. Assume that A = A1 ⊗ U for an infinite dimensional UHF-algebra U. Let α ∈ KKe(C, A)++. Then there exists a unital monomorphism ϕ : C → A such that [ϕ] = α. Moreover we may write ϕ = ϕ′n⊕ϕ′′n, where ϕ′n : C → (1−pn)A(1−pn) is a unital monomorphism, ϕ′′n : C → pnApn is a unital homomorphism with [ϕ′′n] = [Φ] in KK(C, pnApn) for some point evaluation map Φ and lim n→∞ max{τ (1 − pn) : τ ∈ T (A)} = 0 for all τ ∈ T (A), where pn ∈ A is a sequence of projections. Proof. To simplify the matter, we may assume that X is connected. It is also easy to check that the general case can be reduced to the case that C = C(X). It is routine to prove that α ∈ KK(C, A)++ implies that α(kerρC) ⊂ kerρA. 220 / / ? / ? / ? / / / = which we will identify with an element in HomΛ(K(C), K(A)) by a result in ([17]). Since Ki(C) is finitely generated, i = 0, 1, KK(C, A) = KL(C, A). Let α ∈ KLe(C, A)++ Write A = limn→∞(A1⊗Mrn, ın,n+1), where rnrn+1, rn+1 = mnrn and ın,n+1(a) = a⊗1Mmn , n = 1, 2, .... Since K∗(C) is finitely generated and consequently, K(C) is finitely generated modulo Bockstein Λ operations, there is an element α1 ∈ KK(C, A1 ⊗ Mrn) such that α = α1 × [ın], where [ın] ∈ KK(A1⊗Mrn, A) is induced by the inclusion ın : A1⊗Mrn → A. By increasing n, we can assume that α1(kerρC) ⊂ kerρA1⊗Mrn and further that α1 ∈ KKe(C, A1 ⊗ Mrn)++. Replacing A1 by A1 ⊗ Mrn, without lose of generality we can assume that α = α1 × [ı], where α1 ∈ KKe(C, A1)++ and ı : A1 → A is the inclusion. It induces an element α1 ∈ KL(C ⊗ U, A⊗ U ). Let K0(U ) = D, a dense subgroup of Q. Note that Ki(C ⊗ U ) = Ki(C) ⊗ D, i = 0, 1, by the Kunneth formula. We verify that α1(K0(C ⊗ U )+ \ {0}) ⊂ K0(A ⊗ U )+ \ {0}. Consider x = Pm i=1 xi ⊗ di ∈ K0(C ⊗ U )+ \ {0} with xi ∈ K0(C) and di ∈ D, i = 1, 2, ..., m, There is a projection p ∈ Mr(C) for some r ≥ 1 such that [p] = x. Let t ∈ T (C), then t(xi)di > 0. mXi=1 (e 22.1) It should be noted that, since C = C(X) and X is connected, t(xi) ∈ Z and t(xi) = t′(xi) for all t, t′ ∈ T (C). Since α1([1C ]) = [1A1], τ ◦ α1(xi) = t(xi) for any τ ∈ T (A1) and t ∈ T (C). By (e 22.1), τ (α1(x)) = mXi=1 τ ◦ α1(xi)di = mXi=1 t(xi)di > 0 (e 22.2) for all τ ∈ T (A1). This shows that α1 is strictly positive. For any C∗-algebra A′, in this proof, we will use jA′ : A′ → A′ ⊗ U for the homomorphism jA′(a) = a ⊗ 1U for all a ∈ A′. Evidently, (e 22.3) α = α1 ◦ jC = jA1 ◦ α1. Write U = limn→∞(Mrn, ın), where rnrn+1, rn+1 = mnrn and ın(a) = a⊗ 1Mmn , n = 1, 2, .... We may assume that r1 = 1. The UHF algebra U corresponds to the super natural number Π∞i=1mi. Evidently, we can choose mi carefully so that we can write the super natural number Π∞i=1mi in another way Π∞i=1mi = Π∞i=1li with limi and limi→∞ Let {xn} be a sequence of points in X such that {xk, xk+1, ..., xn, ...} is dense in X for each k and each point in {xn} repeats infinitely many times. Let B = limn→∞(Bn =Mrn(C), ψn), where = ∞. mi li ψn(f ) = diag(f, f,··· f , f (x1), f (x2), ..., f (xmn−ln)) for all f ∈ Mrn(C), ln {z } n = 1, 2, .... Note that ψn is injective. Denote en = diag(1Mrn·ln , 0, ..., 0) ∈ Mrn+1(C), n = 1, 2, ... It is standard that B has tracial rank zero and K∗(B) = K∗(C ⊗ U ). Note that B is a unital simple AH-algebra with no dimension growth, with real rank zero and with a unique tracial state. Note that (K0(B), K0(B)+, [1B], K1(B)) = (K0(C ⊗ U ), K0(C ⊗ U )+, [1C⊗U ], K1(C ⊗ U )). Thus we obtain a KK-equivalence κ ∈ KLe(B, C ⊗ U )++ (by the UCT). It is standard to construct a unital homomorphism h : C → B such that KK(h) = κ−1 ◦ jC (see [49]). In particular if N is lager enough, we can choose homomorphism h′ : C → BN such that h = ıN ◦h′, where ıN : BN → B is the inclusion. 221 We also have that α1 ◦ κ ∈ KLe(B, A)++, where recall A = A1 ⊗ U. We also note that B has a unique tracial state. Let γ : T (A) → T (B) by γ(τ ) = t0 where t0 ∈ T (B) is the unique tracial state. It follows that α1 ◦ κ and γ is compatible. By the second part of Theorem 21.9, there is a unital homomorphism H : B → A such that [H] = α1 ◦ κ. Define ϕ : C → A by ϕ = H ◦ h = H ◦ ıN ◦ h′. Then, ϕ is injective and by (e 22.3) and [h] = κ−1 ◦ jC , we have [ϕ] = α. To show the last part, define qn = ϕn+1,∞(en) ∈ B, n = N + 1, N + 2,··· . Define pn = 1 − H(qn), n = N + 1, N + 2,··· . One checks that lim n→∞ max{τ (1 − pn) : τ ∈ T (A)} = lim ln mn =0 (e 22.4) Note that for n > N , qn commutes with image of h and the homomorphism (1 − qn)h(1 − qn) : C → (1 − qn)B(1 − qn) is defined by point evaluation. Define ϕ′n : C → (1 − pn)A(1 − pn) by ϕ′n(f ) = H(qn)H ◦ h(f )H(qn). Define ϕ′′n(f ) = pnH ◦ h(f )pn, which is a point-evaluation map. The lemma follows. We also have the following: Lemma 22.2. Let C = Mk(C(T)) and A be a unital infinite dimensional simple C∗-algebra with stable rank one and with the (SP) property. Then the conclusion of 22.1 also holds. projections e1, e2, ..., ek such that Pk Proof. The existence of α implies that A contains mutually equivalent and mutually orthogonal i=1 ei = 1A. Since eiAei are unital infinite dimensional simple C∗-algebra with stable rank one and with (SP), the general case can be reduced to the case that k = 1. Fix 1 > δ > 0. Choose a non-zero projection p ∈ A such that τ (p) < δ for all τ ∈ T (A). Note that K1(pAp) = K1(A), since A is simple. Let α1 : K1(C(T)) → K1(pAp) be the homomorphism given by α. Let z ∈ C(T) be the standard unitary generator. Let x = α1([z]) ∈ K1(pAp). Since pAp has stable rank one, there is a unitary u ∈ pAp such that [u] = x in K1(pAp) = K1(A). Define ϕ′ : C(T) → pAp by ϕ′(f ) = f (u) for all f ∈ C(T). Define ϕ′′ : C(T) → (1 − p)A(1 − p) by ϕ′′(f ) = f (1)(1 − p) for all f ∈ C(T) (where f (1) is a point-evaluation at 1 on the unit circle). Define ϕ = ϕ′ ⊕ ϕ′ : C(T) → A. The lemma follows. Corollary 22.3. Let X be a connected finite CW complex, C = P Mm(C(X))P, where P ∈ Mm(C(X)) is a projection, let A1 ∈ B0 be a unital separable simple C∗-algebra which satisfies the UCT and let A = A1 ⊗ U, where U is a UHF-algebra of infinite type. Suppose that α ∈ KK(C, A)++ and γ : T (A) → Tf (C(X)) is a continuous affine map. Then there exists a sequence of contractive completely positive linear maps hn : C → A such that (1) limn→∞ khn(ab) − hn(a)hn(b)k = 0, for any a, b ∈ C, (2) for each hn, the map [hn] is well defined and [hn] = α, and (3) limn→∞ max{τ ◦ hn(f ) − γ(τ )(f ) : τ ∈ T (A)} = 0 for any f ∈ C. Proof. By Theorem 21.10, one may assume that A is a unital C-algebra described in 14.8. It follows from Lemma 22.1 that there is a unital homomorphism hn : C → A such that [hn] = α. Moreover, where h′′n : C → pnApn is a homomorphism with [h′′n] = [Φ evaluation map Φ n → ∞. We will modify the map hn = h′n ⊕ h′′n to get our homomorphism. ] in KK(C, pnApn) for some point , where pn is a projection in A with τ (1 − pn) converge to 0 uniformly as ′ hn = h′n ⊕ h′′n, ′ 222 Assert that for any finite subset H ⊆ Cs.a, and ǫ > 0, and any sufficiently large n, there is a unital homomorphism hn : C → pnApn such that [hn] = [Φ] in KK(C, pnApn) for some point-evaluation Φ, and τ ◦ hn(f ) − γ(τ )(f ) < ǫ for all τ ∈ T (A) for all f ∈ H. The corollary then follows by replacing the map h′′n by the map hn—-of course, we use limn→∞ τ (1 − pn) = 0. Let H1,1 (in place of H1,1) be the finite subset of Lemma 17.1 with respect to H (in place of H), ǫ/8 (in the place of σ), and C (in the place of C). Since γ(T (A)) ⊆ Tf (C(X)), there is σ1,1 > 0 such that γ(τ )(h) > σ1,1 for all h ∈ H1,1 for all τ ∈ T (A). Let H1,2 ⊆ C + (in the place of H1,2) be the finite subset of Lemma 17.1 with respect to σ1,1. Since γ(T (A)) ⊆ Tf (C(X)), there is σ1,2 > 0 such that γ(τ )(h) > σ1,2 for all h ∈ H1,2 for all τ ∈ T (A). Let M be the constant of Lemma 17.1 with respect to σ1,2. Using a same argument as that of Lemma 16.12, for sufficiently large n, there is a C-subalgebra D ⊆ pnApn ⊆ A such that D ∈ C0, a continuous affine map γ′ : T (D) → T (C) such that γ′( 1 τ (p) τD)(f ) − γ(τ )(f ) < ǫ/4 for all τ ∈ T (A) for all f ∈ H, where p = 1D, τ (1 − p) < ǫ/(4 + ǫ), for all τ ∈ T (A), γ′(τ )(h) > σ1,1 for all τ ∈ T (D) for all h ∈ H1,1 and γ′(τ )(h) > σ1,2 for all τ ∈ T (D) for all h ∈ H1,2. (e 22.5) (e 22.6) Since A is simple and not of elementary, one may assume that the dimensions of the irreducible representations of D are at least M . Thus, by Lemma 17.1, there is a homomorphism ϕ : C → D such that [ϕ] = [Φ] in KK(C, D) for a point evaluation map Φ, and that τ ◦ ϕ(f ) − γ′(τ )(f ) < ǫ/4 for all f ∈ H for all τ ∈ T (D). Pike a point x ∈ X, and define h : C → pnApn by f 7→ f (x)(pn − p) ⊕ ϕ(f ) for all f ∈ C. Then a calculation as in the proof of Theorem 17.3 shows that the homomorphism h′n⊕h satisfies the assertion. Corollary 22.4. Let C ∈ H (where H is defined in 14.6) and let A1 ∈ B0 be a unital separable simple C∗-algebra which satisfies the UCT and let A = A1 ⊗ U for some UHF-algebra U of infinite type. Suppose that α ∈ KKe(C, A)++, λ : U (C)/CU (C) → U (A)/CU (A) is a continuous homomorphism and γ : T (A) → Tf (C) is a continuous affine map such that α, λ, and γ are compatible. Then there exists a sequence of unital contractive completely positive linear maps hn : C → A such that (1) limn→∞ khn(ab) − hn(a)hn(b)k = 0 for any a, b ∈ C, (2) for each hn, the map [hn] is well defined and [hn] = α, 223 (3) limn→∞ max{τ ◦ hn(f ) − γ(τ )(f ) : τ ∈ T (A)} = 0 for all f ∈ C, and, (4) limn→∞ dist(h‡n(¯u), λ(¯u)) = 0 for any u ∈ U (C). Proof. Let ǫ > 0. Let U be a finite generating set of Jc(K1(C)), where Jc(K1(C)) is as in 2.15. Let δ > 0 and G be the constant and finite subset of Lemma 21.5 with respect to U , ǫ and λ (in the place of α). Without loss of generality, one may assume that δ < ǫ. Let F be a finite subset such that F ⊃ G. Let H ⊆ C be a finite subset of self-adjoint elements with norm at most one. By Corollary 22.3, there is a positive completely linear map h′ : C → A such that h is F-δ-multiplicative, [h′] is well-defined and [h′] = α, and τ (h′(f )) − γ(τ )(f ) < ǫ, τ ∈ T(A), f ∈ H. (e 22.7) By Theorem 21.9, the C-algebra A is isomorphic to one of the model algebras constructed in Theorem 14.8, and therefore there is an inductive limit decomposition A = lim−→(Ai, ϕi), where Ai and ϕi are described as in Theorem 14.8. Without loss of generality, one may assume that h′(C) ⊆ Ai. Therefore, by Theorem 14.8, the map ϕ1,∞ ◦ h has a decomposition ϕ1,∞ ◦ h′ = ψ0 ⊕ ψ1 such that ψ0, ψ1 satisfy the (1)-(4) Lemma 21.5 with the above δ. e0 = ψ0(1C ), such that It then follows from Lemma 21.5 that there is a homomorphism Φ : C → e0Ae0, where (i) Φ is homotopic to a homomorphism with finite dimensional range and (e 22.8) (e 22.9) [Φ]∗0 = [ψ0] and (ii) for each w ∈ U , there is gw ∈ U0(B) with cel(gw) < ǫ such that λ( ¯w)−1(Φ ⊕ ψ1)‡( ¯w) = ¯gw Consider the map h := Φ ⊕ ψ1. Then h is F-ǫ-multiplicative. By (e 22.8), one has [h] = [ψ0] ⊕ [ψ1] = [h′] = α. By (e 22.7) and Condition (4) of Lemma 21.5, one has that, for all f ∈ H, τ (h(f )) − γ(τ )(f ) ≤ τ (h′(f )) − γ(τ )(f ) + δ < ǫ + δ < 2ǫ. It follows from (e 22.9) that, for all u ∈ U , dist(h(u), λ(u)) < ǫ. Since F, H, and ǫ are arbitrary, this shows the Corollary. Corollary 22.5. Let C ∈ H and let A1 ∈ B0 be a unital separable simple C∗-algebra which satisfies the UCT and let A = A1 ⊗ U for some UHF-algebra U of infinite type. Suppose that α ∈ KLe(C, A)++, λ : U (C)/CU (C) → U (A)/CU (A) is a continuous homomorphism and γ : T (A) → Tf (C) is a continuous affine map such that α, λ, and γ are compatible. Then there exists a unital homomorphism h : C → A such that (1) [h] = α, (2) τ ◦ h(f ) = γ(τ )(f ) for any f ∈ C, and, 224 (3) h‡n = λ. in U (C). One may assume that Gn ⊇ G(n)∪G(n−1), Gn ⊇ H1(n)∪H1(n+1)∪H2(n)∪H2(n−1), and Un ⊇ U (n) ∪ U (n − 1). Proof. Let us construct a sequence of homomorphisms hn : C → A which satisfies (1)–(4) of Corollary 22.4, and moreover, the sequence {hn(f )} is Cauchy for any f ∈ C. Then the limit map h = limn→∞ hn is the desired homomorphism. To construct such a sequence of homomorphisms, it is enough to construct a sequence of homomorphisms satisfying (1)-(4) of Corollary 22.4 such that {hn(f )} is Cauchy for any f ∈ C. Let {Fn} be an increasing sequence of the unit ball of C such that its union is dense in the unit ball of C. Define ∆(a) = min{γ(τ )(a) : τ ∈ T(A)}. Since γ is continuous and T(A) is compact, the map ∆ is an order preserving map from C 1,q + \ {0} to (0, 1). Let G(n),H1(n),H2(n) ⊆ C, U (n) ⊆ U∞(C), P(n) ⊆ K(C), γ1(n), γ2(n), δ(n) be the finite subset and constants of Theorem 12.5 with respect to Fn, 1/2n+1 and ∆/2. Without loss of generality, one may assume that δ(n) decrease to 0 if n → ∞ and P(n) ⊂ P(n + 1), n = 1, 2, ..., and ∪∞n=1P(n) = K(C). Let G1 ⊆ G2 ⊆ ··· be an increasing sequence of finite subset of C such thatSGn is dense in C, and let U1 ⊆ U2 ⊆ ··· be an increasing sequence of finite subset of U (C) such thatSUn is dense By Corollary 22.4, there is a G1-δ(1)-multiplicative map h′1 : C → A such that (4) the map [h′1] is well defined and [h1] = α, (5) τ ◦ hn(f ) − γ(τ )(f ) < min{γ1(1), 1 (6) dist(h‡n(¯u), λ(¯u)) < γ2(1) for any u ∈ Un. Define h1 = h′1. Assume that h1, h2, ..., hn : C → A are constructed such that (7) hi is Gi-δ(i)-multiplicative, i = 1, ..., n, (8) the map [hi] is well defined and [hi] = α, i = 1, ..., n, (9) τ ◦ hi(f ) − γ(τ )(f ) < min{ 1 (10) dist(h‡i (¯u), λ(¯u)) < 1 (11) khi−1(g) − hi(g)k < 1 2 γ2(i) for any u ∈ Ui, i = 1, ..., n, and 2i−1 for all g ∈ Gi−1, i = 2, 3, ..., n. 2 ∆(f ) : f ∈ H1} for any f ∈ G1 and, 2 γ1(i), 1 2 ∆(f ) : f ∈ H1(i)} for any f ∈ Gi, i = 1, ..., n, Let us construct hn+1 : C → A such that (12) hn+1 is Gn+1-δ(n + 1)-multiplicative, (13) the map [hn+1] is well defined and [hn+1] = α, (14) τ ◦ hn+1(f ) − γ(τ )(f ) < min{ 1 (15) dist(h‡n+1(¯u), λ(¯u)) < 1 (16) khn(g) − hn+1(g)k < 1 2 γ1(n + 1), 1 2 γ2(n + 1) for any u ∈ U, i = 1, ..., n, and 2n for all g ∈ Fn. 2 ∆(f ) : f ∈ H1(n + 1)} for any f ∈ Gn+1, Then the statement follows. By Corollary 22.4, there is G(n + 1)-δ(n + 1)-multiplicative map h′n+1 : C → A such that h′n+1 is Gn+1-δ(n + 1)-multiplicative, the map [h′n+1] is well defined and [h′n+1] = α, τ ◦ h′n+1(f ) − γ(τ )(f ) < min{ 1 2 γ1(n + 1), 1 2 ∆(f ) : f ∈ H2(n + 1)} (e 22.10) 225 for any f ∈ Gn+1 and 1 2 for any u ∈ U, i = 1, ..., n. In particular, this implies that [h′n+1]P = [hn]P , dist((h′n+1)‡(¯u), λ(¯u)) < γ2(n + 1) and for any f ∈ H2(n) (note that H2(n) ⊆ Gn) τ ◦ hn(f ) − τ ◦ h′n+1(f ) < γ1(n) + γ(τ )(f ) − τ ◦ h′n+1(f ) < γ1(n)/2 + γ1(n + 1)/2 < γ1(n). Also by (e 22.10), for any f ∈ H1(n), one has that 1 2 τ (h′n+1(f )) ≥ γ(τ )(f ) − ∆(f ) > 1 2 ∆(f ). By the inductive hypothesis, a same argument shows that τ (hn(f )) ≥ γ(τ )(f ) − 1 2 ∆(f ) > 1 2 ∆(f ) for all f ∈ H1(n). For any u ∈ U (n), one has dist(h′n+1(u), hn(u)) < < 1 2 1 2 γ2(n + 1) + dist(γ(u), hn(u)) γ2(n + 1) + 1 2 γ2(n) < γ1(n). W ∈ A such that Note that both h′n+1 and hn are G(n)-δ(n)-multiplicative, by Theorem 12.5, there is a unitary kW ∗h′n+1(g)W − hn(g)k < 1/2n for all g ∈ Fn. Then the map hn+1 := AdW ◦ h′n+1 satisfies the desired conditions, and the statement is proved. Lemma 22.6. Let C ∈ C0. Let ε > 0, F ⊂ C be any finite subset. Suppose that B is a unital separable simple C∗-algebra in B0, A = B ⊗ U for some UHF-algebra of infinite type, α ∈ KKe(C ⊗ C(T), A)++. Then there is a unital ε-F-multiplicative contractive completely positive linear map ϕ : C ⊗ C(T) → A such that [ϕ] = α. (e 22.11) Proof. Denote by α0 and α1 the induced maps induced by α on K0-groups and K1-groups. By Theorem 18.2, there exist an F-ǫ-multiplicative map ϕ1 : C ⊗ C(T) → A ⊗ K and a homomorphism ϕ2 : C ⊗ C(T) → A ⊗ K with finite dimensional range such that [ϕ1] = α + [ϕ2] in KK(C, A). In particular, one has that (ϕ1)∗1 = α1. Without loss of generality, one may assume that ϕ1, ϕ2 : C → Mr(A) for some integer r. Since Mr(A) ∈ B0, for any finite subset G ⊆ Mr(A) and any ǫ′ > 0, there are G-ǫ′- multiplicative map L1 : Mr(A) → (1 − p)Mr(A)(1 − p) and L2 : Mr(A) → S0 ⊆ pMr(A)p for a C-subalgebra S0 ∈ C0 with 1S0 = p such that (1) a − L1(a) ⊕ L2(a) < ǫ′ for any a ∈ G, and 226 (2) τ ((1 − p)) < ǫ′ for any τ ∈ T (Mr(A)). Since K1(S0) = {0}, by choosing G sufficiently large and ǫ′ sufficiently small, one may assume that L1 ◦ ϕ1 is F-ε-multiplicative, and [L1 ◦ ϕ1]K1(C⊗C(T)) = (ϕ1)∗1 = α1. Moreover, since the positive cone of K0(C ⊗ C(T)) is finitely generated, by choosing ǫ′ even smaller, one may assume that the map κ := α0 − [L1 ◦ ϕ1]K0(C⊗C(T) : K0(C ⊗ C(T)) → K0(A) is positive. Pick a point x0 ∈ T, and consider the evaluation map π : C ⊗ C(T) ∈ f ⊗ g 7→ f · g(x0) ∈ C. Then π∗0 : K0(C ⊗ C(T)) → K0(C) is an order isomorphism. homomorphism h : C → qAq such that Pick a projection q ∈ A with [q] = κ([1]). Since qAq ∈ B0, by Theorem 18.7, there is a unital [h]0 = κ ◦ π−1 ∗0 on K0(C), and hence one has on K0(C ⊗ C(T)). Put ϕ = (L1 ◦ ϕ1) ⊕ (h ◦ π) : C ⊗ C(T) → A, and it is clear that (h ◦ π)∗0 = κ, ϕ∗0 = [L1 ◦ ϕ1]K0(C⊗C(T)) + κ = [L1 ◦ ϕ1]K0(C⊗C(T)) + α0 − [L1 ◦ ϕ1]K0(C⊗C(T)) = α0 and [ϕ]1 = [L1 ◦ ϕ1]K1(C⊗C(T)) = α1. Since K∗(C ⊗ C(T)) is finitely generated and torsion free, one has that [ϕ] = α in KK(C ⊗ C(T), A). Lemma 22.7. Let C ∈ C0. Let ε > 0, F ⊂ C ⊗ C(T) be any finite subset, σ > 0, H ⊂ (C ⊗ C(T))s.a. be a finite subset. Suppose that A is a unital C∗-algebra in B0, B = A ⊗ U for some UHF-algebra U of infinite type, α ∈ KKe(C ⊗ C(T), B)++, and γ : T (B) → Tf (C ⊗ C(T)) is a continuous affine map such that α and γ are compatible. There is a unital F-ε-multiplicative contractive completely positive linear map ϕ : C ⊗ C(T) → B such that (1) [ϕ] = α and (2) τ ◦ ϕ(h) − γ(τ )(h) < σ for any h ∈ H. Moreover, if A ∈ B1, β ∈ KKe(C, A)++, γ′ : T (A) → Tf (C) is a continuous affine map which is compatible with β and H′ ⊂ Cs.a. is a finite subset, then there is also a unital homomorphism ψ : C → A such that [ψ] = β and τ ◦ ψ(h) − γ′(τ )(h) < σ for all f ∈ H′. (e 22.12) Proof. Since K∗(C ⊗ C(T)) is finitely generated and torsion free, by the UCT, the element α ∈ KK(C⊗ C(T), A) is determined by the induced maps α0 ∈ Hom(K0(C ⊗ C(T)), K0(A)) and α1 ∈ Hom(K1(C ⊗ C(T)), K1(A)). Without loss of generality we may assume that projections in M2(C ⊗ C(T)) generates K0(C ⊗ C(T)). 227 To simplify the proof, without loss of generality, we may assume that khk ≤ 1 for all h ∈ H. Fix a finite generating set G of K0(C ⊗ C(T)). Since γ(τ ) ∈ Tf (C ⊗ C(T)) for all τ ∈ T (B) and τ (B) is compact, one is able to define ∆ : (C ⊗ C(T))q,1 + \ {0} → (0, 1) by ∆(h) = 1 2 inf{γ(τ )(h) : τ ∈ T (B)}. Fix a finite generating set G of K0(C ⊗ C(T)). Let H1 ⊆ C ⊗ C(T), δ > 0, and K ∈ N be the finite subset and the constants of Lemma 16.10 with respect to F, H, ǫ, σ/4 (in the place of σ), and ∆. Since A ∈ B0 and U is of infinite type, for any finite subset G′ ⊆ B and any ǫ′ > 0, there are G′-ǫ′-multiplicative maps L1 : B → (1 − p)B(1 − p) and L2 : B → D ⊗ 1MK ⊂ D ⊗ MK ⊆ pBp for a C-subalgebra D ∈ C0 with 1D⊗MK = p such that (3) a − L1(a) ⊕ L2(a) < ǫ′ for any a ∈ G′, and (4) τ ((1 − p)) < min{ǫ′, σ/4} for any τ ∈ T (B). Put S = D ⊗ MK . By choosing G′ large enough and ǫ′ small enough, one may assume [L1] and [L2] are well-defined on α(K(C ⊗ C(T))), and α = [L1] ◦ α + [j] ◦ [L2] ◦ α, (e 22.13) where j : S → A is the embedding. Note that since K1(S) = {0}, one has that α1 = [L1] ◦ αK1(C⊗C(T)). Define κ′ = [L2] ◦ αK0(C⊗C(T)), which is a homomorphism from K0(C ⊗ C(T)) to K0(D) which mapping [1C⊗C(T)] to [1D]. Let {ei,j : 1 ≤ i, j ≤ K} be a system of matrix units for MK. View ei,j ∈ D ⊗ MK . Then ei,j commutes with the image of L2. Define L′2 : B → D ⊗ e1,1 by L′2(a) = L2(a) ⊗ e1,1 for all a ∈ Mr(A). Put κ = [L′2] ◦ αK0(C⊗C(T)). Put D′ = D ⊗ e1,1. Moreover, by choosing G′ larger and ǫ′ smaller, if necessarily, there is an continuous affine τ (e1,1) τD′)(f ) − γ(τ )(f ) < σ/4 for any f ∈ H, map γ′ : T (D′) → T (C ⊗ C(T)) such that, for all τ ∈ T (A), (5) γ′( (6) γ′(τ )(h) > ∆(h) for any h ∈ H1, and (7) γ′( Then it follows from Lemma 16.10 that there is an F-ǫ-multiplicative map ϕ2 : C ⊗ C(T) → MK (D) = S such that τ (e1,1) τD′)(p) − τ (κ([p])) < δ for all projections p ∈ M2(C ⊗ C(T)). 1 1 (ϕ2)∗0 = Kκ = κ′ and (1/K)t ◦ ϕ2(h) − γ′(t)(h) < σ/4, h ∈ H, t ∈ T (D′). On the other hand, since (1 − p)A(1 − p) ∈ B0, by Lemma 22.6, there is a unital F-ǫ- multiplicative map ϕ1 : C ⊗ C(T) → (1 − p)A(1 − p) such that [ϕ′] = [L1] ◦ α in KK(C ⊗ C(T), A). Define ϕ = ϕ1 ⊕ j ◦ ϕ2 : C ⊗ C(T) → (1 − p)A(1 − p) ⊕ S ⊆ A. Then one has ϕ∗0 = (ϕ1)∗0 + (j ◦ ϕ2)∗0 = ([L1] ◦ α)K0(C⊗C(T)) + ([j ◦ L2] ◦ α)K0(C⊗C(T)) = α0 228 ϕ∗1 = (ϕ1)∗1 + (j ◦ ϕ2)∗1 = ([L1] ◦ αK1(C⊗C(T)) = α1. For any h ∈ H and any τ ∈ T (A), one has (note that we have assumed khk ≤ 1 for all and Hence [ϕ] = α in KK(C ⊗ C(T)). h ∈ H), τ ◦ ϕ(h) − γ(τ )(h) < τ ◦ ϕ(h) − τ ◦ j ◦ ϕ2(h) + τ ◦ j ◦ ϕ2(h) − γ(τ )(h) 1 < σ/4 + τ ◦ j ◦ ϕ2(h) − γ′( < τ ◦ ϕ(h) − γ′( < σ/4 + σ/4 + σ/4 < σ. τ (e1,1) τS)(h) + γ′( τD′)(h) + γ′( τS)(h) − γ(τ )(h) τ (p) 1 τ (p) 1 1 τ (e1,1) τD′)(h) − γ(τ )(h) Hence the map ϕ satisfies the lemma. To see the last part of the lemma holds, we note that, when C ⊗ C(T) is replaced by C and A is assumed to be in B1, the only difference is that we may not use 22.6. But then we can appeal to 18.7 to obtain ϕ1. The semi-projectivity of C allows us actually obtain a unital homomorphism. Corollary 22.8. Let C ∈ C0. Suppose that A is a unital separable simple C∗-algebra in B0, B = A ⊗ U for some UHF-algebra of infinite type, α ∈ KKe(C, B)++ and γ : T (B) → Tf (C)) is a continuous affine map. Suppose that (α, λ, γ) is a compatible triple. Then there is a unital homomorphism ϕ : C → B such that [ϕ] = α and ϕT = γ. In particular, ϕ is a monomorphism. Proof. The proof is exactly the same argument employed in 22.5 by using the second part of 22.7. The reason ϕ is a monomorphism because γ(τ ) is faithful for each τ ∈ T (A). Lemma 22.9. Let B ∈ B0 which satisfies the UCT, A1 ∈ B0, let C = B ⊗ U1 and A = A1 ⊗ U2, where U1 and U2 are unital infinite dimensional UHF-algebras. Suppose that κ ∈ KLe(C, A)++, γ : T (A) → T (C) is a continuous affine map and α : U (C)/CU (C) → U (A)/CU (A) is a continuous homomorphism for which γ, α and κ are compatible. Then there exists a unital monomorphism ϕ : C → A such that (1) [ϕ] = κ in KLe(C, A)++, (2) ϕT = γ and ϕ‡ = α. Proof. The proof follows a same line as that of Lemma 8.5 of [69]. By the classification theorem, one can write where Cn is a direct sum of C-algebras in C0 or in H. Let κn = κ ◦ [ϕn,∞], αn = α ◦ ϕ‡n,∞, and γn = (ϕn,∞)T ◦ γ. By Corollary 22.5 or Corollary 22.8, there are unital monomorphism C = lim−→(Cn, ϕn,n+1) such that [ϕn] = αn, ψ‡n = αn, and (ψn)T = γn. ψn : Cn → A 229 In particular, the sequence of monomorphisms ψn satisfies [ψn+1 ◦ ϕn,n+1] = [ψn], ψ‡n+1 ◦ ϕn,n+1 = ψ‡n, and (ϕn+1 ◦ ϕn,n+1) = (ψn)T . Let Fn ⊆ Cn be a finite subset such that ϕn,n+1(Fn) ⊆ Fn+1 and S ϕn,∞(Fn) is dense in C. Applying Theorem 12.5 with ∆(h) = inf{γ(τ )(ϕn,∞(h)) : τ ∈ T (A)}, h ∈ C + sequence of unitaries un ∈ A such that n , there is a Adun+1 ◦ ψn+1 ◦ ϕn,n+1 ≈1/2n Adun ◦ ψn on Fn. The the maps {Adun ◦ ψn : n = 1, 2, ...} converge to a unital homomorphism ϕ : C → A which satisfies the lemma. Lemma 22.10. Let C be a unital C-algebra. Let p ∈ C be a full projection. Then, for any u ∈ U0(C), there is a unitary v ∈ pCp such that u = v ⊕ (1 − p) in U0(C)/CU (C). If, furthermore, C is separable and has stable rank one, then, for any u ∈ U (C), there is a unitary v ∈ pCp such that u = v ⊕ (1 − p) in U (C)/CU (C). k=1 exp(ick). By 3.1 of [98], u∗(v + (1 − p)) ∈ CU (C). k=1 exp(ibk) for some bk ∈ Cs.a., k = 1, 2, ..., n. Then there are ck ∈ pCp k=1 exp(ick))p. Then v ∈ U0(pCp) and Proof. It is sufficient to prove the first part of the statement. It is essentially contained in the proof of 4.5 and 4.6 of [38]. As in the proof of 4.5 of [38], for any b ∈ Cs.a., there is c ∈ pCp such that b − c ∈ C0, where C0 is the closed subspace of As.a. consisting of elements of the form C. x − y, where x =P∞n=1 c∗ncn and y =P∞n=1 cnc∗n (converge in norm) for some sequence {cn} in Now let u = Qn such that bk − ck ∈ C0, k = 1, 2, ..., n. Put v = p(Qn v + (1 − p) =Qn Lemma 22.11. Let C ∈ C0. Let ε > 0, F ⊂ C be a finite subset, 1 > σ1 > 0, 1 > σ2 > 0, U ⊂ Jc(K1(C ⊗ C(T))) ⊂ U (C ⊗ C(T))/CU (C ⊗ C(T)) be any finite subset (see 2.15) and H ⊂ (C ⊗ C(T))s.a. be a finite subset. Suppose that A is a unital separable simple C∗-algebra in B0, B = A ⊗ U for some UHF-algebra U of infinite type, α ∈ KKe(C ⊗ C(T), B)++, λ : Jc(K1(C ⊗ C(T))) → U (B)/CU (B) is a homomorphism, and γ : T (B) → Tf (C ⊗ C(T)) is a continuous affine map. Suppose that (α, λ, γ) is a compatible triple. Then there is a unital F-ε-multiplicative contractive completely positive linear map ϕ : C ⊗ C(T) → B such that (1) [ϕ] = α, (2) dist(ϕ‡(x), λ(x)) < σ1, for any x ∈ U , and (3) τ ◦ ϕ(h) − γ(τ )(h) < σ2, for any h ∈ H. Proof. Note that K(C ⊗ C(T)) is finitely generated modulo Bockstein operations and K0(C ⊗ C(T))+ is a finitely generated semi group. Using the inductive limit B = limn→∞(A⊗Mrn, ın,n+1), one can find, for n large enough, αn ∈ KKe(C ⊗ C(T), A ⊗ Mrn)++ such that α = αn × [ın] where [ın] ∈ KK(A ⊗ Mrn, B) is induced by the inclusion ın : A ⊗ Mrn → B. Replacing A by A ⊗ Mrn, without lose of generality we can assume that α = α1 × [ı], where α1 ∈ KKe(C ⊗ C(T), A)++ and ı : A → A ⊗ U = B is the inclusion. Note that Jc(K1(C ⊗ C(T))), same argument as above, we know that if the integer n above is large enough, then there is a λn : Jc(K1(C⊗C(T))) → U (A⊗Mrn)/CU (A⊗Mrn) such that ı‡n◦λn(u)−λ(u) arbitrarily small 230 (e.g smaller than σ1 4 ) for all u ∈ U . When we replace A by A ⊗ Mrn, we can assume λ = ı‡ ◦ λn, with λ1 : Jc(K1(C ⊗ C(T))) → U (A)/CU (A) and ı‡ : U (A)/CU (A) → U (B)/CU (B) being induced by the inclusion map. Furthermore, we can assume λ1 being compatible with α1. Without loss of generality, we may assume that khk ≤ 1 for all h ∈ H. Let pi, qi ∈ Mk(C) be projections such that {[p1] − [q1], ..., [pd] − [qd]} forms a set of standard generators of K0(C) (as an abelian group) for some integer k ≥ 1. By choosing a specific Jc, without loss of generality, one may assume that U = {((1k − pi) + pi ⊗ z)((1k − qi) + qi ⊗ z∗) : 1 ≤ i ≤ d}, where z ∈ C(T) is the identity function on the unit circle. Put u′i = (1k − pi) + pi⊗ z)((1k − qi) + qi ⊗ z∗). Hence {[u′1], ..., [u′d]} is a set of standard generators of K1(C ⊗ C(T)) ∼= K0(C) ∼= Zd. Then λ is a homomorphism from Zd to U (B)/CU (B). Let πe : C → F1 =Ll i=1 Mni be defined in 3.1. By 3.5, the map (πe)∗0 induces an embedding of K0(C) to Zl, and the map (πe ⊗ id)∗1 induces an embedding of K1(C ⊗ C(T)) ∼= Zd to K1(Ll i=1 Mni ⊗ C(T)) ∼= Zl. Choose Jc(K1(Ll i=1 Mni ⊗ C(T))) to be the subgroup generated by {ei ⊗ zi ⊕ (1 − ei); i = 1, ..., l}, where ei is a rank one projection of Mni and zi is the standard unitary generator of i-th copy of C(T). Note that the image of Jc(K1(C ⊗ C(T))) under πe is contained in Jc(K1(Ll We write B = B0 ⊗ U2, and B0 = A ⊗ U1, with U = U1 ⊗ U2, both U1 and U2 being UHF algebra of infinite type. Denoted by ı1 : A → B0, ı2 : B0 → B and ı = ı2 ◦ ı1 : A → B the inclusion maps. Recall α = α1 × [ı] ∈ KK(C, B). By applying Lemma 22.7, one obtains a unital F′-ε′-multiplicative contractive completely positive linear map ψ : C ⊗ C(T) → B0 such that i=1 Mni ⊗ C(T))). Denote by wj = ej ⊗ zj ⊕ (1 − ej), 1 ≤ j ≤ l. [ψ] = α1×[ı1] and (e 22.14) (e 22.15) τ ◦ ψ(h) − γ(τ )(h) < min{σ1, σ2}/3 for all h ∈ H and for all τ ∈ T(B0), where ε/2 > ε′ > 0 and F1 ⊃ F. (Note that T (B0) = T (A) = T (B), and the map γ : T (B) → Tf (C ⊗ C(T)) can be regarded as a map with domain T (B0). We may assume that ε′ is sufficiently small and F1 is sufficiently large so that not only (e 22.14) and (e 22.15) make sense but also that ψ‡ is well defined on ¯U and induces a homomorphism from Jc(K1(C ⊗ C(T))) to U (B0)/CU (B0). Let M be the integer in 15.3. For any ε′′ > 0 and any finite subset F′′ ⊂ B0, since B0 has the Popa condition and has (SP) property, there exist a non-zero projection e ∈ B0 and a unital F′′-ε′′-multiplicative contractive completely positive linear map L0 : B0 → F ⊂ eB0e, where F is a finite dimensional and 1F = e and a unital F′′-ε′′-multiplicative contractive completely positive linear map L1 : B0 → (1 − e)B0(1 − e) such that kb − ı ◦ L0(b) ⊕ L1(b)k < ε′′ for all b ∈ F′′, kL0(b)k ≥ kbk/2 for all b ∈ F′′ and τ (e) < min{σ1/2, σ2/2} for all τ ∈ T (B0), (e 22.16) (e 22.17) (e 22.18) where ı : F → eB0e is the embedding and L1(b) = (1 − p)b(1 − p) for all b ∈ B0. Since the positive cone of K0(C ⊗ C(T)) is finitely generated, with sufficiently small ε′′ and sufficiently large F′′, one may assume that [L0 ◦ ψ]K0(C⊗C(T)) is positive. Moreover, one may assume that (L0 ◦ ψ)‡ and (L1 ◦ ψ)‡ are well defined and induce homomorphisms from Jc(K1(C ⊗ C(T))) to U (B0)/CU (B0). One may also assume that [L1 ◦ ψ] is well defined. Moreover, we may assume that Li ◦ ψ is F-ε-multiplicative for i = 0, 1. 231 There is a projection Ec ∈ U2 such that Ec is a direct sum of M copies of some non-zero projections Ec,0 ∈ U2. Put E = 1U2 − Ec. Define ϕ0 : C ⊗ C(T) → F ⊗ EU2E → eB0e ⊗ EU2E by ϕ0(c) = L0 ◦ ψ(c) ⊗ E(∈ B) for all c ∈ C ⊗ C(T) and define ϕ′1 : C → F ⊗ EcU2Ec by ϕ′1(c) = L0 ◦ ψ(c) ⊗ Ec for all c ∈ C. Note that ϕ0 is also F-ε-multiplicative and ϕ‡0 is also well defined as (L0 ◦ ψ)‡ is. Moreover [ϕ′1] is well defined. Define L2 = ı2◦L1 ◦ ψ + ϕ0: C ⊗ C(T) →(cid:0)(1 − e)B0(1 − e) ⊗ 1U2(cid:1) ⊕(cid:0)eB0e ⊗ EU2E(cid:1) (⊂ B). Denote by λ0 = λ − L‡2 = λ − ϕ‡0 − (ı2◦L1 ◦ ψ)‡ : Jc(K1(C ⊗ C(T))) → U (B)/CU (B). Note that L0 factors through the finite dimensional algebra F and therefore [L0] = 0 on K1(B0), consequently [ϕ0]K1(C⊗C(T)) = 0 and [L1 ◦ ψ] = [ı1] ◦ [α1] on K1(C ⊗ C(T)). Hence [ı2 ◦ L1 ◦ ψ] = α on K1. Furthermore α is compatible with λ. We know that the image of λ0 is in U0(B)/CU (B). Note that, by 11.5, the group U0(B)/CU (B) is divisible. It is an injective abelian group. i=1 Mni ⊗ C(T)) → U0(B)/CU (B) such that Therefore there is a homomorphism λ : Jc(Ll Let β = [L0 ◦ ψ]K0(C): K0(C) →K0(F ) = Zn. Let R0 ≥ 1 be the integer given by 15.3. There is a unital C∗-subalgebra MM K ⊂ EcU2Ec such that K ≥ R0 and such that EcU2Ec can be written as MM K ⊗ U3. It follows from 15.3 that there is a positive homomorphism β1 : K0(F1) → K0(F ) such that β1 ◦ (πe)∗0 = M Kβ. Let h : F1 → F ⊗ MM K be the unital homomorphism such that h∗0 = β1. Put ϕ′′1 = h ◦ πe: C → F ⊗ MM K, and then one has that (ϕ′′1)∗0 = M Kβ. Let J : MM K → EcU2Ec be the embedding. One verifies that λ ◦ (πe)‡ = λ0 − L‡2. (e 22.19) (ıF ⊗ J)∗0 ◦ (ϕ′′1)∗0 = (ıF ⊗ J)∗0 ◦ M Kβ = ı∗0 ◦ (ϕ′1)∗0, (e 22.20) where ıF : F → eB0e and ı : F ⊗ EcU2Ec → eB0e ⊗ EcU2Ec be the unital embedding. Choose a unitary yi ∈ (ıF ⊗ J ◦ h)(ej)B(ıF ⊗ J ◦ h)(ej) such that ¯yj = λ(wj), j = 1, 2, ..., l, where we recall that wj = ej ⊗ zj ⊕ (1 − ej) ∈ F1 ⊗ C(T) = ⊕jMnj ⊗ C(T) is the standard generator of K1(Mnj ⊗ C(T). Let 1j be the unit of Mnj ⊂ F1, then 1j = ej ⊕ ej ⊕ ··· ⊕ ej . nj {z } z nj } { Define yj = diag( yj, yj, ..., yj )∈ (ıF ⊗ J ◦ h)(1j )B(ıF ⊗ J ◦ h)(1j ), j = 1, 2, ..., l. Then yj Define ϕ1 : F1⊗C(T) →(cid:0)(ıF ⊗ J) ◦ ϕ′′1(cid:1)(1C )B(cid:0)(ıF ⊗ J) ◦ ϕ′′1(cid:1)(1C ) by ϕ1(cj⊗f ) =(cid:0)(ıF ⊗ J) ◦ ϕ′′1(cid:1)(cj)f (yj) commutes with ıF ⊗ J(F1). for all cj ∈ Mnj and f ∈ C(T). Define ϕ1 = ϕ1◦(πe⊗idC(T)). Then, by identifying K0(C⊗C(T)) with K0(C), one has (e 22.21) (ϕ1)∗0 = ı∗0 ◦ (ϕ′1)∗0 and (ϕ1)‡ = λ. Define ϕ = ϕ0 ⊕ ϕ1 ⊕ ı2◦L1 ◦ ψ. We verify that, by (e 22.15) and (e 22.18), h ∈ H. τ ◦ ϕ(h) − γ(τ )(h) < σ2/3 + σ2/3 = 2σ2/3, It is ready to verify that ϕ∗0 = αK0(C⊗C(T)) and ϕ‡ = λ. (e 22.22) 232 Thus, since λ is compatible with α, Since K∗i(C ⊗ C(T)) ∼= K0(C) is free and finitely generated, one concludes that ϕ∗1 = αK1(C⊗C(T )). (e 22.23) [ϕ] = α. Corollary 22.12. Let C ∈ C0 and C1 = C ⊗ C(T). Suppose that A is a unital separable simple C∗-algebra in B0, B = A ⊗ U for some UHF-algebra of infinite type, α ∈ KKe(C1, B)++, λ : Jc(K1(C)) → U (B)/CU (B) is a homomorphism, and γ : T (B) → Tf (C1)) is a continuous affine map. Suppose that (α, λ, γ) is a compatible triple. Then there is a unital homomorphism ϕ : C1 → B such that [ϕ] = α, ϕ‡ = λ and ϕT = γ. In particular, ϕ is a monomorphism. Proof. The proof is exactly the same argument employed in 22.5 by using 22.11. Corollary 22.13. Let C ∈ C0 and let C1 = C or C1 = C ⊗ C(T). Suppose that A is a unital separable simple C∗-algebra in B0, B = A ⊗ U for some UHF-algebra of infinite type, α ∈ KKe(C1, B)++ and γ : T (B) → Tf (C1) is a continuous affine map. Suppose that (α, γ) is compatible. Then there is a unital homomorphism ϕ : C1 → B such that [ϕ] = α, ϕ‡ = λ and ϕT = γ. In particular, ϕ is a monomorphism. Proof. To apply 22.12, one needs λ. Note that Jc(K1(C1) is isomorphic to K1(C1) which is finitely generated. Let J (1) : K1(B) → U (B)/CU (B) be the splitting map defined in 2.15. Define λ = J (1) ◦ αK1(C1). Then (α, λ, γ) is compatible. This corollary then follows from the previous one. c c Theorem 22.14. Let X be a finite CW complex and let C = P Mn(C(X))P, where n ≥ 1 is an integer and P ∈ Mn(C(X)) is a projection. Let A1 ∈ B′ and let A = A1 ⊗ U for UHF- algebra U of infinite type. Suppose α ∈ KLe(C, A)++, λ : U∞(C)/CU∞(C) → U (A)/CU (A) be a continuous homomorphism and γ : T (A) → Tf (C) be a continuous affine map which are compatible. Then there exists a unital homomorphism h : C → A such that [h] = α, h‡ = λ and hT = γ. (e 22.24) Proof. The proof is similar to that of 6.6 of [69]. To simplify the notation, without loss of generality, we may assume that X is connected. Furthermore, a standard argument shows that the general case case can be reduced to the case that C = C(X). We assume that U (MN (C))/U0(MN (C)) = K1(C). Therefore, in this case, U (MN (C))/CU (MN (C)) = U∞(C)/CU∞(C). Write K1(C) = G1 ⊕ T or(K1(C)), where G1 is the free part of K1(C). Fix a point ξ ∈ X let C0 = C0(X \ {ξ}). Note that C0 is an ideal of C and C/C0∼= C. Write K0(C) = Z · [1C ] ⊕ K0(C0). 233 (e 22.25) Let B ∈ B0 be a unital separable simple C∗-algebra constructed in 14.11 such that (K0(B), K0(B)+, [1B], T (B), rB) = (K0(A), K0(A), [1A], T (A), rA) (e 22.26) and K1(B) = G1⊕T or(K1(A)). Put ∆(g) = inf{γ(τ )(g) : τ ∈ T (A)}. (e 22.27) For each g ∈ C+ \ {0}, since γ(τ ) ∈ Tf (C), γ is continuous and T (A) is compact, ∆(g) > 0. Let ε > 0, F ⊂ C be a finite subset, let 1 > σ1, σ2 > 0, H ⊂ Cs.a. be a finite subset, U ⊂ U (MN (C))/CU (MN (C)) be a finite subset. We assume that U = U0 ∪ U1, where U0 ⊂ U0(MN (C))/CU (MN (C)) and U1 ⊂ Jc(K1(C)) ⊂ U (MN (C))/CU (MN (C)). j=1 exp(√−1ai(u)), where ai(u) ∈ MN (C)s.a.. Write For each u ∈ U0, write u =Qn(u) ai(u) = (a(k,j) i (u))N×N , i = 1, 2, ..., n(u). (e 22.28) Write Put ci,k,j(u) = a(k,j) i (u) + (a(k,j) i )∗ 2 and di,k,j = a(k,j) i (u) − (a(k,j) i 2i )∗ . (e 22.29) M = max{kck,kci,k,j (u)k,kdi,k,j (u)k : c ∈ H, u ∈ U0}. (e 22.30) Choose a non-zero projection e ∈ B such that min{σ1, σ2} τ (e) < 16N 2(M + 1) max{n(u) : u ∈ U0} for all τ ∈ T (B). Let B2 = (1 − e)B(1 − e). In what follows we will use the identification (e 22.26). Define κ0 ∈ Hom(K0(C).K0(B2)) as follows. Define κ0(m[1C ]) = m[1 − e] for m ∈ Z and κ0K0(C0) = αK0(C0). Note that K1(B) = G1 ⊕ T or(K1(A)) and that α induces a map [α]T or(K1(C)) : T or(K1(C)) → T or(K1(A)). Using the given decomposition K1(C) = G1 ⊕ T or(K1(C), we can define κ1 : K1(C) → K1(B) by κ1G1 = id and κ1T or(K1(C)) = [α]T or(K1(C)). By the Universal Coefficient Theorem, there is κ ∈ KL(C, B2) which gives the above two homomorphisms κ0, κ1. Note that κ ∈ KLe(C, B2)++, since K0(C0) = kerρC(K0(C)). Choose H1 = H ∪ {ci,k,j(u), di,k,j(u) : u ∈ U0}. Every tracial state τ′ of B2 has the form τ′(b) = τ (b)/τ (1 − e) for all b ∈ B2 for some τ ∈ T (B). Let γ′ : T (B2) → T (C) be defined as follows. For τ′ ∈ T (B2) as above, define γ′(τ′)(f ) = γ(τ )(f ) for f ∈ C. It follows from 22.3 that there exists a sequence of unital contractive completely positive linear maps hn : C → B2 such that n→∞khn(ab) − hn(a)hn(b)k = 0 for all a, b ∈ C, lim [hn] = κ (−K∗(C) is finitely generated ) and max{τ ◦ hn(c) − γ′(τ )(c) : τ ∈ T (B2)} = 0 for all c ∈ C. lim n→∞ 234 Here we may assume that [hn] is well defined for all n and τ ◦ hn(c) − γ(τ )(c) < min{σ1, σ2} 8N 2 , n = 1, 2, .... (e 22.31) for all c ∈ H1 and for all τ ∈ T (B2). Choose θ ∈ KL(B, A) such that it gives the identification of (e 22.26), and, θG1 = αG1 and θT or(K1(A)) = idT or(K1(A)). Let e′ ∈ A be a projection such that [e′] ∈ K0(A) corresponds [e] ∈ K0(B) under the identification e 22.26. Let β = α − θ ◦ κ. Then β([1C ]) = [e′], βK0(C0) = 0 and βK1(C) = 0. (e 22.32) Then β ∈ KLe(C, e′Ae′). It follows 22.3 that there exists a sequence of unital contractive com- pletely positive linear maps ϕ0,n : C → e′Ae′ such that lim n→∞kϕ0,n(ab) − ϕ0,n(a)ϕ0,n(b)k = 0 and [ϕ0,n] = β. Note that, for each u ∈ U (MN (C)) with ¯u ∈ U0, (e 22.33) (e 22.34) DC (u) = [aj(u), n(u)Xi=1 wherebc(τ ) = τ (c) for all c ∈ Cs.a. and τ ∈ T (C). Since κ and λ are compatible, we compute, for ¯u ∈ U0, dist((hn)‡(¯u), λ(¯u)) < σ2/8. (e 22.35) Fix a pair of large integers n, m, define χn,m : Jc(G1) → Aff(T (A))/ρA(K0(A)) by λJc(G1) − (hn)‡Jc(G1) − (ϕ‡0,mJc(G1). (e 22.36) Viewing Jc(G1) as subgroup of Jc(K1(B))= Jc(K1(B2)), define χn,m on T or(K1(B2)) to be zero, we obtain a homomorphism χn,m : Jc(K1(B2)) → Aff(T (A))/ρA(K0(A)). It follows from 22.9 that there is a unital homomorphism ψ : B2 → (1 − e′)A(1 − e′) such that [ψ] = θ, ψT = idT (A) and ψ‡Jc(K1(B2)) = χn,mJc(K1(B2)) + Jc ◦ θK1(B2), where we identify K1(B2) with K1(B) By (e 22.37), ψ‡Aff(T (B2))/ρB2 (K0(B2)) = id. (e 22.37) (e 22.38) (e 22.39) Define L(c) = ϕ0,m(c) ⊕ ψ ◦ hn(c) for all c ∈ C. It follows, by choosing sufficiently large m and n, one has that L is ε-F-multiplicative, [L] = α, max{τ ◦ ψ(f ) − γ(τ )(f ) : τ ∈ T (A)} < σ1 for all f ∈ H and dist(L‡(¯u), λ(¯u)) < σ2. (e 22.40) (e 22.41) (e 22.42) This implies that that there is a sequence of contractive completely positive linear maps ψn : C → A such that lim n→∞kψn(ab) − ψn(a)ψn(b)k = 0 for all a, b ∈ C, [ψn] = α, max{τ ◦ ψn(c) − γ(τ )(c) : τ ∈ T (A1)} = 0 for all c ∈ Cs.a. and dist(ψ‡n(¯u), λ(¯u)) = 0 for all u ∈ U (MN (C))/CU (MN (C)). lim n→∞ lim n→∞ (e 22.43) (e 22.44) (e 22.45) (e 22.46) 235 Finally, by applying 12.5, as in the proof of 22.5, using ∆/2 above, we obtain a unital homo- morphism h : C → A such that [h] = α, hT = γ and h‡ = λ (e 22.47) as desired. Theorem 22.15. Let C ∈ C0 and let G = K0(C). Write G = Zk with Zk generated by {x1 = [p1] − [q1], x2 = [p2] − [q2], ..., xk = [pk] − [qk]}, where pi, qi ∈ Mn(C) (for some integer n ≥ 1) are projections, i = 1, ..., k. Let A be a simple C-algebra in B0, and let B = A ⊗ U for a UHF algebra U of infinite type. Suppose that ϕ : C → B is a monomorphism. Then, for any finite subsets F ⊆ C and P ⊆ K(C), any ε > 0 and σ > 0, any homomorphism Γ : Zk → U0(B)/CU (B), there is a unitary w ∈ B such that (1) k[ϕ(f ), w]k < ε, for any f ∈ F, (2) Bott(ϕ, w)P = 0, and (3) dist(h((1n − ϕ(pi)) + ϕ(pi) w)((1n − ϕ(qi)) + ϕ(qi) w∗)i, Γ(xi))) < σ, for any 1 ≤ i ≤ k, n where w = diag( w, ..., w). z } { Proof. Write B = limn→∞(A⊗ Mrn, ın,n+1). Using the fact that C is semi-projective (see [21]), one can construct a sequence of homomorphisms ϕn : C → A⊗ Mrn such that ın ◦ ϕn(c) → ϕ(c) for all c ∈ C. Without loss of generality, we can assume ϕ = ı◦ϕ1 for a homomorphism ϕ1C → A (replacing A by A ⊗ Mrn), where ı : A → A ⊗ U = B is the standard inclusion. Without loss of generality, one may assume that f ≤ 1 for any f ∈ F. For any nonzero positive element h ∈ C with norm at most 1, define ∆(h) = inf{τ (ϕ(h)); τ ∈ T (B)}. Since B is simple, one has that ∆(h) ∈ (0, 1). Let H1 ⊆ A+ \ {0}, G ⊆ A, δ > 0, P ⊆ K(A), H2 ⊆ As.a. and γ1 > 0 be the finite subsets and constants of Corollary 12.5 with respect to C (in the place of C), F, ǫ/2 and ∆/2 (since K1(C) = {0}, one does not need U and γ2). Note that B = A ⊗ U. Pick a unitary z ∈ U with sp(u) = T and consider the map ϕ′ : C ⊗ C(T) → B ⊗ U by a ⊗ f 7→ ϕ(a) ⊗ f (z). (Recall that ϕ(a) = ϕ1(a) ⊗ 1U .) Denote by Also define γ = (ϕ′)∗ : T (B) → Tf (C ⊗ C(T)). α := [ϕ′] ∈ KK(C ⊗ C(T), B). Note that K1(C ⊗ C(T)) = K0(C) = Zk. Identifying Jc(K1(C ⊗ C(T))) with Zk, and define a map λ : Jc(K1(U (C ⊗ C(T)))) → U0(B)/CU (B) by λ(a) = Γ(a) for any a ∈ Zk. 236 Denote by U = {(1n − pi + pi z′)(1n − qi + qi z′∗); i = 1, ..., k} ⊆ Jc(U (C ⊗ C(T))), where z′ is the standard generator of C(T), and denote by δ = min{∆(h)/4; h ∈ H1}. Applying Lemma 22.11, one obtains a F-ǫ/4-multiplicative map Φ : C ⊗ C(T) → B such that [Φ] = α, dist(Φ‡(x), λ(x)) < σ, x ∈ U and τ ◦ Φ(h ⊗ 1) − γ(τ )(h ⊗ 1) < min{γ1, δ}, Let ψ denote the restriction of Φ to C ⊗ 1. Then one has h ∈ H1 ∪ H2. (e 22.48) By (e 22.48), one has that for any h ∈ H1, [ψ]P = [ϕ]P . τ (ψ(h)) > γ(τ )(h) − δ = τ (ϕ′(h ⊗ 1)) − δ = τ (ϕ(h)) − δ > ∆(h)/2, and it is also clear that τ (ϕ(h)) > ∆(h)/2 for all h ∈ H1. Moreover, for any h ∈ H2, one has τ ◦ ψ(h) − τ ◦ ϕ(h) = τ ◦ Φ(h ⊗ 1) − τ ◦ ϕ′(h ⊗ 1) = τ ◦ Φ(h ⊗ 1) − γ(τ )(h ⊗ 1) ≤ γ1. Therefore, by Corollary 12.5, there is a unitary W ∈ B such that W ∗ψ(f )W − ϕ(f ) < ǫ/2 for all f ∈ F. The the element is the desired unitary. w = W ∗Φ(1 ⊗ z′)W Theorem 22.16. Let C be a unital C∗-algebra which is a finite direct sum of C∗-algebras in C0 and C∗-algebras with the form P Mn(C(X))P, where X is a finite CW complex, and let G = K0(C). Write G = ZkL Tor(G) with Zk generated by {x1 = [p1] − [q1], x2 = [p2] − [q2], ..., xk = [pk] − [qk]}, where pi, qi ∈ Mn(C) (for some integer n ≥ 1) are projections, i = 1, ..., k. Let A be a simple C-algebra in B0, and let B = A ⊗ U for a UHF algebra U of infinite type. Suppose that ϕ : C → B is a monomorphism. Then, for any finite subsets F ⊆ C and P ⊆ K(C), any ε > 0 and σ > 0, any homomorphism Γ : Zk → U0(Mn(B))/CU (Mn(B)), there is a unitary w ∈ B such that 237 (1) k[ϕ(f ), w]k < ε, for any f ∈ F, (2) Bott(ϕ, w)P = 0, and (3) dist(h((1n − ϕ(pi)) + ϕ(pi) w)((1n − ϕ(qi)) + ϕ(qi) w∗)i, Γ(xi))) < σ, for any 1 ≤ i ≤ k, n where w = diag( w, ..., w). z } { Proof. By 22.15, it suffices to prove the case that C = P Mn(C(X))P, where X is a finite CW complex, n ≥ 1 is an integer and P ∈ Mn(C(X)) is a projection. Proof follows the same lines of the proof as that of 22.15 but one will apply 22.14 instead of 22.11. 23 A pair of almost commuting unitaries Lemma 23.1. Let C ∈ C. There exists a constant MC > 0 satisfying the following: For any ε > 0, any x ∈ K0(C) and any n ≥ MC/ε, if ρC(x)(τ ) < ε for all τ ∈ T (C ⊗ Mn), (e 23.1) then, there are mutually inequivalent and mutually orthogonal minimal projections p1, p2, ..., pk1 and q1, q2, ..., qk2 in C ⊗ Mn and positive integers l1, l2, ..., lk1 , m1, m2, ..., mk2 such that x = [ k1Xi=1 lipi] − [ mjqj] and lipi) < 4ε and τ ( mjqj) < 4ε k2Xj=1 k2Xj=1 τ ( k1Xi=1 (e 23.2) (e 23.3) for all τ ∈ T (C ⊗ Mn). Proof. Let C = C(F1, F2, ϕ1, ϕ2) and F1 =Ll i=1 Mr(i). By 3.15, there is an integer N (C) > 0 such that every projection in C ⊗ K is equivalent to a finite direct sum of projections from a set of finitely many mutually inequivalent minimal projections (some of them may repeat in the direct sum) in MN (C)(C). By enlarge the number N (C), we may assume this set of mutually inequivalent projections are sitting in MN (C (C), orthogonally. We also assume that, as in 3.1, C is minimal. Let Suppose that n ≥ M/ε. With the canonical embedding of K0(C) into K0(F1) ∼= Zl, write M = N (C) + 2(r(1) · r(2)··· r(l)) x = x1 x2 ... xl  ∈ Zl (e 23.4) By (e 23.1), for any irreducible representation π of C and any tracial state t on Mn(π(C)), It follows that t ◦ π(x) < ε. xs/r(s)n < ε, s = 1, 2, ..., l. 238 (e 23.5) (e 23.6) Let Define T = max{xs/r(s) : 1 ≤ s ≤ l}. (e 23.7) . (e 23.8) y = x + T r(1) r(2) ... r(l)  and z = T r(1) r(2) ... r(l)  It follows that y ∈ K0(C). One also computes that It is clear that z ∈ K0(C)+ (see 3.5). y ∈ K0(C)+. It follows that there are projections p, q ∈ ML(C) for some integer L ≥ 1 such that [p] = y and [q] = z. Moreover, x = [p] − [q]. One also computes that One also has τ (q) < T /n < ε for all τ ∈ T (C ⊗ Mn). τ (p) < 2ε for all τ ∈ T (C ⊗ Mn). (e 23.9) (e 23.10) There are two sets of mutually orthogonal minimal projections {p1, p2, ..., pk1} and {q1, q2, ..., qk2} in C ⊗ Mn (since n > N (C)) such that Therefore [p] = k1Xi=1 li[pi] and [q] = mj[qj]. k2Xj=1 x = k1Xi=1 li[pi] − k2Xj=1 mj[qj]. (e 23.11) (e 23.12) Lemma 23.2. Let C ∈ C. There is an integer MC > 0 satisfying the following: For any ε > 0 and for any x ∈ K0(C) with τ (ρC (x)) < ε/24π for all τ ∈ T (C ⊗ Mn), where n ≥ 2MCπ/ε, there exists a pair of unitaries u and v ∈ C ⊗ Mn such that kuv − vuk < ε and τ (bott1(u, v)) = τ (x). (e 23.13) Proof. To simplify the proof, without loss of generality, we may assume that C is minimal. By applying 23.1, there are mutually orthogonal and mutually inequivalent minimal projections p1, p2, ..., pk1 , q1, q2, ..., qk2 ∈ C ⊗ Mn such that k1Xi=1 li[pi] − k2Xj=1 mj[qj] = x, where l1, l2, ..., lk1 , m1, m2, ..., mk2 are positive integers. Moreover, liτ (pi) < ε/6π and k1Xi=1 k2Xj=1 239 mjτ (qj) < ε/6π (e 23.14) for all τ ∈ T (C ⊗ Mn). Choose N ≤ n such that N = [2π/ε] + 1. By (e 23.14), N liτ (pi) + k1Xi=1 k2Xj=1 N mjτ (qj) < 1/2 for all τ ∈ T (C ⊗ Mn). (e 23.15) It follows that there are mutually orthogonal projections di,k, d′j,k∈ C ⊗ Mn, k = 1, 2, ..., N, i = 1, 2, ..., k1, and j = 1, 2, ..., k2 such that [di,k] = li[pi] and [d′j,k] = mj[qj], i = 1, 2, ..., k1, j = 1, 2, ..., k2 (e 23.16) k=1 di,k and D′j =PN and k = 1, 2, ..., N. Let Di =PN There is a partial isometry si,k, s′j,k ∈ C ⊗ Mn such that k=1 d′j,k, i = 1, 2, ..., k1 and j = 1, 2, ..., k2. s∗i,kdi,ksi,k = di,k+1, , (s′j,k)∗d′j,ks′j,k = d′j,k+1 k = 1, 2, ..., N − 1, s∗i,N di,N si,N = di,1 and (s′j,N )∗d′j,N s′j,N = d′j,1, (e 23.17) (e 23.18) i = 1, 2, ..., k1 and j = 1, 2, ..., k2. Thus we obtain unitary ui ∈ Di(C ⊗ Mn)Di and u′j = D′j(C ⊗ Mn)D′j such that u∗i di,kui = di,k+1, u∗i di,N ui = di,1, (u′j)∗d′j,kuj = d′j,k+1 and (u′j)∗d′j,N u′j = d′j,1, (e 23.19) i = 1, 2, ..., k1, j = 1, 2, ..., k2. Define √−1(2kπ/N )di,k and v′j = e vi = NXk=1 √−1(2kπ/N )d′j,k. e NXk=1 We compute that kuivi − viuik < ε and ku′jv′j − v′ju′jk < ε, τ (log viuiv∗i u∗i ) = liτ (pi) and 1 τ (log v′ju′j(v′j)∗(u′j)∗) = mjτ (qj), 2π√−1 2π√−1 1 for τ ∈ T (C ⊗ Mn), i = 1, 2, ..., k1 and j = 1, 2, ..., k2. Now define u = k1Xi=1 v = k2Xi=1 (v′j)∗ + (1C⊗Mn − k2Xj=1 u′j + (1C⊗Mn − k2Xj=1 vi + Di − k2Xi=1 ui + k1Xi=1 k2Xj=1 Di − D′j) and D′j). k2Xj=1 (e 23.20) (e 23.21) (e 23.22) (e 23.23) (e 23.24) (e 23.25) We then compute that τ (bott1(u, v)) = k1Xi=1 1 2π√−1 τ (log(viuiv∗i u∗i )) − for all τ ∈ T (C ⊗ Mn). = 240 k2Xj=1 k1Xi=1 1 2π√−1 liτ (pi) − τ (log v′ju′j(v′j)∗(u′j)∗) mjτ (qj) = τ (x) (e 23.26) k2Xj=1 Lemma 23.3. Let ε > 0. There exists σ > 0 satisfying the following: Let A = A1 ⊗ U, where U is a UHF-algebra of infinite type and A1 ∈ B0, let u ∈ U (A) be a unitary with sp(u) = T, and let x ∈ K0(A) with τ (ρA(x)) < σ for all τ ∈ T (A) and y ∈ K1(A). Then there exists a unitary v ∈ U (A) such that kuv − vuk < ε, bott1(u, v) = x and [v] = y. (e 23.27) Proof. Let ϕ0 : C(T) → A be the unital monomorphism defined by ϕ0(f ) = f (u) for all f ∈ C(T). Let ∆0 : Aq,1 + \ {0} → (0, 1) be defined by ∆0( f ) = inf{τ (f ) : τ ∈ T (A)}. Let ε > 0 be given. Choose 0 < ε1 < ε such that bott1(z1, z2) = bott1(z′1, z′2) if kz1 − z′1k < ε1 and kz2 − z′2k < ε1 for any two pairs of unitaries z1, z2, and z′1, z′2 which also have the property that kz1z2 − z2z1k < ε1 and kz′1z′2 − z′2z′1k < ε1. + \{0} be a finite subset, γ1 > 0, γ2 > 0 and H2 ⊂ C(T)s.a. be a finite subset as required by 12.7 (for ε1/4 and ∆0/2). Let H1 ⊂ C(T)1 Let δ1 = min{γ1/16, γ2/16, min{∆( f ) : f ∈ H1}/4}. Let δ = min{δ1/16, (δ1/16)(ε1/32π)}. Let e ∈ 1 ⊗ U ⊆ A be a non-zero projection such that τ (e) < δ1 for all τ ∈ T (A). Let B = eAe (then B ∼= A ⊗ U′ for some UHF-algebra U′). It follows from 18.10 that there is a unital simple C∗-algebra C′ = limn→∞(Cn, ψn), where Cn ∈ C0 and C = C′ ⊗ U such that (K0(C), K0(C)+, [1C ], T (C), rC ) = (ρA(K0(A)), (ρA(K0(A)))+, [e], T (A), rA). Moreover, we may assume that all ψn are unital. Now suppose that x ∈ K0(A) with τ (ρA(x)) < δ for all τ ∈ T (A) and suppose that y ∈ K1(A). Let z = ρA(x)∈ K0(C). We identify z with the element in K0(C) in the above identification. We claim that, there is n0 ≥ 1 such that there is x′ ∈ K0(Cn0 ⊗ U ) such that z = (ψn0,∞)∗0(x′)∈ K0(C) and t(ρCn0⊗U )(x′) < δ for all t ∈ T (Cn0 ⊗ U ). Otherwise, there is an increasing sequence nk, xk ∈ K0(Cnk ⊗ U ) such that (ψnk,∞)∗0(xk) = z ∈ K0(C) and tk(ρCnk⊗U )(xk) ≥ δ for some tk ∈ T (Cnk ⊗ U ), k = 1, 2, .... Let Lk : C → Cnk ⊗ U such that (e 23.28) lim n→∞kψn,∞ ◦ Ln(c) − ck = 0 for all c ∈ ψk,∞(Cnk ⊗ U ), k = 1, 2, .... It follows that the limit points of tk ◦ Lk is a tracial state of C. Let t0 be one of such limit. Then, by (e 23.28), t0(ρC (z)) ≥ δ. This proves the claim. Write U = limn→∞(Mr(m), ım), where ım : Mr(m) → Mr(m+1) is a unital embedding. By repeating the above argument, we obtain m0 ≥ 1 and y′ ∈ K0(Cn0 ⊗ Mr(m0)) = K0(Cn0) such that (ım0,∞)∗0(y′) = x′ and t(ρCn0 be as the /δ, r(m0)} and let y′′ = (ım0,m1)∗0(y′). Then, we constant in 23.2. Choose r(m1) ≥ max{48MCn0 compute that (y′)) < δ for all t ∈ T (Cn0 ⊗ Mr(m0)). Let MCN0 t(ρCn0 (y′′)) < δ for all t ∈ T (Cn0 ⊗ Mr(m1)). 241 It follows from 23.2 that there exists a pair of unitaries u′1, v′1 ∈ Cn0 ⊗ Mr(m1) such that (e 23.29) (e 23.30) (e 23.31) (e 23.32) ku′1 − v′1k < ε1/4 and bott1(u′1, v′1) = y′′. Put u1 = ım1,∞(u′1) and u2 = ım1,∞(v′1). Then (e 23.29) implies that ku1 − v1k < ε1/4 and bott1(u1, v1) = x′. Let h0 : Cn0 ⊗ U → eAe be a unital homomorphism given by 18.10 such that It follows that ρA ◦ (h0)∗0 = (ψn0,∞)∗0. ρA((h0)∗0(x′) − x) = 0. Let u2 = h0(u1) and v2 = h0(v1). We have that (e 23.33) Choose another non-zero projection e1 ∈ A such that e1e = ee1 = 0 and τ (e1) < δ1/16 for all τ ∈ T (A). It follows from 22.1 that there is a sequence of unital contractive completely positive linear maps Ln : C(T2) → e1Ae1 such that ρA(bott1(u2, v2) − x) = 0. lim n→∞kLn(f g) − Ln(f )Ln(g)k = 0 and [Ln](b) = x − bott1(u2, v2), (e 23.34) where b is the Bott element in K0(C(T 2)). (In fact, we can also apply 22.14 here.) Thus we obtain a pair of unitaries u3, v3 ∈ e1Ae1 such that ku3v3 − v3u3k < ε1/4 and bott1(u3, v3) = x − bott1(u2, v2). (e 23.35) Let e2, e3 ∈ (1− e− e1)A(1− e− e1) be a pair of non-zero mutually orthogonal projections such that τ (e2) < δ1/32 and τ (e3) < δ1/32 for all τ ∈ T (A). It follows from 17.3, applied to X = T, that there is a unitary u4 ∈ (1 − e − e1 − e2 − e3)A(1 − e − e1 − e2 − e3) such that τ ◦ (f (u4)) − τ ◦ f (u)k < δ1/4 for all f ∈ H2 ∪ H1 and for all τ ∈ T (A). (e 23.36) Let w = u2 + u3 + u4 + (1 − e − e1 − e2 − e3). It follows from Theorem 3.10 of [38] that there exists u5 ∈ U (e2Ae2) such that (e 23.37) Since A is simple and has stable rank one, there exists a unitary v4 ∈ e3Ae3 such that[v4] = y − [v2 + v3 + (e2+e3)] ∈ K1(A). Now define u5 = ¯u ¯w∗ ∈ U (A)/CU (A). u6 = u2 + u3 + u4 + u5 + e3 and v6 = v2 + v3 + (1 − e − e1 − e2−e3) + e2+v4. Then Moreover, ku6v6 − v6u6k < ε1/2, bott1(u6, v6) = x and [v6] = y. τ (f (u6)) ≥ ∆( f )/2 for all f ∈ H1 τ (f (u) − τ (f (u6)) < γ1 and ¯u6 = ¯u. It follows from 12.7 that there exists a unitary W ∈ A such that kW ∗u6W − uk < ε1/2. Now let v = W ∗v6W. We compute that kuv − vuk < ε, bott1(u, v) = bott1(u6, v6) = x and [v] = y. 242 (e 23.38) (e 23.39) (e 23.40) (e 23.41) (e 23.42) 24 More existence theorems for Bott elements Using 23.3, 22.1, 21.11, 18.11 and 12.9 we can show the following: Lemma 24.1. Let A = A1 ⊗ U1, where A1 is as in 14.8 and B = B1 ⊗ U2, where B1 ∈ B0 and U1, U2 are two UHF-algebras of infinite type. For any ε > 0, any finite subset F ⊂ A, any finite subset P ⊂ K(A), there exist δ > 0 and a finite subset Q ⊂ K1(A) satisfying the following: Let ϕ : A → B be a unital homomorphism and α ∈ KL(A ⊗ C(T), B) such that τ ◦ ρB(α(β(x))) < δ for all x ∈ Q and for all τ ∈ T (B), (e 24.1) there exists a unitary u ∈ B such that k[ϕ(x), u]k < ε for all x ∈ F and Bott(ϕ, u)P = α(β)P . (e 24.2) (e 24.3) Proof. Let ε1 > 0 and let F1 ⊂ A be a finite subset satisfying the following: If L, L′ : A ⊗ C(T) → B are two unital F′1-ε1-multiplicative contractive completely positive linear maps such that kL(f ) − L′(f )k < ε1 for all f ∈ F′1, (e 24.4) where then F′1 = {a ⊗ g : a ∈ F1 and g ∈ {z, z∗, 1C(T)}}, [L]β(P) = [L′]β(P). (e 24.5) Let B1,n = Mm(1,n)(C(T))⊕Mm(2,n)(C(T))⊕···⊕Mm(k1(1),n)(C(T)), B2,n = P Mr1(n)(C(Xn))P, where Xn is a finite disjoint union of S2, T0,k and T1,k (for various k ≥ 1). Let B3,n be a fi- nite direct sum of C∗-algebras in C0 (with trivial K1 and kerρB3,n = {0}), n = 1, 2, .... Put Cn = B1,n ⊕ B2,n ⊕ B3,n, n = 1, 2, .... We may write that A = limn→∞(Cn, ın) as in 14.8. So we also assume that ın are injective, kerρA ⊂ (ın,∞)∗0(kerρCn) and lim n→∞ sup{τ (1B1,n ⊕ 1B2,n) : τ ∈ T (B)} = 0 (e 24.6) (e 24.7) Let ε2 = min{ε1/4, ε/4} and let F2 = F1 ∪ F. Let P1,1 ⊂ K(B1,n1), P2,1 ⊂ K(B2,n1) and P3,1 ⊂ K(B3,n1) be finite subsets such that P ⊂ [ın1,∞](P1,1) ∪ [ın1,∞](P2,1) ∪ [ın1,∞])(P3,1) for some n1 ≥ 1. Let Q′ be a finite set of generators of K1(Cn) and let Q = [ın1,∞](Q′). F2,1 ⊂ B2,n2 and F3,1 ⊂ B3,n1 be finite subsets such that Without loss of generality, we may assume that F1 ∪ F ⊂ ın1,∞(Cn1). Let F1,1 ⊂ B1,n1, F1 ∪ F ⊂ ın1,∞(F1,1 ∪ F2,1 ∪ F3,1). (e 24.8) Let e1 = ın1,∞(1B1,n1 be defined by ), e2 = ın1,∞(1B2,n1 ) and e3 = 1 − e1 − e2. Let ∆1 : (B2,n1)q,1 + \ {0} → (0, 1) ∆1(h) = (1/2) inf{τ (ϕ(h)) : τ ∈ T (A)} for all h ∈ (B2,n1)1 + \ {0}. 243 Let ∆2 : Bq,1 3,n1 \ {0} → (0, 1) be defined by ∆2(h) = (1/2) inf{τ (ϕ(h)) : τ ∈ T (A)} for all h ∈ (B3,n1)1 + \ {0}. Note that B2,n1 has the form C in 12.5. So we will apply 12.5. Let H2,1 ⊂ (B1 2,n1)+ \ {0} (in place of H1) be a finite subset, γ2,1 > 0 (in place of γ1), δ2,1 > 0 (in place of δ), G2,1 ⊂ B2,n1 (in place of G) be a finite subset, P2,2 ⊂ K(B2,n1) (in place of P), H2,2 ⊂ (B2,n1)s.a. (in place of H2) be a finite subset as required by 12.5 for ε2/16, F2,1 and ∆1 (see also the Remark 12.6 since K1(B2,n1) is torsion or zero). Now let σ > 0 be required by 23.3 for ε2/4 (in place of ε). Let δ = σ · inf{τ (e1) : τ ∈ T (A)}. It follows from 23.3 that if τ ◦ ρB(α(β(x))) < δ for all x ∈ [ın1,∞])(P1,1) then there is a unitary v1 ∈ e1Be1 such that Bott(ϕ ◦ ın1,∞, v1)P1,1 = α ◦ β ◦ [ın1,∞](P1,1). Note that K1(B2,n1) is a finite group. Therefore α(β([ın1,∞])(K1(B2,n1)) ⊂ kerρB (e 24.9) (e 24.10) Define κ1 ∈ KK(B2,n1 ⊗ C(T), A) by κ1K(B2,n1 ) = [ϕ ◦ ın1,∞B2,n1 αβ(K(B2,n1 ). Since ın1,∞ is injective, by (e 24.10), κ1 ∈ KKe(B2,n1 ⊗ C(T), e2Be2)++. ] and κ1β(K(B2,n1 )) = Let σ0 = min{γ2,1/2, min{∆1(h) : h ∈ H2,1} · inf{τ (e2) : τ ∈ T (A)}. Define γ0 : T (e2Ae2) → Tf (B2,n1⊗C(T)) by γ0(τ )(f ⊗ 1C(T)) = τ ◦ ϕ ◦ ın1,∞(f ) for all f ∈ B2,n1 and γ0(1 ⊗ g) =RT g(t)dt for all g ∈ C(T). It follows from 22.14, applied to the space Xn1 × T, that there is a unital monomorphism Φ : B2,n1 ⊗ C(T) → e2Ae2 such that [Φ] = κ1 and ΦT = γ0. (by identifying B2,n1 with B2,n1 ⊗ 1C(T)) and v′2 = Φ(1 ⊗ z), where z ∈ C(T) Put L2 = ΦB2,n1 is the identity function on the unit circle. Then L2 is a unital monomorphism from B2,n1 to e2Ae2. We also have the following: [L2] = [ϕ ◦ ın1,∞], k[L2(f ), v′2]k = 0 Bott(L2, v′2)P2,2 = α(β([ın1,∞]))P2,2 and τ ◦ L2(f ) − τ ◦ ϕ ◦ ın1,∞(f ) = 0 for all f ∈ H2,1 ∪ H2,2 and for all τ ∈ T (e2Ae2). It follows from (e 24.13) that τ (L2(f )) ≥ ∆1( f ) · τ (e2). for all f ∈ H2,1 and τ ∈ T (A). By 12.5(see also 12.6), there exists a unitary w ∈ e2Ae2 such that kAd w ◦ L2(f ) − ϕ ◦ ın1,∞(f )k < ε2/16 for all f ∈ F2,1. Define v2 = w∗v′2w. Then (e 24.11) (e 24.12) (e 24.13) (e 24.14) (e 24.15) k[ϕ ◦ ın1,∞(f ), v2]k < ε2/8 and Bott(ϕ ◦ ın1,∞, v2)P2,1 = α(β([ın2,∞])P2,1 (e 24.16) Note that B3,n1 has the form C in 12.5. Let H3,1 ⊂ (B3,n1)1 + \ {0} (in place of H1) be a finite subset, γ3,1 > 0 (in place of γ1), δ3,1 > 0 (in place of δ), G3,1 ⊂ B3,N1) (in place of G) be a finite subset, P3,2 ⊂ K(B3,n1) (in place of P) be a finite subset and H3,2 ⊂ (B3,n1)s.a. (note that K1(B3,n1) = {0}) be as required by 12.5 for ε2/16, F3,1 and ∆2 (see also Remark 12.6). Let σ1 = (γ3,1/2) min{τ (e3) : τ ∈ T (A)} · min{∆1( f ) : f ∈ H3,1}. 244 = {0} and K1(B3,n1) = {0}. Therefore kerρB3,n1⊗C(T) = kerρB3,n1 = {0}. Note that kerρB3,n1 Define κ2 ∈ KK(B3,n1 ⊗ C(T)) as follows κ2K(B3,n1 ) = [ϕ ◦ ın1,∞]B3,n1 and κ2β(K(B3,n1 ) = α(β(ın1,∞)K(B3,n1 ). Thus κ2 ∈ KKe(B3,n1 ⊗ C(T), e3Ae3)++. It follows from 22.6 that there is a unital G3,1- min{ε2/16, δ3,1/2}-multiplicative contractive completely positive linear map L3 : B3,n1 → e3Ae3 and a unitary v′3 ∈ e3Ae3 such that [L3] = [ϕ], k[L3(f ), v′3]k < ε2/16 for all f ∈ G3,1, Bott(L3, v′3)P3,1 = κ2β(P3,2) and τ ◦ L3(f ) − τ ◦ ϕ ◦ ın1,∞(f ) < σ1 for all f ∈ H3,1 ∪ H3,2 and for all τ ∈ T (e3AE3). (e 24.17) (e 24.18) (e 24.19) (e 24.20) It follows that (e 24.19) that τ (L3(f )) ≥ ∆1( f )τ (e3) for all f ∈ H3,1 and for all τ ∈ T (A). (e 24.21) It follows from 12.5 and its remark that there exists a unitary w1 ∈ e3Ae3 such that kAd w1 ◦ L2(f ) − ϕ ◦ ın1,∞(f )k < ε2/16 for all f ∈ F3,1. (e 24.22) Define v3 = w∗1v′3w1. Then k[ϕ ◦ ın1,∞(f ), v3]k < ε2/8 and Bott(ϕ ◦ ın1,∞, v3)P3,1 = Bott(L3, v′3)P3,1 . (e 24.23) Let v = v1 + v2 + v3. Then Moreover, we compute that k[ϕ(f ), v]k < ε for all f ∈ F. Bott(ϕ, v)P = αβ(P). (e 24.24) (e 24.25) We actually prove the following: Lemma 24.2. Let A = A1 ⊗ U1, where A1 is as in 14.8 and B = B1 ⊗ U2, where B1 ∈ B0 is unital simple C∗-algebra and where U1, U2 are two UHF-algebras of infinite type. Write A = limn→∞(Cn, ın) as described in 14.8. For any ε > 0, any finite subset F ⊂ A, any finite subset P ⊂ K(A), there exists an integer n ≥ 1 such that P ⊂ [ın,∞](K(Cn)) and there is a finite subset Q ⊂ K1(Cn) which generates K1(Cn) and there exists δ > 0 satisfying the following: Let ϕ : A → B be a unital homomorphism and let α ∈ KK(Cn ⊗ C(T), B) such that τ ◦ ρB([ın,∞] ◦ α(β(x))) < δ for all x ∈ Q and for all τ ∈ T (B), there exists a unitary u ∈ B such that k[ϕ(x), u]k < ε for all x ∈ F and Bott(ϕ ◦ [ın,∞], u) = α(β)P . Remark 24.3. Note that, in the above statement, if an integer n works, any integer m ≥ n also works. In the terminology of Definition 3.6 of [75], the above also implies that B has property (B1) and (B2) associated with C. 245 Corollary 24.4. Let B ∈ B0 which satisfies the UCT, A1 ∈ B0, let C = B⊗U1 and A = A1⊗U2, where U1 and U2 are unital infinite dimensional UHF-algebras. Suppose that κ ∈ KKe(C, A)++, γ : T (A) → T (C) is a continuous affine map and α : U (C)/CU (C) → U (A)/CU (A) is a continuous homomorphism for which γ, α and κ are compatible. Then there exists a unital monomorphism h : C → A such that (1) [h] = κ in KKe(C, A)++, (2) hT = γ and h‡ = α. Proof. The proof follows the same line as that of Theorem 8.6 of [69]. Denote by κ ∈ KL(C, A) be the image of κ. It follows from Lemma 22.9 that there is a unital monomorphism ϕ : C → A such that [ϕ] = κ, ϕ‡ = α, and (ϕ)T = γ. Note that it follows from the UCT that (as an element in KK(C, A)) κ − [ϕ] ∈ Pext(K∗(C), K∗+1(A)). By Lemma 24.2, the C-algebra A has Property (B1) and Property (B2) associated with C in the sense of [75]. Since A contains a copy of U2, it is infinite dimensional, simple and antiliminal. It follows from a result in [1] that A contains an element b with sp(b) = [0, 1]. Moreover, A is approximately divisible. By Theorem 3.17 of [75], there is a unital monomorphism ψ0 : A → A which is approximate inner such that [ψ0 ◦ ϕ] − [ϕ] = κ − [ϕ] in KK(C, A). Then the map satisfy the corollary. h := ψ0 ◦ ϕ Lemma 24.5. Let A = A1 ⊗ U1, where A1 is as in 14.8 and B = B1 ⊗ U2, where B1 ∈ B0 is unital simple C∗-algebra and where U1, U2 are two UHF-algebras of infinite type. Let A = limn→∞(Cn, ın) be as described in 14.8. For any ε > 0, any σ > 0, any finite subset F ⊂ A, any finite subset P ⊂ K(A), and any projections p1, p2, ..., pk, q1, q2, ..., qk ∈ A such that {x1, x2, ..., xk} generates a free subgroup G of K0(A), where xi = [pi] − [qi], i = 1, 2, ..., k, there exists an integer n ≥ 1 such that P ⊂ [ın,∞](K(Cn)) and there is a finite subset Q ⊂ K1(Cn) which generates K1(Cn) and there exists δ > 0 satisfying the following: Let ϕ : A → B be a unital homomorphism, let Γ : G → U0(B)/CU (B) be a homomorphism and let α ∈ KK(Cn⊗C(T), B) such that τ ◦ ρB([ın,∞] ◦ α(β(x))) < δ for all x ∈ Q and for all τ ∈ T (B), there exists a unitary u ∈ B such that k[ϕ(x), u]k < ε for all x ∈ F, Bott(ϕ ◦ [ın,∞], u) = α(β)P , and dist(h((1 − ϕ(pi)) + ϕ(pi)u)((1 − ϕ(qi)) + ϕ(qi)u∗)i, Γ(xi)) < σ, i = 1, 2, ..., k. 246 Proof. This follows from 24.2 and 22.16. In fact, for any 0 < ε1 < ε/2 and finite subset F1 ⊃ F, by applying 24.2, there exists δ, n ≥ 1, Q ⊂ K1(Cn) and δ described above, and a unitary u1 ∈ U0(B) such that k[ϕ(x), u1]k < ε1 for all x ∈ F1 and Bott(ϕ ◦ ın,∞, u1) = α(β)P . By choosing smaller ε1 and larger F1, if necessary, we may assume that h((1 − ϕ(pi)) + ϕ(pi)u1)((1 − ϕ(qi)) + ϕ(qi)u∗1)i ∈ U0(B)/CU (B) is well defined for all 1 ≤ i ≤ k. Define a map Γ1 : G → U0(B)/CU (B) by Γ1(xi) = h((1 − ϕ(pi)) + ϕ(pi)u1)(1 − ϕ(qi)) + ϕ(qi)u∗1)i, i = 1, 2, ..., k. (e 24.26) By choosing a large n, without loss of generality, we may assume that there are projections p′1, p′2, ..., p′k, q′1, q2,′ ..., q′k ∈ Cn such that ın,∞(p′i) = pi and ın,∞(q′i) = qi, i = 1, 2, ..., k. Moreover, we may assume that F1 ⊂ ın,∞(Cn). Let Γ2 : G → U0(B)/CU (B) be defined by Γ2(xi) = Γ1(xi)∗Γ(xi), i = 1, 2, ..., k. It follows 22.16 that is a unitary v ∈ U0(B) such that k[ϕ(x), v]k < ε/2 for all x ∈ F, Bott(ϕ ◦ ın,∞, v) = 0 and dist(h((1 − ϕ(pi)) + ϕ(pi)v)((1 − ϕ(qi)) + ϕ(qi)v∗)i, Γ2(xi)) < σ, i = 1, 2, ..., k. Define u = u1v, Xi = h((1 − ϕ(pi)) + ϕ(pi)u1)((1 − ϕ(qi)) + ϕ(qi)u∗1)i and Yi = h((1 − ϕ(pi)) + ϕ(pi)v)((1 − ϕ(qi)) + ϕ(qi)v∗)i, i = 1, 2, ..., k. We then compute that k[ϕ(x), u]k < ε1 + ε/2 < ε for all x ∈ F, Bott(ϕ ◦ ın,∞, u) = Bott(ϕ ◦ ın,∞, u1) = α(β)P dist(h((1 − ϕ(pi)) + ϕ(pi)u)((1 − ϕ(qi)) + ϕ(qi)u∗)i, Γ(xi)) ≤ dist(XiYi, Γ1(xi)Yi) + dist(Γ1(xi)Yi, Γ(xi)) = dist(Xi, Γ1(xi)) + dist(Yi, Γ2(xi)) < σ, and for i = 1, 2, ..., k. (e 24.27) (e 24.28) (e 24.29) (e 24.30) (e 24.31) (e 24.32) (e 24.33) (e 24.34) (e 24.35) (e 24.36) 25 Another Basic Homotopy Lemma Lemma 25.1. Let A be a unital C-algebra and let U be an infinite dimensional UHF-algebra. Then there is a unitary w ∈ U such that for any unitary u ∈ A, one has τ (f (u ⊗ w)) = τ (f (1A ⊗ w)) =ZT f dm, f ∈ C(T), τ ∈ T (A ⊗ U ) (e 25.1) where m is the normalized Lebesgue measure on T. Furthermore, for any a ∈ A and τ ∈ T (A⊗U ), τ (a ⊗ wj) = 0 if j 6= 0. 247 Proof. Denote by τU the unique trace of U . Then any trace τ ∈ T (A ⊗ U ) is a product trace, i.e., τ (a ⊗ b) = τ (a ⊗ 1) ⊗ τU (b), a ∈ A, b ∈ U. Pick a unitary w ∈ U such that the spectral measure of w is the Lebesgue measure (a Haar unitary). Such a unitary always exists (it can be constructed directly; or, one can consider a strictly ergodic Cantor system (Ω, σ) such that K0(C(Ω) ⋊σ Z) ∼= K0(U ), and note that the canonical unitary in C(Ω) ⋊σ Z is a Haar unitary. Then by embedding C(Ω) ⋊σ Z into U , one obtains a Haar unitary in U ). Then one has τU (wn) =(cid:26) 1, if n = 0, 0, otherwise. Hence, for any τ ∈ T (A ⊗ U ), one has τ ((u ⊗ w)n) = τ (un ⊗ wn) = τ (un ⊗ 1)τU (wn) =(cid:26) 1, if n = 0, 0, otherwise; and therefore τ (P (u ⊗ w)) = τ (P (1 ⊗ w)) =ZT P (z)dm for any polynomial P . Since polynomials are dense in C(T), one has τ (f (u ⊗ w)) = τ (f (1 ⊗ w)) =ZT f dm, f ∈ C(T), as desired. Lemma 25.2. Let A be a unital separable amenable C∗-algebra and let L : A ⊗ C(T) → B be a unital completely positive linear map, where B is another unital amenable C∗-algebra. Suppose that C is a unital C∗-algebra and u ∈ C is a unitary. Then, there is a unique unital completely positive linear map Φ : A ⊗ C(T) → B ⊗ C such that ΦA⊗1C(T) = ı ◦ LA⊗1C(T) and Φ(a ⊗ zj) = L(a ⊗ zj) ⊗ uj for any a ∈ A and any integer j, where ı : B → B ⊗ C is the standard inclusion. Furthermore, if δ > 0 and G ⊂ A ⊗ C(T) is a finite subset, there is a δ1 > 0 and finite set G1 ⊂ A ⊗ C(T) (which does not depend on L) such that if L is G1-δ1-multiplicative, then Φ is G-δ-multiplicative. Proof. Denote by C0 the unital C-subalgebra of C generated by u, then the tensor product map L ⊗ idC0 : A ⊗ C(T) ⊗ C0 → B ⊗ C0 is unital and completely positive (see, for example, Theorem 3.5.3 of [8]). Define the homomor- phism ψ : C(T) → C(T) ⊗ C0 by ψ(z) = z ⊗ u. By Theorem 3.5.3 of [8] again, the tensor product map is unital and completely positive. Then the map idA ⊗ψ : A ⊗ C(T) → A ⊗ C(T) ⊗ C0 Φ := (L ⊗ idC0) ◦ (idA ⊗ψ) 248 satisfies the first part of the lemma. Let us consider the second part of the lemma. Let δ > 0 and G ⊂ A⊗ C(T) be a finite subset. Suppose that L is δ-G-multiplicative. To simplify notation, without loss of generality, we may assume that elements in G has the formP−n≤i≤n ai⊗zi. Let N = max{n :P−n≤i≤n ai⊗zi ∈ G}, let δ1 = δ/2N 2 and let G1 ⊃ {ai ⊗ zi : −n ≤ i ≤ n :P−n≤i≤n ai ⊗ zi ∈ G}. Then bi ⊗ zi)) Φ(( X−n≤i≤n = Xi,j = Xi,j ai ⊗ zi)( X−n≤i≤n Φ(aibj ⊗ zi+j) L(aibj ⊗ zi+j) ⊗ ui+j ( X−n≤i≤n = Φ( X−n≤i≤n i=−n cizi, ci ∈ C} be the algebra of all Laurent polynomials. The uniqueness L(ai ⊗ zi) ⊗ ui)( X−n≤i≤n ai ⊗ zi)Φ( X−n≤i≤n bi ⊗ zi), L(bi ⊗ zi) ⊗ ui) (e 25.6) ≈δ (e 25.2) (e 25.3) (e 25.4) (e 25.5) ifP−n≤i≤n ai ⊗ zi,P−n≤i≤n bi ⊗ zi ∈ G. It follows that Φ is G-δ-multiplicative. Let P (T) = {Pn follows from that A ⊗ C(T) is the closure of the algebraic tensor product A ⊗alg P (T). The following follows immediately from 25.2 and 25.1. Corollary 25.3. Let C be a unital C∗-algebra and let U be an infinite dimensional UHF- algebra. For any δ > 0 and any finite subset G ⊂ C ⊗ C(T), there exist δ1 > 0 and a finite subset G1 ⊂ C ⊗ C(T) satisfying the following: For any 1 > σ1, σ2 > 0, any finite subset H1 ⊂ C(T)+ \ {0} and any finite subset H2 ⊂ (C ⊗ C(T))s.a., any unital G1-δ1-multiplicative contractive completely positive linear map L : C ⊗ C(T) → A, where A is another unital C∗- algebra, there exists unitary w ∈ U satisfying the following: τ (L1(f )) − τ (L2(f )) < σ1 for all f ∈ H2, τ ∈ T (B), and τ (g(1A ⊗ w)) ≥ σ2(Z gdm) for all g ∈ H1, τ ∈ T (B), (e 25.7) (e 25.8) where B = A⊗U and m is the normalized Lebesgue measure on T, and L1, L2 : C⊗C(T) → A⊗U are G-δ-multiplicative contractive completely positive linear maps as Φ given by 25.2 such that Li(c⊗1C(T)) = L(c⊗1C(T))⊗1U (i = 1, 2), L1(c⊗zj) = L(c⊗zj )⊗wj, and L2(c⊗zj ) = L(c)⊗wj for all c ∈ C and all integer j. Lemma 25.4. Let A = A1 ⊗ U1, where A1 ∈ B0 which satisfies the UCT and U1 is a UHF- algebra of infinite type. For any 1 > ε > 0 and any finite subset F ⊂ A, there exist δ > 0, σ > 0, a finite subset G ⊂ A, a finite subset {p1, p2, ..., pk, q1, q2, ..., qk} of projections of A such that {[p1] − [q1], [p2] − [q2], ..., [pk] − [qk]} generates a free subgroup Gu of K0(A), and a finite subset P ⊂ K(A), satisfying the following: Let B = B1⊗U2, where B1 ∈ B0 which satisfies the UCT and U2 are UHF-algebras of infinite type. Suppose that ϕ : A → B is a unital homomorphism. 249 For unitary u ∈ U (B) such that k[ϕ(x), u]k < δ for all x ∈ G, Bott(ϕ, u)P = 0, dist(h((1 − ϕ(pi)) + ϕ(pi)u)(1 − ϕ(qi)) + ϕ(qi)u∗)i, ¯1) < σ and dist(¯u, ¯1) < σ, there exists a continuous path of unitaries {u(t) : t ∈ [0, 1]} ⊂ U (B) such that u(0) = u, u(1) = 1B, dist(u(t), CU (A)) < ε for all t ∈ [0, 1], k[ϕ(a), u(t)]k < ε for all a ∈ F and f or all t ∈ [0, 1] and length({u(t)}) ≤ 2π + ε. (e 25.9) (e 25.10) (e 25.11) (e 25.12) (e 25.13) (e 25.14) (e 25.15) (e 25.16) Proof. Without loss of generality, one only has to prove the statement with assumption that u ∈ CU (B). In what follows we will use the fact that every C-algebra in B0 has stable rank one. Define ∆(f ) = (1/2)Z f dm for all f ∈ C(T)1 + \ {0}, where m is the normalized Lebesgue measure on the unit circle T. Let A2 = A⊗ C(T). Let F1 = {x ⊗ f : x ∈ F, f = 1, z, z∗}. To simplify notation, without loss of generality, we may assume that F is a subset of the unit ball of A. Let 1 > δ1 > 0 (in place of δ), G1 ⊂ A2 be a finite subset (in place of G), 1/4 > σ1 > 0, 1/4 > σ2 > 0, P ⊂ K(A2) be a finite subset, H1 ⊂ C(T)1 + \ {0} be a finite subset, H2 ⊂ (A2)s.a. be a finite subset and U ⊂ U (M2(A2))/CU (M2(A)) be a finite subset as required by 12.9 for ε/4 (in place of ε), F1 (in place of F), ∆ and A2 (in place of A). We may assume, without loss of generality, that G1 = {a ⊗ f : a ∈ G2 and f = 1, z, z∗}, where G2 ⊂ A is a finite subset, and P = P1 ∪ β(P2), where P1,P2 ⊂ K(A) are finite subsets. We assume that (2δ1,P,G1) is a KL-triple for A2, (2δ1,P1,G2) is a KL-triple for A, and 1A1 ⊗ H1 ⊆ H2. We may also choose σ1 and σ2 such that max{σ1, σ2} < (1/4) inf{∆(f ) : f ∈ H1}. (e 25.17) Let δ2 (in place of δ1) and a finite subset G3 (in place of G1) be as required by 25.3 for A (in place of C), δ1/4 (in place of δ) and G1 (in place of G). By choosing even smaller δ2, without loss of generality, we may assume that G3 = {a ⊗ f : g ∈ G′2 and f = 1, z, z∗} with a large finite subset G′2 ⊃ G2. We may assume that δ2 < δ1. We may further assume that, U = U1 ∪ {1 ⊗ z} ∪ U2, (e 25.18) where U1 = {a ⊗ 1 : a ∈ U′1 ⊂ U (A)} and U′1 is a finite subset, U2 ⊂ U (A2)/CU (A2) is a finite subset whose elements represent a finite subset of β(K0(A)). So we may assume that U2 ∈ Jc(β(K0(A))). As in 12.10, we may assume that the subgroup of Jc(β(K0(A))) generated by U2 is free. Let U′2 be a finite subset of unitaries such that {¯x : x ∈ U′2} = U2. We may also assume that unitaries in U′2 has the form ((1 − pi) + pi ⊗ z)(1 − qi) + qi ⊗ z∗), i = 1, 2, ..., k. (e 25.19) 250 We further assume that pi ⊗ z ∈ G1, i = 1, 2, ..., k. Choose δ3 > 0 and a finite subset G′4 ⊂ A (and denote G4 := {g⊗f : g ∈ G′4, f = 1, z, z∗}.) such that, for any two unital G4-δ3-multiplicative contractive completely positive linear maps Ψ1, Ψ2 : A ⊗ C(T) → C (any unital C∗-algebra C), any G′4-δ3-multiplicative contractive completely positive linear maps Ψ0 : A → C and unitary V ∈ C (1 ≤ i ≤ k), if kΨ0(g) − Ψ1(g ⊗ 1)k < δ3 for all g ∈ G′4 kΨ1(z) − V k < δ3 and kΨ1(g) − Ψ2(g)k < δ3 for all g ∈ G4, then h(1 − Ψ0(pi) + Ψ0(pi)V )(1 − Ψ0(qi) + Ψ0(qi)V ∗i 210 hΨ1(((1 − pi) + pi ⊗ z)((1 − qi) + qi ⊗ z∗)i, ≈ σ2 khΨ1(x)i − hΨ2(x)ik < σ2/210 for all x ∈ U′2, Ψ1(((1 − pi) + pi ⊗ z)(1 − qi) + qi ⊗ z∗)) Ψ1(((1 − pi) + pi ⊗ z))Ψ1((1 − qi) + qi ⊗ z∗)), ≈ σ2 i = pi, d(2) i = qi,, there are projections ¯d(j) 210 (e 25.20) (e 25.21) (e 25.22) (e 25.23) (e 25.24) (e 25.25) (e 25.26) furthermore for d(1) such that i ∈ C and unitaries ¯z(j) i ∈ ¯d(j) i C ¯d(j) i Ψ1(((1 − d(j) ¯d(j) i ≈ σ2 i ) + d(j) Ψ1(d(j) 212 i ), ¯z(1) i ⊗ z)) ≈ σ2 212 (1 − ¯d(j) i ) + ¯z(j) i and i ≈ σ2 212 Ψ1(pi ⊗ z), and ¯z(2) i ≈ σ2 212 (e 25.27) (e 25.28) Ψ1(qi ⊗ z∗), where 1 ≤ i ≤ k, j = 1, 2. Choose σ > 0 so it is smaller than min{σ1/16, ε/16, σ2/16, δ2/16, δ3/16}. Choose δ5 > 0 and a finite subset G5 ⊂ A satisfying the following: there is a unital G4-σ/8- multiplicative contractive completely positive linear map L : A ⊗ C(T) → B′ such that kL(a ⊗ 1) − ϕ′(a)k < σ/8 for all a ∈ G′4 and kL(1 ⊗ z) − u′k < σ/8 (e 25.29) for any unital homomorphism ϕ′ : A → B′ and any unital u′ ∈ B′ so that kϕ′(g)u′ − u′ϕ′(g)k < δ5 for all g ∈ G5. Let δ = min{δ5/4, σ} and G = G5 ∪ G′4 ∪ G′2. Now suppose that ϕ : A → B is a unital homomorphism and u ∈ CU (B) which satisfy the assumption (e 25.9) to (e 25.11) for the above mentioned δ, σ, G, P, pi, and qi. There is an isomorphism s : U2 ⊗ U2 → U2. Moreover, s ◦ ı is approximately unitarily equivalent to the identity map on U2, where ı : U2 → U2 ⊗ U2 defined by ı(a) = a ⊗ 1 (for all a ∈ U2). To simplify notation, without loss of generality, we may assume that ϕ(A) ⊂ B ⊗ 1 ⊂ B ⊗ U2. Suppose that u ∈ U (B) ⊗ 1U2 is a unitary which satisfies the assumption. As mentioned at the beginning, we may assume that u ∈ CU (B) ⊗ 1U2. Without loss of generality, we may further assume that Let L : A ⊗ C(T) → B be a unital G4-δ2/8-multiplicative contractive completely positive j=1 cjdjc∗j d∗j , where cj, dj ∈ U (B) ⊗ 1U2, 1 ≤ j ≤ m. Let F1 = {cj , dj : 1 ≤ j ≤ m1}. u =Qm1 linear map such that kL(a ⊗ 1) − ϕ(a)k < σ/8 for all a ∈ G′4 and kL(1 ⊗ z) − uk < σ/8. (e 25.30) Since Bott(ϕ, u)P = 0, we can also assume that that [L]P1 = [ϕ]P1 and [L]β(P2) = 0. (e 25.31) 251 Since B is in B0, there is a projection p ∈ B and a unital C∗-subalgebra C ∈ C0 with 1C = p satisfying the following: kL(g) − [(1 − p)L(g)(1 − p) + L1(g)]k < σ2/32(m1 + 1) for all g ∈ G4 and k(1 − p)x − x(1 − p)k < σ2/32(m1 + 1) for all x ∈ F1, (e 25.32) (e 25.33) where L1 : A⊗C(T) → C is a unital G4-min{δ2/8, ε/8}-multiplicative contractive completely positive linear map, τ (1 − p) < min{σ1/16, σ2/16} for all τ ∈ T (B) and, using (e 25.11), (e 25.12), (e 25.30) and (e 25.22) to (e 25.21), we have that dist(L‡2(x), ¯1) < σ2/4 for all x ∈ {1 ⊗ ¯z} ∪ U2 and dist(L‡2(x), ϕ(x′) ⊗ 1C(T)) < σ2/4 for all x ∈ U1, (e 25.34) (e 25.35) (e 25.36) where x′ ⊗ 1C(T) = x, L2(a) = (1 − p)L(a)(1 − p) + L1(a) for all a ∈ A ⊗ C(T). Note that we also have kϕ(g) − L2(g ⊗ 1)k < σ/2 for all g ∈ G′4 and [L2A]P1 = [ϕ]P1 . By (e 25.33) and the choice of F1 there is a unitary v0 ∈ CU (C) and a unitary v00 ∈ CU ((1 − p)B(1 − p)) such that kL1(1 ⊗ z) − v0k < min{δ2/2, ε/8} and k(1 − p)L(1 ⊗ z)(1 − p) − v00k < min{δ2/2, ε/8}. By applying 25.3, we obtain a unitary w ∈ U (U2) = U0(U2) = CU (U2) such that t(L3(g))) − t(Φ(g)) < σ1, t(g(1 ⊗ w)) ≥ 1 2ZT g ∈ H2, and gdm for all g ∈ H1 (e 25.37) (e 25.38) (e 25.39) (e 25.40) (e 25.41) for all t ∈ T (B ⊗ U2), where L3 : A ⊗ C(T) → B ⊗ U2 is a unital G1-δ1/4-completely positive linear map defined by L3(a ⊗ 1) = L2(a ⊗ 1) ⊗ 1U2 and L3(a ⊗ zj) = L2(a ⊗ zj) ⊗ (w)j (e 25.42) for all a ∈ A and all integers j as given by 25.2, w(1−p) = (1−p)w = (1−p), as we consider both w, 1−p as elements in B⊗U2 as 1B⊗w and (1−p)⊗1U2, respectively, and Φ : A⊗C(T) → B⊗U2 is defined by Φ(a ⊗ 1) = ϕ(a) ⊂ B ⊗ 1 for all a ∈ A and Φ(1 ⊗ f ) = f (w) for all f ∈ C(T). Moreover, Φ(1⊗ f ) = f (λ)((1− p)⊗ 1U2) + f (pw) for all f ∈ C(T) and for some λ ∈ T. Note that CU (U2) = U (U2). One obtains a continuous path of unitaries {v(t) : t ∈ [1/4, 1/2]} ⊂ CU (U2) such that v(1/4) = 1U2, v(1/2) = w and length({v(t) : t ∈ [1/4, 1/2]}) ≤ π + ε/256. (e 25.43) Note that ϕ(a)Φ(1 ⊗ z) = Φ(1 ⊗ z)ϕ(a) for all a ∈ A. So, in particular, Φ is a unital homomor- phism and [Φ]β(K(A)) = 0. 252 (e 25.44) Define a unital contractive completely positive linear map Lt : A2 → C([2, 3], B ⊗ U2) by Lt(f ⊗ 1) = L2(f ⊗ 1) and Lt(a ⊗ zj) = L2(a ⊗ zj) ⊗ (v((t − 2)/4 + 1/4))j for all a ∈ A and integers j and t ∈ [2, 3]. Moreover, Lt(1⊗ z) = (v0 ⊕ v00)⊗ v((t− 2)/4 + 1/4)), and, since v(s) ∈ CU (U2), Lt(1⊗ z) ∈ CU (B⊗ U2) for all t ∈ [2, 3]. Note, as in the proof of 25.2, that Lt are G1-δ1/4-multiplicative. Note at t = 2, Lt = L2 and at t = 3, Lt = L3. It follows that (e 25.45) [L3]β(P2) = 0 and [L3]P1 = [L2]P1 = [ϕ]P1 , L‡3(x) = L‡2(x) for all x ∈ U1. If v = (e ⊗ z) + (1 − e) for some projection e ∈ A, then L3(v) = L2(e ⊗ z) ⊗ w + L2((1 − e)). Since w ∈ CU (U2), one computes from (e 25.26) that that, with x = ((1− pi) + pi ⊗ z)((1− qi) + qi ⊗ z∗), hL3(x)i ≈σ2/210 (¯z(1) i ⊗ w + (1 − ¯pi))(¯z(2) i ⊗ w + (1 − ¯qi)) (e 25.48) (e 25.46) (e 25.47) = (¯z(1) = (¯z(1) i + (1 − ¯pi))(¯pi ⊗ w + (1 − ¯pi) ⊗ 1U2)(¯z(2) i + (1 − ¯pi))(¯z(2) i + (1 − ¯qi)) = hL2(x)i, , ¯z(2) i where ¯pi, ¯qi, ¯z(1) that i i + (1 − ¯qi))(¯qi ⊗ w + (1 − ¯qi))(e 25.49) (e 25.50) are as above (see the lines below (e 25.26)), replacing Ψ1 by L2. It follows dist(L‡3(x), ¯1) < σ2/2 for all x ∈ {1 ⊗ z} ∪ U2. (e 25.51) Note that, since w ∈ CU (U2) and ϕ(q) ∈ B ⊗ 1U2, Φ(q ⊗ z + (1 − q) ⊗ 1) = ϕ(q) ⊗ w + ϕ(1 − q) ∈ CU (B ⊗ U2) (e 25.52) for any projection q ∈ A. It follows that Φ‡(x) ∈ CU (B ⊗ U2) for all x ∈ {1 ⊗ z} ∪ U2. Therefore (see also (e 25.44)) [L3]P = [Φ]P and dist(Φ‡(x), L‡3(x)) < σ2 for all x ∈ U . It follows from (e 25.41) that and it follows from (e 25.40) that τ (Φ(f )) ≥ ∆(f ), f ∈ H1, τ ∈ T (B ⊗ U2), τ (Φ(f )) − τ (L3(f )) < σ1, f ∈ H2, τ ∈ T (B ⊗ U2). By applying 12.9, we obtain a unitary w1 ∈ B ⊗ U2 such that kw∗1Φ(f )w1 − L3(f )k < ε/4 for all f ∈ F1. (e 25.53) (e 25.54) (e 25.55) (e 25.56) (e 25.57) Since w ∈ U2, there is a continuous path of unitaries {w(t) : t ∈ [3/4, 1]} ⊂ CU (U2) such that w(3/4) = Φ(1 ⊗ z) = w, w(1) = 1U2 and length({w(t) : t ∈ [3/4, 1]}) ≤ π + ε/256. (e 25.58) 253 Note that (e 25.59) It follows from (e 25.57) that there exists a continuous path of unitaries {u(t) : t ∈ [1/2, 3/4]} ⊂ B ⊗ U2 such that Φ(a)w(t) = w(t)Φ(a) for all a ∈ A and t ∈ [3/4, 1]. u(1/2) = (v00 + (v0)) ⊗ w, u(3/4) = w∗1Φ(1 ⊗ z)w1 and ku(t) − u(1/2)k < ε/4 for all t ∈ [1/2, 3/4]. (e 25.60) (e 25.61) It follows from (e 25.29) and (e 25.39) that there exists a continuous path of unitaries {u(t) : t ∈ [0, 1/4]} ⊂ B such that Now define u(0) = u, u(1/4) = v00 + v0 and ku(t) − uk < ε/4 for all t ∈ [0, 1/4]. u(t) = w∗1w(t)w1 for all t ∈ [3/4, 1]. (e 25.62) (e 25.63) (e 25.64) Then {u(t) : t ∈ [0, 1]} ⊂ B ⊗ U2 is a continuous path of unitaries such that u(0) = u and u(1) = 1. Moreover, by (e 25.57), (e 25.58), (e 25.63), (e 25.43), (e 25.58) and (e 25.61), kϕ(f )u(t) − u(t)ϕ(f )k < ε for all f ∈ F and length({u(t)}) ≤ 2π + ε. (e 25.65) Remark 25.5. Let A be a unital simple separable amenable C∗-algebra with stable rank one. Let G0 ⊂ K0(A) be a finitely generated subgroup containing [1A]. Let Gr = ρA(G0). Then ρA([1A]) 6= 0 and Gr is a finitely generated free group. Then we may write G0 = G0∩kerρA⊕G′r, where ρA(G′r) = Gr and G′r ∼= Gr. Note that G0∩kerρA is a finitely generated subgroup. We may write G0 ∩ kerρA = G00 ⊕ G01, where G00 is a torsion group and G01 is free. Note that G01 ⊕ G′r is free. Therefore G0 = T or(G0) ⊕ F, where F is a finitely generated free subgroup. Note that there is an integer m ≥ 1 such that m[1A] ∈ F. Let z ∈ C(T) be the standard unitary generator. Consider A ⊗ C(T). Then β(G0) ⊂ β(K0(A)) is a subgroup of K1(A ⊗ C(T)). Moreover β([1A]) may be identified with [1 ⊗ z]. If we choose U2 in the above proof which generates β(F ), then, for any σ1 > 0, we may assume that dist(um, ¯1) < σ1/m (e 25.66) provided that (e 25.11) holds for a sufficiently small σ. Since we assume that u ∈ U0(B) as P may be large enough in (e 25.10), by (e 25.66), dist(¯u, ¯1) < σ1. This implies that (with sufficiently small σ) the condition (e 25.12) is redundant and therefore can be omitted. 26 Stably results Lemma 26.1. Let C be a unital separable C∗-algebra which is residually finite dimensional and satisfies the UCT. For any ε > 0, any finite subset F ⊂ C, any finite subset P ⊂ K(C), any unital homomorphism h : C → A, where A is any unital C∗-algebra, and any κ ∈ HomΛ(K(SC), K(A)), there exists an integer N ≥ 1, a unital homomorphism h0 : C → MN (C) ⊂ MN (A) and a unitary u ∈ U (MN +1(A)) such that kH(c), u]k < ε for all c ∈ F and Bott(H, u)P = κ, (e 26.1) where H(c) = daig(h(c), h0(c)) for all c ∈ C. 254 Proof. Define S = {z, 1C(T)}, where z is identity function on the unit circle. Define x ∈ HomΛ(K(C ⊗ C(T)), K(A)) as follows: xK(C) = [h] and xβ(K(C)) = κ. (e 26.2) Fix a finite subset P1 ⊂ β(K(C)). Choose ε1 > 0 and a finite subset F1 ⊂ C satisfying the following: [L′]P1 = [L′′]P1 (e 26.3) for any pair of (F1 ⊗ S) − ε1-multiplicative contractive completely positive linear maps L′, L′′ : C ⊗ C(T) → B (for any unital C∗-algebra B), provided that L′ ≈ε1 L′′ on F1 ⊗ S. (e 26.4) Let a positive number ε > 0, a finite subset F and a finite subset P ⊂ K(C) be given. We may assume, without loss of generality, that Bott(H′, u′)P = Bott(H′, u′′)P (e 26.5) provided ku′ − u′′k < ε for any unital homomorphism H′ from C. Put ε2 = min{ε/2, ε1/2} and F2 = F ∪ F1. Let δ > 0, G ⊂ C be a finite subset and P0 ⊂ K(C) (in place of P) be as required by 4.16 for ε2/2 (in place of ε) and F2 (in place of F). Without loss of generality, we may assume that F2 and G are in the unit ball of C and δ < min{1/2, ε2/16}. Fix another finite subset P2 ⊂ K(C) and defined P3 = P0 ∪ β(P2) (as a subset of K(C ⊗ C(T))). We may assume that P1 ⊂ β(P2). It follows from 18.2 that there are integer N1 ≥ 1, a unital homomorphism h1 : C ⊗ C(T) → MN1(C) ⊂ MN1(A) and a (G ⊗ S)-δ/2-multiplicative contractive completely positive linear map L : C ⊗ C(T) → MN1+1(A) such that [L]P3 = (x + [h1])P3 . We may assume that there is a unitary v0 ∈ MN1+1(A) such that kL(1 ⊗ z) − v0k < ε2/2. Define H1 : C → MN1+1(A) by H1(c) = h(c) ⊕ h1(c ⊗ 1) for all c ∈ C. Define L1 : C → MN1+1(A) by L1(c) = L(c ⊗ 1) for all c ∈ C. Note that [L1]P0 = [H1]P0 . (e 26.6) (e 26.7) (e 26.8) (e 26.9) It follows from 4.16 that there exists an integer N2 ≥ 1, a unital homomorphism h2 : C → MN2(N1+1)(C)⊂ MN2(N1+1)(A) and a unitary W ∈ M(N2+1)(1+N1)(A) such that W ∗(L1(c) ⊕ h2(c))W ≈ε/4 H1(c) ⊕ h2(c) for all c ∈ F2. Put N = N2(N1 + 1) + N1. Now define h0 : C → MN (C) and H : C → MN +1(A) by h0(c) = h1(c ⊗ 1) ⊕ h2(c) and H(c) = h(c) ⊕ h0(c) (e 26.10) (e 26.11) 255 for all c ∈ C. Define u = W ∗(v0 ⊕ 1MN2(N1+1))W. Then, by (e 26.10), and L1 being (G ⊗ S)-δ/2- multiplicative, we have k[H(c), u]k ≤ k(H(c) − Ad W ◦ (L1(c) ⊕ h2(c)))u]k +kAd W ◦ (L1(c) ⊕ h2(c)), u]k + ku(H(c) − Ad W ◦ (L1(c) ⊕ h2(c)))k < ε/4 + δ/2 + ε/4 < ε for all c ∈ F2. (e 26.12) (e 26.13) (e 26.14) Define L2 : C → MN +1(A) by L2(c) = L1(c) ⊕ h2(c) for all c ∈ C. Then, we compute that Bott(H, u)P = Bott(Ad W ◦ L2, u)P = Bott(L2, v0 ⊕ 1MN2(N1+1))P = Bott(L1, v0)P + Bott(h2, 1MN2(N1+1))P = [L]β(P) + 0 = (x + [h])β(P) = κP . (e 26.15) (e 26.16) (e 26.17) Theorem 26.2. Let C be a unital amenable separable C∗-algebra which is residually finite dimensional and satisfies the UCT. For any ε > 0 and any finite subset F ⊂ C, there is δ > 0, a finite subset G ⊂ C, a finite subset P ⊂ K(C) satisfying the following: Suppose that A is a unital C∗-algebra, suppose h : C → A is a unital homomorphism and suppose that u ∈ U (A) is a unitary such that k[h(a), u]k < δ for all a ∈ G and Bott(h, u)P = 0. (e 26.18) Then there exists an integer N ≥ 1 and a continuous path of unitaries {U (t) : t ∈ [0, 1]} in MN +1(A) such that U (0) = u′, U (1) = 1MN+1(A) and k[h′(a), U (t)]k < ε for all a ∈ F, (e 26.19) where u′ = diag(u, H0(1 ⊗ z)) and h′(f ) = h(f ) ⊕ H0(f ⊗ 1) for f ∈ C, where H0 : C ⊗ C(T) → MN (C) (⊂ MN (A)) is a unital homomorphism (with finite dimensional range) and z ∈ C(T) is the identity function on the unit circle. Moreover, Length({U(t)}) ≤ π + ε. (e 26.20) Proof. Let ε > 0 and F ⊂ C be given. Without loss of generality, we may assume that F is in the unit ball of C. Let δ1 > 0, G1 ⊂ C ⊗ C(T), P1 ⊂ K(C ⊗ C(T)) be required by 4.16 for ε/4 and F ⊗ S. Without loss of generality, we may assume that G1 = G′1⊗S, where G′1 is in the unit ball of C and S = {1C(T), z} ⊂ C(T). Moreover, without loss of generality, we may assume that P1 = P2 ∪P3, where P2 ⊂ K(C) and P3 ⊂ β(K(C)). Let P = P2 ∪ β−1(P3) ⊂ K(C). Furthermore, we may assume that any δ1-G1-multiplicative contractive completely positive linear map L′ from C ⊗ C(T) to a unital C∗-algebra well defines [L′]P1 . Let δ2 > 0 and G2 ⊂ C be a finite subset required by 2.8 of [66] for δ1/2 and G′1 above. Let δ = min{δ2/2, δ1/2, ε/2} and G = F ∪ G2. Suppose that h and u satisfy the assumption with above δ, G and P. Thus, by 2.8 of [66], there is δ1/2-G1-multiplicative contractive completely positive linear map L : C ⊗ C(T) → A such that kL(f ⊗ 1) − h(f )k < δ1/2 for all f ∈ G′1 kL(1 ⊗ z) − uk < δ1/2. (e 26.21) (e 26.22) 256 Define y ∈ HomΛ(K(C ⊗ C(T)), K(A)) as follows: yK(C) = [h]K(C) and yβ(K(C)) = 0. It follows from Bott(h, u)P = 0 that [L]β(P) = 0. Then [L]P1 = yP1. (e 26.23) Define H : C ⊗ C(T) → A by for all c ∈ C and g ∈ C(T), where T is identified with the unit circle (and 1 ∈ T). It follows that H(c ⊗ g) = h(c) · g(1) · 1A [H]P1 = yP1 = [L]P1 . (e 26.24) It follows from 4.16 that there is an integer N ≥ 1, a unital homomorphism H0 : C⊗C(T) → MN (C) (⊂ MN (A)) with finite dimensional range and a unitary W ∈ U (M1+N (A)) such that (e 26.25) W ∗(H(c) ⊕ H0(c))W ≈ε/4 L(c) ⊕ H0(c) for all c ∈ F ⊗ S. Since H0 has finite dimensional range and since H0(1 ⊗ z) is in the center of range(H0) it is easy to construct a continuous path {V ′(t) : t ∈ [0, 1]} in a finite dimensional ⊂ MN (C), C∗-subalgebra of MN (C) such that V ′(0) = H0(1 ⊗ z), V ′(1) = 1MN (A) and H0(c ⊗ 1)V ′(t) = V ′(t)H0(c ⊗ 1) for all c ∈ C and t ∈ [0, 1]. Moreover, Length({V ′(t)}) ≤ π. Now define U (1/4 + 3t/4) = W ∗diag(1, V ′(t))W for t ∈ [0, 1] and u′ = u ⊕ H0(1A ⊗ z) and h′(c) = h(c) ⊕ H0(c ⊗ 1) for c ∈ C for t ∈ [0, 1]. Then (e 26.26) (e 26.27) (e 26.28) ku′ − U (1/4)k < ε/4 and k[U (t), h′(a)]k < ε/4 (e 26.29) for all a ∈ F and t ∈ [1/4, 1]. The theorem follows by connecting U (1/4) with u′ with a short path as follows: There is a self-adjoint element a ∈ M1+N (A) with kak ≤ επ 8 such that exp(ia) = u′U (1/4)∗ (e 26.30) Then the path of unitaries U (t) = exp(i(1 − 4t)a)U (1/4) for t ∈ [0, 1/4) satisfy the above. Lemma 26.3. Let C be a unital separable C∗-algebra whose irreducible representations have bounded dimensions and B be a unital C∗-algebra with T (B) 6= ∅. Suppose ϕ1, ϕ2 : C → B are two unital monomorphisms such that [ϕ1] = [ϕ2] in KK(C, B), 257 Let θ : K(C) → K(Mϕ1,ϕ2) be the splitting map defined in (e 2.28). For any 1/2 > ε > 0, any finite subset F ⊂ C and any finite subset P ⊂ K(C), there are integers N1 ≥ 1, an ε/2-F-multiplicative contractive completely positive linear map L : C → M1+N1(Mϕ1,ϕ2), a unital homomorphism h0 : C → MN1(C), and a continuous path of unitaries {V (t) : t ∈ [0, 1 − d]} of M1+N1(B) for some 1/2 > d > 0, such that [L]P is well defined, V (0) = 1M1+N1 (B), [L]P = (θ + [h0])P , πt ◦ L ≈ε ad V (t) ◦ (ϕ1 ⊕ h0) on F for all t ∈ (0, 1 − d], πt ◦ L ≈ε ad V (1 − d) ◦ (ϕ1 ⊕ h0) on F for all t ∈ (1 − d, 1] and π1 ◦ L ≈ε ϕ2 ⊕ h0 on F, (e 26.31) (e 26.32) (e 26.33) (e 26.34) where πt : Mϕ1,ϕ2 → B is the point-evaluation at t ∈ (0, 1). Proof. Let ε > 0 and let F ⊂ C be a finite subset. Let δ1 > 0, G1 ⊂ C be a finite subset and P ⊂ K(C) be a finite subset required by 26.2 for ε/4 and F above. In particular, we assume that δ1 < δP (see 2.13). We may further assume that δ1 is sufficiently small such that Bott(Φ, U1U2U3)P = 3Xi=1 Bott(Φ, Ui)P , (e 26.35) provided that k[Φ, Ui]k < δ1, i = 1, 2, 3. Let ε1 = min{δ1/2, ε/4} and F1 = F∪G1. We may assume that F1 is in the unit ball of C. We may also assume that [L′]P is well defined for any ε1-F1-multiplicative contractive completely positive linear map from C to any unital C-algebra. Let δ2 > 0 and G ⊂ C be a finite subset and P1 ⊂ K(C) be finite subset required by 4.16 for ε1/2 and F1. We may assume that δ2 < δ1/2, G ⊃ F1 and P1 ⊃ P. We also assume that G is in the unit ball of C. It follows from 18.2 that there exists an integer K1 ≥ 1, a unital homomorphism h′0 : C → MK1(C) and a δ2/2-G-multiplicative contractive completely positive linear map L1 : C → MK1+1(Mϕ1,ϕ2) such that [L1]P1 = (θ + [h′0])P1 . Note that [π0] ◦ θ = [ϕ1] and [π1] ◦ θ = [ϕ2] and, for each t ∈ (0, 1), [πt] ◦ θ = [ϕ1]. (e 26.36) (e 26.37) By 4.16, we obtain an integer K0, a unitary V ∈ U (M1+K1+K0((C))) and a unital homomorphism h : C → MK0(C) such that ad V ◦ (πe ◦ L1 ⊕ h) ≈ε1/2 (id⊕h′0 ⊕ h) on F1, (e 26.38) where πe : Mϕ1,ϕ2 → C is the canonical projection. mapped to Mk(A) for any unital C∗ algebra A, without introducing a new notation.) (Here and below, we will identify a homomorphism mapped to Mk(C) as a homomorphism Write V00 = ϕ1(V ) and V ′00 = ϕ2(V ). The assumption that [ϕ1] = [ϕ2] implies that [V00] = [V ′00] in K1(B). By adding another h in (e 26.38), replacing K0 by 2K0 and replac- , if necessary, we may assume that V00 and V ′00 are in the same component ing V by V ⊕ 1MK0 of U (M1+K1+K0(B)). 258 One obtains a continuous path of unitaries {Z(t) : t ∈ [0, 1]} in M1+K1+K0(B) such that Z(0) = V00 and Z(1) = V ′00. (e 26.39) It follows that Z ∈ M1+K1+K0(Mϕ1,ϕ2). By replacing L1 by ad Z ◦ (L1 ⊕ h) and using a new h′0, we may assume that π0 ◦ L1 ≈ε1/2 ϕ1 ⊕ h′0 on F1 and π1 ◦ L1 ≈ε1/2 ϕ2 ⊕ h′0 on F1, i = 1, 2. (e 26.40) Define λ : C → M1+K1+K0(C) by λ(c) = diag(c, h′0(c)), where we also identify MK0+K1(C) is identity ) ◦ h′0 = h′0. Therefore, in this way, one with the scalar matrices in MK0+K1(C). In particular, since ϕi is unital, ϕi⊗idMK1+K0 on MK0+K1(C), i = 1, 2. Consequently (ϕi ⊗ idMK0+K1 may write ϕi(c) ⊕ h′0(c) = (ϕi ⊗ idMK0+K1+1) ◦ λ(c) for all c ∈ C. There is a partition 0 = t0 < t1 < ··· < tn = 1 such that πti ◦ L1 ≈δ2/8 πt ◦ L1 on G for all ti ≤ t ≤ ti+1, i = 1, 2, ..., n − 1. (e 26.41) By applying 4.16 again, we obtain an integer K2 ≥ 1, a unital homomorphism h00 : C → MK2(C), and a unitary Vti ∈ M1+K0+K1+K2(B) such that ad Vti ◦ (ϕ1 ⊕ h′0 ⊕ h00) ≈ε1/2 (πti ◦ L1 ⊕ h00) on F1. (e 26.42) Note that, by (e 26.41), (e 26.42) and (e 26.40), k[ϕ1 ⊕ h′0 ⊕ h00(a), Vti V ∗ti+1]k < δ2/4 + ε1 for all a ∈ F1. Denote by η−1 = 0 and ηk = kXi=0 Bott(ϕ1 ⊕ h′0 ⊕ h00, VtiV ∗ti+1)P , k = 0, 1, ..., n − 1. Now we will construct, for each i, unital homomorphism Fi : C → MJi(C) ⊂ MJi(B) and a (B) such that unitary Wi ∈ M1+K0+K1+K2+Pi k=1 Ji k[Hi, Wi]k < δ2/4 and Bott(Hi, Wi) = ηi−1, (e 26.43) where Hi = ϕ1 ⊕ h′0 ⊕ h00 ⊕ ⊕i homomorphism F1 : C → MJ1(C) and a unitary W0 ∈ U (M1+K0+K1+K2+J1(B)) such that k=1Fi, i = 1, 2, ..., n − 1. . It follows from 26.1 that there is an integer J1 ≥ 1, a unital Let W0 = 1M1+K0+K1+K2 k[H1(a), W1]k < δ2/4 for all a ∈ F1 and Bott(H1, W1) = η0, (e 26.44) where H1 = ϕ1 ⊕ h′0 ⊕ h00 ⊕ F1. Assume that, we have construct required Fi and Wi for i = 0, 1, ..., k < n− 1. It follows from 26.1 that there is an integer Jk+1 ≥ 1, a unital homomorphism Fk+1 : C → MJk+1(C) and a unitary Wk+1 ∈ U (M1+K0+K1+K2+Pk+1 i=1 Ji (B)) such that k[Hk+1(a), Wk+1]k < δ2/4 for all a ∈ F1 and Bott(Hk+1, Wk+1) = ηk (e 26.45) where Hk+1 = ϕ1 ⊕ h′0 ⊕ h00 ⊕ ⊕k+1 i=1 Fi. 259 Now define F00 = h00 ⊕ ⊕n−1 i=0 Fi and define K3 = 1 + K0 + K1 + K2 +Pn−1 vtk = diag(Wkdiag(Vtk , id1M i=1 Ji. Define ), 1M ), Pn−1 i=k+1 Ji Pk i=1 Ji (e 26.46) (e 26.47) (e 26.48) (e 26.49) (e 26.50) ), i = Pn−1 i=1 Ji k = 1, 1, ..., n − 1 and vt0 = 1M 1+K0+K1+K2+Pn−1 i=1 Ji . Then ad vti ◦ (ϕ1 ⊕ h′0 ⊕ F00) ≈δ2+ε1 πti ◦ (L1 ⊕ F00) on F1, k[ϕ1 ⊕ h′0 ⊕ F00(a), vtiv∗ti+1 ]k < δ2/2 + 2ε1 for all a ∈ F1 and Bott(ϕ1 ⊕ h′0 ⊕ F00, vti v∗ti+1) = Bott(ϕ′1, W ′i ) + Bott(ϕ′1, V ′ti(V ′ti+1)∗) + Bott(ϕ′1, (W ′i+1)∗) = ηi−1 + Bott(ϕ′1, VtiV ∗ti+1) − ηi = 0, where ϕ′1 = ϕ1 ⊕ h′0 ⊕ F00, W ′i = diag(Wi, 1M 0, 1, 2, ..., n − 2. MN1(C) and a continuous path of unitaries {wi(t) : t ∈ [ti−1, ti]} such that It follows from 26.2 that there is an integer N1 ≥ 1, another unital homomorphism F ′0 : C → ) and V ′ti = diag(Vti , 1M Pn−1 j=i+1 Ji wi(ti−1) = v′i−1(v′i)∗, wi(ti) = 1, i = 1, 2, ..., n − 1 and k[ϕ1 ⊕ h′0 ⊕ F00 ⊕ F ′0(a), wi(t)]k < ε/2 for all a ∈ F, (e 26.51) (e 26.52) i = 1, 2, ..., n − 1, where v′i = diag(vi, 1MN1 t ∈ [ti−1, ti], i = 1, 2, ..., n − 1. Then V (t) ∈ C([0, tn−1], MN1(B)). Moreover, ad V (t) ◦ (ϕ1 ⊕ h′0 ⊕ F00 ⊕ F ′0) ≈ε/2 πt ◦ L1 ⊕ F00 ⊕ F ′0 on F. (B)), i = 1, 2, ..., n − 1. Define V (t) = wi(t)v′i for (e 26.53) Define h0 = h′0 ⊕ F00 ⊕ F ′0, L = L1 ⊕ F00 + F ′0 and d = 1− tn−1. Then, by (e 26.53), (e 26.32) and (e 26.33) hold. From (e 26.40), (e 26.34) also holds. 27 Asymptotically unitary equivalence Lemma 27.1. Let C1 and A1 be two unital separable simple C∗-algebras in B1, let U1 and U2 be two UHF-algebras of infinite type and let C = C1 ⊗ U1 and A = A1 ⊗ U2. Suppose that ϕ1, ϕ2 : C → A are two unital monomorphisms. Suppose also that [ϕ1] = [ϕ2] in KL(C, A), (ϕ1)T = (ϕ2)T and ϕ‡1 = ϕ‡2. (e 27.1) (e 27.2) Then ϕ1 and ϕ2 are approximately unitarily equivalent. Proof. This follows immediately from 12.9. Note that both A and C are in B1. Lemma 27.2. Let B be a unital C∗-algebra and let u1, u2, ..., un ⊂ U (B) be unitaries. Suppose that v1, v2, ..., vm ⊂ U (B) are also unitaries such that [vj] ⊂ G, where G is the subgroup of K1(B) generated by [u1], [u2], ..., [un]. There exists δ > 0 satisfying the following: For any unital C∗-algebra A and any unital homomorphisms ϕ1, ϕ2 : B → A, if there is a unitary w ∈ U (B) such that kw∗ϕ1(ui)w − ϕ2(ui)k < δ, (e 27.3) 260 then there exists a group homomorphism α : G → Aff(T (A)) such that τ (log(ϕ2(vj)w∗ϕ1(v∗j )w) = α([vj ])(τ ), 1 2πi (e 27.4) any τ ∈ T (A), i = 1, 2, ..., n and j = 1, 2, ..., m. Proof. The proof is essentially contained in the proof of 6.1, 6.2 and 6.3 of [64]. Lemma 27.3. Let C1 be a unital simple C∗-algebra in 14.8, let A1 be a unital separable simple C∗-algebra in B0, and let U1 and U2 be two UHF-algebras of infinite type. Let C = C1 ⊗ U1 and A = A1 ⊗ U2. Suppose that ϕ1, ϕ2 : C → A are two unital monomorphisms. Suppose also that (e 27.5) [ϕ1] = [ϕ2] in KL(C, A), ϕ‡1 = ϕ‡2, (ϕ1)T = (ϕ2)T and Rϕ,ψ(K1(Mϕ1,ϕ2))⊂ρA(K0(A)). (e 27.6) (e 27.7) Then, for any increasing sequence of finite subsets {Fn} of C whose union is dense in C, any increasing sequence of finite subsets Pn of K1(C) with ∪∞n=1Pn = K1(C) and any decreasing sequence of positive numbers {δn} withP∞n=1 δn < ∞, there exists a sequence of unitaries {un} in U (A) such that Adun ◦ ϕ1 ≈δn ϕ2 on Fn and ρA(bott1(ϕ2, u∗nun+1)(x)) = 0 for all x ∈ Pn (e 27.8) (e 27.9) and for all sufficiently large n. Proof. Note that A ∼= A ⊗ U2. Moreover, there is a unital homomorphism s : A ⊗ U2 → A such that s ◦ ı is approximately unitarily equivalent to the identity map on A, where ı : A → A ⊗ U2 defined by a → a ⊗ 1U2 for all a ∈ A. Therefore we may assume that ϕ1(C), ϕ2(C) ⊂ A ⊗ 1U2. It follows from 27.1 that there exists a sequence of unitaries {vn} ⊂ A such that lim n→∞ Ad vn ◦ ϕ1(c) = ϕ2(c) for all c ∈ C. We may assume that Fn are in the unit ball and ∪∞n=1Fn is dense in the unit ball of C. Put ε′n = min{1/2n+1, δn/2}. Let Cn ⊂ C be a unital C∗-subalgebra (in place of Cn) which has finitely generated Ki(Cn) (i = 0, 1), and let Qn be a finite set of generators of K1(Cn), let δ′n > 0 (in place of δ) be as in 24.2 for C (in place of A), ε′n (in place of ε), Fn (in place of F) and [ın](Qn−1) (in place of P), where ın : Cn → C is the embedding. Note that, we assume that (e 27.10) [ın+1](Qn+1) ⊃ Pn+1 ∪ [ın](Qn). Write K1(Cn) = Gn,f ⊕ Tor(K1(Cn)), where Gn,f is a finitely generated free abelian group. Let z1,n, z2,n, ..., zf (n),n be the free generators of Gn,f and z′1,n, z′2,n, ..., z′t(n),n be generators of Tor(K1(Cn)). We may assume that Qn = {z1,n, z2,n, ..., zf (n),n, z′1,n, z′2,n, ..., z′t(n),n}. Let 1/2 > ε′′n > 0 so that bott1(h′, u′)K1(Cn) is well defined group homomorphism, bott1(h′, u′)Qn is well defined and (bott1(h′, u′)K1(Cn))Qn = bott1(h′, u′)Qn for any unital homomorphism h′ : C → A and any unitary u′ ∈ A for which k[h′(c), u′]k < ε′′n for all c ∈ Gn (e 27.11) 261 for some finite subset Gn ⊂ C which contains Fn. Let w1,n, w2,n, ..., wf (n),n, w′1,n, w′2,n, ..., w′t(n),n ∈ C be unitaries (note that C has stable rank one) such that [wi,n] = (ın)∗1(zi,n) and [w′j,n] = (ın)∗1(z′j,n), i = 1, 2, ..., f (n), j = 1, 2, ..., t(n) and n = 1, 2, .... To simplify notation, without loss of generality, we may assume that wi,n ∈ Gn, n = 1, 2, .... Let δ′′1 = 1/2 and, for n ≥ 2, let δ′′n > 0 (in place of δ) be as in 27.2 associated with w1,n, w2,n, ..., wf (n),n, w′1,n, w′2,n, ..., w′t(n),n (in place of u1, u2, .., un) and {w1,n−1, w2,n−1, ..., wf (n−1),n−1, w′1,n−1, w′2,n−1, ..., w′t(n−1),n−1} (in place of v1, v2, ..., vm). Put εn = min{ε′′n/2, ε′n/2, δ′n, δ′′n/2}. We may assume that Ad vn ◦ ϕ1 ≈εn ϕ2 on Gn, n = 1, 2, .... Thus bott1(ϕ2 ◦ ın, v∗nvn+1) is well defined. Since Aff(T (A)) is torsion free, τ(cid:0)bott1(ϕ2 ◦ ın, v∗nvn+1)Tor(K1(Cn))(cid:1) = 0. We have (e 27.12) (e 27.13) kϕ2(wj,n)Ad vn(ϕ1(wj,n)∗) − 1k < (1/4) sin(2πεn) < εn, n = 1, 2, .... (e 27.14) Define hj,n = 1 2πi log(ϕ2(wj,n)Ad vn(ϕ1(wj,n)∗)), j = 1, 2, ..., f (n), n = 1, 2, ... (e 27.15) Then, for any τ ∈ T (A), τ (hj,n) < εn < δ′n, j = 1, 2, ..., f (n), n = 1, 2, .... Since Aff(T (A)) is torsion free, it follows from 27.2 that τ ( 1 2πi log(ϕ2(w′j,n)Ad vn(ϕ1(w ′∗j,n)))) = 0, (e 27.16) j = 1, 2, ..., t(n) and n = 1, 2, .... By the assumption that Rϕ1,ϕ2(K1(Mϕ1,ϕ2))⊂ρA(K0(A)), by the Exel’s formula (see [43]) and by Lemma 3.5 of [65], we conclude that dhj,n(τ ) = τ (hj,n) ∈ Rϕ1,ϕ2(K1(Mϕ1,ϕ2))⊂ρA(K0(A)). Now define α′n : K1(Cn) → ρA(K0(A)) by α′n(zj,n)(τ ) =dhj,n(τ ) = τ (hj,n), j = 1, 2, ..., f (n) and α′n(z′j,n) = 0, j = 1, 2, ..., t(n),(e 27.17) n = 1, 2, .... Since α′n(K1(Cn)) is free, it follows that there is a homomorphism α(1) K0(A) such that n : K1(Cn) → (ρA ◦ α(1) n (zj,n))(τ ) = τ (hj,n), j = 1, 2, ..., f (n), τ ∈ T (A) and n (z′j,n) = 0, j = 1, 2, ..., t(n). α(1) (e 27.18) (e 27.19) Define α(0) κnKi(Cn) = α(i) Ki+1(SCn)). n : K0(Cn) → K1(A) by α(0) n = 0. By the UCT, there is κn ∈ KL(SCn, A) such that n , i = 0, 1, where SCn is the suspension of Cn (here, we also identify Ki(Cn) with 262 By the UCT again, there is αn ∈ KL(Cn ⊗ C(T), A) such that αn ◦ βK(Cn) = κn. In n . It follows from 24.2 that there exists a unitary Un ∈ U0(A) such particular, αn◦ βK1(Cn) = α(1) that k[ϕ2(c), Un]k < ε′′n for all c ∈ Fn and ρA(bott1(ϕ2, Un)(zj,n)) = −ρA ◦ αn(zj,n), (e 27.20) (e 27.21) j = 1, 2, ..., f (n). We also have ρA(bott1(ϕ2, Un)(z′j,n)) = 0, j = 1, 2, ..., t(n). (e 27.22) By the Exel trace formula (see [43]), (e 27.18) and (e 27.21), we have τ (hj,n) = −ρA(bott1(ϕ2, Un)(zj,n)(τ ) = −τ ( 1 2πi log(Unϕ2(wj,n)U∗nϕ2(w∗j,n)))(e 27.23) for all τ ∈ T (A), j = 1, 2, ..., f (n). Define un = vnUn, n = 1, 2, .... By 6.1 of [64], (e 27.23) and (e 27.21), we compute that τ ( 1 2πi = τ ( = τ ( log(Unϕ2(wj,n)U∗nv∗nϕ1(w∗j,n)vn))) log(ϕ2(wj,n)Adun(ϕ1(w∗j,n))))) 1 2πi 1 2πi 1 2πi = τ ( log(Unϕ2(wj,n)U∗nϕ2(w∗j,n)))) + τ ( = ρA(bott1(ϕ2, Un)(zj,n))(τ ) + τ (hj,n) = 0 (e 27.24) (e 27.25) (e 27.26) log(Unϕ2(wj,n)U∗nϕ2(w∗j,n)ϕ2(wj,n)v∗nϕ1(w∗j,n)vn))) 1 2πi log(ϕ2(wj,n)v∗nϕ1(w∗j,n)vn))) (e 27.27) (e 27.28) for all τ ∈ T (A), j = 1, 2, ..., f (n) and n = 1, 2, .... By (e 27.16) and (e 27.22), τ ( 1 2πi log(ϕ2(w′j,n)Adun(ϕ1((w′j,n)∗)))) = 0, (e 27.29) j = 1, 2, ..., t(n) and n = 1, 2, .... Let bj,n = b′j,n = b′′j,n+1 = 1 2πi 1 2πi 1 2πi log(unϕ2(wj,n)u∗nϕ1(w∗j,n)), log(ϕ2(wj,n)u∗nun+1ϕ2(w∗j,n)u∗n+1un) and log(un+1ϕ2(wj,n)u∗n+1ϕ1(w∗j,n)). j = 1, 2, ..., f (n) and n = 1, 2, .... We have, by (e 27.29), τ (bj,n) = τ ( = τ ( 1 2πi 1 2πi log(unϕ2(wj,n)u∗nϕ1(wj,n))) log(ϕ2(w∗j,n)u∗nϕ1(w∗j,n)un)) = 0 (e 27.30) (e 27.31) (e 27.32) (e 27.33) (e 27.34) for all τ ∈ T (A), j = 1, 2, ..., f (n) and n = 1, 2, .... Note that τ (bj,n+1) = 0 for all τ ∈ T (A), j = 1, 2, ..., f (n + 1). It follows from 27.2 and (e 27.10) that τ (b′′j,n+1) = 0 for all τ ∈ T (A), j = 1, 2, ..., f (n), n = 1, 2, .... 263 Note also that une2πib′ j,nu∗n = e2πibj,n · e−2πib′′ j,n+1, j = 1, 2, ..., f (n). Thus, by 6.1 of [64], we compute that τ (b′j,n) = τ (bj,n) − τ (b′′j,n+1) = 0 for all τ ∈ T (A). (e 27.35) By the Exel formula (see [43]) and (e 27.35), ρA(bott1(ϕ2, u∗nun+1))(w∗j,n)(τ ) = τ ( 1 2πi log(u∗nun+1ϕ2(wj,n)u∗n+1unϕ2(w∗j,n)))(e 27.36) = τ ( 1 2πi log(ϕ2(wj,n)u∗nun+1ϕ2(w∗j,n)u∗n+1un)) = 0 for all τ ∈ T (A) and j = 1, 2, ..., f (n). Thus ρA(bott1(ϕ2, u∗nun+1)(wj,n)(τ ) = 0 for all τ ∈ T (A), j = 1, 2, ..., f (n) and n = 1, 2, .... We also have ρA(bott1(ϕ2, u∗nun+1)(w′j,n)(τ ) = 0 for all τ ∈ T (A), j = 1, 2, ..., f (n) and n = 1, 2, .... By 27.2, we have that ρA(bott1(ϕ2, u∗nun+1)(z) = 0 for all z ∈ Pn, n = 1, 2, .... (e 27.37) (e 27.38) (e 27.39) (e 27.40) Theorem 27.4. Let C1 be a unital simple C∗-algebra as in 14.8, let A1 be a unital separable simple C∗-algebra in B0, let C = C1 ⊗ U1 and let A = A1 ⊗ U2, where U1 and U2 are UHF- algebras of infinite type. Suppose that ϕ1, ϕ2 : C → A are two unital monomorphisms. Then they are asymptotically unitarily equivalent if and only if [ϕ1] = [ϕ2] in KK(C, A), ϕ‡ = ψ‡, (ϕ1)T = (ϕ2)T and Rϕ1,ϕ2 = 0. (e 27.41) (e 27.42) Proof. We will prove the “if ” part only. The “only if” part follows from 4.3 of [69]. Note C = C1 ⊗ U1 can be also regarded as a C∗-algebra in 14.8. Let C = limn→∞(Cn, ın) be as in 14.8, where ın : Cn → Cn+1 is injective homomorphism. Let Fn ⊂ C be an increasing sequence of subsets of C such that ∪∞n=1Fn is dense in C. Put Mϕ1,ϕ2 = {(f, c) ∈ C([0, 1], A)⊕C : f (0) = ϕ1(c) and f (1) = ϕ2(c)}. Since C satisfies the UCT, the assumption that [ϕ1] = [ϕ2] in KK(C, A) implies that the following exact sequence splits: 0 → K(SA) → K(Mϕ1,ϕ2) ⇋πe θ K(C) → 0 (e 27.43) for some θ ∈ Hom(K(C), K(A)), where πe : Mϕ1,ϕ2 → C is the projection to C defined in 2.17. Furthermore, since τ ◦ ϕ1 = τ ◦ ϕ2 for all τ ∈ T (A) and Rϕ1,ϕ2 = 0, we may also assume that (e 27.44) Rϕ1,ϕ2(θ(x)) = 0 for all x ∈ K1(C). By [17], we have that (K(Cn), [ın]) = K(C). lim n→∞ 264 (e 27.45) Since Ki(Cn) is finitely generated, there exists K(n) ≥ 1 such that HomΛ(FK(n)K(Cn), FK(n)K(A)) = HomΛ(K(Cn), K(A)) (e 27.46) (see also [17] for the notation Fm there). Let δ′n > 0 (in place of δ), σ′n > 0 (in place of σ), G′n ⊂ C (in place of G), {p′1,n, p′2,n, ..., p′I(n),n), q′1,n, q′2,n, ..., q′I(n),n} (in place of {p1, p2, ..., pk, q1, q2, ..., qk}), P′n ⊂ K(C) (in place of P) corresponding to 1/2n+2 (in place of ε) and Fn (in place of F) as required by 25.4 (see also 25.5). Note that, by the choice as in 25.4, we assume that G′u,n, the subgroup generated by {[p′i,n] − [q′i,n] : 1 ≤ i ≤ I(n)} is free. Without loss of generality, we may assume that G′n ⊂ ın,∞(Gn) and P′n ⊂ [ın,∞](Pn) for some finite subset Gn ⊂ Cn, and for some finite subset Pn ⊂ K(Cn), we may assume that p′i,n = ın,∞(pi,n) and q′i,n = ın,∞(qi,n) for some projections pi,n, qi,n ∈ Cn, i = 1, 2, ..., I(n). Without loss of generality we may assume that the subgroup Gn,u generated by {[pi,n]− [qi,n] : 1 ≤ i ≤ I(n)} is free and pi,n, qi,n ∈ Gn, n = 1, 2, ..., I(n). We may assume that Pn contains a set of generators of FK(n)K(Cn), Fn ⊂ G′n and δ′n < 1/2n+3. We may also assume that Bott(h′, u′)Pn is well defined whenever k[h′(a), u′]k < δ′n for all a ∈ G′n and for any unital homomorphism h′ from Cn and a unitary u′ in the target algebra. Note that Bott(h′, u′)Pn defines Bott(h′ u′). We further assume that Bott(h, u)Pn = Bott(h′, u)Pn (e 27.47) n h′ on G′n. We may also assume that δ′n is smaller than δ/16 for the δ defined provided that h ≈δ′ in 2.15 of [69] for Cn (in place of A) and Pn (in place of P). Let k(n) ≥ n (in place of n), η′n > 0 (in place of δ) and Qk(n) ⊂ K1(Ck(n)) be required by 24.5 for δ′k(n)/4 (in place of ε), ın,∞(Gk(n)) (in place of F), Pk(n) (in place of P) and {pi,n, qi,n, : i = 1, 2, ..., k(n)} (in place of {pi, qi : i = 1, 2, ..., k}), and σ′k(n)/16 (in place of σ). We may assume that Qk(n) form a generator set of K1(Ck(n)). Since P generates FK(n)K(Ck(n+1)), we may assume that Qn ⊂ Pk(n). For Cn, since Ki(Cn) (i = 0, 1) is finitely generated, by (e 27.46), we may further assume that [ık(n),∞] is injective on [ın,k(n)](K(Cn)), n = 1, 2, .... By passing to a subsequence, to simplify notation, we may also assume that k(n) = n + 1. Let δn = min{ηn, σ′n, δ′n/2}. By 27.3, there are unitaries vn ∈ U (A) such that Ad vn ◦ ϕ1 ≈δn+1/4 ϕ2 on ın,∞(Gn+1), ρA(bott1(ϕ2, v∗nvn+1))(x) = 0 k[ϕ2(c), v∗nvn+1]k < δn+1/2 for all a ∈ ın,∞(Gn+1) for all x ∈ (ın,∞)∗1(K1(Cn+1)) and (e 27.48) (e 27.49) (e 27.50) (note that K1(Cn+1) is finitely generated). Note that, by (e 27.47), we may also assume that Bott(ϕ1, vn+1v∗n)[ın,∞](Pn) = Bott(v∗nϕ1vn, v∗nvn+1)[ın,∞](Pn) = Bott(ϕ2, v∗nvn+1])[ın,∞](Pn). In particular, bott1(v∗nϕ1vn, v∗nvn+1)(x) = bott1(ϕ2, v∗nvn+1)(x) (e 27.51) (e 27.52) (e 27.53) for all x ∈ (ın,∞)∗1(K1(Cn+1)). By applying 10.4 and 10.5 of [65], without loss of generality, we may assume that the pair ϕ1 and vn defines an element γn ∈ HomΛ(K(Cn+1), K(Mϕ1,ϕ2)) and [π0] ◦ γn = [idCn+1] (see 10.4 and 10.5 of [65] for the definition of γn). Moreover, by 10.4 and 10.5 of [65], we may assume, without loss of generality, that τ (log(ϕ2 ◦ ın,∞(z∗j )vnϕ1 ◦ ın,∞(zj)vn)) < δn+1, (e 27.54) 265 j = 1, 2, ..., r(n), where {z1, z2, ..., zr(n)} ⊂ U (Mk(Cn+1)) which forms a set of generators of vn, vn, ..., vn). We may assume that zj ∈ Qn ⊂ Pn, j = K1(Cn+1) and where vn = diag( 1, 2, ..., r(n). Let Hn = [ın+1](K(Cn+1)). SinceSn=1[ın+1,∞](K(Cn)) = K(C) and [π0] ◦ γn = [idCn+1], we conclude that z k } { K(Mϕ1,ϕ2) = K(SA) + ∪∞n=1γn(Hn). Thus, by passing to a subsequence, we may further assume that γn+1(Hn) ⊂ K(SA) + γn+2(Hn+1), n = 1, 2, .... (e 27.55) (e 27.56) By identifying Hn with γn+1(Hn), we may write jn : K(SA) ⊕ Hn → K(SA) ⊕ Hn+1 for the inclusion in (e 27.56). By (e 27.55), the inductive limit is K(Mϕ1,ϕ2). From the definition of γn, we note that γn − γn+1 ◦ [ın+1] maps K(Cn+1) into K(SA). By 10.6 of [65], Γ(Bott(ϕ1, vnv∗n+1))Hn = (γn+1 − γn+2 ◦ [ın+2])Hn (see 10.4, 10.5 and 10.6 of [65] for the definition of Γ) gives a homomorphism ξn : Hn → K(SA). Put ζn = γn+1Hn. Then jn(x, y) = (x + ξn(y), [ın+2](y)) (e 27.57) for all (x, y) ∈ K(SA) ⊕ Hn. Thus we obtain the following diagram: Hn ↓[ın+2,∞] Hn+1 ↓[ın+3,∞] Hn+2 K(SA) ⊕ Hn k ւξn ↓[ın+2,∞] K(SA) ⊕ Hn+1 k ւξn+1↓[ın+3,∞] K(SA) ⊕ Hn+2 0 → K(SA) → 0 → K(SA) → 0 → K(SA) → → → → k k → 0 → 0 → 0 By the assumption that ¯Rϕ1,ϕ2 = 0, map θ also gives the following decomposition: Define θn = θ ◦ [ın+2,∞] and κn = ζn − θn. Note that kerRϕ1,ϕ2 = kerρA ⊕ K1(C). We also have that θn = θn+1 ◦ [ın+2]. ζn − ζn+1 ◦ [ın+2] = ξn. Since [π0] ◦ (ζn − θn)Hn = 0, κn maps Hn into K(SA). It follows that κn − κn+1 ◦ [ın+2] = ζn − θn − ζn+1 ◦ [ın+2] + θn+1 ◦ [ın+2] = ζn − ζn+1 ◦ [ın+2] = ξn. (e 27.58) (e 27.59) (e 27.60) (e 27.61) (e 27.62) It follows from 26.3 that there are integers N1 ≥ 1, a δn+1 4 -ın+1(Gn+1)-multiplicative contractive completely positive linear map Ln : ın,∞(Cn+1) → M1+N1(Mϕ1,ϕ2), a unital homomorphism h0 : ın+1,∞(Cn+1) → MN1(C), and a continuous path of unitaries {Vn(t) : t ∈ [0, 3/4]} of M1+N1(A) such that [Ln]P ′ is well defined, Vn(0) = 1M1+N1 (A), n+1 [Ln ◦ ın,∞]Pn = (θ ◦ [ın+1,∞] + [h0 ◦ ın+1,∞])Pn , 266 and πt ◦ Ln ◦ ın+1,∞ ≈δn+1/4 ad Vn(t) ◦ ((ϕ1 ◦ ın+1,∞) ⊕ (h0 ◦ ψn+1,∞)) on ın+1,∞(Gn+1) for all t ∈ (0, 3/4], πt ◦ Ln ◦ ın+1,∞ ≈δn+1/4 ad Vn(3/4) ◦ ((ϕ1 ◦ ın+1,∞) ⊕ (h0 ◦ ın+1,∞)) on ın+1,∞(Gn+1) for all t ∈ (3/4, 1), and π1 ◦ Ln ◦ ın+1,∞ ≈δn+1/4 ϕ2 ◦ ın+1,∞ ⊕ h0 ◦ ın+1,∞ on ın+1,∞(Gn+1), where πt : Mϕ1,ϕ2 → A is the point-evaluation at t ∈ (0, 1). Note that Rϕ1,ϕ2(θ(x)) = 0 for all x ∈ ın+1,∞(K1(Cn+1)). As computed in 10.4 of [65], τ (log((ϕ2(x) ⊕ h0(x)∗Vn(3/4)∗(ϕ1(x) ⊕ h0(x))Vn(3/4))) = 0 (e 27.63) for x = ın+1,∞(y), where y is in a set of generators of K1(Cn+1) and for all τ ∈ T (A). a homomorphism κn ∈ HomΛ(K(Cn+1), K(SA)). By (e 27.54) Define W ′n = daig(vn, 1) ∈ M1+N1(A). Then Bott((ϕ1 ⊕ h0) ◦ ın+1,∞, W ′n(Vn(3/4)∗) defines τ (log((ϕ2 ⊕ h0) ◦ ın+1,∞(zj)∗(W ′n)∗(ϕ1 ⊕ h0) ◦ ın+1,∞(zj)W ′n)) < δn+1, (e 27.64) j = 1, 2, ..., r(n). Put Vn = diag(Vn(3/4), 1). Let bj,n = b′j,n = b′′j,n = 1 2πi 1 2πi 1 2πi log( V ∗n (ϕ1 ⊕ h0)ın+1,∞(zj) Vn(ϕ2 ⊕ h0) ◦ ın+1,∞(zj)∗), log((ϕ1 ⊕ h0) ◦ ın+1,∞(zj ) Vn(W ′n)∗(ϕ1 ⊕ h0) ◦ ın+1,∞(zj)∗W ′n log((ϕ2 ⊕ h0)ın+1,∞(zj)(W ′n)∗(ϕ1 ⊕ h0) ◦ ın+1,∞(zj)∗W ′n), j = 1, 2, ..., r(n). By (e 27.63) and (e 27.64), τ (bj,n) = 0 and τ (b′′j,n) < δn+1 for all τ ∈ T (A). Note that V ∗n e2πib′ j,n Vn = e2πibj,n e2πib′′ j,n. Then, by 6.1 of [64] and by (e 27.68) τ (b′j,n) = τ (bj,n) − τ (b′′j,n) = τ (b′′j,n) for all τ ∈ T (A). It follows from this and (e 27.51) that ρA(κn(zj))(τ ) < δn+1, j = 1, 2, .... for all τ ∈ T (A). It follows from 24.5 that there is a unitary w′n ∈ U (A) such that k[ϕ1(a), w′n]k < δ′n+1/4 for all a ∈ ın+1,∞(Gn+1) and Bott(ϕ1 ◦ ın+1,∞, w′n) = −κn ◦ [ın+1]. By (e 27.47), Bott(ϕ2 ◦ ın+1,∞, v∗nw′nvn)Pn = −κn ◦ [ın+1]Pn . 267 (e 27.65) V ∗n ) and (e 27.66) (e 27.67) (e 27.68) (e 27.69) (e 27.70) (e 27.71) (e 27.72) (e 27.73) (e 27.74) Put wn = v∗nw′nvn. It follows from 10.6 of [65] that Γ(Bott(ϕ1 ◦ ın+1,∞, w′n)) = −κn ◦ [ın+1] and Γ(Bott(ϕ1 ◦ ın+2,∞, w′n+1)) = −κn+1 ◦ [ın+2]. We also have Γ(Bott(ϕ1 ◦ ın+1,∞, vnv∗n+1))Hn = ζn − ζn+1 ◦ [ın+2] = ξn. But, by (e 27.61) and (e 27.62), (−κn + ξn + κn+1 ◦ [ın+2]) = 0. (e 27.75) (e 27.76) (e 27.77) (e 27.78) By 10.6 of [65], Γ(Bott(., .)) = 0 if and only if Bott(., .) = 0. Thus, by (e 27.74), (e 27.75) and (e 27.77), − Bott(ϕ1 ◦ ın+1,∞, w′n) + Bott(ϕ1 ◦ ın+1,∞, vnv∗n+1) + Bott(ϕ1 ◦ ın+1,∞, w′n+1) = 0. (e 27.79) Define un = xnvnw∗n, n = 1, 2, .... Then, by (e 27.48) and (e 27.72), ad un ◦ ϕ1 ≈δ′ n/2 ϕ2 for all a ∈ ın+1,∞(Gn+1). (e 27.80) From (e 27.51), (e 27.47) and (e 27.79), we compute that Bott(ϕ2 ◦ ın+1,∞, u∗nun+1) = Bott(ϕ2 ◦ ın+1,∞, wnv∗nvn+1w∗n+1) (e 27.81) = Bott(ϕ2 ◦ ın+1,∞, wn) + Bott(ϕ2 ◦ ın+1,∞, v∗nvn+1) + Bott(ϕ2 ◦ ın+1,∞, w∗n+1)(e 27.82) = Bott(ϕ1 ◦ ın+1,∞, w′n) + Bott(ϕ1 ◦ ın+1,∞, (w′n+1)∗) (e 27.83) = −[−Bott(ϕ1 ◦ ın+1,∞, w′n) + Bott(ϕ1 ◦ ın+1,∞, vnv∗n+1) (e 27.84) +Bott(ϕ1 ◦ ın+1,∞, w′n+1)] = 0. (e 27.85) Let xi,n = [pi,n]−[qi,n], 1 ≤ i ≤ I(n). Note that we assume that Gu,n is a free group generated by {xi,n : 1 ≤ i ≤ I(n)}. Without loss of generality, we may assume that these are free generators. Define, for each n ≥ 1, a homomorphism Λn : Gu,n → U (A)/CU (A) by Λn(xi,n) = (h((1 − ei,n) + ei,nun)((1 − e′i,n) + e′i,nu∗n)i, where ei,n = ϕ2 ◦ ın+1,∞(pi,n), e′i,n = ϕ2 ◦ ın+1,∞(qi,n), i = 1, 2, ..., I(n). In what follows, we will construct unitaries s1, s2, ..., sn, ... in A such that [ϕ2 ◦ ιn+1,∞(f ), sn] < δ′n+1/4 for all f ∈ Gn+1, Bott(ϕ2 ◦ ιn+1,∞, sn)Pn = 0 and dist(h((1 − ei,n) + ei,nsn)((1 − e′i,n) + e′i,ns∗n)i, Λn(−xi,n)) < σ′n/16, (e 27.86) (e 27.87) (e 27.88) Let s1 = 1, and assume that s2, s3, ..., sn are already constructed. Let us construct sn+1. Note that by (e 27.81), the K1 class of the unitary u∗nun+1 is trivial. In particular, the K1 class of snu∗nun+1 is trivial. Since Λ factors through G′u,n, applying Theorem 24.5 to ϕ2 ◦ ιn+2,∞, one obtains a unitary sn+1 ∈ B such that [ϕ2 ◦ ιn+2,∞(f ), sn] < δ′n+1/4 for all f ∈ Gn+2, Bott(ϕ2 ◦ ιn+2,∞, sn+1)Pn = 0 and (e 27.89) (e 27.90) dist(h((1 − ei,n+1) + ei,n+1sn+1)((1 − e′i,n+1) + e′i,n+1s∗n+1)i, Λn+1(−xi,n+1)) < σ′n/16, (e 27.91) 268 i = 1, 2, ..., I(n + 1). Then s1, s2, ..., sn+1 satisfies (e 27.86), (e 27.87) and (e 27.88). Putfun = uns∗n. Then by (e 27.80) and (e 27.86), one has n ϕ2 for all a ∈ ın+1,∞(Gn+1). By (e 27.81) and (e 27.87), one has adfun ◦ ϕ1 ≈δ′ Note that Bott(ϕ2 ◦ ιn+1,∞, (fun)∗]un+1)Pn = 0. (e 27.92) (e 27.93) h(1 − ei,n) + ei,nsn+1un+1ih(1 − e′i,n) + ei,n+1u∗n+1s∗n+1i = c1c2c3c4 = c1c4c2c3, (e 27.94) where c1 = h(1 − ei,n+1) + ei,n+1sn+1i, c2 = h(1 − ei,n+1) + ei,n+1un+1i, c3 = h(1 − e′i,n+1) + e′i,n+1u∗n+1i, c4 = h(1 − e′i,n+1) + e′i,n+1s∗n+1i. Therefore, by (e 27.91), one has dist(h((1 − ei,n+1) + ei,n+1]un+1)((1 − e′i,n+1) + e′i,n+1]un+1∗)i), ¯1) < σ′n+1/16 + dist(Λ(−xi,n+1)Λ(xi,n+1), ¯1) = σ′n+1/16, (e 27.95) (e 27.96) (e 27.97) (e 27.98) i = 1, 2, ..., I(n). Therefore, by 25.4 (and Remark 25.5), there exists a piece-wise smooth and continuous path of unitaries {zn(t) : t ∈ [0, 1]} of A such that zn(0) = 1, zn(1) = (fun)∗]un+1 and k[ϕ2(a), zn(t)]k < 1/2n+2 for all a ∈ Fn and t ∈ [0, 1]. (e 27.99) (e 27.100) Define Note that u(n) = ]un+1 for all integer n and {u(t) : t ∈ [0,∞)} is a continuous path of unitaries in A. One estimates that, by (e 27.80) and (e 27.100), u(t + n − 1) =funzn+1(t) t ∈ (0, 1]. ad u(t + n − 1) ◦ ϕ1 ≈δ′ n ad zn+1(t) ◦ ϕ2 ≈1/2n+2 ϕ2 on Fn (e 27.101) for all t ∈ (0, 1). It then follows that lim t→∞ u∗(t)ϕ1(a)u(t) = ϕ2(a) for all a ∈ C. (e 27.102) 28 Rotation maps and strong asymptotic equivalence Lemma 28.1. Let A be a unital separable simple C∗-algebra of stable rank one. Suppose that u ∈ CU (A). Then, for any piecewise smooth and continuous path {u(t) : t ∈ [0, 1]} ⊂ U (A) with u(0) = u and u(1) = 1A, DA({u(t)}) ∈ ρA(K0(A)). (recall 2.15 f or DA) (e 28.1) Proof. It follows from 11.10 that the map j : u 7→ diag(u, 1, ..., 1) from U (A) to U (Mn(A)) induces an isomorphism from U (A)/CU (A) to U (Mn(A))/CU (Mn(A)). Then the lemma follows from 3.1 and 3.2 of [98]. 269 Lemma 28.2. Let A be a unital separable simple C∗-algebra of stable rank one. Suppose that B is a unital separable C∗-algebra and suppose that ϕ, ψ : B → A are two unital monomorphisms such that Then [ϕ] = [ψ] in KK(B, A) ϕT = ψT and ϕ‡ = ψ‡. Rϕ,ψ ∈ Hom(K1(B), ρA(K0(A))). (e 28.2) (e 28.3) (e 28.4) Proof. Let z ∈ K1(B) be represented by a unitary u ∈ U (Mm(B)) for some integer m. Then, by (e 28.3), (ϕ ⊗ idMm)(u)(ψ ⊗ idMm)(u)∗ ∈ CU (Mm(A)). Suppose that {u(t) : t ∈ [0, 1]} is a piecewise smooth and continuous path in Mm(U (A)) such that u(0) = (ϕ ⊗ idMm)(u) and u(1) = (ψ ⊗ idMm)(u). Put w(t) = (ψ ⊗ idMm)(u)∗u(t). Then w(0) = (ψ ⊗ idMm)(u)∗(ϕ ⊗ idMm)(u) ∈ CU (A) and w(1) = 1A. Thus Rϕ,ψ(z)(τ ) = = for all τ ∈ T (A). By 28.1, It follows that 0 1 2πiZ 1 2πiZ 1 1 τ ( τ ( du(t) dt u∗(t))dt = dw(t) dt w∗(t))dt 0 1 2πiZ 1 0 τ (ψ(u)∗ du(t) dt u∗(t)ψ(u))dt (e 28.5) Rϕ,ψ(z) ∈ ρA(K0(A)). Rϕ,ψ ∈ Hom(K1(B), ρA(K0(A))). (e 28.6) (e 28.7) (e 28.8) Theorem 28.3. Let C1, C2 ∈ B0 be unital separable simple C∗-algebras, A = C1 ⊗ U1, B = C2 ⊗ U2, where U1 and U2 are UHF-algebras of infinite type. Suppose that B is a unital C∗- subalgebra of A, and denote by ı the embedding. For any λ ∈ Hom(K1(B), ρA(K0(A))), there exists ϕ ∈ Inn(B, A)(see 2.8) such that there are homomorphisms θi : Ki(B) → Ki(Mı,ϕ) with (π0)∗i ◦ θi = idKi(B), i = 0, 1, and the rotation map Rı,ϕ : K1(Mı,ϕ) → Aff(T (A)) given by Rı,ϕ(x) = ρA(x − θ1(π0)∗1(x)) + λ ◦ (π0)∗1(x)) for all x ∈ K1(Mı,ϕ). In other words, [ϕ] = [ı] in KK(B, A) and the rotation map Rı,ϕ : K1(Mı,ϕ) → Aff(T (A)) is given by Rı,ϕ(a, b) = ρA(a) + λ(b) (e 28.9) (e 28.10) (e 28.11) for some identification of K1(Mı,ϕ) with K0(A) ⊕ K1(B). Proof. The proof is exactly the same as that of Theorem 4.2 of [75]. In 4.2 of [75], it is assumed that ρA(K0(A)) is dense in Aff(T (A)). However, it is that λ(K1(B)) ⊂ ρA(K0(A)) is used which we assume here. In 4.2 of [75], it is also assume that A has the property (B1) and (B2) associated with B (defined in 3.6 of [75]). But this follows from 24.1 (see also 24.2). 270 Definition 28.4. Let A be a unital C∗-algebra and let C be a unital separable C∗-algebra. De- note by Mone asu(C, A) the set of all asymptotically unitary equivalence classes of unital monomor- phisms from C into A. Denote by K : Mone asu(C, A) → KK e(C, A)++ the map defined by ϕ 7→ [ϕ] for all ϕ ∈ Mone asu(C, A). Let κ ∈ KK e(C, A)++. Denote by hκi the classes of ϕ ∈ Mone Denote by KKU Te(A, B)++ the set of triples (κ, α, γ) for which κ ∈ KKe(A, B)++, α : U (A)/CU (A) → U (B)/CU (B) is a homomorphism and γ : T (B) → T (A) is an affine continuous map and both α and γ are compatible with κ. Denote by K the map from Mone asu(C, A) into KKU T (C, A)++ defined by asu(C, A) such that K(ϕ) = κ. ϕ 7→ ([ϕ], ϕ‡, ϕT ) for all ϕ ∈ Mone asu(C, A). Denote by hκ, α, γi the subset of ϕ ∈ Mone Theorem 28.5. Let C and A be two unital separable amenable C∗-algebras. Suppose that ϕ1, ϕ2, ϕ3 : C → A are three unital monomorphisms for which asu(C, A) such that K(ϕ) = (κ, α, γ). [ϕ1] = [ϕ2] = [ϕ3] in KK(C, A)) and (ϕ1)T = (ϕ2)T = (ϕ3)T . (e 28.12) Then Rϕ1,ϕ2 + Rϕ2,ϕ3 = Rϕ1,ϕ3. (e 28.13) Proof. The proof is exactly the same as that of Theorem 9.6 of [69]. Lemma 28.6. Let A and B be two unital separable amenable C∗-algebras. Suppose that ϕ1, ϕ2 : A → B are two unital monomorphisms such that [ϕ1] = [ϕ2] in KK(A, B) and (ϕ1)T = (ϕ2)T . Suppose that (ϕ2)T : T (B) → T (A) is an affine homeomorphism. Suppose also that there is α ∈ Aut(B) such that [α] = [idB] in KK(B, B) and αT = idT . Then Rϕ1,α◦ϕ2 = RidB ,α ◦ (ϕ2)∗1 + Rϕ1,ϕ2 (e 28.14) in Hom(K1(A), ρB(K0(B)))/R0. Proof. By 28.5, we compute that Rϕ1,α◦ϕ2 = Rϕ1,ϕ2 + Rϕ2,α◦ϕ2 = Rϕ1,ϕ2 + RidB ,α ◦ (ϕ2)∗1. Theorem 28.7. Let B ∈ N be a unital simple C∗-algebra in B0, let C = B ⊗ U1, where U1 is let A1 is a unital separable amenable simple C∗-algebra in B0 a UHF-algebra of infinite type, and let A = A1 ⊗ U2, where U2 is another UHF-algebra of infinite type. Then the map K : Mone asu(C, A) → KKU T (C, A)++ is surjective. Moreover, for each (κ, α, γ) ∈ KKU T (C, A)++, there exists a bijection η : hκ, α, γi → Hom(K1(C), ρA(K0(A)))/R0. 271 Proof. It follows from 24.4 that K is surjective. Fix a triple (κ, α, γ) ∈ KKT (C, A)++ and choose a unital monomorphism ϕ : C → A such that [ϕ] = κ, ϕ‡ = α and ϕT = γ. If ϕ1 : C → A is another unital monomorphism such that K(ϕ1) = K(ϕ), then by 28.2, Rϕ,ϕ1 ∈ Hom(K1(C), ρA(K0(A)))/R0. (e 28.15) Let λ ∈ Hom(K1(C), ρA(K0(A))) be a homomorphism. It follows from 28.3 that there is a unital monomorphism ψ ∈ Inn(ϕ(C), A) with [ψ ◦ ϕ] = [ϕ] in KK(C, A) such that there exists a homomorphism θ : K1(C) → K1(Mϕ,ψ◦ϕ) with (π0)∗1 ◦ θ = idK1(C) for which Rϕ,ψ◦ϕ ◦ θ = λ. Let β = ψ ◦ ϕ. Then Rϕ,β ◦ θ = λ. Note also since ψ ∈ Inn(ϕ(C), A), β‡ = ϕ‡ and βT = ϕT. In particular, K(β) = K(ϕ). Thus, for each unital monomorphism ϕ, we obtain a well-defined and surjective map ηϕ : h[ϕ], ϕ‡, ϕTi → Hom(K1(A), ρA(K0(A)))/R0. To see it is one to one, consider two monomorphisms ϕ1, ϕ2 : C → A in h[ϕ], ϕ‡, ϕTi such that Rϕ,ϕ1 = Rϕ,ϕ2. Then, by 28.2, Rϕ1,ϕ2 = Rϕ1,ϕ + Rϕ,ϕ2 = −Rϕ,ϕ1 + Rϕ,ϕ2 = 0. (e 28.16) e It follows from 27.4 that ϕ1 and ϕ2 are asymptotically unitarily equivalent. The map ηϕ is the desired bijection η as h[ϕ], ϕ‡, ϕTi = hκ, α, γi. Definition 28.8. Denote by KKU T −1 (A, A)++ the subgroup of those elements (κ, α, γ) ∈ KKU Te(A, A)++ for which κKi(A) is an isomorphism (i = 0, 1), α is an isomorphism and γ is an affine homeomorphism. Recall from the proof of 28.7 ηidA : h[idA], id‡A, (idA)Ti → Hom(K1(A), ρA(K0(A)))/R0 is a bijection. Denote by hidAi the class of those automorphisms ψ which are asymptotically unitarily equivalent to idA—this subset of Aut(A) gives up a single element in Mone asu(A, A) which should not be confused with the subset h[idA], id‡A, (idA)Ti ⊂ Mone asu(A, A). Note that, if ψ ∈ hidAi, then ψ is asymptotically inner, i.e., there exists a continuous path of unitaries {u(t) : t ∈ [0,∞)} ⊂ A such that ψ(a) = lim t→∞ u(t)∗au(t) for all a ∈ A. Note that hidAi is a normal subgroup of Aut(A). Corollary 28.9. Let A1 ∈ N ∩ B0 be a unital simple C∗-algebra and let A = A1 ⊗ U for some UHF-algebra U of infinite type. Then one has the following short exact sequence: 0 → Hom(K1(A), ρA(K0(A)))/R0 In particular, if ϕ, ψ ∈ Aut(A) such that η−1 idA→ Aut(A)/hidAi K → KKU T −1 e (A, A)++ → 0. (e 28.17) Then K(ϕ) = K(ψ) = K(idA), ηidA(ϕ ◦ ψ) = ηidA(ϕ) + ηidA(ψ). 272 Proof. It follows from 24.4 that, for any hκ, α, γi, there is a unital monomorphism h : A → A such that K(h) = hκ, α, γi. The fact that κ ∈ KK−1 e (A, A)++ implies that there is κ1 ∈ KK−1 κ × κ1 = κ1 × κ = [idA]. e (A, A)++ such that By 24.4, choose h1 : A → A such that K(h) = hκ1, α−1, γ−1i. It follows from 27.1 that h1 ◦ h and h ◦ h1 are approximately unitarily equivalent. Applying a standard approximate intertwining argument of G. A. Elliott, one obtains two isomorphisms ϕ and ϕ−1 such that there is a sequence of unitaries {un} in A such that ϕ(a) = lim n→∞ Ad u2n+1 ◦ h(a) and ϕ−1(a) = lim n→∞ Ad u2n ◦ h1(a) for all a ∈ A. Thus [ϕ] = [h] in KL(A, A) and ϕ‡ = h‡ and ϕT = hT . Then, as in the proof of 24.4, there is ψ0 ∈ Inn(A, A) such that [ψ0 ◦ ϕ] = [idA] in KK(A, A) as well as (ψ0 ◦ ϕ)‡ = h‡ and (ψ0 ◦ ϕ)T = hT . So we have ψ0 ◦ ϕ ∈ Aut(A, A) such that K(ψ0 ◦ ϕ) = hκ, α, γi. This implies that K is surjective. Now let λ ∈ Hom(K1(C), Aff(T (A)))/R0. The proof 28.7 says that there is ψ00 ∈ Inn(A, A) (in place of ψ) such that K(ψ00 ◦ idA) = K(idA) and RidA,ψ00 = λ. Note that ψ00 is again an automorphism. The last part of the lemma then follows from 28.6. Definition 28.10. (Definition 10.2 of [65] and see also [70]) Let A be a unital C∗-algebra and B be another C∗-algebra. Recall ([70]) that H1(K0(A), K1(B)) = {x ∈ K1(B) : ϕ([1A]) = x for some ϕ ∈ Hom(K0(A), K1(B))}. Proposition 28.11. (Proposition 12.3 of [65]) Let A be a unital separable C∗-algebra and let B be a unital C∗-algebra. Suppose that ϕ : A → B is a unital homomorphism and u ∈ U (B) is a unitary. Suppose that there is a continuous path of unitaries {u(t) : t ∈ [0,∞)} ⊂ B such that (e 28.18) u(0) = 1B and lim t→∞ ad u(t) ◦ ϕ(a) = ad u ◦ ϕ(a) for all a ∈ A. Then [u] ∈ H1(K0(A), K1(B)). Lemma 28.12. Let C = C′ ⊗ U for some C′ = lim−→(Cn, ψn) and a UHF algebra U of infinite type, where each Cn is a direct sum of C-algebras in C0 and H. Assume that ψn is unital and injective. Let A ∈ B1. Let ϕ1, ϕ2 : C → A be two monomorphisms such that there is an increasing sequence of finite subsets Fn ⊆ C with dense union, an increasing sequence of finite and a sequence of unitaries {un} ⊆ A, such that subsets Pn ⊆ K1(C) with union to be K1(C), a sequence of positive numbers (δn) withP δn < 1 Adun ◦ ϕ1 ≈δn ϕ2 on Fn and ρA(bott1(ϕ2, u∗nun+1)) = 0 for all x ∈ Pn. Suppose that H1(K0(C), K1(A)) = K1(A). Then there exists a sequence of unitaries vn ∈ U0(A) such that Advn ◦ ϕ1 ≈δn ϕ2 ρA(bott1(ϕ2, v∗nvn+1)) = 0, on Fn and x ∈ Pn. (e 28.19) (e 28.20) 273 Proof. Let xn = [un] ∈ K1(A). Since H1(K0(C), K1(A)) = K1(A), there is a homomorphism κn,0 : K0(C) → K1(A) such that κn,0([1C ]) = −xn. Since C satisfies the Universal Coefficient Theorem, there is κn ∈ KL(C ⊗ C(T), A) such that (κn)β(K0(C)) = κn,0 and (κn)β(K1(C)) = 0. Without loss of generality, we may assume that [1C ] ∈ Pn, n = 1, 2, ..... For each δn, choose a positive number ηn < δn, such that Adun ◦ ϕ1 ≈ηn ϕ2 on Fn. By 24.1, there is a unitary wn ∈ U (A) such that k[ϕ2(a), wn]k < (δn − ηn)/2 Put vn = unwn, n = 1, 2, .... Then for all a ∈ Fn and Bott(ϕ2, wn)Pn = κnβ(Pn). Advn ◦ ϕ1 ≈δn ϕ2 on Fn, ρA(bott1(ϕ2, v∗nvn+1))Pn = 0 and, since [1C ] ∈ Pn, [vn] = [un] − xn = 0, as desired. Theorem 28.13. Let B ∈ B1 be a unital separable simple C∗-algebra which satisfies the UCT, let A1 ∈ B1 be a unital separable simple C∗-algebra and let C = B ⊗ U1 and A = A1 ⊗ U2, where U1 and U2 are two unital infinite dimensional UHF-algebras. Suppose that H1(K0(C), K1(A)) = K1(A) and suppose that ϕ1, ϕ2 : C → A are two unital monomorphisms which are asymptotically unitarily equivalent. Then there exists a continuous path of unitaries {u(t) : t ∈ [0,∞)} ⊂ A such that u(0) = 1 and lim t→∞ adu(t) ◦ ϕ1(a) = ϕ2(a) for all a ∈ C. Proof. By 4.3 of [69], one has in KK(C, A), ϕ‡ = ψ‡, (ϕ1)T = (ϕ2)T and Rϕ1,ϕ2 = 0. [ϕ1] = [ϕ2] Then by Lemma 28.12, one may assume that vn ∈ U0(A), n = 1, 2, ... in the proof 27.4. It follows that ξn([1C ]) = 0, n = 1, 2, ..., and therefore κn([1C ]) = 0. This implies that γn ◦ β([1C ]) = 0. Hence wn ∈ U0(A), and un ∈ U0(A). Therefore, the continuous path of unitaries {u(t)} constructed in 27.4 is in U0(A), and then one may require that u(0) = 1A. 29 The classification theorem Lemma 29.1. Let A1 ∈ B0 be a unital separable simple C∗-algebra, let A = A1 ⊗ U for some infinite dimensional UHF-algebra and let p be a supernatural number of infinite type. Then the homomorphism ı : a 7→ a ⊗ 1 induces an isomorphism from U0(A)/CU (A) to U0(A ⊗ Mp)/CU (A ⊗ Mp). 274 Proof. There are sequences of positive integers {m(n)} and {k(n)} such that A⊗Mp = limn→∞(A⊗ Mm(n), ın), where ın : Mm(n)(A) → Mm(n+1)(A) is defined by ı(a) = diag( Mm(n)(A1) ⊗ U and Mm(n)(A1) ∈ B0. Let k(n) } { z a, a, ..., a) for all a ∈ Mm(n)(A), n = 1, 2, .... Note, Mm(n)(A) = jn : U (Mm(n)(A))/CU (Mm(n)(A))) → U (Mm(n+1)(A))/CU (Mm(n+1)(A)) be defined by jn(¯u) = diag(u, 1, 1, ..., 1 ) for all u ∈ U ((Mm(n)(A)). k(n)−1 {z } It follows from Theorem 11.10 that jn is an isomorphism. By 11.7, U0(Mm(n)(A))/CU (Mm(n)(A)) is divisible. For each n and i, there is a unitary Ui ∈ Mm(n+1)(A) such that U∗i E1,1Ui = Ei,i, i = 2, 3, ..., k(n), ın(u) = u′U∗2 u′U2 ··· U∗k(n)u′Uk(n), j=(i−1)m(n)+1 ej,j and {ei,j} is a matrix unit for Mm(n+1). Then where Ei,i =Pim(n) where u′ = diag(u,z It follows that ı‡nU0(Mm(n)(A))/CU (Mm(n) (A)) is injective, since U0(Mm(n+1)(A))/CU (Mm(n+1)(A)) is torsion free (see 11.5). For each z ∈ U0(Mm(n+1)(A)/CU (Mm(n+1)), there is a unitary v ∈ Mm(n+1)(A) such that 1, 1, ..., 1), for all u ∈ Mm(n)(A). Thus ı‡n(¯u) = k(n)jn(¯u). } { since jn is an isomorphism. By the divisibility of U0(Mm(n)(A)/CU (Mm(n)), there is u ∈ Mm(n)(A) such that jn(¯v) = z, As above, uk(n) = uk(n) = v. ı‡n(¯u) = k(n)jn(¯v) = z. So ı‡nU0(Mm(n)(A))/CU (Mm(n)(A)) is a surjective. It follows that ı‡n,∞U0(Mm(n)(A))/CU (Mm(n)(A)) is also an isomorphism. One then concludes that ı‡U0(A)/CU (A) is an isomorphism. Lemma 29.2. Let A1 and B1 be two unital separable simple C∗-algebras in B0, let A = A1 ⊗ U1 and let B = B1⊗U2, where U1 and U2 are two infinite dimensional UHF-algebras. Let ϕ : A → B be an isomorphism and let β : B⊗Mp → B⊗Mp be an automorphism such that β∗1 = idK1(B⊗Mp) for some supernatural number p of infinite type. Then ψ‡(U (A)/CU (A)) = (ϕ0)‡(U (A)/CU (A)) = U (B)/CU (B), where ϕ0 = ı ◦ ϕ, ψ = β ◦ ı ◦ ϕ and where ı : B → B ⊗ Mp is defined by ı(b) = b ⊗ 1 for all b ∈ B. Moreover there is an isomorphism µ : U (B)/CU (B) → U (B)/CU (B) with µ(U0(B)/CU (B)) ⊂ U0(B)/CU (B) such that where q1 : U (B)/CU (B) → K1(B) is the quotient map. ı‡ ◦ µ ◦ ϕ‡ = ψ‡ and q1 ◦ µ = q1, 275 Proof. The proof is exactly the same as that of Lemma 11.3 of [69]. Lemma 29.3. Let A1 and B1 be two unital simple amenable C∗-algebras in N ∩ B0, let A = A1 ⊗ U1 and let B = B1 ⊗ U2, where U1 and U2 are UHF-algebras of infinite type. Suppose that ϕ1, ϕ2 : A → B are two isomorphisms such that [ϕ1] = [ϕ2] in KK(A, B). Then there exists an automorphism β : B → B such that [β] = [idB] in KK(B, B) and β◦ϕ2 is asymptotically unitar- ily equivalent to ϕ1. Moreover, if H1(K0(A), K1(B)) = K1(B), they are strongly asymptotically unitarily equivalent. Proof. It follows from 28.7 that there is an automorphism β1 : B → B satisfying the following: (e 29.1) β‡1 = ϕ‡1 ◦ (ϕ−1 [β1] = [idB] in KK(B, B), 2 )‡ and (β1)T = (ϕ1)T ◦ (ϕ2)−1 T . (e 29.2) (e 29.3) (e 29.4) (e 29.5) (e 29.6) (e 29.7) (e 29.8) By 28.9, there is automorphism β2 ∈ Aut(B) such that [β2] = [idB] in KK(B, B), β‡2 = id‡B, (β2)T = (idB)T and RidB ,β2 = −Rϕ1,β1◦ϕ2 ◦ (ϕ2)−1 ∗1 . Put β = β2 ◦ β1. It follows that [β ◦ ϕ2] = [ϕ1] in KK(A, B), (β ◦ ϕ2)‡ = ϕ‡1 and (β ◦ ϕ2)T = (ϕ1)T . Moreover, by 28.6, Rϕ1,β◦ϕ2 = RidB ,β2 ◦ (ϕ2)∗1 + Rϕ1,β1◦ϕ2 = (−Rϕ1,β1◦ϕ2 ◦ (ϕ2)−1 In the case that H1(K0(A), K1(B)) = K1(B), it follows from 28.13 that β ◦ ϕ2 and ϕ1 are ∗1 ) ◦ (ϕ2)∗1 + Rϕ1,β1◦ϕ2 = 0. It follows from 28.7 that β ◦ ϕ2 and ϕ1 are asymptotically unitarily equivalent. strongly asymptotically unitarily equivalent. Lemma 29.4. Let A1 and let B1 be two unital simple amenable C∗-algebras in N ∩ B0 and let A = A⊗ U1 and B = B1 ⊗ U2 for some UHF-algebras U1 and U2 of infinite type. Let ϕ : A → B be an isomorphism. Suppose that β ∈ Aut(B ⊗ Mp) for which [β] = [idB⊗Mp] in KK(B ⊗ Mp, B ⊗ Mp) and βT = (idB⊗Mp)T for some supernatural number p of infinite type. Then there exists an automorphism α ∈ Aut(B) with [α] = [idB] in KK(B, B) such that ı ◦ α ◦ ϕ and β ◦ ı ◦ ϕ are asymptotically unitarily equivalent, where ı : B → B ⊗ Mp is defined by ı(b) = b ⊗ 1 for all b ∈ B. Proof. It follows from 29.2 that there is an isomorphism µ : U (B)/CU (B) → U (B)/CU (B) such that Note that ıT : T (B ⊗ Mp) → T (B) is an affine homeomorphism. It follows from 28.7 that there is an automorphism α : B → B such that ı‡ ◦ µ ◦ ϕ‡ = (β ◦ ı ◦ ϕ)‡. [α] = [idB] in KK(B, B), α‡ = µ, αT = (β ◦ ı ◦ ϕ)T ◦ ((ı ◦ ϕ)T )−1 = (idB⊗Mp)T and RidB ,α(x)(τ ) = −Rβ◦ı◦ϕ, ı◦ϕ(ϕ−1 ∗1 (x))(ıT (τ )) for all x ∈ K1(A) (e 29.9) (e 29.10) (e 29.11) 276 and for all τ ∈ T (B). Denote by ψ = ı ◦ α ◦ ϕ. Then we have, by 28.6, [ψ] = [ı ◦ ϕ] = [β ◦ ı ◦ ϕ] in KK(A, B ⊗ Mp) ψ‡ = ı‡ ◦ µ ◦ ϕ‡ = (β ◦ ı ◦ ϕ)‡, ψT = (ı ◦ α ◦ ϕ)T = (ı ◦ ϕ)T = (β ◦ ı ◦ ϕ)T . Moreover, for any x ∈ K1(A) and τ ∈ T (B ⊗ Mp), Rβ◦ı◦ϕ,ψ(x)(τ ) = Rβ◦ı◦ϕ,ı◦ϕ(x)(τ ) + Rı,ı◦α ◦ ϕ∗1(x)(τ ) = Rβ◦ı◦ϕ,ı◦ϕ(x)(τ ) + RidB ,ı◦α ◦ ϕ∗1(x)(ı−1 = Rβ◦ı◦ϕ,ı◦ϕ(x)(τ ) − Rβ◦ı◦ϕ,ı◦ϕ(ϕ−1 ∗1 )(ϕ∗1(x))(τ ) = 0 T (τ )) (e 29.12) (e 29.13) (e 29.14) (e 29.15) (e 29.16) (e 29.17) It follows from 28.7 that ı ◦ α ◦ ϕ and β ◦ ı ◦ ϕ are asymptotically unitarily equivalent. Theorem 29.5. Let A and B be two unital separable simple C∗-algebras in N . Suppose that there is an isomorphism Suppose also that, for some pair of relatively prime supernatural numbers p and q of infinite type such that Mp ⊗ Mq ∼= Q, A ⊗ Mp ∈ B0, B ⊗ Mp ∈ B0, A ⊗ Mq ∈ B0 and B ⊗ Mq ∈ B0 Then, Γ : Ell(A) → Ell(B). A ⊗ Z ∼= B ⊗ Z. Proof. The proof is almost identical to that of 11.7 of [69] with a few necessary modifications. Note that Γ induces an isomorphism Γp : Ell(A ⊗ Mp) → Ell(B ⊗ Mp). Since A⊗ Mp ∈ B0 and B ⊗ Mp ∈ B0, by Theorem 21.10, there is an isomorphism ϕp : A⊗ Mp → B ⊗ Mp. Moreover, (by the proof of 21.10), ϕp carries Γp. For exactly the same reason, Γ induces an isomorphism Γq : Ell(A ⊗ Mq) → Ell(B ⊗ Mq) and there is an isomorphism ψq : A ⊗ Mq → B ⊗ Mq which induces Γq. Put ϕ = ϕp ⊗ idMq : A ⊗ Q → B ⊗ Q and ψ = ψq ⊗ idMp : A ⊗ Q → B ⊗ Q. Note that (ϕ)∗i = (ψ)∗i (i = 0, 1) and ϕT = ψT (they are induced by Γ). Note that ϕT and ψT are affine homeomorphisms. Since K∗i(B ⊗ Q) is divisible, we in fact have [ϕ] = [ψ] (in KK(A ⊗ Q, B ⊗ Q)). It follows from 29.3 that there is an automorphism β : B ⊗ Q → B ⊗ Q such that [β] = [idB⊗Q] KK(B ⊗ Q, B ⊗ Q) and such that ϕ and β ◦ ψ are asymptotically unitarily equivalent. Since K1(B ⊗ Q) is divisible, H1(K0(A⊗ Q), K1(B⊗ Q)) = K1(B⊗ Q). It follows that ϕ and β◦ ψ are strongly asymptotically unitarily equivalent. Note also in this case βT = (idB⊗Q)T . Let ı : B ⊗ Mq → B ⊗ Q defined by ı(b) = b ⊗ 1 for b ∈ B. We consider the pair β ◦ ı ◦ ψq and ı ◦ ψq. By applying 29.4, there exists an automorphism α : B ⊗ Mq → B ⊗ Mq such that 277 ı ◦ α ◦ ψq and β ◦ ı ◦ ψq are asymptotically unitarily equivalent (in B ⊗ Q). So they are strongly asymptotically unitarily equivalent. Moreover, [α] = [idB⊗Mq] in KK(B ⊗ Mq, B ⊗ Mq). We will show that β ◦ ψ and α ◦ ψq ⊗ idMp are strongly asymptotically unitarily equivalent. Define β1 = β ◦ ı ◦ ψq ⊗ idMp : B ⊗ Q ⊗ Mp → B ⊗ Q ⊗ Mp. Let j : Q → Q ⊗ Mp be defined by j(b) = b ⊗ 1. There is an isomorphism s : Mp → Mp ⊗ Mp with (idMq ⊗ s)∗0 = j∗0. In this case [idMq ⊗ s] = [j]. Since K1(Mp) = 0, by 27.4, idMq ⊗ s is strongly asymptotically unitarily equivalent to j. It follows that α ◦ ψq ⊗ idMp and β ◦ ı ◦ ψq ⊗ idMp are strongly asymptotically unitarily equivalent. Consider the C∗-subalgebra C = β ◦ ψ(1⊗ Mp)⊗ Mp ⊂ B ⊗ Q⊗ Mp. In C, β ◦ ϕ1⊗Mp and j0 are strongly asymptotically unitarily equivalent, where j0 : Mp → C is defined by j0(a) = 1⊗a for all a ∈ Mp. There exists a continuous path of unitaries {v(t) : t ∈ [0,∞)} ⊂ C such that lim t→∞ ad v(t) ◦ β ◦ ϕ(1 ⊗ a) = 1 ⊗ a for all a ∈ Mp. (e 29.18) It follows that β ◦ ψ and β1 are strongly asymptotically unitarily equivalent. Therefore β ◦ ψ and α ◦ ψq ⊗ idMp are strongly asymptotically unitarily equivalent. Finally, we conclude that α ◦ ψq ⊗ idp and ϕ are strongly asymptotically unitarily equivalent. Note that α ◦ ψq is an isomorphism which induces Γq. Let {u(t) : t ∈ [0, 1)} be a continuous path of unitaries in B ⊗ Q with u(0) = 1B⊗Q such that lim t→∞ ad u(t) ◦ ϕ(a) = α ◦ ψq ⊗ idMp(a) for all a ∈ A ⊗ Q. One then obtains a unitary suspended isomorphism which lifts Γ along Zp,q (see [105]). It follows from Theorem 7.1 of [105] that A ⊗ Z and B ⊗ Z are isomorphic. Definition 29.6. Denote by N0 the class of those unital simple C∗-algebras A in N for which A ⊗ Mp ∈ N ∩ B0 for any supernatural number p of infinite type. Of course N0 contains all unital simple amenable C∗-algebras in B0 which satisfy the UCT. It contains all unital simple inductive limits of C∗-algebras in C0. Corollary 29.7. Let A and B be two C∗-algebras in N0. Then A ⊗ Z ∼= B ⊗ Z if and only if Ell(A ⊗ Z) ∼= Ell(B ⊗ Z). Proof. This follows from 29.5 immediately. Theorem 29.8. Let A and B be two unital separable simple amenable Z-stable C∗-algebras which satisfy the UCT. Suppose that gT R(A ⊗ Q) ≤ 1 and gT R(B ⊗ Q) ≤ 1. Then A ∼= B if and only if Ell(A) ∼= Ell(B). Proof. It follows from 19.3 that A ⊗ U, B ⊗ U ∈ B0 for any UHF-algebra U of infinite type. Therefore the theorem follows immediately from 29.7. Corollary 29.9. Let A and B be two unital separable simple C∗-algebras which satisfy the UCT. Suppose that gT R(A) ≤ 1 and gT R(B) ≤ 1. Then A ∼= B if and only if Ell(A) ∼= Ell(B). Corollary 29.10. Let A and B be two unital simple C∗-algebras in B1 ∩N . Then A ∼= B if and only if Ell(A) ∼= Ell(B). 278 Proof. It follows from 10.7 that A ⊗ Z ∼= A and B ⊗ Z ∼= B. Thus the corollary follows from 29.8. Added in Proof: As this paper was revised in 2015, it has been shown (see [28]) that all unital separable simple C∗-algebras A with finite decomposition rank which satisfy the UCT have gT R(A ⊗ U ) ≤ 1 for all UHF-algebras of infinite type. Therefore, by Theorem 29.8, they are classified by the Elliott invariant. References [1] C. Akemann and F. Shultz, Perfect C ∗-algebras, Mem. Amer. Math. Soc. 55 (1985), no. 326, xiii+117 pp [2] B. Blackadar, M. Dadarlat, and M. Rørdam, The real rank of inductive limit C ∗ -algebras. Math. Scand. 69 (1991), 211–216 (1992). [3] B. Blackadar and E. Kirchberg, Generalized inductive limits of finite-dimensional C-algebras, Math. Ann. 307 (1997), 343–380. [4] B. Blackadar, A. Kumjian, and M. Rørdam, Approximately central matrix units and the structure of noncommutative tori, K-Theory 6 (1992), 267–284. [5] B. Blackadar and M. Rørdam, Extending states on preordered semigroups and existence of the qua- sitrace on C-algebras, J. Algebra 152 (1992), 240–247. [6] L. G. Brown, Stable isomorphism of hereditary subalgebras of C-algebras, Pacific J. Math. 71 (1977), 335–348 [7] L. G. Brown, P. Green and M. Rieffel, Stable isomorphism and strong Morita equivalence of C- algebras, Pacific J. Math. 71 (1977), 349–363. [8] N. Brown and N. Ozawa, C-algebras and finite-dimensional approximations, American Mathemat- ical Society, Inc., Providence, RI, 2008. xv+509 pp. ISBN: 0-8218-4381. [9] N. Brown, F. Perera and A. Toms, The Cuntz semigroup, the Elliott conjecture, and dimension functions on C ∗-algebras, J. Reine Angew. Math. 621 (2008), 191–211. [10] M. D. Choi and G. A. Elliott, Density of the selfadjoint elements with finite spectrum in an irrational rotation C ∗-algebra, Math. Scand. 67 (1990), 73–86. [11] K. T. Coward, G. A. Elliott, and C. Ivanescu, The Cuntz semigroup as an invariant for C-algebras, J. Reine Angew. Math. 623 (2008), 161–193. [12] J. Cuntz, The structure of multiplication and addition in simple C ∗-algebras, Math. Scand. 40 (1977), 215–233. [13] J. Cuntz, Dimension functions on simple C ∗-algebras, Math. Ann. 233 (1978), 145–153. [14] J. Cuntz and G. K. Pedersen, Equivalence and traces on C ∗-algebras, J. Funct. Anal. 33 (1979), 135–164 [15] de la Harpe, P. and Skandalis, G., D´eterminant associ´e `a une trace sur une alg`ebre de Banach, Ann. Inst. Fourier, Grenoble, 34-1 (1984), 169–202. [16] M. Dadarlat, Approximately unitarily equivalent morphisms and inductive limit C ∗-algebras, K- Theory 9 (1995), 117–137. [17] M. Dadarlat and T. Loring, A universal multicoefficient theorem for the Kasparov groups, Duke Math. J. 84 (1996), 355–377. [18] M. Dadarlat and S. Eilers, On the classification of nuclear C ∗-algebras, Proc. London Math. Soc. 85 (2002), 168–210 [19] M. Dadarlat, G. Nagy and A. N´emethi and C. Pasnicu, Reduction of topological stable rank in inductive limits of C-algebras, Pacific J. Math. 153 (1992), 267–276. 279 [20] E. G. Effros, D. E. Handelman, and C. L. Shen, Dimension groups and their affine representations, Amer. J. Math 102 (1980), 385–407. [21] S. Eilers, T. Loring and G. K. Pedersen, Stability of anticommutation relations: an application of noncommutative CW complexes, J. Reine Angew. Math. 499 (1998), 101–143. [22] S. Eilers, T. Loring and P.K. Pedersen, Fragility of subhomogeneous C-algebras with one- dimensional spectrum, Bull. London Math. Soc. 31 (1999), 337–344. [23] G. A. Elliott, On the classification of inductive limits of sequences of semisimple finite-dimensional algebras, J. Algebra, 38 (1976), 29–44. [24] G. A. Elliott, On the classification of C-algebras of real rank zero, J. Reine Angew. Math, 443 (1993), 179–219. [25] G. A. Elliott. An invariant for simple C-algebras. Canadian Mathematical Society. 1945–1995, Vol. 3, 61–90, Canadian Math. Soc., Ottawa, ON, 1996. [26] G. A. Elliott and G. Gong, On the classification of C-algebras of real rank zero. II, Ann. of Math. (2) 144 (1996), 497–610. [27] G. A. Elliott, G. Gong, and L. Li, On the classification of simple inductive limit C-algebras. II. The isomorphism theorem, Invent. Math. 168 (2007), 249–320. [28] G. A. Elliott, G. Gong, H. Lin and Z. Niu, On the classification of simple C ∗-algebras with finite decomposition rank, II , preprint, arXiv:1507.03437 [29] G. A. Elliott and T. Loring, AF embeddings of C(T2) with a prescribed K-theory, J. Funct. Anal. 103 (1992), 1–25. [30] G. A. Elliott and Z. Niu, On tracial approximation, J. Funct. Anal. 254 (2008), 396–440. [31] G. A. Elliott, Z. Niu, L. Santiago, A. P. Tikuisis and W. Winter Decomposition rank of minimal crossed products, draft preprint, 2015. [32] G. A. Elliott and K. Thomsen, The state space of the K0-group of a simple separable C-algebra, Geom. Funct. Anal. 4 (1994), no. 5, 522–538. [33] Q. Fan and X. Fang, C-algebras of tracially stable rank one, Acta Math. Sinica (Chin. Ser.) 48 (2005), 929–934. [34] H. Freudenthal, Entwicklungen yon Rdumen und ihren Gruppen, Compositio Math., 4, (1937), 145- 234. [35] G. Gong, On the classification of simple inductive limit C∗-algebras. I. The reduction theorem, Doc. Math. 7 (2002), 255–461. [36] G. Gong, X. Jiang, and H. Su, Obstructions to Z -stability for unital simple C∗-algebras. Canad. Math. Bull. 43 (2000), 418–426. [37] G. Gong and H. Lin, Almost multiplicative morphisms and K-theory, Internat. J. Math. 11 (2000), no. 8, 983–1000. [38] G. Gong, H. Lin, and Y. Xue, Determinant rank of C-algebras, Pacific J. Math. 274 (2015), 405- 436. [39] K. R. Goodearl, Partially ordered abelian groups with interpolation, Mathematical Surveys and Monographs, no. 20, American Mathematical Society, Providence, RI, 1986. [40] K. R. Goodearl and D. Handelman, Rank functions and K0 of regular rings, J. Pure Appl. Algebra 7 (1976), 195–216. [41] V. Guillemin and A. Pollack, Differential topology, Prentice-Hall, Inc., Englewood Cliffs, N.J., 1974. [42] Uffe Haagerup, Quasitraces on exact C∗-algebras are traces, C. R. Math. Acad. Sci. Soc. R. Can. 36 (2014), no. 2-3, 67–92. [43] J. Hua and H. Lin, Rotation algebras and Exel trace formula, Canad. J. Math. 67 (2015), 404–423. 280 [44] X. Jiang and H. Su, On a simple unital projectionless C ∗-algebra, Amer. J. Math. 121 (1999), 359–413. [45] G. Kasparov, Hilbert C ∗-modules: theorems of Stinespring and Voiculescu, J. Operator Theory 4 (1980), 133–150. [46] E. Kirchberg, The classification of purely infinite C-algebras using Kasparov’s theory, Preprint (1994). [47] E. Kirchberg and N. C. Phillips, Embedding of exact C-algebras in the Cuntz algebra O2, J. Reine Angew. Math. 525 (2000), 17–53. [48] E. Kirchberg and M. Rørdam, Non-simple purely infinite C ∗-algebras, Amer. J. Math. 122 (2000), 637–666. [49] L. Li, C ∗-algebra homomorphisms and KK-theory, K-theory 18 (1999), 167-172. [50] L. Li, Simple inductive limit C-algebras: spectra and approximations by interval algebras, J. Reine Angew. Math. 507 (1999), 57–79. [51] H. Lin, Tracially AF C-algebras., Trans. Amer. Math. Soc. 353 (2001), 693–722. [52] H. Lin, Classification of simple tracially AF C-algebras., Canad. J. Math. 53 (2001), no. 1, 161–194. [53] H. Lin, Simple C-algebras with locally bounded irreducible representation, J. Funct. Anal. 187 (2001), no. 1, 42–69. [54] H. Lin, The tracial topological rank of C ∗-algebras, Proc. London Math. Soc. 83 (2001), 199–234. [55] H. Lin, An introduction to the classification of amenable C ∗-algebras., World Scientific Publishing Co., Inc., River Edge, NJ, 2001. xii+320 pp. ISBN: 981-02-4680-3. [56] H. Lin, Stable approximate unitary equivalence of homomorphisms, J. Operator Theory 47 (2002), 343–378. [57] H. Lin, Classification of simple C-algebras and higher dimensional noncommutative tori, Ann. of Math. (2). 157 (2003), 521–544. [58] H. Lin, Traces and simple C ∗-algebras with tracial topological rank zero, J. Reine Angew. Math. 568 (2004), 99–137. [59] H. Lin, Classification of simple C-algebras of tracial topological rank zero, Duke Math. J. 125 (2004), no. 1, 91–119. [60] H. Lin, A separable Brown-Douglas-Fillmore theorem and weak stability, Trans. Amer. Math. Soc. 356 (2004), no. 7, 2889–2925 [61] H. Lin, An approximate universal coefficient theorem, Trans. Amer. Math. Soc. 357 (2005), 3375– 3405. [62] H. Lin, Classification of homomorphisms and dynamical systems, Trans. Amer. Math. Soc. 359 (2007), 859–895. [63] H. Lin, Simple nuclear C-algebras of tracial topological rank one, J. Funct. Anal. 251 (2007), 601– 679. [64] H. Lin, AF-embedding of crossed products of AH-algebras by Z and asymptotic AF-embedding, Indi- ana Univ. Math. J. 57 (2008), 891–944. [65] H. Lin, Asymptotically unitary equivalence and asymptotically inner automorphisms, Amer. J. Math. 131 (2009), 1589–1677. [66] H. Lin, Approximate homotopy of homomorphisms from C(X) into a simple C-algebra, Mem. Amer. Math. Soc. 205 (2010), no. 963, vi+131. [67] H. Lin, Homotopy of unitaries in simple C-algebras with tracial rank one, J. Funct. Anal. (2010), 1822–1882. 281 [68] H. Lin, Inductive limits of subhomogeneous C ∗-algebras with Hausdorff spectrum, J. Funct. Anal. 258 (2010), 1909–1932. [69] H. Lin, Asymptotic unitary equivalence and classification of simple amenable C-algebras, Invent. Math. 183 (2011), 385–450. [70] H. Lin, Approximate unitary equivalence in simple C ∗-algebras of tracial rank one, Trans. Amer. Math. Soc. 364 (2012), 2021–2086. [71] H. Lin, Approximately diagonalizing matrices over C(Y ), Proc. Natl. Acad. Sci. USA 109 (2012), 2842–2847. [72] H. Lin, Localizing the Elliott conjecture at strongly self-absorbing C-algebras, II, J. Reine Angew. Math. 692 (2014), 233–243. [73] H. Lin, On local AH algebras, Memoirs of AMS, 235 no. 1107, ix+109. DOI: http://dx.doi.org/10.1090/memo/1107. [74] H. Lin, Homomorphisms from AH-algebras, preprint, arXiv: 1102.4631v1 (2011). [75] H. Lin and Z. Niu, Lifting KK-elements, asymptotic unitary equivalence and classification of simple C-algebras, Adv. Math. 219 (2008), no. 5, 1729–1769. [76] H. Lin and Z. Niu, The range of a class of classifiable separable simple amenable C ∗-algebras, J. Funct. Anal. 260 (2011), 1–29. [77] H. Lin and N. C. Phillips, Crossed products by minimal homeomorphisms, J. Reine Angew. Math. 641 (2010), 95–122. [78] H. Lin and M. Rørdam, Extensions of inductive limits of circle algebras, J. London Math. Soc. (2) 51 (1995), 603–613. [79] H. Lin and W. Sun, Tensor products of classifiable C ∗-algebras, arXiv: 1203.3737.v2, preprint. [80] K.E. Nielsen and K. Thomsen, Limit of circle algebras, Expo. Math. 14 (1996), 17-56. [81] Z. Niu, On the classification of TAI algebras, C. R. Math. Acad. Sci. Soc. R. Can. 26 (2004), 18–24. [82] Z. Niu, A classification of tracially approximate splitting interval algebras. I. The building blocks and the limit algebras., C. R. Math. Acad. Sci. Soc. R. Can. 35 (2014), 1–34. [83] Z. Niu, A classification of tracially approximate splitting interval algebras. II. Existence theorem., C. R. Math. Acad. Sci. Soc. R. Can. (2014). [84] Z. Niu, A classification of tracially approximate splitting interval algebras. III. Uniqueness theorem and isomorphism Theorem., C. R. Math. Acad. Sci. Soc. R. Can. (2015). [85] Z. Niu, Mean dimension and AH-algebras with diagonal maps, J. Funct. Anal. 266 (2014), no. 8, 4938–4994. [86] N. C. Phillips, Reduction of exponential rank in direct limits of C-algebras, Canad. J. Math. 46 (1994), 818–853. [87] N. C. Phillips, A classification theorem for nuclear purely infinite simple C-algebras, Doc. Math. 5 (2000), 49–114 (electronic). [88] G. K. Pedersen, Unitary extensions and polar decompositions in a C ∗-algebra, J. Operator Theory 17 (1987), 357–364. [89] M. Rieffel, The homotopy groups of the unitary groups of noncommutative tori, J. Operator Theory 17 (1987), 237–254. [90] J. R. Ringrose, Exponential length and exponential rank in C-algebras, Proc. Roy. Soc. Edinburgh Sect. A 121 (1992), 55–71. [91] L. Robert, The Cuntz semigroup of some spaces of dimension at most two, C. R. Math. Acad. Sci. Soc. R. Can. 35 (2013), 22–32. 282 [92] M. Rørdam, On the structure of simple C-algebras tensored with a UHF-algebra. II, J. Funct. Anal. 107 (1992), 255–269. [93] M. Rørdam, Classification of certain infinite simple C-algebras, J. Funct. Anal. 131 (1995), 415– 458. [94] M. Rørdam, A simple C-algebra with a finite and an infinite projection, Acta Math. 191 (2003), 109–142. [95] M. Rørdam, The stable rank and real rank of Z-absorbing C ∗-algebras, Inter. J. Math., 15, (2004) 1065-1084. [96] K. Thomsen, Homomorphisms between finite direct sums of circle algebras, Linear and Multilinear Algebra 32 (1992), 33–50. [97] K. Thomsen, Inductive limits of interval algebras: unitary orbits of positive elements, Math. Ann. 293 (1992), no. 1, 47–63. [98] K. Thomsen, Traces, unitary characters and crossed products by Z, Publ. Res. Inst. Math. Sci. 31 (1995), 1011–1029. [99] A. S. Toms, On the classification problem for nuclear C ∗-algebras. Ann. of Math. 167 (2008), 1029– 1044. [100] A. S. Toms, On the classification problem for nuclear C-algebras, Ann. of Math. (2) 167 (2008), 1029–1044. [101] J. Villadsen, Simple C-algebras with perforation, J. Funct. Anal. 154 (1998), no. 1, 110–116. [102] J. Villadsen, On the stable rank of simple C-algebras, J. Amer. Math. Soc. 12 (1999), no. 4, 1091–1102. [103] W. Winter. Decomposition rank and Z-stability. Invent. Math., 179 (2010), 229–301. [104] W. Winter. Nuclear dimension and Z-stability of pure C-algebras. Invent. Math., 187 (2012), 259–342. [105] W. Winter, Localizing the Elliott conjecture at strongly self-absorbing C-algebras, J. Reine Angew. Math. 692 (2014), 193–231. [email protected], [email protected], [email protected]. 283
1809.09134	3	1809	2019-07-14T06:42:09	Free products with amalgamation over central C*-subalgebras	[ "math.OA" ]	Let A and B be C-algebras whose quotients are all RFD, and let C be a central C-subalgebra in both A and B. We prove that the full amalgamated free product of A and B over C is then RFD. This generalizes Korchagin's result that amalgamated free products of commutative C-algebras are RFD. When applied to the case of a trivial amalgam, our methods recover the result of Exel-Loring for separable C-algebras. As corollaries to our theorem, we give sufficient conditions for amalgamated free products of maximally almost periodic (MAP) groups to have RFD C*-algebras and hence to be MAP.	math.OA	math	FREE PRODUCTS WITH AMALGAMATION OVER CENTRAL C∗-SUBALGEBRAS KRISTIN COURTNEY AND TATIANA SHULMAN Abstract. Let A and B be C∗-algebras whose quotients are all RFD, and let C be a central C∗-subalgebra in both A and B. We prove that the full amalgamated free product A ∗C B is then RFD. This generalizes Korchagin's result that amalgamated free products of commutative C∗-algebras are RFD. When applied to the case of trivial amalgam, our methods recover the result of Exel and Loring for separable C∗-algebras. As corollaries to our theorem, we give sufficient conditions for amalgamated free products of maximally almost periodic (MAP) groups to have RFD C∗-algebras and hence to be MAP. 1. Introduction a faithful embedding into a direct product of matrix algebras Qi∈I We say a C∗-algebra A is residually finite dimensional (RFD) if it has a separating family of finite dimensional representations. The direct sum of such a family yields Mki , and so RFD C∗-algebras can be thought of as those which are "block diagonalizable". In addition to this, various other characterizations of the property have been obtained over the years (notably [1], [14], [18], [10]), and numerous classes of C∗-algebras have been shown to be RFD. Residual finite dimensionality and its permanence properties can be found at the heart of some of the most important questions in operator algebras. Perhaps most famously, Kirchberg proved in [22] that Connes' Embedding Problem ([9]) is equiv- alent to the question of whether or not C∗(F2 × F2) is RFD. Outside of F2 × F2, new examples of groups whose full group C∗-algebra is RFD have become particularly welcome due to their relevance to problems of finding decidability algorithms for groups (see [15]). In the interest of finding more examples (and non-examples) of RFD C∗-algebras, C∗-algebraists have explored various permanence properties of residual finite dimen- sionality. In particular, when is it preserved under amalgamated free products? Given two C∗-algebras A and B, each containing a copy of another C∗-algebra C, their amalgamated free product A ∗C B is the unique C∗-algebra such that there exist maps ιA : A → A ∗C B and ιB : B → A ∗C B whose images generate A ∗C B and whose restrictions to C agree and such that A ∗C B is universal in this regard, meaning any pair of maps ψA : A → D and ψB : B → D into another C∗-algebra 2010 Mathematics Subject Classification. Primary 46L05; Secondary 47A67. The research of the first-named author was supported by the Deutsche Forschungsgemeinschaft (SFB 878 Groups, Geometry & Actions). The research of the second-named author was supported by the Polish National Science Centre grant under the contract number DEC- 2012/06/A/ST1/00256, by the grant H2020-MSCA-RISE- 2015-691246-QUANTUM DYNAMICS and Polish Government grant 3542/H2020/2016/2, and from the Eric Nordgren Research Fellowship Fund at the University of New Hampshire. 1 2 KRISTIN COURTNEY AND TATIANA SHULMAN D that agree on C must factor through ιA and ιB respectively. We call C the amalgam. The question of when the amalgamated product of two RFD C∗-algebras is again RFD is quite difficult in full generality. For instance, Connes' Embedding Problem can be reformulated as a question of whether a certain amalgamated free product of full group C∗-algebras is RFD. Indeed, for any discrete groups Λ ≤ G1, G2, we have that C∗(G1 ∗Λ G2) ≃ C∗(G1) ∗C∗(Λ) C∗(G2) (see [13, Lemma 3.1] for an argument). So we can write C∗(F2 × F2) ≃ C∗(F2 × Z) ∗C∗(F2) C∗(F2 × Z), where it follows from [8] and the fact that residual finite dimensionality is preserved by minimal tensor products that C∗(F2 × Z) ≃ C∗(F2) ⊗ C∗(Z) is RFD. In the case where the amalgam is trivial, i.e., when the amalgamated product is the full free product A ∗ B (or unital full free product A ∗C B), Exel and Loring showed in [14] that the amalgamated product of RFD C∗-algebras is RFD. For non-trivial amalgams, such a nice result is too much to ask. In fact this can fail even in the case of amalgamated products of matrix algebras, as [7, Example 2.4] shows. Nonetheless, necessary and sufficient conditions have been given for when amalgamated products of two separable RFD C∗-algebras over finite dimensional amalgams are RFD, first for matrix algebras by Brown and Dykema in [7], then for finite dimensional C∗-algebras by Armstrong, Dykema, Exel, and Li in [2], and finally for all separable RFD C∗-algebras by Li and Shen in [24]. Moving beyond finite dimensional amalgams, group theoretic results and restictions, which we outlilne below, indicate that the next natural class to study is amalgamated products of RFD C∗-algebras over central amalgams. In [23], Korchagin proved that any amalgamated product of two commutative C∗-algebras is RFD. This was the first and, until now, the only positive result on amalgamated products of RFD C∗-algebras over infinite dimensional C∗-algebras. In this paper we substantially generalize Korchagin's statement. Let us say that a C∗-algebra is strongly RFD if all its quotients are RFD. Here we prove Theorem. Let A and B be separable strongly RFD C∗-algebras and let C be a central C∗-subalgebra in both A and B. Then the amalgamated free product A ∗C B is RFD. In particular, since all commutative C∗-algebras are clearly strongly RFD, this gives the result of [23] with a different and shorter proof. Moreover, when applied to the case of trivial amalgam, our methods recover the result of Exel-Loring for separable C∗-algebras. In fact for a non-trivial C and strongly RFD A and B, or a trivial C and RFD A and B, given an irreducible representation ρ of A ∗C B and a sequence pn ↑ 1 of projections on a Hilbert space, we can construct finite dimensional representations of A ∗C B living on subspaces of pnH and ∗-strongly converging to ρ. As corollaries to our main theorem, we give sufficient conditions for amalgamated free products of discrete and locally compact groups to have RFD C∗-algebras. Corollary. Let G1 and G2 be virtually abelian discrete groups and let Λ be a central subgroup in both G1 and G2. Then the full group C∗-algebra of the amalgamated free product G1 ∗Λ G2 is RFD. FREE PRODUCTS WITH AMALGAMATION OVER CENTRAL C∗-SUBALGEBRAS 3 Corollary. Let G1 and G2 be separable locally compact groups, and let Λ be an open central subgroup in both. Assume moreover that G1 and G2 are Lie groups each containing a closed subgroup of finite index that is compact modulo its center or are projective limits of such Lie groups. Then the full group C∗-algebra of the amalgamated free product G1 ∗Λ G2 is RFD. To better understand why amalgamated products over central amalgams are par- ticularly favorable candidates for RFD C∗-algebras, we should pay heed to related results from group theory. The property of being RFD can be considered as a C∗- analogue of maximal almost periodicity and residual finiteness for groups. We say a discrete group is maximally almost periodic (MAP) if its finite dimensional unitary representations separate its elements, and we say it is residually finite (RF) if the same can be said for homomorphisms of the group into finite groups. Any discrete RF group is MAP, and by Mal'cev's theorem [26] the converse is true when the group is finitely generated. If the full group C∗-algebra C∗(G) of a discrete group G is RFD, then G is clearly MAP. (In fact, this also holds for locally compact groups, as was shown by Spronk and Wood in [30]; here the definition of MAP is only changed by the addition of the word "continuous".) Though the converse does not always hold (e.g. for SL3(Z) as shown in [4]), Bekka and Louvet show in [5] that when G is amenable, C∗(G) is RFD exactly when G is MAP. In particular, this means that a finitely generated amenable group is RF if and only if its full group C∗-algebra is RFD. This gives us a wealth of examples of RFD C∗-algebras coming from discrete groups. Beyond these, examples of groups with RFD C∗-algebras in- clude full group C∗-algebras of nonabelian free groups [8], virtually abelian groups [31], surface groups and fundamental groups of closed hyperbolic 3-manifolds that fiber over the circle [25], and many 1-relator groups with non-trivial center [19]. Permanence (or lack thereof) of RF and MAP under amalgamation has been well studied in group theory, and results and examples coming from this are valu- able guides for the analogous study in C∗-algebras. For instance, in [20], Higman constructed a pair of finitely generated metabelian groups G1 and G2 with com- mon cyclic subgroup Λ such that G1 ∗Λ G2 is not RF, and Baumslag proved in [3] that any two finitely generated, torsion free, nilpotent groups G1 and G2 that are not abelian have some common subgroup Λ such that G1 ∗Λ G2 is not RF. Since, in both cases, the amalgamated product is still a finitely generated group, we conclude that C∗(G1 ∗Λ G2) ≃ C∗(G1) ∗C∗(Λ) C∗(G2) cannot be RFD. These examples give an indication of how restrictive we must be in our choice of algebras and amalgam in the C∗-setting. On the other hand, in [3], Baumslag proved that the amalgamated product of polycyclic groups over a common central subgroup must be RF, and in [21] Kahn and Morris proved that the amalgamated product of two topological groups G1 and G2 over a common compact central subgroup Λ is MAP if and only if both groups are MAP. These point to central amalgams as the next promising frontier for amalgamated products of RFD C∗-algebras, now that that finite dimensional amalgams are completely understood. As informative as results for RF and MAP groups are for the study C∗-algebras, it is nice when we can return the favor. For G1, G2, and Λ as in either of our corollaries above, G1 ∗Λ G2 is MAP. To our knowledge, this fact is new in group theory and gives new examples outside of the result of Kahn and Morris. 4 KRISTIN COURTNEY AND TATIANA SHULMAN Acknowledgements. We would like to thank Mikhail Ershov and Ben Hayes for useful discussions on amalgamated free products of groups, Don Hadwin for teaching us his unitary orbits result, and Anton Korchagin and Wilhelm Winter for discussions that led to a better exposition of the paper. 2. Proofs Definition 2.1. A C∗-algebra is called strongly RFD if all its quotients are RFD. Here are two examples of classes of C∗-algebras that are strongly RFD: Example 2.2 (FDI C∗-algebras). In [10], we call C∗-algebras whose irreducible representations are all finite dimensional FDI. Since irreducible representations separate the elements of a C∗-algebra, any FDI C∗-algebra is RFD. Moreover any quotient of an FDI C∗-algebra is again FDI since an irreducible representation of a quotient of a C∗-algebra gives rise to an irreducible representation of the C∗-algebra. Particular cases of FDI C∗-algebras are subhomogeneous C∗-algebras (i.e., those whose irreducible representations are all of dimension no more than some fixed n < ∞) and continuous fields of finite dimensional C∗-algebras (as defined in [12]). Example 2.3 (RFD just-infinite C∗-algebras). In [16], Grigorchuk, Musat, and Rørdam defined a C∗-algebra to be just-infinite when it is infinite dimensional and all of its proper quotients are finite dimensional. In the same paper, they demonstrate the existence of just-infinite RFD C∗-algebras. Moreover, they prove that there are examples of non-exact, just-infinite RFD C∗-algebras, which means, in particular, that strongly RFD C∗-algebras need not be nuclear. It is worth noting that no C∗-algebra can be both FDI and just-infinite. Indeed, by [16, Lemma 5.4], no RFD just-infinite C∗-algebra is of type I; while, on the other hand, all FDI algebras are type I. The following result of Hadwin will be crucial for the proof of the main theorem. It is very close to Theorem 4.3 in [17]. The particular formulation below was given in private communication, and so we include a brief proof here. Lemma 2.4 (Hadwin [17]). Let {en} be an orthonormal basis in a separable Hilbert space H, a, b, c, d ∈ B(H) and x = (cid:18) a b d (cid:19). Then for any unitary wn : H → c H ⊕ H such that wnek = (ek, 0), 1 ≤ k ≤ n, w∗ operator topology. nxwn converge to a in the weak Moreover we have convergence in the strong operator topology if and only if c = 0 and in the ∗-strong operator topology if and only if c = b = 0. Proof. We will show here only the second claim, that w∗ nxwn converges to a strongly if and only if c = 0. The third claim, which is the only one we use in this paper, follows directly from the second one. The first claim is proved similarly. Let ǫ > 0 and ξ = P∞ and write ξ′ = PN j=1 ξjej ∈ H. Choose N > 0 so that kPj>N ξjejk < ǫ 8kxk , j=1 ξjej and ξ′′ = ξ − ξ′. Then for each n ≥ N , we have FREE PRODUCTS WITH AMALGAMATION OVER CENTRAL C∗-SUBALGEBRAS 5 (2.1) kw∗ nxwnξ − aξk = kw∗ = kw∗ = kw∗ nx(ξ′, 0) − aξ′ + (w∗ n(aξ′, cξ′) − aξ′ + (w∗ n(0, cξ′) + w∗ nxwn − a)ξ′′ nxwn − a)ξ′′k n(aξ′, 0) − aξ′ + (w∗ nxwn − a)ξ′′k. Choose M ≥ N so that kpM aξ′ − aξ′k < ǫ/8 where pM is the projection onto span{e1, ..., eM }. Then we have kw∗ nxwn − a)ξ′′k < ǫ/4, and kw∗ n(0, cξ′)k = kcξ′k for all n ≥ M . So (2.1) gives us n(aξ′, 0) − aξ′k < ǫ/4, k(w∗ kcξ′k − ǫ/2 < kw∗ nxwnξ − aξk < kcξ′k + ǫ/2 for all n ≥ M . If c = 0, then we have kw∗ there exists an n > M such that kw∗ nxwnξ − aξk < ǫ/4, then nxwnξ − aξk < ǫ. On the other hand, if kcξk < kcξ′k + ǫ 4 < kw∗ nxwnξ − aξk + 3ǫ 4 < ǫ. Since ǫ and ξ were arbitrary, we conclude that w∗ exactly when cξ = 0 for all ξ ∈ H. nxwn converges strongly to a (cid:3) We will need one more lemma, which is essentially the statement 5 in [18, Lemma 1], where it is formulated in slightly different terms. For the reader's convenience we give a proof of it here. Lemma 2.5 (Hadwin [18]). Let u ∈ B(H) be a unitary operator and let pn ∈ B(H), n ∈ N, be a sequence of finite rank projections such that pn ↑ 1 in ∗-strong operator topology. Then for each n ∈ N there is a unitary operator un ∈ B(pnH) such that un → u in ∗-strong operator topology, where we view pnH as a subspace of H in the obvious way. Proof. Write u as u = e2πia, where −1 ≤ a ≤ 1, and let un = e2πipnapn . Since pnapn → a in the ∗-strong topology and since the functional calculus is continuous with respect to the ∗-strong operator topology, one has un → u ∗-strongly. (cid:3) Our main theorem is for both unital and non-unital cases, meaning that C- algebras can be either unital or non-unital, and in the case both of them are unital, the amalgamated free product can be either unital or non-unital. Theorem 2.6. Let A and B be separable strongly RFD C∗-algebras and let C be a central C∗-subalgebra in both A and B. Then the amalgamated free product A ∗C B is RFD. Proof. Let 0 6= x ∈ A ∗C B. We construct a finite dimensional representation σ of A ∗C B such that σ(x) 6= 0. There exist an irreducible representation ρ of A ∗C B on a separable Hilbert space H and a unit vector ξ ∈ H such that kρ(x)ξk ≥ kxk 2 . Let iA and iB denote the standard embeddings of A and B into A ∗C B. Choose K ∈ N, a(k) i ∈ B, i = 1, . . . , N (k), k = 1, . . . , K such that for i ∈ A, b(k) x = K Xk=1 (2.2) we have (2.3) iA(a(k) 1 )iB(b(k) 1 ) . . . iA(a(k) N (k) )iB(b(k) N (k) ), kx − xk ≤ kxk 8 . 6 KRISTIN COURTNEY AND TATIANA SHULMAN (The sum in (2.2) might also contain monomials starting with an element from iB(B) or ending with an element from iA(A), but we will assume that it does not. This does not change anything in the proof and we do it just to avoid notational nightmare.) Then (2.4) kρ (x) ξk ≥ 3kxk 8 . We denote the representations of A and B induced by ρ with ρA and ρB respectively, i.e., ρA(a) = ρ(iA(a)), ρB(b) = ρ(iB(b)), for each a ∈ A, b ∈ B. Let N = max 1≤k≤K N (k), k = 1, . . . , K, i = 1, . . . , N (k)}, k = 1, . . . , K, i = 1, . . . , N (k)}, E = {a(k) F = {b(k) i i and i G = {ξ}[{ρB(b(k) [{ρA(a(k) i )ρA(a(k) i+1)ρB(b(k) i+1) . . . ρA(a(k) N (k) )ρB(b(k) )ρB(b(k) i )ρA(a(k) i+1) . . . ρA(a(k) N (k) )ρB(b(k) N (k) )ξ i = 1, . . . , N (k), k = 1, . . . , K} N (k) )ξ i = 2, . . . , N (k), k = 1, . . . , K}. Since ρ is irreducible and C is a central subalgebra in A ∗C B, for each c ∈ C there is λ(c) ∈ C such that ρ(c) = λ(c)1. Since A and B are strongly RFD, ρA(A) and ρB(B) are RFD. Note that, unless ρ(C) is zero, ρA(A) and ρB(B) are both unital, regardless of whether or not A ∗C B is. In this case let ¯π1, ¯π2, . . . be a countable separating family of unital finite dimensional representations of ρA(A) and let πi = ¯πi ◦ ρA. Let ¯π′ 2, . . . be a countable separating family of unital finite dimensional representations of ρB(B) and let π′ 1, ¯π′ i = ¯π′ i ◦ ρB. In the case ρ(C) is zero, we replace unital representations by nondegenerate ones. Let Ni = dim πi, N ′ i = dim π′ i. Let πi = π(N ′ i ) i i. Notice that for each c ∈ C, i , π′ i = π′(Ni) . Then dim πi = dim π′ (2.5) πi(c) = λ(c)1 = π′ i(c). Let π be a direct sum of all πi's where each one is repeated infinitely many times, say π = π1 ⊕ (π1 ⊕ π2) ⊕ (π1 ⊕ π2 ⊕ π3) ⊕ . . . . We consider π as a representation of A on B(H) with respect to some decomposition H = H1 ⊕ H2 ⊕ H3 ⊕ . . . where H1 ∼= CN1N ′ 1, H2 ∼= CN1N ′ 1 ⊕ (π′ π′ = π′ 1+N2N ′ 1 ⊕ π′ 2 , . . .. Let π′ : B → B(H) be defined by 2) ⊕ (π′ 3) ⊕ . . . 1 ⊕ π′ 2 ⊕ π′ with respect to the same decomposition of H. Let pi ∈ B(H) be the orthogonal projection onto H1⊕. . .⊕Hi. It is easy to see that each piH is an invariant subspace for π and π′. FREE PRODUCTS WITH AMALGAMATION OVER CENTRAL C∗-SUBALGEBRAS 7 Now, for any a ∈ A, we have that π(a) = 0 iff ρA(a) = 0, and rank(π(a)) = ∞ when π(a) 6= 0. So by Voiculescu's theorem, ρA ⊕ π is approximately unitarily equivalent to π. Hence there exists a unitary u : H ⊕ H → H such that for all d ∈ E we have where (cid:13)(cid:13)(cid:13)(cid:13) (cid:18) ρA(d) π(d) (cid:19) − u∗π(d)u(cid:13)(cid:13)(cid:13)(cid:13) ≤ δ, δ = kxk . 80 · KN By Lemma 2.4 there exist unitaries wm : H → H ⊕ H such that for all a ∈ A, w∗ m(cid:18) ρA(a) π(a) (cid:19) wm → ρA(a) in the ∗-strong topology. In particular there exists a unitary w : H → H ⊕ H such that for all d ∈ E, η ∈ G (cid:13)(cid:13)(cid:13)(cid:13) ρA(d)η − w∗(cid:18) ρA(d) Hence for all d ∈ E, η ∈ G, π(d) (cid:19) wη(cid:13)(cid:13)(cid:13)(cid:13) < δ. (2.6) kρA(d)η − w∗u∗π(d)uwηk ≤ 2δ. Similarly we find unitaries u′, w′ such that (2.7) kρB(d)η − w′∗u′∗π′(d)u′w′ηk ≤ 2δ. for all d ∈ F , η ∈ G. Applying Lemma 2.5 to uw and u′w′, we find M ∈ N and unitaries v and v′ on pmH, which is identified with a subspace of H, such that kη − pmηk ≤ δ, k(uw − v)pmηk ≤ δ, k(u′w′ − v′)pmηk ≤ δ, for all η ∈ G, for all d ∈ E, η ∈ G and for all d ∈ F , η ∈ G. k(w∗u∗ − v∗)π(d)uwpmηk ≤ δ, k(w′∗u′∗ − v′∗)π′(d)u′w′pmηk ≤ δ, Now we define finite dimensional representations σA : A → pmB(H)pm and σB : B → pmB(H)pm by σA(a) = v∗(pmπ(a)pm)v, σB(b) = v′∗(pmπ′(b)pm)v′, for any a ∈ A, b ∈ B. Then by (2.5) σA(c) = λ(c)1pmH = σB(c), for any c ∈ C. As σA and σB agree on C, we obtain a finite dimensional represen- tation σ : A ∗C B → pmB(H)pm. 8 KRISTIN COURTNEY AND TATIANA SHULMAN Since pmH is invariant subspace for π, we have pmπ(a)pmv = π(a)v, for each a ∈ A. Using this we obtain (2.8) kw∗u∗π(d)uwη − σA(d)pmηk = kw∗u∗π(d)uwη − v∗(pmπ(d)pm)vpmηk ≤kw∗u∗π(d)uw(η − pmη)k + kw∗u∗π(d)uwpmη − v∗π(d)vpmηk ≤kw∗u∗π(d)uw(η − pmη)k + k(w∗u∗ − v∗)π(d)uwpmηk + kv∗π(d)(uw − v)pmηk ≤3δ, for any d ∈ E, η ∈ G. By (2.6) and (2.8), we have (2.9) kρA(d)η − σA(d)pmηk ≤ 5δ, for any d ∈ E, η ∈ G. Similarly we obtain (2.10) kρB(d)η − σB(d)pmηk ≤ 5δ, for any d ∈ F , η ∈ G. By repeated uses of (2.9) and (2.10), kσA(a(k) 1 ) . . . σA(a(k) 1 )σB(b(k) N (k) )σB(b(k) N (k) )pmξ ≤ kσA(a(k) 1 )σB(b(k) 1 )ρB(b(k) − ρA(a(k) 1 ) . . . σA(a(k) 1 ) . . . ρA(a(k) N (k) )pm(cid:16)σB(b(k) N (k) )pmρB(b(k) N (k) )ξ N (k) )ρB(b(k) N (k) )ξk N (k) )pmξ − ρB(b(k) N (k) )(cid:17) ξk + kσA(a(k) 1 )σB(b(k) 1 ) . . . σA(a(k) ≤ 5δ + kσA(a(k) 1 )σB(b(k) 1 )ρB(b(k) − ρA(a(k) 1 ) . . . σA(a(k) − ρA(a(k) 1 ) . . . ρA(a(k) N (k) )ξ N (k) )pmρB(b(k) 1 )ρB(b(k) 1 ) . . . ρA(a(k) N (k) )ρB(b(k) N (k) )ξk N (k) )ρB(b(k) N (k) )ξk ≤ 5δ + kσA(a(k) + kσA(a(k) 1 )σB(b(k) 1 )σB(b(k) 1 ) . . . σB(b(k) 1 ) . . . σB(b(k) − ρA(a(k) N (k)−1)pm(cid:16)σA(a(k) N (k)−1)pmρA(a(k) 1 )ρB(b(k) N (k) ) − ρA(a(k) N (k) )(cid:17) (ρB(b(k) N (k) )ξ))k N (k) )ρB(b(k) N (k) )ξ 1 ) . . . ρA(a(k) N (k) )ρB(b(k) N (k) )ξk ≤ . . . ≤ 2N (k)5δ ≤ 10N δ. Hence (2.11) kσ(x)pmξ − ρ(x)ξk ≤ 10N Kδ = kxk 8 . Combining (2.11), (2.4) and (2.3) we obtain kσ(x)pmξk ≥ kσ (x) pmξk − kσ (x − x) pmξk Thus σ(x) 6= 0. ≥ kρ (x) ξk − kσ(x)pmξ − ρ (x) ξk − kσ (x − x) pmξk ≥ kxk 8 . (cid:3) Remark 2.7. When the amalgam C is trivial (that is C = C1 in the unital case and C = 0 in the non-unital case), the assumption that the C∗-algebras are strongly RFD can be omitted. One needs them only to be RFD because starting with any pair of separating families of unital (or nondegenerate, in the non-unital case) FREE PRODUCTS WITH AMALGAMATION OVER CENTRAL C∗-SUBALGEBRAS 9 representations of A and B would lead to representations coinciding on C. Thus our proof recovers the Exel-Loring result that free products of separable RFD C- algebras are RFD. Remark 2.8. It follows from the proof that given an irreducible representation ρ of A ∗C B on a separable Hilbert space H and an arbitrary sequence pn ↑ 1 of projections on H, we actually can construct finite dimensional representations of A ∗C B living on subspaces of pnH and ∗-strongly converging to ρ. Corollary 2.9. (Korchagin [23]) The amalgamated free product of commutative C∗-algebras is RFD. Corollary 2.10. Let G1 and G2 be virtually abelian discrete groups and let Λ be a central subgroup in both G1 and G2. Then C∗(G1 ∗Λ G2) ≃ C∗(G1) ∗C∗(Λ) C∗(G2) is RFD. Proof. It was proved in [31] that C∗-algebras of discrete virtually abelian groups are subhomogeneous. Hence they are strongly RFD and Theorem 2.6 applies. The isomorphism is well-known (see e.g. [13, Lemma 3.1]). (cid:3) We also have an application for amalgamated free products of locally compact groups. However, to our knowledge, the isomorphism from Corollary 2.10 has not been addressed in the case of locally compact groups. Thus, before we can proceed, we must prove that this isomorphism holds, at least in the case where the common subgroup is open and central. The following useful fact was communicated to us by Ben Hayes. Proposition 2.11. Suppose G is a locally compact Hausdorff group and Λ ≤ G is an open subgroup. Then the restriction of any nondegenerate representation π : L1(G) → B(H) to the natural copy of L1(Λ) inside L1(G) is also nondegenerate. Proof. Since Λ ⊆ G is open, we can naturally identify L1(Λ) as a subalgebra of L1(G) by extending compactly supported functions on Λ to be zero off their sup- ports on G. Moreover, there exists an approximate unit of L1(G) that is contained in L1(Λ). Indeed, take a neighborhood basis of the unit of G consisting of open sets in Λ with compact closure. Since Λ is open in G, this forms a neighborhood basis of the unit in G, and the normalized characteristic functions on these sets give an approximate unit. Now, under a nondegenerate representation π : L1(G) → B(H), any approximate unit of L1(G) will converge strongly to the identity. Since our particular approximate unit was contained in L1(Λ), the same will hold when π is restricted to L1(Λ). It follows that πL1(Λ) is also nondegenerate. (cid:3) Proposition 2.12. Let G1 and G2 be locally compact Hausdorff groups and Λ ≤ G1, G2 an open central subgroup of both. Then Γ = G1 ∗Λ G2 is a locally compact Hausdorff group and C∗(Γ) ≃ C∗(G1) ∗C∗(Λ) C∗(G2). Proof. By [21, Theorem 5], since Λ is open and central, we know that Γ is locally compact and Hausdorff. Moreover, by the remark following Corollary 3 in [21], since Λ is open in G1 and G2, it follows that G1 and G2 are open in Γ. Choose the Haar measures on G1 and G2, inherited from the embeddings ιi : Gi → Γ, i = 1, 2. As in Proposition 2.11, we have natural embeddings L1(Λ) ֒→ L1(Gi) for i = 1, 2, and similarly, we can embed L1(G1), L1(G2) ֒→ L1(Γ) so that the 10 KRISTIN COURTNEY AND TATIANA SHULMAN embeddings agree on the respective copies of L1(Λ). The same arguments as in the discrete setting show that these embeddings extend to the full group C∗-algebras, and moreover, the induced embeddings ιi : C∗(Gi) ֒→ C∗(Γ) agree on C∗(Λ) (see e.g. [29, Proposition 8.8]). To verify that C∗(Γ) ≃ C∗(G1) ∗C∗(Λ) C∗(G2), it suffices to check that C∗(Γ) = C∗(ι1(C∗(G1))∪ι2(C∗(G2))) and that C∗(Γ) satisfies the desired universal property. Let B = C∗(ι1(C∗(G1)) ∪ ι2(C∗(G2))). To show B = C∗(Γ), it suffices to show that for every nondegenerate representation π : C∗(Γ) → B(H), W ∗(π(C∗(Γ))) = W ∗(π(B)). To that end, let π : C∗(Γ) → B(H) be a nondegenerate representation, and let πi = π ◦ ιi : C∗(Gi) → B(H), for i = 1, 2. Since ιi(Gi) are both open in Γ, π1 and π2 are still nondegenerate by Proposition 2.11. Since there is a 1-1 correspondence between nondegenerate representations of a full group C∗-algebra of a locally compact group and strongly continuous unitary representations of the group (see e.g. [12, 13.3.5] or [11, p. 183-184]), π, πi, i = 1, 2, correspond to unique strongly continuous unitary representations π : Γ → U (H) and πi : Gi → U (H), i = 1, 2. Moreover, W ∗(π(C∗(Γ))) = W ∗(π(Γ)) and W ∗(πi(C∗(Gi))) = W ∗( πi(Gi)), i = 1, 2 (see again [12, 13.3.5] or [11, p. 183-184]). Since ι1(G1) ∪ ι2(G2) generates Γ, we compute W ∗(π(B)) = W ∗ (π1(C∗(G1)) ∪ π2(C∗(Gi))) = W ∗ (W ∗(π1(C∗(G1))), W ∗(π2(C∗(G2)))) = W ∗ (W ∗(π1(G1)), W ∗(π2(G2))) = W ∗(π(Γ)) = W ∗(π(C∗(Γ))). Now, suppose φi : C∗(Gi) → B(H), i = 1, 2 are nondegenerate representations that agree on C∗(Λ). Again, these correspond to unique strongly continuous unitary representations φi : Gi → U (H), i = 1, 2, which agree on Λ. The universal property of Γ gives a unique strongly continuous unitary representation ψ : Γ → U (H) such that ψιi = φi for i = 1, 2. This induces a nondegenerate representation ψ : C∗(Γ) → B(H). Moreover, for i = 1, 2 and f ∈ L1(Gi) ιi(f )(t) ψ(t)dt = Zιi(Gi) ιi(f )(ιi(s)) ψ(ιi(s))dιi(s) ιi(f )(t) ψ(t)dt ψιi(f ) = ZΓ = ZG1 = ZG1 f (s) φi(s)ds = φi(f ). Thus C∗(Γ) has the universal property of C∗(G1) ∗C∗(Λ) C∗(G2). (cid:3) Corollary 2.13. Let G1 and G2 be separable locally compact groups, and let Λ be an open central subgroup in both. Assume moreover that G1 and G2 are Lie groups each containing a closed subgroup of finite index that is compact modulo its center or are projective limits of such Lie groups. Then the full group C∗-algebra of the amalgamated free product G1 ∗Λ G2 is RFD. Proof. From [21], we know G1 ∗Λ G2 exists and is a locally compact Hausdorff topological group (actually even a Lie group). As was proved in [27], the groups described are exactly the locally compact groups whose irreducible representations FREE PRODUCTS WITH AMALGAMATION OVER CENTRAL C∗-SUBALGEBRAS 11 are all finite dimensional. Hence, their full group C∗-algebras are FDI and, by assumption, separable. So, Theorem 2.6 again applies. (cid:3) Moreover, the amalgamated groups from each of these corollaries are MAP (using results from [30] in the locally compact case). But groups with FDI C∗-algebras are not the only ones covered by Theorem 2.6. For example, in [6], the authors prove that there exist infinite discrete groups with just-infinite RFD full group C∗-algebras. On the other hand, it follows from [31] that a discrete group has FDI full group C∗-algebra if and only if it is virtually abelian. Just as their C∗-algebras form disjoint classes, so must these groups. This means, Corollaries 2.10 and 2.13 can be phrased in more generality, but the generalized statements sound oblique without knowing which groups yield strongly RFD C∗-algebras, which leads us to the following question(s). Question 2.14. Is there a nice characterization of all (discrete) groups that have strongly RFD C∗-algebras? The questions are particularly curious because the class of all such groups does not seem to fit neatly into any well-known classes of MAP or RF groups. Consider the discrete case. Clearly every quotient of such a group is MAP. Moreover, it follows from Rosenberg's theorem that all discrete groups with strongly RFD full group C∗-algebras must be amenable since their reduced group C∗-algebras must be RFD and hence quasidiagonal. With Bekka and Louvet's aforementioned result from [5] in mind, one may make the naive guess that it is sufficient to be amenable with all quotients MAP, but this is wrong. Every quotient of a finitely generated nilpotent group is also finitely generated, and hence RF and amenable. However, if the full group C∗-algebra of a finitely generated nilpotent group is strongly RFD, then the group must be virtually abelian. This is because the C∗-algebra gener- ated by any irreducible representation of a nilpotent group is simple ([28]) and, as Thoma has shown ([31]), having only finite dimensional irreducible representations is equivalent to being virtually abelian. References [1] R. J. Archbold, On residually finite-dimensional C∗-algebras, Proc. Amer. Math. Soc. 123 (1995), no. 9, 2935 -- 2937. MR 1301006 [2] Scott Armstrong, Ken Dykema, Ruy Exel, and Hanfeng Li, On embeddings of full amal- gamated free product C∗-algebras, Proc. Amer. Math. Soc. 132 (2004), no. 7, 2019 -- 2030. MR 2053974 [3] Gilbert Baumslag, On the residual finiteness of generalised free products of nilpotent groups, Trans. Amer. Math. Soc. 106 (1963), 193 -- 209. MR 144949 [4] M. B. Bekka, On the full C∗-algebras of arithmetic groups and the congruence subgroup problem, Forum Math. 11 (1999), no. 6, 705 -- 715. MR 1725593 [5] M. B. Bekka and N. Louvet, Some properties of C∗-algebras associated to discrete linear groups, C∗-algebras (Munster, 1999), Springer, Berlin, 2000, pp. 1 -- 22. MR 1796907 [6] V. Belyaev, R. Grigorchuk, and P. Shumyatsky, On just-infiniteness of locally finite groups and their C∗-algebras, Bull. Math. Sci. 7 (2017), no. 1, 167 -- 175. MR 3625854 [7] Nathanial P. Brown and Kenneth J. Dykema, Popa algebras in free group factors, J. Reine Angew. Math. 573 (2004), 157 -- 180. MR 2084586 [8] Man Duen Choi, The full C∗-algebra of the free group on two generators, Pacific J. Math. 87 (1980), no. 1, 41 -- 48. MR 590864 [9] A. Connes, Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1, Ann. of Math. (2) 104 (1976), no. 1, 73 -- 115. MR 0454659 12 KRISTIN COURTNEY AND TATIANA SHULMAN [10] Kristin Courtney and Tatiana Shulman, Elements of C∗-algebras attaining their norm in a finite-dimensional representation, Canad. J. Math. 71 (2019), no. 1, 93 -- 111. MR 3928257 [11] Kenneth R. Davidson, C∗-algebras by example, Fields Institute Monographs, vol. 6, American Mathematical Society, Providence, RI, 1996. MR 1402012 [12] Jacques Dixmier, C∗-algebras, North-Holland Publishing Co., Amsterdam-New York-Oxford, 1977, Translated from the French by Francis Jellett, North-Holland Mathematical Library, Vol. 15. MR 0458185 [13] Søren Eilers, Tatiana Shulman, and Adam P.W. Sørensen, C∗-stability of discrete groups, preprint, arXiv:1808.06793. [14] Ruy Exel and Terry A. Loring, Finite-dimensional representations of free product C∗- algebras, Internat. J. Math. 3 (1992), no. 4, 469 -- 476. MR 1168356 [15] Tobias Fritz, Tim Netzer, and Andreas Thom, Can you compute the operator norm?, Proc. Amer. Math. Soc. 142 (2014), no. 12, 4265 -- 4276. MR 3266994 [16] Rostislav Grigorchuk, Magdalena Musat, and Mikael Rø rdam, Just-infinite C∗-algebras, Comment. Math. Helv. 93 (2018), no. 1, 157 -- 201. MR 3777128 [17] Don Hadwin, An operator-valued spectrum, Indiana Univ. Math. J. 26 (1977), no. 2, 329 -- 340. MR 0428089 [18] , A lifting characterization of RFD C∗-algebras, Math. Scand. 115 (2014), no. 1, 85 -- 95. MR 3250050 [19] Don Hadwin and Tatiana Shulman, Stability of group relations under small Hilbert-Schmidt perturbations, J. Funct. Anal. 275 (2018), no. 4, 761 -- 792. MR 3807776 [20] Graham Higman, A finitely related group with an isomorphic proper factor group, J. London Math. Soc. 26 (1951), 59 -- 61. MR 0038347 [21] M. S. Khan and Sidney A. Morris, Free products of topological groups with central amalga- mation. II, Trans. Amer. Math. Soc. 273 (1982), no. 2, 417 -- 432. MR 667154 [22] Eberhard Kirchberg, On nonsemisplit extensions, tensor products and exactness of group C∗-algebras, Invent. Math. 112 (1993), no. 3, 449 -- 489. MR 1218321 [23] Anton Korchagin, Amalgamated free products of commutative C∗-algebras are residually finite-dimensional, J. Operator Theory 71 (2014), no. 2, 507 -- 515. MR 3214649 [24] Qihui Li and Junhao Shen, A note on unital full amalgamated free products of RFD C∗- algebras, Illinois J. Math. 56 (2012), no. 2, 647 -- 659. MR 3161345 [25] Alexander Lubotzky and Yehuda Shalom, Finite representations in the unitary dual and Ramanujan groups, Discrete geometric analysis, Contemp. Math., vol. 347, Amer. Math. Soc., Providence, RI, 2004, pp. 173 -- 189. MR 2077037 [26] A. Malcev, On isomorphic matrix representations of infinite groups, Rec. Math. [Mat. Sbornik] N.S. 8 (50) (1940), 405 -- 422. MR 0003420 [27] Calvin C. Moore, Groups with finite dimensional irreducible representations, Trans. Amer. Math. Soc. 166 (1972), 401 -- 410. MR 302817 [28] Calvin C. Moore and Jonathan Rosenberg, Groups with T1 primitive ideal spaces, J. Func- tional Analysis 22 (1976), no. 3, 204 -- 224. MR 0419675 [29] Gilles Pisier, Introduction to operator space theory, London Mathematical Society Lecture Note Series, vol. 294, Cambridge University Press, Cambridge, 2003. MR 2006539 [30] Nico Spronk and Peter Wood, Diagonal type conditions on group C∗-algebras, Proc. Amer. Math. Soc. 129 (2001), no. 2, 609 -- 616. MR 1800241 [31] Elmar Thoma, Eine Charakterisierung diskreter Gruppen vom Typ I, Invent. Math. 6 (1968), 190 -- 196. MR 0248288 Mathematical Institute, WWU Munster, Einsteinstr. 62, Munster E-mail address: [email protected] Department of Mathematical Physics and Differential Geometry, Institute of Math- ematics of Polish Academy of Sciences, Warsaw E-mail address: [email protected]
1902.00195	1	1902	2019-02-01T06:17:39	A note on irreducible quadrilaterals of $II_1$ factors	[ "math.OA" ]	Given any finite index quadrilateral $(N, P, Q, M)$ of $II_1$-factors, the notions of interior and exterior angles between $P$ and $Q$ were introduced in \cite{BDLR2017}. We determine the possible values of these angles when the quadrilateral is irreducible and the subfactors $N \subset P$ and $N \subset Q$ are both regular in terms of the cardinalities of the Weyl groups of the intermediate subfactors. For a more general quadruple, an attempt is made to determine the values of angles by deriving expressions for the angles in terms of the common norm of two naturally arising auxiliary operators and the indices of the intermediate subfactors of the quadruple. Finally, certain bounds on angles between $P$ and $Q$ are obtained, which enforce some restrictions on the index of $N \subset Q$ in terms of that of $N \subset P$.	math.OA	math	A NOTE ON IRREDUCIBLE QUADRILATERALS OF II1 FACTORS KESHAB CHANDRA BAKSHI AND VED PRAKASH GUPTA Abstract. Given any finite index quadrilateral (N, P, Q, M ) of II1-factors, the notions of in- terior and exterior angles between P and Q were introduced in [1]. We determine the possible values of these angles when the quadrilateral is irreducible and the subfactors N ⊂ P and N ⊂ Q are both regular in terms of the cardinalities of the Weyl groups of the intermediate subfactors. For a more general quadruple, an attempt is made to determine the values of angles by deriving expressions for the angles in terms of the common norm of two naturally arising auxiliary opera- tors and the indices of the intermediate subfactors of the quadruple. Finally, certain bounds on angles between P and Q are obtained, which enforce some restrictions on the index of N ⊂ Q in terms of that of N ⊂ P . 9 1 0 2 b e F 1 ] . A O h t a m [ 1 v 5 9 1 0 0 . 2 0 9 1 : v i X r a 1. Introduction A quadrilateral is a quadruple (N, P, Q, M ) of II1-factors such that N ⊂ P, Q ⊂ M , N = P ∧ Q, M = P ∨ Q and [M : N ] < ∞; it is called irreducible if N ⊂ M is irreducible. The Weyl group of a finite index II1-subfactor N ⊂ M is the quotient group G := NM (N )/U(N ). This article concentrates mainly on the analysis of such quadrilaterals from the perspectives of (a) calculating the interior and exterior angles between P and Q as was introduced in [1], (b) understanding the Weyl group of N ⊂ M in terms of those of N ⊂ P and N ⊂ Q, and (c) establishing a relationship between the above two aspects. Unlike the notion of set of angles by Sano and Watatani ([13]), the interior and exterior angles are both single entities and are seemingly more calculable, as we show in Section 2.2 by making some explicit calculations. As an important application of the notion of interior angle, the authors in [1] were able to improve a result of Longo [9] by providing a better bound for the number of intermediate subfactors of a given subfactor. A natural question that struck us, after the appearance of [1], was to determine the possible set of values that interior and exterior angles can attain. This article is devoted to this theme. In general, it looks like a tough nut to crack. However, in the irreducible set up, we see that these angles take definitive values. In Section 2, we discuss various generalities and formulae related to the interior and exterior angles and employ them to compute angles between two intermediate subfators associated with a quadruple of crossed product algebras. In Section 3, our main focus is on irreducible quadrilaterals (N, P, Q, M ) for which N ⊂ P and N ⊂ Q are both regular. Jones, in [5], had asked whether an irreducible regular subfactor is always a group subfactor. Making use of a theorem of Sutherland [14] on vanishing of cohomologies, Popa [11] and Kosaki [8] (for properly infinite case) answered Jones' question in the affirmative, which was announced earlier for the hyperfinite case by Ocnenanu in 1986. Later, Hong gave an explicit realization of the same in [4]. Using Hong's technique, we deduce (in Theorem 3.5) that an irreducible quadrilateral (N, P, Q, M ) with regular N ⊂ P and N ⊂ Q can be realized as a quadrilateral of crossed product algebras through outer actions of Weyl groups. Using this realization and the calculations of Section 2.2, we provide a direct relationship between the interior and exterior angles between P and Q and the Weyl groups of N ⊂ P and N ⊂ Q in: The first named author was supported by a postdoctoral fellowship of the National Board for Higher Mathematics (NBHM), India. 1 2 K C BAKSHI AND V P GUPTA Theorem 3.9 Let (N, P, Q, M ) be an irreducible quadrilateral such that N ⊂ P and N ⊂ Q are both regular. Then, α(P, Q) = π/2, i.e., (N, P, Q, M ) is a commuting square, and cos β(P, Q) = G HK − 1 , p[G : H] − 1p[G : K] − 1 where H, K and G denote the Weyl groups of N ⊂ P , N ⊂ Q and N ⊂ M , respectively. In particular, (N, P, Q, M ) is a cocommuting square if and only if G = HK. Section 4 dwells around the main theme of this article, viz., to determine the possible values of the interior and exterior angles. We first derive expressions for the angles in terms of the common norm λ of two naturally arising auxiliary operators and the indices of the intermediate subfactors of the quadruple (in Proposition 4.5). Then, in the irreducible setup, we exploit these expressions to obtain some definitive values for angles by making use of above relationship between angles and Weyl groups, a theorem of Popa [12] wherein he determines the possible values taken by the set Λ(M, N ) of relative dimensions of projections, and the values attained by the polynomials Pn(x), n ≥ 0, as introduced by Jones in [5]. The results that we prove are: Theorem 4.11 Let (N, P, Q, M ) be a quadruple with N ⊂ M irreducible and let 0 < t ≤ 1/2 be such that t(1 − t) = τ . If r/λ ≥ t, then, for some k ≥ 0, where r := [Q:N ] [M:P ] . Theorem 4.12 Let (N, P, Q, M ) be an irreducible quadrilateral such that N ⊂ P and N ⊂ Q , where are both regular and suppose [P : N ] = 2. Then, cos(β(P, Q)) = P2(m/2) pP2(δ2/2)P3(m/2) m = [M : Q] ∈ N and, as usual δ :=p[M : N ]. As a 'geometric' consequence, in Corollary 4.13, we see that if both N ⊂ P and N ⊂ Q have index 2, then the exterior angle β(P, Q) > π/3. Finally, while analyzing a quadrilateral intuitively as a picture in the plane (Figure 1), loosely speaking, we realize in Section 5 that the angles impose some sort of rigidity on the lengths of its sides. This could be inferred as a direct consequence of certain bounds on interior and exterior angles that we obtain in: Theorem 5.1 Let (N, P, Q, M ) be a finite index irreducible quadruple such that N ⊂ P is regular. Then, and cos(cid:0)α(P, Q)(cid:1) ≤ s [P : N ] − 1 [Q : N ] − 1! cos(β(P, Q)) ≤ s [P : N ] − r [Q : N ] − r!. and And, if r/λ < t, then, and cos(α(P, Q)) ≤ cos(β(P, Q)) ≤ cos(α(P, Q)) = cos(β(P, Q)) = . 1 t − 1 [P : N ][Q : N ](1 − t) − 1 p[P : N ] − 1p[Q : N ] − 1 p[M : P ] − 1p[M : Q] − 1 Pk−1(τ ) − 1 p[P : N ] − 1p[Q : N ] − 1 p[M : P ] − 1p[M : Q] − 1 [P : N ][Q : N ] Pk(τ ) τ Pk−1(τ ) − 1 Pk(τ ) A NOTE ON IRREDUCIBLE QUADRILATERALS OF II1 FACTORS 3 The flow of the article revolves around the results mentioned above, more or less in the same order. 2. Interior and Exterior angles between intermediate subfactors In this section, we first recall the notions of interior and exterior angles between intermediate subfactors of a given subfactor as introducted by Bakshi et al. in [1] and some useful formulae related to them. This will be followed by some further generalities and explicit calculations related to these angles. In this article, we will be dealing only with subfactors and quadruples of type II1 with finite Jones' index. Given any such quadruple consider the basic constructions N ⊂ M ⊂ M1, P ⊂ M ⊂ P1 and Q ⊂ M ⊂ Q1. As is standard, we denote by e1 the Jones projection eM N . It is easily seen that, as II1-factors acting on L2(M ), both P1 and Q1 are contained in M1. In particular, if eP : L2(M ) → L2(P ) denotes the orthogonal projection, then eP ∈ M1. Likewise, eQ ∈ M1. Thus, we naturally obtain a dual quadruple Q ⊂ M ∪ ∪ N ⊂ P, P1 ⊂ M1 ∪ ∪ M ⊂ Q1. Definition 2.1. [1] Let P and Q be two intermediate subfactors of a subfactor N ⊂ M . Then, the interior angle αN M (P, Q) between P and Q is given by where vP := eP −e1 keP −e1k2 between P and Q is given by βN , hx, yi2 := tr(y∗x) and kxk2 := (tr(x∗x))1/2. And, the exterior angle αN M (P, Q) = cos−1 hvP , vQi2, M (P, Q) = αM M1 (P1, Q1). We will avoid being pedantic and often drop the superscript N and the subscript M when the subfactor N ⊂ M is clear from the context. The following useful relationship between α(P, Q) and β(P, Q) was mentioned in [1], following Definition 3.6, without any proof. For the sake of completeness, we include a proof (though, only for the extremal case, which will be enough for our requirements). Lemma 2.2. For an extremal subfactor N ⊂ M with intermediate subfactors P and Q, we have M (P, Q) = βM αN M1 (P1, Q1). Proof. We have a tower N ⊂ P ⊂ M ⊂ P1 ⊂ M1 ⊂ P2 ⊂ M2 where Pi ⊂ Mi ⊂ Pi+1 = hMi, ePii is a basic construction with Jones projection ePi = e0,i+1 : L2(Mi) → L2(Pi), and P0 := P - see [2, § 3]. Likewise, we have another tower N ⊂ Q ⊂ M ⊂ Q1 = hM, eQi ⊂ M1 ⊂ Q2 = hM1, eQ1i ⊂ M2. From [2, Lemma 4.2], we have eP2 = PSfrag replacements eP + + eP We have a similar figure for eQ2 with respect to eQ. From this pictorial description, it is readily seen through pictures that . tr(eP eQ) = tr(eP2 eQ2 ), tr(eP ) = tr(eP2 ) and tr(eQ) = tr(eQ2 ). 4 K C BAKSHI AND V P GUPTA From Equation (2.1), we have cos(cid:0)αN tr(eP2 eQ2 )−τ √τP2−τ√τQ2−τ . Finally, employing the above equalities obtained through pictures, we obtain M (P, Q)(cid:1) = tr(eP eQ)−τ √τP −τ√τQ−τ and, similarly, cos(cid:0)αM1 M2 (P2, Q2)(cid:1) = as was desired. PSfrag replacements cos(cid:0)βM M1 (P1, Q1)(cid:1) = cos(cid:0)αM1 M2 (P2, Q2)(cid:1) = cos(cid:0)αN M (P, Q)(cid:1), Q1 M1 (cid:3) P P1 α N − M β Q Figure 1. Intuitive picture of a quadrilateral along with its dual 2.1. Some useful formulae related to interior and exterior angles. We first list some plausible facts from [1] that make computations of α(P, Q) and β(P, Q) more amenable. Theorem 2.3. [1] For a quadruple (N, P, Q, M ), let τP = tr(eP ) and τQ = tr(eQ). Then, the interior angle α(P, Q) satisfies (2.1) cos α(P, Q) = which, then, yields that (2.2) cos(cid:0)α(P, Q)(cid:1) = Pi,j trM(cid:0)EM for any two Pimsner-Popa bases {λi} and {µj} of P/N and Q/N , respectively. And, if the quadru- ple is extremal, i.e., N ⊂ M is extremal, then the exterior angle β(P, Q) satisfies (2.3) cos β(P, Q) = . tr(eP eQ) − τ √τP − τ√τQ − τ , N (λ∗i µj)µ∗j λi(cid:1) − 1 p[P : N ] − 1p[Q : N ] − 1 tr(eP eQ) − τP τQ PqτQ − τ 2 pτP − τ 2 Q The following useful expression for tr(eP eQ) is quite evident from Equation (2.1) and Equa- tion (2.2); the details can be readily extracted from the proof of [1, Proposition 2.14]. Lemma 2.4. Let N ⊂ M be a subfactor and P and Q be two intermediate subfactors. Then, trM1 (eP eQ) = τXi,j kEN (λ∗i µj)k2 2 for any two Pimsner-Popa bases {λi} and {µj} of P/N and Q/N , respectively, where τ := [M : N ]−1. Recall that two subfactors N ⊂ M and N ⊂ M are said to be isomorphic (denoted as (N ⊂ M ) ∼= (N ⊂ M)) if there exists a ∗-isomorphism ϕ from M onto M such that ϕ(N ) = N . Likewise, two quadruples (N, P, Q, M ) and (N ,P,Q,M) are said to be isomorphic if there is an isomorphism ϕ between the subfactors N ⊂ M and N ⊂ M such that ϕ(P ) = P and ϕ(P ) = P. Remark 2.5. Since Pimsner-Popa bases are preserved by isomorphisms of subfactors, in view of Theorem 2.3 and Lemma 2.4, we observe that an isomorphism between two quadruples preserves interior and exterior angles. A NOTE ON IRREDUCIBLE QUADRILATERALS OF II1 FACTORS 5 Recall that a quadruple (N, P, Q, M ) is said to be a commuting square if eP eQ = e1 = eQeP . It is said to be a cocommuting square if the dual quadruple (M, P1, Q1, M ) is a commuting square. It is said to be non-degenerate (resp., irreducible) if spanP Q = M (resp., N′∩ M = C). Further, it is said to be a parallelogram if τP τQ = τ or, equivalently, if [M : P ] = [Q : N ] or [M : Q] = [P : N ]. And, a quadruple (N, P, Q, M ) is said to be a quadrilateral if P ∨ Q = M and P ∧ Q = N. Remark 2.6. Commuting and cocommuting conditions have very natural interpretations in terms of above angles, viz., a quadruple (N, P, Q, M ) is a commuting (resp., co-commuting) square if and only if α(P, Q) (resp., β(P, Q)) equals π/2 - see [1, §2]. 2.2. Computation of angles for quadruples of crossed product algebras. Proposition 2.7. Let G be a finite group acting outerly on a II1-factor S. Let H, K and L be subgroups of G such that H ⊆ K ∩ L and K and L are non-trivial. Consider the quadruple (N = S ⋊ H, P = S ⋊ K, Q = S ⋊ L, M = S ⋊ G). Then, tr(eP eQ) = K ∩ L (2.4) , G (2.5) and (2.6) cos(cid:16)α(P, Q)(cid:17) = K ∩ L − H pK − 1pL − 1 G KL − 1 cos(cid:16)β(P, Q)(cid:17) = p[G : K] − 1p[G : L] − 1 . In particular, as is well known, (N, P, Q, M ) is a commuting (resp., cocommuting) square if and only if K ∩ L = H (resp., G = KL). Proof. Note that if α : G → Aut(S) denotes the action of G on S, then there is a unitary representation G ∋ t 7→ ut ∈ B(L2(S)), such that ut(xΩ) = αt(x)Ω for all x ∈ S - see [7, § A.4]. Fix left coset representatives {ki : 1 ≤ i ≤ [K : H]} and {lj : 1 ≤ j ≤ [L : H]} of H in K and L, respectively. Since EN (Pg xgug) = Ph xhuh, it follows that {ki : 1 ≤ i ≤ [K : H]} and {lj : 1 ≤ j ≤ [L : H]} are (right) orthonormal bases for P/N and Q/N , respectively. So, by Lemma 2.4, we obtain tr(eP eQ) = [G : H]−1Xi,j ki lj uki(cid:17) uki(cid:17) (cid:16)since EN(cid:0)Xg∈G ulj u−1 lj xgug(cid:1) = Xh∈H xhuh(cid:17) trM(cid:16)EN(cid:0)u−1 ulj(cid:1)u−1 trM(cid:16)u−1 = [G : H]−1 X{i,j:k−1 = [G : H]−1{(i, j) : lj ∈ kiH} = [G : H]−1 {(i, j) : kiH ∩ ljH 6= ∅}; lj∈H} ki i and note that the map {(i, j) : kiH ∩ ljH 6= ∅} ∋ (i, j) 7→ kiH = ljH ∈ (K ∩ L)/H is a natural bijection; so that, tr(eP eQ) = K∩L G obtain . Then, from Equation (2.1), we immediately and, from Equation (2.3), through an elementary simplification, we deduce that cos(cid:16)α(P, Q)(cid:17) = K ∩ L − H pK − 1pL − 1 G KL − 1 cos(cid:16)β(P, Q)(cid:17) = p[G : K] − 1p[G : L] − 1 . 6 K C BAKSHI AND V P GUPTA The commuting and cocommuting conditions follow from Remark 2.6. (cid:3) Corollary 2.8. Let K, L, G and S be as in Proposition 2.7. Consider the quadruple (N = SG, P = SK, Q = SL, M = S). Then, cos(α(P, Q)) = and cos(β(P, Q)) = G KL − 1 p[G : K] − 1p[G : L] − 1 pK − 1pL − 1 K ∩ L − 1 . In particular, (N, P, Q, M ) is a commuting square if and only if KL = G. And, it is a cocommuting square if and only if K ∩ L = {e}. Proof. Since SG ⊂ S is extremal, by Lemma 2.2, we have αN M (P, Q) = βM M1 (P1, Q1). Outhere, we have M = S, P1 = S ⋊ K, Q1 = S ⋊ L and M1 = S ⋊ G; so, by Proposition 2.7 (taking H to be the trivial subgroup), we obtain cos(cid:0)αN M (SK, SL)(cid:1) = cos(cid:0)βM M1 (S ⋊ K, S ⋊ L)(cid:1) = On the other hand, by definition, we have βN we obtain M (P, Q) = αM G KL − 1 . p[G : K] − 1p[G : L] − 1 M1 (P1, Q1). Hence, by Proposition 2.7, cos(β(P, Q)) = K ∩ L − 1 pK − 1pL − 1 . (cid:3) It was shown in [1, § 5] that the notion of Sano-Watatani's set of angles does not agree with the notion of interior angle. Using Corollary 2.8, we add to that list and show that the Sano-Watatani's set of angles and the interior angle may not be equal even if the former is a singleton. Example 2.9. Consider the quadruple (N = RG, P = RH , Q = RK, M = R) with the assumption that H ∩ K = {e}, H\G/H = 2, H and K are both non-trivial subgroups. Then, the Sano- Watatani's set of angles AngM (P, Q) is a singleton and {α(P, Q)} 6= AngM (P, Q). Proof. From [13, Lemma 5.3 and Proposition 5.2], we have AngM (P, Q) is a singleton, namely, AngM (P, Q) =(cos−1(cid:18) G − HK K(G − H)(cid:19)1/2) . And, by Corollary 2.8, we have cos(cid:0)α(P, Q)(cid:1) = where the second equality follows because H ∩ K = {e} gives HK = HK. Thus, {α(P, Q)} = AngM (P, Q) if and only if p[G : H] − 1p[G : K] − 1 pG/H − 1pGK − 1 G HK − 1 = G HK − 1 , (2.7) Note that RHS =(cid:16) G−HK G HK − 1 pG/H − 1pGK − 1 . = K(G − H)(cid:19)1/2 (cid:18) G − HK (G−H) (cid:17)1/2 (G−K)HK(cid:17)1/2 ·(cid:16) G−HK = G − HK 1 K (G − K)HK , . Hence (2.7) is true if and only if A NOTE ON IRREDUCIBLE QUADRILATERALS OF II1 FACTORS 7 which is then true if and only if H = 1, which is not true since H is not the trivial subgroup. (cid:3) We conclude this subsection by deducing the following well known fact. Example 2.10. Let N ⊂ M be a subfactor and G be a finite group acting outerly (through α) on M . Then, (N, N ⋊ G, M, M ⋊ G) is a commuting square. Proof. Note that EM ⋊G N N (ae). Indeed, for any b ∈ N , we have (cid:0)Pg agug(cid:1) = EM tr(cid:16)(cid:0)X agug(cid:1)b(cid:17) = tr(cid:16)X agαg(b)ug(cid:17) = tr(aeb), and, on the other hand, tr(cid:0)EM Let {λi} be a (right) basis for M/N . Then, from Equation (2.1), we obtain N (aeb)(cid:1) = tr(aeb). N (ae)b(cid:1) = tr(cid:0)EM cos(cid:16)α(N ⋊ G, M )(cid:17) = Pi,g tr(cid:0)EN (λ∗i ug)u∗gλi(cid:1) − 1 p[M : N ] − 1pG − 1 tr(cid:0)Pi EN (λ∗i )λi(cid:1) − 1 p[M : N ] − 1pG − 1 because Pi λiEN (λ∗i ) = 1. Thus, α = π/2, and hence, (N, N ⋊ G, M, M ⋊ G) is a commuting square. = 0, = (cid:3) 3. Weyl group, Quadrilaterals and regularity In this section we focus on the analysis of irreducible subfactors and quadrilaterals from the perspectives of Weyl group and interior and exterior angles between intermediate subfactors. First, we make some useful observations related to Pimsner-Popa bases and regularity. instance, the proof of [1, Proposition 2.14]. This proves necessity. Proposition 3.1. Let P be an intermediate II1-factor of a subfactor N ⊂ M . Let eP denote the canonical Jones projection for the basic construction P ⊂ M ⊂ P1 and {λi} be a finite set in P . Then, {λi} is a Pimsner-Popa basis for P/N if and only if Pi λie1λ∗i = eP Proof. If {λi} is a Pimsner-Popa basis for P/N , then we know that Pi λie1λ∗i = eP - see , for To prove sufficiency, consider the basic construction N ⊂ P ⊂ N1 with Jones projection eP N . Recall, from [2], that this tower is isomorphic to the tower N eP ⊂ eP M eP = P eP ⊂ eP M1eP via a map φ : N1 → eP M1eP satisfying φ(x) = xeP for all x ∈ P . The Jones projection for the second tower is given by eP e1 = e1. Note that Pi λieP e1eP λ∗i =Pi λie1λ∗i = eP . Thus, we obtain This implies that P λieP Proposition 3.2. Let N ⊂ M be an irreducible subfactor and P := (cid:8)NM (N )(cid:9)′′. If {ug : g = [ug] ∈ G} denotes a set of coset representatives of G in NM (N ), then {ug : g ∈ G} ⊂ NP (N ) and it forms a two sided orthonormal basis for P/N. Proof. Since N ⊂ M is irreducible, it follows that P is a II1-factor. By definition, we have NP (N ) ⊂ NM (N ). On the other hand, NM (N ) ⊂ P . So, if u ∈ NM (N ), then u ∈ NP (N ). Hence, NP (N ) = NM (N ). Therefore, we conclude that N ⊂ P is a regular subfactor and also that the Weyl group of N ⊂ P is the same as that of N ⊂ M . Then, since N ⊂ P is regular and irreducible, we conclude, from [4, Lemma 3.1], that {ug : g ∈ G} forms a two-sided orthonormal basis for P/N . φ(cid:0)Xi N λ∗i = 1 and, hence, {λi} is a Pimsner-Popa basis for P/N . (λieP )e1(eP λ∗i ) = eP = φ(1). N λ∗i(cid:1) =Xi λieP (cid:3) (cid:3) Above two propositions yield the following improvement of [4, Lemma 3.1]: Theorem 3.3. Let N ⊂ M be an irreducible subfactor and {ug : g ∈ G} be a set of coset representatives of G in NM (N ). Then, the following are equivalent: (1) [M : N ] = G. (2) {ug : g ∈ G} is a two sided orthonormal basis for M/N . 8 K C BAKSHI AND V P GUPTA (3) N ⊂ M is regular. Proof. (1) ⇔ (2) : Let p := Pg uge1u∗g. Then, tr(p) = Pg∈G tr(uge1u∗g) = G[M : N ]−1. Thus, {ug : g ∈ G} is an orthonormal basis for M/N if and only if [M : N ] = G. (2) ⇔ (3) : This equivalence follows immediately from Proposition 3.1 and Proposition 3.2. (cid:3) Next, we move to Weyl groups. Let N ⊂ M be a subfactor and let U(N ) (resp., U(M )) denote the group of unitaries of N (resp., M ) and NM (N ) := {u ∈ U(M ) : uN u∗ = N} denote the group of unitary normalizers of N in M . Clearly, U(N ) is a normal subgroup of NM (N ). For a finite index subfactor N ⊂ M , one associates the so-called Weyl group, which we shall denote by G, defined as the quotient group NM (N )/U(N ) (see, for example [4, 6, 8, 3, 11]). Example 3.4. Let G be a finite group acting outerly on a II1-factor N and H be a normal subgroup of G. Then, the Weyl group of the subfactor N ⋊ H ⊂ N ⋊ G is isomorphic to the quotient group G/H. Proof. Fix a set of coset representatives {gi : 1 ≤ i ≤ n = [G : H]} of H in G. Then, {ugi : 1 ≤ i ≤ n} forms a two sided orthonormal basis for (N ⋊ G)/N (where ug's are as in Proposition 2.7). Clearly, the map G/H ∋ giH 7−→ [ugi] ∈ G is a bijection. Then, note that ϕ ϕ(giH)ϕ(gjH) = [ugi ][ugj ] = [ugi ugj ] = [ugigj ]. On the other hand, if gigjH = gkH, then gigk = gkh for some h ∈ H, which implies that ugigj = ugk uh, i.e., [ugigj ] = [uk] in G. Thus, ϕ(giH · gjH) = ϕ(gkH) = [ugk ] = [ugigj ] = ϕ(giH)ϕ(gjH) for all 1 ≤ i, j ≤ n. Hence, K/H ∼= G. (cid:3) Analogous to the well known Goldman's Theorem for a subfactor with index 2, it is known that an irreducible regular subfactor N ⊂ M can be realized as the group subfactor N ⊂ N ⋊ G, where G is the Weyl group of N ⊂ M which acts outerly on N - see [4, 11, 8] and the references therein. As a consequence, we deduce the following version of Goldman's type Theorem for irreducible quadrilaterals. Theorem 3.5. Let (N, P, Q, M ) be an irreducible quadrilateral such that N ⊂ P and N ⊂ Q are both regular. Then, G acts outerly on N and (N, P, Q, M ) = (N, N ⋊ H, N ⋊ K, N ⋊ G), where H, K and G are the Weyl groups of N ⊂ P , N ⊂ Q and N ⊂ M , respectively. Proof. First, note that N ⊂ M is regular because M = P ∨ Q = NP (N )′′ ∨ NQ(N )′′ ⊆ {NP (N ) ∪ NQ(N )}′′ ⊆ NM (N )′′ ⊆ M. Then, since N ⊂ M is regular and irreducible, Hong [4] had shown that if N−1 ⊂ N ⊂ M is an instance of downward basic construction with Jones projection e−1, then there is a representation G ∋ g 7→ vg ∈ U(N′ −1 ∩ M ) such that vg ∈ NM (N ) for all g ∈ G, M = {N, vg : g ∈ G}′′ and G ∈ g 7→ Advg ∈ Aut(N ) is an outer action of G on N , i.e., (N ⊂ M ) = (N ⊂ N ⋊ G) - see [4, Lemma 3.3 and Theorem 3.1]. Also, for each g ∈ G, the coset vg U(N ) = g in G. So, by Galois correspondence, P = N ⋊ H′ and Q = N ⋊ K′ for unique subgroups H′ and K′ of G. We assert that H = H′ and K = K′. We have P = N ⋊ H′ = {N, vh′ : h′ ∈ H′}′′. So, for each h′ ∈ H′, vh′ ∈ P ; thus, {vh′ : h′ ∈ H′} ⊂ NP (N ). Also, h′ = [vh′] ∈ NP (N )/U(N ) = H, so that H′ ⊂ H. As seen in Example 3.4, H′ ∼= H; so, we must have H′ = H and hence H′ = H. Likewise, we obtain K = K′. Hence, (N, P, Q, M ) = (N, N ⋊ H, N ⋊ K, N ⋊ G). (cid:3) Corollary 3.6. Let (N, P, Q, M ) be an irreducible quadrilateral such that N ⊂ P and N ⊂ Q are both regular. Then, the Weyl groups of N ⊂ P and N ⊂ Q together generate the Weyl group of N ⊂ M . A NOTE ON IRREDUCIBLE QUADRILATERALS OF II1 FACTORS 9 Proof. Let G′ be the subgroup of G generated by H and K, then N ⋊ G′ ⊆ N ⋊ G. Also, since M = P ∨ Q, we have N ⋊ G = (N ⋊ H) ∨ (N ⋊ K) ⊆ N ⋊ G′ ⊆ N ⋊ G. Hence, by Galois correspondence again, we must have G = G′, i.e., G is generated by its subgroups H and K. (cid:3) We have the following partial converse of Corollary 3.6. Proposition 3.7. Let (N, P, Q, M ) be an irreducible quadruple such that N ⊂ P and N ⊂ Q are both regular. If N ⊂ M is regular and the Weyl groups of N ⊂ P and N ⊂ Q together generate the Weyl group of N ⊂ M , then M = P ∨ Q. Proof. Fix any set of coset representatives {ug : g ∈ G} of G in NM (N ). Since N ⊂ M is regular, {ug : g ∈ G} forms a two sided orthonormal basis for M/N , by Theorem 3.3. Note that each g in G is a word in H ∪ K and for any pair g, g′ ∈ G, we have [ugg′ ] = gg′ = [ug][ug′ ] = [ugug′], so that ugg′ = vugu′g for some v ∈ U(N ). Thus, M =Pg N ug ⊆ (Ph N uh) ∨ (Pk uk) = P ∨ Q. Following corollary first appeared implicitly in the proof of [13, Theorem 6.2]. We include it (cid:3) here, as an application of Theorem 3.5 and Corollary 3.6. Corollary 3.8. Let (N, P, Q, M ) be an irreducible quadrilateral with [P : N ] = 2 = [Q : N ]. Then, [M : N ] is an even integer and the Weyl group of N ⊂ M is isomorphic to the Dihedral group of order 2n, where n = [M : P ] = [M : Q]. Proof. By Goldman's Theorem, we know that (N ⊂ P ) ∼= (N ⊂ N ⋊σ Z2) and (N ⊂ Q) ∼= (N ⊂ N ⋊τ Z2) for some outer actions σ and τ of Z2 on N and hence both N ⊂ P and N ⊂ Q are regular. Then, by Theorem 3.3, we obtain H = [P : N ] = 2 and K = [Q : N ] = 2, where H and K are as in Theorem 3.5. So, H and K are both cyclic of order 2. By Theorem 3.5, N ⊂ M is regular and hence [M : N ] = G by Theorem 3.3. Also, by Theorem 3.5, G is a finite group generated by H and K. Thus, G is generated by two elements which are both of order 2. Hence, by [15, Theorem 6.8], G is isomorphic to the Dihedral group of order 2n. (cid:3) We conclude this section with the demonstration of a direct relationship between angles and Weyl groups of interemdiate subfactors of an irreducible quadruple. As above, for a quadruple (N, P, Q, M ), we denote by G, H and K the Weyl groups of N ⊂ M, N ⊂ P and N ⊂ Q, respectively. First, we deduce the relationship for an irreducible quadrilateral. Theorem 3.9. Let (N, P, Q, M ) be an irreducible quadrilateral such that N ⊆ P and N ⊆ Q are both regular. Then, α(P, Q) = π/2, i.e., (N, P, Q, M ) is a commuting square, and In particular, (N, P, Q, M ) is a commuting square if and only if G = HK. cos β(P, Q) = G HK − 1 p[G : H] − 1p[G : K] − 1 . . Proof. It follows from Theorem 3.5 and Proposition 2.7. (cid:3) More generally, we have the following relationship. Theorem 3.10. Let (N, P, Q, M ) be an irreducible quadruple such that N ⊆ P and N ⊆ Q are both regular. Then, and cos β(P, Q) = cos α(P, Q) = H ∩ K − 1 pH − 1pK − 1 [M:N ] HK − 1 p[M : P ] − 1p[M : Q] − 1 10 K C BAKSHI AND V P GUPTA In particular, (N, P, Q, M ) is a commuting square, if and only if, H ∩ K is trivial, if and only, if P ∩ Q =. And, (N, P, Q, M ) is a cocommuting square if and only if G = HK = [M : N ]. Proof. Since N ⊂ M is irreducible, P ∨ Q is a II1-factor. Consider the irreducible quadruple (N, P, Q, P ∨ Q). Then, by Theorem 3.5(2), (N, P, Q, P ∨ Q) = (N, N ⋊ H, N ⋊ K, N ⋊ G′) where G′ is the Weyl group of N ⊂ P ∨ Q. Hence, by Proposition 2.7, we obtain cos(cid:0)αN P∨Q(P, Q)(cid:1) = H ∩ K − 1 pH − 1pK − 1 . And, it is known that αN P∨Q(P, Q) = αN M (P, Q) - see [1, Proposition 2.16]. On the other hand, being irreducible, N ⊂ M is extremal. So, by Equation (2.3), the exterior angle between P and Q is given by cos β(P, Q) = Q = [M : N ]−1H ∩ K − [M : P ]−1[M : Q]−1 tr(eP eQ) − τP τQ pqτQ − τ 2 qτP − τ 2 p[M : P ]−1 − [M : P ]−2p[M : Q]−1 − [M : Q]−2 = H ∩ K[P : N ]−1[M : Q] − 1 p[M : P ] − 1p[M : Q] − 1 p[M : P ] − 1p[M : Q] − 1 HK − 1 [M:N ] = , 2 = [M : N ]−1H ∩ K where we have used the equalities tr(eP eQ) = τPh∈H,k∈K kEN (u∗huk)k2 by Lemma 2.4 and the well known formula HK = HKH ∩ K−1. Hence, by [1, Proposition 2.7], (N, P, Q, M ) is a cocommuting square if and only if β(P, Q) = π/2, i.e., if and only if HK = [M : N ]. Note that G ≤ [M : N ] (see [4, 11]) and HK ⊆ G. So, HK = [M : N ] if and only if G = HK = [M : N ]. Remark 3.11. Sano and Watatani ([13, Theorem 6.1]) had proved that an irreducible quadrilateral (N, P, Q, M ) with [P : N ] = 2 = [Q : N ] is always a commuting square. Thus, Theorem 3.9 can also be thought of as a generalization of that result. (cid:3) 4. Possible values of interior and exterior angles It is a very natural curiousity to know the possible values of interior and exterior angles between intermediate subfactor. As a first attempt in this direction, we make some calculations in the irreducible set up. Prior to that we recall two auxiliary positive operators associated to a quadruple (from [1, 13]) whose norms are equal, and show that this common entity has a direct relationship with the possible values of interior and exterior angles. 4.1. Two auxiliary operators associated to a qudruple. Consider a quadruple (N, P, Q, M ). Let {λi : i ∈ I} and {µj : j ∈ J} be (right) Pimsner-Popa bases for P/N and Q/N , respectively. Consider two positive operators p(P, Q) and p(Q, P ) given by p(P, Q) =Xi,j λiµje1µ∗j λ∗i and p(Q, P ) =Xi,j µjλie1λ∗i µ∗j . Remark 4.1. By [1, Lemma 2.18], p(P, Q) and p(Q, P ) are both independent of choices of bases. And, by [1, Proposition 2.22], Jp(P, Q)J = p(Q, P ), where J is the usual modular conjugation operator on L2(M ); so that, kp(P, Q)k = kp(Q, P )k. Notation 4.2. For a quadruple (N, P, Q, M ), let r := [P :N ] kp(Q, P )k. [M:P ] and λ := kp(P, Q)k = [M:Q] = [Q:N ] A NOTE ON IRREDUCIBLE QUADRILATERALS OF II1 FACTORS 11 Recall that for a self adjoint element x in a von Neumann algebra M, its support is given by s(x) := inf{p ∈ P(M) : px = x = xp}. We will need the following useful lemma which follows from [1, Proposition 2.25 & Lemma 3.2]. Lemma 4.3. [1] If (N, P, Q, M ) is a quadruple such that N ⊂ M is irreducible, then λ = [Q : N ] tr(p(P, Q)eQ) and s(p(P, Q)) = p(P, Q)/λ. In particular, tr(s(p(P, Q))) = r/λ and p(P, Q) is a projection if and only if λ = 1. It turns out that s(p(P, Q)) is a minimal projection in P ′ ∩ Q1 which is central as well. λ Proposition 4.4. Let (N, P, Q, M ) be quadruple such that N ⊂ M irreducible. Then, p(P,Q) (resp., p(Q,P ) ) is a minimal projection in P ′ ∩ Q1 (resp., Q′ ∩ P1) which is also central. P ′ (eQ) ∈ P ′ ∩ Q1. We first show that 1 Proof. By Lemma 4.3, 1 λ p(P, Q) is a projection. Further, by [1, Proposition 2.25], we have p(P, Q) = [P : N ]EN ′ λ p(P, Q) is minimal in P ′ ∩ Q1. Consider any projection q ∈ P ′ ∩ Q1 satisfying 0 ≤ q ≤ 1 λ p(P, Q). We also have qeQ = [M : Q]EQ1 M (qeQ)eQ (by the Pushdown Lemma [10, Lemma 1.2]). Clearly, EM (qeQ) ∈ N′ ∩ M . Thus, irreduciblility of N ⊆ M implies that qeQ = teQ for the scalar t = [M : Q]EQ1 λ p(P, Q). Then, q = q M (qeQ). Therefore, λ q = = = = = p(P, Q) q λ q λ [P : N ] λ [P : N ] λ t λ p(P, Q). [P : N ]EN ′ P ′ (eQ) EN ′ P ′ (qeQ) tEN ′ P ′ (eQ) Since q and 1 q was arbitrary, this proves the minimality of p(P,Q) λ p(P, Q) are projections we conclude that t2 = t. Therefore, q = 0 or p(P, Q). Since . We now prove that 1 λ p(P, Q) is a central projection in P ′ ∩ Q1. For this, we first show that eQ is a minimal central projection in N′ ∩ Q1. Let u be an arbitrary unitary in N′ ∩ Q1. Then, by the Pushdown Lemma again, we have ueQ = [M : Q]EM (ueQ)eQ. But clearly EM (ueQ) ∈ N′∩M = C. Thus, ueQu∗ = teQ for some scalar t. Since teQ is a non-zero projection, we must have t = 1; so that, ueQ = eQu for all u ∈ U(N′ ∩ Q1), thereby implying that eQ is central in N′ ∩ Q1. This shows that λ vp(P, Q)v∗ = [P : N ]EN ′ P ′ (veQv∗) = [P : N ]EN ′ P ′ (eQ) = p(P, Q) for all v ∈ U(P ′ ∩ Q1) and we are done. sition 2.22]. Assertion about p(Q, P ) then follows from the fact that p(Q, P ) = Jp(P, Q)J - see [1, Propo- (cid:3) As asserted above, we now present the direct relationship that exists between the values of the interior and exterior angles and the common norm of the above two auxiliary operators. Proposition 4.5. Let (N, P, Q, M ) be a finite index quadruple of II1-factors with N ⊂ M irre- ducible. Then, (4.1) and (4.2) cos(α(P, Q)) = cos(β(P, Q)) = (λ − 1) p[P : N ] − 1p[Q : N ] − 1 p[P : N ] − rp[Q : N ] − r (λ − r) . 12 K C BAKSHI AND V P GUPTA Proof. From Equation (2.1), we have cos(α(P, Q)) = tr(eP eQ) − τ ptr(eP ) − τptr(eQ) − τ = [M : N ]tr(eP eQ) − 1 p[P : N ] − 1p[Q : N ] − 1 . It can be shown that p(P, Q) = [Q : N ]EM1 (eP ) - see the proof of [1, Proposition 2.25]; hence, Q1 p(P, Q)eQ = [Q : N ]EM1 (eP eQ). Thus, tr(p(P, Q)eQ) = [Q : N ]tr(eP eQ). This, along with the Q1 fact that p(P, Q)eQ = λeQ (because N ⊂ M is irreducible - see the proof of [1, Lemma 3.2]), yields tr(eP eQ) = λtr(eQ) [Q:N ] = λτ, that is [M : N ]tr(eP eQ) = λ. Thus, we obtain On the other hand, being irreducible, N ⊂ M is extremal. So, from Equation (2.3), we have cos(α(P, Q)) = cos(β(P, Q)) = = = λ − 1 . Q Q τ λ − τP τQ p[P : N ] − 1p[Q : N ] − 1 tr(eP eQ) − τP τQ PqτQ − τ 2 pτP − τ 2 PqτQ − τ 2 pτP − τ 2 λ − r [M:P ]q[Q : N ] − [Q:N ] q[P : N ] − [P :N ] [M : Q](cid:19) − [P : N ][Q : N ] [M : P ] [Q : N ] + [M:Q] . Then, note that (cid:18)[P : N ] − [P : N ] [M : P ](cid:19)(cid:18)[Q : N ] − [P : N ][Q : N ] = [P : N ][Q : N ] − = [P : N ][Q : N ] − r[Q : N ] − r[P : N ] + r2 = ([Q : N ] − r)([P : N ] − r), [M : Q] [P : N ][Q : N ] [M : P ][M : Q] and we are done. Proposition 4.6. Let (N, P, Q, M ) be a commuting square with N ⊂ M irreducible. Then, (cid:3) And, if (N, P, Q, M ) is a a cocommuting square with N ⊂ M irreducible, then cos(β(P, Q)) = cos(α(P, Q)) = r−1 − 1 r − 1 p[M : P ] − 1p[M : Q] − 1 p[P : N ] − 1p[Q : N ] − 1 . . Proof. The formula for β(P, Q) is easy and is left to the reader. Cocommuting square implies β(P, Q) = π/2. Thus, tr(eP eQ) = τP τQ. Now simply use τP τQ τ = r and the formula follows from the definition of α(P, Q). (cid:3) 4.2. Values of angles in the irreducible setup. In order to determine the values of interior and exterior angles between intermeidate subfactors, as is evident from Proposition 4.5, it becomes im- portant to know the possible values of r and λ. Recall, from [12], Popa's set of relative-dimensions of projections in M relative to N given by Λ(M, N ) = {α ∈ R : ∃ a projection f0 ∈ M such that EN (f0) = α1N}. Lemma 4.7. For an irreducible quadruple (N, P, Q, M ), tr(cid:0)s(cid:0)p(P, Q)(cid:1)(cid:1) = r (1 − r λ ) ∈ Λ(M1, M ). λ ∈ Λ(M1, M ). Also, A NOTE ON IRREDUCIBLE QUADRILATERALS OF II1 FACTORS 13 = [M : Q] operator 1 [P : N ] [M : Q] = r. Thus, EM1 M (eQ)λ∗i = P λiλ∗i Proof. Let {λi} be a basis for P/N . Then, p(P, Q) :=P λieQλi∗ ∈ M1 and clearly EM1 P λiEM1 For a commuting square (N, P, Q, M ) with N ⊂ M irreducible, it is known that λ = 1 - see [1, λ p(P, Q) = s(p(P, Q)) is a projection. This completes the proof. Proposition 2.20]. Thus, we deduce the following: Corollary 4.8. If (N, P, Q, M ) is a commuting square with N ⊂ M irreducible, then r ∈ Λ(M1, M ). And, if (N, P, Q, M ) is a parallelogram, then λ−1 ∈ Λ(M1, M ). M (cid:0)p(P, Q)(cid:1) = λ . And, by Lemma 4.3, the λ p(P, Q)) = r M ( 1 (cid:3) 1 Consider the polynomials Pn(x) for n ≥ 0 (introduced by Jones in [5] and) defined recursively by P0 = 1, P1 = 1, Pn+1(x) = Pn(x) − xPn−1(x), n > 0. Thus, P2(x) = 1 − x, P3(x) = 1 − 2x and so on. From [5], we know that Pk(cid:0) 4cos2π/(n+2)(cid:1) = 0. Furthermore, Pk(ǫ) > 0 for all ǫ ≤ 1/4 and k ≥ 0. Also, by definition, we have τ Pn−1(τ ) 1 − Pn+1(τ ) theorem: Theorem 4.9. [12] Let N ⊂ M be a subfactor of finite index. 4cos2π/(n+2)(cid:1) > 0 for 0 ≤ k ≤ n − 1, and Pn(cid:0) for all n ≥ 1, where, as is standard, τ := [M : N ]−1. While trying to determine the possible entries of the set Λ(M, N ), Popa [12] proved the following Pn(τ ) = Pn(τ ) 1 (1) If [M : N ] = 4cos2(cid:0) π n+2(cid:1) for some n ≥ 1, then Pk(τ ) : 0 ≤ k ≤ n(cid:27). (2) If [M : N ] ≥ 4 and t ≤ 1/2 is so that t(1 − t) = τ , then Λ(M, N ) =(cid:8)0(cid:9) ∪(cid:26) τ Pk−1(τ ) : 0 ≤ k ≤ n − 1(cid:27) =(cid:26) Pk(τ ) Pk−1(τ ) : k ≥ 0(cid:27). non-trivial quadrilateral (N, P, Q, M ), we always have [M : N ] ≥ 4. Proposition 4.10. Let (N, P, Q, M ) be an irreducible quadruple and let 0 < t ≤ 1/2 be such that t(1 − t) = τ . Then, either r Proof. This follows from Theorem 4.9 and Lemma 4.7. Since a subfactor with index less than 4 does not admit any intermediate subfactor, for any Λ(M, N ) ∩ (0, t) =(cid:26) τ Pk−1(τ ) for some k ≥ 0. λ ≥ t or r λ = τ Pk−1(τ ) Pk(τ ) Pk(τ ) (cid:3) We may thus compute the interior and exterior angles in this specific situation, as follows. Theorem 4.11. Let (N, P, Q, M ) be a quadruple with N ⊂ M irreducible and let 0 < t ≤ 1/2 be such that t(1 − t) = τ . If r/λ ≥ t, then, cos(α(P, Q)) ≤ cos(β(P, Q)) ≤ cos(α(P, Q)) = cos(β(P, Q)) = and And, if r/λ < t, then, and for some k ≥ 0. . 1 t − 1 [P : N ][Q : N ](1 − t) − 1 p[P : N ] − 1p[Q : N ] − 1 p[M : P ] − 1p[M : Q] − 1 Pk−1(τ ) − 1 p[P : N ] − 1p[Q : N ] − 1 p[M : P ] − 1p[M : Q] − 1 [P : N ][Q : N ] Pk(τ ) τ Pk−1(τ ) − 1 Pk(τ ) 14 K C BAKSHI AND V P GUPTA Proof. First, suppose that r/λ ≥ t. Thus, λ ≤ r/t. Observe that r/t = [P : N ][Q : N ](1 − t); so, from Equation (4.1), we obtain the first inequality. Also, λ − r ≤ r(1/t − 1). Thus, from Equation (4.2), we obtain cos(β(P, Q)) ≤ r(1/t − 1) = (1/t − 1) . p[P : N ] − rp[Q : N ] − r p[M : Q] − 1p[M : P ] − 1 Next, suppose that N ⊂ M is irreducible and r/λ < t. By Proposition 4.10, we have λ = Pk−1(τ ) , for some k ≥ 0. Observe that r/τ = [P : N ][Q : N ]. Thus, by Equation (4.1), we obtain Pk(τ ) r τ cos(α(P, Q)) = and, by Equation (4.2), we obtain [P : N ][Q : N ] Pk(τ ) Pk−1(τ ) − 1 p[P : N ] − 1p[Q : N ] − 1 cos(β(P, Q)) = τ Pk−1(τ ) − 1(cid:19) r(cid:18) Pk(τ ) p[P : N ] − rp[Q : N ] − r = Pk(τ ) τ Pk−1(τ ) − 1 p[M : P ] − 1p[M : Q] − 1 . (cid:3) Theorem 4.12. Let (N, P, Q, M ) be an irreducible quadrilateral such that N ⊂ P and N ⊂ Q , where are both regular and suppose [P : N ] = 2. Then, cos(β(P, Q)) = P2(m/2) m = [M : Q] ∈ N and, as usual, δ :=p[M : N ]. Proof. Let H, K and G denote the Weyl groups of N ⊂ P , N ⊂ Q and N ⊂ M , respectively. We have m = [M : Q] = [M:N ] (by Theorem 3.5 and Theorem 3.3) and H = 2. Thus, by Theorem 3.9, we obtain [Q:N ] = G K pP2(δ2/2)P3(m/2) cos(β(P, Q)) = = = m K − 1 G HK − 1 q G H − 1q G 2 − 1 p(δ2/2) − 1√m − 1 pP2(δ2/2)pP3(m/2) P2(m/2) . (cid:3) Corollary 4.13. Let (N, P, Q, M ) be an irreducible quadrilateral such that [P : N ] = 2 = [Q : N ]. Then, α(P, Q) = π/2 and β(P, Q) = cos−1(cid:16) P2(m/2) In particular, β(P, Q) > π/3. P3(m/2)(cid:17), where m = [M : P ] = [M : Q] is an integer. Proof. This follows from Theorem 4.12 and the fact that P2(δ2/2) = P2(m) = P3(m/2), where [M : P ] = [M : Q] = m. (cid:3) Corollary 4.14. Let (N, P, Q, M ) be an irreducible quadrilateral such that N ⊆ P and N ⊆ Q are both regular with [P : N ] = 2 and [Q : N ] = 3. Then, α(P, Q) = π/2 and β(P, Q) > cos−1(cid:18) 1√6(cid:19). Proof. By Theorem 3.9, we have α(P, Q) = π/2 and taking m = [G : K], we obtain cos2(β(P, Q)) = m2/4 − m + 1 3m2/2 − 5m/2 + 1 < 1/6, where the inequality follows from a routine comparison using the fact that m ≥ 2. (cid:3) A NOTE ON IRREDUCIBLE QUADRILATERALS OF II1 FACTORS 15 5. Certain bounds on angles and their implications In this section, we observe that when one leg of an irreducible quadruple is assumed to be regular, it enforces certain bounds on interior and exterior angles, which then imposes some bounds on the index of the other leg. Theorem 5.1. Let (N, P, Q, M ) be an irreducible quadruple such that N ⊂ P is regular. Then, and cos(cid:0)α(P, Q)(cid:1) ≤ s [P : N ] − 1 [Q : N ] − 1! cos(β(P, Q)) ≤ s [P : N ] − r [Q : N ] − r!. Proof. For any u ∈ NM (P ) we have uEQ(u∗) ∈ N′ ∩ M = C. To see this, let n ∈ N be arbitrary. Then, uEQ(u∗).n = uEQ(u∗n) = uEQ(u∗nu.u∗) = n.uEQ(u∗). Let uEQ(u∗) = t ∈ C. If t 6= 0 we get EQ(u∗) = tu∗ so that u ∈ Q, which implies that t = uEQ(u∗) = uu∗ = 1. Thus, uEQ(u∗) ∈ {0, 1} for all u ∈ NM (P ). Using Theorem 3.3, fix an orthonormal basis {ui} ⊂ NP (N ) for P/N . Consider the auxiliary operator p(P, Q) = Pi uieQu∗i . Since N ⊂ M is irreducible, we have p(P, Q)eQ = λeQ (see [1, Lemma 3.2]), where λ = kp(P, Q)k. Also, p(P, Q)eQ = Pi uieQu∗i eQ = Pi uiEQ(u∗i )eQ ∈ CeQ, which yields Pi uiEQ(u∗i ) = λ. In particular, we obtain 0 ≤ λ ≤ NP (N )/U(N ) = [P : N ]. Thus, λ − 1 ≤ [P : N ] − 1 and the lower bound for α follows from Proposition 4.5(1). We also have λ − r ≤ [P : N ] − r. So the lower bound for β follows from Proposition 4.5. (cid:3) Remark 5.2. Note that in above proof, we also observed that kp(P, Q)k is an integer less than or equal to [P : N ]. Above theorem imposes some immediate bounds on [Q : N ] in terms of [P : N ], as follows. Corollary 5.3. Let (N, P, Q, M ) be as in Theorem 5.1. Then, we have the following: (1) If α(P, Q) ≤ π/3, then [Q : N ] ≤ 4[P : N ] − 3. (2) If α(P, Q) ≤ π/4, then [Q : N ] ≤ 2[P : N ] − 1. (3) If α(P, Q) ≤ π/6, then [Q : N ] ≤ 4/3[P : N ] − 1/3. Proof. If α(P, Q) ≤ π/3, then cos α(P, Q) ≥ cos(π/3) = 1/2. So, [Q : N ] ≤ 4[P : N ] − 3. Others follow similarly. [P :N ]−1 [Q:N ]−1 ≥ 1/4 and hence (cid:3) As a consequence, when we intuitively try to visualize an irreducible quadruple (N, P, Q, M ) with [P : N ] = 2 as a 4-sided structure in plane (as in Figure 1), then it seems that the smaller is the interior angle between P and Q the shorter is the length (or index) of N ⊂ Q. This assertion is supported by the following observations: If [Q : N ] > 7/3, then it follows from above corollary that α(P, Q) > π/6. Likewise, If [Q : N ] > 3, then α(P, Q) > π/4. And, if [Q : N ] > 5, then α(P, Q) > π/3. In particular, since P is a minimal subfactor, i.e., N ⊂ P admits no intermediate subfactor, in the last scenario, Q cannot be a minimal subfactor because, by [1], we know that the interior angle between two minimal intermedite subfactors is always less than π/3. Also, observe that if α(P, Q) ≤ π/4, then [Q : N ] ≤ 3 and hence N ⊂ Q must be a Jones' subfactor. The reader can get a better feeling of above assertion by making similar calculations for an irreducible quadruple (N, P, Q, M ) such that P ⊂ N is regular with [P : N ] = n, for arbitrary n. 16 K C BAKSHI AND V P GUPTA References [1] K. C. Bakshi, S. Das, Z. Liu and Y. Ren, Angle between intermediate subfactors and its rigidity, Trans. Amer. Math. Soc., to appear. [2] B. Bhattacharya and Z. Landau, Intermediate standard ivariants and intermediate planar algebras, preprint. [3] M. Choda, A characterization of crossed product of factors by discrete outer automorphism groups, J. Math. Soc. Japan, 31, 1979, 257 - 261. [4] J. Hong, Characterization of crossed product without cohomology, J. Korean Math. Soc., 32 (2), 1995, 183 - 192. [5] V. F. R. Jones, Index for subfactors, Invent. Math., 72 (1), 1983, 1 - 25. [6] V. F. R. Jones, and S. Popa, Some properties of MASA's in factors, Invariant Subspaces and Other Topics, Operator Theory: Advances and Applications, Birkhauser, Basel, vol 6., 1982, 89 - 102. [7] V. F. R. Jones and V. S. Sunder, Introduction to Subfactors, London Mathematical Society Lecture Note Series, Cambridge University Press, vol. 234, 1997. [8] H. Kosaki, Characterization of crossed product (properly infinite case), Pacific J. Math., 137 (1), 1989, 159-167. [9] R. Longo, Conformal subnets and intermediate subfactors, Comm. Math. Phys., 237 (1-2), 2003, 7-30. [10] M. Pimsner and S. Popa, Entropy and index for subfactors, Ann. Sci. Ecole Norm. Sup., Series 4, 19 (1), 1986, 57 - 106. [11] M. Pimsner, S. Popa, Finite dimensional approximation of algberas and obstructions for the index, J. Funct. Anal. 98, 1991, 270 - 291. [12] S. Popa, Relative dimension, towers of projections and commuting squares of subfactors, Pacific J. Math., 137(1), 1989, 181-207. [13] T. Sano and Y. Watatani, Angles between two subfactors, J. Operator Theory, 32 (2), 1994, 209 - 241. [14] C. E. Sutherland, Cohomology and extensions of von Neumann algebras. II, Publ. RIMS, Kyoto Univ., 16, 135 - 174, 1980. [15] M. Suzuki, Group Theory I, Grundlehren der mathematischen Wissenschaften, Springer, 1982. Chennai Mathematical Institute, Chennai, INDIA E-mail address: [email protected],[email protected] School of Physical Sciences, Jawaharlal Nehru University, New Delhi, INDIA E-mail address: [email protected]
1703.07953	1	1703	2017-03-23T07:39:49	Fredholm conditions on non-compact manifolds: theory and examples	[ "math.OA", "math-ph", "math.AP", "math.FA", "math-ph", "math.SP" ]	We give explicit Fredholm conditions for classes of pseudodifferential operators on suitable singular and non-compact spaces. In particular, we include a "user's guide" to Fredholm conditions on particular classes of manifolds including asymptotically hyperbolic manifolds, asymptotically Euclidean (or conic) manifolds, and manifolds with poly-cylindrical ends. The reader interested in applications should be able read right away the results related to those examples, beginning with Section 5. Our general, theoretical results are that an operator adapted to the geometry is Fredholm if, and only if, it is elliptic and all its limit operators, in a sense to be made precise, are invertible. Central to our theoretical results is the concept of a Fredholm groupoid, which is the class of groupoids for which this characterization of the Fredholm condition is valid. We use the notions of exhaustive and strictly spectral families of representations to obtain a general characterization of Fredholm groupoids. In particular, we introduce the class of the so-called groupoids with Exel's property as the groupoids for which the regular representations are exhaustive. We show that the class of "stratified submersion groupoids" has Exel's property, where stratified submersion groupoids are defined by glueing fibered pull-backs of bundles of Lie groups. We prove that a stratified submersion groupoid is Fredholm whenever its isotropy groups are amenable. Many groupoids, and hence many pseudodifferential operators appearing in practice, fit into this framework. This fact is explored to yield Fredholm conditions not only in the above mentioned classes, but also on manifolds that are obtained by desingularization or by blow-up of singular sets.	math.OA	math	FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS: THEORY AND EXAMPLES CATARINA CARVALHO, VICTOR NISTOR, AND YU QIAO Abstract. We give explicit Fredholm conditions for classes of pseudodifferen- tial operators on suitable singular and non-compact spaces. In particular, we include a "user's guide" to Fredholm conditions on particular classes of mani- folds including asymptotically hyperbolic manifolds, asymptotically Euclidean (or conic) manifolds, and manifolds with poly-cylindrical ends. The reader interested in applications should be able read right away the results related to those examples, beginning with Section 5. Our general, theoretical results are that an operator adapted to the geometry is Fredholm if, and only if, it is elliptic and all its limit operators, in a sense to be made precise, are invert- ible. Central to our theoretical results is the concept of a Fredholm groupoid, which is the class of groupoids for which this characterization of the Fred- holm condition is valid. We use the notions of exhaustive and strictly spectral families of representations to obtain a general characterization of Fredholm groupoids. In particular, we introduce the class of the so-called groupoids with Exel's property as the groupoids for which the regular representations are ex- haustive. We show that the class of "stratified submersion groupoids" has Exel's property, where stratified submersion groupoids are defined by glueing fibered pull-backs of bundles of Lie groups. We prove that a stratified submer- sion groupoid is Fredholm whenever its isotropy groups are amenable. Many groupoids, and hence many pseudodifferential operators appearing in practice, fit into this framework. This fact is explored to yield Fredholm conditions not only in the above mentioned classes, but also on manifolds that are obtained by desingularization or by blow-up of singular sets. Contents Introduction 1. 1.1. Background and main result 1.2. Exhaustive families of representations 1.3. Fredholm Lie groupoids 1.4. Examples and applications 1.5. Contents of the paper 2. Groupoids and their C∗-algebras 2.1. Locally compact groupoids 2 2 3 4 4 5 6 6 Date: April 11, 2018. 1991 Mathematics Subject Classification. 58J40 (primary) 58H05, 46L60, 47L80, 47L90. Key words and phrases. pseudodifferential operator, differential operator, Fredholm opera- tor, groupoid, Lie group, Sobolev space, C ∗-algebras, compact operators, non-compact manifold, singular manifold. V.N. has been partially supported by ANR-14-CE25-0012-01. Qiao was partially supported by NSF of China (11301317, 11571211). Carvalho was partially supported by Fundao para a Cincia e Tecnologia UID/MAT/04721/2013 (Portugal). Manuscripts available from http://iecl.univ-lorraine.fr/Victor.Nistor/. 1 2 C. CARVALHO, V. NISTOR, AND Y. QIAO 2.2. Haar systems and C∗-algebras 2.3. Manifolds with corners and Lie groupoids 2.4. Examples of Lie groupoids 3. Exhaustive families of representations and Exel's question 3.1. Exhaustive families of representations of C∗-algebras 3.2. Exel's question and properties 4. Fredholm conditions 4.1. Fredholm groupoids and their characterization 4.2. Pseudodifferential operators 5. Examples 5.1. Examples related to group actions 5.2. The edge groupoid 5.3. Desingularization groupoids 5.4. Desingularization and singular spaces References 7 9 11 12 13 14 17 17 20 24 25 29 31 32 34 1. Introduction We obtain in this paper necessary and sufficient conditions for classes of opera- tors to be Fredholm. Our results specialize to yield Fredholm conditions for pseu- dodifferential operators on manifolds with cylindrical and poly-cylindrical ends, on manifolds that are asymptotically Euclidean, and on manifolds that are asymp- totically hyperbolic. Other examples of non-compact manifolds covered by our results include operators obtained by desingularization of suitable singular spaces by successively blowing up the lowest dimensional singular strata. 1.1. Background and main result. Let M0 be a Riemannian manifold and let P : H s(M0; E) → H s−m(M0; F ) be an order m pseudo-differential operator acting between Sobolev sections of two smooth, Hermitian vector bundles E, F . Recall that P is called elliptic if its principal symbol σm(P ) is invertible outside the zero section. When M0 is compact, a classical, well-known result [18, 93, 94] states that P is Fredholm if, and only if, it is elliptic. This classical Fredholm result has many applications, so a natural question to ask is to what extent it extends to (suitable) non-compact manifolds. The example of constant coefficient differential operators on Rn shows that that this classical result might be no longer true as stated if M0 is not compact. On certain classes of manifolds, however, it is possible to reformulate this classical result as follows. Let M0 be a non-compact manifold with "amenable ends" (see Subsection 4.2 for the precise definition). Then we can associate to M0 the following data: (1) Smooth manifolds Mα, α ∈ I, and (2) Lie groups Gα acting on Mα, α ∈ I, (for a suitable index set I), satisfying the following theorem. FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 3 Theorem 1.1. Let P be an order m pseudodifferential operator on M0 compatible with the geometry. Then one can associate to P certain Gα-invariant pseudodiffer- ential operators Pα on Mα with the following property: P : H s(M0; E) → H s−m(M0; F ) is Fredholm ⇔ P is elliptic and Pα: H s(Mα; E) → H s−m(Mα; F ) is invertible for all α ∈ I . Many results of this kind were obtained for psedudodifferential operators before. See for instance [24, 52, 58, 56, 91, 92] and the references therein for a very small sample. Similar results for differential operators were also obtained earlier, see [23, 41, 42, 67, 76] and the references therein, again for a very small sample. Other related results appeared in the context of localization principles, often in relation to the N -body problem. See [21, 29, 30, 46, 53, 78, 85] and the references therein. Inspired by these works, we shall call the operators Pα limit operators (of P ). We shall refer to results similar to Theorem 1.1 as non-local Fredholm conditions. The class of "manifolds with amenable ends" is introduced such that, almost by definition, it is close to being the largest class of manifolds for which Theorem 1.1 is valid. The challenge then becomes to provide a large enough class of manifolds with amenable ends, which we do in this paper. The term "amenable" comes from the fact that certain isotropy groups at infinity must be amenable. In many cases, the manifolds Mα are obtained from the orbits of certain natural vector fields acting on a compactification M of M0 and from their isotropies. This is the case for manifolds with cylindrical and poly-cylindrical ends, for manifolds that are asymptotically Euclidean, and for manifolds that are asymptotically hyperbolic, which are all particular cases of manifolds with amenable ends. See the last section of the paper for explicit statements covering in detail the case of these classes of manifolds. It is interesting to notice that the case of asymptotically hyperbolic manifolds leads to the study of certain operator on solvable Lie groups. Theorem 1.1 is almost a consequence of the results in [44, 45], but the results of those papers turned out to be difficult to use by non-specialists. We have thus tried in the last section of the paper to provide a presentation such that the reader can understand the main results without any knowledge of groupoids or C∗-algebras. 1.2. Exhaustive families of representations. It was realized by some of the authors mentioned above that the non-local Fredholm conditions of Theorem 1.1 are related to the representation theory of certain C∗-algebras A. See also [17, 18, 76, 98] and the references therein. These C∗-algebras are such that they contain the operators of the form a = (1 + ∆)(s−m)/2D(1 + ∆)−s/2 (where ∆ is the positive Laplacian) acting on L2(M0). The reason for considering the operator a is that D is Fredholm if, and only if, a is Fredholm; moreover, a is bounded. In most practical applications, these C∗-algebras can be chosen to be groupoid C∗-algebras. These ideas are used in the proof of Theorem 4.12, which contains Theorem 1.1 as a particular case. The operators Pα in Theorem 1.1 then can be obtained as homomorphic images of the operator P , that is Pα = πα(P ). The relevant families of representations in this setting are those of strictly spectral (and strictly norming) families of representations, introduced by Roch in [84] and recently studied in detail by Nistor and Prudhon in [72]. In that paper, the concept of an exhaustive family of representations has emerged as a useful concept to study strictly spectral families of representations. From a technical point of view, these families of representations form the backbone of the theoretical results in this paper. 4 C. CARVALHO, V. NISTOR, AND Y. QIAO 1.3. Fredholm Lie groupoids. The key point of our approach is to start with a general study of Fredholm conditions for pseudodifferential operators in the frame- work of Fredholm Lie groupoids. A Fredholm Lie groupoid G is, by definition, a Lie groupoid for which the Fredholm property of its associated operators is equivalent to the invertibility of the principal symbol and of its fiberwise boundary restric- tions. More precisely, if M is the set of units of G and if G is the pair groupoid above M0 := M r ∂M , we have that G is a Fredholm Lie groupoid if it has the following property: (1) "a ∈ 1 + C∗ r (G) is Fredholm ⇔ πx(a) is invertible for all x ∈ ∂M ." The reader will recognize in condition (1) the type of Fredholm conditions that appeared in Theorem 1.1 (and which are typically used in practice). It turns out that if the defining condition (1) is satisfied (for all a ∈ 1 + C∗ r (G)) then it will be satisfied for many other operators associated to G, in particular it will be satisfied for a = P ∈ Ψm(G; E, F ), a pseudodifferential operators on G acting between the sections of the vector bundles E, F → M (however, for pseudodifferential operators, one has to add the ellipticity condition). The extension of the Fredholm conditions in Equation (1) from operators in 1+C∗ r (G) to operators in Ψm(G; E, F ) turns out to be almost automatic and is based on the use of strictly spectral and exhaustive families of representations of groupoid C∗-algebras (see Section 3 for definitions and references). Thus, in our paper, we shift the study of the Fredholm conditions from the study of a single operator to the study of suitable algebras containing it. In particular, it will be enough to study the properties and the representations of regularizing pseudodifferential operators. The Fredholm properties of higher order pseudodifferential operators will then follow simply by including the ellipticity condition. Thus the Fredholm conditions established in this paper will be formulated in terms of representations of groupoid C∗-algebras. We give various characterizations of Fredholm groupoids and provide methods to prove that individual groupoids are Fredholm. We introduce the class of stratified submersion groupoids, which is built out by glueing fibered pull-backs of bundles of Lie groups. We show that a stratified submersion groupoid with amenable sta- bilizers is Fredholm and hence that it yields a manifold with amenable ends. 1.4. Examples and applications. The main significance of the class of stratified submersion Lie groupoids is that many of the groupoids that appear in practice exhibit this structure naturally. Typically, using the notations as above, we have that M identifies with a compactification of M0 to a manifold with corners, where the behavior at F = M \ M0 models the behavior at infinity ('at the ends'). Often the corresponding isotropy groups are amenable. A class of manifolds that fits this perspective is the class of Lie manifolds [44, 45], a few examples of which appear in Section 5. Moreover, stratified submersion Lie groupoids are also tailored to applications to singular spaces obtained by (iterative) desingularization procedures, desingular- ization being the analog in the category of groupoids of the blow-up construction in the category of manifolds with corners. In fact, it can be seen that desingular- ization preserves Fredholm groupoids, and that the class of stratified submersion groupoids is closed under desingularization. While we do not pursue this approach FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 5 here in full generality, we outline the key ideas involved and give the construction of such a groupoid yielding Fredholm conditions for the edge calculus. Our results can be extended to products of such manifolds, or to manifolds that locally at infinity are products of such manifolds. 1.5. Contents of the paper. The paper consists of roughly two parts: the the- oretical part and applications. The applications are included in the last section, which we tried to write in such a way that they can, to a large extent, be read independently of the rest of the paper. The reader interested only in applications can thus start immediately with Section 5. At least the main results of that section should be understandable without any knowledge of groupoids. We now describe in detail the main contents of the paper. We start with review- ing the relevant topics related to groupoids and groupoid C∗-algebras, so Section 2 is mostly background material on locally compact groupoids G, on their Haar systems, and their C∗-algebras. We then review manifolds with corners and tame submersions, and we recall the definitions of Lie groupoids in the framework that we need, that is, that of manifolds with corners. Note that all our Lie groupoids will be second countable and Hausdorff. We finish this section with a review of some examples of Lie groupoids that will play a role in what follows. Section 3 contains preliminaries on exhaustive families of representations from [72] and some general results on groupoid C∗-algebras. We recall the notions of strictly spectral and strictly norming families [84] and the main results from [72]. We then introduce groupoids with Exel's property, respectively, strong Exel's prop- erty, when the induced family of regular representations is exhaustive for the re- duced C∗-algebra, respectively, for the full C∗-algebra. We prove that groupoids given by fibered pull-backs of bundles of amenable Lie groups always have Exel's strong property, and we introduce the class of stratified submersion groupoids, given essentially by glueing fibered pull-backs of bundles of Lie groups. Our main result here is that stratified submersion Lie groupoids with amenable isotropy groups always have Exel's strong property. In Section 4, we define Fredholm Lie groupoids. Note that in the following sec- tions, we will work always in the setting of Lie groupoids. We provide a characteri- zation of Fredholm Lie groupoids using strictly spectral families of regular represen- tations and show that groupoids that have the strong Exel's property are Fredholm. It follows that stratified submersion Lie groupoids with amenable isotropy groups are Fredholm. We then specialize to algebras of pseudodifferential operators on Lie groupoids and obtain the crucial Theorems 4.12 and 4.17. The last section of the paper, Section 5, contains examples and applications of our results, namely of Theorem 4.17. We start with group actions, define the associated transformation groupoid, and use this construction to describe Fredholm conditions related to several pseudodifferential calculi: the b-groupoid that models operators on manifolds with cylindrical and poly-cylindrical ends; the scattering groupoid that models operators on manifolds that are asymptotically Euclidean, and also operators on asymptotically hyperbolic spaces. We then consider the edge calculus and construct a suitable Fredholm stratified submersion groupoid that will recover Fredholm conditions. This is a particular case of a desingularization groupoid, whose construction is outlined in the following subsection. We then give more explicit examples of the desingularization and blow-up process. A first example deals with the blow-up of a smooth submanifold along another smooth, 6 C. CARVALHO, V. NISTOR, AND Y. QIAO compact manifold. The second example extends this construction to manifolds with boundary. Finally, the last example deals with the iterated blow-up of a singular stratified subset of dimension one. See the main body of the paper for specific references to the existing literature. For simplicity, in view of the applications considered in this paper, we shall work almost exclusively with Lie groupoids (and their reductions), although some defini- tions and results are valid in greater generality. For the most part, our manifolds will be Hausdorff and second countable. 2. Groupoids and their C∗-algebras We recall in this section some basic definitions and properties of groupoids. Although our interest in applications lies mainly in Lie groupoids, we have found it convenient to consider also the general case of locally compact groupoids, so we shall discuss these two cases in parallel. We refer to Mackenzie's books [49, 50] for more details and, in general, for a nice introduction to the subject of Lie groups and Lie groupoids, as well as to further references and historical comments. See also [12, 60, 81, 106] for the more specialized issues relating to analytic applications. Most of the needed results can be found also in [69], whose approach we also use here. 2.1. Locally compact groupoids. Let us introduce groupoids as in [12, 50, 81]. Definition 2.1. A groupoid is a pair G = (G(1), G(0)), where G(i), i = 0, 1, are sets, together with structural morphisms (1) d, r : G(1) → G(0) (the "domain" and "range"), (2) ι : G(1) → G(1) (the "inverse"), (3) u : G(0) → G(1) (the inclusion of units), and (4) µ : G(2) := {(g, h) ∈ G(1) × G(1) d(g) = r(h)} → G(1) (the product), with the following properties (we write gh := µ(g, h), for simplicity): (1) d(gh) = d(h), r(gh) = r(g) if d(g) = r(h); (2) g1(g2g3) = (g1g2)g3 for all gi ∈ G such that d(gi) = r(gi+1); (3) d(u(x)) = x = r(u(x)), for all x ∈ M . (4) gu(d(g)) = g and u(r(g))g = g for all g ∈ G. (5) gι(g) = u(r(g)) and ι(g)g = u(d(g)) for all g ∈ G. Moreover, we see that G(0) can be regarded as the set of objects of a category with morphisms G(1). The objects of G will also be called units and the morphisms of G will also be called arrows. Actually, a groupoid G is simply a small category in which every morphism is invertible. (A small category is one whose objects form a set.) For convenience, we shall identify G = G(1) and denote M := G(0). We shall write G ⇒ M for a groupoid G with units M . In this paper, G always will denote a groupoid. Definition 2.2. A locally compact groupoid is a groupoid G ⇒ M such that: (1) G and M are locally compact spaces, with M Hausdorff; (2) the structural morphisms d, r, ι, u, and µ are continuous; (3) d is surjective and open. FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 7 See Subsection 2.4 for examples of groupoids. Let us notice that the definition In this paper, however, all requires only the space of units M to be Hausdorff. spaces will be assumed or proved to be Hausdorff. In what follows, we denote by GA := d−1(A), GA = r−1(A) and GB A := d−1(A) ∩ r−1(B). We call GA A the reduction of G to A. If A is G-invariant, in the sense that A = GA = GA = r−1(A), then GA is also a groupoid, called the restriction of G to GA x = d−1(x) ∩ r−1(x) is a group, called the isotropy group at x. A. For any x ∈ X, Gx We will see several examples of groupoids in more detail when we introduce Lie groupoids. 2.2. Haar systems and C∗-algebras. We now recall the definition of a Haar system of a locally compact groupoid and we use this opportunity to fix some more notations to be used throughout the rest of the paper. We refer to [12, 39, 81] for more information on the topics discussed in this subsection. If G is a locally compact groupoid, we shall denote by Cc(G) the space of contin- uous, complex valued, compactly supported functions on G. Definition 2.3. A right Haar system for a locally compact groupoid G is a family λ = {λx}x∈M , where λx is a Borel regular measure on G with support supp(λx) = d−1(x) =: Gx for every x ∈ M = G(0), satisfying (i) The continuity condition: M ∋ x 7→ λx(ϕ) := ZGx is continuous for ϕ ∈ Cc(G). ϕ(g)dλx(g) ∈ C (ii) The invariance condition: ZGr(g) ϕ(hg)dλr(g)(h) = ZGd(g) ϕ(h)dλd(g)(h) for all g ∈ G and ϕ ∈ Cc(G). We shall assume from now on that all our locally compact groupoids are endowed with a (right) Haar system. Let thus G be a locally compact groupoid and {λx} be the Haar system associated to it. The space Cc(G) has a natural product given by the formula (ϕ1 ∗ ϕ2)(g) := ϕ1(gh−1)ϕ2(h)dλd(g)(h) . Z d−1(d(g)) This makes Cc(G) into an associative ∗-algebra with the involution defined by ϕ∗(g) := ϕ(g−1) for all g ∈ G and ϕ ∈ Cc(G). There also exists a natural algebra norm on Cc(G) defined by kf kI := maxn sup x∈MZ ϕdλx, sup x∈MZ ϕ∗dλxo. The completion of Cc(G) with respect to this norm k · kI will be denoted by L1(G). Recall [26] that a C∗-algebra is a complex algebra A together with a conjugate linear involution ∗ and a complete norm k k such that (ab)∗ = b∗a∗, kabk ≤ kakkbk, and ka∗ak = kak2, for all a, b ∈ A. Let H be a Hilbert space and denote by L(H) the space of linear, bounded operators on H. Then L(H) is a C∗-algebra. A 8 C. CARVALHO, V. NISTOR, AND Y. QIAO representation of a C∗-algebra A on the Hilbert space Hπ is a ∗-morphism π : A → L(Hπ). To a locally compact groupoid G (endowed with a Haar system), there are as- r (G), sociated two basic C∗-algebras, the full and reduced C∗-algebras C∗(G) and C∗ whose definition we now recall. Definition 2.4. The (full) C∗-algebra associated to G, denoted C∗(G), is defined as the completion of Cc(G) with respect to the norm kπ(ϕ)k , kϕk := sup π where π ranges over all contractive ∗-representations of Cc(G). Let us define as usual for any x ∈ M the regular representation πx : C∗(G) → L(L2(Gx, λx)) by the formula πx(ϕ)ψ(g) := ϕ ∗ ψ(g) := ZGd(g) ϕ(gh−1)ψ(h)dλd(g)(h) , φ ∈ Cc(G) . We then define similarly the reduced C∗-algebra C∗ with respect to the norm r (G) as the completion of Cc(G) kϕkr := sup x∈M kπx(ϕ)k . The groupoid G is said to be metrically amenable if the canonical surjective ∗- homomorphism C∗(G) → C∗ r (G), induced by the definitions above, is also injective. Also, for further use, we note that if G is second countable, then C∗(G) is a separable C∗-algebra. Remark 2.5. For any G-invariant, locally closed subset A ⊆ M , the reduced groupoid GA = GA A is locally compact and has a Haar system λA obtained by restricting the Haar system λ of G to GA. In particular, we can construct as above the corresponding C∗-algebra C∗(GA) and the reduced C∗-algebra C∗ r (GA). For any closed subset A ⊆ M , the subset d−1(A) ⊆ G is also closed, so the restriction map Cc(G) → Cc(d−1(A)) is well defined. If A is also G-invariant then the restriction extends by continuity to both a ∗-homomorphism ρA : C∗(G) → C∗(GA) and a ∗-homomorphism (ρA)r : C∗ r (G) → C∗ r (GA). We have the following well known, but important result [64, 65, 82] that we record for further reference. Proposition 2.6. Let G ⇒ M be a second countable, locally compact groupoid with a Haar system. Let U ⊂ M be an open G-invariant subset, F := M r U . (i) C∗(GU ) is a closed two-sided ideal of C∗(G) that yields the short exact sequence 0 → C∗(GU ) → C∗(G) ρF−→ C∗(GF ) → 0 . (ii) If GF is metrically amenable, then one has the exact sequence 0 → C∗ r (GU ) → C∗ r (G) (ρF )r −−−−−→ C∗ r (GF ) → 0 . (iii) If the groupoids GF and GU (respectively, G) are metrically amenable, then G (respectively, GU ) is also metrically amenable. Proof. The first assertion is well known, see for instance [65, Lemma 2.10]. The second statement is in [82, Remark 4.10]. The last part follows from (ii) and the Five Lemma. (cid:3) FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 9 Remark 2.7. We notice that the exact sequence of Proposition 2.6 corresponds to a disjoint union decomposition G = GF ⊔ GU . 2.3. Manifolds with corners and Lie groupoids. Even if one is interested only in analysis on smooth manifolds (that is, in manifolds without corners or boundary), one important class of applications will be to spaces obtained as the blow-up with respect to suitable submanifolds, which leads to manifolds with corners, as each blow-up increases the highest codimension of corners by one. To model the analysis on these spaces, we need Lie groupoids. The advantage of using Lie groupoids is that they have a distinguished class of Haar systems defined using the smooth structure. We begin with some background material following [69]. It will be impor- tant to distinguish between smooth manifolds without corners (or boundaries) and (smooth) manifolds with corners. The former will be smooth manifolds, while the later will be simply manifolds. Thus, in this paper, a manifold M is a second count- able, Hausdorff topological space locally modeled by open subsets of [−1, 1]n with smooth coordinate changes. In particular, our manifolds may have corners. By con- trast, a smooth manifold will not have corners (or boundary). A point p ∈ M of a manifold (with corners) is called of depth k if it has a neighborhood Vp diffeomorphic to [0, a)k × (−a, a)n−k, a > 0, by a diffeomorphism φp : Vp → [0, a)k × (−a, a)n−k with φp(p) = 0. The set of inward pointing tangent vectors in v ∈ Tx(M ) define a closed cone denoted T + x (M ). A function f : M → M1 between two manifolds with corners will be called smooth if its components are smooth in all coordinate charts. A little bit of extra care is needed here in defining the derivatives. This is illustrated if f : [0, 1] → R, then f ′(0) := limh→0,t>0 h−1(f (h) − clearly in one dimension: f (0)), whereas f ′(1) := limh→0,t>0 h−1(f (1) − f (1 − h)). Since limh→0 h−1(f (x + h) − f (x)) = limh→0 h−1(f (x) − f (x − h)) =: f ′(x), for 0 < x < 1, this defines unambiguously f ′(x). A similar comment is in order in higher dimensions as well. In local coordinates, a smooth function [0, 1]n → [0, 1]k is one that is the restriction of a smooth function Rn → Rk. Let M and M1 be manifolds with corners and f : M1 → M be a smooth map. Then f induces a vector bundle map df : T M1 → T M such that df (T + z (M1)) ⊂ T + If the smooth map f : M1 → M is injective, has injective differential f (z)M . df , and has locally closed range, then we say that f (M1) is a submanifold of M . We stress the condition that submanifolds be locally closed (that is, the intersec- tion of an open and a closed subset). Other than this conditions, our concept of submanifold is the most general possible. Definition 2.8. A tame submersion h between two manifolds with corners M1 and M is a smooth map h : M1 → M such that its differential dh is surjective everywhere and (dhx)−1(T + h(x)M ) = T + x M1 . (That is, dh(v) is an inward pointing vector of M if, and only if, v is an inward pointing vector of M1.) Clearly, if h : M1 → M is a tame submersion of manifolds with corners, then x and h(x) will have the same depth. We have the following well known lemma (see for instance [69]). 10 C. CARVALHO, V. NISTOR, AND Y. QIAO Lemma 2.9. Let h : M1 → M be a tame submersion of manifolds with corners. (i) For m1 ∈ M1, there exists an open neighborhood U of m1 in M1 such that h(U ) is open and the restriction of h to U is a C∞ fibration with basis h(U ). (ii) Let L ⊂ M be a submanifold, then L1 := h−1(L) is a submanifold of M1 of rank ≤ the rank of L. We now define Lie groupoids roughly by replacing our spaces with manifolds with corners and the continuity conditions with smoothness conditions. The reason why Lie groupoids play an important role in applications is that they model the distribution kernels of many interesting classes of pseudodifferential operators. Definition 2.10. A Lie groupoid is a groupoid G ⇒ M such that (i) G and M are manifolds with corners, (ii) the structural morphisms d, r, ι, and u are smooth, (iii) d is a tame submersion of manifolds with corners, and (iv) µ is smooth. In particular, the Lie groupoids used in this paper are locally compact groupoids and are second countable and Hausdorff. In different contexts in other papers, it is sometimes useful not to assume the set of arrows G of a Lie groupoid to be Hausdorff, although the space of units M of a Lie groupoid G is always assumed to be Hausdorff. Lie groupoids were introduced by Ehresmann. A more general class also useful in applications is that of continuous family groupoids, where we assume continuity only along the units, keeping smoothness along the fibers. See [44, 75]. For the rest of this section, G will always be a Lie groupoid. In fact, most of the groupoids that we will consider in what follows will be Lie groupoids. However, one has to be careful since the restriction of a Lie groupoid is not necessarily a Lie groupoid itself. Moreover, we shall assume that all our Lie groupoids are Hausdorff. Since d and r are tame submersions, it follows from Lemma 2.9 that the fibers Gx := d−1(x), x ∈ M , are smooth manifolds (that is, they have no corners). Sim- ilarly, the same lemma implies that the set G(2) ⊂ G × G of composable units is a manifold as well (but it may have corners). In particular, it makes sense to consider smooth maps G(2) → G(1). Remark 2.11. Let G ⇒ M be a Lie groupoid and let A(G) := ∪x∈M TxGx, where Gx := d−1(x), as usual. Let us denote by f∗ : T M1 → T M2 the differential of a smooth map f : M1 → M2. Then A(G) is the restriction to units of the vector bundle ker(d∗), and hence it has a natural structure of vector bundle over M . It is called the Lie algebroid of G. The differential r∗ of the range map r : G → M then induces a map := r∗A(G) → T M , called the anchor map of A(G). We let Lie(G) := (Γ(M ; A(G))) denote the image by of the space of sections of A(G). It is coincides with the image by r∗ of the space d-vertical (i.e. tangent to the fibers Gx of d) vector fields on G that are invariant for the action of G on itself by right multiplication. Therefore both Γ(M ; A(G)) and Lie(G) := (Γ(M ; A(G))) are Lie algebras for the Lie bracket. Both Lie algebras will play an important role in the analysis of differential operators. Remark 2.12. We can use the vector bundle A(G) to define a Haar system. Let D := ΛnA(G), where n is the dimension of the Lie algebroid of G. The pull-back vector bundle r∗(D) is the bundle of 1-densities along the fibers of d. A trivialization FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 11 of D will hence give rise to a right invariant set of measures on Gx and hence to a right Haar system. We conclude that a Lie groupoid always has a (right) Haar system, and we shall always assume that it is obtained as in the remark above. 2.4. Examples of Lie groupoids. We continue with various examples of con- structions of Lie groupoids that will be needed in what follows. Recall that we only consider Hausdorff groupoids in this paper. Also, recall that a Lie groupoid G is metrically amenable if the canonical morphism C⋆(G) → C⋆ r (G) is an isomorphism. Example 2.13. Any Lie group G gives rise to a Lie groupoid with set of units reduced to one point: the identity element of G. Thus d, r : G → M := {e} are constant, µ : G × G → G is the group multiplication, ι(g) := g−1 is the (usual) inverse, and u : M → G is the inclusion of the unit. We have that the Lie algebroid A(G) of G coincides with the Lie algebra of G (the definition of a Lie algebroid was recalled in Remark 2.11). Hence Lie(G) := Γ(A(G)) is also the Lie algebra of G. We have that G is metrically amenable if, and onl if, it is amenable in the usual sense of groups. In particular, if G is solvable, then G is amenable. Example 2.14. Let M be a manifold with corners (hence Hausdorff by our conven- tions), let G = M , and d = r = u = ι = idM . Then G is a Lie groupoid with only units. We shall call a Lie groupoid with these properties a manifold. We have that A(M ) = M × {0}, that is, the zero vector bundle. We have Lie(G) = 0. Example 2.15. Let Gi → Mi, i = 1, 2, be two Lie groupoids. Then G1 × G2 is a Lie groupoid with units M1 × M2 and A(G1 × G2) = A(G1) × A(G2). We have that Lie(G1 × G2) is a suitable topological tensor product of Lie(G1) and Lie(G2). The product of a manifold with a Lie group is thus again a Lie groupoid. We can "twist" this example to obtain a (smooth) "bundle of Lie groups." Example 2.16. Let G be a Lie group with automorphism group Aut(G) and let P → M be a smooth, locally trivial, principal Aut(G)-bundle, with M and P manifolds with corners (hence Hausdorff, by our conventions). Then the associated fiber bundle G := P ×Aut(G) G with fiber G is a Lie groupoid with units M . A Lie groupoid of this form will be called a smooth bundle of Lie groups. It satisfies d = r. If P → M is also differentiable (with P and M manifolds, possibly with corners), then G is a Lie groupoid. We have A(G) ≃ P ×Aut(G) Lie(G) is a smooth bundle of Lie algebras and : A(G) → T M is the zero map. The Lie algebra Lie(G) identifies thus with the family of sections of the bundle of Lie algebras P ×Aut(G) Lie(G). If G is amenable, then G is metrically amenable. The fact that G is metrically amenable when G is amenable will be of crucial importance to us and will be used when the group G is a solvable Lie group. The following simple example of the "pair groupoid" will be fundamental in what follows. Example 2.17. Let M be a smooth manifold. (Thus M has no corners, according to our terminology). Then the pair groupoid of M is G := M ×M . It is a groupoid with units M . The structural morphisms are as follows: d is the second projection, r is the first projection, and the product µ is given by (m1, m2)(m2, m3) = (m1, m3). This determines u(m) = (m, m) and ι(m, m′) = (m′, m). We thus have that the 12 C. CARVALHO, V. NISTOR, AND Y. QIAO c (G) = C∞ product on C∞ c (M × M ) is the product of operators with integral kernels. We have A(M × M ) = T M , with anchor map the identity map. One crucial feature of the pair groupoid is that, for any x, the regular representation πx defines an isomorphism between C∗(M ×M ) and the ideal of compact operators in L(L2(M )). In particular, all pair groupoids are metrically amenable. Example 2.18. Let M be a smooth manifold (so without corners) and M be it universal covering. Let π1(M ) be the fundamental group of M associated to some fixed point of M . A related example to the pair groupoid is that of the path groupoid P(M ) := ( M × M )/π1(M ) of M , which will have the same Lie algebroid as the pair groupoid: A(P(M )) = T M . The analysis associated to P(M ) is that of π1(M )- invariant operators on the covering space M , thus quite different to that of the pair groupoid. This underscores the importance of choosing the right integrating groupoid when interested in Analysis. We extend the example of the pair groupoid by defining fibered pull-back groupoids [36, 37] (we use the terminology in [69], however). Example 2.19. Let again M and L be manifolds and f : M → L be continuous map. Let H ⇒ L be a Lie groupoid. The fibered pull-back groupoid is then f ↓↓(H) := { (m, g, m′) ∈ M × H × M, f (m) = r(g), d(g) = f (m′) } , with units M and product (m, g, m′)(m′, g′, m′′) = (m, gg′, m′′). We shall also sometimes write M ×f H×f M = f ↓↓(H) for the fibered pull-back groupoid. Assume now that f is a tame submersion (in particular, that it is smooth). Then f ↓↓(H) is a Lie groupoid and its Lie algebroid is A(f ↓↓(H)) = {(ξ, X) ξ ∈ A(H) , X ∈ T M , (ξ) = f∗(X)} , the thick pull-back Lie algebroid f ↓↓A(H) of A(H), see [49, 50]. The following particular example of a fibered pull-back will be useful for the study of the b-groupoid Gb in Section 5. Example 2.20. Let us assume that M is a smooth manifold and let B be its set of connected components. Let G be a Lie group. We let H := B × G, the product of a manifold and a Lie group, and f : M → B be the map that associates to a point its connected component. Then f ↓↓(H) is the topological disjoint union of the groupoids (F × F ) × G (product of the pair groupoid and a Lie group) for F ranging through the connected components of M . 3. Exhaustive families of representations and Exel's question We next recall some basic facts on exhaustive families of representations following [72]. A groupoid has Exel's property when its set of regular representations is exhaustive. Exel [27] has asked which groupoids have this property. Establishing that a groupoid has Exel's property will turn out to be important in establishing Fredholm conditions in the next section. At the end of the section, we prove that groupoids given by fibered pull-backs of bundles of Lie groups always have Exel's property, which leads us to introduce, and extend this property to, the more general class of stratified submersion groupoids. FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 13 3.1. Exhaustive families of representations of C∗-algebras. A two-sided ideal I ⊂ A of a C∗-algebra A is called primitive if it is the kernel of an irreducible rep- resentation of A. We shall denote by Prim(A) the set of primitive ideals of A. The trivial ideal A (of A) is not considered a primitive ideal (of A), so A /∈ Prim(A). For any representation φ of A, we define its support supp(φ) ⊂ Prim(A) as the set of primitive ideals of A containing ker(φ). Let J be a closed, two-sided ideal of A. Then Prim(J) = {I ∈ Prim(A) J 6⊂ I} and Prim(A/J) identifies with {I ∈ Prim(A) J ⊂ I} = Prim(J)c. We endow Prim(A) with the hull-kernel topol- ogy, which is the topology whose open sets are those of the form Prim(J), with J a two-sided ideal of A. We shall need the following definition [72]. Definition 3.1. Let F be a set of representations of a C∗-algebra A. We say that F is exhaustive if Prim(A) =Sφ∈F supp(φ). In practice, we rather need families with the following definition [84]. Definition 3.2. Let F be a set of representations of a unital C∗-algebra A. (i) We say that F is strictly norming if, for any a ∈ A, there exists φ ∈ F such that kφ(a)k = kak. (ii) We say that F is strictly spectral if, for any a ∈ A, we have that a is invertible in A if, and only if, φ(a) is invertible for all φ ∈ F . The two definitions above can be formulated for sets of morphisms or sets of primitive ideals. We also have that F∗ := {π ∈ F π 6= 0} has the same properties ("exhaustive," "strictly norming," ... ) as F . If A is non-unital, we modify the last definition as follows [72]. Let A+ := A ⊕ C and χ0 : A+ → C be the morphism defined by χ0 = 0 on A and χ0(1) = 1. We then replace A with A+ and F with F + := F ∪ {χ0}. This works also for exhaustive families since F is exhaustive for A if, and only if, F + is exhaustive for A+. Sometimes it is convenient to use the following alternative characterization of strictly spectral sets of representations. The set of representations F of A is strictly spectral if it satisfies the following property: 1 + a ∈ A+, a ∈ A, is invertible if, and only if, 1 + φ(a) is invertible for any φ ∈ F . The set of all equivalence classes of irreducible representations of a unital C∗- algebra is strictly norming (see [26, 27]). Therefore any exhaustive family is strictly norming [72]. We recall the following results from [72, 84], which establish the relations between these notions. Theorem 3.3. Let F be a set of non-degenerate representations of a C∗-algebra A. Then F is strictly norming if, and only if, it is strictly spectral. Every exhaustive set of representations is strictly spectral. If A is furthermore separable, then the converse is also true. Let A be a C∗-algebra and I ⊂ A be a closed two-sided ideal. Recall that any nondegenerate representation π : I → L(H) extends to a unique representation π : A → L(H). (See [26, Proposition 2.10.4].) This leads to the following results (see [72], Proposition 3.15 and Corollary 3.16). Proposition 3.4. Let I ⊂ A be an ideal of a C∗-algebra. Let FI be a set of nondegenerate representations of I and FA/I be a set of representations of A/I. Let F := FI ∪ FA/I , regarded as a family of representations of A. If FI and FA/I 14 C. CARVALHO, V. NISTOR, AND Y. QIAO are both exhaustive, then F is also exhaustive. The same result holds by replacing exhaustive with strictly norming. In the following corollary, one should think of the invertibility of a in A/I as an "ellipticity" condition. Corollary 3.5. Let I ⊂ A be an ideal of a unital C∗-algebra A and let FI be a strictly spectral set of nondegenerate representations of I. Let a ∈ A. Then a is invertible in A if, and only if, it is invertible in A/I and φ(a) is invertible for all φ ∈ FI. We now address Morita equivalence [83] in a very simple form (see also [75, 81]). It is known that there exists a homeomorphism between the primitive ideal spectra of Morita equivalent C∗-algebras, so exhaustive families of representations will correspond to exhaustive families of representations under Morita equivalence. In particular, we have the following result. Proposition 3.6. Let F be a set of representations of a C∗-algebra A. (i) If Fn := {π ⊗ 1 π ∈ F} is the corresponding family of representations of Mn(A) = A ⊗ Mn(C), then F is exhaustive if, and only if, Fn is exhaustive. (ii) Let I ⊂ A be a closed, two-sided ideal. If F is exhaustive, then FI := {πI π ∈ F} is an exhaustive family of representations of I. (iii) Let e ∈ Mn(A) be a projection. Then Fn defines, by restriction, an exhaustive family of representations of eMn(A)e. Proof. The first part follows by Morita equivalence. The second part follows since supp(φI ) = supp(φ) ∩ Prim(I). The last part is proved by combining the first two parts. Indeed, Fn defines an exhaustive set of representations of Mn(A), and hence an exhaustive set of representations of Mn(A)eMn(A). Since the algebra eMn(A)e is Morita equivalent to Mn(A)eMn(A), by Proposition 4.27 in [31] (see also [83]), the result follows. (cid:3) 3.2. Exel's question and properties. We will let G denote a Lie groupoid and let R(G) will denote its set of regular representations. The discussion below usu- ally makes sense in the more general setting of locally compact groupoids, but we restrict, nevertheless to Lie groupoids, for simplicity. In [27], Exel has asked for which groupoids G is the set R(G) a strictly norming set of representations of C∗ r (G). An example due to Voiculescu, see [72], shows that not every locally compact groupoid has this property. (The question whether all Lie groupoids have Exel's property is still open.) It is easier to work with exhaustive families of representations, so we introduce the following definition. Definition 3.7. We say that G has Exel's property if R(G), the set of regular representations of G, is an exhaustive set of representations for C∗ r (G). We say that G has Exel's strong property if R(G) is an exhaustive set of representations for C∗(G). We have that [φ∈F supp(φ) = [φ∈F Prim(A)\ Prim(ker(φ)) = Prim(A)\ \φ∈F Prim(ker(φ)). Hence, if F is an exhaustive family of representations of a C∗-algebra A, then Tφ∈F Prim(ker(φ)) = ∅, and hence F is a faithful family of representations of A. FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 15 In particular, G has Exel's strong property if, and only if, G has Exel's property and is metrically amenable. We shall need the following consequence of Propositions 2.6 and 3.4 of the pre- vious subsection. Corollary 3.8. Let U ⊂ M be an open, G-invariant subset and assume G to be second countable. We let F := M r U . Let us assume that GU and GF have Exel's strong property. Then G also has Exel's strong property. r (GU ), A := C∗(G). Then A/I ≃ C∗(GF ) ≃ C∗ Proof. Let I := C∗(GU ) = C∗ r (GF ) by Proposition 2.6(ii). Since GF and GU are metrically amenable, it follows that G is metrically amenable as well, by Proposition 2.6(iii). Proposition 3.4 then gives that R(G) = R(GU ) ∪ R(GF ) is an exhaustive family of representations of C∗(G) ≃ C∗ (cid:3) r (G), that is, G has Exel's strong property, as claimed. More generally, we can consider groupoids given by a filtration. Corollary 3.9. Let G ⇒ M be a second countable groupoid and assume that Ui are open, G-invariant subsets of M , such that ∅ := U−1 ⊂ U0 ⊂ . . . ⊂ Ui−1 ⊂ Ui ⊂ . . . ⊂ UN := M , and for each S := Ui r Ui−1, GS has Exel's strong property. Then G also has Exel's strong property. Proof. Corollary 3.8 yields the result by induction, as follows. If N = 1, it reduces to Corollary 3.8. Assuming that the result holds for a filtration in N − 1 sets, then GUN −1 has Exel's strong property. Since by assumption GM\UN −1 also has, the result follows again by Corollary 3.8. (cid:3) Here is a simple, but very important example of a class of Lie groupoids G for which the set R(G) of regular representations is an exhaustive set of representations of C∗(G). The idea of the proof below is that, locally, the C∗-algebra of G is of the form C∗(G) ⊗ K, where K are the compact operators on the (typical) fiber of f . We found it convenient to formalize the proof of this well-known result using Morita equivalence of groupoids. Proposition 3.10. Let H ⇒ L be a bundle of Lie groups with fiber G and let f : M → L be a tame submersion. Define G := f ↓↓(H). Then G is a Lie groupoid with Exel's property. If G is amenable, then G has Exel's strong property. Proof. We have already discussed that G := f ↓↓(H) is a Lie groupoid (because f is a tame submersion). We endow G with one of the standard Haar measures coming from the Lie groupoid structure. Let X = M ×L H := {(m, γ) ∈ M × Hf (m) = r(γ)} . Then X defines a Morita equivalence between G and H, and hence a Morita equiva- lence between their full and reduced C∗-algebras of these groupoids, by the classical results of [64, 96]. See also [100]. In particular, Prim(C∗(G)) and Prim(C∗(H)) are homeomorphic by natural homeomorphisms [83]. The same is true for Prim(C∗ r (G)) and Prim(C∗ All irreducible representations of C∗ r (H) factor through an evaluation morphism ex : C∗ r (G), for some x ∈ M , where Gx is the fiber of H → M above x. We have that the regular representation πx is obtained from the regular representation of Gx via ex, the evaluation at x. Hence G has Exel's property. r (H)). r (H) → C∗ r (Gx) ≃ C∗ 16 C. CARVALHO, V. NISTOR, AND Y. QIAO Let us assume now that G is amenable. Hence the fibers of H → M are amenable, and therefore H is metrically amenable. Hence G is also metrically amenable. Since G was already proved to have Exel's property, it follows that it has also Exel's strong property. (cid:3) We notice that if G is (isomorphic to a groupoid) as in the proposition above, then H ⇒ L and f : M → L are uniquely determined, up to equivalence. Indeed, L is diffeomorphic to the set of orbits of G (acting on M ), f becomes the quotient map M → M/G and H is obtained from the isotropy groupoid of G. The above results suggest the following definition. Definition 3.11. Let H be a groupoid with units M . We say that H is a stratified submersion groupoid with filtration (Ui) if Ui are open, H-invariant subsets of M , ∅ := U−1 ⊂ U0 ⊂ . . . ⊂ Ui−1 ⊂ Ui ⊂ . . . ⊂ UN := M , and each S := Ui r Ui−1 is a manifold, possibly with corners, such that there exist a Lie group bundle GS → BS and a tame submersion fS : S → BS of manifolds with corners such that HS ≃ f ↓↓ S (GS) . A groupoid H will be called a stratified submersion groupoid if it is a stratified submersion groupoid for some filtration. Remark 3.12. It is clear from the definition that if H is a stratified submersion groupoid, then, with the notation of the definition, each HUk rUj , j < k, is also a stratified submersion groupoid. We have to notice, however, that HUkrUj may not be a Lie groupoid, even if H is a Lie groupoid, because Uk r Uj may not be a manifold (even if it is a union of manifolds). As we shall see, the class of stratified submersion groupoids is large enough to comprise many classes of groupoids and operators that arise in practice. Combining Corollary 3.9 with Proposition 3.10, we now obtain the following result. Theorem 3.13. Let G ⇒ M be a stratified submersion Lie groupoid and assume that all the isotropy groups Gx x , x ∈ M , are amenable. Then the set R(G) of regular representations of G is an exhaustive family of representations of C∗(G). In particular, G is metrically amenable and has Exel's property. Proof. Let us use the notation of Definition 3.11. Indeed then, the isotropy group Gx x is the fiber of GS → BS above fS(x), for x ∈ S. Thus, by Proposition 3.10, all the groupoids GS are metrically amenable and have Exel's property. Using Corollary 3.9, we obtain that G has Exel's strong property, that is, that it is metrically amenable and has Exel's property. (cid:3) In the next section, we will see how this property plays a role in establishing Fredholm conditions for operators on groupoids. Remark 3.14. Let H be a stratified submersion Lie groupoid as in Definition 3.11. In view of the remark 2.7, we have then the following disjoint union decomposition: H = ⊔S f ↓↓ S (GS) = ⊔S S ×B S ×B GS , where ×B is the fibered product over the base (that is, over BS). We notice also that two regular representations πx and πy are unitarily equivalent if there is g ∈ G such that d(g) = x and r(g) = y. For a stratified submersion groupoid, this is FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 17 then the case exactly if x, y ∈ S := Ui−1 r Ui and fS(x) = fS(y) = b ∈ Bs, in which case, these two representations act, up to a unitary equivalence, on the set of square integrable functions on the space Zb := f −1 S (b) × (GS)b. We denote by πb the corresponding representation on L2(Zb), b ∈ BS. 4. Fredholm conditions We now study Fredholm conditions for operators in algebras Ψ containing a reduced groupoid C∗-algebra C∗ r (G) as an essential ideal. (Recall that an ideal of an algebra is essential if its annihilator vanishes.) The groupoids for which we obtain the kind of Fredholm conditions that we want (the kind that are typically used in practice) will be called "Fredholm groupoids." They are introduced and discussed next. Examples of Fredholm groupoids will be provided in the next section. Throughout the rest of this paper, G ⇒ M will denote a second countable Lie groupoid, which is hence a locally compact groupoid and the choice of a metric on A(G) gives rise to a Haar system λx. Also, recall that all our Lie groupoids are assumed to be Hausdorff. 4.1. Fredholm groupoids and their characterization. We now introduce Fred- holm groupoids and give a first characterization of these groupoids. Recall that πx : C∗(G) → L(L2(Gx, λx)) denotes be the regular representation on Gx, given by left convolution, πx(φ)ψ := φ ∗ ψ (Definition 2.4). We shall use the following notation throughout the rest of the paper. Notations 4.1. Let us assume that U ⊂ M is an open, G-invariant subset with GU ≃ U × U (the pair groupoid, see Example 2.17). For any x0 ∈ U , the range map then defines a bijection r : Gx0 → U and hence a measure µ on U corresponding to λx0. This measure does not depend on the choice of x0 ∈ U and leads to isometries L2(Gx0 , λx0 ) ≃ L2(U ; µ) that commute with the action of G. We then denote by π0 the corresponding representation of C∗(G) on L2(U ; µ). It is often called the vector representation of C∗(G). We shall usually write L2(U ) := L2(U ; µ). The representation π0 then defines an isomorphism C∗ r (GU ) = K, the algebra of compact operators on L2(U ). r (GU ) ≃ π0(C∗ Remark 4.2. An important remark is that, since we only consider Hausdorff Lie groupoids, the vector representation π0 is injective, by a result of Khoshkam and Skandalis [40]. In particular, we have that 1 + π0(a) is invertible in π0(C∗ r (G))/K if, and only if, 1 + a is invertible in C∗ r (GU ). We shall thus identify C∗ r (G) with its image under π0, that is, with a class of operators on L2(U ), without further comment. r (G)/C∗ In the following definition, we shall use that C∗ r (GU ) ⊂ ker πx, x /∈ U , by defini- tion, and hence that πx descends to the quotient algebra. Definition 4.3. We say that G is a Fredholm Lie groupoid if: (i) There is an open, dense, G-invariant subset U ⊂ M such that GU ≃ U × U . (ii) Given a ∈ C∗ r (G), we have that 1 + a is Fredholm if, and only if, all 1 + πx(a), x ∈ F r U , are invertible (see 4.1 for notation). The set F := M r U will be called the set of boundary units of G and a set U as in this definition will be called a manifold with amenable ends. 18 C. CARVALHO, V. NISTOR, AND Y. QIAO For a Fredholm Lie groupoid G ⇒ M , we shall always denote by U ⊂ M the G-invariant open subset from the definition of a Lie Fredholm groupoid. Since U is a dense orbit, it is uniquely determined by G. Hence, F := M r U is closed and G-invariant. We shall need the following result in the matricial case. Proposition 4.4. Let us assume that G is a Fredholm Lie groupoid, that M is compact, and that e ∈ Mn(C(M )) is a projection. (i) The set of representations πx, x ∈ F defines a strictly spectral family of representations of eMn(cid:0)C∗ (ii) For any a ∈ eMn(C∗ r (G)/C∗ r (GU )(cid:1)e. r (G))e, we have that 1 + π0(a) is Fredholm if, and only if, all 1 + πx(a), x ∈ F := M r U , are invertible (see 4.1 for notation). Proof. By Atkinson's theorem (which states that an operator is Fredholm if, and only if, it is invertible modulo the compacts) and by the definition of strictly spec- tral families of representations in the non-unital case, one can see that condition (ii) in the definition of Fredhom groupoid is equivalent to the fact that R(GF ) defines a strictly spectral family of representations of the quotient C∗ r (G)/K ≃ r (GU ). Since G is second countable, its associated C∗-algebras are separa- C∗ ble, hence the set {πxx ∈ F } is strictly spectral if, and only if, it is exhaustive, by Theorem 3.3. The first part is then a consequence of Proposition 3.6. r (G)/C∗ The second part follows again from the definition of strictly spectral families of r (GU ) ≃ K(L2(U )), and (cid:3) representations in the non-unital case, from the fact that C∗ from Atkinson's theorem. In the next results, we give an abstract characterization of Lie Fredholm groupoids, using some of the ideas of the previous section. We obtain in particular conditions that will make it simpler to check whether a given Lie groupoid is Fredholm. Recall that, in this paper, all groupoids are Hausdorff. Theorem 4.5. Assume G is a Lie groupoid. If G is Fredholm, then following two conditions are satisfied: (i) The canonical projection C∗ r (G) → C∗ r (GF ) induces an isomorphism r (G)/C∗ (ii) GF has Exel's property. C∗ r (GU ) ≃ C∗ r (GF ) , F := M r U . (Recall that condition (ii) means that R(GF ) is an exhaustive set of representa- tions of C∗ r (GF ).) Proof. Let us assume that G is a Fredholm groupoid and check the two conditions of the statement. Let p : C∗ r (G) → C∗ r (GF ) be the natural projection induced by restriction. It is known that C∗ r (GU ) ⊆ ker(p) (see the proof of [81, Prop.II.4.5 (a)]). To prove (i), we need to show that we have equality C∗ r (GU ) = ker(p). Let us proceed by contradiction, that is, let us assume that C∗ r (GU ) 6= ker(p). Then we can choose a = a∗ ∈ ker(p) r C∗ r (GU ). Note that since πx = πx ◦ p, x ∈ F = M \ U , we have ker p ⊂ ker πx, in particular, 1 − πx(a) = 1 is invertible for all x /∈ U . On the other hand, since a is self-adjoint and non-zero, in the quotient ker(p)/C∗ r (GU ), there is 0 6= λ ∈ R such that λ − a is not invertible in ker(p)/C∗ r (GU ). By rescaling, we may assume λ = 1 and thus 1 − a is not invertible in C∗ r (G)/K. We conclude that in this case R(GF ) is not a strictly spectral family of representations r (GU ) ∼= C∗ r (G)/C∗ FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 19 r (G)/C∗ of C∗ is invertible for all x /∈ U . This is a contradiction. r (GU ). It then follows that 1 − a is not Fredholm, however 1 − πx(a) = 1 The first part shows that there is an isomorphism C∗ r (GF ) ≃ C∗ r (G)/C∗ r (GU ) ≃ C∗ r (G)/K(L2(U )). By Fredholmness of G, we have that R(GF ) defines a strictly spectral family of representations of C∗ r (G)/K, so R(GF ) is also a strictly spectral family of represen- tations of C∗ r (GF ). Since G is second countable, it is exhaustive, by Theorem 3.3, so that GF has Exel's property. This proves (ii) and hence the direct implication. (cid:3) We remark that we proved in fact that, if R(GF ) defines a strictly spectral family r (GF ) r (GU ), then necessarily C∗ r (GU ) ≃ C∗ r (G)/C∗ r (G)/C∗ of representations of C∗ and GF has Exel's property. A stronger form of the converse of Theorem 4.5 is contained in the following theorem. Theorem 4.6. Let G ⇒ M be a Lie groupoid. Let us assume that U ⊂ M is a dense G-invariant open subset such that GU ≃ U × U . Also, let us assume that the two conditions of Theorem 4.5 are satisfied. Then, for any unital C∗-algebra Ψ containing Mn(C∗ r (G)) as an essential ideal and for any a ∈ Ψ, we have that π0(a) if Fredholm if, and only if, the image of a in Ψ/Mn(C∗ r (G)) is invertible and all πx(a), x ∈ M r U , are invertible. Proof. The proof is the same for any n (using also Proposition 3.6, so we let n = 1 for notational simplicity). Let us assume that conditions (i) and (ii) of Theorem 4.5 are satisfied and let Ψ be a C∗-algebra containing C∗ r (G) as an essential ideal. Property (i) implies that π0 is injective on Ψ since C∗(G) is an essential ideal of Ψ. By our assumptions on G, we obtain that the algebra Ψ/C∗ r (GU ) ≃ π0(Ψ)/K =: B contains B0 := π0(C∗ r (G))/K as an ideal and B/B0 ≃ Ψ/C∗ r (G). Moreover, B0 ≃ C∗ r (GF ) by (ii). Let now a ∈ Ψ be arbitrary. By Atkinson's Theorem, we know that π0(a) is Fredholm if, and only if, its image in B is invertible. But by (iii) and by Corollary 3.5, c ∈ B is invertible if, and only if, the image of c in Ψ/C∗ r (G) and all πx(c) are invertible for all x ∈ F . (cid:3) Corollary 4.7. Let G ⇒ M be a Lie groupoid and ∅ 6= U ⊂ M be a dense, G- invariant, open subset such that GU ≃ U × U , and F = M \ U . Then G is Fredholm if, and only if, C∗ r (GF ) and R(GF ) is a strictly spectral set of representations of GF (that is, if GF has Exel's property). r (GU ) ≃ C∗ r (G)/C∗ Recalling the exact sequence in Proposition 2.6, it follows that, in particular, if G is metrically amenable, then G is Fredholdm if, and only if, GF has Exel's property. Remark 4.8. It is immediate to observe that, if the set R(GF ) of regular represen- tations πx, x ∈ F , forms a strictly spectral set of representations of C∗(GF ), then GF has Exel's strong property, and conditions (ii) and (iii) of the Theorem 4.5 are satisfied (using Proposition 2.6). This remark and the above results then give the following sufficient conditions for a Lie groupoid to be Fredholm. Corollary 4.9. Let G ⇒ M be a Lie groupoid and U ⊂ M be a dense G-invariant open subset such that GU ≃ U × U , and F = M \ U . Assume that R(GF ) forms a 20 C. CARVALHO, V. NISTOR, AND Y. QIAO strictly spectral set of representations of C∗(GF ) or, equivalently, that GF has Exel's strong property. Then G is a Fredholm groupoid. See also [7]. Together with Theorem 3.13, we obtain an important class of Fredholm groupoids. Recall that we assume that all our groupoids are Hausdorff. Theorem 4.10. Let G ⇒ M be a stratified submersion Lie groupoid with filtration Ui such that U0 is dense and GU0 ≃ U0 × U0. Assume that all isotropy groups Gx x are amenable. Then G is a Fredholm groupoid. Proof. By Remark 3.12 applied to F := M r U0 and by Theorem 3.13, we know that the family R(GF ) of regular representations of GF is an exhaustive family of representations of C∗(GF ), that is, GF has Exel's strong property. The result follows from the previous corollary. (cid:3) We notice that we can replace πx, x ∈ F , with πb, b ∈ BS, for the strata corresponding to the boundary. We can further restrict to the set of bs in a dense set of orbits, since they give rise to a faithful family of representations of GF , again by [40]. 4.2. Pseudodifferential operators. So far we have discussed mostly groupoids and operators in their (unitized) C∗-algebras. The study of differential operators on suitable non-compact spaces can be reduced to the study of groupoid algebras using pseudodifferential operators on Lie groupoids. In this subsection, we will explain this reduction procedure. See [38, 68, 87, 95, 105] for introductions to pseudodifferential operators. Let G be a Lie groupoid and consider the algebra Ψ∗(G) of pseudodifferential operators on the groupoid G, whose definition we now briefly recall [4, 61, 73]. Then, for m ∈ Z ∪ {±∞}, Ψm(G) consists of smooth families (Px)x∈M of classical pseudo-differential operators Px ∈ Ψm(Gx) of order m, that are right invariant with respect to the action of G and have compactly supported distribution kernels. In particular, Ψ−∞(G) is nothing but the convolution algebra of smooth, compactly supported function on G, that is, Ψ−∞(G) ≃ C∞ c (G). This construction extends right away to operators between smooth sections of some vector bundles E, F → M . There are two methods of doing this. The first method is to consider families (Px)x∈M of operators Px ∈ Ψm(Gx; r∗(E), r∗(F )). We denote by Ψm(G; E, F ) the resulting set of operators. By embedding E ⊕ F into a trivial bundle of dimension N , we can identify Ψm(G; E, F ) with a subspace of MN (Ψm(G)), when convenient. Similar considerations yield differential opera- tors generated by V acting between vector bundles E and F endowed with metric preserving connections. The second method to deal with operators between two vector bundles E, F → M is as follows. Let us choose an embedding of E ⊕ F → CN into a trivial vector bundle and denote by e and f the corresponding orthogonal projections onto the images of e and f . Then we can identify Ψm(G; E, F ) := f MN (Ψm(G))e . This idea applies to other types of similar operators (norm closures of operators in Ψm(G), differential operators, ... ). FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 21 One of the main points of considering pseudodifferential operators on groupoid is that, if we denote V := Lie(G) := (Γ(A(G))), then Diff m(V; E, F ) = Ψm(G; E, F ) ∩ Diff(M ; E, F ) , (2) where Diff m(V; E, F ) is generated by C∞(M ; End(E ⊕ F )) and V (using compatible connections on E and F ) and Diff(M ; E, F ) denotes all differential operators on M acting between sections of E and F . See also 4.14. Another main point for introducing the algebra Ψ∗(G) is related to parametri- ces and inverses of differential operators. To discuss this, let us denote by Ψ(G) the C∗-algebra obtained as the closure of Ψ∗(G) with respect to all contractive ∗- representations of Ψ−∞(G), as in [45]. As we will see below, this definition extends to operators acting between vector bundles. Let us denote, as usual, by S∗A the set of unit vectors in the Lie algebroid A∗(G) associated with G (as in Remark 2.11), with respect to some fixed metric on A∗(G). Then Ψ(G) fits into the following exact sequence (3) 0 → C∗(G) → Ψ(G) σ0 −→ C0(S∗A) → 0 . (See for instance [44] and the references therein.) Moreover, if P ∈ Ψm(G; E, F ), m ≥ 0, is elliptic and invertible in L2, then its inverse will be in Ψ(G; F, E). Typ- ically, Fredholm conditions are obtained for Fredholm Lie groupoids by applying our results to the algebra Ψ = MN (Ψ(G)), for some N . Let us see how this is done. We first need to recall the definition of Sobolev spaces. Let us fix in this subsec- tion a Lie groupoid G ⇒ M . For the purpose of the next result, let us assume that its space of units M has an open, dense, G-invariant subset U ⊂ M0 := M r ∂M such that the restriction GU is isomorphic to the pair groupoid U × U . Let us also assume that the space of units M is compact. This is needed in order to construct canonical Sobolev spaces on the interior of M . Indeed, there is an essentially unique class of metrics on A(G), which, by restriction, gives rise to a class of metrics on U , that are called compatible (with the groupoid G), see also [3]. All these metrics are Lipschitz equivalent and complete [2, 45]. In fact, the Sobolev spaces of all these metrics will coincide. They are given as the domains of the powers of 1 + ∆, where ∆ is the (geometer' s, i.e. positive) Laplacian. We shall denote by H s(U ) = H s(M ) these Sobolev spaces. The Sobolev spaces H s(M ) are discussed in detail in [2]. See also [71] for a review. The operators in Ψm(G) (and their vector bundle analogues) model differential operators associated to (any) compatible metric (see also Re- mark 4.14 below). By considering a vector bundle E → M and ∆E the associated Laplacian on sections of E, we obtain the spaces H s(M ; E). Note that the Sobolev spaces H s(M ) are not defined using the compact manifold structure on M , but rather using the complete metric on U ⊂ M . Recall that if the set M \ U is as in Definition 4.3, then the set U is called a manifold with amenable ends. If P ∈ Ψm(G; E, F ), then P : H s(M ; E) → H s−m(M ; F ) is continuous. Let Ψm(G; E, F ) denote the pseudodifferential operators acting on the lifts of E and F to G via d. We let Ψm(G; E) := Ψm(G; E, E). Then if P : H s(M ; E) → H s−m(M ; F ) is an order m pseudodifferential operator, we replace the study of its Fredholm properties with those of (4) P1 :=   0 P ∗ 0 P 0 0 0 0 Qm   , 22 C. CARVALHO, V. NISTOR, AND Y. QIAO where Qm is an invertible order m pseudodifferential operator acting on the com- plement of E ⊕ F . Let G denote the complement of E ⊕ F in CN . We fix the smooth vector bundles E, F, G → M in what follows. We denote by ∆E, ∆F , and ∆G the associated Laplacians acting on sections of E, respectively, F and G. For instance, Qm ∈ Ψm(G; G) could be (1 + ∆G)m/2 (or a suitable approximation it, see [45]). One of the main drawbacks of the algebras Ψm(G) is that they are too small to contain resolvents. This issue is easily fixed by considering completions with respect to suitable norms. Let us consider the norm k · km,s defined by. (5) kP km,s := k(1 + ∆F )(s−m)/2P (1 + ∆E)−s/2kL2→L2 We then let (6) Lm s (G; E, F ) := k · km,s-closure of Ψm(G; E, F ) . Recall that H s(M ; E) is the domain of (1 + ∆E)s/2, if s ≥ 0, whenever M is compact (see [2, 33, 45]. Hence Lm s (G; E, F ) is the norm closure of Ψm(G; E, F ) in the topology of continuous operators H s(M ; E) → H s−m(M ; F ). We note that, if P ∈ Ψm(G; E, F ), then (7) (1 + ∆F )(s−m)/2P (1 + ∆E)−s/2 ∈ Ψ(G; E, F ) =: L0 0(G; E, F ) by [45] (see also [5]). Moreover, let (8) W m(G) := Ψm(G) + ∩sL−∞ s (G) . Then W m(G) ⊂ Lm that contains the inverses of its L2-invertible operators. s (G) and W ∞(G) is an algebra of pseudodifferential operators Recall that we denote by π0 : C∗(G) → L(L2(M )) the vector representation and that M is compact. It is unitarily equivalent to the regular representations πx, x ∈ U . Moreover, since G is Hausdorff, π0 is injective. We have the following result from [45]. Proposition 4.11. Let G ⇒ M be a Lie groupoid with GU = U × U and M 0(G)) contains MN (C∗(G)) as an essential compact. Then MN (Ψ(G)) = MN (L0 ideal. Let s ∈ R and P ∈ Lm s (G; E, F ) ⊃ Ψm(G; E, F ). Then a := (1 + ∆F )(s−m)/2P (1 + ∆E)−s/2 ∈ Ψ(G; E, F ), . We have that P : H s(M ; E) → H s−m(M ; F ) is Fredholm if, and only if, a : L2(M ; E) → L2(M ; F ) is Fredholm. This proposition applies, in particular, if G is a Fredholm Lie groupoid with M compact. Proof. We can assume N = 1. The fact that Ψ(G) contains C∗(G) as an essential ideal is a general fact–true for any Lie groupoid. This general fact is true because it is true for any non-compact manifold, in particular, for each of the manifolds Gx. We have, by the definitions of Sobolev spaces and of Fredholm operators, that P is Fredholm if, and only if, a is Fredholm. The fact that a := (1 + ∆F )(s−m)/2P (1 + ∆E)−s/2 ∈ Ψ(G; E, F ) follows from the definition of Lm s (G; E, F ), Equation (6) and the results in [45], as noticed already above. (cid:3) FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 23 This proposition then gives right away the following result (for technical reasons, we may have to replace P with a in certain proofs). Recall that an operator P ∈ Ψm(G; E, F ) consists of a right invariant family P = (Px), x ∈ M , with Px acting between the sections of r∗(E) and r∗(F ) on Gx. For simplicity, we shall drop r∗ from the notation below, since there is no danger of confusion. We have Px = πx(P ). The operators Px, x ∈ M \ U , will be called limit operators (of P ). They go back to [28]. Theorem 4.12. Let G ⇒ M be a Fredholm Lie groupoid with M compact and ∅ 6= U ⊂ M , open such that GU = U × U . Let s ∈ R and P ∈ Lm s (G; E, F ) ⊃ Ψm(G; E, F ). We have P : H s(M ; E) → H s−m(M ; F ) is Fredholm ⇔ P is elliptic and each Px : H s(Gx; E) → H s−m(Gx; F ) , x ∈ M r U , is invertible . Proof. Let us use the notation of Proposition 4.11. By replacing P with a and using Proposition 4.4, we may assume that we work with N × N matrices of operators. The proof is the same for all N , so we let N = 1. Since G is Fredholm, Theorem 4.6, applied to Ψ := Ψ(G), gives that a ∈ Ψ(G) is Fredholm if, and only if, its image in Ψ(G)/C∗(G) is invertible and all the operators πx(a) are invertible. We then notice that πx(a) = (1 + ∆x)(s−m)/2Px(1 + ∆x)−s/2 since the extension of πx to operators affiliated to Ψ(G) is given by πx(P ) = Px since this is true for P ∈ Ψ0(G) and πx(∆) = ∆x, the Laplacian on Gx by [45]. (cid:3) Remark 4.13. To obtain the result as formulated in Theorem 1.1, we let α be the set of orbits of G acting on F := M r U , M0 := U , Mα := Gx, for some x in the orbit α, Gα := Gx x , and Pα := πx(P ). It may be useful to notice here that often the bundle Mα → Mα/Gα is trivial. This is the case, for example, for stratified submersion groupoids. Remark 4.14. Let V := r∗(Γ(A(G)) be the vector fields on M coming from the infinitesimal action of G ⇒ M on M . We denote by Diff m(V) the set of order m differential operators on M generated by V and multiplication with functions in C∞(M ) and Diff(V) = ∪m Diff m(V). We denote by Diff m(V; E, F ) the analogous differential operators acting between sections of vector bundles E, F → M . We have that all geometric operators (Laplace, Dirac, Hodge, ... ) associated to a compatible metric on M (one that comes by restriction from A(G)) belong to Diff m(V; E, F ) for suitable E, F → M ; moreover, πx(D) for a geometric operator is of the same type as D. [3]. Remark 4.15. If G is d-connected (in the sense that the fibers of d : G → M are connected), then the orbits of the vector fields V are the same as the orbits of G. If x ∈ M and α ⊂ M is its orbit, then this orbit (with its intrinsic manifold topology) is diffeomorphic to the set Mα/Gα. This makes more explicit the data in Remark 4.13. Although we shall not use this in the present paper, let us record the consequence for essential spectra. Notice that we do not need the closure of the unions for the essential spectra. We denote by σ(Q) the spectrum of an operator Q, defined as the set of λ ∈ C such that Q − λ is not invertible, and by σess(Q) its essential spectrum, defined as the set of λ ∈ C such that Q − λ is not Fredholm. Here Q may be unbounded. If P ∈ Ψm(G; E), m > 0, we consider it to be defined on H m. 24 C. CARVALHO, V. NISTOR, AND Y. QIAO Corollary 4.16. Consider the framework of Theorem 4.12 and P ∈ Ψm(G; E), m ≥ 0, elliptic. We have σess(P ) = ∪x∈∂M σ(Px) ∪ ∪ξ∈S ∗Aσ(σ0(P )(ξ)) , if m = 0 , σess(P ) = ∪x∈∂M σ(Px) , if m > 0 . Thus, in the scalar, order zero case, we have (9) σess(P ) = ∪x∈∂M σ(Px) ∪ Im(σ0(P )) , Combining Theorems 4.12 and 4.10, we obtain the following result (recall that all our groupoids are Hausdorff and second countable). Theorem 4.17. Let M be compact and G ⇒ M be a stratified submersion Lie groupoid with filtration Ui such that U0 is dense in M and GU0 ≃ U0 × U0. Assume that all isotropy groups Gx x are amenable. Then, for any smooth vector bundles s (G; E, F ) ⊃ Ψm(G; E, F ), we have E, F → M and any P ∈ Lm P : H s(M ; E) → H s−m(M ; F ) is Fredholm ⇔ P is elliptic and all Px : H s(Gx; E) → H s−m(Gx; F ) , x ∈ M r U , are invertible . We actually have Px : H s(Gx; r∗(E)) → H s−m(Gx; r∗(F )), but, as we have ex- plained above, we drop r∗ from the notation, for the sake of simplicity, since there is no danger of confusion. We will continue to do that in all related statements below. Remark 4.18. If G be as in the above theorem (a stratified submersion groupoid with filtration Ui), and assume that it is also d-connected, as in Remark 4.13. Then, when P = D is a differential operator, we can identify the limit operators Dα (and hence also Mα and Gα) in a direct way. Indeed, let α be the V-orbit = the G-orbit through some x ∈ M r U0. We have Gα = Gx x , as discussed in Remark 4.13. Then Mα = α×Gα. Let D ∈ Diff(V). By linearity, we may assume D = aXi1Xi2 . . . Xim, with Xi ∈ V a local basis of V near x and a ∈ C∞(M ). (We may also assume that ij form a non decreasing sequence, but that is not necessary.) Since α is an orbit, we have that this basis (Xj) consists of vectors tangent to α. We can choose this basis such that X1, . . . , Xk form a local basis of Txα and Xj, j > k, are zero on the orbit α (where they are defined). Consequently, the vector fields Xj, j > k, come from a basis Xj, j > k, of Lie(Gα). We then have (10) near {x} × Gα ⊂ α × Gα =: Mα , Dα = aαY1Y2 . . . Ym , where Yj = Xij α, if ij ≤ k, and Yj = eXij if ij > k. We shall say that the variables Yj (or Xj) with j > k are ghost derivatives at the considered orbit α (see Remark 4.18). See also [9, 8, 97, 103]. It would be interesting to extend the results of this section to Lp-spaces in view of [1, 78]. 5. Examples In this section, we use the results of the previous section to obtain Fredholm conditions for operators on some standard non-compact manifolds: manifolds with (poly)cylindrical ends, asymptotically Euclidean manifolds, asymptotically hyper- bolic manifolds, boundary fibration structures, and others. We tried to write this section in such a way that it can to a large extent be read independently of the FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 25 rest of the paper. The reader interested only in applications can start reading the paper with this section. The general setting of this section is that of a (non-compact) smooth manifold M0 whose geometry is determined by a compactification M to a manifold with cor- ners and a Lie algebra of vector fields V on M . (Although we shall not use this, let us mention for people familiar with the concept, that (M, V) will be a Lie manifold with some additional properties.) The differential (and pseudodifferential) opera- tors considered and for which we obtain Fredholm conditions are the ones generated by V and C∞(M ), that is, the ones in Diff(V) and its variants for vector bundles. (Recall that Diff(V) was introduced in Remark 4.14.) The proof of the Fredholm conditions in this section are obtained as particular cases of Theorem 4.12 by show- ing that appropriate groupoids are Hausdorff stratified submersion Lie groupoids with amenable isotropy groups and hence that they are Fredholm groupoids, in view of Theorem 3.13. Many of the results below can also be obtained from the results in [44, 45], but the approach followed here aims to be more convenient for non-specialists. In this section, we continue to denote by G a Lie groupoid with units M , a In the theorem yielding Fredholm conditions, M will manifold of dimension n. be assumed compact and endowed with a smooth metric h. We assume that h is defined everywhere on M , in particular, that it extends to a smooth manifold containing M as a submanifold. Also, usually it will be no loss of generality to assume M connected. Our results are formulated for operators in Lm s (G; E, F ), which is a suitable completion of Ψm(G; E, F ), see Equations (5) and (6). This allows us to greatly enlarge the scope of our results since the completion procedure defining the spaces Lm s leads to algebras that are closed under taking the inverses of L2-invertible elements. 5.1. Examples related to group actions. We include first some examples that are closely related to group actions. 5.1.1. The action of a group on a space. Let us assume that Lie group G acts smoothly on a manifold M . This yields the transformation (or group action) Lie groupoid G := M ⋊ G, which, as a set, consists of M × G and has units M a. We have d(x, g) := x , r(x, g) := gx , and (hx, g)(x, h) := (x, gh) . If M = G with G acting by translations, then M ⋊G ≃ G×G, the product groupoid. Therefore, if G ⊂ M as a dense, G-invariant open set, and G acts on itself by left translation, we are in the setting in which we can ask if the resulting groupoid is Fredholm Lie groupoid. This groupoid was used in [30, 51, 63]. The groupoids used in [30, 63] turn out to be stratified submersion groupoids. As discussed in those papers, this recovers the classical HVZ-theorem [20, 80, 99] as a particular case of Theorem 4.12 (with U = G). We note that M × G is always Hausdorff (since M and G are Hausdorff). If V ⊂ M is an open subset, then the reduction groupoid (M ⋊ G)V V will be called a local transformation (or action) groupoid, and will also be Hausdorff. Many related results (including the non Lie case), were also obtained by Bottcher, Chandler-Wilde, Karlovich, Lindner, Rabinovich, Roch, Rozenblyum, Silberman, and many others. See [6, 10, 11, 15, 55, 77, 78, 79, 85] and the references therein. 26 C. CARVALHO, V. NISTOR, AND Y. QIAO This groupoid models pseudodifferential operators compatible with any G in- variant metric on G. Let g be the Lie algebra of G. Then A(G) = M × g, with g acting on M via the infinitesimal action of G. The associated Lie algebra of vector fields is Lie(G) := (Γ(A(G))) = V := C∞(M )g ⊂ Γ(T Ω). We have that Diff(V) is generated by C∞(M ) and g. 5.1.2. The b-groupoid. Let M be a manifold with corners. Then the b-groupoid of [45, 57, 61, 73] is defined as a set as the disjoint union (11) Gb := ⊔F (F × F ) × (R∗ +)kF ≃ ⊔F (F × F ) × RkF ⇒ ⊔F F = M , where F ranges through the open, connected faces of M , kF is the codimension of the face F , F × F is the pair groupoid and (R∗ +)kF is a group for componentwise multiplication (and hence also a Lie groupoid). To obtain the smooth structure on this groupoid, we notice that, locally, it is +)k. More precisely, let us choose g ∈ Gb. a transformation groupoid [0, ∞)k ⋊ (R∗ Then g = (x, y, v) ∈ (F × F ) × RkF for some connected open face F ⊂ M . We can choose a coordinate system V ⊂ M such that x, y ∈ V , V is compact in F , and we have a tubular neighborhood V × [0, ǫ)kF ⊂ M . Then, a neighborhood of g in Gb is diffeomorphic to the local transformation groupoid obtained by reducing to V × [0, ǫ)kF the action groupoid (RdF × [0, ∞)kF ) ⋊ (RdF × (0, ∞)kF ) , where dF = n − kF is the dimension of F . Here RdF acts by translations on itself and R∗ + = (0, ∞) acts by multiplication on [0, ∞). In applications, it will be, in fact, more convenient to notice that RdF ⋊ RdF is the pair groupoid, and hence to identify a neighborhood of g with a reduction of (V × V ) × [0, ∞)kF ⋊ (0, ∞)kF , the product of the pair groupoid V × V and the transformation groupoid [0, ∞)kF ⋊ (0, ∞)kF . If M has embedded faces, that is, if each hyperface H has a defining function rH , then we can identify Gb with an open subset of Monthubert's realization of the b-groupoid [62, Proposition 4.5] M := { (x, y, t) ∈ M × M × [−1, 1]H (1 − tK)xK(x) = (1 + tH )xH (y) } , where H denotes the set of hyperfaces of M and H, K ∈ H. See also [47, 48, 86, 13, 101]. We see that the b-groupoid is a stratified submersion gropoid as follows. The set Uk, k ≤ n, is defined as the union of the open faces of codimension k of M . Then (Gb)Uk\Uk−1 is isomorphic to the topological disjoint union of the groupoids (F × F ) × Rk, where F ranges through the set of open faces of codimension k. In particular, (Gb)Uk\Uk−1 is isomorphic to the fibered pull-back of a bundle of Lie ∼= Rn−k is amenable, for all x ∈ M . groups, by Example 2.20. In particular, (Gb)x x 5.1.3. Manifolds with poly-cylindrical ends. The Lie algebroid of Gb is identified by (12) Γ(A(Gb)) ≃ Vb := {X ∈ Γ(M ; T M ) X tangent to all faces of M } . FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 27 Let h be an ordinary metric on M . The general form of a compatible metric on M is then (13) gb := h + XH∈H xH (cid:19)2 (cid:18) dxH . Manifolds with metrics of this form will be called manifolds with poly-cylindrical ends, following [59]. The Lie algebroid A(Gb) is often denoted T bM . The groupoid Gb models pseudodifferential operators compatible with the metric gb. Thus, by definition, the metric gb comes from restriction from the Lie algebroid A(Gb), since A(Gb)M0 = T M0. Since the base M is compact, all metrics on A(Gb) will be equivalent, so the constructions will not depend (essentially) on the choice of the metric. In particular, all geometric operators associated to the metric gb (Laplace, Dirac, Hodge, ... ) will belong to Ψm(Gb; E, F ) (which is independent of the metric), for suitable E and F . Moreover, it turns out that Theorem 4.17 applies to Gb and Ψm(Gb; E, F ), provided that M is compact. (See also Remark 4.14. Further details can be found in [3].) The statement of Theorem 4.5 can be (slightly) simplified in this case by noticing the following. The representations πx and πy are unitarily equivalent if x and y are in the same open face F , in which case, they will act on F × RkF ≃ Gx. Let πF be the associated representation. For Fredholm conditions, it is enough to consider the invertibility of the operators PH := πx(P ), x ∈ H, for the faces of maximal dimension (that is, for hyperfaces) in order to obtain Fredholm conditions, since πx is weakly contained in πy if x is contained in the closure of the face containing y. Let H denote the set of hyperfaces of M , as before. Recall that Lm s (G; E, F ) is a suitable completion of Ψm(G; E, F ), see Equations (5) and (6). We obtain the following result [58]. Corollary 5.1. Let P ∈ Lm s (Gb; E, F ) ⊃ Ψm(Gb; E, F ), M compact. We have P : H s(M ; E) → H s−m(M ; F ) is Fredholm ⇔ P is elliptic and all PH : H s(H × R; E) → H s−m(H × R; F ) , H ∈ H , are invertible . All geometric differential operators associated to the metric gb belong to Diff m(Vb; E, F ) ⊂ Ψm(Gb; E, F ), for suitable vector bundles E, F → M . If P ∈ Diff m(Vb; E, F ), the limit operators Px, x ∈ ∂M , are obtained as ex- plained in Remark 4.18. They are invariant with respect to the action of the isotropy group Gx x (this is always the case for operators of the form πx(a)). In this example, only derivatives of the form xH ∂xH are ghost derivatives (see Remark 4.18). Remark 5.2. Carvalho and Qiao have constructed in [14] a similar groupoid to the b-groupoid in order to study layer potentials. Their groupoid, however, was not d- connected, in general. Nevertheless, the above corollary generalizes to their setting, after some obvious modifications. 5.1.4. Asymptotically Euclidean spaces. Let us assume that M has a smooth bound- ary ∂M with defining function x = x∂M and let Vsc := rVb. The resulting differ- ential operators Diff(Vsc) and the associated pseudodifferential ooperators are the SG-operators of [19, 74, 88, 90, 89] (called "scattering operators" in [59]). They 28 C. CARVALHO, V. NISTOR, AND Y. QIAO can be obtained by considering the groupoid (14) Gsc := T M ∂M ⊔ (M0 × M0) ⇒ ∂M ⊔ M0 = M . To obtain a manifold structure on Gsc, let us consider first G = Rn and M = the radial compactification of G with the induced action of G. Then Gsc = M ⋊ G is Hausdorff and a stratified submersion groupoid. In general, Gsc is locally of this form (and can be obtained by glueing reductions of such groupoids), and hence it is Hausdorff. It satisfies Lie(Gsc) := Γ(A(Gsc)) = Vsc. Thus, if gb denotes a b-metric on M (or rather, on M0, then the natural metric associated to Vsc is (15) where gb is as in Equation (13). gsc := x−2gb := x−2(cid:0)h + x−2dx2(cid:1) , We have that the orbits of G on F := M r U0 = ∂M are reduced to points and that each stabilizer Gx x , for x ∈ ∂M and all derivatives at the boundary are ghost derivatives. If P ∈ Diff m(Vsc; E, F ), the limit operators Px are obtained as explained in Remark 4.18. Theorem 4.17 becomes in our case: x = TxM ≃ Rn, for x ∈ F . In particular, Gx = Gx Corollary 5.3. Let P ∈ Lm s (Gsc; E, F ) ⊃ Ψm(Gsc; E, F ), M compact. We have P : H s(M ; E) → H s−m(M ; F ) is Fredholm ⇔ P is elliptic and all Px : H s(TxM ; E) → H s−m(TxM ; F ) , x ∈ ∂M , are invertible . All geometric differential operators associated to the metric gsc belong to Diff m(Vsc; E, F ) ⊂ Ψm(Gsc; E, F ), for suitable vector bundles E, F → M . We have Px = πx(P ) and Px is translation invariant (i.e. constant coefficient, in this case), and hence can be studied using the Fourier transform. In this case, all the vector fields yield ghost derivatives. 5.1.5. Asymptotically hyperbolic manifolds. We continue to assume that M is a manifold with smooth boundary ∂M . The groupoid Gah modeling asymptotically hyperbolic spaces is chosen such that Lie(Gah) ≃ Γ(A(Gah)) = V0 := {X ∈ Γ(M ; T M ) X∂M = 0} . Let Lx := Tx(∂M )⋊(0, ∞) be the semi-direct product as in the previous subsection, with (0, ∞) acting by dilations on Tx(∂M ). Let L → ∂M be the bundle of Lie groups with fiber Lx. Then Gah := L ⊔ M0 × M0 ⇒ ∂M ⊔ M0 . The topology is again locally given by a transformation groupoid. This can be seen in the case of M = Rn−1 × [0, ∞) with the natural action of Gn := Rn−1 ⋊ (0, ∞) obtained by recalling that, as smooth manifolds, we have Rn−1 ⋊ (0, ∞) = Rn−1 × (0, ∞). This groupoid is a particular case of the edge groupoid, following next, so the reader can consult to the next section for more details. We again have that Gx = Gx x = Lx, if x ∈ ∂M , and that all derivatives are ghost derivatives at the boundary. The metric is gah := x−2h , where x is the distance to the boundary (close to the boundary) and h is an every- where smooth metric on M , as before. FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 29 Corollary 5.4. Let s ∈ R and P ∈ Lm compact. Denote Mx := Tx∂M × R. We have s (Gah; E, F ) ⊃ Ψm(Gah; E, F ), with M P : H s(M ; E) → H s−m(M ; F ) is Fredholm ⇔ P is elliptic and all Px : H s(Mx; E) → H s−m(Mx; F ) , x ∈ ∂M , are invertible . All geometric differential operators associated to the metric gah belong to Diff m(Vah; E, F ) ⊂ Ψm(Gah; E, F ), for suitable vector bundles E, F → M . We thus need to study the invertibility of certain (right) invariant operators on the group Gn−1 := Rn−1 ⋊ (0, ∞), for which standard methods of representation theory can be used. 5.2. The edge groupoid. This example is motivated by the results in [32, 34, 54, 69, 92], where the original Fredholm results on the edge calculus can also be found. It is a particular case of the next example, that of a desingularization groupoid, but we nevertheless treat it separately, for the benefit of the reader. See also [22, 43, 104]. We consider the following framework. First, M is a manifold with smooth boundary ∂M . We assume that we are given a smooth fibration π : ∂M → B, where B is a smooth manifold (thus without boundary). We fix a tubular neighborhood U of ∂M in M : U ≃ ∂M × [0, 1). Let H := B × B be the pair groupoid. We now construct the so called edge groupoid Ge that will turn out to be a groupoid with Lie algebroid given by the set Ve of vector fields on M that are tangent to the fibers of π : ∂M → B (in particular, these vector fields are tangent to the boundary). If B = ∂M , then we recover the groupoid that models asymptotically hyperbolic spaces. We denote by Diff(Ve) the algebra of differential operators generated by Ve and by multiplication with functions in C∞(M ). Let L := T B ⋊ R∗ + → B be the bundle of Lie groups obtained from T B → B (regarded as a bundle of commutative Lie groups) by taking the semi-direct product with R∗ + := (0, ∞) acting by dilations on the fibers of T B → B. Its pullback π↓↓(L) via π : ∂M → B is hence a Lie groupoid with units ∂M . Let M0 := M r ∂M be the interior of M . Then, as a set, the edge groupoid Ge is the disjoint union (16) Ge := π↓↓(L) ⊔ (M0 × M0) ⇒ ∂M ⊔ M0 . To define the smooth structure on this groupoid, we could use either the results in [70] or proceed directly in four steps as follows. Step 1. We first consider the adiabatic groupoid Had of H := B × B (this is the tangent groupoid of [16]; see [69] for more details and references). The adiabatic groupoid Had is a Lie groupoid with units B × [0, ∞) and Lie algebroid A(Had) = T B × [0, ∞) → B × [0, ∞), which, as a vector bundle, is the fibered pull-back of A(H) = T B → B to B × [0, ∞) via the projection B × [0, ∞) → B. The Lie algebroid structure on the sections of A(Had) is not that of a fibered pull-back Lie algebroid, but is given by [X, Y ](t) = t[X(t), Y (t)]. As a set, Had is the disjoint union Had := A(H) × {0} ⊔ H × (0, ∞) . The groupoid structure of Had is such that A(H) × {0} has the Lie groupoid struc- ture of a bundle of Lie groups and H×(0, ∞) has the product Lie groupoid structure with (0, ∞) the groupoid associated to a space (that is (0, ∞) has only units, and 30 C. CARVALHO, V. NISTOR, AND Y. QIAO all orbits are reduced to a single point). The smooth structure is obtained using the exponential map. See also [102]. Step two. Let π : ∂M → B be the given fibration map. We denote also by π the resulting map ∂M × [0, ∞) → B × [0, ∞). Then we consider the pullback Lie groupoid π↓↓(Had). Step three. Let R∗ + = (0, ∞) act by dilations on the [0, ∞) variable on π↓↓(Had) + [25]. As a set, it is the disjoint and consider the semi-direct product π↓↓(Had) ⋊ R∗ union of π↓↓(L) and of the pair groupoid of ∂M × (0, ∞). Step four. Let us identify the tubular neighborhood U ⊂ M of ∂M with ∂M × [0, 1). Then we can consider the reduction H′ U . By the previous step, this reduction H′ is the disjoint union of π↓↓(L) and of the pair groupoid of ∂M ×(0, 1), which we can view as a subset of M0×M0, the pair groupoid U with M0 × M0 by identifying the reduction of H′ to U r∂M (which is the pair groupoid of ∂M ×(0, 1), as we have seen), with its image in M0 ×M0. This glueing construction is, of course, nothing but a particular case of the glueing construction in [35] and [69]. +(cid:1)U of M0. We then glue the reduction H′ := (cid:0)π↓↓(Had) ⋊ R∗ +(cid:1)U := (cid:0)π↓↓(Had) ⋊ R∗ In any case, we obtain right away from the definition that the edge groupoid Ge is a Hausdorff stratified submersion groupoid. The set of units of Ge is M and the representations πx, x ∈ ∂M , are equivalent precisely when they map to the same point in B and they act on π−1(b) × Lb ≃ π−1(b) × TbB × R. If P ∈ Diff(Ve) and b ∈ B, we can obtain the limit operators Pb := πb(P ) as follows. The restriction of P to an infinitesimal neighborhood of π−1(b) in M will have some hidden (ghost) derivatives coming from the Lie algebra of the group Lb := TbB ⋊ R∞ + . Let us choose local coordinates (y, z) on ∂M that are compatible with ∂M → B, in the sense that z comes from a coordinate system on B. Let x denote the defining function of ∂M . Then locally, Ve is generated by x∂x, ∂yj , x∂zk , with j and k (always) ranging through a suitable index set: (17) Ve = C∞(M )x∂x +Xj C∞(M )∂yj +Xj C∞(M )x∂zj . With this notation, the Lie algebra Lie(Lb) is generated by the vector fields x∂x and x∂zk . These come from non-zero vector fields in Ve that restrict to 0 on π−1(b), they are the ghost derivatives. We let the ghost derivatives act on Lb, and thus we obtain a differential operator Pb on π−1(b) × Lb. The metric on M is then an edge metric in the sense of Mazzeo [54]. It can be obtained by patching together metrics for which x∂x, ∂yj , x∂zk are an orthonormal set of vectors. Theorem 4.12 then becomes. Corollary 5.5. Let P ∈ Lm s (Ge; E, F ) ⊃ Ψm(Ge; E, F ), M compact. We have P : H s(M ; E) → H s−m(M ; F ) is Fredholm ⇔ P is elliptic and all Pb : H s(Mb; E) → H s−m(Mb; F ) , b ∈ B , are invertible , where Mb := π−1(b) × TbB × R. All geometric differential operators associated to ge belong to Diff m(Ve; E, F ) ⊂ Ψm(Ge; E, F ), for suitable E, F → M . If B is reduced to a point and ∂M is connected, the groupoid Ge constructed in the last subsection recovers the groupoid of the b-calculus: Ge = Gb. It models in FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 31 this case manifolds with cylindrical ends. If B = ∂M , the corresponding groupoid models "asymptotically hyperbolic" spaces. 5.3. Desingularization groupoids. One of the nice features of the class of strat- ified submersion groupoids is that it is invariant with respect to desingularization along suitable submanifolds. In this subsection, we will give an ad hoc argument for this statement. Recall the fibered pull-back π↓↓(B) of Example 2.19. We proceed in a slightly greater generality than in [69], to which we refer for more details and for the unexplained arguments. Let us assume that M is a manifold with corners, that H is a hyperface of M and that we are given a tame submersion of manifolds with corners π : H → B. Let U ≃ H × [0, 1) be a tubular neighborhood of H in M . The hyperface H will play the role played by the boundary in the previous examples. Typically, M will be the result of a desingularization procedure, such as a blow-up (see Remark 5.6 and the following section). We assume that we are given a Lie groupoid G on M r H and a Lie groupoid H → B such that the following is satisfied. Local fibered pull-back structure assumption: The reduction of G to H×(0, 1) is isomorphic to the pullback p↓↓(H) via the map p := π ◦ p1 : H × (0, 1) → B. We fix H as in the above assumption throughout this subsection. We define then a groupoid K ⇒ M such that, as a set, it is the disjoint union K := π↓↓(A(H) ⋊ R∗ +) ⊔ G ⇒ H ⊔ M r H. The groupoid H replaces the pair groupoid B × B of the example of the edge groupoid. We proceed as in that example to consider the adiabatic groupoid Had of H, which has units B×[0, ∞). We pull back this groupoid to a groupoid π↓↓(Had) using the map π1 : H × [0, ∞) → B × [0, ∞), and then we consider the semi-direct product π↓↓ + acting by dilations on [0, ∞). 1 (Had) ⋊ R∗ + = π↓↓ 1 (Had ⋊ R∗ +) with R∗ As in the previous example of the edge groupoid, we view U as an open subset of H × [0, ∞) and we consider the groupoid H′ defined as the reduction to U of π↓↓(Had) ⋊ R∗ +, with H as in the local fibered pull-back assumption. The construc- tion is completed as in the fourth step of the edge groupoid. That is, we use the invariant subset H × {0} of the units of H′ to write the groupoid as a disjoint union using Remark 2.7. The reduction of H′ to the complement of H × {0}, that is, to H × (0, 1), is (isomorphic to) the fibered pull-back p↓↓(H) of H to H × (0, 1), by the local fibered pull-back structure assumption. We can view this fibered pull-back p↓↓(H) as a subset of G. This gives that we can glue H′ and G along the common open groupoid p↓↓(H) to obtain a groupoid K. It can be proved that if G is Fredholm, then K is also Fredholm. Moreover, if G is a stratified submersion groupoid, then K will also be a stratified submersion groupoid. Remark 5.6. Typically, we start with a Lie groupoid G′ ⇒ M ′ and L ⊂ M ′ is a submanifold with corners. Then M := [M ′ : L], the blow-up of M ′ with respect to L and H is the hyperface corresponding to L in this blow-up, with π : H → L the blow-down map. Finally, G is the reduction of G′ to the complement of L. Then we denote [[G′ : L]] := K. This is the desingularization construction from [69]. 32 C. CARVALHO, V. NISTOR, AND Y. QIAO 5.4. Desingularization and singular spaces. Let us show how to use the desin- gularization construction in some typical examples. By Mk and Lk we will denote manifolds with corners of depths k. Thus M0 and L0 will have, in fact, no corners or boundary (hence they will be "smooth"). 5.4.1. The desingularization of a smooth submanifold. The simplest example of a desingularization groupoid is the desingularization of the pair groupoid G0 := M0 × M0 with respect to a smooth submanifold L0 ⊂ M0. Recall that M0 is also smooth. Thus neither M0 nor L0 have corners. The example of this subsection is a particular case of the edge groupoid, and hence it is related to the edge calculus. [34, 54, 66, 92]. Let N be the normal bundle of L0 in M0 and denote by S ⊂ N the set of unit vectors in N , that is, S is the unit sphere bundle of the normal bundle of L in M0. We let π : S → L0 be the natural projection. Then the blow-up M1 := [M0 : L0] of M0 with respect to L0 is the disjoint union M1 := [M0 : L0] := (M0 r L0) ⊔ S , with the structure of a manifold with smooth boundary S. We let G1 to be the associated edge groupoid introduced in Subsection 5.2. It is a Lie groupoid with base M1 := [M0 : L0]. We are moreover in the framework of Remark 5.6, so G1 = [[G0 : L0]]. The filtration of M1 has two sets, namely, U1 = M1 and U0 := M0 r L0 ⊂ M1, both of which are open and invariant for G1 (but U0 is not invariant for G0, in general). It turns out that G1 is a stratified submersion Lie groupoid if we consider the = U0 × U0 and GS is the fibered pull-back of a two strata U0 and S. Also, (G0)U0 U0 bundle of Lie groups L → S, as described in Subsection 5.2. 5.4.2. The desingularization of a submanifold with boundary. We now generalize he last construction to manifolds with boundary. Let M1 be a compact manifold with smooth boundary. We denote by F := ∂M1 its boundary and by G := M1 r F its interior. On M1 we consider the Lie algebra of vector fields Vb tangent to ∂M1, as before. It is the space of sections of T bM1, the "b-tangent bundle" [57, 59] of M1. Let G1 := Gb, as defined in the previous section. Let L1 ⊂ M1 be an embedded smooth submanifold assumed to be such that its boundary is ∂L1 = L1 ∩ ∂M1 and such that L1 intersects ∂M1 transversely. Then L1 has a tubular neighborhood U in M1, and hence we can consider the blow-up M2 := [M1 : L1], which will be a manifold with corners of codimension 2. Moreover, the reduction of G1 = Gb to U r L1 satisfies the local fibered pull-back structure assumption, since it is the b-groupoid of U r L1. In view of Remark 5.6, we can then define G2 := [[G1 : L1]]. Again it turns out that G2 is a stratified submersion Lie manifold; indeed, this is seen by choosing the following filtration of M2 with three sets: U0 := M1 r (L1 ∪ ∂M1) ⊂ U1 := M1 r L1 ⊂ U2 := M2 . The sets Uj are open and G2 invariant (but not G1 invariant). The sets U0 and U1 are G2-invariant by construction. Assume they are connected, for simplicity. The restriction (G2)U1 coincides with the reduction (G1)U1 by the definition of the desin- U1 gularization groupoid (Remark 5.6). In particular, (G2)U0 = (G2)U0 = U 2 0 U0 and = (G1)U0 U0 (G2)U1rU0 := (G1)U1rU0 U1rU0 = (U1 r U0)2 × R∗ + , FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 33 where U 2 0 and (U1 r U0)2 are pair groupoids. In the general case, if U1 r U0 is not connected, we write U1 r U0 = ⊔Vj as the disjoint union decomposition of its connected components, then we have (G2)U1rU0 := ⊔jV 2 +. Let us denote as in the boundaryless case by S the unit sphere bundle of the normal bundle to L1 in M1. (The set S is the last stratum U2 r U1 in M2.) Then the restriction of G1 to S := M2 r U1 is the following fibered fibered pull-back groupoid. Let π : S → L be the natural projection, as before. Let again GS := T bL ⋊ R∗ + → L be the group bundle over L obtained by taking the direct product of the b-tangent bundle to L with the action of R∗ + by dilation on the fibers of T L → L. Then j × R∗ GS := π↓↓(GS) . 5.4.3. Desingularization of a stratified subset of dimension one. We now deal with a slightly more complicated example by combining the two previous examples. We thus introduce the groupoid that is obtained from the desingularization of a stratified subset of dimension one. Full details as well as applications will be included in a forthcoming paper with Mihai Putinar. Let M0 be a smooth, compact manifold (so no corners). Let L0 := {P1, P2, . . . , Pk} ⊂ M0 and let us assume that we are given a subset S ⊂ M0 such that (18) S = L0 ∪ ∪l j=1γj, where each γj is the image of a smooth map cj : [0, 1] → M0, with the following properties: j(t) 6= 0, (i) c′ (ii) cj(0), cj(1) ∈ L0 := {P1, P2, . . . , Pk}, (iii) cj((0, 1)) are disjoint and do not intersect L0 and (iv) the vectors c′ j(0) and c′ j(1), j = 1, . . . , l, are all distinct. We now introduce the desingularization of G0 := M0 × M0 (the pair groupoid) with respect to S. The set L0 := {P1, P2, . . . , Pk} ⊂ M0 satisfies the assumptions of 5.4.1. We can first define G1 to be the desingularization of G0 with respect to L0 as in that example: (19) G1 := [[G0 : L0]] , which is a groupoid with units M1 := [M0 : L0]. The smooth maps cj then lift to smooth maps cj : [0, 1] → [M0 : L0] . The assumption that the vectors c′ j(1), j = 1, . . . , l, are all distinct then gives that the sets γj := cj([0, 1]) are all disjoint and intersect the boundary of M1 transversally. Let L1 be the disjoint union of the embedded curves γj. Then we can perform a further desingularization along L1, as in 5.4.2, thus obtaining j(0) and c′ G2 := [[G1 : L1]] , which is a boundary fibration Lie groupoid. The Lie groupoid G2 is the desingular- ization groupoid of M0 with respect to S. Its structure is given as in the previous subsection. This example can be extended to higher dimensional cases by using clean inter- secting families. Acknowledgements. We thank Vladimir Georgescu, Marius Mantoiu, and Wolfgang Schulze for useful discussions. 34 C. CARVALHO, V. NISTOR, AND Y. QIAO References [1] B. Ammann and N. Gross e. Lp-spectrum of the Dirac operator on products with hyperbolic spaces. Calc. Var. Partial Differential Equations, 55(5):127–163, 2016. [2] B. Ammann, Alexandru D. Ionescu, and V. Nistor. Sobolev spaces on Lie manifolds and regularity for polyhedral domains. Doc. Math., 11:161–206 (electronic), 2006. [3] B. Ammann, R. Lauter, and V. Nistor. On the geometry of Riemannian manifolds with a Lie structure at infinity. Int. J. Math. Math. Sci., (1-4):161–193, 2004. [4] B. Ammann, R. Lauter, and V. Nistor. Pseudodifferential operators on manifolds with a Lie structure at infinity. Ann. of Math. (2), 165(3):717–747, 2007. [5] B. Ammann, R. Lauter, V. Nistor, and A. Vasy. Complex powers and non-compact mani- folds. Comm. Partial Differential Equations, 29(5-6):671–705, 2004. [6] M.A. Bastos, C. Fernandes, and Yu.I. Karlovich. A C ∗-algebra of singular integral operators with shifts admitting distinct fixed points. J. Math. Anal. Appl., 413(1):502–524, 2014. [7] I. Belti¸ta, D. Belti¸ta, and M. M antoiu. Symbol calculus of square-integrable operator- valued maps. Rocky Mountain J. Math., 46(6):1795–1851, 2016. [8] K. Bohlen. Boutet de Monvel operators on Lie manifolds with boundary. Preprint arXiv:1507.01543. [9] K. Bohlen. Boutet de Monvel operators on singular manifolds. C. R. Math. Acad. Sci. Paris, 354(3):239–243, 2016. [10] A. Bottcher, Yu. I. Karlovich, and I. M. Spitkovsky. The C ∗-algebra of singular integral operators with semi-almost periodic coefficients. J. Funct. Anal., 204(2):445–484, 2003. [11] A. Bottcher and B. Silbermann. Analysis of Toeplitz operators. Springer Monographs in Mathematics. Springer-Verlag, Berlin, second edition, 2006. Prepared jointly with Alexei Karlovich. [12] M. Buneci. Groupoid C ∗-algebras. Surv. Math. Appl., 1:71–98 (electronic), 2006. [13] P. Carrillo-Rouse and J.-M. Lescure. Geometric obstructions for fredholm boundary condi- tions for manifolds with corners. ArXiv preprint, 2017. [14] C. Carvalho and Yu Qiao. Layer potentials C ∗-algebras of domains with conical points. Cent. Eur. J. Math., 11(1):27–54, 2013. [15] S. Chandler-Wilde and M. Lindner. Limit operators, collective compactness, and the spectral theory of infinite matrices. Mem. Amer. Math. Soc., 210(989):viii+111, 2011. [16] A. Connes. Non commutative differential geometry. Publ. Math. IHES, 62:41–144, 1985. [17] H. Cordes. Spectral theory of linear differential operators and comparison algebras, vol- ume 76 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 1987. [18] H. O. Cordes and E. A. Herman. Gel′fand theory of pseudo differential operators. Amer. J. Math., 90:681–717, 1968. [19] S. Coriasco and L. Maniccia. On the spectral asymptotics of operators on manifolds with ends. Abstr. Appl. Anal., pages Art. ID 909782, 21, 2013. [20] H. L. Cycon, R. G. Froese, W. Kirsch, and B. Simon. Schrodinger operators with application to quantum mechanics and global geometry. Texts and Monographs in Physics. Springer- Verlag, Berlin, study edition, 1987. [21] M. Damak and V. Georgescu. Self-adjoint operators affiliated to C ∗-algebras. Rev. Math. Phys., 16(2):257–280, 2004. [22] A. Dasgupta and M.W. Wong. Spectral theory of SG pseudo-differential operators on Lp(Rn). Studia Math., 187(2):185–197, 2008. [23] M. Dauge. Elliptic boundary value problems on corner domains, volume 1341 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1988. Smoothness and asymptotics of solu- tions. [24] C. Debord, J.-M. Lescure, and F. Rochon. Pseudodifferential operators on manifolds with fibred corners. preprint arXiv:1112.4575, to appear in Annales de l'Institut Fourier. [25] C. Debord and G. Skandalis. Adiabatic groupoid, crossed product by R∗ + and pseudodiffer- ential calculus. Adv. Math., 257:66–91, 2014. [26] J. Dixmier. Les C ∗-alg`ebres et leurs repr´esentations. Les Grands Classiques Gauthier- Villars. [Gauthier-Villars Great Classics]. ´Editions Jacques Gabay, Paris, 1996. Reprint of the second (1969) edition. FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 35 [27] R. Exel. Invertibility in groupoid C ∗-algebras. In Operator theory, operator algebras and applications, volume 242 of Oper. Theory Adv. Appl., pages 173–183. 2014. [28] J. Favard. Sur les ´equations diff´erentielles lin´eaires `a coefficients presque-p´eriodiques. Acta Math., 51(1):31–81, 1928. [29] V. Georgescu and A. Iftimovici. Crossed products of C ∗-algebras and spectral analysis of quantum Hamiltonians. Comm. Math. Phys., 228(3):519–560, 2002. [30] V. Georgescu and V. Nistor. On the essential spectrum of N -body hamiltonians with asymp- totically homogeneous interactions. http://arxiv.org/abs/1506.03267 [math.SP], to appear in J. Oper. Theory. [31] J. Gracia-Bond´ıa, J. V´arilly, and H. Figueroa. Elements of noncommutative geometry. Birkhauser Advanced Texts: Basler Lehrbucher. [Birkhauser Advanced Texts: Basel Text- books]. Birkhauser Boston, Inc., Boston, MA, 2001. [32] D. Grieser and E. Hunsicker. Pseudodifferential operator calculus for generalized Q-rank 1 locally symmetric spaces. I. J. Funct. Anal., 257(12):3748–3801, 2009. [33] N. Grosse and C. Schneider. Sobolev spaces on Riemannian manifolds with bounded geom- etry: general coordinates and traces. Math. Nachr., 286(16):1586–1613, 2013. [34] V. V. Grusin. A certain class of elliptic pseudodifferential operators that are degenerate on a submanifold. Mat. Sb. (N.S.), 84 (126):163–195, 1971. [35] M. Gualtieri and Songhao Li. Symplectic groupoids of log symplectic manifolds. Int. Math. Res. Not. IMRN, (11):3022–3074, 2014. [36] Ph. Higgins and K. Mackenzie. Algebraic constructions in the category of Lie algebroids. J. Algebra, 129(1):194–230, 1990. [37] Ph. Higgins and K. Mackenzie. Fibrations and quotients of differentiable groupoids. J. Lon- don Math. Soc. (2), 42(1):101–110, 1990. [38] L. Hormander. The analysis of linear partial differential operators. III. Classics in Mathe- matics. Springer, Berlin, 2007. Pseudo-differential operators, Reprint of the 1994 edition. [39] M. Ionescu and D. Williams. The generalized Effros-Hahn conjecture for groupoids. Indiana Univ. Math. J., 58(6):2489–2508, 2009. [40] M. Khoshkam and G. Skandalis. Regular representation of groupoid C ∗-algebras and appli- cations to inverse semigroups. J. Reine Angew. Math., 546:47–72, 2002. [41] V. A. Kondrat′ev. Boundary value problems for elliptic equations in domains with conical or angular points. Transl. Moscow Math. Soc., 16:227–313, 1967. [42] V. Kozlov, V. Maz′ya, and J. Rossmann. Spectral problems associated with corner singular- ities of solutions to elliptic equations, volume 85 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2001. [43] T. Krainer. A calculus of abstract edge pseudodifferential operators of type (ρ, δ). arXiv:math/1403.6100 [math.AP], 2014. [44] R. Lauter, B. Monthubert, and V. Nistor. Pseudodifferential analysis on continuous family groupoids. Doc. Math., 5:625–655 (electronic), 2000. [45] R. Lauter and V. Nistor. Analysis of geometric operators on open manifolds: a groupoid approach. In Quantization of singular symplectic quotients, volume 198 of Progr. Math., pages 181–229. Birkhauser, Basel, 2001. [46] M. Lein, M. Mantoiu, and S. Richard. Magnetic pseudodifferential operators with coefficients in C ∗-algebras. Publ. Res. Inst. Math. Sci., 46(4):755–788, 2010. [47] M. Lesch. Operators of Fuchs type, conical singularities, and asymptotic methods, volume 136 of Teubner-Texte zur Mathematik [Teubner Texts in Mathematics]. B. G. Teubner Ver- lagsgesellschaft mbH, Stuttgart, 1997. [48] M. Lesch and B. Vertman. Regularizing infinite sums of zeta-determinants. Math. Ann., 361(3-4):835–862, 2015. [49] K. Mackenzie. Lie groupoids and Lie algebroids in differential geometry, volume 124 of LMS Lect. Note Series. Cambridge U. Press, Cambridge, 1987. [50] K. Mackenzie. General theory of Lie groupoids and Lie algebroids, volume 213 of LMS Lect. Note Series. Cambridge U. Press, Cambridge, 2005. [51] M. Mantoiu. Essential spectrum and Fredholm properties for operators on locally compact groups. to appear in Journal of Operator Theory, preprint http://arxiv.org/abs/1510.05308 [math.SP]. [52] M. Mantoiu. C ∗-algebras, dynamical systems at infinity and the essential spectrum of gen- eralized Schrodinger operators. J. Reine Angew. Math., 550:211–229, 2002. 36 C. CARVALHO, V. NISTOR, AND Y. QIAO [53] M. Mantoiu, R. Purice, and S. Richard. Spectral and propagation results for magnetic Schrodinger operators; a C ∗-algebraic framework. J. Funct. Anal., 250(1):42–67, 2007. [54] R. Mazzeo. Elliptic theory of differential edge operators. I. Commun. Partial Differ. Equa- tions, 16(10):1615–1664, 1991. [55] M. Melgaard and G. Rozenblum. Spectral estimates for magnetic operators. Math. Scand., 79(2):237–254, 1996. [56] S. T. Melo, R. Nest, and E. Schrohe. C ∗-structure and K-theory of Boutet de Monvel's algebra. J. Reine Angew. Math., 561:145–175, 2003. [57] R. Melrose. The Atiyah-Patodi-Singer index theorem. Research Notes in Mathematics (Boston, Mass.). 4. Wellesley, MA: A. K. Peters, Ltd.. xiv, 377 p. , 1993. [58] R. Melrose and P. Piazza. Analytic K-theory on manifolds with corners. Adv. Math., 92(1):1– 26, 1992. [59] R. B. Melrose. Geometric scattering theory. Stanford Lectures. Cambridge University Press, Cambridge, 1995. [60] I. Moerdijk and J. Mrcun. Introduction to foliations and Lie groupoids, volume 91 of Cam- bridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2003. [61] B. Monthubert. Pseudodifferential calculus on manifolds with corners and groupoids. Proc. Amer. Math. Soc., 127(10):2871–2881, 1999. [62] B. Monthubert. Groupoids and pseudodifferential calculus on manifolds with corners. J. Funct. Anal., 199(1):243–286, 2003. [63] J. Mougel, V. Nistor, and N. Prudhon. A refined HVZ-theorem for asymptotically homoge- neous interactions and finitely many collision planes. Hal preprint 2017, to appear in Revue Romaine de Math´ematiques Pures et Appliqu´es. [64] P. Muhly, J. Renault, and D. Williams. Equivalence and isomorphism for groupoid C ∗- algebras. J. Operator Theory, 17(1):3–22, 1987. [65] P. S. Muhly, J. Renault, and D. Williams. Continuous-trace groupoid C ∗-algebras. III. Trans. Amer. Math. Soc., 348(9):3621–3641, 1996. [66] V. E. Nazaıkinskiı, A. Yu. Savin, B. Yu. Sternin, and B.-W. Shulze. On the index of elliptic operators on manifolds with edges. Mat. Sb., 196(9):23–58, 2005. [67] S. A. Nazarov and B. A. Plamenevsky. Elliptic problems in domains with piecewise smooth boundaries, volume 13 of de Gruyter Expositions in Mathematics. Walter de Gruyter & Co., Berlin, 1994. [68] F. Nicola and L. Rodino. Global pseudo-differential calculus on Euclidean spaces, volume 4 of Pseudo-Differential Operators. Theory and Applications. Birkhauser Verlag, Basel, 2010. [69] V. Nistor. Desingularization of Lie groupoids and pseudodifferential operators on singular spaces. http://arxiv.org/abs/1512.08613 [math.DG], to appear in Communications in Anal- ysis and Geometry. [70] V. Nistor. Groupoids and the integration of Lie algebroids. J. Math. Soc. Japan, 52(4):847– 868, 2000. [71] V. Nistor. Analysis on singular spaces: Lie manifolds and operator algebras. J. Geom. Phys., 105:75–101, 2016. [72] V. Nistor and N. Prudhon. Exhausting families of representations and spectra of pseudo- differential operators. preprint [math.OA], http://arxiv.org/abs/1411.7921, to appear in J. Oper. Theory. [73] V. Nistor, A. Weinstein, and Ping Xu. Pseudodifferential operators on differential groupoids. Pacific J. Math., 189(1):117–152, 1999. [74] C. Parenti. Operatori pseudodifferentiali in Rn e applicazioni. Annali Mat. Pura ed App., 93:391–406, 1972. [75] A. Paterson. Groupoids, inverse semigroups, and their operator algebras, volume 170 of Progress in Mathematics. Birkhauser Boston, Inc., Boston, MA, 1999. [76] B. A. Plamenevskiı. Algebras of pseudodifferential operators, volume 43 of Mathematics and its Applications (Soviet Series). Kluwer Academic Publishers Group, Dordrecht, 1989. Translated from the Russian by R. A. M. Hoksbergen. [77] V. Rabinovich and S. Roch. Essential spectrum and exponential decay estimates of solutions of elliptic systems of partial differential equations. Applications to Schrodinger and Dirac operators. Georgian Math. J., 15(2):333–351, 2008. [78] V. Rabinovich, S. Roch, and B. Silbermann. Limit operators and their applications in oper- ator theory, volume 150 of Operator Theory: Advances and Applications. Birkhauser, 2004. FREDHOLM CONDITIONS ON NON-COMPACT MANIFOLDS 37 [79] V. Rabinovich, B.-W. Schulze, and N. Tarkhanov. C ∗-algebras of singular integral operators in domains with oscillating conical singularities. Manuscripta Math., 108(1):69–90, 2002. [80] M. Reed and B. Simon. Methods of modern mathematical physics. IV. Analysis of operators. Academic Press [Harcourt Brace Jovanovich Publishers], New York, 1978. [81] J. Renault. A groupoid approach to C ∗-algebras, volume 793 of LNM. Springer, 1980. [82] J. Renault. The ideal structure of groupoid crossed product C ∗-algebras. J. Operator Theory, 25(1):3–36, 1991. With an appendix by G. Skandalis. [83] M. Rieffel. Induced representations of C ∗-algebras. Advances in Math., 13:176–257, 1974. [84] S. Roch. Algebras of approximation sequences: structure of fractal algebras. In Singular integral operators, factorization and applications, volume 142 of Oper. Theory Adv. Appl., pages 287–310. Birkhauser. [85] S. Roch, P. Santos, and B. Silbermann. Non-commutative Gelfand theories. Universitext. Springer-Verlag London, Ltd., London, 2011. [86] P. Carrillo Rouse, J. M. Lescure, and B. Monthubert. A cohomological formula for the Atiyah-Patodi-Singer index on manifolds with boundary. J. Topol. Anal., 6(1):27–74, 2014. [87] M. Ruzhansky and V. Turunen. Pseudo-differential operators and symmetries, volume 2 of Pseudo-Differential Operators. Theory and Applications. Birkhauser Verlag, Basel, 2010. Background analysis and advanced topics. [88] E. Schrohe. The symbols of an algebra of pseudodifferential operators. Pacific J. Math., 125(1):211–224, 1986. [89] E. Schrohe. A Ψ∗ algebra of pseudodifferential operators on noncompact manifolds. Arch. Math. (Basel), 51(1):81–86, 1988. [90] E. Schrohe. Fr´echet algebra techniques for boundary value problems on noncompact man- ifolds: Fredholm criteria and functional calculus via spectral invariance. Math. Nachr., 199:145–185, 1999. [91] E. Schrohe and B.-W. Schulze. Boundary value problems in Boutet de Monvel's algebra for manifolds with conical singularities. I. In Pseudo-differential calculus and mathematical physics, volume 5 of Math. Top., pages 97–209. Akademie Verlag, Berlin, 1994. [92] B.-W. Schulze. Pseudo-differential operators on manifolds with singularities, volume 24 of Studies in Mathematics and its Applications. North-Holland Publishing Co., 1991. [93] R. T. Seeley. Singular integrals on compact manifolds. Amer. J. Math., 81:658–690, 1959. [94] R. T. Seeley. The index of elliptic systems of singular integral operators. J. Math. Anal. Appl., 7:289–309, 1963. [95] S. R. Simanca. Pseudo-differential operators, volume 171 of Pitman research notes in math- ematics. Longman Scientific & Technical, Harlow, Essex, 1990. [96] A. Sims and D. Williams. Renault's equivalence theorem for reduced groupoid C ∗-algebras. J. Operator Theory, 68(1):223–239, 2012. [97] Bing Kwan So. On the full calculus of pseudo-differential operators on boundary groupoids with polynomial growth. Adv. Math., 237:1–32, 2013. [98] M. Taylor. Gelfand theory of pseudo differential operators and hypoelliptic operators. Trans. Amer. Math. Soc., 153:495–510, 1971. [99] G. Teschl. Mathematical methods in quantum mechanics, volume 157 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, second edition, 2014. With applications to Schrodinger operators. [100] J. Tu. Non-Hausdorff groupoids, proper actions and K-theory. Doc. Math., 9:565–597, 2004. [101] E. Van Erp and R. Yuncken. A groupoid approach to pseudodifferential operators. http://arxiv.org/abs/1511.01041 [math.DG], 2015. [102] E. Van Erp and R. Yuncken. On the tangent groupoid of a filtered manifold. http://arxiv.org/abs/1611.01081 [math.DG], 2016. [103] S. Vassout. Unbounded pseudodifferential calculus on Lie groupoids. J. Funct. Anal., 236(1):161–200, 2006. [104] B. Vertman. Heat-trace asymptotics for edge Laplacians with algebraic boundary conditions. J. Anal. Math., 125:285–318, 2015. [105] M. I. Visik and V. V. Grusin. Degenerate elliptic differential and pseudodifferential operators. Uspehi Mat. Nauk, 25(4(154)):29–56, 1970. [106] D. Williams. Crossed products of C ∗-algebras, volume 134 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2007. 38 C. CARVALHO, V. NISTOR, AND Y. QIAO Dep. Matem´atica, Instituto Superior T´ecnico, University of Lisbon, Av. Rovisco Pais, 1049-001 Lisbon, Portugal E-mail address: [email protected] Universit´e de Lorraine, UFR MIM, Ile du Saulcy, CS 50128, 57045 METZ, France and Inst. Math. Romanian Acad. PO BOX 1-764, 014700 Bucharest Romania E-mail address: [email protected] School of Mathematics and Information Science,, Shaanxi Normal University, Xi'an, 710119, China E-mail address: [email protected]
1112.4584	1	1112	2011-12-20T07:03:59	Equivariant semiprojectivity	[ "math.OA" ]	We define equivariant semiprojectivity for C-algebras equipped with actions of compact groups. We prove that the following examples are equivariantly semiprojective: arbitrary finite dimensional C-algebras with arbitrary actions of compact groups; the Cuntz algebras ${\mathcal{O}}_d$ and extended Cuntz algebras $E_d,$ for finite $d,$ with quasifree actions of compact groups; the Cuntz algebra ${\mathcal{O}}_{\infty}$ with any quasifree action of a finite group. For actions of finite groups, we prove that equivariant semiprojectivity is equivalent to a form of equivariant stability of generators and relations. We also prove that if $G$ is finite, then $C^* (G)$ is graded semiprojective.	math.OA	math	EQUIVARIANT SEMIPROJECTIVITY N. CHRISTOPHER PHILLIPS Abstract. We define equivariant semiprojectivity for C-algebras equipped with actions of compact groups. We prove that the following examples are equivariantly semiprojective: • Arbitrary finite dimensional C-algebras with arbitrary actions of com- pact groups. • The Cuntz algebras Od and extended Cuntz algebras Ed, for finite d, with quasifree actions of compact groups. • The Cuntz algebra O∞ with any quasifree action of a finite group. For actions of finite groups, we prove that equivariant semiprojectivity is equiv- alent to a form of equivariant stability of generators and relations. We also prove that if G is finite, then C ∗(G) is graded semiprojective. Semiprojectivity has become recognized as the "right" way to formulate many approximation results in C-algebras. The standard reference is Loring's book [20]. The formal definition and its basic properties are in Chapter 14 of [20], but much of the book is really about variations on semiprojectivity. Also see the more recent survey article [5]. There has been considerable work since then. In this paper, we introduce an equivariant version of semiprojectivity for C- algebras with actions of compact groups. (The definition makes sense for actions of arbitrary groups, but seems likely to be interesting only when the group is compact.) The motivation for the definition and our choice of results lies in applications which will be presented elsewhere. We prove that arbitrary actions of compact groups on finite dimensional C-algebras are equivariantly semiprojective, that quasifree actions of compact groups on the Cuntz algebras Od and the extended Cuntz al- gebras Ed, for finite d, are equivariantly semiprojective, and that quasifree actions of finite groups on O∞ are equivariantly semiprojective. We also give, for finite group actions, an equivalent condition for equivariant semiprojectivity in terms of equivariant stability of generators and relations. In a separate paper [26], we prove the following results relating equivariant semiprojectivity and ordinary semiprojectivity. If G is finite and (G, A, α) is equiv- ariantly semiprojective, then C∗(G, A, α) is semiprojective. If G is compact and second countable, A is separable, and (G, A, α) is equivariantly semiprojective, then A is semiprojective. Examples show that finiteness of G is necessary in the first statement, and that neither result has a converse. We do not address equivariant semiprojectivity of actions on Cuntz-Krieger alge- bras, on C([0, 1])⊗Mn, C(S1)⊗Mn, or dimension drop intervals (except for a result for C(S1) which comes out of our work on quasifree actions; see Remark 3.14), or Date: 20 December 2011. 2000 Mathematics Subject Classification. Primary 46L55. This material is based upon work supported by the US National Science Foundation under Grants DMS-0701076 and DMS-1101742. It was also partially supported by the Centre de Recerca Matem`atica (Barcelona) through a research visit conducted during 2011. 1 2 N. CHRISTOPHER PHILLIPS on C∗(Fn). We presume that suitable actions on these algebras are equivariantly semiprojective, but we leave investigation of them for future work. We also presume that there are interesting and useful equivariant analogs of weak stability of relations (Definition 4.1.1 of [20]), weak semiprojectivity (Definition 4.1.3 of [20]), projectivity (Definition 10.1.1 of [20]), and liftability of relations (Definition 8.1.1 of [20]). Again, we do not treat them. (Equivariant projectivity will be discussed in [26].) Finally, we point out work in the commutative case. It is well known that C(X) is semiprojective in the category of commutative C-algebras if and only if X is an absolute neighborhood retract. Equivariant absolute neighborhood retracts have a significant literature; as just three examples, we refer to the papers [15], [3], and [2]. (I am grateful to Adam P. W. Sørensen for calling my attention to the existence of this work.) This paper is organized as follows. Section 1 contains the definition of equivariant semiprojectivity, some related definitions, and the proofs of some basic results. Section 2 contains the proof that any action of a compact group on a finite dimensional C-algebra is equivariantly semiprojective. As far as we can tell, tra- ditional functional calculus methods (a staple of [20]) are of little use here. We use instead an iterative method for showing that approximate homomorphisms from compact groups are close to true homomorphisms. For a compact group G, we also prove that equivariant semiprojectivity is preserved when tensoring with any finite dimensional C-algebra with any action of G. In Section 3, we prove that quasifree actions of compact groups on the Cuntz algebra Od and the extended Cuntz algebras Ed, for d finite, are equivariantly semiprojective. We use an iterative method similar to that used for actions of finite dimensional C-algebras, but this time applied to cocycles. Section 4 extends the result to quasifree actions on O∞, but only for finite groups. The method is that of Blackadar [5], but a considerable amount of work needs to be done to set this up. We do not know whether the result extends to quasifree actions of general compact groups on O∞. In Section 5, we show that the universal C-algebra given by a bounded finite equivariant set of generators and relations is equivariantly semiprojective if and only if the relations are equivariantly stable. This is the result which enables most of the current applications of equivariant semiprojectivity. It is important for these applications that an approximate representation is only required to be approximately equivariant. We give one application here: we show that in the Rokhlin and tracial Rokhlin properties for an action of a finite group, one can require that the Rokhlin projections be exactly permuted by the group. Section 6 contains a proof that for a finite group G, the algebra C∗(G), with its natural G-grading, is graded semiprojective. This result uses the same machin- ery as the proof that actions on finite dimensional C-algebras are equivariantly semiprojective. We do not go further in this direction, but this result suggests that there is a much more general theory, perhaps of equivariant semiprojectivity for actions of finite dimensional quantum groups. I am grateful to Bruce Blackadar, Ilijas Farah, Adam P. W. Sørensen, and Hannes Thiel for valuable discussions. I also thank the Research Institute for Mathematical Sciences of Kyoto University for its support through a visiting professorship. EQUIVARIANT SEMIPROJECTIVITY 3 1. Definitions and basic results The following definition is the analog of Definition 14.1.3 of [20]. Definition 1.1. Let G be a topological group, and let (G, A, α) be a unital G- algebra. We say that (G, A, α) is equivariantly semiprojective if whenever (G, C, γ) is a unital G-algebra, J0 ⊂ J1 ⊂ · · · are G-invariant ideals in C, J =S∞ and πn : C/Jn → C/J κn : C → C/Jn, κ : C → C/J, n=0 Jn, are the quotient maps, and ϕ : A → C/J is a unital equivariant homomorphism, then there exist n and a unital equivariant homomorphism ψ : A → C/Jn such that πn ◦ ψ = ϕ. When no confusion can arise, we say that A is equivariantly semiprojective, or that α is equivariantly semiprojective. Here is the diagram: C κn κ C/Jn =④ πn ④ C/J. ψ ④ A ϕ ④ The solid arrows are given, and n and ψ are supposed to exist which make the diagram commute. We suppose that Definition 1.1 is probably only interesting when G is compact. Blackadar has shown that, in the nonunital category, the trivial action of Z on C is not equivariantly semiprojective [6]. This is equivalent to saying the the trivial action of Z on C ⊕ C is not equivariantly semiprojective in the sense defined here. In the unital category, the trivial action of any group on C is equivariantly semipro- jective for trivial reasons, but there are no other known examples of equivariantly semiprojective actions of noncompact groups. We will also need the following form of equivariant semiprojectivity for homo- morphisms. Our definition is not an analog of the definition of semiprojectivity for homomorphisms given before Lemma 14.1.5 of [20]. Rather, it is related to the second step in the idea of two step lifting as in Definition 8.1.6 of [20], with the caveat that lifting as there corresponds to projectivity rather than semiprojectiv- ity of a C-algebra. It is the equivariant version of a special case of conditional semiprojectivity as in Definition 5.11 of [8]. Definition 1.2. Let G be a topological group, let (G, A, α) and (G, B, β) be unital G-algebras, and let ω : A → B be a unital equivariant homomorphism. We say that ω is equivariantly conditionally semiprojective if whenever (G, C, γ) is a unital G-algebra, J0 ⊂ J1 ⊂ · · · are G-invariant ideals in C, J =S∞ κn : C → C/Jn, κ : C → C/J, and πn : C/Jn → C/J n=0 Jn, are the quotient maps, and λ : A → C and ϕ : B → C/J are unital equivariant homomorphisms such that κ ◦ λ = ϕ ◦ ω, then there exist n and a unital equivariant homomorphism ψ : B → C/Jn such that πn ◦ ψ = ϕ and κn ◦ λ = ψ ◦ ω. ✤ ✤ ✤ ✤ ✤ ✤ / / = 4 N. CHRISTOPHER PHILLIPS Here is the diagram: >⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥ ④ ψ λ ④ B ϕ ④ A ω C κn C/Jn =④ πn C/J. κ The part of the diagram with the solid arrows is assumed to commute, and n and ψ are supposed to exist which make the whole diagram commute. Remark 1.3. (1) Definition 1.1 is stated for the category of unital G-algebras. Without the group, a unital C-algebra is semiprojective in the unital cat- egory if and only if it is semiprojective in the nonunital category. (See Lemma 14.1.6 of [20].) The same is surely true here, and should be essen- tially immediate from what we do, but we don't need it and do not give a proof. (2) In the situations of Definition 1.1 and Definition 1.2, we say that ψ equiv- ariantly lifts ϕ. (3) In proofs, we will adopt the standard notation πn,m : C/Jm → C/Jn, for m, n ∈ Z>0 with n ≥ m, for the maps between the different quo- tients implicit in Definition 1.1 and Definition 1.2. Thus πn ◦ πn,m = πm and πn,m ◦ πm,l = πn,l for suitable choices of indices. We further let γ(n) : G → Aut(C/Jn) and γ(∞) : G → Aut(C/J) be the induced actions on the quotients. Lemma 1.4. Let G be a topological group, let (G, B, β) be a unital G-algebra, let A ⊂ B be a unital G-invariant subalgebra, and let ω : A → B be the inclusion. If A is equivariantly semiprojective and ω is equivariantly conditionally semiprojective, then B is equivariantly semiprojective. Proof. Let the notation be as in Definition 1.1 and Remark 1.3(3). Suppose ϕ : B → C/J is an equivariant unital homomorphism. Then equivariant semiprojectivity of A implies that there are n0 and an equivariant unital homomorphism λ : A → C/Jn0 such that πn0 ◦ λ = ϕA. Now apply equivariant conditional semiprojectivity of ω, with C/Jn0 in place of C and the ideals Jn/J0, for n ≥ n0, in place of the ideals Jn. We obtain n ≥ n0 and an equivariant unital homomorphism ψ : B → C/Jn such that πn ◦ ψ = ϕ (and also πn,n0 ◦ λ = ψA). (cid:3) Notation 1.5. Let (G, A, α) be a G-algebra. We denote by AG the fixed point algebra AG = {a ∈ A : αg(a) = a for all g ∈ G}. In case of ambiguity of the action, we write Aα. Further, if (G, B, β) is another G-algebra and ϕ : A → B is an equivariant ho- momorphism, then ϕ induces a homomorphism from AG to BG, which we denote by ϕG. We need the following two easy lemmas. ✤ ✤ ✤ ✤ ✤ ✤ / / > / / = EQUIVARIANT SEMIPROJECTIVITY 5 Lemma 1.6. Let G be a compact group, and let (G, C, γ) be a G-algebra. Let J ⊂ C be a G-invariant ideal. Then the obvious map ρ : CG/J G → C/J is injective and has range exactly (C/J)G. Proof. Injectivity is immediate from the relation J ∩ CG = J G. It is obvious that ρ(CG/J G) ⊂ (C/J)G. For the reverse inclusion, let x ∈ (C/J)G. Let π : C → C/J be the quotient map. Choose c ∈ C such that π(c) = x. Let µ be Haar measure on G, normalized so that µ(G) = 1. Set a =ZG γg(c) dµ(g). Then a ∈ CG and π(a) = x. Therefore a + J G ∈ CG/J G and ρ(a + J G) = x. (cid:3) Lemma 1.7. Let G be a compact group, and let (G, A, α) be a G-algebra. Let A0 ⊂ A1 ⊂ · · · be an increasing sequence of G-invariant subalgebras of A such that n=0 AG n . n ⊂ AG. For the reverse inclusion, let a ∈ AG and let ε > 0. Choose n and x ∈ An such that kx − ak < ε. Let µ be Haar measure on G, n and satisfies (cid:3) S∞ n=0 An = A. Then AG =S∞ Proof. It is clear thatS∞ normalized such that µ(G) = 1. Then b = RG γg(x) dµ(g) is in AG kb − ak < ε. n=0 AG Now we are ready to prove equivariant semiprojectivity of some G-algebras. Lemma 1.8. Let G be a compact group, let N ⊂ G be a closed normal subgroup, and let ρ : G → G/N be the quotient map. Let A be a unital C-algebra, and let α : G/N → Aut(A) be an equivariantly semiprojective action of G/N on A. Then (G, A, α ◦ ρ) is equivariantly semiprojective. Proof. We claim that there is an action γ : G/N → Aut(CN ) such that for g ∈ G and c ∈ CN we have γgN (c) = γg(c). One only needs to check that γ is well defined, which is easy. Let the notation be as in Definition 1.1 and Remark 1.3(3). Then ϕ(A) ⊂ (C/J)N , which by Lemma 1.6 is the same as CN /J N . Let ϕ0 : A → CN /J N n , so semiprojectivity of (G/N, A, α) provides n and a unital G/N -equivariant homomorphism ψ0 : A → CN /J N n which lifts ϕ0. We take ψ to be the following composition, in which the middle map comes from Lemma 1.6 and the last map is the inclusion: be the corestriction. Lemma 1.7 implies J N = S∞ n=0 J N ψ0−→ CN /J N Then ψ is G-equivariant and lifts ϕ. A n −→ (C/Jn)N −→ C/Jn. (cid:3) Corollary 1.9. Let G be a compact group, let A be a unital C-algebra, and let ι : G → Aut(A) be the trivial action of G on A. If A is semiprojective, then (G, A, ι) is equivariantly semiprojective. Proof. In Lemma 1.8, take N = G. (cid:3) Corollary 1.10. Let G be a compact group, and let (G, A, α) be a unital G- algebra. Then A is equivariantly semiprojective if and only if the inclusion of C · 1 in A is equivariantly conditionally semiprojective in the sense of Definition 1.2. Proof. The subalgebra C · 1 is equivariantly semiprojective by Corollary 1.9, so we may apply Lemma 1.4. (cid:3) 6 N. CHRISTOPHER PHILLIPS Proposition 1.11. Let G be a compact group, and let (cid:0)(cid:0)G, Ak, α(k)(cid:1)(cid:1)m l ∈ {0, 1, . . . , m − 1}. Set A = (cid:16)Ll k=1 be a finite collection of equivariantly semiprojective unital G-algebras. Suppose that k=1 Ak, with the obvious direct sum actions α : G → Aut(A) (with G acting trivially on C) and β : G → Aut(B). Define ω : A → B by k=1 Ak(cid:17) ⊕ C and set B = Lm ω(a1, a2, . . . , al, λ) =(cid:0)a1, a2, . . . , al, λ · 1Al+1, λ · 1Al+2, . . . , λ · 1Am(cid:1) for a1 ∈ A1, a2 ∈ A2, . . . , al ∈ Al, and λ ∈ C. Then ω is equivariantly conditionally semiprojective. let fk ∈ B be the identity of the summand Ak ⊂ B. Set q = 1 −Pl Proof. Let the notation be as in Definition 1.2 and Remark 1.3(3). For k = 1, 2, . . . , l let ek ∈ A be the identity of the summand Ak ⊂ A, and for k = 1, 2, . . . , m k=1 µ(ek). Let P ⊂ B be the subalgebra generated by fl+1, fl+2, . . . , fm. Then P is semiprojec- tive and G acts trivially on it. Therefore Corollary 1.9 provides n0 and a unital equivariant homomorphism ψ0 : P → qCq/qJn0 q such that πn0 ◦ ψ0 = ϕP . For k = l + 1, l + 2, . . . , m, set pk = ψ0(ek). Use equivariant semiprojectivity of Ak, with pk(C/Jn0)pk in place of C and with pk(Jn/Jn0)pk in place of Jn (for n ≥ n0) to find nk ≥ n0 and a unital equivariant lifting ψk : Ak → πnk,n0(pk)(C/Jnk )πnk,n0(pk) of ϕAk . Define n = max(n1, n2, . . . , nm), and define ψ : A → C/Jn by ψ(a1, a2, . . . , am) = (κn ◦ µ)(a1, a2, . . . , al) + mXk=l+1 πn,nk (ψk(ak)). Then ψ is an equivariant lifting of ϕ. (cid:3) Corollary 1.12. Let G be a compact group, and let(cid:0)(cid:0)G, Ak, α(k)(cid:1)(cid:1)m collection of equivariantly semiprojective unital G-algebras. Then A =Lm with the direct sum action α : G → Aut(A), is equivariantly semiprojective. k=1 be a finite k=1 Ak, Proof. Proposition 1.11 (with l = 0) implies that the unital inclusion of C in A is equivariantly conditionally semiprojective, so Corollary 1.10 implies that A is equivariantly semiprojective. (cid:3) We can use traditional methods to give an example of a nontrivial action which is equivariantly semiprojective. This result will be superseded in Theorem 2.6 below, using more complicated methods, so the proof here will be sketchy. Proposition 1.13. Let G be a finite cyclic group. Let G act on C(G) by the trans- lation action, τg(a)(h) = a(g−1h) for g, h ∈ G and a ∈ C(G). Then (G, C(G), τ ) is equivariantly semiprojective. Proof. Let the notation be as in Definition 1.1 and Remark 1.3(3). Take G = Z/dZ =(cid:8)1, e2πi/d, e4πi/d, . . . , e2(d−1)πi/d(cid:9) ⊂ S1. Let u be the inclu- sion of G in S1, which we regard as a unitary in C(G). Then u generates C(G) and τλ(u) = λ−1u for λ ∈ G. Therefore it suffices to find n and a unitary z ∈ C/Jn such that πn(z) = ϕ(u), sp(z) ⊂ G, and γ(n) λ (z) = λ−1z for all λ ∈ G. EQUIVARIANT SEMIPROJECTIVITY 7 Since C(G) is semiprojective (in the nonequivariant sense), there are n0 and a 0 = 1. Moreover, for all λ ∈ G, unitary v0 ∈ C/Jn0 such that πn(v0) = ϕ(u) and vd we have lim n→∞(cid:13)(cid:13)πn,n0(cid:0)γ(n0) λ (v0)(cid:1) − λ−1v0(cid:13)(cid:13) = 0. 1 Choose ε > 0 such that ε < 1 C-algebra and b ∈ B satisfies kb − 1k < ε, then (cid:13)(cid:13)b(b∗b)−1/2 − 1(cid:13)(cid:13) < 1 Choose n so large that v = πn,n0(v0) satisfies(cid:13)(cid:13)γ(n) 2(cid:12)(cid:12)1 − eπi/d(cid:12)(cid:12), and such that whenever B is a unital 2(cid:12)(cid:12)1 − eπi/d(cid:12)(cid:12). λ (v) − λ−1v(cid:13)(cid:13) < ε for all λ ∈ G. d Xλ∈G Define a ∈ C/Jn by Then one checks that γ(n) so a is invertible. Set w = a(a∗a)−1/2, and check that γ(n) λ ∈ G. A calculation, using the choice of ε, shows that kw − vk < 1 So eπi/dG ∩ sp(w) = ∅. Let f : S1 \ eπi/dG → S1 be the function determined by λ (a) = λ−1a for all λ ∈ G and that ka − vk < ε < 1, λ (w) = λ−1w for all 2(cid:12)(cid:12)1 − eπi/d(cid:12)(cid:12). d (cid:1). Then f (λζ) = λf (ζ) for all λ ∈ G and ζ ∈ S1 \ eπi/dG, and f is continuous on sp(w). Define z = f (w). The verification that z satisfies the required conditions is a calculation. (cid:3) g(eit) = e2πik/d when t ∈ (cid:0) 2k−1 λ (v). , 2k+1 λγ(n) a = d 2. Equivariant semiprojectivity of finite dimensional C-algebras The main result of this section is that actions of compact groups on finite di- mensional C-algebras are equivariantly semiprojective. The main technical tool is a method for replacing approximate homomorphisms to unitary groups by nearby exact homomorphisms, in such a way as to preserve properties such as being equivariant. (In Section 6, we will also need to preserve the property of being graded.) The method used here has been discovered twice before, in Theorem 3.8 of [13] (most of the work is in Section 4 of [12], but the result in [12] uses the wrong metric on the groups) and in Theorem 1 of [18]. It is not clear from either of these proofs that the additional properties we need are preserved. We will instead follow the proofs of Theorem 5.13 and Proposition 5.14 of [1]. (We are grateful to Ilijas Farah for pointing out these references.) Notation 2.1. For a unital C-algebra A, we let U (A) denote the unitary group of A. The following lemmas give an estimate whose proof is omitted in [1]. We will need this estimate again, in the proof of Lemma 3.9 below. (We don't get quite the same estimate as implied in [1].) Lemma 2.2. Let Γ be a compact group with normalized Haar measure µ. Let A function such that ku(g) − 1k ≤ r for all g ∈ G. Then be a unital C-algebra. Suppose r ∈(cid:2)0, 1 u(g) dµ(g) − exp(cid:18)ZΓ (cid:13)(cid:13)(cid:13)(cid:13)ZΓ 2(cid:3), and let u : Γ → U (A) be a continuous log(u(g)) dµ(g)(cid:19)(cid:13)(cid:13)(cid:13)(cid:13) ≤ u(g) dµ(g)(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1. (cid:13)(cid:13)(cid:13)(cid:13)ZΓ 2(1 − 2r) and 5r2 8 N. CHRISTOPHER PHILLIPS Proof. The second statement is obvious. For the first, we require the following estimates (compare with Lemma 5.15 of [1]): for u ∈ U (A) with ku − 1k < 1, we have (cid:13)(cid:13) log(u) − (u − 1)(cid:13)(cid:13) ≤ (cid:13)(cid:13) exp(a) − (1 + a)(cid:13)(cid:13) ≤ and for a ∈ A with kak < 1, we have Both are obtained from power series: (2.1) (2.2) and (2.3) Since r ≤ 1 ∞Xn=2 (cid:13)(cid:13) log(u) − (u − 1)(cid:13)(cid:13) ≤ (cid:13)(cid:13) exp(a) − (1 + a)(cid:13)(cid:13) ≤ (cid:13)(cid:13)(cid:13)(cid:13)ZΓ log(u(g)) dµ(g) −ZΓ log(u(g)) dµ(g)(cid:13)(cid:13)(cid:13)(cid:13) ≤ log(u(g)) dµ(g)(cid:19) −ZΓ (cid:13)(cid:13)(cid:13)(cid:13)ZΓ (cid:13)(cid:13)(cid:13)(cid:13)exp(cid:18)ZΓ log(u(g)) dµ(g)(cid:19) −ZΓ 2 , we also get r2 ku − 1k2 n ≤ ≤ 1 2 1 2 kak2 ku − 1kn kakn n! 2(cid:0)1 − ku − 1k(cid:1) , 2(cid:0)1 − kak(cid:1) . ∞Xn=2 ∞Xn=2 ∞Xn=2 u(g) dµ(g) − 1(cid:13)(cid:13)(cid:13)(cid:13) ≤ +ZΓ log(u(g)) dµ(g) − 1(cid:13)(cid:13)(cid:13)(cid:13) ≤ kakn. ku − 1kn r2 . 2(1 − r) (2r)2 2(1 − 2r) . Apply (2.1) to the condition ku(g) − 1k ≤ r and integrate, getting We therefore get, integrating and using (2.2), 2(1 − r) ku(g) − 1k dµ(g) ≤ 2r. Combining this estimate with (2.3) gives as desired. (cid:13)(cid:13)(cid:13)(cid:13)exp(cid:18)ZΓ u(g) dµ(g)(cid:13)(cid:13)(cid:13)(cid:13) ≤ be a unital C-algebra. Suppose r ∈(cid:2)0, 1 function such that for all g, h ∈ Γ we have r2 (2r)2 + 2(1 − r) 2(1 − 2r) ≤ 5r2 2(1 − 2r) , (cid:3) 5(cid:3), and let ρ : Γ → U (A) be a continuous Lemma 2.3. Let Γ be a compact group with normalized Haar measure µ. Let A kρ(gh) − ρ(g)ρ(h)k ≤ r. For g ∈ Γ define σ(g) = exp(cid:18)ZΓ log(cid:16)ρ(k)∗ρ(kg)ρ(g)∗(cid:17) dµ(k)(cid:19) ρ(g). Then σ is a continuous function from Γ to U (A) which satisfies kσ(gh) − σ(g)σ(h)k ≤ 17r2 and kσ(g) − ρ(g)k ≤ 2r for all g, h ∈ Γ. EQUIVARIANT SEMIPROJECTIVITY 9 Proof. For g ∈ Γ, define σ0(g) =ZΓ ρ(k)∗ρ(kg) dµ(k). The first part of the proof of Proposition 5.14 of [1] shows that for g, h ∈ Γ, we have kσ0(gh) − σ0(g)σ0(h)k ≤ 2r2 and kσ0(g) − ρ(g)k ≤ r. The rest of the proof in [1] uses a Lie algebra valued logarithm, called "ln" there. We replace statements in [1] involving the Lie algebra of the codomain with the use of the logarithm coming from holomorphic functional calculus. Rewriting σ0(g) =(cid:18)ZΓ ρ(k)∗ρ(kg)ρ(g)∗ dµ(k)(cid:19) ρ(g) and applying Lemma 2.2, we get kσ(g) − σ0(g)k ≤ 5r2 2(1 − 2r) ≤ 5r2 ≤ r for all g ∈ Γ. This implies kσ(g) − ρ(g)k ≤ 2r for all g ∈ Γ, which is the second of the required estimates. Clearly kσ(g)k ≤ 1 for all g ∈ Γ. Lemma 2.2 implies kσ0(g)k ≤ 1 for all g ∈ Γ. Therefore kσ(gh) − σ(g)σ(h)k ≤ kσ(gh) − σ(gh)k + kσ(g) − σ(g)k + kσ(h) − σ(h)k + kσ0(gh) − σ0(g)σ0(h)k ≤ 5r2 + 5r2 + 5r2 + 2r2 = 17r2. This is the first of the required estimates. (cid:3) Lemma 2.4. Let Γ be a compact group with normalized Haar measure µ. Let A and B be unital C-algebras, and let κ : A → B be a unital homomorphism. Suppose 0 ≤ r < 1 17 , and let ρ0 : Γ → U (A) be a continuous map such that for all g, h ∈ Γ, we have kρ0(gh) − ρ0(g)ρ0(h)k ≤ r and (κ ◦ ρ0)(gh) = (κ ◦ ρ0)(g)(κ ◦ ρ0)(h). Inductively define functions ρm : Γ → A by (following Lemma 2.3) ρm+1(g) = exp(cid:18)ZΓ log(cid:16)ρm(k)∗ρm(kg)ρm(g)∗(cid:17) dµ(k)(cid:19) ρm(g) for g ∈ Γ. Then for every m ∈ Z>0 the function ρm is a well defined continuous function from Γ to U (A) such that κ ◦ ρm = κ ◦ ρ0. Moreover, the functions ρm converge uniformly to a continuous homomorphism ρ : Γ → U (A) such that kρ(g) − ρ0(g)k ≤ sup g∈Γ 2r 1 − 17r and κ ◦ ρ = κ ◦ ρ0. Proof. We claim that for all m ∈ Z≥0, the function ρm is well defined, continuous, take values in U (A), and satisfies κ ◦ ρm = κ ◦ ρ0, and that for g, h ∈ Γ we have (2.4) and (2.5) kρm(gh) − ρm(g)ρm(h)k ≤ r(17r)m kρm(g) − ρm−1(g)k ≤ 2r(17r)m−1. The proof of the claim is by induction on m. The case m = 1 is Lemma 2.3 and 5 . Assume the result is known for m. Since the estimates (2.4) and (2.5) hold r ≤ 1 10 N. CHRISTOPHER PHILLIPS for m, and by Lemma 2.3 and because r(17r)m < r ≤ 1 defined, continuous, take values in U (A), and for g, h ∈ Γ we have 5 , the function ρm+1 is well and, also using 17r < 1 at the last step, kρm+1(g) − ρm(g)k ≤ 2r(17r)m kρm+1(gh) − ρm+1(g)ρm+1(h)k ≤ 17(cid:0)r(17r)m(cid:1)2 It remains to prove that κ ◦ ρm+1 = κ ◦ ρ0. Let g ∈ Γ. Using κ ◦ ρm = κ ◦ ρ0 at the second step and the fact that κ ◦ ρ0 is a homomorphism at the last step, we get = r(17r)m+1(17r)m < r(17r)m+1. κ(cid:18)exp(cid:18)ZΓ = exp(cid:18)ZΓ = exp(cid:18)ZΓ log(cid:16)ρm(k)∗ρm(kg)ρm(g)∗(cid:17) dµ(k)(cid:19)(cid:19) log(cid:16)(κ ◦ ρm)(k)∗(κ ◦ ρm)(kg)(κ ◦ ρm)(g)∗(cid:17) dµ(k)(cid:19) log(cid:16)(κ ◦ ρ0)(k)∗(κ ◦ ρ0)(kg)(κ ◦ ρ0)(g)∗(cid:17) dµ(k)(cid:19) = 1. For g ∈ Γ we get, changing variables at the second step, (2.6) aρ(g) =ZΓ σ(h)∗ρ(hg) dµ(h) =ZΓ σ(cid:0)hg−1(cid:1)∗ ρ(h) dµ(h) = σ(g)a. Since kσ(h)∗ρ(h) − 1k < 1 for all h ∈ Γ, we have ka − 1k < 1. Therefore u = a(a∗a)−1/2 is a well defined unitary in A. Taking adjoints in (2.6), we get a∗σ(g) = ρ(g)a∗ for all g ∈ Γ, so a∗a commutes with ρ(g). Thus (a∗a)−1/2 commutes with ρ(g). Applying (2.6) again, we get uρ(g) = σ(g)u for all g ∈ Γ. The hypotheses imply that κ(a) = 1, so also κ(u) = 1. (cid:3) Therefore κ(ρm+1(g)) = κ(ρm(g)) = κ(ρ0(g)). This completes the induction, and proves the claim. The estimate (2.5) implies that there is a continuous function ρ : Γ → U (A) such that ρm → ρ uniformly, and in fact for g ∈ Γ we have kρ(g) − ρ0(g)k ≤ 2r(17r)m−1 = 2r 1 − 17r . ∞Xm=1 The estimate (2.4) and convergence imply that ρ is a homomorphism. Continuity of κ implies that κ ◦ ρ = κ ◦ ρ0. (cid:3) The following proposition is a variant of the fact that two close homomorphisms from a finite dimensional C-algebra are unitarily equivalent. Proposition 2.5. Let Γ be a compact group, let A and B be unital C-algebras, and let κ : A → B be a unital homomorphism. Let ρ, σ : Γ → U (A) be two contin- uous homomorphisms such that kρ(g) − σ(g)k < 1 and κ ◦ ρ(g) = κ ◦ σ(g) for all g ∈ Γ. Then there exists a unitary u ∈ A such that uρ(g)u∗ = σ(g) for all g ∈ Γ, and such that κ(u) = 1. Proof. Let µ be normalized Haar measure on Γ. Define a =ZΓ σ(h)∗ρ(h) dµ(h). EQUIVARIANT SEMIPROJECTIVITY 11 Theorem 2.6. Let α : G → Aut(A) be an action of a compact group G on a finite dimensional C-algebra A. Then (G, A, α) is equivariantly semiprojective. Proof. Set ε0 = 1 6·34 , and choose ε > 0 such that ε ≤ ε0 and such that whenever A is a unital C-algebra, u ∈ U (A), and a ∈ A satisfies ka − uk < ε, then we have (cid:13)(cid:13)a(a∗a)−1/2 − u(cid:13)(cid:13) < ε0. Let the notation be as in Definition 1.1 and Remark 1.3(3). Let ϕ : A → C/J be a unital equivariant homomorphism. Since finite dimensional C-algebras are semiprojective, there exist n0 and a unital homomorphism (not necessarily equi- variant) ψ0 : A → C/Jn0 which lifts ϕ. For n ≥ n0, define fn : G × U (A) → [0, ∞) by for g ∈ G and x ∈ U (A). The functions fn are continuous and satisfy fn(g, x) =(cid:13)(cid:13)πn,n0(cid:0)ψ0(αg(x)) − γ(n0) g (ψ0(x))(cid:1)(cid:13)(cid:13) fn0 ≥ fn0+1 ≥ fn0+2 ≥ · · · . n=1 Jn at the first step and equivariance of ϕ at the second step, for Using J =S∞ g ∈ G and x ∈ U (A) we have lim n→∞ fn(g, x) = kϕ(αg(x)) − γg(ϕ(x))k = 0. Since G × U (A) is compact, Dini's Theorem (Proposition 11 in Chapter 9 of [30]) implies that fn → 0 uniformly. Therefore there exists n ≥ n0 such that for all g ∈ G and x ∈ U (A), we have Set ψ1 = πn,n0 ◦ ψ. Then this estimate becomes (2.7) for every g ∈ G and x ∈ U (A). Let ν be normalized Haar measure on G. For x ∈ A define g g (cid:13)(cid:13)πn,n0(cid:0)ψ0(αg(x)) − γ(n0) (cid:13)(cid:13)ψ1(αg(x)) − γ(n) T (x) =ZG(cid:0)γ(n) (T (x)) =ZG(cid:0)γ(n) =ZG(cid:0)γ(n) (ψ0(x))(cid:1)(cid:13)(cid:13) < ε. (ψ1(x))(cid:13)(cid:13) < ε. h (cid:1)(x) dν(h). h (cid:1)(x) dν(h) g−1h(cid:1)(x) dν(h) = T (αg(x)). h ◦ ψ1 ◦ α−1 gh ◦ ψ1 ◦ α−1 h ◦ ψ1 ◦ α−1 Then for g ∈ G we have γ(n) g So T is equivariant. Also, since πn ◦ψ1 = ϕ is equivariant, we have πn(T (x)) = ϕ(x) for all x ∈ A. It follows from (2.7) that kT (x) − ψ1(x)k < ε for all x ∈ U (A). Since ε < 1, we may define ρ0 : U (A) → U (C/Jn) by for x ∈ U (A). Then γ(n) g ρ0(x) = T (x)(cid:0)T (x)∗T (x)(cid:1)−1/2 (ρ0(x)) = ρ0(αg(x)) and πn(ρ0(x)) = ϕ(x) for all g ∈ G and x ∈ U (A). By the choice of ε, we have kρ0(x) − T (x)k < ε0, whence kρ0(x) − ψ1(x)k ≤ kρ0(x) − T (x)k + kT (x) − ψ1(x)k < ε0 + ε ≤ 2ε0. 12 N. CHRISTOPHER PHILLIPS Let x, y ∈ U (A). Since ψ1(x), ψ1(y) ∈ U (C/Jn) and ψ1(xy) = ψ1(x)ψ1(y), it follows that kρ0(xy) − ρ0(x)ρ0(y)k < 6ε0. Let µ be normalized Haar measure on the compact group U (A). Inductively define functions ρm : Γ → U (C/Jm) by (following Lemma 2.4) ρm+1(x) = ρm(x) exp ZU(A) log(cid:16)ρm(x)∗ρm(xy)ρm(y)∗(cid:17) dµ(y)! for x ∈ U (A). Since 6ε0 = 1 17 , Lemma 2.4 implies that each function ρm is a well defined continuous function from U (A) to U (C/Jn) and that ρ(x) = limm→∞ ρm(x) defines a continuous homomorphism from U (A) to U (C/Jn) satis- fying 34 < 1 kρ(x) − ρ0(x)k ≤ 2 · 6ε0 1 − 17 · 6ε0 = 2 17 and πn(ρ(x)) = ϕ(x) for all x ∈ U (A). Since homomorphisms respect functional calculus, an induction argument shows that (ρm(x)) = ρm(αg(x)) for all m ∈ Z≥0, g ∈ G, and x ∈ U (A). Therefore also γ(n) g (2.8) γ(n) g (ρ(x)) = ρ(αg(x)) for all g ∈ G and x ∈ U (A). For x ∈ U (A) we have kρ(x) − ψ1(x)k ≤ kρ(x) − ρ0(x)k + kρ0(x) − ψ1(x)k < 2 17 + 6ε0 < 1. Since πn(ρ(x)) = ϕ(x) = πn(ψ1(x)) for x ∈ U (A), and since U (A) is compact, Proposition 2.5 provides a unitary w ∈ C/Jn such that πn(w) = 1 and such that wψ1(x)w∗ = ρ(x) for all x ∈ U (A). Define a homomorphism ψ : A → C/Jn by ψ(a) = wψ1(x)w∗ for a ∈ A. Then ψ lifts ϕ because πn(w) = 1. Furthermore, ψ is equivariant by (2.8) and because U (A) spans A. (cid:3) As an immediate application, one can require that the projections in the defi- nitions of the Rokhlin and tracial Rokhlin properties for finite groups be exactly orthogonal and exactly permuted by the group action, rather than merely being approximately permuted by the group action. We postpone the proof until after dis- cussion equivariant stability of relations. See Proposition 5.26 and Proposition 5.27. We can now show that tensoring with finite dimensional G-algebras preserves equivariant semiprojectivity. The proof is essentially due to Adam P. W. Sørensen, and Hannes Thiel and we are grateful to them for their permission to include it here. We begin with a lemma. Lemma 2.7. Let A1 and A2 be unital C-algebras, and let ϕ : A1 → A2 be a surjec- tive homomorphism. Let F be a finite dimensional C-algebra, and let λ1 : F → A1 be a unital homomorphism. Set λ2 = ϕ ◦ λ1. For s = 1, 2 define Bs =(cid:8)a ∈ As : a commutes with λ(x) for all x ∈ F(cid:9). Then ϕB1 is a surjective homomorphism from B1 to B2. EQUIVARIANT SEMIPROJECTIVITY 13 Proof. There are n, r(1), r(2), . . . , r(n) ∈ Z>0 such that F =Ln Mr(l) for l = 1, 2, . . . , n. Let(cid:0)e(l) For s = 1, 2 define Es : As → As by j,k=1 be a system of matrix units for Fl. l=1 Fk and Fl ∼= j,k(cid:1)r(l) r(l)Xk=1 nXl=1 λs(cid:0)e(l) 1,k(cid:1) k,1(cid:1)aλs(cid:0)e(l) (2.9) Es(a) = for a ∈ As. We claim that Es(a) commutes with λs(x) for a ∈ As, x ∈ F, and s = 1, 2. It suffices to take x = e(m) i,j . In the product Ea(a)λs(x), the terms coming from (2.9) with l 6= m vanish, leaving r(m)Xk=1 1,k e(m) Es(a)λs(x) = λs(cid:0)e(m) k,1(cid:1)aλs(cid:0)e(m) Similarly, we also get λs(x)Es(a) = λs(cid:0)e(m) λs(cid:0)e(l) i,1 (cid:1)aλs(cid:0)e(m) k,k(cid:1) = λs(1) = 1 r(l)Xk=1 nXl=1 The relation i,1 (cid:1)aλs(cid:0)e(m) i,j (cid:1) = λs(cid:0)e(m) 1,j (cid:1). 1,j (cid:1), proving the claim. implies that for s = 1, 2 and a ∈ Bs, we have Es(a) = a. It is clear that E2 ◦ ϕ = ϕ ◦ E1. Now let b ∈ B2. Choose a ∈ A1 such that ϕ(a) = b. Then E1(a) ∈ B1, and (cid:3) ϕ(E1(a)) = E2(b) = b. This completes the proof. Theorem 2.8. Let G be a compact group, let A be a unital C-algebra, and let F be a finite dimensional C-algebra. Let α : G → Aut(A) be an equivariantly semiprojective action, and let β : G → Aut(F ) be any action. Then β ⊗ α : G → Aut(F ⊗ A) is equivariantly semiprojective. Proof. Define ω : F → F ⊗ A by ω(x) = x ⊗ 1. By Theorem 2.6 and Lemma 1.4, it suffices to prove that ω is equivariantly conditionally semiprojective. Let the notation be as in Definition 1.2, except with F in place of A and F ⊗ A in place of B. We have the diagram C κn ;✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇✇ ψ ✈ ✈ ✈ ✈ λ F ⊗ A ϕ ω C/Jn ;✈ πn C/J, F κ in which the solid arrows correspond to given equivariant unital homomorphisms. We must find n and an equivariant unital homomorphism ψ which make the whole diagram commute. Define D =(cid:8)c ∈ C : c commutes with λ(x) for all x ∈ F(cid:9), which is a G-invariant subalgebra of C. Define In = Jn ∩ D for n ∈ Z≥0, and n=0 In. set I = J ∩ D. Then I, I0, I1, . . . are G-invariant ideals in D, and I = S∞ ✤ ✤ ✤ ✤ ✤ ✤ / / ; / / ; 14 N. CHRISTOPHER PHILLIPS Moreover, Lemma 2.7 implies that for n ∈ Z≥0, and that (2.10) D/In =(cid:8)c ∈ C/Jn : c commutes with (κn ◦ λ)(x) for all x ∈ F(cid:9) D/I =(cid:8)c ∈ C/J : c commutes with (κ ◦ λ)(x) for all x ∈ F(cid:9). Define an equivariant homomorphism ϕ0 : A → C/J by ϕ0(a) = ϕ(1 ⊗ a) for a ∈ A. By (2.10), the range of ϕ0 is contained in D/I. Since A is equivariantly semipro- jective, there are n and a unital equivariant homomorphism ψ0 : A → D/In ⊂ C/Jn such that πn ◦ ψ0 = ϕ0. By construction, the ranges of ψ0 and κn ◦ λ commute, so there is a unital homomorphism ψ0 : A → C/Jn, necessarily equivariant, such that ψ(x ⊗ a) = (κn ◦ λ)(x)ψ0(a) for all x ∈ F and a ∈ A. Using πn ◦ κn ◦ λ = ϕ ◦ ω, we get πn ◦ ψ = ϕ. (cid:3) 3. Quasifree actions on Cuntz algebras The purpose of this section is to prove that quasifree actions of compact groups on the Cuntz algebras Od and the extended Cuntz algebras Ed, for d finite, are equivariantly semiprojective. We begin by defining and introducing notation for quasifree actions. Notation 3.1. Let d ∈ Z>0. (We allow d = 1, in which case Od = C(S1).) We write s1, s2, . . . , sd for the standard generators of the Cuntz algebra Od. That is, we take Od to be generated by elements s1, s2, . . . , sd satisfying the relations s∗ j sj = 1 for j = 1, 2, . . . , d andPd j=1 sjs∗ j = 1. Notation 3.2. Let d ∈ Z>0. We recall the extended Cuntz algebra Ed. It is the universal unital C-algebra generated by d isometries with orthogonal range pro- jections which are not required to add up to 1. (For d = 1, we get the Toeplitz algebra, the C-algebra of the unilateral shift, which here is called r1.) We call these isometries r1, r2, . . . , rd, so that the relations are r∗ j rj = 1 for j = 1, 2, . . . , d and rj r∗ k = 0 for j, k = 1, 2, . . . , d with j 6= k. When d must be specified, we write r(d) 2 , . . . , r(d) d . We further let η : Ed → Od be the quotient map, defined, following Notation 3.1, by η(rj) = sj for j = 1, 2, . . . , d. j rkr∗ 1 , r(d) Notation 3.3. For λ = (λ1, λ2, . . . , λd) ∈ Cd, we further define sλ ∈ Od and rλ ∈ Ed by sλ = dXj=1 λjsj and rλ = λj rj. dXj=1 Notation 3.4. Let d ∈ Z>0. We let (ej,k)n j,k=1 be the standard system of matrix units in Mn. We denote by µ : Md → Od the injective unital homomorphism deter- mined by µ(ej,k) = sjs∗ k for j, k = 1, 2, . . . , d. We further denote by µ0 : Md ⊕ C → k for Ed the injective unital homomorphism determined by µ0(cid:0)(ej,k, 0)(cid:1) = rjr∗ j, k = 1, 2, . . . , d and µ0(0, 1) = 1 −Pd Recall (Notation 2.1) that U (A) is the unitary group of A. j=1 rjr∗ j . EQUIVARIANT SEMIPROJECTIVITY 15 Notation 3.5. Let A be a unital C-algebra, and let u ∈ U (A). We denote by Ad(u) the automorphism Ad(u)(a) = uau∗ for a ∈ A. Further let G be a topological group, and let ρ : G → U (A) be a continuous homomorphism. We denote by Ad(ρ) the action α : G → Aut(A) given by Ad(ρ)g = Ad(ρ(g)) for g ∈ G. For ρ : G → U (Md), we also write Ad(ρ ⊕ 1) for the action g 7→ Ad(ρ(g), 1) on Mn ⊕ C. We always take Md ⊕ C to have this action. We now give the basic properties of quasifree actions on Ed. Lemma 3.6. Let G be a topological group, let d ∈ Z>0, and let ρ : G → U (Md) be a unitary representation of G on Cd. Then there exists a unique action αρ : G → Aut(Ed) such that, with µ0 as in Notation 3.4, for j = 1, 2, . . . , d and g ∈ G we have αρ g(rj ) = µ0(ρ(g), 1)rj . Moreover, this action has the following properties: (1) For all g ∈ G, if we write ρj,k(g)ej,k = ρ1,1(g) ρ1,2(g) ρ2,1(g) ρ2,2(g) ... ... ρd,1(g) ρd,2(g) · · · · · · . . . · · · ρ1,d(g) ρ2,d(g) ... ρd,d(g)  , ρ(g) = dXj,k=1 then αρ g(rk) = ρj,k(g)rj dXj=1 for k = 1, 2, . . . , d. (2) Following Notation 3.3, for every λ ∈ Cd and g ∈ G, we have αρ g(rλ) = rρ(g)λ. (3) The homomorphism µ0 is equivariant. (Recall the action on Md ⊕ C from Notation 3.5.) (4) The projection 1 −Pd j=1 rj r∗ j ∈ Ed is G-invariant. Proof. For g ∈ G, we claim that the elements µ0(ρ(g), 1)rj satisfy the relations defining Ed. Because µ0(ρ(g), 1) is unitary, we have (cid:0)µ0(ρ(g), 1)rj(cid:1)∗(cid:0)µ0(ρ(g), 1)rk(cid:1) = r∗ j rk for j, k = 1, 2, . . . , d. Using the definition of µ0 at the second step, we also get (3.1) dXj=1(cid:2)µ0(ρ(g), 1)rj(cid:3)(cid:2)µ0(ρ(g), 1)rj(cid:3)∗ = µ0(ρ(g), 1)(cid:18)Xd rjr∗ j=1 j(cid:19) µ0(ρ(g), 1) = rjr∗ j . dXj=1 It is now easy to prove the claim. It follows that there is a unique homomorphism αρ g(rj) = µ0(ρ(g), 1)rj for j = 1, 2, . . . , d. Part (4) follows from (3.1). Part (1) is just a calculation, and implies part (2) when λ ∈ Cd is a standard basis vector. The general case of part (2) follows by linearity. g : Ed → Ed such that αρ 16 N. CHRISTOPHER PHILLIPS Part (2) implies that αρ αρ 1 = idEd , so g 7→ αρ a ∈ Ed follows from the fact that it holds whenever a = rj for some j. h(sλ) for g, h ∈ G and λ ∈ C, and also g is a homomorphism to Aut(Ed). Continuity of g 7→ αg(a) for gh(sλ) = αρ g ◦ αρ g(cid:0)µ0(ej,k, 0)(cid:1) = αρ It remains to prove (3). For j, k = 1, 2, . . . , d, we have αρ k) = µ0(ρ(g), 1)rj r∗ g(rj )αρ g(r∗ kµ0(ρ(g), 1)∗ = µ0(cid:0)(ρ(g), 1)(ej,k, 0)(ρ(g), 1)∗(cid:1) = µ0(cid:0)Ad(ρ ⊕ 1)g(ej,k, 0)(cid:1), as desired. We must also to check the analogous equation with (0, 1) in place of (ej,k, 0), but this is immediate from part (4). (cid:3) Here are the corresponding properties for quasifree actions on Od. These are mostly well known, and are stated for reference and to establish notation. They also follow from Lemma 3.6. Lemma 3.7. Let G be a topological group, let d ∈ Z>0, and let ρ : G → U (Md) be a unitary representation of G on Cd. Then there exists a unique action βρ : G → Aut(Od) such that, with µ as in Notation 3.4, for j = 1, 2, . . . , d and g ∈ G we have βρ g (sj ) = µ(ρ(g), 1)sj. Moreover, this action has the following properties: (1) For all g ∈ G, if we write ρ(g) = ρj,k(g)ej,k, βρ g (sk) = ρj,k(g)sj dXj,k=1 dXj=1 then for k = 1, 2, . . . , d. (2) Following Notation 3.3, for every λ ∈ Cd and g ∈ G, we have βρ g (sλ) = sρ(g)λ. (3) When Md is equipped with the action Ad(ρ), the homomorphism µ is equi- variant. (4) The quotient map η from (G, Ed, αρ) to (G, Od, βρ) is equivariant. Proof. Lemma 3.6(4) implies that the ideal in Ed generated byPd j is invari- ant. Therefore the quotient is a G-algebra. It is well known that we may identify η : Ed → Od with this quotient map. So we have an action βρ : G → Aut(Od). It is clear from the construction and the fact that η(µ0(ej,k, 0)) = µ(ej,k) for j, k = 1, 2, . . . , d that this action satisfies βρ g (sj) = µ(ρ(g), 1)sj for j = 1, 2, . . . , d and g ∈ G. Uniqueness of βρ is clear. Similarly, parts (1), (2), and (3) follow from the corresponding formulas in Lemma 3.6. (cid:3) j=1 rjr∗ The algebraic computations we need for equivariant semiprojectivity of quasifree actions are contained in the following lemma. Lemma 3.8. Let G be a topological group. Let d ∈ Z>0, let ρ : G → U (Md) be a unitary representation of G on Cd, and let αρ : G → Aut(Ed) be the quasifree action of Lemma 3.6. Let µ0 : Md ⊕ C → Ed be as in Notation 3.4, and recall (Notation 3.5) the action Ad(ρ ⊕ 1) on Md ⊕ C. Let (G, C, γ) be a unital G-algebra, EQUIVARIANT SEMIPROJECTIVITY 17 and let ϕ : Ed → C be a unital homomorphism such that ϕ ◦ µ0 is equivariant. For g ∈ G, define w(g) = ϕ(αρ g(r1))∗γg(ϕ(r1)). Then: (1) g 7→ w(g) is a continuous function from G to U (C). (2) For j = 1, 2, . . . , d and g ∈ G, we have ϕ(αρ (3) For every g, h ∈ G, we have w(gh) = w(g)γg(w(h)). g(rj ))w(g) = γg(ϕ(rj )). (4) For every g ∈ G, we have kw(g) − 1k =(cid:13)(cid:13)ϕ(αρ (5) If v ∈ U (C) satisfies vγg(v)∗ = w(g) for all g ∈ G, then there is a unique uni- tal homomorphism ψ : Ed → C such that ψ(rj ) = ϕ(rj )v for j = 1, 2, . . . , d. Moreover, ψ is equivariant and ψ ◦ µ0 = ϕ ◦ µ0. g(r1)) − γg(ϕ(r1))(cid:13)(cid:13). (6) If κ : C → D is an equivariant homomorphism from C to some other G- algebra D, and κ ◦ ϕ is equivariant, then κ(w(g)) = 1 for all g ∈ G. Proof. We use the usual notation for matrix units, as in Notation 3.4. We also recall (Lemma 3.6(3)) that µ0 is equivariant. We prove (1) by showing that ϕ(αρ g(r1)) and γg(ϕ(r1)) are isometries with the same range projection. It is clear that both are isometries. The range projections are ϕ(αρ g(r1))ϕ(αρ g(r1))∗ = ϕ(αρ g(r1r∗ and γg(ϕ(r1))γg(ϕ(r1))∗ = γg(ϕ(r1r∗ 1)) = (ϕ ◦ µ0)(cid:0)Ad(ρ ⊕ 1)g(e1,1, 0)(cid:1) 1)) = γg(cid:0)(ϕ ◦ µ0)(e1,1, 0)(cid:1). These are equal because ϕ ◦ µ0 is equivariant. So (1) follows. For (2), we have, using equivariance of both µ0 and ϕ ◦ µ0 at the third step, ϕ(αρ g(rj))wg = ϕ(αρ g(rj ))ϕ(αρ g(r1))∗γg(ϕ(r1)) = (ϕ ◦ αρ = (γg ◦ ϕ ◦ µ0)(ej,1, 0)(γg ◦ ϕ)(r1) = (γg ◦ ϕ)(rj r∗ g ◦ µ0)(ej,1, 0)(γg ◦ ϕ)(r1) 1r1) = (γg ◦ ϕ)(rj ), as desired. For (3), we simplify the notation by defining (3.2) for g ∈ G. Then u(g) is unitary. By the definition of αρ, we have u(g) = (ϕ ◦ µ0)(ρ(g), 1) (3.3) ϕ(αρ g(rj )) = u(g)ϕ(rj ) for g ∈ G and j = 1, 2, . . . , d. By equivariance of ϕ ◦ µ0, we have (3.4) (γg ◦ ϕ ◦ µ0)(x) =(cid:0)Ad(u(g)) ◦ ϕ ◦ µ0(cid:1)(x) for g ∈ G and x ∈ Md ⊕ C. Using (3.3) at the first step, (3.2) and (3.4) at the second step, ϕ(r1r∗ 1r1 = r1, (3.3), and u(g)u(h) = u(gh) at the last step, for g, h ∈ G we get 1) = µ0(e1,1, 0) and (3.4) at the third step, and r1r∗ w(g)γg(w(h)) =(cid:2)ϕ(r1)∗u(g)∗γg(ϕ(r1))(cid:3)γg(cid:0)ϕ(r1)∗u(h)∗γh(ϕ(r1))(cid:1) 1))(cid:2)u(g)u(h)∗u(g)∗(cid:3)γgh(ϕ(r1)) = ϕ(r1)∗u(g)∗γg(ϕ(r1r∗ = ϕ(r1)∗ϕ(r1r∗ 1)u(h)∗u(g)∗γgh(ϕ(r1)) = w(gh). This proves (3). For (4), use the fact that ϕ(αρ g(r1)) is an isometry at the first step and part (2) at the second step to write kw(g) − 1k =(cid:13)(cid:13)ϕ(αρ g(r1))w(g) − ϕ(αρ g(r1))(cid:13)(cid:13) =(cid:13)(cid:13)γg(ϕ(r1)) − ϕ(αρ g(r1))(cid:13)(cid:13). 18 N. CHRISTOPHER PHILLIPS We prove (5). Existence and uniqueness of ψ are true for any unitary v, because the elements ϕ(rj )v are isometries with orthogonal ranges. For j, k = 1, 2, . . . , d, we have (ψ ◦ µ0)(ej,k, 0) = ψ(rj )ψ(r∗ k) =(cid:0)ϕ(rj )v(cid:1)(cid:0)v∗ψ(r∗ k)(cid:1) = (ϕ ◦ µ0)(ej,k, 0). Since also (ψ ◦ µ0)(1) = (ϕ ◦ µ0)(1), it follows that ψ ◦ µ0 = ϕ ◦ µ0. It remains to prove that ψ is equivariant. In the following calculation, we let u(g) be as in (3.2). We use (2) and (3.3) at step 3, and (3.2) and ψ ◦ µ0 = ϕ ◦ µ0 at step 5, to get, for g ∈ G and j = 1, 2, . . . , d, γg(ψ(rj )) = γg(ϕ(rj ))γg(v) = γg(ϕ(rj ))w(g)∗v = u(g)ϕ(rj )v = u(g)ψ(rj ) = ψ(µ0(ρ(g), 1)rj) = ψ(αρ g(rj )). Equivariance of ψ follows. Part (6) is immediate. (cid:3) Lemma 3.8 will be used to produce cocycles which are close to 1. To deal with them, we need results similar to Lemma 2.3 and Lemma 2.4. Lemma 3.9. Let G be a compact group with normalized Haar measure µ. Let (G, A, α) be a unital G-algebra, and let w : G → U (A) be a continuous function such that for all g, h ∈ G we have w(gh) = w(g)αg(w(h)). Suppose r ∈(cid:2)0, 1 5(cid:3), and let v ∈ U (A) satisfy kvαg(v)∗ − w(g)k ≤ r for all g ∈ G. Define z = v exp(cid:18)ZG log(cid:16)v∗α−1 h (cid:0)w(h)∗v(cid:1)(cid:17) dµ(h)(cid:19) . Then z ∈ U (A) and satisfies for all g ∈ G and kzαg(z)∗ − w(g)k ≤ 10r2 kz − vk ≤ 2r. Proof. For every h ∈ G, we have (3.5) (cid:13)(cid:13)v∗α−1 h (cid:0)w(h)∗v(cid:1) − 1(cid:13)(cid:13) =(cid:13)(cid:13)(cid:2)w(h) − vαh(v)∗(cid:3)∗(cid:13)(cid:13) ≤ r. Since r < 1, the logarithm in the formula for v exists, so v is well defined. Moreover, for h ∈ G, so v ∈ U (A) implies z ∈ U (A). Define log(cid:16)v∗α−1 z0 =ZG h (cid:0)w(h)∗v(cid:1)(cid:17) ∈ iAsa h (cid:0)w(h)∗v(cid:1) dµ(h). α−1 Using kvαh(v)∗ − w(h)k ≤ r at the third step, we get Then, making the change of variables h to hg at the first step and using w(hg) = w(h)αh(w(g)) at the second step, for g ∈ G we have kz0 − vk ≤ZG(cid:13)(cid:13)α−1 αg(z0) =ZG h (cid:0)w(h)∗v(cid:1) − v(cid:13)(cid:13) dµ(h) =ZG(cid:13)(cid:13)w(h) − αh(v)v∗(cid:13)(cid:13) dµ(h) ≤ r. h (cid:0)w(h)∗v(cid:1) dµ(h) =ZG h (cid:0)w(h)∗v(cid:1) dµ(h) = w(g)∗z0. w(g)∗α−1 α−1 EQUIVARIANT SEMIPROJECTIVITY 19 Rewriting and using (3.5), we can apply Lemma 2.2 to get z0 = vZG v∗α−1 h (cid:0)w(h)∗v(cid:1) dµ(h) kz − z0k ≤ 5r2 2(1 − 2r) ≤ 5r2. It follows from the equation αg(z0) = w(g)∗z0 that kαg(z) − w(g)∗zk ≤ 10r2. The first estimate in the conclusion follows. Since r ≤ 1 is the second estimate in the conclusion. 5 and we already know kz0 − vk ≤ r, we also get kz − vk ≤ 2r. This (cid:3) Lemma 3.10. Let G be a compact group with normalized Haar measure µ. Let (G, A, α) and (G, B, β) be unital G-algebras, and let κ : A → B be a unital equivari- ant homomorphism. Let w : G → U (A) be a continuous function such that for all g, h ∈ G we have w(gh) = w(g)αg(w(h)). Suppose 0 ≤ r < 1 10 , and let v0 ∈ U (A) satisfy and κ(v0)βg(κ(v0))∗ = κ(w(g)) for all g ∈ G. Inductively define vm ∈ U (A) by (following Lemma 3.9) kv0αg(v0)∗ − w(g)k ≤ r vm+1 = vm exp(cid:18)ZG log(cid:16)v∗ mα−1 h (cid:0)w(h)∗vm(cid:1)(cid:17) dµ(h)(cid:19) . Then for every m ∈ Z>0 the element vm is a well defined unitary in A such that κ(vm) = κ(v). Moreover, v = limm→∞ vm exists and satisfies vαg(v)∗ = w(g) for all g ∈ G, and also kv − v0k ≤ 2r 1 − 10r and κ(v) = κ(v0). Proof. The proof is essentially the same as the proof of Lemma 2.4. One proves by induction that for all m ∈ Z≥0, the element vm is well defined, in U (A), and satisfies κ(vm) = κ(v) and kvm − vm−1k ≤ 2r(10r)m−1, and that for g ∈ G we have kvmαg(vm)∗ − w(g)k ≤ r(10r)m. We omit further details. (cid:3) Theorem 3.11. Let G be a compact group, let d ∈ Z>0, and let ρ : G → U (Md) be a unitary representation of G on Cd. Then the quasifree action αρ : G → Aut(Ed) of Lemma 3.6 is equivariantly semiprojective. Proof. Let µ0 : Md ⊕ C → Ed be as in Notation 3.4. Recall (Notation 3.5) the action Ad(ρ⊕1) on Md ⊕C. Then µ0 is equivariant by Lemma 3.6(3). The action Ad(ρ⊕1) is equivariantly semiprojective by Theorem 2.6. By Lemma 1.4, it therefore it suffices to prove that µ0 is equivariantly conditionally semiprojective in the sense of Definition 1.2. We adopt the notation of Definition 1.2 and Remark 1.3. Thus, assume that λ : Md ⊕ C → C and ϕ : B → C/J are unital equivariant homomorphisms such that κ ◦ λ = ϕ ◦ ω. Since Ed is semiprojective without the group, there exists n0 ∈ Z>0 and a unital homomorphism ν0 : Ed → C/Jn0 such that πn0 ◦ ν0 = ϕ. In particular, πn0 ◦ ν0 ◦ µ0 = πn0 ◦ κn0 ◦ λ. 20 N. CHRISTOPHER PHILLIPS For k ≥ n0 define fk : U (Md ⊕ C) → [0, ∞) by for x ∈ U (Md ⊕ C). The functions fn are continuous, and satisfy fk(x) =(cid:13)(cid:13)πk,n0(cid:0)(ν0 ◦ µ0)(x) − (κn0 ◦ λ)(x)(cid:1)(cid:13)(cid:13). fn0 ≥ fn0+1 ≥ fn0+2 ≥ · · · and fk → 0 pointwise. Since U (Md⊕C) is compact, Dini's Theorem (Proposition 11 in Chapter 9 of [30]) implies that fk → 0 uniformly. Therefore there exists n1 ≥ n0 such that for all x ∈ U (A), we have Proposition 2.5 provides a unitary u ∈ U (C/Jn1) such that πn1 (u) = 1 and for all x ∈ U (Md ⊕ C). Define ν1 : Ed → C/Jn1 by ν1 = Ad(u) ◦ πn1,n0 ◦ ν0. Then πn1 ◦ ν1 = ϕ and κn1 ◦ λ = ν1 ◦ µ0. For k ≥ n1, the functions 2 . (cid:13)(cid:13)πn1,n0(cid:0)(ν0 ◦ µ0)(x) − (κn0 ◦ λ)(x)(cid:1)(cid:13)(cid:13) < 1 u(cid:0)πn1,n0 ◦ ν0 ◦ µ0(cid:1)(x)u∗ =(cid:0)πn1,n0 ◦ κn0 ◦ λ(cid:1)(x) g 7→(cid:13)(cid:13)πk,n1(cid:0)(γ(k) g)(r1)(cid:1)(cid:13)(cid:13) ◦ ν1)(r1) − (ν1 ◦ αρ g are continuous and pointwise nonincreasing as k → ∞. Since πn1 and πn1 ◦ ν1 = ϕ are equivariant, these functions converge pointwise to zero. Another application of Dini's Theorem provides n ≥ n1 such that, with ν = πn,n1 ◦ ν1 and using equivariance of πn,n1 , we have sup g∈G(cid:13)(cid:13)(γ(n) g ◦ ν)(r1) − (ν ◦ αρ 1 20 . g)(r1)(cid:13)(cid:13) < Now let w(g) be as in Lemma 3.8, with C/Jn in place of C and ν in place of ϕ. Then supg∈G kw(g) − 1k < 1 20 by Lemma 3.8(4) and πn(w(g)) = 1 for all g ∈ G by Lemma 3.8(6). Using these facts, Lemma 3.8(1), and the cocycle condition of Lemma 3.8(3), we can apply Lemma 3.10 with v0 = 1 to find v ∈ U (C/Jn) such that πn(v) = 1 and vαg(v)∗ = w(g) for all g ∈ G. Let ψ : Ed → C/Jn be as in Lemma 3.8(5) with this choice of v. Then ψ is equivariant and ψ ◦ µ0 = ν ◦ µ0 by Lemma 3.8(5). Since ν1 ◦ µ0 = κn1 ◦ λ, we get ψ ◦ µ0 = κn ◦ λ. Since πn(v) = 1, we get πn(ψ(rj )) = πn(ν(rj )) = ϕ(rj ) for j = 1, 2, . . . , d. Therefore πn ◦ ψ = ϕ. This completes the proof that µ0 is equivariantly conditionally semiprojective. (cid:3) Corollary 3.12. Let G be a compact group, let d ∈ Z>0, and let ρ : G → U (Md) be a unitary representation of G on Cd. Then the quasifree action βρ : G → Aut(Od) of Lemma 3.7 is equivariantly semiprojective. Proof. By Theorem 3.11 and Lemma 1.4, it suffices to prove that the quotient map η : Ed → Od is equivariantly conditionally semiprojective in the sense of Defini- tion 1.2. Let the notation be as in Definition 1.2, except that the map called ω there is η. Set f = λ(cid:16)1 −Pd j=1 rj r∗ j(cid:17) , which is a projection in C. Then j(cid:19) = 0. κ(f ) = ϕ(cid:18)1 −Xd sjs∗ j=1 Therefore there is n ∈ Z>0 such that kκn(f )k < 1. Since κn(f ) is a projection, this means that κn(f ) = 0, that is, (κn ◦ λ)(cid:16)1 −Pd j=1 rj r∗ j(cid:17) = 0. Therefore there is EQUIVARIANT SEMIPROJECTIVITY 21 ψ : Od → C/Jn such that κn ◦ λ = ψ ◦ η. Since η is surjective, equivariance of ψ follows from equivariance of η and κn ◦ λ. Similarly, from πn ◦ ψ ◦ η = κ ◦ λ = ϕ ◦ η we get πn ◦ ψ = ϕ. (cid:3) Remark 3.13. An important example of a quasifree action is the one coming from the regular representation of a finite group. In this case, one can prove equivariant semiprojectivity without using any of the machinery developed in this section. Let d = card(G). We discuss only Ed, but the result for Od can be treated the same way, or reduced to the result for Ed as in Corollary 3.12. Label the generators of Ed as rg for g ∈ G, and consider the action α : G → Aut(Ed) determined by αg(rh) = rgh for g, h ∈ G. Following the beginning of the proof of Theorem 3.11, we reduce to the situation in which we have a nonequivariant lifting ν : Ed → C/Jn of ϕ and an isometry v1 ∈ C/Jn such that v1v∗ 1 = (κn ◦ µ0)(e1,1, 0) and kπn(v1) − ϕ(r1)k is small. Functional calculus arguments can be used to replace v1 by a nearby partial isometry w1 such that w1w∗ 1 = (κn ◦ µ0)(e1,1, 0) and πn(w1) = ϕ(r1). Then there is a unique homomorphism ψ : Ed → C/Jn such that ψ(rg) = γ(n) (r1), and this homomorphism is easily seen to be an equivariant lifting of ϕ which satisfies κn ◦ λ = ψ ◦ µ0. g Remark 3.14. We describe what happens when d = 1. In this case, Od becomes C(S1) and Ed becomes the C-algebra C∗(s) of the unilateral shift s. Quasifree actions are those that factor through the action of S1 on C(S1) coming from the translation action of S1 on S1, and those that factor through the action of S1 on C∗(s) coming from the automorphisms determined by βζ(s) = ζs for ζ ∈ S1. Thus, for example, we conclude that the translation action of S1 on C(S1) is equivariantly semiprojective. This, however, is easy to prove directly. A unital equivariant homomorphism from C(S1) with translation to C/J with the action γ(∞) is just a unitary u ∈ C/J such that γ(∞) (u) = ζu for all ζ ∈ S1. This unitary can be partially lifted to a unitary v in some C/Jn such that(cid:13)(cid:13)γ(n) for all ζ ∈ S1. To get an exactly equivariant lift w, set ζ (v) − ζv(cid:13)(cid:13) is small ζ a =ZS1 ζ−1γ(n) ζ (v) dζ (using normalized Haar measure on S1), and take w = a(a∗a)−1/2. 4. Quasifree actions on O∞ The purpose of this section is to prove that quasifree actions of finite groups on the Cuntz algebras O∞ are equivariantly semiprojective. We begin with a discussion of quasifree actions on O∞. We will need to include a point of view different from that of Lemma 3.6 and Lemma 3.7, primarily to take advantage of the KK-theory computations in [27]. Notation 4.1. For d ∈ Z>0, we make Cd into a Hilbert space in the standard way. For d = ∞, we take Cd = l2(Z>0). We let δj ∈ Cd (for j = 1, 2, . . . , d or for d ∈ Z>0) denote the jth standard basis vector, and we let Ud be the group of unitary operators on Cd, with the strong operator topology. For convenience, we also define E∞ = O∞, and denote its generators by r(∞) as well as by sj. j Of course, when d is finite, the topology on Ud is the same as the norm topology. We warn that the notation C∞ conflicts with notation often used for the product 22 N. CHRISTOPHER PHILLIPS or the algebraic direct sum (we are using the Hilbert direct sum), and that U∞ conflicts with notation sometimes used for the (much smaller) algebraic direct limit of the groups Ud. We summarize various results from [27], and relate them to the viewpoint of Lemma 3.6 and Lemma 3.7. For a C-algebra A, a Hilbert A -- A bimodule F is as described at the beginning of Section 1 of [27]: F is a right Hilbert A-module, with A-valued scalar product which is conjugate linear in the first variable, together with an injective homomorphism ϕ : A → L(F ). Theorem 4.2 (Pimsner). Let d ∈ Z>0 ∪ {∞}, and follow Notation 4.1. Make Cd into a Hilbert C -- C bimodule Fd, as described above in the obvious way, with ϕ(λ) = λ · 1L(Fd) for λ ∈ C. Let Td be the associated Toeplitz algebra TFd as described in Definition 1.1 of [27], and call its generators Tξ as there. Then: (1) There is a unique continuous action γ(d) : Ud → Aut(Td) which satisfies γ(d) u (Tξ) = Tuξ for all ξ ∈ Cd. (2) There is a unique isomorphism σd : Ed → Td such that σd(cid:0)r(d) all j. (Recall that for d = ∞ this means σ∞ : O∞ → T∞.) j (cid:1) = Tδj for (3) If d < ∞, then σd is equivariant for the action of Ud on Ed gotten by taking ρ = idUd in Lemma 3.6 and the action of part (1). (4) If d = ∞, when u ∈ U∞ is written as a matrix u = (uj,k)∞ j,k=1, we have d ◦ γ(d) (cid:0)σ−1 u ◦ σd(cid:1)(cid:0)r(d) k (cid:1) = dXj=1 uj,kr(d) j for all k ∈ Z>0, with convergence in the norm topology on the right. (5) Let d1 ∈ Z>0 and d2 ∈ Z>0∪{∞} satisfy d1 ≤ d2, and set G = Ud1 ×Ud2−d1. Let G act on Ed1 by projection to the first factor followed by the action on Ed1 corresponding to γ(d), and let G act on Ed2 by the inclusion of G in Ud2 as block diagonal matrices followed by the action on Ed2 corresponding to γ(d). Then the standard inclusion of Ed1 in Ed2 is equivariant. Proof. The group action of (1) is obtained as in Remark 1.2(2) of [27]. In [27], for a general Hilbert bimodule F, only the action on the quotient OF of TF is described, but the same reasoning also gives an action on TF . Continuity of the action is easily checked on the generators Tξ for ξ ∈ Cd, and continuity on the algebra follows by a standard argument. For part (2), relations giving Td as a universal C-algebra are described at the beginning of Section 3 of [27]. By comparing these relations with those for Ed, one sees that the maps σd exist and are are isomorphisms. Part (3) is a computation. For part (4), orthogonality of the ranges of the sj shows that if λ ∈ C∞ satisfies λj = 0 for all but finitely many j ∈ Z>0, then this implies convergence on the right in the formula in (4). The validity of the formula is now a computation like that for part (3). Since for all k ∈ Z>0, λjsj(cid:13)(cid:13)(cid:13) = kλk2. uj,k2 < ∞, j=1 (cid:13)(cid:13)(cid:13)X∞ ∞Xj=1 EQUIVARIANT SEMIPROJECTIVITY 23 Part (5) now follows by comparing the formulas for the actions from parts (3) (cid:3) and (4) with the definitions of σd1 and σd2. Remark 4.3. Following Theorem 4.2(2), we will identify T∞ with E∞ = O∞, and for finite d we will identify Td with Ed. We then write the action γ(d) of Theorem 4.2(1) as an action on Ed. Definition 4.4. Let G be a topological group. A quasifree action of G on O∞ is an action of the form γ(∞) ◦ ρ with γ(∞) as in Remark 4.3 and for some continuous homomorphism ρ : G → U∞ (that is, for some unitary representation ρ of G on l2(Z>0)). Remark 4.5. Theorem 4.2(3) implies that for d ∈ Z>0, an action of a topological group G on Ed is quasifree in the sense of Lemma 3.6 if and only if it factors through γ(d) : Ud → Aut(Ed) in a similar way. Proposition 4.6. Let G be a topological group, and let ρ1, ρ2 : G → U∞ be uni- tarily equivalent representations. Then the corresponding quasifree actions of G on O∞ are conjugate. Proof. This is immediate from parts (1) and (2) of Theorem 4.2. (cid:3) Theorem 4.7 (Pimsner). Let d ∈ Z>0 ∪ {∞}, let G be a second countable locally compact group, and let ρ : G → Ud be a unitary representation. With the action of G on Ed as in Remark 4.3, the inclusion of C in Ed via λ 7→ λ · 1 is a KK G- equivalence. Proof. See Theorem 4.4 and Remark 4.10(2) in [27]. (cid:3) Definition 4.8. Let G be a topological group, and let ρ : G → U∞ be a unitary representation of G on l2(Z>0). We say that ρ is filtered if there are d(1) < d(2) < · · · in Z>0 such that for each k, the projection pk on the span of the first d(k) standard basis vectors in l2(Z>0) is G-invariant. We call the d(k)-dimensional representations ρk given by g 7→ ρ(g)pkl2(Z>0) the filtering representations. (They are, of course, not uniquely determined by ρ.) We say that a quasifree action of G on O∞ filtered if the corresponding repre- sentation ρ as in Definition 4.4 is filtered. Remark 4.9. Let G be a topological group, and let α : G → Aut(O∞) be a quasifree action of G on O∞ coming from a filtered representation ρ : G → U∞. Let the notation for a sequence of filtering representations be as in Definition 4.8. Let α(n) : G → Aut(Ed(n)) be the quasifree action of Lemma 3.6 coming from the representation g 7→ ρ(g)pkl2(Z>0). Then O∞ is the equivariant direct limit lim Ed(n). −→ This follows from Theorem 4.2 (using all its parts). The following special version of a filtered representation is introduced for tech- nical convenience. Definition 4.10. Let G be a topological group, and let ρ : G → U∞ be a filtered unitary representation of G on l2(Z>0). We say that a collection (ρk)n∈Z>0 of filter- ing representations is almost even if there exist N0, N ∈ Z>0 and representations σ0 : G → UN0 and σ : G → UN such that, following the notation of Definition 4.8, (1) d(n) = N0 + nN for all n ∈ Z>0. (2) ρ1 = σ0 ⊕ σ. 24 N. CHRISTOPHER PHILLIPS (3) ρn+1 = ρn ⊕ σ for all n ∈ Z>0. It is important that we have equality in parts (2) and (3) of Definition 4.10, not merely unitary equivalence. Lemma 4.11. Let G be a compact group, and let α : G → Aut(O∞) be a quasifree action of G on O∞. Then: (1) The action α is conjugate to a filtered quasifree action. (2) If G is in fact finite, then α is conjugate to the quasifree action coming from a representation with an almost even filtration. Part (2) can fail if the group is not finite. The regular representation of a second countable infinite compact group does not have an almost even filtration. Proof of Lemma 4.11. For both parts, we use Proposition 4.6. Part (1) is immediate from the fact that every unitary representation of a com- pact group is a direct sum of finite dimensional representations. For part (2), we need to show that every representation π : G → U∞ is unitarily equivalent to a representation ρ with an almost even filtration. Let τ1, τ2, . . . , τl be a set of representatives of the unitary equivalence classes of irreducible representa- tions of G. We may assume that τ1, τ2, . . . , τl0 occur in π with finite multiplicities m1, m2, . . . , ml0 ∈ Z≥0, and that τl0+1, τl0+2, . . . , τl occur with infinite multiplicity. Then l0 < l. Take σ0 = mk · τk, σ = l0Mk=1 This completes the proof. lMk=l0+1 τk, and ρ = σ0 ⊕ σ ⊕ σ ⊕ · · · . (cid:3) Proposition 4.12. Let G be a topological group, and let ρ : G → L(l2(Z>0)) be an injective filtered unitary representation of G. Then the corresponding quasifree action α : G → Aut(O∞) is pointwise outer, that is, αg is outer for all g ∈ G \ {1}. Proof. Adopt the notation of Definition 4.8. Also let δ1, δ2, . . . be the standard basis vectors of l2(Z>0). Let g ∈ G \ {1}; we will show that αg is outer. Choose k so large that ρ(g)pkl2(Z>0) is nontrivial. Replacing d(1), d(2), . . . by d(k), d(k +1), . . . , we may assume that k = 1. Since ρ(g)p1l2(Z>0) is unitary and nontrivial, and since p1l2(Z>0) is finite dimensional, there exists a unitary u ∈ L(l2(Z>0)), of the form u = u0 + (1 − p1) with u0 a unitary in L(p1l2(Z>0)), such that δ1 is an eigenvector of uρ(g)u∗ with eigenvalue ζ 6= 1. Let σ : G → L(l2(Z>0)) be the representation σ(g) = uρ(g)u∗, and let β : G → Aut(O∞) be the corresponding quasifree action. It follows from Proposition 4.6 that β is conjugate to α. Therefore it suffices to show that βg is outer. Note that βg(s1) = ζs1. We follow the proof of Theorem 4 of [9]. Suppose βg is inner, and let v ∈ O∞ be a unitary such that βg = Ad(v). Define f : Z>0 → Z>0 × Z>0 by f (j, l) = 2j−1(2l − 1) for j, l ∈ Z>0. Define isometries tj ∈ L(l2(Z>0)) by tjδl = δf (j,l) for j, l ∈ Z>0. Since f is injective, there is a unital representation π : O∞ → L(l2(Z>0)) such that π(sj) = tj for all j ∈ Z>0. Since π(v)δ1 ∈ l2(Z>0) and has norm 1, we can write k=1 λk2 = 1. Computations similar to those in the π(v)δ1 = P∞ k=1 λkδk with P∞ proof of Theorem 4 of [9] show that ∞Xk=1 λkδk = π(cid:0)βg(s1)(cid:1)π(v)δ1 = ζλkδ2k−1. ∞Xk=1 EQUIVARIANT SEMIPROJECTIVITY 25 Compare coefficients. For k = 1, we get λ1 = 0 since ζ 6= 1. For k > 1, we get contradiction. λk = ζ−1λ2k−1 = ζ−2λ2(2k−1)−1 = · · · . l=1 λl2 < ∞, this implies λk = 0. But then π(v)δ1 = 0, a (cid:3) Since ζ = 1 and P∞ quasifree action. Then (O∞)G is purely infinite and simple, and K1(cid:0)(O∞)G(cid:1) = 0. Lemma 4.13. Let G be a topological group, let ρ : G → L(l2(Z>0)) be an injective filtered unitary representation of G, and let α : G → Aut(O∞) be the corresponding Proof. Since α is pointwise outer (Proposition 4.12), it follows from Theorem 3.1 of [19] that C∗(G, O∞, α) is simple, and from Corollary 4.6 of [16] that C∗(G, O∞, α) is purely infinite. The Proposition in [29] and its proof imply that (O∞)G is iso- morphic to a corner in C∗(G, O∞, α), necessarily full. Therefore (O∞)G is purely infinite and simple. Theorem 4.7 implies that K G 1 (O∞) = 0. From [17] or Theorem 2.8.3(7) of [22], we get K1(cid:0)C∗(G, O∞, α)(cid:1) = 0. Since (O∞)G is a full corner in C∗(G, O∞, α), it follows from Proposition 1.2 of [21] that K1(cid:0)(O∞)G(cid:1) = 0. Lemma 4.14. Let G be a finite group. Let ρ : G → U∞ be an injective representa- tion with an almost even filtration, for which we use the notation of Definition 4.10, and let α(n) : G → Aut(Ed(n)) be as in Remark 4.9. For m ∈ Z>0, set (cid:3) em = N0+mNXj=N0+(m−1)N +1 sjs∗ j . Then there exists M ∈ Z>0 such that for all n ≥ M, there are two isometries in en(Ed(n))Gen with orthogonal ranges. Proof. Let α : G → Aut(O∞) be the corresponding quasifree action of G on O∞. Following Remark 4.9, we regard Ed(n) as a subalgebra of O∞. Lemma 4.13 implies that e2(O∞)Ge2 is purely infinite and simple. It follows from Lemma 1.7 that e1(O∞)Ge1 = e1(Ed(n))Ge1. ∞[n=0 Therefore there is M ∈ Z>0 such that there are isometries t1, t2 ∈ e1(Ed(M))Ge1 with orthogonal ranges. Now let n ≥ M. Recall from Definition 4.10 that ρn is the direct sum of σ0 and n copies of σ. Let u ∈ UN0+mN be the permutation unitary which exchanges the first and last copies of σ. Then u commutes with ρn(g) for all g ∈ G. Applying Lemma 3.6 to the group Z×G, we see that u induces a quasifree automorphism ψ of Ed(n) which commutes with the action α(n). Moreover, ψ(e1) = en. Since Ed(M) ⊂ Ed(n), the elements ψ(t1) and ψ(t2) are defined and are G-invariant isometries in en(Ed(n))Gen with orthogonal ranges. (cid:3) The following result is the equivariant analog of (a special case of) Lemma 3.3 of [5]. Our statement is more abstract; the concrete version, analogous to that given in [5], is rather long. Lemma 4.15. Let G be a finite group, let ρ : G → U∞ be an injective representa- tion with an almost even filtration, and let α : G → Aut(O∞) be the corresponding 26 N. CHRISTOPHER PHILLIPS quasifree action of G on O∞. Let the notation be as in Definition 4.10 and Re- mark 4.9. In particular, O∞ = lim −→n Ed(n); call the maps of the system ιn,m : Ed(m) → Ed(n) and ι∞,m : Ed(m) → O∞. Let (G, A, α) be a unital G-algebra, and let π : A → O∞ be a surjective equivariant homomorphism. Then there exists M ∈ Z>0 such that for all n ≥ M, the following holds. Let ϕ : Ed(n) → A be a unital equivariant homomorphism such that π ◦ ϕ = ι∞,n. Then there exists a unital equivariant homomorphism ψ : Ed(n+1) → A such that π ◦ ψ = ι∞, n+1 and ψ ◦ ιn+1, n−1 = ϕ ◦ ιn, n−1. Here is the diagram: Ed(n−1) ιn+1, n−1 9sssssssss %❑❑❑❑❑❑❑❑❑ ιn, n−1 ❊ ❊ "❊ ❊ ψ ι∞, n+1 ❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯❯ 4✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐ <②②②②②②②②② ι∞, n A ϕ π / O∞. Ed(n+1) ❊ Ed(n) The solid arrows are given, and ψ is supposed to exist which makes the diagram commute. Proof of Lemma 4.15. We use the names r(d) in Notation 3.2 for the generators of Ed, and we denote the standard generators of O∞ by s1, s2, . . . . Also recall that d(m) = N0 + mN for m ∈ Z>0. Choose M as in Lemma 4.14. j j=1 sjs∗ j , which is the projection in O∞ associated with the rep- resentation σ0 : G → UN0. For m ∈ Z>0, set Define e0 =PN0 em = N0+mNXj=N0+(m−1)N +1 sjs∗ j and qm = N0+mNXj=1 sjs∗ j . Thus, em is the projection in O∞ associated with the mth copy of σ in the direct sum decomposition ρ = σ0 ⊕ σ ⊕ σ ⊕ · · · , k=0 ek is similarly associated with ρm. and qm =Pm For k, l ∈ Z>0, define One easily checks that ck,l = NXj=1 sd(k−1)+j(sd(l−1)+j)∗. ck,lcl,k = ek and ck,l = c∗ l,k for k, l ∈ Z>0. We claim that ck,l is G-invariant. To prove the claim, for m ∈ Z>0 let j,k=1 be the standard system of matrix units in Md(m), and let µm : Md(m) ⊕ C → Ed(m) be the homomorphism called µ0 in Notation 3.4. Recall that Ed(m) (cid:0)e(m) j,k (cid:1)d(m) " * 9 % / < 4 EQUIVARIANT SEMIPROJECTIVITY 27 has the action α(m) = αρm , and equip Md(m) ⊕ C with the action Ad(ρm ⊕ 1) (Notation 3.5). Then µm is equivariant by Lemma 3.6(3). Now let g ∈ G. Set m = max(k, l) and w = ed(k−1)+j, d(l−1)+j ∈ Md(m) ⊕ C. Then w is G-invariant since it is a partial isometry which intertwines the kth and lth copies of σ in the direct sum decomposition of ρm. Therefore ck,l = (ι∞,m ◦ µm)(w) is also G-invariant. The claim is proved. NXj=1 Let n ∈ Z>0 satisfy n ≥ M. By the choice of M using Lemma 4.14, there exist isometries t1, t2 ∈ en(Ed(n))Gen with orthogonal ranges. Define partial isometries in (O∞)G by v1 = ι∞,n(t1)∗ and v2 = cn+1, nι∞,n(t2)∗. (For G-invariance of v2, use the claim above.) We now follow the proof of Lemma 3.3 of [5]. One checks that v1v∗ 1 = en, v∗ 1 v1 = t1t∗ 1, v2v∗ 2 = en+1, and v∗ 2 v2 = t2t∗ 2. Thus, qn−1, v∗ 1v1, v∗ 2 v2 and qn−1, v1v∗ 1 , v2v∗ 2 are two sets of mutually orthogonal projections in (O∞)G, and the projections 1 − qn−1 − v∗ 1v1 − v∗ 2 v2 and 1 − qn−1 − v1v∗ 1 − v2v∗ 2 are both nonzero and have the same class in K0(cid:0)(O∞)G(cid:1). Therefore, by Lemma 4.13, we can find v3 ∈ (O∞)G such that 1 v1 − v∗ v∗ 3v3 = 1 − qn−1 − v∗ and v3v∗ 3 = 1 − qn−1 − v1v∗ 1 − v2v∗ 2 . 2v2 Set w = v1 + v2 + v3 and v = q + v1 + v2 + v3. Then w is a unitary in (1 − qn−1)(O∞)G(1 − qn−1). Define p = d(n−1)Xj=1 ϕ(cid:0)r(d(n)) j (cid:1)ϕ(cid:0)r(d(n)) j (cid:1)∗ , which is a G-invariant projection in A such that π(p) = qn−1. Proposition 1.2 of [21] and Lemma 4.13 imply that K1(cid:0)(1 − qn−1)(O∞)G(1 − qn−1)(cid:1) = 0. The map AG → (O∞)G is surjective by Lemma 1.6. So there exists a unitary y ∈ (1 − p)AG(1 − p) such that π(y) = w. Set u = p + y, which is a unitary in AG such that π(u) = v. Theorem 1.9 of [7] now implies that U(cid:0)(1 − qn−1)(O∞)G(1 − qn−1)(cid:1) is connected. (cid:1) for j = 1, 2, . . . , d(n − 1). It is then easy to We have uϕ(cid:0)r(d(n)) (cid:1) ψ(cid:0)r(d(n+1)) check that there is a unital homomorphism ψ : Ed(n+1) → A satisfying (cid:1) = ϕ(cid:0)r(d(n)) ϕ(cid:0)r(d(n)) uϕ(t1)ϕ(cid:0)r(d(n)) uϕ(t2)ϕ(cid:0)r(d(n)) (cid:1) = (cid:1) d(n − 1) + 1 ≤ j ≤ d(n) j−N (cid:1) d(n) + 1 ≤ j ≤ d(n + 1). Clearly ψ ◦ ιn+1, n−1 = ϕ ◦ ιn+1, n−1. j ≤ d(n − 1) j j j j j 28 N. CHRISTOPHER PHILLIPS For j = 1, 2, . . . , d(n), it is easily checked that For j = d(n) + 1, d(n) + 2, . . . , d(n + 1), we have (π ◦ ψ)(cid:0)r(d(n+1)) j (cid:1) = ι∞, n+1(cid:0)r(d(n+1)) j (cid:1). π(cid:0)uϕ(t2)ϕ(cid:0)r(d(n)) j−N (cid:1)(cid:1) = vι∞,n(t2)sj−N = cn+1, nι∞,n(t∗ It follows that π ◦ ψ = ι∞, n+1. 2t2)sj−N = cn+1, nsj−N = sj. To finish the proof, we must check that ψ is equivariant. It is enough to check equivariance on the generators. Since u is G-invariant and ϕ is equivariant, it is enough to check equivariance of the homomorphism ψ0 : Ed(n+1) → Ed(n) deter- mined by ψ0(cid:0)r(d(n+1)) j Define b, f ∈ E(d(n+1)) by (cid:1) = r(d(n)) j t1r(d(n)) t2r(d(n)) j−N j j ≤ d(n − 1) d(n − 1) + 1 ≤ j ≤ d(n) d(n) + 1 ≤ j ≤ d(n + 1). b = NXj=1 r(d(n+1)) d(n)+j (cid:0)r(d(n+1)) d(n−1)+j(cid:1)∗ and f = d(n−1)Xj=1 r(d(n+1)) j j (cid:0)r(d(n+1)) (cid:1)∗ . Then ι∞, n+1(b) = cn+1, n and ι∞, n+1 is injective and equivariant, so b is G- invariant. Similarly, ι∞, n+1(f ) = qn−1, so f is G-invariant. Also, br(d(n+1)) = r(d(n+1)) j for j = d(n) + 1, d(n) + 2, . . . , d(n + 1). Thus j−N (ιn+1, n ◦ ψ0)(cid:0)r(d(n+1)) j (cid:1) = (f + t1 + t2b)r(d(n+1)) j for j = 1, 2, . . . , d(n + 1). Since f + t1 + t2b is G-invariant, and since ιn+1, n is injective and equivariant, it follows that ψ0 is equivariant, as desired. This completes the proof. (cid:3) Theorem 4.16. Let G be a finite group. Let α : G → Aut(O∞) be a quasifree action. Then α is equivariantly semiprojective. Proof. We follow the proof of Theorem 3.2 of [5]. Let ρ : G → U∞ be the repre- sentation which gives rise to α. Using Lemma 1.8, we may reduce to the case in which ρ is injective. By Lemma 4.11(2), we may assume that ρ has an almost even filtration as in Definition 4.10. Let the notation be as in Lemma 4.15, and choose M as there. is the quotient map. We follow the notation in Remark 1.3(3): C is a unital G-algebra with an increas- n=1 Jn. The map πn : C/Jn → C/J ing sequence of invariant ideals Jn, and J =S∞ Let ϕ : O∞ → C/J be a unital equivariant homomorphism. First suppose that ϕ is an isomorphism. From Theorem 3.11 we get n ∈ Z>0 and a unital equivariant homomorphism ψM : Ed(M) → C/Jn such that πn ◦ ψM = ϕ ◦ ι∞,M . Applying Lemma 4.15 to π = ϕ−1 ◦ πn : C/Jn → O∞, for m ≥ M we inductively construct unital equivariant homomorphisms ψm : Ed(m) → C/Jn such that π ◦ ψm+1 = ϕ ◦ ι∞,m+1 and ψm+1 ◦ ιm+1, m−1 = ψm ◦ ιm,m−1. EQUIVARIANT SEMIPROJECTIVITY 29 j (cid:1) exists for all g ∈ G and j ∈ Z>0, because when (cid:1). So there is a unital equivariant homomor- j phism ψ : O∞ → C/Jn such that ψ(sj) = rj for all j ∈ Z>0. Clearly π ◦ ψ = ϕ. Then rj = limm→∞ ψm(cid:0)rd(m) d(m) ≥ j it is equal to ψm+2(cid:0)rd(m+2) of Q, set In = D ∩ Jn for n ∈ Z>0, and set I = D ∩ J. Then I =S∞ For the general case, set Q = ϕ(O∞) ⊂ C/J, let D ⊂ C be the inverse image n=1 In. (In [20], see Proposition 13.1.4, Lemma 13.1.5, and the discussion afterwards.) So Q = D/I. Since O∞ is simple, the corestriction ϕ0 : O∞ → D/I of ϕ is an isomorphism. The result follows by applying the special case above with D in place of C, with In in place of Jn, and with ϕ0 in place of ϕ. (cid:3) Problem 4.17. Let G be an infinite compact group. Is a quasifree action of G on O∞ necessarily equivariantly semiprojective? As a test case, consider the quasifree action coming from the left regular repre- sentation of S1. 5. Equivariantly stable relations We relate equivariant semiprojectivity to equivariant stability of relations be- cause, in the applications we have in mind [25], equivariant stability of relations is what we actually use. Weak stability of relations (Definition 4.1.1 of [20]) also has an equivariant ver- sion. Since equivariant stability holds for the examples we care about, we only consider equivariant stability. We follow Section 13.2 of [20] for our definition of generators and relations. For reference, we give the version of the definition without the group action, except that we give a version for unital C-algebras. This is a variant of Definition 13.2.1 of [20]. Definition 5.1. Let S be a set. We denote by FS the universal unital C-algebra generated by the elements of S subject to the relations ksk ≤ 2 for all s ∈ S. A set of relations on S is a subset R ⊂ FS. We refer to (S, R) as a set of generators and relations. We say that (S, R) is finite if S and R are finite. We define IR ⊂ FS to be the ideal in FS generated by R. Since we are asking for unital algebras and homomorphisms, we make the fol- lowing definition. Definition 5.2. A set (S, R) of generators and relations as in Definition 5.1 is admissible if IR 6= FS. When (S, R) is admissible, we let τR : FS → FS/IR be the quotient map. The C-algebra on the generators and relations (S, R), which we write C∗(S, R), is by definition FS/IR. We say that (S, R) is bounded if for every s ∈ S, we have kτR(s)k ≤ 1. The choices ksk ≤ 2 and kτR(s)k ≤ 1 are convenient normalizations. By scaling, every set of generators and relations can be fit in this framework. The following is essentially Definition 13.2.2 of [20], but for the unital situation. By convention, we declare (except in a few places where we explicitly allow it) that the zero C-algebra is not unital. Definition 5.3. Let the notation be as in Definition 5.1, let A be a unital C- algebra, and let ρ : S → A be a function such that kρ(s)k ≤ 2 for all s ∈ S. In 30 N. CHRISTOPHER PHILLIPS this situation, we write ϕρ : FS → A for the corresponding homomorphism. We say that ρ is a representation of (S, R) in A if ϕρ(x) = 0 for all x ∈ R. For δ ∈ [0, 1), we say that ρ is a δ-representation of (S, R) in A if kϕρ(x)k ≤ δ for all x ∈ R. (Sometimes, we will also allow the map to the zero C-algebra as a representation.) If (S, R) is admissible, then the universal representation ρR is obtained by taking A = C∗(S, R) and ρR = τRS. Remark 5.4. It is clear that the universal representation, as defined above, really has the appropriate universal property. Lemma 5.5. Let (S, R) be a set of generators and relations as in Definition 5.1. Then (S, R) is admissible if and only if there exists a representation in a (nonzero) unital C-algebra. Proof. This is immediate. (cid:3) Remark 5.6. We make some general remarks. (1) The relation corresponding to an element x ∈ FS is really just the statement x = 0. Here x could be any -polynomial in the noncommuting variables S, but in fact we are allowing arbitrary elements of the C-algebra FS. The framework we describe in fact allows much more general relations. For example, suppose R0 ⊂ FS, M : R0 → [0, ∞) is a function, and we want the relations to say kxk ≤ M (x) for all x ∈ R0. We simply take the intersection I ⊂ FS of the kernels of all unital homomorphisms ϕ : FS → A, for arbitrary unital C-algebras A, such that kϕ(x)k ≤ M (x) for all x ∈ R0. Then we take as relations all elements of I, that is, we take R = I. Positivity conditions on elements of FS can be handled the same way. (2) If S is countable, we may always take R to be finite. Choose a countable subset {x1, x2, . . .} of the unit ball of IR whose span is dense in IR. Then we can take the relations to consist of the single element a = 2−nx∗ nxn. ∞Xn=1 (This change does, however, change the meaning of a δ-representation.) (3) It follows from (1) and (2) that if (S, R) is finite and bounded, and δ ∈ [0, 1), then the universal C-algebra generated by a δ-representation of (R, S) is again the universal C-algebra on a finite and bounded set of generators and relations. (4) We have made a choice in the definition of a δ-representation: we still require kρ(s)k ≤ 1 for all s ∈ S. By suitable scaling and application of (1) above, it is also possible to get a version in which we merely require kρ(s)k ≤ 1 + δ for all s ∈ S. We now give equivariant versions of these definitions. We restrict to discrete groups, and to finite groups in practice. If G is not discrete, but the universal C-algebra is supposed to carry a continuous action of G, then the relations must demand that the action of G on each generator defines a continuous function from G to the universal C-algebra. There are many kinds of conditions on elements of a C-algebra which can be made into relations which determine a universal C- algebra, but continuity of functions from the set of generators isn't one of them. The universal algebra will in general only be an inverse limit of C-algebras. See EQUIVARIANT SEMIPROJECTIVITY 31 Definition 1.3.4 and Proposition 1.3.6 of [23]. There do exist examples of universal G-algebras on generators and relations when G is not discrete. See Example 5.18 and Example 5.19 below. However, we leave the development of the appropriate theory for elsewhere. Notation 5.7. Let S be a set, let G be a discrete group, and let σ be an action of G on S, written (g, s) 7→ σg(s). We denote by µσ the action of G on FS induced by σ. Definition 5.8. Let G be a discrete group. A G-equivariant set of generators and relations is a triple (S, σ, R) in which (S, R) is a set of generators and relations as in Definition 5.1, σ is an action of G on S (just as a set), and R is invariant under the action µσ of Notation 5.7. We say that (S, σ, R) is admissible if (S, R) is admissible in the sense of Definition 5.2. We say that (S, σ, R) is bounded if (S, R) is, and is finite if G and (S, R) are finite. It may seem better to omit σ and the requirement of G-invariance, and to allow the group action in the relations. We address this formulation starting with Defi- nition 5.13 below. However, doing so does not give anything new, and the version we have given above is technically more convenient. Definition 5.9. Let G be a discrete group. Let (S, σ, R) be a G-equivariant set of generators and relations in the sense of Definition 5.8. Let α : G → Aut(A) be an action of G on a unital C-algebra A. An equivariant representation of (S, σ, R) in A is a representation of (S, R) in the sense of Definition 5.3 such that for ev- ery g ∈ G and s ∈ S, we have ρ(σg(s)) = αg(ρ(s)). For δ1, δ2 ∈ [0, 1), a δ1- equivariant δ2-representation of (S, σ, R) in A is a δ2-representation ρ of (S, R) such that kρ(σg(s)) − αg(ρ(s))k ≤ δ1 for all g ∈ G and s ∈ S. When δ1 = 0, we speak of an equivariant δ2-representation of (S, σ, R) in A. If (S, R) is admissible, then the universal equivariant representation ρR is ob- tained by taking A = C∗(S, R), with the action µσ : G → Aut(C∗(S, R)) com- ing from the fact that IR is an invariant ideal for µσ : G → Aut(FS), and taking ρR = τRS. We write C∗(S, σ, R) for the algebra equipped with this action. We show that we have the right definition of admissibility. Lemma 5.10. Let G be a discrete group, and let (S, σ, R) be a G-equivariant set of generators and relations. Then (S, σ, R) is admissible if and only if there exists an equivariant representation in a (nonzero) unital G-algebra. Proof. If there is an equivariant representation, then Lemma 5.5 implies that (S, R) is admissible, so that (S, σ, R) is admissible. For the reverse, since IR 6= FS, the universal equivariant representation of Defi- (cid:3) nition 5.9 is an equivariant representation in a unital G-algebra. The universal equivariant representation, as in Definition 5.9, really is universal. Lemma 5.11. Let G be a discrete group, and let (S, σ, R) be a G-equivariant set of generators and relations. Let α : G → Aut(A) be an action of G on a unital C-algebra A, and let ρ : S → A be an equivariant representation of (S, σ, R) in A. Then there exists a unique equivariant homomorphism ϕ : C∗(S, σ, R) → A such that ϕ ◦ ρR = ρ. 32 N. CHRISTOPHER PHILLIPS Proof. As an algebra, we have C∗(S, σ, R) = C∗(S, R). So Remark 5.4 provides a unique homomorphism ϕ : C∗(S, σ, R) → A such that ϕ ◦ ρR = ρ. Let µσ : G → Aut(C∗(S, σ, R)) be as in Definition 5.9. Since ρ is equivariant, for all g ∈ G and s ∈ S we have ϕ(cid:0)µσ g (τR(s))(cid:1) = ϕ(τR(µσ g (s))) = ρ(σg(s)) = αg(ρ(s)) = αg(ϕ(cid:0)τR(s))). Since τR(S) generates C∗(S, σ, R), equivariance of ϕ follows. (cid:3) We have equivariant analogs of the first two parts of Remark 5.6. Remark 5.12. (1) Let G be a discrete group, let S be a set, and let σ be an action of G on S. For any proper G-invariant ideal I ⊂ FS, we can get FS/I as a universal G-algebra C∗(S, σ, R) simply by taking R = I. As an example, let R ⊂ FS be G-invariant, and let M : R → [0, ∞) be a function such that M (σg(s)) = M (s) for all g ∈ G and s ∈ S. We take I ⊂ FS to be the intersection of the kernels of all unital equivariant homomorphisms ϕ : FS → A, for arbitrary unital G-algebras (G, A, α), such that kϕ(x)k ≤ M (x) for all x ∈ R. (2) If S is countable and G is finite, we always take R to be finite. Choose a countable subset {x1, x2, . . .} of the unit ball of IR whose span is dense in IR. Then we can take the relations to consist of the single G-invariant element a = ∞Xn=1Xg∈G 2−nµσ g (x∗ nxn). (3) If (S, σ, R) is finite and bounded, and δ ∈ [0, 1), then the universal C- algebra generated by an equivariant δ-representation of (S, σ, R) is again the universal C-algebra on a finite and bounded set of generators and relations. However, for δ0 > 0, there is no obvious action of G on the universal C- algebra generated by a δ0-equivariant δ-representation of (S, σ, R). If we want to allow the action of G to appear in the relations, we can use the following alternate definition. We omit the word "equivariant" in the name. Definition 5.13. Let G be a discrete group. A set of generators and relations for a G-algebra is a pair (S, R) in which S is a set and R is a subset of FG×S. Define an action σ of G on G × S by σg(h, s) = (gh, s) for g, h ∈ G and s ∈ S, and let µσ : G → Aut(FG×S) be as in Notation 5.7. The associated G-equivariant set of generators and relations to (S, R) is then (cid:16)G × S, σ, [g∈G g (R)(cid:17) . We let IG,R ⊂ FG×S be the ideal generated bySg∈G µσ g (R). We say that (S, R) is admissible if IG,R 6= FG×S, and in this case we define the universal G-algebra gener- ated by (S, R) to be C∗(S, R) = FG×S/IG,R, with the action µ : G → Aut(C∗(S, R)) induced by the action µσ : G → Aut(FG×S). Let τG,R : FG×S → C∗(S, R) be quo- tient map. We say that (S, R) is bounded if for every s ∈ S, we have kτG,R(1, s)k ≤ 1. We say that (S, R) is finite if G, S, and R are all finite. µσ Definition 5.14. Let G be a discrete group, and let (S, R) be a set of generators and relations for a G-algebra in the sense of Definition 5.13. Let α : G → Aut(A) be an action of G on a unital C-algebra A. A representation of (S, R) in A is a function EQUIVARIANT SEMIPROJECTIVITY 33 ρ : S → A such that the function π : G × S → A, defined by π(g, s) = αg(ρ(s)) for g ∈ G and s ∈ S, is an equivariant representation, in the sense of Definition 5.9, of the associated G-equivariant set of generators and relations. For δ ∈ [0, 1), we say that ρ is a δ-representation of (S, R) in A if, using the notation of Definition 5.3, we have kϕπ(x)k ≤ δ for all x ∈ R. Remark 5.15. Let G be a discrete group, let (S, R) be a set of generators and relations for a G-algebra in the sense of Definition 5.13, and let the notation be as there. Set Q =Sg∈G µσ of Definition 5.8. g (R). Then: (1) (S, R) is admissible if and only if (G × S, σ, Q) is admissible in the sense (2) (S, R) is bounded if and only if (G × S, σ, Q) is bounded in the sense of Definition 5.8. (Use the fact that for g ∈ G and s ∈ S, we have τG,R(g, s) = µg(τG,R(1, s)).) (3) (S, R) is finite if and only if (G × S, σ, Q) is finite in the sense of Defini- tion 5.8. (4) There is a unique equivariant isomorphism ψ : C∗(S, R) → C∗(G × S, σ, Q) such that ψ(τG,R(g, s)) = τR(g, s) for all g ∈ G and s ∈ S. We then get the following universal property for C∗(S, R). The proof is clear, and is omitted. Lemma 5.16. Let G be a discrete group, and let (S, R) be a set of generators and relations for a G-algebra in the sense of Definition 5.13. Let α : G → Aut(A) be an action of G on a unital C-algebra A, and let ρ : S → A be a representation of (S, R) in A. Then there exists a unique equivariant homomorphism ϕ : C∗(S, R) → A such that ϕ ◦ ρR = ρ. Remark 5.17. Analogously to Remark 5.6(1) and Remark 5.12(1), we can now speak of the universal G-algebra generated by a set S with relations given by norm bounds and positivity conditions on -polynomials in the noncommuting variables `g∈G σg(S), that is, polynomials in the noncommuting variables consisting of the generators, their formal adjoints, and the formal images of all these under an action of G. We now present examples to show that there are some cases in which there is a reasonable universal C-algebra, with continuous action of G, even with G not discrete. Example 5.18. Let G be any topological group, and let (G, A, α) be any G-algebra. Take the generating set S to be the closed unit ball of A, take R to be the collection of all algebraic relations that hold among elements of S and their adjoints, and take σ = αS. Then the universal C-algebra generated by (S, σ, R) is just A, with the representation being the identity map and the action of G being α. The algebra A is universal when G is given the discrete topology, but the action is in fact continuous when G is given its original topology. One can make a slightly more interesting example as follows. Example 5.19. Take A = Mn, and take α to be any action of G on Mn. Let (ej,k)n j,k=1 be the standard system of matrix units in Mn. Take the generators to consist of elements vg,j,k for g ∈ G and j, k ∈ {1, 2, . . . , n}. Set σg(vh,j,k) = vgh,j,k. 34 N. CHRISTOPHER PHILLIPS The universal representation is intended to be ρ(vg,j,k) = αg(ej,k). To make this happen, take the relations to say that for each g ∈ G, the collection (vg,j,k)n j,k=1 is a system of matrix units, and also to include, for all g, h ∈ G and j, k ∈ {1, 2, . . . , n}, the relation corresponding to the (unique) expression of αg(αh(ej,k)) as a linear combination of the matrix units αh(el,m). The following is the equivariant analog of Definition 14.1.1 of [20]. Following [20], we restrict to finite sets of generators and relations. Accordingly, we take the group to be finite. Definition 5.20. Let G be a finite group, and let (S, σ, R) be a finite admissible G-equivariant set of generators and relations. Then we say that (S, σ, R) is sta- ble if for every ε > 0 there is δ > 0 such that the following holds. Suppose that (G, A, α) and (G, B, β) are unital G-algebras (except that we allow B = 0), that ω : A → B is an equivariant homomorphism, that ρ0 : S → A is a δ-equivariant δ-representation of (S, σ, R) (in the sense of Definition 5.9), and that ω ◦ ρ0 is an equivariant representation of (S, σ, R). Then there exists an equivariant represen- tation ρ : S → A of (S, σ, R) such that ω ◦ ρ = ω ◦ ρ0 and such that for all s ∈ S we have kρ(s) − ρ0(s)k < ε. We allow B = 0 to incorporate the possibility that we are merely given a δ- equivariant δ-representation of (S, σ, R) but no homomorphism ω such that ω ◦ ρ0 is an equivariant representation. Lemma 5.21. Let G be a finite group, and let (S, σ, R) be a bounded finite ad- missible G-equivariant set of generators and relations. Then for every η > 0 there is δ > 0 such that whenever (G, A, α) and (G, B, β) are unital G-algebras (with possibly B = 0), ω : A → B is equivariant, and ρ0 : S → A is a δ-equivariant δ-representation of (S, σ, R) such that ω ◦ ρ0 is an equivariant representation of (S, σ, R), then there exists an (exactly) equivariant η-representation ρ : S → A such that ω ◦ ρ = ω ◦ ρ0 and kρ(s) − ρ0(s)k < η for all s ∈ S. Proof. Since S and R are finite, there is δ0 > 0 such that whenever C is a C-algebra and ψ1, ψ2 : FS → C are two unital homomorphisms such that kψ1(s) − ψ2(s)k < δ0 for all s ∈ S, then kψ1(r) − ψ2(r)k < 1 2 η for all r ∈ R. Set δ = min(cid:0)δ0, 1 2 η(cid:1). Now let ρ0 be as in the hypotheses. For s ∈ S, define ρ(s) = (αg ◦ ρ0 ◦ σ−1 g )(s). 1 card(G)Xg∈G and, since ρ0 is δ-equivariant, kρ(s) − ρ0(s)k ≤ δ ≤ δ0. Therefore, in the notation 2 η, whence Then ρ is exactly equivariant. Also, for all s ∈ S, we have kρ(s)k ≤ max(cid:0)(cid:8)kρ0(t)k : t ∈ S(cid:9)(cid:1) ≤ 2 of Definition 5.3, for all r ∈ R we have(cid:13)(cid:13)ϕρ(r) − ϕρ0 (r)(cid:13)(cid:13) ≤ 1 kϕρ(r)k ≤ 1 2 η + δ ≤ η. 2 η +(cid:13)(cid:13)ϕρ0 (r)(cid:13)(cid:13) ≤ 1 Thus ρ is an η-representation. From ω ◦ αg ◦ ρ0 ◦ σ−1 g = βg ◦ ω ◦ ρ0 ◦ σ−1 g = ω ◦ ρ0, we get ω ◦ ρ = ω ◦ ρ0, completing the proof. (cid:3) EQUIVARIANT SEMIPROJECTIVITY 35 Theorem 5.22. Let G be a finite group, and let (S, σ, R) be a bounded finite admissible G-equivariant set of generators and relations. Then (S, σ, R) is stable if and only if C∗(S, σ, R) is equivariantly semiprojective. Proof. Proposition 13.2.5 of [20] holds equally well, and with the same proof, for unital algebras, for a bounded finite admissible G-equivariant set (S, σ, R) of gener- ators and relations (with, in particular, G finite), for an equivariant direct system of unital G-algebras with unital maps, and for a δ-equivariant δ-representation of (S, σ, R). Therefore stability of (S, σ, R) implies equivariant semiprojectivity of C∗(S, σ, R). The prooof of the reverse implication roughly follows the proof for the nonequiv- ariant case, as, for example, in the proof of Theorem 14.1.4 of [20]. For n ∈ Z>0 let Jn ⊂ FS be the intersection of the kernels of the homomorphisms ϕρ as ρ runs through all equivariant 2−n-representations of (S, σ, R). Then Jn is a G-invariant ideal in FS, J1 ⊂ J2 ⊂ · · · , and Jn = IR. ∞[n=1 The quotient FS/Jn is the universal G-algebra generated by an equivariant 2−n- representation of (S, σ, R). We will apply the definition of equivariant semiprojec- tivity to C∗(S, σ, R), with C = FS, with Jn as given, with J = IR, and with ϕ = idC ∗(S,σ,R). We use the same names κ : FS → FS/IR, κn : FS → FS/Jn, πn : FS/Jn → FS/IR, etc. for the maps as in Definition 1.1 and Remark 1.3(3). By equivariant semiprojectivity, we can choose n0 ∈ Z>0 and a unital equivariant homomorphism ψ0 : C∗(S, σ, R) → FS/Jn0 such that πn0 ◦ ψ0 = idC ∗(S,σ,R). For s ∈ S we have (πn0 ◦ ψ0 ◦ κ)(s) = (πn0 ◦ ψ0 ◦ πn0 ◦ κn0 )(s) = (πn0 ◦ κn0)(s). Since S is finite, there is n ≥ n0 such that for all s ∈ S we have 2 ε. (cid:13)(cid:13)(πn,n0 ◦ ψ0 ◦ κ)(s) − κn(s)(cid:13)(cid:13) =(cid:13)(cid:13)(πn,n0 ◦ ψ0 ◦ κ)(s) − (πn,n0 ◦ κn0)(s)(cid:13)(cid:13) < 1 2 ε. Define ψ = πn,n0 ◦ ψ0, getting πn ◦ ψ = We may also require that 2−n < 1 idC ∗(S,σ,R) and (5.1) for all s ∈ S. k(ψ ◦ κ)(s) − κn(s)k < 1 2 ε Choose δ > 0 as in Lemma 5.21 for η = 2−n. Let (G, A, α) and (G, B, β) be unital G-algebras (with possibly B = 0), let ω : A → B be equivariant, and let ρ0 : S → A be a δ-equivariant δ-representation of (S, σ, R) such that ω ◦ ρ0 is an equivariant representation of (S, σ, R). By the choice of δ, there is an equivariant 2−n-representation ρ1 : S → A such that ω ◦ ρ1 = ω ◦ ρ0 and (5.2) for all s ∈ S. kρ1(s) − ρ0(s)k < 2−n The following diagram (in which the triangle and the square will be shown to commute, and we already know that πn ◦ κn = κ) shows some of the maps we have 36 N. CHRISTOPHER PHILLIPS or which will be constructed: κ ψ ❘❴❧ κn / FS/Jn FS/IR S FS (◗◗◗◗◗◗◗◗◗◗◗◗◗◗◗◗◗ ρ1 ϕ A πn ω λ / B. By the definition of Jn, there is a unital equivariant homomorphism ϕ : FS/Jn → A such that ϕ(κn(s)) = ρ1(s) for all s ∈ S. Define ρ(s) = (ϕ ◦ ψ ◦ κ)(s) for s ∈ S. Then ρ is an equivariant representation of (S, σ, R). Moreover, there is an equivariant homomorphism λ : FS/IR → B such that λ(κ(s)) = ω(ρ1(s)) for all s ∈ S. By construction, for s ∈ S we have (ω ◦ ϕ ◦ κn)(s) = (ω ◦ ρ1)(s) = (λ ◦ πn ◦ κn)(s). Since κn is surjective and S generates FS, we get ω ◦ ϕ = λ ◦ πn. For s ∈ S we now have (ω ◦ ρ)(s) = (ω ◦ ϕ ◦ ψ ◦ κ)(s) = (λ ◦ πn ◦ ψ ◦ κ)(s) = (λ ◦ κ)(s) = (ω ◦ ρ1)(s) = (ω ◦ ρ0)(s). That is, ω ◦ ρ = ω ◦ ρ0. It remains only to show that kρ(s) − ρ0(s)k < ε for s ∈ S. Using (5.2) and 2−n < 1 2 ε at the second step, we have kρ(s) − ρ0(s)k ≤ kρ(s) − ρ1(s)k + kρ1(s) − ρ0(s)k < kρ(s) − ρ1(s)k + 1 2 ε, and, by (5.1), kρ(s) − ρ1(s)k = k(ϕ ◦ ψ ◦ κ)(s) − (ϕ ◦ κn)(s)k ≤ k(ψ ◦ κ)(s) − κn(s)k < 1 2 ε. The required estimate follows, and the theorem is proved. (cid:3) We now consider the version of stability in which the group action is allowed in the relations. Definition 5.23. Let G be a finite group, and let (S, R) be a finite admissible set of generators and relations for a G-algebra, in the sense of Definition 5.13. We say that (S, R) is stable if for every ε > 0 there is δ > 0 such that the following holds. Suppose that (G, A, α) and (G, B, β) are unital G-algebras (except that we allow B = 0), that ω : A → B is an equivariant homomorphism, that ρ0 : S → A is a δ-representation of (S, R) (in the sense of Definition 5.14), and that ω ◦ ρ0 is a representation of (S, R). Then there exists a representation ρ : S → A of (S, R) such that ω ◦ ρ = ω ◦ ρ0 and such that for all s ∈ S we have kρ(s) − ρ0(s)k < ε. Lemma 5.24. Let G be a finite group, and let (S, R) be a finite bounded admissible set of generators and relations for a G-algebra in the sense of Definition 5.13. Let g (R), so that the the action σ of G on G × S be as there, and set Q = Sg∈G µσ associated G-equivariant set of generators and relations is (G × S, σ, Q). Then: (1) For every η > 0 there is δ > 0 such that whenever (G, A, α) is a unital G-algebra and λ : G × S → A is a δ-equivariant δ-representation of (G × S, σ, Q), then the function s 7→ λ(1, s) is an η-representation of (S, R). / / ( / / / v v / EQUIVARIANT SEMIPROJECTIVITY 37 (2) For every η > 0 there is δ > 0 such that whenever (G, A, α) is a unital G-algebra and ρ : S → A is a δ-representation of (S, R), then the function (g, s) 7→ αg(ρ(s)) is an equivariant η-representation of (G × S, σ, Q). Proof. We prove part (1). Suppose the conclusion fails. Apply Definition 5.14 and use finiteness of R to find x ∈ R, η > 0, and for each n ∈ Z>0 a unital G-algebra n -representation λn : G × S → An such that, if g (ρn(s)) for g ∈ G and s ∈ S, we define ρn(s) = λn(1, s) for s ∈ S and πn(g, s) = α(n) then, following the notation of Definition 5.3, we have kϕπn(x)k > η. n -equivariant 1 n=1 An be the C-algebraic product (the set of sequences (an)n∈Z>0 in the (cid:0)G, An, α(n)(cid:1) and a 1 LetQ∞ algebraic product such that supn∈Z>0 kank is finite), and define A = An, ∞Mn=1 ∞Yn=1 An. The obvious coordinatewise definitions, followed by the quotient map, give an action α : G → Aut(A) and functions λ : G × S → A, ρ : S → A, and π : G × S → A. One checks that λ is an equivariant representation of (G × S, σ, Q). Clearly ρ(s) = λ(1, s) for s ∈ S and π(g, s) = αg(ρ(s)) for g ∈ G and s ∈ S. Therefore π = λ. Since x ∈ Q, we have ϕπ(x) = 0. This contradicts the fact that kϕπn(x)k > η for all n ∈ Z>0. Part (1) is proved. and for each n ∈ Z>0 a unital G-algebra (cid:0)G, An, α(n)(cid:1) and a 1 Now suppose part (2) is false. Since Q is finite, there exist x ∈ Q, η > 0, n -representation ρn : S → An such that, if we define πn(g, s) = α(n) (ρn(s)) for g ∈ G and s ∈ S, then kϕπn(x)k > η. The functions πn are equivariant. Define A, α, ρ, and π as in the proof of part (1). Then ρ is a representation of (S, R), π is an equivariant representation of (G × S, σ, Q), and π(g, s) = αg(ρ(s)) for all g ∈ G and s ∈ S. Therefore ϕπ(x) = 0, contradicting kϕπn(x)k > η for all n ∈ Z>0. (cid:3) g Theorem 5.25. Let G be a finite group, and let (S, R) be a bounded finite admis- sible set of generators and relations for a G-algebra. Then C∗(S, R) is equivariantly semiprojective if and only if (S, R) is stable in the sense of Definition 5.23. Proof. Define σ, µ, and Q as in Lemma 5.24. It follows from Remark 5.15 that (G× S, σ, Q) is bounded, finite, and admissible. By Theorem 5.22 and Remark 5.15(4), it therefore suffices to prove that (S, R) is stable if and only if (G × S, σ, Q) is stable in the sense of Definition 5.20. Assume that (S, R) is stable. Let ε > 0. Choose η > 0 as in Definition 5.23 2 ε in place of ε. Choose δ0 > 0 following 2 ε. (where the number is called δ), for 1 Lemma 5.24(1) (where the number is called δ). We may also require that δ0 ≤ 1 Apply Lemma 5.21, with δ0 in place of η, to get a number δ > 0. Let (G, A, α) and (G, B, β) be unital G-algebras (except that we allow B = 0), and let ω : A → B be an equivariant homomorphism. Let λ0 : G × S → A be a δ-equivariant δ-representation of (G×S, σ, Q) such that ω◦λ0 is an equivariant rep- resentation of (S, σ, R). By the choice of δ, there is an equivariant δ0-representation λ1 : G × S → A such that ω ◦ λ1 = ω ◦ λ0 and kλ1(g, s) − λ0(g, s)k < δ0 for all g ∈ G and s ∈ S. Define ρ1 : S → A by ρ1(s) = λ1(1, s) for s ∈ S. Since λ1 is equivariant, ρ1 is a δ0-representation of (S, R). Clearly ω ◦ ρ1 is a representation of (S, R). By the 38 N. CHRISTOPHER PHILLIPS choice of δ0, there exists a representation ρ : S → A of (S, R) such that ω ◦ρ = ω ◦ρ1 and such that for all s ∈ S we have kρ(s) − ρ1(s)k < 1 2 ε. Define λ : G × S → A by λ(g, s) = αg(ρ(s)) for g ∈ G and s ∈ S. Then, using equivariance of ω at the first step, we have ω ◦λ = ω ◦λ1 = ω ◦λ0. Moreover, for g ∈ G and s ∈ S, by equivariance of λ and λ1, we have λ(g, s) = αg(ρ(s)) and λ1(g, s) = αg(ρ1(s)). Therefore kλ(g, s) − λ0(g, s)k ≤ kαg(ρ(s)) − αg(ρ1(s))k + kλ1(g, s) − λ0(g, s)k < 1 This completes the proof that (G × S, σ, Q) is stable. 2 ε + δ0 ≤ ε. For the reverse, assume that (G × S, σ, Q) is stable. We prove that (S, R) is stable. Let ε > 0. Choose η > 0 as in Definition 5.20 (where the number is called δ). Choose δ > 0 as in Lemma 5.24(2). Let (G, A, α) and (G, B, β) be unital G-algebras (except that we allow B = 0), and let ω : A → B be an equivariant homomorphism. Let ρ0 : S → A be a δ-representation of (S, R) such that ω ◦ ρ0 is a representation of (S, R). Define π0 : G × S → A by π0(g, s) = αg(ρ(s)) for g ∈ G and s ∈ S. Then π0 is an equivariant δ-representation of (G × S, σ, Q). Therefore there exists an equivariant representation π : G × S → A of (S, σ, R) such that ω ◦ π = ω ◦ π0 and such that for all g ∈ G and s ∈ S we have kπ(g, s) − π0(g, s)k < ε. Define ρ : S → A by ρ(s) = π(1, s) for s ∈ S. Then kρ(s) − ρ0(s)k < ε for all s ∈ S. Also clearly ω ◦ ρ = ω ◦ ρ0. This completes the proof of the theorem. (cid:3) As an immediate application, we can derive stronger versions of the Rokhlin property for actions of finite groups (Definition 3.1 of [14]; formulated without the central sequence algebra in Definition 1.1 of [24]) and the tracial Rokhlin property (Definition 1.2 of [24]). Proposition 5.26. Let A be a separable unital C-algebra, and let α : G → Aut(A) be an action of a finite group G on A. Then α has the Rokhlin property if and only if for every finite set F ⊂ A and every ε > 0, there are mutually orthogonal projections eg ∈ A for g ∈ G such that: (1) αg(eh) = egh for all g, h ∈ G. (2) kega − aegk < ε for all g ∈ G and all a ∈ F. (3) Pg∈G eg = 1. The definition of the Rokhlin property differs in that in condition (1), one merely requires kαg(eh) − eghk < ε for all g, h ∈ G. The proof is very similar to, but simpler than, the proof of Proposition 5.27, and is omitted. Proposition 5.27. Let A be an infinite dimensional simple separable unital C- algebra, and let α : G → Aut(A) be an action of a finite group G on A. Then α has the tracial Rokhlin property if and only if for every finite set F ⊂ A, every ε > 0, and every positive element x ∈ A with kxk = 1, there are mutually orthogonal projections eg ∈ A for g ∈ G such that: (1) αg(eh) = egh for all g, h ∈ G. (2) kega − aegk < ε for all g ∈ G and all a ∈ F. (3) With e =Pg∈G eg, the projection 1 − e is Murray-von Neumann equivalent to a projection in the hereditary subalgebra of A generated by x. (4) With e as in (3), we have kexek > 1 − ε. The definition of the tracial Rokhlin property differs in that in condition (1) one merely requires kαg(eh) − eghk < ε for all g, h ∈ G. EQUIVARIANT SEMIPROJECTIVITY 39 We give the details of the proof to demonstrate how our machinery works, and in particular to show why we do not want to require our δ-representations to be exactly equivariant. Proof of Proposition 5.27. Let F ⊂ A be finite and let ε > 0. By scaling, without loss of generality kak ≤ 1 for all a ∈ F. Set n = card(G) and ε0 = min(cid:18) 1 n , ε 2n + 1(cid:19) . Let S consist of distinct elements pg for g ∈ G. Define an action σ of G on S by σg(ph) = pgh for g, h ∈ G. Define R =(cid:8)pgph − δg,hpg : g, h ∈ G(cid:9) ∪(cid:8)p∗ g − pg : g ∈ G(cid:9). Then (S, σ, R) is an equivariant set of generators and relations, and C∗(S, σ, R) is equivariantly isomorphic to C(G) ⊕ C, with the action on C(G) coming from the translation action of G on itself and the trivial action on C, in such a way that pg is sent to (χ{g}, 0). This action is equivariantly semiprojective by Theorem 2.6. So (S, σ, R) is stable by Theorem 5.22. Choose δ0 > 0 as in the definition of stability for ε0 in place of ε. Set δ = min(δ0, ε0). Apply the tracial Rokhlin property with δ in place of ε, obtaining mutually orthogonal projections e(0) g ∈ A for g ∈ G such that: g a − ae(0) g , the projection 1 − e(0) is Murray-von Neumann equivalent to a projection in the hereditary subalgebra of A generated by x. h (cid:1) − e(0) Define ρ0 : S → A by ρ0(pg) = e(0) g gh(cid:13)(cid:13) < δ for all g, h ∈ G. (5) (cid:13)(cid:13)αg(cid:0)e(0) (6) (cid:13)(cid:13)e(0) g (cid:13)(cid:13) < δ for all g ∈ G and all a ∈ F. (7) With e(0) = Pg∈G e(0) (8) With e(0) as in (7), we have(cid:13)(cid:13)e(0)xe(0)(cid:13)(cid:13) > 1 − ε0. (S, σ, R) such that (cid:13)(cid:13)ρ(pg) − e(0) g (cid:13)(cid:13) < ε0 ≤ 1 for g ∈ G and a ∈ F. for g ∈ G. Then ρ0 is a δ-equivariant δ- representation of (S, σ, R). Therefore there is an equivariant representation ρ of By the definition of an equivariant representation, the eg are mutually orthogonal projections satisfying condition (1). Condition (2) follows from the estimates g (cid:13)(cid:13) < ε0 for all g ∈ G. Set eg = ρ(pg) for g ∈ G. kak ≤ 1, (cid:13)(cid:13)eg − e(0) g (cid:13)(cid:13) < δ ≤ ε0 ≤ 1 It remains to prove conditions (3) and (4). Set e =Pg∈G eg. First, we have g a − ae(0) 3 ε, 3 ε and (cid:13)(cid:13)e(0) (cid:13)(cid:13)e − e(0)(cid:13)(cid:13) < nε0. Since nε0 ≤ 1, the projection e is Murray-von Neumann equivalent to e(0), and therefore also to a projection in the hereditary subalgebra of A generated by x. This is (3). Moreover, kexek ≥(cid:13)(cid:13)e(0)xe(0)(cid:13)(cid:13) − 2(cid:13)(cid:13)e − e(0)(cid:13)(cid:13) > 1 − ε0 − 2nε0 ≥ 1 − ε. This is (4). (cid:3) 40 N. CHRISTOPHER PHILLIPS 6. Graded semiprojectivity of the C-algebra of a finite group In this section, we show that if G is a finite group then C∗(G), with its natural G- grading, is semiprojective in the graded sense. This is an application of Lemma 2.4, the same result that played a key role in the proof that finite dimensional C- algebras are equivariantly semiprojective. Presumably much more general results are possible. Indeed, the appropriate setting may be actions of finite dimensional Hopf algebras or compact quantum groups on finite dimensional C-algebras. The following definition is a special case of Definitions 3.1 and 3.4 of [10], of a C-algebra (topologically) graded by a discrete group G. In [10], the group is not of the projection to A1 (as in Definition 3.4 of [10]) is automatic when the group is necessarily finite, and one only requires that Lg∈G Ag be dense in A. Continuity finite andLg∈G Ag = A. Definition 6.1. Let G be a finite group, and let A be a C-algebra. A G-grading on A is a direct sum decomposition as Banach spaces A =Mg∈G Ag such that if g, h ∈ G, a ∈ Ag, and b ∈ Ah, then ab ∈ Agh and a∗ ∈ Ag−1 . (We do not say anything about the direct sum norm except that it is equivalent to the usual norm on A.) A subspace E ⊂ A is graded if E =Pg∈G(E ∩ Ag). We denote by Pg, or P A direct sum decomposition. g , the projection map from A to Ag associated with this To put this definition in context, we make three remarks. First, when G is finite, a G-grading of A is the same as an identification of A with the C-algebra of a Fell bundle over G. The basic correspondence is given in VIII.16.11 and VIII.16.12 of [11], but in general it is not bijective. It is bijective for Fell bundles over discrete groups which are amenable in the sense of Definition 4.1 of [10], and in particular for all Fell bundles and topological gradings when G is amenable. (This follows from Theorem 4.7 of [10].) Since our groups are finite, the correspondence is bijective in our case. Second, for discrete groups G, a normal coaction on a C-algebra A (as defined before Definition 1.1 of [28]) is the same as an identification of A with the C-algebra of a Fell bundle over G. See Proposition 3.3 and Theorem 3.8 of [28]. Finally, if G is abelian, then a G-grading on A is the same as an action α : bG → Aut(A). Given a G-grading on A and τ ∈ bG, we define ατ ∈ Aut(A) by ατ (a) = τ (g)a for a ∈ Ag. Given α, for g ∈ G we set Ag =(cid:8)a ∈ A : ατ (a) = τ (g)a for all τ ∈ bG(cid:9), that is, Ag is the spectral subspace for g when g is regarded as an element of the second dual of G. Remark 6.2. Let G be a finite group, and let A =Lg∈G Ag be a G-grading of A. Then the summand A1 is a C-algebra. (This is clear.) Let Pg : A → Ag be as in Definition 6.1. Then P1 is a conditional expectation onto A1, and kPgk ≤ 1 for all g ∈ G. (See Theorem 3.3 and Corollary 3.5 of [10].) EQUIVARIANT SEMIPROJECTIVITY 41 Definition 6.3. Let G be a finite group, let A and B be a C-algebras with G- We say that ϕ is graded if for every g ∈ G we have ϕ(Ag) ⊂ Bg. gradings A =Lg∈G Ag and B =Lg∈G Bg, and let ϕ : A → B be a homomorphism. A = Lg∈G Ag, and let I ⊂ A be a graded ideal. Then A/I becomes a graded Remark 6.4. Let G be a finite group, let A be a C-algebra with G-grading C-algebra with the grading (A/I)g = Ag/(Ag ∩ I) = (Ag + I)/I, and the quotient map A → A/I is a graded homomorphism. Remark 6.5. Let G be a finite group. Then the direct limit of a direct system of G-graded C-algebras with graded maps is a G-graded C-algebra in an obvious way. Remark 6.6. Let G be a finite group, let A be a C-algebra, and let α : G → Aut(A) be an action of G on A. Then the crossed product C∗(G, A, α) is graded in the following way. Let ug ∈ C∗(G, A, α) (or in M (C∗(G, A, α)) if A is not unital) be the standard unitary corresponding to g ∈ G. Then C∗(G, A, α)g = {aug : a ∈ A}. (This is the dual coaction.) Remark 6.7. In Remark 6.6, take A = C and take α to be the trivial action. This gives a canonical G-grading on C∗(G). If ug ∈ C∗(G) is the unitary corresponding to g ∈ G, then C∗(G)g = Cug. The following definition is the analog of Definition 14.1.3 of [20]. Definition 6.8. Let G be a finite group, and let A be a C-algebra with G-grading A =Lg∈G Ag. We say that the grading is graded semiprojective if whenever C is a a C-algebra with G-grading C =Lg∈G Cg, J0 ⊂ J1 ⊂ · · · are graded ideals in C, J = S∞ n=0 Jn, and ϕ : A → C/J is a graded homomorphism, then there exists n and a graded homomorphism ψ : A → C/Jn such that the composition A ψ −→ C/Jn −→ C/J is equal to ϕ. When no confusion can arise, we say that A is graded semiprojective. Here is the diagram: C C/Jn =④ C/J. ④ ψ ④ A ϕ ④ Theorem 6.9. Let G be a finite group. Then C∗(G), with the G-grading in Remark 6.7, is graded semiprojective. ✤ ✤ ✤ ✤ ✤ ✤ / / = 42 N. CHRISTOPHER PHILLIPS Proof. Set ε0 = 1 6·34 , and choose ε > 0 such that ε ≤ ε0 and such that whenever A is a unital C-algebra, u ∈ U (A), and a ∈ A satisfies ka − uk < ε, then we have (cid:13)(cid:13)a(a∗a)−1/2 − u(cid:13)(cid:13) < ε0. Let the notation be as in Definition 6.8 and Remark 1.3(3). Further, for g ∈ G let P (n) : C/Jn → (C/Jn)g and Pg : C/J → (C/J)g be the projection maps associated g to the gradings, as in Definition 6.1. Also let ug ∈ C∗(G) be the unitary associated with the group element g ∈ G, as in Remark 6.7. Let ϕ : A → C/J be a unital graded homomorphism. Since finite dimensional C-algebras are semiprojective, there exist n0 and a unital homomorphism (not necessarily graded) ψ0 : A → C/Jn0 which lifts ϕ. Since πn0 and ϕ are graded, for g ∈ G we have πn0(cid:0)ψ0(ug) − P (n0) g (ψ0(ug))(cid:1) = ϕ(ug) − Pg(ϕ(ug)) = 0. Therefore there exists n ≥ n0 such that for all g ∈ G we have (6.1) (cid:13)(cid:13)πn,n0(cid:0)ψ0(ug) − P (n0) g (ψ0(ug))(cid:1)(cid:13)(cid:13) < ε. Set ψ1 = πn,n0 ◦ ψ0 and for g ∈ G set cg = P (n) Then (6.1) becomes g (ψ1(ug)), which is in (C/J)g. for all g ∈ G. kψ1(ug) − cgk < ε Since ε < 1, we can define ρ0 : G → U (C/Jn) by ρ0(g) = cg(c∗ cg ∈ (C/J)g, we have c∗ (C/J)g. Moreover, the choice of ε ensures that kcg − ρ0(g)k < ε0. Therefore gcg ∈ (C/J)1, whence (c∗ gcg)−1/2. Since gcg)−1/2 ∈ (C/J)1, so that ρ0(g) ∈ kρ0(g) − ψ1(ug)k ≤ kρ0(g) − cgk + kcg − ψ1(ug)k < ε0 + ε ≤ 2ε0. Let g, h ∈ G. Since ψ1(ug), ψ1(uh) ∈ U (C/Jn) and ψ1(ugh) = ψ1(ug)ψ1(uh), it follows that Since πn(cg) = ϕ(ug) is unitary, we also get kρ0(gh) − ρ0(g)ρ0(h)k < 6ε0. πn(ρ0(g)) = πn(cg) = ϕ(ug). Inductively define functions ρm : G → U (C/Jn) by (following Lemma 2.4) ρm+1(g) = exp 1 card(G) Xh∈G log(cid:16)ρm(h)∗ρm(hg)ρm(g)∗(cid:17)! ρm(g) for g ∈ G. Since 6ε0 = 1 17 , Lemma 2.4 implies that the functions ρm are well defined maps ρm : G → U (C/Jn) such that ρ(g) = limm→∞ ρm(g) defines a homomorphism ρ : G → U (C/Jn) satisfying 34 < 1 kρ(g) − ρ0(g)k ≤ 2 · 6ε0 1 − 17 · 6ε0 and πn(ρ(g)) = ϕ(ug) for all g ∈ G. We claim that ρm(g) ∈ (C/Jn)g for all g ∈ G and m ∈ Z≥0. The proof is by induction on m. We know this is true for m = 0. Assume it is true for m. For all g, h ∈ G we have ρm(h)∗ρm(hg)ρm(g)∗ ∈ (C/Jn)∗ h(C/Jn)hg(C/Jn)∗ g ⊂ (C/Jn)1. EQUIVARIANT SEMIPROJECTIVITY 43 Therefore also exp 1 card(G) Xh∈G log(cid:16)ρm(h)∗ρm(hg)ρm(g)∗(cid:17)! ∈ (C/Jn)1, and the induction step follows. This proves the claim. Taking limits, we get ρ(g) ∈ (C/Jn)g for all g ∈ G. By the universal property of C∗(G), there is a unital homomorphism ψ : C∗(G) → C/J such that ψ(ug) = ρ(g) for all g ∈ G. By construction, ψ is graded. Moreover, πn ◦ ψ(ug) = ϕ(ug) for all g ∈ G, so the universal property of C∗(G) implies that πn ◦ ψ = ϕ. Thus ψ lifts ϕ. (cid:3) References [1] M. A. Alekseev, L. Y. Glebskiı, and E. I. Gordon, On approximation of groups, group ac- tions, and Hopf algebras, Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 256(1999), Teor. Predst. Din. Sist. Komb. i Algoritm. Metody. 3, 224 -- 262, 268; English trans- lation in: J. Math. Sci. (New York) 107(2001), 4305 -- 4332. [2] S. A. Antonyan, Compact group actions on equivariant absolute neighborhood retracts and their orbit spaces, Topology Appl. 158(2011), 141 -- 151. [3] S. A. Antonyan and E. Elfving, The equivariant homotopy type of G-ANR's for compact group actions, Manuscripta Math. 124(2007), 275 -- 297. [4] B. Blackadar, Shape theory for C-algebras, Math. Scand. 56(1985), 249 -- 275. [5] B. Blackadar, Semiprojectivity in simple C-algebras, pages 1 -- 17 in: Operator Algebras and Applications (Adv. Stud. Pure Math. vol. 38), Math. Soc. Japan, Tokyo, 2004. [6] B. Blackadar, private communication. [7] J. Cuntz, K-theory for certain C-algebras, Ann. Math. 113(1981), 181 -- 197. [8] S. Eilers, T. A. Loring, and G. K. Pedersen, Stability of anticommutation relations. An application of noncommutative CW complexes, J. reine angew. Math. 499(1998), 101 -- 143. [9] M. Enomoto, H. Takehana, and Y. Watatani, Automorphisms on Cuntz algebras, Math. Japonica 24(1979), 231 -- 234. [10] R. Exel, Amenability for Fell bundles, J. reine angew. Math. 492(1997), 41 -- 74. [11] J. M. G. Fell and R. S. Doran, Representations of -Algebras, Locally Compact Groups, and Banach -Algebraic Bundles. Vol. 2. Banach -Algebraic Bundles, Induced Representations, and the Generalized Mackey Analysis, Pure and Applied Mathematics, no. 126. Academic Press, Boston MA, 1988. [12] K. Grove, H. Karcher, and E. A. Ruh, Group actions and curvature, Invent. Math. 23(1974), 31 -- 48. [13] K. Grove, H. Karcher, and E. A. Ruh, Jacobi fields and Finsler metrics on compact Lie groups with an application to differentiable pinching problems, Math. Ann. 211(1974/75), 7 -- 21. [14] M. Izumi, Finite group actions on C-algebras with the Rohlin property. I , Duke Math. J. 122(2004), 233 -- 280. [15] J. Jaworowski, An equivariant extension theorem and G-retracts with a finite structure, Manuscripta Math. 35(1981), 323 -- 329. [16] J. A. Jeong and H. Osaka, Extremally rich C-crossed products and the cancellation property, J. Austral. Math. Soc. (Series A) 64(1998), 285 -- 301. [17] P. Julg, K-th´eorie ´equivariante et produits crois´es, C. R. Acad. Sci. Paris S´er. I Math. 292(1981), 629 -- 632. [18] D. Kazhdan, On ε-representations, Israel J. Math. 43(1982), 315 -- 323. [19] A. Kishimoto, Outer automorphisms and reduced crossed products of simple C-algebras, Commun. Math. Phys. 81(1981), 429 -- 435. [20] T. A. Loring, Lifting Solutions to Perturbing Problems in C-Algebras, Fields Institute Mono- graphs no. 8, American Mathematical Society, Providence RI, 1997. [21] W. L. Paschke, K-theory for actions of the circle group on C-algebras, J. Operator Theory 6(1981), 125 -- 133. 44 N. CHRISTOPHER PHILLIPS [22] N. C. Phillips, Equivariant K-Theory and Freeness of Group Actions on C-Algebras, Springer-Verlag Lecture Notes in Math. no. 1274, Springer-Verlag, Berlin, Heidelberg, New York, London, Paris, Tokyo, 1987. [23] N. C. Phillips, Inverse limits of C-algebras and applications, pages 127 -- 186 in: Operator Algebras and Applications Volume 1: Structure Theory: K-Theory, Geometry and Topology, London Math. Soc. Lecture Note Series No. 135, Cambridge University Press, Cambridge, New York, New Rochelle, Melbourne, Sydney, 1988. [24] N. C. Phillips, The tracial Rokhlin property for actions of finite groups on C-algebras, Amer. J. Math., to appear (arXiv: math.OA/0609782). [25] N. C. Phillips, in preparation. [26] N. C. Phillips, A. P. W. Sørensen, and H. Thiel, in preparation. [27] M. Pimsner, A class of C-algebras generalizing both Cuntz-Krieger algebras and crossed products by Z, pages 189 -- 212 in: Free Probability Theory (Waterloo, ON, 1995), Fields Inst. Commun. vol. 12, Amer. Math. Soc., Providence, RI, 1997. [28] J. C. Quigg, Discrete C-coactions and C*-algebraic bundles, J. Australian Math. Soc. 60(1996), 204 -- 221. [29] J. Rosenberg, Appendix to O. Bratteli's paper on "Crossed products of UHF algebras", Duke Math. J. 46(1979), 25 -- 26. [30] H. L. Royden, Real Analysis, 3rd. ed., Macmillan, New York, 1988. Department of Mathematics, University of Oregon, Eugene OR 97403-1222, USA, and Research Institute for Mathematical Sciences, Kyoto University, Kitashirakawa- Oiwakecho, Sakyo-ku, Kyoto 606-8502, Japan. E-mail address: [email protected]
1709.08938	1	1709	2017-09-26T10:55:57	Locally free actions of groupoids and proper topological correspondences	[ "math.OA" ]	Let $(G,\alpha)$ and $(H,\beta)$ be locally compact Hausdorff groupoids with Haar systems, and let $(X,\lambda)$ be a topological correspondence from $(G,\alpha)$ to $(H,\beta)$ which induce the ${C}^$-correspondence $\mathcal{H}(X)\colon {C}^(G,\alpha)\to {C}^(H,\beta)$. We give sufficient topological conditions which when satisfied the ${C}^$-correspondence $\mathcal{H}(X)$ is proper, that is, the ${C}^$-algebra ${C}^(G,\alpha)$ acts on the Hilbert ${C}^(H,\beta)$-module ${H}(X)$ via the comapct operators. Thus a proper topological correspondence produces an element in ${KK}({C}^(G,\alpha),{C}^*(H,\beta))$.	math.OA	math	LOCALLY FREE ACTIONS OF GROUPOIDS AND PROPER TOPOLOGICAL CORRESPONDENCES ROHIT DILIP HOLKAR Abstract. Let (G, α) and (H, β) be locally compact Hausdorff groupoids with Haar systems, and let (X, λ) be a topological correspondence from (G, α) to (H, β) which induce the C∗-correspondence H(X): C∗(G, α) → C∗(H, β). We give sufficient topological conditions which when satisfied the C∗-correspondence H(X) is proper, that is, the groupoid C∗(G, α) acts on the C∗(H, β)-Hilbert module H(X) via the comapct operators. Thus a proper topological corre- spondence produces an element in KK(C∗(G, α), C∗(H, β)). Contents Introduction 1. Preliminaries 1.1. Measures concentrated on sets 1.2. Free, proper and transitive actions of groupoids 1.3. Locally free actions of groupoids 1.4. Proper groupoids and cutoff functions 1.5. Compact operators 2. Proper topological correspondences 2.1. Some examples 2.2. Proper correspondences 3. Examples 3.1. The case of spaces 3.2. The case of étale groupoids References 1 6 7 15 17 22 23 26 26 28 35 37 40 43 Introduction Given C∗-algebras A and B, a C∗-correspondence from A to B is a pair (M, φ) where M is a Hilbert B-module, and φ : A → B(M) is a nondegenerate rep- resentation. We call the C∗-correspondence (M, φ) proper if the representation φ : A → K(M). Proper correspondences are important and studied for various pur- poses. In this article, we shall denote a C∗-correspondence merely by the Hilbert module in its definition. We shall not come across an occasion where the represen- tation needs to be explicitly spelled out. In [9], topological correspondences of locally compact groupoids equipped with Haar systems are defined. Let (G, α) and (H, β) be locally compact groupoids equipped with Haar systems. A topological correspondence from (G, α) to (H, β) is a pair (X, λ) where X is a G-H-bispace with the H-action proper, λ := {λu}u∈H(0) is a proper H-invariant family of measures along the right momentum map sX : X → Key words and phrases. C∗-correspondences, topological correspondences, proper C∗- correspondences, proper topological correspondences, groupoid KK-theory. Subject class. 22D25, 22A22, 46H35, 47L30, 46L08, 19K35, 19L99. 1 2 ROHIT DILIP HOLKAR H (0) and each measure λu in λ is (G, α)-quasi-invariant. The main theorem in [9, Theorem 2.10] asserts that a certain completion of Cc(X) gives a C∗-correspondence H(X, λ) from C∗(G, α) to C∗(H, β). rX←−− X For the above topological correspondence (X, λ), let G(0) sX−−→ H (0) be the momentum maps. The C∗-algebra K(H((X, λ)) of compact operators on the Hilbert C∗(H, β)-module H(X, λ) can be described using a topological, namely, a hypergroupoid equipped with a Haar system. To do this, one observes that the right diagonal action of H on the fibre product X ×sX , X is proper. In [19], Renault shows that the quotient space (X×sX ,H(0),sX )/H is a hypergroupoid equipped with a Haar system λ. The Haar system λ is induced by λ. The C∗-algebra of this hypergroupoid, C∗(X×sX ,H(0),sX )/H), λ is the C∗-algebra of compact operators on H(X, λ). We revise this result in Section 1.5; this was written for locally compact, Hausdorff and second countable groupoids and spaces in [7]. The notion of topological correspondence in [9] generalises many existing no- tions of topological correspondences in the literature including Jean-Louis Tu's notion of locally proper generalised morphism in [20]. In [20], Tu defines a proper locally proper generalised morphism which is a proper topological correspondences. That is, he adds an extra conditions to his topological correspondence so that the C∗-correspondence H(X, λ) it produces is proper. This work of Tu raises a question, when is a topological correspondence defined in [9] proper? To be precise, under what extra hypothesis on the topological correspondence in [9] is the corresponding C∗-correspondence proper? Apart from Tu's above-mentioned work, this question has been discussed and answered for special cases. Marth-Stadler and O'uchi define a topological corre- spondence and a proper one in [12]. This work is a special case of the work in [20]. In [15], Muhly and Tomford define topological quivers, which are special types of topological correspondences (see [9, Example 3.3]). In the same article, they also characterise proper topological quivers, see [15, Theorem 3.11]. Now we briefly elaborate the above works putting Tu's at the end. Let Y be a locally compact Hausdorff space. Slightly modifying [15, Definition 3.1], we may say that a topological quiver from Y to itself is a quadruple (X, b, f, λ) where b, f : X → Y are continuous maps and λ is a continuous family of measures along f . [15, Theorem 3.11] says that the quiver is proper if and only if f is a local homeomorphism and b is proper. Let Y and Z be locally compact Hausdorff spaces. Then as in [3], in general, one defines that a quiver from Y to Z is a quadruple (X, b, f, λ) where b : X → Y and f : X → Z are continuous maps, and λ is a continuous family of measures along f . One may check that [15, Theorem 3.11] is valid for this slightly generalised notion of quivers also. If one thinks of the spaces Y and Z above as the trivial groupoids, then both the above of quivers are topological correspondences, see [9, Example 3.3]. Tu [20] works with locally compact groupoids equipped with Haar systems and the space involved in the (proper) topological correspondence is locally compact but not necessarily Hausdorff. Marta-Stadler and O'uchi's work [12] involves Hausdorff groupoids equipped with Haar systems and Hausdorff spaces; their proper corre- spondences is a special case of Tu's work, hence we focus on [20]. Let (G, α) and (H, β) be locally compact groupoids equipped with Haar systems. A topological correspondence from (G, α) to (H, β) in the sense of Tu is a G-H-bispace X such that i) both the actions are proper, ii) the action of G is free, PROPER TOPOLOGICAL CORRESPONDENCES 3 iii) the right momentum map sX : X → H (0) induces an isomorphism [sX ] : X/H → H (0). Property iii) above is equivalent to iii') the map msX : G ×sG,G(0),rX X → X ×sX ,H(0),sX X, msX (γ, x) 7→ (γx, x), is a homeomorphism where sG is the source map of G, rX : X → G(0) and sX : X → H (0) are the momentum maps for the action of G and H, respectively, and the domain and codomain of the function msX are the obvious fibre products. [9, Example 3.7] shows that this is a topological correspondence in our sense, that is, in the sense of [9, Definition 2.1]. [20, Definition 7.6] says that the above corre- spondence is proper if the map [rX ] : X/H → G(0) induced by the momentum map rX : X → G(0) for the action of G is proper. [20, Theorem 7.8] proves that if the above topological correspondence X is proper, then so is the C∗-correspondence H(X) : C∗ r (G, α) → C∗ r (H, β). Now we discuss the three conditions in the definition of a proper correspondence which is proposed in this article. Let (X, λ) be a topological correspondence from a locally compact groupoid equipped with a Haar system (G, α) to another one, say (H, β). Let rX and sX be the momentum maps for the actions of G and H, respectively, on X. First of all, one may expect from preceding literature survey that the map [rX ] : X/H → G(0) should be proper. This is true; this is one of the conditions we need. However, this is not a sufficient condition while dealing a general topological correspondence. For example, if G = H = {∗}, the trivial group, and (X, λ) is a compact measure space, then [rX ] is proper. In this case, H(X, λ) = L2(X, λ) and C∗(G) = C∗(H) = C. The action of C on H(X) is by scalar multiplication which has the identity operator 1. Hence, in general, H(X, λ) : C → C is not a proper correspondence. Secondly, since there are families measures involved one may expect that there should be condition(s) that relate the family of measures on the bispace and the Haar system on the left groupoid. This is a technical condition; this is the third condition in Definition 2.5. Thirdly and finally, an interesting and not-at-all-obvious condition is the rephras- ing iii') of the property iii) appearing in the definition of Tu's proper correspondence above. It is a classical condition that appears in the definition of a groupoid equiv- alence ([14]) and may other notions of groupoid morphisms. This property gives a family of measures on X as observed in [9, Exmaple 3.7]. What is so interesting about this property? In its other form, namely iii'), the property proves very useful. With the same notations as in the last paragraph, let b ∈ Cc(G) and ξ ∈ Cc(X). Then the action of b on ξ that gives the representation of C∗(G, α) on H(X, λ) is given by bξ(x) =ZG b(γ)ξ(γ−1x)∆(γ, γ−1x) dαrX (x)(γ) where ∆ is the adjoining function for the correspondence (X, λ). Now, we aim to assign b an element b in Cc((X ×sX ,H(0),sX X)/H) such that bξ = bξ. For this purpose, using the fact that [rX ] is proper, we choose adummy function t in Cc(X). Then b⊗G(0) t ∈ Cc(G×G(0) X). If msX were a homeomorphism, then (b⊗G(0) t)◦m−1 sX is an element of Cc(X ×sX ,H(0),sX X). If one chooses the dummy function t carefully, then after averaging over H, the image of b⊗G(0) t in Cc((X×sX ,H(0),sX X)/H) serves as the required function b. Note that if msX is a homeomorphism, then the action of G on X is free as well as proper. A bit of more work shows that the above argument goes through when 4 ROHIT DILIP HOLKAR msX is a homeomorphism onto its image. An action of G on X for which msX is a homeomorphism onto its image are called basic in [13]. However, the following simple example shows that for us even basic actions too much to ask for. Let S1 denote the unit circle; consider this compact space as the trivial groupoid. Let the multiplicative group of order two, Z/2Z := {1, −1}, act on the space S1 from left by (−1) · z = ¯z for −1 ∈ Z/2Z and z ∈ S1. Let S1 act on itself trivially from right. The momentum map for this action is the identity map IdS1 : S1 → S1. Let τIdS1 be the family of measures along IdS1 which consists of point masses. Then (S1, τIdS1 ) is a topological correspondence from Z/2Z to S1; the adjoining function of this correspondence is the constant function 1. The topological correspondence (S1, τIdS1 ) gives the C∗-correspondence C(S1) : C∗(Z/2Z) → C(S1). Note that here we make the C∗-algebra C(S1) a Hilbert module over itself in the obvious way. Since this C∗-algebra is unital, we have K(C(S1)) = M(C(S1)) = C(S1). Therefore, the C∗-correspondence C(S1) : C∗(Z) → C(S1) is proper. In above example, msS1 : Z/2Z×S1 → S1 ×S1 is the map (±1, z) 7→ (z, z) or (¯z, z) which is not a homeomorphism, even, onto its image as (−1) · (1 + 0i) = 1 · (1 + 0i) where 1, −1 ∈ Z/2Z and 1 + 0i ∈ S1. However, on may check that this map is for ±1 ∈ Z/2Z, mS1 {±1}×S1 is a home- a local homeomorphism onto its image: omorphism onto its mage; z , z) where (z′, z) ∈ Im(mS1 {±1}×S1 ). inverse of this restriction is (z′, z) 7→ ( z′ This and similar examples motivate us to add the condition that msX is a local homeomorphism onto its image. And this serves our purpose well. Once the maps which are local homeomorphisms onto their images are in pic- ture, we have to study these local homeomorphisms: we have to study how does a map of this type behave (i) with respect extensions of continuous function along it, (ii) with respect to induction of measures along it. We study these technical issues in Sections 1.1. One may put all these conditions and above study together and state the main result in this article. However, we investigate when is the map msX above a local homeomorphism. This question breaks down into two pieces: (i) when is msX locally one-to-one? (ii) when is msX open onto its image? The second question has no concrete answer and examples show that msX is open has to be a part of data. The first question leads us to the study of the locally free action of groupoids; this study is done in Section 1.3. While studying the locally free actions, we also define transitive actions of groupoids in Section 1.2 which facilitate rephrasing a classical condition more theoretically, see Remark 1.26. Since Muhly and Tomford [15] not only define a proper topological correspon- dence but also characterise it, we wish to see that up to what level can our definition of a proper correspondence reproduce their results, we study measures which are concentrated on certain sets (Section 1.1). What do we finally achieve? We define a proper topological correspondence (Definition 2.5) and show that a proper topological correspondence of groupoids equipped with Haar systems produce a proper C∗-correspondence of full groupoid C∗-algebras (Theorem 2.11). We study locally free actions of groupoid: we define locally free and strongly locally free actions of groupoids and show that they do not mean the same (Example 1.32). However, in the case of groups, these notions are same. While proving this, the example, Example 1.32, shows that a groupoid with discrete fibres need not be étale. Moreover, while studying measures concentrated on a set, we prove Lemma 1.12 which gives different characterisations that when is a measure concentrated on a set; this lemma is not a deep work, but it comfortable PROPER TOPOLOGICAL CORRESPONDENCES 5 to have it. In Section 3, we show that all the definitions of various proper topo- logical correspondences mentioned earlier fit in Definition 2.5. Following is a detail discussion of results in this section. We discuss the (étale) proper correspondences of spaces and étale groupoids in detail in Section 3.1 and 3.2, respectively. We show that in a proper topological correspondence of spaces in the sense of Definition 2.5, the family of measures on the middle space consists of atomic measures and has full support, see 3.8. Reader may compare this with [15, Theorem 3.11]. This result generalises to étale groupoids also, Lemma 3.19. Let X, Y and Z be spaces, and Y b←− X f −→ Z be continuous maps where f is a local homeomorphism. Then X carries a continuous family of counting measures, τsX , along the étale map f which makes X into a topological correspondence from Y to Z. Definition 2.5 implies that X is proper if b is a proper map. Section 3.2 discusses the case of étale groupoids. In this section, we study étale topological correspondence of étale groupoids. Let X be a G-H-bispace, where G and H are étale groupoids. Assume that the right momentum map sX is a local homeomorphism which gives X a continuous family of atomic measure, τsX , along sX . If (X, τsX ) is a topological correspondence from G to H, then we call X an étale correspondence. Example of such a morphism is the Hilsum-Skandalis mor- phism, see [6]. Proposition 3.14 shows that any (proper) topological correspondence obtained from the above G-H-bispace X by changing the family of measures on X is isomorphic to (X, τsX ). Thus the KK-class in KK(C∗(G), C∗(H)) determined by such a bispace does not depend on the family of measures the bispace carries. We show that X is a proper correspondence if the momentum map for the right action is a proper, see Proposition 3.18. What could we not achieve? The proof of Theorem 2.11 uses the cutoff func- tion on a proper groupoid which needs that the space of units is Hausdorff. This forces that the bispace involved in a topological correspondence is Hausdorff. Thus, though we write results for locally compact groupoids when it comes to proving Theorem 2.11, the fact that the groupoids are non-Hausdorff does not play a great role. We do not know examples of groupoid actions which is locally strictly locally free but not strictly locally free, and locally free but not locally locally free, see Figure 1 and the discussion below it. Before proceeding to the summary of the article, we would recommend the reader to assume the function D and an adjoining function of a topological correspondence to be the constant function 1, especially in Section 2 and the proof of Theorem 2.11 during the first reading. This would reduce the complexity in the proofs and ideas. Following is the sectionwise summary. Section 1: we fix some important notation in this section. This section has five subsections. The first one, namely, Section 1.1, discusses extensions of functions along local homeomorphisms which are not necessarily surjective and then some measure theoretic preliminaries. The continuous extensions of a function along a local homeomorphism, which need not be surjective, are used to verbalise one of the conditions for a proper correspondence, Definition 2.5(iii). In the measure theoretic preliminaries, firstly we discuss constructing measures using local data and restriction of measures. Then, in Lemma 1.12, we discuss various equivalent ways of saying that a measure is concentrated on a subset. This is used to prove the next lemma, Lemma 1.14, which is one of important result for this article. We use Lemma 1.14 to interpret a condition in the definition of a proper correspondence in certain cases. The last part of this section discusses absolute continuity of families of measures. In the next subsection, Section 1.2, we discuss free, proper and transitive actions. 6 ROHIT DILIP HOLKAR In Section 1.3, we introduce different notions of locally free actions of groupoids and examples of some of them. In Proposition 1.35, we show that any action of an étale groupoid is strongly locally free. Section 1.4 is a revision of some well-known results about proper groupoids and cutoff functions. Let (H, β) be a locally compact groupoid with a Haar system. Let X be a proper H-space and λ an H-invariant family of measures on X. In Section 1.5, we describe KT(X), the C∗-algebra of compact operators on the C∗(H, β). Section 2: This section starts with three examples which, we expect, may prove helpful to understand the definition of a proper correspondence. The examples are followed by the definition of a proper correspondence, Definition 2.5. Then we make few remarks. The rest of the section is the proof our main theorem, namely, Theorem 2.11. Section 3: In this section, we give examples of proper correspondences and show that some well-known proper topological correspondences are proper in our sense. Example 3.1 shows that a proper locally proper generalised morphism defined by Jean -- Louis Tu in [20] is a proper topological correspondence. Example 3.4 shows that an equivalence of groupoid, Example 3.12 shows that a proper quiver defined by Muhly and Tomford in [15] is a proper topological correspondence. This section discusses étale topological correspondence of spaces and étale groupoids. Propo- sition 3.18 shows that for an étale correspondences, one of the groupoids can be reduced to a space. 1. Preliminaries Topological conventions. We assume that reader is familiar with basic theory of locally compact groupoids ([17], [20]), continuous families of measures ([18], [7, Sec- tion 2.5.2]) and topological correspondences of locally compact groupoids equipped with Haar systems ([9]). Let X be a topological space. A subset A ⊆ X is called quasi-compact if every open cover of A has a finite subcover, compact if it is quasi-compact and Hausdorff. The space X is called locally compact if every point has a compact neighbourhood. We call a groupoid locally compact if it is a locally compact topological space and its space of units is Hausdorff. For locally compact space X Cc(X)0 denotes the set of functions f such that f vanishes outside a compact set V and f v ∈ Cc(V ). And Cc(X) is defined as the linear span of Cc(X)0. The main definition and theorem in this article are stated for locally compact groupoids and Hausdorff spaces, so reader may simply assume that groupoids are also Hausdorff. For a function f : X → Y of sets, Im(f ) denotes the image of f . Assume that X and Y are spaces, and f : X → Y a continuous map. We say f is open, if f (U ) is open in Y for every open U ⊆ X. We call f open onto its image, if f (U ) is open in f (X) for very open U ⊆ X. The function f above is called locally one-to-one if for each x ∈ X there is a neighbourhood U ⊆ X of x with f U : U → f (U ) is a one-to-one function. The function f is called a local homeomorphism if f is surjective and for each x ∈ X there is a neighbourhood U ⊆ X of x with f U : U → f (U ) is a homeomorphism. In this case, f is an open map. We call f a local homeomorphism onto its image if x ∈ X there is a neighbourhood U ⊆ X of x with f U : U → f (U ) is a homeomorphism; this is equivalent to saying that the function f ′ : X → f (X) obtained from f by restricting its codomain is a local homeomorphism. Let X be a space. If ∼ is an equivalence relation on X, then [x] ∈ X/ ∼ denotes the equivalence class of x ∈ X under ∼, and qX : X → X/ ∼ denotes the quotient map qX (X) = [x]. Let Y be another space and f : X → Y a continuous map. If PROPER TOPOLOGICAL CORRESPONDENCES 7 f preserves ∼, that is, f (x) = f (y) for all x, y ∈ X with x ∼ y, then the universal property of quotient induces a map X/ ∼→ Y , [x] → f (x). We denote this map by [f ]. Let X, Y and Z be spaces, and let a : X → Z and b : Y → Z be functions. Then we denote the set {(x, y) ∈ X × Y : a(x) = b(y)}, which is the fibre product of X and Y over Z along a and b, by X ×a,Z,b Y most often. Sometimes, when the map a and b are clear, we write X ×Z Y . For U ⊆ X and V ⊆ Y , U ×a,Z,b V denotes the set (U × V ) ∩ (X ×Z Y ). Note that U ×a,Z,b V ⊆ X ×a,Z,b Y is open if and only if U × V ⊆ X ×a,Z,b Y is open, which is, in turn, open if and only if U ⊆ X and V ⊆ Y are open. Let X, Y, Z, a and b be as in the last paragraph. Let f be a function on X and g on Y . Then f ⊗ g is the function on X × Y (x, y) 7→ f (x)g(x). The restriction of f ⊗ g to the fibre product X ×a,Z,b Y is written as f ⊗Z g. Let X, Y and Z be spaces, and f : X → Z and g : Y → Z continuous maps. Then a basic open set for the product topology on X ×Y is a product U ×V where U ⊆ X and V ⊆ Y are open. Hence the sets of type U ×f,Z,g V = (U × V ) ∩ (X ×f,Z,g Y ), where U ⊆ X and V ⊆ Y are open, are basic open sets for the subspace topology of the fibre product X ×f,Z,g Y ⊆ X × Y . For a groupoid G, rG and sG denote its range and source maps, respectively. For a left (right) G-space, we always assume that the corresponding momentum map is rX (respectively, sX ), with an exception of Example 3.12. We denote the fibre product G ×sG,G(0),rX X (respectively, X ×sX ,G(0),rG G) by G ×G(0) X (respectively, X ×G(0) G). We say γ and x are composable or (γ, x) is a composable pair if (γ, x) ∈ G ×G(0) X; similarly for the right action. Let X be a left G-space and right H-space for groupoids G and H. We call X a G-H-bispace if the actions of G and H commute in the usual sense, that is, for every composable pairs (γ, x) ∈ G ×G(0) X and (x, η) ∈ X ×H(0) H we have that (i) (γx, η) ∈ X ×H(0) H and (γ, xη) ∈ G ×G(0) X, (ii) (γx)η = γ(xη). Let (H, β) be a locally compact groupoid with a Haar system and X a right H-space. Then there is continuous family of measures with full support βX = {β[x] X }[x]∈X/H along the quotient map X → X/H which is given by (1.1) ZX f dβ[x] X :=ZG f (xη) dβsX (x)(η) for [x] ∈ X/H and f ∈ Cc(X). See [7, Proposition 1.3.21] for the proof. Let A be a category and let A0 denote its class of objects. By x ∈∈ A we mean that x is an object in A. The symbols R∗ and C∗ denote, respectively, the set of nonzero real and complex numbers. And R− and R+ denote is the set of negative and positive real numbers, respectively. 1.1. Measures concentrated on sets. Our reference for measure theory is Bour- baki [2]. For a topological space X, let Bor(X) denote the set of Borel functions on X. Given Y ⊆ X, χY denotes the characteristic function of the set Y . In this article, all the measures are assumed to be positive, Radon and σ-finite. Now we introduce some notation. Notation 1.2 (Notation and remarks). (1) Let X be a space and V, U ⊆ X with U ⊆ V . Let f be a function on U . Then we write 0V (f ) for the function on V which equals f on U and zero on V − U . Let X be locally compact space and U an open Hausdorff subset of it. Then define the set Cc(X, U ) = {f ∈ Cc(X) : supp(f ) ⊆ U }. 8 ROHIT DILIP HOLKAR When X is Hausdorff one may identify Cc(X, U ) with Cc(U ) as follows: if g is in Cc(X, U ), then the restriction of g to U lies in Cc(U ). On the other hand, if f is in Cc(U ), then 0X (f ) is in Cc(X, U ). These two process are inverses of each other. We shall frequently use this identification. (2) Let X and Y be locally compact Hausdorff spaces, and φ : X → Y a map which is open onto its image. For an open set U ⊆ X, define E Set φ (U ) = {U ′ ⊆ Y : U ′ is open in Y and U ′ ∩ φ(X) = φ(U )} (here an everywhere else, E stands for 'extension'). Note that for U 6= ∅, E Set φ (U ) is not empty: since φ(U ) ⊆ φ(X) is open, there is an open set U ′ ⊆ Y with φ(U ) = U ′ ∩ φ(X). We call an element of E Set φ (U ) an extension of U via φ, or simply an extension of U . (3) A set is called cocomapct if it has compact closure. Let X and Y be locally compact Hausdorff spaces, and φ : X → Y a local homeomorphism onto its image. Then φ is an open map onto its image. Let φHo := {U ⊆ X : U is open and φU is a homeomorphism} φHo,Co := {U ⊆ X : U is cocompact, U ∈ φHo and φU is a homeomorphism} Then φHo 6= ∅ is an open cover of X for X 6= ∅. Given U ∈ φHo, we write φ−1 U U : φ(U ) → U is a homeomorphism. The map φ−1 for (φU )−1. Note that φ−1 U induces an obvious isomorphism Cc(U ) → Cc(φ(U )) which sends a function f ∈ Cc(U ) to the function f ◦ φ−1 φ(U) in Cc(φ(U )). Since X is locally compact, for any U ∈ φHo and x ∈ U , there is an open cocompact neighbourhood V of x with V ⊆ U . Hence φHo,Co 6= ∅. Moreover, φHo and φHo,Co are open covers of X which are closed under intersections of sets. (4) Let X, Y and φ be as in (3) above. Let U ∈ φHo. For f ∈ Cc(X; U ) define the set Eφ(f ) = {f ′ ∈ Cc(Y, U ′) : U ′ ∈ E Set φ (U ) and f ′ = f ◦ φ−1 Since φU is a homeomorphism onto its image, we may write: U on φ(U )}. Eφ(f ) = {f ′ ∈ Cc(Y, U ′) : U ′ ∈ E Set φ (U ) and f ′ ◦ φU = f on U }. Lemma 1.5 says that for U ∈ φHo,Co and f ∈ Cc(X, U ), Eφ(f ) is nonempty. We call an element of Eφ(f ) an extension of f to Y via φ or, sometimes, simply an extension. Clearly, any two extensions of f agree on φ(U ) and equal f ◦ φ−1 U . Since the intersection of any two sets in E Set φ with φ(X) is φ(U ), we have f ′φ(X) = f ′′φ(X) = 0φ(X)(f ◦ φ−1 U ) for any two extensions f ′, f ′′ ∈ Eφ(f ). For U ∈ φHo , let U ′ ∈ E Set φ (U ). For g ∈ Cc(Y, U ′) we define the function Rφ(g) on X to be 0Y (g ◦ φU ), that is, Rφ(g) =(g ◦ φU 0 on U, on X − U The function Rφ(g) is continuous on U . Lemma 1.3 says that Rφ(g) ∈ Cc(X; U ) when U ∈ φHo,Co. Here R stands for 'restriction'. (5) Let X, Y and φ be as in (3) above. Analogous to Cc(X, U ), for U ⊆ X, we define Bor(X, U ) as the set of Borel measurable functions on X whose support lies in U . For U ∈ φHo and f ∈ Bor(X, U ), we define E Borel (f ) analogous to Eφ(f ) in (4) above but with Cc(Y, U ′) replaced by Bor(Y, U ′). Finally, observe that Rφ(g) makes sense for a Borel function g ∈ Bor(Y, U ′) where U ′ ∈ E Set φ (U ). φ PROPER TOPOLOGICAL CORRESPONDENCES 9 Now we start preparing for Lemma 1.3. Let X be a space, and Y and A its subspaces with A ⊆ Y . The set A is (quasi-)compact in Y if and only if it is so in ′, there X. To prove the if part, let {Uj ′} cover A, so do {Uj}. Let ′ of X with Uj is an open set Uj ′ is a finite subcover of Uj 1, . . . , Uj n be a finite cover of A in X, then Uj 1 A in Y . ′} be an open cover of A in Y . For each Uj ′ = Uj ∩ Y . Since {Uj ′, . . . , Uj 1 Conversely, let {Uj} be an open cover of A in X. Then Uj ′ := Uj ∩ Y are open sets in Y which cover A. Since A is compact in Y , we may choose a finite subcover, ′. Then Uj 1, . . . , Uj n is the finite subcover of A in X. We need if say Uj 1 part observation in the final argument of Lemma 1.3. Even later this observation shall be used. ′ . . . , Uj n Y Let X be a space, Y and W its subspaces with Y ⊇ W . In the next lemma, W and W denote the closures of W in X and Y , respectively. In general, W denotes the closure of W in the biggest space in the discussion that contains W . It is a basic result in point-set topology that W = W ∩ Y . Y Lemma 1.3. Let X and Y be locally compact Hausdorff spaces, and φ : X → Y be a local homeomorphism onto its image. Let U ∈ φHo,Co and V ∈ E Set φ (U ). Then for g ∈ Cc(Y, V ), Rφ(g) ∈ Cc(X; U ). A similar claim holds for g ∈ Bor(Y, V ), that is, for g ∈ Bor(Y, V ), Rφ(g) ∈ Bor(X; U ). Proof. The Borel part of the lemma follows easily from the definition of Rφ(g). We deal with the continuous case. Assume that gφ(U) ∈ Cc(φ(U )). Then Rφ(g) ∈ Cc(X; U ). Hence we shall show that gφ(U) ∈ Cc(φ(U )). We write Z for φ(X) in the discussion that follows. We need the following observation: for any subset A of V , A ∩ φ(U ) = A ∩ φ(U ). The proof of the observation is as follows: it is obvious that A ∩ φ(U ) ⊇ A ∩ φ(U ). One the other hand, A ∩ φ(U ) = (A ∩ V ) ∩ (φ(U ) ∩ Z) ⊆ A ∩ (V ∩ Z) = A ∩ φ(U ). Now the fact that g−1(C∗) ⊆ supp(g) ⊆ V and the observation in the last paragraph together give us that g−1(C∗) ∩ φ(U ) = g−1(C∗) ∩ φ(U ). But g−1(C∗) ∩ φ(U ) = {y ∈ φ(U ) : g(y) 6= 0} = g−1 g−1(C∗) ∩ φ(U ) = {y ∈ φ(U ) : g(y) 6= 0} = g−1 φ(U)(C∗) and (C∗). φ(U) Thus g−1 φ(U)(C∗) = g−1 φ(U) (C∗); denote either of these sets by B. Then supp(gφ(U)) = φ(U) B (1.4) and supp(gφ(U)) = B φ(U) . Furthermore, supp(gφ(U)) = B φ(U ) ∩ φ(U ). Now we claim that B φ(U) ⊆ V . This is because g−1 φ(U) (C∗) ⊆ g−1(C∗) and hence φ(U ) B = g−1 φ(U) (C∗) ∩ φ(U ) ⊆ g−1 φ(U) (C∗) ⊆ g−1(C∗) = supp(g) ⊆ V. Now Equation (1.4) and the observation made at the beginning of the proof together yield that supp(gφ(U)) = B φ(U ) ∩ φ(U ) = B φ(U) ∩ φ(U ) = B φ(U) = supp(gφ(U)). φ(U ) φ(U ) Note that since B is compact in φ(U ). This along with the observation prior to the statement of this lemma is a closed subset of the compact set φ(U ), B implies that B φ(U) = supp(gφ(U)) is compact in Y as well as φ(U ). (cid:3) 10 ROHIT DILIP HOLKAR Lemma 1.5. Let X and Y be locally compact spaces with Y Hausdorff. Let φ : X → Y be a local homeomorphism onto its image. Let U ∈ φHo,Co and U ′ ∈ E Set φ (U ). Then i) for f ∈ Cc(X; U ), Eφ(f ) is nonempty; ii) for f ∈ Cc(X; U ) and f ′ ∈ Eφ(f ), Rφ(f ′) = f ; iii) for g ∈ Cc(Y, U ′), g ∈ E(Rφ(g)). Similar claims for hold the measurable case. U ) lies (f ). (ii) and (iii) can be proved on the similar lines as the continuous case. Proof. Consider the measurable case first. i): For f ∈ Bor(X; U ), 0Y (f ◦ φ−1 in E Borel Now we prove the claims for the continuous case. i): We know that φ(U ) ⊆ Y is compact, therefore, φ f ◦ φ−1 φ(U ) : φ(U ) → C is a well-defined continuous function. We extend f ◦ φ−1 with f = f on φ(U ) using Tietze's extension theorem. φ(U) to a function f ∈ Cc(Y ) Let V ∈ E Set φ (U ). Then φ(supp(f )) ⊆ V is compact, since φ(supp(f )) is compact in the subspace φ(U ) of V . Let w ∈ Cc(Y ) be a function which is 1 on the compact set φ(supp(f )) and has support in V . Then f ′ := f w ∈ Cc(Y ) is an element of Eφ(f ): it is plain that f ′φ(U) = f φ(U) = f ◦ φ−1 U . Moreover, (supp(f ′) ∩ φ(X)) ⊆(cid:0)(supp( f ) ∩ supp(w)) ∩ φ(X)(cid:1) ⊆ supp(w) ∩ φ(X) ⊆ φ(U ). Thus f ′ ∈ Eφ(f ). ii): If f ′ ∈ Eφ(f ), then by definition f ′ ◦φU = f on φ(U ); hence Rφ(f ′) = f ′ ◦φU = f on φ(U ). Since f ∈ Cc(X; U ), f = 0 on X − U , and so is Rφ(f ′). iii): Lemma 1.4 says that Rφ(g) ∈ Cc(X; U ), hence we may define Eφ(Rφ(g)). The rest follows directly from the definitions of Rφ(g) and Eφ(f ). (cid:3) Restriction of a measure, and construction of a measures using local data: Let X be a locally compact, Hausdorff space and U an open subset of X. Identify Cc(X, U ) with Cc(U ) in the obvious way, see 1.2(1). Recall from elementary measure theory that the measure λ : Cc(X) → R on X, when restricted to the subspace Cc(X, U ) ≃ Cc(U ) gives a measure on U which is called the restriction of the λ to U . We denote this restriction by λU . In general, the measure λ can be restricted not only to an open set but also to any Borel set. Let (X, λ) be a Borel measure space and Y a Borel subset of X. Then any Borel subset U ⊆ Y is of type U = U ′ ∩ Y for a Borel U ′ ⊆ X. For Y ⊆ X as above, define λY , the restriction of λ to Y , by λY (U ) = λ(U ′ ∩ Y ) where U ⊆ Y and U ′ ⊆ X are Borel sets with U = U ′ ∩ Y . We shall often use the following result from elementary measure theory: let f be an integrable function on X then Let X and λ be as in the last paragraph, and let U and V be Borel subsets of X with V ⊆ U . It is easy to check that the restricted measures satisfy the equality λV = (λU )V . We will not need such a generalised notion of restricted measures, but we shall need the notion of restriction of a measure to a locally compact space, see [2, Nr. 7, §5, IV. 74]. The following proposition gives a criterion to extend measures defined on ele- ments of an open cover of X to whole of the space. Proposition 1.6 (Proposition [2, Chpater III, §2, 1, Proposition 1]). Let {Yα}α∈A be an open cover of X, X a locally compact Hausdorff space, and suppose given, ZX f Y dλ :=ZY f dλ =ZY f dλY . PROPER TOPOLOGICAL CORRESPONDENCES 11 on each subspace Yα, a measure µα, in such a way that for every pair (α, β), the restriction of µα and µβ to Yα ∩ Yβ are identical. Under these conditions, there is one and only one measure µ on X whose restriction to Yα is equal to µα for every index α. Lemma 1.7. Let X and Y be locally compact, Hausdorff spaces and let λ be a measure on Y . If φ : X → Y be a local homeomorphism (onto) Y , then λ induces a unique measure φ∗(λ) on X such that for every f ∈ Cc(X, U ), where U ∈ φHo, we have (1.8) ZX f φ∗(λ) =ZY 0Y (f ◦ φ−1 U ) dλ. Proof. We know that φHo is an open cover of X. For each U ∈ φHo define a measure φ∗(λ)U on U by φ∗(λ)U (f ) =Zφ(U) f ◦ φ−1 U dλ. If U, V ∈ φHo, then U ∩ V ∈ φHo. Furthermore, for f ∈ Cc(U ∩ V ), φ∗(λ)U (f ) = φ∗(λ)U∩V (f ) = φ∗(λ)V (f ). Now Proposition 1.6 says that there is a unique measure on X, which we denote by φ∗(λ), which has the property that restriction of φ∗(λ) to each U in φHo equals φ∗(λ)U . The result, now, follows from the observation that λφ(U)(f ◦ φ−1 U ) = λ(0Y (f ◦ φ−1 (cid:3) U )) for f ∈ Cc(X, U ) where U ∈ φHo. One may observe that Lemma 1.7 can be restated when the function f in Equa- tion (1.8) is an integrable function in Bor(X, U ) for U ∈ φHo. The proof is similar to the one above. Let Y be a locally compact, Hausdorff space and λ a measure on it. Then [1, Proposition 12, §9. 7, Chapter I] says that any locally compact Hausdorff subspace Z of Y is locally closed. Moreover, [1, Proposition 5, §3. 3, Chapter I] implies that Z is an intersection of a closed and open subsets of Y . Hence Z is a Borel subspace of Y and the restriction of λ to Z is defined. Corollary 1.9. Let X and Y be locally compact Hausdorff spaces, and λ a measure on Y . If φ : X → Y is a local homeomorphism onto its image, then λ induces a unique measure φ∗(λ) on X such that for every f ∈ Cc(X, U ) where U ∈ φHo,Co we have (1.10) ZX f dφ∗(λ) =Zφ(U ) f ′ dλ =Zφ(X) f ′ dλ for any extension f ′ ∈ Eφ(f ). A similar statement holds for f ∈ Bor(X, U ) and f ′ ∈ E Borel (f ). φ Proof. Let φ′ : X → φ(X) be the map obtained from φ by restricting the codomain; here φ(X) ⊆ Y is equipped with subspace topology. Then φ′ is a local homeomor- phism and hence open. Since the image of a locally compact space under an open map is locally compact, φ(X) is locally compact when equipped with the subspace topology from Y . Thus φ(X) is a locally compact subspace of Y . Now we bestow φ(X) with the restricted measure λφ(X) and apply Lemma 1.7 to φ′ : X → φ(X). This gives us the measure φ′∗(λφ(X)) which we denote by φ∗(λ). Equation (1.8) tells us that for f ∈ Cc(X; U ), where U ∈ φHo, (1.11) ZX f dφ∗(λ) =Zφ(X) 0φ(X)(f ◦ φ−1 φ(U)) dλ. 12 ROHIT DILIP HOLKAR Since 0φ(X)(f ◦ φ−1 φ(U)) ∈ Cc(φ(X), φ(U )), we can write ZX f dφ∗(λ) =Zφ(X) 0φ(X)(f ◦ φ−1 φ(U)) dλ =Zφ(U) 0φ(X)(f ◦ φ−1 φ(U)) dλ. Let f ′ ∈ Eφ(f ). Then Notation 1.2(4) tells us that f ′φ(X) = 0φ(X)(f ◦ φ−1 f ′φ(X) = (f ′φ(U))0 that U ) and φ(X). From this discussion and Equation (1.11) we may write ZX f dφ∗(λ) =Zφ(U) (f ◦ φ−1 U )0 φ(X) dλ =ZY f ′φ(X) dλ =ZY f ′φ(U) dλ. The claim of the lemma follows by observing that the last two terms in the above equation are Rφ(X) f ′ dλ and Rφ(U) f ′ dλ , respectively. The proof for the measurable case can be written along the same lines as above. (cid:3) Lemma 1.12. Let Y be a locally compact, Hausdorff space, λ a measure on it and Z a locally compact subspace of Y . Then the following are equivalent: i) For every nonnegative f ∈ Cc(Y ), RY f dλ =RZ f dλ. ii) For every f ∈ Cc(Y ), RY f dλ =RZ f dλ. iii) For every f ∈ L1(Y ), RY f dλ =RZ f dλ. iv) For every nonnegative f ∈ L1(Y ), RY f dλ =RZ f dλ. v) Every y ∈ Y has a neighbourhood V ⊆ Y such that λ(V ∩ (Y − Z)) = 0. Proof. (i) =⇒ (ii): Let f ∈ Cc(Y ). Put f + = max{f, 0} and f − = − min{f, 0} where 0 is the constant function 0 on Y . Then f + and f − are continuous, and they are compactly supported since supp(f +) and supp(f −) are contained in supp(f ). Furthermore, f = f + − f − . Hence ZY f dλ =ZY f + dλ −ZY f − dλ =ZZ f + dλ −ZZ f − dλ =ZZ f dλ. (ii) =⇒ (iii): Let f ∈ L1(Y ). Then f − f Z = f − f χZ is also in L1(Y ) and we may choose g ∈ Cc(Y ) such that RY (f − f Z) − g dλ < ǫ/2 for given ǫ > 0. Now (f − f Z) dλ (1.13) ZY g dλ =ZZ g dλ ≤ZZ =ZZ g − (f − f Z) dλ +ZZ g − (f − f Z) dλ ≤ZY g − (f − f Z) dλ < ǫ/2. In above computation, the first equality is due to the hypothesis, the second equality f dλ −ZY f Z dλ ≤ZY last inequality is because of the fact that hχZ ≤ h for any integrable function h ∈ L1(Y ). Now is due to the fact that RZ (f − f Z) dλ = RY χZ (f − f χZ) dλ = 0 and the second 0 ≤ ZY equation holds for any ǫ > 0, we have RY f dλ =RY f Z dλ. (iii) =⇒ (iv): Clear. (iv) =⇒ (v): For y ∈ Y , let V ⊆ Y be a relatively compact neighbourhood. Then λ(χV ) = λ(V ) ≤ λ(V ) < ∞; thus χV ∈ L1(Y ) is a nonnegative function. Using the hypothesis we compute The last inequality above is due to the choice of g and Equation (1.13). Since above (f − f Z) − g dλ +ZY (f − f Z) dλ ≤ZY g dλ < ǫ. λ(V ∩(Y −Z)) =ZY as χZ · χ(Y −Z) = 0. χV · χ(Y −Z) dλ =ZZ χV · χ(Y −Z) dλ =ZY χV · χZ · χ(Y −Z) dλ = 0, PROPER TOPOLOGICAL CORRESPONDENCES 13 (v) =⇒ (i): Let y ∈ Y and V a neighbourhood of y in Y such that λ(V ∩(Y −Z)) = 0. Let g ∈ Cc(Y, V ) be a nonnegative function then χV · χY −Z dλ = g∞λ(V ∩(Y −Z)) = 0. g dλ =ZY g· χV · χY −Z dλ ≤ g∞ZY ZY −Z Now the equality RY g dλ =RZ g dλ +RY −Z g dλ gives us that RY g dλ =RZ g dλ. Let f be a nonnegative function in Cc(Y ). For every y ∈ supp(f ), let Vy be a neighbourhood of y such that λ(Vy ∩ (Y − Z)) = 0. Because the support of f is compact and {Vy}y∈supp(f ) cover supp(f ), we may choose finitely many neighbour- hoods Vy1 , . . . , Vyn which cover supp(f ). Let u1, . . . , un be a partition of unity on supp(f ) subordinated to this finite cover. Then for every i ∈ {1, . . . , n}, f ui is a i=1 f ui using the result proved in nonnegative function in Cc(Y, Vxi ). Since f =Pn the previous paragraph we see that ZY f dλ =ZY n f ui dλ = Xi=1 n Xi=1ZY f ui dλ = n Xi=1ZZ f ui dλ =ZZ f dλ. (cid:3) Let Y be a locally compact, Hausdorff space and λ a measure on it. Let Z be locally compact subspace of Y . We say λ is concentrated on Y , if any one of the equivalent conditions in Lemma 1.12 holds. Condtion (v) in Lemma 1.12 is [2, Definition 4, Nr. 7, §5, V. 55]. Lemma 1.14. Let Y, Z and λ be as in Lemma 1.12 above. Let {Ui} be an open cover of Y . Then, for each i, λ is concentrated on Z if and only if the restricted measure λUi is concentrated on Ui ∩ Z for each i. Proof. If λ is concentrated on Z, then for every Ui and every f ∈ Cc(Y, Ui) ≃ Cc(Ui) we have ZUi f dλ =ZY f dλ f χUi dλ =ZY =ZZ f dλ =ZZ f χUi dλ =ZY f χZ∩Ui dλ =ZZ∩Ui f dλ, which proves that λUi is concentrated on Z ∩ Ui. Conversely, for given f ∈ Cc(Y ), cover its support by finitely many elements in {Ui}, say Ui1 , . . . , Uin where n ∈ N. Let u1, . . . , un be the partition of unity on supp(g) subordinated to the open cover {Ui1, . . . , Uin}. Then an argument as in the proof of the part (v) =⇒ (i) in Lemma 1.12 shows that RY f dλ =RZ f dλ. (cid:3) Notation 1.15. Let X, Y be spaces and φ : X → Y a continuous map which is also a local homeomorphism onto its image. Let A be an open cover of X. An extension of A via φ is a collection of open sets A′ of Y with the following properties (i) if U ′ ∈ A′, then either there is U ∈ A such that φ(U ) = φ(X) ∩ U ′, or φ(X) ∩ U ′ = ∅; (ii) for each U ∈ A, E Set φ (U ) ∩ A′ 6= ∅, that is, there is at least one set U ′ ∈ A′ such that U ′ ∩ φ(X) = φ(U ). That is, each set in A′ which intersects Im(φ) is an extension of a set in A via φ, and for given U ∈ A′, A′ contains at least one extension of U . Note that an extension via φ always exits for any cover of X. The collection A′ is called exhaustive if it covers Y . If φ is an inclusion map of a subspace, we shall simply say an "extension" or "exhaustive extension". 14 ROHIT DILIP HOLKAR Example 1.16. Let φ : X → Y be a local homeomorphism onto its image where X and Y are spaces. Assume that Im(φ) ⊆ Z is closed, then every open cover {Ui}i∈I of X has an exhaustive extension in Y which is {U ′ i ⊆ Y : i ∈ I, U ′ i is open in Y, and φ(Ui) = U ′ i ∩ Z} ∪ {Y − Im(X)}. As the next example shows, it is not true that any open cover of a subspace admits an exhaustive extension. Example 1.17. We show that there is a cover of the open subspace R − {0} of R which does not admit an exhaustive extension to R. Let {R−, R+} be a cover of R∗. Then this cover does not have an exhaustive extension to R. Given any open cover {Ui}i∈I of R, choose an index j ∈ I such that 0 ∈ Uj. Then Uj ∩ R− 6= ∅ and Uj ∩ R+ 6= ∅. Hence Uj ∩ R∗ * R− as well as Uj ∩ R∗ * R+. In fact, one may show that no open cover of R − {0} admits an exhaustive extension to R; the proof uses an argument similar to the one above. Lemma 1.18. Let X and Y be locally compact Hausdorff spaces, and λ a measure on Y . Let φ : X → Y be a local homeomorphism onto its image. Let φ∗(λ) be as in Corollary 1.9. Let A be an open of X consisting of sets in φHo,Co, and let A′ be an extension of A via φ. Assume that for every V ∈ A, f ∈ Cc(X, V ), every V ′ ∈ A′ which is an extension of V via φ, and any extension f ′ ∈ Cc(Y, V ′) of f via φ, we have (1.19) φ∗(λ)(f ) = λ(f ′). Then the following statements are true: i) For every V ∈ A and its extension V ′ ∈ A′, the measure λV ′ is concen- trated on φ(V ). ii) If B := ∪V ′∈A′ V ′, then B ⊆ Y is open and the measure λB is concentrated on φ(X). iii) If the open cover A′ is exhaustive, then λ is concentrated on φ(X). Proof. (i): Let g ∈ Cc(V ′). Then Lemma 1.5(iii) says that g is in Eφ(Rφ(g)) and the hypothesis tells us that ZX Rφ(g) dφ∗(λ) =ZY g dλ. ZX Rφ(g) dφ∗(λ) =Zφ(V ) g dλ Note that the last term equals RV ′ g dλ since g ∈ Cc(Y, V ′). Using Equation (1.10) we see that The above two equations imply thatRV ′ g dλ =Rφ(V ) g dλ for any g ∈ Cc(Y, V ′) ≃ Cc(V ′). Hence using Lemma 1.12(ii) we infer that λV ′ is concentrated on φ(V ). (ii): By definition, the space B is covered by V ′ ∈ A′ where V ′ is an extension of some V ∈ φHo,Co. For every V ′ like this, V ′ ∩ φ(X) = φ(V ). Now (i) of this lemma and Lemma 1.14 together yield the desired result. (iii): In this case, B = Y and the result follows from (ii) above. (cid:3) Let X be a space, and λ and µ measures on X. To say that λ is absolutely continuous with respect to µ, we write "λ ≪ µ" and to say that λ and µ are equivalent we write "λ ∼ µ". Lemma 1.20. Let X, Y, φ and λ be as in Lemma 1.18. Let µ be another measure on Y . i) If µ ≪ λ, then φ∗(µ) ≪ φ∗(λ). Moreover, dφ∗(µ) dφ∗(λ) = dµ dλ ◦ φ. PROPER TOPOLOGICAL CORRESPONDENCES 15 ii) If µ ∼ λ, then φ∗(µ) ∼ φ∗(λ). Proof. (i): Due to Corollary 1.9, for U ∈ φHo,Co and f ∈ Cc(X; U ), we have that ZX (1.21) for f ′ ∈ Eφ(f ). Let f (x) dφ∗(µ)(x) =Zφ(X) =Zφ(X) f ′(y) dµ(y) f ′(y) dµ dλ (y) dλ(y) F = f · dµ dλ ◦ φ be a function on X. Then F ∈ Bor(X; U ) is integrable. Now note that f ′ dµ E Borel φ (F ) and compute dλ ∈ ZX(cid:18)f dµ dλ ◦ φ(cid:19) dφ∗(λ)(x) =ZX =Zφ(X) =ZX F dφ∗(λ)(x) f ′(y) dµ dλ (y) dλ(y) f (x) dφ∗(µ)(x) (due to Equation (1.21)). The second equality in the above computation uses the measurable version of Equa- tion (1.10)). This shows that φ∗(µ) ≪ φ∗(λ) and dφ∗(µ) (ii): Proof of (i) above shows that dφ∗(µ) dµ dλ ◦ φ > 0 which is equivalent to saying that φ∗(λ) ∼ φ∗(µ). 1.2. Free, proper and transitive actions of groupoids. Let G be a topological groupoid and X a left G-space. Let Y be a space. We call a map f : X → Y G-invariant if f (γx) = f (x) for all (γ, x) ∈ G ×G(0) X. dλ ◦ φ. If µ ∼ λ, then dµ dλ > 0. Hence (cid:3) dφ∗(λ) = dµ dφ∗(λ) = dµ dλ ◦ φ. A map f : X → Y of locally compact spaces is called proper if the inverse image of a (quasi)-compact set in Y under f is quasi-compact. A proper map has various characterisations, for example, see [20, Proposition 1.6]. Definition 1.22. For a groupoid G, a left G-space X, and a G-invariant map f from X to a space Y , define the function mf : G ×G(0) X → X ×f,Y,f X by mf (γ, x) = (γx, x). A similar definition can be written for a right G-space. Let G be a groupoid acting on a space X. Then the map (γ, x) 7→ (γx, x) from G ×G(0) X → X × X can be realised as the one in Definition 1.22 by taking the space Y to be the singleton {∗} and f the constant map; in this case, we denote the map by m0. We call the map in Definition 1.22 "the map mf ". Let G be a topological groupoid, X a left G-space and m0 as in the above para- graphs. We call the action of G on X (i) free if the map m0 is one-to-one; (ii) proper if m0 is proper; (iii) basic if m0 is a homeomorphism onto its image (see [13]). It can be checked that the action of G on X is free if and only if for any two composable pairs (γ, x) and (γ′, x) in G ×G(0) X, γx = γ′x implies γ = γ′. Remark 1.23. Talking about a proper action, a proper map is closed, see [1, Propo- sition I, §10.1, Chapter I]. Thus for a groupoid G and a G-space X (i) the image of G ×G(0) X under m0 is a closed subset of X × X if the action of G on X is proper; (ii) the action of G on X is free as well as proper if and only if the image of m0 is a closed subset of X × X, and m0 is a homeomorphisms onto its image; the action is basic. The inclusion of the open interval (0, 1) ֒→ [0, 1] is a map which is homeomorphism onto its image but is not proper. in (0, 1) is not compact. 16 ROHIT DILIP HOLKAR Let G be a topological groupoid and X a left G-space. Let Y be a space, and let f : X → Y be a G-invariant map. Let [f ] : G\X → Y be the continuous map induced by f ; thus [f ]([x]) = f (x) for all x ∈ X. We always bestow G\X with the quotient topology. Lemma 1.24. Let G be a topological groupoid and X a left G-space. Let Y be a space and f : X → Y a G-invariant map. Then the following statements hold: i) The action of G on X is free if and only if the map mf : G ×G(0) X → X ×f,Y,f X is one-to-one. ii) The map [f ] : G\X → Y is surjective if and only if f : X → Y is surjective. If f is open, then so is [f ]. iii) [f ] : G\X → Y is one-to-one if and only in the map mf is surjective. Proof. (i): Note that the images of mf and m0 are the same subsets of X × X, since f is invariant. Let ι : X ×f,Y,f X ֒→ X × X be the inclusion map. Then mf = m0 ◦ ι. Since ι is one-to-one, mf is one-to-one if and only if m0 is one-to-one. (ii): The first claim follows from the fact that f and [f ] have the same images. The second claim follows from the universal property of the quotient map of topological spaces. (iii): Firstly, assume that [f ] is one-to-one. Let (x, z) ∈ X ×f,Y,f X be given, then f (x) = f (z). Now the fact that [f ]([x]) = f (x) = f (z) = [f ]([z]) along with the injectivity of [f ] implies that [x] = [z], that is, x = γz for some γ ∈ G. Thus (x, z) = (γz, z) = mf (γ, z). Conversly, assume that mf is surjective. Let [x], [z] ∈ G\X be such that [f ]([x]) = [f ]([z]). Then f (x) = f (z) and we see that (x, z) ∈ X ×f,Y,f X. Now the surjectivity of mf gives γ ∈ G such that (γz, z) = mf (γ, z) = (x, z), thus [x] = [z]. (cid:3) Definition 1.25. Let G be a groupoid, X a G-space, and Y a space. For a G-invariant map f : X → Y we say that the action of G on X is transitive over f if one of the following equivalent conditions hold: i) the map mf is surjective, that is, for all (x, x′) ∈ X ×f,Y,f X, there is γ ∈ G with γx′ = x; ii) the map [f ] : G\X → Y induced by f is one-to-one. Lemma 1.24 (iii) shows that the two conditions in Definition 1.25 are equivalent. We say that an action of a groupoid G on a space X is transitive if it is transitive over the constant map X → {∗}. A groupoid is transitive if and only if the obvious action of the groupoid on its space of units is transitive. Let G and H be groupoids, and X a G-H-bispace. The conditions that msX : G×G(0) X → X ×sX ,H(0),sX X is a homeomorphism, or the map [sX ] : G/X → H (0) is one-to-one (or a homeomorphism) appear in many works concerning equivalences, generalised morphisms, and actions of groupoids. We revisit these conditons in the following remark. Remark 1.26. Let X be a G-H-bispace for groupoids G and H. Let sX : X → H (0) be an open surection. Recall the following one-sided classical conditions in for a groupoid equivalence in [14], namely, i) The action of G on X is free. ii) The action of G on X is proper. iii) The map sX induces an isomorphism [sX ] : G/X → H (0). 1) Since sX is an open surjection, due to Lemma 1.24(ii). Hence Condition (iii) above is equivalent to, and can be replaced by, the one that the action of G on X is transitive over sX . PROPER TOPOLOGICAL CORRESPONDENCES 17 2) As the action of G on X is free and proper, Remark 1.23 implies that mf is an isomorphism onto its image Im(G ×G(0) X). But the action of G on X is transitive over sX , hence Im(G ×G(0) X) = X×sX ,H(0),sX , that is, msX is an isomorphism -- this is one of the classical conditions that is given as an alternate to the third one above, and as we see, in present situation, this condition also is equivalent to that the action of G on X is transitive over sX . 1.3. Locally free actions of groupoids. Let X be a left G-space where G is a groupoid. For x ∈ X, the isotropy group of x, which we denote by Fix(x)G, is the group {γ ∈ G : γx = x}. The element rX (x) is the unit in Fix(x)G. Let G and X be as above, and let γ ∈ G. We say that the action of G is locally free at γ if there is a neighbourhood U ⊆ G of γ with the property that for all η ∈ U and x ∈ X, γx = ηx implies γ = η. In this case, we say that ' γ acts freely on X in the neighbourhood U '. Caution: this nomenclature does not mean that for any two elements η, δ ∈ U and any x ∈ X with ηx = δx =⇒ η = δ; the statement is true if and only if one of η or η′ is γ. Lemma 1.27. Let G be a groupoid and X a G-space. Assume that the momentum map rX is surjective. Then the following statements are equivalent. i) The action of G on X is locally free at every u ∈ G(0). ii) The action of G on X is locally free at every γ ∈ G. iii) The isotropy group of every x ∈ X is discrete. Proof. We prove that (i) ⇐⇒ (ii) and (i) ⇐⇒ (iii). (i) ⇐⇒ (ii): The implication (ii) =⇒ (i) is clear. To prove the converse, let γ ∈ G be given. Let Z ⊆ G be an open neighbourhood of sG(γ) in which sG(γ) acts freely on X. Let multG : G ×sG,G(0),rG G → G denote the multiplication map, multG(η, η′) = ηη′. Then multG(γ−1, γ) = sG(γ). Using the continuity of multG, choose open neighbourhoods U1 ⊆ G of γ−1, and U2 ⊆ G of γ with multG(U1 ×G(0) U2) ⊆ Z. Let invG : G → G denote the inverting function η 7→ η−1. Then U = invG(U1) ∩ U2 is an open neighbourhood containing γ, to see this, we observe that γ = (γ−1)−1 ∈ invG(U1) and γ ∈ U2, hence γ ∈ U . And U is open because invG is a homeomorphism. We claim that γ acts freely on X in U . Let η ∈ U and x ∈ X be such that γx = ηx, that is, η−1γx = multG(η−1, γ)x = x. Since η ∈ U = invG(U1) ∩ U2, we have that η−1 ∈ U1. Then multG(η−1, γ) ∈ Z. But Z acts freely on X. Hence η−1γx = x implies η−1γ = rX (x), that is, γ = η. (i) ⇐⇒ (iii): Assume that the action of G is locally free at every γ ∈ G. Then, for each x ∈ X, we produce a neighbourhood of the unit rX (x) in Fix(x)G which contains only the unit. Let x ∈ X be given. Let U ⊆ G be a neighbourhood of the unit rX (x) ∈ Fix(x)G in which rX (x) acts freely on X. Then U ∩Fix(x)G = {rX (x)} because for every η ∈ U , ηx = x is equivalent to ηx = rX (x)x = x which implies that η = rX (x). Conversely, assume that Fix(x)G is discrete for all x in X. Let u ∈ G(0) be given. Then u = rX (x) for some x ∈ X. Let U ⊆ G be an open neighbourhood with U ∩ Fix(x)G = {rX (x)}. Then u acts freely on X in the neighbourhood U . Because if γ ∈ U with γy = y, then γ ∈ Fix(x)G. Thus γ ∈ (U ∩ Fix(x)G) = {rX (x)}. Hence γ = rX (x). (cid:3) Assume that G and X are as in Lemma 1.27. Then the proof of (i) =⇒ (ii) above shows that for x ∈ X the isotropy group Fix(x)G is discrete if and only if the action of G is locally free at rX (x) ∈ G(0), or equivalently, for x ∈ X the group Fix(x)G is discrete if and only if the action of G is locally free at γ ∈ G for some γ ∈ GrX (x). 18 ROHIT DILIP HOLKAR Definition 1.28. i) An action of a groupoid G on a space X is called locally free if any of (i) -- (iii) in Lemma 1.27 holds. ii) An action of a groupoid G on a space X is called locally locally free if every γ ∈ G has a neighbourhood U ⊆ G and every y ∈ X sG(γ) has a neighbourhood V ⊆ X such that for all η ∈ U and x ∈ V , γx = ηx implies γ = η. Clearly, a locally free action is locally locally free. For a group, the converse holds. To see it, one may use an argument as in the proof of Lemma 1.27(i) =⇒ (iii), and observe that all isotropy groups are discrete. Example 1.29. The action of R on the unit circle S1, r · e2πiθ = e2πi(θ+r) is not free; here r ∈ R and e2πiθ ∈ S1. But this action is locally free since Fix(e2πiθ)R ≃ Z is discrete for all e2πiθ ∈ S1. Let G be a groupoid acting on a space X. For γ ∈ G, we say that the action of G is strongly locally free at γ if there is a neighbourhood U ⊆ G of γ with the property that for all η, δ ∈ U and x ∈ X, ηx = δx implies that η = δ. We say that 'the neighbourhood U (of γ)' acts freely on X. Lemma 1.30. Let G be a groupoid and X a G-space. Assume that the momentum map rX is surjective. Let Y be a space and f : X → Y a G-invariant map. Then the following statements are equivalent. i) The action of G on X is strongly locally free at every u ∈ G(0). ii) Every unit u ∈ G(0) has a neighbourhood U in G such that such that the restriction of the map mf to U ×G(0) X is one-to-one. iii) The action of G on X is strongly locally free at every γ ∈ G. iv) Every element γ ∈ G has a neighbourhood U in G such that the restriction of the map mf to U ×G(0) X is one-to-one. Proof. It can be readily seen that (i) and (ii) are different phrasing of the same fact, and so are (iii) and (iv). The implication (iii) =⇒ (i) is obvious and the proof of its converse can be written on exactly the same lines as that of the proof of (ii) =⇒ (i) of Lemma 1.27. We sketch the proof roughly: let γ ∈ G. Let Z ⊆ G be an open neighbourhood of sX (γ) that acts freely on X. Let U1 and U2 be as in the proof of Lemma 1.27. Then as shown in the proof of Lemma 1.27, U = invG(U1) ∩ U2 is an open neighbourhood of γ in G. We claim that U acts freely on X. Let η, δ ∈ U , then ηx = δx ⇐⇒ δ−1ηx = x. As in the same proof above, we observe that δ−1η ∈ Z. And since Z acts freely on X, we must have δ−1η = rG(η), that is, η = δ. (cid:3) Definition 1.31. i) An action of a groupoid G on a space X is called strongly locally free if any one of (i)-(iv) in Lemma 1.30 above holds. ii) An action of a groupoid G on a space X is called locally strongly locally free if every γ ∈ G has a neighbourhood U ⊆ G and every y ∈ X sG(γ) has a neighbourhood V ⊆ X such that for all δ, η ∈ U and x ∈ V , ηx = δx implies η = δ. Clearly, a strongly locally free action is locally free as well as locally strongly locally free. As discussed on page 20, for groups, all these four notions of free action coincide. Example 1.32. Let G = {(0, 0)} ∪ {(1/n, i/n2) ∈ R2 : n ∈ N and 0 ≤ i ≤ n}, G(0) := {(0, 0)} ∪ {(1/n, 0) ∈ R2 : n ∈ N} PROPER TOPOLOGICAL CORRESPONDENCES 19 and equip both sets with the subspace topology coming from R2; then G(0) is a subspace of G. We think of G as a group bundle over the space G(0): for n ∈ N, the fibre over (1/n, 0) is {(1/n, i/n2) : 0 ≤ i ≤ n} = {(1/n, 0), (1/n, 1/n2), . . . , (1/n, n/n2)} which we identify with the finite cyclic group Z/(n + 1)Z of order n by the map i/n2 7→ [i] ∈ Z/(n + 1)Z. The fibre over the origin (0, 0) is the trivial group. One may see that this group bundle is a continuous group bundle. We make this bundle into a groupoid in the standard way: G(0) is a the space of units, the projection onto the first factor, G → G(0), (x, y) 7→ (x, 0), is the source as well as range map for G. The composite of (1/n, i/n2), (1/n, j/n2) ∈ G is (1/n, [i + j]/n2) and the inverse of (1/n, i/n2) is (1/n, (n + 1 − i)/n). For the fibre over zero, the operations are defined in the obvious way. Let X = G(0) and sX : G(0) → G the inclusion map. Then we claim that the trivial left action of G on X is locally free but not strongly locally free. The action is locally free because for every (x, 0) ∈ G(0) the isotropy group FixG((x, 0)) is discrete, to be precise, for n ∈ N FixG((1/n, 0)) ≃ Z/(n + 1)Z, and FixG((0, 0)) is the trivial group. Now Lemma 1.27 tells us that action of G is locally free. We claim that given no neighbourhood of (0, 0) in G acts freely on X. Let U be an open neighbourhood of (0, 0) and let U ′ ⊆ R2 be an open set such that U = U ′ ∩ G. Let ǫ > 0 be sufficiently small so that, W ′ := (−ǫ, ǫ) × (−ǫ, ǫ), an open square centred at the origin is contained in U ′. Put W = W ′ ∩ G, then W ⊆ U is an open set containing (0, 0). Since the sequences {(1/n, 1/n)}n∈N and {(1/n, 0)}n∈N converge to (0, 0) in G as well as R2, for a sufficiently large n we have (1/n, 0), (1/n, 1/n) ∈ W ⊆ U . Thus for (1/n, 0), (1/n, 1/n) ∈ G and for x := (1/n, 0) ∈ X, we have x = (1/n, 0) · x = (1/n, 1/n) · x; here · is the trivial action of G on X. Thus U does not act strongly freely on X. Furthermore, the contrapositive of Proposition 1.35 says that G is not étale. Thus G is a groupoid in which the fibre over every unit is discrete but the groupoid is not étale. In this article, locally strongly locally free actions are what interest us the most. Lemma 1.33. Let G be a groupoid and X a G-space. Assume that the momentum map rX is surjective. Let f : X → Y be a G-invariant map to a space Y . Then the following statements are equivalent. i) The action of G on X is locally strongly locally free at every γ ∈ G. ii) The map mf is locally one-to-one. Additionally, if mf is an open map onto its image, then the following statements are equivalent. iii) The action of G on X is locally strongly locally free at every γ ∈ G. iv) The map mf is a local homeomorphism onto its image. Proof. Assume that the action of G is locally strongly locally free at every element γ. Given a point (γ, x) ∈ G×G(0) X, we need to find a neighbourhood of it such that the restriction of mf to this neighbourhood is one-to-one. Let U be a neighbourhood of γ and V ⊆ X be the one of x with the property that for all δ, η ∈ U and x ∈ V , ηx = δx =⇒ η = δ. Then U ×G(0) V is the neighbourhood to which the restriction mf U× G(0) X is one-to-one. Conversely, let γ ∈ G is given. Choose x in X which is composable with γ; this can be done since rX is assumed to be surjective. Let W be an open neighbourhood 20 ROHIT DILIP HOLKAR of (γ, x) in G ×G(0) X such that mf W is one-to-one. Let U ⊆ G and V ⊆ X be open neighbourhoods of γ and x, respectively, such that U ×G(0) V is a basic open neighbourhood of (γ, x) and U ×G(0) B ⊆ V . Then U and V are the required neighbourhoods which prove the claim. The rest of the proof is an easy exercise. (cid:3) One can see that a free action is locally free, locally locally free, strictly locally free and locally strictly locally free. It is well-known that action of a group on a space is called locally free if the isotropy group of each point of the space is discrete, for example, see [4, Definition 11.3.7] which discusses the special case of locally free action of a group on a foliation. It can be shown that for groups a locally free action as in Definition 1.28 and a strictly locally free one as in Definition 1.31 are equivalent: assume that G is a group which acts on a space X and the action is locally free. Thus for each point in X the corresponding isotropy group is discrete. Let γ ∈ G be given. Choose x ∈ X and let U ′ ⊆ G be a neighbourhood of e, the unit of G, that intersects Fix(x)G only at e. Let U ⊆ U ′ be a neighbourhood such that U −1U ⊆ U ′. Then γU is the neighbourhood of γ which acts freely on X. This is because, if η, δ ∈ γU , then η′ := γ−1η, δ′ := γ−1δ and δ′−1η′ are in U ′. And then η′x = δ′x ⇐⇒ δ′−1η′x = x. Which implies η′ = δ′ and hence η = δ. The neighbourhood U of U ′ above can be constructed, for example, see [5, Proposition 2.1 b]. The proof above also shows that, for groups, a locally locally free action is locally strongly locally free. Let G and X be as above, and assume that the action of G on X is locally locally free. Let U ′ × V ⊆ G × X a basic neighbourhood of (e, x) ∈ G × X such that m0 is a one-to-one when restricted to U ×. Now the above proof goes through with X replaced by V . Thus for groups, the notion of locally free action is equivalent to a locally locally free action (see page 18) and a strongly locally free action, and a locally locally free action is same as a locally strongly free action. This allows us to conclude that for groups, a strongly locally free action is same as a locally strongly free action. This shows that the four notions of locally free action coincide for groups. The proofs above, however, do not work for groupoids. The reason is that there may be more than a single unit in a groupoid. As mentioned earlier, locally strictly locally free actions interest us the most. Since, in the case of groups, every locally free action is a locally strictly free action, we pocket a large class of examples of locally strictly locally free actions. SLF LF SLF LF LSLF LLF LSLF LLF Figure 1. The different types of free actions and their interrela- tions; the first square shows the case of groupoids, and the second one shows the case of groups; here S stands for strognly, L for locally and F for free. Speaking of the square for groupoid in Figure 1, Example 1.32 shows that a locally free action need not be strongly locally free. But as-of-now, we do not know if a locally locally free should be locally free, or if a locally strictly locally free actions should be strictly locally free. Now we turn our attention to a special class of groupoids, namely, étale groupoids. Here is an example first: PROPER TOPOLOGICAL CORRESPONDENCES 21 Example 1.34. Let Z/2Z = {−1, 1} act on [−1, 1] as ±1 · t = ±1 for all t ∈ [−1, 1]. This action is not free since ±1 ·0 = 0. But Fix(t)Z/2Z is discrete for each t ∈ [−1, 1]. Hence, this action is locally free. In general, if G is a discrete group, then any G action is locally free. Because if X is a G space, then for each x ∈ X, Fix(x)G ⊆ G is discrete. This example generalises very well to étale groupoids. A groupoid is called étale if its range map is a local homeomorphism. Recall from [17, Definition 1.2.6] that a locally compact groupoid is called r-discrete if its space of units is an open subset. A groupoid is étale if and only if it is r-discrete and has a Haar system (Proposition [17, 1.2.8]). Proposition 1.35. Let G be an r-discrete groupoid. Then the action of G on X is strongly locally free. Proof. We claim that the restriction of mf to G(0) ×G(0) X is a homeomorphism onto its image. Because the map (rX , Id) : X → G(0) ×G(0) X, x 7→ (rX (x), x), is a homeomorphism, and so is the diagonal embedding diaX : X ֒→ X ×f,Y,f X, x 7→ (x, x). Now observe that mf is the composite diaX ◦ (rX , Id)−1; this proves that claim. In this case, Lemma 1.30 implies that the action of G on X is strongly locally free. (cid:3) We conclude this section by introducing a notation. Let G be a locally compact groupoid, X a G-space, and let the G action be locally strongly locally free. Let Y be a space and f : X → Y a G-invariant map. Assume that mf is open onto its image. Then every (γ, x) ∈ G ×G(0) X has a basic neighbourhood A ×G(0) B ⊆ G ×G(0) X restricted to which mf is one-to-one. Thus G ×G(0) X has an open cover consisting of basic open sets restricted to which mf is a homeomorphism onto its image. One may choose an open cover where each basic open set is cocompact. Note that since X is Hausdorff, the fibre product G ×G(0) X is Hausdorff. Notation 1.36 (Notation and discussion). Let G be a locally compact groupoid and X a left G-space. Let Y be a space and f : X → Y be a G-invariant map. Assume that mf : G ×G(0) X → X ×f,Y,f X, (γ, x) 7→ (γx, y), is a local homeomor- phism onto its image , that is, the action of G on X is locally strongly locally free and m0 is open onto its image. Let U ⊆ G and V ⊆ X be open, Hausdorff sets with mU× G(0) V a homeomor- phism onto its image. Then for each (x, y) ∈ m(U ×G(0) V ), there is a unique γ ∈ U with γy = x. We define the function Sm(U× G(0) V ) : m(U ×G(0) V ) → G which sends (x, y) ∈ m(U ×G(0) V ) to the unique γ ∈ U such that x = γy. Note that if pr1 : G ×G(0) X → G is the projection on the first factor, then Sm(U× G(0) V ) = pr1 ◦ m−1 G(0) V ). This shows that Sm(U× G(0) V ) is continuous. m(U× We call Sm(U× may write m−1 m(U× the local section at U ×G(0) V . Using Sm(U× G(0) V ) we G(0) V ) G(0) V ) : m(U ×G(0) V ) → U ×G(0) V as m−1 m(U× G(0) V )(x, y) = (Sm(U× = (Sm(U× G(0) V )(x, y), y) G(0) V )(x, y), Sm(U× G(0) V )(x, y)−1x). 22 ROHIT DILIP HOLKAR If one writes γ(x,y) for Sm(U× (1.37) (1.38) m−1 m(U× G(0) V )(x, y), then the above formula looks better: G(0) V )(x, y) = (γ(x,y), y) = (γ(x,y), γ−1 (x,y)x). Equation (1.37) expresses the local section Sm(U× y whereas Equation (1.38) expreses it in terms of x. G(0) V ) of U ×G(0) V in terms of 1.4. Proper groupoids and cutoff functions. This subsection is based on Sec- tion 1.2 of [8]. Let G be a groupoid. We call G proper if the map sG × rG : G → G(0) × G(0), sG × rG(γ) = (sG(γ), rG(γ)) is proper. An action of G on a space X is proper if the transformation groupoid G ⋉ X is proper; this definition is equivalent to the one given after Definition 1.22 on page 15. It is well-known that if the action of G on X is proper, then G/X inherits many good topological properties from X, that is, if X is locally compact (Hausdorff or paracompact or second-countable), then G/X is also locally compact (respectively, Hausdorff or paracompact or second-countable). For a groupoid G, the space of units G(0) carries an action of G for which sG is the momentum map, and the action is that for u ∈ G(0) and γ ∈ Gu γu = rG(γ). It can be checked that G is proper if and only if this action of G on G(0) is proper. We call this action the left action of G on its space of units. The right action of G on its space of units is defined similarly. Let G be a groupoid. For nonempty subsets V, W ⊆ G(0), we denote the set {γ ∈ G : rG(γ) ∈ V and sG(γ) ∈ W } by GV V is a groupoid called the restriction groupoid for the subset V ⊆ G(0). The space of units of the restriction groupoid GV V is a topological subgroupoid of G when V bestowed with the subspace topology of G. W . If W = V , then GV is V . GV (0) Lemma 1.39. Let G be a proper groupoid and V a subset of G(0). i) If G is proper the so is the restriction groupoid GV V . ii) The inclusion map V ֒→ G(0) induces a topological embedding G/V ֒→ G/G(0). V = (rG ×sG)−1(V ×V ). Now the claim of the lemma Proof. (i) Firstly, note that GV follows from [1, Chapter I, §10 No. 1, Proposition 3(a)]. (ii) First of all observe that for every v ∈ V the equivalence class of v in G/V and G/G(0) are the same sets. Thus the inclusion V ֒→ G(0) does induce a well-defined one-to-one map G/V ֒→ G/G(0). This map is continuous because of the universal property of the quotient space, and is open onto its image because V ֒→ G(0) is so. (cid:3) Let G and V be as in Lemma 1.39. Assume that rG is an open map. If V is an open subset of G(0), then (ii) of Lemma 1.39 says that G/V is homeomorphic to an open subset of G/G(0) since the quotient map G(0) → G/G(0) is open in this case. Thus if G/G(0) is paracompact, then G/V is paracompact. Lemma 1.40. Let G be a locally compact proper groupoid with the range map open. Assume that G/G(0) is paracompact for the action of G on its space of units G(0). Then there is a continuous real-valued function F on G(0) which has the following properties i) F is not identically zero on any equivalence class for the action of G on G(0); ii) for every compact subset K of G/G(0), the intersection of r−1 G (K) with supp(F ) is compact. PROPER TOPOLOGICAL CORRESPONDENCES Proof. This proof follows from that of [8, Lemma 1.7]. 23 (cid:3) Let (G, α) be a locally compact proper groupoid with a Haar system. Then rG is open. Let F be a function on G(0) as in Lemma 1.40. Now the proof of [8, Lemma 1.7] shows that the function ¯F on G(0) defined as ¯F (u) =ZG F ◦ sG(γ) dαu(γ) cG := ◦ sG F ¯F is continuous and 0 < ¯F (u) < ∞ for all u ∈ G(0). The discussion in the same proof also shows that the function on G is continuous has the property that for all u ∈ G(0). We call cG a cutoff function for (G, α). cG(γ) dαu(γ) = 1 ZGu Example 1.41. Let (H, β) be a locally compact groupoid with a Haar system and X a proper rigt H-space with X/H paracompact. Then X ⋊ H is a locally compact proper groupoid. Recall that for (x, η) ∈ X ⋊ H, (x, η)−1 = (xη, η−1), rX⋊H (x, η) = x, and sX⋊H (x, η) = xη. It is well known that the Haar system β induces a Haar system β for X ⋊ H: for x ∈ X = (X ⋊ H)(0) and f ∈ Cc(X ⋊ H) Let c be a function on X as in Lemma 1.40. Then f (x, η) dβsX (x)(η). ZX⋊H ¯c(u) =ZX⋊H f d βu :=ZH c ◦ sX⋊H(x, η) d βu(η) =ZH c(xη) dβu(η), where u ∈ X, is the function similar to ¯F above and cX⋊H = c ¯c ◦ sX⋊H is a cutoff function like cG above. Note that ¯c is H and X ⋊ H-invariant. In this example, we call the function c a pre-cutoff function and c/¯c the normalised pre- cutoff function corresponding to c. Thus, the cutoff function cX⋊H is the composite of the the normalised pre-cutoff function corresponding to c and the source map of the groupoid X ⋊ H. We shall use this example and notation in (4) of the proof of the main theorem, Theorem 2.11. 1.5. Compact operators. Let (H, β) be a locally compact groupoid with a Haar system. Let (X, λ) be a pair consisting of a proper right H-space X and an H-invariant family of measures on λ along the momentum map sX . For ζ, ξ ∈ Cc(X) and g ∈ Cc(H) define (1.42) (ζ · g)(x) :=ZHsX (x) hζ, ξi(η) :=ZXrH (η) ζ(xη) g(η−1) dβsX (x)(η) and ζ(x)ξ(xη) dλrH (η)(x).   Using the above equations one may make Cc(X) into an inner product Cc(H)-module which completes to a Hilbert C∗(H, β)-module; see [18, Corollaire 5.2] for details. We denote this Hilbert C∗(H, β)-module by H(X). Let (X, λ) and (H, β) be as above. Then X ×sX ,H(0),sX X carries the diagonal action of H which is proper; for (x, y) ∈ X ×sX ,H(0),sX X and η ∈ H sX (x) the action 24 ROHIT DILIP HOLKAR is (x, y)η = (xη, yη). Let G denote the quotient space (X ×sX ,H(0),sX X)/H. For f, g ∈ Cc(G) define (1.43) (1.44) f ∗ g([x, z]) =ZX f ∗([x, y]) = f [y, x]. f ([x, y])g([y, z]) dλsX (z)(y) Then f ∗ is in Cc(G) and [10] shows that f ∗g is well-defined and lies in Cc(G). When equipped with the convolution in Equation (1.43) and the involution (1.44), Cc(G) becomes a ∗-algebra. [10, Theorem 2.2] shows the representations of the ∗-algebra Cc(H) induces representations of Cc(G). Using these induced representations we complete Cc(G) to a C∗-algebra which we denote by C∗(G); see [10] for details. In [19], Renault shows that the quotient space G and the above C∗-algebra corresponding to it carry a well-defined mathematical structure. He shows that G is the spatial hypergroupoid associated with the proper H-space X. The H-invariant family of measures λ induces a Haar system λG on G and the algebra C∗(G) is the C∗-algebra C∗(G, λG) of the hypergroupoid G equipped with the Haar system λG. Remark 1.45. [A long remark that explains the origins of Equations (1.43),(1.44) and (1.47).] We advise reader to have look at Proposition 1.46 and its proof until Equa- tion (1.47) below before reading this remark further. In this remark, we discuss a ∗-category studied in [10] and its relation to Equations (1.43),(1.44) and (1.47). For a locally compact groupoid H equipped with a Haar system β, one may form a ∗-category Cc(H, β) as follows: the objects of Cc(H, β) are the vector spaces Cc(X, λ) where (X, λ) is a proper H-space with an H-invariant family of measures. The arrows from Cc(X, λ) to Cc(Y, µ), where Cc(X, λ), Cc(Y, µ) ∈∈ Cc(H, β)0, are func- tions in Cc((X ×sX ,H(0),sY Y )/H). The involution of f ∈ Cc((X ×sX ,H(0),sY Y )/H) is f ∗([x, y]) = f ([y, x]). Let Cc(Z, ν) be one more object in Cc(H, β). If f : Cc(X) → Cc(Y ) and g : Cc(Y ) → Cc(Z) are arrows in Cc(H, β), then their composite arrow, which goes from (X, λ) to (Z, ν), is denoted by f ∗ g and defined as f ∗ g[x, z] =ZY f [x, y]g[y, z] dµsz(z)(ψ). When equipped with the involution in Equation (1.44), Cc(H) becomes a ∗-category. See [10] for details. [10, Theorem 2.2] shows that that the representations of (H, β) induce the ones of Cc(H) which makes Cc(H) a normed ∗-category. Using these induced represen- tations, we complete Cc(H) to a C∗-category C(H, β). Let (X, λ) ∈∈ Cc(H, β). Note that Cc(H, β−1) is also an object in Cc(H, β). Now it can be checked that Equation (1.43) gives the composition of two arrows from Cc(X, λ) to itself and Equation (1.44) is an involution of an arrow from Cc(X, λ) to itself. Furthermore, [7, Table 3.2] shows that Equation (1.47) below is equivalent to the composition of certain arrows in Cc(H). This explains the origins and well- definedness of Equations (1.43),(1.44) and (1.47). Proposition 1.46. Let (X, λ) be a proper H-space equipped with an H-invariant family of measures. Let C∗(G) be the C∗-algebra obtained by completing the ∗-algebra Cc(G) as in [10]. Then C∗(G) ≃ K(H(X)). Proof. We show that C∗(G) and C∗(H, β) are Morita equivalent and H(X) is the imprimitivity bimodule; this implies that C∗(G) ≃ K(H(X)). The isomorphism C∗(G) ≃ K(H(X)) is given by the representation of C∗(G) on H(X). PROPER TOPOLOGICAL CORRESPONDENCES 25 For f ∈ Cc(G), ξ, ζ ∈ Cc(X) and g ∈ Cc(H) let ζa and hζ, ξi be as in Equa- tion 1.42. Furthermore, define (1.47) f ζ(x) =ZX Ghζ, ξi =ZH   f [x, y] ζ(y) dλsX (x)(y) and ζ(xη) ξ(yη) dβsX (x)(η). The integrals in Equation 1.47 are well-defined, in addition to which, f ζ ∈ Cc(X) and Ghζ, ξi ∈ Cc(G), see Remark 1.45 above. Equation (1.47) defines a representa- tion of the ∗-algebra Cc(G) on Cc(X) and a Cc(G)-valued inner produce on Cc(X), respectively. Completing the ∗-category Cc(H, β) to the C∗-category, C(H, β), we get a representation of C∗(G) on H(X) and the Cc(G)-valued inner product makes H(X) into a left Hilbert C∗(G)-module. It is a standard computation to check that the actions of Cc(G) and Cc(H) on Cc(X) commute in the usual sense, that is (f ξ)g = f (ξg) for all f ∈ Cc(G), ξ ∈ Cc(X) and g ∈ Cc(H). Thus H(X) is a C∗(G)- Cst(H, β)-bimodule which carries C∗(G) and C∗(H, β)-valued inner products h, i and Gh, i, respectively. Since λ is a proper family of measures along sX , the set {hζ, ξiζ, ξ ∈ Cc(X)} is dense in Cc(H, β), see [18, Lemme 1.1]. The set {Ghζ, ξi : ζ, ξ ∈ Cc(X)}, that is, the image of Cc(X) ⊗C Cc(X) in Cc(G) under the map (ζ, ξ) 7→ Ghζ, ξi is dense: this follows from the a standard argument that uses the family of measures in Equation (1.1). The the argument is as follows: The restriction map mapping ζ ⊗ξ ∈ Cc(X)⊗CCc(X) to ζ ⊗sX ,H(0),sX ξ ∈ Cc(X ×sX ,H(0),sX X) has a dense image due to the Stone-Weirstrass theorem. The Haar system β induces the family of measures βX× X along the quotient map X ×sX ,H(0),sX X → G similar to the one in Equation (1.1). Hence the image X : Cc(X ×sX ,H(0),sX X) → Cc(G), induced by of the integration map BX× X , is dense. the family of measures βX× sX ,H(0) ,sX sX ,H(0) ,sX sX ,H(0) ,sX Let ι be the bijection from Cc(X) ⊗C Cc(X) itself given by ι : ζ ⊗C ξ 7→ ζ ⊗C ξ. sX ,H(0) ,sX Then we observe that {Ghζ, ξi : ζ, ξ ∈ Cc(X)} ⊆ Cc(G) is the image of the compos- ite BX× X ◦ ι. Hence {Ghζ, ξi : ζ, ξ ∈ Cc(X)} ⊆ Cc(G) is dense. To prove that H(X) implements Morita equivalence between the two C∗-algebras, we need to check that the equality Ghζ, ξi · θ = ζ · hξ, θi holds for every ζ, ξ and θ in H(X). To check this let ζ, ξ, θ ∈ Cc(X), then for any x ∈ X Ghζ, ξi[x, y] θ(y) dλsX (x)(y) (1.48) ζ(xη) ξ(yη) θ(y) dβsX (x)(η) dλsX (x)(y), and Ghζ, ξi · θ (x) =ZX =ZXZG ζ · hξ, θi (x) =ZG =ZGZX ζ(xη) hξ, θi(η−1) dβsX (x)(η) ζ(xη) ξ(y) θ(yη−1) dλsX (x)(y) dβsX (x)(η). Change the variable y 7→ yη in the last term above and use the right invariance of the family of measures λ to see that the equation transforms to (1.49) ζ · hξ, θi (x) =ZGZX ζ(xη) ξ(yη) θ(y) dλsX (x)(y) dβsX (x)(η). 26 ROHIT DILIP HOLKAR Now use Fubini's theorem in the above equation to interchange the integrals to see that Equations (1.48) and (1.49) are same. Thus Ghζ, ξi · θ = ζ · hξ, θi for all ζ, ξ and θ ∈ Cc(X). One may see that the C∗-category in [9] allows to extend above equality to C∗(G), H(X) and C∗(H, β) which proves that H(X) is the imprimitivity bimodule between C∗(G) and C∗(H, β). (cid:3) Notation 1.50. Let X, H, β and λ as in Proposition 1.46. Then we denote the hypergroupoid G = (X ×sX ,H(0),sX X)/H in the proof of the proposition by KT(X); the suffix 'T' stands for topological. Hence we change σG and Gh, i change to σKT(X) and KT(X)h, i, respectively, in the further writing. 2. Proper topological correspondences In this section, we define a proper topological correspondence, see Definition 2.5. Section 2.1 discusses some details that may proof helpful to understand the defini- tion of a proper topological correspondence. Without these details, the definition may look artificial or technical. We hope that Example 2.3 in Section 2.1 should setup the background for Definition 2.5. 2.1. Some examples. Example 2.1 revises some well-known facts about transfor- mation groupoids; Example 2.2 discusses the groupoid associated with a continuous equivalence relation and Example 2.3 puts these two examples together to elaborate the background in the definition of a proper correspondence. Example 2.1 (Transformation groupoid). Let G be a locally compact groupoid and X a left G-space. For (γ, x) ∈ G ⋉ X, rG⋉X (γ, x) = (rG(γ), γx) and sG⋉X (γ, x) = (sG(γ), x). We often identify the space of units G(0) ×Id G(0) ,G(0),rX X of this transfor- mation groupoid with X in which case, the range and source of the above element are rG⋉X (γ, x) = γx and sG⋉X (γ, x) = x. Given a unit (rX (x), x) ∈ G ⋉ X (0), r−1 G⋉X (rX (x), x) = {(γ, γ−1x) : rG(γ) = rX (x)}. It is well known that a Haar sys- tem on G induces one on the transformation groupoid G ⋉ X; if α is a Haar system on G, then the Haar system on G ⋉ X, which we denote by α2, is given by ZG⋉X f dα2 x :=ZG f (γ, γ−1x) dαrX (x)(γ) for f ∈ Cc(G ⋉ X) and x ∈ X ≈ (G ⋉ X)(0). Let H be another groupoid. Assume that H acts on X from right. Then (γ, x)η 7→ (γ, xη), for (γ, x) ∈ G ⋉ X and (x, η) ∈ X ×H(0) H, is a right action of groupoid H on groupoid G ⋉ X by invertible functors, see [9, Definition 1.8 and Example 1.11]. Example 2.2 (The groupoid of an equivalence relation). Let X and Y be spaces, and φ : X → Y a continuous map. Then the fibre product X ×φ,Y,φ X := {(x, x′) : φ(x) = φ(x′)} carries a groupoid structure: (x, x′) and (x′′, x′′′) in X ×φ,Y,φ X are composable if and only if x′ = x′′ and the composite (x, x′)(x′, x′′′) = (x, x′′′). The space of units of Xφ is the diagonal {(x, x) ∈ X ×φ,Y,φ X : x ∈ X} in X ×φ,Y,φ X which can be identified with X. For the element (x, x′) as above, its (i) source is x′, (ii) range is x and (iii) inverse (x, x′)−1 = (x′, x). We denote this groupoid by Xφ; this is the groupoid of the equivalence relation x ∼ x′ if and only if φ(x) = φ(x′). In addition to the above data, let λ = {λy}y∈Y be a continuous family of mea- sures with full support along φ. Then φ induces a Haar system λ1 on Xφ; for PROPER TOPOLOGICAL CORRESPONDENCES 27 f ∈ Cc(Xφ) and z ∈ X ≃ Xφ (0), ZXφ z :=ZX f dλ1 f (z, x) dλφ(z)(x). Note that λz 1 = δz ⊗ λφ(z) where δx is the point mass at x ∈ X. The family of measures λ1 is a continuous family of measures due to [18, Lemme 1.2] and has full support because λ has full support. It is easy to see that λ1 is left invariant: for f ∈ Cc(Xφ) and (v, w) ∈ Xφ f ((v, w)(w, x)) dλ1 f ((v, w)(w, x)) dλ1 w(w, x) rXφ (v,w)(w, x) =ZXφ ZXφ =ZX f (v, x) dλφ(w)(x) =ZX f (v, x) dλφ(v)(x) =ZXφ f (v, x) dλ1 rXφ (v,w)(v, x). In the above computation, the third equality due to the fact that φ(v) = φ(w). If λ does not have full support, then λ1 is a continuous family of left invariant measures and not a Haar system. Additionally, if we assume that X and Y are right H-spaces for a groupoid H, and φ is an H-equivariant map, then (x, y)η 7→ (xη, yη), for (x, y) ∈ X ×φ,Y,φ X and η ∈ X ×H(0) H, is an action of H on the groupoid Xφ in the sense of [9, Definition 1.8]. The momentum map rXφ,H : Xφ → H (0) of this action is rXφ,H(x, y) = sX (x). If λ is an H-invariant family of measures, then a routine computation shows that the family of measures λ1 is also H-invariant. The example ends here. We need the following remark: it is well-known that the right action rH (η)η = sH (η) of a groupoid H on its space of units H (0) is well-defined continuous ac- tion. Moreover, for any right H-space X, the momentum map sX : X → H (0) is H-equivariant for this action. Example 2.3. Let G and H be locally compact groupoids, and X a G-H-bispace. Let α be a Haar system on G and λ an H-invariant continuous family of measures on X along sX . (1) Then sX : X → H (0) is a G-invariant map. Let XsX be the groupoid associated It is a routine with the continuous map sX : X → H (0) as in Example 2.2. calculation to check that the map msX : G ⋉ X → XsX , (γ, x) 7→ (γx, x), is a homomorphism of groupoids. Furthermore, since X is a G-H-bispace, both groupoids, G⋉X and XsX , carry actions of H described in Examples 2.2 and 2.1 and the map msX is an H-equivariant homomorphisms of groupoids. Note that msX is the restriction of rG⋉X × sG⋉X : G ⋉ X → X × X to the closed subspace X ×sX ,H(0),sX X ⊆ X × X. Identify X with the space of units of G ⋉ X as well as XsX . Then (a) msX G⋉X (0) = IdX : X → X, (b) rG⋉X = rXsX The last equality above implies that m−1 ◦ msX . G⋉X (x) for all x ∈ X. (2) The Haar system α on G induces the Haar system α2 on G ⋉ X (see Exam- ple 2.1). As in Example 2.2, λ induces a continuous invariant family of measures λ1 on the groupoid XsX . Since λ is H-invariant, so is λ1. If λ has full support, then λ1 is a Haar system for XsX . sX ◦ r−1 XsX (x) = r−1 (3) In addition to the current hypotheses, assume that the action of G on X is locally strongly locally free and msX is open onto its image, that is, msX is a local homeomorphism onto its image (see Lemma 1.33). Let D : X ×sX ,H(0),sX X → R be a nonnegative H-invariant continuous function. Then, for each 28 ROHIT DILIP HOLKAR 1 on X x x ∈ X ≈ X (0) to λx D(x, y). Let Dλ1 := {D(x, _)λx X ≈ X (0) on the fibre X x x ∈ X ≈ X (0) be precise, on r−1 sX , let D(x, _)λx sX , D(x, _)λ1 sX . Furthermore, the Radon-Nikodym derivative dD(x,_)λx x is an absolutely continuous measure with respect (x, y) = 1 }x∈X. We abuse the notation: for each x ∈ 1 denote the measure on XsX which equals D(x, _)λx 1 sX and zero outside it. Now Corollary 1.9 says that for each 1 ) on G ⋉ X, to x induces a measure m∗ sX (D(x, _)λx sX , D(x, _)λ1 G⋉X (x) ⊆ G ⋉ X. dλx 1 1 Since D is H-equivariant, so is the family of measures Dλ1. Recall from (1) above that msX is an H-equivariant map. The H-equivariance of msX and Dλ1 implies that the family of measures m∗ sX (Dλ1) := {m∗ sX (D(x, _)λx 1 )}x∈X is also H-invariant -- this claim follows from a direct computation. Lemma 1.33 describes this family of measures: for U ∈ mHo,Co and f ∈ Cc(G ⋉ X, U ), sX (2.4) m∗ sX (D(x, _)λx 1 )(f ) = (D(x, _)λx 1 )(f ′) =ZX f ′(x, y)D(x, y) dλsX (x)(y) for any extension f ′ ∈ EmsX (f ). (4) In Definition 2.5, one of the conditions demands that (m∗ sX (D(x, _)λx 1 ) = αx 2 for each x ∈ X. Recall from (3) above that D(x, _)λx 1 ≪ λx 1 and the Radon-Nikodym deriva- tive d(D(x,_)λx 1 ) m∗ sX (λx dλx 1 1 ), and, the Radon-Nikodym derivative (x, _) = D(x, _). Lemma 1.20 says that α2x = m∗ sX (D(x, _)λx 1 ) ≪ dαx 2 sX (λx 1 ) dm∗ = D ◦ msX . (5) Note that, for U ∈ mHo sX and f ∈ Cc(G ⋉ X, U ), we may write f ◦ msX −1 U (x, y) = f (γ−1 (x,y), y) using Equation 1.37 or f ◦ msX −1 U (x, y) = f (γ(x,y), γ−1 (x,y)x) using Equation 1.38. We use the second form of f ◦ msX −1 U in Definition 2.5. 2.2. Proper correspondences. As advised in the Introduction, reader may elim- inate the functions D and the adjoinig function ∆ in this section during the first reading -- one may consider D and ∆ are the constant function 1, and skip the remarks discussing these functions. Definition 2.5. Let (G, α) and (H, β) be locally compact groupoids with Haar systems, and (X, λ) a topological correspondence from (G, α) to (H, β). Let X be Hausdorff and X/H paracompact. We call (X, λ) proper if there is an H-invariant nonnegative continuous function D : X ×sX ,H(0),sX X → R and following holds: i) the left action of G on X is locally strongly locally free, and the map msX is open onto its image, ii) the map [rX ] : X/H → G(0) induced by the left momentum map is proper, iii) there is an open cover A of G ×G(0) X consisting of sets in mHo,Co and there is an extension A′ of A via msX with the property that for all U ∈ A, any sX PROPER TOPOLOGICAL CORRESPONDENCES 29 extension U ′ ∈ A′ of it via msX , given f ∈ Cc(G ⋉ X, U ), and given any extension f ′ ∈ Cc(X×sX ,H(0),sX , U ′) of f , the following equality holds (2.6) ZX f ′(x, y) D(x, y) dλsX (x)(y) =ZG for every x ∈ X. f (γ(x,y), γ−1 (x,y)x) dαrX (x)(γ) We make few remarks. In the following remarks, we denote restriction of a measure µ to a set U by µU instead of µU . We adopt this convention only for the following discussion. Lemma 1.33 implies that the first condition in Definition 2.5 is equivalent to msX is a local homeomorphism. Thus (iii) in Definition 2.5 makes sense, see Example 2.3 for details. Equation (2.4) in that example says that the left side of Equation (2.6) is m∗ 1 )(f ), and thus Equation (2.6) says that sX (D(x, _)λx (2.7) m∗ sX (D(x, _)λx 1 )(f )(x) = αx 2 (f )(x) for all f ∈ Cc(G ⋉ X, U ) where U ∈ A ⊆ sX Now Equation (2.7) along with Lemma 1.18 tell us that for U ′ ∈ E Set the measure D(x, _)λx is the measure D(x, _)λx discussed in Example 2.3, the measure αx Radon-Nikodym derivative Ho,Co and x ∈ X ≈ G ⋉ X (0) ≈ X (0) sX . sX (U ) ∩ A′, 1 U ′ is concentrated on msX (U ); Lemma 2.8 discusses when 1 is concentrated on msX (G ×G(0) {x}). Furthermore, as 1 ) and the sX (D(x, _)λx 1 ) ≪ m∗ 2 = m∗ sX (λx dm∗ sX (λx 1 ) dαx 2 = D(x, _) ◦ msX ◦ invG⋉X , see the composite map in (4) of Example 2.3. In Definition 2.5, if the open cover {msX (U )}U∈mHo,Co of msX (G ⋉ X) admits an extension to X ×sX ,H(0),sX X, then, due to Lemma 1.18, we may conclude that the measure D(x, _)λx 1 is concentrated on the image of msX . sX Lemma 2.8. In Definition 2.5, if the image of msX is a closed subset of X×sX ,H(0),sX X, then the measure D(x, _)λx 1 is concentrated on the image of msX . Proof. This follows from Example 1.16 and the remark immediately above this lemma. (cid:3) A particular instance when the hypothesis of Lemma 2.8 is fulfilled is when msX is a proper map. We encounter examples of this type when msX is a homeomorphism or homeomorphism onto a closed subspace of X ×sX ,H(0),sX X. Remark 2.9. Let (X, λ) be a topological correspondence from a groupoid with a Haar system (G, α) to another one (H, β). If X is H-compact, that is X/H is compact, then the map [rX ] : X/H → G(0) induced by the right momentum map rX is proper. On the other hand, assume that the space of the units of the groupoid G is compact, for example, G is a group. If [rX ] is a proper map, then X/H = [rX ]−1(G(0)) is a compact space. Thus when G(0) is a compact space, Condition (ii) in Definition 2.5 is equivalent to that the quotient space X/H is compact. Now we prepare to state the main theorem. Let (G, α) and (H, β) be locally com- pact groupoids with Haar systems, and (X, λ) a topological correspondence from (G, α) to (H, β) with ∆ as the adjoining function. This gives a C∗-correspondence H(X) : C∗(G, α) → C∗(H, β) which is a certain completion of Cc(X). From [9] , recall the formulae for the actions of Cc(G) and Cc(H) on Cc(X), and the 30 ROHIT DILIP HOLKAR Cc(H)-valued inner product on Cc(X) which gives this C∗-correspondence: b ∈ Cc(G), ζ, ξ ∈ Cc(X) and a ∈ Cc(H) for b(γ) ζ(γ−1x) ∆1/2(γ, γ−1x) dαrX (x)(γ), ζ(xη) a(η−1) dβsX (x)(η) and ζ(x)ξ(xη) dλrH (η)(x). (2.10) (b · ζ)(x) :=ZGrX (x) (ζ · a)(x) :=ZHsX (x) hζ, ξi(η) :=ZXrH (η)   Denote the ∗-representation of the pre-C∗-algebra Cc(G), defined in Equation (2.10), on the dense subspace Cc(X) of the Hilbert C∗(H, β)-module H(X) by σG, that is, σG(b)(ζ) = b · ζ. Similarly, let σH denote the representation of the pre-C∗-algebra Cc(H) on the pre-Hilbert C∗(H, β)-module Cc(X). Theorem 2.11. Let (G, α) and (H, β) be locally compact groupoids with Haar systems. If (X, λ) is a proper topological correspondence from (G, α) to (H, β), then the C∗-correspondence H(X) : C∗(G, α) → C∗(H, β) is proper. The rest of this section is devoted to prove Theorem 2.11. The proof is broken into three steps: (i) first we show that if h ∈ Cc(G) and ξ ∈ Cc(X), then the G(0) X ∈ Cc(G ×G(0) X); this is a well know fact that we re- restriction (h ⊗ ξ)G× vise. (ii) For given b ∈ Cc(G), we find suitable functions ǫi ∈ Cc(G) and δj ∈ Cc(X), where 1 ≤ i ≤ m and 1 ≤ j ≤ n, and m and n are natural numbers, such that fij := bǫi ⊗ c ¯c δj is in Cc(G ×G(0) X, Dij) for some Dij ∈ A, here c and ¯c are fixed pre-cutoff and cutoff functions. Recall from Definition 2.5 that A is a fixed cover of G ⋉ X by sets in mHo,Co ij ∈ Cc(X ×sX ,H(0),sX X) be an extension of fij via msX which is supported in A′. We show that the operator . (iii) Finally, let F ′ sX σG(b) on H(X) is same as the operator σKT(X)(cid:16)Pi=n,j=m See the proof of Proposition 1.46 for the meaning of BX× by proving follwing lemmas. i=1,j=1 BX× sX ,H(0) ,sX sX ,H(0) ,sX X (F ′ ij ). We start X (F ′ ij )(cid:17). Lemma 2.12. Let X be a topological space and L, K ⊆ X subspaces. If L is closed in X and K is (quasi-)compact in X, then L ∩ K is (quasi-)compact in L as well as in X. i ⊆ X be an open set such that Ui = U ′ Proof. Let {Ui}i∈I be an open cover of K ∩ L in L where I is an index set. For each i ∈ I let U ′ i }i∈I ∪ {X − L} is an open cover of K in X. Since K ⊆ X is (quasi-)compact, choose and open subcover of this cover which covers K; thus there is a finite natural number n such that U ′ in and X − L cover K. Now one may see that Ui1 , . . . , Uin , is an open cover K ∩ L. i ∩ K. Then {U ′ i1, . . . , U ′ Moreover, since L ∩ K is (quasi-)compact in L, it is (quasi-)compact in any (cid:3) subspace of X containing L, in particular, X itself. Lemma 2.13. Let X be a locally compact left G-space for a locally compact groupoid G. Let K ⊆ G and Q ⊆ X be nonempty subsets with sG(K) ∩ rX (Q) 6= ∅. If K and Q are quasi-compact (or compact) in G and X, respectively, then so is K ×G(0) Q ⊆ G ×G(0) X. Proof. Since all the sets K, Q and sG(K) ∩ rX (Q) are nonempty, so is K ×G(0) Q. Assume that K ⊆ G and Q ⊆ X are quasi-compact. We observe that G×G(0) X ⊆ G × X is closed; this is proved in the last paragraph below. Then K ×G(0) Q := (K × Q) ∩ (G ×G(0) X) is quasi-compact in G ×G(0) X due to Lemma 2.12. This proves the assertion about quasi-compactness of K ×G(0) Q. Similar is the argument for the claim regarding compactness. PROPER TOPOLOGICAL CORRESPONDENCES 31 A proof that G×G(0) X ⊆ G×X is closed: recall from basic topology that a space Y is Hausdorff if and only if the diagonal dia(Y ) := {(y, y) : y ∈ Y } ⊆ Y × Y is closed. By hypothesis, G(0), the space of units of the groupoid G, is Hausdorff hence dia(G(0)) ⊆ G(0) × G(0) is closed. Now the function sG × rX : G × X → G(0) × G(0), (γ, x) 7→ (sG(γ), rX (x)), is continuous, and G ×G(0) X = (sG × rX )−1(dia(G(0))). Being the inverse image of a closed set under a continuous function G ×G(0) X ⊆ G × X is closed. (cid:3) Lemma 2.14. Let Y and X be locally compact spaces, and let K ⊆ Y and Q ⊆ X be quasi-compact spaces. Let {Wd}d∈D be an open cover of K × Q where D is an index set and each Wd is open Y × X. Then there are finite open covers {U1, . . . , Um} of {K}, and {Vi, . . . , Vn} of {Q} such that for each (i, j) ∈ {1, . . . , m} × {1, . . . , n} there is an index d ∈ D with Ui × Vj ⊆ Wd, and Ui ⊆ Y and Vj ⊆ X are open. Proof. The proof of this lemma is exactly same as the one of Lemma [20, 7.15] except that one replaces all the open covers in K and Q by open covers in Y and X, respectively. (cid:3) Now we write the proof of Theorem 2.11; during the first reading of this proof, reader may assume that D and the adjoining function ∆ are the constant function 1. Proof of Theorem 2.11. 1) Let (G, α), (H, β) and (X, λ) be as in the statement of the theorem. Let ∆ be the adjoining function of the topological correspondence. Let σG and σKT have the meaning as the discussion following Equation (2.10) and Notation 1.50, respectively. sX ,H(0) ,sX Recall from the proof of Proposition 1.46 that the Haar system β of H in- X along the quotient map X ×sX ,H(0),sX duces a family of measures βX× X → (X ×sX ,H(0),sX X)/H similar to the one in Equation (1.1). The image X : Cc(X ×sX ,H(0),sX X) → of the corresponding integration map BX× Cc((X ×sX ,H(0),sX X)/H) is dense. 2) Let b ∈ Cc(G) be a given nonzero function. In (6) of this proof, we show that there are natural numbers m and n, and a function Fij ∈ Cc(X ×sX ,H(0),sX X) for 0 ≤ i ≤ m and 0 ≤ j ≤ n with sX ,H(0) ,sX σG(b)ξ = i=m,j=n Xi,j=1 σKT(X)(BX× sX ,H(0) ,sX X (Fij ))ξ for all ξ ∈ Cc(X). Since every function in Cc(G) is a finite linear combination of functions in Cc0(G), we may assume that b ∈ Cc0(G), see [11, Lemma 1.3] for details. Then K = supp(b) ⊆ T where T ⊆ G is open and Hausdorff. Consequently, sG(K) ⊆ G(0) is compact. Use Condition (ii) in Definition 2.5 to see that [rX ]−1(sG(K)) ⊆ X/G is compact. Since the quotient map q : X → X/G is open, we choose a compact set Q ⊆ X with q(Q) = [rX ]−1(K). Then sG(K) = rX (X) 6= ∅. Now apply Lemma 2.13 to the sets K ⊆ G and Q ⊆ X to see that K ×G(0) Q ⊆ G ×G(0) X is compact. 3) Let A = { A ⊆ G × X : A is open in G × X, and A ∩ (G ×G(0) X) ∈ A}; this is a collection of open sets in G× K which covers G×G(0) X. Since K × Q = supp(b)× Q and G ×G(0) X are closed in G × X, so is K ×G(0) Q := (K × Q) ∩ (G ×G(0) X). Therefore, A ∪ {(G × X) − (K ×G(0) Q)} is an open cover of K × Q consisting of open sets in G × X. Apply Lemma 2.14 to this cover to get finitely many open sets U1, . . . , Um in G and finitely many open sets V1, . . . , Vm in X which cover K and 32 ROHIT DILIP HOLKAR Q, respectively, and for each (γ, x) ∈ K ×G(0) Q, there are indices i ∈ {1, . . . , n} and j ∈ {1, . . . , m} such that (γ, x) ∈ Ui × Vj ⊆ A for some A ∈ A. Therefore, for 1 ≤ i ≤ m and 1 ≤ j ≤ n and any functions f ∈ Cc(G, Ui) and g ∈ Cc(X, Vj), the restriction f ⊗G(0) g is supported in a set in A. Let {ǫi}n i=1 be a partition of unity of K subordinate to the open cover {Ui}n i=1 and {δj}m j=1 be the one of Q subordinate to {Vj}m j=1. 4) Recall Example 1.41. As in this example, let c : X → ¯R+ be a pre-cutoff function, that is, a function similar to the function F in Lemma 1.40. Let ¯c be same as in the example; ¯c is obtained by averaging c over H: ¯c(x) =ZH c(xη) dβsX (x)(η) for u ∈ X. Therefore, as discussed in the example, ¯c is an H-invariant function. Fix (i, j) ∈ {1, . . . , m} × {1, . . . , n}. Then the product bǫi ∈ Cc(G, Ui). Similarly, (c/¯c) δj ∈ Cc(X, Vj) and the function (2.15) lij :=(cid:16)bǫi ⊗G× G(0) X c ¯c δj(cid:17) · ∆ is in Cc(G ×G(0) X, Ui ×H(0) Vj), due to Lemma 2.13. Recall from (3) above that each lij is supported in a set in A ⊆ mHo,Co . Fix an (lij ) which is supported in a set in A′. Using the local sections extension Fij ∈ EmsX in Equation (1.38) we see that for (x, y) ∈ msX (Ui ×G(0) Vj), sX Fij (x, y) = lij ◦ msX −1 Ui× (x, y) G(0) Vj = b(γ(x,y)) ǫ(γ(x,y)) δ(γ−1 (x,y) x) c(γ−1 ¯c(γ−1 (x,y) x) (x,y) x) ∆(γ(x,y), γ−1 (x,y) x). 5) Let F ′ ij = Fij D where D : X ×sX ,H(0),sX X → R∗ is the H-invariant nonnegative continuous function associated to the proper correspondence (X, λ). Since Fij ∈ EmsX ij is also continuous and compactly supported. (lij) is continuous function with compact support, and D is continuous, F ′ Let ξ ∈ Cc(X) and recall from (1) above that BX× X is the family of measures along the quotient map X ×sX ,G(0),sX X → (X ×sX ,G(0),sX X)/H as in Equation (1.1). Then sX ,H(0) ,sX σKT(X)(BX× sX ,H(0) ,sX X (F ′ ij ))ξ(x) BX× sX ,H(0) ,sX X (F ′ ij )[x, y]ξ(y) dλsX (x)(y) F ′ ij (xη, yη)ξ(y) dβsX (x)(η) dλsX (x)(y) Fij (xη, yη) D(xη, yη) ξ(y) dβsX (x)(η) dλsX (x)(y). =ZX =ZXZH =ZXZH ZXZH Now use the H-invariance of D which is built in its definition (see Definition 2.5) to see that the above term equals Fij (xη, yη) D(x, y) ξ(y) dβsX (x)(η) dλsX (x)(y). Using Fubini's theorem we write the above term as ZH(cid:18)ZX Fij(xη, yη) D(x, y) dλsX (x)(y)(cid:19) dβsX (x)(η). PROPER TOPOLOGICAL CORRESPONDENCES 33 (lij ) and Fij is supported in an element of A′. Recall from (4) above that Fij ∈ EmsX Therefore, Equation (2.6) in Definition 2.5 allows us to change the above integral on X to the one on G as ZHZG lij(γ(xη,yη), γ−1 (xη,yη)xη) ξ(y) dαrX (x)(γ) dβsX (x)(η). Now substitute the value of lij in above equation from Equation (2.15); we write γ instead of γ(xη,yη) for simplicity and then the previous term looks ZHZG b(γ)ǫi(γ) δj(γ−1xη)ξ(γ−1xη) c(γ−1xη) ¯c(γ−1xη) ∆(γ, γ−1xη) dαrX (x)(γ) dβsX (x)(η). 6) Recall from Example 1.41 that ¯c is H-invariant, and recall from [9] that ∆ is also H-invariant. We incorporate these changes and compute:   = = = i=m,j=n i=m i=m,j=n Xi,j=1 Xi,j=1 ZHZG Xi=1 ZG Xi=1 ZG i=m σKT(X)(cid:16)BX× sX ,H(0) ,sX X (F ′ ij )(cid:17)  ξ(x) c(γ−1xη) ¯c(γ−1x) b(γ) ǫi(γ) δj(γ−1xη) ξ(γ−1x) ∆(γ, γ−1x) dαrX (x)(y)dβsX (x)(η) n b(γ) ǫi(γ) ξ(γ−1x)∆(γ, γ−1x) Xj=1ZH  ZH b(γ) ǫi(γ) ξ(γ−1x)∆(γ, γ−1x) Xj=1  n δj(γ−1xη) c(γ−1xη) ¯c(γ−1x) dβsX (x)(η) δj(γ−1xη)  c(γ−1xη) ¯c(γ−1x) dβsX (x)(η)  dαrX (x)(y) dαrX (x)(y). We use Fubini's theorem during the second step above. Now use the fact that {δj}n j=1 of Q. Then the last term equals j=1 is a partition of unity subordinate to the cover {Vj}n i=m Xi=1 ZG b(γ) ǫi(γ) ξ(γ−1x)∆(γ, γ−1x)(cid:18)ZH c(γ−1xη) ¯c(γ−1x) dβsX (x)(η)(cid:19) dαrX (x)(y). Note that the integrant in the second integral, dβsX (x), is the cutoff function (c/¯c) ◦ sX⋊H , see Example 1.41. With this observation we continue the computing the last term: i=m b(γ) ǫi(γ) ξ(γ−1x)∆(γ, γ−1x)(cid:18) 1 ¯c(γ−1x)ZH c(γ−1xη) dβsX (x)(η)(cid:19) dαrX (x)(y) b(γ)ǫi(γ) ξ(γ−1x)∆(γ, γ−1x) dαrX (x)(y) ǫi(γ)! ξ(γ−1x)∆(γ, γ−1x) dαrX (x)(y) = σG(b)ξ(x). To get the last equality we use the fact that {ǫi}m the cover {Ui}m i=1 of K. i=1 is a partition of unity subodinate Thus, for given b ∈ Cc(G), we have found finite natural numbers m and n, and function F ′ ij ∈ Cc(X ×sX ,G(0),sX X), where 0 ≤ i ≤ m and 0 ≤ j ≤ n, such that = i=m Xi=1 ZG Xi=1 ZG b(γ) i=m =ZG Xi=1 σG(b) = i=m,j=n Xi,j=1 σKT(X)(BX× sX ,H(0) ,sX X (F ′ ij )) 34 ROHIT DILIP HOLKAR where each σKT(X)(BX× operator on H(X) . sX ,H(0) ,sX X (Fij )), as per Proposition 1.46, is a compact (cid:3) Corollary 2.16. Let (G, α) and (H, β) be localy compact goupoids equipped with Haar systems. Then a proper topological correspondence (X, λ) from (G, α) to (H, β) defines an element in KK(C∗(G, α), C∗(H, β)). Proof. Theorem 2.11 says that H(X) is a proper C∗-correspondence from C∗(G, α) to C∗(H, β). Hence [0, H(X)] ∈ KK(C∗(G, α), C∗(H, β)) where 0 ∈ BC∗(H,β)(H(X)) is the zero operator. (cid:3) Recall Section 2.5.2 in [7] that discusses isomorphism of topological correspon- dence. Proposition 2.17. Let (G, α) and (H, β) be locally compact groupoids, and let (X, λ) and (Y, µ) be two correspondences from (G, α) to (H, β). Let φ : X → Y be a G-H-equivariant homeomorphism which implements an isomorphism of correspon- dences. Then, if (X, λ) is proper, then so is (Y, β). Proof. The isomorphism of correspondences φ induces a H-equivariant isomorphism of topological groupoids φ′ : G ⋉ X → G ⋉ Y, (γ, x) 7→ (γ, φ(x)). We also get the following commutative square In this square, the vertical arrows G ×G(0) X φ′ G ×G(0) Y msX msY X ×sX ,H(0),sX X φ× H(0) φ Y ×sY ,H(0),sY Y Figure 2. are isomorphisms. We leave it as excercise to reader to prove that if msX is a local homeomorphism onto its image, then so is msY . The map φ induces a G-equivariant isomorphism [φ] : X/H → Y /H. Further- more, [rX ] = [rY ] ◦ [φ], that is, [rX ] ◦ [φ]−1 = [rY ]. Therefore, [rX ] is a proper map if and only if[rY ] is so. By hypothesis [rX ] is proper, therefore [rY ] is proper. Finally, let A and A′ be open covers of G×G(0) X and X×sX,H(0),sX , respectively, which satisfies the last condition in Definition 2.5. Then B :={φ′(U ) : U ∈ A}, B′ :={φ ×H(0) φ(V ) : V ∈ A′} are open covers of G ×G(0) Y and Y ×sX ,H(0),sX , respectively, which satisfy the last condition in Definition 2.5. For u ∈ H (0), let φ∗(λu) be the measure induced by λu on Y via φ. Let DX deonote the function D appearing in Definition 2.5. Define DY (x, y) = DX ◦ φ−1(x, y) dφ∗(λ)sX ((x)) dµsY (x) (y). PROPER TOPOLOGICAL CORRESPONDENCES 35 Let f ∈ Cc(G ×G(0) Y, U ) where U ∈ A′, and let f ′ be an extension of it in EmY (f ). Observe that f ′ ◦ φ′ is an extension of f ◦ φ′. Now f ′(x, y) DY (x, y) dλsY (x)(y) f ′(x, y) DX ◦ φ−1(x, y) dφ∗(λ)sX ((x)) dµsY (x) (y) dλsY (x)(y) f ′ ◦ φ′(a, b) DY (a, b) dµsX (a)(b) ZY =ZY =ZX Since (X, λ) is proper correspondence, using Equation (2.6) we see that the last term equals ZG f ◦ φ′(γ, γ−1a) dαrX (a)(γ) =ZG f (γ, γ−1x) dαrY (x)(γ). The last equailty is due to the fact that φ is a G-equivariant isomorphism. This proves that (Y, µ) is proper. (cid:3) 3. Examples Example 3.1. In [20], Tu defines locally proper generalised morphism and proper locally proper generalised morphism from a locally compact groupoid G to another one, say, H. This morphism is a locally compact G-H-bispace X satisfying certain properties. We repeat Tu's definition of a locally proper generalised morphism and proper locally proper generalised morphism when X is Hausdorff; we modify the verbalisation to suit our setting, for the original definitions see [20, Definitions 7.3 and 7.6]. Definition 3.2 ([20, Definitions 7.3 and 7.6] when X is Hausdorff). A locally proper generalised morphism from a locally compact groupoid with Haar system (G, α) to a groupoid with Haar system (H, β) is a G-H-bispace X such that i) the action of G is proper and free, ii) the right momentum map induces a homeomorphism G\X → H (0) iii) the action of H is proper. Furthermore, the locally proper generalised morphism X is called proper if iv) [rX ] : X/H → G(0) is a proper map. [20, Theorem 7.8] proves that a proper correspondence X from (G, α) to (H, β) produces a proper C∗-correspondence from C∗ r (H, β). The discussion on page 219 in the sixth paragraph in [9] and Example 3.8 in the same article show that a locally proper generalised morphism is a topological correspondence; this example also shows that the H-invariant family of measures λ on X is given by r (G, α) to C∗ (3.3) ZX f dλu :=ZG f (γ−1x) dαrX (x)(γ) where f ∈ Cc(X), u ∈ H (0) and x ∈ s−1 is well-defined and is independent of the choice of y. X (u). The same computation shows that λu Let X be a proper locally proper generalised morphism from a locally compact groupoid with Haar system (G, α) to (H, β). We show that when X is Hausdorff and X/H is paracompact, (X, λ) is a proper topological correspondence; here λ is the family of measures discussed above. In this case, Remark 1.26 shows that the map msX is a homeomorphism. Hence the first condition in Definition 2.5 is satisfied. Condition (iv) in Definition 3.2 above agres with Definition 2.5(ii). 36 ROHIT DILIP HOLKAR Finally, to see that Definition 2.5(iii) is satisfied, let A{G ×G(0) X} be the open cover of G ×G(0) X, and A′ = {X ×sX ,G(0),sX X} be its extension via msX ; recall from Remark 1.26 that in present situation msX is a homeomorphism. Let f ∈ Cc(G ×G(0) X), then {f ◦ m−1 (f ). Take D to be the constant function 1 SX on X ×sX ,H(0),sX X. We compute the left and right side terms in Equation (2.6): } = EmsX ZX f ′(x, y) D(x, y) dλsX (x)(y) =ZX f ◦ m−1 sX (x, y) dλsX (x)(y). Since msX is a homeomorphism, there is unique element γ(x,y) ∈ G such that γ(x,y)y = x, that is γ−1 (x,y)x = y; we remove the suffix of γ for the sake of simplicity. Now we may write the right side term of the previous equation as f ◦ m−1 sX (x, γ−1x) dλsX (x)(γ−1x). ZX Using the definition of the measure λsX (x) in Equation (3.3) we write the right side term in the last equation as f ◦ m−1 sX (x, η−1γ−1x) dαsG(γ)(η). ZX ZX Now use Equation 1.38 to see f ◦ m−1 value of f ◦ m−1 sX in the last integral, then it becomes sX (x, η−1γ−1x) = f (γη, η−1γ−1x). Put this f (γη, η−1γ−1x) dαsG(γ)(η). Now change the variable γη 7→ γ and use the left invariance of the the Haar system α to see that the above term equals which is the right hand side of Equation (2.6). f (γ, γ−1x) dαrX (x)(γ) ZX Example 3.4. (Groupoid equivalence) Recall the definition of groupoid equivalence [14, Definition 2.1]. Let X be a groupoid equivalence between locally compact groupoids G and H. Let α and β be the Haar systems on G and H, respectively. Then X is a proper locally proper generalised morphism; a straightforward comparison be- tween [14]Definition 2.1 and Definition 3.2. Example 3.9 shows this. Example 3.1 now shows that when equipped with the family of measures λ in Equation 3.3, (X, λ) is a proper correspondence. The function D on X ×sX ,H(0),sX X is the constant function 1. Example 3.5. Let G be a locally compact group and {∗} the trivial group(oid). Let α be the Haar system on G, then (G, α) is a topological correspondence from G to {∗}: G acts on itself with the left multiplication and {∗} acts on G trivially; α is (G, α)-quasi-invariant measure on G. This is a special case of Example [9, Example 3.5] with the measure on the bisapce inverted; the map is the zero map {∗} → G. Assume that G is compact, then (G, α) is a proper correspondence. In this case, the first two conditions in Definition 2.5 are clearly satisfied. Finally, observe that the map m0 : G × G → G × G is an isomorphism. To see that the last condition in Definition 2.5 holds, the computation is similar to the one in Example 3.1 with X is replaced by G. To elaborate the algebraic side of this example, the topological correspondence (G, α) from G to {∗} gives the left regular representation of C∗(G) on L2(G, α). Now Theorem 2.11 reproduces the basic result that in this case the the left regaular representation takes values in K(L2(G, α)). PROPER TOPOLOGICAL CORRESPONDENCES 37 Example 3.6. Let (G, α) be a locally compact groupoid with a Haar system, and H ⊆ G a closed subgroupoid. Let β be a Haar system on H. Let X := {γ ∈ G : rG(γ) ∈ H (0)}. Popularly, X is denoted by GH(0) . Then H and G act on X from left and right, respectively, by multiplication. The momentum maps for these actions are rX = rGX and sX = sGX . Both the actions are free, and as [9, Example 3.14] shows, the actions are proper. In this case, X is a topological correspondence from H to G where the family of measures is given by Equation (3.3); note that in this case the map [sX ] : H\X → G(0) is need not be surjective. This is a correspondence similar to the one in [9, Example 3.14] with the direction inverted. Since the action of H on X is proper, the map m : H ×H(0) X → X × X, (η, x) 7→ (ηx, x), is proper. As m−1(X ×sX ,H(0),sX X) = H ×H(0) X, Proposition 10.1.3(a) in [1, Chapter 1]) implies that msX is proper. Now the fact that msX is one-to-one and proper, using [1, Chapter 1, Proposition 10.1.2], we conclude that msX is a homeomorphism onto a closed subspace of X ×sX ,H(0),sX X. This shows that the first condition in Definition 2.5 holds. We observe that [rX ] : X/G → H (0) is the identity map. Thus the second condi- tion in Definition 2.5 is satisfied. The third condition can be checked as in Exam- ple 3.1. Example 3.7. Let G be a locally comapct group and H a closed subgroup of it. Assume that G is H-compact, that is, G/H is a compact space. Let α and β be the Haar measure on G and H, respectively. Then with the left and right multiplication actions, (G, α−1) is a topological correspondence from G to H. Reader may check that this is an example of correspondences in Definition 3.2. In this case, m0 : G × G → G × G is the map (γ, γ′) 7→ (γγ′, γ′) which is a homeomorphism; this fulfils the first condition in Definition 2.5. Since G/H is com- pact, the constant map G/H → {∗} is proper which fulfils the second condition for a proper correspondences. A computation as in Example 3.1 shows that the third condition in Definition 2.5 holds. Thus (G, α−1) : G → H is a proper corre- spondence. A concrete example of this case is when G = Z and H = nZ where n = 1, 2, . . .. x }x∈X where δX 3.1. The case of spaces. A locally compact space X is a locally compact groupoid; in this groupoid every arrow is a unit arrow and the source as well as the range maps equal IdX the identity map on X. Equip the groupoid X with the Haar system δX := {δX y is the point mass at x ∈ X. Now on, whenever we think of a space as a locally compact groupoid equipped with the (or standard or obvious) Haar system, we shall mean this setting. Note that X is an étale groupoid. Let Y be another space. Then the only possible action of X on Y is the trivial action, that is, there is a momentum map b : X → Y and for (x, y) ∈ X ×IdX ,X,b Y , the action is x · y = y. The actions of X on Y are in one-to-one correspondence with the continuous maps from X to Y . Lemma 3.8. Let Y b←− X a space Z. If (X, λ) is proper in the sense of Definition 2.5, then f −→ λ Z be a topological correspondence from a space Y to i) for each z ∈ Z, the measure λz is atomic, ii) for each z ∈ Z, each atom in λz is nonzero and the function D appearing in Definition 2.5(3) positive. Thus the family of measure λ has full support. Proof. i): The discussion regarding Figure 3 in Example 3.10 shows that the image of mf is the diagonal {(x, x) ∈ X ×f,Z,f X : x ∈ X} which is closed in X ×f,Z,f X (since X is Hausdorff). Fix x ∈ X. Then using Lemma 2.8 we see that the measure λx 1 is concentrated on the diagonal in X ×f,Z,f X. Note that the measure λx 1 is defined on π−1 2 (x) where π2 is the projection onto the second factor of X ×f,Z,f X. 38 ROHIT DILIP HOLKAR 1 is concentrated on π−1 Thus λx {(x, x′) ∈ X ×f,Z,f X : f (x) = f (x′)}∩{(x′′, x′′) ∈ X ×f,Z,f X : x′′ ∈ X} = {(x, x)}. 2 (x) ∩ diaX (X) which equals Thus for each x ∈ X. Now λx 1 = δx ⊗ λf (x) hence supp(λx 1 ) ⊆ {(x, x)} = {x} × {x} supp(λx 1 ) = supp(δx) × supp(λf (x)) = {x} × supp(λf (x)), due to [2, Chapter III, §4.2, Proposition 2, page III.44]. Hence supp(λf (x)) ⊆ {x} for each x ∈ X. ii): Let x0 ∈ X be given. Let g ∈ Cc(Y ×IdY ,Y,b X) be a function which is nonnegative and g(b(x0), x0) > 0. Let D be the function in the third condition of the proper correspondence. Then Equation 3.11 says that g(y, y−1 · x) dδb(x0) Y = g(b(x0), x0) > 0. ZY Let g′ ∈ Ef (g) be a nonnegative extension of g. Since the third condition in Defini- tion 2.5 holds, we get ZX g′(x0, x) D(x0, x′)dλf (x0)(x) =ZY g(y, y−1 · x) dδb(x0) Y = g(b(x0), x0) > 0. Thus the measure λf (x0) 6= 0, that is, λz > 0 for each each z ∈ Z. Furthermore, since the function g′ and the measure λz in the left hand side of the above equation are nonnegative, and the whole term is positive, we get D(x0, x′) > 0. (cid:3) Let X and Y be spaces and f : X → Y be a (surjective) local homeomorphism. Then there is a canonical continuous family of measure along f which we denote by τ φ: (3.9) ZX g τ φ y = Xx∈supp(g)∩f −1(y) g(x) where g ∈ Cc(X) and y ∈ Y . f Example 3.10. Let X, Y and Z be locally compact Hausdorff spaces, and Y b←− X −→ Z continuous maps. Assume that f is a local homeomorphism. Then (X.τ f ) is a topological correspondence from the étale groupoid X to Y . The adjoining function of this correspondence is the constant function 1; this follows directly of one writes the Equation in Definition 2.1 (iv) in [9], and keep in mind that the actions are trivial. Additionally, assume that b is a proper map. Then (X, λ) is a proper correspon- dence. Following are the details: in this case, Y ×Y X = Y ×IdY ,Y,b X ≈ X. Hence mf : Y ×Y X → X ×f,Z,f X can be identified with the diagonal embedding (see Figure 3), thus the first condition in Definition 2.5 is satisfied. Since b is proper, and the action of Z on X is the trivial action, the second condition is also fulfilled. To see that the third condition is also satisfied, one identifies Y ×Y X with X, mf with the diagonal embedding X ֒→ X ×f,Z,f X, and use this embedding to prove the required claim. Let θ : Y ×Y X → X be the the homeomorphism θ : (b(x), x) 7→ x; inverse of θ is x 7→ (b(x), x). Let diaX : X → X ×f,Z,f X be the diagonal embedding. We draw the commutative triangle in Figure 3. Take the cover A of Y ×Y X to be {θ−1(U ) : U ∈ f Ho,Co}. Let A′ := {U ×f,Z,f U : U ∈ f Ho,Co}. We observe that A′ is an extension of f Ho,Co via diaX , and hence it is an extension of A via mf . Note that for U ′, U ⊆ X, (U ×f,Z,f U )∩(X ×f,Z,f X) = U ′ PROPER TOPOLOGICAL CORRESPONDENCES 39 Y ×Y X mf X ×f,Z,f X ≈θ X diaX Figure 3. if and only if U = U ′. Thus every element in f Ho,Co or A has a unique extension in A′. Let D be the constant function 1. Let θ−1(U ) ∈ A be given where U ∈ f Ho,Co. Let g ∈ Cc(Y ×Y X, θ−1(U )); then g := g ◦ θ−1 ∈ Cc(X; U ). Let g′ ∈ Cc(X ×f,Z,f X; U ×f,Z,f U ) be an extension of g via diaX .Then g′ is an extension of g also. For x ∈ X, the right side of Equation (2.6) for g is (3.11) g(y, y−1 · x) dδb(x) Y g(y, x) dδb(x) Y (y) = g(b(x), x). ZY (y) =ZY For the same x ∈ X as above, the left side of Equation (2.6) for g′ reads g′(x, x′) dτ f f (x)(x′). ZX note that g′ is supported in U ×f,Z,f U , and f U is a homeomorphism. Therefore, the only element in U which hits f (x) under f is x itself. Now recall that g′ is an extension of g via mf , hence g′ ◦ diaX = g. Thus g′(x, x′) dτ f f (x)(x′) = g′(x, x) = g′ ◦ mf (b(x), x) = g(b(x), x); ZX the claim follows by comparing the above equation with Equation (3.11). dence from y to z, see [9, example 3.3]. Example 3.12. Let Y and Z be spaces considered as locally compact groupoids equipped with the canonical Haar system. let (X, λ) be a topological correspon- thus there are momentum maps Y b←− f −→ Z and a continuous family of measures λ along f . The topological correspon- X dence (X, λ) gives a C∗-correspondence H(X) from C0(Y ) to C0(Z). We use the shorthand notation Y b←− X Z to write this correspondence. f −→ λ When the spaces Y and Z above are same and the family of measures λ has full support, Muhly and Tomford call this topological correspondence a topolog- ical quiver, see [15, Definition 3.1]. They call the quiver (X, λ) proper if b is a proper map, and f is a local homeomorphism. [15, Theorem 3.11] asserts that the C∗-correspondence H(X) from C0(Y ) to itself is proper if and only if the quiver (X, λ) is proper. One can readily see that that [15, Theorem 3.11] is valid for a topological correspondence of spaces if the family of measures on the middle space has full support. Returning to our example of the topological correspondences Y b←− X f −→ λ Z, assume that λ has full support, f is a local homeomorphism and b is proper. Then Example 3.10 shows that (X, τ f ) is a proper correspondence from Y to Z. We show that the topological correspondences (X, λ) and (X, τ f ) are isomorphic. Then Proposition 2.17 implies that (X, λ) is proper. We claim that the identity map, IdX : X → X, gives the isomorphism (X, λ) → (X, τ f ) of correspondences. Define D : X → R, D(x) = λf (x)(x) . 40 ROHIT DILIP HOLKAR Then D is positive since λ has full support and continuous since λ is continuous. We note that D(x′) = dλf (x) dδf (x) (x′) where the latter term is the Radon-Nikodym derivative of λf (x) with respect to δf (x) = (IdX )∗(δf (x)). This implies that (IdX )∗(δz) ∼ λz for every z ∈ Z and the Radon-Nikodym derivative is continuous. Thus IdX gives the isomorphism of the correspondences (X, λ) → (X, τ f ), and the corresponding positive Radon-Nikodym derivative is given by the function D above. The isomorphism of correspondences in this example generalises to étale groupoids, see Proposition 3.14. [15, Theorem 3.11] implies that if Y b←− X Z is a topological correspon- dence, then the corresponding C∗-correspondence is proper if and only if the map [rX ] : X/H → G(0) is proper, and the right momentum map sX is a local homeo- morphism. f −→ λ f Proposition 3.13. Let X, Y and Z be spaces, Y b←− X −→ Z maps and λ and λ′ two continuous families of measures along f having full support. Assume that f is a local homeomorphism and b a proper map. Thus (X, λ) and (X, λ′) are topo- logical correspondences from Y → Z (Example 3.12). Then in KK(C0(Y), C0(Z)), [(H(X)), 0] = [(H(X ′)), 0]. Proof. Example 3.12 shows that (X, λ) and (X ′, λ′) are isomorphic to (X, τ f ). Hence the required KK-classes are same. (cid:3) 3.2. The case of étale groupoids. Let G be an étale groupoid. Assume that G(0) is locally compact Hausdorff subspace of G and that G is covered by countably many compact sets the interiors of which form a basis for the topology of G. Then G is a locally compact groupoid in the sense of [16, Definition 2.2.1]. In this section, the term étale groupoid shall stand for groupoids which are étale and fulfil the above conditions. In this case, G is a locally compact groupoid equipped with a Haar system, and it has a nice approximate identity: one writes G(0) as an increasing union of compact sets, Kn where n ∈ N, whose interiors cover G(0). Then using Urysohn's lemma, one gets functions un ∈ Cc(G(0)) for each n ∈ N such that each 0 ≤ fn ≤ 1 and fn = 1 on Kn. The sequence {fn}n∈N is an approximate identity for the ∗-algebra Cc(G), for details see [16, pages 47 -- 49]. For an étale groupoid G, C∗(G(0)) = C0(G(0)) is a subalgebra of C∗(G), and the map Cc(G) → Cc(G(0)), f 7→ f G(0) , of ∗-algebras induces an expectation from C∗(G) → C0(G(0)), see [17, page 61] for details. In this section, during computations the canonical Haar system on the étale groupoid G shall be denoted by α and the one on H by β. Let G and H be étale groupoids and X a G-H-bispace. Assume that the mo- mentum map sX : X → H (0) is a local homeomorphism. Let τ sX be the family of point-mass measures along sX (see Equation (3.9)). Assume that (X, τ sX ) is a topological correspondence, that is, the action of H is proper and each measure in τ sX is G-quasi-invariant. Then we call X, or to be precise (X, τsX ), an étale correspondence from G to H. The adjoining function of this correspondence is the constant function 1. This is because for each u ∈ H (0), the measure τ sX on the u space of units of the transformation groupoid G ⋉ X is quasi-invariant; here G ⋉ X is an étale groupoid. And the adjoining function is the modular function of G ⋉ X for τ sX u which is constant function 1 due to [17, 1.3.22]. PROPER TOPOLOGICAL CORRESPONDENCES 41 Proposition 3.14. Let G and H be étale groupoids and X a G-H-bispace. Assume that the right momentum map sX : X → H (0) is a local homeomorphism. Let λ be a continuous family of measures with full support along sX . Assume that (X, λ) is a topological correspondence from G to H. Let τ sX be the family of measures consist- ing of point-masses along sX . Then (X, τ sX ) is a topological correspondence from G to H and the identity map IdX : X → X induces an isomorphism of topological correspondences. Proof. Since the family of measures λ is continuous and has full support, the func- tion is continuous. Since λ is H-invariant, so is D. Then D : X → R+ D : x 7→ λ(x) D : Cc(X) → Cc(X) : f 7→ f D is an isomorphism of complex vector spaces. Now we note that for f ∈ Cc(X) and every u ∈ H (0), (3.15) λ(f ) = τ sX (f λ(x))(u) = τ sX (f D)(u). Thus λu ∼ τ sX u τ for every u ∈ H (0), and the Radon-Nikodym derivative λu sX u (x) = λ(x) is continuous. Now it follows from [17, Proposition 1.3.3(ii)] that τ sX is u G ⋉ X-quasi-invariant for each u ∈ H (0). This shows that (X, τ sX ) is a topological correspondence from G to H. Since λu ∼ τ sX u , which follows from Equation (3.15), we have also proved that the identity map X → X is an isomorphism of the topological correspondences (X, λ) and (X, τ sX ). (cid:3) Converse of Proposition 3.14 clearly holds. Let G and H be étale groupoids, X a G-H-bispace and sX : X → H (0) a local homeomorphism such that (X, τ sX ) : G → H is a topological correspondences. If D : X → R is an H-invariant positive contin- uous function, then Dτ sX := {Dτ sX u }u∈H(0) is an H-invariant family of measures, and since each measure Dτ sX u , [17, Proposition u 1.3.3 (ii)] implies that Dτ sX is G ⋉ X-quasi-invariant measure on its space of units. u in this family is equivalent to τ sX Let X be a G-H-bispace and let sX : X → H (0) be a local homeomorphism. Assume that (X, τ sX ) : G → H is a topological correspondence. Then, Proposi- tion 1.35 implies that, this correspondence is proper if msX : G×G(0) X → X×sX ,H(0),sX X is open and the rest two conditions in Definition 2.5 hold. Due to Proposi- tion 2.17, the correspondence (X, λ) in Proposition 3.14 above is proper if and only if (X, τ sX ) is proper. Thus for a correspondence (X, λ) as in Proposition 3.14, the space X determines the correspondence upto isomorphism of correspondences. Clearly, the correspondences of spaces in Examples 3.10 and 3.12 are étale cor- respondences. Following are a few more examples of étale correspondences. Example 3.16. Let Y be a space considered as groupoid, H an étale groupoid and X a Y -H-bispace. Let sX : X → H (0) be a local homeomorphism. Then (X, τsX ) : Y → H is a topological correspondence. This correspondence is proper if the right momentum map rX : X → Y is proper. To confirm that (X, τsX ) is a topological correspondence, one needs to check that τ u sX is Y -quasi-invariant. This follows directly if one substitutes appropriate values in the required equation in [9, Definition 2.1]. Now assume that rX is proper. To see that this correspondence is proper, we need to check that Definition 2.5(iii) holds as other conditions obviously hold. The map msX : Y ×IdY ,Y,rX X : X ×sX ,H(0),sX X is (f (x), x) 7→ (x, x), and checking the 42 ROHIT DILIP HOLKAR necessary condition is same as in Example 3.10. Thus the proof follows from the same computation as in Example 3.10. This examples ends here. Let G and H be étale groupoids and X an étale correspondence from G to H. Then H(X) is the Hilbert C∗(H)-module; let σG : C∗(G) → C∗(H) be the representation which gives the C∗-correspondence. One may restrict the codomain of rX to G(0) and consider X as an étale correspondence from G(0) to H; we call this correspondence as the one obtained from X : G → H by restricting to G(0). Note that in this case, H(X) is the Hilbert C∗(H)-module. And since C0(G(0)) ⊆ C∗(G) is a C∗-subalgebra, the representation σG(0) : C∗(G) → H(X) is obtained by restricting σG to C0(G(0)). Thus (3.17) ( X : G → H X : G(0) → H gives gives (H(X), σG) : C∗(G) → C∗(H), (H(X), σG(0) ) : C0(G(0)) → C∗(H). Proposition 3.18. Let G and H be étale groupoids and X an étale correspondence from G to H. Then the following hold: i) The C∗-correspondence (H(X), σG) : C∗(G) → C∗(H) is proper if and only if X : G(0) → H obtained by restricting X : G → H to G(0) is proper in the sense of Definition 2.5. ii) If rX : X → G(0) is a proper map, then the C∗-correspondence (H(X), σG) from C∗(G) to C∗(H) is proper. Proof. (i): Let ι : C0(G(0)) ֒→ C∗(G) be the inclusion. Then the representations in Equation (3.17) are related as σG(0) = σG ◦ ι; this can be checked by writing formulae for the representations of the pre-C∗-algebras Cc(G(0)) and Cc(G) on the pre-Hilbert C∗(H)-module Cc(X). Therefore, if σG(C∗(G)) ⊆ K(H(X)), then σG(0) (C0(G(0))) ⊆ K(H(X)). Conversely, assume that X : G(0) → H is proper. Then Theorem 2.11 says that σG(0) (C0(G(0))) ⊆ K(H(X)). Let {ui} be the approximate unit of the pre-C∗- algebra Cc(G) where each ui ∈ Cc(G(0)); such an approximate unit is discussed at the beginning of this section. Note that for u ∈ Cc(G(0)), σG(u)(ξ) = σG(0) (u)(ξ) for every ξ ∈ Cc(X). Let f ∈ Cc(G). Then for every ξ ∈ Cc(X), we have σG(f )(ξ) = lim i σG(f ui)(ξ) = lim i σG(f )σG(ui)(ξ) where each σG(f )σG(ui) ∈ K(H(X)). Therefore, σG(f ) = limi σG(f )σG(ui) is also in K(H(X)). (ii): If rX : X → G(0) is a proper map, then Examples 3.16 says that X : G(0) → H is a proper correspondence in the sense of Definition 2.5. Theorem 2.11 implies that (H(X), σG) : C∗(G) → C∗(H) is a proper C∗-correspondence. (cid:3) Lemma 3.19. Let G and H be étale groupoids. Let (X, λ) : Y → H be a proper topological correspondence in the sense of Definition 2.5, then i) for each z ∈ Z, the measure λz is atomic, ii) for each z ∈ Z, each atom in λz is nonzero and the function D appearing in Definition 2.5(3) positive. Thus the family of measure λ has full support. Proof. In this case, since rX : X → G(0) is a proper map, Examples 3.16 says that the correspondence X : G(0) → H obtained by restricting X : G → H is proper. We prove the claims of present lemma for the restricted correspondence. If one writes the map msX for the restricted correspondence, then one observes that the proof of both the claims run on the same lines as the one of Lemma 3.8. (cid:3) PROPER TOPOLOGICAL CORRESPONDENCES 43 Acknowledgement: We are grateful to Ralf Meyer, Suliman Albandik and Al- cides Buss for many fruitful discussions. Special thanks to S. Albandik for many motivating discussions. We thank Ralf Meyer and Jean Renault for checking the manuscript, pointing out mistakes and suggesting improvements. This work is done during the author's stay in the Federal University of Santa Caterina, Florianopólis, Brazil and Indian Institute of Science Education and Research, Pune, India. The author is thankful to both the institutes for their hospitality. We are thankful to CNPq, Brazil and SERB, India whose funding made this work possible. References [1] Nicolas Bourbaki, Elements of mathematics. General topology. Part 1, Hermann, Paris; Addison-Wesley Publishing Co., Reading, Mass.-London-Don Mills, Ont., 1966. MR0205210 (34 #5044a) [2] , Integration. I. Chapters 1 -- 6, Elements of Mathematics (Berlin), Springer-Verlag, Berlin, 2004. Translated from the 1959, 1965 and 1967 French originals by Sterling K. Berbe- rian. MR2018901 (2004i:28001) [3] Alcides Buss, Rohit Dilip Holkar, and Ralf Meyer, A universal property for groupoid c∗- algebras (2016). Preprint- arXiv:1612.04963v1. [4] Alberto Candel and Lawrence Conlon, Foliations. I, Graduate Studies in Mathematics, vol. 23, American Mathematical Society, Providence, RI, 2000. MR1732868 [5] Gerald B. Folland, A course in abstract harmonic analysis, Studies in Advanced Mathematics, CRC Press, Boca Raton, FL, 1995. MR1397028 (98c:43001) [6] Michel Hilsum and Georges Skandalis, Morphismes K-orientés d'espaces de feuilles et fonc- torialité en théorie de Kasparov (d'après une conjecture d'A. Connes), Ann. Sci. École Norm. Sup. (4) 20 (1987), no. 3, 325 -- 390. MR925720 (90a:58169) [7] Rohit Dilip Holkar, Topological construction of C∗-correspondences for groupoid C ∗-algebras, Ph.D. Thesis, 2014. [8] [9] , Composition of topological correspondences, Journal of Operator Theory 78.1 (2017), 89 -- 117. , Topological construction of C ∗-correspondences for groupoid C ∗-algebras, Journal of Operator Theory 77:1 (2017), no. 23-24, 217 -- 241. [10] Rohit Dilip Holkar and Jean Renault, Hypergroupoids and C ∗-algebras, C. R. Math. Acad. Sci. Paris 351 (2013), no. 23-24, 911 -- 914. MR3133603 [11] Mahmood Khoshkam and Georges Skandalis, Regular representation of groupoid C ∗- algebras and applications to inverse semigroups, J. Reine Angew. Math. 546 (2002), 47 -- 72. MR1900993 [12] Marta Macho Stadler and Moto O'uchi, Correspondence of groupoid C ∗-algebras, J. Operator Theory 42 (1999), no. 1, 103 -- 119. MR1694789 (2000f:46077) [13] R. Meyer and C. Zhu, Groupoids in categories with pretopology, ArXiv e-print (August 2014), available at 1408.5220. [14] Paul S. Muhly, Jean N. Renault, and Dana P. Williams, Equivalence and isomorphism for groupoid C ∗-algebras, J. Operator Theory 17 (1987), no. 1, 3 -- 22. MR873460 (88h:46123) [15] Paul S. Muhly and Mark Tomforde, Topological quivers, Internat. J. Math. 16 (2005), no. 7, 693 -- 755. MR2158956 (2006i:46099) [16] Alan L. T. Paterson, Groupoids, inverse semigroups, and their operator algebras, Progress in Mathematics, vol. 170, Birkhäuser Boston, Inc., Boston, MA, 1999. MR1724106 (2001a:22003) [17] Jean Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, vol. 793, Springer, Berlin, 1980. MR584266 (82h:46075) [18] [19] , Représentation des produits croisés d'algèbres de groupoïdes, J. Operator Theory 18 (1987), no. 1, 67 -- 97. MR912813 (89g:46108) , Induced representations and hypergroupoids, SIGMA Symmetry Integrability Geom. Methods Appl. 10 (2014), Paper 057, 18. MR3226993 [20] Jean-Louis Tu, Non-Hausdorff groupoids, proper actions and K-theory, Doc. Math. 9 (2004), 565 -- 597 (electronic). MR2117427 (2005h:22004) E-mail address: [email protected] Department of Mathematics, Indian Institute of Science Education and Research, Dr. Homi Bhabha Road, NCL Colony, Pashan, Pune 411008, India.
1803.10446	2	1803	2018-11-08T13:56:06	On tensors of factorizable quantum channels with the completely depolarizing channel	[ "math.OA", "math-ph", "math-ph" ]	In this paper, we obtain results for factorizability of quantum channels. Firstly, we prove that if a tensor $T\otimes S_k$ of a quantum channel $T$ on $M_n(\mathbb{C})$ with the completely depolarizing channel $S_k$ is written as a convex combination of automorphisms on the matrix algebra $M_n(\mathbb{C})\otimes M_k(\mathbb{C})$ with rational coefficients, then the quantum channel $T$ has an exact factorization through some matrix algebra with the normalized trace. Next, we prove that if a quantum channel has an exact factorization through a finite dimensional von Neumann algebra with a convex combination of normal faithful tracial states with rational coefficients, then it also has an exact factorization through some matrix algebra with the normalized trace.	math.OA	math	On tensors of factorizable quantum channels with the completely depolarizing channel Yuki Ueda Abstract In this paper, we obtain results for factorizability of quantum channels. Firstly, we prove that if a tensor T ⊗ Sk of a quantum channel T on Mn(C) with the completely depolarizing channel Sk is written as a convex combination of automorphisms on the matrix algebra Mn(C) ⊗ Mk(C) with rational coefficients, then the quantum channel T has an exact factorization through some matrix algebra with the normalized trace. Next, we prove that if a quantum channel has an exact factorization through a finite dimensional von Neumann algebra with a convex combination of normal faithful tracial states with rational coefficients, then it also has an exact factorization through some matrix algebra with the normalized trace. keywords: Markov maps, factorizable quantum channels, free group von Neumann al- gebras, completely depolarizing channels 1 Introduction In [1], Anantharaman-Delaroche defined the class of factorizable Markov maps to study the noncommutative analogue of ergodic theory. After the works of [1], Haagerup and Musat proved in [3, Theorem 6.1] that every non-factorizable quantum channel on Mn(C) (n ≥ 3) fails the asymptotic quantum Birkhoff conjecture which was raised by Smolin, Verstraete and Winter (see [7]) as one of the most important problems in quantum infor- mation theory. In [3, 4], they also approached the Connes embedding problem by using factorizable quantum channels, in particular, tensors of factorizable quantum channels with the completely depolarizing channel. In this paper, we focus on the relation between the factorizability of quantum channels and the property of tensors of factorizable quan- tum channels with the completely depolarizing channel. Haagerup and Musat proved in [4, Corollary 3.5] that if a quantum channel T on Mn(C) has an exact factorization through a tracial W ∗-probability space (Mn(C) ⊗ Mk(C), τn ⊗ τk), that is, there exists a unitary matrix u in Mn(C) ⊗ Mk(C) such that T is written in the following form T x = (idn ⊗ τk)(u∗(x ⊗ 1k)u), x ∈ Mn(C), (1.1) then T ⊗ Sk ∈ conv(Aut(Mn(C) ⊗ Mk(C))). We raise the natural problem of whether the converse claim of the statement is true or not (see below). Problem 1.1. Let n be a positive integer and T a quantum channel on Mn(C). Is it true that the following properties are equivalent? (1) There exists a positive integer k such that T has an exact factorization through 1 (Mn(C) ⊗ Mk(C), τn ⊗ τk). (2) There exists a positive integer k such that T ⊗ Sk ∈conv(Aut(Mn(C) ⊗ Mk(C))). The implication (1) ⇒ (2) is true, but the implication (2) ⇒ (1) is unknown. We obtain a partial result for Problem 1.1 as follows. Theorem 1.2. Let T be a quantum channel on Mn(C). If there exists a positive integer k i=1 ciad(ui) ∈conv(Aut(Mn(C) ⊗ Mk(C))) for some positive integer d(k), unitary matrices u1, · · · , ud(k) ∈ U (Mn(C) ⊗ Mk(C)) and positive rational numbers i=1 ci = 1, then there exists a positive integer L such that T has an such that T ⊗ Sk =Pd(k) c1, · · · , cd(k) with Pd(k) exact factorization through (Mn(C) ⊗ ML(C), τn ⊗ τL). Moreover we also raise the following problem for the quantum channels which have an exact factorization through a finite dimensional W ∗-probability space (see below). Problem 1.3. Let T be a quantum channel on Mn(C). Is it true that if there exists a finite dimensional W ∗-probability space (N , φ) (i.e. a pair of a finite dimensional von Neumann algebra N and a normal faithful state φ on N ) such that T has an exact factorization through (Mn(C) ⊗ N , τn ⊗ φ) then there exists a positive integer k such that T also has an exact factorization through (Mn(C) ⊗ Mk(C), τn ⊗ τk)? Note that every finite dimensional von Neumann algebra is ∗-isomorphic to a direct sum of some matrix algebras. For this problem, we obtain the following partial result. Theorem 1.4. Let T be a quantum channel on Mn(C). If for any α := (α1, · · · , αd) ∈ Qd + with α1 + · · · + αd = 1 there exist positive integers k1, · · · , kd such that T has an exact factorization through (Mn(C) ⊗ (Mk1(C) ⊕ · · · ⊕ Mkd(C)), τn ⊗ τα), where τα is a normal faithful tracial state on Mk1(C) ⊕ · · · ⊕ Mkd(C) defined by τα(x1, · · · , xd) := α1τk1(x1) + · · · + αdτkd(xd) (1.2) for all (x1, · · · , xd) ∈ Mk1(C) ⊕ · · · ⊕ Mkd(C), then there exists a positive integer k such that T has an exact factorization through (Mn(C) ⊗ Mk(C), τn ⊗ τk). In section 2, we set notations and definitions in this paper. In section 3, we recall and discuss for the concepts and basic properties of quantum channels, factorizable quantum channels and completely depolarizing channels. In section 4 and 5, we prove that Theorem 1.2 and Theorem 1.4 hold, respectively. 2 Notations and Definitions In this paper, we use the following notations: • N := {1, 2, 3, · · · } and Q+ is the set of all positive rational numbers. • Mn(C) is the set of all n × n matrices with complex entries. • U (n) is the set of all n × n unitary matrices with complex entries. • τn is the normalized trace on Mn(C), i.e. τn((xij )1≤i,j≤n) := x11+···xnn n . 2 • idn is the identity map on Mn(C). • Define ad(u)(x) := u∗xu for all x ∈ Mn(C), u ∈ U (n). • conv(Aut(Mn(C))) is the convex hull of the set Aut(Mn(C)) := {ad(u) : u ∈ U (n)}. • 1M is the unit in von Neumann algebra M, in particular, 1n := 1Mn(C). A pair (M, φ) is called a W ∗-probability space if M is a von Neumann algebra and φ is a normal faithful state on M. In particular, we call (M, φ) a tracial W ∗-probability space when φ is tracial, that is φ(xy) = φ(yx) for all x, y ∈ M. 3 Basic properties of factorizable quantum channels In [1], Anantharaman-Delaroche considered factorizable Markov maps to prove a noncom- mutative analogue of Rota's theorem. We first recall the definition of Markov maps on a W ∗-probability space. The concept is a noncommutative analogue of measure-preserving Markov operator on a probability space. Definition 3.1. Let (M, φ) and (N , ψ) be W ∗-probability spaces. A linear map T : M → N is called a (φ, ψ)-Markov map if (1) T is completely positive, (2) T is unital, (3) T is (φ, ψ)-preserving, i.e. ψ ◦ T = φ, (4) T ◦ σφ In particular, we call it φ-Markov map when (M, φ) = (N , ψ). t }t∈R denotes the automorphism group of the state φ. t = σψ t ◦ T , where {σφ If (M, φ) = (N , ψ) = (Mn(C), τn) in Definition 3.1, the fourth condition is removed since the operator στ is trivial for any normal faithful tracial states τ on von Neumann t algebras and t ∈ R, so that a τn-Markov map means a unital completely positive trace- preserving map (quantum channel) on Mn(C). Denote by Q(n) := {T : Mn(C) → Mn(C) : T is a quantum channel}. (3.1) In [1, Definition 6.2], Anantharaman-Delaroche defined the class of factorizable Markov maps in the following sense. Definition 3.2. A (φ, ψ)-Markov map T : M → N is called factorizable if there exists a W ∗-probability space (L, χ) and ∗-monomorphisms α : M → L and β : N → L such that α is (φ, χ)-Markov, β is (ψ, χ)-Markov and T = β∗ ◦ α, where β∗ : L → M is the adjoint of β (see [3, Remark 1.2]). The set of all factorizable (φ, ψ)-Markov maps is closed under composition, the adjoint operation, taking convex combinations and w∗-limits (See [6, Proposition 2]). Haagerup and Musat proved in [3, Theorem 2.2] the following statement for the class of factorizable quantum channels. Proposition 3.3. Consider T ∈ Q(n). Then the following properties are equivalent: (1) T is factorizable, (2) There exists a tracial W ∗-probability space (M, φ) and a unitary u in Mn(C) ⊗ M such that T x = (idn ⊗ φ)(u∗(x ⊗ 1M)u), x ∈ Mn(C). (3.2) 3 In this case, we say that T has an exact factorization through (Mn(C) ⊗ M, τn ⊗ φ). A factorization of quantum channels is not unique. We have two examples of a factorization of quantum channels. Firstly we show the following statement. Lemma 3.4. If T ∈ Q(n) has an exact factorization through a tracial W ∗-probability space (Mn(C) ⊗ M, τn ⊗ φ) and there exist a tracial W ∗-probability space (N , ψ) and a (φ, ψ)-Markov -homomorphism S : (M, φ) → (N , ψ), then T also has an exact factoriza- tion through (Mn(C) ⊗ N , τn ⊗ ψ). Proof. Since T has an exact factorization through (Mn(C) ⊗ M, τn ⊗ φ), there exists a unitary u ∈ Mn(C) ⊗ M such that T x = (idn ⊗ φ)(u∗(x ⊗ 1M)u), x ∈ Mn(C). (3.3) Since S is a -homomorphism, (idn ⊗ S)(u) is a unitary in Mn(C) ⊗ N . Hence, (idn ⊗ ψ)(cid:0)(idn ⊗ S)(u)∗(x ⊗ 1N )(idn ⊗ S)(u)(cid:1) = (idn ⊗ ψ)(idn ⊗ S)(u∗(x ⊗ 1M)u) = (idn ⊗ φ)(u∗(x ⊗ 1M)u) = T x, (3.4) for all x ∈ Mn(C). Therefore T has an exact factorization through (Mn(C)⊗N , τn⊗ψ). As the second example, we consider a linear map T defined by a∗ i xai, 2 T x = Xi=1 a1 :=(cid:18)1 0 0 0(cid:19) , x ∈ M2(C), a2 :=(cid:18)0 0 0 1(cid:19) , (3.5) (3.6) where i = 12. Hence T is a quantum channel on M2(C) by [2]. Let g1, g2 be generators of the free i=1 ai ⊗ λgi ∈ U (M2(C) ⊗ LF2), where λg is the left has an exact factorization through M2(C) ⊗ LF2. It is clear thatP2 group F2 of degree 2 and set u :=P2 representation of g ∈ F2, that is, i ai =P2 i=1 aia∗ i=1 a∗ λg(f )(h) := f (g−1h), f ∈ l2F2, g, h ∈ F2, (3.7) and LF2 is the free group von Neumann algebra. For all x ∈ M2(C), (id2 ⊗ τLF2)(u∗(x ⊗ 1LF2)u) = τLF2(λ∗ gi λgj )a∗ i xaj = 2 Xi,j=1 2 Xi,j=1 δij a∗ i xaj = T x, (3.8) where τLF2(λ) :=< λδe, δe >l2F2, λ ∈ LF2, (3.9) and e ∈ F2 is the unit of F2. Therefore T has an exact factorization through (M2(C) ⊗ LF2, τ2 ⊗ τLF2). On the other hand, T also has an exact factorization through (M2(C) ⊗ M4(C), τ2 ⊗ τ4) since a1, a2 ∈ M2(C) are self-adjoint, a2 2 = 12 and a1a2 = a2a1 and therefore we can apply [3, Corollary 2.5] to the quantum channel T . Similarly we have the following statement. 1 + a2 4 Proposition 3.5. Consider d ≥ 2. Let T be a quantum channel on Md(C) defined by T x := d Xi=1 E∗ iixEii, x ∈ Md(C), (3.10) where {Eij}1≤i,j≤d is the set of standard matrix units in Md(C). Then the following conditions hold. (1) T has an exact factorization through (Md(C) ⊗ LFd, τd ⊗ τLFd). (2) T has an exact factorization through (Md(C) ⊗ M2d(C), τd ⊗ τ2d). Denote by F(n) := {T ∈ Q(n) : T is factorizable}. (3.11) By the statements before Proposition 3.3, we have that conv(Aut(Mn(C)))⊂ F(n) for all positive integers n. Haagerup and Musat found a quantum channel in F(n)\conv(Aut(Mn(C))) in [3, Example 3.3]. In particular, Kummerer proved in [5] that conv(Aut(M2(C)))=F(2). Haagerup and Musat pointed out the important relations between the factorizable quan- tum channels and the Connes embedding problem in [3, 4]. Recall the completely depolarizing channels. Let Sk : Mn(C) → Mn(C) be a linear map defined by Sk(x) := τk(x)1k, x ∈ Mk(C). (3.12) The map Sk is called the completely depolarizing channel on Mn(C). Note that Sk is in conv(Aut(Mk(C))) and therefore it is a factorizable quantum channel on Mn(C). It is clear that Sk ⊗ Sl = Skl for all k, l ∈ N. It is very important to understand tensors T ⊗ Sk of a quantum channel T with the completely depolarizing channel Sk to know what T has an exact factorization through some W ∗-probability space. By [3, Corollary 2.5], Haagerup and Musat found a quantum channel T ∈ F(n)\ conv(Aut(Mn(C))) (n ≥ 3) such that it has an exact factorization through (Mn(C) ⊗ M2d(C), τn ⊗ τ2d) for some d ≥ 3. By [4, Corollary 3.5], we have that T ⊗ S2d ∈ conv(Aut(Mn(C) ⊗ M2d(C))). 4 Proof of Theorem 1.2 We prove that Theorem 1.2 holds. We first introduce the following sets. T ∈ F(n)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) In :=  Jn :=(cid:26)T ∈ F(n)(cid:12)(cid:12)(cid:12)(cid:12) i=1 ciad(ui) for some positive integer d(k), ∃k ∈ N s.t. T ⊗ Sk =Pd(k) c1, · · · , cd(k) ∈ Q+ with Pd(k) ,   i=1 ci = 1, and u1, · · · , ud(k) ∈ U (Mn(C) ⊗ Mk(C)) ∃k ∈ N s.t. T has an exact factorization through (Mn(C) ⊗ Mk(C), τn ⊗ τk) (4.1) (cid:27) . By using these notations, we can rewrite as the statement of Theorem 1.2. Theorem 4.1. Let n ∈ N. Then In ⊂ Jn holds. 5 Proof. Suppose that n ∈ N. If T ∈ In, then there is a positive integer k > 0 such that i zui, for all z ∈ Mn(C) ⊗ Mk(C) for some positive integer d(k) > 0, unitaries u1, · · · , ud(k) ∈ Mn(C) ⊗ Mk(C), and positive rational numbers c1, · · · , cd(k) > 0 i=1 ciu∗ T ⊗ Sk(z) =Pd(k) with Pd(k) i=1 ci = 1. Therefore T x = (idn ⊗ τk)(T ⊗ Sk)(x ⊗ 1k) = ci(idn ⊗ τk)(u∗ i (x ⊗ 1k)ui), x ∈ Mn(C). (4.2) d(k) Xi=1 Since ci is a rational number for each 1 ≤ i ≤ d(k), we suppose that ci = li Li , 1 ≤ i ≤ d(k), (4.3) where li and Li are relatively prime positive numbers for each i. Let L be the least common multiple of L1, L2, · · · , Ld(k). Then we can rewrite as: ci = li × L Li L , 1 ≤ i ≤ d(k) (4.4) and rewrite as Ci := li × L Li Then we set ∈ N for each i. Note that C1 + C2 + · · · + Cd(k) = L. C1 C2 Cd(k) U := diag( u1, · · · , u1, u2 · · · , u2, · · · , ud(k), · · · , ud(k)) ∈ Mn(C) ⊗ Mk(C) ⊗ ML(C). (4.5) z } { z } { z { Clearly the block matrix U is a unitary in Mn(C) ⊗ Mk(C) ⊗ ML(C), and we have that (idn ⊗ τk ⊗ τL)(U ∗(x ⊗ 1k ⊗ 1L)U ) } C1 } u∗ 1(x ⊗ 1k)u1, · · · , u∗ { 1(x ⊗ 1k)u1, · · ·   { d(k)(x ⊗ 1k)ud(k)(cid:1) d(k)(x ⊗ 1k)ud(k) ⊗ EL−cd(k)+i,L−cd(k)+i ! u∗ d(k)(x ⊗ 1l)ud(k), · · · , u∗ } Cd(k) i=1 u∗ 1(x ⊗ 1k)u1) + · · · i=1 τL(EL−cd(k)+i,L−cd(k)+i)(idn ⊗ τk)(u∗ d(k)(x ⊗ 1k)ud(k)) · · · , z · · · +PCd(k) i=1 u∗ 1(x ⊗ 1k)u1 ⊗ Eii + · · · z = (idn ⊗ τk ⊗ τL) diag(cid:0)  = (idn ⊗ τk ⊗ τL) PC1 = PC1 Xi=1 Xi=1 ci(idn ⊗ τk)(u∗ (idn ⊗ τk)(u∗ Ci L d(k) d(k) = = i=1 τL(Eii)(idn ⊗ τk)(u∗ · · · +PCd(k) i (x ⊗ 1k)ui) i (x ⊗ 1k)ui) = T x, for all x ∈ Mn(C), where {Eij}1≤i,j≤L is the set of standard matrix units in ML(C). Therefore T has an exact factorization through (Mn(C)⊗(Mk(C)⊗ML(C)), τn ⊗(τk ⊗τL)). Hence we have the inclusion In ⊂ Jn, for each positive integer n ∈ N. Thus the proof is complete. 6 (4.6) 5 Proof of Theorem 1.4 We prove that Theorem 1.4 holds. We also denote the following set. Kn := {T ∈ F(n) : ∃k ∈ N s.t. T ⊗ Sk ∈ conv(Aut(Mn(C) ⊗ Mk(C)))}. (5.1) Clearly conv(Aut(Mn(C)))⊂ Kn. The statement before the condition (1.1) in section 1 or [4, Corollary] implies that Jn ⊂ Kn holds for each positive integer n. By the last statements in section 3, the set Kn\conv(Aut(Mn(C)))) is not empty for all n ≥ 3. We can rewrite as the statement of Theorem 1.4 by the new notations in sections 4 and 5. Theorem 5.1. Let n ∈ N. If for any α := (α1, · · · , αd) ∈ Qd + with α1 + · · · + αd = 1 there exist some positive integers k1, · · · , kd such that T ∈ F(n) has an exact factorization through (Mn(C) ⊗ (Mk1(C) ⊕ · · · ⊕ Mkd(C)), τn ⊗ τα), where τα is defined by (1.2), then T ∈ Jn, and therefore T ∈ Kn. Proof. (Step 1) Suppose that α := (α1, · · · , αd) ∈ Qd exist k ∈ N and the normal faithful tracial state τα defined by + with α1 + · · · + αd = 1 and there d τα(x1, · · · , xd) := α1τk(x1) + · · · + αdτk(xd), (x1, · · · , xd) ∈ Mk(C) ⊕ · · · ⊕ Mk(C) } { (5.2) z d } z such that T has an exact factorization through (Mn(C)⊗( Since α1, · · · , αd are rational numbers, we can write as Mk(C) ⊕ · · · ⊕ Mk(C)), τn ⊗ τα). αi = li L , 1 ≤ i ≤ d, (5.3) where li and L are positive integers and l1 + · · · + ld = L. We will define the following map: z d } φ : Mk(C) ⊕ · · · ⊕ Mk(C) → MkL(C), By the definition of φ, homomorphism. We consider xi = (xi { xd, · · · , xd) (5.4) it is easy to check that φ is a unital completely positive - (x1, · · · , xd) 7→ diag( x1, · · · , x1, · · · , } { z } { l1 ld sl)1≤s,l≤k ∈ Mk(C) (1 ≤ i ≤ d). Then { z τkL ◦ φ(x1, · · · , xd) = τkL(diag( x1, · · · , x1, · · · , xd, · · · , xd) = z } { z } { l1 ld 1 kL d Xi=1 k Xj=1 lixi jj. (5.5) On the other hand, τα(x1, · · · , xd) = αiτk(xi) = d Xi=1 d Xi=1 li L(cid:16) 1 k k Xj=1 xi jj(cid:17) = 1 kL d Xi=1 k Xj=1 lixi jj. (5.6) Therefore φ is a unital completely positive (τα, τkL)-preserving -homomorphism from d z } { Mk(C) ⊕ · · · ⊕ Mk(C), τα) to (MkL(C), τkL). By Lemma 3.4, T also has an exact factoriza- ( tion through (Mn(C)⊗MkL(C), τn⊗τkL). By [4, Corollary 3.5], T ⊗SkL ∈ conv(Aut(Mn(C)⊗ 7 MkL(C))). Therefore T ∈ Jn ⊂ Kn. (Step 2) Suppose that α := (α1, α2) ∈ Q2 + with α1 + α2 = 1 and assume that T has an exact factorization through (Mn(C) ⊗ (Mk1(C) ⊕ Mk2(C)), τn ⊗ τα) (In general, k1 6= k2), where τα is a normal faithful tracial state on Mk1(C) ⊗ Mk2(C) defined by τα(x, y) := α1τk1(x) + α2τk2(y), (x, y) ∈ Mk1(C) ⊗ Mk2(C). (5.7) We define the following map: k2 k1 ψ : Mk1(C)⊕Mk2(C) → Mk1k2(C)⊕Mk1k2(C), y, · · · , y)). (5.8) It is clear that ψ is a unital completely positive -homomorphism. Consider two matrices x = (xij)1≤i,j≤k1 ∈ Mk1(C) and y = (yij)1≤i,j≤k2 ∈ Mk2(C). Then x, · · · , x), diag( (x, y) 7→ (diag( z } { z } { k2 k1 (α1τk1k2 ⊕ α2τk1k2) ◦ ψ(x, y) = (α1τk1k2 ⊕ α2τk1k2)(diag( x, · · · , x), diag( k1 y, · · · , y)) z } { z } { y, · · · , y)) (5.9) = α1τk1k2(diag( x, · · · , x)) + α2τk1k2(diag( z } { yjj(cid:17) Xj=1 k2 k1k2 k2 z } { xii(cid:17) + α2(cid:16) k1 Xj=1 α2 k2 yjj k2 k1 Xi=1 xii + k1k2 = α1(cid:16) k2 Xi=1 α1 k1 = k1 = α1τk1(x) + α2τk2(y) = τα(x, y), where α1τk1k2 ⊕ α2τk1k2 is a normal faithful tracial state on Mk1k2(C) ⊕ Mk1k2(C) defined by α1τk1k2 ⊕ α2τk1k2(x, y) := α1τk1k2(x) + α2τk1k2(y), (x, y) ∈ Mk1k2(C) ⊕ Mk1k2(C). (5.10) Therefore ψ is a unital completely positive (τα, α1τk1k2⊕α2τk1k2)-preserving -homomorphism. By using Lemma 3.4, T has an exact factorization through (Mn(C)⊗(Mk1k2(C)⊕Mk1k2(C)), τn⊗ (α1τk1k2 ⊕ α2τk1k2)). Denote by α1 = l1 L , α2 = l2 L , l1, l2, L ∈ N with l1 + l2 = L. (5.11) By using the first step, T has an exact factorization through (Mn(C) ⊗ Mk1k2L(C), τn ⊗ τk1k2L), and therefore T ∈ Jn ⊂ Kn by [4, Corollary 3.5]. (Step 3) Assume that α := (α1, · · · , αd) ∈ Qd + with α1 + · · ·+ αd = 1 and there exist some positive integers k1, · · · , kd and a normal faithful tracial state τα on Mk1(C)⊕· · · ⊕Mkd(C) such that T has an exact factorization through (Mn(C)⊗(Mk1(C)⊕· · ·⊕Mkd(C)), τn ⊗τα). By using repeatedly first step or second step, there exists positive integer k such that T also has an exact factorization through (Mn(C) ⊗ Mk(C), τn ⊗ τk). Therefore T ∈ Jn ⊂ Kn by [4, Corollary 3.5] again. 8 Acknowledgment This paper is a revised version of the master's thesis of the author. The author would like to express his heartly thanks to Professor Benoıt Collins for carefully reading this paper and pointing out some inaccuracies. The author had a chance to visit Profes- sor Magdalena Musat and Professor Mikael Rørdam (in Copenhagen university) during November-December, 2016 with the support of the Top Global University project for Ky- oto University. The author would appreciate their hospitality when the author was staying at Copenhagen University. In particular, thanks to their important advices, the author could advance researches related to the sections 4 and 5. References [1] C. Anantharaman-Delaroche, On ergodic theorems for free group actions on noncom- mutative spaces, Probab. Theory Related Fields 135 (2006), no. 4, 520-546. [2] M. D. Choi, Completely positive linear maps on complex matrices, Linear Algebra and Appl. 10, Issue 3, (1975), 285-290. [3] U. Haagerup, M. Musat, Factorization and dilation problems for completely positive maps on von Neumann algebras. Comm. Math. Phys. 303 (2011), no. 2, 555-594. [4] U. Haagerup, M. Musat, An asymptotic property of factorizable completely positive maps and the Connes embedding problem. Comm. Math. Phys. 338 (2015), no. 2, 721-752. [5] B. Kummerer, Markov dilations on the 2 × 2 matrices, In Operator algebras and their connections with topology and ergodic theory (Busteni, 1983), Lect. Notes Math. 1132, Berlin, Springer, 1985, 312-323. [6] E. Ricard, A Markov dilation for self-adjoint Schur multipliers, Proc. Amer. Math. Soc. 136 (2008), no. 12, 4365-4372. [7] J. A. Smolin, F. Verstraete, A. Winter, Entanglement of assistance and multipartite state distillation, Phys. Rev. A 72, 052317 (2005) 1-10. Yuki Ueda Department of Mathematics, Hokkaido University, Kita 10, Nishi 8, Kita-Ku, Sapporo, Hokkaido, 060-0810, Japan email: [email protected] 9
1703.04256	1	1703	2017-03-13T05:19:36	Connes integration formula for the noncommutative plane	[ "math.OA" ]	Our aim is to prove the integration formula on the noncommutative (Moyal) plane in terms of singular traces {\it a la} Connes.	math.OA	math	CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE F. SUKOCHEV AND D. ZANIN Abstract. Our aim is to prove the integration formula on the noncommuta- tive (Moyal) plane in terms of singular traces a la Connes. 1. Introduction Let M be a compact Riemannian manifold. The following formula can be found in p. 34 in [1] and in Corollary 7.21 in [9]. (1) Trω(Mf (1 − ∆)− d 2 ) = ZM f dvol, f ∈ C∞(M ). Here, Mf is the multiplication operator, ∆ is the Hodge-Laplacian operator on L2(M, vol) and Trω is the Dixmier trace on the ideal L1,∞ (see Section 2). Also, Corollary 7.22 in [9] wrongly extends this result to f ∈ L1(M, vol) (in fact, f ∈ L2(M, vol) is the necessary and sufficient condition for this formula to hold; see [14] or the book [15] for detailed proofs). According to [1], formula (1) "led Connes to introduce the Dixmier trace as the correct operator theoretical substitute for integration of infinitesimals of order one in non-commutative geometry." It appears suitable to refer to (1) and similar results as the "Connes Integration Formula". Compactness of the (resolvent of the) Hodge-Dirac operator plays a crucial role in the proofs of Connes Integration Formula for unital spectral triples (see [1] and [9]). For non-unital spectral triples (including non-compact manifolds), the proofs become radically harder. Even the case of the simplest non-compact manifold Rd required a substantial effort and the first reasonable answer was very recently given in [11] (see the book [15] for detailed proofs). In this paper, we investigate the validity of Connes Integration Formula for the noncommutative (Moyal) plane Rd θ (here, θ is a non-degenerate antisymmetric matrix). Earlier attempts in this direction can be found in [8] (see Proposition 4.17 there), [2] and [3]. We substantially strengthen corresponding results from these papers and present a completely different approach to Connes Integration Formula. The novelty of our approach is in the consistent use of Cwikel estimates for the noncommutative plane (obtained in a recent paper [12]) -- see Section 2. Our main result is the following theorem. Theorem 1.1. If x ∈ W d,1(Rd 2 ∈ L1,∞ and θ), then x(1 − ∆)− d ϕ(x(1 − ∆)− d 2 ) = τθ(x) for every normalised continuous trace ϕ on L1,∞. 1 2 CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE Here, W d,1(Rd θ) is a Sobolev space on Rd trace on L∞(Rd θ). θ and τθ is the faithful normal semifinite Section 2 involves the preliminaries necessary to prove Theorem 1.1. In Section 3, we prove that ϕ(x(1 − ∆)− d 2 ) = cϕτθ(x), x ∈ W d,1(Rd θ), for every normalised trace on L1,∞. In Section 4, we construct one particular x ∈ W d,1(Rd 2 ) does not depend on the choice of a normalised continuous trace ϕ. The combination of these results yield Theorem 1.1. θ) such that ϕ(x(1 − ∆)− d 2. Preliminaries 2.1. General notation. Fix throughout a separable infinite dimensional Hilbert space H. We let L(H) denote the algebra of all bounded operators on H. For a compact operator T on H, let µ(k, T ) denote k−th largest singular value (these are the eigenvalues of T ). The sequence µ(T ) = {µ(k, T )}k≥0 is referred to as to the singular value sequence of the operator T. The standard trace on L(H) is denoted by Tr. Fix an orthonormal basis in H (the particular choice of a basis is inessential). We identify the algebra l∞ of bounded sequences with the subalgebra of all diagonal operators with respect to the chosen basis. For a given sequence α ∈ l∞, we denote the corresponding diagonal operator by diag(α). 2.2. Schatten ideals Lp and Lp,∞, p > 0. For every p > 0, we set Lp = {T ∈ L(H) : Tr(T p) < ∞}. We set kT kp = (cid:0)Tr(T p)(cid:1) 1 p , T ∈ Lp. For every p > 0, k · kp is a quasi-norm1 and (Lp, k · kp) is a quasi-Banach space. For p ≥ 1, k · kp is a norm. For p < 1, the space (Lp, k · kp) is not Banach -- that is, its quasi-norm is not equivalent to any norm. For a given 0 < p ≤ ∞, we let Lp,∞ denote the principal ideal in L(H) generated by the operator diag({(k + 1)− 1 p }k≥0). Equivalently, Lp,∞ = {T ∈ L(H) : µ(k, T ) = O((k + 1)−1/p)}. We set kT kp,∞ = sup k≥0 (k + 1)1/pµ(k, T ), T ∈ Lp,∞. For every p > 0, k · kp,∞ is a quasi-norm and (Lp,∞, k · kp,∞) is a quasi-Banach space. For p > 1, k · kp,∞ is equivalent to a (unitarily invariant Banach) norm. For p ≤ 1, the space (Lp,∞, k · kp,∞) is not Banach -- that is, its quasi-norm is not equivalent to any norm. In [17], the Banach envelope of L1,∞ was thoroughly investigated. 1A quasinorm satisfies the norm axioms, except that the triangle inequality is replaced by x + y ≤ K(x + y) for some uniform constant K > 1. CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE 3 2.3. Traces on L1,∞. Definition 2.1. If I is an ideal in L(H), then a unitarily invariant linear functional ϕ : I → C is said to be a trace. Since U −1T U − T = [U −1, T U ] for all T ∈ I and for all unitaries U ∈ L(H), and since the unitaries span L(H), it follows that traces are precisely the linear functionals on I satisfying the condition ϕ(T S) = ϕ(ST ), T ∈ I, S ∈ L(H). The latter may be reinterpreted as the vanishing of the linear functional ϕ on the commutator subspace which is denoted [I, L(H)] and defined to be the linear span of all commutators [T, S] : T ∈ I, S ∈ L(H). It is shown in Lemma 5.2.2 in [15] that ϕ(T1) = ϕ(T2) whenever 0 ≤ T1, T2 ∈ I are such that the singular value sequences µ(T1) and µ(T2) coincide. For p > 1, the ideal Lp,∞ does not admit a non-zero trace [7], while for p = 1, there exists a plethora of traces on L1,∞ (see e.g. [18] or [15]). A standard example of a trace on L1,∞ is a Dixmier trace introduced in [6] that we now explain. Definition 2.2. Let ω be a free ultrafilter on Z+. The functional Trω : A → lim n→ω 1 log(2 + n) n Xk=0 µ(k, A), 0 ≤ A, is finite and additive on the positive cone of L1,∞. Therefore, it extends to a trace on L1,∞. We call such traces Dixmier traces. These traces clearly depend on the choice of the ultrafilter ω on Z+. Using a slightly different definition, this notion of trace was applied by Connes [4] in noncommutative geometry. An extensive discussion of traces, and more recent developments in the theory, may be found in [15] including a discussion of the following facts. We refer the reader to an alternative approach to the theory of traces on L1,∞ suggested in [18] (based on the fundamental paper [16] by Pietsch). (1) All Dixmier traces on L1,∞ are positive. (2) All positive traces on L1,∞ are continuous in the quasi-norm topology. (3) There exist positive traces on L1,∞ which are not Dixmier traces (see [18]). (4) There exist traces on L1,∞ which fail to be continuous (see [15]). Definition 2.3. We say that an operator A ∈ L1,∞ is measurable if ϕ(A) does not depend on the choice of the continuous normalised trace ϕ on L1,∞. 2.4. Noncommutative plane: algebra. Each assertion in this subsection is rig- orously established in Section 6 in [12]. Our approach to the noncommutative plane is to introduce the von Neumann algebra generated by a strongly continuous family of unitary operators {U (t)}t∈Rd, d ∈ N, satisfying the commutation relation (2) U (t + s) = exp(− ht, θsi)U (t)U (s), t, s ∈ Rd, i 2 where θ is a fixed antisymmetric real d × d matrix. Namely, we set (3) (U (t)ξ)(u) = e− i 2 ht,θuiξ(u − t), ξ ∈ L2(Rd), u, t ∈ Rd. 4 CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE Definition 2.4. Let d ∈ N and let θ be a fixed non-degenerate2 antisymmetric real d × d matrix. The von Neumann subalgebra in L(L2(Rd)) generated by {U (t)}t∈Rd, introduced in (3), is called the noncommutative plane and denoted by L∞(Rd θ). Example 2.5. If d = 2, then L∞(Rd t → U2(t), t ∈ R satisfying the condition θ) is generated by 2 unitary groups t → U1(t), U1(t1)U2(t2) = eiαt1t2 U2(t2)U1(t1), t1, t2 ∈ R. Here, U1(t1) = U ((t1, 0)) and U2(t2) = U ((0, t2)). The following assertion is well-known. In [12], a spatial isomorphism is con- structed. Theorem 2.6. For every non-degenerate antisymmetric real matrix θ, the algebra L∞(Rd θ) is isomorphic to L(L2(R 2 )). d Having established the isomorphism between r : L∞(Rd θ) → L(L2(R d 2 )) we now equip L∞(Rd θ) with a faithful normal semifinite trace τθ = Tr ◦ r. We can now define Lp−spaces on L∞(Rd θ). θ) : τθ(xp) < ∞o. θ) = nx ∈ L∞(Rd Lp(Rd Lemma 2.7. An operator x ∈ L∞(Rd θ) is in L2(Rd (2π)d/4 ZRd for some unique f ∈ L2(Rd) with kxk2 = kf k2. x = Op(f ) def = 1 θ) if and only if3 f (s)U (s)ds Note that our picture is the Fourier dual of the one considered in [8]. More precisely, the paper [8] deals with operators of the form Op(F f ), where f is Schwartz (in [8], these operators are written simply as f ). 2.5. Noncommutative plane: calculus. Each assertion in this subsection is rigorously established in Section 6 in [12]. Let Dk, 1 ≤ k ≤ d be multiplication operators on L2(Rd) (Dkξ)(t) = tkξ(t), ξ ∈ L2(Rd). For brevity, we denote ∇ = (D1, · · · , Dd). For every 1 ≤ k ≤ d, we have (4) Moreover, we have [Dk, U (s)] = skU (s), s ∈ Rd. (5) eiht,∇iU (s)e−iht,∇i = eiht,siU (s), s, t ∈ Rd. If [Dk, x] ∈ L(L2(Rd)) for some x ∈ L∞(Rd θ), then [Dk, x] ∈ L∞(Rd θ). This crucial fact allows us to introduce mixed partial derivative ∂αx of x ∈ L∞(Rd θ). 2A non-degenerate antisymmetric matrix is automatically of even order. 3To be precise, where the limit is taken in L2(Rd limit in order to lighten the notations. x = lim 1 (2π)d/4 Z[−N,N]d f (s)U (s)ds, N→∞ θ ). In what follows, we write the integral over Rd instead of the CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE 5 Definition 2.8. Let α be a multiindex and let x ∈ L∞(Rd θ). If every repeated com- mutator [Dαj , [Dαj +1, · · · , [Dαn, x]]], 1 ≤ j ≤ n, is a bounded operator on L2(Rd), then the mixed partial derivative ∂αx of x is defined as ∂αx = [Dα1, [Dα2 , · · · , [Dαn , x]]]. In this case, we have that ∂αx ∈ L∞(Rd θ). As usual, ∂0x = x. Therefore, we can introduce the Sobolev space W m,p(Rd θ) associated with the noncommutative plane in the following way. Definition 2.9. For m ∈ Z+ and p ≥ 1, the space W m,p(Rd x ∈ Lp(Rd This space is equipped with the norm, θ) such that every partial derivative of x up to order m is also in Lp(Rd θ) is the space of θ). kxkW m,p = Xα≤m k∂αxkp, x ∈ W m,p(Rd θ). The following assertion is one of the main results in [12]. Theorem 2.10. If x ∈ W d,1(Rd (a) x(1 − ∆)− d+1 2 ∈ L1 and θ), then (b) x(1 − ∆)− d 2 ∈ L1,∞ and kx(1 − ∆)− d+1 2 k1 ≤ cdkxkW d,1. kx(1 − ∆)− d 2 k1,∞ ≤ cdkxkW d,1. 3. Integration formula modulo a constant factor For every φ ∈ L∞(Rd), we define a bounded operator Tφ : L2(Rd θ) → L2(Rd θ) by the formula Tφ : ZRd f (s)U (s)ds → ZRd f (s)φ(s)U (s)ds, f ∈ L2(Rd). Lemma 3.1. If φ is a Schwartz function, then Tφ : L1(Rd θ) → L1(Rd θ). Proof. We claim that Tφx = ZRd (F φ)(u)U (−θ−1u)xU (θ−1u)du, x ∈ L2(Rd θ). Since both sides above define bounded operators on L2(Rd θ) and since the set {Op(f ) : f is Schwartz} is dense in L2(Rd θ), it suffices to establish the claim for x = ZRd f (s)U (s)ds, f ∈ S(Rd). Using the inverse Fourier transform, we write φ(s) = ZRd (F φ)(u)eihu,sidu, s ∈ Rd. Since both f and F φ are Schwartz functions, it follows that Tφx = ZZRd×Rd f (s)(F φ)(u)eihu,siU (s)dsdu. It follows from (2) that eihu,siU (s) = U (−θ−1u)U (s)U (θ−1u). 6 CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE Therefore, Tφx = ZRd (F φ)(u)(cid:16)ZRd f (s)U (−θ−1u)U (s)U (θ−1u)ds(cid:17)du. Using the definition of x, we obtain ZRd f (s)U (−θ−1u)U (s)U (θ−1u)ds = U (−θ−1u)xU (θ−1u). This proves the claim. Now, we prove the assertion of the lemma as follows. kTφxk1 ≤ ZRd (F φ)(u) · kU (−θ−1u)xU (θ−1u)k1du = kF φk1kxk1. (cid:3) Lemma 3.2. For every x ∈ W d,1(Rd θ), the mapping t → U (−t)xU (t), t ∈ Rd, is a continuous W d,1(Rd θ)−valued function. Moreover, kU (−t)xU (t)kW d,1 = kxkW d,1 . Proof. It follows from Leibniz rule that [Dk, U (−t)xU (t)] = [Dk, U (−t)]·xU (t)+U (−t)·[Dk, x]·U (t)+U (−t)x·[Dk, U (t)] = = −tkU (−t)xU (t) + U (−t)[Dk, x]U (t) + tkU (−t)xU (t) = U (−t)[Dk, x]U (t). Iterating the latter inequality, we obtain ∂α(U (−t)xU (t)) = U (−t)∂α(x)U (t). Thus, kU (−t)xU (t)kW d,1 = Xα≤d k∂α(U (−t)xU (t))k1 = = Xα≤d kU (−t)∂α(x)U (t)k1 = Xα≤d k∂α(x)k1 = kxkW d,1. We now establish the continuity. For every y ∈ L1, the mapping t → V (−t)yV (t), t ∈ Rd, is continuous in the L1−norm whenever the mapping t → V (t) is strongly con- tinuous. Recall that (L∞(Rd θ), τθ) is ∗−isomorphic (so that trace is preserved) to (L(L2(R 2 )), Tr). Thus, the mapping d t → U (−t)∂α(x)U (t) = ∂α(U (−t)xU (t)) is continuous in L1−norm. This completes the proof. (cid:3) Lemma 3.3. (a) If f is Schwartz, then Op(f ) ∈ W d,1(Rd (b) The set {Op(f ) : f is Schwartz} is dense in L1(Rd θ). θ). In particular, W d,1(Rd θ) is dense in L1(Rd θ). Proof. There exists a sequence {ekl}k,l≥0 ⊂ L∞(Rd θ) such that (i) ek1l1 ek2l2 = δl1,k2 ek1l2 and e∗ (ii) τθ(ekk) = 1. kl = elk. (iii) Pk≥0 ekk = 1 in strong operator topology. (iv) for every k, l ≥ 0, there exists a Schwartz function fkl such that ekl = Op(fkl). CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE 7 The existence of such a sequence is established in Lemma 2.4 in [8] (see also addi- tional references therein). A particular formula for fkl can be found on p. 618 in [8] in terms of Laguerre polynomials. We prove (a). Let f be a Schwartz function. By Proposition 2.5 in [8], one can write f as Thus, f = Xk,l≥0 cklfkl, Xk,l≥0 ckl < ∞. Op(f ) = Xk,l≥0 cklekl, where the series converges in L1−norm. Thus, Op(f ) ∈ L1(Rd θ). Let fα(t) = tαf (t), t ∈ Rd. By (4), ∂α(Op(f )) = Op(fα). Since fα is also a Schwartz function, it follows that ∂α(Op(f )) ∈ L1(Rd θ). This proves (a). To prove (b), note that, for every x ∈ L1(Rd θ), Xk,l≤N ekkxell = (Xk≤N ekk)x(Xl≤N ell) → x in L1−norm as N → ∞. Note that ekkxell is a scalar multiple of ekl = Op(fkl). Since a linear combination of Schwartz functions is again a Schwartz function, it follows that ekkxell ∈ {Op(f ) : f is Schwartz} ⊂ W d,1(Rd θ). Xk,l≤N This proves (b). (cid:3) Lemma 3.4. If F is a continuous functional on W d,1(Rd θ) such that F (x) = F (U (−t)xU (t)), x ∈ W d,1(Rd θ), t ∈ Rd, then F = τθ (up to a constant factor). Proof. Let T : W d,1(Rd θ) be defined by setting θ) → W d,1(Rd T x = ZRd U (−θ−1t)xU (θ−1t)e− 1 2 t2 dt. The integral is understood as a Bochner integral of a continuous W d,1(Rd θ)−valued function (the continuity and convergence of the integral follow from Lemma 3.2). For every x ∈ W d,1(Rd θ), we have F (T x) = ZRd Thus, F (U (−θ−1t)xU (θ−1t))e− 1 2 t2 dt = ZRd F (x)e− 1 2 t2 dt = (2π) d 2 F (x). x ∈ W d,1(Rd θ). We claim that kT xkW d,1 ≤ cdkxk1 for every x ∈ W d,1(Rd θ). To see this, let F (x) = (2π)− d 2 F (T x), x = ZRd f (s)U (s)ds, f ∈ L2(Rd). If, in the proof of Lemma 3.1, we select φ(t) = e− 1 given there yields 2 t2 , t ∈ Rd, then the argument T x = ZRd f (s)U (s)e− 1 2 s2 ds. 8 CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE By (4), we have ∂α(T x) = ZRd , s ∈ Rd. We have that ∂α ◦ T = Tφα. By Lemma 3.1, f (s)U (s)sαe− 1 2 s2 ds. θ) is a bounded operator. This proves the claim. Let φα(s) = sαe− 1 Tφα : L1(Rd 2 s2 θ) → L1(Rd For every x ∈ W d,1(Rd F (x) = (2π)− d 2 F (T x) ≤ (2π)− d θ), we have 2 kF k(W d,1)∗ kT xkW d,1 ≤ cdkF k(W d,1)∗kxk1. Thus, a functional F on W d,1(Rd Banach Theorem, F extends to a bounded functional on L1(Rd exists y ∈ L∞(Rd θ) is bounded in k · k1−norm. By the Hahn- θ). Hence, there θ) such that F (x) = τθ(xy), x ∈ W d,1(Rd θ). Clearly, F (U (−t)xU (t)) = τθ(U (−t)xU (t)y) = τθ(xU (t)yU (−t)). Comparing the last 2 equalities, we obtain τθ(xU (t)yU (−t)) = τθ(xy), x ∈ W d,1(Rd θ). θ) is dense in L1(Rd Since W d,1(Rd θ), it follows that y = U (t)yU (−t) for every t ∈ Rd. In other words, y commutes with every U (t) and, therefore, with every element in L∞(Rd θ) is a factor (see Theorem 2.6), it follows that y is a scalar operator. This completes the proof. (cid:3) θ). Since L∞(Rd The following proposition is a light version of Theorem 1.1. Proposition 3.5. If x ∈ W d,1(Rd θ), then x(1 − ∆)− d 2 ) = cϕτθ(x) ϕ(x(1 − ∆)− d 2 ∈ L1,∞ and for every continuous trace on L1,∞ and for some constant cϕ. Proof. By Theorem 2.10 (b), the functional F : x → ϕ(x(1 − ∆)− d 2 ), x ∈ W d,1(Rd θ), is a well defined bounded linear functional on W d,1(Rd θ). Since ϕ is unitarily invariant, it follows that ϕ(x(1 − ∆)− d 2 ) = ϕ(eiht,∇ix(1 − ∆)− d 2 e−iht,∇i), t ∈ Rd. By the Spectral Theorem, we have (1 − ∆)− d 2 e−iht,∇i = e−iht,∇i(1 − ∆)− d 2 , and so ϕ(x(1 − ∆)− d 2 ) = ϕ(eiht,∇ixe−iht,∇i(1 − ∆)− d 2 ). For every s ∈ Rd, we have (see (5)) eiht,∇iU (s)e−iht,∇i = eiht,siU (s). On the other hand, it follows from (2) that U (−θ−1t)U (s)U (θ−1t) = eiht,siU (s). Comparing preceding equalities, we arrive at eiht,∇iU (s)e−iht,∇i = U (−θ−1t)U (s)U (θ−1t). CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE 9 It follows that eiht,∇ixe−iht,∇i = U (−θ−1t)xU (θ−1t), x ∈ L∞(Rd θ). Combining the preceding paragraphs, we obtain ϕ(x(1 − ∆)− d 2 ) = ϕ(U (−θ−1t)xU (θ−1t)(1 − ∆)− d 2 ). Applying Lemma 3.4 to our functional F, we conclude the argument. (cid:3) 4. Proof of measurability Lemma 4.1. If K ∈ W 2d+2,1([0, 1]d ×[0, 1]d) and if T : L2((0, 1)d) → L2((0, 1)d) is an integral operator with integral kernel K, then T ∈ L1 and kT k1 ≤ cdkKkW 2d+2,1. Proof. Let K ∈ W 2d+2,1([−π, π]d × [−π, π]d) be an extension of K such that kKkW 2d+2,1([−π,π]d×[−π,π]d) ≤ cdkKkW 2d+2,1([0,1]d×[0,1]d) and such that K vanishes on and near the boundary. Thus, K ∈ W 2d+2,1(Td × Td). Let S : L2(Td) → L2(Td) be an integral operator with integral kernel K. We have T = Mχ(0,1)d SMχ(0,1)d . Thus, kT k1 ≤ kSk1. Let us write Fourier series K(t, s) = Xm1,m2∈Zd cm1,m2em1(t)em2 (s), t, s ∈ Td. Set Sm1,m2ξ = hξ, e−m2iem1 , ξ ∈ L2(Td). It is an integral operator on L2(Td) with the integral kernel (t, s) → em1(t)em2 (s). Hence, S = Xm1,m2∈Zd cm1,m2Sm1,m2. By triangle inequality, we have kSk1 ≤ Xm1,m2∈Zd cm1,m2 ≤ ≤ sup m1,m2∈Zd (1 + m12 + m22)d+1cm1,m2 · Xm1,m2∈Zd (1 + m12 + m22)−d−1. Observe that (1 + m12 + m22)d+1cm1,m2 is the (m1, m2)−th Fourier coefficient of the function (1 − ∆T2d )d+1(K) (here, ∆T2d is the Laplacian on the torus T2d). Taking into account that Fourier coefficients do not exceed the L1−norm, we infer that (1 + m12 + m22)d+1cm1,m2 ≤ (2π)−2dk(1 − ∆T2d )d+1Kk1 ≤ cdkKkW 2d+2,1. Here, the last inequality follows from the definition of a Sobolev space. (cid:3) In what follows, we consider the tensor product of 2 bounded operators on a Hilbert space H as a bounded operator on the Hilbert space H ¯⊗H. Lemma 4.2. If T ∈ L1,∞ and S ∈ L1, then S ⊗ T ∈ L1,∞ and (6) ϕ(S ⊗ T ) = Tr(S) · ϕ(T ) for every continuous trace ϕ on L1,∞. 10 CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE Proof. Firstly, we show that S ⊗ T ∈ L1,∞. Let z(t) = t−1, t > 0. By definition, we have µ(T ) ≤ kT k1,∞z. The crucial fact that µ(S ⊗ z) = kSk1z is proved on p. 211 in [13]. Thus, kS ⊗ T k1,∞ = kS ⊗ µ(T )k1,∞ ≤ kT k1,∞kS ⊗ zk1,∞ = kT k1,∞kSk1. We now turn to the proof of (6). If S is a rank one projection, then there is nothing to prove. If S is a positive finite rank operator, then the assertion follows by linearity. If S is an arbitrary finite rank operator, then the assertion again follows by linearity. Let S ∈ L1 be arbitrary. Fix ǫ > 0 and choose S1, S2 ∈ L1 such that S = S1 +S2, S1 is finite rank and kS2k1 ≤ ǫ. Clearly, ϕ(S ⊗ T ) − Tr(S) · ϕ(T ) = = (ϕ(S1 ⊗ T ) − Tr(S1) · ϕ(T )) + (ϕ(S2 ⊗ T ) − Tr(S2) · ϕ(T )). By the preceding paragraph, the summand in the first bracket vanishes. Thus, ϕ(S ⊗ T ) − Tr(S) · ϕ(T ) = ϕ(S2 ⊗ T ) − Tr(S2) · ϕ(T ). Hence, ϕ(S ⊗ T ) − Tr(S) · ϕ(T ) ≤ ϕ(S2 ⊗ T ) + Tr(S2) · ϕ(T ) ≤ ≤ kϕkL∗ 1,∞ · (kS2 ⊗ T k1,∞ + Tr(S2)kT k1,∞). By the norm estimate in the first paragraph and by the assumption on S2, we have ϕ(S ⊗ T ) − Tr(S) · ϕ(T ) ≤ 2ǫkϕkL∗ 1,∞kT k1,∞. Since ǫ > 0 is arbitrarily small, the assertion follows. (cid:3) In the following lemma, we consider the direct sum of bounded operators on a Hilbert space H as a bounded operator on a Hilbert space Lm≥0 H. Lemma 4.3. If the operators {Tm}m≥0 are pairwise orthogonal, i.e. Tm1Tm2 = m1Tm2 = 0 for m1 6= m2, then Pm≥0 Tm is unitarily equivalent4 to Lm≥0 Tm. T ∗ Here, the sums are taken in the weak operator topology. Proof. Let p1 and p2 be projections on H. Since t → t monotone function for every n ≥ 1, it follows that 1 n , t > 0, is an operator p1 = p 1 n 1 ≤ (p1 + p2) 1 n sot→ supp(p1 + p2). Similarly, p2 ≤ supp(p1 + p2) and, therefore, p1 ∨ p2 ≤ supp(p1 + p2). This simple fact can be also found in Proposition 2.5.14 in [10]. Let pm = supp(Tm) and qm = supp(T ∗ m). It follows from the assumption that pm1pm2 = pm1qm2 = qm1 qm2 = 0, m1 6= m2. Set rm = pm ∨ qm. We have (pm1 + qm1 )(pm2 + qm2) = 0, m1 6= m2. Thus, supp(pm1 + qm1 ) · supp(pm2 + qm2 ) = 0, m1 6= m2. where rm is the projection defined in the proof of Lemma 4.3. Clearly, Tm is unitarily equivalent 4To be pedantic, Pm≥0 Tm is unitarily equivalent to the direct sum Lm≥0 Tmrm(H)→rm(H), to the direct sum Tmrm(H)→rm(H) L 0(1−rm )(H)→(1−rm)(H). Thus, a direct sum Lm≥0 Tm is unitarily equivalent to (Pm≥0 Tm)L 0. In what follows, we ignore this subtle difference and write unitary equivalence as stated in Lemma 4.3. CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE 11 By the preceding paragraph, we have rm1rm2 = 0, m1 6= m2. If T = Pm≥0 Tm, then rmT = Tm and T rm = Tm for every m ≥ 0. Thus, T = Lm≥0 Tm, where Tm acts on the Hilbert space rm(H). (cid:3) Let h(t) = (1 + d Xk=1 ⌊tk⌋2)− d 2 , t ∈ Rd. The following proposition yields a special case of Theorem 1.1. Proposition 4.4. If f is a Schwartz function supported on [−1, 1]d and if x = Op(f ), then xh(∇) is measurable. Proof. Step 1: We have that xh(∇) is an integral operator with the kernel K : (t, s) → f (t − s)h(s)e i 2 hs,θti, t, s ∈ R2. By assumption on f, we have that s ∈ m1 + [0, 1]d, f (s − t) = 0, Thus, t ∈ m2 + [0, 1]2, m1 − m2 /∈ {−1, 0, 1}d. xh(∇) = Xl1,l2∈{−1,0,1}d Tl1,l2, Tl1,l2 = Xm∈Zd m=l2mod3 h(m)Tm,l1, where Tm,l1 is an integral operator whose integral kernel is given by the formula i (t, s) → f (t − s)e 2 hs,θtiχm+l1+[0,1]d (t)χm+[0,1]d (s), Step 2: We claim that Tl1,l2 ∈ L1,∞ and is measurable. Note that the operators {Tm,l1} m∈Zd t, s ∈ Rd, are pairwise orthogonal. Therefore, we have (∼ denotes unitary equivalence) m=l2mod3 Tl1,l2 ∼ Mm∈Zd m=l2mod3 (1 + m2)− d 2 Tm,l1. By definition, Tm,l1 : L2(m + [−1, 2]d) → L2(m + [−1, 2]d). Define a unitary operator by setting Um : L2([−1, 2]d) → L2(m + [−1, 2]d) (Umξ)(t) = e i 2 hm,θtiξ(t − m), ξ ∈ L2([−1, 2]d), t ∈ m + [−1, 2]d. Define an operator Sl1 : L2([−1, 2]d) → L2([−1, 2]d) to be an integral operator with the integral kernel (t, s) → f (t − s)e i 2 hs,θtiχl1+[0,1]d (t)χ[0,1]d (s), t, s ∈ [−1, 2]d. A direct computational argument shows that5 Tm,l1 = UmSl1 U −1 m . 5Indeed, Thus, (U −1 m ξ)(t) = e− i 2 hm,θtiξ(t + m), ξ ∈ L2(m + [−1, 2]d), t ∈ [−1, 2]d. (Sl1 U −1 m ξ)(t) = χl1+[0,1]d (t) · Z[0,1]d f (t − s)e i 2 hs,θ(t+m)iξ(s + m)ds. 12 CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE Hence, Tl1,l2 ∼ Mm∈Zd m=l2mod3 (1 + m2)− d 2 Sl1 ∼ Sl1 ⊗n(1 + m2)− d 2o m∈Zd . m=l2mod3 By Lemma 4.1, Sl1 ∈ L1. The claim follows now from Lemma 4.2. Proof of Theorem 1.1. Choose a Schwartz function f0 supported on [−1, 1]d such that f0(0) 6= 0 and let x0 = Op(f0). Set (cid:3) k(t) = (1 + t2) d+1 2 · (cid:16)(1 + t2)− d 2 − (1 + d Xk=1 2(cid:17), ⌊tk⌋2)− d t ∈ Rd. Clearly, k is a bounded function on Rd. By Lemma 3.3 (a), we have x0 ∈ W d,1(Rd θ). Using the obvious equality x0(1 − ∆)− d 2 − x0h(∇) = x0(1 − ∆)− d+1 2 · k(∇) and Theorem 2.10 (a), we infer that x0(1 − ∆)− d 2 − x0h(∇) ∈ L1. By Proposition 4.4, we have that x0h(∇) is measurable and, hence, so is the oper- ator x0(1 − ∆)− d 2 . Let now x ∈ W d,1(Rd θ) be arbitrary. Since f0 is a Schwartz function, it follows that Without loss of generality, τθ(x0) = 1. Let z = x − τθ(x)x0 ∈ W d,1(Rd τθ(z) = 0. We have θ). Clearly, τθ(x0) = f0(0) 6= 0. ϕ(x(1 − ∆)− d 2 ) = ϕ(z(1 − ∆)− d 2 ) + τθ(x) · ϕ(x0(1 − ∆)− d 2 ). By Proposition 3.5, the first summand vanishes. By the preceding paragraph, the second summand does not depend on ϕ. This completes the proof. (cid:3) References [1] Benameur M., Fack T. Type II non-commutative geometry. I. Dixmier trace in von Neumann algebras. Adv. Math. 199 (2006), no. 1, 29 -- 87. [2] Carey A., Gayral V., Rennie A., Sukochev F. Integration on locally compact noncommutative spaces. J. Funct. Anal. 263 (2012), no. 2, 383 -- 414. [3] Carey A., Gayral V., Rennie A., Sukochev F. Index theory for locally compact noncommuta- tive geometries. Mem. Amer. Math. Soc. 231 (2014), no. 1085. [4] Connes A. Noncommutative Geometry. Academic Press, San Diego, 1994. [5] Connes A. The action functional in noncommutative geometry. Comm. Math. Phys. 117 (1988), no. 4, 673 -- 683. [6] Dixmier J. Existence de traces non normales. (French) C. R. Acad. Sci. Paris Ser. A-B 262 (1966) A1107 -- A1108. [7] Dykema K., Figiel T., Weiss G., Wodzicki M. Commutator structure of operator ideals. Adv. Math. 185 (2004), no. 1, 1 -- 79. Thus, (UmSl1 U −1 m ξ)(t) = χl1+[0,1]d (t − m) · Z[0,1]d = χm+l1+[0,1]d (t) · Zm+[0,1]d e i 2 hm,θtif (t − s − m)e i 2 hs,θtiξ(s + m)ds = f (t − s)e i 2 hs,θtiξ(s)ds. CONNES INTEGRATION FORMULA FOR THE NONCOMMUTATIVE PLANE 13 [8] Gayral V., Gracia-Bondia J., Iochum B., Schucker T., Varilly J. Moyal planes are spectral triples. Comm. Math. Phys. 246 (2004), no. 3, 569 -- 623. [9] Gracia-Bondia J., Varilly J., Figueroa H. Elements of noncommutative geometry. Birkhauser Advanced Texts: Basel Textbooks, Birkhauser Boston, Inc., Boston, MA, 2001. [10] Kadison R., Ringrose J. Fundamentals of the theory of operator algebras. Vol. I. Elemen- tary theory. Reprint of the 1983 original. Graduate Studies in Mathematics, 15. American Mathematical Society, Providence, RI, 1997. [11] Kalton N., Lord S., Potapov D., Sukochev F. Traces of compact operators and the noncom- mutative residue. Adv. Math. 235 (2013), 1 -- 55. [12] Levitina G., Sukochev F., Zanin D. Cwikel estimates revisited. submitted manuscript. [13] Lindenstrauss J., Tzafriri L. Classical Banach spaces. II. Function spaces. Ergebnisse der Mathematik und ihrer Grenzgebiete, 97. Springer-Verlag, Berlin-New York, 1979. [14] Lord S., Potapov D., Sukochev F. Measures from Dixmier traces and zeta functions. J. Funct. Anal. 259 (2010), no. 8, 1915 -- 1949. [15] Lord S., Sukochev F., Zanin D. Singular traces. Theory and applications. De Gruyter Studies in Mathematics, 46, De Gruyter, Berlin, 2013. [16] Pietsch A. Traces and shift invariant functionals. Math. Nachr. 145 (1990), 7 -- 43. [17] Pietsch A. About the Banach envelope of l1,∞. Rev. Mat. Complut. 22 (2009), no. 1, 209 -- 226. [18] Semenov E., Sukochev F., Usachev A., Zanin D. Banach limits and traces on L1,∞. Adv. Math. 285 (2015), 568 -- 628. School of Mathematics and Statistics, University of New South Wales, Kensington, 2052, Australia E-mail address: [email protected] School of Mathematics and Statistics, University of New South Wales, Kensington, 2052, Australia E-mail address: [email protected]
1610.06956	1	1610	2016-10-21T21:10:13	Compact and "compact" operators on the standard Hilbert module over a $W^*$ algebra	[ "math.OA", "math.FA" ]	We construct a topology on the standard Hilbert module $l^2(\mathcal A)$ over a unital $W^*$-algebra $\mathcal A$ such that any "compact" operator, (i.e.\ any operator in the norm closure of the linear span of the operators of the form $x\mapsto\left<y,x\right>z$) maps bounded sets into totally bounded sets.	math.OA	math	COMPACT AND "COMPACT" OPERATORS ON THE STANDARD HILBERT MODULE OVER A W ∗ ALGEBRA DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Abstract. We construct a topology on the standard Hilbert module l2(A) over a unital W ∗-algebra A such that any "compact" operator, (i.e. any op- erator in the norm closure of the linear span of the operators of the form x 7→ hy, xi z) maps bounded sets into totally bounded sets. 1. Introduction Given a unital W ∗-algebra A we consider the standard Hilbert module denoted by l2(A), (the notation HA is also widespread) +∞ l2(A) = {x = (ξ1, ξ2, . . . ) ξj ∈ A, X j=1 ξ∗ j ξj converges in the norm topology}, equipped with the A-valued inner product ξ∗ j ηj ∈ A, l2(A) × l2(A) ∋ (x, y) 7→ X +∞ j=1 x = (ξ1, ξ2, . . . ), y = (η1, η2, . . . ). Since an arbitrary A-linear bounded operator on l2(A) does not need to have an adjoint, the natural algebra of operators is Ba(l2(A)) - the algebra of all A-linear bounded operators on l2(A) having an adjoint. It is known that Ba(l2(A)) is a C∗-algebra and also that it is a W ∗-algebra whenever A is of that kind. Among all operators in Ba(l2(A)), those that belong to the linear span of the operators of the form x 7→ Θy,z(x) = z hy, xi (y, z ∈ l2(A)) are called finite rank operators. The norm closure of finite rank operators is known as the algebra of all "compact" operators. The quotation marks are usually written in order to emphasize the fact that "compact" operators does not maps bounded sets into relatively compact sets, as it is the case in the framework of Hilbert (and also Banach) spaces, though they share many properties of proper compact operators on a Hilbert space, [5], [6] For general literature concerning Hilbert modules over more general C∗ algebras, including the standard Hilbert module, the reader is referred to [4] or [7]. The aim of this note is to introduce a locally convex topology on l2(A), where A is a unital W ∗-algebra, such that any "compact" operator maps bounded sets (in the norm) into totally bounded in the introduced topology. In a very special case, where A ∼= B(H) the algebra of all bounded operators on a Hilbert space, 2010 Mathematics Subject Classification. Primary: 46L08, Secondary: 47B07, 54E15. Key words and phrases. Hilbert module, uniform spaces, compact operators. The authors was supported in part by the Ministry of education and science, Republic of Serbia, Grant #174034. 1 2 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ the converse is also true. Namely, any operator T ∈ Ba(l2(A)) that maps bounded into totally bounded sets is "compact". Therefore, speaking freely, we can omit the quotation marks. 2. Preliminaries Let us recall some basic definitions and facts concerning uniform spaces. For more details see [1] or [3]. Uniform spaces are those topological spaces in which one can deal with notions such as Cauchy sequence, Cauchy net or uniform continuity. Although it is usual to define them as spaces endowed with a family of sets in X × X given as some kind of neighborhoods of the diagonal, so called entourages, for our purpose it is more convenient to give an equivalent definition, via a family of semimetrics. Definition 2.1. A nonempty set endowed with a family of semimetrics, functions dα : X × X → [0, +∞) satisfying (i) dα(x, y) ≥ 0; (ii) dα(x, y) = dα(y, x); (iii) dα(x, z) ≤ dα(x, y) + dα(y, z) is called a uniform space. All dα are metrics except they do not distinguish points, i.e. there might be dα(x, y) = 0 for some x 6= y. However it is provided that for all x 6= y there is an α such that dα(x, y) > 0. The family of sets Bdα(x; ε) = {y ∈ X dα(x, y) < ε} makes a basis for some topology. It is well known that a topological space X is a uniform space if and only if it is completely regular. Let X be a uniform space. For a net xi ∈ X we say that it is a Cauchy net if it is a Cauchy net with respect to all dα, i.e. if for all α and for all ε > 0 there is i0 such that for all i, j > i0 we have dα(xi, xj) < ε. The notion of a complete uniform space is defined in an obvious way. A set A ⊆ X is called totally bounded, if for all ε > 0 and all α there is a finite set c1, c2, . . . , cm ∈ X such that Bα(cj ; ε) = {y ∈ X dα(cj, y) < ε} cover A. It is well known that any relatively compact set is totally bounded, and that the converse is true provided that X is complete. If X is not complete then there are totally bounded sets that are not relatively compact, for instance, Q ∩ [0, 1] as a subset of Q. (See also [1, Remark 4.2.2]) Any locally convex topological vector space is a uniform space. Indeed, there is a family of seminorms generating its topology. This family can be obtained by Minkowski functionals of basic neighborhoods of zero. And an arbitrary seminorm define a semimetric in a natural way. Conversely, any family of seminorms that distinguishes points leads to a locally convex Hausdorff topological vector space. Hence a family of seminorms allows to deal with notions: totally bounded set, complete space, Cauchy net, etc. 3. Topology For an arbitrary Hilbert W ∗-module M, Paschke [8], [9] in his initial works on Hilbert C∗-modules and Frank [2], introduced two topologies, τ1 and τ2, the first of them generated by functionals x 7→ ϕ(hy, xi), y ∈ M, ϕ normal state, and the second by seminorms p(x) = ϕ(hx, xi)1/2, ϕ normal state. Frank proved that M is self-dual if and only if the unit ball in M is complete in τ1 (and this is equivalent to the completeness in τ2). Therefore, if M = l2(A) is a standard Hilbert module, it is not complete neither in τ1 nor in τ2, since l2(A) is never self-dual, except in COMPACT AND "COMPACT" OPERATORS 3 the case where A is finite dimensional algebra. Since obviously τ1 ⊂ τ2, we shall refer to these topologies as to weak PF and strong PF topologies. However, we need a topology which is between weak and strong PF topology. Namely, on a standard Hilbert module l2(A) where A is a unital W ∗ algebra we define a locally convex topology τ by the family of seminorms (3.1) +∞ pϕ,y(x) = vuut X j=1 ϕ(η∗ j ξj)2, where ϕ is a normal state, and y = (η1, η2, . . . ) is a sequence of elements in A such that (3.2) ϕ(η∗ j ηj) = 1. sup j≥1 Proposition 3.1. Seminorms (3.1) are well defined. Also τ1 ⊂ τ ⊂ τ2. Proof. Since (ξ, η) 7→ ϕ(η∗ξ) is a semi inner product, we have ϕ(η∗ By this, and by (3.2) we have j ξj)2 ≤ ϕ(ξ∗ j ξj )ϕ(η∗ j ηj). (3.3) pϕ,y(x)2 = +∞ X j=1 ϕ(η∗ j ξj)2 ≤ +∞ X j=1 ϕ(ξ∗ j ξj)ϕ(η∗ j ηj ) ≤ +∞ X j=1 ϕ(ξ∗ j ξj ) = ϕ(hx, xi). This proves that seminorms (3.1) are well defined, and also that τ ⊂ τ2. ζj = ηj/ϕ(η∗ Hence To prove τ1 ⊂ τ , pick y ∈ l2(A), y = (η1, η2, . . . ). The sequence ζj given by j ηj ) 6= 0, and ζj = 0 otherwise obviously fulfils (3.2). j ηj)1/2 if ϕ(η∗ ϕ(hy, xi) = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ  +∞ X j=1   finishing the proof. ≤ +∞ X j=1 η∗ j ξj (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)   j ηj)  ϕ(η∗ +∞ X j=1 = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = ϕ(hy, yi)1/2pϕ,z(x), (cid:3) ϕ(η∗ j ηj)1/2ϕ(ζ∗ j ξj)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1/2 ≤ 1/2   +∞ X j=1 ϕ(ζ∗ j ξj)2  Remark 3.1. The dual module of the module M is defined as the module of all A-linear and A-valued bounded functionals. It is denoted by M′. The module M always can be embedded in M′ via M ∋ y 7→ Λy ∈ M′, Λy(x) = hy, xi. If this embedding is onto, the module M is called self-dual. It is well known that l2(A) is not self-dual, except the algebra A is finite dimensional. Namely, l2(A)′ can be described as the module of all sequences x = (ξ1, ξ2, . . . ) such that the sequence of sums Pn j ξj is norm bounded, [7, Proposition 2.5.5]. j=1 ξ∗ Reading carefully the proof of the preceding proposition, one can see that nothing is changed if we replace l2(A) by l2(A)′. Indeed, the entire proof does not depend on the norm convergence of the series P+∞ Proposition 3.2. The unit ball in l2(A) is not complete with respect to τ . Its completion is the unit ball in the dual module l2(A)′. j=1 ξ∗ j ξj. 4 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Let x ∈ l2(A)′, x = (ξ1, ξ2, . . . ). Since the sequence of sums Pn Proof. First, we prove that the unit ball in l2(A) is dense in the unit ball in l2(A)′. j ξj is bounded it is convergent in strong (or weak, or ultraweak etc.) topology. By normality of ϕ we have j=1 ξ∗ +∞ ϕ X  j=1 ξ∗ j ξj   = +∞ X j=1 ϕ(ξ∗ j ξj), implying that ϕ(P+∞ j=n ξ∗ j ξj ) → 0, as n → +∞. the form (3.1). Thus, by the inequality (3.3) (ξ1, ξ2, . . . , ξn, 0, 0, . . . ) → x in each seminorm of Next, we prove that l2(A)′ is complete. Let xα = (ξα 2 , . . . ) be a Cauchy net in the unit ball. Choosing an arbitrary normal state, and ηk = 1, ηj = 0 for j 6= k we obtain that ξα k is a Cauchy net in weak-∗ topology in the unit ball in A. Hence, it is convergent, say ξα Since multiplying is ultraweakly continuous, for any n ∈ N and for all ηj which k → ξk in the weak-∗ topology. 1 , ξα satisfy (3.2) we have k j )2 → j ξα ϕ(η∗ X j=1 j ξj )1/2 if ϕ(ξ∗ k X j=1 ϕ(η∗ j ξj)2. Choosing ηj = ξj/ϕ(ξ∗ j ξj) 6= 0 and ηj = 0 otherwise, we get k X j=1 ϕ(ξ∗ j ξj ) = X k=1 ϕ(η∗ j ξj)2 = lim α k X j=1 ϕ(η∗ j ξα j )2 ≤ x ≤ 1. Taking the limit as k → +∞ we conclude that x = (ξ1, ξ2, . . . ) ∈ l2(A)′. To see that x is the limit of the Cauchy net xα it is enough to take the limit over β in k X j=1 ϕ(η∗ j ξα j ) − ϕ(η∗ j ξβ j )2 ≤ +∞ X j=1 ϕ(η∗ j ξα j ) − ϕ(η∗ j ξβ j )2 < ε, and finally the limit as k → +∞. (cid:3) Next, we want to study the restriction of τ to module An seen as a submodule of l2(A) consisting of those x for which ξj = 0 for all j > n. Proposition 3.3. a) On An the weak PF and our topology coincide, i.e. we have τ1An = τAn ; b) The embedding i : An → l2(A), i(ξ1, . . . , ξn) = (ξ1, . . . , ξn, 0, . . . ) is continu- ous with respect to (τAn , τ ). Proof. a) We already have τ1 ⊆ τ . Let us prove the converse. An arbitrary semi- norm of the form (3.1) restricted to An has the form (3.4) n pϕ,y(x) = vuut X j=1 ϕ(η∗ j ξj)2. Consider the vectors yj = (0, . . . , 0, ηj, 0, . . . , 0), where ηj is the j-th entry. Then pϕ,y(x) = vuut n X j=1 ϕ(hyj, xi)2 ≤ n X j=1 ϕ(hyj, xi), COMPACT AND "COMPACT" OPERATORS from which we conclude that pϕ,y is continuous with respect to τ1; b) One can easily check that i−1({x pϕ,η1,...,ηn,...(x) < ε}) = {(ξ1, . . . , ξn) pϕ,η1,...,ηn (ξ1, . . . , ξn) < ε}. 5 (cid:3) Proposition 3.4. The unit ball in An is compact with respect to τAn . Since An is self-dual, the unit ball is also complete and hence totally bounded. Proof. In the case n = 1, both topologies τ and τ1 are generated by seminorms ξ 7→ ϕ(η∗ξ), η ∈ A, ϕ normal state. It is easy to verify that these topologies are exactly the weak-∗ topology on A. Therefore, in this special case the conclusion follows by Banach-Alaoglu theorem. To obtain the result in general case, consider the product topology on An = A × ··· × A. Basic neighborhoods of zero have the form {(ξ1, ξ2, . . . , ξn) ∀j = 1, 2, . . . , n ϕj (η∗ j ξj ) < εj}. Due to the inequalities 1≤j≤n ϕ(η∗ max j ξj) ≤ vuut n X j=1 j ξj)2 ≤ √n max 1≤j≤n ϕ(η∗ ϕ(η∗ j ξj), the topology τ is weaker then the product topology. Since the product of unit balls is compact in stronger, product topology, and since τ is Hausdorff, we conclude that τ coincides with the product topology on the product of unit balls. Therefore, it remains to show that the unit ball in An is closed in the product of n unit balls in A, i.e. that its complement is open. Let z = (ζ1, . . . , ζn) ∈ An, z > 1 be arbitrary. Let ε > 0 be a number less then (z2 − z)/√n, and let ϕ be the normal state that attains its norm at nζn up to ε√n, i.e. ϕ(hz, zi) > z2 − ε√n. Consider the hz, zi = ζ∗ seminorm 1 ζ1 + ··· + ζ∗ pϕ,z(x) = qϕ(ζ∗ z2 − 2ε√n < ϕ(ζ∗ (3.5) However, ϕ(hz, xi) is a semi inner product and it satisfy Cauchy Schwartz inequality (3.6) nξn) = ϕ(hz, xi). ϕ(hz, xi)2 ≤ ϕ(hz, zi)ϕ(hx, xi) ≤ z2x2. 1 ξ1 + ··· + ζ∗ 1 ξ1)2 + ··· + ϕ(ζ∗ nξn)2, x = (ξ1, ξ2, . . . , ξn). We claim that the open set does not intersect the unit ball B. G = {x pϕ,z(x − z) < ε} Indeed, let x ∈ G. Then by classic Cauchy-Schwartz inequality we have ε2 > pϕ,z(x − z)2 = ϕ(ζ∗ 1 ξ1) − ϕ(ζ∗ 1 nϕ(ζ∗ 1 ξ1) + ··· + ϕ(ζ∗ ε√n > ϕ(ζ∗ 1 ξ1 + ··· + ζ∗ ≥ z2 − ε√n − ϕ(ζ∗ nζn)2 ≥ nξn) − ϕ(ζ∗ nξn) − ϕ(ζ∗ 1 ζ1)2 + ··· + ϕ(ζ∗ 1 ζ1) − ··· − ϕ(ζ∗ nζn)2, nξn) − z2 + ε√n ≥ 1 ξ1 + ··· + ζ∗ nξn), ≥ or i.e. 6 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ From (3.5) and (3.6) we obtain and taking into account how ε is chosen, we have zx > z2 − 2ε√n, (z2 − 2ε√n) > 1. x > 1 z Therefore, x /∈ B, implying B is a closed set. The proof is complete. Proposition 3.5. The unit ball in l2(A) is not totally bounded in τ . Proof. Let ej = (0, . . . , 0, 1, 0, . . . ), where 1 the unit of the algebra A stands at the j-th entry. Let ϕ be an arbitrary normal state and consider a seminorm p = pϕ,1,1,... given by p(x)2 = P+∞ Indeed p(ei − ej)2 = ϕ(1)2 + ϕ(−1)2 = 2, i.e. p(ei − ej) = √2. Hence, the set {ej j ≥ 1} is not totally bounded in p and also in τ . The same is valid for a larger set - the unit ball. (cid:3) We claim that the sequence ej is totally discrete in p. j=1 ϕ(ξj )2. (cid:3) 4. "Compact" operators Let y, z ∈ l2(A). The operator l2(A) → l2(A), x 7→ z hy, xi is adjointable (its adjoint is x 7→ y hz, xi) and bounded. The closed linear hull of such operators is called the algebra of "compact" operators. We say that the operator T ∈ Ba(l2(A)) is compact if its image of any (norm) bounded set is a totally bounded set in topology τ described in the previous section. For the operator T ∈ Ba(l2(A)) it is enough to maps the unit ball into a totally bounded set to be a compact operator. Remark 4.1. Totally bounded and relatively compact sets differ in general case (whenever the unit ball is not complete). Also, throughout the literature, there is a certain ambiguity between terms completely continuous, compact and precompact operators. Although it seems that terms completely continuous and precompact are more accurate, we found that compact is more convenient for our purpose. Before we prove that any "compact" operator is compact, we need a few lemmata. Lemma 4.1. For S, T ⊆ l2(A) and a seminorm p denote dp(S, T ) = sup x∈S inf y∈T p(x − y) (and note that dp is not symmetric). Let S ⊆ l2(A). If for all seminorms p of the form (3.1) and all ε > 0 there is a totally bounded set Sp,ε such that (4.1) d(S, Sp,ε) < ε. then S is also totally bounded. Proof. Denote The condition (4.1) gives Bp(x; ε) = {y ∈ l2(A) p(x − y) < ε}. (4.2) for all ε > 0. S ⊆ [ x∈Sp,ε Bp(x; ε/2), COMPACT AND "COMPACT" OPERATORS 7 Let ε > 0 be arbitrary. The set Sp,ε/2 is totally bounded in p and hence there is a finite set {c1, . . . , cm} such that the union of balls Bp(cj; ε/2) covers Sε/2. By (4.2) the union of balls Bp(cj ; ε) covers S. Lemma 4.2. Let Tα : l2(A) → l2(A) be a net of compact operators such that Tαx → T x in τ uniformly with respect to x < 1. Then T is also compact. Proof. For any ε > 0 and any seminorm p of the form (3.1) there is α such that supx<1 p(T x − Tαx) < ε. Therefore, (cid:3) dp(T (B·(0; 1)), Tα(B·(0; 1))) ≤ ε and the conclusion follows from the previous Lemma. Corollary 4.3. Let S ⊆ l2(A) be a set such that for all ε > 0 there is a totally bounded (in τ ) set Sε such that (cid:3) (4.3) d(S, Sε) = sup x∈S inf y∈Sε x − y < ε. Then S is also totally bounded in τ . Also, let Tn : l2(A) → l2(A) be a sequence of compact operators that converges to T in the operator norm. Then T is also compact. Proof. Both conclusions follows from the fact that τ is coarser then the norm topol- ogy. (cid:3) Lemma 4.4. Let T1 and T2 be compact operators, and let u1, u2 ∈ A. Then T1u1 + T2u2 is also compact. Proof. Let ε > 0 be arbitrary. Since T1 and T2 are compact there is a finite ε/2u1 net for T1(B·(0; 1)), say c1, c2, . . . , cn, and a finite ε/2u2 net for T2(B·(0; 1)), say d1, . . . , dm. Then the set {ciu1 + dju2 1 ≤ i ≤ i, 1 ≤ j ≤ m} is a finite ε net for (T1u1 + T2u2)(B·(0; 1)). Indeed, if x ∈ B·(0; 1), then there is i and j such that T1x − ci < ε/2u1 and T2x − dj < ε/2u2. Hence (T1xu1 + T2xu2) − (ciu1 + dju2) ≤ T1x − ciu1 + T2x − dju2 < ε. (cid:3) Theorem 4.5. Let T : l2(A) → l2(A) be a "compact" operator. Then T is compact. Proof. In view od Lemmata 4.2 and 4.4, it is enough to prove that operators of the form x 7→ Θy,z(x) = z hy, xi are compact. In the special case, where z = ejζ, ζ ∈ A it immediately follows from Proposition 3.4. Indeed, then Θy,ej ζ(B·(0; 1)) is contained in the ball of radius T in A1 which is totally bounded. j=1 ejζj where the series In general case, let z = (ζ1, ζ2, . . . ). Then z = P+∞ converges in the norm. Since Θy,z − Θy,z ′ ≤ yz − z′, we have n Θy,z = lim n→+∞ Θy,ejζj X j=1 and the required follows from the special case and Lemmata 4.2 and 4.4. (cid:3) The converse is true in the special case where A = B(H) is the full algebra of all bounded linear operators on a Hilbert space H. Before we prove such result we need a technical Lemma. 8 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ j ajuj) > δ. Lemma 4.6. Let A = B(H) and let aj ∈ A, j ≥ 1 be positive elements with aj > δ. There is a normal state ϕ and unitary elements uj, vj ∈ A such that ϕ(v∗ Remark 4.2. Actually, we can choose ϕ to be a vector state, and also we can choose uj = vj. Proof. Let ψ ∈ H be a unit vector, and let ϕ be the corresponding vector state, i.e. ϕ(a) = haψ, ψi. For all aj let hj be a unit vector such that hajhj, hji > δ. As it is easy to see, there is a unitary uj such that ujψ = hj. Thus, we have ϕ(u∗ (cid:3) j ajuj) = (cid:10)u∗ j ajujψ, ψ(cid:11) = hajhj, hji > δ. Theorem 4.7. Let A = B(H) and let T : l2(A) → l2(A) be a compact operator. Then T is "compact". Proof. Let Pk denote the projection to the first k coordinates, i.e. Pk(ξ1, ξ2, . . . ) = (ξ1, . . . , ξk, 0, 0, . . . ). It is well known that all Pk are "compact". Suppose T is not "compact". Then δ = inf n≥1(I − Pk)T > 0. Indeed, otherwise either for some k we have (I − Pk)T = 0 and hence T = PkT is "compact", or there is a sequence of positive integers kn such that T − Pkn T → 0 from which it follows that T is "compact". To simplify the calculations assume T = 1. Then immediately, δ ≤ 1. Define the sequence of projections Qn ∈ {P1, P2, . . .} and the sequences of vec- tors xn, yn and zn ∈ l2(A) in the following way. Let Q0 = 0. If Qn−1 is already defined, there is xn ∈ l2(A) such that xn = 1 and (I − Qn−1)T xn > δ/2. Denote yn = T xn. Then, by I − Qn−1 = 1, yn ≥ (I − Qn−1)yn > δ 2 . Since limk→+∞ (I − Pk)(I − Qn−1)yn = 0, there is a positive integer kn such that (I − Pkn )(I − Qn−1)yn < δ2/8 ≤ δ/8. Define Qn = Pkn and (4.4) zn = Qn(I − Qn−1)yn. The sequences yn and zn have the following properties: Firstly, by definition, there hold the inequalities (4.5) (4.6) (4.7) (I − Qn)(I − Qn−1)yn < δ2 8 ≤ zn ≤ yn ≤ Txn = 1, δ 8 , zn ≥ (I − Qn−1)yn − (I − Qn)(I − Qn−1)yn > δ 2 − δ 8 = 3δ 8 . Secondly, (4.8) Indeed, since zn = Qn(I − Qn−1)yn, we have hzn, yni = hQn(I − Qn−1)yn, yni = hzn, yni = hzn, zni . = hQn(I − Qn−1)yn, (I − Qn−1)Qnyni = hzn, zni . COMPACT AND "COMPACT" OPERATORS 9 Thirdly, for m > n we have (4.9) hzm, yni < δ 8 . Indeed, for such m and n we have Qn−1 ≤ Qn ≤ Qm−1, i.e. I − Qm−1 ≤ I − Qn ≤ I − Qn−1, implying I − Qm−1 = (I − Qm−1)(I − Qn)(I − Qn−1), and thus hzm, yni = h(I − Qm−1)zm, yni = = hzm, (I − Qm−1)(I − Qn)(I − Qn−1)yni = = hzm, (I − Qn)(I − Qn−1)yni . Therefore, by (4.5) and (4.7) hzm, yni ≤ zm(I − Qn)(I − Qn−1)yn ≤ δ2 8 . Let us construct a seminorm p, continuous in τ , and a totally discrete sequence n hzn, zni νn > (3δ/8)2, we can from T (B·(0; 1)). Since by (4.7) zn2 = υ∗ choose ϕ and υj, νj ∈ A according to Lemma 4.6, such that (4.10) ϕ(υ∗ n hzn, zni νn) > 9δ2 64 . Consider the seminorm p given by p(x) = vuut +∞ X j=1 ϕ(hzjυj, xi)2 By (4.4) there is a sequence ζj ∈ A such that zk = (0, . . . , 0, ζkn−1+1, . . . , ζkn , 0, . . . ). Define ωn = ζnυn/ϕ(υ∗ (ξ1, ξ2, . . . ) we have nζ∗ nζnυn)1/2. Obviously ϕ(ω∗ nωn) = 1. Also, for x = ϕ(υ∗ nζ∗ j ζj υn)1/2ϕ(ω∗ 2 ≤ j ξj)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ϕ(hznυn, xi)2 = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) kn X j=kn−1+1 kn kn ≤ X j=kn−1+1 ϕ(υnζ∗ j ζjυn) X j=kn−1+1 ϕ(ω∗ j ξj)2 = = ϕ(υ∗ n hzn, zni υn) kn X j=kn−1+1 ϕ(ω∗ j ξj)2 Including (4.6) we obtain ϕ(υ∗ hence n hzn, zni υn) ≤ υ∗ n hzn, zni υn = zn2 ≤ 1 and p(x)2 = +∞ X n=1 ϕ(hzn, xi)2 ≤ +∞ X j=1 ϕ(ω∗ j ξj)2 = pϕ,ω1,...,ωn,...(x)2. Thus, we conclude that p is well defined and also that it is continuous with respect to τ . 10 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Also, xnνn = xn, i.e. ynνn = T xnνn ∈ T (B(0; 1)). Finally we shall prove that ynνn is a totally discrete sequence. Indeed, for m > n we have p(ymνm − ynνn) ≥ ϕ(hzmυm, ymνm − ynνni) ≥ m hzm, ymi νm) − ϕ(υ∗ ≥ ϕ(υ∗ m hzm, yni νn). However, by (4.8) and (4.10), ϕ(υm hzm, zmi νm) > 9δ2 64 and, by (4.9) Therefore ϕ(υ∗ m hzm, yni νn) ≤ hzm, yni < δ2 64 p(ymνm − ynνn) > 9δ2 64 − δ2 8 = δ2 8 . . (cid:3) 5. An example and a comment The proof of Theorem 4.7 depends on Lemma 4.6. Hence it is valid for all unital W ∗-algebras that satisfy the mentioned Lemma. We do not know how to describe such algebras, but it should be mentioned that Lemma 4.6 does not hold for infinite dimensional commutative W ∗-algebras. Example 5.1. In any infinite dimensional commutative W ∗ algebra A, there is a sequence pj of nontrivial mutually orthogonal projections. Since Pn j=1 pj is an increasing sequence, p = P+∞ j=1 pj ∈ A. Therefore, for an arbitrary normal state ϕ the series P+∞ j=1 ϕ(pj) is convergent. The algebra is commutative, and for all unitary υj, νj we have ϕ(υj pjνj) = ϕ(pj υjνj) ≤ ϕ(pj)1/2ϕ(ν∗ j υ∗ j υjνj)1/2 → 0. Thus, Lemma 4.6 is not valid for commutative W ∗ algebras. Moreover, we can use this sequence of projection to construct an operator which is compact but is not "compact". Indeed, let T : l2(A) → l2(A) be the operator defined by (5.1) T x = T (ξ1, ξ2, . . . ) = (p1ξ1, p2ξ2, . . . ). Then, T is not "compact". Indeed, if it is "compact", for all ε > 0 there is an operator S of the form S = Pn j=1 λjΘyj ,zj such that T − S < ε/3. Since Pkzj − zj → 0, as k → +∞ for all 1 ≤ j ≤ n implies PkS−S → 0, there is k large enough such that PkS− S < ε/3 and then T − PkT ≤ T − S +S− PkS +Pk(S − T ) < ε. However, as it is easy to see T − PkT ≥ T ek− PkT ek = 1− pk = 1. On the other hand, for an arbitrary semi norm of the form (3.1) we have p((T − PkT )x) → 0, uniformly with respect to x ∈ B·(0, 1). Indeed, A is commutative and therefore ξ∗ j ηj ) ≤ 1, by x < 1 and (3.2). Thus, we have j ηj) ≤ ξj2 supj ϕ(η∗ j ηj, and further ϕ(ξ∗ j ηj ≤ ξj2η∗ j ξjη∗ j ξj η∗ +∞ p((T − PkT )x)2 = X j=k+1 ϕ(η∗ j pjξj)2 ≤ X j>k Hence, T is compact by Lemma 4.2 ϕ(pj)ϕ(ξ∗ j ξjη∗ j ηj) ≤ X j>k ϕ(pj) → 0. COMPACT AND "COMPACT" OPERATORS 11 Remark 5.1. Topology τ defined in this note highly depends on coordinates, and therefore it is inappropriate for Hilbert modules other then l2(A). One might try to define a topology by semi norms (5.2) +∞ pϕ,zj (x) = vuut X j=1 ϕ(hzj, xi)2, where ϕ is a normal state, and zj is an orthogonal sequence, that satisfies supj≥1 ϕ(hzj, zji) = 1. These semi norms are generalization of those given by (3.1). Indeed, semi norm (5.2) become semi norms (3.1) by choosing zj = ejηj. However, such new topology is in the case of l2(A) larger then τ , even if we suppose that zj is even more orthonormal. Namely, if A = B(H), H infinite dimensional, there is a Cuntz ∞-tuple, i.e. a sequence of isometries vj satisfying j vj = 1 and P+∞ v∗ j = 1. Then, it is easy to see that xj = (vj , 0, 0, . . . ) is orthonormal. But, in the semi norm pϕ,xj of the form (5.2) the sequence xj itself is totally discrete. j=1 vjv∗ Acknowledgement. The authors was supported in part by the Ministry of edu- cation and science, Republic of Serbia, Grant #174034. References [1] N. Bourbaki, General Topology, chapters 1 -- 4, Springer, 1987, ISBN-13: 978-3-540-64241-1 [2] M. Frank, Self-duality and C ∗-reflexivity of Hilbert C ∗-modules, Z Anal. Anwendungen, 9 (1990), 165 -- 176 [3] J. L. Kelley General Topology D. Van Nostrand Company Inc. 1955 [4] E. C. Lance, Hilbert C ∗-Modules: A toolkit for operator algebraists, Cambridge University Press, 1995 [5] V. M. Manuilov, Diagonalization of compact operators in Hilbert modules over finite W ∗- algebras, Ann. Global Anal. Geom., 13-3 (1995), 207 -- 226. [6] V. M. Manuilov, Diagonalization of compact operators on Hilbert modules over C ∗-algebras of real rank zero, Math. Notes, 62-6 (1997), 726 -- 730. [7] V. M. Manuilov, E. V. Troitsky, Hilbert C ∗-Modules, American Mathematical Society, 2005 [8] W. L. Paschke, Inner product modules over B∗-algebras, Trans. Amer. Math. Soc., 182 (1973), 443 -- 468 [9] W. L. Paschke, Inner product modules arising from compact automorphism groups of von Neumann algebras, Trans. Amer. Math. Soc., 224 (1976), 87 -- 102 University of Belgrade, Faculty of Mathematics, Studentski trg 16-18, 11000 Beograd, Serbia E-mail address: [email protected] University of Belgrade, Faculty of Mathematics, Studentski trg 16-18, 11000 Beograd, Serbia E-mail address: [email protected]
1602.03709	2	1602	2016-02-12T16:28:15	Invariance of the Cuntz splice	[ "math.OA" ]	We show that the Cuntz splice induces stably isomorphic graph $C^*$-algebras.	math.OA	math	1 1(cid:19) and  (cid:18)1 1 1 1 0 0 1 0 1 1 1 1 0 1 0 0 1 1  , INVARIANCE OF THE CUNTZ SPLICE SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN Abstract. We show that the Cuntz splice induces stably isomorphic graph C ∗-algebras. 1. Introduction Cuntz and Krieger introduced the Cuntz-Krieger algebras in [CK80], and Cuntz showed in [Cun81] that if we restrict to the matrices satisfying the modest condition (II), then the stabilized Cuntz-Krieger algebras are an invariant of shifts of finite type up to flow equivalence. Shortly after Franks had made a successful classifi- cation of irreducible shifts of finite type up to flow equivalence ([Fra84]), Cuntz raised the question of whether this invariant or the K0-group alone classifies simple Cuntz-Krieger algebras up to stable isomorphism. He sketched in [Cun86] that it was enough to answer whether O2 and O2− are isomorphic, where O2 and O2− are the Cuntz-Krieger algebras associated with the matrices respectively. This question remained open until Rørdam in [Rør95] showed that O2 and O2− are in fact isomorphic and elaborated on the arguments of Cuntz to show that the K0-group is a complete invariant of the stabilized simple Cuntz-Krieger algebras. This procedure of gluing the graph corresponding to the former matrix above onto another graph has since been known as Cuntz splicing a graph at a certain vertex. Knowing that when we Cuntz splice a graph (on a vertex that supports two return paths), we get stably isomorphic Cuntz-Krieger algebras, has been important for classifying Cuntz-Krieger algebras ([Rør95, Hua95, Res06]), as well as under- standing the connection between the dynamics of the underlying shift spaces and the Cuntz-Krieger algebras. With the recent work on the relation between move equivalence of graphs and stable isomorphism of the corresponding graph C∗-alge- bras, the question of whether Cuntz splicing yields stably isomorphic C∗-algebras has become of great interest. Bentmann has shown that this is in fact the case for purely infinite graph C∗-algebras with finitely many ideals ([Ben15]), while Gabe recently has generalized this to also cover general purely infinite graph C∗-algebras ([Gab16]). Their methods depend heavily on the result of Kirchberg on lifting in- vertible ideal-related KK-elements to equivariant isomorphisms for strongly purely infinite C∗-algebras ([Kir00]). In this paper we show in general that Cuntz splicing a vertex that supports two distinct return paths yields stably isomorphic graph C∗-algebras -- only assuming that the graph is countable. The results, the proofs and methods of this paper are Date: September 4, 2018. 2010 Mathematics Subject Classification. 46L35, 46L80 (46L55, 37B10). Key words and phrases. Graph C ∗-algebras, Cuntz splice. 1 2 SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN important for recent development in the geometric classification of general Cuntz- Krieger algebras and of unital graph C∗-algebras ([ERRS16a], [ERRS16b]) as well as for the question of strong classification of general Cuntz-Krieger algebras and of unital graph C∗-algebras ([CRR16], [ERRS16b]). We proved invariance of the Cuntz splice in the special case of unital graph C∗-algebras in an arXiv preprint (1505.06773) posted in May 2015. Bentmann's recent paper showed us how to reduce the general question to the row-finite case, and we proceeded to discover that our arguments applied with only minor changes to that case. Since most of the results of our preprint have since been superseded by other forthcoming work, we do not intend to publish it, whereas this work is intended for publication. 2. Preliminaries Definition 2.1. A graph E is a quadruple E = (E0, E1, r, s) where E0 and E1 are sets, and r and s are maps from E1 to E0. The elements of E0 are called vertices, the elements of E1 are called edges, the map r is called the range map, and the map s is called the source map. When working with several graphs at the same time, to avoid confusion, we will denote the range map and source map of a graph E by rE and sE respectively. All graphs considered will be countable, i.e., there are countably many vertices and edges. Definition 2.2. A loop is an edge with the same range and source. A path µ in a graph is a finite sequence µ = e1e2 · · · en of edges satisfying r(ei) = s(ei+1), for all i = 1, 2, . . . , n − 1, and we say that the length of µ is n. We extend the range and source maps to paths by letting s(µ) = s(e1) and r(µ) = r(en). Vertices in E are regarded as paths of length 0 (also called empty paths). A cycle is a nonempty path µ such that s(µ) = r(µ). We call a cycle e1e2 · · · en a vertex-simple cycle if r(ei) 6= r(ej ) for all i 6= j. A vertex-simple cycle e1e2 · · · en is said to have an exit if there exists an edge f such that s(f ) = s(ek) for some k = 1, 2, . . . , n with ek 6= f . A return path is a cycle µ = e1e2 · · · en such that r(ei) 6= r(µ) for i < n. For a loop, cycle or return path, we say that it is based at the source vertex of its path. We also say that a vertex supports a certain loop, cycle or return path if it is based at that vertex. Note that in [BHRS02, Szy02], the authors use the term loop where we use cycle. Definition 2.3. A vertex v in E is called regular if s−1(v) is finite and nonempty. We denote the set of regular vertices by E0 reg. A vertex v in E is called a sink if s−1(v) = ∅. A graph E is called row-finite if for each v ∈ E0, v is either a sink or a regular vertex. It is essential for our approach to graph C∗-algebras to be able to shift between a graph and its adjacency matrix. In what follows, we let N denote the set of positive integers, while N0 denotes the set of nonnegative integers. Definition 2.4. Let E = (E0, E1, r, s) be a graph. We define its adjacency matrix AE as a E0 × E0 matrix with the (u, v)'th entry being (cid:12)(cid:12)(cid:8)e ∈ E1(cid:12)(cid:12) s(e) = u, r(e) = v(cid:9)(cid:12)(cid:12) . As we only consider countable graphs, AE will be a finite matrix or a countably infinite matrix, and it will have entries from N0 ⊔ {∞}. Let X be a set. If A is an X × X matrix with entries from N0 ⊔ {∞}, we let EA be the graph with vertex set X such that between two vertices x, x′ ∈ X we have A(x, x′) edges. INVARIANCE OF THE CUNTZ SPLICE 3 It will be convenient for us to alter the adjacency matrix of a graph in a very specific way, subtracting the identity, so we introduce notation for this. Notation 2.5. Let E be a graph and AE its adjacency matrix. Let BE denote the matrix AE − I. 2.1. Graph C∗-algebras. We follow the notation and definition for graph C∗-al- gebras in [FLR00]; this is not the convention used in Raeburn's monograph [Rae05]. Definition 2.6. Let E = (E0, E1, r, s) be a graph. The graph C∗-algebra C∗(E) is defined as the universal C∗-algebra generated by a set of mutually orthogonal projections (cid:8)pv(cid:12)(cid:12) v ∈ E0(cid:9) and a set (cid:8)se(cid:12)(cid:12) e ∈ E1(cid:9) of partial isometries satisfying the relations • s∗ • s∗ • ses∗ esf = 0 if e, f ∈ E1 and e 6= f , ese = pr(e) for all e ∈ E1, e ≤ ps(e) for all e ∈ E1, and, • pv =Pe∈s−1(v) ses∗ e for all v ∈ E0 with 0 < s−1(v) < ∞. Whenever we have a set of mutually orthogonal projections (cid:8)pv(cid:12)(cid:12) v ∈ E0(cid:9) and a set(cid:8)se(cid:12)(cid:12) e ∈ E1(cid:9) of partial isometries in a C∗-algebra satisfying the relations, then we call these elements a Cuntz-Krieger E-family. We will also need moves on graphs as defined in [Sør13]. In the case of graphs with finitely many vertices the basic moves are outsplitting (Move (O)), insplitting (Move (I)), reduction (Move (R)), and removal of a regular source (Move (S)). It turns out that in the general setting, move (R) must be replaced by the following Definition 2.7 (Collapse a regular vertex that does not support a loop, Move (Col)). Let E = (E0, E1, r, s) be a graph and let v be a regular vertex in E that does not support a loop. Define a graph ECOL by COL = E0 \ {v}, E0 E1 COL = E1 \ (r−1(v) ∪ s−1(v)) ⊔(cid:8)[ef ](cid:12)(cid:12) e ∈ r−1(v) and f ∈ s−1(v)(cid:9) , the range and source maps extends those of E, and satisfy rECOL ([ef ]) = r(f ) and sECOL ([ef ]) = s(e). Move (Col) was defined in [Sør13, Theorem 5.2] for graphs with finitely many vertices as an auxiliary move, and proved there to be realized by moves (I), (O) and (R). Definition 2.8. The equivalence relation generated by the moves (O), (I), (S), (Col) together with graph isomorphism is called move equivalence, and denoted ∼ME. Let X be a set and let A and A′ be X × X matrices with entries from N0 ⊔ {∞}. If EA ∼ME EA′, then we say that A and A′ are move equivalent, and we write A ∼ME A′. Remark 2.9. By [Sør13, Theorem 5.2], the above definition is equivalent to the definition in [Sør13, Section 4] for graphs with finitely many vertices. These moves have been considered by other authors, and were previously noted to preserve the Morita equivalence class of the associated graph C∗-algebra. The moves (O) and (I) induce stably isomorphic C∗-algebras due to the results in [BP04], and by [CG06], moves (R), (S), (Col) preserve the Morita equivalence class of the associated graph C∗-algebras (see also [Sør13, Propositions 3.1, 3.2 and 3.3 and Theorem 3.5]). Therefore, we get the following theorem. 4 SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN Theorem 2.10. Let E1 and E2 be graphs such that E1 ∼ME E2. Then C∗(E1) ⊗ K ∼= C∗(E2) ⊗ K. We now recall the definiton of the Cuntz splice (see Notation 4.1 and Example 4.2 for illustrations). Definition 2.11 (Move (C): Cuntz splicing at a regular vertex supporting two return paths). Let E = (E0, E1, r, s) be a graph and let v ∈ E0 be a regu- lar vertex that supports at least two return paths. Let Ev,− denote the graph (E0 v,−, rv,−, sv,−) defined by v,−, E1 E0 E1 v,− := E0 ⊔ {u1, u2} v,− := E1 ⊔ {e1, e2, f1, f2, h1, h2}, where rv,− and sv,− extend r and s, respectively, and satisfy sv,−(e1) = v, sv,−(e2) = u1, sv,−(fi) = u1, sv,−(hi) = u2, and rv,−(e1) = u1, rv,−(e2) = v, rv,−(fi) = ui, rv,−(hi) = ui. We call Ev,− the graph obtained by Cuntz splicing E at v, and say Ev,− is formed by performing Move (C) to E. The aim of this paper is to prove that C∗(E) ⊗ K ∼= C∗(Ev,−) ⊗ K for any graph E. In fact, we prove slightly more, since our proof allows for Cuntz splicing also at infinite emitters supporting at least two return paths. 3. Elementary matrix operations preserving move equivalence In this section we perform row and column additions on BE without changing the move equivalence class of the associated graphs. Our setup is slightly different from what was considered in [Sør13, Section 7], so we redo the proofs from there in our setting. There are no substantial changes in the proof techniques, which essentially go back to [Fra84]. Lemma 3.1. Let E = (E0, E1, rE , sE) be a graph. Let u, v ∈ E0 be distinct vertices. Suppose the (u, v)'th entry of BE is nonzero ( i.e., there is an edge from u to v), and that the sum of the entries in the u'th row of BE is strictly greater than 0 ( i.e., u emits at least two edges). If B′ is the matrix formed from BE by adding the u'th column into the v'th column, then AE ∼ME B′ + I. Proof. Fix an edge f from u to v. Form a graph G from E by removing f but adding for each edge e ∈ r−1 E (u) an edge ¯e with sG(¯e) = sE(e) and rG(¯e) = v. We claim that B′ = BG. At any entry other than the (u, v)'th entry the two matrices have the same values, since we in both cases add entries into the v'th column that are exactly equal to the number of edges in E. At the (u, v)'th entry of BG we have (s−1 E (u) ∩ r−1 E (v) − 1) + s−1 E (u) ∩ r−1 E (u) = BE(u, v) + BE(u, u) = B′(u, v). Thus to prove this lemma it suffices to show E ∼ME G. Partition s−1 E (u) as E1 = {f } and E2 = s−1 E (u) \ {f }. By assumption E2 is not empty, so we can use Move (O). Doing so yields a graph just as E but where u is replaced by two vertices, u1 and u2. The vertex u1 receives a copy of everything u did and it emits only one edge. That edge has range v. The vertex u2 also receives a copy of everything u did, and it emits everything u did, except f . Since u1 is regular and not the base of a loop, we can collapse it. The resulting graph is G (after we relabel u2 as u), so G ∼ME E. (cid:3) INVARIANCE OF THE CUNTZ SPLICE 5 We can also add columns along a path. Proposition 3.2. Let E = (E0, E1, rE , sE) be a graph and let u, v ∈ E0 be distinct vertices with a path from u to v going through distinct vertices u = u0, u1, u2, . . . , un = v (labelled so there is an edge from ui to ui+1 for i = 0, 1, 2 . . . , n − 1). Suppose further that u supports a loop. If B′ is the matrix formed from BE by adding the u'th column into the v'th column, then AE ∼ME B′ + I. Proof. That u supports a loop guarantees that B′ + I is the adjacency matrix of a graph E′ = EB ′+I . The vertex ui emits exactly one edge in E if and only if it emits exactly one edge in E′, for i = 1, . . . , n − 1. So by collapsing all regular vertices ui, i = 1, 2, . . . , n − 1 emitting exactly one edge both in E and in E′, we get two new graphs E1 ∼ME E 1 ∼ME E′. In E1, there is a path from u to v through vertices that all emit and E′ at least two edges. Moreover, BE ′ 1 is obtained from BE1 by adding the u'th column into the v'th column. Therefore, we may without loss of generality assume that all the vertices ui, i = 0, 1, 2, . . . , n − 1 emit at least two edges. By repeated applications of Lemma 3.1, we first add the un−1'th column into the un'th column of BE, which we can since there is an edge from un−1 to un. Then we add the un−2'th column into the un'th column, which we can since there now is an edge from un−2 to un. Continuing this way, we end up with a matrix C which is formed from BE by adding all the columns ui, for i = 0, 1, 2, . . . , n − 1, into the the un'th column. We have that AE ∼ME C + I. Now consider the matrix B′ = BE ′ . By repeated applications of Lemma 3.1, we first add the un−1'th column into the un'th column of B′ = BE ′ , which we can since there is an edge from un−1 to un. Then we add the un−2'th column into the un'th column, which we can since there now is an edge from un−2 to un. Continuing this way, we end up with a matrix D which is formed from B′ = BE ′ by adding all the columns ui, for i = 1, 2, . . . , n − 1, into the the un'th column. We have that B′ + I = AE ′ ∼ME D + I. But it is clear from the construction that C = D. (cid:3) Remark 3.3. Similar to how we used Lemma 3.1 in the above proof, we can use Proposition 3.2 "backwards" to subtract columns in BE as long as the addition that undoes the subtraction would be legal. We now turn to row additions. Lemma 3.4. Let E = (E0, E1, rE , sE) be a graph. Let u, v ∈ E0 be distinct vertices. Suppose the (v, u)'th entry of BE is nonzero ( i.e., there is an edge from v to u), and that u is a regular vertex. If B′ is the matrix formed from BE by adding the u'th row into the v'th row, then AE ∼ME B′ + I. Proof. Let E′ = EB ′+I denote the graph with adjacency matrix B′ + I. First assume that u only receives one edge in E (which necessarily is the edge from v). Then u is a regular vertex not supporting a loop, so we can collapse it obtaining a graph E′′. Note that the vertex u is a regular source in E′, so we may remove it. It is clear that the resulting graph is exactly E′′. Now assume instead that u receives at least two edges. Fix an edge f from v to u. Form a graph G from E by removing f but adding for each edge e ∈ s−1 E (u) an edge ¯e with sG(¯e) = v and rG(¯e) = rE(e). We claim that E ∼ME G. Arguing as in the proof of Lemma 3.1 we see that this is equivalent to proving AE ∼ME B′ + I. 6 SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN Partition r−1 E (u) as E1 = {f } and E2 = r−1 E (u) \ {f }. By our assumptions on u, E2 is nonempty, and u is regular, so we can use Move (I). Doing so replaces u with two new vertices, u1 and u2. The vertex u1 only receives one edge, and that edge comes from v, the vertex u2 receives the edges u received except f . Since u1 is regular and not the base of a loop we can collapse it. The resulting graph is G (after we relabel u2 as u), so G ∼ME E. (cid:3) We can also add rows along a path of vertices. Proposition 3.5. Let E = (E0, E1, rE , sE) be a graph and let u, v ∈ E0 be distinct vertices with a path from v to u going through distinct vertices v = v0, v1, v2, . . . , vn = u (labelled so there is an edge from vi to vi+1 for i = 0, 1, 2, . . . , n − 1). Suppose further that the vertex u is regular and supports at least one loop. If B′ is the matrix formed from BE by adding the u'th row into the v'th row, then AE ∼ME B′ + I. Proof. That u supports a loop guarantees that B′ + I is the adjacency matrix of a graph E′ = EB ′+I . First we prove the special case where all the vertices v1, . . . , vn are regular. By repeated applications of Lemma 3.4, we first add the v1'st row into the v0'th row of BE, which we can since there is an edge from v0 to v1 and v1 is regular. Then we add the v2'nd row into the v0'th row, which we can since there now is an edge from v0 to v2 and v2 is regular. Continuing this way, we end up with a matrix C which is formed from BE by adding all the rows vi, for i = 1, 2, . . . , n, into the the v0'th column. We have that AE ∼ME C + I. Now consider the matrix B′ = BE ′ . By repeated applications of Lemma 3.4, we first add the v1'st row into the v0'th row of B′ = BE ′ , which we can since there is an edge from v0 to v1. Then we add the v2'nd row into the v0'th row, which we can since there now is an edge from v0 to v2. Continuing this way, we end up with a matrix D which is formed from B′ = BE ′ by adding all the rows vi, for i = 1, 2, . . . , n − 1, into the the v0'th row. We have that B′ + I = AE ′ ∼ME D + I. But it is clear from the construction that C = D. i and v2 i = s−1 i = {ei} and E 2 Now we prove that the general case when only u is assumed to be regular can be reduced to the case where v1, . . . , vn are regular. Choose a path e0e1 · · · en−1 going through the distinct vertices v1, . . . , vn. For each singular vertex vi, i = 1, . . . , n−1, we outsplit according to the partition E 1 E (vi) and call the corresponding vertices v1 i , respectively. Denote the split graph by E1, and denote the vertices vi, i = 1, . . . , n − 1 that were not split by v1 i . Note that we now have a path from v to u through distinct regular vertices. Note also that since all vertices along the path are distinct, what happens to the vi'th entry of row u and v is that it gets doubled for each vertex ui that gets split and stays unchanged for the vertices ui = u1 i ∈ E0 that are regular. Let E′ be the graph EB ′+I , and let E′ 1 be the graph constructed using exactly the same outsplittings as in the graph above. Now it is clear that the graph we get from E1 by adding row u into row v is exactly E′ (cid:3) 1. Thus the general case now follows from the above. Remark 3.6. We can also use Proposition 3.5 "backwards" to subtract rows in BE ( cf. Remark 3.3). 4. Cuntz splice implies stable isomorphism In this section, we show that the Cuntz splice gives stably isomorphic graph C∗-algebras. We first set up some notation. INVARIANCE OF THE CUNTZ SPLICE 7 Notation 4.1. Let E∗ and E∗∗ denote the graphs: e1 •v1 e4 * •v2 e2 e3 E∗ = f10 f7 f1 f4 E∗∗ = •w4 f9 f8 * •w3 f6 f5 •w1 f2 f3 •w2 The graph E∗ is what we attach when we Cuntz splice. If we instead attach the graph E∗∗, we have Cuntz spliced twice. Let E = (E0, E1, rE , sE) be a graph and let u be a vertex of E. Then Eu,− can be described as follows (up to canonical isomorphism): u,− = E0 ⊔ E0 ∗ u,− = E1 ⊔ E1 ∗ ⊔ {d1, d2} with rEu,− E 1 = rE, sEu,−E 1 = sE, rEu,− E1 ∗ = rE∗ , sEu,−E1 E0 E1 ∗ = sE∗ , and sEu,− (d1) = u sEu,− (d2) = v1 rEu,− (d1) = v1 rEu,− (d2) = u. Moreover, Eu,−− can be described as follows (up to canonical isomorphism): E0 E1 u,−− = E0 ⊔ E0 ∗∗ u,−− = E1 ⊔ E1 ∗∗ ⊔ {d1, d2} with rEu,−− E 1 = rE, sEu,−−E 1 = sE, rEu,−− E1 ∗∗ = rE∗∗ , sEu,−−E1 ∗∗ = sE∗∗, and sEu,−−(d1) = u sEu,−−(d2) = w1 rEu,−− (d1) = w1 rEu,−− (d2) = u. Example 4.2. Consider the graph Then and E = •u e1 •v1 e4 * •v2 e2 e3 Eu,− = d1 d2 •u f10 f7 f1 f4 Eu,−− = •w4 f9 f8 * •w3 f6 f5 •w1 f2 f3 •w2 d1 d2 •u The strategy for obtaining the result is as follows. By [Rør95], the graph C∗-al- gebras C∗(E∗) and C∗(E∗∗) are isomorphic. We first show in Proposition 4.3 that C∗(E∗) and C∗(E∗∗) are still isomorphic if we do not enforce the summation re- lation at v1 and w1 respectively, by appealing to general classification results. In * j j * * * j j * * j j j j % % g g * j j H H % % g g * * * j j * * j j j j H H % % g g 8 SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN fact, we need to establish (Lemma 4.4) that they are isomorphic in a way send- ing prescribed elements of the nonstable K-theory to other prescribed elements. Using this, we prove in Theorem 4.5 by use of Kirchberg's Embedding Theorem that Cuntz splicing once and twice yields isomorphic graph C∗-algebras. Finally, we establish in Proposition 4.7 that the graph obtained by Cuntz splicing twice is move equivalent to the original, and the desired conclusion follows. Proposition 4.3. The relative graph C∗-algebras (in the sense of Muhly-Tomforde [MT04]) C∗(E∗, {v2}) and C∗(E∗∗, {w2, w3, w4}) are isomorphic. Proof. Following [MT04, Definition 3.6] we define a graph (E∗){v2} = e4 * •v2 e2 e3 e′ 4 e1 •v1 e′ 1 •v′ 1 Then by [MT04, Theorem 3.7] we have that C∗(E∗, {v2}) ∼= C∗((E∗){v2}). Similarly we define a graph f10 f7 f1 f4 (E∗∗){w2,w3,w4} = •w4 f9 f8 * •w3 f6 f5 f ′ 6 f2 f3 •w2 f ′ 3 •w1 f ′ 1 •w′ 1 Using [MT04, Theorem 3.7] again, we have that C∗(E∗∗, {w2, w3, w4}) is isomorphic to C∗((E∗∗){w2,w3,w4}). Both the graphs (E∗){v2} and (E∗∗){w2,w3,w4} satisfy Condition (K). Using the well developed theory of ideal structure and K-theory for graph C∗-algebras, we see that both have exactly one nontrivial ideal, that this ideal is the compact operators, and that their six-term exact sequences are Zhv′ 1i 0 / Z 0 / 0 0 Zhw′ 1i 0 / Z 0 / 0 0 Furthermore, in K0(C∗((E∗){v2})) we have [pv1 ] = −[pv′ 1] = [pv2], and in K0(C∗((E∗∗){w2,w3,w4})) we have [pw1 ] = −[pw′ [pw3 ] = 0 = [pw4 ]. 1] = [pw2], Therefore the class of the unit is −[pv′ 1], respectively. It now follows from [BD96, Theorem 2] (see also [ERR13, Corollary 4.20]) that C∗((E∗){v2}) ∼= C∗((E∗∗){w2,w3,w4}) and hence that C∗(E∗, {v2}) ∼= C∗(E∗∗, {w2, w3, w4}). 1] and −[pw′ (cid:3) We also need a technical result about the projections in E = C∗(E∗, {v2}). Lemma 4.4. Let E = C∗(E∗, {v2}) and choose an isomorphism between E and C∗(E∗∗, {w2, w3, w4}) according to the previous proposition. Let pv1, pv2 , se1 , se2 , se3 , se4 be the canonical generators of C∗(E∗, {v2}) = E and let pw1, pw2, pw3, pw4, * j j w w * ' ' * * j j * * j j w w j j / / O O o o o o / / O O o o o o INVARIANCE OF THE CUNTZ SPLICE 9 sf1 , sf2 , . . . , sf10 denote the image of the canonical generators of C∗(E∗∗, {w2, w3, w4}) in E under the chosen isomorphism. Then e2 ∼ sf1 s∗ f1 + sf2 s∗ se1 s∗ e1 + se2 s∗ e1 + se2 s∗ pv1 −(cid:0)se1 s∗ e2(cid:1) ∼ pw1 −(cid:0)sf1 s∗ 1E − pv1 = pv2 ∼ pw2 + pw3 + pw4 = 1E − pw1 f2 + sf5 s∗ f1 + sf2 s∗ f5 , f2 + sf5 s∗ f5(cid:1) , in E, where ∼ denotes Murray-von Neumann equivalence. Thus there exists a uni- tary z0 in E such that z0(cid:0)se1 s∗ z0(cid:0)pv1 −(cid:0)se1 s∗ e1 + se2 s∗ e1 + se2 s∗ e2(cid:1) z∗ e2(cid:1)(cid:1) z∗ z0pv1 z∗ z0pv2 z∗ 0 = sf1 s∗ f1 + sf2 s∗ 0 = pw1 −(cid:0)sf1 s∗ 0 = pw1 0 = pw2 + pw3 + pw4 . f2 + sf5 s∗ f1 + sf2 s∗ f5 , f2 + sf5 s∗ f5(cid:1) , Proof. By [AMP07, Corollary 7.2], row-finite graph C∗-algebras have stable weak cancellation, so by [MT04, Theorem 3.7], E has stable weak cancellation. Hence any two projections in E are Murray-von Neumann equivalent if they generate the same ideal and have the same K-theory class. As in the proof of Proposition 4.3, we will use [MT04, Theorem 3.7] to realize our relative graph C∗-algebras as graph C∗-algebras of the graphs (E∗){v2} and (E∗∗){w2,w3,w4}. Denote the image of the vertex projections of C∗((E∗){v2}) in- side E under this isomorphism by qv1 , qv2 , qv′ 1 and denote the image of the vertex ∼= projections of (E∗∗){w2,w3,w4} inside E under the isomorphisms (E∗∗){w2,w3,w4} C∗(E∗∗, {w2, w3, w4}) ∼= E by qw1 , qw2 , qw3 , qw4 , qv′ 1. Using the description of the isomorphism in [MT04, Theorem 3.7], we see that we need to show that qv1 ∼ qw1 , qv′ 1 and qv2 ∼ qw2 + qw3 + qw4 . {v2} satisfies Condition (K) and the smallest hereditary and saturated subset containing v1 is all of (E∗)0 {v2} we have that qv1 is a full projection ([BHRS02, Theorem 4.4]). Similarly qw1 , qv2 and qw2 + qw3 + qw4 are full. In K0(E) we have, using our calculations from the proof of Proposition 4.3, that 1 ∼ qw′ Since (E∗)0 [qv1 ] = [1] = [qw1 ], [qv2 ] = [1] = [qw2 ] = [qw2 ] + [qw3 ] + [qw4 ]. So by stable weak cancellation qv1 ∼ qw1 and qv2 ∼ qw2 + qw3 + qw4. Both qv′ 1 and qw′ 1 generate the only nontrivial ideal I of E ([BHRS02, Theo- rem 4.4]). Since that ideal is isomorphic to the compact operators and both [qv′ 1 ] 1 ] are positive generators of K0(I) ∼= K0(K) ∼= Z, they must both represent and [qw′ the same class in K0(I), and thus also in K0(E). Therefore qv′ 1 ∼ qw′ 1 . Let u, v and w be partial isometries realizing the Murray-von Neumann equiva- lences. Then z0 = u + v + w is a unitary that satisfies the required conditions. (cid:3) Theorem 4.5. Let E be a graph and let u be a vertex of E. Then C∗(Eu,−) ∼= C∗(Eu,−−). Proof. As above, we let E denote the C∗-algebra C∗(E∗, {v2}), and we choose an isomorphism between E and C∗(E∗∗, {w2, w3, w4}), which exists according to Proposition 4.3. Since C∗(Eu,−) and E are separable, nuclear C∗-algebras, by the Kirchberg Em- bedding Theorem [KP00], there exist an injective ∗-homomorphism We will suppress this embedding in our notation. C∗(Eu,−) ⊕ E ֒→ O2. 10 SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN u,− and se, e ∈ E1 In O2, we denote the vertex projections and the partial isometries coming from C∗(Eu,−) by pv, v ∈ E0 u,−, respectively, and we denote the ver- tex projections and the partial isometries coming from E = C∗(E∗, {v2}) by p1, p2 and s1, s2, s3, s4, respectively. Since we are dealing with an embedding, it follows from Szymański's Generalized Cuntz-Krieger Uniqueness Theorem ([Szy02, Theo- rem 1.2]) that for any vertex-simple cycle α1α2 · · · αn in Eu,− without any exit, we have that the spectrum of sα1 sα2 · · · sαn contains the entire unit circle. We will define a new Cuntz-Krieger Eu,−-family. We let qv = pv qv1 = p1, qv2 = p2. for each v ∈ E0, Since any two nonzero projections in O2 are Murray-von Neumann equivalent, we can choose partial isometries x1, x2 ∈ O2 such that x1x∗ x2x∗ 1 = sd1 s∗ 2 = p1 − (s1s∗ d1 1 + s2s∗ 2) We let te = se tei = si td1 = x1, td2 = x2. x∗ 1x1 = p1 x∗ 2x2 = pu. for each e ∈ E1, for each i = 1, 2, 3, 4, By construction(cid:8)qv(cid:12)(cid:12) v ∈ E0 u,−(cid:9) is a set of orthogonal projections, and(cid:8)te(cid:12)(cid:12) e ∈ E1 u,−(cid:9) a set of partial isometries. Furthermore, by choice of {te e 6= d1, d2} the relations are clearly satisfied at all vertices other than v1 and u. The choice of x1, x2 ensures that the relations hold at u and v1 as well. Hence {qv, te} does indeed form a Cuntz-Krieger Eu,−-family. Denote this family by S. Using the universal property of graph C∗-algebras, we get a ∗-homomorphism from C∗(Eu,−) onto C∗(S) ⊆ O2. Let α1α2 · · · αn be a vertex-simple cycle in Eu,− without any exit. Since u is where the Cuntz splice is glued on, no vertex-simple cycle without any exit uses edges connected to u, v1 or v2. Hence tα1 tα2 · · · tαn = sα1 sα2 · · · sαn and so its spectrum contains the entire unit circle. It now follows from [Szy02, Theorem 1.2] that the ∗-homomorphism from C∗(Eu,−) to C∗(S) is in fact a ∗-isomorphism. Let A0 be the C∗-algebra generated by(cid:8)pv(cid:12)(cid:12) v ∈ E0(cid:9), and let A be the subalge- bra of O2 generated by(cid:8)pv(cid:12)(cid:12) v ∈ E0(cid:9) and E. Note that A = A0 ⊕ E. Let us denote by(cid:8)rwi , yfj (cid:12)(cid:12) i = 1, 2, 3, 4, j = 1, 2, . . . , 10(cid:9) the image of the canon- so choose a unitary z0 ∈ E according to this lemma, and set z = z0 +Pv∈E 0 pv ∈ ical generators of C∗(E∗∗, {w2, w3, w4}) in O2 under the chosen isomorphism be- tween C∗(E∗∗, {w2, w3, w4}) and E composed with the embedding into O2. M(A). Clearly z is a unitary in M(A). Since the approximate identity of A given by By Lemma 4.4, certain projections in E are Murray-von Neumann equivalent, ( nXk=1 pvk + 1E)n∈N , where (cid:8)pv(cid:12)(cid:12) v ∈ E0(cid:9) = {pv1 , pv2 , . . . }, is an approximate identity of C∗(S), we have a canonical unital ∗-homomorphism from M(A) to M(C∗(S)) which, when restricted to A, gives the embedding of A into C∗(S). So we can consider z as INVARIANCE OF THE CUNTZ SPLICE 11 a unitary in M(C∗(S)). Hence, for all x ∈ C∗(S), we have that zx and xz are elements of C∗(S). By construction of z, we have that z(cid:0)te1 t∗ z(cid:0)qv1 −(cid:0)te1 t∗ e1 + te2 t∗ e1 + te2 t∗ zqv = qvz = qv, for all v ∈ E0, zte = tez = te, for all e ∈ E1, f2 + yf5 y∗ f5 , f1 + yf2 y∗ e2(cid:1) z∗ = yf1 y∗ e2(cid:1)(cid:1) z∗ = rw1 −(cid:0)yf1 y∗ zqv1 z∗ = rw1 , zqv2 z∗ = rw2 + rw3 + rw4 . f1 + yf2 y∗ f2 + yf5 y∗ f5(cid:1) , We will now define a Cuntz-Krieger Eu,−−-family in O2. We let Pv = qv = pv Pwi = rwi for each v ∈ E0, for each i = 1, 2, 3, 4. Moreover, we let Se = te = se Sfi = yfi Sd1 = ztd1 z∗ = zx1z∗, Sd2 = ztd2 z∗ = zx2z∗. for each e ∈ E1, for each i = 1, 2, . . . , 10, Denote this family by T . By construction(cid:8)Pv(cid:12)(cid:12) v ∈ E0 a set of partial isometries satisfying u,−−(cid:9) is a set of orthogonal projections, and(cid:8)Se(cid:12)(cid:12) e ∈ E1 u,−−(cid:9) for all e ∈ E1 and i = 1, 2, . . . , 10. From this, it is clear that T will satisfy the Cuntz-Krieger relations at all vertices in E0. Similarly, we see that since (cid:8)rwi , yfj (cid:12)(cid:12) i = 1, 2, 3, 4, j = 1, 2, . . . , 10(cid:9) is a Cuntz-Krieger (E∗∗, {w2, w3, w4})-family, XsEu,−− (e)=w1 f5(cid:1) T will satisfy the relations at the vertices w2, w3, w4. It only remains to check the summation relation at w1, for that we compute f2 + Sf5 S∗ f2 + yf5 y∗ e = Sf1 S∗ = yf1 y∗ = rw1 = Pw1 . f5 + rw1 −(cid:0)yf1 y∗ f1 + Sf2 S∗ f1 + yf2 y∗ f5 + Sd2 S∗ f1 + yf2 y∗ f2 + yf5 y∗ SeS∗ d2 Hence T is a Cuntz-Krieger Eu,−−-family. The universal property of C∗(Eu,−−) provides a surjective ∗-homomorphism from C∗(Eu,−−) to C∗(T ) ⊆ O2. Let α1α2 · · · αn be a vertex-simple cycle in Eu,−− without any exit. We see that all the edges αi must be in E1, and hence we have Sα1 Sα2 · · · Sαn = tα1 tα2 · · · tαn = sα1 sα2 · · · sαn and so its spectrum contains the entire unit circle. Theorem 1.2] that C∗(Eu,−−) is isomorphic to C∗(T ). It now follows from [Szy02, Recall that z ∈ M(C∗(S)). Therefore, T ⊆ C∗(S) since A ⊆ C∗(S) and since rwi , yfj ∈ E ⊆ C∗(S), for i = 1, 2, 3, 4, j = 1, 2, . . . , 10. So C∗(T ) ⊆ C∗(S). Since the approximate identity of A given by ( nXk=1 pvk + 1E)n∈N , ese = pr(e), yfi = rfi , e Se = s∗ S∗ Sfi = y∗ S∗ fi fi S∗ d1 Sd1 = rw1 , S∗ d2 Sd2 = pu, e = ses∗ e, fi = yfi y∗ fi d1 = sd1 s∗ d1 , , SeS∗ Sfi S∗ Sd1 S∗ Sd2 S∗ d2 = rw1 −(cid:0)yf1 y∗ f1 + yf2 y∗ f2 + yf5 y∗ f5(cid:1) , 12 SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN where(cid:8)pv(cid:12)(cid:12) v ∈ E0(cid:9) = {pv1 , pv2 , . . . }, is an approximate identity of C∗(T ), we get that for all x ∈ C∗(T ), zxz∗ and z∗xz are elements of C∗(T ). But since A is also contained in C∗(T ) and E ⊆ C∗(T ), we have that S ⊆ C∗(T ), and hence C∗(S) ⊆ C∗(T ). Therefore C∗(Eu,−) ∼= C∗(S) = C∗(T ) ∼= C∗(Eu,−−). (cid:3) The next two results will show that E ∼ME Eu,−− for a row-finite graph E and a vertex u ∈ E0 which supports two distinct return paths. This will be enough to prove our main result since by [Ben15, Lemma 5.1], there exists a row-finite graph F and a vertex v supporting two distinct return paths such that C∗(Eu,−) ⊗ K ∼= C∗(Fv,−) ⊗ K and C∗(E) ⊗ K ∼= C∗(F ) ⊗ K. Proposition 4.6. Let E be a row-finite graph and let u be a vertex which supports two distinct return paths. Then there exists a row-finite graph F and a vertex v ∈ F 0 which supports two distinct loops such that E ∼ME F and Eu,−− ∼ME Fv,−−. Proof. Throughout the proof, we will freely use the following fact: Let G be a graph and let w be a vertex and let w′ 6= w be a regular vertex that does not support a loop. Let G′ be the resulting graph after collapsing w′. Then G ∼ME G′ and Gw,−− ∼ME G′ w,−−. Suppose u ∈ E0 supports two loops. Then set E = F and v = u. Suppose u does not support two loops. Then there exists a return path µ = e1e2 · · · en with n ≥ 2. Starting at r(e1), if r(e1) does not support a loop, we collapse the vertex r(e1). Doing this will result in reducing the length on µ. Note that we may have also added a loop at u. Continuing this procedure, we have obtained a graph E′ with u in (E′)0 such that E ∼ME E′, Eu,−− ∼ME E′ u,−−, and either u supports two loops or u supports a return path ν = f1f2 . . . fm with m ≥ 2, with r(f1) supporting a loop. If u supports two loops, set F = E′ and v = u. Suppose u supports a return path ν = f1f2 . . . fm with m ≥ 2, with r(f1) supporting a loop. Then by Proposition 3.2, we add the r(f1)'th column to the u'th column twice, to get a graph F with u ∈ F 0 supporting two loops such that F ∼ME E′. Note that we may perform the same matrix operations to BE ′ u,−− ∼ME Fu,−−. Set v = u. and get that E′ u,−− We have just obtained the desired graph F and the desired vertex v ∈ F 0 since (cid:3) E ∼ME E′ ∼ME F and Eu,−− ∼ME E′ u,−− ∼ME Fv,−−. We now show that performing the Cuntz splice twice is a legal move for a row- finite graph. Proposition 4.7. Let E be a row-finite graph and let v be a vertex that supports at least two distinct return paths. Then E ∼ME Ev,−−. Proof. According to Proposition 4.6, we can assume that E satisfies the conditions of that proposition -- so we assume that v is a regular vertex that supports at least two loops. For a given matrix size N ∈ N∪{∞} and i, j ∈ {1, 2, . . . , N }, we let E(i,j) denote the N × N matrix that is equal to the identity matrix everywhere except for the (i, j)'th entry, that is 1. If B is a N × N matrix, then E(i,j)B is the matrix obtained from B by adding j'th row into the i'th row, and BE(i,j) is the matrix obtained from B by adding i'th column into the j'th column. Using E−1 (i,j) instead will yield subtraction. In what follows we will make extensive use of the results from Section 3, but we do so implicitly since we feel it will only muddle the exposition if we add all the references in. INVARIANCE OF THE CUNTZ SPLICE 13 Note that BEv,−− can be written as 0 0 1 0 0 0 ... Now let B2 = E(3,2)B1 and B3 = B2E−1 (2,1). Then B1+I ∼ME B2 +I ∼ME B3 +I. 1 0 0 1 1 0 0 1 0 1 0 0 ... ... 0 0 0 0 1 0 0 0 · · · · · · · · · · · · B1 = BE We have that B3 = The 1st vertex in EB3+I does not support a loop, so it can be collapsed yielding · · · · · · · · · · · ·      · · ·  · · · · · · 0 1 0 0 0 0 ...              1 1 0 0 0 ... 2 1 0 1 0 ... B10 = B4 = B9 =      1 0   1 0 0 0 ... ... 0 0 1 0 0 0 ... 1 0 1 1 −1 1 0 0 1 1 0 1 1 0 1 0 0 1 0 0 0 0 ... ... ... 0 0 0 0 1 0 0 0  BE 0 1 0 0 · · · 0 · · · 0 · · · BE 1 0 1 0 0 0 BE 0 1 1 0 0 0 0 0 ... ... 1 −1   (cid:18)2 1 1 1(cid:19) (cid:18)1 0   1 0 0 0 ... ... 1 0 BE · · · · · ·(cid:19)  With B4 + I ∼ME B3 + I. Now we let B5 = E−1 E−1 (4,3)E−1 B5 + I ∼ME B6 + I ∼ME B7 + I ∼ME B8 + I ∼ME B9 + I. We have that (4,3)B6, B8 = E(1,2)B7 and B9 = B8E−1 (2,3)B4, B6 = E(4,1)B5, B7 = (2,3). We then have B4 + I ∼ME In EB9+I the 3rd vertex does not support a loop, so it can be collapsed to yield with B9 + I ∼ME B10 + I. Now we look at the graph E again, and and let BE = (bij). Since the vertex v (number 1) has at least two loops, we have b11 ≥ 1. Now we can insplit by partitioning r−1(v) into two sets, one with a single set consisting of a loop based at v, and the other the rest. In the resulting graph, v is split into two vertices v1 and v2, and let E′ denote the rest of the graph. The vertex v1 has the same edges in and out of E′ as v had, but it has only b11 loops. There is one edge from v1 14 SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN to v2 and v2 has one loop and there are b11 edges from v2 to v1 as well as all the same edges going from v2 into E′ as originally from v. Use the inverse collapse move to add a new vertex u to the middle of the edge from v1 to v2 and call the resulting graph F . Label the vertices such that v2, u and v1 are the 1st, 2nd and 3rd vertices, then BF is: BF =  0 0 (cid:18)0 1 −1(cid:19) (cid:18)b11   0 1 0 0 ... ... b12 0 eB · · · · · ·(cid:19)  C6 = 1 0 (cid:18)1 1 1 2(cid:19) (cid:18)1 0    0 1 0 0 ... ... BE · · · · · ·(cid:19) .  where eB is BE except for on the (1, 1)'th entry, which is b11−1. Note that b11−1 ≥ 0, so that there is still a loop based at the 3rd vertex. Also, note that since E is a row-finite graph, the b1k's are eventually zero. This is important since it allows us to do the following matrix manipulations. Let C2 = BF E(1,2)E(1,2), C3 = E(1,2)C2, C4 = E−1 (1,3)C3, C5 = C4E(2,3) and C6 = C5E(1,2). We have that C1 + I ∼ME C2 + I ∼ME C3 + I ∼ME C4 + I ∼ME C5 + I ∼ME C6 + I. The matrix C6 Therefore, C6 is in fact equivalent to B10 upon relabeling of the first two vertices, thus it follows, that E ∼ME Ev,−−. (cid:3) Thus we have the following fundamental result. Theorem 4.8. Let E be a graph and let v be a vertex that supports at least two distinct return paths. Then C∗(E) ⊗ K ∼= C∗(Ev,−) ⊗ K. Proof. By [Ben15, Lemma 5.1], we may assume that E is a graph with no singular vertices, in particular, E is a row-finite graph. By Theorem 4.5, C∗(Ev,−) ∼= C∗(Ev,−−) and hence, C∗(Ev,−)⊗K ∼= C∗(Ev,−−)⊗K. By Proposition 4.7, C∗(E)⊗ K ∼= C∗(Ev,−−) ⊗ K. Thus, C∗(E) ⊗ K ∼= C∗(Ev,−) ⊗ K. (cid:3) Acknowledgements This work was partially supported by the Danish National Research Founda- tion through the Centre for Symmetry and Deformation (DNRF92), by VILLUM FONDEN through the network for Experimental Mathematics in Number The- ory, Operator Algebras, and Topology, by a grant from the Simons Foundation (# 279369 to Efren Ruiz), and by the Danish Council for Independent Research Natural Sciences. Part of this work was initiated while all four authors were attending the research program Classification of operator algebras: complexity, rigidity, and dynamics at the Mittag-Leffler Institute, January -- April 2016. The authors would also like to thank Aidan Sims for many fruitful discussions. References [AMP07] Pere Ara, M. Angeles Moreno, theory no. doi:10.1007/s10468-006-9044-z . MR 2310414 (2008b:46094) and Enrique Pardo, Nonstable K- (2007), http://dx.doi.org/10.1007/s10468-006-9044-z, Algebr. Represent. graph 157 -- 178, algebras, Theory URL: for 10 2, INVARIANCE OF THE CUNTZ SPLICE 15 [BD96] [Ben15] J. London Math. Soc. Lawrence G. Brown and Marius Dadarlat, Extensions of C ∗-algebras and quasidiagonality, 582 -- 600, URL: http://dx.doi.org/10.1112/jlms/53.3.582, doi:10.1112/jlms/53.3.582. MR 1396721 (97d:46086) Rasmus Bentmann, Cuntz splice invariance for purely infinite graph algebras, ArXiv e-prints (2015), to appear in Math. Scand., arXiv:1510.06757v2. (2) 53 (1996), no. 3, of 2, [BP04] (2004), [CG06] Flow equivalence no. 24 and David Pask, Systems [BHRS02] Teresa Bates, Jeong Hee Hong, Iain Raeburn, and Wojciech Szymański, The ideal structure of the C ∗-algebras of infinite graphs, Illinois J. Math. 46 (2002), no. 4, 1159 -- 1176, URL: http://projecteuclid.org/euclid.ijm/1258138472. MR 1988256 (2004i:46105) Teresa Bates algebras, graph Ergodic Theory Dynam. 367 -- 382, URL: http://dx.doi.org/10.1017/S0143385703000348, doi:10.1017/S0143385703000348. MR 2054048 (2004m:37019) Tyrone Crisp and Daniel Gow, Contractible subgraphs and Morita equiv- alence of graph C ∗-algebras, Proc. Amer. Math. Soc. 134 (2006), no. 7, 2003 -- 2013, http://dx.doi.org/10.1090/S0002-9939-06-08216-5, doi:10.1090/S0002-9939-06-08216-5. MR 2215769 (2006k:46083) Joachim Cuntz and Wolfgang Krieger, A class of C ∗-algebras and topo- logical Markov chains, 251 -- 268, URL: http://dx.doi.org/10.1007/BF01390048, doi:10.1007/BF01390048. MR 561974 (82f:46073a) Toke Meier Carlsen, Gunnar Restorff, and Efren Ruiz, Strong classification of purely infinite Cuntz-Krieger algebras, In preparation, 2016. J. Cuntz, A class of C ∗-algebras and topological Markov chains. II. Reducible chains and the Ext-functor for C ∗-algebras, Invent. Math. 63 (1981), no. 1, 25 -- 40, URL: http://dx.doi.org/10.1007/BF01389192, doi:10.1007/BF01389192. MR 608527 (82f:46073b) Invent. Math. 56 (1980), no. 3, [CRR16] [Cun81] [CK80] URL: [Cun86] [ERR13] , The classification problem for the C ∗-algebras OA, Geometric methods in operator algebras (Kyoto, 1983), Pitman Res. Notes Math. Ser., vol. 123, Longman Sci. Tech., Harlow, 1986, pp. 145 -- 151. MR 866492 (88a:46081) Søren Eilers, Gunnar Restorff, and Efren Ruiz, Strong classification of extensions of classifiable C ∗-algebras, ArXiv e-prints (2013), arXiv:1301.7695v1. [ERRS16a] , Geometric classification of C ∗-algebras over finite graphs, in preparation, 2016. and [Fra84] [HS03] infinite URL: Soc. 128 no. (2000), graphs, [Gab16] [FLR00] Proc. Amer. Math. J. Fowler, Marcelo Laca, [ERRS16b] Søren Eilers, Gunnar Restorff, Efren Ruiz, and Adam P. W. Sørensen, The complete classification of unital graph C ∗-algebras: Geometric and strong, in preparation, 2016. Iain Raeburn, The C ∗-algebras Neal 8, of 2319 -- 2327, http://dx.doi.org/10.1090/S0002-9939-99-05378-2, doi:10.1090/S0002-9939-99-05378-2. MR 1670363 (2000k:46079) John Franks, Flow equivalence of subshifts of finite type, Ergodic Theory Dynam. Sys- tems 4 (1984), no. 1, 53 -- 66, URL: http://dx.doi.org/10.1017/S0143385700002261, doi:10.1017/S0143385700002261. MR 758893 (86j:58078) James Gabe, Classifying purely infinite C ∗-algebras: The obstruction class, In prepa- ration, 2016. Jeong Hee Hong and Wojciech Szymański, Purely infinite Cuntz-Krieger algebras of directed graphs, Bull. London Math. Soc. 35 (2003), no. 5, 689 -- 696, URL: http://dx.doi.org/10.1112/S0024609303002364, doi:10.1112/S0024609303002364. MR 1989499 (2005c:46097) Danrun Huang, Flow equivalence of type and Cuntz-Krieger algebras, J. Reine Angew. Math. 462 (1995), 185 -- 217, URL: http://dx.doi.org/10.1515/crll.1995.462.185, doi:10.1515/crll.1995.462.185. MR 1329907 (96m:46123) Eberhard Kirchberg, Das nicht-kommutative Michael-Auswahlprinzip und die Klas- sifikation nicht-einfacher Algebren, C ∗-algebras (Münster, 1999), Springer, Berlin, 2000, pp. 92 -- 141. MR 1796912 (2001m:46161) Eberhard Kirchberg and N. Christopher Phillips, Embedding of exact C ∗- algebras in the Cuntz algebra O2, J. Reine Angew. Math. 525 (2000), 17 -- 53, URL: http://dx.doi.org/10.1515/crll.2000.065, doi:10.1515/crll.2000.065. MR 1780426 (2001d:46086a) Paul S. Muhly and Mark Tomforde, Adding tails to C ∗-correspondences, Doc. Math. 9 (2004), 79 -- 106. MR 2054981 (2005a:46117) shifts of finite reducible [Hua95] [MT04] [KP00] [Kir00] 16 SØREN EILERS, GUNNAR RESTORFF, EFREN RUIZ, AND ADAM P. W. SØRENSEN [Rae05] [Res06] [Rør95] [Sør13] [Szy02] Classification up to of Cuntz-Krieger algebras J. Reine Angew. Math. 598 (2006), sta- 185 -- 210, URL: doi:10.1515/CRELLE.2006.074. Iain Raeburn, Graph algebras, CBMS Regional Conference Series in Mathematics, vol. 103, Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2005. MR 2135030 (2005k:46141) Gunnar Restorff, isomorphism, ble http://dx.doi.org/10.1515/CRELLE.2006.074, MR 2270572 (2007m:46090) Mikael Rørdam, Classification of Cuntz-Krieger algebras, K-Theory 9 (1995), no. 1, 31 -- 58, URL: http://dx.doi.org/10.1007/BF00965458 , doi:10.1007/BF00965458. MR 1340839 (96k:46103) simple graph algebras, Adam P. W. Sørensen, Geometric classification of Ergodic Theory Dynam. Systems 33 (2013), 1199 -- 1220, URL: 4, no. http://dx.doi.org/10.1017/S0143385712000260, doi:10.1017/S0143385712000260. MR 3082546 Wojciech Szymański, General Cuntz-Krieger uniqueness theorem, Internat. J. Math. 13 (2002), no. 5, 549 -- 555, URL: http://dx.doi.org/10.1142/S0129167X0200137X, doi:10.1142/S0129167X0200137X. MR 1914564 (2003h:46083) Department of Mathematical Sciences, University of Copenhagen, Universitets- parken 5, DK-2100 Copenhagen, Denmark E-mail address: [email protected] Department of Science and Technology, University of the Faroe Islands, Nóatún 3, FO-100 Tórshavn, the Faroe Islands E-mail address: [email protected] Department of Mathematics, University of Hawaii, Hilo, 200 W. Kawili St., Hilo, Hawaii, 96720-4091 USA E-mail address: [email protected] Department of Mathematics, University of Oslo, PO BOX 1053 Blindern, N-0316 Oslo, Norway E-mail address: [email protected]
1810.10272	1	1810	2018-10-24T09:49:48	On pseudodifferential operators on filtered and multifiltered manifolds	[ "math.OA" ]	This memoir is a summary of recent work, including collaborations with Erik van Erp, Christian Voigt and Marco Matassa, compiled for the "Habilitation \`a diriger des recherches". We present various different approaches to constructing algebras of pseudodifferential operators adapted to filtered and multifiltered manifolds and some quantum analogues. A general goal is the study of index problems in situations where standard elliptic theory is insufficient. We also present some applications of these constructions. We begin by presenting a characterization of pseudodifferential operators on filtered manifolds in terms of distributions on the tangent groupoid which are essentially homogeneous with respect to the natural $\mathbb{R}^\times_+$-action. Next, we describe a rudimentary multifiltered pseudodifferential theory on the full flag manifold $\mathcal{X}$ of a complex semisimple Lie group $G$ which allows us to simultaneously treat longitudinal pseudodifferential operators along every one of the canonical fibrations of $\mathcal{X}$ over smaller flag manifolds. The motivating application is the construction of a $G$-equivariant $K$-homology class from the Bernstein-Gelfand-Gelfand complex of a semisimple group. Finally, we discuss pseudodifferential operators on two classes of quantum flag manifolds: quantum projective spaces and the full flag manifolds of $SU_q(n)$. In particular, on the full flag variety of $SU_q(3)$ we obtain an equivariant fundamental class from the Bernstein-Gelfand-Gelfand complex.	math.OA	math	UNIVERSITÉ CLERMONT AUVERGNE Laboratoire de Mathématiques Blaise Pascal UMR 6620 CNRS TRAVAUX DE RECHERCHE présentés en vue de l'obtention de L'HABILITATION À DIRIGER DES RECHERCHES EN MATHÉMATIQUES ON PSEUDODIFFERENTIAL OPERATORS ON FILTERED AND MULTIFILTERED MANIFOLDS par Robert Yuncken Maître de conférences en Mathématiques à l'Université Clermont Auvergne Rapporteurs : • Georges Skandalis, Université Paris 7 Denis Diderot • Pierre Julg, Université d'Orleans • Sergey Neshveyev, Univeristé d'Oslo Habilitation soutenue publiquement le 28 septembre 2018 devant le jury com- posé de : • Georges Skandalis, Université Paris 7 Denis Diderot • Ryszard Nest, University de Copenhague • Pierre Julg, Université d'Orléans • Claire Debord, Université Paris 7 Denis Diderot • Julien Bichon, Université Clermont Auvergne UNIVERSITÉ CLERMONT AUVERGNE Laboratoire de Mathématiques Blaise Pascal UMR 6620 CNRS Campus des Cézeaux 63178 Aubière cedex France Téléphone : +33 (0)4 73 40 76 97 Courriel : [email protected] Abstract We present various different approaches to constructing algebras of pseudod- ifferential operators adapted to particular geometric situations. A general goal is the study of index problems in situations where standard elliptic theory is insufficient. We also present some applications of these constructions. We begin by presenting a characterization of pseudodifferential operators in terms of distributions on the tangent groupoid which are essentially ho- mogeneous with respect to the natural R× +-action. This is carried out in the generality of filtered manifolds, which are manifolds that are locally modelled on nilpotent groups, generalizing contact, CR and parabolic geometries. We describe the tangent groupoid of a filtered manifold, and use this to construct a pseudodifferential calculus analogous to the unpublished calculus of Melin. Next, we describe a rudimentary multifiltered pseudodifferential theory on the full flag manifold X of a complex semisimple Lie group G which allows us to simultaneously treat longitudinal pseudodifferential operators along ev- ery one of the canonical fibrations of X over smaller flag manifolds. The mo- tivating application is the construction of a G-equivariant K-homology class from the Bernstein-Gelfand-Gelfand complex of a semisimple group. This construction been completely resolved for only a few groups, and we will discuss the remaining obstacles as well as the successes. Finally, we discuss pseudodifferential operators on two classes of quantum flag manifolds. First, we consider quantum projective spaces CPn q , where we can generalize the abstract pseudodifferential theory of Connes and Moscovici to obtain a twisted algebra of pseudodifferential operators associated to the Dolbeault-Dirac operator. Secondly, we look at the full flag manifolds of SUq(n), where we instead generalize the multifiltered construction of the classical flag manifolds, thus obtaining an equivariant fundamental class for the full flag variety of SUq(3) from the Bernstein-Gelfand-Gelfand complex. As applications, we obtain equivariant Poincaré duality of the quantum flag manifold of SUq(3) and the Baum-Connes Conjecture for the discrete dual of SUq(3) with trivial coefficients. Thanks There are a lot of people to thank. Let me start at the beginning by again thanking my PhD adviser, Nigel Higson, for his guidance and good taste. He has been a constant source of inspiration. Thanks also to the late great John Roe, another big inspiration, who has had far more influence on my work than may be apparent. Huge thanks to my collaborators -- Heath, Christian, Erik, Julien, Marco and Robin -- who have all made my mathematics better, and more fun. Thanks to the French operator algebra community for adopting me. I couldn't ask for a nicer clan. The list is long, but let me particularly thank Hervé, Jean-Marie, Claire, Georges, Saad, Jean and Julien. Thanks very much to the referees and jury members who are offering their precious time for assessing this work. And last but not least, thanks to my family: to my parents for their support, to my kids for keeping me happy with their bottomless spirit, and to HJ, who owns a piece of everything I have done and who deserves more thanks than I could fit on this page. List of publications presented The following are articles which will be presented in this memoir. Chapter 2: Pseudodifferential operators from tangent groupoids [vEY17] Erik van Erp and Robert Yuncken. On the tangent groupoid of a filtered manifold. Bull. Lond. Math. Soc., 49(6):1000 -- 1012, 2017 [vEY] Erik van Erp and Robert Yuncken. A groupoid approach to pseudodif- ferential operators. J. Reine Agnew. Math. To appear. https://doi.org/10.1515/crelle-2017-0035 Chapter 3: Pseudodifferential operators on multifiltered man- ifolds and equivariant index theory for complex semisimple groups [Yun13] Robert Yuncken. Foliation C∗-algebras on multiply fibred manifolds. J. Funct. Anal., 265(9):1829 -- 1839, 2013 This work is heavily motivated by the earlier publication: [Yun11] Robert Yuncken. The Bernstein-Gelfand-Gelfand complex and Kas- parov theory for SL(3, C). Adv. Math., 226(2):1474 -- 1512, 2011 Chapter 4: Pseudodifferential operators on quantum flag man- ifolds [MY] Marco Matassa and Robert Yuncken. Regularity of twisted spectral triples and pseudodifferential calculi. J. Noncommut. Geom. To appear. Preprint at https://arxiv.org/abs/1705.04178 [VY15] Christian Voigt and Robert Yuncken. Equivariant Fredholm modules for the full quantum flag manifold of SUq(3). Doc. Math., 20:433 -- 490, 2015 Material from the following preprint will also be used: [VY17] Christian Voigt and Robert Yuncken. Complex semisimple quantum groups and representation theory. Preprint. https://arxiv.org/abs/1705.05661 (226 pages), 2017 Contents 1 Introduction 1.1 Elliptic operators and their generalizations . . . . . 1.2 Pseudodifferential operators . . . . 1.3 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Pseudodifferential operators from tangent groupoids . . . . . . . . . . . . . . . . . . . . . . . 2.1 Filtered manifolds . . . . 2.2 Schwartz kernels and fibred distributions on Lie groupoids . . . . . 2.3 The osculating groupoid . . . . . . 2.4 The tangent groupoid of a filtered manifold . 2.5 The R× . . . . . . . . . . . . . . 2.6 Pseudodifferential operators . . . . 2.7 Principal symbols . . . . . . . . . . . . . 2.8 Algebra structure, conormality, regularity . . . 2.9 Hypoellipticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . +-action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Pseudodifferential operators on multifiltered manifolds and equiv- ariant index theory for complex semisimple groups . . . . . . . . . . . . . 3.1 Complex flag manifolds . . . . . . . . . . . . . . . 3.2 Motivation: The BGG complex . . . . . . . . . . . . . . 3.3 Longitudinal pseudodifferential operators . . . . . . 3.4 Multifiltered manifolds . . . . . . . . . . . . . . . . 3.5 Locally homogeneous structures . . . . . . . . . . . . 3.6 Products of longitudinally smoothing operators . . . . . 3.7 Application: The KK-class of the BGG complex of rank-two . . . . . . . . . . . . . . . . . . . . . . . . . . complex semisimple groups . . . . . . . . . . . . . . . . . . . 3 4 5 6 8 9 10 12 13 14 15 16 17 19 21 22 23 25 25 26 28 29 . . . . . . . . 4 Pseudodifferential operators on quantum flag manifolds . . . . 4.1 Notation and conventions . . . . . 4.2 Pseudodifferential calculus and twisted spectral triples . . . . . . . . . . . 33 34 36 36 4.2.1 Twisted regularity . . . . . 4.2.2 Differential and pseudodifferential operators . 36 4.2.3 Pseudodifferential calculus on quantum projective spaces 39 41 4.3 Pseudodifferential operators on the full flag manifold of SUq(n) 4.3.1 Algebras of longitudinal pseudodifferential operators on . . quantum flag manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 1 4.3.2 Principal series representations of SLq(n, C) . . 4.3.3 Algebras of longitudinal pseudodifferential operators, . . 4.3.4 The BGG class of the quantum flag manifold of SUq(3) . cont'd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 45 45 2 Chapter 1 Introduction The goal of this memoir is to describe a handful of different approaches to constructing algebras of pseudodifferential operators on manifolds and non- commutative spaces with particular geometries. In each case, the methods will be rather different, but the goals similar. We will also present some of the applications of these pseudodifferential algebras. Pseudodifferential operators are, of course, a tool for studying differential operators, particularly elliptic differential operators and their generalizations. Our motivations come from index theoretic problems where the underlying geometry demands more exotic pseudodifferential theories than ordinary el- liptic theory. We will be particularly interested in: (1) Rockland operators, including subelliptic operators on contact and Heisenberg manifolds; (2) the Bernstein-Gelfand-Gelfand complex, a canonical complex of equiv- ariant differential operators on the flag manifold of a semisimple Lie group; (3) analogues of the above operators on quantum homogeneous spaces. As mentioned, the techniques of analysis for the various examples will be very different: in (1) we shall use groupoid techniques, in (2) we use convo- lution operators on nilpotent Lie groups, and in (3) methods of noncommu- tative geometry and noncommutative harmonic analysis. But in each case, the general strategy -- the requirements of the associated pseudodifferential calculus -- are roughly similar. We shall dedicate this introduction to a very broad discussion of this strat- egy. Our goal here is to isolate the general properties of pseudodifferential operators which are essential to index theory. Finally, at the end of the in- troduction, we will give a brief overview of the particular problems to be discussed in the following chapters. 3 1.1 Elliptic operators and their generalizations Index theory begins with the study of analytic properties of elliptic differential operators on a closed manifold M. The crucial properties of elliptic operators, which for the moment we will state imprecisely1 , are: (1) Elliptic differential operators on a closed manifold are Fredholm. (2) If two elliptic differential operators have the same principal symbol then their difference is a compact operator. These two properties imply that the Fredholm index of an elliptic differential operator depends only upon the principal symbol of the operator. Moreover, the index is unchanged by smooth perturbations of the principal symbol. In other words, the index of an elliptic operator depends only on some topolog- ical data associated to the principal symbol. The appropriate topological data is the K-theory class of the principal symbol σ(D), and for elliptic operators, the index is calculated by the famous Atiyah-Singer Formula: Ind(D) =ZM ch(σ(D))Td(M). Properties (1) and (2) are valid for a much larger class of operators than elliptic differential operators. The examples we consider will all be Rockland (see Definition 2.9.3). This is a generalization of ellipticity to filtered mani- folds, i.e. manifolds with a filtration of the tangent bundle compatible with the Lie bracket of vector fields. Such operators appear naturally in applica- tions, particularly in CR and contact geometry, and more recently in parabolic geometry, see e.g. [FS74, Roc78, BG88, Tay, Mel82, EMM91, CSS01, DH17]. One example of particular interest to us is the Bernstein-Gelfand-Gelfand complex. This is a canonical equivariant differential complex on a flag man- ifold of a semisimple Lie group, or more generally on a parabolic manifold [ CSS01]. The Bernstein-Gelfand-Gelfand complex has gained much attention recently, originally due to its appearance in twistor theory [BE89]. From our point of view, the BGG complex is an obligatory replacement of a Dirac-type operator for the equivariant index theory of semisimple Lie groups and their quantizations; see Sections 3 and 4. Another important property of elliptic differential operators, also shared by Rockland operators, is hypoellipticity. Let E be a vector bundle over a manifold M without boundary. We write C∞(M; E) for the space of smooth sections of E, and D′(M; E) for the space of distributional sections. Definition 1.1.1. A linear differential operator P : D′(M; E) → D′(M; E) is hypoelliptic if, for any open set U ⊆ M and any distribution u ∈ D′(M; E) we have PuU ∈ C∞(U; E) =⇒ uU ∈ C∞(U; E). 1For a precise interpretation of these statements, one should consider an elliptic operator of order m as a bounded operator between Sobolev spaces Hs+m(M) → Hs(M) for any s ∈ R. 4 In other words, hypoellipticity guarantees smooth solutions u to the par- tial differential equation Pu = f whenever the right-hand-side f is smooth. Both hypoellipticity and Fredholmness are typically proven by constructing a pseudodifferential calculus adapted to P. 1.2 Pseudodifferential operators Let M be a smooth manifold without boundary. For simplicity, we will sup- press coefficient vector bundles in this section. We will also be lazy about the difference between functions and densities. We will be more precise in later chapters. Definition 1.2.1. A distribution p ∈ D′(M × M) is called • properly supported if the restriction to supp(p) of each of the two coordi- nate projections pr1, pr2 : M × M ։ M is a proper map; • semiregular in both variables if u ∈ C∞(M) ¯⊗D′(M) ∩ D′(M) ¯⊗C∞(M). We write D′ p(M × M) for the space of distributions on M × M which are properly supported and semiregular in both variables -- for a better context see Definition 2.2.3. Via the Schwartz kernel theorem, D′ p(M × M) is in bijection with the al- gebra of operators on D′(M) which preserve each of the subspaces E ′(M), C∞(M) and C∞ c (M), see [Trè67]. By abuse of notation, we will often identify such an operator with its kernel in D′ p(M × M). The product of such operators corresponds to the convolution product of distributions: p ∗ q(x, z) =ZM p(x, y)q(y, z) dy, p, q ∈ D′ p(M × M), which makes sense thanks to semiregularity. In particular, linear differential operators on M correspond to elements of D′ p(M × M) with support on the diagonal. Now let Eℓℓ be some set of linear differential operators on M. We are imagining a class of operators for which we hope to prove hypoellipticity or Fredholmness. The following definition is intended only as a guiding philos- ophy. Definition 1.2.2. An algebra of pseudodifferential operators adapted to Eℓℓ will mean a Z-filtered algebra Ψ• ⊂ D′ p(M × M) which has the following proper- ties: 1. Ψ• contains (Schwartz kernels of) all linear differential operators on M. 2. Pseudolocality: Every p ∈ Ψ• is equal to a smooth function off the diagonal. 3. Ψ−∞ = Tm∈Z Ψm = C∞ smoothing kernels. p (M × M) is the algebra of properly supported 5 4. Existence of asymptotic limits: Let m ∈ Z and pi ∈ Ψm−i for all i ∈ N Then there exists p ∈ Ψm such that for every n ∈ Z, p − k ∑ i=0 pi ∈ Ψ−n for all k ≫ 0. 5. Existence of parametrices: If p ∈ Eℓℓ, then there exists q ∈ Ψ• such that I − p ∗ q, I − q ∗ p ∈ Ψ−1, where I is the Schwartz kernel of the identity operator. For instance, the algebra of classical (step one polyhomogeneous) pseu- dodifferential operators on M is adapted to the class of elliptic differential operators in this sense. If M is a contact manifold then the Heisenberg calcu- lus of Beals and Greiner [BG88] is adapted to the maximally hypoelliptic (i.e., Rockland) operators. The point of the five properties above is that they imply the hypoellipticity of operators in Eℓℓ. Theorem 1.2.3. Let Eℓℓ be a class of linear differential operators on a manifold M and suppose that there exists an algebra of pseudodifferential operators Ψ• adapted to Eℓℓ in the sense of Definition 1.2.2. Then the operators in Eℓℓ are hypoelliptic. Proof. Let P ∈ Eℓℓ of order m. Let Q ∈ Ψ−m be a parametrix, and put R = I − PQ ∈ Ψ−1. The series Q ∑∞ k=0 Rk admits an asymptotic limit, which we denote by Q. One checks that I − P Q, I − QP ∈ Ψ−∞. Now, for any u ∈ D′(M) we have u = (I − QP)u + Pu, and it follows that sing-supp(u) ⊆ sing-supp(Pu). Thus P is hypoelliptic. A similar philosophy could be applied to Fredholmness. By adding slightly more structure to Definition 1.2.2, replicating basic elements of Sobolev theory, we could deduce Fredholmness of the operators in Eℓℓ. We will not formulate a precise statement here. 1.3 Overview The goal of the following chapters, roughly speaking, will be to obtain some version of the properties stated in Definition 1.2.2 for algebras of pseudodif- ferential operators which are adapted to particular differential operators. This will be achieved with more or less success depending on the examples. In Chapter 2, we will describe an approach to pseudodifferential theory based on the tangent groupoid of a filtered manifold [vEY]. The main result is a simple groupoid definition of pseudodifferential kernels which repro- duces the classical calculus, the Heisenberg calculus [BG88, Tay], and more 6 generally, Melin's unpublished pseudodifferential calculus on an arbitrary fil- tered manifold [Mel82]. Philosophically, this result is supposed to indicate a general principal: in order to define a pseudodifferential calculus, it is suffi- cient to construct a tangent groupoid adapted to the geometry, from which the pseudodifferential calculus follows automatically. In Chapter 3, we consider longitudinal pseudodifferential operators on manifolds equipped with multiple foliations [Yun13]. The examples of inter- est are the flag manifolds of complex semisimple Lie groups, which admit a family of canonical fibrations over smaller flag manifolds. The results pre- sented are designed to simultaneously treat longitudinally pseudodifferential operators along the fibres of different fibrations. The theory we present is far from being a pseudodifferential calculus in the sense of Definition 1.2.2, since it essentially distinguishes only between longitudinal pseudodifferential operators of order 0 and those of negative or- der. Still, it allows us to interpret the Bernstein-Gelfand-Gelfand complex as an equivariant fundamental class in the Kasparov K-homology of the full flag manifold of SL(3, C). As has been shown in [Yun11], this leads an explicit construction of the γ element from the Baum-Connes Conjecture. Finally, In Chapter 4, we pass to pseudodifferential theory in noncommu- tative geometry -- specifically, on quantized flag manifolds. These noncom- mutative spaces would seem to deserve the status of "noncommutative man- ifolds", but it has been very difficult to incorporate them into Connes' frame- work. Much of the difficulty is due to the absence of a reasonable notion of pseudodifferential calculus for these spaces. We will consider two classes of quantum flag manifolds, with rather differ- ent behaviours. Firstly, we look at the quantized projective spaces CPn q , where things are relatively simple. We show that the existing spectral triples on CPn q [Krä04, KTS15, DD10] are twisted regular, in the sense of [CM08]. This is achieved by means of a generalization of the abstract pseudodifferential calculus of Connes and Moscovici [CM95, CM08, Mos10]. Secondly, we consider the full flag manifold Xq of SUq(n), for which we have no candidates for a spectral triple. Instead, we show that essential ele- ments of the multi-filtered harmonic analysis of Chapter 3 can be carried over to quantum flag manifolds [VY15]. As a result, the quantum analogue of the Bernstein-Gelfand-Gelfand complex for SUq(3) yields a fundamental class in equivariant K-homology KSLq(3,C)(Xq, C), which allows us to prove equivari- ant Poincaré Duality for Xq and the Baum-Connes Conjecture (with trivial coefficients) for the discrete dual of SUq(3). This document is intended as a survey. Proofs, when given, will be sketched, with references to the original publications for details. 7 Chapter 2 Pseudodifferential operators from tangent groupoids Lie groupoids are a fusion of differential geometry and algebra. This structure allow us to describe linear operators on smooth manifolds as convolution operators. In particular, pseudodifferential operators on a smooth manifold M can be interpreted as convolution operators by their Schwartz kernels on the pair groupoid M × M, while the principal symbol, up to Fourier transform, is a convolution operator on the tangent bundle TM seen as a bundle of abelian Lie groups. Connes introduced the tangent groupoid [Con94] as a tool for smoothly deforming a pseudodifferential operator to its principal symbol. Debord and Skandalis realized that classical pseudodifferential operators could be characterized in terms of the tangent groupoid [DS14]. Inspired by this, we showed in [vEY] that tangent bundles can be used to define pseu- dodifferential calculi adapted to a wide variety of geometric situations. The context we work in is that of a filtered manifold M, which is a structure gen- eralizing contact manifolds, Heisenberg manifolds, and parabolic manifolds. Pseudodifferential calculi on such manifolds have been developed by many authors [FS74, Tay, BG88, EM, Mel82]. One can construct a tangent groupoid TH M for such manifolds [vE05, Pon06, vEY17, CP]. It is a deformation of the pair groupoid M × M to the bundle of osculating groups of M. As in [DS14], it admits a one-parameter group of automorphisms (αλ)λ∈R× . We define a pseudodifferential kernel of order m to be a properly supported semiregular Schwartz kernel p which is the restriction to M × M of a smooth family p of distributions on the r-fibres of TH M satisfying + αλ∗p − λmp ∈ C∞(TH M), ∀λ ∈ R× +. A precise statement will be given in Section 2.6. We showed in [vEY] that this coincides with the usual definitions for the classical and Heisenberg calculi. This chapter is dedicated to explaining this definition and its consequences. 8 2.1 Filtered manifolds Definition 2.1.1. A filtered manifold (also called a Carnot manifold in [CP]) is a C∞ manifold M equipped with a filtration of its tangent bundle by subbundles 0 = H0 ≤ H1 ≤ · · · ≤ HN = TM, such that the space of smooth vector fields Γ∞(TM) becomes a filtered Lie algebra: [Γ∞(Hi), Γ∞(Hj)] ⊆ Γ∞(Hi+j) ∀i, j. Here we are using the convention Hk = TM for all k ≥ N. The number N will be called the depth of the filtration. Smooth sections of Hk are called vector fields of H-order ≤ k. The filtration of vector fields by H-order generates an algebra filtration on the differential operators1 DO(M). The set of differential operators of H-order ≤ m will be denoted DOm H(M). Example 2.1.2. Any manifold M can be equipped with the trivial filtration of depth 1, where H1 = TM. The resulting filtration on DO(M) corresponds to the usual notion of order of a differential operator. Example 2.1.3. Let M be the Heisenberg group of dimension 3. We write X, Y, Z for the usual left-invariant vector fields which satisfy [X, Z] = [Y, Z] = 0 and [X, Y] = Z. We obtain a filtration of TM of depth 2 by defining H1 to be the 2-dimensional subbundle spanned by X and Y. Then the vector fields X and Y are order 1, while Z = XY − YX counts as order 2. This example generalizes to the notion of Heisenberg order for differential operators on a contact manifold, see [BG88]. Example 2.1.4. Any subbundle H of the tangent bundle of M gives rise to a filtered manifold of depth 2, 0 ≤ H ≤ TM, since the condition on Lie brackets is trivial. This includes contact mani- folds, where H is the contact hyperplane bundle, as well as foliated manifolds, where H is the tangent space to the leaves. Example 2.1.5. Flag manifolds and, more generally, parabolic geometries are examples of filtered manifolds [ CS09]. The principal part of a differential operator on a filtered manifold is de- fined via the passage from the filtered algebra DOH(M) to its associated graded algebra. We fix some general notation. Notation 2.1.6. For any N-filtered vector space (or vector bundle) V•, we write gr V =Mm grm V :=Mm Vm/Vm−1 1If M is noncompact, we will insist that DO(M) consists of the differential operators of finite order. 9 for the associated graded space (or graded bundle) and σm : Vm → grm V ֒→ gr V for the canonical quotient maps. Definition 2.1.7. If P ∈ DOm principal part is defined as σm(P). H(M) is a differential operator of H-order m, its As in elliptic theory, it is the principal part of a differential operator which will be used to prove hypoellipticity and Fredholmness. The usual criterion of ellipticity needs to be replaced by the Rockland condition, which we will state in Section 2.9. 2.2 Schwartz kernels and fibred distributions on Lie groupoids Definition 2.2.1. The pair groupoid of a smooth manifold M is the Cartesian product M × M equipped with range and source maps r(x, y) = x s(x, y) = y and groupoid laws (x, y)(y, z) = (x, z), (x, y)−1 = (y, x). This is the appropriate structure for describing Schwartz kernels of con- tinuous operators on C∞(M); see below. Recall, from Section 1.2 that we are interested in Schwartz kernels which are properly supported and semiregular in both variables. These notions have groupoid interpretations, which will important in the sequel. Definition 2.2.2. Let G be a Lie groupoid with base space G(0). Consider C∞(G) as a C∞(G(0))-module via the range fibration: a. f = r∗(a) f , a ∈ C∞(G(0)), f ∈ C∞(G). (2.2.1) Then an r-fibred distribution2 on G means a continuous C∞(G(0))-linear opera- tor u : C∞(G) → C∞(G(0)). (2.2.2) The space of r-fibred distributions on G is denoted by E ′ r(G). Likewise, we define the space Es(G) of s-fibred distributions on G, where we replace the C∞(G(0))-module structure (2.2.1) with its analogue using the source fibration. 2 Strictly speaking, we are defining here the r-fibred distributions with compact vertical support (or r-proper support). We will not need any more general support conditions. 10 It is possible to integrate an r- or s-fibred distribution to an ordinary dis- tribution on G by composing with a fixed choice of nowhere zero smooth density µ on the base space G(0). This yields maps µr :E ′ µs :E ′ r(G) → D(G); s(G) → D(G); u 7→ µ ◦ u, v 7→ µ ◦ v. These maps depend on the choice of µ, but their images do not. In this way, we may see the r-fibred or s-fibred distributions as those distributions on G which can be disintegrated as a smooth family of distributions on the r-fibres or s-fibres, respectively. We will want to have both. Definition 2.2.3. A distribution w ∈ D′(G) will be called proper if it lies in the image of both µr and µs. We write D′ p(G) for the space of proper distributions. We also define the spaces of proper r-fibred and proper s-fibred distributions, re- spectively, as r,s(G) = µ−1 E ′ r D′ p(G) and s,r(G) = µ−1 E ′ s D′ p(G). To see the importance of these definitions, consider again the pair groupoid G = M × M. We will fix, once and for all, a nowhere vanishing smooth density µ on M, which allows us to identify C∞(M) as a subspace of D′(M). Recall that if p ∈ D′(M × M) is an arbitrary distribution, the Schwartz kernel c (M) → D′(M), and so these operator with kernel p defines a linear map C∞ operators do not form an algebra. In this respect, the proper distributions are better. Proposition 2.2.4. The following are equivalent for a distribution p ∈ D′(M × M): (1) The Schwartz kernel operator with kernel p extends to an operator preserving each of the spaces D′(M), E ′(M), C∞(M) and C∞ c (M); (2) p is properly supported and semiregular in both variables (see Section 1.2); (3) p ∈ D′ p(M × M). In a similar fashion, the convolution product of distributions on a groupoid G, p ∗ q(γ) =Zβ∈Gr(γ) p(β)q(β−1γ), does not make sense for arbitrary distributions p, q ∈ D′(G), but it does make sense for p, q ∈ E ′ r(G), via the interpretation hp ∗ q, ϕi = hp(β), hq(β−1γ), ϕ(γ)ii, ϕ ∈ C∞(G). (See [LMV17, vEY].) Moreover, this product restricts to the subspace E ′ r,s(G). In particular, under the equivalences of Proposition 2.2.4, composition of Schwartz kernel operators corresponds to groupoid convolution on E ′ r,s(M × M) ∼= D′ p(M × M). We next need to introduce the Lie groupoid analogue of the ideal of smoothing operators on M. 11 Definition 2.2.5. Let Ωr denote the bundle of 1-densities along ker(dr), the tangent bundle of the r-fibres. We denote by C∞ p (G; Ωr) the space of sections f of Ωr with proper support, meaning that r, s : supp( f ) → G(0) are both proper maps. Lemma 2.2.6. The space C∞ p (G; Ωr) is a two-sided ideal of E ′ r,s(G). Remark 2.2.7. Note that the subalgebra of smooth densities in E ′ ideal but not a left ideal. Likewise, the algebra of smooth densities in E ′ a left ideal but not a right ideal. r(G) is a right s(G) is When G = M × M, this corresponds to the fact that the properly sup- ported smoothing operators on M are a left ideal in the algebra of continuous operators on E ′(M), but not a right ideal, and a right ideal in the algebra of continuous operators on C∞ c (M) but not a left ideal. For this reason, it is crucial that we work with proper r-fibred distributions on G. 2.3 The osculating groupoid Note that the tangent bundle TM of a smooth manifold is a Lie groupoid, viewed as a bundle of abelian Lie groups. The osculating groupoid TH M is a replacement for TM in the world of filtered manifolds. As usual, we start with the Lie algebroid. Definition 2.3.1. The osculating Lie algebroid tH M of a filtered manifold M is the associated graded bundle of the filtered tangent bundle: tH M = gr TM = Mm∈N Hm/Hm−1. This is indeed a Lie algebroid, with anchor zero, thanks to the following calculation. Let X ∈ Γ∞(Hm), Y ∈ Γ∞(Hn) be vector fields on M of H-order m and n respectively. Then for any f , g ∈ C∞(M) we have [ f X, gY] = f g[X, Y] + f (Xg)Y − g(Y f )X ≡ f g[X, Y] mod Γ∞(Hm+n−1). Thus the Lie bracket of vector fields induces a C∞(M)-linear Lie bracket on sections of the associated graded bundle, and hence a pointwise Lie bracket on the fibres tH Mx (x ∈ M). In this way, tH M is a smooth bundle of nilpotent Lie algebras. Definition 2.3.2. The osculating groupoid TH M of a filtered manifold M is the bundle of connected, simply connected nilpotent Lie groups which integrates tH M. In other words, TH Mx = tH Mx with product defined by the Baker-Campbell- Hausdorff formula. The principal part σm(P) of a differential operator of H-order m (see Def- inition 2.1.7) can now be interpreted in groupoid language. Let us spell this out. 12 The classical tangent bundle TM ։ M is the Lie algebroid of the pair Its section space Γ∞(TM) is the space of vector fields, groupoid M × M. and its universal enveloping algebra U (TM) is the algebra DO(M) of linear differential operators on M (see [NWX99]). A Lie filtration on M is equivalent to a Lie algebroid filtration on TM, and this induces an algebra filtration on the enveloping algebra DO(M). This is precisely the filtration DO• H(M) by H-order. Passing to the associated graded spaces is functorial for all these construc- tions. The osculating groupoid tH M = gr TM is the associated graded of TM, and thus the universal enveloping algebra of tH M is U (tH M) = gr DO• H(M). The principal part of a differential operator of H-order m is given by the canonical grading map σm : DOm H(M) → U m(tH M). (2.3.1) 2.4 The tangent groupoid of a filtered manifold The tangent groupoid TH M of a filtered manifold M was introduced indepen- dently in [CP] and [vEY17]. It is a smooth one-parameter family of groupoids which deforms the pair groupoid to the osculating groupoid Algebraically, we have M × M ⇒ M TH M ⇒ M. TH M = (TH M ⊗ {0}) ⊔ (M × M × R×) ⇒ M × R. (2.4.1) The Lie groupoid structures on the two disjoint components in (2.4.1) are the standard ones, namely two elements of M × M × R× are composable if and only if their R× components are equal and their M × M components are composable in the pair groupoid, and two elements of TH M × {0} are composable if and only if their TH M components are. The difficulty lies in giving the correct global smooth structure on TH M. As usual, this is most easily achieved by beginning with the Lie algebroid tH M. Note that tH M is a deformation of TM to its associated graded tH M = gr(TM). Notation 2.4.1. When dealing with a space X which is fibred over R, we will write Xt for the fibre over t. Likewise if f is a function or section of a bundle over X we write f t for its restriction to Xt. Lemma 2.4.2. There is a unique C∞ structure on the disjoint union tH M = (tH M × {0}) ⊔ (TM × R×) which makes it into a smooth vector bundle over M × R with the the following prop- for any smooth vector field X ∈ Γ∞(Hm), the section X : M × R → tH M erty: 13 defined by X(x, t) =(tmX(x), σm(X(x)), if t 6= 0, if t = 0 (2.4.2) is a smooth section of tH M ։ M × R. With this smooth structure, tH M becomes a Lie algebroid with anchor and Lie bracket defined fibrewise on each tH Mt. Proof. See [vEY17] The sections X will play a major role in the sequel. We will write σm for the map X 7→ X. σm : Γ∞(Hm) → Γ∞(tH M); (2.4.3) Having defined a smooth Lie algebroid structure on tH M, there are many results which permit us to integrate it to a Lie groupoid [Deb01, CF03]. In this case, we already have smooth groupoid structures on the two components tH M × {0} and TM × R×, so the most easily applicable result from the liter- ature is a theorem of Nistor [Nis00] (with its correction in [BN03]). Modulo the technical issue of s-simply connectedness -- which we will sweep under the rug here -- we can immediately deduce the existence of a C∞-structure on TH M making it into a Lie groupoid with Lie algebroid tH M. See [vEY17] for full details. 2.5 The R× +-action The tangent groupoid TH M admits a crucial extra piece of structure, namely an action of R× + by Lie groupoid automorphisms, cf. [DS14]. We begin with an R× +-action on the bundle of osculating groups TH M, which generalizes the action of R× + on TM by homotheties. Definition 2.5.1. Let V = Lm∈N Vm be a finite dimensional N-graded vector space (or vector bundle). The dilations δλ of V are the linear automorphisms defined for each λ ∈ R× + by δλv = λmv for all v ∈ Vm. of tH M = + by Lie algebroid automorphisms. This integrates +-action by groupoid automorphisms on the osculating groupoid TH M, Lemma 2.5.2. Let M be a filtered manifold. The dilations (δλ)λ∈R× gr(TM) define a smooth action of R× to a smooth R× which we again denote by δλ. Definition 2.5.3. We define an R×-action α on TH M by + αλ :(x, y, t) 7→ (x, y, λ−1t), (x, ξ, 0) 7→ (x, δλξ, 0), t 6= 0, (x, y) ∈ M × M t = 0, ξ ∈ TH Mx. We also use the same notation αλ for the corresponding automorphisms of the Lie algebroid tH M: αλ :(x, v, t) 7→ (x, v, λ−1t), (x, ξ, 0) 7→ (x, δλξ, 0), t 6= 0, v ∈ TMx t = 0, ξ ∈ tH Mx. 14 Note that the sections X = σ(X) of tH M which we defined in Lemma 2.4.2 are homogeneous with respect to this action, in the sense that αλ∗X = λmX, for all λ ∈ R× +. (2.5.1) It follows that the maps αλ are smooth, both on the Lie algebroid tH M and the tangent groupoid TH M. Since the α-action on TH M is by Lie groupoid automorphisms, it induces an action on the spaces of r- and s-fibred distributions αλ∗ : E ′ r(TH M) → E ′ r(TH M), αλ∗ : E ′ s(TH M) → E ′ s(TH M). These actions restrict to the subalgebras of proper r- and s-fibred distribu- tions, and to the ideals C∞ p (TH M; Ωs) of properly sup- ported smooth r- and s-fibred 1-densities. p (TH M; Ωr) and C∞ 2.6 Pseudodifferential operators We can now give the characterization of pseudodifferential operators on a filtered manifold M. Let p ∈ E ′ r(TH M) be a properly supported r-fibred distribution on the tangent groupoid. Recall that we will write pt for the restriction of p ∈ E ′ r(TH M) to the fibre TH Mt. In particular, p1 ∈ E ′ r(M × M), p0 ∈ E ′ r(TH M). Thus p is a smooth deformation of a properly supported Schwartz kernel, semiregular in both variables, to a smooth family of properly supported dis- tributions on the osculating groups. Definition 2.6.1. Let M be a filtered manifold. An r-fibred distribution p ∈ E ′ r(TH M) is called • homogeneous of weight m if αλ∗p = λmp for all λ ∈ R× +; • essentially homogeneous of weight m if αλ∗p − λmp ∈ C∞ p (TH M; Ωr) for all λ ∈ R× +. The crucial definition is the following. Definition 2.6.2. Let p ∈ E ′ r(M × M) be a properly supported Schwartz kernel on M, semiregular in the first variable. We say p is an H-pseudodifferential kernel of order ≤ m if p = p1 for some properly supported p ∈ E ′ r(TH M) which is essentially homogeneous of order ≤ m. The associated Schwartz kernel operator P : C∞(M) → C∞(M) will be called an H-pseudodifferential operator. Remark 2.6.3. The proper support condition implies that P restricts to an op- erator on C∞ c (M). Moreover, we will shortly see that this definition forces p to be proper, so that P also extends to D′(M) and E ′(M) by Proposition 2.2.4. 15 Amongst the H-pseudodifferential kernels are those p = p1 where p ∈ E ′ r(TH M) is homogeneous of weight m, and not just essentially homogeneous. In this case, P is a differential operator. Proposition 2.6.4. Let p ∈ E ′ r(TH M) which is homogeneous of weight m if and only if p is the Schwartz kernel of a differ- ential operator of H-order ≤ m on M. r(M × M). Then p = p1 for some p ∈ E ′ r(G)(0) for the r-fibred distributions on a Lie groupoid Proof. Let us write E ′ r(G)(0) is isomorphic, as G with support on the unit space, and note that E ′ an algebra, to the universal enveloping algebra of the Lie algebroid AG, see [NWX99, vEY]. This isomorphism sends the homogeneous section X = σ(X) ∈ Γ∞(tH M) from Lemma 2.4.2 to a homogeneous r-fibred distribution r(TH M)(0). This proves that the Schwartz kernels of vector fields extend in E ′ to homogeneous elements of E ′ H(M) by multiplicativity. r(TH M). The result extends to DO• Conversely, note that the only orbits of α on TH M which are r-proper are the orbits contained in the unit space. Therefore, if p is r-properly supported and homogeneous of weight m, we must have supp(p) ⊂ G(0). The result follows. We likewise see that the singular support of an essentially homogeneous +-action, and so contained r-fibred distribution on TH M is invariant for the R× in TH M(0). Proposition 2.6.5. If p ∈ E ′ r(TH M) is essentially homogeneous of weight r, then sing-supp(p) ⊆ TH M(0). Therefore, H-pseudodifferential operators are pseudolocal. We have thus obtained the first two of the five desirable properties (Def- inition 1.2.2) for an algebra of pseudodifferential operators. The other three require some more serious analysis, which we will outline in Section 2.8. But let us pre-empt this by stating the relation of our H-pseudodifferential calculus with the classical calculus in the case of an unfiltered manifold. The proof depends upon the machinery to follow. Theorem 2.6.6. Let M be a manifold with the trivial filtration. Then Ψm H(M) is the space of Schwartz kernels of properly supported classical (polyhomogeneous step one) pseudodifferential operators on M of order ≤ m. 2.7 Principal symbols Definition 2.7.1. We introduce the notation Ψm tributions p ∈ E ′ neous of weight m. Thus Ψm H(M) for the set of r-fibred dis- r(TH M) that are properly supported and essentially homoge- H(M) = Ψm Next, we consider the restriction Ψm H(M)1. H(M)0. Lemma 2.7.2. If p and p′ ∈ Ψm C∞ p (TH M; Ωr). H(M) satisfy p1 = p′1, then p0 − p′0 ∈ 16 Proof. See [vEY]. The suggests the following definition. Definition 2.7.3. The space of principal cosymbols is defined as Σm H (M) := {a ∈ E ′ r(TH M)/C∞ p (TH M; Ωr) δλ∗a = λma ∀λ ∈ R× +}. To avoid clutter in the notation, we will typically use the same notation H(M) and an r-fibred distribution to denote a principal cosymbol class a ∈ Σm a ∈ E ′ r(TH M) which represents it. Definition 2.7.4. Lemma 2.7.2 shows that we have a well-defined map σm : Ψm H(M) → Σm H (M); p1 7→ p0. This is called the principal cosymbol map. Remark 2.7.5. We had previously used the notation σm to denote the principal part of a differential operator, σm : DOm H(M) = U m(TM) → U m(tH M). Under the isomorphism U (AG) ∼= E ′ 2.6.4, this becomes a map r(G)(0) from the proof of Proposition σm : E ′ r(M × M)(0) → E ′ r(TH M)(0). In this way, one can check that Definition 2.7.4 is consistent with the earlier notation. Proposition 2.7.6. For every m ∈ R, we have a short exact sequence 0 −→ Ψm−1 H (M) −→ Ψm H(M) σm−→ Σm H −→ 0. (2.7.1) Proof. Multiplication by the real parameter t yields a linear map ×t : Ψm−1 H (M) → Ψm H(M). The restriction of this map to t = 1 yields the inclusion on the left of (2.7.1). For exactness in the middle, suppose p ∈ Ψm(M) has σm(p) = 0. Then we can find p ∈ Ψm H(M) with p1 = p and p0 = 0. Since p is a smooth family of distributions on the r-fibres of TH M, t−1p is well-defined and is essentially homogeneous of weight m − 1, so p ∈ Ψm−1 H (M). Finally, a construction in groupoid exponential coordinates on TH M shows that any a ∈ Σm H(M) can be extended to a ∈ Ψm H(M) with a0 = a, which proves the surjectivity of σm. 2.8 Algebra structure, conormality, regularity Let p, q ∈ E ′ r(TH M). If p is homogeneous of weight m and q is homogeneous of weight n, then a simple calculation shows that their convolution product p ∗ q is homogeneous of weight m + n. It is tempting to say the same of essentially 17 homogeneous elements, but we are obstructed by the fact that C∞ is only a right ideal in E ′ r(TH M), not a two-sided ideal. p (TH M; Ωr) For this reason, the following technical lemma is crucially important for further progress. Its proof, which also serves to give the regularity results to follow, is one of main pieces of analysis in this work. We won't give complete details, but we will sketch the main ideas. For full details, see [vEY]. Lemma 2.8.1. Let p ∈ E ′ neous of weight m for some m, then p ∈ E ′ r,s(TH M). r(TH M) be properly supported. If p is essentially homoge- Sketch of proof. • We first linearize the problem. The groupoid exponen- tial allows us to identify a neighbourhood of the unit space TH M(0) with a neighbourhood of the zero section in tH M. By fixing a splitting ∼=→ TM, this can in turn be identified with a neigh- tH M = gr(TM) bourhood of the zero section in tH M × R. Under this identification, p is identified with a smooth family of compactly supported distributions on the fibres of the vector bundle tH M × R ։ M × R which is essentially homogeneous with respect to the R× +-action αλ : tH M × R → tH M × R; (x, ξ, t) 7→ (x, δλξ, λ−1t). • Now apply the fibrewise Fourier transform: r(tH M × R) → C∞(t∗ H M × R); p 7→ p, b : E ′ where t∗ H M ։ M is the dual bundle of tH M. Then p is a smooth function on t∗ H M × R with its own essential homogeneity property. Specifically, let δ′ λ denote the canonical dilations on t∗ H M with N-grading dual to that of tH M, and let βλ denote the action on t∗ H M × R given by (x, η, t) 7→ (x, δ′ H M × R → t∗ H M × R; λη, λt). βλ : t∗ Then for all λ ∈ R× + we have λ p − λm p ∈ Sr(t∗ β∗ H M × R), where Sr(t∗ H M × R which are Schwartz-class on each r-fibre (see [vEY] for the precise definition). H M × R) denotes the smooth functions on t∗ • Essential homogeneity of smooth functions near infinity is better be- haved than essential homogeneity of distributions near 0. Specifically, a bundle version of [Tay, Lemma 2.2] shows that, outside any neighbour- hood of the zero section, p is equal to a genuinely homogeneous smooth function plus a function of rapid decay on the fibres. As a consequence, we obtain bounds on all derivatives of p(x, η, t) in terms of powers of kηk. • The above bounds on p imply that the wavefront set of p is conormal to the unit space. This is sufficient to deduce that p is both r- and s-fibred, thanks to [LMV17, Proposition 7]. 18 Corollary 2.8.2. Convolution of r-fibred distributions induces an R-filtered algebra structure on Ψ• H(M). H(M) ⊆ E ′ r,s(TH M). Since C∞ r,s(TH M), the convolution product induces maps Ψm Proof. We have Ψm p (TH M; Ωr) is a two-sided ideal in E ′ H(M) → Ψm+n H (M) for every m, n ∈ R. Restriction to t = 1 gives the product on Ψ• H(M). Going further, a careful analysis of the bounds on p allows us to deduce H(M) × Ψn the degree of regularity of p on the unit space of TH M. Lemma 2.8.3. Let N be the depth of the filtration on TM, and let dH = ∑N k=1 k dim(Hk/Hk−1) be the homogeneous dimension of tH M. Then Ψ−dH−kN−1 H (M) ⊆ Ck p(M × M; Ωr). Proof. See [vEY]. Corollary 2.8.4. We haveTm∈R Ψm H(M) = Ψ∞(M), the algebra of smooth kernels. This is Property (3) from our list in Definition 1.2.2. We also obtain Prop- erty (4), the existence of asymptotic limits. Corollary 2.8.5. Given any sequence (pi)i∈N of H-pseudodifferential kernels, with pi ∈ Ψm−i H (M) for each i ∈ N, there exists p ∈ Ψm H(M) such that for any l ∈ N, p − l ∑ i=0 pi ∈ Ψm−l−1 H (M). Proof. Let pi ∈ Ψm−i ciently large we have pi ∈ Ck(i) After modifying each pi by an element of C∞ pi vanishes on the unit space TH M(0) to order k(i). H (M) with pi1 = pi. Lemma 2.8.3 implies that for i suffi- p (TH M) for some k(i) with k(i) → ∞ as i → ∞. p (TH M; Ωr), we may assume that Then we have tipi ∈ Ψm i=0 tipi ∈ tl+1Ck(l+1) k(i) on TH M(0). An analogue of Borel's Lemma yields p ∈ E ′ that p − ∑l is homogeneous of weight m modulo C∞ mogeneous of weight m modulo Ck p (TH M; Ωr). That is, p ∈ Ψm C∞ Ψm−l−1 H H(M), vanishing to order i on TH M0 and to order r(TH M) such i=0 tipi p (TH M; Ωr), it follows that p is ho- p(TH M; Ωr) for every k, and hence modulo i=0 tip) ∈ H(M). Put p = p1. Then t−l−1(p − ∑l H for all l ≫ 0. Since the finite sum ∑l . This completes the proof. i=1 pi ∈ Ψm−l−1 (M), so p − ∑l p 2.9 Hypoellipticity Given the above structure, the following definition is the natural abstract ana- logue of ellipticity for H-pseudodifferential operators. Definition 2.9.1. An H-pseudodifferential kernel p ∈ Ψm elliptic if its principal cosymbol σm(p) admits a convolution inverse in Σ−m H(M) is called H- H (M). 19 By the surjectivity of the principal cosymbol map, this means that there exists q ∈ Ψ−m such that p ∗ q − I ∈ Ψ−1 H (M). Thus, H-elliptic operators satisfy axiom (5) of Definition 1.2.2. The following is now an immediate consequence of Theorem 1.2.3. Theorem 2.9.2. Any H-elliptic operators on a filtered manifold M is hypoelliptic. The H-ellipticity condition is difficult to verify in practice. The correct analogue for ellipticity on a filtered manifold should be the Rockland property [Roc78, HN79, Mel82]. The relationship between these two properties has been clarified by Dave and Haller [DH17], based on results of Folland-Stein [FS82] and Ponge [Pon08]. We begin by recalling the definition. Definition 2.9.3. A differential operator P ∈ DOm H(M) on a filtered manifold M is Rockland if, for every x ∈ M and for every non-trivial irreducible unitary representation π of the osculating group tH Mx on a Hilbert space Hπ, the operator π(σm(P)) is injective on the space of smooth vectors in Hπ. We will say P is two-sided Rockland if both P and its transpose Pt are Rockland. Theorem 2.9.4. [DH17] If P ∈ DO(M) two-sided Rockland then it is H-elliptic. Therefore, every Rockland operator on a filtered manifold is hypoelliptic. It is also shown in [DH17] that there is a Sobolev theory adapted to the pseudodifferential calculus Ψ• H. It follows that if M is compact without bound- ary, then H-elliptic operators are Fredholm operators between the appropriate Sobolev spaces [DH17, Corollary 3.28]. 20 Chapter 3 Pseudodifferential operators on multifiltered manifolds and equivariant index theory for complex semisimple groups Flag manifolds of semisimple Lie groups G are important examples of filtered manifolds. In differential geometry, they are the model spaces of a class of spaces called parabolic geometries, see e.g. [ CS09], which generalize CR man- ifolds and play an important role in twistor theory [BE89], amongst others. Our interest, however, will come from equivariant index theory. Kasparov and Julg [JK95, Jul02] showed that when G = SU(n, 1) or Sp(n, 1), one can use natural subelliptic operators on the flag manifold of G to define a G-equivariant K-homology class which represents the notorious element γ ∈ KG(C, C) at the heart of the Baum-Connes conjecture. On the other hand, when G has real rank greater than 1, a result of Puschnigg [Pus11] implies that any construction based on the filtered calculus of Chapter 2.1 can only result in multiples of the trivial class. Nonetheless, at least in the case of G = SL(3, C), it is possible to represent the γ element as an equivariant K-homology class for the flag manifold by using the Bernstein-Gelfand-Gelfand complex [Yun10, Yun11]. The Bernstein- Gelfand-Gelfand (BGG) complex is a G-equivariant differential complex built from longitudinally elliptic differential operators along the fibres of the nu- merous canonical fibrations of the flag manifold. The analysis in those articles was based on Gelfand-Tsetlin theory for the representations of SU(3), but these methods generalize badly, since analogues of Gelfand-Tsetlin theory for Lie algebras other than type An are difficult. Here we shall focus on the article [Yun13], which describes a rudimentary pseudodifferential theory adapted to the analysis of longitudinal differential 21 operators on manifolds with multiple foliations. It is rudimentary in the sense that it essentially only distinguishes between operators of order 0 and opera- tors of negative order. Nevertheless, this suffices to distinguish bounded and compact operators, and their analogues for foliations, which will be enough for application to index theory. 3.1 Complex flag manifolds Throughout, G will denote a connected, simply connected complex semisim- ple Lie group. We will use the following notation: • g is the Lie algebra of G • h is a Cartan subalgebra • ∆, ∆+, Σ denote the sets of roots, positive roots, and simple roots, re- spectively. • ρ = 1 2 ∑β∈∆+ β is the half-sum of the positive roots, • P and Q are the integral weight and root lattices, respectively. • k, a, n are the components of the Iwasawa decomposition, with n = Lα∈∆+ gα the sum of the positive root spaces. • ¯n =Lα∈∆+ g−α is the opposite nilpotent subgroup. • b = h ⊕ n and ¯b = h ⊕ ¯n are the standard upper and lower Borel sub- groups. A connected Lie subgroup of G will always be denoted by the same letter as its Lie algebra, in upper case. The standard parabolic Lie subalgebras of g are in bijection with subsets I ⊆ Σ of the simple roots. To make this precise this, we introduce the notation hIi = {β ∈ ∆+ β = ∑ α∈I nαα for some nα ≥ 0} for the set of positive roots in the span of I. Then we put1 pI = b ⊕ ∑ α∈hIi g−α. In particular, pΣ = g and p∅ = b. The homogeneous spaces are the flag manifolds of G. XI = G/PI 1 Unfortunately, our notation here is the opposite of standard one, in that we are writing pI for what would often be denoted pΣ\I. We choose this convention to be consistent with [Yun11, Yun13]. 22 In the case where I = {α} is a singleton, we will write pα, Pα, Xα, etc. for p{α}, P{α}, X{α}, etc. We will also write X = X∅ = G/B for the full flag manifold. The various flag manifolds are connected by smooth G-equivariant fibra- tions XJ ։ XI whenever J ⊆ I. In particular we have G-equivariant maps qI : X ։ XI. Thus, for each I ⊆ Σ there is a foliation of the full flag manifold by the fibres of qI. We will denote the tangent bundle to the fibres by FI = ker dqI ⊆ TX . 3.2 Motivation: The BGG complex As mentioned above, this chapter will be dedicated to describing the results of [Yun13], concerning longitudinal pseudodifferential operators on manifolds with multiple foliations. But this work was entirely motivated by the equivari- ant index theory of the differential Bernstein-Gelfand-Gelfand complex, which consists of longitudinal differential operators along the various fibrations of the flag manifold just described. For context, therefore, we will start with a very rapid overview of the Bernstein-Gelfand-Gelfand complex. Let µ ∈ P be an integral weight of g, and let λ ∈ a∗ C = h∗. These exponen- tiate to a character eµ of the compact torus M = K ∩ H and a character eλ of A (not generally unitary), via the formulas eµ(exp(X)) := eµ(X), eµ(exp(Y)) := eµ(Y), X ∈ m, Y ∈ a. We obtain a character of the Borel subgroup B = MAN by χµ,λ(man) = eµ(m)eλ(a). Definition 3.2.1. We write Eµ,λ = G ×B Cµ,λ+2ρ for the G-equivariant line bundle over the flag manifold G/ ¯B which is induced from the shifted character χµ,λ+2ρ. This means that smooth sections of Eµ,λ are defined by Γ∞(Eµ,λ) = {ξ ∈ C∞(G) ξ(gb) = χµ,λ+2ρ(b−1)ξ(g) ∀g ∈ G, b ∈ B}. (3.2.1) The space of L2-sections of Eµ,λ, which we denote by L2(Eµ,λ), is the com- pletion of Γ∞(Eµ,λ) with respect to the K-invariant inner product hξ, ηi =ZK ξ(k)η(k) dk. 23 The left regular representation of G on L2(G) restricts to a representation on L2(Eµ,λ) called the principal series representation of parameter (µ, λ) ∈ P × h∗. It is unitary when λ ∈ ia∗, see e.g. [Kna01]. The Bernstein-Gelfand-Gelfand complex is a complex of G-equivariant dif- ferential operators between the section spaces of certain very particular line bundles Eµ,λ over X . These operators are obtained by the right action of par- ticular elements of U (g). Such operators, if well-defined, obviously commute with the principal series actions of G. The details are as follows. Let W denote the Weyl group of G. We write ℓ(w) for the length of w ∈ W, i.e., the minimal length as a word in the simple reflections. We recall that the Bruhat graph is a directed graph with vertices W and edges w → w′ whenever ℓ(w′) = ℓ(w) + 1 and w′ = sαw for some simple reflection sα; see, e.g., [Hum08]. Note that, with the parametrization of Definition 3.2.1, the trivial bundle is E0,−2ρ. We have an inclusion C ֒→ Γ∞(X0,−2ρ) of the trivial G-module as constant functions. The Bernstein-Gelfand-Gelfand complex is a resolution of this inclusion by direct sums of principal series representations. Theorem 3.2.2. Let w, w′ ∈ W be connected by an edge w → w′ in the Bruhat graph. There exists X ∈ U (g) such that the right regular action of X defines a G-equivariant differential operator X : Γ∞(E−wρ+ρ,−wρ−ρ) → Γ∞(E−w′ρ+ρ,−w′ρ−ρ). (3.2.2) In particular, if w′ = sαw where sα is the simple reflection associated to α ∈ Σ, then X = En α where Eα ∈ g is the simple root vector of weight α and n ∈ N is determined by wρ − w′ρ = nα. With an appropriate choice of signs, these operators X form a resolution of the trivial G-module C ֒→ Γ∞(E0,−2ρ) → Mℓ(w)=1 Γ∞(E−wρ+ρ,−wρ−ρ) → Mℓ(w)=2 Γ∞(E−wρ+ρ,−wρ−ρ) → · · · (3.2.3) Remark 3.2.3. With other choices of parameters (µ, λ) ∈ P × h∗, one can ob- tain more general Bernstein-Gelfand-Gelfand complexes giving resolutions of other finite dimensional G-modules, see e.g. [BE89]. The key observation for what follows is that the differential operators En α corresponding to simple edges in the Bruhat graph are longitudinally elliptic differential operators along the fibres of the fibration X ։ Xα. Therefore, in order to convert the BGG complex into an equivariant KK-cycle, we need some kind of pseudodifferential theory which simultaneously contains longitudinal pseudodifferential operators along each of these canonical fibrations. This is the goal of [Yun13]. Remark 3.2.4. The Dolbeault complex of X is also a G-equivariant resolution of the trivial module, but being elliptic it cannot easily be converted into a nontrivial KK-cycle, because of the previously mentioned result of Puschnigg [Pus11]. The main problem is that the action of G on the spinor bundle of X 24 is not isometric, nor even conformal. For the importance of conformality in equivariant KK-theory, compare [Kas84, JK95, Jul02]. The advantage of the BGG complex is that the induced bundles are di- rect sums of line bundles, each of which admits a Hermitian metric which is conformal for the G-action and with explicit formulas for the conformality constants. This is a key point in [Yun11]. 3.3 Longitudinal pseudodifferential operators Definition 3.3.1. Let E, E ′ be vector bundles over the full flag variety X of G. For each I ⊆ Σ, we write Ψm I (E, E ′) for the set of longitudinal pseudodifferen- tial operators T : Γ∞(E ) → Γ∞(E ′) which are of order ≤ m (polyhomogeneous step 1) along the leaves of the foliation FI. If E = E ′ we will simplify the notation to Ψ−m (E ). But in general, we will abbreviate all of these by writing Ψm I , where the bundles involved can be inferred from the context. In [Yun13] we used the notation Ψm(FI ) instead of Ψm I . For each I, the elements of Ψ0 I and thus the norm closure Ψ0 of the longitudinal smoothing operators is an ideal in Ψ0 Ψm I (X ) is a C∗-algebra. The norm closure Ψ−∞ I (X ) extend to bounded operators on L2(X ), (X ) I (X ) which contains I (X ) for all m < 0. It is sometimes denoted C∗(FI). In particular, if I = Σ then the foliation of X by FΣ consists of a single I leaf, and we have Ψm σ (X ) = Ψm(X ). Therefore Ψ−∞(FΣ) = K(L2(X )). If I 6= J, the product of an element of Ψm I (X ) and one of ΨJ(X ) is typically not a pseudodifferential operator in any reasonable sense. In this respect, the C∗-closures are much better behaved. Theorem 3.3.2 ([Yun13]). For any I, J ⊆ Σ, we have Ψ−∞ I .Ψ−∞ J ⊆ Ψ−∞ I∪J . In particular, if I1, . . . , In ⊆ Σ withSn k=1 Ik = Σ, and if Pk is a longitudinal pseu- dodifferential operator of negative order along the leaves of FIk for each k = 1, . . . n, then the product P1 . . . Pn is a compact operator. This theorem is a key point for the construction of an equivariant K- homology class from the Bernstein-Gelfand-Gelfand complex, as we shall show in Section 3.7. But first we will describe the proof of Theorem 3.3.2 in the next three sections. 3.4 Multifiltered manifolds Let us fix some notation. Fix r ∈ N×, usually signifying the rank of G. We denote multi-indices by a = (a1, . . . , ar) ∈ Nr. We write a ∨ b and a ∧ b for the entry-wise maximum and entry-wise minimum, respectively, of two multi-indices. 25 Definition 3.4.1. Let V be a finite dimensional vector space (or a finite di- mensional vector bundle). An r-multifiltration on V will mean a family of subspaces (or subbundles) Va ≤ V indexed by a ∈ Nr satisfying • V0 = 0 and Vm = V for some m ∈ Nr, • Va ∩ Vb = Va∧b for all a, b ∈ Nr. Definition 3.4.2. An r-multifiltered manifold is a manifold M equipped with a r-multifiltration of its tangent bundle by subbundles Ha (a ∈ Nr) such that [Γ∞(Ha), Γ∞(Hb)] ⊆ Γ∞(Ha+b), ∀a, b ∈ Nr. Example 3.4.3. The Cartesian product of two manifolds, M = M1 × M2 is a 2-multifiltered manifold, with H(1,0) = pr∗ 1TM1, H(0,1) = pr∗ 2TM2, H(1,1) = TM, where pri : M ։ Mi is the coordinate projection. This multifiltered structure is relevant to the study of bisingular operators [Rod75, Boh15] and the exterior product in analytic K-homology. Example 3.4.4. The full flag manifold X of a complex semisimple Lie group G is an r-multifiltered manifold, where r is the rank of G. To see this, recall that the tangent space TX is isomorphic to the induced bundle G ×B g/b, where g/b is equipped with the adjoint action of B. A G-equivariant multifiltration on TX can therefore be determined by a multifiltration on the B-module g/b. Let α1, . . . , αr be an enumeration of the simple roots, and for each a = (ai, . . . , ar) ∈ Nr put Va = M(gβ β = biαi with bi ≥ −ai ∀i)!, b r ∑ i=1 and Ha = G ×BVa. This makes X into an r-multifiltered manifold. This multifiltration on X is related to the tangent bundles FI of the canon- ical fibrations X ։ XI as follows. Multi-indices a are in bijection with the positive elements of the root lattice, by associating a to ∑i aiαi. Then FI = Hb, where b corresponds to the maximal β ∈ ∆+ with support in I. 3.5 Locally homogeneous structures In the proof of Theorem 3.3.2 we will take advantage of the fact that multi- filtration of the flag manifold is locally diffeomorphic to a multifiltration of ¯N induced by a family of nilpotent subgroups ¯NI. Let us begin with some abstract definitions. We use the notation ǫi = (0, . . . , 1, . . . , 0) for the multi-index whose only nonzero coefficient is the ith, which is 1. 26 Definition 3.5.1. An r-multigraded nilpotent Lie algebra will mean a finite di- mensional nilpotent Lie algebra n with an Nr-grading n = Ma∈Nr na that is compatible with the Lie bracket, such that n0 = 0 and n is generated as a Lie algebra by the subspaces nǫi. The r-multigrading on n• induces an r-multifiltration, denoted n•, by putting na = Mb≤a nb. Let N be the associated simply connected nilpotent Lie group and identify TN = N × n via left translations. Then TN inherits an r-multifiltration, and so N is a multifiltered manifold. Definition 3.5.2. Let n be a multigraded Lie algebra. A multifiltered manifold M will be called locally homogeneous of type n if there exists an atlas of local charts ϕ : U → M from open subsets of U ⊆ N satisfying ϕ∗(Ha) = U × na, ∀a ∈ Nr, where the Ha are the subbundles defining the multifiltration of M, as in Defi- nition 3.4.2. Of course, we are only really interested in this structure because it encap- sulates the structure of flag manifolds. Example 3.5.3. Let ¯n be the nilpotent radical of the lower Borel subgroup of g. Then ¯n is an r-multigraded lie algebra via ¯na = g−α, where, α = ∑i aiαi. We claim that the multifiltration of X in Example 3.5.3 is locally homoge- neous of type ¯n. For each x ∈ G, we can define a chart ϕx : ¯N → X ; ϕx( ¯n) = x ¯nB, which is a diffeomorphism of ¯N onto a left translate of the big Bruhat cell in X . From Example 3.5.3, we see that ϕ∗ x(Ha) = ¯N × ¯na, as desired. The following definition, which we will not make much use of, allows one to generalize the foliations FI of the flag manifold to any locally homogeneous multifiltered manifold. Definition 3.5.4. Let n be an r-multigraded Lie algebra. For any set I ⊆ {1, . . . , r}, we will denote by nI the Lie subalgebra of n generated byLi∈I nǫi. The associated Lie subgroup is denoted NI. If M is any locally homogeneous multifiltered manifold of type n, then the foliation of N by cosets of NI induces, via the local charts, a foliation of M. We denote the tangent bundle to the leaves by FI. 27 3.6 Products of longitudinally smoothing operators The point of the above abstract nonsense is merely to show that Theorem 3.3.2 can be reduced to a theorem about smoothing operators along the coset foliations of a family of subgroups of a nilpotent group. Indeed, by using a partition of unity subordinate to the charts ϕ of Definition 3.5.2, and by putting H1 = NI, H2 = NJ and H = NI∪J it suffices to prove the following result. Theorem 3.6.1. Let H be a nilpotent Lie group and let H1, H2 be connected sub- groups of H such that their Lie algebras h1, h2 generate the Lie algebra h of H. Let A1, A2 : C∞ c (H) be compactly supported longitudinally smoothing oper- ators along the coset foliations of H by H1 and H2, respectively. Then A1A2 extends to a compact operator on L2(H). Sketch proof. The longitudinally smoothing operators Ai are given by a convo- lution formula of the following form: c (H) → C∞ Aiu(x) =Zh∈Hi ai(x, h)u(h−1x) dh, u ∈ C∞ c (H), (3.6.1) for some ai ∈ C∞ c (H × Hi). To prove that A1A2 is compact, it suffices to prove that A∗ 1 A1A2 is com- 1 A1A2)n is compact for some n ∈ N. Let us take n pact, or indeed that (A∗ large enough that the space of products 2 A∗ 2 A∗ H2H1H2H1H2 . . . H1H2 2n+1 terms {z } has positive measure in H. Repeated application of Equation (3.6.1) shows that (A∗ 1 A1A2)n has the form 2 A∗ 1 A1A2)nu(x) (A∗ 2 A∗ =ZH2 × H1 × · · · × H2 } {z 2n+1 terms a(x, h1, h2, . . . , h2n+1)u((h1h2 · · · h2n+1)−1x) dh1 dh2 . . . dh2n+1 for some smooth function a ∈ C∞ c (H × H2 × H1 × · · · × H2). The product map φ : H2 × H1 × · · · × H1 → H is real analytic, so its deriva- tive is surjective outside a set of measure zero. The proof is then completed by the following Lemma, which is essentially a consequence of the implicit function theorem. z 2n+1 terms } { Lemma 3.6.2. ([Yun13]) Let M be a smooth manifold, φ : M → H a smooth map whose derivative is surjective outside a set of zero measure, and let a ∈ C∞ c (M × H). Then the operator A defined by Au(x) =ZM a(x, m)u(φ(m)−1x) dx is a compact operator on L2(H). 28 3.7 Application: The KK-class of the BGG complex of rank-two complex semisimple groups To conclude this chapter, we return to the motivating application of construct- ing a G-equivariant K-homology class for the full flag manifold from the Bernstein-Gelfand-Gelfand complex. At present, this construction has only been completely carried out for the group G = SL(3, C) (although unpub- lished work shows that the argument can be extended to the group Sp(4, C) of type B2 / C2). Since we would like to point out the technical point which remains to be resolved for the construction in general, we will begin this dis- cussion in the generality of an arbitrary connected, simply connected, complex semisimple Lie group G. Let w → w′ be a an edge in the Bruhat graph with w′ = sαw for some In this case, we will call it a simple edge. According to simple root α ∈ Σ. Theorem 3.2.2, we obtain a G-invariant BGG operator α : L2(E−wρ+ρ,−wρ−ρ) → L2(E−w′ρ+ρ,−w′ρ−ρ) En for some n, and this is a longitudinally elliptic differential operator along the leaves of Fα. We will replace this operator by its bounded operator phase Tw→w′ := ph(En α ) : L2(E−wρ+ρ,0) → L2(E−w′ρ+ρ,0), (3.7.1) acting between unitary principal series. We call this the normalized BGG opera- tor. It belongs to Ψ0 α and satisfies w→w′Tw→w′ − I ∈ Ψ−1 T∗ α , [a, Tw→w′] ∈ Ψ−1 α , gTw→w′g−1 − Tw→w′ ∈ Ψ−1 α , ∀a ∈ C∞(X ), ∀g ∈ G, where a ∈ C∞(X ) is acting on Γ∞(E−wρ+ρ,0) and Γ∞(E−w′ρ+ρ,0) by multiplica- tion, and g ∈ G is acting by the unitary principal series representations given in (3.7.1). For the rank-two complex semisimple groups, the system of normalized BGG operators are indicated in Figure 3.1. Here α and β are the two simple roots, with β the longer root in the cases B2 / C2 and G2. The arrows pointing α ) for some n, and those pointing south-east are ph(En north-east are ph(En β) for some n. We have not yet defined the horizontal arrows (associated to the non-simple edges). In order to continue we will need to slightly enlarge the C∗-algebras Ψ−∞ . Let E = Lw∈W E−wρ+ρ,−wρ−ρ denote sum of the line bundles in the BGG complex, with Z/2Z-grading according to the parity of ℓ(w). Put also H = L2(E ). Definition 3.7.1. For each I ⊆ Σ, we define KI to be the hereditary C∗- subalgebra of L(H) generated by Ψ−∞ (E ). We also define A to be the si- multaneous multiplier algebra of the KI, I I A = {A ∈ L2(H) AKI ⊆ KI, ∀I ⊆ Σ}. Finally, we put JI = KI ∩ A. 29 A1 ⊔ A1 : L2(E−α,0) L2(E0,0) 7♦♦ ❖❖ L L2(E−β,0) ❘❘ L2(E−α−β,0) 6♠♠ A2 : L2(E−α,0) L2(E−2α−β,0) L2(E0,0) 7♦♦ ❖❖ L ■■■■■■■■ :✉✉✉✉✉✉✉✉ )❙❙ L2(E−2α−2β,0) 5❦❦ L2(E−β,0) / L2(E−α−2β,0) B2/C2 : L L L2(E−α,0) L2(E−3α−β,0) L2(E−4α−2β,0) L2(E0,0) 7♦♦ ❖❖ L ■■■■■■■■ :✉✉✉✉✉✉✉✉ ❑❑❑❑❑❑❑❑❑ 9sssssssss L )❚❚ L2(E−4α−3β,0) 5❥❥ L2(E−β,0) / L2(E−α−2β,0) / L2(E−3α−3β,0) G2 : L2(E−α,0) L2(E−4α−β,0) L2(E−6α−2β,0) L2(E−9α−4β,0) L2(E−10α−5β,0) L2(E0,0) 7♦♦ ❖❖ L ■■■■■■■■ :✉✉✉✉✉✉✉✉ L ❑❑❑❑❑❑❑❑ 9ssssssss L ❑❑❑❑❑❑❑❑❑ 9sssssssss L ▲▲▲▲▲▲▲▲▲ 9rrrrrrrrr L L2(E−β,0) / L2(E−α−2β,0) / L2(E−4α−4β,0) / L2(E−6α−5β,0) / L2(E−9α−6β,0) )❚❚ 5❦❦ L2(E−10α−6β,0) Figure 3.1: The Bernstein-Gelfand-Gelfand complexes of the rank 2 complex semisim- ple groups. One advantage of the algebras KI over Ψ−∞ is that they are nested: KI ⊆ KJ whenever J ⊆ I. Therefore, the algebras JI (I ⊆ Σ) form a nested family of ideals of A. I The construction of the BGG class in K-homology will take place entirely within the C∗-algebra A. It is easy to check that multiplication operators by functions on X belong to A, as do the principal series actions of g ∈ G. We will also need that the normalized BGG operators belong to A. The results of [Yun11] give this result only for the simply-laced case. Lemma 3.7.2. Let G be a simply-laced complex semisimple Lie group. For any α, β ∈ Σ we have ph(Eβ).Kα ⊂ Kα. It follows that the normalized BGG operators belong to A. Remark 3.7.3. The proof of Lemma 3.7.2 depends only on the relative position of two roots α and β, and so its proof reduces to the rank 2 groups. For type A1 ⊔ A1 the result is elementary. For A2 it was proven in [Yun11] using Gelfand-Tsetlin theory. In unpublished work we have also generalized this to type B2, which shows that the lemma can be extended to all groups other than G2. For type G2 it remains unproven, although obviously one expects it to be true as well. Next we define normalized BGG operators associated to the non-simple edges. 30 ( ( 7 ' ' 6 $ $ / / ) 7 ' ' : / 5 $ $ / / % % / / ) 7 ' ' : / 9 / 5 $ $ / / % % / / % % / / % % / / ) 7 ' ' : / 9 / 9 / 9 / 5 Lemma 3.7.4. Suppose G is a group for which Lemma 3.7.2 holds. Then one can associate to each non-simple edge w → w′ in the Bruhat graph a bounded operator Tw→w′ : L2(E−wρ,0) → L2(E−w′ρ,0) (3.7.2) such that the operators Tw→w′ form a complex modulo the sum of the ideals ∑α Jα. Proof. We compare the normalized BGG operators associate to the simple edges, Equation (3.7.1), with the same operators shifted by −ρ: Iw→w′ := ph(En α ) : L2(E−wρ,0) → L2(E−w′ρ,0). (3.7.3) The operators Iw→w′ are actually intertwiners of irreducible unitary principal series representations (see, e.g. [Duf75]) so by Schur's Lemma the diagram of the operators Iw→w′ will commute on the nose. Using a partition of unity, we can find sections f1, . . . , fn ∈ Γ∞(E−ρ) such that ∑n i=1 fi fi = 1. By considering principal symbols, we see that ∑ i fiIw→w′ fi! − Tw→w′ ∈ Ψ−1 α ⊂ Kα (3.7.4) for every edge w → w′ associated to a simple root α. Let us define the nor- malized BGG operators associated to non-simple edges w → w′ by Tw→w′ := ∑ i fiIw→w′ fi. Note that Iw→w′ can be written as a composition of intertwiners associated to simple roots (or their inverses). Using Lemma 3.7.2, we infer that the diagram of normalized BGG operators Tw→w′ commutes modulo ∑α Jα. By introducing signs, as in [BGG75], we obtain a complex modulo ∑α Jα. Until this point, the entire construction works for an arbitrary simply laced G, and using Remark 3.7.3 it can be generalized to any complex semisimple group with no factor of type G2. The final step, however, is an analogue of the Kasparov product, and this has only been achieved in rank 2. We must therefore restrict our attention now to the group SL(3, C). We also consider the group SL(2, C) × SL(2, C) in order to make a comparison with the classical Kasparov product. Theorem 3.3.2 implies that Jα.Jβ ⊆ K(H) for the rank-two groups. The Kasparov Technical Theorem yields a pair of bounded self-adjoint opera- tors Nα, Nβ ∈ L(H) that are diagonal with respect to the direct sum H = Lw∈W L2(Ew·(0,−2ρ)), that commute modulo compacts with the normalized BGG operators, the multiplication action of C(X ) and the principal series rep- resentations of G on each L2(Ew·(0,−2ρ)), and that satisfy NαJα ⊆ K(H), NβJβ ⊆ K(H) and N2 α + N2 β = 1. (3.7.5) We multiply each of the normalized BGG operators by a power of Nα and Nβ as indicated in Figure 3.2 where the symbol T is shorthand for the normalized 31 A1 ⊔ A1 : A2 : L2(E0,0) L2(E0,0) Nα T 6 6♥♥♥ PPP Nβ T Nα T 6 6♥♥♥ PPP Nβ T L2(E−α,0) Nβ T ❙❙❙ L L L2(E−α−β,0) 5❧❧❧ Nα T L2(E−β,0) Nα Nβ T L2(E−α,0) ❑❑❑❑❑❑ ssssss L2(E−2α−β,0) Nβ T ❚❚❚ N2 α T N2 β T / L2(E−α−2β,0) 4❥❥❥ Nα T L L2(E−β,0) Nα Nβ T L2(E−2α−2β,0) Figure 3.2: Modification of the normalized BGG operators by the Kasparov Technical Theorem. BGG operator Tw→w′ associated to the given edge. Finally, we define F to be the sum of all these operators and their adjoints. A direct check using the properties of Nα, Nβ and the normalized BGG operators shows that F2 ≡ 1 modulo K(H). We arrive at the following result. Theorem 3.7.5. Let G = SL(2, C) × SL(2, C) or SL(3, C) and let X be the full flag manifold of G. The operator F described above defines a G-equivariant K-homology class for X , which we denote by [BGG] ∈ KKG(C(X ), C). Its image under the forgetful map to KKK(C, C) is the class of the trivial representation, and its image in KKG(C, C) is Kasparov's γ element. For the group G = SL(2, C) × SL(2, C), the flag manifold is X = CP1 × CP1 and the four induced bundles in Figure 3.2 are isomorphic to L2(E−iα−jβ) ∼= L2(Λ0,iCP1 ⊗ Λ0,jCP1), i, j ∈ {0, 1}. In other words, Figure 3.2 for A1 ⊔ A1 depicts precisely the exterior Kasparov product of two copies of the Dolbeault-Dirac class for CP1. Thus the BGG class for SL(3, C) is literally a generalization of the Kasparov product. If we allow ourselves to use Lemma 3.7.2 for the group of type B2, as indicated in Remark 3.7.3, then we can obtain Theorem 3.7.5 also for the group Sp(4, C) of type B2 / C2 by modifying the normalized BGG complex as in Figure 3.3. B2/C2 : L2(E−α,0) L2(E0,0) Nα Nβ T Nα Nβ T Nα T 6 6♥♥♥ PPP Nβ T ❑❑❑❑❑❑ ssssss L L2(E−β,0) L2(E−3α−β,0) N2 α T N2 β T / L2(E−α−2β,0) L ▼▼▼▼▼▼ qqqqqq L2(E−4α−2β,0) Nβ T ❚❚❚ N2 α T N2 β T / L2(E−3α−3β,0) 4❥❥❥ Nα T L ▼▼ Nα Nβ T Nα Nβ T L2(E−4α−3β,0) Figure 3.3: Modification of the normalized BGG operators in type B2 / C2. 32 ) ) ( ( 5 % % / / * ( ( 9 9 / 4 % % / / & & / / * ( ( 9 9 / 8 8 / 4 Chapter 4 Pseudodifferential operators on quantum flag manifolds Quantum groups and Connes' noncommutative geometry arose independently, but it was soon realized that algebras of functions on quantized Lie groups and their homogeneous spaces should, morally, be examples of noncommuta- tive manifolds. Unfortunately, making this precise has turned out to be much more difficult than expected. There is no definitive list of axioms of a noncommutative manifold, rather a list of desirable properties which should be adapted to suit the various examples. The basic object, however, is always a spectral triple (A, H, D) where A is a ∗-algebra of bounded operators on the Hilbert space H, and D is an unbounded self-adjoint operator on H satisfying axioms designed to mimic the properties of a Dirac-type operator. This needs to be supplemented, at minimum, with some version of the regularity axiom [Con94, CM95]. See below for definitions. In practice, regularity amounts to the existence of a pseudodifferential cal- culus, in which D is elliptic. This was made into an explicit theorem by Higson [Hig04] and Uuye [Uuy11]. Not only does regularity imply the existence of an abstract pseudodifferential calculus, but proving the regularity axiom is typically achieved by exhibiting that pseudodifferential calculus. In this chapter, we shall discuss the noncommutative geometry of two classes of quantum homogeneous spaces by studying analogues of pseudod- ifferential operators upon them. We first look at the quantum projective spaces CPn q , chosen as a case where things work relatively easily. Quantum projective spaces admit spec- tral triples which are q-deformations of the classical Dirac-type operators ([Krä04],[DD10]). They fail to be regular, but only very mildly. In the case of the Podle´s sphere CP1 q, Neshveyev and Tuset produced a twisted algebra of pseudodifferential operators and thus proved a local index formula for the Dabrowski-Sitarz spectral triple [NT05]. We will present some recent work with M. Matassa giving the first steps towards a similar result for general CPn q . We will generalize the pseudodifferential calculus of Connes and Moscovici to 33 twisted spectral triples, in a sense slightly more general than [CM08, Mos10]. With this definition, the q-deformed Dirac operators of [KTS15] yield twisted regular spectral triples. In a second example, we consider the quantized full flag manifold Xq of Kq = SUq(3). Here, the usual axioms of a noncommutative manifold fail more spectacularly. We don't even have candidates for Dirac operators on Xq which are analogous the Dabrowski-Sitarz operator on CP1, and there are good rea- sons to be pessimistic about the existence of such operators (although see [NT10] for a very different approach to defining spectral triples on quantum groups). On the other hand, there is a q-deformation of the Bernstein-Gelfand- Gelfand complex on Xq. We shall describe the results of [VY15], which show that the rudimentary multifiltered pseudodifferential theory of Chapter 3 sur- vives the quantization process. Once again, this is far from being as powerful as a full pseudodifferential calculus, but it still suffices to obtain some useful results. As in the classical case, we can obtain an Gq-equivariant Bernstein- Gelfand-Gelfand class in KKGq (C(Xq), C), where Gq denotes the Drinfeld dou- ble Kq ⊲⊳ Kq. As a result, following earlier work of Meyer, Nest and Voigt [MN06, MN07, NV10, Voi11], we obtain equivariant Poincaré duality for Xq and a proof of the Baum-Connes Conjecture for the discrete quantum dual of SUq(3). 4.1 Notation and conventions We continue to use the notation of Section 3.1 for complex semisimple Lie groups and their Lie algebras. For their quantizations, we follow the conven- tions of [VY17], which we rapidly recap here. Unless otherwise stated, we assume 0 < q < 1. • Uq(g) is the Hopf algebra with -- generators: Eα, Fα (for α ∈ Σ) and Kλ (λ ∈ P), -- algebra relations: K0 = 1, KλEαK−1 λ = q(λ,α)Eα, [Eα, Fβ] = δα,β Kα − K−1 α q − q−1 , KλKµ = Kλ+µ, KλFαK−1 λ = q−(λ,α)Eα, and the quantum Serre relations, which we shall not write out here, -- coalgebra relations: ∆Eα = Eα ⊗ Kα + 1 ⊗ Eα, ∆Kλ = Kλ ⊗ Kλ, ǫ(Eα) = ǫ(Fα) = 0, ∆Fα = Fα ⊗ 1 + K−1 α ⊗ Fα, ǫ(Kλ) = 1, -- antipode: S(Kλ) = K−λ, S(Eα) = −EαK−1 α , S(Fα) = −KαFα. 34 • Uq(k) denotes Uq(g) with the ∗-structure E∗ α = KαFα, α = EαK−1 F∗ α , K∗ α = Kα. • A(Kq) denotes the algebra of matrix coefficients of finite dimensional Uq(k)-modules of type 1 (meaning that the Kλ act as positive operators), with Hopf ∗-algebra structure dual to Uq(k)cop, i.e., for all X, Y ∈ Uq(k), a, b ∈ A(Kq), (X, ab) = (X(1), b)(X(2), a), (X, a∗) = ( S−1(X)∗, a). (XY, a) = (X, a(1))(Y, a(2)), • The correspondence between left A(Kq)-comodules and left Uq(k)-modules (of type 1) is made by equipping a left A(Kq)-comodule V with the left Uq(k)-action X.v := (S(X), v(−1))v(0) ∀X ∈ Uq(k), v ∈ V. • For X ∈ Uq(k), a ∈ A(Kq), we will write X ⇀ a = a(1)(X, a(2)), (4.1.1) for the right regular representation of U (k) on C∞(K). • φ denotes the bi-invariant Haar state on A(Kq), L2(Kq) is the corre- sponding GNS space, and C(Kq) is the C∗-closure of the GNS repre- sentation of A(Kq). If Lq is a quantum subgroup of Kq, defined via a surjective morphism resLq : A(Kq) ։ A(Lq), then the quantum homogeneous space Kq/Lq is de- fined by its algebra of polynomial functions, A(Kq/Lq) = {a ∈ A(Kq) (id ⊗ resLq )∆a = a ⊗ 1}. The closures of A(Kq/Lq) in L2(Kq) and C(Kq) are denoted by L2(Kq/Lq) and C(Kq/Lq), respectively. If V is a finite-dimensional unitary left A(Lq)-comodule, with coaction α : V → A(Lq) ⊗ V, then the induced bundle Kq ×Lq V is defined via its space of polynomial sections: A(Kq ×Lq V) = {ξ ∈ A(Kq) ⊗ V (id ⊗ resLq ⊗ id)(∆ ⊗ id)ξ = (id ⊗α)ξ}. Equivalently, an element ξ = ∑i ai ⊗ vi ∈ A(Kq) ⊗ V belongs to A(Kq ×Lq V) if and only if (X(1) ∑ i ⇀ ai) ⊗ (X(2).vi) = ǫ(X) ∑ i ai ⊗ vi ∀X ∈ Uq(l). (4.1.2) The space A(Kq ×Lq V) is a left A(Kq/Lq)-module by left multiplication on the first leg, and a left A(Kq)-comodule by comultiplication on the first leg. It's closure in L2(Kq) ⊗ V is denoted L2(Kq ×Lq V). 35 4.2 Pseudodifferential calculus and twisted spec- tral triples 4.2.1 Twisted regularity In this section and the next, a twisting of an algebra A will mean a linear automorphism θ : A → A (not necessarily an algebra automorphism). We will use the notation for the twisted commutator of x and y. [x, y]θ = xy − θ(y)x Definition 4.2.1. ([CM08, Mos10, MY]) A (unital) twisted spectral triple (called a type III spectral triple in [CM08]) is (A, H, D, θ) where • H is a Hilbert space, • A is a unital ∗-algebra, represented as bounded operators on H, with twisting θ, • D is an unbounded self-adjoint operator on H, such that 1. D has compact resolvent, 2. [D, a]θ is densely defined and bounded for all a ∈ A. It is regular if θ can be extended to a ∗-algebra B ⊆ L(H) of bounded operators containing both A and [D, A]θ such that the twisted derivation preserves B. δθ = [D, · ] Our definition of a twisted spectral triple is weaker than the definitions of Connes and Moscovici, thanks our allowing θ to be merely a linear isomor- phism, not an algebra automorphism. This relaxation of the twisting will be necessary for application to quantum homogeneous spaces. Our main goal will be to prove a twisted analogue of the results of [Hig04, Uuy11] which show the equivalence of regularity and the existence of a pseu- dodifferential calculus. 4.2.2 Differential and pseudodifferential operators The following definitions are standard. • ∆ = D2 + 1 is referred to as the Laplace operator, • H∞ =Tn∈N Dom(∆n), is the space of smooth elements in H, • The s-Sobolev space Hs (for s ∈ R) is the completion of H∞ with respect to the norm kvks = k∆ s 2 vk, 36 • Opt (for t ∈ R) is the set of operators T on H∞ which extend to bounded operators T : Hs+t → Hs for every s ∈ R; these are called the operators of analytic order t. 2 ≤ ∆− 1 2 ∈ Op− 1 1 Remark 4.2.2. The Laplace operator ∆ is intended as an invertible replacement of D2. Note that D − ∆ 2 . This will mean that D and 1 2 are essentially interchangeable in the analysis that follows, and we will ∆ occasionally do so without further comment. Definition 4.2.3. An algebra of differential operators adapted to a twisted spectral triple (A, H, D, θ) is an N-filtered algebra DO• of operators on H∞ which con- tains both A and [D, A]θ in DO0, is equipped with an extension of θ, and satisfies: 1. Elliptic estimates: for every X ∈ DOn, there exists C > 0 such that kXvkH ≤ Ck∆ 2. [∆, DOn]θ2 ⊆ DOn+1 for all n. n 2 kH for all v ∈ H∞, Definition 4.2.4. An algebra of pseudodifferential operators adapted to (A, H, D, θ) is an R-filtered subalgebra Ψ• of Op• containing both A and [D, A]θ in Ψ0, equipped with a 1-parameter group (Θz)z∈C of twistings, and satisfying 1. θ(a) − Θ1(a) ∈ Ψ−1 for all a ∈ A, 2. ∆zΨ• ⊆ Ψ• and Ψ•∆z ⊆ Ψ• for all z ∈ C, 3. [∆ z 2 , Ψs]Θz ⊆ ΨRe(z)+s−1 for all z ∈ C, s ∈ R. Theorem 4.2.5. ([MY]) Let (A, H, D, θ) be a twisted spectral triple. 1. If (A, H, D, θ) is regular, then there exist algebras of differential operators and pseudodifferential operators adapted to it. 2. Conversely, if there exists an algebra of differential operators DO adapted to (A, H, D, θ), with θ diagonalizable -- meaning that DO is the direct sum of its θ-eigenspaces -- then (A, H, D, θ) is regular. Proof. The first statement is straightforward. Using δθ-stability, one can prove that the algebra B from the definition of regularity lies in Op0. Then we can define Ψ• as the subalgebra of Op• generated by B and ∆z for all z ∈ C, equipped with the family of twistings Θz(X) = ∆zX∆−z, X ∈ Ψ, z ∈ C. This yields an algebra of pseudodifferential operators adapted to (A, H, D, θ). For the algebra of differential operators, put DOm = Ψm. We have Θ1(b) − θ(b) ∈ Ψ−1 for all b ∈ B, and consequently we can extend θ to a perturbation of Θ1 on DO which satisfies the axioms of an algebra of differential operators. The second statement is more profound. Suppose DO• is an algebra of dif- ferential operators adapted to (A, H, D, θ), with θ diagonalizable. It suffices to define an adapted algebra of pseudodifferential operators, since then we can use B = Ψ0 to obtain regularity. To define the pseudodifferential operators, we start with the following definition. 37 Definition 4.2.6. Let T be an operator on H∞. We say that T has an asymptotic i=1 Ti, where each Ti ∈ Op•, if for any r ∈ R we have T − expansion T ∼ ∑∞ i=1 Ti ∈ Op−r for all sufficiently large N. ∑N One defines a basic pseudodifferential operator to be an operator T on H∞ that admits an asymptotic expansion of the form T ∼ N ∑ k=0 Xk∆ z 2 −k, with Xk ∈ DO. We then define Ψ as the space of finite linear combinations of basic pseudodifferential operators. It is nontrivial to prove that Ψ is an algebra. The key point is to prove that if X ∈ DO we have ∆zX ∈ Ψ. The standard trick here is use the Cauchy Integral Formula to write, when Re(z) < 0, ∆z =IΓ λz(λ − ∆)−1 dλ, where Γ is a vertical line separating the spectrum of ∆ from 0. In the untwisted case, one can then repeatedly apply the commutator formula [(λ − ∆)−1, Y] = (λ − ∆)−1[∆, Y](λ − ∆)−1, (Y ∈ DO) to develop the product ∆zX as an asymptotic expansion. In the twisted case, the commutators with the resolvent are more delicate, and it's here we take advantage of the diagonalizability of θ. Let us write W = Sp(θ) for the set of eigenvalues of θ. We write ∇(X) = [∆, X]θ2 for the twisted commutator of X ∈ DO. We will define a decomposition of the iterated twisted commutator ∇n(X) indexed by sequences µ = (µ0, . . . , µn) ∈ Wn, as follows. If n = 0 then ∇(µ0)(X) denotes the µ0-eigencomponent of X. If n > 0 we define recursively Thus ∇µ(X) = ∇(∇(µ0,...,µn−1)(X))µn. ∇n(X) = ∑ ∇µ(X). µ∈Wn+1 Lemma 4.2.7. For any X ∈ DO and z ∈ C, we have the asymptotic expansion ∆zX ∼ ∞ ∑ n=0 µ∈Wn+1(cid:18)z n(cid:19)µ ∑ ∇µ(X)∆z−k, where the (z n) The n = 0 term is equal to θz(X)∆z−k. µ are the quantized binomial coefficients defined in [MY, Appendix A]. From this, we obtain that Ψ• is an algebra. We can define a one parameter family of twistings on Ψ by Θz(T) = ∆z/2T∆−z/2, and it is easy to check that Ψ• satisfies all the axioms of an algebra of pseudodifferential operators. 38 4.2.3 Pseudodifferential calculus on quantum projective spaces The framework above was designed with the following application in mind. Let G = SL(n + 1, C) with simple roots {α1, . . . , αn}. Let ¯P = ¯P{α1,...,αn−1} be the lower parabolic subgroup corresponding to I = {α1, . . . , αn−1} ⊆ Σ -- see Section 3.1 -- so that G/ ¯P = CPn. Let K = SU(n + 1) and L = ¯P ∩ K = S(U(n) × U(1)), and note that K/L = CPn. We denote the nilpotent radical of ¯p by ¯n, and observe that ¯n ∼= (g/¯p)∗, which is an irreducible L-module via the coadjoint action. The anti-holomorphic exterior bundle of CP1 is isomorphic to K ×L Λ(¯n) Now let A(Kq), A(Lq), A(CPn q ) be the algebras of polynomial functions on Kq, Lq and CPn q . Equip ¯n with the left Uq(l) representation which cor- responds to the L-module structure above. Krähmer and Tucker-Simmons [KTS15] showed that one can define a q-deformation Λq(¯n) of the exterior algebra of ¯n. They use this to define a natural q-deformation of the Dirac- Dolbeault operator D = ¯∂ + ¯δ∗ on the sections of the quantum exterior bundle Ωq := Kq ⊗Lq Λq(¯n). Theorem 4.2.8. Let A = A(CPn operator of [KTS15]. The twisted spectral triple (A, H, D, id) is regular. q ), H = L2(Ωq) and D be the Dirac-Dolbeault Remark 4.2.9. The fact that (A, H, D) has compact resolvent was proved in [DD10]. Note that the twisting is trivial on the algebra A(CPn q ), but the spectral triple is not regular in the untwisted sense. The associated algebras of differ- ential and pseudodifferential operators will be twisted, as seen for instance in [NT05]. In the rest of this section, we will sketch the proof of Theorem 4.2.8. We have to define a twisted algebra of differential operators. We begin by defining the algebra of differential operators on Kq (with polynomial coeffi- cients) as the smash product DO(Kq) = A(Kq)#Uq(k). That is, DO(Kq) = A(Kq) ⊗ Uq(k) as a space, with algebra structure given by the usual products on A(Kq) and Uq(k) and the commutation relation Xa = (X(2) ⇀ a)X(1), ∀X ∈ Uq(k), a ∈ A(Kq). These relations are defined so that the multiplication action of A(Kq) and the ⇀ action (4.1.1) of Uq(k) define an action of DO(Kq) on A(Kq). We define an algebra filtration on Uq(k) by declaring Uq(l) ⊂ U 0 q (k), Eαn, Fαn ∈ U 1 q (k). This is a Hopf algebra filtration, meaning that ∆U m (k). We can therefore extend the filtration to DO(Kq) by putting DOm(Kq) = A(Kq)#U m q(k) ⊗ U m−i q q (k) ⊆ ∑m i=0 U i q (k). Now let V be a finite dimensional Lq-module and let E = Kq ×Lq V be q . Define an adjoint action of Uq(l) on the associated induced bundle over CPn 39 DO(Kq) ⊗ End(V) by ad(X)(A ⊗ T) = X(1)AS(X(4)) ⊗ X(2)TS(X(3)), for X ∈ Uq(l), A ⊗ T ∈ DO(Kq) ⊗ End(V). From Equation (4.1.2) it follows that the invariant subalgebra DO(E ) := (DO(Kq) ⊗ End(V))ad(Uq(l)) = {P ∈ DO(Kq) ⊗ End(V) ad(X)P = ǫ(X)P ∀X ∈ Uq(l)} preserves the section space A(E ) ⊆ A(Kq) ⊗ V. We refer to DO(E ) as the algebra of differential operators on E, with filtration inherited from DO(Kq). The Dirac-Dolbeault operator D of [KTS15] belongs to DO1(Ωq). D'Andrea- Dabrowski proved that it satisfies the following Parthasarathy formula. Proposition 4.2.10. [DD10] There is a central element C ∈ U 2 q (k) such that D2 − C ⊗ IΛq( ¯n) ∈ DO1(Ωq). The element C is called the (order 2) Casimir element. D'Andrea-Dabrowski show that, as an unbounded operator on L2(Ωq), C ⊗ I has compact resolvent, so (A, H, D) is a spectral triple. From these calculations, we also see that E∗ αn Eαn, F∗ αn Fαn ≤ C as unbounded operators, from which we can deduce the elliptic estimates of Definition 4.2.3. The coproduct of C has the particularly nice form ωn ⊗ C + Q + C ⊗ K−2 ωn ∆(C) = K2 for some Q ∈ U 1 ing (ωn, αi) = δin for all i. It follows that, for any a ∈ A(Kq), q (k), where ωn ∈ P is the fundamental weight satisfy- q (k) ⊗ U 1 Ca − (K−2 ωn ⇀ a)C ∈ DO1(Kq) and for any X ∈ Uq(k), CX − XC = 0. We are led to define the following twisting. Definition 4.2.11. Define a twisting θ on DO(Kq) ⊗ End(Λq(¯n)) via the for- mula θ(aX ⊗ T) = (K−2 ωn ⇀ a)X ⊗ T for all a ∈ A(Kq), X ∈ Uq(k), T ∈ End(Λq(¯n)). Note also that K2 ωn commutes with all elements of Uq(l). We obtain the following. Lemma 4.2.12. The twisting θ on DO(Kq) ⊗ End(Λq(¯n)) preserves the subspace DO(Ωq). With this twisting, DO(Ωq) is a twisted algebra of differential operators adapted to (A, H, D, id). This completes the proof of Theorem 4.2.8. 40 4.3 Pseudodifferential operators on the full flag manifold of SUq(n) We now pass to our final example, the q-deformation of the full flag manifold of SUq(n). Here we will emulate the algebras of longitudinal pseudodifferen- tial operators defined in Section 3. On the full quantum flag manifold of a quantized compact semisimple Lie group Kq, there are two reasons why the Bernstein-Gelfand-Gelfand complex seems to be the appropriate point of departure for equivariant index theory. The first, as already mentioned, is that there is no reasonable candidate for a Dirac operator analogous to those just discussed for CPn q . More profoundly, as Nest and Voigt pointed out in [NV10], the Kasparov product is not well-defined in Kq-equivariant KK-theory. In order to define a product of two C(Kq)-comodule algebras, one of them needs to be equipped with a Yetter-Drinfeld module structure (see also [Vae05]), or equivalently, a representation of the Drinfeld double Kq ⊲⊳ Kq. The Drinfeld double of a q-deformed complex semisimple Lie group Kq is often referred to as its com- plexification Gq because it behaves as a q-deformation of the complexification G of K. Thus in order to have a KK-product, we need to work in Gq-equivariant KK-theory. In light of the results of Chapter 3, this again suggests the BGG complex. In the remaining sections, we will show that the BGG class in the SL(3, C)- equivariant K-homology of the full flag manifold X does indeed admit a q- deformation to a SLq(3, C)-equivariant K-homology class for Xq. Again, we will not define a full pseudodifferential calculus, but merely a family of al- gebras of "operators of negative order" associated to each of the canonical fibrations of the quantum flag manifold. This time, we will rely upon har- monic analysis of the compact quantum subgroups associated to the various fibrations, rather than the nilpotent subgroups. 4.3.1 Algebras of longitudinal pseudodifferential operators on quantum flag manifolds We will continue with the notation from Sections 3.1 and 4.1, although we will specialize to the case K = SU(n + 1). We introduce the following notation for quantum flag manifolds. Here, I, J ⊆ Σ are sets of simple roots. • KI = K ∩ PI denotes the compact part of the parabolic subgroup PI, and KI,q its q deformation. The latter is a quantum subgroup of Kq via a surjective morphism resKI,q : A(Kq) ։ A(KI,q). In particular, KΣ,q = Kq and K∅,q = T, the classical torus subgroup. • XI,q = Kq/KI,q is the quantized partial flag manifold of Kq. In particular, we write Xq = X∅,q = Kq/T for the full quantum flag manifold. • If µ ∈ P is an integral weight, we write eµ ∈ C(T) for the corresponding unitary character of T, and Cµ for the associated one-dimensional left corepresentation of C(T), given by z 7→ eµ ⊗ z for all z ∈ C. 41 • Eµ = Kq ×T Cµ is the induced line bundle over Xq, with section space A(Eµ) = {ξ ∈ A(Xq) (id ⊗ resT)∆ξ = ξ ⊗ eµ}. The Hilbert completion with respect to the Haar state on A(Kq) is de- noted L2(Eµ). To characterize the negative order pseudodifferential operators, we are obliged to use harmonic analysis with respect to the subgroups KI,q. We in- troduce further notation: • Irr(KI,q) is the set of finite dimensional irreducible KI,q-representations, up to equivalence. • If σ ∈ Irr(KI,q), Vσ denotes the vector space underlying σ. • The algebras of functions and compactly supported functions on the discrete dual of KI,q are, respectively, A( KI,q) = ∏ End(Vσ), σ∈Irr(KI,q) Ac( KI,q) = Mσ∈Irr(KI,q) End(Vσ). Note that there are canonical embeddings A( KI,q) ֒→ A( Kq). • For any I ⊂ Σ and any σ ∈ Irr(KI,q), pσ denotes the central projection in Ac( KI,q), which acts on any unitary Kq-representation as the projection onto the σ-isotypical subspace. • In the particular case I = ∅, we identify eµ ∈ Irr(T) with µ ∈ P. Then pµ acts as the projection onto the µ-weight space of any unitary Kq- representation. In particular, we have L2(Eµ) = pµ ⇀ L2(Kq), where ⇀denotes the obvious extension of the right regular representation (4.1.1) to A( Kq). The following lemma is elementary. Lemma 4.3.1. If I ⊆ J, then for any σ ∈ Irr(Kq,I) and τ ∈ Irr(Kq,J), the projections pσ and pτ commute. In particular, pσ commutes with pµ for any integral weight µ ∈ P. Note that the action ⇀ of A( Kq) on A(Kq) does not restrict to an action on the subspace A(Eµ). But Lemma 4.3.1 implies that the projection pσ ⇀ does restrict to a projection on A(Eµ). In other words, it is still meaningful to talk of the decomposition of the section space L2(Eµ) into isotypical components for the right regular action of the subgroup KI,q. This observation will allow us to define analogues of the algebras KI for the classical groups (Definition 3.7.1) which contain the longitudinal pseudodifferential operators of negative order along the foliations FI. In the sequel, we write pσ for the operator pσ ⇀ on L2(Eµ). 42 Definition 4.3.2. Fix q ∈ (0, 1], let µ, ν ∈ P and let I ⊆ Σ. We define KI (L2(Eµ), L2(Eν)) to be the norm-closure of the set of bounded operators T : L2(Eµ) → L2(Eν) satisfying pσT = Tpσ = 0 for all but finitely many σ ∈ Irr(KI,q) This definition generalizes in an obvious way to operators between finite direct sums of line bundles E = Li Eµi. As in Chapter 3, we will often sup- press the bundles in the notation, when they are clear from the context. As a technical convenience, we can view the spaces KI (L(E ), L2(E ′)) as the mor- phism spaces of a C∗-category KI with objects L2(E ) for any E direct sum of induced line bundles. Example 4.3.3. When q = 1, the above definition of KI coincides with the algebra KI of Definition 3.7.1, that is, the hereditary C∗-algebra generated by the longitudinal smoothing operators Ψ−∞ along the fibres of the fibration X ։ XI. I The following theorem is a quantum analogue of Theorem 3.3.2 for prod- ucts of longitudinal pseudodifferential operators on classical flag manifolds. Theorem 4.3.4 ([VY15]). Let Kq = SUq(n). For any I, J ⊆ Σ, we have KIKJ ⊆ KI∪J. Moreover, KΣ = K, the compact operators. The proof of this theorem uses explicit calculations in terms of the Gelfand- Tsetlin bases of simple SUq(n)-modules, see [VY15]. This type of calculation is not easy to generalize to other semisimple groups. Although, one obviously expects that Theorem 4.3.4 should hold for all compact semisimple quantum groups, this remains an open problem. 4.3.2 Principal series representations of SLq(n, C) As in Chapter 3, the above analysis is motivated by the equivariant index theory. The Bernstein-Gelfand-Gelfand complex of Theorem 3.2.2 admits a q-deformation to a complex of unbounded intertwiners between non-unitary principal series representations of a quantized complex semisimple Lie group, see [VY17]. In this section, we will describe the normalized BGG complex, for which we will only need the base of principal series representations, meaning those principal series with continuous parameter λ = 0. We begin with some general notation. Definition 4.3.5. Let W ∈ M(A(Kq) ⊗ Ac( Kq)) denote the multiplicative unitary of Kq, which is characterized by the following property. For f ∈ A(Kq) we write [ f ] ∈ L2(Kq) for the corresponding element in the GNS space. We equip L2(Kq) the left actions λ of A(Kq) and λ of Ac( Kq) defined by λ(a)[ f ] = [a f ], λ(x)[ f ] = ( S(x), f(1))[ f(2)], a, f ∈ A(Kq), x ∈ Ac( Kq), f ∈ A(Kq). Then W satisfies (λ ⊗ λ)(W) ([ f ] ⊗ [g]) = [S−1(g(1)) f ] ⊗ [g(2)], ∀ f , g ∈ A(Kq). 43 Definition 4.3.6. The complex semisimple quantum group Gq = SLq(n, C) is defined as the Drinfeld double of Kq = SUq(n). This means that it has function algebra Ac(Gq) = A(Kq) ⊲⊳ Ac( Kq), which is equal to A(Kq) ⊗ Ac( Kq) as an algebra, and equipped with the coal- gebra operations ∆Gq = (id ⊗Flip ⊗ id)(id ⊗ ad(W) ⊗ id)(∆ ⊗ ∆), ǫGq = ǫ ⊗ ǫ. Note that Gq is an algebraic quantum group in the sense of van Daele [VD98]. We write D(Gq) for the dual multiplier Hopf algebra, with operations dual to Ac(Gq)op. Definition 4.3.7. Let µ ∈ P. The base of principal series representation with parameter µ is the representation πµ,0 of D(Gq) on the Hilbert space L2(Eµ) which corresponds to the following Kq-Yetter-Drinfeld structure: • coaction of A(Kq) on L2(Eµ) by the left regular corepresentation, • action of A(Kq) on L2(Eµ) by the twisted adjoint action: a · ξ = (K2ρ, a(2)) a(1)ξS(a(3)) The resulting representation of D(Gq) is a ∗-representation, i.e. corresponds to a unitary representation of Gq. The following theorem summarizes some key properties of these representations, which are analogues of well-known results for the unitary principal series of the classical group G. Theorem 4.3.8. 1. The representations πµ,0 are all irreducible. 2. Two representations πµ,0 and πµ′,0 are unitarily equivalent if and only if µ′ = wµ for some w ∈ W. 3. Suppose w → w′ is an edge in the Bruhat graph associated to a simple root α and µ is a dominant weight. Then the unitary intertwiner between π−wµ,0 and π−w′µ,0 is given by Iw→w′ := ph(En α ) ⇀ : L2(E−wµ,0) → L2(E−w′µ,0). where n is such that wµ − w′µ = nα. The first two results are essentially due to Joseph and Letzter [JL95, Jos95]. The third is quantum analogue of the formula (3.7.3) when µ = ρ. For proofs of these, and many other related results, see [VY17]. 44 4.3.3 Algebras of longitudinal pseudodifferential operators, cont'd Morally speaking, the intertwiners of Theorem 4.3.8(3) are longitudinal pseu- dodifferential operators of order 0 along the fibres of Xq ։ Xα,q. Let us give a precise meaning to this. Following Definition 3.7.1, we define A to be simulta- neous multiplier category of all KI (I ⊆ Σ). In other words, A ∈ AI(L2(E, E ′)) if for any I ⊆ Σ and any direct sum of line bundles E ′′ we have A.KI ⊆ KI and KI.A ⊆ KI. We also put All of the operators we are interested in belong to A. JI = KI ∩ A. Lemma 4.3.9 ([VY15]). Let Kq = SUq(n). For any µ ∈ P, the following bounded operators belong to A: (i) left or right multiplication by any a ∈ C(Eν), as an operator L2(Eµ) → L2(Eµ+ν); (ii) the principal series representation πµ,0( f ) of any f ∈ D(Gq) acting on L2(Eµ); (iii) the operator phases of the ⇀ action of Eα, Fα for any simple root α ∈ Σ, viewed as operators ph(Eα) : L2(Eµ) → L2(Eµ+α) and ph(Fα) : L2(Eµ) → L2(Eµ−α). Remark 4.3.10. Parts (i) and (ii) of this lemma are straightforward, but the proof of part (iii) relies on some rather delicate analysis using Gelfand-Tsetlin bases and q-orthogonal polynomials; see [VY15]. Recall from Lemma 3.7.2 that this was one of the most difficult parts of the classical analysis, as well. α ) : L2(Eµ) → L2(Eµ+nα) and Keeping in mind that the operators ph(En α ) : L2(Eµ) → L2(Eµ−nα) are to be thought of as order 0 longitudinal ph(Fn pseudodifferential operators along the foliation Fα, the following commutator relations should not be surprising. Lemma 4.3.11. Let µ, ν ∈ P, n ∈ N. For any section a ∈ A(Eν) we have ph(En ph(Fn α )a − a ph(En α )a − a ph(Fn α ) ∈ Jα(L2(Eµ), L2(Eµ+ν+nα)), α ) ∈ Jα(L2(Eµ), L2(Eµ+ν−nα)). 4.3.4 The BGG class of the quantum flag manifold of SUq(3) Having obtained the key technical results -- Theorem 4.3.4 and Lemma 4.3.9 -- we can now translate the construction of the Kasparov module in Section 3.7 almost word-for-word to the quantum context. We will give an extremely brief summary. Let α, β and ρ = α + β be the three positive roots of SUq(3). 45 Lemma 4.3.12. The diagramme of normalized BGG operators L2(E0,0) ph(Eα) 6 6♥♥♥ PPP ph(Eβ) L2(Eα,0) ⊕ ❑❑❑❑❑❑❑❑❑❑❑ 9sssssssssss L2(Eβ,0) T1 T2 L2(E2α+β,0)ph(Eβ) ph(F2 α ) ph(F2 β) ❙❙❙ ⊕ 5❦❦❦ / L2(Eα+2β,0)ph(Eα) L2(E2ρ,0), (4.3.1) with T1 = − ph(E2 β), defines a complex modulo Jα + Jβ. Moreover, the arrows which are pointed north-east (respec- tively, south-east, east) are invertible modulo Jα (respectively, Jβ, Jα + Jβ). α) and T2 = − ph(E2 β) ph(Eα) ph(E∗ α) ph(Eβ) ph(E∗ Proof. By Hopf-Galois theory, we can find sections f1, . . . , fn ∈ A(E−ρ) such that ∑n i fi = 1 ∈ A(X ). From Lemma 4.3.11 we obtain that i=1 f ∗ ∑ i i ph(Eα) fi! − ph(Eα) ∈ Jα(L2(Eµ), L2(Eµ+α), f ∗ for any µ ∈ P. The result then follows, as in the classical case of Lemma 3.7.4, by comparing the diagram of normalized BGG operators (4.3.1) with the diagram of intertwiners Iw→w′ from Theorem 4.3.8. We now have all the elements that we used in the construction of the BGG class for classical SL(3, C). We can repeat the application of the Kasparov Technical Theorem to obtain operators Nα, Nβ analogous to those in Equation (3.7.5), and then modify the normalized BGG complex as in Figure 3.2. The result is an Kasparov K-homology class [BGG] for Xq which is equivariant with respect to the Drinfeld double SLq(3, C) of SUq(3). The details can be found in [VY15]. Theorem 4.3.13. ([VY15]) The above construction yields an SLq(3, C)-equivariant K-homology class for Xq, which we denote by [BGG] ∈ KKSLq(3,C)(C(Xq), C). Its image under the forgetful map to KKSUq(3)(C, C) is the class of the trivial representa- tion. Let us finish with two immediate applications of the BGG class. The first application is the equivariant Poincaré duality of the flag man- ifold Xq. The notion of KK-theoretic Poincaré duality was introduced by Kasparov, but its generalization to quantum group equivariant KK-theory re- quires some considerable technical refinement. This is due to the observation of Nest and Voigt [NV10] that the Kasparov product in KKKq requires the use of the braided tensor product ⊠. Therefore, even if we want to prove only Kq-equivariant Poincaré duality, we are obliged to define a Poincaré duality class in KKGq, where Gq is the Drinfeld double of Kq. Nest and Voigt [NV10] showed how to prove SUq(2)-equivariant Poincaré duality for the Podle´s sphere, using an SLq(2, C)-equivariant duality class. The same argument goes through for the full flag variety of SUq(3) using the class [BGG], and yields the following. 46 % % / / ) ) ( ( 9 / 5 Theorem 4.3.14. The quantum flag manifold Xq is SUq(3)-equivariantly Poincaré dual to itself. That is, there is a natural isomorphism KKSLq(3,C) ∗ (C(Xq) ⊠ A, B) ∼= KKSLq(3,C) ∗ (A, C(Xq) ⊠ B) for all SLq(3, C)-C∗-algebras A and B. The second application is to the Baum-Connes Conjecture. Once again, making sense of the Baum-Connes Conjecture for discrete quantum groups requires some serious technical machinery, this time developed by Meyer and Nest [MN06, MN07]. The Baum-Connes Conjecture for the discrete dual of SUq(2) was proven by Voigt in [Voi11]. An analogous argument using the class [BGG] yields the following. Theorem 4.3.15. The discrete quantum dual of SUq(3) with q ∈ (0, 1] satisfies the Baum-Connes Conjecture with trivial coefficients. 47 Bibliography [BE89] Robert J. Baston and Michael G. Eastwood. The Penrose transform. Oxford Mathematical Monographs. The Clarendon Press, Oxford University Press, New York, 1989. Its interaction with representa- tion theory, Oxford Science Publications. [BG88] Richard Beals and Peter Greiner. Calculus on Heisenberg manifolds, volume 119 of Annals of Mathematics Studies. Princeton University Press, Princeton, NJ, 1988. [BGG75] I. N. Bernstein, I. M. Gelfand, and S. I. Gelfand. Differential op- erators on the base affine space and a study of g-modules. pages 21 -- 64, 1975. [BN03] Moulay-Tahar Benameur and Victor Nistor. Homology of algebras of families of pseudodifferential operators. J. Funct. Anal., 205(1):1 -- 36, 2003. [Boh15] Karsten Bohlen. The K-theory of bisingular pseudodifferential al- gebras. J. Pseudo-Differ. Oper. Appl., 6(3):361 -- 382, 2015. [CF03] Marius Crainic and Rui Loja Fernandes. Integrability of Lie brack- ets. Ann. of Math. (2), 157(2):575 -- 620, 2003. [CM95] A. Connes and H. Moscovici. The local index formula in noncom- mutative geometry. Geom. Funct. Anal., 5(2):174 -- 243, 1995. [CM08] Alain Connes and Henri Moscovici. Type III and spectral triples. In Traces in number theory, geometry and quantum fields, Aspects Math., E38, pages 57 -- 71. Friedr. Vieweg, Wiesbaden, 2008. [Con94] Alain Connes. Noncommutative geometry. Academic Press, Inc., San Diego, CA, 1994. [CP] [ CS09] Woocheol Choi and Raphael Ponge. nates and tangent groupoid for carnot manifolds. http://arxiv.org/abs/1510.05851. Andreas Cap and Jan Slovák. Parabolic geometries. I, volume 154 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2009. Background and general theory. Privileged coordi- Preprint. 48 [ CSS01] Andreas Cap, Jan Slovák, and Vladimír Soucek. Bernstein-Gelfand- Gelfand sequences. Ann. of Math. (2), 154(1):97 -- 113, 2001. [DD10] Francesco D'Andrea and Ludwik D abrowski. Dirac operators on ' quantum projective spaces. Comm. Math. Phys., 295(3):731 -- 790, 2010. [Deb01] Claire Debord. Local integration of Lie algebroids. In Lie algebroids and related topics in differential geometry (Warsaw, 2000), volume 54 of Banach Center Publ., pages 21 -- 33. Polish Acad. Sci., Warsaw, 2001. [DH17] [DS14] Shantanu Dave and Stefan Haller. Graded hypoellipticity of bgg sequences. Preprint. http://arxiv.org/abs/1705.01659, 2017. Claire Debord and Georges Skandalis. Adiabatic groupoid, crossed product by R∗ + and pseudodifferential calculus. Adv. Math., 257:66 -- 91, 2014. [Duf75] Michel Duflo. Représentations irréductibles des groupes semi- simples complexes. pages 26 -- 88. Lecture Notes in Math., Vol. 497, 1975. [EM] Charles Epstein and Richard Melrose. gebra, index theory and homology. http://www-math.mit.edu/~rbm/book.html. The heisenberg al- Unpublished book, [EMM91] C. L. Epstein, R. B. Melrose, and G. A. Mendoza. Resolvent of the Laplacian on strictly pseudoconvex domains. Acta Math., 167(1- 2):1 -- 106, 1991. G. B. Folland and E. M. Stein. Estimates for the ¯∂b complex and analysis on the Heisenberg group. Comm. Pure Appl. Math., 27:429 -- 522, 1974. [FS74] [FS82] G. B. Folland and Elias M. Stein. Hardy spaces on homogeneous groups, volume 28 of Mathematical Notes. Princeton University Press, Prince- ton, N.J.; University of Tokyo Press, Tokyo, 1982. [Hig04] Nigel Higson. The local index formula in noncommutative geom- etry. In Contemporary developments in algebraic K-theory, ICTP Lect. Notes, XV, pages 443 -- 536. Abdus Salam Int. Cent. Theoret. Phys., Trieste, 2004. [HN79] B. Helffer and J. Nourrigat. Caracterisation des opérateurs hy- poelliptiques homogènes invariants à gauche sur un groupe de Lie nilpotent gradué. Comm. Partial Differential Equations, 4(8):899 -- 958, 1979. [Hum08] James E. Humphreys. Representations of semisimple Lie algebras in the BGG category O, volume 94 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2008. 49 [JK95] [JL95] [Jos95] Pierre Julg and Gennadi Kasparov. Operator K-theory for the group SU(n, 1). J. Reine Angew. Math., 463:99 -- 152, 1995. Anthony Joseph and Gail Letzter. Verma module annihilators for quantized enveloping algebras. Ann. Sci. École Norm. Sup. (4), 28(4):493 -- 526, 1995. Anthony Joseph. Quantum groups and their primitive ideals, vol- ume 29 of Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Re- sults in Mathematics and Related Areas (3)]. Springer-Verlag, Berlin, 1995. [Jul02] Pierre Julg. La conjecture de Baum-Connes à coefficients pour le groupe Sp(n, 1). C. R. Math. Acad. Sci. Paris, 334(7):533 -- 538, 2002. [Kas84] G. G. Kasparov. Lorentz groups: K-theory of unitary representa- tions and crossed products. Dokl. Akad. Nauk SSSR, 275(3):541 -- 545, 1984. English translation: Soviet Math. Dokl. 29 (1984), no. 2, 256 -- 260. [Kna01] Anthony W. Knapp. Representation theory of semisimple groups. Princeton Landmarks in Mathematics. Princeton University Press, Princeton, NJ, 2001. An overview based on examples, Reprint of the 1986 original. [Krä04] Ulrich Krähmer. Dirac operators on quantum flag manifolds. Lett. Math. Phys., 67(1):49 -- 59, 2004. [KTS15] Ulrich Krähmer and Matthew Tucker-Simmons. On the Dolbeault- Dirac operator of quantized symmetric spaces. Trans. London Math. Soc., 2(1):33 -- 56, 2015. [LMV17] Jean-Marie Lescure, Dominique Manchon, and Stéphane Vassout. About the convolution of distributions on groupoids. J. Noncommut. Geom., 11(2):757 -- 789, 2017. [Mel82] Anders Melin. Lie filtrations and pseudo-differential operators. Preprint, 1982. [MN06] Ralf Meyer and Ryszard Nest. The Baum-Connes conjecture via localisation of categories. Topology, 45(2):209 -- 259, 2006. [MN07] Ralf Meyer and Ryszard Nest. An analogue of the Baum-Connes isomorphism for coactions of compact groups. Math. Scand., 100(2):301 -- 316, 2007. [Mos10] Henri Moscovici. Local index formula and twisted spectral triples. In Quanta of maths, volume 11 of Clay Math. Proc., pages 465 -- 500. Amer. Math. Soc., Providence, RI, 2010. [MY] Marco Matassa and Robert Yuncken. Regularity of twisted spec- tral triples and pseudodifferential calculi. J. Noncommut. Geom. To appear. Preprint at https://arxiv.org/abs/1705.04178. 50 [Nis00] Victor Nistor. Groupoids and the integration of Lie algebroids. J. Math. Soc. Japan, 52(4):847 -- 868, 2000. [NT05] [NT10] [NV10] Sergey Neshveyev and Lars Tuset. A local index formula for the quantum sphere. Comm. Math. Phys., 254(2):323 -- 341, 2005. Sergey Neshveyev and Lars Tuset. The Dirac operator on compact quantum groups. J. Reine Angew. Math., 641:1 -- 20, 2010. Ryszard Nest and Christian Voigt. Equivariant Poincaré duality for quantum group actions. J. Funct. Anal., 258(5):1466 -- 1503, 2010. [NWX99] Victor Nistor, Alan Weinstein, and Ping Xu. Pseudodifferential op- erators on differential groupoids. Pacific J. Math., 189(1):117 -- 152, 1999. [Pon06] Raphaël Ponge. The tangent groupoid of a Heisenberg manifold. Pacific J. Math., 227(1):151 -- 175, 2006. [Pon08] Raphaël S. Ponge. Heisenberg calculus and spectral theory of hy- poelliptic operators on Heisenberg manifolds. Mem. Amer. Math. Soc., 194(906):viii+ 134, 2008. [Pus11] Michael Puschnigg. Finitely summable Fredholm modules over higher rank groups and lattices. J. K-Theory, 8(2):223 -- 239, 2011. [Roc78] Charles Rockland. Hypoellipticity on the Heisenberg group- representation-theoretic criteria. Trans. Amer. Math. Soc., 240:1 -- 52, 1978. [Rod75] Luigi Rodino. A class of pseudo differential operators on the prod- uct of two manifolds and applications. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4), 2(2):287 -- 302, 1975. [Tay] Michael crolocal http://www.unc.edu/math/Faculty/met/NCMLMS.pdf. Taylor. analysis, Noncommutative part I (revised mi- version). [Trè67] François Trèves. Topological vector spaces, distributions and kernels. Academic Press, New York-London, 1967. [Uuy11] Otgonbayar Uuye. Pseudo-differential operators and regularity of spectral triples. In Perspectives on noncommutative geometry, vol- ume 61 of Fields Inst. Commun., pages 153 -- 163. Amer. Math. Soc., Providence, RI, 2011. [Vae05] Stefaan Vaes. A new approach to induction and imprimitivity re- sults. J. Funct. Anal., 229(2):317 -- 374, 2005. [VD98] A. Van Daele. An algebraic framework for group duality. Adv. Math., 140(2):323 -- 366, 1998. 51 [vE05] [vEY] Erik van Erp. The Atiyah-Singer index formula for subelliptic operators on contact manifolds. ProQuest LLC, Ann Arbor, MI, 2005. Thesis (Ph.D.) -- The Pennsylvania State University. Erik van Erp and Robert Yuncken. A groupoid approach to pseudodifferential operators. J. Reine Agnew. Math. To appear. https://doi.org/10.1515/crelle-2017-0035. [vEY17] Erik van Erp and Robert Yuncken. On the tangent groupoid of a filtered manifold. Bull. Lond. Math. Soc., 49(6):1000 -- 1012, 2017. [Voi11] [VY15] [VY17] Christian Voigt. The Baum-Connes conjecture for free orthogonal quantum groups. Adv. Math., 227(5):1873 -- 1913, 2011. Christian Voigt and Robert Yuncken. Equivariant Fredholm mod- ules for the full quantum flag manifold of SUq(3). Doc. Math., 20:433 -- 490, 2015. Christian Voigt and Robert Yuncken. ple quantum groups and representation theory. https://arxiv.org/abs/1705.05661 (226 pages), 2017. Complex semisim- Preprint. [Yun10] Robert Yuncken. Products of longitudinal pseudodifferential oper- ators on flag varieties. J. Funct. Anal., 258(4):1140 -- 1166, 2010. [Yun11] Robert Yuncken. The Bernstein-Gelfand-Gelfand complex and Kas- parov theory for SL(3, C). Adv. Math., 226(2):1474 -- 1512, 2011. [Yun13] Robert Yuncken. Foliation C∗-algebras on multiply fibred mani- folds. J. Funct. Anal., 265(9):1829 -- 1839, 2013. 52
1907.05215	1	1907	2019-07-11T14:08:48	Pure infiniteness and paradoxicality for graph $C^*$-algebras	[ "math.OA" ]	We obtain necessary and sufficient conditions for pure infiniteness of the path groupoid $C^*$-algebra of a row-finite graph without sinks. In particular we show that for such a path groupoid $\mathcal{G}_E$, the properties of being essential principal and the existence of a basis of $(\mathcal{G}_E^a, 2, 1)$-paradoxical sets for the topology are not only sufficient, but also necessary.	math.OA	math	PURE INFINITENESS AND PARADOXICALITY FOR GRAPH C∗-ALGEBRAS FRANCESCA ARICI, BAUKJE DEBETS, AND KAREN R. STRUNG Abstract. We obtain necessary and sufficient conditions for pure infiniteness of the path groupoid C ∗-algebra of a row-finite graph without sinks. In par- ticular we show that for such a path groupoid GE, the properties of being essential principal and the existence of a basis of (G a E , 2, 1)-paradoxical sets for the topology are not only sufficient, but also necessary. 1. Introduction One of the most useful aspects of C∗-algebraic theory lies in its ability to interpret other mathematical or physical systems. Typically this involves input information from the system in question and an output C∗-algebra. Analysing the resulting C∗-algebra should then give information about the original system, and vice versa. The theory has been especially successful when the input is a directed graph. In this case, one constructs a universal C∗-algebra given by generators and relations determined by the vertex and edge sets of the graph. That one could construct a C∗-algebra for a directed graph was first shown in [24], where directed graphs were associated to the incidence matrices of topological Markov chains and more general shifts of finite type; from these one constructs C∗-algebras using the methods of Cuntz and Krieger [9]. Properties of a graph C∗-algebra A are quite tractable in comparison to arbitrary C∗-algebras, thanks to the combinatorial techniques available for the underlying graph. For example, one can read off the ideal structure of A directly from the graph [2]. Conversely, one can also use the structure of the C∗- algebra to identify information about the graph, for example, if A is approximately finite (AF), then the graph cannot have loops [15]. In addition to the Cuntz -- Krieger construction, one can start with a row-finite directed graph and first construct a locally compact ´etale groupoid coming from shift-equivalence of infinite paths [16]. Then, using the construction of Renault [21], one can construct the (reduced) groupoid C∗-algebra. Once again, it is possible to determine information about the graph C∗-algebra from its groupoid and vice versa. In this paper, we are interested in determining properties of the groupoid of a row-finite directed graph that imply that the corresponding C∗-algebra is purely infinite. Pure infiniteness was introduced for simple C∗-algebras by Cuntz [7, 6] as a C∗-algebraic analogue of the behaviour of type III von Neumann algebra factors. In the simple case, pure infiniteness implies a number of interesting structural properties. For example, if A is purely infinite and simple then it cannot admit Date: July 12, 2019. 2010 Mathematics Subject Classification. 22A22, 46L05, 46L35, 37B10. Key words and phrases. purely infinite C ∗-algebra, graph C ∗-algebra, ample groupoid, para- doxical decomposition. FA was partially supported by NWO under the VIDI-grant 016.133.326 and under the VENI- grant 016.192.237. BD was partially supported by the European Research Council Consolidator Grant 614195. KS was funded by Sonata 9 NCN grant 2015/17/D/ST1/02529 and a Radboud Excellence Initiative Postdoctoral Fellowship. 1 2 FRANCESCA ARICI, BAUKJE DEBETS, AND KAREN R. STRUNG any non-trivial traces. This can be seen by the fact that the nonzero projections in a purely infinite C∗-algebras are always infinite, which is to say, they are always Murray -- von Neumann equivalent to a proper subprojection [17]. Furthermore, a simple purely infinite C∗-algebra always has an abundant supply of projections -- every nonzero hereditary C∗-subalgebra contains an infinite projection. In fact, for simple C∗-algebras this property is equivalent to being purely infinite and can be taken as the definition [25]. In the non-simple case, it is not immediately clear what the appropriate notion of pure infiniteness should be. Cuntz's original formulation implies simplicity, so one is tempted to alternatively define a non-simple C∗-algebra to be purely infinite if every nonzero hereditary C∗-subalgebra contains an infinite projection. Although this property, which we call Condition (IH) in the sequel, is interesting in its own right, it is not enough to guarantee some of the more important consequences of pure infiniteness that one sees in the simple case. In particular, one would like that a (simple or otherwise) purely infinite C∗-algebra cannot admit nonzero traces and that the tensor product of any (simple or otherwise) C∗-algebra with the Cuntz algebra O∞ is purely infinite. A satisfactory definition was given in [13], which asks for a certain infinite condition for positive elements via Cuntz comparison. For the Cuntz -- Krieger construction of non-simple graph C∗-algebras of row- finite graphs, sufficient conditions guaranteeing pure infiniteness are already known [10]. However, if one forgets the Cuntz -- Krieger construction and relies only on the groupoid model, such conditions are only known in the case of a simple graph C∗-algebras. In this paper, we prove the following. Theorem (see Theorem 3.14). Let E be a row-finite directed graph without sinks. The following are equivalent: (1) C∗(GE) is purely infinite; (2) GE is essentially principal and for every finite path α, the cylinder set Z(α) is paradoxical; (3) E satisfies Conditions (K) and (DI); (4) E satisfies Conditions (K) and (DL). Condition (K) was introduced in [16] where it was used to generalise previous results from finite to infinite graphs. Conditions (I), (DI), and (DL) are condi- tions on the behaviour of paths in the graph and their relation to cylinder sets in the groupoid (see Definition 3.8). The notion of paradoxicality of a cylinder set (Definition 3.6 below) comes from to the notion of paradoxicality for more general groupoids, introduced in [4], which in turn was inspired by similar definitions for group actions (see for example [23, 11]). Essentially, it means that we are able to see the peculiar Banach -- Tarksi-like behaviour witnessed by infinite projections -- that they can be decomposed into subprojections of equal size -- at the level of the groupoid. The paper is organised as follows. In Section 2 we recall the constructions of graph C∗-algebras from a Cuntz -- Krieger E-family and from the path groupoid, as well as some of the properties we will need in the sequel. In Section 3 we sketch the ideas behind the characterisation of Condition (IH) in the groupoid model and discuss the difficulties of a straightforward generalisation to the non-simple case. Finally, we prove the main result, Theorem 3.14. Acknowledgments. The authors wish to thank Christian Bonicke, Kang Li, Bram Mesland, Bartosz Kwa´sniewski, and Adam Rennie for helpful discussions. The main theorem of this note appears as part of the second author's Master's thesis, for which the first and third authors were supervisors. A special thanks goes out to Klaas Landsman, the third supervisor of the project, for his support. PARADOXICALITY FOR GRAPH ALGEBRAS 3 2. Preliminaries A directed graph E is a quadruple (E0, E1, t, s) consisting of countable sets E0 and E1, called the set of vertices and set of edges, respectively, and two functions t, s : E1 → E0 called the target and source maps. Note that some authors, such as [20], use the opposite convention for the maps t and s. In what follows, we are only interested in graphs which are row-finite, as these are the graphs which allow for a groupoid model. A graph E is called row-finite if s−1(v) is a finite set for all v ∈ E0. A sink is a vertex v ∈ E0 such that s−1(v) = ∅ and a source is a vertex v ∈ E0 such that t−1(v) = ∅. 2.1. Cuntz -- Krieger algebras. The typical way of constructing a graph C∗-algebra is to realise it as a universal C∗-algebra from a Cuntz -- Krieger E-family. Definition 2.1. Let E be a row-finite directed graph and H a Hilbert space. A Cuntz -- Krieger E-family {S, P } on H consists of a set (cid:8)Pv : v ∈ E0(cid:9) of mutually orthogonal projections on H and a set (cid:8)Se : e ∈ E1(cid:9) of partial isometries on H, such that the following two conditions, called the Cuntz -- Krieger relations, hold: (1) S∗ e Se = Pt(e) for all e ∈ E1; (2) Pv = P{e∈E1:s(e)=v} SeS∗ e whenever v is not a sink. The graph C∗-algebra C∗(E) is then defined to be the universal C∗-algebra gener- ated by {S, P }, subject to the Cuntz -- Krieger relations. A finite path in E is a sequence µ = (µ1, ..., µk) of k edges, k ∈ N \ {0}, with s(µi+1) = t(µi) for 1 ≤ i ≤ k − 1. We extend the source and target maps by defining s(µ) = s(µ1) and t(µ) = t(µk) and we denote the length of µ by µ = k. If we denote by Ek the set of paths of length k in E, then the elements of E0 (the vertices of E) can be regarded as paths of length 0. A loop (or cycle) is a finite path µ such that s(µ) = t(µ). We say that µ has an exit if there exists 1 ≤ i ≤ µ and e ∈ E1 such that s(αi) = s(e) and αi 6= e. Define E∗ := [n≥0 En, the set of finite paths in E. If µ ∈ E∗ is a finite path, then we denote by Sµ the element Sµ1 Sµ2 · · · Sµk ∈ C∗(E), and we have C∗(E) = span {SµS∗ ν : µ, ν ∈ E∗, t(µ) = t(ν)} . 2.2. The path groupoid. We can also associate a C∗-algebra to a row-finite di- rected graph E via a groupoid. Let G be a locally compact and Hausdorff groupoid with locally compact unit space G(0). Denote the range and domain maps by r, d : G → G(0). The ordered pair (g, h) ∈ G × G is composable if d(g) = r(h) and if so the composition is denoted gh. The set of composable pairs is given by G(2) = {(g, h) ∈ G × G d(g) = r(h)} . The inverse of g ∈ G is denoted g−1. A groupoid is ´etale if r and d are local homeomorphisms. In this case G(0) is an open subset of G and there is a canonical Haar system is given by counting measures. An open subset U ⊂ G is called an open bisection if both dU and rU are homeomorphisms onto their ranges. If G is ´etale and G(0) is totally disconnected, then the groupoid G is said to be ample. Equivalently, an ´etale groupoid is ample if it has a basis of compact open bisections. Let E be a row-finite directed graph. An infinite path is an infinite sequence of edges x1, x2, . . . with the property that s(xi+1) = t(xi) for every i ≥ 1. The infinite path space is defined to be E∞ = (cid:8)(x1, x2, ...) : xi ∈ E1, t(xi) = s(xi+1) ∀i ≥ 1(cid:9) . with µ ∈ E∗, form a basis of open sets. Observe that the cylinder sets are also the product topology for which the cylinder sets The infinite path space is a subset of the product spaceQ∞ Z(µ) = (cid:8)x ∈ E∞ : x1 = µ1, ..., xµ = µµ(cid:9) , (E1 \ {µi}) is open in Q∞ E∞ \ Z(µ) = E∞ ∩  i=1 E1 and (E1 \ {µi})  . closed since Sµ i=1 π−1 i µ [i=1 π−1 i 4 FRANCESCA ARICI, BAUKJE DEBETS, AND KAREN R. STRUNG i=1 E1 and thus inherits The cylinder sets form a basis for a locally compact, σ-compact, totally discon- nected, Hausdorff topology on E∞ [16, Corollary 2.2]. Definition 2.2. For a directed graph E, define the associated path groupoid by GE = {(x, k, y) x, y ∈ E∞, k ∈ Z, ∃N ∈ N with xi = yi+k ∀i ≥ N } , with unit space G(0) ∼= E∞, and domain and range maps d, r : G → G(0) given by d((x, k, y)) = y and r((x, k, y)) = x. Composition and inverse are given by (x, k, y)(y, l, z) = (x, k + l, z), (x, k, y)−1 = (y, −k, x). Observe that (x, k, y) ∈ GE if and only if x = wz and y = vz where w, v are finite paths whose lengths satisfy w + k = v. We could also describe the path groupoid by using the shift map σ : E∞ → E∞ defined by (σx)i = xi+1 for all i ∈ Z+. This shift map is a local homeomorphism and hence (E∞, σ) is a one-sided subshift of finite type over the alphabet E1. The groupoid GE thus arises from equivalence with lag of infinite paths, that is, GE = (cid:8)(x, k, y) x, y ∈ E∞, k ∈ Z, ∃N ∈ N with σN (x) = σN +k(y)(cid:9) . The groupoid GE can be endowed with a topology with respect to which it is a second countable locally compact ´etale Hausdorff groupoid. A basis of this topology is given by Z(α, β) = {(x, k, y) x ∈ Z(α), y ∈ Z(β), k = β − α, xi = yi+k for i > α} . where α, β ∈ E∗ are (possibly empty) paths with t(α) = t(β). It is not hard to check that Z(α, β) is a compact open bisection, hence GE is ample. To any locally compact ´etale groupoid G, we associate its reduced groupoid C∗- algebra as follows. Let Cc(G) denote the (vector space of) compactly supported continuous functions on G. For f1, f2, f ∈ Cc(G) we define multiplication and involution by (f1 · f2)(g) = Xh1h2=g f1(h1)f2(h2), for all g ∈ G, and f ∗(g) = f (g−1). With these operations Cc(G) is a ∗-algebra. For every x ∈ G(0), let ℓ2(d−1(x)) denote the Hilbert space of square-summable functions on d−1(x). From this we can define a ∗-representation by πx : Cc(G) → B(ℓ2(d−1(x))) (πx(f )ξ)(g) = Xh1h2=g f (h1)ξ(h2), PARADOXICALITY FOR GRAPH ALGEBRAS 5 for f ∈ Cc(G), ξ ∈ ℓ2(d−1(x)), g ∈ d−1(x). The reduced groupoid C∗-algebra, denoted C∗ r (G) is the completion of Cc(G) with respected to the norm kf k = sup x∈G(0) kπx(f )k. One may also define a full groupoid C∗-algebra C∗(G). In the case that the groupoid is amenable, as is the case for the groupoid associated to a row-finite directed graph [19, Theorem 4.2], the two coincide. Let GE be the path groupoid of the row-finite directed graph E with no sinks. It is easy to check that for two basis sets Z(α, β), Z(γ, δ) their intersection is Z(α, β), Z(γ, δ) or the empty set. The assumption that E has no sinks implies that for all α, β ∈ E∗ with t(α) = t(β), the sets Z(α), Z(β) and Z(α, β) are nonempty. Thus for any row-finite directed graph E without sinks, C∗(GE) is generated by a Cuntz -- Krieger E-family obtained by looking at characteristic functions on bisections of the path groupoid. Furthermore, as was mentioned above, by [19, Theorem 4.2] the groupoid GE is always amenable. In summary, we have the following: Theorem 2.3 ([16, Section 4]). Let E be a row-finite directed graph without sinks, and let GE be the corresponding path groupoid. Then GE is an ample amenable locally compact Hausdorff groupoid and C∗(GE) ∼= C∗(E). 3. Purely infinite path groupoid C∗-algebras A projection p in a C∗-algebra A is said to be infinite if it is Murray -- von Neu- mann equivalent to a proper subprojection of itself, which is to say, there exists a partial isometry v ∈ A such that v∗v = p and vv∗ (cid:8) p. If A is a simple unital C∗-algebra and 1A is an infinite projection, then A is called purely infinite. The first examples of purely infinite simple C∗-algebras were the Cuntz algebras, On, n ∈ N, and O∞. Purely infinite simple C∗-algebras boast a number of interesting properties: they are traceless, always have real rank zero (and hence always con- tain many projections), every nonzero hereditary C∗-subalgebra contains an infinite projection, every unitary can be approximated by a unitary of finite spectrum, and so [25, 17] (see also [22, Proposition 4.1.1] for a further list). The Cuntz algebra O∞ is particularly important among simple, separable, unital, purely infinite C∗- algebras. If A is any simple, separable, nuclear, unital C∗-algebra, then A ⊗ O∞ is always purely infinite. Moreover if A is a separable, nuclear, unital, purely infinite C∗-algebra, then A ∼= A ⊗ O∞ [12]. 3.1. Groupoid C∗-algebras and Property (IH). A groupoid G is called topo- logically principal (or essentially free) if the set of points with trivial isotropy is dense in G(0). We say that an ´etale groupoid G is locally contracting if, for every nonempty open subset U ⊂ G(0), there exist an open subset V ⊂ U and an open bisection S with V ⊂ d(S) and d(V S−1) $ V . If G is a topologically principal ´etale groupoid which is locally contracting, then r (G) has property (IH): every nonzero hereditary r (G) contains an infinite projection [1, Proposition 2.4]. When the reduced groupoid C∗-algebra C∗ C∗-subalgebra of C∗ a C∗-algebra A is simple, property (IH) is equivalent to pure infiniteness [25]. In the original statement of the theorem below, the graph E is also assumed to be locally finite. This was required at the time for the groupoid to be amenable. Since this is no longer the case and local finiteness is not required elsewhere in their proof, we use the reformulation below. See also [3, Proposition 5.3], where the statement is proved in the Cuntz -- Krieger model. Theorem 3.1 ([15, Theorem 3.9]). Let E be a directed graph with no sinks. Then C∗(E) has property (IH) if and only if every vertex connects to a loop and every loop has an exit. 6 FRANCESCA ARICI, BAUKJE DEBETS, AND KAREN R. STRUNG The proof amounts to showing that the path groupoid is topologically principal and locally contracting if and only if every vertex connects to a loop and every loop has an exit. For a unit x ∈ G(0) E = E∞, one shows that non-trivial isotropy corresponds to eventual periodicity [15, Lemma 3.2]. It follows that to show that GE is topologically principal, it is enough to show that every vertex is the source of an aperiodic path, since this in turn implies that every basic set contains an aperiodic path. By [15, Lemma 3.4] this occurs exactly when every loop has an exit. This allows one to construct a path starting at a vertex v that never passes through the same vertex twice, hence is aperiodic. Conversely, if there is a loop without an exit, then it is straightforward to find an eventually periodic path. The locally contracting condition is satisfied if every vertex connects to a vertex that has a return path with an exit. If there is a vertex which does not connect to a loop, one shows that C∗(GE) has a hereditary C∗-subalgebra which is approximately finite (AF). Similarly, if there is a loop that does not have an exit one can show that C∗(GE) has a hereditary C∗-subalgebra that is Morita equivalent to a commutative C∗-algebra and hence does not have property (IH). In particular, C∗(GE) does not have property (IH). 3.2. Non-simple purely infinite C∗-algebras. If C∗(GE) is simple, which is the case if and only if every loop in E has an exit and E is cofinal (see [3, Proposition 5.1]), then the above implies C∗(GE) is purely infinite. However, for the non-simple case, we require a stronger condition. Extending the definition of pure infiniteness from simple C∗-algebras to non- simple C∗-algebras is a subtle matter. One needs to decide which key properties of pure infiniteness should hold in the non-simple case, such as O∞-absorption and admitting no nonzero traces. To do so, one needs to consider subequivalence of positive elements, as introduced by Cuntz in [8]. Let A be a C∗-algebra and a, b ∈ A with a, b ≥ 0. We say that a is Cuntz-subequivalent to b, written a - b, if there exists a sequence (rn)n∈N ⊂ A such that krnbr∗ n − ak → 0 as n → ∞. More generally, if a ∈ Mn(A)+ and b ∈ Mm(A)+, then we define a - b if there exists a sequence of rectangular matrices (rn)n∈N ⊂ Mn,m(A) such that kr∗ nbrn − ak → 0 as n → ∞. If a, b ∈ A then we define the direct sum of a and b by a ⊕ b = (cid:18) a 0 b (cid:19) ∈ M2(A). 0 The definition below is due to Kirchberg and Rørdam [13, Definition 4.1, Theo- rem 4.16]. Definition 3.2. Let A be a C∗-algebra. We say A is purely infinite if a ⊕ a - a for every nonzero positive element a ∈ A. This is already enough to imply that if A is purely infinite, then A admits no nonzero traces and that A ⊗ O∞ is always purely infinite. On the other hand, some other properties one might expect are not automatic. However, if A is separable, nuclear, and has real rank zero then the situation begins to look a lot more like the simple case; for example A is purely infinite if and only if A ⊗ O∞ is purely infinite, if and only if all nonzero projections in A are properly infinite [18]. We will see below that real rank zero is automatic in the setting of graph algebras. Let E be a row-finite directed graph with no sinks. For all vertices v, w ∈ E0, we write v ≥ w if there exists a finite path α ∈ E∗ such that s(α) = v and t(α) = w. Obviously for all v ∈ E0 we have v ≥ v (letting α be the empty path) and for all u, v, w ∈ E0 it follows that u ≥ v and v ≥ w imply u ≥ w. Thus ≥ is a preorder. Note that it is not a partial order as there might be two vertices v 6= w on a cycle such that v ≥ w ≥ v. PARADOXICALITY FOR GRAPH ALGEBRAS 7 Definition 3.3. Let M ⊆ E0 be a non-empty subset. Then we call M a maximal tail if the following three conditions hold: (1) If v ∈ E0, w ∈ M , and v ≥ w, then v ∈ M ; (2) If v ∈ M and s−1(v) 6= ∅, then there exists e ∈ E1 with s(e) = v and t(e) ∈ M ; (3) For every v, w ∈ M there exists y ∈ M such that v ≥ y and w ≥ y. Let E be a row-finite directed graph. We define V 2 to be the set of vertices v for which there are at least two distinct finite cycles based at v, that is, V 2 = {v ∈ E0 there are cycles µ 6= ν with t(µi) = t(νj ) = v if and only if i = µ and j = ν}. Similarly, we define V 1 to be the set of vertices that lie exactly in one cycle, and V 0 to be the set of vertices v for which there is no cycle based at v. Definition 3.4 ([16, Section 6]). A graph E satisfies Condition (K) if V 1 = ∅, or, equivalently, if E0 = V 0 ∪ V 2. Note that condition (K) generalises Condition (II) in [6]. Indeed, a finite directed graph satisfies condition (K) if and only if the associated incidence matrix satisfies condition (II). For a row-finite directed graph E we have the following result, due to Hong and Szymanski. Note that it does not require that every loop in E has an exit or that E is cofinal; in particular, the result holds for non-simple graph C∗-algebras. Theorem 3.5 ([10, Theorems 2.3, 2.5]). Let E be a row-finite directed graph. Then the following are equivalent: (1) C∗(E) is purely infinite; (2) C∗(E) is purely infinite and has real rank zero; (3) all loops in each maximal tail M have exits in M and each vertex in every maximal tail of M connects to a loop in M ; (4) E satisfies Condition (K) and each vertex in every maximal tail of M con- nects to a loop in M . The goal of this section is to interpret the theorem above in terms of the path groupoid. Unlike for graph C∗-algebras, for an arbitrary ´etale groupoid necessary and sufficient conditions for pure infiniteness are not known. This question was recently addressed in [4]. There, the authors establish a sufficient condition on an ample groupoid G that ensures pure infiniteness of the reduced C∗-algebra C∗ r (G) by extending the notion of paradoxical decompositions for actions of discrete groups on totally disconnected spaces to the setting of ´etale groupoids. Here, we define paradoxicality for the path groupoid of a row-finite directed graph with no sinks. Definition 3.6. For a graph E and a finite path µ ∈ E∗, we say that the cylinder set Z(µ) is paradoxical if there exist numbers m, n ∈ N and compact open bisections Z(α1, β1), Z(α2, β2), . . . , Z(αn, βn), Z(γ1, δ1), Z(γ2, δ2), . . . , Z(γm, δm), such that: n [i=1 Z(βi) = m [j=1 Z(δi) = Z(µ), and the sets Z(αi), Z(γj) ⊆ Z(µ) are pairwise disjoint. In the language of [4] this says that Z(µ) is "(Ga, 2, 1)-paradoxical". It is easy to see that for all µ ∈ E∗, Z(µ) is paradoxical if and only if Z(r(µ)) is paradoxical. We will need a stronger version of topological freeness. Let G be a locally compact groupoid. A subset D ⊂ G(0) is called invariant if for any g ∈ G, d(g) ∈ G implies 8 FRANCESCA ARICI, BAUKJE DEBETS, AND KAREN R. STRUNG r(g) ∈ G. Let GD := {g ∈ G d(g) ∈ D}. When G is ´etale and D ⊂ G(0) is closed and invariant, then GD is a closed ´etale subgroupoid. We say G is essentially principal if, for every closed invariant subset D ⊂ G(0), the subgroupoid GD is topologically principal. Proposition 3.7. Let E be a row-finite graph without sinks. Suppose that GE is essentially principal. Then C∗(GE) is purely infinite if the cylinder set Z(v) is paradoxical for every v ∈ E0. Proof. The path groupoid GE is ample and by assumption essentially principal. Furthermore, since GE is amenable, [5, Lemma 6.1] and [5, Remark 6.2] imply that it is inner exact in the sense of [4, Definition 3.5]. Since (cid:8)Z(v) v ∈ E0(cid:9) is a basis for the topology of G(0) ∼= E∞, it follows from [4, Corollary 4.12] that C∗(GE) is purely infinite if Z(v) is paradoxical for every v ∈ E0. (cid:3) We would like a condition on the graph E that implies paradoxicality of Z(v) for every v ∈ V . To that end, we introduce the following conditions. Definition 3.8. Let E be a row-finite directed graph with no sinks. (1) The graph E satisfies Condition (I) if, for every v ∈ E0, there exists a finite path α ∈ E∗ with s(α) = v and t(α) ∈ V 2. (2) The graph E satisfies Condition (DI) if, for every v ∈ E0, there exists a decomposition Z(v) = ⊔n i=1Z(βi) such that for every i = 1, . . . , n there exists a path αi ∈ E∗ with t(αi) = t(βi), s(αi) = v, and αi passes through V 2. (3) The graph E satisfies Condition (DL) if, for every v ∈ E0, there exists a decomposition Z(v) = ⊔n i=1Z(βi) such that for every i = 1, . . . , n there exists a path αi ∈ E∗ with t(αi) = t(βi), s(αi) = v, and αi passes through a loop. Condition (I) was first defined in [9] by Cuntz and Krieger for the Cuntz -- Krieger algebras associated to an n × n {0, 1}-matrix. Note that (DI) implies (I) as well as (DI) implies (DL), but neither converse need hold. Let E be a row-finite directed graph without sinks. We have already seen that Condition (K) holds whenever C∗(GE) ∼= C∗(E) is purely infinite. We show in the sequel that (I), (DI), and (DL) are also necessary, and that when combined with Condition (K), either Condition (DI) or condition (DL) is enough to establish that C∗(GE) is purely infinite. Con- dition (I), on the other hand, is not enough, even in the presence of Condition (K). Condition (K) implies that GE is essentially principal [16, Proposition 6.3] and thus is required for our main result, Theorem 3.14. In fact, Condition (K) is also necessary to ensure that for every v ∈ E0, the cylinder set Z(v) is paradoxical. Indeed, suppose that (K) does not hold for E. Then there exists a cycle µ in some v with no other return path. If Z(v) has a paradoxical decomposition, then x = µµµ · · · ∈ Z(βi) ∩ Z(δj), where βi = µ . . . µ µ1 . . . µl and δj = µ . . . µ µ1 . . . µm, k times {z } n times {z } for some k, l, n, m. Then αi, γj must be paths from v to vertices on µ, but these paths must be subpaths of x, for otherwise we would have two distinct return paths in v. Thus Z(αi) ∩ Z(γj) = Z(µ1 . . . µp) 6= ∅ for some p ∈ N. Hence Z(v) cannot have a paradoxical decomposition. In view of this, let us examine Conditions (I), (DI) and (DL) in the presence of Condition (K). PARADOXICALITY FOR GRAPH ALGEBRAS 9 Proposition 3.9. Let E be a row-finite directed graph without sinks. Suppose that E satisfies Condition (K). Then if the cylinder set Z(v) is paradoxical for every v ∈ E0, the graph E satisfies Condition (I). Proof. Assume that the cylinder set Z(v) is paradoxical for every v ∈ E0 but that Condition (I) does not hold. Then there must be a vertex v that is not connected to V 2, and because Condition (K) holds, v is not connected to V 1 either. Following the proof of [15, Theorem 3.9] (see also the discussion following Theorem 3.1 in the previous subsection), let H be the subgraph of E formed by those vertices that can be reached from v. It is clear that H has no loops and no exits. By [16, Theorem 6.6], there exists an isomorphism between the lattice of ideals in the groupoid C∗ algebra C∗(GE) and the lattice of saturated subsets of E0 (see [16, Section 6]). In particular, by [15, Theorem 3.7] and [15, Proposition 2.1], C∗(H) is a hereditary C∗-subalgebra of C∗(GE). Since H has no loops it is AF, and so C∗(GE) cannot be purely infinite. Thus E has no paradoxical decomposition in every vertex. Therefore Condition (I) must hold. (cid:3) On the other hand, Condition (I) is not sufficient. Let E be the following graph: f1 • v f2 • • e1 f3 • • e2 . Both Conditions (I) and (K) hold for this graph. Suppose now that Z(v) has a paradoxical decomposition, where v is the vertex on the bottom left (as denoted). Then ¯e = e1e2 · · · ∈ Z(βi) ∩ Z(δj) for some i and j, but then we must have βi = e1 · · · en and δj = e1 · · · em for some n, m ∈ N. Thus αi = e1 · · · en and βj = e1 · · · em as well, so Z(αi) ∩ Z(γj) 6= ∅. Hence Z(v) has no paradoxical decomposition, so (I) is not a sufficient condition for paradoxicality. Condition (DI) is a strengthening of Condition (I). In the example above, we constructed a graph with a path containing only vertices in V 0 to show that (I) was not sufficient to paradoxical decompositions of the cylinder sets. Thus, one might suppose that a suitable strengthening of (I) might be given by the property that every path must pass through V 2. However, all cylinder sets can be paradoxical in absence of this condition, as we see in the following example. Consider the graph E pictured below. f e w α1 β1 v β2 • • • Observe that E contains paths that do not pass through V 2. However E satisfies (DL). Indeed, for every vertex u ∈ E0, the set (cid:8)x ∈ E≤∞ s(x) = u(cid:9) is either isomorphic to (cid:8)x ∈ E≤∞ s(x) = v(cid:9) or to (cid:8)x ∈ E≤∞ s(x) = w(cid:9). Thus we need only show that there are paradoxical decompositions for Z(v) and Z(w), and these are easily seen to exist: For Z(v) take Z(β1e, β1), Z(α1, α1) and Z(β1f, β1), Z(β1β2, α1); 10 FRANCESCA ARICI, BAUKJE DEBETS, AND KAREN R. STRUNG and for Z(w) take Z(ee, e), Z(f f, f ), Z(β2, β2) and Z(ef, e), Z(f e, f ), Z(eβ2, β2). Note however, E does satisfy (DI), and indeed, as the next proposition shows, this is precisely the condition we are after. Proposition 3.10. Let E be a row-finite directed graph without sinks that satisfies (K). Then E satisfies Condition (DI) if and only if Z(v) is paradoxical for all v ∈ E0. i , βi) such that Z(αk Proof. Suppose E satisfies Condition (DI). There are decompositions Z(α1 i , βi) and Z(α2 j) = ∅ for all i, j ≤ n and k, l = 1, 2. Since all αi pass through V 2 and for w ∈ V 2 we know that for all m ∈ N there exists µ1, . . . , µm ∈ E∗ with s(µi) = t(µi) = w and Z(µi) ∩ Z(µj) = ∅ for all i 6= j. i ) ∩ Z(αl For the other direction, suppose E does not satisfy (DI). Then for all n ∈ N the set Pn defined by (cid:8)γ ∈ En : s(γ) = v and ∄α that passes through V 2 with s(α) = v, t(α) = t(γ)(cid:9) is non-empty and for all µ ∈ Pn+1, we have µ1 . . . µn ∈ Pn. Thus by Zorn's Lemma, there exists an infinite path x ∈ E∞ such that there is no path from v via a vertex in V 2 to t(xn) for some n ∈ N. However, that means that M = (cid:8)u ∈ E0 : u ≥ t(xn) for some n ∈ N(cid:9) is a maximal tail, where v does not connect to a loop in M . Therefore the algebra C∗(E) ∼= C∗(GE) is not purely infinite, and hence by [4, Corollary 4.12] the cylinder set Z(v) is not paradoxical for all v ∈ E0. (cid:3) The next proposition will allow us to examine the role of Condition (DL). Proposition 3.11. Let E be a row-finite directed graph without sinks, and let v ∈ E0. The following are equivalent. (1) There exists a decomposition Z(v) = ⊔n i=1Z(βi) such that for every i = 1, . . . , n there exists a path αi ∈ E∗ with t(αi) = t(βi), s(αi) = v, and αi passes through a loop. (2) If M is a maximal tail containing v, then there is a loop in M that v is connected to. Proof. Suppose (1) holds and let M be a maximal tail with v ∈ M . Then by Definition 3.3 (2), there exists an infinite path x ∈ E∞ with s(x) = v and t(xn) ∈ M for all n ∈ N. We know that Z(v) = n Gi=1 Z(βi), so x ∈ Z(βi) for some 1 ≤ i ≤ n. Hence t(βi) = t(αi) ∈ M , and then by Definition 3.3 (1), every vertex on αi is in M and αi passes through a loop. So that loop is in M as well, and lastly v connects to that loop. Thus (1) implies (2). For the converse, (1) does not hold, that is, there is no decomposition Z(v) = ⊔n i=1Z(βi) satisfying the requirements of Definition 3.8 (3). Using Zorn's lemma, once again, there exists an infinite path x ∈ E∞ with s(x) = v, such that there is no path from v via a loop to t(xn) for some n ∈ N. Then M = (cid:8)u ∈ E0 : u ≥ t(xn) for some n ∈ N(cid:9) is again a maximal tail, where v does not connect to a loop. Thus (2) implies (1). (cid:3) Corollary 3.12. Let E be a row-finite directed graph without sinks. The graph E satisfies Conditions (K) and (DL) if and only if C∗(GE) is purely infinite. PARADOXICALITY FOR GRAPH ALGEBRAS 11 Proof. By the previous proposition (DL) is equivalent to the property that every vertex in every maximal tail of M connects to a loop in M . Thus the result follows from the equivalence of (1) and (4) of Theorem 3.5, (cid:3) As we observed earlier, Condition (DI) implies Condition (DL), but the reverse does not necessarily hold. However, in the presence of Condition (K), they are equivalent. Corollary 3.13. Let E be a row-finite directed graph without sinks that satisfies Condition (K). Then E satisfies Condition (DI) if and only if E satisfies (DL). Proof. We only need to show that Condition (DL) implies (DI), but this is imme- diate since (K) implies that at any vertex with a loop, there is a second loop. (cid:3) Finally, as a summary of the above, we come to the main theorem. Theorem 3.14. Let E be a row-finite directed graph without sinks. The following are equivalent: (1) C∗(GE) is purely infinite; (2) GE is essentially principal and for every finite path α, the cylinder set Z(α) is paradoxical; (3) E satisfies Conditions (K) and (DI); (4) E satisfies Conditions (K) and (DL). Proof. The implication (2) implies (1) is given by Proposition 3.7. The equivalence of (3) and (4) is Corollary 3.13. If (3) holds, then Condition (K) implies that GE is essentially principal and since Condition (DI) also holds, the cylinder set of any finite path is paradoxical by Proposition 3.10. Thus (3) implies (2). Finally, the equivalence of (4) and (1) follows directly from Corollary 3.12. (cid:3) Observe that if C∗(GE) satisfies any of the equivalent conditions of Theorem 3.14, then it automatically has real rank zero by Theorem 3.5. Thus each of (1) -- (4) is also equivalent to the following properties: (5) C∗(GE) ⊗ O∞ ∼= C∗(GE); (6) C∗(GE) is strongly purely infinite (see [14, Definition 5.1]); (7) every nonzero hereditary C∗-subalgebra in any quotient of C∗(GE) contains an infinite projection [18, Proposition 2.11]. Unlike the Hong -- Szymanski result of Theorem 3.5, we require the assumption that E has no sinks. This comes from the fact that the path groupoid is only constructed for a row-finite graph without sinks. However, one can deal with graphs with sinks by adding tails, a technique introduced in [3]. Given a graph E with sinks, we denote by F the graph obtained by adding a tail to every sink. Then C∗(GE) ∼= C∗(E) is isomorphic to a full corner in C∗(F ) [3]. In particular, C∗(E) is purely infinite if and only if C∗(F ) is purely infinite. Finally, it also worth noting that the equivalence of (1) and (2) in the main theorem is a stronger result than what one has for an arbitrary ample amenable ´etale groupoid G. In that case, if G is essentially principal and has a basis for the topology consisting of (Ga, 2, 1)-paradoxical sets, then C∗(G) is (strongly) purely infinite by [4, Corollary 4.12]. However, unlike for the path groupoid, the converse is not known. References [1] Claire Anantharaman-Delaroche. Purely infinite C ∗-algebras arising from dynamical systems. Bull. Soc. Math. France, 125(2):199 -- 225, 1997. [2] Teresa Bates, Jeong Hee Hong, Iain Raeburn, and Wojciech Szyma´nski. The ideal structure of the C ∗-algebras of infinite graphs. Illinois J. Math., 46(4):1159 -- 1176, 2002. 12 FRANCESCA ARICI, BAUKJE DEBETS, AND KAREN R. STRUNG [3] Teresa Bates, David Pask, Iain Raeburn, and Wojciech Szyma´nski. The C ∗-algebras of row- finite graphs. New York J. Math., 6:307 -- 324, 2000. [4] Christian Bonicke and Kang Li. Ideal structure and pure infiniteness of ample groupoid C ∗ -algebras. Ergodic Theory Dynam. Systems, pages 1 -- 30, 2018. [5] Jonathan Brown, Lisa Orloff Clark, and Adam Sierakowski. Purely infinite C ∗-algebras as- sociated to ´etale groupoids. Ergodic Theory Dynam. Systems, 35(8):2397 -- 2411, 2015. [6] Joachim Cuntz. Simple C ∗-algebras generated by isometries. Comm. Math. Phys., 57(2):173 -- 185, 1977. [7] Joachim Cuntz. The structure of multiplication and addition in simple ∗-algebras. Math. Scand., 40(2):215 -- 233, 1977. [8] Joachim Cuntz. Dimension functions on simple Cspec∗-algebras. Math. Ann., 233(2):145 -- 153, 1978. [9] Joachim Cuntz and Wolfgang Krieger. A class of C ∗-algebras and topological Markov chains. Invent. Math., 56(3):251 -- 268, 1980. [10] Jeong Hee Hong and Wojciech Szyma´nski. Purely infinite Cuntz-Krieger algebras of directed graphs. Bull. London Math. Soc., 35(5):689 -- 696, 2003. [11] David Kerr and Piotr W. Nowak. Residually finite actions and crossed products. Ergodic Theory Dynam. Systems, 32(5):1585 -- 1614, 2012. [12] Eberhard Kirchberg and N. Christopher Phillips. Embedding of exact C ∗-algebras in the Cuntz algebra O2. J. Reine Angew. Math., 525:17 -- 53, 2000. [13] Eberhard Kirchberg and Mikael Rørdam. Non-simple purely infinite C ∗-algebras. Amer. J. Math., 122(3):637 -- 666, 2000. [14] Eberhard Kirchberg and Mikael Rørdam. Infinite non-simple C ∗-algebras: absorbing the Cuntz algebras O∞. Adv. Math., 167(2):195 -- 264, 2002. [15] Alex Kumjian, David Pask, and Iain Raeburn. Cuntz -- Krieger algebras of directed graphs. Pacific J. Math., 184(1):161 -- 174, 1998. [16] Alex Kumjian, David Pask, Iain Raeburn, and Jean Renault. Graphs, groupoids, and Cuntz- Krieger algebras. J. Funct. Anal., 144(2):505 -- 541, 1997. [17] Hua Xin Lin and Shuang Zhang. On infinite simple C ∗-algebras. J. Funct. Anal., 100(1):221 -- 231, 1991. [18] Cornel Pasnicu and Mikael Rørdam. Purely infinite C ∗-algebras of real rank zero. J. Reine Angew. Math., 613:51 -- 73, 2007. [19] Alan L. T. Paterson. Graph inverse semigroups, groupoids and their C ∗-algebras. J. Operator Theory, 48(3, suppl.):645 -- 662, 2002. [20] Iain Raeburn. Graph algebras, volume 103 of CBMS Regional Conference Series in Mathe- matics. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2005. [21] Jean Renault. A groupoid approach to C ∗-algebras, volume 793 of Lecture Notes in Mathe- matics. Springer, Berlin, 1980. [22] M. Rørdam. Classification of nuclear, simple C ∗-algebras. In Classification of nuclear C ∗- algebras. Entropy in operator algebras, volume 126 of Encyclopaedia Math. Sci., pages 1 -- 145. Springer, Berlin, 2002. [23] Mikael Rørdam and Adam Sierakowski. Purely infinite C ∗-algebras arising from crossed prod- ucts. Ergodic Theory Dynam. Systems, 32(1):273 -- 293, 2012. [24] Yasuo Watatani. Graph theory for C ∗-algebras. In Operator algebras and applications, Part I (Kingston, Ont., 1980), volume 38 of Proc. Sympos. Pure Math., pages 195 -- 197. Amer. Math. Soc., Providence, R.I., 1982. [25] Shuang Zhang. A property of purely infinite simple C ∗-algebras. Proc. Amer. Math. Soc., 109(3):717 -- 720, 1990. Institute for Mathematics, Astrophysics, and Particle Physics, Radboud University, Postbus 9010, 6500 GL Nijmegen, The Netherlands E-mail address: [email protected] Section of Analysis, Department of Mathematics, KU Leuven, Celestijnenlaan 200b - box 2400, 3001 Leuven, Belgium E-mail address: [email protected] Institute for Mathematics, Astrophysics, and Particle Physics, Radboud University, Postbus 9010, 6500 GL Nijmegen, The Netherlands E-mail address: [email protected]
1904.00292	1	1904	2019-03-30T21:25:24	Inductive Limits for Systems of Toeplitz Algebras	[ "math.OA", "math-ph", "math.FA", "math-ph" ]	This article deals with inductive systems of Toeplitz algebras over arbitrary directed sets. For such a system the family of its connecting injective $$-homomorphisms is defined by a set of natural numbers satisfying a factorization property. The motivation for the study of those inductive systems comes from our previous work on the inductive sequences of Toeplitz algebras defined by sequences of numbers and the limit automorphisms for the inductive limits of such sequences. We show that there exists an isomorphism in the category of unital $C^$-algebras and unital $$-homomorphisms between the inductive limit of an inductive system of Toeplitz algebras over a directed set defined by a set of natural numbers and a reduced semigroup $C^$-algebra for a semigroup in the group of all rational numbers. The inductive systems of Toeplitz algebras over arbitrary partially ordered sets defined by sets of natural numbers are also studied.	math.OA	math	INDUCTIVE LIMITS FOR SYSTEMS OF TOEPLITZ ALGEBRAS R.N. GUMEROV Abstract. This article deals with inductive systems of Toeplitz algebras over arbitrary directed sets. For such a system the family of its connecting injective ∗-homomorphisms is defined by a set of natural numbers satisfying a factoriza- tion property. The motivation for the study of those inductive systems comes from our previous work on the inductive sequences of Toeplitz algebras defined by sequences of numbers and the limit automorphisms for the inductive limits of such sequences. We show that there exists an isomorphism in the category of unital C ∗-algebras and unital ∗-homomorphisms between the inductive limit of an inductive system of Toeplitz algebras over a directed set defined by a set of natural numbers and a reduced semigroup C ∗-algebra for a semigroup in the group of all rational numbers. The inductive systems of Toeplitz algebras over arbitrary partially ordered sets defined by sets of natural numbers are also studied. 1. Introduction The main part of motivation for the present article comes from our work on inductive sequences of Toeplitz algebras and limit automorphisms of C ∗-algebras generated by isometric representations for semigroups of rational numbers (see [1]). We transfer some results from [1] concerning inductive sequences of Toeplitz alge- bras defined by sequences of numbers to the case of inductive systems over arbitrary directed sets defined by sets of natural numbers satisfying a factorization property. In turn, the results in [1] are closely related to those in [2, 3, 4, 5, 6, 7] which are devoted to mappings of compact topological groups. A part of motivation for studying inductive systems of C ∗-algebras comes from algebraic quantum field theory [8, 9, 10, 11]. The general framework of algebraic quantum field theory is given by a covariant functor. Usually that functor acts from a category associated to a partially ordered set into a category describing the alge- braic structure of observables. The standard assumption in quantum physics is that the second category consists of unital C ∗-algebras and unital ∗-homomorphisms be- tween C ∗-algebras. Thus one has an inductive system of C ∗-algebras. A simple example of an inductive system F = (K, {Aa}, {σba}) of C ∗-algebras over a directed set ( K, ≤ ) is that in which {Aa a ∈ K} is a net of C ∗-subalgebras of a given C ∗-algebra A. By this, one means that each Aa is a C ∗-subalgebra containing the unit IA of the algebra A, Aa ⊂ Ab and σba : Aa −→ Ab is the inclusion mapping whenever a, b ∈ K and a ≤ b. Given such a net F , the norm 2010 Mathematics Subject Classification. 46L05, 47L40, 81T05 . Key words and phrases. C ∗-algebra, directed set, inductive limit, inductive system, partially ordered set, reduced semigroup C ∗-algebra, semigroup, Toeplitz algebra. The research was funded by the subsidy allocated to Kazan Federal University for the state assignment in the sphere of scientific activities, project 1.13556.2019/13.1. 1 2 R.N. GUMEROV closure of the union of all Aa is itself a C ∗-subalgebra of A which is a simple example of an inductive limit in the category of C ∗-algebras. The basic tool of the algebraic approach to quantum fields over a spacetime is a net of C ∗-algebras over a set defined as a suitable set of regions of the spacetime ordered under inclusion [8, 9, 10]. In [12, 13, 14] nets consisting of C ∗-algebras of quantum observables for the case of curved spacetimes are studied. The paper [15] deals with a net that is constructed by means of the semigroup C ∗-algebra generated by the path semigroup for a partially ordered. In [16] the authors consider a net of C ∗-algebras associated to a net over a partially ordered set consisting of Hilbert spaces. In this article we study inductive limits for systems of Toeplitz algebras over arbitrary directed sets. Here, by the Toeplitz algebra we mean the reduced semi- group C ∗-algebra for the additive semigroup of non-negative integers. We recall that the reduced semigroup C ∗-algebra can be constructed for an arbitrary left cancellative semigroup. This algebra is a very natural object because it is gener- ated by the left regular representation of the left cancellative semigroup.The study of such semigroup C ∗-algebras goes back to L. A. Coburn [18, 19], R. G. Douglas [20], G. J. Murphy [21, 22] and is continued at the present time (see, for example, [23, 24] and references there in). To define a family of connecting injective ∗-homomorphisms for an inductive system of Toeplitz algebras over a directed set we make use of a set of natural num- bers satisfying factorization equalities (see Section 3, Definition 3.1) and Coburn's Theorem [17, Theorem 3.5.18]. The main result, Theorem 4.1, states that the inductive limit for an inductive system of Toeplitz algebras over a directed set defined by a set of natural numbers is isomorphic in the category of unital C ∗-algebras and their unital ∗-homomorphisms to a reduced semigroup C ∗-algebra for a semigroup in the group of all rational numbers. The article is organized as follows. It consists of Introduction, Preliminaries and three more sections containing the results. Section 3 deals with the auxiliary statements that are used for proving the main result. In Section 4 is devoted to the proof of Theorem 4.1. Section 5 contains the theorem on inductive systems of Toeplitz algebras over arbitrary partially ordered sets defined by sets of natural numbers satisfying factorization equalities. The results in this article were announced without proofs in [25] and are closely related to those in [26]. 2. Preliminaries As usual, N, Z, Q and C denote the set of all natural numbers, the additive group of all integers, the additive group of all rational numbers and the field of all complex numbers respectively. For a sequence of numbers M = (m1, m2, . . .), where ms ∈ N, s ∈ N, we shall consider the subgroup QM of the group Q defined as follows: QM :=(cid:26) m m1m2 . . . ms m ∈ Z, s ∈ N(cid:27) . It is known (see, for example, [22, Proposition 1], [28, Lemma 1]) that the group QM is the inductive (direct) limit in the category of groups and their homomorphisms INDUCTIVE LIMITS FOR SYSTEMS OF TOEPLITZ ALGEBRAS 3 for the following inductive (direct) sequence (2.1) Z τ1−−−−→ Z τ2−−−−→ Z τ3−−−−→ . . . , where the connecting homomorphisms τs are given by τs(m) = msm, m ∈ Z, s ∈ N. By a directed set we mean an upward directed partially ordered set. We recall the definition of reduced semigroup C ∗-algebras for semigroups in Q. To do this we assume that Γ is an arbitrary subgroup in the group Q. The positive cone in the ordered group Γ is denoted by the symbol Γ+ := Γ ∩ [0, +∞) . As usual, the symbol l2(Γ+) stands for the Hilbert space of all square summable complex-valued functions on the additive semigroup Γ+: l2(Γ+) := {f : Γ+ → C : Xγ∈Γ+ f (γ)2 < +∞}. Recall that the inner product in the space l2(Γ+) is given by the formula < f, g >:= Pγ∈Γ+ f (γ)g(γ). The canonical orthonormal basis in the Hilbert space l2(Γ+) is denoted by { eg g ∈ Γ+ }. That is, for arbitrary elements g, h ∈ Γ+, we set eg(h) = δg,h, where δg,h :=(1, 0, if g = h ; if g 6= h . Let us consider the C ∗-algebra of all bounded linear operators B(l2(Γ+)) in the Hilbert space l2(Γ+). For every element g ∈ Γ+, we define the isometry Vg ∈ B(l2(Γ+)) by Vgeh := eg+h, where h is an element of the semigroup Γ+. We denote by C ∗ r (Γ+) the C ∗-subalgebra in the C ∗-algebra B(l2(Γ+)) generated r (Γ+) is called the reduced by the set of isometries { Vg g ∈ Γ+ }. The algebra C ∗ semigroup C ∗-algebra of the semigroup Γ+, or the Toeplitz algebra generated by Γ+. As was mentioned in Introduction, in the similar way a semigroup C ∗-algebra can be defined for an arbitrary left cancellative semigroup (see, for example, [24, Section 2]). In the case when Γ is the group Z, we also denote the semigroup r (Z+) by T and use the symbols T and T n instead of V1 and Vn, C ∗-algebra C ∗ respectively, where n ∈ Z+. The following statement is an immediate consequence of Coburn's theorem [17, Theorem 3.5.18]. Lemma 2.1. For every number n ∈ N, there exists a unique unital ∗-homo- morphism of C ∗-algebras ϕ : T −→ T such that ϕ(T ) = T n. Moreover, ϕ is isometric. In the sequel, we abbreviate those self-homomorphisms of the Toeplitz algebra as follows: ϕ : T −→ T : T 7−→ T n. We note that a straightforward proof of Lemma 2.1 is also contained in [29, Propo- sition 3]. For necessary results in the theory of C ∗-algebras we refer the reader, for example, to books [17] and [30, Ch. 4, § 7]. Further, we recall the definition and some facts concerning the inductive limits for inductive systems of C ∗-algebras (see, for example, [31, Section 11.4]). Necessary facts from the theory of categories and functors are contained, for example, in [30, Ch. 0, § 2] and [32]. 4 R.N. GUMEROV In what follows, up to Section 5, we shall consider a directed set ( K, ≤ ). Note that in Section 5 we will deal with an arbitrary partially ordered set ( K, ≤ ) that is not necessarily directed. The category associated to the set ( K, ≤ ) is denoted by the same letter K. We recall that the objects of this category are the elements of the set K, and for any pair a, b ∈ K the set of morphisms M orK(a, b) from a to b is defined as follows: M orK(a, b) =({(a, b)}, ∅, if a ≤ b ; otherwise . Let us take a covariant functor F from the category K into the category of unital C ∗-algebras and their unital ∗-homomorphisms. Such a functor is called an inductive system in the category of C ∗-algebras over the set (K, ≤ ). It may be given by a collection (K, {Aa}, {σba}) satisfying the properties from the definition of a functor. We shall write F = (K, {Aa}, {σba}). Here, {Aa a ∈ K} is a family of unital C ∗-algebras. We also suppose that all morphisms σba : Aa −→ Ab, where a ≤ b, are embeddings of C ∗-algebras, i. e., unital injective ∗-homomorphisms. Recall that the diagram Aa ❆❆❆❆❆❆❆❆ σba σca Ab Ac >⑥⑥⑥⑥⑥⑥⑥⑥ σcb is commutative for all elements a, b, c ∈ K satisfying the conditions a ≤ b and b ≤ c, that is, the following equation holds: (2.2) σca = σcb ◦ σba. Furthermore, for each element a ∈ K the morphism σaa : Aa −→ Aa is the identity In the case of a countable set K the system F is called an inductive mapping. sequence of C ∗-algebras. The inductive limit of the system F = (K, {Aa}, {σba}) over a directed set K is a pair (A, {σK a : Aa → A a ∈ K} is a family of unital injective ∗-homomorphisms such that the following two properties are fulfilled [31, Proposition 11.4.1]: a }), where A is a C ∗-algebra and {σK 1) For every pair of elements a, b ∈ K satisfying the condition a ≤ b the diagram Aa Ab σba ❆❆❆❆❆❆❆❆ σK a ~⑦⑦⑦⑦⑦⑦⑦⑦ σK b A is commutative, that is, the equality for mappings (2.3) σK a = σK b ◦ σba holds. Moreover, we have the following equality: (2.4) σK a (Aa), A = [a∈K where the bar means the closure of the set with respect to the norm topology in the C ∗-algebra A. / / > / / ~ INDUCTIVE LIMITS FOR SYSTEMS OF TOEPLITZ ALGEBRAS 5 2) If B is a C ∗-algebra, ψa : Aa −→ B is an injective ∗-homomorphism for each a ∈ K, and conditions analogous to those in (2.3) and (2.4) are satisfied, then there exists a unique ∗-isomorphism θ from A onto B such that the diagram Aa a σK ~⑥⑥⑥⑥⑥⑥⑥⑥ θ A ψa ❆❆❆❆❆❆❆ / B is commutative for every a ∈ K, that is, the equality ψa = θ ◦ σK a holds. The C ∗-algebra A itself is often called the inductive limit. It is denoted as follows: F := A. As is well known, the inductive limit can always be constructed for an lim−→ inductive system in the category of C ∗-algebras and their ∗-homomorphisms. For the details we refer the reader to Proposition 11.4.1 in [31]. 3. Auxiliary results The inductive sequences of Toeplitz algebras defined by arbitrary sequences of prime numbers are the objects for studying in [1]. Now, by analogy with the definition of such a sequence (see [1, Definition 1]), we are going to give the definition of the inductive system of Toeplitz algebras over a partially ordered set defined by a set of natural numbers possessing an additional property. Let ( K, ≤ ) be a directed set. In what follows, we consider a set of natural numbers (3.1) N = {nba ∈ N a, b ∈ K, a ≤ b} such that the factorization equalities (3.2) nca = ncb · nba hold for all elements a, b, c ∈ K satisfying the conditions a ≤ b and b ≤ c. It follows immediately from (3.2) that the equality naa = 1 holds for every a ∈ K. Further, using Lemma 2.1, for every number nba ∈ N we define the isometric ∗-homomorphism by the formula (3.3) σba : T −→ T : T 7−→ T nba . It is clear that the equalities (2.2) are valid for all elements a, b, c ∈ K whenever the conditions a ≤ b and b ≤ c hold, and for each a ∈ K the ∗-homomorphism σaa : Aa −→ Aa is the identity mapping. Thus we can give the definition of an inductive system of Toeplitz algebras (compare with [1, Definition 1]) over a directed set defined by a set of natural numbers satisfying factorization equalities. Definition 3.1. Let ( K, ≤ ) be a directed set and N be a set of natural numbers (3.1) satisfying (3.2). An inductive system of C ∗-algebras (3.4) F = (K, {Ta}, {σba}), where Ta = T for all a ∈ K and the connecting ∗-homomorphisms σba are given by (3.3), is called the inductive system of Toeplitz algebras over K defined by N . To obtain the main result of the article we shall prove in this section several auxiliary assertions. For formulations of these assertions we introduce additional notation. Firstly, for a directed set K and its arbitrary element a ∈ K the cofinal subset K a of K is defined as follows: K a := {b ∈ K a ≤ b}. ~ / 6 R.N. GUMEROV Secondly, for a subset S in the set K a we shall deal with the set of natural numbers N (S) := {nba ∈ N b ∈ S} . Throughout this section (K, {Tb}, {σcb}) is an inductive system of Toeplitz algebras over K defined by set (3.1) satisfying (3.2). Moreover, for an element a ∈ K we consider the inductive system (3.5) (K a, {Tb}, {σcb}) of Toeplitz algebras over K a defined by the set (3.6) {ncb ∈ N b, c ∈ K a} . Using property 2) of the inductive limit for the system (3.5) (see Preliminaries), it is straightforward to prove the following statement. Lemma 3.2. For every element a ∈ K there exists an isomorphism between the inductive limits lim−→(K a, {Tb}, {σcb}) ≃ lim−→(K, {Tb}, {σcb}) in the category of unital C ∗-algebras and unital ∗-homomorphisms. Lemma 3.3. The following two assertions are valid : 1) if for elements b, c ∈ K a the equality for numbers (3.7) nba = nca holds then we have the equality for the images of Toeplitz algebras (3.8) σK a b (Tb) = σK a c (Tc); 2) if for some k ∈ N and b, c ∈ K a the equality for numbers (3.9) nca = k · nba holds then we have the inclusion for the images of Toeplitz algebras (3.10) σK a b (Tb) ⊂ σK a c (Tc). Proof. 1) Assume that for some elements b, c ∈ K a equality (3.7) holds. Since K is a directed set there exists an element d ∈ K a such that b ≤ d and c ≤ d. Then factorization equalities (3.2) yield the following equalities for the corresponding natural numbers: (3.11) ndb · nba = nda = ndc · nca. Therefore, using condition (3.7), one gets immediately the equality ndb = ndc, which implies the following equality for the images of Toeplitz algebras: (3.12) σdb(Tb) = σdc(Tc). By (2.3) and (3.12), we obtain desired relation (3.8): σK a b (Tb) = σK a d (σdb(Tb)) = σK a d (σdc(Tc)) = σK a c (Tc). 2) Again, we choose an element d ∈ K a such that the conditions b ≤ d and c ≤ d are satisfied. Then, by (3.11) and (3.9), we get the equality ndb · nba = ndc · k · nba. Consequently, we have the equality ndb = ndc · k that together with (3.3) imply the following inclusion for the images of Toeplitz algebras: (3.13) σdb(Tb) ⊂ σdc(Tc). INDUCTIVE LIMITS FOR SYSTEMS OF TOEPLITZ ALGEBRAS 7 Thus, using equality (2.3) and inclusion (3.13), we obtain required inclusion (3.10): σK a b (Tb) = σK a d (σdb(Tb)) ⊂ σK a d (σdc(Tc)) = σK a c (Tc). (cid:3) Lemma 3.4. There exists a totally ordered countable subset Λa in the set K a satisfying the following property. For every element b ∈ K a there is an element c ∈ Λa and a number k ∈ N such that the equality (3.14) nca = k · nba holds, where nca ∈ N (Λa), nba ∈ N (K a). Proof. First of all, in the set of natural numbers N (K a) we consider the subset (3.15) {nbsa ∈ N s ∈ N} which is uniquely determined by the following three properties: • b1 = a; • the inequality nbsa < nbs+1a is valid for every s ∈ N; • for every number nba ∈ N (K a) there exists a natural number nbsa in set (3.15) such that the equality nba = nbsa holds. In other words, we throw out repeating numbers from the set N (K a). Moreover, the elements of set (3.15) constitute an increasing sequence of natural numbers indexed by s. To construct the desired set Λa we use set (3.15) and proceed as follows. As the first element of the set Λa denoted by c1 we choose the element a. We note that the condition b1 = a ≤ c1 = a is satisfied. Then equality (3.14) holds with c = c1 = a, b = b1 = a and k = nc1b1 = naa = 1. As the second element of the set Λa we take any element c2 ∈ K a satisfying the conditions c1 ≤ c2 and b2 ≤ c2. Such an element exists because K a is a directed set. In this case we have equality (3.14) with c = c2, b = b2 and k = nc2b2 . Continuing to argue in this way, we shall construct a countable totally ordered subset (3.16) Λa := {cs ∈ K a s ∈ N} in the set K a possessing the required property. It is worth noting that the condition cs ≤ cs+1 holds for every s ∈ N. (cid:3) In the next lemma we consider the inductive sequence of Toeplitz algebras (3.17) (Λa, {Tcs}, {σctcs}) over the set Λa defined by the subset {ncb ∈ N b, c ∈ Λa} in the set N . The proof of the following statement is similar to that of Proposition 1 in [1]. Lemma 3.5. There exists a subgroup QM in the group Q such that the reduced semigroup C ∗-algebra of the semigroup Q+ M is isomorphic to the inductive limit of the inductive sequence (3.17): r (Q+ M ) ≃ lim−→ (Λa, {Tcs}, {σctcs}) . (3.18) C ∗ Remark. In [21, Theorem 1] it is shown that the functor sending a partially ordered group to the corresponding Toeplitz algebra is continuous. Consider the inductive sequence of groups (2.1) with the connecting homomorphisms τs given by τs(m) = ncs+1cs m, m ∈ Z, s ∈ N. Applying the above-mentioned functor to 8 R.N. GUMEROV this inductive sequence of groups and making use of Theorem 1 in [21], we obtain another proof of Lemma 3.5. 4. The main result In this section we prove the following statement concerning the inductive limits for inductive systems of Toeplitz algebras over directed sets defined by sets of natural numbers (3.1) satisfying factorization equalities (3.2). To do this we shall use the results from the previous section. Theorem 4.1. Let F be an inductive system of Toeplitz algebras over a directed set K defined by a set of natural numbers N satisfying factorization equalities. Then there exists a subgroup QM in the group of all rational numbers such that the reduced semigroup C ∗-algebra of the semigroup Q+ M is isomorphic to the inductive limit of the inductive system F : (4.1) C ∗ r (Q+ M ) ≃ lim−→ F . Proof. As in the previous section we use the notation F = (K, {Ta}, {σba}). Let us fix an element a ∈ K. Then we take a totally ordered countable subset Λa in the set K a which is constructed in the proof of Lemma 3.4 (see (3.16)). We claim that there exists an isomorphism between the inductive limits of the inductive system F and inductive sequence (3.17) of Toeplitz algebras, that is, (4.2) lim−→ F ≃ lim−→ (Λa, {Tcs}, {σctcs}) in the category of unital C ∗-algebras and unital ∗-homomorphisms. Indeed, we consider the inductive system (3.5) of Toeplitz algebras over the set K a defined by the subset (3.6) in the set N . By (2.4) we have the following equality for the inductive limit of this system: (4.3) lim−→ (K a, {Tb}, {σcb}) = [b∈K a σK a b (Tb). By Lemma 3.2, to show the existence of isomorphism (4.2) it is enough to prove that one has the isomorphism between the inductive limits of systems (3.17) and (3.5) in the category of unital C ∗-algebras and their unital ∗-homomorphisms: (4.4) lim−→ (Λa, {Tcs}, {σctcs}) ≃ lim−→ (K a, {Tb}, {σcb}) . To construct an isomorphism (4.4) we shall use the universal property and prop- erty 2) of the inductive limit for system (3.17) (see Preliminaries). To this end, for elements cs, ct ∈ Λa satisfying the condition cs ≤ ct we consider the following INDUCTIVE LIMITS FOR SYSTEMS OF TOEPLITZ ALGEBRAS 9 commutative diagram (firstly without the dashed arrow): (4.5) Tcs σct cs Tct a σΛ cs 'PPPPPPPPPPPPPPPPPPPPPPPPPPP ❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅ σK cs a θ a σΛ ct w♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦ ⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦⑦ σK ct a lim−→ (Λa, {Tcs}, {σctcs}) lim−→ (K a, {Tb}, {σcb}) The universal property of the inductive limits for systems of C ∗-algebras and the in- jectivity of ∗-homomorphisms in diagram (4.5) yield the injective ∗-homomorphism θ : lim−→ (Λa, {Tcs}, {σctcs}) −→ lim−→ (K a, {Tb}, {σcb}) such that diagram (4.5) complemented by θ is also commutative. To show that the homomorphism θ is surjective it is sufficient (see property 2) in Preliminaries) to prove the equality (4.6) lim−→ (K a, {Tb}, {σcb}) = σK a cs (Tcs ). +∞ [s=1 Since we have the inclusion of sets Λa ⊂ K a and representation (4.3), the space on the left-hand side of (4.6) contains the space on its right-hand side. To prove the reverse inclusion for the spaces in (4.6) we use (4.3) again. For this (Tb) is aim let us fix an arbitrary element b ∈ K a. We state that the image set σK a contained in the set σK a cs (Tcs) for an element cs ∈ Λa. Really, by Lemma 3.4, there exists an element cs ∈ Λa and a number k ∈ N such that the factorization equality (3.14) is valid for natural numbers ncsa ∈ N (Λa), nba ∈ N (K a). By assertion 2) in Lemma 3.3, for the image sets σK a (Tb) and σK a cs (Tcs ) we obtain inclusion (3.10) with cs instead of c. It follows from (4.3) that the inclusion for C ∗-algebras b b lim−→ (K a, {Tb}, {σcb}) ⊂ σK a cs (Tcs) +∞ [s=1 holds. Hence, equality (4.6) is proved. Therefore, the ∗-homomorphism θ between the inductive limits is an isomorphism of C ∗-algebras. Thus, isomorphism (4.4) exists. Furthermore, we obtain the existence of isomorphism (4.2), as claimed. Finally, by Lemma 3.5, there exists a subgroup QM in the group Q for which we have isomorphism (3.18). Thus, using isomorphism (4.2), we obtain isomorphism (4.1), as required. (cid:3) / / ' w ✤ ✤ ✤ ✤ ✤ ✤ 10 R.N. GUMEROV 5. Inductive systems of Toeplitz algebras over arbitrary partially ordered sets Throughout this section a pair ( K, ≤ ) denotes a partially ordered set that is not necessarily directed. By analogy with Definition 3.1, one can define the inductive system of Toeplitz algebras over ( K, ≤ ) defined by set (3.1) satisfying factoriza- tion equalities (3.2). Below we shall consider such an inductive system F and use notation (3.4). Taking the family of all directed subsets of the set ( K, ≤ ) and making use of Zorn's lemma, one can easily prove the following statement. Lemma 5.1. Let ( K, ≤ ) be a partially ordered set. Then the following equality holds: (5.1) Ki, K = [i∈I where { Ki i ∈ I } is the family of all maximal directed subsets of the set ( K, ≤ ). Now, for a given inductive system of Toeplitz algebras (3.4) over the partially ordered set ( K, ≤ ) defined by set (3.1) satisfying (3.2) we consider representation (5.1). Then for each index i ∈ I we can construct the inductive system of Toeplitz algebras (5.2) Fi = (Ki, {Aa}, {σba}) over the directed set ( Ki, ≤ ) defined by the set of natural numbers { nba ∈ N a, b ∈ Ki } and its inductive limit lim−→ Fi. We consider the direct product of C ∗-algebras lim−→ Fi. That is, the C ∗-algebra Yi∈I lim−→ Fi :=(cid:26)(fi)(cid:12)(cid:12) k(fi)k = sup i kfik < +∞(cid:27) relative to the pointwise operations and the supremum norm. To formulate the result of this section it is convenient for us to give the definition. Definition 5.2. The inductive system Fi defined by (5.2) is called the inductive subsystem of F over the set Ki. The following statement is an immediate consequence of Theorem 4.1. Theorem 5.3. Let K be a partially ordered set and let {Ki ⊂ K i ∈ I} be a family of all maximal directed subsets in the set K. Let F be an inductive system of Toeplitz algebras over K defined by a set of natural numbers N satisfying factorization equalities. Let Fi, where i ∈ I, denote the inductive subsystem of F over the set Ki. Then there exists a family {QMi ⊂ Q i ∈ I} consisting of subgroups in the group of all rational numbers Q and an isomorphism between the direct products of C ∗-algebras in the category of unital C ∗-algebras and their unital ∗-homomorphisms. Yi∈I lim−→ Fi ≃Yi∈I C ∗ r (Q+ Mi) The author is grateful to S. A. Grigoryan and E. V. Lipacheva for helpful dis- Acknowledgments cussions of the results. The research was funded by the subsidy allocated to Kazan Federal University for the state assignment in the sphere of scientific activities, project 1.13556.2019/13.1. INDUCTIVE LIMITS FOR SYSTEMS OF TOEPLITZ ALGEBRAS 11 References 1. R. N. Gumerov, "Limit Automorphisms of C ∗-algebras Generated by Isometric Representa- tions for Semigroups of Rationals," Sib. Math. J. 59 (1), 73 -- 84 (2018). 2. S. A. Grigoryan and R. N. Gumerov, "On a covering group theorem and its applications," Lobachevskii J. Math. 10, 9 -- 16 (2002). 3. R. N. Gumerov, "On finite-sheeted covering mappings onto solenoids," Proc. Amer. Math. Soc. 133 (9), 2771 -- 2778 (2005). 4. R. N. Gumerov, "On the existence of means on solenoids," Lobachevskii J. Math. 17, 43 -- 46 (2005). 5. R. N. Gumerov, "Weierstrass Polynomials and Coverings of Compact Groups, "Sib. Math. J. 54 (2), 243-246 (2013). 6. R. N. Gumerov, "Characters and Coverings of Compact Groups," Russian Math. (Izvestiya VUZ. Matematika) 58 (4), 7 -- 13 (2014). 7. R. N. Gumerov, "Coverings of solenoids and automorphisms of semigroup C*-algebras," Uchenye Zapiski Kazanskogo Universiteta. Seria Fiziko-Matematicheskie Nauki 160 (2), 275 -- 286 (2018). 8. R. Haag and D. Kastler "An algebraic approach to quantum field theory," J. Math. Phys. 5, 848 (1964). 9. R. Haag, Local quantum physics: fields, particles, algebras, Springer Texts and Monographs in Physics, 2nd. rev. and enlarged ed. (1996). 10. H. Araki, Mathematical Theory of Quantum Fields (Oxford University Press, Oxford 2009). 11. S. S. Horuzhy, Introduction to algebraic quantum field theory. Mathematics and its Applica- tions (Soviet Series) (Kluwer, Dordrecht, 1990). 12. G. Ruzzi, "Homotopy of posets, net-cohomology and superselection sectors in globally hyper- bolic space-times," Rev. Math. Phys. 17 (9), 1021 -- 1070 (2005). 13. G. Ruzzi and E. Vasselli, "A new light on nets of C ∗-algebras and their representations," Comm. Math. Phys. 312 (3), 655 -- 694 (2012). 14. E. Vasselli, "Presheaves of symmetric tensor categories and nets of C ∗-algebras," J. Non- commut. Geometry 9 (1), 121 -- 159 (2015). 15. S. A. Grigoryan, T. A. Grigoryan, E. V. Lipacheva and A. S. Sitdikov, " C ∗-algebra generated by the path semigroup," Lobachevskii J. Math. 37 (6), 740 -- 748 (2016). 16. S. A. Grigoryan, E. V. Lipacheva and A. S. Sitdikov, "Nets of graded C ∗-algebras over partially ordered sets," Algebra and Analysis 30 (6), 1 -- 19 (2018) (in Russian). 17. G. J. Murphy, C ∗-algebras and operator theory (Academic Press, New York, 1990). 18. L. A. Coburn, "The C ∗-algebra generated by an isometry, " Bull. Amer. Math. Soc. 73 (5), 722 -- 726 (1967). 19. L. A. Coburn, "The C ∗-algebra generated by an isometry.II," Trans. Amer. Math. Soc. 137, 211 -- 217 (1969). 20. R. G. Douglas, "On the C ∗-algebra of a one-parameter semigroup of isometries," Acta Math. 128, 143 -- 152 (1972). 21. G. J. Murphy, "Ordered groups and Toeplitz algebras," J. Oper. Theory 18, 303 -- 326 (1987). 22. G. J. Murphy, "Toeplitz operators and algebras, "Math. Z. 208, 355 -- 362 (1991). 23. E. V. Lipacheva and K. H. Hovsepyan, "Automorphisms of some subalgebras of the Toeplitz algebra," Sib. Math. J. 57 (3), 525 -- 531 (2016). 24. X. Li,"Semigroup C ∗-algebras," https://arxiv.org/pdf/1707.05940.pdf. 25. R. N. Gumerov, E. V. Lipacheva and T. A. Grigoryan, "On Inductive Limits for Systems of C ∗ - algebras," Russian Math. (Izvestiya VUZ. Matematika) 62 (7), 68 -- 73 (2018). 26. R. N. Gumerov, E. V. Lipacheva and T. A. Grigoryan, "On a Topology and Limits for In- ductive Systems of C ∗-Algebras over Partially Ordered Sets, " Int. J. Theor. Phys. (2019). https://doi.org/10.1007/s10773-019-04048-0 27. G. J. Murphy, "Simple C ∗-algebras and subgroups of Q," Proc. Amer. Math. Soc. 107, 97 -- 100 (1989). 28. S. A. Bogatyi and O. D. Frolkina, "Classification of generalized solenoids," Trudy seminara on vector and tensor analysis. Moscow: MSU, XXVI, 31 -- 59 (2005) (in Russian). 12 R.N. GUMEROV 29. R. N. Gumerov, by "On Norms Shift Transforma- tions Arising in Signal and Image Processing on Meshes Supplied with Semi- groups 012042 Sci. Eng. http://china.iopscience.iop.org/article/10.1088/1757-899X/158/1/012042/pdf. of Operators Generated IOP Conf. Ser.: Structures," 2016, Mater. 158 30. A. Ya. Helemskii, Banach and locally convex algebras ( Oxford Science Publications. The Clarendon Press Oxford University Press, New York, 1993). 31. R. V. Kadison and J. R. Ringrose, Fundamentals of the theory of operator algebras, Volume II, Advanced theory (Academic Press. Inc., London, 1986). 32. I. Bucur and A. Deleanu A., with the collaboration of Hilton P.J. Introduction to the Theory of Categories and Functors, Pure and Appl. Math., V. XIX, ( Wiley -- Interscience Publ., London -- New York -- Sydney 1968). Chair of Mathematical Analysis, N.I. Lobachevskii Institute of Mathematics and Mechanics, Kazan (Volga region) Federal University, Kremlevskaya 35, Kazan,420008, Tatarstan, Russian Federation E-mail address: [email protected]; [email protected]
0907.5473	4	0907	2011-03-26T04:27:18	Conditionally monotone independence I: Independence, additive convolutions and related convolutions	[ "math.OA", "math.PR" ]	We define a product of algebraic probability spaces equipped with two states. This product is called a conditionally monotone product. This product is a new example of independence in non-commutative probability theory and unifies the monotone and Boolean products, and moreover, the orthogonal product. Then we define the associated cumulants and calculate the limit distributions in central limit theorem and Poisson's law of small numbers. We also prove a combinatorial moment-cumulant formula using monotone partitions. We investigate some other topics such as infinite divisibility for the additive convolution and deformations of the monotone convolution. We define cumulants for a general convolution to analyze the deformed convolutions.	math.OA	math	Conditionally monotone independence I: Independence, additive convolutions and related convolutions Takahiro Hasebe∗ Graduate School of Science, Kyoto University, Kyoto 606-8502, Japan Email: [email protected] Abstract We define a product of algebraic probability spaces equipped with two states. This product is called a conditionally monotone product. This product is a new example of independence in non-commutative probability theory and unifies the monotone and Boolean products, and moreover, the orthogonal product. Then we define the associated cumulants and calculate the limit distributions in central limit theorem and Poisson's law of small numbers. We also prove a combinatorial moment-cumulant formula using monotone partitions. We investigate some other topics such as infinite divisibility for the additive convolution and deformations of the monotone convolution. We define cumulants for a general convolution to analyze the deformed convolutions. Mathematics Subject Classification: 46L53; 46L54 Keywords: Conditionally free independence; monotone independence; Boolean independence; free independence; cumulants 1 Introduction Non-commutative probability theory lays the foundation of quantum mechanics and has many math- ematical branches. The basic framework consists of a (unital) ∗-algebra A and a state ϕ on it. The pair (A, ϕ) is called an algebraic probability space or a non-commutative probability space. When A has structure of a C∗-algebra (resp. von Neumann algebra), we call the pair a C∗- (resp. von Neumann) algebraic probability space. Many kinds of independence have been studied as an aspect of non-commutative probability theory. The usual independence in probability theory is called tensor independence from the non- commutative probabilistic viewpoint. Other famous ones are free independence defined by Voiculescu [43], Boolean independence by Speicher and Woroudi [42] and monotone independence by Muraki [31]. These kinds of independence can canonically be realized by using products of states on the free product of algebras (with or without the identification of units): there are two "universal products" (tensor and free) defined on the free product of algebras with the identification of units [3, 41]; there are three universal products (tensor, free and Boolean) defined on the free product of algebras without the identification of units; there are five "natural products" (tensor, free, Boolean, monotone and anti-monotone) defined on the free product of algebras without the identification of units [32, 33]. These results can also be understood in terms of tensor structures with inclusions [14]. In particular, monotone and Boolean products are important in this paper. Moreover, the conditionally (c- for ∗This work was supported by Grant-in-Aid for JSPS Fellows. 1 short) free product of states was introduced by Bozejko, Leinert and Speicher [8, 9]. This product can be seen as a universal product of pairs of states which can be defined similarly to the single state case. As such a concept has not been defined in the literature, we shall systematically study it elsewhere. In [8] it was proved that the c-free product and cumulants unify the free and Boolean products and their cumulants introduced in [40, 42, 44]. In addition, the c-free product also unifies the monotone product as proved in [15]. While the latter is quite nontrivial, some complication appears in its application: it is difficult to repeat the calculation of monotone products in terms of c-free products; it is difficult to identify monotone cumulants [21] in terms of c-free cumulants. The latter difficulty is essentially the same as the former. The solution of these difficulties is a purpose of this paper. To this end, in Section 3 we introduce a c-monotone product analogously to the c-free product. Once it is introduced, the monotone and Boolean products can be formulated in terms of it. The concept of c-monotone independence can also be extracted from the c-monotone product since the product is associative. In terms of probability measures, we can also define (additive) c-monotone convolutions. We prove that c-monotone independence and c-free one include orthogonal indepen- dence [26] as special cases. Therefore, the additive c-monotone convolution unifies the additive monotone, Boolean and orthogonal convolutions. As a result, c-monotone convolutions can give the characterization of orthogonal convolutions (Theorem 6.2 of [26]). In Section 4 we introduce c-monotone cumulants rn(µ, ν) to linearize powers of probability mea- sures. A moment-cumulant formula is proved by using combinatorics of monotone partitions; this formula is naturally expected from the monotone case [21]. An important point is that c-monotone cumulants generalize monotone and Boolean cumulants. Here we achieve a purpose of this paper. The remaining sections are roughly divided into two parts; one is devoted to infinitely divisible distributions, and the other is to deformations of the monotone convolution. In Sections 6 -- 8, we investigate convolution semigroups and infinitely divisible distributions. Ad- ditive monotone and Boolean infinitely divisible distributions were first studied in [31] and [42], respectively. The results in this paper generalize these studies: we prove the L´evy-Khintchine for- mula, and the correspondence among a convolution semigroup, an infinitely divisible distribution, a pair of vector fields and a positive definite sequence of cumulants. Moreover, we construct convolution semigroups from monotone and Boolean ones. As a result, c-monotone cumulants rn(·, ν), for a fixed ν, turn out to linearize the Boolean convolution. We note that infinite divisibility was introduced and studied in [22] for c-free independence. Results on c-monotone infinite divisibility however do not follow from the c-free case. In Sections 10 -- 15 we work on deformations of the monotone convolution. This topic may impress the reader as specialized at first sight; however this clarifies how the structure of c-monotone indepen- dence behaves analogously to that of c-free independence. A deformation of the free convolution can be defined in a graph of probability measures [10, 11, 23, 24, 35, 36]. More precisely, if T is a map from the set of probability measures to itself, we can define the graph {(µ, T µ); µ is a probability measure}. If this graph is closed under the c-free convolution, we can define an associative convolution. For the details, the reader is referred to Section 10. Analogously, a deformed convolution arises from the c-monotone convolution of a graph of probability measures. These kinds of convolutions include the monotone and Boolean convolutions. We show many examples of such deformed convolutions. A remarkable point is that such maps T , found in the context of c-free convolutions, give associative convolutions also in the c-monotone case. The Boolean and monotone convolutions preserve the sets {µ; supp µ ⊂ [0,∞)} and {µ; µ is symmetric}. The former property can be proved easily in terms of the operator-theoretic approach in [17]; the latter can be proved by using the complex-analytic characterizations of the convolutions. As an extension of these properties, we give necessary and sufficient conditions under which the deformed convolution explained above preserves the two sets. We introduce the cumulants for the convolution deformed by a map T and then limit distributions are calculated for some class of such convolutions. When we introduce the cumulants of the deformed 2 convolutions, the axiom of homogeneity for cumulants does not hold in general (see (13.5)). For this reason we consider the uniqueness and the existence of cumulants of a general convolution in Section 13. Let us mention a few topics which are not covered in this paper. Multiplicative convolutions and the infinite divisibility were studied in [5, 6, 15, 16, 27] in the Boolean, monotone and orthogonal cases. Multiplicative c-monotone convolutions can be similarly defined to generalize the monotone, Boolean and orthogonal convolutions. We do not treat these in this paper; these aspects will be studied in [20]. 2 Preliminaries 2.1 Reciprocal Cauchy transform We use the notation C+ := {z ∈ C; Imz > 0}. The Cauchy transform of a probability measure µ is defined by The reciprocal Cauchy transform of a probability measure µ is defined by Gµ(z) =ZR 1 z − x dµ(x), z ∈ C\R. Hµ(z) = 1 Gµ(z) , z ∈ C\R. (2.1) (2.2) This is an analytic map from C+ to C+. Since limy→∞ iyGµ(iy) = 1, Hµ has the following form: Hµ(z) = b + z +ZR 1 + xz x − z dη(x), (2.3) where b ∈ R and η is a positive finite measure. Conversely, any function of the form of the right hand side of (2.3) is a reciprocal Cauchy transform of a probability measure (see [2, 30] for details). 2.2 Monotone independence Muraki defined the concept of monotone independence in [31]. A definition is as follows. Let (A, ϕ) be an algebraic probability space and let I be a linearly ordered set. A family of subalgebras {Ai}i∈I is said to be monotone independent if the equality ϕ(a1a2 · · · an) = ϕ(aj)ϕ(a1a2 · · · aj−1aj+1 · · · an) (2.4) holds for ak ∈ Aik with i1, i2,· · · , in ∈ I, ij−1 < ij > ij+1 and 1 ≤ j ≤ n. When j = 1 (resp. j = n), the condition ij−1 < ij > ij+1 is understood to be i1 > i2 (resp. in−1 < in). The monotone convolution µ ⊲ ν is defined for two probability measures µ, ν and is characterized by the formula Hµ⊲ν = Hµ ◦ Hν. (2.5) The monotone convolution is non-commutative and associative. Let NC(n) be the set of all non-crossing partitions [34]. Let M(n) be the set of all monotone partitions defined by M(n) := {(π, λ) : π ∈ NC(n), if V, W ∈ π and V is in the inner side of W , then V >λ W}, (2.6) where λ denotes a linear ordering of the blocks of π. V >λ W means that V is larger than W under the linear ordering λ (see [32, 33], and also [28, 29]). 3 In the paper [21] the concept of monotone cumulants has been defined. Monotone cumulants do not satisfy the additivity for general probability measures, but satisfy the power additivity: rn(µ⊲N ) = Nrn(µ). The moment-cumulant formula is described as mn(µ) = X(π,λ)∈M(n) 1 π!YV ∈π rV (µ). (2.7) Example 2.1. We exhibit the moment-cumulant formula until the forth order. m1(µ) = r1(µ), m2(µ) = r2(µ) + r1(µ)2, m3(µ) = r3(µ) + 5 2 r1(µ)r2(µ) + r1(µ)3, m4(µ) = r4(µ) + 3r1(µ)r3(µ) + 3 2 r2(µ)2 + 13 3 r1(µ)2r2(µ) + r1(µ)4. 2.3 Conditionally free independence Let I be an index set. Let Ai be a unital ∗-algebra and let ϕi, ψi be states on Ai for i ∈ I. The c-free product of triples (Ai, ϕi, ψi)i∈I was defined by Bozejko and Speicher in [9]. We define (A, ϕ, ψ) = ∗i∈I(Ai, ϕi, ψi) by setting A := ∗i∈IAi (the free product with the identification of units) and ψ := ∗i∈Iψi (the free product of states). ϕ is defined by the following condition: the equality ϕ(a1 · · · an) = nYk=1 ϕik(ak) (2.8) holds if ak ∈ Aik with i1 6= · · · 6= in and ψik (ak) = 0 for all 1 ≤ k ≤ n. If the index set I consists of two elements, that is, I = 2, ϕ is denoted by ϕ1ψ1∗ψ2ϕ2. Let µ, ν be probability measures on R with compact supports. Define the R-transform of ν and the c-free R-transform of (µ, ν) by 1 Gν(z) 1 Gµ(z) = z − Rν(Gν(z)), = z − R(µ,ν)(Gν(z)). (2.9) (2.10) We expand R(µ,ν)(z) = P∞n=1 Rn(µ, ν)zn−1 as a formal power series. Rn(µ, ν) are called c-free cu- mulants. Similarly, we expand Rν(z) = P∞n=1 Rn(ν)zn−1 and Rn(ν) are called free cumulants. In this paper, Hµ(z) is more useful than Gµ(z), and correspondingly, we use φ(µ,ν)(z) := R(µ,ν)( 1 φν(z) := Rν( 1 z ). Then (7.4) and (7.7) can be written as z ) and Hν(z) = z − φν(Hν(z)), Hµ(z) = z − φ(µ,ν)(Hν(z)). A moment-cumulant formula for c-free independence is mn(µ) = Xπ∈NC(n)(cid:16) YV ∈π, V : outer RV (µ, ν)(cid:17)(cid:16) YV ∈π, V : inner (2.11) (2.12) (2.13) RV (ν)(cid:17), which generalizes the free and Boolean moment-cumulant formulae. The c-free convolution of (µ1, ν1) and (µ2, ν2) is the pair (µ, ν) = (µ1, ν1) ⊞ (µ2, ν2), where µ and ν are characterized by φν(z) = φν1(z) + φν2(z), φ(µ,ν)(z) = φ(µ1,ν1)(z) + φ(µ2,ν2)(z). (2.14) (2.15) Let (µ1ν1 ⊞ν2 µ2, ν1 ⊞ ν2) denote the c-free convolution of (µ1, ν1) and (µ2, ν2). 4 2.4 Technical facts We summarize the notation and several lemmata which will be used in Sections 10, 11 and 12. Let P, P 2, Pm, Pc, P+ and Psym be the set of probability measures, the set of probability measures with finite variance, the set of probability measures with finite moments of all orders, the set of probability measures with compact supports, the set of probability measures on [0,∞) and the set of symmetric probability measures, respectively. The following lemma was proved in [30]. Lemma 2.2. A probability measure µ belongs to P 2 if and only if Hµ has the representation Hµ(z) = a + z +ZR 1 x − z dρ(x), (2.16) where a ∈ R, ρ a positive finite measure. a and ρ are determined uniquely. Furthermore, we have ρ(R) = σ2(µ) and a = −m(µ), where m(µ) is the mean of µ and σ2(µ) is the variance of µ. We define a(µ) := inf{x ∈ supp µ} and b(µ) := sup{x ∈ supp µ}. We note that −∞ ≤ a(µ) < ∞ and −∞ < b(µ) ≤ ∞. The following lemmata 2.3-2.6 were proved in [18]. Lemma 2.3. Let ν and µ be probability measures. Then the following inequalities hold: (1) If supp ν ∩ (−∞, 0] 6= ∅ and supp ν ∩ [0,∞) 6= ∅, then a(µ) ≥ a(ν ⊲ µ) and b(µ) ≤ b(ν ⊲ µ). (2) If supp ν ⊂ (−∞, 0], then a(µ) ≥ a(ν ⊲ µ) and b(ν) + b(µ) ≤ b(ν ⊲ µ). (3) If supp ν ⊂ [0,∞), then a(ν) + a(µ) ≥ a(ν ⊲ µ) and b(µ) ≤ b(ν ⊲ µ). Lemma 2.4. We use the notation in (2.3). For µ ∈ P, the condition µ ∈ P+ is equivalent to supp η ⊂ [0,∞) and Hµ(−0) ≤ 0. Moreover, if supp η ⊂ [0,∞), the condition Hµ(−0) ≤ 0 is equivalent to η({0}) = 0, dη(x) < ∞, Z ∞ 0 1 x b +Z ∞ 0 1 x dη(x) ≤ 0. Lemma 2.5. Let {µt}t≥0 be a weakly continuous ⊲-convolution semigroup with µ0 = δ0. Then the following statements are equivalent: (1) there exists t0 > 0 such that supp µt0 ⊂ [0,∞); (2) supp µt ⊂ [0,∞) for all 0 ≤ t < ∞; Lemma 2.6. We assume that the support of each µt is compact (or equivalently, the support of µt is compact for some t > 0). Then the following statements are equivalent. (1) There exists t0 > 0 such that µt0 is symmetric. (2) µt is symmetric for all t > 0. Lemma 2.7. P+ and Psym are closed subsets of P under the weak topology. Proof. Let {µn} ⊂ P+ be a sequence converging to µ ∈ P. The weak convergence implies that µ((−∞, 0)) ≤ lim inf µn((−∞, 0)) = 0; therefore, µ ∈ P+. For a probability measure ν, ν ∈ Psym is equivalent to the condition ZR g(x)dν(x) = 0 for all g ∈ Cb(R), g(−x) = −g(x), (2.17) where Cb(R) is the set of bounded continuous functions on R. This equivalence can be proved with a simple approximation argument. Then the conclusion is not difficult. 5 3 Conditionally monotone independence It is known that the free, Boolean and monotone products of states (denoted by ∗,⋄ and ⊲, respec- tively) can be expressed in terms of the c-free product [8, 15], as follows. We consider triples of algebras and states (A1, ϕ1, ψ1) and (A2, ϕ2, ψ2). We assume that Ai has a decomposition Ai = C1 ⊕ A0 i (3.1) with a subalgebra A0 λ ∈ C and a0 ∈ A0 from Ai to C, then A0 i (i = 1, 2). Then the delta state δi (i = 1, 2) is defined by δi(λ1 + a0) = λ for i . δi is a homomorphism from Ai to C. Conversely, if there exists a homomorphism i can be defined to be its kernel. We have the following relations. (ϕ, ϕ) ∗ (ψ, ψ) = (ϕ ∗ ψ, ϕ ∗ ψ) on A1 ∗ A2, 1 ∗ A0 (ϕ, δ1) ∗ (ψ, δ2) = (ϕ ⋄ ψ, δ1 ∗ δ2) on A0 2, (ϕ, δ1) ∗ (ψ, ψ) = (ϕ ⊲ ψ, ψ) on A0 1 ∗ A2. (3.2) (3.3) (3.4) In terms of additive convolutions of (compactly supported) probability measures, the equalities (3.2)- (3.4) can be written as (µ, µ) ⊞ (ν, ν) = (µ ⊞ ν, µ ⊞ ν), (µ, δ0) ⊞ (ν, δ0) = (µ ⊎ ν, δ0), (µ, δ0) ⊞ (ν, ν) = (µ ⊲ ν, ν). (3.5) (3.6) (3.7) We can understand the associative laws of the Boolean and free convolutions from (3.5) and (3.6) since the c-free convolution is associative. The associative law of the monotone convolution is, however, not easy to understand from (3.7) since we cannot repeat (3.7) more than twice. We note here that the associative law of the monotone convolution was proved rigorously first in [13]. We show that the monotone convolution for the second component solves this problem. We follow the setting of [14] on unitization. Definition 3.1. (1) Let (A1, ϕ1, ψ1) and (A2, ϕ2, ψ2) be triples consisting of algebras and two linear functionals and let A1 ∗nu A2 be the free product without the identification of units. We define a c-monotone product (A1, ϕ1, ψ1) ⊲ (A2, ϕ2, ψ2) := (A1 ∗nu A2, ϕ1 ⊲ψ2 ϕ2, ψ1 ⊲ ψ2), (3.8) by setting ϕ1 ⊲ψ2 ϕ2 as follows. Denote the unitization of each Ai byfAi := C1fAi⊕Ai. Then the linear functionals ϕi and ψi are naturally extended to eϕi and eψi on fAi by eϕi(1fAi There is a natural isomorphism ^A1 ∗nu A2 ∼= fA1 ∗fA2. We let eδ1 be the delta state associated to the decomposition fA1 = C1 ⊕ A1. We define ϕ1 ⊲ψ2 ϕ2 to be the restriction offϕ1 eδ1∗fψ2fϕ2 on A1 ∗nu A2. (2) For pairs of probability measures (µ1, ν1) and (µ2, ν2), we define an additive c-monotone convo- lution (µ1, ν1) ⊲ (µ2, ν2) := (µ1δ0 ) = 1 and eψi(1fAi ⊞ν2 µ2, ν1 ⊲ ν2). We write µ1 ⊲ν2 µ2 for µ1δ0 ⊞ν2 µ2. ) = 1. Remark 3.2. We defined the c-monotone product on the free product without the identification of units. Considering the context of category theory [14], it is natural to treat products of linear functionals. This is why we considered linear functionals instead of states. The c-monotone product, however, preserves the positivity of linear functionals since it is defined to be the restriction of the c-free product. Proposition 3.3. µ1 ⊲ν2 µ2 is characterized by Hµ1⊲ν2 µ2 = Hµ1 ◦ Hν2 + Hµ2 − Hν2. (3.9) 6 Proof. We immediately obtain the equality Hµ1◦Hν2(z) = Hν2(z)−φ(µ1,δ0)(Hν2(z)) from (2.12). Since a c-free convolution is characterized by the sum of φ(,), we have Hµ1 ◦ Hν2(z) − Hν2(z) + Hµ2(z) = z − φ(µ1 ⊲ν2 µ2,ν2) ◦ Hν2(z). Remark 3.4. For any probability measures µi, νi (i = 1, 2), we can prove that (1) Hµ1 ◦ Hν2 + Hµ2 − Hν2 is an analytic map from C+ to C+, (2) inf z∈C+ Therefore, the definition of the c-monotone convolution of compactly supported probability measures can be extended to arbitrary probability measures [30]. Im(Hµ1◦Hν2 (z)+Hµ2 (z)−Hν2 (z)) = 1. Im z We can easily check with Proposition 3.3 that the c-monotone convolution of probability measures is associative, i.e.,(cid:0)(µ1, ν1) ⊲ (µ2, ν2)(cid:1) ⊲ (µ3, ν3) = (µ1, ν1) ⊲(cid:0)(µ2, ν2) ⊲ (µ3, ν3)(cid:1). Therefore, the c- monotone product of pairs of linear functionals is also expected to be associative. To prove this, we need to know how to compute mixed moments under the c-monotone product of linear functionals. We use the following notation: for a linearly ordered index set I = {i1, i2,· · · , in} with i1 < i2 < · · · < in, we set −→Yi∈I ai := ai1ai2 · · · ain. Lemma 3.5. Let xj and yk be (possibly non-commutative) elements in an algebra over C with unit 1 and let pj ∈ C. Then the following identity holds: x1y1x2 · · · yn−1xn = XS⊂{1,··· ,n−1}(cid:16)Yj /∈S pj(cid:17)(cid:16)xS1(yk1−pk11)xS2(yk2−pk21)· · · (ykm−pkm1)xSm+1(cid:17), (3.10) := −→Qk∈Sj xk. {Sj} is a partition of {1,· · · , n} determined by S as follows: where xSj if S = {k1,· · · , km} with 1 ≤ k1 < · · · < km ≤ n− 1, then Sj := {kj−1 + 1,· · · , kj} for 1 ≤ j ≤ m + 1, where k0 := 0 and km+1 := n. If S = ∅, then S1 = {1,· · · , n}. A product over the empty set is defined to be 1. Proof. (3.10) can be proved easily by induction on the number n. Theorem 3.6. Let (A1, ϕ1, ψ1) and (A2, ϕ2, ψ2) be triples consisting of algebras and two linear func- tionals. The calculation rule for mixed moments under ϕ1 ⊲ψ2 ϕ2 is what follows. (1) ϕ1 ⊲ψ2 ϕ2(bax) = ϕ2(b)ϕ1 ⊲ψ2 ϕ2(ax) for a ∈ A1, b ∈ A2 and bax ∈ A1 ∗nu A2. This is also true if x is absent. (2) ϕ1 ⊲ψ2 ϕ2(xab) = ϕ1 ⊲ψ2 ϕ2(xa)ϕ2(b) for a ∈ A1, b ∈ A2 and xab ∈ A1 ∗nu A2. This is also true if x is absent. (3) ϕ1⊲ψ2ϕ2(a1b1 · · · bn−1an) = (ϕ2(bj)−ψ2(bj))ϕ1⊲ψ2ϕ2(a1b1a2 · · · bj−1aj)ϕ1⊲ψ2ϕ2(aj+1bj+1 · · · bn−1an) + ψ2(bj)ϕ1 ⊲ψ2 ϕ2(a1b1a2 · · · bj−1ajaj+1bj+1 · · · bn−1an) for ak ∈ A1, bk ∈ A2, 1 ≤ j ≤ n − 1, n ≥ 2. Proof. Since (1) and (2) can be proved similarly, we only prove (3). Moreover, we assume that j = 1 for simplicity. since the proof for general j is essentially the same. We denote by 1 the unit 1 ^A1∗nuA2 First we obtain fϕ1eδ1∗fψ2fϕ2(a1b1a2 · · · bn−1an) =fϕ1eδ1∗fψ2fϕ2(a1(b1 − ψ2(b1)1)a2 · · · bn−1an) + ψ2(b1)fϕ1eδ1∗fψ2fϕ2(a1a2 · · · bn−1an). 7 (3.11) Using the identity (3.10) and the definition of c-free products, we have fϕ1eδ1∗fψ2fϕ2(a1(b1 − ψ2(b1)1)a2 · · · bn−1an) =fϕ1eδ1∗fψ2fϕ2(cid:16)a1(b1 − ψ2(b1)1) XS⊂{2,··· ,n−1}(cid:16)Yj /∈S = (ϕ2(b1) − ψ2(b1)) XS⊂{2,··· ,n−1}(cid:16)Yj /∈S ψ2(bj)(cid:17)(cid:16)Yj∈S = ϕ1(a1)(ϕ2(b1) − ψ2(b1))fϕ1 eδ1∗fψ2fϕ2(a2b2 · · · bn−1an). ψ2(bj)(cid:17)(cid:16)aS1(bk1 − ψ2(bk1)1)· · · aSm+1(cid:17)(cid:17) (ϕ2(bj) − ψ2(bj))(cid:17)ϕ1(a1) S+1Yj=1 ϕ1(aSj ) Theorem 3.7. The c-monotone product is associative. Proof. Let (Ai, ϕi, ψi) (1 ≤ i ≤ 3) be triples consisting of algebras and two linear functionals. We can naturally identify (A1 ∗nu A2) ∗nu A3 with A1 ∗nu (A2 ∗nu A3) and denote them simply by A1 ∗nu A2 ∗nu A3. What should be proved is the equality ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3) = (ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3 on A1 ∗nu A2 ∗nu A3. We call x ∈ A1 ∗nu A2 ∗nu A3 an elementary word if x is of such a form as x = x1x2 · · · xn, xj ∈ Aij , i1 6= · · · 6= in. Each si is simply called an element of x. We give a proof by induction on the number of elements of A3 contained in an elementary word. In this proof we use the notation a, a′ and aj for elements of A1, b, b′ and bj for elements of A2 and c and cj for elements of A3. For x ∈ A2 ∗nu A3 ⊂ A1 ∗nu A2 ∗nu A3, we can easily prove that ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(x) = (ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3(x). We assume that the statement is the case for all elementary words containing not more than n elements of A3. Let x be an elementary word containing n + 1 elements of A3. Let c be the (n + 1)-th element of A3 contained in x and let a, if exists, be the last element of A1 which appears before c. Except for some cases, x can be written as x = uavcw, where each component satisfies the following property: (i) u is an elementary word in A1 ∗nu A2 ∗nu A3 containing not more than n elements of A3; (ii) v is an elementary word in A2∗nuA3 containing not more than n elements of A3. The last element of v belongs to A2, i.e., v = v′b1; (iii) w is an elementary word in A1 ∗nu A2. Here we note some remarks on the exceptional cases. For x of such a form as x = uacw, the following proof is applicable by understanding that v plays the role of a unit. For x of such a form as x = uavc, the proof follows easily from the property (1) in Theorem 3.6. If a does not appear, again from the property (1) in Theorem 3.6 the proof is easy. Therefore, we only prove the claim for x of the form uavcw. We consider the following two cases: (1) The first element in w belongs to A1, i.e., w = a′w′; (2) The first element in w belongs to A2, i.e., w = b2w′. Case (1): We obtain (ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3(uavcw) = (ϕ3(c) − ψ3(c))(ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3(uav)(ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3(w) + ψ3(c)(ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3(uavw) = (ϕ3(c) − ψ3(c))ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uav)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uavw) = (ϕ3(c) − ψ3(c))ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uav)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)(ϕ2 ⊲ψ3 ϕ3(v) − ψ2 ⊲ ψ3(v))ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)ψ2 ⊲ ψ3(v)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw), (3.12) 8 where we have used the assumption of induction and the property (2) in Theorem 3.6. On the other hand, we obtain ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uavcw) = (ϕ2 ⊲ψ3 ϕ3(vc) − ψ2 ⊲ ψ3(vc))ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(w) + ψ2 ⊲ ψ3(vc)ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw) = ϕ2 ⊲ψ3 ϕ3(v)ϕ3(c)ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(w) − ψ2 ⊲ ψ3(v)ψ3(c)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(w) + ψ3(c)ψ2 ⊲ ψ3(v)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw) = (ϕ3(c) − ψ3(c) + ψ3(c))ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uav)ϕ1 ⊲ψ2 ϕ2(w) − ψ2 ⊲ ψ3(v)ψ3(c)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)ψ2 ⊲ ψ3(v)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw) = (ϕ3(c) − ψ3(c))ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uav)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)(cid:16)ϕ2 ⊲ψ3 ϕ3(v) − ψ2 ⊲ ψ3(v)(cid:17)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)ψ2 ⊲ ψ3(v)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw). Therefore, they coincide with each other. Case (2): Carrying out a similar calculation, we have (ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3(uavcw) = (ϕ3(c) − ψ3(c))ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uav)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uavw) = (ϕ3(c) − ψ3(c))ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ2 ⊲ψ3 ϕ3(v)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)(ϕ2 ⊲ψ3 ϕ3(vb2) − ψ2 ⊲ ψ3(vb2))ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w′) + ψ3(c)ψ2 ⊲ ψ3(vb2)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw′). On the other hand we have ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uavcw) = (ϕ2 ⊲ψ3 ϕ3(vcb2) − ψ2 ⊲ ψ3(vcb2))ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w′) + ψ2 ⊲ ψ3(vcb2)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw′) = ϕ2 ⊲ψ3 ϕ3(v)(ϕ3(c) − ψ3(c))ϕ2(b2)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w′) + ψ3(c)ϕ2 ⊲ψ3 ϕ3(vb2)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w′) − ψ2 ⊲ ψ3(vcb2)ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w′) + ψ2 ⊲ ψ3(vcb2)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw′) = ϕ2 ⊲ψ3 ϕ3(v)(ϕ3(c) − ψ3(c))ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w) + ψ3(c)ϕ2 ⊲ψ3 ϕ3(vb2)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w′) − ψ3(c)ψ2 ⊲ ψ3(vb2)ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(ua)ϕ1 ⊲ψ2 ϕ2(w′) + ψ3(c)ψ2 ⊲ ψ3(vb2)ϕ1 ⊲ψ2 ⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(uaw′). (3.13) (3.14) (3.15) Then ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(x) = (ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3(x) for any x containing n + 1 elements of A3. By induction, ϕ1 ⊲ψ2⊲ψ3 (ϕ2 ⊲ψ3 ϕ3)(x) = (ϕ1 ⊲ψ2 ϕ2) ⊲ψ3 ϕ3(x) holds for all elementary words x in A1 ∗nu A2 ∗nu A3. Using the associativity of the c-monotone product, we can define the product ⊲i∈I (Ai, ϕi, ψi) for a linearly ordered set I. Now we come to define the concept of c-monotone independence. 9 Definition 3.8. Let (A, ϕ, ψ) be an algebraic probability space equipped with two states. We assume that A is unital. Let I be a linearly ordered set. A family of subalgebras {Ai}i∈I is said to be c-monotone independent if the following properties are satisfied: (1) ϕ(a1a2 · · · an) = ϕ(a1)ϕ(a2 · · · an) for ak ∈ Aik with i1, i2,· · · , in ∈ I and i1 > i2; (2) ϕ(a1a2 · · · an) = ϕ(an)ϕ(a1 · · · an−1) for ak ∈ Aik with i1, i2,· · · , in ∈ I and in > in−1; (3) ϕ(a1a2 · · · an) = (ϕ(aj) − ψ(aj))ϕ(a1 · · · aj−1)ϕ(aj+1 · · · an) + ψ(aj)ϕ(a1a2 · · · aj−1aj+1 · · · an) for ak ∈ Aik with i1, i2,· · · , in ∈ I and ij−1 < ij > ij+1 with 2 ≤ j ≤ n − 1; (4) Ai are monotone independent with respect to ψ. Remark 3.9. (1) This definition includes monotone (resp. Boolean) independence in the special case ϕ = ψ (resp. ψ = 0) on Ai for all i ∈ I. (2) One can define the independence for random variables by considering the subalgebra generated by each random variable without the unit. (3) Let a1 and a2 be c-monotone independent and self-adjoint. If ai has a pair of probability distri- butions (µi, νi) under a pair of states, then the distribution of a1 + a2 is (µ1, ν1) ⊲ (µ2, ν2). This fact follows from the definition of the c-monotone convolution. We can prove that Ai (i ∈ I) are c-monotone independent in (A, ϕ, ψ) = ⊲i∈I (Ai, ϕi, ψi). Here we mention a relation between c-monotone/free independence and orthogonal independence. The reader is referred to [26] for the definition and properties of orthogonal independence and the orthogonal (additive) convolution. Proposition 3.10. With the notation in Definition 3.1, A1 is orthogonal to A2 with respect to the linear functional (ϕ1 ⊲ψ2 02, ψ1 ⊲ ψ2) in the algebra A := A1 ∗nu A2 for arbitrary linear functionals ϕ1, ψ1, ψ2. 02 denotes the 0 linear functional on A2. Remark 3.11. Orthogonal independence has been defined in a unital algebra [26]. To state this proposition strictly along the line of the original definition, we only need to unitize A and extend the linear functionals to eA. Proof. This fact follows from Theorem 3.6 immediately. In terms of convolutions of probability measures, we obtain the following relations. (µ, δ0) ⊞ (δ0, ν) = (µ ⊢ ν, ν), (µ, λ) ⊲ (δ0, ν) = (µ ⊢ ν, λ ⊲ ν). (3.16) (3.17) We give another application: The orthogonal convolution is characterized by Hµ⊢ν(z) = Hµ(Hν(z)) − Hν(z) + z (see Theorem 6.2 of [26]). If we use the relation in Proposition 3.3, this characterization follows immediately. the equality µ ⊲ ν = (µ ⊢ ν) ⊎ ν (see Corollary 6.6 in [26]) can be understood in terms of the associativity of the c-monotone convolution. We first note that ⊞δ0 ν, ν) = (ν, ν) and (µ, µ) ⊲(δ0, ν) = (µ ⊢ ν, µ). Using the associativity we can (δ0, ν) ⊲(ν, δ0) = (δ0δ0 calculate (µ, µ)⊲(δ0, ν)⊲(ν, δ0) in different two ways and we obtain ((µ ⊢ ν)⊎ν, µ⊲ν) = (µ⊲ν, µ⊲ν). 4 Conditionally monotone cumulants 4.1 Power additivity of cumulants In the paper [21], monotone cumulants were introduced. The crucial concept is the dot operation N.X which is defined as the sum of independent random variables with the same distributions as 10 X. If X is a self-adjoint element with the distribution µ, the random variable N.X has the proba- bility distribution µ⋆N , where ⋆ is the additive convolution associated to a concept of independence. Moreover, the uniqueness of cumulants holds also in the setting of two states (we explain this briefly at the end of this subsection). While the dot operation is fundamental for cumulants, we use a method based on the character- ization in Proposition 3.3 to define cumulants. This method is useful to obtain relations between generating functions. The c-monotone convolution consists of two components. While the first component comes from a c-free convolution, c-free cumulants are not useful to find c-monotone cumulants. For the second component, we use the monotone cumulants. We define c-monotone cumulants for the first component. Let Dλ be the dilation operator defined so thatRR f (x)(Dλµ)(dx) =RR f (λx)µ(dx) for all bounded continuous functions f . Lemma 4.1. Let µ ∈ Pm. We define bn = bn(µ) by Hµ(z) = z(cid:16)1 +P∞n=1 asymptotic expansion. Then bn is of the form bn zn(cid:17) in the sense of bn = −mn + Wn(m1,· · · , mn−1), (4.1) where Wn is a polynomial. Moreover, bn(Dλµ) = λnbn(µ) for λ > 0. In particular, Wn does not contain a constant term or linear terms. Proof. We have Hµ(z) = z 1 1 +P∞k=1 = z(cid:16)1 − ∞Xk=1 mk zk mk zk + ( ∞Xk=1 mk zk )2 − · · ·(cid:17), (4.2) so that the polynomials Wn exist. The last statement follows from the relation HDλµ(z) = λHµ(λ−1z). Lemma 4.2. For any n ≥ 3, there exists a polynomial Yn of 2n−3 variables such that bn(µ1 ⊲ν2 µ2) = bn(µ1) + bn(µ2) + Yn(b1(µ1),· · · , bn−1(µ1), b1(ν2),· · · , bn−2(ν2)). In addition, Yn does not contain a constant term or linear terms. For n = 1 and 2, we have b1(µ1 ⊲ν2 µ2) = b1(µ1) + b1(µ2) and b2(µ1 ⊲ν2 µ2) = b2(µ1) + b2(µ2). Proof. We have the equality Hµ1⊲ν2 µ2(z) = Hµ1(Hν2(z)) + Hµ2(z) − Hν2(z) = b1(µ1) + = b1(µ1) + ∞Xn=1 ∞Xn=1 bn+1(µ1)Gν2(z)n + z + b1(µ2) + bn+1(µ1) zn (cid:16)1 + ∞Xk=1 mk(ν2) zk (cid:17)n ∞Xn=1 bn+1(µ2) zn (4.3) + z + b1(µ2) + bn+1(µ2) zn . ∞Xn=1 We can express mk as a polynomial of rn (1 ≤ p ≤ k) from Lemma 4.1: mk(ν2) = −bk(ν2) + Xk(b1(ν2),· · · , bk−1(ν2)), where Xk is a polynomial without a constant term or linear terms. There- fore, we have the conclusion. Let rn(ν) be the monotone cumulants of µ and Aν(z) = −P∞n=1 rn(ν) zn−1 be their generating function. Let {(µt, νt)}t≥0 be a c-monotone convolution semigroup with (µ0, ν0) = (δ0, δ0). Then {νt}t≥0 is a monotone convolution semigroup with ν0 = δ0. We define µ := µ1 and ν = ν1. Proposition 3.3 11 implies that Hµs+t = Hµs ◦ Hνt + Hµt − Hνt. We assume that mn(µt) is a differentiable function of t. Differentiating the equality with s formally and putting s = 0, we obtain d dt Hµt(z) = A(µ,ν)(Hνt(z)), where A(µ,ν)(z) = d dtHµt(z)t=0. We expand A(µ,ν)(z) as a formal power series: A(µ,ν)(z) = − rn(µ, ν) zn−1 . ∞Xn=1 (4.4) is equivalent to d dt Gµt(z) = ∞Xn=1 rn(µ, ν)Gµt(z)2Gνt(z)n−1 (4.4) (4.5) (4.6) (4.7) as a formal power series. Thus there exists a polynomial Mn for any n ≥ 3 such that mn(µt) = rn(µ, ν) + Mn(r1(µ, ν),· · · , rn−1(µ, ν), m1(µt),· · · , mn−1(µt), d dt m1(νt),· · · , mn−2(νt)). For n = 1 and 2, we can understand that M1 = 0 and that M2 only depends on r1(µ, ν) and m1(µt). Since mk(νt) is a polynomial of r1(ν),· · · , rk(ν), we can prove inductively that mn(µt) is of the form (4.8) mn(µt) = rn(µ, ν)t + t2Cn(t, r1(µ, ν),· · · , rn−1(µ, ν), r1(ν),· · · , rn−2(ν)), where Cn is a polynomial for any n ≥ 3 and rk(µ) is the k-th monotone cumulant of µ. From the construction, Cn is a universal polynomial in the sense that Cn does not depend on µ or ν; Cn is determined only by the definition of c-monotone independence. It is easy to show that m1(µt) = r1(µ, ν)t and m2(µt) = r2(µ, ν)t + r1(µ, ν)2t2. If we set t = 1, rn(µ, ν) turns out to be expressed as a polynomial of mk(µ) and mk(ν) (1 ≤ k ≤ n). This relation can be generalized to any probability measures µ, ν ∈ Pm since Cn is universal. Definition 4.3. For probability measures µ, ν ∈ Pm, we define the c-monotone cumulants rn(µ, ν) (n ≥ 1) by the equations mn(µ) = rn(µ, ν) + Cn(1, r1(µ, ν),· · · , rn−1(µ, ν), r1(ν),· · · , rn−2(ν)) for n ≥ 3. For n = 1, 2, we define r1(µ, ν) := m1(µ) and r2(µ, ν) := m2(µ) − m1(µ)2. Now we prove the power additivity of cumulants. The proof is based on the relation of generating functions in Proposition 3.3. We note, however, that we can give a proof without the use of generating functions as shown in [21]. Theorem 4.4. rn((µ, ν)⊲N ) = Nrn(µ, ν) holds for µ, ν ∈ Pm and N ∈ N. Proof. We define µN by (µN , ν ⊲N ) = (µ, ν)⊲N . We define mn(ν, t) := nXk=1 X 1=i0<i1<···<ik−1<ik=n+1 tk k! kYl=1 il−1ril−il−1(ν) and mn(µ, ν, t) := rn(µ, ν)t + t2Cn(t, r1(µ, ν),· · · , rn−1(µ, ν), r1(ν),· · · , rn−2(ν)), which may not be moments of a probability measure for general t. It is noted that mn(µ, ν, 1) = mn(µ) and mn(µ, µ, t) = mn(µ, t). The latter equality comes from the relation (µ, µ) ⊲ (ν, ν) = (4.9) (4.10) 12 (µ ⊲ ν, µ ⊲ ν). We shall show that mn(µ, ν, N) = mn(µN , ν ⊲N , 1)(= mn(µN )) for any n, N ≥ 1. We and Hν(t, z) := define formal power series A(µ,ν)(z) using (4.5), H(µ,ν)(t, z) := (cid:16)P∞n=0 (cid:16)P∞n=0 zn+1 (cid:17)−1 zn+1 (cid:17)−1 . Now we prove the equalities mn(µ,ν,t) mn(ν,t) H(µ,ν)(t + s, z) = H(µ,ν)(t, Hν(s, z)) − Hν(s, z) + H(µ,ν)(s, z), Hν(t + s, z) = Hν(t, Hν(s, z)), (4.11) (4.12) as formal power series. To do so, we define K 1 K 2 s (t, z) := Hν(t, Hν(s, z)). With the arguments just before Definition 4.3, we have s (t, z) := H(µ,ν)(t, Hν(s, z))− Hν(s, z) + H(µ,ν)(s, z) and d dt d dt H(µ,ν)(t, z) = A(µ,ν)(Hν(t, z)), Hν(t, z) = Aν(Hν(t, z)). (4.13) The latter equality follows from the former by setting µ = ν. Therefore it is clear that (H(µ,ν)(t + s, z), Hν(t + s, z)) satisfies the two-dimensional complex differential equation d dt d dt H(µ,ν)(t + s, z) = A(µ,ν)(Hν(t + s, z)), Hν(t + s, z) = Aν(Hν(t + s, z)) (4.14) for a fixed s with the initial value (H(µ,ν)(s, z), Hν(s, z)). On the other hand, (4.13) implies that d dt d dt K 1 s (t, z) = A(µ,ν)(K 2 s (t, z)), K 2 s (t, z) = Aν(K 2 s (t, z)) (4.15) s (t, z), K 2 s (0, z), K 2 with the initial value (K 1 s (0, z)) = (H(µ,ν)(s, z), Hν(s, z)). Thus both (H(µ,ν)(t + s, z), Hν(t + s, z)) and (K 1 s (t, z)) satisfy the same differential equation with the same initial value. It is not difficult to prove the uniqueness of solution of an ordinary differential equation as a formal power series; therefore (H(µ,ν)(t + s, z), Hν(t + s, z)) = (K 1 s (t, z)). Setting t = s = 1, we obtain Inductively, we can show that mn(µ, ν, N) = mn(µN ). mn(µ, ν, 2) = mn(µ2, ν ⊲ ν, 1) = mn(µ2). By using the (power) associativity of the c-monotone convolution, we also obtain mn(µ, ν, MN) = mn(µN , ν ⊲N , M) for M, N ∈ N. Then we have s (t, z), K 2 MNrn(µ, ν) + M 2N 2Cn(MN, r1(µ, ν),· · · , rn−1(µ, ν), r1(ν),· · · , rn−2(ν)) = Mrn(µN , ν ⊲N ) + M 2Cn(M, r1(µN , ν ⊲N ),· · · , rn−1(µN , ν ⊲N ), r1(ν ⊲N ),· · · , rn−2(ν ⊲N )). (4.16) Since the coefficients of M coincide, it holds that rn(µN , ν ⊲N ) = Nrn(µ, ν). Before closing this subsection, we note some remarks. First we summarize the basic relations between generating functions: d dt H(µ,ν)(t, z) = A(µ,ν)(Hν(t, z)), H(µ,ν)(1, z) = Hµ(z), H(µ,ν)(0, z) = z, d dt Hν(t, z) = Aν(Hν(t, z)), Hν(1, z) = Hν(z), Hν(0, z) = z. These relations are analogous to Hµ(z) = z − φ(µ,ν)(Hν(z)), Hν(z) = z − φν(Hν(z)), 13 (4.17) (4.18) (4.19) (4.20) which are basic for c-free convolutions. We note that monotone cumulants and Boolean cumulants are special cases of c-monotone cumulants: A(µ,µ) is a generating function of monotone cumulants and A(µ,δ0) is a generating function of Boolean cumulants. Second, we note that c-monotone cumulants rn(µ, ν) satisfy the following conditions. (K1') Power additivity: rn((µ, ν)⊲N ) = Nrn(µ, ν). (K2) Homogeneity: for any λ > 0 and any n, rn(Dλµ, Dλν) = λnrn(µ, ν), (4.21) where Dλ is defined by (Dλµ)(B) = µ(λ−1B) for any Borel set B. (K3) For any n, there exists a universal polynomial Qn (in the sense that Qn does not depend on µ or ν) of 2n − 2 variables such that rn(µ, ν) = mn(µ) + Qn(mp(µ), mq(ν) (1 ≤ p, q ≤ n − 1)). (4.22) The usual additivity rn((µ1, ν1) ⊲ (µ2, ν2)) = rn(µ1, ν1) + rn(µ2, ν2) does not hold due to the non- commutativity of the convolution. We can easily prove the uniqueness of cumulants satisfying (K1'), (K2) and (K3) (see [21] and also Section 13 of the present paper). 4.2 Moment-cumulant formula and monotone partitions Next we show a combinatorial moment-cumulant formula for the c-monotone convolution. Let LNC(n) be the set of all linearly ordered non-crossing partitions defined by LNC(n) := {(π, λ) : π ∈ NC(n), λ is a linear ordering of the blocks of π}. (4.23) We prove that the n-th moment can be described as mn(µ) = X(π,λ)∈M(n) 1 π!(cid:16) YV ∈π, V : outer rV (µ, ν)(cid:17)(cid:16) YV ∈π, V : inner rV (ν)(cid:17). This formula is analogous to the c-free formula mn(µ) = X(π,λ)∈LNC(n) 1 π!(cid:16) YV ∈π, V : outer RV (µ, ν)(cid:17)(cid:16) YV ∈π, V : inner RV (ν)(cid:17) (4.24) (4.25) if we impose the linear order structure. Clearly the role of the linear order structure is trivial in (4.25), but crucial in (4.24). When we defined c-monotone cumulants, we used the Taylor expansion of the equality d dt Hµt(z) = A(µ,ν)(Hνt(z)) to prove the power additivity rn((µ, ν)⊲N ) = Nrn(µ, ν). By contrast, the following is of use in proving the moment-cumulant formula. Proposition 4.5. Let rn(µ) and rn(µ, ν) be the momotone cumulants and c-monotone cumulants re- spectively, and let mn(µ, ν, t) be the quantity defined in (4.10). Then we have the recursive differential equations d dt mn(µ, ν, t) = n−1Xk=0 (k + 1)rn−k(ν)mk(µ, ν, t) − n−1Xk=0 kXl=0 rn−k(ν)ml(µ, ν, t)mk−l(µ, ν, t) (4.26) + n−1Xk=0 kXl=0 rn−k(µ, ν)ml(µ, ν, t)mk−l(µ, ν, t). 14 d dt mn(µ, δ0, t) = n−1Xk=0 kXl=0 rn−k(µ, δ0)ml(µ, δ0, t)mk−l(µ, δ0, t), (4.27) Remark 4.6. In the special case µ = ν, the above equation becomes d k=0(k + 1)rn−k(µ)mk(µ, t) which has been obtained in the case of monotone convolutions [21]. Moreover, we have dt mn(µ, t) = Pn−1 since rn(δ0) = 0. In this case rn(µ, δ0) is an n-th Boolean cumulant. It might be interesting that this differential equation describes the structure of interval partitions. Proof. We use the same notation as used in the subsection 4.1. Differentiating the equality (4.11) with s and putting s = 0 we obtain ∂H(µ,ν) ∂t (t, z) = Aν(z) ∂H(µ,ν) ∂z (t, z) − Aν(z) + A(µ,ν)(z). (4.28) We define G(µ,ν)(t, z) := 1 H(µ,ν)(t,z) . Then we get the equality ∂G(µ,ν) ∂t (t, z) = Aν(z) ∂G(µ,ν) ∂z (t, z) + G(µ,ν)(t, z)2Aν(z) − G(µ,ν)(t, z)2A(µ,ν)(z). (4.29) (4.26) follows from the expansion of (4.29) in formal power series. We say a block V in a partition π ∈ P(n) is of interval type if there exist j and k (1 ≤ j ≤ n, 0 ≤ k ≤ n − j) such that V = {j, j + 1,· · · , j + k}. Theorem 4.7. The moment-cumulant formula for c-monotone independence is mn(µ) = X(π,λ)∈M(n) 1 π!(cid:16) YV ∈π, V : outer rV (µ, ν)(cid:17)(cid:16) YV ∈π, V : inner rV (ν)(cid:17). (4.30) Proof. This proof gives the combinatorial meaning of the differential equations (4.26). The claim is proved by induction on n: assume that the formula mn(µ, ν, t) = X(π,λ)∈M(n) tπ π!(cid:16) YV ∈π, V : outer rV (µ, ν)(cid:17)(cid:16) YV ∈π, V : inner rV (ν)(cid:17) (4.31) holds for 1 ≤ n ≤ N, where mn(µ, ν, t) is defined in (4.10). Let π = {V1 < · · · < Vπ} denote a monotone partition (π, λ) ∈ M(N + 1); this notation means that a subscript of a block itself expresses the linear ordering. Note that Vπ is always a block of interval type. Let k be defined as Vπ = N + 1 − k (0 ≤ k ≤ N). Two cases are possible: (1) Vπ is outer; (2) Vπ is inner. (1) If Vπ is an outer block, π is of such a form as π σ Vπ τ where σ and τ are arbitrary non-crossing partitions with σ ∈ NC(l) and τ ∈ NC(k − l) (0 ≤ l ≤ k). We understand that σ = ∅ for l = 0 and τ = ∅ for l = k. Next we need to consider the linear orderings of σ and τ . A linear ordering of the blocks of σ ∪ τ can be described by distributing natural numbers {1,· · · ,σ + τ} to them such that σ and τ , equipped with the natural numbers, become monotone partitions. Once σ and τ are fixed, there exist (σ+τ)! σ!τ! ways to divide the set {1,· · · ,σ + τ} into two subsets C and D with C = σ and D = τ. After such a division, we 15 can choose linear orderings of σ from C and τ from D independently with each other, and hence, we can take arbitrary (σ, ρ) ∈ M(l) and (τ, µ) ∈ M(k − l). The above arguments imply that = tπ V : outer Vπ: outer X(π,λ)∈M(N +1); NXk=0 NXk=0 π!(cid:16) YV ∈π, kXl=0 X(σ,ρ)∈M(l) rN +1−k(µ, ν)Z t kXl=0 (τ,µ)∈M(k−l) = 0 rV (µ, ν)(cid:17)(cid:16) YV ∈π, V : inner rV (ν)(cid:17) ml(µ, ν, s)mk−l(µ, ν, s)ds, (σ + τ)! σ!τ! tσ+τ+1rN +1−k(µ, ν) (σ + τ + 1)! (cid:16) YV ∈σ∪τ, V : outer rV (µ, ν)(cid:17)(cid:16) YV ∈σ∪τ, V : inner rV (ν)(cid:17) (4.32) where we have used the assumption of induction in the last line. (2) Next we consider the case where Vπ is inner. Since sums with Vπ inner blocks are difficult to compute directly, first we sum over all monotone partitions and later we remove the redundant sums, i.e., the sums over monotone partitions such that Vπ are outer. When Vπ = N + 1− k, there are k + 1 ways to choose a position of Vπ as a subset of {1,· · · , N}. After the choice of Vπ, we can take arbitrary monotone partition of M(k) as if the highest block Vπ were absent. The prototype of this idea appeared in [38] and was used in [21]. The above arguments then amount to Vπ: inner X(π,λ)∈M(N +1); NXk=0 = tπ π!(cid:16) YV ∈π, V : outer rV (µ, ν)(cid:17)(cid:16) YV ∈π, V : inner rV (ν)(cid:17) (σ + 1)!(cid:16) YV ∈σ, tσ+1 V : outer (k + 1)rN +1−k(ν) X(σ,ρ)∈M(k) NXk=0 kXl=0 X(σ,ρ)∈M(l) (k + 1)rN +1−k(ν)Z t (τ,λ)∈M(k−l) 0 (σ + τ)! σ!τ! − = NXk=0 tσ+τ+1rN +1−k(ν) (σ + τ + 1)! (cid:16) YV ∈σ∪τ, rN +1−k(ν)Z t V : outer 0 NXk=0 kXl=0 π!(cid:16) YV ∈π, tπ V : outer mk(µ, ν, s)ds − ml(µ, ν, s)mk−l(µ, ν, s)ds. Combining the results from (1), (2) and Proposition 4.5, we have mN +1(µ, ν, t) = X(π,λ)∈M(N +1) rV (µ, ν)(cid:17)(cid:16) YV ∈π, V : inner rV (ν)(cid:17). (4.34) rV (µ, ν)(cid:17)(cid:16) YV ∈σ, V : inner rV (ν)(cid:17) rV (µ, ν)(cid:17)(cid:16) YV ∈σ∪τ, V : inner (4.33) rV (ν)(cid:17) Example 4.8. The moment-cumulant formula for n = 1, 2, 3, 4 is calculated as m1(µ) = r1(µ, ν), m2(µ) = r2(µ, ν) + r1(µ, ν)2, m3(µ) = r3(µ, ν) + 2r2(µ, ν)r1(µ, ν) + 1 2 r2(µ, ν)r1(ν) + r1(µ, ν)3, m4(µ) = r4(µ, ν) + 2r3(µ, ν)r1(µ, ν) + r3(µ, ν)r1(ν) + r2(µ, ν)2 + + 3r2(µ, ν)r1(µ, ν)2 + r2(µ, ν)r1(µ, ν)r1(ν) + 1 3 16 1 2 r2(µ, ν)r2(ν) r2(µ, ν)r1(ν)2 + r1(µ, ν)4. 5 Limit theorems We show the central limit theorem and Poisson's law of small numbers for c-monotone independence. The t-transformation Ut [10], defined by HUt(µ)(z) = (1 − t)z + tHµ(z), appears in the limit distributions. Theorem 5.1. Let (A, ϕ, ψ) be a C∗-algebraic probability space with two states. (1) Let {Xi}∞i=1 be identically distributed (with respect to each state), c-monotone independent self- i ) = β2, then the adjoint random variables in A. distribution of X1+···+Xn with respect to (ϕ, ψ) converges to (Uα2/β2(νβ2), νβ2) weakly, where νβ2 is the arcsine law with variance β2. Uα2/β2(νβ2) is a Kesten distribution. (2) Let X (n) If ϕ(Xi) = ψ(Xi) = 0, ϕ(Xi) = α2 and ψ(X 2 N (1 ≤ n ≤ N, 1 ≤ N < ∞) be self-adjoint random variables such that √n (a) for each N, X (n) N (1 ≤ n ≤ N) are identically distributed with respect to each state, c-monotone independent self-adjoint random variables; (b) Nϕ((X (1) N )k) → λ > 0 and Nψ((X (1) N )k) → ρ > 0 as N → ∞ for all k ≥ 1. Then the distribution of XN := X (1) where pρ is the monotone Poisson distribution with parameter ρ. N +· · ·+X (N ) N with respect to (ϕ, ψ) converges weakly to (Uλ/ρ(pρ), pρ), Proof. We discuss the problems in terms of probability measures. Let µ, ν ∈ Pc such that m1(µ) = m1(ν) = 0, m2(µ) = α2 and m2(ν) = β2 with α2, β2 > 0. We define (µN , νN ) := (D 1√N ν)⊲N . The convolution for the second component is the usual monotone convolution, so that the limit distribution νβ2 exists and it is the centered arcsine law with variance β2. A simple calculation shows that r1(µN , νN ) = 0, r2(µN , νN ) = α2 and rn(µN , νN ) → 0 as N → ∞. Therefore, there exists a pair of limit distributions (να2,β2, νβ2) at least in the sense of moments. The limit distributions are characterized by µ, D 1√N d dt H(να2,β2 ,νβ2 )(t, z) = − α2 Hνβ2 (t, z) , H(να2,β2 ,νβ2 )(1, z) = Hνα2,β2 (z), d dt Hνβ2 (t, z) = − β2 Hνβ2 (t, z) , Hνβ2 (1, z) = Hνβ2 (z). (5.1) (5.2) 1 We obtain Hνβ2 (t, z) =pz2 − 2β2t from (5.2). Setting t = 1, we obtain the limit measure νβ2(dx) = dx, x ∈ (−√2β,√2β). For the first component, we can easily prove that H(να2,β2 ,νβ2 )(1, z) = π√2β2−x2 β2pz2 − 2β2, and hence (1 − α2 β2 )z + α2 Gνα2,β2 (z) = α2pz2 − 2β2 + (α2 − β2)z (2α2 − β2)z2 − 2α4 . (5.3) This is the Stieltjes transform of a Kesten distribution. In terms of the t-transformation Ut, the limit distribution can be written as Uα2/β2(νβ2). It is compactly supported, so that the convergence is in the sense of weak convergence (see Theorem 4.5.5 of [12]). We note that νβ2,β2 = νβ2. Next we consider Poisson's law of small numbers. Let µ(N ), ν(N ) ∈ Pc such that Nmn(µ(N )) → λ > 0 and Nmn(ν(N )) → ρ > 0 as N → ∞ for any n ≥ 1. We define (µN , νN ) := (µ(N ), ν(N ))⊲N . It is not difficult to prove that rn(µN , νN ) → λ as N → ∞ for any n ≥ 1. It is known that νN 17 converges weakly to the monotone Poisson distribution pρ with parameter ρ, which is characterized by the differential equation d dt Hpρ(t, z) = ρHpρ(t, z) 1 − Hpρ(t, z) , Hpρ(1, z) = Hpρ(z), (5.4) where Hpρ(1, z) = Hpρ(z). We let pλ,ρ denote the limit distribution limN→∞ µN in the sense of moments. Then pλ,ρ is characterized by the differential equation d dt H(pλ,ρ,pρ)(t, z) = λHpρ(t, z) 1 − Hpρ(t, z) , H(pλ,ρ,pρ)(1, z) = Hpλ,ρ(z). (5.5) From (5.4) and (5.5) we have d dt H(pλ,ρ,pρ)(t, z) = λ ρ d dtHpρ(t, z), which implies Hpλ,ρ(z) =(cid:16)1 − λ ρ(cid:17)z + λ ρ Hpρ(z). (5.6) It is not difficult to see that pλ,ρ has a compact support. Therefore, µN converges weakly to pλ,ρ. Also in this case, pλ,ρ can be written as pλ,ρ = Uλ/ρ(pρ). 6 Convolution semigroups We investigate convolution semigroups and infinite divisibility from this section. First we establish the equivalence between a pair of vector fields and a weakly continuous c-monotone convolution semigroup. We follow the method used by Muraki. We use the notation Ht(z) and Ft(z) for the reciprocal Cauchy transforms of the left component of a semigroup and of the right one, respectively. Theorem 6.1. Let {Ft(z)}t≥0 be a composition semigroup of reciprocal Cauchy transforms with F0(z) = z. We assume that Ft(z) is continuous with respect to t ∈ [0,∞) for each fixed z ∈ C\R. Let {Ht(z)}t≥0 be a family of reciprocal Cauchy transforms with H0(z) = z, satisfying the same continuity condition as Ft(z) and satisfying the equality Ht+s(z) = Ht(Fs(z)) − Fs(z) + Hs(z). Then there exist analytic vector fields A1 and A2 such that Ht and Ft satisfy the differential equations d dt d dt Ht(z) = A1(Ft(z)), Ft(z) = A2(Ft(z)). (6.1) (6.2) (6.3) Moreover, the vector fields A1(z) = d dt0Ht(z) and A2(z) = d 1 + xz x − z Aj(z) = −γj +ZR dt0Ft(z) have the representations dτj(x) for j = 1, 2, where τj is a positive finite measure and γj ∈ R (This is the L´evy-Khintchine formula for the c-monotone convolution). Conversely, given analytic vector fields A1 and A2 on the upper halfplane of the forms (6.3), by solving the equations (6.1) and (6.2) we obtain {Ft(z)}t≥0 and {Ht(z)}t≥0 which are C ω functions with respect to (t, z) and satisfy F0(z) = z, H0(z) = z, Ft+s = Ft ◦ Fs and Ht+s(z) = Ht(Fs(z)) − Fs(z) + Hs(z). Proof. We use the representations Ht(z) = at + z +ZR Ft(z) = bt + z +ZR 18 1 + xz x − z 1 + xz x − z dηt(x), dξt(x). (6.4) (6.5) We follow the method of [31]. The claims for Ft were proved in [31]; we only have to prove the claims for Ht. However, the proof below actually includes that for Ft when we put Ht = Ft. A direct calculation leads to the following equality: Ht+s(z) − Ht(z) = as +ZR 1 + xz x − z ηs(dx) + (Fs(z) − z)ZR 1 + x2 (x − Fs(z))(x − z) ηt(dx) (6.6) = (Hs(z) − z) + (Fs(z) − z)ZR 1 + x2 (x − Fs(z))(x − z) ηt(dx). Step (1): the right differentiability of Ht and Ft at 0. Take δ > 0, n ∈ N and k ∈ N (0 ≤ k ≤ n). We set s = δ n and t = k nδ. Summing the equality (6.6) over k, we obtain Hδ(z) − z = n(H δ n (z) − z) + (F δ n (z) − z) Then Hδ(z) − z δ = H δ n (z) − z δ/n + δ−1 · F δ n (z) − z δ/n · n−1Xk=0ZR n−1Xk=0ZR δ n 1 + x2 (z))(x − z) n (x − F δ n δ(dx). η k (6.7) 1 + x2 (z))(x − z) n (x − F δ η k n δ(dx). (6.8) Since Ht(i) = at +(1+ηt(R))i, at is a continuous function. If Fs(z) = z for some s > 0, then Fs(z) = z for all s > 0, so that the parameter t of Ht expresses the Boolean time evolution. In this case, the (x−Fs(z))(x−z) ηt(dx) is a continuous function of t since it can be expressed as (Ht(Fs(z)) − Ht(z) − Fs(z) + z)/(Fs(z) − z). Therefore, claims are trivial. Therefore, we assume that Fs(z) 6= z for all s > 0. ThenRR 1+x2 δ n lim n→∞ n−1Xk=0ZR 1 + x2 (z))(x − z) n (x − F δ η k n δ(dx) =Z δ 0 dtZR We note that limtց0 ηt(R) = 0, and hence, for small enough δ > 0 the integral 1 is small. If we put Ht equal to Ft, (6.9) implies the existence of 1 + x2 (x − z)2 ηt(dx). δR δ 0 dtRR (6.9) 1+x2 (x−z)2 ηt(dx) (6.10) Aδ,2(z) := lim n→∞ F δ n (z) − z δ/n = (Hδ(z) − z)/δ 1 + 1 δR δ 0 dtRR 1+x2 (x−z)2 ηt(dx) for small δ, which has been obtained in [31]. The limit does not depend on δ > 0 as shown in [31]; we summarize the proof here. We can prove easily that Arδ,2 = Aδ,2 for any positive rational number r. The equality (6.10) implies that Aδ,2 depends on δ continuously. Therefore, Arδ,2 = Aδ,2 for every positive real number r > 0. Then it may be simply denoted by A2(z). We note again that the equality A2(z) = (Fδ(z) − z)/δ 1 + 1 δR δ 0 dtRR 1+x2 (x−z)2 ηt(dx) (6.11) Fδ(z)−z holds. Taking δ → 0 in the above equality we obtain A2(z) = limδց0 . Again for gen- eral (Ht, Ft), by taking the limit n → ∞ in (6.8) we have the existence of the limit A1,δ(z) = limn→∞ (z)−z δ/n which satisfies H δ n δ Hδ(z) − z δ = A1,δ(z) + A2(z)δ−1Z δ 0 dtZR 1 + x2 (x − z)2 ηt(dx). (6.12) With the same argument as for A2,δ, A1,δ does not depend on δ, so that it may be denoted by A1. Taking the limit δ ց 0 in (6.12), we have the right differentiability limδց0 = A1(z). Hδ(z)−z δ 19 HT (z) − HT−δ(z) δ = H δ n Taking the limit n → ∞, we obtain HT (z) − HT−δ(z) · F δ n δ n + δ−1 · (x − F δ (z) − z δ/n (z) − z δ/n nXk=1ZR = A1(z) + δ−1A2(z)Z δ dtZR D−t Ht(z) = A1(z) + A2(z)(cid:16) ∂Ht (z) − 1(cid:17) (t > 0) 1 + x2 ∂z δ 0 (x − z)2 ηT−t(dx). Moreover, let δ ց 0, to know that the limit Step (2): the differentiability of Ht and Ft. We do not refer to the claims for Ft, since it is known in [31], or since the proof below is true for Ft if we put Ht = Ft. We define the right and left derivatives by setting D+ t Ht(z) = limδց0 Ht+δ(z)−Ht(z) δ and D−t Ht(z) = limδց0 Ht−δ(z)−Ht(z) δ respectively. Then D+ t Ht(z) = lim δց0 Hδ(Ft(z)) − Ft(z) δ = A1(Ft(z)). (6.13) Take T > 0, δ > 0, n ∈ N and k ∈ N (0 ≤ k ≤ n). From an argument similar to the derivation of (6.8), we have 1 + x2 (z))(x − z) n ηT− k n δ(dx). (6.14) exists. Finally, the differentiability of Ht at t > 0 follows from the calculations D+ t Ht(z) = lim δց0 Ht(Fδ(z)) − Ht(z) − Fδ(z) + Hδ(z) δ Ht(Fδ(z)) − Ht(z) Fδ(z) − z (z) + A1(z) − A2(z) δ = lim δց0 ∂Ht = A2(z) ∂z = D−t Ht(z). Fδ(z) − z · + lim δց0 Hδ(z) − z δ Fδ(z) − z δ − lim δց0 (6.15) (6.1) follows from (6.13). Then Ht(z) = z +R t Step (3): the representation of A1(z). It is sufficient to prove that (i) A1(z) is a Pick function, i.e., A1(z) is analytic in C+ and maps C+ into C+ ∪R; (ii) limy→∞ Now we know the relation ImA1(iy) = 0. y 0 A1(Fs(z))ds, which implies that Ht(z) is in C ω([0,∞)). A1(z) = Hδ(z) − z δ − A2(z)δ−1Z δ 0 dtZR 1 + x2 (x − z)2 ηt(dx), (6.16) from which A1 is analytic. Since Hδ(z)−z the property (ii) by using (6.16). δ is a Pick function, its limit A1 is also. It is easy to prove Step (4): the converse statement. We note that the solution of the differential equation Ft(z) = A2(Ft(z)), d dt F0(z) = z, 20 (6.17) does not explode in a finite time [4]. Then the equation defines a flow of reciprocal Cauchy transforms indexed by the non-negative real numbers. We can define Ht by setting Ht(z) = z +Z t 0 A1(Fs(z))ds. (6.18) We need to prove the equality Ht+s(z) = Ht(Fs(z)) − Fs(z) + Hs(z). We fix s ≥ 0, z ∈ C+ and put Jt(z) := Ht+s(z), Kt(z) := Ht(Fs(z)) − Fs(z) + Hs(z). Then d dt Kt(z) = A1(Ft ◦ Fs(z)) with J0 = K0 by definition, so that d dt Kt. Therefore, Jt = Kt for all t ≥ 0. dt Jt(z) = A1(Ft+s(z)) and d dt Jt = d 7 Constructions of convolution semigroups Several examples of monotone convolution semigroups are known in [19, 31]. In this section, we show several ways to construct c-monotone convolution semigroups from monotone and Boolean convolution semigroups. (1) Let {µt}t≥0 be a weakly continuous Boolean convolution semigroup with µ0 = δ0. Then {(µt, δ0)}t≥0 is a c-monotone convolution semigroup. The vector fields A1 and A2 are given by A1(z) = −z + Hµ1(z), A2(z) = 0. {(µt, µt)}t≥0 is a c-monotone convolution semigroup. The vector fields A1 and A2 coincide. (2) Let {µt}t≥0 be a weakly continuous monotone convolution semigroup with µ0 = δ0. Then (3) Let Ut be the t-transformation [10]. We recall that Ut(µ) is characterized by HUt(µ)(z) = (1 − t)z + tHµ(z). Let {µt}t≥0 be a weakly continuous monotone convolution semigroup with µ0 = δ0. Then {(Ur(µt), µt)}t≥0 is a c-monotone convolution semigroup for any r ≥ 0 since, under the notation Ht(z) := HUr(µt), we have Ht(Fs(z)) − Fs(z) + Hs(z) = (1 − r)Fs(z) + rFt(Fs(z)) − Fs(z) + (1 − r)z + rFs(z) = (1 − r)z + rFt+s(z) = Ht+s(z). In this case, A1(z) = d when r = 1. This property can be understood in terms of cumulants as follows. dtt=0Ht(z) = rA2(z). We note that this construction includes the example (2) Proposition 7.1. The relation rn(µ⊎t, µ) = trn(µ) (7.1) holds for all µ with the finite moments of all orders. This means formally "rn(µ⊎t, µ) = rn(µ⊲t)", which connects the Boolean convolution with the monotone convolution. We will generalize this relation in Corollary 7.4. (4) Let {µt}t≥0 be a weakly continuous monotone convolution semigroup with µ0 = δ0 and a ∈ R. Then {(δat, µt)}t≥0 is a c-monotone convolution semigroup. The vector field A1 is a constant a. (5) The Cauchy distribution λtb(dx) := π(x2+(tb)2)dx is infinitely divisible in the tensor, free, Boolean and monotone cases and becomes a convolution semigroup with the time parameter t ≥ 0. Moreover, λtb is a strictly 1-stable distribution (see [7, 19, 31, 37, 42]) for any one of the four kinds of convolutions. λtb plays a special role also in the case of c-monotone independence; for in- stance, {(λbt, µt)}t≥0 is a c-monotone convolution semigroup for any monotone convolution semigroup {µt}t≥0. A1(z) is a constant ib. (6) We introduce an operation to construct a new c-monotone convolution semigroup for given c-monotone convolution semigroups. tb 21 Definition 7.2. A convolution κu,v: P × P → P is defined by κu,v(µ, ν) := µ⊎u ⊎ ν⊎v for u, v ≥ 0. In terms of reciprocal Cauchy transforms, we have Hκu,v(µ,ν)(z) := uHµ(z) + vHν(z) + (1 − u − v)z for u, v ≥ 0. (7.2) (7.3) Proposition 7.3. (1) Let {(µt, λt)}t≥0 and {(νt, λt)}t≥0 be c-monotone convolution semigroups with the same right component. We let Aµ,λ denote the vector fields for the left component of {(µt, νt)}t≥0 and {(νt, λt)}t≥0, respectively, and also Aλ 2 denote the vector field for the common right component. Then {(κu,v(µt, νt), λt)}t≥0 is a c-monotone convolution semigroup. A pair of the vector fields is given by (uAµ,λ (2) In terms of cumulants, we can understand the convolution κu,v in the form 1 + vAν,λ and Aν,λ 1 1 1 , Aλ 2). rn(κu,v(µ, ν), λ) = urn(µ, λ) + vrn(ν, λ) for all probability measures µ, ν, λ with finite moments of all orders. Proof. Denote by H µ Hκu,v(µt,νt)(Fs) − Fs + Hκu,v(µs,νs) = uH µ t and H ν t the reciprocal Cauchy transforms of µt and νt, respectively. t (Fs) + vH ν + (1 − u − v)z = u(H µ t (Fs) − Fs + H µ = uH µ t+s + vH ν = Hκu,v(µt+s,νt+s). s ) + v(H ν t+s + (1 − u − v)z t (Fs) + (1 − u − v)Fs − Fs + uH µ s + vH ν t t (Fs) − Fs + H ν s ) + (1 − u − v)z The relation for the vector fields follows immediately. The second statement for cumulants follows from the definition of c-monotone cumulants. We obtain a nontrivial property of the c-monotone cumulants. Corollary 7.4. We have the following relation between the additive Boolean convolution and the c-monotone cumulants: rn(µ ⊎ ν, λ) = rn(µ, λ) + rn(ν, λ) for all probability measures µ, ν, λ with finite moments of all orders. (7.4) Remark 7.5. (1) Since (µ, δ0) ⊲ (ν, δ0) = (µ⊎ ν, δ0) and rn(µ, δ0) is the n-th Boolean cumulant, this corollary is trivial for λ = δ0. (2) The uniqueness of Boolean cumulants follows from axioms (C1) and (C2') (see Section 13). In the present case, {rn(·, λ)}n≥1 satisfies (C1) and (C2), but do not satisfy (C2') for any λ 6= δ0. Example 7.6. With the above methods (1)-(6) and with examples of Boolean or monotone con- volution semigroups in the literature, we can construct many examples of c-monotone convolution semigroups. We give an important example among them. Consider the Kesten distribution µσ2,r characterized by Gµσ2,r (z) := √z2 − 2σ2 + (1 − r)z (2 − r)z2 − 2σ2r . (7.5) 1 π√2σ2−x2 dx is the centered arcsine law with variance σ2. We note Ur(µσ2,1) = µσ2,r. µσ2,1(dx) = Then {(µt,r, µt,1)}t≥0 is a c-monotone convolution semigroup since {µt,1}t≥0 is a monotone convolution semigroup. This example is connected with the central limit measure for the c-monotone convolution (see Section 5). 22 In this section, we have shown constructions of c-monotone convolution semigroups and properties of c-monotone cumulants. By the way, we can prove similar results in the conditionally free case. Among them, the following is interesting. We omit the proof. Proposition 7.7. Let Rn(µ, ν) be the c-free cumulants of (µ, ν). Then the identity Rn(µ ⊎ ν, λ) = Rn(µ, λ) + Rn(ν, λ) (7.6) holds. Finally we obtain the following relation between rn(µ, ν) and Rn(µ, ν). We note that we need a result from Section 13. Theorem 7.8. There exist polynomials Pn,k of n − k variables for 2 ≤ k ≤ n − 1, n ≥ 3 such that rn(µ, ν) = Rn(µ, ν) + n−1Xk=2 Pn,k(m1(ν),· · · , mn−k(ν))Rk(µ, ν). (7.7) We note that r1(µ, ν) = R1(µ, ν) = m1(µ) and r2(µ, ν) = R2(µ, ν) = m2(µ) − m1(µ)2 for n = 1, 2. Roughly, rn(µ, ν) is expressed by a linear combination of Rn(µ, ν) with polynomial coefficients mk(ν). Proof. The existence of Pn,k follows from Corollary 7.4, Proposition 7.7 and the argument used in Proposition 13.3 (1). We replace (µ, ν) with (Dλµ, Dλν) and compare the powers of λ, to conclude that Pn,k only depends on m1(ν),· · · , mn−k(ν). The reason why R1(µ, ν) does not appear is that a term of the form Qn(m1(ν),· · · , mp(ν))m1(µ) never appears in the moment-cumulant formulae in both c-monotone and c-free cases, except for the first order r1(µ, ν) = R1(µ, ν) = m1(µ). Example 7.9. The third cumulants for c-monotone and c-free cases are given by m1(ν)(m2(µ) − m1(µ)2) + m1(µ)3, r3(µ, ν) = m3(µ) − 2m2(µ)m1(µ) − R3(µ, ν) = m3(µ) − 2m2(µ)m1(µ) − m1(ν)(m2(µ) − m1(µ)2) + m1(µ)3. 1 2 Therefore, r3(µ, ν) = R3(µ, ν) + 1 2 m1(ν)R2(µ, ν). (7.8) 8 Infinite divisibility In this section, we define infinite divisibility for the c-monotone convolution and we characterize the infinite divisible distributions with compact supports. Definition 8.1. A pair of probability measures (µ, ν) on R is said to be (additively) c-monotone infinitely divisible if for any n ≥ 1 there exists a pair of probability measures (µn, νn) such that (µ, ν) = (µn, νn)⊲n. The above definition might become easy to understand in terms of random variables in a C∗- algebra: Definition 8.2. In a C∗-algebraic probability space (A, ϕ, ψ) equipped with two states, we say that a self-adjoint operator X has a c-monotone infinitely divisible distribution if for any n ≥ 1 there exist a C∗-algebraic probability space (An, ϕn, ψn) and identically distributed, c-monotone independent random variables X1,· · · , Xn ∈ An such that X has the same distribution as X1 + · · · + Xn with respect to the two states. 23 These definitions are the same for compactly supported probability measures, since we know a canonical realization of c-monotone independence in Section 3 and since µn, νn are compactly sup- ported whenever µ, ν are (see Lemma 8.3). In this paper, we focus on the convolution of probability measures, and hence, use the former definition. Lemma 8.3. Let (µ3, ν3) be the c-monotone convolution of (µ1, ν1) and (µ2, ν2). (1) Let (ai, ηi) and (bi, ξi) respectively denote the pairs of real numbers and finite measures appearing in the representations in (2.3) for µi and νi. Then supp η2 ⊂ supp η3 and supp ξ2 ⊂ supp ξ3. (2) If µi and νi (i = 1, 2) are compactly supported, also µ3 and ν3 are. Proof. (1) We only prove the claim for the first component since the fact for the second component is know in [31]. From a simple calculation, we obtain Hµ3(z) = aµ1 + aµ2 + z +ZR 1 + xHν2(z) x − Hν2(z) dη1(x) +ZR 1 + xz x − z dη2(x). (8.1) Applying the Stieltjes inversion formula, we have limvց0R b a ImHµ3(u + iv)du = 0 whenever [a, b] ∩ (2) If Gµi and Gνi (i = 1, 2) are analytic outside a ball, then Gµ1⊲ν2 µ2 and Gν1⊲ν2 are also analytic supp η3 = ∅, which implies that supp η2 ⊂ supp η3. outside a ball. Corollary 8.4. Let µ, ν be probability measures with compact supports. An n-th root of (µ, ν) for the c-monotone convolution is unique for any n ≥ 1. Proof. Let (µn, νn) be an n-th roof of (µ, ν), i.e., (µ, ν) = (µn, νn)⊲n. µn and νn are compactly supported from Lemma 8.3. With the power additivity of the monotone cumulants and the c- monotone cumulants, we have rk(µn, νn) = 1 k (ν) for all k ≥ 1. This implies the uniqueness. n rk(µ, ν) and rM k (µn) = 1 nrM We prove the c-monotone analogue for Theorem 13.6 of [34]. Theorem 8.5. Let µ, ν be probability measures with compact support. The following statements are equivalent. (1) (µ, ν) is c-monotone infinitely divisible. (2) There exists a compactly supported, weakly continuous c-monotone convolution semigroup {(µt, νt)}t≥0 with (µ0, ν0) = (δ0, δ0) such that (µ1, ν1) = (µ, ν). (3) Both {rn(µ, ν)}n≥2 and {rn(ν)}n≥2 are positive definite sequences. (4) There exist compactly supported probability measures µN , νN for each N such that (µN , νN )⊲N converges to (µ, ν) weakly. Proof. The implications (2) ⇒ (1) ⇒ (4) follow from Lemma 8.3. Now we prove (4) ⇒ (3). We note that For a1,· · · , an ∈ C, we have nXj,k=1 aj ¯akrk+j(µ, ν) = lim N→∞ xk+jµN (dx) aj ¯akZR ajZR xjµN (dx)(cid:12)(cid:12)(cid:12) 2 rn(ν) = lim N→∞ rn(µ, ν) = lim N→∞ xnνN (dx), xnµN (dx). NZR NZR nXj,k=1 N(cid:12)(cid:12)(cid:12) nXj=1 N = lim N→∞ ≥ 0. 24 (8.2) (8.3) (8.4) Next we prove the implication (3) ⇒ (2). We do not know a priori the existence of R > 0 such that rn(µ, ν) ≤ Rn and rn(ν) ≤ Rn. At least, however, there exist positive finite measures τ1, τ2 such that A1(z) := −r1(µ, ν) +ZR A2(z) := −r1(ν) +ZR 1 + xz x − z 1 + xz x − z τ2(dx) = − τ1(dx) = − rn(µ, ν) zn−1 , ∞Xn=1 ∞Xn=1 rn(ν) zn−1 (8.5) (8.6) in the sense of asymptotic expansion. We define two functions Ht, Ft and a weakly continuous c- monotone convolution semigroup {(µt, νt)}t≥0 using Theorem 6.1. We obtain rn(µt, νt) = trn(µ, ν) and rn(νt) = trn(ν) from the power additivity of the monotone cumulants and the c-monotone cumulants. Therefore, µ1 and ν1 have the same moments as µ and ν, respectively. Since µ and ν are compactly supported, (µ1, ν1) = (µ, ν); moreover, η1 and ξ1 are compactly supported. From Lemma 8.3 (1), ηt and ξt are compactly supported for 0 ≤ t ≤ 1. Recalling the equalities (6.11) and (6.16), we conclude that supp τ1 ⊂ supp ηδ and supp τ2 ⊂ supp ξδ for sufficiently small δ > 0. Therefore, supp τj is compact. It is immediate that µt and νt are compactly supported for all 0 ≤ t < ∞ from Lemma 8.3 (2). 9 Remarks and discussions on infinite divisibility Let D be the unit disc in the complex plane. Theorem 1.1 of [4] says that if a semigroup of analytic maps φt(z) (t ≥ 0) defined on D with φ0(z) = z is continuous in [0,∞) × D, there exists an analytic vector field as a generator of the semigroup. As a result, the semigroup also belongs to C ω([0,∞)×D). There is a different proof of Theorem 6.1 based on the above (and its generalization). This approach was used by Bercovici in [5] for multiplicative monotone convolutions (see also Franz's argument in [15]). Weak convergence of probability measures on the real line is equivalent to pointwise convergence of the reciprocal Cauchy transforms as shown in [30]. On the unit circle, the weak convergence is equivalent to pointwise convergence of ηµ(z) := 1 − z z ), z ∈ D. This fact can be proved in the same idea as [30]. Therefore, a given weakly continuous convolution semigroup µt on R (resp. on T) with µ0 = δ0 (resp. µ0 = δ1) has a continuous Hµt(z) (resp. ηµt(z)) for each z. The following fact is needed to apply Theorem 1.1 of [4]. Proposition 9.1. Let {φt}t∈I be a family of analytic maps on D parametrized by t ∈ I, where I is an interval. We assume that the map t 7→ φt(z) is continuous for each z ∈ D. Then the map φ : I × D → D defined by φ(t, z) = φt(z) is continuous. Gµ( 1 The proof is not difficult; we will give a proof in [20]. Since D is analytically isomorphic to C+, we can apply this to both additive and multiplicative convolutions. In the case of the c-monotone convolution, the functional relation for reciprocal Cauchy transforms is not only a composition semigroup: Fs+t = Fs ◦ Ft and Ht+s = Ht◦ Fs − Fs + Hs. We used Muraki's method to prove Theorem 6.1, but it is also possible to use a method similar to Theorem 1.1 in [4] for (Ht, Ft). This method is useful especially for the multiplicative convolution and we will show it in [20]. We proved in Theorem 8.5 the equivalence between infinite divisibility and the embedding of a measure into a convolution semigroup for compactly supported probability measures. We also proved the positive definiteness of cumulants. This result is new even in the monotone case. 25 10 Convolutions arising from the conditionally monotone convolution The c-monotone convolution unifies the monotone and Boolean convolutions: (µ, µ) ⊲ (ν, ν) = (µ ⊲ ν, µ ⊲ ν), (µ, δ0) ⊲ (ν, δ0) = (µ ⊎ ν, δ0). (10.1) (10.2) These are analogous to (3.5) and (3.6). We generalize these relations in this section in analogy with the c-free case. Therefore, we first explain the c-free case. For a map T : P → P, a new convolution ⊞T can be defined by (µ ⊞T ν, T µ ⊞ T ν) = (µ, T µ) ⊞ (ν, T ν). A problem of Bozejko is to find all maps T : P → P such that T (µ ⊞T ν) = T µ ⊞ T ν. (10.3) (10.4) This relation exactly says that the graph {(µ, T µ); µ ∈ P} is closed under the c-free convolution. Once such T is found, the new convolution is associative and commutative. Moreover, this generalizes the free and Boolean convolutions. Many maps satisfying (10.4) were found in [10, 11, 23, 24, 35, 36]. Motivated by these works, we consider the following type of convolution: (µ ⊲T ν, T ν) = (µ, δ0) ⊞ (ν, T ν). This relation is parallel to (10.3) in terms of the c-monotone convolution: (µ ⊲T ν, T µ ⊲ T ν) = (µ, T µ) ⊲ (ν, T ν). (10.5) (10.6) Clearly, this convolution includes Boolean and monotone convolutions if we take T as T µ = δ0 for all µ and T = Id, respectively. Now we characterize the associativity of the convolution ⊲T . Proposition 10.1. (1) ⊲T is characterized by the equality (2) The convolution ⊲T becomes associative if and only if Hµ ⊲T ν = Hµ ◦ HT ν + Hν − HT ν. T (µ ⊲T ν) = T µ ⊲ T ν for all µ and ν. (10.7) (10.8) Remark 10.2. (1) In many cases T is only defined in a subset of P such as Pm. In such a case, the above Proposition holds if the subset is closed under the convolution ⊲T . Henceforth, we often state results about the convolution ⊲T only for T : P → P if such a generalization of the domain of T is trivial. (2) It seems to be not known whether the condition (10.4) is a necessary condition for the associativity of ⊞T . Proof. (1) The proof is easy. (2) First we calculate H(µ ⊲T ν) ⊲T λ as H(µ ⊲T ν) ⊲T λ = Hµ ⊲T ν ◦ HT λ + Hλ − HT λ = Hµ ◦ HT ν ◦ HT λ + Hν ◦ HT λ − HT ν ◦ HT λ + Hλ − HT λ. (10.9) 26 Hµ ⊲T (ν ⊲T λ) is calculated as follows. Hµ ⊲T (ν ⊲T λ) = Hµ ◦ HT (ν ⊲T λ) + Hν ⊲T λ − HT (ν ⊲T λ) = Hµ ◦ HT (ν ⊲T λ) + Hν ◦ HT λ + Hλ − HT λ − HT (ν ⊲T λ). (10.10) Then the associativity of the convolution implies that Hµ◦HT ν◦HT λ−Hµ◦HT (ν ⊲T λ) only depends on ν and λ. If HT ν ◦ HT λ were not equal to HT (ν ⊲T λ) for some ν and λ, then there would exist w ∈ C\R such that HT ν ◦ HT λ(w) 6= HT (ν ⊲T λ)(w). In this case, Hµ ◦ HT ν ◦ HT λ(w)− Hµ ◦ HT (ν ⊲T λ)(w) clearly depends on µ, which is a contradiction. Therefore, we conclude that for all ν, λ ∈ P. Conversely, if (10.11) holds, it is not difficult to see that the convolution is associative. HT ν ◦ HT λ = HT (ν ⊲T λ) (10.11) We show the additivity of mean and variance; this will be used in the proof of Theorem 11.2. Proposition 10.3. Let T : P → P be an arbitrary map. Then we have the following properties. (1) P 2 is closed under the convolution ⊲T . (2) m(µ) and σ2(µ) are additive with respect to the convolution ⊲T considered in P 2: m(µ ⊲T ν) = m(µ) + m(ν), σ2(µ ⊲T ν) = σ2(µ) + σ2(ν). Proof. We use the notation Then we have Hµ(z) = −m(µ) + z +Z 1 x − z dρµ(x). Hµ ⊲T ν(z) = Hµ ◦ HT ν(z) + Hν(z) − HT ν(z) 1 = −m(µ) + HT ν(z) +ZR = −m(µ) + Hν(z) +ZR = −m(µ) − m(ν) + z +ZR x − HT ν(z) 1 x − z dρµ(x) + Hν(z) − HT ν(z) d(cid:16)ZR (T ν)y(x)dρµ(y)(cid:17) d(cid:16)ZR (T ν)y(x)dρµ(y) + ρν(x)(cid:17), 1 x − z (10.12) (10.13) (10.14) (10.15) where λy ∈ P is defined by Hλy = Hλ − y for λ ∈ P. With Lemma 2.2, P 2 is closed under the convolution; moreover, mean and variance are additive. Thus, P 2 is closed under the convolution ⊲T for arbitrary map T . On the contrary, P+ and Psym are not closed in general under the convolution ⊲T . We show the necessary and sufficient conditions for P+ and Psym. Proposition 10.4. (1) The following two conditions are equivalent. (1a) T (P+) ⊂ P+, (1b) µ ⊲T ν ∈ P+ for all µ, ν ∈ P+. (2) The following two conditions are equivalent. 27 (2a) T (Psym) ⊂ Psym, (2b) µ ⊲T ν ∈ Psym for all µ, ν ∈ Psym. Proof. (1) We assume (1a). From Lemma 2.4, Hλ is analytic in C\[0,∞) and Hλ < 0 in (−∞, 0) for λ = µ, ν, T ν. Then the composition Hµ ◦ HT ν is analytic in C\[0,∞), and hence, Hµ ⊲T ν is also analytic in the same region. This implies the condition (1) in Lemma 2.4. We use the notation (10.16) (10.17) Hµ(z) = bµ + z +ZR 1 + xz x − z dηµ(x) and similarly for Hν. We put gµ(z) :=RR 1+xz x−z dηµ(x). Proposition 10.1 implies that Hµ ⊲T ν(z) = Hµ ◦ HT ν(z) + Hν(z) − HT ν(z) = bµ + gµ(HT ν(z)) + Hν(z). Since gµ and HT ν are non-decreasing, gµ ◦ HT ν is also non-decreasing. Then bµ + gµ ◦ HT ν(−0) ≤ bµ + gµ(−0) = Hµ(−0) ≤ 0. Therefore, we obtain Hµ ⊲T ν(−0) ≤ 0. Next we assume (1b). We shall prove the fact by reductio ad absurdum; we assume that there exists ν ∈ P+ such that T ν /∈ P+. In the notation (10.15), we have Hµ ⊲T ν(z) = Hµ ◦ HT ν(z) + Hν(z) − HT ν(z) = −m(µ) + Hν(z) +ZR 1 x − z = −m(µ) + Hν(z) − Gρµ⊲T ν(z) d(cid:16)ZR (T ν)y(x)dρµ(y)(cid:17) (10.18) for all µ ∈ P 2, where we defined ρµ ⊲ T ν using the affinity of the left component of the monotone convolution. We can construct µ ∈ P+ such that a(ρµ) = 0. Therefore, we have a(ρµ ⊲T T ν) ≤ a(T ν) < 0 from Lemma 2.3 (3); this inequality means that Gρµ⊲T ν is not analytic in C\[0,∞). On the contrary, both Hµ ⊲T ν(z) and Hν(z) are analytic in C\[0,∞) by assumption; this is a contradiction (we note that Lemma 2.3 is applicable to all positive finite measures). (2) We assume (2a). Since a probability measure µ is symmetric if and only if Hµ(−z) = −Hµ(z) Conversely, we assume (2b). We take µ to be the arcsine law with mean 0 and variance 1. Clearly, for z ∈ C\R, the proof is not difficult. µ ∈ Psym. (2b) implies that Hµ ⊲T ν(−z) = −Hµ ⊲T ν(z) for all ν ∈ Psym. Using (10.7) we have pHT ν(z)2 − 2 −pHT ν(−z)2 − 2 = HT ν(z) + HT ν(−z) for z ∈ C+. After some calculations we obtain HT ν(−z) = −HT ν(z), which means T ν ∈ Psym. Remark 10.5. The above property unifies the properties of Boolean and monotone convolutions [18]. In the cases of Boolean and monotone convolutions, we can moreover prove that ν (cid:3)n ∈ P+ implies ν ∈ P+, where (cid:3) is the Boolean or monotone convolution. This property was used to characterize the subordinators in terms of the L´evy-Khintchine representations [18]. (10.19) Sometimes the limit distribution of Poisson's law of small numbers concerning a deformed con- volution does not belong to P+ [23]. We can prove a sufficient condition for this problem. We also show a condition for the central limit measure to be contained in Psym. Corollary 10.6. (1) We assume that T (P+) ⊂ P+. If Poisson's law of small numbers holds, then the limit distribution belongs to P+. (2) We assume that T (Psym) ⊂ Psym. If the central limit theorem holds, then the limit distribution belongs to Psym. Remark 10.7. Poisson's law of small numbers and the central limit theorem mean the statements as in Theorem 15.6. Proof. If we take µ(N ) := (1− λ using Lemma 2.7. For the central limit theorem, we take µ := 1 and the limit distribution also belongs to Psym again from Lemma 2.7. N δ1 ∈ P+, then the limit distribution limN→∞(µ(N )) ⊲T N ∈ P+ by µ) ⊲T N ∈ Psym 2 (δ−1 + δ1). Then (D 1√N N )δ0+ λ 28 11 Transformations Vt,u,a We introduce a family of transformations denoted by Vt,u,a and prove that the transformations satisfy the condition (10.8). Moreover, this family unifies the generalized t-transformation [24] and the Va- transformation [23]. Let f : Q → R, where Q is a subset of P. Typically Q is chosen to be P 2, Pm or Pc. Motivated by the generalized t-transformation in [24] and Va-transformation in [23], we look for a transform Vt,f of the form (11.1) If f (µ) = (t−u)m(µ), this is the same as the generalized t-transformation. If t = 1 and f (µ) = aσ2(µ), this is the same as Va-transformation. µ 7→ Vt,f µ, HVt,f µ(z) = tHµ(z) + (1 − t)z + f (µ). Lemma 11.1. Assume that Q is closed under the convolution ⊲Vt,f . Vt,f satisfies the associativity condition (10.8) considered in Q if and only if f (µ ⊲Vt,f ν) = f (µ) + f (ν) for all µ, ν ∈ Q. Proof. Denote ⊲Vt,f by ⊲t,f for simplicity. Applying (10.7) we obtain HVt,f (µ⊲t,f ν)(z) = tHµ⊲t,f ν(z) + (1 − t)z + f (µ ⊲t,f ν) = tHµ ◦ HVt,f ν(z) + tHν(z) − tHVt,f ν(z) + (1 − t)z + f (µ ⊲t,f ν) = tHµ ◦ HVt,f ν(z) + tHν(z) + (1 − t)z + f (ν) − tHVt,f ν(z) + f (µ ⊲t,f ν) − f (ν) = tHµ ◦ HVt,f ν(z) + (1 − t)HVt,f ν(z) + f (µ ⊲t,f ν) − f (ν). (11.2) On the other hand, we have HVt,f µ⊲Vt,f ν(z) = HVt,f µ ◦ HVt,f ν(z) = tHµ ◦ HVt,f ν(z) + (1 − t)HVt,f ν(z) + f (µ). Therefore, the associativity condition (10.8) is equivalent to f (µ ⊲t,f ν) = f (µ) + f (ν). We define transformations Vt,u,a by letting More clearly, we define f (µ) = ft,u,a(µ) = (t − u)m(µ) + aσ2(µ). HVt,u,aµ(z) = tHµ(z) + (1 − t)z + (t − u)m(µ) + aσ2(µ). (11.3) (11.4) (11.5) We expect that higher order moments for f have nontrivial structure, but we do not treat them in this article. We use the notation ⊲t,u,a for the convolution defined by Vt,u,a. Theorem 11.2. The convolution ⊲t,u,a defined on P 2 is associative. Proof. This fact follows from Lemma 11.1 and Proposition 10.3. In order to calculate the inverse transformation of Vt,u,a, we show the following facts. Lemma 11.3. We have the following equalities. m(Vt,u,aµ) = um(µ) − aσ2(µ), σ2(Vt,u,aµ) = tσ2(µ). (11.6) (11.7) 29 Proof. A direct computation leads to HVt,u,aµ(z) = −um(µ) + aσ2(µ) + z + tZ 1 x − z dρµ(x), (11.8) from which and Lemma 2.16 the conclusion follows. Proposition 11.4. We have the following equality: In particular, we have Vt′,u′,a′Vt,u,a = Vt′t,u′u,u′a+a′t for t ≥ 0, u, a ∈ R. V −1 t,u,a, = Vt−1,u−1,− a tu for t > 0, u 6= 0, a ∈ R. (11.9) (11.10) Proposition 11.5. (1) P+ ∩P 2 is closed under the convolution ⊲t,u,a if and only if u ≥ t and a = 0. (2) Psym ∩ P 2 is closed under the convolution ⊲t,u,a if and only if a = 0. Proof. These facts are easy consequences of Proposition 10.4. Example 11.6. between the Fermi convolution and the c-free convolution: (1) Oravecz introduced the Fermi convolution • in [35]. He mentioned a relation (µ • ν, δm(µ) ⊞ δm(ν)) = (µ, δm(µ)) ⊞ (ν, δm(ν)), (11.11) where m(µ) denotes the mean of µ. We can easily extend the Fermi convolution to the convo- lution coming from the map Fu defined by Fuµ = δum(µ). Clearly V0,u,0 = Fu. An associative convolution ⊲Fu arises from Fu: (µ ⊲Fu ν, δum(µ)+um(ν)) = (µ, δum(µ)) ⊲ (ν, δum(ν)). (11.12) relation (2) The t-transformation is realized as Ut = Vt,t,0. An associative convolution ⊲t arises from the (11.13) (µ ⊲t ν,Ut(µ) ⊲ Ut(ν)) = (µ,Ut(µ)) ⊲ (ν,Ut(ν)). We note that the t-transformation interpolates the Boolean and monotone convolutions: they appear when t = 0 and t = 1, respectively. (3) The Va-transformation is equal to V1,1,a. In the following we make the meaning of the results in this section clearer. It is known that the t-transformation Ut (t > 0) satisfies the condition (10.4), so that a new convolution ⊞Ut [10] can be defined. This convolution can also be written as µ ⊞Ut ν = U1/t(Ut(µ) ⊞ Ut(ν)) (11.14) for t > 0. Apart from the context of the c-free convolution, it seems interesting to study the deformation of Boolean and tensor convolutions defined by the right hand side of (11.14), with ⊞ replaced by ⊎ and ∗, respectively. The new convolutions were studied in [11]. By definition, the deformed convolutions are associative and commutative. We can also define the same deformations in the monotone case. The results in this section show that the deformation has a natural meaning in terms of the c-monotone convolution as in the case of the free convolution (cf. (10.3) and (10.6)). The above discussion is meaningful for any T which is invertible such as some class of the t-transformation. 30 12 Deformations related to monotone infinitely divisible dis- tributions Krystek and Wojakowski have introduced a deformation connected to a ⊞-infinitely divisible distribu- tion in [23], which we explain now. For a ⊞-infinitely divisible distribution ϕ with a compact support, there corresponds a unique weakly continuous ⊞-convolution semigroup {ϕt}t≥0 with ϕ0 = δ0 and ϕ1 = ϕ. Define a transformation Φϕ t by Φϕ t µ = ϕσ2(µ)t. (12.1) This map satisfies the condition (10.4). We introduce the monotone analog of Φϕ t . Definition 12.1. For a ⊲-infinitely divisible distribution ξ ∈ P 2, let {ξt}t≥0 be the corresponding weakly continuous ⊲-convolution semigroup with ξ0 = δ0 and ξ1 = ξ. Let f : P 2 → R. We define a transformation Ξξ (12.2) f by setting Ξξ f µ := ξf (µ). f satisfies the associativity condition (10.8) in P 2 if and only if f (µ ⊲ Ξξ f ν) = Lemma 12.2. Ξξ f (µ) + f (ν) for all µ, ν ∈ P 2. Proof. This fact follows from the equality Ξξ f µ ⊲ Ξξ f ν = ξf (µ)+f (ν). Theorem 12.3. The map Ξξ t (t ≥ 0) defined by f (µ) = tσ2(µ) satisfies the condition (10.8). fu,a = V0,u,a. Proof. The fact follows from Proposition 10.3. Remark 12.4. (1) If ξ = δa, the map fs,t(µ) = −sm(µ) + tσ2(µ) (s, t ∈ R) is also possible. In this case we have Ξδ1 (2) Higher order moments may be possible for f , which we do not consider in this paper. Proposition 12.5. (1) P+ ∩P 2 is closed under the convolution ⊲ (2) We assume that ξ ∈ Pc. Then Psym ∩ Pc is closed under the convolution ⊲ ξ ∈ Psym ∩ Pc. Proof. This is an immediate consequence of Lemma 2.5, Lemma 2.6 and Proposition 10.4. if and only if ξ ∈ P+ ∩P 2. if and only if Ξξ t Ξξ t 13 Cumulants for a general convolution Only in this section, rn(µ) denote general cumulants, not only the monotone cumulants. To define cumulants for deformed convolutions in the section 14, we consider what are cumulants of a convolution product. We have clarified three axioms of cumulants in [21] for random variables; however, we need more general axioms to treat convolutions appearing in Sections 11, 12. Results in this section are quite general and will be applicable to other convolutions which do not appear in this paper. Let (cid:3) be a convolution defined on Pm. We shall treat convolutions which are not necessarily commutative for the later applications. All results in this section hold for both Pm and Pc , except for Theorem 13.9. Then we use the set Pm mainly. Definition 13.1. (1) We define recursively ν (cid:3)n := ν(cid:3)ν (cid:3)n−1 for ν ∈ Pm. (cid:3) is said to be power associative if ν (cid:3)(n+m) = ν (cid:3)n(cid:3)ν (cid:3)m for all m, n ≥ 0. Let mn(µ) be the n-th moment of µ ∈ Pm. We put the following assumptions. 31 (M1) There exists a universal polynomial Pn of 2n − 2 variables for each n ≥ 1 such that mn(µ(cid:3)ν) = mn(µ) + mn(ν) + Pn(m1(µ),· · · , mn−1(µ), m1(ν),· · · , mn−1(ν)). (13.1) (M2) The polynomial Pn contains no constants for any n ≥ 1. Remark 13.2. The condition (M2) is equal to the condition δ0(cid:3)δ0 = δ0. Let rn(µ) be a polynomial of {mk(µ)}k≥1 for any n ≥ 1. We consider the following properties. (C1) Power additivity: for any n, N ≥ 1, rn(µ(cid:3)N ) = Nrn(µ). (C2) There exists a universal polynomial Qn of n − 1 variables such that rn(µ) = mn(µ) + Qn(m1(µ),· · · , mn−1(µ)). (13.2) (13.3) (C2') In addition to the condition (C2), the polynomial Qn never contains linear terms mk(µ), 1 ≤ k ≤ n − 1 for any n. Q1 is understood to be a constant which turns out to be 0 in Proposition 13.3. If a sequence {rn} satisfies (C2), we can write mn in terms of rn as mn(µ) = rn(µ) + Rn(r1(µ),· · · , rn−1(µ)), where Rn is a polynomial of n − 1 variables. We note that in many important examples the condition of homogeneity rn(Dλµ) = λnrn(µ) (13.4) (13.5) holds. Indeed, this condition holds for tensor, free, Boolean and monotone cumulants. Clearly (C2) and (13.5) imply (C2'). We do not assume this condition since the uniqueness of cumulants follows from only (C1) and (C2') (see Proposition 13.4). Moreover, there are examples which satisfy (C2') but do not satisfy (13.5) such as cumulants for a convolution deformed by the Va-transformation [23]. If there exists a sequence {rn} satisfying (C1) and (C2), we consider a transformation of the form rn 7→ r′n := rn + an,krk n−1Xk=1 (13.6) for real numbers an,k, 1 ≤ k ≤ n−1, 2 ≤ n < ∞. This transformation clearly preserves the properties (C1) and (C2). Moreover, we obtain the following property (1). Proposition 13.3. We assume (M1) and (M2) and assume that (cid:3) is power associative. (1) If there are two sequences {rn} and {r′n} satisfying (C1) and (C2), there exists a unique trans- formation of the form (13.6) which maps {rn} to {r′n}. (2) The polynomial Qn in (C2) never contains a constant term. Proof. (1) There exists a polynomial An of variables n − 1 for each n ≥ 1 such that r′n = rn + An(r1,· · · , rn−1) by using (13.3) and (13.4). Replacing µ by µ(cid:3)N , we obtain Nr′n = Nrn+An(Nr1,· · · , Nrn−1) for any N. This is an equality between polynomials of N, and hence, An is of the form An(r1,· · · , rn−1) = Pn−1 (2) We show the fact inductively. For n = 1, there exists b1 ∈ R such that r1 = m1 + b1. Since Pn does not contain a constant term in (13.1), we have m1(µ(cid:3)µ) = 2m1(µ), which implies 2r1(µ) − b1 = 2r1(µ) − 2b1. Therefore, b1 = 0. We assume that Qn does not contain a constant term for n ≤ k. Using a similar argument, we can prove that Qk+1 does not contain a constant term. k=1 an,krk. 32 Proposition 13.4. We assume (M1) and (M2) and assume that (cid:3) is power associative. The fol- lowing statements are equivalent: (a) There exists a sequence {rn}n≥1 satisfying (C1) and (C2); (b) There exists a sequence {rn}n≥1 satisfying (C1) and (C2'); (c) mn(µ(cid:3)N ) is a polynomial of m1(µ),· · · , mn(µ) and N for any n. Moreover, the sequence {rn} in (b) is unique and is given by rn(µ) = ∂ ∂N mn(µ(cid:3)N )(cid:12)(cid:12)(cid:12)N =0 . (13.7) Remark 13.5. We can see from (13.7) that cumulants are strongly related to a convolution semi- group {µt}t≥0 with µ0 = δ0 and to infinite divisibility. Proof. (a) ⇒ (c): if there exists a sequence {rn}n≥1 satisfying (C1) and (C2), we have mn(µ(cid:3)N ) = rn(µ(cid:3)N ) + Rn(r1(µ(cid:3)N ,· · · , rn−1(µ(cid:3)N )) = Nrn(µ) + Rn(Nr1(µ),· · · , Nrn−1(µ)) = Nmn(µ) + NQn(m1(µ),· · · , mn−1(µ)) + Rn(Nm1(µ),· · · , Nmn−1(µ) + NQn−1(m1(µ),· · · , mn−2(µ))). Therefore, mn(µ(cid:3)N ) is a polynomial of N and mk(µ). (c) ⇒ (a): by using (M1), (M2) and the assumption (c), mn(µ(cid:3)N ) has such a form as mn(µ(cid:3)N ) = Nmn(µ) + N lSl(m1(µ),· · · , mn−1(µ)) (13.8) (13.9) (13.10) for polynomials Sl and an L ∈ N. We define rn(µ) := ∂ ∂N LXl=0 mn(µ(cid:3)N )(cid:12)(cid:12)(cid:12)N =0 = mn(µ) + S1(m1(µ),· · · , mn−1(µ)). The power associativity of (cid:3) implies (C1). (C2) follows from (13.10). (a) ⇒ (b): mn +Pn−1 r′n = rn −Pn−1 k=1 an,kr′k for n ≥ 2. Then r′n do not contain linear terms mk. for a sequence {rn} satisfying (C1) and (C2), we can write rn in the form rn = k=1 bn,kmk + Tn(m1,· · · , mn−1), where Tn is a polynomial which does not contain linear terms mk, 1 ≤ k ≤ n − 1. We define a new sequence {r′n} inductively as follows: r′1 := r1, r′2 = r2 − b2,1r1, We note that Qn does not contain a constant term from Proposition 13.3 (2). If there exists a sequence {rn} satisfying (C1) and (C2'), the corresponding polynomial Rn in (13.4) also does not contain linear terms mk, 1 ≤ k ≤ n − 1 or a constant term. Therefore, the equality mn(µ(cid:3)N ) = Nrn(µ) + Rn(Nr1(µ),· · · , Nrn−1(µ)) implies that rn = ∂ Definition 13.6. Let (cid:3) be a power associative convolution defined on Pm satisfying (M1) and (M2). Then the polynomials rn satisfying (C1) and (C2') are called the cumulants for the convolution (cid:3). Cumulants are unique. ∂N mn(µ(cid:3)N )(cid:12)(cid:12)(cid:12)N =0 . Remark 13.7. This definition extends the cumulants for the tensor, free, Boolean and monotone convolutions. We can prove the existence of cumulants. 33 Theorem 13.8. We assume the conditions (M1) and (M2) for a power associative convolution (cid:3). Then cumulants of (cid:3) exist. Proof. It is sufficient to prove that mn(µ(cid:3)N ) is a polynomial of N due to Proposition 13.4. Then the proof is the same as in [21], which we omit here. We discuss when the additivity of cumulants holds. In the proof of the following theorem, we assume the convolution is defined on Pc so that moments determine a unique probability measure. Theorem 13.9. Let (cid:3) be a power associative convolution defined on Pc satisfying (M1) and (M2). Let rn be the cumulants. Then the following conditions are equivalent. (1) rn(µ(cid:3)ν) = rn(µ) + rn(ν) for all n and µ, ν ∈ Pc. (2) (cid:3) is associative and commutative, and moreover, Pn in (13.1) does not contain linear terms mk(µ) or mk(ν), 1 ≤ k ≤ n − 1. Proof. (1) ⇒ (2): the associativity and commutativity follow immediately since a probability measure with compact support is determined by the cumulants. From (M1) and (13.4) we obtain the identity Pn(m1(µ),· · · , mn−1(µ), m1(ν),· · · , mn−1(ν)) = Qn(m1(µ),· · · , mn−1(µ)) + Qn(m1(ν),· · · , mn−1(ν)) + Rn(r1(µ) + r1(ν),· · · , rn−1(µ) + rn−1(ν)). It follows from (C2') that Qn and Rn do not contain linear terms. (2) ⇒ (1): Using (M1), (C2') and (13.4) we have rn(µ(cid:3)ν) = mn(µ(cid:3)ν) + Qn(m1(µ(cid:3)ν),· · · , mn−1(µ(cid:3)ν)) = rn(µ) + rn(ν) + Pn(m1(µ),· · · , mn−1(µ), m1(ν),· · · , mn−1(ν)) + Rn(r1(µ),· · · , rn−1(µ)) + Rn(r1(ν),· · · , rn−1(ν)) + Qn(m1(µ(cid:3)ν),· · · , mn−1(µ(cid:3)ν)). Therefore, there exists a polynomial Un which does not contain linear terms such that rn(µ(cid:3)ν) = rn(µ) + rn(ν) + Un(r1(µ),· · · , rn−1(µ), r1(ν),· · · , rn−1(ν)). We replace µ and ν by µ(cid:3)N and ν (cid:3)N , respectively. The associativity and commutativity implies that rn(µ(cid:3)N (cid:3)ν (cid:3)N ) = rn((µ(cid:3)ν)(cid:3)N ) = Nrn(µ(cid:3)ν). Then Nrn(µ(cid:3)ν) = Nrn(µ)+Nrn(ν)+Un(Nr1(µ),· · · , Nrn−1(µ), Nr1(ν),· · · , Nrn−1(ν)). This can be seen as an identity between polynomials of N; therefore, we have Un = 0. Limit theorems can be formulated in terms of moments and cumulants. The proofs are easy. Theorem 13.10. Let (cid:3) be a power associative convolution defined on Pm satisfying (M1) and (M2). Let rn be the cumulants. µ)(cid:3)N . (1) (Central limit theorem) For µ ∈ Pm with m1(µ) = 0 and m2(µ) = 1, we define µN := (D 1√N Then r1(µN ) → 0, r2(µN ) → 1 and rn(µN ) → 0 as N → ∞ for any n ≥ 3. (2) (Poisson's law of small numbers) Let {µ(N )} be a sequence such that for any n ≥ 1 Nmn(µ(N )) → λ > 0 as N → ∞. We define µN := (µ(N ))(cid:3)N . Then rn(µN ) → λ as N → ∞ for any n ≥ 1. 14 Cumulants for deformed convolutions We define rT of Definition 13.6. n (µ) := rn(µ, T µ). rT n (µ) turn out to be cumulants for the convolution ⊲T in the sense 34 Proposition 14.1. We assume that there exists a polynomial Vn of n + 1 variables, which does not contain a constant term, such that mn(T µ) = Vn(m1(µ),· · · , mn+1(µ)) (14.1) for any n ≥ 1. Then the conditions (M1) and (M2) hold for the convolution ⊲T . Proof. (M1) follows from (4.1) and (14.1); (M2) follows from the fact that both Wn and Yn do not contain constant terms nor linear terms. Theorem 14.2. Let T : Pm → Pm be a map satisfying (10.8) and (14.1). Then rn(µ, T µ) satisfy the conditions (C1) and (C2') for the convolution ⊲T . Proof. (C2') follows from the definition of c-monotone cumulants and (14.1). (C1) can be proved showed as follows: rT n (µ ⊲T N ) = rn(µ ⊲T N , T (µ ⊲T N )) = rn((µ, T µ)⊲N) = Nrn(µ, T µ) = NrT n (µ). The c-free cumulants Rn(µ, T µ) satisfy the conditions (C1) and (C2') under similar conditions. Proposition 14.3. Let T : Pm → Pm be a map satisfying the condition (10.4). We assume that the n-th moment of T µ is of the form mn(T µ) = Vn(m1(µ),· · · , mn+1(µ)) (14.2) for any n ≥ 1, where Vn is a polynomial which does not contain a constant term. Then the convolution ⊞T satisfies the conditions (M1) and (M2), and Rn(µ, T µ) satisfies the conditions (C1) and (C2'). Remark 14.4. All the convolutions studied in [10, 11, 23, 24, 35] satisfy the condition (14.2). 15 Limit theorems for deformed convolutions We can apply Theorem 13.10 to the convolution ⊲T under the conditions (10.8) and (14.1). We summarize the statements combining Theorem 13.10 and Theorem 14.2. Theorem 15.1. Let T : Pm → Pm be a map which satisfies (10.8) and (14.1). (1) (Central limit theorem) Let µ be a probability measure in Pm with mean 0 and variance 1. We ) are characterized . Then mn(µN ) converges to mn(ν(T ) ), where mn(ν(T ) t define µN :=(cid:0)D 1√N by µ(cid:1) ⊲T N 1 ∂ ∂t Hν(T ) t (z) = − 1 HT ν(T ) t . (z) (15.1) (2) (Poisson's law of small numbers) Let {µ(N )}∞N =1 be a sequence of probability measures in Pm such that Nmn(µ(N )) → λ > 0 as N → ∞ for all n ≥ 1. We define µN := (µ(N )) ⊲T N . Then mn(µN ) converges to mn(p(T ) λ ) are characterized by λ ), where mn(p(T ) ∂ ∂λ Hp(T ) λ (z) = λ HT p(T ) (z) 1 − HT p(T ) λ (z) . (15.2) In this section we calculate the limit distributions for T constructed in Sections 11 and 12. If T is invertible, we can use monotone cumulants to calculate the limit distributions since µ ⊲T ν = T −1(T µ ⊲ T ν). Cumulants introduced in Section 14, however, enable us to calculate the limit distributions for even non-invertible T . In this section, we always use cumulants introduced in Section 14. 35 15.1 Transformations Vt,u,a n (µ) denote the cumulants rVt,u,a We now calculate the central limit measure for the convolution ⊲t,u,a. We only calculate the two cases a = 0 and t = 0; otherwise explicit expressions of the limit measures are difficult. For simlicity, let r(t,u,a) (µ). In this section we use two logarithms log[1] and log[2]: log[1](z) is defined by log[1](z) := logz + i arg(z), arg(z) ∈ (−π, π), z ∈ C\(−∞, 0]; log[2] is defined by logz + i arg(z), arg(z) ∈ (0, 2π), z ∈ C\[0,∞). Let √z be exp( 1 2 log[2] z) for z ∈ C\[0,∞). Then, for instance, the Cauchy transform of the normalized arcsine law becomes n 1√z2−2 for z ∈ C+. Theorem 15.2. (1) Let µ be a probability measure in Pm with mean 0 and variance 1. Then µ)⊲t,u,0N converges weakly to a Kesten distribution ν(t,0). The absolutely continuous part µN := (D 1√N 2 )x2 dx on [−√2t,√2t]. There is no singular part for t ≥ 1, but ν(t,0) contains atoms at √2t−x2 is 1 1−(1− t 2π x = ± 1√1− t µ)⊲0,0,aN (2) Let µ be a probability measure in Pm with mean 0 and variance 1. Then µN := (D 1√N converges weakly to a probability measure ν(0,a). The absolutely continuous part of ν(0,a) is given by for t < 1. 2 ν(0,a)ac =( a (log 1+ a (log 1+ a x−ax)2+π2 dx, x ∈ [−a, 0], a > 0, a x−ax)2+π2 dx, x ∈ [0,a], a < 0. (15.3) ν(0,a) contains two atoms: one in (−∞,−a) and the other in (0,∞) if a > 0; one in (−∞, 0) and the other in (a,∞) if a < 0. Remark 15.3. (1) Kesten distributions also appear in the central limit theorem of ⊞Ut [10, 11] with the parameter t replaced by 2t. (2) The limit distribution of (2) is symmetric only in the case of a = 0 where the convolution becomes a Boolean convolution (cf. Proposition 11.5). Proof. (1) Let {ν(t,0) }s≥0 be a (formal) convolution semigroup which is a solution of (15.1) for T = Vt,u,0. (The word "formal" means that the limit moments might not be deterministic. Therefore, we consider ν(t,0) ) = 0. Then s (z) for simplicity. (15.1) can be HVt,u,aν(t,a) integrated and we obtain t as a sequence of moments.) We note that m1(ν(t,0) (z) = tHν(t,a) ) = sr(t,u,0) (ν(t,0) 1 1 s s 2)z2, which implies s s (z) + (1 − t)z. We let Hs(z) denote Hν(t,0) 2Hs(z)2 + (1 − t)zHs(z) = −s + (1 − t √z2 − 2st 2)z2 − s 2) + 1 (1 − t ( 1 2 − t Gs(z) = . 2 s G1 is the Cauchy transform of a Kesten distribution (see [11]), whose support is compact. Then the weak convergence holds (see Theorem 4.5.5 of [12]). (2) Let {ν(0,a) We note that m1(ν(0,a) s Hs(z) denote Hν(0,a) (z) for simplicity. We have }s≥0 be a (formal) convolution semigroup which is a solution of (15.1) for T = V0,0,a. (z) = z + as. Let ) = s. Then HV0,0,aν(0,a) ) = 0 and σ2(ν(0,a) ) = r(0,0,a) (ν(0,a) 2 s s s s s (15.4) (15.5) Hs(z) = −Z s = z − 1 dr + z 0 1 a z + ar log[1](cid:16)1 + as z (cid:17). x−ax)2+π2 dx supported Case a > 0: the absolutely continuous part of the limit distribution is on the interval {x ∈ R : G1(x + i0) < 0} = [−a, 0]. We can show that the limit distribution contains an atom in (0,∞) and the other in (−∞,−a). (log 1+ a a 36 Case a < 0: the absolutely continuous part is We can show that the limit distribution contains an atom in (a,∞) and the other in (−∞, 0). distribution is 1 We note that the case a = 0 corresponds to the Boolean convolution, and hence, the limit a x−ax)2+π2 dx supported on the interval [0,a]. (log 1+ a 2 (δ−1 + δ1). We calculate the limit distribution for Poisson's law of small numbers. We consider only the case T = V0,u,a; otherwise, the explicit form is difficult to obtain. Theorem 15.4. Let {µ(N )}∞N =1 be a sequence of probability measures in Pm such that Nmn(µ(N )) → λ > 0 as N → ∞ for all n ≥ 1. Then µN := (µ(N ))⊲0,u,aN converges weakly to a compactly supported distribution p(u,a) . The absolutely continuous part of p(u,a) is given by λ dx, x ∈ [1 − (a − u)λ, 1], a > u, dx, x ∈ [1, 1 + (u − a)λ], a < u. +π2 +π2 (15.6) p(u,a) λ λ ac = a−u (cid:0) log 1+ (a−u)λ (cid:0) log 1+ (a−u)λ x−1 −(a−u)(x−λ)(cid:1)2 x−1 −(a−u)(x−λ)(cid:1)2 a−u p(u,a) λ (−∞, 1) and the other in (1 + (u − a)λ,∞) for a < u. Remark 15.5. One can see that p(u,a) Proof. We note that m1(p(u,a) tial equation ) = r(0,u,a) λ λ λ 1 is in P+ if and only if u ≥ a (cf. Proposition 11.5). (p(u,a) ) = λ and σ2(p(u,a) ) = λ. Then we obtain the differen- ∂ ∂λ Hp(u,a) λ (z) = −1 − z − 1 + (a − u)λ λ 1 . (15.7) contains two atoms: one in (−∞, 1 − (a − u)λ) and the other in (1,∞) for a > u; one in The remaining arguments are similar to Theorem 15.2 and we omit the proof. 15.2 Deformations related to ⊲-infinitely divisible distributions For a compactly supported ⊲-infinitely divisible distribution ξ, let {ξt}t≥0 be the corresponding weakly continuous ⊲-convolution semigroup with ξ0 = δ0 and ξ1 = ξ. Then ξt is compactly supported for every t > 0 [31]. Let {ν[ξ,t] s }s≥0 be the (formal) convolution semigroups defined by (15.1) and (15.2), respectively. Let r[ξ,t] ) = s we obtain t (ν[ξ,t] Ξξ ) = ξst. Similarly, we obtain Ξξ λ ) = ξλt. Therefore, (15.1) and (15.2) become n (µ). Since r[ξ,t] ) = σ2(ν[ξ,t] (ν[ξ,t] 2 s s s t s }s≥0 and {p[ξ,t] n (µ) denote rΞξ t (p[ξ,t] ∂ ∂s ∂ ∂λ Hν[ξ,t] Hp[ξ,t] (z) = − (z) = λ s 1 Hξst(z) Hξtλ(z) 1 − Hξtλ(z) , . (15.8) (15.9) These equations have been defined in the sense of formal power series. However, once equations (15.8) and (15.9) are understood to be ordinary differential equations, the solutions are analytic outside a ball for every s > 0 and λ > 0. As a result, (15.8) and (15.9) give moments of compactly supported probability measures for each s > 0 and λ. Therefore, ν[ξ,t] λ make sense as uniquely determined probability measures. Moreover, the convergence of moments in Theorem 15.1 becomes the weak convergence. We summarize the above arguments. Let ⊲[ξ,t] denote ⊲ Theorem 15.6. Let ξ be a ⊲-infinitely divisible distribution in Pc. (1) (Central limit theorem) Let µ be a probability measure in Pm with mean 0 and variance 1. Then µN :=(cid:0)D 1√N (2) (Poisson's law of small numbers) Let {µ(N )}∞N =1 be a sequence of probability measures in Pm such that Nmn(µ(N )) → λ > 0 as N → ∞ for all n ≥ 1. Then µN := (µ(N ))⊲[ξ,t]N converges to p[ξ,t] weakly. µ(cid:1)⊲[ξ,t]N converges to ν[ξ,t] 1 weakly. and p[ξ,t] Ξξ t λ . s 37 We calculate the limit distributions explicitly when ξ is the normalized arcsine law. is the Kesten distribution ν(t,0). Theorem 15.7. Let η be the normalized arcsine law. (1) The limit distribution ν[η,t] (2) The absolutely continuous part of p[η,t] is supported on [−√2λt,√2λt]∪[1,√2λt + 1]. The singular part consists of atoms: an atom exists in (√2λt + 1,∞); another exists in (√2λt, 1) if 0 < t < 1 t )√2λt − 1 t − 1)√2λt − and (1 − 1 t log(√2λt + 1) > 0. t log(1 − √2λt) − λ < 0; the other exists in (−∞,−√2λt) if ( 1 2λ λ 1 1 Remark 15.8. It is remarkable that the limit distribution in (1) also appears in Theorem 10 of [23] with the parameter t replaced by 2t. Proof. (1) The differential equation (15.8) becomes ∂ ∂s 1 which implies Hν[η,t] distribution. (2) The differential equation (15.9) becomes (z) = (1 − 1 t )z + 1 t s s , (z) = − √z2 − 2ts Hν[η,t] √z2 − 2ts. Therefore, the limit distribution is the Kesten (15.10) ∂ ∂λ Hp[ξ,t] λ (z) = √z2 − 2λt 1 − √z2 − 2λt . (15.11) We can solve this and obtain λ 1 Hp[ξ,t] √z2 − 2λt + (z) =(cid:16)1 − t(cid:17)z + (z) > 0 if and only if Rez ∈ (−√2λt,√2λt) ∪ [1,√2λt + 1]. We re- (x) is strictly increasing in the intervals (−∞,−√2λt), (√2λt, 1) and (√2λt + 1,∞). log[1](cid:16)√z2 − 2λt − 1 One can see that limImzց0 Hp[ξ,t] mark that Hp[ξ,t] Then it is not difficult to show the existence of atoms. (cid:17) − λ. z − 1 (15.12) 1 t 1 t λ λ Acknowledgements This paper owes much to the joint work with Mr. Hayato Saigo on cumulants. The author would like to thank Mr. Hayato Saigo for many discussions about quantum probability, independence, umbral calculus and in particular, cumulants. He is grateful to Professor Izumi Ojima for reading the manuscript, suggesting improvements of many sentences and discussions about independence. He thanks Professor Marek Bozejko for guiding him to the notion of conditionally free independence and an important reference [15]. He also thanks Professor Uwe Franz for fruitful discussions and for giving a seminar on the categorical treatment of independence during the visit to Kyoto. He also thanks Professor Shogo Tanimura, Mr. Ryo Harada, Mr. Hiroshi Ando and Mr. Kazuya Okamura for their comments and encouragement. This work was supported by Japan Society for the Promotion of Science, KAKENHI 21-5106. The author also thanks the support by Global COE Program at Kyoto University. References [1] L. Accardi, R. Lenczewski and R. Sa lapata, Decompositions of the free product of graghs, Infin. Dim. Anal. Quantum Probab. Rel. Topics 10, no. 3 (2007), 303 -- 334. 38 [2] N. I. Akhiezer, The Classical Moment Problem (English transl.), Oliver and Boyd, 1965. [3] A. Ben Ghorbal and M. Schurmann, Non-commutative notions of stochastic independence, Math. Proc. Comb. Phil. Soc. 133 (2002), 531 -- 561. [4] E. Berkson and H. Porta, Semigroups of analytic functions and composition operators, Michigan Math. J. 25 (1978), 101 -- 115. [5] H. Bercovici, Multiplicative monotonic convolution, Illinois J. Math. 49, no. 3 (2005), 929 -- 951. [6] H. Bercovici, On Boolean convolutions, Operator Theory 20, 7 -- 13, Theta. Ser. Adv. Math. 6, Theta, Bucharest, 2006. [7] H. Bercovici and D. Voiculescu, Free convolution of measures with unbounded support, Indiana Univ. Math. J. 42, No. 3 (1993), 733 -- 773. [8] M. Bozejko, M. Leinert and R. Speicher, Convolution and limit theorems for conditionally free random variables, Pac. J. Math. 175 (1996), no. 2, 357 -- 388. [9] M. Bozejko and R. Speicher, ψ-independent and symmetrized white noises, Quantum Proba- bility and Related Topics (L. Accardi, ed.), World Scientific, Singapore, VI (1991), 219 -- 236. [10] M. Bozejko and J. Wysocza´nski, New examples of convolutions and non-commutative central limit theorems, Banach Center Publ., 43 (1998), 95 -- 103. [11] M. Bozejko and J. Wysocza´nski, Remarks on t-transformations of measures and convolutions, Ann. I. H. Poincar´e-PR 37 (2001), 737 -- 761. [12] K. L. Chung, A Course in Probability Theory, Harcourt, Brace & World, Inc., 1968. [13] U. Franz, Monotone independence is associative, Infin. Dim. Anal. Quantum Probab. Rel. Topics 4, no. 3 (2001), 401 -- 407. [14] U. Franz, L´evy processes on quantum groups and dual groups, in Quantum independent incre- ment processes II, Lecture Notes in Math., vol. 1866, Springer-Verlag, 2006. [15] U. Franz, Multiplicative monotone convolutions, Banach Center Publ., 73 (2006), 153 -- 166. [16] U. Franz, Boolean convolution of probability measures on the unit circle, Analyse et proba- bilit´es, S´eminaires et Congr`es 16 (2009), 83 -- 93. [17] U. Franz, Monotone and boolean convolutions for non-compactly supported probability mea- sures, Indiana Univ. Math. J. 58, no. 3 (2009), 1151 -- 1186. [18] T. Hasebe, Monotone convolution semigroups, Studia Math. 200 (2010), 175 -- 199. arXiv:1002.3430v2. [19] T. Hasebe, Monotone convolution and monotone infinite divisibility from complex analytic viewpoint, Infin. Dim. Anal. Quantum Probab. Rel. Topics 13, No. 1 (2010), 111 -- 131. arXiv:1002.3430v2. [20] T. Hasebe, Conditionally monotone independence II: Multiplicative convolutions and infinite divisibility, Complex Analysis and Operator Theory, to appear. arXiv:0910.1319v3. [21] T. Hasebe and H. Saigo, The monotone cumulants, to appear in Ann. Inst. Henri Poincar´e Probab. Stat. arXiv:0907.4896v3. 39 [22] A. D. Krystek, Infinite divisibility for the conditionally free convolution, Infin. Dim. Anal. Quantum Probab. Rel. Topics 10, no. 4 (2007), 499 -- 522. [23] A. D. Krystek and L. J. Wojakowski, Associative convolutions arising from conditionally free convolution, Infin. Dim. Anal. Quantum Probab. Rel. Topics 8, no. 3 (2005), 515 -- 545. [24] A. D. Krystek and H. Yoshida, Generalized t-transformations of probability measures and deformed convolution, Probab. Math. Stat. 24 (2004), 97 -- 119. [25] F. Lehner, Cumulants in noncommutative probability theory I, Math. Z. 248 (2004), 67 -- 100. [26] R. Lenczewski, Decompositions of the additive free convolution, J. Funct. Anal. 246 (2007), 330 -- 365. [27] R. Lenczewski, Operators related to subordination for free multiplicative convolutions, Indiana Univ. Math. J. 57, no. 3 (2008), 1055 -- 1103. [28] R. Lenczewski and R. Sa lapata, Discrete interpolation between monotone probability and free probability, Infin. Dim. Anal. Quantum Probab. Rel. Topics 9, no. 1 (2006), 77 -- 106. [29] R. Lenczewski and R. Sa lapata, Noncommutative Brownian motions associated with Kesten distributions and related Poisson processes, Infin. Dim. Anal. Quantum Probab. Rel. Topics 11, no. 3 (2008), 351 -- 375. [30] H. Maassen, Addition of freely independent random variables, J. Funct. Anal. 106 (1992), 409 -- 438. [31] N. Muraki, Monotonic convolution and monotonic L´evy-Hincin formula, preprint, 2000. [32] N. Muraki, The five independences as quasi-universal products, Infin. Dim. Anal. Quantum Probab. Rel. Topics 5, no. 1 (2002), 113 -- 134. [33] N. Muraki, The five independences as natural products, Infin. Dim. Anal. Quantum Probab. Rel. Topics 6, no. 3 (2003), 337 -- 371. [34] A. Nica and R. Speicher, Lectures on the Combinatorics of Free Probability, London Math. Soc. Lecture Note Series, vol. 335, Cambridge Univ. Press, 2006. [35] F. Oravecz, Fermi convolution, Infin. Dim. Anal. Quantum Probab. Rel. Topics 5, no. 2 (2002), 235 -- 242. [36] F. Oravecz, The number of pure convolutions arising from conditionally free convolution, Infin. Dim. Anal. Quantum Probab. Rel. Topics 8, no. 3 (2005), 327 -- 355. [37] K. Sato, L´evy Processes and Infinitely Divisible Distributions, Cambridge University Press, Cambridge, 1999. [38] H. Saigo, A simple proof for monotone CLT, Infin. Dim. Anal. Quantum Probab. Rel. Topics 13, no. 2 (2010), 339 -- 343. [39] A. N. Shiryayev, Probability, Springer-Verlag, New York, 1984. [40] R. Speicher, Multiplicative functions on the lattice of non-crossing partitions and free convo- lution, Math. Ann. 298 (1994), 611 -- 628. [41] R. Speicher, On universal products, in Free Probability Theory, papers from a Workshop on Random Matrices and Operator Algebra Free Products, Toronto, Canada 1995, ed. D. V. Voiculescu, Fields Inst. Commun. 12 (Amer. Math. Soc., 1997), 257 -- 266. 40 [42] R. Speicher and R. Woroudi, Boolean convolution, in Free Probability Theory, papers from a Workshop on Random Matrices and Operator Algebra Free Products, Toronto, Canada 1995, ed. D. V. Voiculescu, Fields Inst. Commun. 12 (Amer. Math. Soc., 1997), 267 -- 280. [43] D. Voiculescu, Symmetries of some reduced free product algebras, Operator algebras and their connections with topology and ergodic theory, Lect. Notes in Math. 1132, Springer (1985), 556 -- 588. [44] D. Voiculescu, Addition of certain non-commutative random variables, J. Funct. Anal. 66 (1986), 323 -- 346. 41
1110.6227	1	1110	2011-10-28T01:06:58	Noncommutative Solenoids	[ "math.OA", "math.FA", "math.KT" ]	A noncommutative solenoid is the C-algebra $C^\ast(\Q_N^2,\sigma)$ where $\Q_N$ is the group of the $N$-adic rationals twisted and $\sigma$ is a multiplier of $\Q_N^2$. In this paper, we use techniques from noncommutative topology to classify these C-algebras up to -isomorphism in terms of the multipliers of $\Q_N^2$. We also establish a necessary and sufficient condition for simplicity of noncommutative solenoids, compute their K-theory and show that the $K_0$ groups of noncommutative solenoids are given by the extensions of $\Z$ by $\Q_N$. We give a concrete description of non-simple noncommutative solenoids as bundle of matrices over solenoid groups, and we show that irrational noncommutative solenoids are real rank zero AT C-algebras.	math.OA	math	NONCOMMUTATIVE SOLENOIDS FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER ABSTRACT. A noncommutative solenoid is the C-algebra C ∗(Q2 N , σ) where QN is the group of the N -adic rationals twisted and σ is a multiplier of Q2 N . In this paper, we use techniques from noncommutative topology to classify these C- algebras up to -isomorphism in terms of the multipliers of Q2 N . We also establish a necessary and sufficient condition for simplicity of noncommutative solenoids, compute their K-theory and show that the K0 groups of noncommutative solenoids are given by the extensions of Z by QN . We give a concrete description of non- simple noncommutative solenoids as bundle of matrices over solenoid groups, and we show that irrational noncommutative solenoids are real rank zero AT C- algebras. 1. INTRODUCTION Since the early 1960's, the specific form of transformation group C ∗-algebras given by the action of Z on the circle generated through a rotation that was an irrational multiple of 2π has sparked interest in the classification problem for C ∗- algebras in particular and the theory of C ∗-algebras in general. When first intro- duced by Effros and Hahn in [9], it was thought that these C ∗-algebras had no non-trivial projections. This was shown not to be the case by M. Rieffel in the late 1970's [19], when he constructed a whole family of projections in these C ∗- algebras, and these projections played a key role one of Pimsner's and Voiculescu's methods of classifying these C ∗-algebras up to ∗-isomorphism, achieved in 1980 ([18]) by means of K-theory. Since then these C ∗-algebras were placed into the wider class of twisted Zn-algebras by M. Rieffel in the mid 1980's ([20]) and from this point of view were relabeled as non-commutative tori. The Zn-analogs have played a key role in the non-commutative geometry of A. Connes [3], and the class of C ∗-algebras has been widened to include twisted C ∗-algebras associated to ar- bitrary compactly generated locally compact Abelian groups [8]. However, up to this point, the study of twisted group C ∗-algebras associated to Abelian groups that are not compactly generated has been left somewhat untouched. There are a variety of reasons for this lack of study, perhaps the foremost be- ing that Abelian groups that cannot be written as products of Lie groups Rn and finitely generated Abelian groups are much more complicated and best under- stood by algebraists; furthermore, the study of extensions of such groups can touch on logical conundrums. One could also make the related point that such groups require more technical algebraic expertise and are of less overall interest in applications than their compactly generated counterparts. On the other hand, Date: May 11, 2018. 2000 Mathematics Subject Classification. Primary: 46L05, 46L80 Secondary: 46L35. Key words and phrases. Twisted group C-algebras, solenoids, N -adic rationals, N -adic integers, rotation C-algebras, K-theory, -isomorphisms. 1 2 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER it can also be said that discrete Abelian groups that are not finitely generated have begun to appear more frequently in the literature, including in algebra in the study of the two-relation Baumslag-Solitar groups, where they appear as nor- mal Abelian subgroups, in the study of wavelets, where these groups and their duals, the solenoids, have appeared increasingly often in the study of wavelets [6, 7, 5, 1, 2]. We thus believe it is timely to study the twisted C-algebras of the groups Q2 N where QN is the group of N -adic rational numbers for arbitrary nat- ural number N > 1 and in homage to M. Rieffel, we call such C ∗-algebras non- commutative solenoids. In this paper, we present the classification of noncommutative solenoids up to ∗- isomorphism using methods from noncommutative topology. They are interesting examples of noncommutative spaces, and in particular, they can be seen as non- commutative orbit spaces for some actions of the N -adic rationals on solenoids, some of them minimal. Thus, our classification provides a noncommutative topo- logical approach to the classification of these actions as well. Our work is a first step in the study of the topology of these new noncommutative spaces. Our clas- sification result is based on the computation of the K-theory of noncommutative solenoids. We prove that the K0 groups of noncommutative solenoids are exactly the groups given by Abelian extensions of Z by QN , which follows from a care- ful analysis of such extensions. We relate the class of noncommutative solenoids with the group Ext(QN , Z), which is isomorphic to ZN /Z where ZN is the addi- tive group of N -adic integers [12], and we make explicit the connection between N -adic integers and our classification problem. We also partition the class of non- commutative solenoids into three distinct subclasses, based upon their defining twisting bicharacter: rational periodic noncommutative solenoids, which are the nonsimple noncommutative solenoids, and the only ones of of type I, and are fully described as bundles of matrices over a solenoid group; irrational noncommuta- tive solenoids, which we show to be simple and real rank zero AT-algebras in the sense of Elliott; and last rational aperiodic noncommutative solenoids, which give very intriguing examples. We build our work from the following family of groups: Definition 1.1. Let N ∈ N with N > 1. The group of N -adic rationals is the group: (1.1) endowed with the discrete topology. QN =n p N k ∈ Q : p ∈ Z, k ∈ No An alternative description of the group QN is given as the inductive limit of the sequence: (1.2) Z z7→N z−−−−→ Z z7→N z−−−−→ Z z7→N z−−−−→ Z z7→N z−−−−→ · · · From this latter description, we obtain the following result. We denote by T the unit circle {z ∈ C : z = 1} in the field C of complex numbers. Proposition 1.2. Let N ∈ N with N > 1. The Pontryagin dual of the group QN is the N -solenoid group, given by: SN =(cid:8)(zn)n∈N ∈ TN : ∀n ∈ N zN n+1 = zn(cid:9) , NONCOMMUTATIVE SOLENOIDS 3 endowed with the induced topology from the injection SN ֒→ TN. The dual pairing between QN and SN is given by: where p N k ∈ QN and (zn)n∈N ∈ SN . N k , (zn)n∈NE = zp D p k, Proof. The Pontryagin dual of QN is given by taking the projective limit of the sequence: (1.3) · · · z7→zN −−−−→ T z7→zN −−−−→ T z7→zN −−−−→ T z7→zN −−−−→ T. using the co-functoriality of Pontryagin duality and Sequence (1.2). We check that this limit is (up to a group isomorphism) the group SN , and the pairing is easily computed. (cid:3) Using Proposition (1.2), we start this paper with the computation of the sec- ond cohomology group of Q2 N . We then compute the symmetrizer group for any skew-bicharacter of Q2 N , as it is the fundamental tool for establishing simplicity of twisted group C-algebras. The second section of this paper studies the basic structure of quantum solenoids, defined as C ∗(Q2 N ). We thus establish conditions for simplicity, and isolate the three subclasses of noncommu- tative solenoids. We then compute the K-theory of noncommutative solenoids and show that they are extensions of Z by QN . We then prove that the K0 groups of noncommutative solenoids are given exactly by all possible Abelian extensions of Z by QN . This section presents self-contained computations of the Z-valued 2-cocycles of QN corresponding to K0 groups of noncommutative solenoids and a careful analysis of Ext(QN , Z). We then compute an explicit presentation of ratio- nal noncommutative solenoids. N , σ) for σ ∈ H 2(Q2 In our last section, we classify all noncommutative solenoids in terms of their defining T-valued 2-cocycles. Our technique, inspired by the work of [21] on ratio- nal rotation C-algebras, uses noncommutative topological methods, namely our computation of the K-theory of noncommutative solenoids. We also connect the theory of Abelian extensions of Z by QN with our -isomorphism problem. Our work is a first step in the process of analyzing noncommutative solenoids. Questions abound, including queries about Rieffel-Morita equivalence of non- commutative solenoids and the structure of their category of modules, additional structure theory for aperiodic rational noncommutative solenoids, higher dimen- sional noncommutative solenoids and to what extent the Connes' noncommuta- tive geometry can be extended to these noncommutative solenoids. 2. MULTIPLIERS OF THE N -ADIC RATIONALS We first compute the second cohomology group of Q2 N . A noncommutative solenoid will mean, for us, a twisted group C-algebra of QN × QN for some N ∈ N, N > 1. We shall apply the work of Kleppner [15] to determine the group H 2(Q2 N ) for N ∈ N, N > 1. Theorem 2.1. Let N ∈ N, N > 1. We let: ΞN = {(νn) : ν0 ∈ [0, 1) ∧ (∀n ∈ N ∃k ∈ {0, . . . , N − 1} N νn+1 = νn + k)} . 4 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER The set ΞN is a group for the pointwise modulo-one addition operation. As a group, ΞN is isomorphic to SN . Let B(2)(Q2 N ) be the group of skew-symmetrized bicharacters defined by: N ) = {(x, y) ∈ Q2 B(2)(Q2 N ) is the group of bicharacters of Q2 N 7→ ϕ(x, y)ϕ(y, x)−1 : ϕ ∈ B(Q2 N )} where B(Q2 α ∈ ΞN such that, for all p1, p2, p3, p4 ∈ Z and k1, k2, k3, k4 ∈ N, we have N . Then ϕ ∈ B(2) if and only if there exists ϕ(cid:16)(cid:16) p1 N k1 , p2 N k2(cid:17) ,(cid:16) p3 N k3 , p4 N k4(cid:17)(cid:17) = exp(2iπ(α(k1+k4)p1p4 − α(k2+k3)p2p3)). Moreover, α is uniquely determined by ϕ. Proof. If α ∈ ΞN then αk ∈ [0, 1) for all k ∈ N. Indeed α0 ∈ [0, 1) and if αk ∈ [0, 1) then αk+1 = αk+j N with 0 ≤ j ≤ N − 1 so 0 ≤ αk+1 < 1, so our claim holds by induction. With this observation, it becomes straightforward to check that ΞN is a group for the operation of entry-wise addition modulo one. By definition of ΞN , the map e : ΞN 7→ SN defined by e(α)k = exp(2iπαk) for any α ∈ ΞN is a bijection, which is easily checked to be a group isomorphism. Following [15], let B be the group of bicharacters of Q2 N and denote the group B(2)(Q2 N ) simply by B(2). The motivation for this computation is that, as a group, B(2)(Q2 N ) is isomorphic N ) by [15, Theorem 7.1] since QN is discrete and countable. However, N ) in our next theorem using the to H 2(Q2 we will find a more convenient form of H 2(Q2 following computation: Let Ψ ∈ B(2). Fix ϕ ∈ B such that: Now, the dual of Q2 N × Q2 N 7−→ ϕ(x, y)ϕ(y, x)−1. Ψ : x, y ∈ Q2 N is S 2 N with pairing given in Proposition (1.2). The map: p is a character of QN , so there exists a unique ζ ∈ SN such that: N k , 0(cid:17)(cid:17) k N k ∈ QN 7−→ ϕ(cid:16)(1, 0),(cid:16) p N k , 0(cid:17)(cid:17) = ζ p N k , 0(cid:17)(cid:17) = ηp N k(cid:17)(cid:17) = χp N k(cid:17)(cid:17) = ξp ϕ(cid:16)(1, 0),(cid:16) p ϕ(cid:16)(0, 1),(cid:16) p ϕ(cid:16)(0, 1),(cid:16)0, ϕ(cid:16)(1, 0),(cid:16)0, p p k k k for all p ∈ Z, k ∈ N. Similarly, there exists η, χ, ξ ∈ SN such that for all p ∈ Z, k ∈ N we have: Using the bicharacter property of ϕ again, we arrive at: Now, since ϕ(cid:0)(cid:0) 1 ν0 = 1 such that: χp2p4 k4 ξp1p4 k4 . , k3 p4 N k3 ηp2p3 k3 ϕ(cid:16)(p1, p2) ,(cid:16) p3 N k , 0(cid:1) ,(cid:0) p N k3 , 0(cid:1)(cid:1)(N k) N k , 0(cid:19) ,(cid:16) p ϕ(cid:18)(cid:18) 1 N k4(cid:17)(cid:17) = ζ p1p3 = ϕ(cid:0)(1, 0),(cid:0) p , 0(cid:17)(cid:19) = νkζ p N k3 k+k3 , N k3(cid:1)(cid:1), there exists ν ∈ SN with , , p4 p2 N k3 N k1 γk1+k4 µk2+k3 N k4(cid:17)(cid:17) N k2(cid:17) ,(cid:16) p3 p4 (cid:21)(cid:19) . p2 (cid:3)(cid:20) βk1+k3 ρk2+k4 (cid:21)(cid:20) p3 N k2(cid:1)(cid:1)−1 p2 (cid:3)(cid:20) βk1+k3 µk1+k4 ρk2+k4 (cid:21)(cid:20) p3 p4 (cid:21)(cid:19) N k4(cid:17)(cid:17) N k2(cid:17) ,(cid:16) p3 is given by: γk2+k3 N k1 N k3 p2 p4 , , N k1 , q1 ϕ(cid:16)(cid:16) p1 exp(cid:18)2iπ(cid:2) p1 N k4(cid:1) ,(cid:0) p1 exp(cid:18)−2iπ(cid:2) p1 Ψ(cid:16)(cid:16) p1 p2 (cid:3)(cid:20) 0 (γ − µ)(k1+k4) (µ − γ)(k2+k3) 0 (cid:21)(cid:20) p3 p4 (cid:21)(cid:19) is given by: Thus, ϕ(cid:0)(cid:0) p3 N k3 , p4 is: (2.1) exp(cid:18)2iπ(cid:2) p1 after transposing the matrix multiplication as the product is a scalar. So NONCOMMUTATIVE SOLENOIDS 5 where we use the property that ζ(N k) have ν = 1. By the same method, we deduce: k+k3 = ζk3 . Since ϕ(g, 0) = 1 for any g ∈ Q2 N , we ϕ(cid:16)(cid:16) p1 N k1 , p2 N k2(cid:17) ,(cid:16) p3 N k3 , p4 N k4(cid:17)(cid:17) = ζ p1p3 k1+k3 ηp2p3 k2+k3 χp2p4 k2+k4 ξp1p4 k1+k4 . Now, by setting all but one of p1, p2, p3, p4 to zero, we see that ϕ determines (η, ζ, χ, ξ) ∈ S 4 of Q2 forward that this map is a bijection. N uniquely. Thus, we have defined an injection ι from the group of bicharacters N by setting, with the above notation: ι(ϕ) = (ζ, ξ, η, χ). It is straight- N into S 4 Thus, ϑ : ι−1 ◦ e⊗4 : Ξ4 N ) is a bijection, so there exists a unique N such that for all p1, p2, p3, p4 ∈ Z and k1, k2, k3, k4 ∈ N, the value N → B(Q2 (β, γ, µ, ρ) ∈ Ξ4 though it is not in our chosen canonical form, i.e. γ −µ may not lie in ΞN - it takes values in (−1, 1) instead of [0, 1). Let us find the unique element of Ξ4 N which is mapped by ϑ to Ψ. Observe that we can add any integer to the entries of the matrix in Expression (2.1) without changing Ψ. Let n ∈ N. Set ǫn to be 1 if γn − νn < 0, or to be 0 otherwise. Let ω1 n = (1 − ǫn) + µn − γn. We check that ω1, ω2 ∈ ΞN and that ω1 n = 1 for all n ∈ N. We can moreover write: n = ǫn + γn − µn and ω2 n + ω2 as: (2.2) N k1 Ψ(cid:16)(cid:16) p1 p2 (cid:3)(cid:20) , p2 N k2(cid:17) ,(cid:16) p3 N k3 , p4 N k4(cid:17)(cid:17) 0 (ω2)(k2+k3) (ω1)(k1+k4) 0 exp(cid:18)2iπ(cid:2) p1 p4 (cid:21)(cid:19) (cid:21)(cid:20) p3 i.e. Ψ = ϑ(0, ω1, ω2, 0). Since ω1 + ω2 is the constant sequence (1)n∈N, we have in fact constructed a bijection from ΞN onto B2(Q2 N ) as desired. The form for Ψ proposed in the Theorem is more convenient. We obtain it by n for all n ∈ N, which does not change the value of simply subtracting 1 from ω2 Expression(2.2). We thus get: , N k1 Ψ(cid:16)(cid:16) p1 p2 (cid:3)(cid:20) exp(cid:18)2iπ(cid:2) p1 p2 N k2(cid:17) ,(cid:16) p3 N k3 0 , p4 N k4(cid:17)(cid:17) = α(k1+k4) −α(k2+k3) 0 p4 (cid:21)(cid:19) . (cid:21)(cid:20) p3 6 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER (cid:3) While [15] shows that B(2)(Q2 of subtlety is that several elements of B(2)(Q2 are in general two non-cohomologous multipliers of Q2 isomorphism to two distinct but cohomologous multipliers in B(2)(Q2 N ), a point N ) may be cohomologous, i.e. there N which are mapped by this N ) is, as a group, isomorphic to H 2(Q2 Example 2.2. If N = 3, then one checks that α = (cid:0) 1 corresponds to the element (−1)n∈N in S3. Now, if φ is given by Theorem (2.1), then ϕ ∈ B(2)(Q2 3) is symmetric. Hence it is cohomologous to the trivial multiplier 1 ∈ B(2)(Q2 3 which are not cohomologous, and map, respectively, to ϕ and 1, since [15] shows that there is a bijection from H 2(Q2 3). However, there exists two multipliers σ1, σ2 of Q2 2(cid:1)n∈N ∈ Ξ3. This element 3) onto B(2)(Q2 3). N ). This is quite inconvenient, and we prefer, for this reason, the description of multipliers of Q2 N up to equivalence given by our next Theorem (2.3). Theorem 2.3. Let N ∈ N, N > 1. There exists a group isomorphism ρ : H 2(Q2 such that if σ ∈ H 2(Q2 cohomologous to: N ) → ΞN N ) and α = ρ(σ), and if f is a multiplier of class σ, then f is p2 Ψα(cid:16)(cid:16) p1 N k2(cid:17) ,(cid:16) p3 N k4(cid:17)(cid:17) = exp(2iπα(k1+k4)p1p4). N k3 N ) be the epimorphism from the group of bichar- N k1 N ) → B(2)(Q2 p4 , , Proof. Let δ : B(Q2 acters of Q2 N onto B(2)(Q2 N ) defined by: δ(ϕ) : (x, y) ∈ Q2 N 7→ ϕ(x, y)ϕ(y, x)−1 for all ϕ ∈ B(Q2 map such that δ ◦ µ is the identity on B(2)(Q2 N ). We shall define a cross-section µ : B(2)(Q2 N ). N ) → B(Q2 N ), i.e. a For ϕ ∈ B(2)(Q2 N ), by Theorem (2.1) there exists a unique α ∈ ΞN such that: , N k1 ϕ(cid:16)(cid:16) p1 p2 (cid:3)(cid:20) exp(cid:18)2iπ(cid:2) p1 p2 N k2(cid:17) ,(cid:16) p3 N k3 0 , p4 N k4(cid:17)(cid:17) = α(k1+k4) −α(k2+k3) 0 (cid:21)(cid:20) p3 p4 (cid:21)(cid:19) . Define µ(ϕ) = Ψα. We then check immediately that δ ◦ µ is the identity. Now, denote by ζ : H 2(Q2 N ) → B(2)(Q2 g are two multipliers of Q2 cohomologous. So µ(ζ(f )) is cohomologous to f as desired. N , then ζ(f ) = ζ(g) ∈ B2(Q2 N ) the isomorphism from [15]. If f and N ) if and only if f, g are (cid:3) Remark 2.4. We thus have shown that H 2(Q2 N, N > 1. However, we find the identification of H 2(Q2 in our proofs. N ) is isomorphic to SN for all N ∈ N ) with ΞN more practical The simplicity of twisted group C-algebras is related to the symmetrizer sub- group of the twisting bicharacter. We thus establish, using the notations intro- duced in Theorem (2.1), a necessary and condition for the triviality of the sym- metrizer group of multipliers of QN for N ∈ N, N > 1. As our work will show, it is in fact fruitful to invest some effort in working with a generalization of the group ΞN based upon certain sequences of prime numbers. Definition 2.5. The set of all sequences of prime numbers with finite range is de- noted by P. NONCOMMUTATIVE SOLENOIDS 7 As a matter of notation, if Λ ∈ P then its nth entry is denoted by Λn, so that Λ = (Λn)n∈N. Definition 2.6. Let Λ ∈ P. For all k ∈ N, K > 0 we define πk(Λ) asQk−1 j=0 Λj, and π0(Λ) = 1. The set {πk(Λ) : k ∈ N} is denoted by Π(Λ). Note that π defines a strictly increasing map from N into Π(Λ), whose inverse will be denoted by δ. Periodic sequences form a subset of P, and if we impose a specific ordering on the prime numbers appearing in the smallest period of such a periodic sequence, we can define a natural embedding of N \ {0, 1} in P. We shall use: Notation 2.7. Given two integers n and m, the remainder for the Euclidean divi- sion of n by m in Z is denoted by n mod m. On the other hand, given H ⊂ G and x, y ∈ G, then x ≡ y mod H means that x and y are in the same H-coset in G. Definition 2.8. Let Λ ∈ P be a periodic sequence. If T is the minimal period of Λ ∈ n=0 Λn. Conversely, if N ∈ N and N > 1, we define Λ(N ) ∈ P as the sequence (λn mod Ω(N ))n∈N where Ω(N ) is the number of primes in the decomposition of N , λ0 ≤ . . . ≤ λΩ(N )−1 are P, we define ν(Λ) to be the natural number πT −1(Λ) =QT −1 prime and N =QΩ(N )−1 A central family of objects for our work is given by: j=0 λj. Thus in particular, ν(Λ(N )) = N . Definition 2.9. Let Λ ∈ P. The group ΞΛ is defined as a set by: ΞΛ = {(αn)n∈N : ∀n ∈ N ∃k ∈ {0, . . . , Λn − 1} Λnαn+1 = αn + k}, and with the operation of pointwise addition modulo 1. The group ΞΛ(N ) is isomorphic to ΞN , as defined in Theorem (2.1). An explicit construction of an isomorphism is given by: Proposition 2.10. Let N ∈ N with N > 1. Let Ω(N ) be the minimal period of Λ(N ), i.e. the number of prime factors in the decomposition of N . The map: ω :(cid:26) ΞΛ(N ) −→ (νn)n∈N 7−→ (νnΩ(N ))n∈N ΞN is a group isomorphism. Proof. Let α ∈ ΞΛ(N ). Define ω(α)k = αkΩ(N ) for all k ∈ N. It is immediate to check that ω(α) ∈ ΞN and, thus defined, ω is a group monomorphism. We shall now prove it is also surjective. Let us denote Λ(N ) simply by Λ. Let (νn∈N) ∈ ΞN . Let ηnΩ(N ) = νn for all n ∈ N. Let n ∈ N. By definition of ΞN , there exists m ∈ {0, . . . , N − 1} such that N νn+1 = νn + m. Let r0, m0 be the remainder and quotient for the Euclidean division of m by Λ0. More generally, we construct mj+1, rj+1 as respectively the quotient and remainder of The Euclidean division of mj by Λj for j = 0, . . . , Ω(N ) − 1. Set: ηnΩ(N )+j = ΛjηnΩ(N )+j+1 − rj for all j = 0, Ω(N ) − 1. We have given two definitions of ηnΩ(N ) and need to check they give the same values: N η(n+1)Ω(N ) = Λ0 · · · ΛΩ(N )−1η(n+1)Ω(N ) = Λ0 · · · ΛΩ(N )−2(η(n+1)Ω(N )−1 + rΩ(N )−1) · · · = ηnΩ + r0 + Λ0(r1 + Λ1(r2 + · · · )) = ηnΩ(N ) + k 8 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER so our construction leads to a coherent result. Now, by construction, η ∈ ΞΛ(N ), and ω(η) = ν. Hence ω is a group isomorphism. This completes our proof. (cid:3) Remark 2.11. The group ΞΛ can be topologized as a subspace of ([0, 1]/ ∼)N where ∼ is the equivalence relation defined by x ∼ y ⇐⇒ (x = y) ∨ (x = 0 ∧ y = 1). With this topology, the natural isomorphism e is of course an homeomorphism, so that ΞΛ is isomorphic to: SΛ = {(zn)n∈N ∈ TN : ∀n ∈ N zΛn n+1 = zn} as a topological group, though we shall not need this. We are now ready to establish a necessary and sufficient condition for the sym- metrizer group of a given multiplier to be nontrivial. Theorem 2.12. Let N ∈ N, N > 1. Let α ∈ ΞN . The symmetrizer subgroup in Q2 Ψα is defined by: N for Sα =ng =(cid:16) p1 N k1 , p2 N k2(cid:17) ∈ Q2 N : Ψα(g, ·) = Ψ(·, g)o . The following assertions are equivalent: (1) The symmetrizer group Sα is non-trivial, (2) The sequence α has finite range (i.e. {αn : n ∈ N} is finite). (3) There exists j < k ∈ N such that αj = αk, (4) There exists k ∈ N such that (N k − 1)α0 ∈ Z, (5) The sequence α is periodic. (6) The group Sα is either Q2 N (which is equivalent to α = 0) or there exists a nonzero b ∈ N such that: Sα =(cid:26)(cid:18) p1b N m , p2b N n(cid:19) : p1, p2 ∈ Z, n.m ∈ N(cid:27) . Proof. Let us assume that sα is nontrivial and prove that the range of α is finite. The result is trivial if α = (0)n∈N, so we assume that there exists s ∈ N such that αs 6= 0. By definition of ΞΛ, we then have αn 6= 0 for all n ≥ s. Let Θα : (x, y) ∈ Q2 N 7→ Ψα(x, y)Ψα(y, x)−1. Now, given p1, p2, p3, p4 ∈ Z and N k1 , p2 N k3 , p4 The symmetrizer group sα is now given by: k1, k2, k3, k4 ∈ N, we have Θα(cid:0)(cid:0) p1 N k4(cid:1)(cid:1) given by: exp(cid:0)2iπ(cid:0)α(k1+k4)p1p4 − α(k2+k3)p2p3(cid:1)(cid:1) . ng =(cid:16) p1 N : Θα(g, ·) = 1o . N k4(cid:1) ∈ Q2 N k4(cid:17)(cid:17) = 1. Θα(cid:16)(cid:16) n N k2(cid:1) ,(cid:0) p3 N k2(cid:17) ∈ Q2 N k2(cid:1) ∈ Sα, so that for all(cid:0) p3 N k2(cid:17) ,(cid:16) p3 Then, by Theorem (2.3), for all p3, p4 ∈ Z and k3, k4 ∈ N: Fix(cid:0) n N we have: N k3 , p4 N k1 , m N k1 , N k1 , N k3 p2 m , p4 (2.3) α(k1+k4)np4 ≡ α(k2+k3)mp3 mod Z. Since Congruence (2.3) only depends on k1 + k4 and must be true for all k4 ∈ N, we can and shall henceforth assume that k1 ≥ s. Without loss of generality, we NONCOMMUTATIVE SOLENOIDS 9 assume n 6= 0 (if n = 0, then m 6= 0 and the following argument can be easily adapted). Denote by β the unique extension of α in ΞΛ(N ) and denote Λ(N ) simply by Λ. Congruence (2.3) implies that for all k3, k4 ∈ N: (2.4) βΩ(N )(k1+k4)np4 ≡ βΩ(N )(k2+k3)mp3 mod Z. SinceQr−1 k3, k4 ∈ N we have: l=j βj+r ≡ Λj mod Z for all j, r ∈ N, r > 0, we conclude that for any (2.5) βΩ(N )(k1)+k4 np4 ≡ βΩ(N )(k2)+k3 mp3 mod Z, or, more generally, for any l1 ≥ Ω(N )k1, we have: (2.6) βl1+k4 np4 ≡ βΩ(N )(k2)+k3 mp3 mod Z, for all k3, k4 ∈ N. We shall now modify Λ and β so that we may assume that n in Congruence (2.6) may be chosen so that n is relatively prime with N . To do so, we write n = n1Q with n1 ∈ Z relatively prime with N and the set of prime factors of Q ∈ N is a subset of the set of prime factors of N . Let k ∈ N be the smallest integer such that Q divides πkΩ(N )(Λ) and k ≥ k1. Such a natural number exists by definition of Q and Λ. Let j1 < j2 < · · · < jr ∈ N such that jr < Ω(N )k l=1 Λjl: such a choice of integers j1, . . . , jr exists by definition of k. We also note that r = Ω(Q) − 1. Let z1 < z2 < · · · < zt ∈ N be chosen so that: and Q =Qr {z1, . . . , zt, j1, . . . , jr} = {0, . . . , Ω(N )k − 1}. We now define the following permutation of N: Let Λ′ ∈ P be defined by Λ′ j = Λs(j) for all j ∈ N. By construction, Λ and Λ′ agree for indices greater or equal than Ω(N )k. Let α be the unique sequence in ΞΛ′ such that α′ kΩ(N )+j = βkΩ(N )+j for all j ∈ N. By construction, for all k3, k4 ∈ N, we have: (2.7) α′ Ω(N )k+k4 np4 ≡ βΩ(N )(k2)+k3mp3 mod Z. Yet n = n1Q and by construction, α′ Ω(N )k+k4 Q ≡ α′ Ω(N )k+k4 −rn1 mod Z. Thus, we have shows that if Sα is not trivial, then there exists Λ′ ∈ P and a supersequence α′ ∈ ΞΛ′ of (a truncated subsequence of) α, as well as n1 ∈ Z with the set of prime factors of n1 disjoint from the range of Λ′ and k, k2 ∈ N, such that for all j, j ′ ∈ N and p, q ∈ Z, we have: k+j n1p ≡ α′ k2+j ′ mq mod Z. (2.8) α′ We now set q = 0. This relation can only be satisfied if α′ k ∈ Q, in which j ∈ Q for all j ∈ Q by definition of ΞΛ′ . Since Congruence (2.8) b with for some b ∈ Z such that b n1 and equivalent to α′ implies that α′ kn ∈ Z, we write α′ b ∧ a = 1, where a ∈ {1, . . . , b − 1}. k = a Now, by definition of ΞΛ′ , there exists x ∈ {0, . . . , Λ′ k − 1} such that: α′ k+1 = α′ k + x Λ′ k = a + xb bΛ′ k . s(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) N −→ N x 7−→  Ω(N )k − l l x if x = jl if x = zl otherwise. 10 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER We now must have: α′ k+1n1 = a + bx Λ′ k n1 b ∈ Z which implies a+bx Λ′ k ∈ N since Λ′ k and n1 are relatively prime. Hence we have: α′ k+1 ∈(cid:26) 1 b , . . . , b − 1 b (cid:27) . By induction, using the same argument as above, we thus get that we must have: (2.9) {α′ k+j : j ∈ N} ⊆(cid:26) 1 b , . . . , b − 1 b (cid:27) . Hence if Sα is nontrivial, then α′ (and therefore α) must have finite range. Remark 2.13. Condition (2.9) implies that in fact, there exists b, k ∈ N such that for all n ≥ k, there exists a ∈ {1, . . . , b − 1} with a ∧ b = 1 such that α′ b . Indeed, since α′ has finite range, there exists K ∈ N such that α′ m occurs infinitely often r = a in α′ for all m > K. Let r = max{K, k} and write α′ b for some a, b ∈ N with b′ with a∧b′ = 1 and b′ b, then Condition a∧b = 1. if for any n > r, we have α′ (2.9) implies that b′α′ r occurs again for some r′ > n. Condition (2.9) then implies that b b′, so b = b′. m ∈ Z for all m > n. By assumption on r, α′ n = a n = a It is obvious that if α has finite range, then there exists j < k such that αj = αk. Let us now prove that if α takes the same value at least twice, then there exists k ∈ N such that (N k − 1)α0 ∈ Z. Thus there exist j, k ∈ N such that αj+k = αj, yet by definition of ΞN we have N kαj+k ≡ αj mod Z, so (N k − 1)αj ≡ 0 mod Z, and since (N k − 1)α0 ≡ N j(N k − 1)αj mod Z ≡ 0 mod Z , we conclude that (N k − 1)α0 ∈ Z. Let us now assume that there exists k ∈ N such that (N k − 1)α0 ∈ Z and show b for some j ∈ N and some a, b ∈ N nonzero and b for some d ∈ {1, . . . , b − 1} and: that α is periodic. If αj = a relatively prime, then αk+j = d N kαk+j = N kd b = (N k − 1)d b + d b ≡ d b mod Z, while we must have N kαk+j ≡ a tion, αk+j = αj for all j ∈ N, as desired. b mod Z, which implies d = a. Hence by induc- Let us assume that α is periodic, which of course implies α0 = a relatively prime a, b ∈ Z, or α = 0. In the former case, we simply have: b for some Ψα(cid:16)(cid:16) n N k1 , m N k2(cid:17) ,(cid:16) p2 N k3 , q2 N k4(cid:17)(cid:17) = exp(cid:18) 2iπ b ak1+k4 nq2(cid:19) where αj = aj b for aj ∈ {1, . . . , b − 1} and all j ∈ N, using Remark (2.13). The computation of Sα is now trivial. It is also immediate, of course, if α = 0. In particular, this computation shows that Sα is not trivial if α is periodic, which concludes our equivalence. (cid:3) NONCOMMUTATIVE SOLENOIDS 11 Remark 2.14. We note that if the symmetrizer group of the multiplier Ψα for α ∈ ΞN is nontrivial, then α is rational valued. The converse is false, as it is easy to construct an aperiodic α ∈ ΞN which is rational valued: for instance, given any N > 1 we can set αn = 1 N n for all n ∈ N. Then sα = {0}. Example 2.15. For an example of a periodic multiplier, one can choose N = 5 and 62 , 25 62 , 5 62 , 1 α =(cid:0) 1 62 , . . .(cid:1). The symmetrizer group is then given by 5q (cid:19) : n, m ∈ Z, p, q ∈ N(cid:27) . (cid:26)(cid:18) 62n 5p , 62m 3. THE NONCOMMUTATIVE SOLENOID C ∗-ALGEBRAS We now start the analysis of the noncommutative solenoids, defined by: Definition 3.1. Let N ∈ N with N > 1 and let α ∈ ΞN . Let Ψα be the skew bicharacter defined in Theorem (2.3). The twisted group C-algebra C ∗(Q2 N , Ψα) is called a noncommutative solenoid and is denoted by A S α . The main purpose of this and the next section is to provide a classification result for noncommutative solenoids based upon their defining multipliers. The key ingredient for this analysis is the computation of the K-theory of noncommutative solenoids, which will occupy most of this section. However, we start with a set of basic properties one can read about noncommutative solenoids from their defining multipliers. It is useful to introduce the following notations, and provide an alternative de- scription of our noncommutative solenoids. Notation 3.2. Let α ∈ ΞN for some N ∈ N, N > 1. By definition, A S universal C-algebra for the relations α is the W p1 N k1 , p2 N k2 W p3 N k3 , p4 N k4 where Wx,y are unitaries for all (x, y) ∈ Q2 N. = Ψα(cid:16)(cid:16) p1 N k1 , p2 N k2(cid:17) ,(cid:16) p3 N k3 , p4 N k4(cid:17)(cid:17) W p1 N k1 + p3 N k3 , p2 N k2 + p4 N k4 N , and p1, p2, p3, p4 ∈ Z and k1, k2, k3, k4 ∈ Proposition 3.3. Let N ∈ N, N > 1 and α ∈ ΞN . Let θα be the action of QN on SN defined by: θα p N k ((zn)n∈N) = (exp(2iπαk+np)zn)n∈N . The C-crossed-product C(SN ) ⋊θα QN is -isomorphic to A S α . Proof. The C-algebra C(SN ) of continuous functions on SN is the group C- it is generated by unitaries Up for p ∈ QN such algebra of the dual of SN , i.e. that UpUp′ = Up+p′ . Equivalently, it is the universal C-algebra generated by n+1 = un, with the natural -isomorphism ϕ extending unitaries un such that uN (cid:16)∀n ∈ N un 7→ U 1 N n(cid:17). The C-crossed-product C(SN ) ⋊θα QN is generated by a copy of C(SN ) and unitaries Vq, for q ∈ QN , such that VqunV ∗ q = θα V p1 N k1 U p2 N k2 = θα p1 N k1 (cid:16) p2 N k2(cid:17) U p2 N k2 = exp(2iπαδ(N k1 )+δ(N k2 )p1p2)(cid:16) p2 1 N q (cid:0) 1 V p1 N k1 N n(cid:1) un. Thus: N k2(cid:17) U p2 N k2 V p1 N k1 12 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER for all p1, p2 ∈ Z and k1, k2 ∈ N. Now, the following map (using Notation (3.2)): ∀p ∈ Z, k ∈ N ( U p N k V p N k 7−→ W0, p N k 7−→ W p N k ,0 can be extended into a -epimorphism using the universal property of C(SN ) ⋊θα QN . The universal property of A S by showing the inverse of this -epimorphism is a well-defined -epimorphism. (cid:3) α implies that this -morphism is a -isomorphism, Remark 3.4. Let N ∈ N, N > 1 and α ∈ ΞN . The action θ of QN on SN defined in Proposition (3.3) is minimal if and only if α is irrational-valued. However, if α has infinite range, the orbit space of θ is still a single topological point. We start our study of noncommutative solenoids by establishing when these C-algebras are simple: Theorem 3.5. Let N ∈ N with N > 1. Let α ∈ ΞN . The following statements are equivalent: α is simple, (1) The C-algebra A S (2) The set {αn : n ∈ N} is infinite, (3) For all k ∈ N with k > 0, we have (N k − 1)α0 6∈ Z, (4) Given any j, k ∈ N with j 6= k we have αj 6= αk. Proof. The symmetrizer group Sα of Ψα is trivial if and only if the asserted con- dition holds, by Theorem (2.12). Since Q2 Λ is Abelian, and since the dual of Sα is trivial, the action of Q2 α is simple by [17, Theorem 1.5]. (cid:3) Λ oncSα is free and minimal. Thus A S As our next observation, we note that noncommutative solenoids carry a trace, Λ/Q2 which will be a useful tool for their classification. Theorem 3.6. Let N ∈ N, N > 1 and α ∈ ΞN . The C-algebra A S tracial state for the dual action of S 2 tracial state of A S α . N . Moreover, if A S α has an invariant α is simple, then this is the only Proof. For any α ∈ ΞN for N ∈ N, N > 1, the group S 2 strongly continuously on A S α by setting, for all (z, w) ∈ SN and (x, y) ∈ Q2 N : N acts ergodically and (z, w) · Wx,y = hz, xi hw, yi Wx,y and extending · by universality of A S dual action of S 2 ant tracial state τ is due to [13]. Moreover, A S only for g = 0, by Theorem (3.5). If τ ′ is any tracial state on A S (using Notation (3.2)): α , using Notation (3.2). This is of course the N is compact, the existence of an invari- α(g, ·) = 1 α , we must have α is simple if and only if Ψ2 N , Ψα). Since S 2 N on C ∗(Q2 τ ′(WgWh) = Ψ2 α(g, h)τ ′(WhWg) N . Hence if A S for all g, h ∈ Q2 α is simple, we have τ (WgWh) = 0 for all g, h ∈ Q2 N , except for h ∈ {g, g−1}. So ker τ = ker τ ′ and τ (1) = 1 = τ ′(1), so τ = τ ′ as desired. (cid:3) NONCOMMUTATIVE SOLENOIDS 13 As our next observation, the C-algebras A S α (α ∈ ΞN , N ∈ N, N > 1) are inductive limit of rotation algebras. Rotation C-algebras have been extensively studied, with [19, 10] being a very incomplete list of references. We recall that given θ ∈ [0, 1), the rotation C-algebra Aθ is the universal C-algebra for the re- lation V U = exp(2iπθ)U V with U, V unitaries. It is the twisted group C-algebra C ∗(Z2, Θ) where Θ((n, m), (p, q)) = exp(iπθ(nq − mp)). The unitaries associated to (1, 0) and (0, 1) in C ∗(Z2, Θ) will be denoted by Uθ and Vθ and referred to as the canonical unitaries of Aθ. Of course, {Uθ, Vθ} is a minimal generating set of Aθ. We now have: Theorem 3.7. Let N ∈ N with N > 1 and α ∈ ΞN. For all n ∈ N, let ϕn be the unique -morphism from Aα2n into Aα2n+2 extending: (cid:26) Uα2n Vα2n 7−→ U N 7−→ V N α2n+2 α2n+2 Then: Aα0 ϕ0−−−−→ Aα2 ϕ1−−−−→ Aα4 ϕ2−−−−→ · · · converges to A S α , where Aθ is the rotation C-algebra for the rotation of angle 2iπθ. Proof. We use Notations (3.2). Consider the given sequence of irrational C-algebra. Fix k ∈ N. Define the map: υk :( Uα2k Vα2k 7→ W 1 N k ,0 7→ W0, 1 N k N k ,0 = e2iπα2k W 1 By definition of Ψα, we have W0, 1 N k By universality of Aα2k , the map υk extends to a unique -morphism, which we α . It is straightforward to check that the diagram: still denote υk, from Aα2k into A S N k ,0W0, 1 W 1 N k . ϕ0−−−−→ Aα2 ϕ1−−−−→ Aα4 Aα0 yυ0 A S α yυ1 A S α ϕ2−−−−→ · · · · · · · · · yυ2 A S α (Aα2k , ϕk)k∈N to A S commute. So by universality of the inductive limit, there is a morphism from lim −→ we conclude that A S α is in fact generated bySk∈N υk(Aα2k ), (Aα2k , ϕk)k∈N, as desired. α . Now, since A S α is in fact lim −→ (cid:3) We can use Theorem (3.7) to compute the K-theory of the C-algebras A S α for N ∈ N, α ∈ ΞN . Theorem 3.8. Let N ∈ N with N > 1, and let α ∈ ΞN . Define the subgroup Kα of Q2 N by: where (J α k )k∈N = (N kαk − α0)k∈N and by convention, Jk = 0 for k ≤ 0. We then have: Kα =(cid:26)(cid:18)z + pJ α k N k , p N k(cid:19) : z, p ∈ Z, k ∈ N(cid:27) K0(A S α ) = Kα, and K1(A S α ) = Q2 N . K0(τ ) :(cid:18)z + α , then we have: pJ α k N k , N k(cid:19) ∈ Kα 7−→ z + pαk. p Moreover, if τ is a tracial state of A S (3.1) 14 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER In particular, all tracial states of A S α lift to the same state of K0(A S α ) given by (3.1). Proof. Define jα definition: n ∈ {0, . . . , N − 1} for n ∈ N by N αn+1 = αn + jα n , so that by J α n = N kjα k . n−1Xk=0 To ease notations, we also introduce for all n ∈ N the integer rα such that N 2α2n+2 = α2n + rα for all n ∈ N. n . Thus rα 2n+1 + jα n = N jα 2n and J α As a preliminary step, we check that Kα is a group. It is a nonempty subset of Q2 of Kα. Let n = max(k, r), and m1, m2 ∈ N be given so that N km1 = N n and N rm2 = N n. We then have: N since it contains (0, 0). Now, let(cid:16)z + pJ α (3.2)(cid:18)z + N r(cid:19) =(cid:18)z − y + N k(cid:19) −(cid:18)y + k(cid:17) and(cid:16)y + qJ α k − m2qJ α r N n pJ α k N k , qJ α r N r , N k , p J α N r , q m1pJ α p q k r k n ∈ {0, . . . , N 2 − 1} k=0 N 2krα 2n =Pn−1 N r(cid:17) be elements (cid:19) . m1p − m2q N n , Now, assume k < n, so r = n. By definition, J α m1J α N n−1jα n = J α k+1 + · · · + N 2n−1jα k + N n+1jα n − (N njα k = m1J α n−1 + m1pJ α N n . In this case, Expression (3.2) becomes: (m1p − m2q)J α n n k − · · · − N n−1jα n−1 + (cid:18)z − y − jα N n , m1p − m2q N n (cid:19) k + · · · + N n−1jα k + N kjα n−1), so m1pJ α k N n = −jα n−1 so k − · · · − which lies in Kα. The computations are similar if we assume instead r < n and k = n. Thus Kα is a subgroup of Q2 N . We remark here that the sequences (jk)k∈N and (Jk)k∈N are closely related to the group of N -adic integers ZN ; we shall discuss this relationship in detail at the conclusion of the proof. We simplify our notations in this proof and denote the canonical unitaries of the rotation C-algebra Aα2k as Uk and Vk for all k ∈ N. It is well known that: K0(Aα2k ) = Z2 and K1(Aα2k ) = Z2. Moreover, K0(Aα2k ) is generated by the classes of the identity and a Rieffel pro- jection P of trace α2k, which we denote by (1, 0) and (0, 1) respectively. We also know that K1(Aα2k ) is generated by the classes of Uk and Vk, denoted respectively by (1, 0) and (0, 1). Aα2k , then it is of the form g(Uk)Vk + f (Uk) + h(Uk)V ∗ We start with a key observation. Let P be a Rieffel projection of trace α2k in k with f, g, h ∈ C(T) and k+1 + k+1 whose trace is again α2k. We recall that with our notation: α2k = RT f . Hence P is mapped by ϕk to the Rieffel projection g(U N k+1) + h(U N k+1)V N k+1)V N f (U N N 2α2k+2 = α2k + rα k , where we note that α2k+2 is the trace of the generator of K0(A2k+2). Let k ∈ N and let ϕk be the -morphism defined in Theorem (3.7). The maps K0(ϕk) and K1(ϕk) are thus completely determined, as morphisms of Z2, by the relations: K1(ϕk) :(cid:26) (1, 0) (0, 1) 7→ (N, 0) 7→ (0, N ) and K0(ϕk) :(cid:26) (1, 0) (0, 1) 7→ (1, 0) 7→ (rα 2k, N ) NONCOMMUTATIVE SOLENOIDS 15 We now use the continuity of K-theory groups to conclude: K1(A S α ) = lim −→ (cid:18) Z2 K1(ϕ0) / Z2 K1(ϕ2) / Z2 K1(ϕ2) / · · ·(cid:19) = Q2 N , and K0(A S α ) = lim K0(ϕ0) / Z2 K0(ϕ2) / Z2 K0(ϕ2) rα 0 " 1 0 N 2 # rα 1 " 1 0 N 2 # / Z2 / Z2 rα 2 " 1 0 N 2 # / · · ·(cid:19) / · · · . = lim −→ Z2 −→ (cid:18) Z2  (cid:20) 1 − J α N 2k 0 k N 2k 1 We claim that the group K0(A S the multiplication by the matrix: α ) is Kα. For k ∈ N we define υk : Z2 → K to be We now check the following diagram is commutative: 0 N 2 (cid:21)−1 n−k . (cid:21) = kYn=0(cid:20) 1 rα " 1 0 N 2 # rα 1 rα 0 " 1 0 N 2 # Z2 υ0 Kα Z2 υ1 Kα rα 2 " 1 0 N 2 # · · · · · · Z2 υ2 Kα It is now easy to check that K is indeed K0(A S α ). Let τ be a tracial state of A S α . First, we note that (1, 0) ∈ K is the image of (1, 0) ∈ Z2 for all υk, with k ∈ N. Since τ (1) = 1 in A2k for all k ∈ N, we image of (0, 1) ∈ Z2 by υk. The generator (0, 1) of K0(Aα2k ) has trace α2k, so conclude that K0(τ )(1, 0) = 1. On the other hand, the element(cid:16) J α K0(τ )(cid:16) J α N 2k(cid:17) is the N 2k , 2k 1 1 2k 1 N 2k , N 2k−1 , and since K0(τ ) is a group morphism, we get: N 2k(cid:17) = α2k for all k ∈ N. Now, since: (cid:18) J2k−1 K0(τ )(cid:18) J2k−1 N 2k−1(cid:19) =(cid:18)−jα N 2k−1(cid:19) = −jα for all k ∈ N, k > 1. In summary, K0(τ ) maps(cid:0) Jk N 2k−1 , 2k−1 + 1 2k−1 + N α2k = α2k−1 N k , 1 N k(cid:1) to αk for all k ∈ N. Using (cid:3) the morphism property of K0(τ ) again, we obtain the desired formula. The group Kα defined in Theorem (3.8) is in fact an extension of QN given by: (3.3) 0 −−−−→ Z ι−−−−→ Kα π−−−−→ QN −−−−→ 0, J2kN N 2k−1 , N N 2k−1(cid:19) / / / / / / / / / / / / / / / 16 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER where ι : z ∈ Z 7→ (z, 0) is the canonical injection and π : (cid:16)z + pJ α N k is easily checked to be a group morphism such that the above sequence is exact. The class of this extension in H 2(QN , Z) is however not in general an invariant of the -isomorphism problem for noncommutative solenoids: as we shall explain in the next section, we must consider a weaker form of equivalence for Abelian extensions to construct such an invariant. It will translate into an equivalence relation on Ext(QN , ZN ) to be detailed after Theorem (4.2). N k(cid:17) 7→ p k N k , p We now proceed to provide a description of the Z-valued 2-cocycle of QN as- sociated to Extension (3.3) and provide a different, more standard picture for Kα. Remarkably, we shall see that every element of Ext(QN , Z) is given by the K- theory of A S α for some α ∈ ΞN . As a first indication of this connection, we note that for a given α ∈ ΞN , the sequence (J α k )k∈N can be seen an element of the group ZN of N -adic integers [14]. For our purpose, we choose the following description of ZN : Definition 3.9 ([14]). Let N ∈ N, N > 1. Set: ZN =(cid:26)(Jk)k∈N : ∧(cid:26) J0 = 0, ∀k ∈ N ∃j ∈ {0, . . . , N − 1} Jk+1 = Jk + N kj (cid:27) . This set is made into a group with the following operation. If J, K ∈ ZN then J + K is the sequence (Lk)k∈N where Lk is the remainder of the Euclidean division of Jk + Kk by N k for all k ∈ N. This group is the group of N -adic integers. The connection between ΞN (or equivalently, SN ), ZN , Ext(QN , Z) and K0 groups of noncommutative solenoids is the matter of the next few theorems. We start by observing that the following is a short exact sequence: 0 −−−−→ ZN ι−−−−→ ΞN q −−−−→ T −−−−→ 0 where q : α ∈ ΞN 7→ exp(2iπα0) and ι is the natural inclusion given by: ι : (Jn)n∈N ∈ ZN 7−→(cid:18) Jn N n(cid:19)n∈N . Thus, for any element α of ΞN , the sequence (J α to α is easily checked to be the unique element in ZN such that αk = q(α) + J α all k ∈ N. k )k∈N of Theorem (3.8) associated k for We shall use the following terminology: Definition 3.10. Let N ∈ N, N > 1. The N -reduced form of q ∈ QN is (p, N k) ∈ Z × N such that q = p N k where k is the smallest element of {n ∈ N : ∃p ∈ Z q = q N n }. By standard abuse of terminology, we say that p N k is q written in its reduced form. A fraction in N -reduced form in QN may not be irreducible in Q, so this notion depends on our choice of N . Namely, even if QN = QM for N 6= M , and p N k ∈ QN is in N -reduced form, it may not be in M -reduced form. We shall however drop the prefix N when the context allows it without introducing any confusion. We now prove the following lemma: ξJ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) QN × QN (cid:0) p1 N k1 , p2 N k2(cid:1) 7−→ Z 7−→  − p1 − p2 N k1 (Jk2 − Jk1 ) N k2 (Jk1 − Jk2 ) q N r (Jk1 − Jr) if k2 > k1 if k1 > k2 if ∧(cid:26) k1 = k2 N k1 + p2 p1 N k2 = q N r NONCOMMUTATIVE SOLENOIDS 17 Lemma 3.11. Let N ∈ N, N > 1 and α ∈ ΞN . Let J = (Jk)k∈N ∈ ZN . Writing all elements of QN in their N -reduced form only, the map: is a Z-valued symmetric 2-cocycle of QN . Proof. We introduce some useful notations for this proof. We defined jk ∈ {0, . . . , N − 1} for all k ∈ N by: We also define Jk,m for all m, k ∈ N, m > k by: Jk+1 − Jk = N kjk. Jk,k = 0 ∧ Jk,m = Note that Jk−Jr N r = Jr,k for all r ≤ k by definition. With this definition, we have ξJ(cid:0) p1 N k1 , p2 −p2Jk2,k1 when k2 < k1 and qJr,k1 if k1 = k2 and p1 + p2 = N k1−rq, with q and N relatively prime, and with all fractions written in their reduced form in QN . N r−kjr. m−1Xr=k N k2(cid:1) equal to −p1Jk1,k2 if k1 < k2, to By construction, ξJ is a symmetric function. Let x, y, z ∈ QN . We wish to show that: (3.4) Let us write x = px We proceed by checking various cases. ξJ (x + y, z) + ξJ (x, y) = ξJ (y + z, x) + ξJ (y, z). N kx in its reduced form, and use similar notations for y and z. Case 3.11.1. Assume x, y, z have the same denominator N k in their reduced form, and that x + y = q N r in its reduced form, with r < k. Then by definition, ξJ (x, y) = qJr,k and ξJ (x + y, z) = −qJr,k so the left hand side of Identity (3.4) is zero. Let y + z = q′ N n in its reduced form. If, again, n < k, the right hand side of Identity If n = k then (3.4) is zero again and we have shown that Identity (3.4) holds. ξJ (y, z) = 0 by definition. Moreover, x + y + z must have denominator N k in its reduced form. Indeed, since x, y have the same denominator N k in reduced form, yet their sum does not, px + py is a multiple of N . If moreover, px + py + pz is also a multiple of N , then pz is a multiple of N , which contradicts the definition of reduced form. Hence, x + y + z has denominator N k in its reduced form and ξJ (x, y + z) = 0 by definition. Case 3.11.2. Assume now that kx > ky > kz. Then by definition: (3.5) while (3.6) ξJ (x, y) + ξJ (x + y, z) = −pyJky ,kx − pzJkz ,kx ξJ (y, z) + ξJ (y + z, x) = −pzJkz ,ky − (N ky −kz pz + py)Jky ,kx . By definition, Jkz ,ky + N ky −kz Jky ,kx = Jkz ,kx. We then easily check that the left and right hand side of Identity (3.4) which are given by Identities (3.5) and (3.6) agree. 18 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER This case also handles the situation kz > ky > kz by switching the left and right hand side of Identity (3.4). Case 3.11.3. Assume now that ky > kx > kz. Then the left hand side of Identity (3.4) is given by: ξJ (x, y) + ξJ (x + y, z) = −pxJkx,ky − pzJkz ,ky On the other hand, the right hand side becomes: ξJ (y, z) + ξJ (y + z, x) = −pzJkz ,ky − pxJkx,ky and thus Identity (3.4) is satisfied again. We also get by symmetry the case ky > kx > kz. Case 3.11.4. Assume kx > kz > ky. Then the left hand side of Identity (3.4) is: ξJ (x, y) + ξJ (x + y, z) = −pyJky ,kx − pzJkz ,kx while the right hand side is: ξJ (y, z) + ξJ (y + z, x) = −pyJky ,kz − (N kz −ky py + pz)Jkz ,kx and as in Case 1, both side agree. The last possible strict inequality kz > kx > ky is handle by symmetry again. One similarly verifies that ξJ is a cocycle for the cases ky > kx > kz, kx > kz > ky, kx = ky > kz and kx = ky > kz. (cid:3) Theorem 3.12. Let N ∈ N, N > 1 and α ∈ ΞN . Let ξα be the Z-valued 2-cocycle of QN given by ξJ α as defined in Lemma (3.11), where J α k = N kαk − α0 for all k ∈ N. Let us define the group Qα as the set Z × QN together with the operation: p2 p2 p2 p1 p1 N k1 + N k2(cid:17) for all z, y, p1, p2 ∈ Z, k1, k2 ∈ N. The map: (cid:16)z, , k Qα −→ N k1 N k2(cid:17) =(cid:16)z + y + ξα(cid:16) p1 N k1(cid:17) ⊞(cid:16)y, ω(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 7−→ (cid:16)z + pJ α (cid:0)z, p N k(cid:1) K0(τ ) : (1, 0) 7→ 1, (cid:18)0, N k2(cid:17) , N k(cid:17) . N k(cid:19) 7→ αk. Kα N k , p 1 is a group isomorphism. Thus K0(A S these groups, we have: α ) is isomorphic to Qα and, using ω to identify Proof. It is immediate that ω is a bijection. It remains to show that it is a group morphism. Let x = px N ky with px, py ∈ Z, kx, ky ∈ N. Let z, t ∈ Z. We consider three distinct cases. N kx , y = py Case 3.12.1. The easiest case is when kx = ky and px + py is not a multiple of N . Then ξα(x, y) = 0 so ⊞ reduces to the usual addition and we have: ω(z, x) + ω(t, y) = (z + x, x) + (t + y, y) = ω(z + t, x + y) = ω((z, x) ⊞ (t, y)), as needed. NONCOMMUTATIVE SOLENOIDS 19 Case 3.12.2. Now, assume kx = ky yet px + py = N k−rq for some q not divisible by N and some r ∈ N, r > 0. Then: ω(z, x) + ω(t, y) = (cid:18)z + t + = (cid:18)z + t + (px + py)J α kx N kx q N r(cid:19) . qJ α kx N r , , px + py N kx (cid:19) Now, J α (3.11). Hence: kx = J α r + N rJ α r,kx by definition, as given in Theorem (3.8) and Lemma ω(z, x) + ω(t, y) = (cid:18)z + t + qJr,kx + = (cid:16)z + t + ξα(cid:16) px N kx = ω((z, x) ⊞ (t, y)), q qJ α r N r , py N r(cid:19) N ky(cid:17) + , q N r , q N r(cid:17) Case 3.12.3. Last, assume kx 6= ky. Without loss of generality, since our groups are Abelian, we may assume kx < ky. Now: as desired. ω(z, x) + ω(y, t) = (cid:18)z + t + = z + t + = z + t + = z + t − pxJ α + pyJ α pxJ α ky kx N ky N kx pxN ky −kxJ α kx N ky , pxN ky −kx + py N ky (cid:19) + pyJ α ky N ky , pxN ky −kx(J α ky N ky − N kx J α kx,ky ! pxN ky −kx + py N ky ) + pyJ α ky N ky , pxN ky −kx + py N ky ! kx,ky +(cid:0)pxN ky −kx + py(cid:1) J α N ky ky , pxN ky −kx + py N ky ! = ω((z, x) ⊞ (t, y)), as expected. This completes the proof of that ω is an isomorphism. Now, ω(1, 0) = (1, 0) and N k(cid:1) for all k ∈ N. Using Theorem (3.8), we conclude that tracial (cid:3) ω(cid:0)0, 1 N k(cid:1) =(cid:0) 1 states lift to the given map in our theorem. N k , 1 Thus, to α ∈ ΞN , we can associate a cocycle ξα in H 2(QN , Z) such that K0(A S α ) is given by the extension of QN by Z associated with ξα. It is natural to ask how much information the class of ξα in H 2(QN , Z) contains about noncommutative solenoids. This question will be fully answered in the next section, yet we start here by showing that the map J ∈ ZN 7→ [ξJ ] ∈ Ext(QN , Z) is surjective with kernel Z, where [ξ] is the class of the extensions of QN by Z (which is Abelian for our cocycles) for the equivalence of extension relation. First, we recall: Lemma 3.13. Let N ∈ N, N > 1. Let z ∈ N. For all n ∈ N we define ι(z)n to be the remainder for the Euclidean division of z by N n in Z. Then ι(z) ∈ ZN by construction, and there exists Kz ∈ N such that ιn(z) = ιKz (z) for all n ≥ Kz. Conversely, given any J ∈ ZN which is eventually constant, we can associate the natural number ζ(J) = JK 20 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER where K is the largest natural number such that JK < JK+1. One checks easily that ζ ◦ ι is the identity on N. The map ι extends to a group monomorphism from Z to ZN . Moreover, if z < 0 then there exists Kz ∈ N such that ιn+1 − ιn = N k(N − 1) for all n ≥ Kz. Conversely, if, for some J ∈ ZN , there exists K ∈ N such that Jk+1 − Jk = (N − 1)N k for all k ≥ K, then there exists a unique z ∈ Z, z < 0 such that ι(z) = J. Proof. This is well known. (cid:3) We now compute the cohomology relation for our Z-valued cocycles given by K0 groups of noncommutative solenoids: Theorem 3.14. Let N ∈ N, N > 1. Let J = (Jk)k∈N ∈ ZN and R = (Rk)k∈N ∈ ZN . Let ι : Z → ZN be the monomorphism of Lemma (3.13). Let ξJ , ξR be the respective Z-valued 2-cocycle of QN given by Lemma (3.11). Then ξJ and ξR are cohomologous if and only if J − R ∈ ι(Z). This is equivalent to one of the following condition holding: • There exists M ∈ N such that Jn − Rn = N n−1(N − 1) for all n ≥ M , • There exists M ∈ N such that Jn − Rn = N n−1(1 − N ) for all n ≥ M , • There exists M ∈ N such that Jn = Rn for all n ≥ M . In particular, if N > 1 then there exists nontrivial cocycles of the form ξα for some α ∈ ΞN . Proof. Let σ = ξJ − ξR. For all n ∈ N, we define jn and rn as the unique integers in {0, . . . , N − 1} such that N njn = Jn+1 − Jn and N nrn = Rn+1 − Rn. Assume there exists ψ : QN → Z such that for all x, y ∈ QN , we have: σ(x, y) = ψ(x + y) − ψ(x) − ψ(y). in reduced form in QN . Hence, under this condition, we have: N k(cid:1) = 0 if p + q is not a multiple of N , with all fractions written Note that σ(cid:0) p N k , q for all k ∈ N. Hence, for all k ∈ N we have: Now: ψ(1) + (Jk − Rk) ∈ N kZ. Jk − Rk = N n(jn − rn) k−1Xn=0 Since J1 = j0 < N and if Jk < N k then Jk+1 = Jk + N kjk < N k + N k+1 − N k = N k+1, we conclude by induction that Jk < N k for all k ∈ N. Hence, ψ(1) + Jk − Rk ∈ Zk implies that either ψ(1) + Jk − Rk = 0 or ψ(1) + Jk − Rk ≥ N k. We now get: so: q N k + N k(cid:17) = ψ(cid:16) p ψ(cid:16) p N k(cid:17) + ψ(cid:16) q N k(cid:17) . −jk + rk = ψ(cid:18) 1 N k(cid:19) − N ψ(cid:18) 1 N k+1(cid:19) , N k(cid:1) jk − rk − ψ(cid:0) 1 N k+1(cid:19) ∈ Z = ψ(cid:18) 1 N NONCOMMUTATIVE SOLENOIDS 21 Case 3.14.1. Assume first that for all m ∈ N there exists k ∈ N with k > m such that ψ(1) + Jk − Rk ≥ N k. Then, for all k ∈ N such that ψ(1) + Jk − Rk ≥ N k: (3.7) ψ(1) ≥ N k − Jk + Rk = 1 + N n(−jn + rn + N − 1) for infinitely many k ∈ N. Since −jn + rn > −N , we have −jn + rn + N − 1 ≥ 0. If −jn + rn + N − 1 > 0 then the right hand side of Inequality (3.7) is unbounded as k is allowed to go to infinity, which is absurd since the left hand side is ψ(1). This implies that there exists M ∈ N such that for all k ≥ M , we have jk − rk = N − 1. Conversely, if there exists M ∈ N such that jn − rn = N − 1 for all n ≥ M , then set n=0 N n(N − 1 − jn + rn). We then have: ψ(1) = 1 +PM −1 ψ(1) + Jk − Rk = 1 + k−1Xn=0 N n(N − 1 − jn + rn) N n(jn − rn) + M −1Xn=0 M −1Xn=0 k−1Xn=M k−1Xn=0(cid:0)N k+1 − N k(cid:1) = N k + = 1 + N n(N − 1) as desired. The cases for ψ(1)+Jk −Rk ≤ N k for infinitely many k ∈ N, and ψ(1) = Rk −Jk for infinitely many k ∈ N, are proved similarly. (cid:3) Remark 3.15. Let α ∈ ΞN be given such that there exists ψ ∈ Z such that ψ + Jk = N k for all n ∈ N. Then define the map: (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Z × QN −→ N k(cid:1) (cid:0)z, p Kα 7−→ (cid:16)z − pψ + pJk N k , p N k(cid:17) is easily checked to be a group isomorphism. Similar constructions may be used for the other two cases of Theorem (3.14). The following theorem shows that K0 groups of noncommutative solenoids give all possible Abelian extensions of QN by Z. Theorem 3.16. Given any Abelian extension: (3.8) 0 −−−−→ Z −−−−→ Q −−−−→ QN −−−−→ 0 there exists J ∈ ZN such that the extension of Z by QN given by the cocycle ξJ of Lemma (3.11) is equivalent to Extension (3.8). In particular, fixing any c ∈ [0, 1), there exists α ∈ ΞN with α0 = c and such that Q is isomorphic as a group to K0(A S α ). Proof. The Pontryagin dual of ZN is given by the Pr ufer N -group Z(N ∞) defined as the subgroup of T of all elements of order a power of N : endowed with the discrete topology. Z(N ∞) is also the inductive limit of: Z (N ∞) =nexp(cid:16)2iπ p N k(cid:17) : p ∈ Z, k ∈ No Z/N Z ⊂ Z/N 2Z ⊂ Z/N 3Z ⊂ · · · 22 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER and its dual pairing with ZN is given by: DJ, p N kE = exp(cid:18)2iπ p(Jk) N k (cid:19) where J ∈ ZN and p N k ∈ Z(N ∞). We note that from the theory of infinite Abelian groups [12, p. 219], to the short exact sequence of Abelian groups: 0 −−−−→ Z −−−−→ QN −−−−→ Z(N ∞) −−−−→ 0 there corresponds the Cartan-Eilenberg long exact sequence in Ext theory for groups: Hom(Z(N ∞), Z) / Hom(QN , Z) / Hom(Z, Z) sffffffffffffffffffffffff Ext(Z(N ∞), Z) / Ext(QN , Z) / Ext(Z, Z). Since Hom(QN , Z) = 0 and Ext(Z, Z) = 0, we deduce that we have a short exact sequence: 0 −−−−→ Hom(Z, Z) −−−−→ Ext(Z(N ∞), Z) −−−−→ Ext(QN , Z) −−−−→ 0. Since Z(N ∞) is a torsion group, the group Ext(Z(N ∞), Z) can be identified with Hom(Z(N ∞), Q/Z) ∼= \Z(N ∞) [12, p. 224], which in turn can be identified with the Pontryagin dual of Z(N ∞), namely ZN . The identification between Hom(Z(N ∞), Q/Z) and Ext(Z(N ∞), Z) is constructed as follows. Let s be a cross-section of π∗ with s(0) = 0Q in the short exact sequence: 0 −−−−→ Z −−−−→ Q π∗−−−−→ Q/Z −−−−→ 0 where π∗ is the natural projection. Any such choice will do, and we take s(z) = x with x ∈ Q ∩ [0, 1) uniquely defined by x ≡ z mod Z. We can then define the two-cocycle: Q/Z × Q/Z −→ Z (z1, z2) 7−→ s(z1) + s(z2) − s(z1 + z2). We can now identify ZN and Ext(Z(N ∞), Z) as follows. For J = (Jn)n∈N,n>0, we define the Z-valued 2-cocycle of QN by: ω(cid:12)(cid:12)(cid:12)(cid:12) ζJ :(cid:16) p1 , N k1 p2 N k2(cid:17) ∈ Q2 N 7−→ ω(cid:16)Jhπ∗(cid:16) p1 N k1(cid:17)i , Jhπ∗(cid:16) p2 N k2(cid:17)i(cid:17) . We then compute that: s ◦ π∗(x) = [x] mod 1, x ∈ Q, where for x ∈ Q, [x] mod 1 is defined to be that unique element of [0, 1) congruent to x modulo 1. Let us now fix J ∈ ZN . As before, we define (jn)n∈N by requiring for all n ∈ k=0 N kjk and jn ∈ {0, 1, · · · , N − 1}. We now calculate that N, n > 0 that Jn =Pn−1 / / s / / NONCOMMUTATIVE SOLENOIDS 23 the two-cocycle ζJ is given as follows: , p2 N k1 ζJ(cid:16) p1 ZJ(cid:16) p1 where ZJ(cid:0) p1 N k2(cid:17) = N k2(cid:17) − N k1 mod 1i +h p2Jk2 N k2(cid:1) = h p1Jk1 p1+p2 N k Jk mod 1 N k1 , p2 (p1N k2 −k1 +p2) (p2N k1 −k2 +p1) N k1 N k2 N k1 p2 , Jk1 mod 1 Jk2 mod 1 if k1 > k2, if k1 < k2, if k1 = k2 = k N k2 mod 1i. We remark that al- though each term in the expression defining the cocycle may not be an integer, the combination turns into an integer. We now claim that if J ∈ ZN is in the image of ι : Z → ZN described in the Lemma (3.13), then ζJ is a coboundary. This is to be expected from the short exact sequence giving Ext(QN , Z) as a quotient of Ext(Z(N ∞), Z). In this case, we recall that for (Jn)n∈N,n>0 = ι(P ) for P ≥ 0, there is M ∈ N such that Jn = P for all n ≥ M . In that case for all k1, k2 ≥ M and p1, p2 ∈ Z: ζJ ( p1 N k1 , p2 N k2 )) = (cid:20) p1P N k1 mod 1(cid:21)+(cid:20) p2P N k2 mod 1(cid:21)−h(cid:16) p1 N k1 + p2 N k2(cid:17) P mod 1i . But this eventually constant sequence is a coboundary, since defining µJ : QN → Z by: we check that: N k mod 1(cid:21) − pP N k , µJ : p N k 7→(cid:20) pP N k2(cid:17) − µJ(cid:16) p1 N k1 + p2 N k2(cid:17) = ζJ(cid:16) p1 N k1 , p2 N k2(cid:17) µJ(cid:16) p1 N k1(cid:17) + µJ(cid:16) p2 N k1 , p2 for all p1 N k2 ∈ QN . Similarly if J = ι(P ) for a negative integer P , the statement of Lemma (3.13) shows that jn = N − 1 for all n ≥ M, and one proves in a similar fashion that ζJ is a coboundary. We now claim that the two-cocycle of Lemma (3.11) (denoted hereafter by ξJ ) is cohomologous to ζJ . Recall that ξJ is defined by : − p1 − p2 N k1 (Jk2 − Jk1 ) N k2 (Jk1 − Jk2 ) q N r (Jk1 − Jr) if k2 > k1 if k1 > k2 if ∧(cid:26) k1 = k2 N k1 + p2 p1 N k2 = q N r (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 7−→ Z QN × QN ξJ N k1 , p2 N k2(cid:1) 7−→  (cid:0) p1 N k · Jk mod 1(cid:3) = h pN m have (cid:2) p that each Jk =Pk−1 computation. To establish this, we first remark that for all p N k ∈ QN and for all m ≥ 0, we N k+m · Jk+m mod 1i. We establish this by recalling jiN i, and the result is an easy i=0 jiN i so that Jk+m =Pk+m−1 i=0 Now consider the following one-cochain, generalizing our definition given ear- lier on this page: µJ(cid:12)(cid:12)(cid:12)(cid:12) QN −→ Z p N k 7−→ (cid:2) p N k Jk mod 1(cid:3) − p N k Jk 24 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER where the p the following two-coboundary δ(−µJ ) on QN × QN → Z given by: N k is taken in reduced form. Then the cobounding map takes −µJ to δ(−µJ )( p1 N k1 , p2 N k2 ) = p1 N k1 Jk1 − h p1 Jk1 mod 1i + )Jr +h( p2 N k2 + − ( N k1 p1 N k1 N k2 = q N r in reduced form. Then one verifies that p2 N k2 + Jk2 −h p2 N k2 Jk2 mod 1i )Jr mod 1i , p2 N k2 p1 N k1 where we want p1 N k1 + p2 ζJ δ(−µJ ) = ξJ , so that the cocycles ζJ and ξJ are cohomologous. (cid:3) Remark 3.17. Using Theorem (3.16) and Theorem (3.14), we have shown that Ext(QN , Z) is isomorphic to ZN /Z where we identified Z with ι(Z) ⊆ ZN . We now turn our attention to some properties of the C-algebras A S α for some special classes of α. There are three distinct subclasses of noncommutative solenoids based upon their basic structure: Definition 3.18. Let N ∈ N, N > 1. Let α ∈ ΞN . (1) If α is a periodic sequence (and thus in particular rational), we call A S α a periodic rational noncommutative solenoid. These are exactly the nonsim- ple noncommutative solenoids. (2) If α is a sequence of rationals, though not periodic, then we call A S α an aperiodic rational noncommutative solenoid. (3) If α is a sequence of irrationals (and thus can never be periodic), then we call A S α an irrational noncommutative solenoid. We note that simplicity is associated to a form of finiteness, or rationality con- dition: we need both the (eventual) periodicity of the decimal expansion of the entries of α and the periodicity of α itself. The aperiodic rational case is the more mysterious of the three and an interesting surprise. We start with the case where α is irrational. We use the following well known result [11] (see also [4] for a similar argument used for AF-algebras, which can be applied for AT-algebras as well), whose proof is included for the reader's conve- nience. We refer to [16] for the foundation of the theory of AT-algebras. A circle algebra is the C-algebra of n × n matrix - valued continuous functions on some connected compact subset of T. An AT-algebra is the inductive limit of a sequence of direct sums of circle algebras. Lemma 3.19. The inductive limit of AT-algebras is AT. Proof. Let (An)n∈N be a sequence of AT-algebras of inductive limit A. To simplify notations, we identify An with a subalgebra of A for all n ∈ N. Let ε > 0, k ∈ N with n > 0 and a1, . . . , ak ∈ A. Since A is an inductive limit, there exists K ∈ N and b1, . . . , bk ∈ AK such that kaj − bjk ≤ 1 2 ε for j = 1, . . . , k. Now, since AK is an AT-algebra, there exists L ∈ N, a finite direct sum C of circle algebras and c1, . . . , ck ∈ C such that kbj − cjk < 1 2 ε for j = 1, . . . , k. Hence, kaj − cjk < ε. By [16, Theorem 4.1.5], we have characterized A as an AT-algebra. (cid:3) NONCOMMUTATIVE SOLENOIDS 25 Proposition 3.20. Let N ∈ N, N > 1 and α ∈ ΞN . If α0 6∈ Q (or equivalently, if there exists k ∈ N such that αk 6∈ Q), then A S α is a simple AT-algebra of real rank 0. Proof. This follows from [10], Theorem (3.5), Lemma (3.19) and [16]. (cid:3) We now consider α ∈ ΞN (N ∈ N, N > 1) with α0 rational periodic. By The- orem (3.5), A S α is not simple. It is possible to provide a full description of the C-algebra A S α . We denote by Mq(C) the C-algebra of q × q matrices with com- plex entries, and we denote by C(X, A) the C-algebra of continuous function from a compact space X to a C-algebra A. Theorem 3.21. Let N ∈ N with N > 1 and α ∈ ΞN . Let α0 = p q with p, q ∈ N, nonzero, p and q relatively prime. Assume there exists k ∈ N nonzero such that (N k − 1)α0 ∈ Z, and that k is the smallest such nonzero natural. Let λ = exp(cid:16)2iπ p following two unitaries: q(cid:17). We define the uλ = 1 λ λ2 . . . λq−1  vλ =  0 · · · 0 1 . . . 0 0 · · · 0 1 0 · · · 0 1  N k and fiber Mq(C). More precisely, A S α is the C-algebra of continuous sections of a α is the fixed point of and observe that vλuλ = λuλvλ. Then A S bundle with base space S 2 C(S 2 N k , Mq(C)) for the action ρ of Z/qZ2 given by: λ u−n ρ(n,m)(ζ) : (z, w) ∈ S 2 for (n, m) ∈ (Z/qZ)2 and ζ ∈ C(S 2 N 7→ v−m N k , Mq(C)). λ ζ(λ−nz, λ−mw)un λvm λ Proof. By Theorem (2.12), our assumption implies that α is k-periodic. Let (βn)n∈N = (α0)n∈N - i.e. β is constant, and moreover β ∈ ΞN k . Let θk = ϕnk ◦ . . . ϕ(n+1)k−1 for all n ∈ N where we use the notations of Theorem (3.7). We have: A S (Aβ2k , θk) = A S β α = lim −→ (Aα2k , ϕk) = lim −→ as desired. We shall henceforth write β, by abuse of language, to mean the constant value the sequence β takes - namely α0. Let E = C(S 2 N k , Mq(C)) and let Eτ be the fixed point C-subalgebra of E for the action τ of (Z/qZ)2. It is well known that the fixed point C-algebra Eτ of τ is -isomorphic to Aβ. Let ϕ : E → E be defined by setting: ϕ(ζ) : (z, w) ∈ T2 7→ ζ(z(N k), w(N k)) for all ζ ∈ E. Now, using our assumption that (N k − 1)α0 ∈ Z so λ(N k) = 1, we show that ϕ and τ commute: τ(1,0)(ϕ(f ⊗ A)) : (z, w) ∈ T2 7→ f ((λ−1z)(N k), w(N k)) ⊗ A = f (λ−1z(N k), w(N k)) ⊗ A = ϕ(τ(1,0)(f ⊗ A)), for all f ∈ C(T2) and A ∈ Mq(C). Hence τ(1,0) ◦ ϕ = ϕ ◦ τ(1,0) ◦ ϕ by extending (3.9) linearly and by continuity. A similar computation would show that τ(0,1) ◦ ϕ = 26 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER ϕ ◦ τ(0,1). Hence, ϕ restricts to an endomorphism of Eτ . Now, the inductive limit of: E ϕ −−−−→ E ϕ −−−−→ E ϕ −−−−→ · · · is C(S 2 limit by: N k , Mq(C)). Since ϕ and τ commute, the action τ extends to the inductive for all ζ ∈ C(S 2 λ ζ(λ−pz, λ−qw)up ρ(p,q)(ζ) : (z, w) ∈ S 2 N k , Mq(C)) and moreover, the inductive limit of: N 7→ v−q λ u−p λvq λ ϕ ϕ ϕ which is A S action ρ of (Z/qZ)2 on C(S 2 Aα = Eτ −−−−→ · · · α by Theorem (3.7) is also the fixed point of C(S 2 −−−−→ Aα −−−−→ Aα N k , Mq(C)). Hence our theorem. N k , Mq(C)) by the (cid:3) We note that the proof of Theorem (3.21) shows that the embeddings from The- orem (3.7) map from and to the centers of the rotation C-algebras. This is in contrast with the situation when α0 is rational but α is not pseudo-periodic, which illustrates why the associated noncommutative solenoids are simple. 4. THE ISOMORPHISM PROBLEM Our classification of noncommutative solenoids is based on our computation of their K-theory. We start with the following simple observation: Lemma 4.1. Let σ : QN → QN be a group isomorphism. Then there exists p ∈ Z with p N and p 6∈ {−N, N } and k ∈ N such that σ(1) = p for all r ∈ N. Proof. Let us write σ(1) = pq N k in its reduced form, with q relatively prime with N and nonnegative. Note that as σ is an isomorphism, pq 6= 0 and moreover, there exists x ∈ QN such that σ(x) = p N k and we must have qx = 1. This contradicts the relative primality of N and q. (cid:3) N k . Consequently σ(cid:0) 1 N r(cid:1) = p N k+r We now obtain the main result of our paper. We fully characterize the iso- morphism classes of noncommutative solenoids based on the multipliers of adic rationals. Theorem 4.2. Let N, M ∈ N with N > 1 and M > 1. Let α ∈ ΞN and β ∈ ΞM . The following assertions are equivalent: α and A S β are -isomorphic, (1) The C-algebras A S (2) The integers N and M have the same set of prime factor. Let R be the the greatest common divisor of N and M , and let µ = N n = µnαn n = νnβn mod 1 for all n ∈ N and note α′, β′ ∈ ΞR. There mod 1 and β′ exists Λ ∈ P and γ ∈ ΞΛ such that both α′ and γ have a common subsequence, and β′ or −β′ = (1 − β′ n)n∈N has a common subsequence with γ. Moreover, {Λn : n ∈ N} is the set of prime factors of R. R and ν = M R . Set α′ Proof. Assume that there exists Λ ∈ P and γ ∈ ΞΛ such that α and β have subse- quences which are also subsequences of γ. Then a standard intertwining argument shows that A S γ . Moreover, for any irrational ro- tation algebra Aθ, we have that Aθ is -isomorphic to A−θ. Hence, A S β and A S −β are -isomorphic as well. β are -isomorphic to A S α and A S Now, let N = µR and assume the set of prime factors in µ is a subset of the set of prime factors of N . Set α′ n = µnαn for all n ∈ N. NONCOMMUTATIVE SOLENOIDS 27 First, it is straightforward to show that if N and R have the same set of prime factors, then QN and QR are isomorphic. Second, for all n ∈ N we have: Rα′ n+1 ≡ Rµn+1αn+1 mod Z ≡ µnN αn+1 mod Z ≡ µnαn mod Z ≡ α′ n mod Z. Hence α′ ∈ ΞR. Third, given pj Rkj = pj µkj N kj ∈ QR for j = 1, 2, 3, 4, we have: p4 Rk4(cid:17)(cid:17) = , , p2 Rk3 Rk1 Rk2(cid:17) ,(cid:16) p3 Ψα′(cid:16)(cid:16) p1 k1+k4 p1p4(cid:1)(cid:1) = exp(cid:0)2iπ(cid:0)α′ = exp(cid:0)2iπ(cid:0)αk1+k4 (µk1 p1µk4 p4(cid:1)(cid:1) = Ψα(cid:18)(cid:18) p1µk1 N k2 (cid:19) ,(cid:18) p3µk3 p2µk2 N k1 N k3 , , p4µk4 N k4 (cid:19)(cid:19) . Hence, Ψα = Ψ′ implies (1). α. Consequently, A S α′ = A S α . This concludes the proof that (2) α → A S Conversely, let θ : A S introduced in Theorem (3.8). If τ is a tracial state of A S on A S A S α . Denote, respectively, by τα and τβ the lift of a tracial state of A S β , and note that by Theorem (3.6), the choices of tracial state is irrelevant. By functoriality of K-theory, we obtain the following commutative diagram: β be a -isomorphism. We shall use the notations β then τ ◦ θ is a tracial state α and (4.1) K0(A S α ) K0(θ) K0(A S β ) #GGGGGGGGG τα {wwwwwwwww τβ R where K0(θ) is the group isomorphism induced by θ. To ease notations, let us write σ = K0(θ). Our first observation is that τβ ◦ σ(1, 0) = τα(1, 0) = 1, which implies that M k . It is easily checked that πβ is a group epimorphism. Moreover, ker πβ = {(z, 0) : z ∈ Z}. Conse- k σ(1, 0) = (1, 0). Let πβ : Kβ → QM be defined by πβ(cid:16)z + pJ β quently, if z, z ′ ∈ Z, since σ(cid:16)z + pJ α N k , p that: k M k , p M k(cid:17) = p N k(cid:17) = σ(z, 0) + σ(cid:16) pJ α N k(cid:19)(cid:19) = 0. pJ α k N k , N k , p p k p N k(cid:19) − σ(cid:18)z ′ + πβ(cid:18)σ(cid:18)z + pJ α k N k , N k(cid:17), we observe / / # { 28 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER Consequently, we have the following commuting diagram: Kα πα QN σ f Kβ πβ / QM . with f : QN → QM defined by setting f(cid:0) p N k(cid:1) = πβ ◦ σ(cid:16) pJ α k N k , p N k(cid:17). In particular, f is a group isomorphism, so the set of prime factors of N and M are the same and QN = QM . As we showed in the first half of this proof, and using the definition of our Theorem, α′, β′ ∈ ΞR and QN = QR where R is the greatest common divisor of N, M and A S β . We shall henceforth work within QR with α′ and β′. α while A S α′ = A S β ′ = A S Rk and p Let p ∈ Z, k ∈ N be defined so that f (1) = p Rn(cid:1) = p f(cid:0) 1 Rk is in reduced form, with p R and p 6∈ {−R, R} by Lemma (4.1) . Since f is an isomorphism, we have Rk+n for all n ∈ N. Using the notation Ω(R) for the number of prime factors of R, let Λ ∈ P be defined as a periodic sequence of period Ω(R) such that ΛΩ(R)−1−j = Λ(p)Ω(R)−1−j for j = 0, . . . , Ω(p) − 1 and πΩ(R)(Λ) = R. Any of the (Ω(R) − Ω(p))! possible choices of order for the first Ω(R) − Ω(p) values of Λ can be used, and we assume we pick one in the rest of this proof. We can visualize Λ as: ΛΩ(R)−Ω(p), · · · , ΛΩ(R)−1 , ΛΩ(R), · · · , Λ2Ω(R)−1 product = R equal to previous Ω(R) terms Let γ be the (unique) extensions of β′ to ΞΛ. Thus γΩ(R)n = β′ n for all n ∈ N. , · · · } z Λ =Λ0, Λ1, · · · , σ J α′ n product = p } {z Rn! = pn + 1 Now, for any n ∈ N, there exists pn ∈ Z such that pJ β ′ Rn+k , Rn , n+k { } {z Rn+k! . p Using the computation of the traces on K0 in Theorem (3.8) and the commutative diagram (4.1), and noting that if r = Ω(p) then pβ′ n = pγnΩ(R) = γnΩ(R)−r by definition of p, Λ and γ, we thus have: α′ n = pβ′ n+k + pn ≡ sign(p)γ(n+k)Ω(N )−r mod Z. Thus, sign(p)α′ is a subsequence of γ, and β is a subsequence of γ by construc- (cid:3) tion. This concludes the proof of (1) implies (2). Corollary 4.3. Let N, M be prime numbers. Let α ∈ ΞN and β ∈ ΞM . Then the following assertions are equivalent: (1) The noncommutative solenoids A S (2) We have N = M and one of the sequence α or β is a truncated subsequence of the β are -isomorphic, α and A S other. Proof. If N = M and α is a truncated subsequence of β then A S trivially -isomorphic. The same holds if β is a truncated subsequence of α. α and αβ are Conversely, assume A S are -isomorphic. Then as N and M are prime, so by Theorem (4.2) we have N = M . Moreover, there exists a sequence γ ∈ ΞN such that both α and β are subsequences of γ. Now, since α, β, γ ∈ ΞN , α and A S β / / / NONCOMMUTATIVE SOLENOIDS 29 this implies that for some n, n′ ∈ N we have αj = γn+j and βj = γn′+j for all j ∈ N. This shows that either α is a truncated subsequence of β (if n′ ≤ n) or β is a truncated subsequence of α. (cid:3) Theorem (4.2) relies on the invariant A S τα is the unique map given by lifting any tracial state of A S α to its K0 group. We would like to add an observation regarding the information on noncommutative solenoids one can read from the K0 group seen as an Abelian extension of Z by QN rather than as a group alone. We fix N ∈ N, N > 1. α (α ∈ ΞN ) 7→ (cid:0)K0(A S α ), τα(cid:1) where First, note that given α ∈ ΞN , the pair (K0(A S α ), [1]), where [1] is the K0-class of α , we can construct an Abelian extension of Z by QN by defining the identity of A S ι : z ∈ Z 7→ z[1] and noting that K0(A S α )/i(Z) is isomorphic to QN . Now, consider α, β ∈ ΞN such that there exists a (unital) -isomorphism ψ : α → A S β . Then the following diagram commutes: A S (4.2) 0 0 / Z / Z ι/ / Qα = K0(A S α ) K0(ψ) ι/ / Qβ = K0(A S β ) QN σ / QN 0 / 0 since ψ is unital, and where the arrow σ is defined and proven to be an isomor- phism by standard diagram chasing arguments. Conversely, we say that two Abelian extensions of Z by QN such that there exists a commutative diagram of the form Diagram (4.2) are weakly equivalent (note that Theorem (3.16) shows that any such extension can be obtained using the K- theory of noncommutative solenoids). Note that weakly equivalent extensions are isomorphic but not necessarily equivalent as extensions. The difference is that we allow for an automorphism σ of QN . This reflects, informally, that according to Theorem (4.2), the noncommutative solenoid A S α only partially determines α. Now, given two equivalent Abelian extensions of Z by QN , if one is weakly equivalent to some other extension, then so is the other. Hence, weakly equiva- lence defines an equivalence relation ≡ on Ext(QN , Z) such that if α, β ∈ ΞN give rise to -isomorphic noncommutative solenoids, then the associated Abelian ex- tensions of cocycle ξα and ξα (see Lemma (3.11)) in Ext(QN , Z) are equivalent for ≡. According to Theorem (3.16), the group Ext(QN , Z) is isomorphic to the quo- tient ZN /Z of the group ZN of N -adic integers by the group Z of integers. Using Theorem (4.2), and for N prime, we easily see that the relation induced by ≡ on ZN /Z is given by: where [J] is the class in ZN /Z of J ∈ ZN and N kJ is the sequence (N kJn)n∈N for any J ∈ ZN . [Jn]n∈N ≡ [Rn]n∈N ⇐⇒ ∃k ∈ N (cid:0)N kJ − R ∈ Z(cid:1) ∨(cid:0)N kR − J ∈ Z(cid:1) Hence, in conclusion, for a given α ∈ ΞN , the data(cid:0)K0(A S α ), [1], α0(cid:1) where α0 is the trace of any Rieffel-Powers projection in A S for A S equivalence ≡, and we can recover α up to a shift using the value α0. α , is a complete invariant α ), [1]) determines a cocycle ξJ in H 2(QN , Z), up to the α . Indeed, (K0(A S Our classification result has the following interesting dynamical application. The following is easily seen: / / / / / / / / 30 FR ´ED ´ERIC LATR ´EMOLI `ERE AND JUDITH PACKER Corollary 4.4. Let N, M ∈ N with N, M > 1. Let α ∈ ΞN and β ∈ ΞM . If any of the following assertion holds: (1) N and M have distinct set of prime factors, (2) For any Λ ∈ P such that {Λn : n ∈ N} is the set of prime factors of N , there is no γ ∈ ΞΛ such that α′ and β′ are subsequences of γ and no γ ∈ ΞΛ such that α′ and −β′ are subsequences of γ, where α′, β′ ∈ ΞR are defined as in Theorem (4.2), then the actions θα and θβ of, respectively, QN on SN and QM on SM are not topologi- cally conjugate. REFERENCES 1. L. Baggett, N. Larsen, J. Packer, I. Raeburn, and A. Ramsay, Direct limits, multiresolution analyses, and wavelets, Journal of Funct. Anal. 2010 (258), 2714–2738. 2. L. Baggett, K. Merrill, J. Packer, and A. Ramsay, Probability measures on solenoids corresponding to fractal wavelets, Trans. Amer. Math. Soc. (To appear). 3. A. Connes, C–alg`ebres et g´eom´etrie differentielle, C. R. de l'academie des Sciences de Paris (1980), no. series A-B, 290. 4. J. Dixmier, Some C-algebras considered by Glimm, Journal of Funct. Anal. 1 (1967), 182–203. 5. D. Dutkay, Low-pass filters and representations of the baumslag solitar group, Trans. Amer. Math. Soc. 358 (2006), 5271–5291. 6. D. Dutkay and P. Jorgensen, Wavelets on fractals, Rev. Mat. Iberoam. (2006), 131–180. 7. , Analysis of orthogonality and of orbits in affine iterated function systems, Math. Z. 256 (2007), 801–823. 8. S. Echterhoff and J. Rosenberg, Fine structure of the Mackey machine for actions of Abelian groups with constant Mackey obstruction, Pacific J. Math. 170 (1995), 17–52. 9. E. Effros and F. Hahn, Locally compact transformation groups and C ∗- algebras, vol. 75, American Mathematical Society, 1967. 10. G. Elliott and D. Evans, Structure of the irrational rotation C ∗-algebras, Annals of Mathematics 138 (1993), 477–501. 11. G. A. Elliott, On the classification of C ∗-algebras of real rank zero, J. Reine Angew. Math. 443 (1993), 179–219. 12. L. Fuchs, Infinite abelian groups, Volume I, Academic Press, 1970. 13. R. Hoegh-Krohn, M. B. Landstad, and E. Stormer, Compact ergodic groups of automorphisms, Annals of Mathematics 114 (1981), 75–86. 14. I. Kaplansky, Infinite abelian groups, University of Michigan Press, 1969. 15. A. Kleppner, Multipliers on Abelian groups, Mathematishen Annalen 158 (1965), 11–34. 16. H. Lin, An introduction to the classification of amenable C ∗-algebras, World Scientific, 2001. 17. J. Packer and I. Raeburn, On the structure of twisted group C ∗-algebras, Trans. Amer. Math. Soc. 334 (1992), no. 2, 685–718. 18. M. Pimsner and D. V. Voiculescu, Exact sequences for K-groups and Ext-groups of certain crossed product C-algebras, Journal of Operator Theory 4 (1980), no. 1, 93–118. 19. M. A. Rieffel, C-algebras associated with irrational rotations, Pacific Journal of Mathematics 93 (1981), 415–429. 20. M. A. Rieffel, Projective modules over higher-dimensional non-commutative tori, Can. J. Math. XL (1988), no. 2, 257–338. 21. H.-S. Yin, A simple proof of the classification of rational rotation C-algebras, Proc. Amer. Math. Soc. 98 (1986), 469–470. E-mail address: [email protected] URL: http://www.math.du.edu/frederic DEPARTMENT OF MATHEMATICS, UNIVERSITY OF DENVER, DENVER CO 80208 E-mail address: [email protected] URL: http://spot.colorado.edu/packer DEPARTMENT OF MATHEMATICS, UNIVERSITY OF COLORADO, BOULDER CO 80309
1510.06757	2	1510	2015-11-30T20:35:33	Cuntz Splice invariance for purely infinite graph algebras	[ "math.OA", "math.KT" ]	We show that the Cuntz Splice preserves the stable isomorphism class of a purely infinite graph C*-algebra with finitely many ideals.	math.OA	math	CUNTZ SPLICE INVARIANCE FOR PURELY INFINITE GRAPH ALGEBRAS RASMUS BENTMANN Abstract. We show that the Cuntz Splice preserves the stable isomorphism class of a purely infinite graph C∗-algebra with finitely many ideals. 1. Introduction The assignment E 7→ C∗(E) associates to a countable directed graph E a C∗-algebra C∗(E) given as the universal C∗-algebra with certain generators and relations encoded by the graph E. This generalizes a construction of Cuntz and Krieger exhibiting close ties to symbolic dynamics: the stabilized Cuntz -- Krieger algebra of a subshift of finite type is an invariant of flow equivalence [7, 8]. It is therefore natural to ask when two graphs E and F give rise to Morita equivalent C∗-algebras. In particular, it is desirable to find modifications that can be applied to a graph E, such that the C∗-algebra of the resulting graph F is always stably isomorphic to C∗(E). Various such modifications, or moves, have been established (see [2, 6, 10, 19]) and, in some cases, it has even been shown that two graphs E and F give rise to stably isomorphic C∗-algebras if and only if F is the result of a finite number of permitted modifications applied to E (see [10, 19]). One important example of such a graph modification is the Cuntz Splice. This move does not preserve the flow equivalence class of a subshift of finite type because it reverses the sign of the determinant of the matrix 1 − A, where A is the adjacency matrix defining the subshift (see [5]). However, the Cuntz Splice does preserve the stable isomorphism class of the associated Cuntz -- Krieger algebra OA (this follows, for instance, from the classification theorem in [17]). The more general question of Cuntz Splice invariance for the class of all graph C∗-algebras is currently open. A vital step towards the classification of unital graph C∗-algebras (that is, C∗-algebras of graphs with finitely many vertices) established in [10] was to prove Cuntz Splice invariance for this class (see [10, Proposition 5.8], or [18] for an earlier result along such lines). On the other hand, if a complete classification invariant has already been es- tablished by other means, one would hope to be able to determine whether Cuntz Splice invariance holds for the class of graphs under consideration by comparing the values of the invariant. In this article, we achieve this for the class of graphs whose C∗-algebra is purely infinite and has finitely many ideals. In [4], we have found a complete stable isomorphism invariant, denoted by XKδ, for the class of purely infinite graph C∗-algebras with finitely many ideals. Let C∗(E) belong to this class and let EC be the result of an application of a Cuntz Splice to the graph E. Then C∗(EC ) also belongs to the classified class, and we will show that C∗(E) and C∗(EC ) are stably isomorphic by verifying that XKδ(cid:0)C∗(E)(cid:1) ∼= XKδ(cid:0)C∗(EC )(cid:1). 2010 Mathematics Subject Classification. 46L35, 46L55, 46L80, 18G35. This work was supported by the German Research Foundation (DFG) through the project grant "Classification of non-simple purely infinite C∗-algebras". 1 2 RASMUS BENTMANN We would like to regard our result as positive evidence towards the hope that Cuntz Splice invariance holds more generally for all (non-simple, non-unital) graph C∗-algebras, although it is not clear to us how the methods from [10] should be adapted to the non-unital case. James Gabe recently announced results that lead to a generalization of the clas- sification theorem from [4] to cases of certain infinite primitive ideal spaces. A computation as we perform in the present article should then show that Cuntz Splice invariance also holds for purely infinite graph C∗-algebras with infinitely many ideals. We give a short description of the flavour of our computations. Let AE denote the adjacency matrix of the graph E. Recall that the K-theory groups of C∗(E) are computed as the kernel and the cokernel of the matrix (AE − 1)t regarded as a map on the free Abelian group on the set of vertices of E (assuming that all vertices of E are regular). An application of some row and column manipula- tions preserving the kernel and cokernel of a given matrix then easily shows that K∗(cid:0)C∗(E)(cid:1) ∼= K∗(cid:0)C∗(EC )(cid:1), that is, the Cuntz Splice preserves the K-theory of the graph C∗-algebra. As we shall describe below, the arguments in this article are a refinement of the above computation. In [3], we defined the homology theory XK for C∗-algebras over a finite space X taking values in Z/2-graded modules over the integral incidence algebra ZX of the partially ordered set (associated to) X. Let E be a graph with an identifica- tion Prim(cid:0)C∗(E)(cid:1) ∼= X. Then, just like above, XK(cid:0)C∗(E)(cid:1) is the homology of a 0 ) of projective ZX-modules canonically length-one chain complex PE associated to E. • = (P E 1 → P E • Our main result is that there exists an explicit quasi-isomorphism between the complexes PE ; this is a stronger statement than the two complexes merely having isomorphic homology modules. While the latter statement says that • and PEC XK(cid:0)C∗(E)(cid:1) ∼= XK(cid:0)C∗(EC )(cid:1), the former implies that XKδ(cid:0)C∗(E)(cid:1) ∼= XKδ(cid:0)C∗(EC )(cid:1), so that the classification results from [4] become applicable. The article is organized as follows. After the two preliminary Sections 2 and 3 discussing some aspects of graphs C∗-algebras and the classification invariant XKδ, respectively, we prove the main result under regularity assumptions in Section 4 and in full generality in Section 5. Acknowledgement. The author would like to thank James Gabe and Adam P. W. Sørensen for helpful discussions. 2. Graph C∗-algebras In this section, we gather definitions related to graph C∗-algebras and establish a few results needed later on. Throughout this article, a graph will always be a countable directed graph: Definition 2.1 (Graph notions). A graph E is a four-tuple E = (E0, E1, r, s) where E0 and E1 are countable sets, and r and s are maps from E1 to E0. The elements of E0 are called vertices, the elements of E1 are called edges, the map r is called the range map, and the map s is called the source map. We say that an edge e ∈ E1 is an edge from s(e) to r(e) and that s(e) emits e. A path in E is a sequence µ = e1e2 · · · en of edges ei ∈ E1 with n ≥ 1 and r(ei) = s(ei+1) for all i = 1, 2, . . . , n − 1. Extending the range and source maps to paths by setting s(µ) = s(e1) and r(µ) = r(en), a cycle is a path µ such that s(µ) = r(µ) and a return path is a cycle µ = e1e2 · · · en such that r(ei) 6= r(µ) for i < n. For v, w ∈ E0, we write v ≥ w if there is a path from v to w. CUNTZ SPLICE INVARIANCE FOR PURELY INFINITE GRAPH ALGEBRAS 3 A vertex v ∈ E0 is called regular if the set s−1(v) is finite and non-empty, it is called a sink if s−1(v) is empty, and it is called an infinite emitter if s−1(v) is infinite. A breaking vertex is an infinite emitter v ∈ E0 such that the number of emitted edges, whose range is equal to v or is the source of a path to v, is finite and non-zero. The graph E satisfies Condition (K) if no vertex v ∈ E0 supports precisely one return path. The adjacency matrix AE of E is the E0 × E0-matrix whose entry at (u, v) is the number of egdes from u to v. Definition 2.2 (Cuntz Splice). Let E = (E0, E1, r, s) be a graph and let v ∈ E0 be a vertex that supports at least two (distinct) return paths. The Cuntz Splice Ev C , rC , sC ) of E at v is defined by C = EC = (E0 C , E1 E0 E1 C := E0 ⊔ {u1, u2} C := E1 ⊔ {f1, f2, h1, h2, k1, k2}, where rC (e) = r(e) and sC (e) = s(e) for e ∈ E1 and sC (f1) = v, rC (f1) = u1, sC (f2) = u1, rC (f2) = v, C is made from the graph E by adding a segment of sC (ki) = u2, rC (ki) = ui. sC(hi) = u1, rC (hi) = ui, In other words, the graph Ev the form v * u1 u2 . Proposition 2.3. Let E be a graph satisfying Condition (K). Let v ∈ E0 support at least two return paths. Then Ev C satisfies Condition (K). Proof. In the graph Ev C the distinguished vertex v as well as the two added ver- C )0 \ {v, u1, u2} = E0 \ {v}. tices u1 and u2 support two return paths. Let w ∈ (Ev If w supports two return paths in E then these are also return paths in Ev C . If w supports no return path in E then in particular we cannot have v ≥ w ≥ v, so that w supports no return path in E either. (cid:3) We will use the more common graph C∗-algebra convention, which is opposite to the one in Raeburn's monograph [15]: Definition 2.4 (Graph C∗-algebra). Let E = (E0, E1, r, s) be a graph. The graph C∗-algebra C∗(E) is defined as the universal C∗-algebra generated by mutually orthogonal projections pv, v ∈ E0, and partial isometries se, e ∈ E1, satisfying • s∗ • s∗ • ses∗ esf = 0 for all e, f ∈ E1 with e 6= f , ese = pr(e) for all e ∈ E1, e ≤ ps(e) for all e ∈ E1, e for all regular vertices v ∈ E0. • pv =Pe∈s−1(v) ses∗ The lattice of closed ideals in C∗(E) will be denoted by I(cid:0)C∗(E)(cid:1), its primitive ideal space by Prim(cid:0)C∗(E)(cid:1). Definition 2.5 (Subsets of graphs). A subset K ⊆ E0 is hereditary if v ∈ K and v ≥ w implies w ∈ K. A subset K ⊆ E0 is saturated if the following holds for every regular vertex v ∈ E0: if r(e) ∈ K for every e ∈ E1 with s(e) = v then v ∈ K. The set of hereditary and saturated subsets of E is denoted by HS(E). Given a hereditary and saturated subset H ⊆ E0, one defines the set H fin ∞ to consist of all infinite emitters in E that belong to E0 \ H and emit a finite, non-zero number of edges to E0 \ H. An admissible pair for E is a pair (H, B) consisting of a hereditary and saturated set H ⊆ E0 and an arbitrary subset B ⊆ H fin ∞ . The set of admissible pairs for E is denoted by AP(E). * i i * * j j g g 4 RASMUS BENTMANN A maximal tail is a non-empty subset M ⊆ E0 such that the following conditions hold: • if v ≥ w and w ∈ M then v ∈ M ; • if v ∈ M is a regular vertex then there exists e ∈ E1 with s(e) = v and r(e) ∈ M ; • if v, w ∈ M there exists y ∈ M such that v ≥ y and w ≥ y. Theorem 2.6 ([11, Theorem 2.3]). The C∗-algebra C∗(E) is purely infinite if and only if the following conditions hold: • the graph E satisfies Condition (K); • there are no breaking vertices in E; • each vertex in each maximal tail M connects to a cycle in M . Proposition 2.7. Let E be a graph such that C∗(E) is purely infinite. Let v ∈ E0 support at least two return paths. Then C∗(Ev C ) is purely infinite. Proof. We must check that the three conditions in Theorem 2.6 are passed on from E to Ev In Proposition 2.3, we have seen that the Cuntz Splice inherits C . Condition (K). Clearly no vertex in Ev C except v has a chance of being a breaking vertex. But if v is regular in E then it is also regular in Ev C , and if v is an infinite emitter in E then it emits infinitely many edges, both in E and hence in Ev C , whose range is equal to v or supports a path to v. Thus there are no breaking vertices C . Then M ∩ E0 is a maximal tail in E in Ev (in order to check the second condition for the distinguished vertex v, notice that M ∩ E0 must completely contain the return paths in E based at v). This implies that every vertex in M ∩ E0 connects to a cycle in M . If ui ∈ M for one (hence both) i ∈ {1, 2}, then ui obviously connects to a cycle in M as well. (cid:3) C . Finally, let M be a maximal tail in Ev To an admissible pair (H, B) for E, one associates a gauge-invariant ideal J(H,B) in C∗(E) given as the closed span of a certain set of elements (see [1, page 6]). Theorem 2.8 ([1, Theorem 3.6, Corollary 3.10]). Let E be a graph satisfying Condition (K). The assignment (H, B) 7→ J(H,B) is a bijection from AP(E) to I(cid:0)C∗(E)(cid:1). One has J(H,B) ⊆ J(H ′,B ′) if and only if H ⊆ H ′ and B ⊆ H ′ ∪ B′. We equip the set AP(E) with the partial ordering ≤ described in the theorem. Proposition 2.9. Let E be a graph satisfying Condition (K). Let v ∈ E0 support at least two return paths. Then (H, B) 7→((H, B) (H ∪ {u1, u2}, B) if v 6∈ H, if v ∈ H defines an order isomorphism AP(E) → AP(Ev C ). Proof. Firstly, the assignment H 7→ ¯H :=(H if v 6∈ H, H ∪ {u1, u2} if v ∈ H C . Moreover, the set ¯H fin ∞ (formed in Ev is a bijection from the hereditary (and saturated) subsets of E to the hereditary (and saturated) subsets of Ev C ) coincides with the set H fin ∞ (formed in E). This gives the desired bijection. Let (H, B) and (H ′, B′) be admissible pairs for E. Then, by definition, ( ¯H, B) ≤ ( ¯H ′, B) if and only if ¯H ⊆ ¯H ′ and B ⊆ ¯H ′ ∪B′. The condition ¯H ⊆ ¯H ′ is equivalent to H ⊆ H ′. Moreover, since B∩{u1, u2} = ∅, the condition B ⊆ ¯H ′∪B′ is equivalent to B ⊆ H ′ ∪ B′. Hence ( ¯H, B) ≤ ( ¯H ′, B) if and only if (H, B) ≤ (H ′, B). (cid:3) CUNTZ SPLICE INVARIANCE FOR PURELY INFINITE GRAPH ALGEBRAS 5 Corollary 2.10. Let E be a graph satisfying Condition (K). Let v ∈ E0 support at least two return paths. Then J(H,B) 7→(J(H,B) J(H∪{u1 ,u2},B) if v 6∈ H, if v ∈ H defines an order isomorphism I(cid:0)C∗(E)(cid:1) → I(cid:0)C∗(Ev C )(cid:1). Proof. Combine Theorem 2.8 and Proposition 2.9. (cid:3) 3. The classification invariant for purely infinite graph C∗-algebras In this section, we discuss the invariant XKδ, which was shown to be a complete (strong) classification invariant for purely infinite graph C∗-algebras with finitely many ideals in [4] (using results of Kirchberg [12]). For this purpose, we first need to recall the invariant XK introduced in [3]. Let X be a finite T0-space; it carries a partial ordering called the spezialization preorder defined such that x ≥ y if and only if Ux ⊆ Uy, where Ux denotes the smallest open neighbourhood of the point x ∈ X. Recall that if A is a C∗-algebra over X as defined in [14] (that is, it is equipped with a continuous map Prim(A) → X) then every open subset U of X gives rise to an ideal A(U ) of A. Moreoever, if U ⊆ V ⊆ X are open subsets, we have an ideal inclusion ιV Definition 3.1. For a C∗-algebra A over X, the invariant XK(A) consists of the U : A(U ) ⊆ A(V ). collection of Z/2-graded Abelian groups K∗(cid:0)A(Ux)(cid:1) for x ∈ X together with the collection of graded group homomorphisms K∗(cid:0)ιUy The assignment XK becomes a functor from the category of (separable) C∗-al- gebras over X to the category of X-diagrams in (countable) Z/2-graded Abelian groups or, equivalently, to the category of (countable) Z/2-graded modules over the integral incidence algebra of the partially ordered set X (see [3, §4] for more details). If M is an X-diagram as above, we will denote the Z/2-graded Abelian group corresponding to the point x ∈ X by Mx and the homomorphism Mx → My Ux(cid:1) for x ≥ y. for x ≥ y by Mx→y. Thus Definition 3.1 says that XK(A)x = K∗(cid:0)A(Ux)(cid:1) and XK(A)x→y = K∗(cid:0)ιUy Ux(cid:1). The invariant XKδ(A) is only defined for C∗-algebras A over X such that the projective dimension of the module XK(A) is at most 2 (see [4, §2]). This was shown in [4, §5.3] to be the case whenever A is a graph C∗-algebra with primitive ideal space X. Then XKδ(A) consists of the invariant XK(A) equipped with the additional structure of a so-called obstruction class δA, which is an element of the Z/2-graded Abelian group (3.2) where [−1] signifies a degree-shift just like in the Ext1-term in the universal co- efficient sequence. Notice that this Ext2-group is formed in the Abelian target category of XK described above. An isomorphism XKδ(A) ∼= XKδ(A′) is then simply an isomorphism XK(A) ∼= XK(A′) such that the induced identification Ext2(cid:0)XK(A), XK(A)[−1](cid:1), takes the element δA to the element δA′ . Ext2(cid:0)XK(A), XK(A)[−1](cid:1) ∼= Ext2(cid:0)XK(A′), XK(A′)[−1](cid:1) We will omit the general, intrinsic definition of the obstruction class δA given in [4, Definition 2.16] giving preference to a more explicit description resulting from [4, §5] in the case that A = C∗(E) is a graph C∗-algebra. Since the module XK1(cid:0)C∗(E)(cid:1) is always projective by [4, Lemma 5.18], one summand of the group (3.2) vanishes. Elements of the second summand Ext2(cid:0)XK0(A), XK1(A)(cid:1) can be 6 RASMUS BENTMANN described as equivalence classes of length-two extensions of XK0(A) by XK1(A) (see for instance [13, III. §5]). As was shown in [4, Theorem 5.19], the obstruction class δC∗(E) is represented by the dual Pimsner -- Voiculescu sequence 0 → XK1(A) → XK0(A ⋊γ T) ∗ −id γ(1)−1 −−−−−−−→ XK0(A ⋊γ T) → XK0(A) → 0, where γ denotes the canonical gauge action on A = C∗(E). (As long as one is only interested in an exact sequence, the map in the middle of this sequence is only determined up to a factor of ±1. Changing the sign however also replaces the represented Ext2-class by its additive inverse. Hence there is good reason for our choice of convention in the given setting.) Recall that C∗(E) ⋊γ T is the graph C∗-algebra of the skew-product graph E ×1 Z (see [16, § 3] and [1, § 6]). The discussion in [4, §5] also shows that we can replace C∗(E ×1 Z) with the subalgebra C∗(E ×1 N ) where N = {n ∈ Z n ≤ 0}. This is an improvement because the K-theory of C∗(E ×1 N ) is more manageable than the one of C∗(E ×1 Z). By [4, Theorem 5.3 and (5.17)], the resulting length-two extension XK1(cid:0)C∗(E)(cid:1) ֌ XK0(cid:0)C∗(E ×1 N )(cid:1) S−id−−−→ XK0(cid:0)C∗(E ×1 N )(cid:1) ։ XK0(cid:0)C∗(E)(cid:1) still represents the class δC∗(E) ∈ Ext2(cid:16)XK0(cid:0)C∗(E)(cid:1), XK1(cid:0)C∗(E)(cid:1)(cid:17). Here S de- notes the map induced by the shift map (e, n) 7→ (e, n − 1) on E ×1 N . Definition 3.3. We define the length-one chain complex PE • = (P E 1 ϕE −−→ P E We observe that the chain complex PE 0 ) :=(cid:16)XK0(cid:0)C∗(E ×1 N )(cid:1) S−id−−−→ XK0(cid:0)C∗(E ×1 N )(cid:1)(cid:17) • carries all information about the invariant XKδ(cid:0)C∗(E)(cid:1): we can recover XK1(cid:0)C∗(E)(cid:1) and XK0(cid:0)C∗(E)(cid:1) as the first and zeroth homology modules of this complex; the obstruction class is then represented by the sequence ker(ϕ) ֌ P E 1 ϕE −−→ P E 0 ։ coker(ϕ). As a consequence, we obtain the following criterion for isomorphism on the invari- ant XKδ. Recall that a quasi-isomorphism is a chain map inducing isomorphisms on homology in each degree. If the chain complexes PE1 • Proposition 3.4. Let E1, E2 be graphs such that Prim(cid:0)C∗(Ei)(cid:1) ∼= X for both i. are quasi-isomorphic then XKδ(cid:0)C∗(E1)(cid:1) ∼= XKδ(cid:0)C∗(E2)(cid:1). Proof. A quasi-isomorphism ψ• : PE1 • gives rise to a commutative diagram • → PE2 and PE2 • ∼= XK1(cid:0)C∗(E1)(cid:1) XK1(cid:0)C∗(E2)(cid:1) ϕE1 ϕE2 P E1 1 ψ1 P E2 1 P E1 0 ψ0 P E2 0 ∼= XK0(cid:0)C∗(E1)(cid:1) XK0(cid:0)C∗(E2)(cid:1) The dotted arrows thus yield an isomorphism XK(cid:0)C∗(E1)(cid:1) ∼= XK(cid:0)C∗(E2)(cid:1). As we have seen above, the two rows represent the obstruction classes δC∗(E1) and δC∗(E2), respectively. The commutative diagram thus shows that the obtained dotted iso- morphism identifies the obstruction classes (the two Ext-classes coincide by [13, III. Proposition 5.2]). (cid:3) CUNTZ SPLICE INVARIANCE FOR PURELY INFINITE GRAPH ALGEBRAS 7 4. The regular case In this section, we establish the desired result for the case that every vertex in the graph is regular: Theorem 4.1. Let E be a graph with no sinks and no infinite emitters. Assume that C∗(E) is purely infinite and has finitely many ideals. Let v be a vertex of E supporting at least two return paths. Then C∗(E) ⊗ K ∼= C∗(Ev Proof. The C∗-algebra C∗(Ev C ) is purely infinite by Proposition 2.7. Corollary 2.10 provides a homeomorphism Prim(cid:0)C∗(Ev rem 5.19], it suffices to show that XKδ(cid:0)C∗(E)(cid:1) ∼= XKδ(cid:0)C∗(Ev C )(cid:1) → Prim(cid:0)C∗(E)(cid:1) =: X. By [4, Theo- C )(cid:1), which follows from Proposition 3.4 once we establish a quasi-isomorphism P Ev • → PE • . C C ) ⊗ K. and Ev We describe the chain complexes PE in more detail. Both graphs E C satisfy Condition (K) by Theorem 2.6. Since there are no infinite emitters, C ) → Theorem 2.8 provides order isomorphisms HS(E) → I(cid:0)C∗(E)(cid:1) and HS(Ev I(cid:0)C∗(Ev C )(cid:1) given by H 7→ JH := J(H,∅), and Corollary 2.10 shows that Ev • and P C • JH 7→(JH JH∪{u1,u2} if v 6∈ H, if v ∈ H is a well-defined order isomorphism I(cid:0)C∗(E)(cid:1) → I(cid:0)C∗(Ev C )(cid:1). Given a point x ∈ X, we let Hx ∈ HS(E) and H ′ x ∈ HS(Ev i )x = XK0(cid:0)C∗(E ×1 N )(cid:1)x = K0(cid:0)C∗(E ×1 N )(Ux)(cid:1) ∼= K0(cid:0)C∗(Hx ×1 N )(cid:1) ∼= ZHx itary and saturated subsets such that C∗(E)(Ux) = JHx and C∗(Ev By [16, Theorem 3.2], for i ∈ {0, 1}, we have (P E and, similarly, (P Ev ZHx ֒→ ZHy and analogously for (P Ev x : (P E the components ϕE identifications above, by i )x→y is simply the inclusion )x→y. Finally, again by [16, Theorem 3.2], 0 )x of the module map ϕE are given, under the x . If x ≥ y, the map (P E 1 )x → (P E )x ∼= ZH ′ C C i i C ) denote the hered- C )(Ux) = JH ′ x . ZHx x −1)t (AE −−−−−→ ZHx , where AE x denotes the restriction of the adjacency matrix AE to the index set Hx ⊆ E0 (in both the rows and the columns), and 1 denotes the identity matrix of the appropriate size. Again, the analogous formula holds for the map ϕEv x . Here x − 1)t as we think of elements of ZHx as column vectors and view the matrix (AE a map via matrix multiplication from the left, et cetera. Hence we have implicitly chosen enumerations of the sets Hx and H ′ x. In the following, we assume, if v ∈ Hx, that these enumerations have been chosen to be of the form Hx = (v, v1, v2, . . .) and H ′ x = (u2, u1, v, v1, v2, . . .), where (v1, v2, . . .) is an arbitrary enumeration of Hx \ {v}. Thus the matrix A Ev x has the following specific block form: C C Ev C x = A 0 1 1 1 0 0 ... 1 1 1 0 ...   0 · · · 0 · · · AE x .   Ev • → PE C Having fully described the chain complexes PE , we will now define a chain map ψ• : P • and verify that it is a quasi-isomorphism. For this, we need to specify, for every x ∈ X and i ∈ {0, 1}, a group homomorphism (ψi)x : (P Ev i )x such that all face squares of the cube in Figure 1 com- mute whenever x ≥ y. We know that the front and back square commute because )x → (P E C i Ev • and P C • 8 RASMUS BENTMANN ϕEv x C )x (P Ev C 1 (ψ1)x (P Ev C 0 )x ϕE x )x→y (ψ0)x (P E 0 )x (P Ev C 0 )x→y (P E 1 )x (P Ev C 1 (P E 1 )x→y (P Ev C 1 (P E 0 )x→y )y ϕEv y C (P Ev C 0 )y (P E 1 )y (ψ1)y ϕE y (ψ0)y (P E 0 )y Figure 1. A commuting cube ϕE and ϕEv C are module maps. The left and right square commuting means that ψ1 and ψ0 are module maps, and the top and bottom square commuting says that ψ• is a chain map. For x ∈ X such that Hx ∋ v, we define the maps (ψi)x from (P Ev C )x ∼= ZH ′ x ∼= i Z{u1,u2} ⊕ ZHx to (P E i )x ∼= ZHx by the following block matrices: (ψ1)x =  0 0 ... 0 0 ... , 1  (ψ0)x =  −1 0 0 0 ... ... . 1  i C )x ∼= ZH ′ x ∼= ZHx to If Hx 6∋ v, we simply let (ψi)x be the identity map from (P Ev (P E i )x ∼= ZHx for both values of i. We check the commutativity of the left- and right-hand square. If Hx ∋ v then also Hy ∋ v and the left-hand square commutes because projecting from Z{u1,u2} ⊕ ZHx to ZHx and then including into ZHy is the same thing as including Z{u1,u2} ⊕ ZHx into Z{u1,u2} ⊕ ZHy and then projecting onto ZHy . In the right-hand square, if Hx ∋ v, the two composite maps Z{u1,u2} ⊕ ZHx → ZHy again restrict to the inclusion ZHx ֒→ ZHy on the second summand. On the first summand Z{u1,u2}, both take an element (a, b) to the vector (−a, 0, 0, . . .). Now we assume that Hx 6∋ v. In the case Hy 6∋ v, commutativity of both squares is trivial. Otherwise, the left- and right-hand square commute because composing (ψi)y with the inclusion ZHx ֒→ Z{u1,u2} ⊕ ZHy simply gives the inclusion ZHx ֒→ ZHy for both values of i. This shows that the left- and right-hand square always commute. Next we consider the top square. Commutativity is clear when Hx 6∋ v. Other- wise, it comes down to the identity of the two matrix products (AE x − 1)t  0 0 0 0 ... ... −1 0 0 0 ... ... 1  ,   1  0 1 1 0 0 1 0 0 ... ...   0 0 1 0 · · · · · · · · · · · · (AE x − 1)t ,   CUNTZ SPLICE INVARIANCE FOR PURELY INFINITE GRAPH ALGEBRAS 9 both of which are indeed equal to   0 0 0 0 ... ... So far, we have defined a chain map ψ• : P (AE x − 1)t  . Ev • → PE C • and it remains to check that it induces isomorphisms on first and zeroth homology. Consider the diagram ker(ϕEv C ) ψ1 ker(ϕE) ϕEv C ϕE P Ev C 1 ψ1 P E 1 P Ev C 0 ψ0 P E 0 coker(ϕEv C ) ¯ψ0 coker(ϕE), where ψ1 is the restriction of ψ1 and ¯ψ0 is the map induced by ψ0. We wish to show that, for every given point x ∈ X, the components (ψ1)x and ( ¯ψ0)x are invertible. )x ∼= Z{u1,u2} ⊕ ZHx satisfy (ψ1)x(m) = 0. Then m x (m) = (b, a, b, 0, 0 . . .)t shows that must be of the form (a, b, 0, 0, . . .)t, and 0 = ϕEv a = b = 0 and thus m = 0. Hence (ψ1)x is injective. Let m ∈ ker(ϕEv x ) ⊆ (P Ev 1 C C C Given m = (m1, m2, . . .)t ∈ ker(ϕE ∼= ZHx , we define m′ := (−m1, 0, m1, m2, . . .)t ∈ Z{u1,u2} ⊕ ZHx ∼= (P Ev x ) ⊆ (P E 1 )x C )x. 1 Then m′ ∈ ker(ϕEv Let ¯m ∈ coker(ϕEv m = (a, b, m1, m2, . . .)t ∈ P Ev C x ) and (ψ1)x(m′) = m. Hence (ψ1)x is surjective. C x ) satisfy ( ¯ψ0)x( ¯m) = 0. Then ¯m is represented by an element C 0 ∼= Z{u1,u2} ⊕ ZHx such that (ψ0)x(m) = (−a + m1, m2, m3, . . .)t belongs to the image of ϕE (n1, n2, . . .)t ∈ ZHx . We define the element x , that is, (ψ0)x(m) = (AE x − 1)t(n) for some n = n′ = (b − n1, a, n1, n2, . . .)t ∈ Z{u1,u2} ⊕ ZHx ∼= (P Ev C 1 )x. Its image under the map ϕEv x C is 0 0 1 0 · · · · · · · · · · · · b − n1 (AE x − 1)t     a n1 n2 ... =     a b a 0 0 ...     0 0 m2 m3 ...   −a + m1 + = m. 0 1 1 0 0 1 0 0 ... ...   This shows that ¯m = 0. Hence ( ¯ψ0)x is injective. Surjectivity of ( ¯ψ0)x follows immediately from surjectivity of (ψ0)x. This completes the proof that ψ• is a quasi-isomorphism. (cid:3) 5. The general case In this section, we remove the regularity assumptions on the graph E from the previous section. We will use the so-called Drinen -- Tomforde Desingularization procedure introduced in [9]. It proceeds by "adding a tail" to every sink and every infinite emitter in E (see [9, Definition 2.1]). It is shown in [9, Theorem 2.11] that this procedure does not change the stable isomorphism class of the graph C∗-algebra. As we shall see, the desired theorem reduces to Theorem 4.1 because Drinen -- Tomforde Desingularization commutes with the Cuntz Splice up to stable isomorphism of the associated graph C∗-algebra. 10 RASMUS BENTMANN Lemma 5.1. Let E be a graph. Let v be a vertex of E supporting at least two return paths. Let F be a Drinen -- Tomforde Desingularization of E. Then v also supports two return paths in F . Moreover, the Cuntz Splice F v C is stably isomorphic to (any Drinen -- Tomforde Desingularization of ) Ev C . Proof. The first claim is straightforward. Desingularizing a singular vertex different from v commutes with the Cuntz Splice at v even in the strong sense that compos- ing the two procedures in both possible orders results in exactly the same graph (provided, in the case of an infinite emitter, that the same enumeration of outgo- ing edges is chosen in E and in Ev C -- differing choices will give different graphs but result in stably isomorphic graph C∗-algebras by [9, Theorem 2.11]). Hence we may assume that v is an infinite emitter and that all other vertices in E are regular. If the set of edges emitted from v in E is enumerated as (e1, e2, . . .), we choose the enumeration (f1, e1, e2, . . .) for the set of edges emitted from v in Ev C . We will show that F v C is stably isomorphic to the corresponding Drinen -- Tomforde Desingularization of Ev C . This implies the second claim of the lemma. The relevant segments of the two resulting graphs we need to compare look as follows: (5.2) · · · · · · · · · · · · v2 v1 f1 + u1 v r(e3) r(e2) r(e1), v2 v1 f1 * u1 v u2 u2 r(e2) r(e1). Up to omitted vertices and edges, the first graph represents the Cuntz Splice F v C of F , while the second graph is the desingularization of Ev C . Notice that a vertex of the form r(ei) may agree with v, although this is not reflected in our drawing. We observe that the subgraph T = {v1} in the second graph is contractible in the sense of [6] (although the criteria in [6] are complicated, they are trivially satisfied for our choice of T consisting of only one vertex and no edges), and that the contraction procedure described in [6, Theorem 3.1] yields the graph · · · · · · v3 v2 v + u1 u2 r(e3) r(e2) r(e1), which is isomorphic to the first graph above. Hence the two graphs partially shown in (5.2) have stably isomorphic graph C∗-algebras by [6, Theorem 3.1]. Notice that the assumption of no tails in [6, Theorem 3.1] does no harm here: we may simply replace all eventual tails with sinks before applying the Crisp -- Gow Contraction and desingularize these sinks afterwards. We remark that it is also possible to write the particular contraction move above as a combination of simpler moves (an out-splitting followed by two reversed out- delays) whose preserving the stabilized graph C∗-algebra was already established in [2]; see also [19, Theorem 5.2]. (cid:3) o o o o o o + j j * * X X j j g g o o o o o o * i i * * X X j j g g o o o o o o + j j * * X X j j g g CUNTZ SPLICE INVARIANCE FOR PURELY INFINITE GRAPH ALGEBRAS 11 Theorem 5.3. Let E be a graph. Assume that C∗(E) is purely infinite and has finitely many ideals. Let v be a vertex of E supporting at least two return paths. Then C∗(E) ⊗ K ∼= C∗(Ev C ) ⊗ K. Proof. Combining [9, Theorem 2.11], Theorem 4.1 and Lemma 5.1, we get C∗(E) ⊗ K ∼= C∗(F ) ⊗ K ∼= C∗(F v C ) ⊗ K ∼= C∗(Ev C ) ⊗ K, where F denotes some Drinen -- Tomforde Desingularization of E. (cid:3) References [1] Teresa Bates, Jeong Hee Hong, Iain Raeburn, and Wojciech Szymański, The ideal structure of the C ∗-algebras of infinite graphs, Illinois J. Math. 46 (2002), no. 4, 1159 -- 1176. [2] Teresa Bates and David Pask, Flow equivalence of graph algebras, Ergodic Theory Dynam. Systems 24 (2004), 367 -- 382. [3] Rasmus Bentmann, Kirchberg X-algebras with real rank zero and intermediate cancellation, J. Noncommut. Geom. 8 (2014), no. 4, 1061 -- 1081. [4] Rasmus Bentmann and Ralf Meyer, A more general method to classify up to equivariant KK-equivalence (2014). arXiv: 1405.6512. [5] Rufus Bowen and John Franks, Homology for Zero-Dimensional Nonwandering Sets, Ann. of Math. (2) 106 (1977), no. 1, 73 -- 92. [6] Tyrone Crisp and Daniel Gow, Contractible subgraphs and Morita equivalence of graph C ∗-algebras, Proc. Amer. Math. Soc. 134 (2006), no. 7, 2003 -- 2013. [7] Joachim Cuntz, A class of C ∗-algebras and topological Markov chains. II. Reducible chains and the Ext-functor for C ∗-algebras, Invent. Math. 63 (1981), no. 1, 25 -- 40. [8] Joachim Cuntz and Wolfgang Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56 (1980), no. 3, 251 -- 268. [9] Douglas John Drinen and Mark Tomforde, The C ∗-algebras of arbitrary graphs, Rocky Moun- tain J. Math. 35 (2005), no. 1, 105 -- 135. [10] Søren Eilers, Gunnar Restorff, Efren Ruiz, and Adam Sørensen, Geometric classification of unital graph C ∗-algebras of real rank zero (2015). arXiv: 1505.0677. [11] Jeong Hee Hong and Wojciech Szymański, Purely Infinite Cuntz -- Krieger Algebras of Directed Graphs, Bull. London Math. Soc. 35 (2003), no. 5, 689 -- 696. [12] Eberhard Kirchberg, Das nicht-kommutative Michael-Auswahlprinzip und die Klassifikation nicht-einfacher Algebren, C ∗-Algebras (Münster, 1999), Springer, Berlin, 2000, pp. 92 -- 141. [13] Saunders Mac Lane, Homology, Classics in Mathematics, Springer, Berlin, 1995. Reprint of the 1975 edition. [14] Ralf Meyer and Ryszard Nest, C ∗-Algebras over topological spaces: the bootstrap class, Mün- ster J. Math. 2 (2009), 215 -- 252. [15] Iain Raeburn, Graph algebras, CBMS Regional Conference Series in Mathematics, vol. 103, Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2005. [16] Iain Raeburn and Wojciech Szymanski, Cuntz -- Krieger algebras of infinite graphs and matri- ces, Trans. Amer. Math. Soc. 356 (2004), 39-59. [17] Gunnar Restorff, Classification of Cuntz -- Krieger algebras up to stable isomorphism, J. Reine Angew. Math. 598 (2006), 185 -- 210. [18] Mikael Rørdam, Classification of Cuntz -- Krieger algebras, K-Theory 9 (1995), no. 1, 31 -- 58. [19] Adam P. W. Sørensen, Geometric classification of simple graph algebras, Ergodic Theory Dynam. Systems 33 (2013), no. 4, 1199 -- 1220. E-mail address: [email protected] Mathematisches Institut, Georg-August Universität Göttingen, Bunsenstrasse 3 -- 5, 37073 Göttingen, Germany
1801.00767	2	1801	2019-07-18T20:31:02	The local-triviality dimension of actions of compact quantum groups	[ "math.OA", "math.QA" ]	We define the local-triviality dimension for actions of compact quantum groups on unital C-algebras. The resulting compact quantum principal bundle is said to be locally trivial when this dimension is finite. For commutative C-algebras, this notion recovers the standard definition of local triviality of compact principal bundles. We prove that actions with finite local-triviality dimension are automatically free. Then we apply this new notion to prove the noncommutative Borsuk-Ulam-type conjecture under the assumption that a compact quantum group admits a non-trivial classical subgroup whose induced action has finite local-triviality dimension. This is a noncommutative extension of the Borsuk-Ulam-type theorem for locally trivial principal bundles.	math.OA	math	THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS EUSEBIO GARDELLA, PIOTR M. HAJAC, MARIUSZ TOBOLSKI, AND JIANCHAO WU Abstract. We define the local-triviality dimension for actions of compact quan- tum groups on unital C-algebras. The resulting compact quantum principal bundle is said to be locally trivial when this dimension is finite. For commuta- tive C-algebras, this notion recovers the standard definition of local triviality of compact principal bundles. We prove that actions with finite local-triviality dimension are automatically free. Then we apply this new notion to prove the noncommutative Borsuk -- Ulam-type conjecture under the assumption that a com- pact quantum group admits a non-trivial classical subgroup whose induced action has finite local-triviality dimension. This is a noncommutative extension of the Borsuk -- Ulam-type theorem for locally trivial principal bundles. Contents Introduction 1. 2. Different types of equivariant noncommutative joins 2.1. Preliminaries on joins and actions 2.2. The free noncommutative join 3. Locally trivial compact quantum principal bundles 3.1. The (weak) local-triviality dimension 3.2. Examples of locally trivial compact quantum principal bundles 3.3. The n-universal bundles and the strong local-triviality dimension 4. Locally trivial compact principal bundles 5. Relations with piecewise triviality and the Rokhlin dimension 5.1. Piecewise triviality 5.2. Rokhlin dimension 6. The noncommutative Borsuk -- Ulam-type conjectures Acknowledgement References 1 5 5 8 12 12 15 18 21 24 25 26 32 34 35 1. Introduction The noncommutative geometry program of Connes [14], taking inspiration from a variety of areas including representation theory, quantum mechanics, and index theory, aims at building new mathematical tools by providing noncommutative (or "quantum") generalizations of classical mathematical theories. One starting point of this program is the Gelfand -- Naimark natural duality between commutative C-algebras and locally compact Hausdorff spaces [22], which leads to the percep- tion of the study of general, possibly noncommutative C-algebras as the theory of 1 2 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU noncommutative topology. Many other important mathematical subjects have since found their noncommutative analogues, e.g., measure theory, spin geometry, and topological K-theory. The resulting new tools have made far-reaching impact in topology, representation theory, ergodic theory, mathematical physics, etc. In this paper, we are interested in the concept of a principal bundle which is pivotal in algebraic topology and provides a rigorous mathematical description of gauge field theories in physics. For a compact group G, a compact Hausdorff principal G-bundle in the sense of H. Cartan may be taken as a compact Hausdorff space with a free continuous G-action. However, it is often desirable in applications to impose the condition of local triviality, namely that each point has a neighborhood of the form G × W , where W is some open subset of the orbit space and G acts only on the first coordinate by translation. Such a neighborhood is called trivializable, as it represents a trivial bundle. Local triviality is incorporated into the definition of principal bundles per Steenrod, but for the sake of clarity, we will stick to Cartan's formulation in this paper and explicitly state local triviality when needed. Attempts to extend locally trivial principal bundles to the realm of noncommu- tative topology go back already to the 1990's [10, 43]. Unlike the notion of a vector bundle, which immediately entered noncommutative topology as a finitely generated projective module via the Serre -- Swan theorem, the notion of a locally trivial com- pact Hausdorff principal bundle resisted generalization to noncommutative topology, largely due to the latter's global nature. Already the concept of a free action on C-algebras proved to be difficult to formulate in a satisfactory manner [4]. A key problem in imposing the local-triviality condition in this setting is the lack of a C-algebraic formulation of the notion of an open cover. Therefore, there came first the definition of a piecewise trivial compact quantum principal bundle, using appropriate families of ideals to define a noncommutative finite closed cover [24, 23]. However, as shown in [5], a piecewise trivial compact Hausdorff principal bundle need not be locally trivial, so the problem of introducing local triviality to noncommutative geometry remained unsolved. In this paper, we introduce a notion of locally trivial compact quantum princi- pal bundles. Our local triviality is characterized by the finiteness of a dimension concept which we call the local-triviality dimension, defined for actions of compact quantum groups on unital C-algebras. Our approach is inspired by the theory of the Rokhlin dimension [26] used in and around the classification program of unital simple separable nuclear C-algebras and is also motivated by the noncommutative Borsuk -- Ulam-type conjecture [3]. To explain how we circumvent the need for open covers in our definition, we first describe an equivalent characterization of local triviality in the classical setting. We begin by noting that a prominent example of locally trivial compact Hausdorff prin- cipal bundles for a compact group G is given by applying Milnor's join construction to G. More precisely, the join G ∗ G is defined to be the quotient topological space of G × G × [0, 1] by collapsing each of the two copies of G at the endpoints of [0, 1], one at a time. When equipped with the diagonal translation action, the join G ∗ G becomes a locally trivial compact Hausdorff principal G-bundle. We may iterate THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 3 this construction to obtain the multi-joins EnG := G ∗ G ∗ . . . ∗ G (cid:124) (cid:123)(cid:122) n+1 (cid:125) for any natural number n. For example, when G is the two-element group, the multi- join EnG is identified with the n-sphere with the antipodal action. These multi- joins provide the desired equivalent characterization of local triviality: a compact Hausdorff principal bundle is locally trivial if and only if it admits an equivariant continuous map into such a multi-join. This fact is proved by a partition-of-unity argument. Observe that this equivalent characterization is global in nature, in the sense that it avoids the need to talk about open neighborhoods. Therefore, it serves as a perfect point of departure in our quest to generalize local triviality to the noncommutative setting. In addition, it turns out that the smallest n needed for the existence of an equivariant continuous map into EnG is exactly one less than the smallest number of trivializable open sets needed to cover the compact Hausdorff principal bundle. As we shall see below, it is particularly meaningful to keep track of this number n as a measurement of the complexity of the principal bundle. This is our local- triviality dimension in the classical setting1. Thus in summary, the local triviality of a compact Hausdorff principal bundle is characterized by the finiteness of its local-triviality dimension. To generalize this dimension concept to the noncommutative setting, we just need to invoke the Gelfand -- Naimark duality: the local-triviality dimension of an action by a compact quantum group G on a unital C-algebra A is the smallest number n such that there exists an equivariant ∗-homomorphism from the C-algebra of the "n-th multi-join" of G into A. There are, however, some subtleties in constructing multi- joins for compact quantum groups. A naive approach is to dualize the construction in the classical setting and define the C-algebra of the join of G with itself as a C- subalgebra of C(G) ⊗ C(G) ⊗ C([0, 1]), where ⊗ denotes the minimal or maximal tensor product of C-algebras. Unfortunately, unlike classical groups, there is in general no well-defined diagonal action by G on C(G) ⊗ C(G). There are two ways to fix this issue: (1) Replacing the tensor product C(G) ⊗ C(G) by an amalgamated free product C(G)∗CC(G). This results in a construction that we call the free noncommu- tative join, denoted by C(G)+×C(G) and, for the iterated version, C(E+× G). Using this as the C-algebra of the "n-th multi-join" in the above statement leads to our definition of the local-triviality dimension. (2) Replacing the tensor product C(G) ⊗ C(G) by a braided tensor product, which admits a diagonal action. Equivalently, we may "twist" the above "naive join" by altering one of the two endpoints on [0, 1] in a suitable way and replace the diagonal action by the action on the second tensor factor alone, which is well defined. Either way, the resulting C-algebra is called ∆(cid:126) C(G) and, for the the equivariant noncommutative join, denoted by C(G) 1The terms G-index and Schwarz genus have also been used for this number in the literature. n 4 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU G), where ∆ stands for the comultiplication in C(G). iterated version, C(E∆ n Using this in place of the free noncommutative join, we arrive at the definition of the strong local-triviality dimension, so named because it assumes greater values than the local-triviality dimension in general. Both approaches generalize the local-triviality dimension in the classical setting, though they differ in general, even when the acting group is classical. An example is given by the antipodal action on the free spheres, which may be identified with the free noncommutative multi-joins C(E+× n G) for G being the two-element group. Compared to the strong local-triviality dimension, there seem to be fewer tools for giving lower bounds of the local-triviality dimension, as it turns out all the free noncommutative multi-joins, for n > 1, has the same (equivariant) K-theory as the complex numbers. This K-theoretic computation generalizes a result of Nagy on the free 1-sphere [37]. We remark that there is an equivalent definition of the local-triviality dimension using order zero maps, in a way similar to how the Rokhlin dimension is defined [26]. This reformulation gives rise to another variant called the weak local-triviality dimension. On the other hand, the Rokhlin dimension may also be reformulated in terms of the free noncommutative multi-joins. Thus the aforementioned K-theoretic computation may be seen as a reason why we have currently very few means to give lower bounds for the Rokhlin dimension. Moreover, from this perspective, the strong local-triviality dimension is analogous to the Rokhlin dimension with commuting towers (for which we do have obstructions that help give lower bounds). See Section 5.2 for more on these connections between the dimensions as well as a somewhat surprising computation of the Rokhlin dimension for actions by p-adic groups on commutative C-algebras. As an evidence that our notion of local triviality dimension behaves in a desired manner, we prove that actions with finite local-triviality dimension are automatically free. We also illustrate our definition by calculating with a variety of examples. A major motivation for our definition of local triviality is the Borsuk -- Ulam- type conjecture. Recall that the classical Borsuk -- Ulam theorem says there is no antipodal-equivariant continuous map from an (n + 1)-sphere to an n-sphere, for any n. Using the join construction, this theorem is reformulated as: for any n, there is no equivariant continuous from EnG ∗ G to EnG, where G is the two-element group. Analogous statements have been proved for other compact groups G. In [3], Baum, D (cid:44)abrowski, and Hajac stated a noncommutative Borsuk -- Ulam-type conjecture: for any action δ of a compact quantum group G on a unital C-algebra A, there is no equivariant ∗-homomorphism from A δ(cid:126) denotes an equivariant noncommutative join as above. This conjecture, if true even just in the commutative setting, would imply a weak version of the Hilbert-Smith conjecture in topology. On the other hand, progress in the noncommutative setting would give us new insight into the complexity of quantum principal bundles. In fact, from our perspective, this conjecture points to the key difficulty in giving lower bounds on the (strong) local-triviality dimension. δ(cid:126)C(G) to A, where THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 5 As an application of our local-triviality dimension, we prove the noncommutative Borsuk -- Ulam-type conjecture under the assumption that a compact quantum group admits a non-trivial classical subgroup whose induced action has finite local-triviality dimension. The idea is that the finiteness of the dimension allows us to reduce the problem to one where the compact quantum group G is replaced by its classical subgroup H and A is replaced by EnH, but this latter case is already resolved by classical methods. Convention: All tensor products of C-algebras are assumed to be minimal. All C-algebras and ∗-homomorphisms are assumed to be unital unless otherwise stated. 2. Different types of equivariant noncommutative joins 2.1. Preliminaries on joins and actions. To begin with, we recall the definition of a join (e.g., see [25, Chapter 0])) of two topological spaces. Definition 2.1. Let X and Y be topological spaces. The topological join X ∗ Y of X and Y is defined as the following quotient X ∗ Y := (X × Y × [0, 1]) / ∼ , where the equivalence relation is given by (x, y, 1) ∼ (x(cid:48), y, 1) (x, y, 0) ∼ (x, y(cid:48), 0) for all x, x(cid:48) ∈ X, y ∈ Y, for all x ∈ X, y, y(cid:48) ∈ Y. The topological join construction is associative. It is also functorial in the fol- lowing sense: given two continuous maps f : X → W and g : Y → Z between topological spaces, there exists a continuous map f ∗ g : X ∗ Y → W ∗ Z. If X and Y are equipped with a continuous free action of a topological group G, then the diagonal action on the join is again free. This action is also continuous if G is locally compact Hausdorff by a classical result due to Whitehead [55, Lemma 4] (see [32] for a formulation of the result using modern terminology). This turns X∗Y into a free G-space. Recall that, for a locally compact Hausdorff group G, the space G ∗ G is the first step in the Milnor construction [33] of a universal principal G-bundle EG with its base space being a model of the classifying space BG (it is true for an arbitrary group G if, instead of the quotient topology, we put the Milnor topology [33] on the join). We introduce a concise notation for the multi-join of a topological group G: E0G := G, EnG := G ∗ . . . ∗ G , n > 0. (cid:124) (cid:123)(cid:122) n+1 (cid:125) Definition 2.2. Let X be a topological space. The unreduced cone of X is defined as the quotient CX := (X × [0, 1]) / ∼ , where the equivalence relation is given by (x, 0) ∼ (x(cid:48), 0) for all x, x(cid:48) ∈ X. 6 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU We can view the join of X and Y as a subspace of the product of cones: X ∗ Y ∼= {([(t, x)], [(s, y)]) ∈ CX × CY t + s = 1} ⊆ CX × CY. (2.1) Let now X and Y be compact Hausdorff spaces. One can describe the unital com- mutative C-algebras C(CX) and C(X ∗ Y ) of continuous complex-valued functions on the unreduced cone and the join respectively, in the following way: C(CX) ∼= {f ∈ C([0, 1], C(X)) f (0) ∈ C}, C(X ∗ Y ) ∼= {f ∈ C([0, 1], C(X) ⊗ C(Y )) f (0) ∈ C(X) ⊗ C, f (1) ∈ C ⊗ C(Y )}. In the spirit of the celebrated Gelfand -- Naimark theorem [22, Lemma 1] and [21, Theorems 8', Theorem 10], one can think of unital commutative C-algebras as being equivalent to compact Hausdorff topological spaces. Then, the study of noncommu- tative C-algebras can be viewed as noncommutative topology. Having this in mind, we now recall the unreduced cone and the generalization of the join construction to the noncommutative setting. Definition 2.3. Let A be a unital C-algebra. The noncommutative unreduced cone of A is defined as follows CA := {f ∈ C([0, 1], A) f (0) ∈ C}. Definition 2.4. Let A and B be unital C-algebras. One defines the noncommuta- tive join A (cid:126) B of A and B as follows A (cid:126) B := {f ∈ C([0, 1], A ⊗ B) f (0) ∈ A ⊗ C, f (1) ∈ C ⊗ B}. The noncommutative join construction is associative and functorial. Indeed, given two ∗-homomorphisms ϕ : A → C and ψ : B → D of unital C-algebras, there exists a ∗-homomorphism ϕ (cid:126) ψ : A (cid:126) B → C (cid:126) D. As in the case of topological spaces, we introduce group symmetry into the picture. An action of a topological group G on a C-algebra A is a jointly continuous group homomorphism α : G → Aut(A) and any C-algebra equipped with an action of a group G is called a G-C-algebra. We also use the following notation AG := {a ∈ A ∀g ∈ G : αg(a) = a} for the fixed-point subalgebra of A. Notice that, for a compact Hausdorff space X, there is an isomorphism C(X)G ∼= C(X/G). One can verify that if A and B are two G-C-algebras, then A (cid:126) B is again a G-C-algebra with the diagonal action of G defined using the functoriality of the join. In the same way as C-algebras generalize topological spaces, quantum groups generalize topological groups. Let us now proceed to actions of compact quantum groups on unital C-algebras. First, we recall basic definitions. Definition 2.5 ([57]). A compact quantum group G is a unital C-algebra C(G) together with a unital injective ∗-homomorphism ∆ : C(G) → C(G) ⊗ C(G) that is THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 7 coassociative, i.e., (id ⊗ ∆)∆ = (∆ ⊗ id)∆, and such that the two-sided cancellation property holds {(a ⊗ 1)∆(b) a, b ∈ C(G)}cls = C(G) ⊗ C(G) = {(1 ⊗ a)∆(b) a, b ∈ C(G)}cls, where cls denotes the closed linear span. If G is a compact Hausdorff group, then C(G) is a compact quantum group. This example is the reason for the suggestive notation C(G) for the C-algebra of the compact quantum group G. The group C-algebra C∗ r (Γ) of a discrete group Γ is another example. Here the coproduct is given by ∆(χg) = χg ⊗ χg , where χg is the generating unitary corresponding to g ∈ Γ. Note that we assume that the coproduct ∆ is injective. However, to the best of our knowledge, there is no proof of its injectivity nor an example of a compact quantum group with a non-injective coproduct. Definition 2.6. Let G be a compact quantum group and let A be a unital C-algebra. A unital injective ∗-homomorphism δ : A → A ⊗ C(G) is called a coaction of C(G) on A (or an action of G on A) if and only if (1) (δ ⊗ id) ◦ δ = (id ⊗ ∆) ◦ δ (coassociativity), (2) {(1 ⊗ h)δ(a) a ∈ A, h ∈ C(G)}cls = A ⊗ C(G) (counitality). A C-algebra A equipped with a coaction of C(G) is called a G-C-algebra. In the case of a coaction, the fixed-point subalgebra is defined as follows G A := {a ∈ A δ(a) = a ⊗ 1}. Similarly as for the coproduct, we assume that the coaction δ is injective, but here the situation differs as there are examples of non-injective coactions (see, e.g. [51, Proposition 4.1]). Nevertheless, such coactions in the classical case would correspond to actions in which the neutral element of the group does not act as the identity. To exclude these examples, one introduces minimal reduced coactions [51]. Suppose that δ : A → A⊗ C(G) is a non-injective action of a compact quantum group G on a unital C-algebra A. One defines the minimal reduced coaction by δ : A → A⊗C(G), where A := A/ ker δ. This coaction is well defined (since ker δ is G-invariant) and injective due to the injectivity of the coproduct [51, Theorem 3.3]. Throughout the paper, by a coaction we always mean the minimal reduced coaction. Definition 2.7 ([18]). Let δ : A → A ⊗ C(G) be an action of G on A. We say that δ is free if and only if {(a ⊗ 1C(G))δ(a) a ∈ A}cls = A ⊗ C(G). (2.2) A basic example of a free action in the above sense is the canonical translation action given by the coproduct ∆ : C(G) → C(G) ⊗ C(G). Then the Ellwood con- dition for ∆ is satisfied due to the left-sided cancellation property. One can show that (2.7) generalizes free actions of groups on spaces (e.g., see [18, Theorem 2.9]). Let us now recall the notion of an equivariant noncommutative join of C-algebras that plays a crucial role in the noncommutative Borsuk -- Ulam-type conjecture [3]. We shall return to this conjecture in Section 6. 8 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU Definition 2.8 ([15]). Let A be a unital G-C-algebra with a coaction δ : A → A ⊗ C(G). The equivariant noncommutative join A (cid:126)δ C(G) of A and C(G) is defined as follows δ(cid:126) C(G) := {f ∈ C([0, 1], A ⊗ C(G)) f (0) ∈ δ(A), f (1) ∈ C ⊗ C(G)}. A The above type of a join was introduced as a remedy for the lack of diagonal actions of compact quantum groups on tensor products of C-algebras. As an alter- native, one could follow [39] and use braided tensor products. In [4, Corollary 5.6], it was shown that, given a unital free G-C-algebra A, an analog of the diagonal action on A (cid:126)δ C(G) is again free. Let n be a nonnegative integer. In analogy with the classical case, we fix the following notation: C(E∆ 0 G) := C(G), C(E∆ n G) := C(G) (cid:124) ∆(cid:126) . . . (cid:123)(cid:122) n+1 ∆(cid:126) C(G) (cid:125) . 2.2. The free noncommutative join. In line with the above constructions, we introduce a new type of a noncommutative join of C-algebras using the notion of the amalgamated free product A ∗D B (e.g., see [6, Definition II.8.3.5]) of unital C-algebras A and B over their unital ∗-subalgebra D. This new type of a join will be used to build an n-universal locally trivial quantum principal bundle in Section 3. Definition 2.9. Let A and B be unital C-algebras. The free noncommutative join of A and B is defined by A+×B := {f ∈ C([0, 1], A ∗C B) : f (0) ∈ A, f (1) ∈ B}. The above construction is associative and functorial by the associativity and uni- versality of the amalgamated free product respectively. Notice that, if A and B are G-C-algebras with coactions δA and δB respectively, we can define a diagonal coaction δA∗CB. Indeed, let ιA and ιB be the inclusions of A and B in A ∗C B respectively. We define the following ∗-homomorphisms (ιA ⊗ id) ◦ δA : A −→ (A ∗C B) ⊗ C(G), (ιB ⊗ id) ◦ δB : B −→ (A ∗C B) ⊗ C(G). Then, by the universal property of the amalgamated free product, we obtain the ∗-homomorphism δA∗CB : (A ∗C B) −→ (A ∗C B) ⊗ C(G). It is straightforward to check that δA∗CB satisfy the coassociativity and counitality conditions. In general, this coaction might not be injective. In such cases we always consider the minimal reduced coaction. Using the above coaction, one can also define a (minimal reduced) diagonal coaction of C(G) on A+×B. We introduce the following notation: C(E+× 0 G) := C(G), C(E+× n G) := C(G)+× . . . +×C(G) , n ∈ N \ {0}. (cid:124) (cid:123)(cid:122) n+1 (cid:125) THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 9 Let us now prove an analogous result to (2.1) for the free noncommutative join. Since we were not able to find it in the literature, we first provide a proof in the case of the ordinary noncommutative join. Theorem 2.10. Let A and B be unital C-algebras. There is an isomoprhism A (cid:126) B ∼= (CA ⊗ CB) /(cid:104)t1 ⊗ 1 + 1 ⊗ t2 − 1 ⊗ 1(cid:105), (2.3) where t1 and t2 denote the inclusion of the half-open interval (0, 1] into C in CA and CB respectively. Proof. Using Definition 2.3, we obtain an isomorphism CA ⊗ CB ∼= {f ∈ C([0, 1]2, A ⊗ B) f (t1, 0) ∈ A ⊗ C, f (0, t2) ∈ C ⊗ B}. Observe that CA⊗CB has a natural structure of a C(X)-algebra, where X = [0, 1]2, in the sense of [27, Definition 1.5]. Consider the ideal I = (cid:104)F(cid:105) = (cid:104)t1 ⊗ 1 + 1 ⊗ t2 − 1 ⊗ 1(cid:105). Function F vanishes exactly at the points of the standard one-dimensional simplex ∆1 = {(s, t) ∈ [0, 1]2 s + t = 1}. Therefore, we have that (CA ⊗ CB)/I ∼= {f ∈ C(∆1, A ⊗ B) f (t1, 0) ∈ A ⊗ C, f (0, t2) ∈ C ⊗ B} ∼= {f ∈ C([0, 1], A ⊗ B) f (0) ∈ A ⊗ C, f (1) ∈ C ⊗ B} = A (cid:126) B. (cid:4) Next, we prove a similar result for the free join A+×B. In fact we prove even more, namely that A+×B is also isomorphic to the amalgamated free product CA∗C([0,1])CB by means of the unital ∗-homomorphisms given on generators by f1 : C([0, 1]) → CA : t (cid:55)→ 1 − t1, (2.4) Here t is the inclusion function from (0, 1] into C generating C0((0, 1]) as a C-algebra and ti , i = 1, 2, are defined as in Theorem 2.10. f2 : C([0, 1]) → CB : t (cid:55)→ t2. Theorem 2.11. Let A and B be unital C-algebras. We have the following isomor- phisms A+×B ∼= (CA ∗C CB) /(cid:104)t1 + t2 − 1(cid:105) ∼= CA ∗C([0,1]) CB. Proof. First, note that the second isomorphism follows from the universal properties of CA∗CCB and CA∗C([0,1])CB. Therefore, it suffices to show that A+×B is isomorphic to CA ∗C([0,1]) CB. Define two ∗-homomorphisms ψ1 : CA → A+×B, (ψ1(a))(t) := a(1 − t), ψ2 : CB → A+×B, (ψ1(b))(t) := b(t). Let f1 and f2 be defined as in (2.4). Then, ψ1 ◦ f1 = ψ2 ◦ f2, and by the universal property of CA∗C([0,1])CB, the maps ψ1 and ψ2 give rise to a unital ∗-homomorphism ψ : CA ∗C([0,1]) CB → A+×B. 10 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU We will show that the above map is an isomorphism. For this purpose, observe that both the domain and the codomain of ψ can be viewed as C(X)-algebras (see [27, Definition 1.5]), where X = [0, 1]. Indeed, if we denote by ι and ι(cid:48) the inclusions of C([0, 1]) in CA∗C([0,1])CB and A+×B respectively, their images land in the centers, namely ι(C([0, 1])) ⊆ Z(CA ∗C([0,1]) CB), ι(cid:48)(C([0, 1])) ⊆ Z(A+×B). (2.5) Any isomorphism of C(X)-algebras over the same space is given by an isomor- phism of the C-algebras over the fibers via the fiber-preserving map. Let us now describe the fibers of the C(X)-algebras under consideration. Fix a t0 ∈ [0, 1]. One defines It0 := ι(C0([0, 1] \ {t0})) ·(cid:0)CA ∗C([0,1]) CB(cid:1) (cid:0)CA ∗C([0,1]) CB(cid:1)(cid:12)(cid:12)t0 :=(cid:0)CA ∗C([0,1]) CB(cid:1) /It0 . A ∼= CA1−t0 ∗C CBt0 ∼= (cid:0)CA ∗C([0,1]) CB(cid:1)(cid:12)(cid:12)t0 t0 = 0 A ∗C B t0 ∈ (0, 1) B . t0 = 1 t0 (using the inclusion ι(cid:48)) and (A+×B)t0 . It is straight- This quotient is well defined because of the first inclusion in (2.5). Taking advantage of (2.4), we observe that and subsequently Similarly, one can define I(cid:48) forward to see that ∼= {f ∈ C([0, 1], A ∗C B) f (0) ∈ A, f (1) ∈ B, f (t0) = 0} I(cid:48) t0 and subsequently (A+×B)t0 ∼= t0 = 1 Now, ψ induces a ∗-homomorphism between fibers A t0 = 0 A ∗C B t0 ∈ (0, 1) B . ψt0 : CA1−t0 ∗C CBt0 → (A+×B)t0 and it suffices to show that ψt0 is an isomorphism for every t0 ∈ [0, 1]. This can be (cid:4) done using a standard partition-of-unity argument for [0, 1]. Example 2.12 (Free spheres). The n-dimensional free sphere C(Sn +) is defined as the unital universal C-algebra generated by n + 1 elements x0, x1, . . ., xn subject to relations x∗ i = xi, i = 0, . . . , n, x2 i = 1. n(cid:88) Our aim is to show that any free noncommutative sphere is isomorphic to an iterated free noncommutative join of C(Z/2Z), namely that Z/2Z). +) ∼= C(E+× C(Sn n i=0 THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 11 By Theorem 2.11, we have that C(E+× (cid:125) Z/2Z) ∼= (CC(Z/2Z) ∗C . . . ∗C CC(Z/2Z ) (cid:123)(cid:122) (cid:124) )/I, n+1 where I is the ideal generated by 1 −(cid:80)n n the half-open interval (0, 1] into C in a different copy of CC(Z/2Z). Note that i=0 ti. Here each ti denotes the inclusion of CC(Z/2Z) ∼= C(CZ/2Z) ∼= C([−1, 1]), where we view C([−1, 1]) as the unital universal algebra generated by one self-adjoint element p such that (cid:107)p(cid:107) ≤ 1. There is a Z/2Z-action on C([−1, 1]) given by p (cid:55)→ −p. If we denote by t the inclusion of the half-open interval (0, 1] into C in CC(Z/2Z), then p2 = t. We have a natural well-defined Z/2Z-equivariant ∗-homomorphism +) → (C([−1, 1])∗Cn) /(cid:10)Σn i=0 p2 i − 1(cid:11) , Φ : C(Sn xi (cid:55)→ pi, i = 0, 1, . . . , n, where each pi is the generator of the ith copy of C([−1, 1]) in the iterated amalga- mated free product. By the universality of the free product, we find the inverse of (cid:5) the above map and we conclude the isomorphism. We end this section by showing that the K-theory of any free noncommutative join is the same as that of C. Theorem 2.13. Let A and B be unital C-algebras. The embedding C (cid:44)→ A+×B induces isomorphisms of the K-theory groups: K0(A+×B) = Z, K1(A+×B) = 0, where K0(A+×B) is generated by [1]. Proof. Without loss of generality, we restrict to separable C-algebras, since every C-algebra is a direct limit of separable C-algebras and K-theory functor is con- tinuous with respect to direct limits. Recall that, by Theorem 2.11, we have the isomorphism A+×B ∼= CA ∗C([0,1]) CB. By [53, Theorem 6.4], there is the following six-term exact sequence: / K0(A+×B) / K0(CA) ⊕ K0(CB) K0(C([0, 1])) ((f1)∗,(f2)∗) K1(A+×B) K1(CA) ⊕ K1(CB) K1(C([0, 1])) from which the statement follows, since K0(C([0, 1])) = K0(CA) = K0(CB) = Z (generated by [1]) and K1(C([0, 1])) = K1(CA) = K1(CB) = 0 for any A and B. (cid:4) Let us remark, that the above result is also true at the level of the equivariant K-theory. As a corollary, we obtain the K-theory of free noncommutative spheres C(Sn +). / / O O o o o o 12 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU Corollary 2.14. Let C(Sn +) be the noncommutative free sphere. We have that K0(C(Sn +)) = Z, K1(C(Sn +)) = 0, where K0(C(Sn +)) is generated by [1]. As far as the authors know, for n > 1 (see [37] for n = 1), this result was not stated in the literature before. 3. Locally trivial compact quantum principal bundles In this section, we introduce three different incarnations of the local-triviality dimension for actions of compact quantum groups on unital C-algebras and intro- duce the concept of a locally trivial compact quantum principal bundle. We also prove that local triviality in that sense implies freeness of the action and introduce n-universal locally trivial compact quantum principal bundles. Throughout this section t denotes the inclusion function from the half-open in- terval (0, 1] into C, which generates C0((0, 1]) as a C-algebra. 3.1. The (weak) local-triviality dimension. Definition 3.1. Let A be a unital C-algebra equipped with an action δ of a compact quantum group G. If d is the minimal nonnegative integer such that there exist G-equivariant ∗-homomorphisms ρ0, . . . , ρd : C0((0, 1]) ⊗ C(G) → A satisfying the condition that: (1) (cid:80)d (2) (cid:80)d j=0 ρj(t ⊗ 1) is invertible, we say that δ has the weak local-triviality di- mension d, written dimWLT(δ) = d, and we set dimWLT(δ) = ∞ if no such d exists, j=0 ρj(t ⊗ 1) = 1 (joint-unitality), we say that δ has the local-triviality dimension d, written dimLT(δ) = d, and we set dimLT(δ) = ∞ if no such d exists. Notation 3.2. If there is no ambiguity about the G-action on A, we will also denote (cid:5) the (weak) local-triviality dimension by dim G (W)LT(A). It is immediate from Definition 3.1 that for any coaction δ, we have that (3.1) Next, let A and B be G-C-algebras with actions δA and δB respectively. If there exists a G-equivariant ∗-homomorphism A → B, then dimWLT(δ) ≤ dimLT(δ) . dimWLT(δA) ≥ dimWLT(δB), dimLT(δA) ≥ dimLT(δB). (3.2) The above property of the local-triviality dimensions is useful for Borsuk -- Ulam-type problems (see Section 6), where one usually tries to establish nonexistence of certain equivariant maps. Moreover, if A is a unital G-C-algebra with an action δA and I is a G-invariant ideal, i.e. δA(I) ⊆ I ⊗ C(G), then it follows from (3.2) that dimWLT(δA/I) ≤ dimWLT(δ), dimLT(δA/I) ≤ dimLT(δ). (3.3) THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 13 Here δA/I is the action induced on the quotient by δA. Remark 3.3. Few comments regarding the minimal reduced coactions are in order. Recall that if δA : A → A⊗ C(G) is not injective we use δA : A → A⊗ C(G) instead, where A = A/ ker δ. Adapting our definition of the local-triviality dimension to non-injective coactions, by (3.3), we obtain dimLT(δA) ≤ dimLT(δA). Hence considering the minimal coaction does not change the finiteness of the local- triviality dimension. Next, let B be another G-C-algebra with a non-injective coaction δB and suppose that there is a G-equivariant ∗-homomorphism A → B. Then, there is a G-equivariant ∗-homomorphism A → B, so that an analog of (3.2) (cid:5) is satisfied. We list some elementary facts about the local-triviality dimensions. Proposition 3.4. Let δ : A → A ⊗ C(G) be an action of a compact quantum group G on a unital C-algebra A. Then dimLT(δ) = 0 if and only if there exists a G-equivariant -homomorphism C(G) → A. Proof. Let dimLT(δ) = 0, so we have a ∗-G-homomorphism ρ : C0((0, 1])⊗C(G) → A such that ρ(t ⊗ 1) = 1. Consider the map ϕ : C(G) −→ A : h (cid:55)−→ ρ(t ⊗ h). The above assignment is clearly unital, G-equivariant and respects the ∗-structure. Moreover, since √ ϕ(ab) = ρ(t ⊗ ab) = ρ( √ t ⊗ a)ρ(t ⊗ 1)ρ( t ⊗ b) = ρ(t ⊗ a)ρ(t ⊗ b) = ϕ(a)ϕ(b), (cid:4) it is also an algebra homomorphism. The other implication is analogous. Proposition 3.5. Let δ : A → A ⊗ C(G) be an action of a compact quantum group G on a unital C-algebra A. Then we have that dimWLT(δ) = 0 ⇐⇒ dimLT(δ) = 0. Proof. For the not immediate implication, let dimWLT(δ) = 0, so that we have a G-equivariant -homomorphism ρ : C0((0, 1]) ⊗ C(G) → A, such that ρ(t ⊗ 1) is invertible. Every element of C0((0, 1]) ⊗ C(G) can be approximated by linear combinations of elements of the form tk ⊗ h, where k ∈ N\{0} and h ∈ C(G). Now consider a map (cid:101)ρ : C0((0, 1]) ⊗ C(G) → A defined by (cid:101)ρ(tk ⊗ h) = ρ( √ √ t ⊗ 1)−kρ(t ⊗ h)ρ( t ⊗ 1)−k (cid:101)ρ(t ⊗ g)(cid:101)ρ(t ⊗ h) = ρ( for any k ∈ N \ {0} and h ∈ C(G). Since ρ(t ⊗ 1) is invertible and G-invariant, this map is well defined and G-equivariant. For any g, h ∈ C(G), we have that √ √ √ t ⊗ 1)−1ρ(t ⊗ h)ρ( t ⊗ 1)−1ρ( t ⊗ 1)−1ρ(t ⊗ g)ρ( √ √ √ t ⊗ 1)−1ρ( t ⊗ g)ρ( t ⊗ h)ρ( t ⊗ 1)−1 √ t ⊗ 1)−2 t ⊗ h)−2ρ(t2 ⊗ gh)ρ( √ √ = ρ( √ = ρ( t ⊗ 1)−1 =(cid:101)ρ(t2 ⊗ gh). 14 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU Hence, (cid:101)ρ can be extended to a G-equivariant ∗-homomorphism One can verify that (cid:101)ρ(t ⊗ 1) = 1 and therefore dimLT(δ) = 0. (cid:101)ρ : C0((0, 1]) ⊗ C(G) → A. (cid:4) Corollary 3.6. Let δ : A → A ⊗ C(G) be an action of a compact quantum group G on a unital C-algebra A. Then dimLT(δ) = 1 implies that dimWLT(δ) = 1. Proof. Let dimLT(δ) = 1. This implies that dimWLT(δ) ≤ 1 by (3.1). Since dimLT(δ) (cid:54)= 0 implies that dimWLT(δ) (cid:54)= 0 by Proposition 3.5, the claim follows. (cid:4) Proposition 3.7. Let δ : A → A ⊗ C(G) be an action of a compact quantum group j=0 ρj(t ⊗ 1) commutes with the images of the maps ρj for all j = 0, 1, . . . , d. Then, dimLT(δ) = d. j=0 ρj(t ⊗ 1). Similarly as in the G on a unital C-algebra A. Suppose that dimWLT(δ) = d and that (cid:80)d Proof. Suppose that dimWLT(δ) = d and let e :=(cid:80)d proof of Proposition 3.5, we define the maps and, since e is invertible, G-invariant, and it commutes with the images of the maps ρj, we extend each (cid:101)ρi to a G-equivariant ∗-homomorphism i = 0, 1, . . . , d, It only remains to check that the joint-unitality condition is satisfied: (cid:101)ρi(tk ⊗ h) := ρi(tk ⊗ h)e−k, (cid:101)ρi : C0((0, 1]) ⊗ C(G) → A, d(cid:88) (cid:101)ρi(t ⊗ 1) = d(cid:88) i = 0, 1, . . . , d. ρi(t ⊗ 1)e−1 = ee−1 = 1. i=0 i=0 (cid:4) Next, we show that finiteness of the weak local-triviality dimension implies free- ness in the sense of Definition 2.7, in complete generality. Theorem 3.8. Let G be a compact quantum group, let A be a unital C-algebra, and let δ be an action of G on A. If dimWLT(δ) < ∞, then δ is free. Proof. Set B = {(A ⊗ 1C(G))δ(A)}cls. To show that B = A ⊗ C(G), it is enough to show that B contains all simple tensors. In addition, since B is a left (A ⊗ 1C(G))- module, it suffices to show that 1A ⊗ x belongs to B for all x ∈ C(G). Let x ∈ C(G) be fixed, and set d = dimWLT(δ) < ∞. Let ε > 0. We write f ≈ε g to mean that (cid:107)f − g(cid:107) < ε. Since the action ∆ : C(G) → C(G) ⊗ C(G) is free, there are m ∈ N and y1, . . . , ym, z1, . . . , zm ∈ C(G) such that m(cid:88) 1C(G) ⊗ x ≈ε (yk ⊗ 1C(G))∆(zk). Using the fact that dimWLT(δ) = d, we find G-equivariant ∗-homomorphisms k=1 ρ0, . . . , ρd : C0((0, 1]) ⊗ C(G) → A THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 15 such that the element e := (cid:80)d (cid:101)ρj = ρj⊗idC(G) : C0((0, 1])⊗C(G)⊗C(G) → A⊗C(G), which are also G-equivariant j=0 ρj(t ⊗ 1C(G)) is invertible. For j = 0, . . . , d, set ∗-homomorphisms. Then, (cid:101)ρj(t ⊗ 1C(G) ⊗ x) e ⊗ x = j=0 j=0 d(cid:88) ρj(t ⊗ 1C(G)) ⊗ x = d(cid:88) d(cid:88) m(cid:88) (cid:101)ρj(t ⊗ (yk ⊗ 1C(G))∆(zk)) d(cid:88) m(cid:88) t ⊗ (yk ⊗ 1C(G)))(cid:101)ρj( (cid:101)ρj( m(cid:88) d(cid:88) √ t ⊗ yk) ⊗ 1C(G))δ(ρj( √ (ρj( √ √ j=0 k=1 j=0 k=1 ≈ε = = t ⊗ ∆(zk)) t ⊗ zk)), j=0 k=1 where we used the equivariance of the maps ρj. well. Hence B = A ⊗ C(G) and we conclude that δ is free. This shows that e ⊗ x belongs to B. Since e is invertible, 1A ⊗ x belongs to B as (cid:4) Next, we introduce the notion of a locally trivial compact quantum principal bundle. Let δ : A → A ⊗ C(G) be a free action of a compact quantum group G , G) is called a compact quan- on a unital C-algebra A. Then the triple (A, A tum principal bundle (cf. [3, Definition 3.1]). Hence, by Theorem 3.8, any action of a compact quantum group G with finite local-triviality dimension gives rise to a compact quantum principal bundle. We arrive at the following definition, which is a noncommutative analog of a locally trivial compact principal bundle (see Sec- tion 4). G Definition 3.9. A compact quantum principal bundle (A, A trivial if and only if dim LT(A) < ∞. G G , G) is said to be locally A compact quantum principal bundle (A, A , G) is called trivializable if there exists a G-equivariant ∗-homomorphism C(G) → A [3, Definition 3.1]. Hence, by Proposition 3.4, (A, A , G) is trivializable if and only if dim G LT(A) = 0. G G 3.2. Examples of locally trivial compact quantum principal bundles. We start with antipodal Z/2Z-actions on two different kinds of noncommutative spheres and show that they have finite local-triviality dimension. Example 3.10 (Antipodal action on the free spheres). There is a natural antipodal action of Z/2Z on C(Sn +) (see Example 2.12) given on generators by xi (cid:55)→ −xi, i = 0, 1, 2, . . . , n. Recall that C(Z/2Z) can be viewed as the universal C-algebra generated by a single self-adjoint element γ such that γ2 = 1. One can find equivariant ∗-homomorphisms ϕi : C0((0, 1]) ⊗ C(Z/2Z) → C(Sn +) : t ⊗ γ (cid:55)→ xi, i = 0, 1, 2, . . . , n. √ 16 Observe that E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU n(cid:88) ϕi(t ⊗ 1) = n(cid:88) (cid:16) √ t ⊗ γ) ϕi( i=0 i=0 The above considerations imply that dim Z/2Z LT (C(Sn (cid:17)2 n(cid:88) +)) ≤ n. i=0 = x2 i = 1. (cid:5) Example 3.11 (Antipodal action on the equatorial Podle´s sphere). Recall that the equatorial Podle´s sphere C(S2 q∞) [47] is the unital universal C-algebra generated by B and self-adjoint A satisfying the relations B∗B = 1 − A2, BB∗ = 1 − q4A2. AB = q2BA, The antipodal Z/2Z-action is defined by A (cid:55)→ −A, Define Z/2Z-equivariant ∗-homomorphisms B (cid:55)→ −B. ρ0 : C0((0, 1]) ⊗ C(Z/2Z) → C(S2 ρ1 : C0((0, 1]) ⊗ C(Z/2Z) → C(S2 q ) : ρ2 : C0((0, 1]) ⊗ C(Z/2Z) → C(S2 q ) : q ) : √ t ⊗ γ (cid:55)→ i 2 √ t ⊗ γ (cid:55)→ 1 2 √ t ⊗ γ (cid:55)→ (cid:114) (B∗ − B), (B∗ + B), 1 + q4 2 A. Using the defining relations of C(S2 q∞), one can check that 2(cid:88) i=0 ρi(t ⊗ 1) = (cid:105)2 √ ρi( t ⊗ γ) (cid:104) 2(cid:88) q∞)) ≤ 2. i=0 B∗B + = 1 2 1 2 BB∗ + 1 + q4 2 A2 = 1. Hence, dim Z/2Z LT (C(S2 (cid:5) Note that finiteness of the local-triviality dimension of any Z/2Z-action on a uni- tal C-algebra A is tantamount to the existence of finitely many odd self-adjoint elements in A whose squares add up to one. We will show in Section 6, that dim Z/2Z LT (C(Sn In the next example we consider a U (1)-action for which we not only bound the local-triviality dimension but also obtain its actual value. Then, we present an example of an SU (2)-action of a similar flavour. +)) = n for all n. Example 3.12 (Noncommutative Matsumoto -- Hopf fibration). Let us consider the θ [30, 31], where θ ∈ (0, 1). It is defined Matsumoto noncommutative three-sphere S3 as the universal C-algebra generated by two normal elements Z and W subject to relations Define the action of U (1) on C(S3 ZW = e2πiθW Z, θ ) by W (cid:55)→ eiϕW, Z (cid:55)→ eiϕZ, Z∗Z + W ∗W = 1. for all eiϕ ∈ U (1). Since Matsumoto showed that C(S3 this example is sometimes called the noncommutative Matsumoto -- Hopf fibration (or the noncommutative Matsumoto -- Dirac monopole bundle). θ )U (1) ∼= C(S2) [31], the action considered in THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 17 θ ) : θ ) : Recall that C(U (1)) can be viewed as the universal C-algebra generated by one unitary U . Define two U (1)-equivariant ∗-homomorphisms Assume now that dimU (1) ρ0 : C0((0, 1]) ⊗ C(U (1)) → C(S3 ρ1 : C0((0, 1]) ⊗ C(U (1)) → C(S3 √ t ⊗ U (cid:55)→ Z, √ t ⊗ U (cid:55)→ W. Note that ρ0(t ⊗ 1) + ρ1(t ⊗ 1) = 1, and hence dimU (1) θ )) ≤ 1. θ )) = 0. By Proposition 3.4, there is a U (1)- equivariant ∗-homomorphism C(U (1)) → C(S3 θ ). This map is in particular Z/2Z- equivariant and the image of U under this map would contradict [41, Proposition 3.9] (which states that no element of C(S3 θ ) can be both odd and invertible). Hence, we (cid:5) have that dimU (1) Similarly as for the case of Z/2Z, the finiteness of the local-triviality dimension of any U (1)-action on a C-algebra A can be stated in a different way. Recall that LT (A) < ∞ can be any U (1)-action gives a Z-grading on A. Then the fact that dimU (1) translated into existence of finitely many normal elements xi of degree 1 in A such LT (C(S3 LT (C(S3 θ )) = 1. LT (C(S3 that(cid:80) i x∗ i xi = 1. Example 3.13 (Noncommutative Hopf principal SU (2)-bundle). Let θ = (θij) be an n × n Hermitian matrix such that θij = 1 for all i, j and with θii = 1 for each i. The C-algebra C(S2n−1 [38] is defined as the universal unital C-algebra generated by n normal elements Z1, . . . , Zn satisfying the relations ) of the odd Natsume -- Olsen quantum sphere S2n−1 θ θ n(cid:88) Z∗ i Zi = 1, ZjZi = θijZiZj, i, j = 1, 2, . . . , n. i=1 Recall that C(SU (2)) can be viewed as the unital universal C-algebra generated by two normal elements α and γ commuting with each other and satisfying α∗α + γ∗γ = 1. We write the coproduct of C(SU (2)) on generators as follows ∆(α) = α ⊗ α − γ ⊗ α∗, ∆(γ) = α ⊗ γ + γ ⊗ γ∗. In [28], Landi and Suijlekom consider C(S7 θ13 = θ24, together with an action of SU (2) given by θ ), with θ12 = θ34 = 1 and θ14 = θ23 = δ(Z1) = Z1 ⊗ α − Z2 ⊗ α∗, δ(Z3) = Z3 ⊗ α − Z4 ⊗ α∗, δ(Z2) = Z1 ⊗ γ + Z2 ⊗ γ∗, δ(Z4) = Z3 ⊗ γ + Z4 ⊗ γ∗. Note that this coaction is well defined because Z1 commutes with Z2 and Z3 com- mutes with Z4. We define two SU (2)-equivariant -homomorphisms ρ0 : C0((0, 1]) ⊗ C(G) → C(S7 ρ1 : C0((0, 1]) ⊗ C(G) → C(S7 θ ) : θ ) : t ⊗ α (cid:55)→ Z1, t ⊗ α (cid:55)→ Z3, √ √ √ √ t ⊗ γ (cid:55)→ Z2, t ⊗ γ (cid:55)→ Z4. 18 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU These maps are well defined because all generators are normal. Observe that 4(cid:88) (ρ0 + ρ1)(t ⊗ 1) = (ρ0 + ρ1)(t ⊗ (α∗α + γ∗γ)) = Z∗ i Zi = 1. Hence, dimSU (2) LT (C(S7 θ )) ≤ 1. i=1 (cid:5) Next example shows that both dimensions differ in general. Example 3.14 (Z/2Z-action on 3 × 3 complex matrices). Let us view M3(C) as a graph C-algebra of the graph v1 v3 v2 e12 e23 (e.g., see [2] for a definition of a graph C-algebra). This C-algebra is generated by three orthogonal projections p1, p2, p3, associated to vertices, and two partial isometries s12 and s23, associated to edges, satisfying s12s∗ s∗ 23s23 = p3. 23, s12s23, s∗ 12, s∗ We consider a Z/2Z-action given on the generators by It has the following linear basis: p1, p2, p3, s12, s23, s∗ s∗ 12s12 = p2 = s23s∗ 23, 12 = p1, 23s∗ 12. pi (cid:55)→ pi, i = 1, 2, 3, s12 (cid:55)→ −s12, s23 (cid:55)→ −s23. The above action gives a Z/2Z-grading and any odd self-adjoint element in M3(C) is of the form 12 + k∗ 1s∗ x = k1s12 + k2s23 + k∗ 2s∗ 23, k1, k2 ∈ C. that(cid:80)n j=0 x2 Suppose that we have odd self-adjoint elements xj ∈ M3(C), j = 0, 1, 2, . . . , n, such j = 1. Then the relation p1 + p2 + p3 = 1 (sum of the vertex projections equals 1 for any graph C-algebra of a graph with finitely many vertices) leads to a contradiction. Hence we get that dim It is equally straightforward to find two odd self-adjoint elements of M3(C) whose sum of squares is invertible. Since there is no Z/2Z-equivariant ∗-homomorphism (cid:5) C(Z/2Z) → M3(C), we conclude that dim 3.3. The n-universal bundles and the strong local-triviality dimension. Motivated by [52, § 19.2], we introduce the following definition. LT (M3(C)) = ∞. Z/2Z Z/2Z WLT(M3(C)) = 1. G Definition 3.15. We say that a locally trivial compact quantum principal bundle , G) is n-universal if and only if for any other compact quantum principal (A, A LT(B) ≤ n there exists a G-equivariant ∗-homomorphism bundle (B, B A → B. , G) with dim G G finitely many odd self-adjoint elements yi such that(cid:80) Recall that, in the case of Z/2Z-actions, to check if a given Z/2Z-C-algebra A gives rise to a locally trivial compact quantum principal bundle one needs to find i = 1. This in turn means that there is a unital Z/2Z-equivariant ∗-homomorphism C(Sn +) → A, for some n, given by xi (cid:55)→ yi. Hence C(Sn +) is an n-universal compact quantum Z/2Z-bundle. i y2 THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 19 n n +) and C(E+× In Section 2.9, using the notion of the free noncommutative join, we introduced G) for any compact quantum group G. We also showed that Z/2Z) are isomorphic as Z/2Z-C-algebras. In Theorem 3.19 we G) is an n-universal G-bundle for any G, but first we need the the C-algebra C(E+× C(Sn will prove that C(E+× following lemma. Lemma 3.16. Let A be a G-C-algebra with a coaction δ : A → A ⊗ C(G). Then, LT(A) + 1, where we consider A+×C(G) with the diagonal dim action of G. LT(A+×C(G)) ≤ dim G G n LT(A) = n. We have jointly-unital G-equivariant ∗-homomorphisms Proof. Let dim ρi : C0((0, 1])⊗ C(G) → A, i = 0, 1, . . . , n. Let us define the following G-equivariant G ∗-homomorphisms(cid:101)ρi : C0((0, 1]) ⊗ C(G) → A+×C(G), (cid:101)ρi(t ⊗ h)(s) := (1 − s)ρi(t ⊗ h), (cid:103)ρn+1(t ⊗ h)(s) := s · h, i(cid:101)ρi(t ⊗ 1) = 1, and hence we conclude that LT(A+×C(G)) ≤ n + 1 = dim It is evident that(cid:80) h ∈ C(G), h ∈ C(G), G LT(A) + 1. i = 0, 1, . . . , n, s ∈ [0, 1]. dim G i = 0, 1, . . . , n + 1, s ∈ [0, 1], (cid:4) Corollary 3.17. Let n be a nonnegative integer and let G be a compact quantum group with the coproduct ∆ : C(G) → C(G) ⊗ C(G). Then, dim G)) ≤ n LT(C(E+× G for the diagonal action of G on C(E+× G). n n Proof. As dimLT(∆) = 0, one can proceed by induction using Lemma 3.16. Example 3.18 (Compact quantum matrix groups). Let G be a compact quantum matrix group and let the matrix (ujk) ∈ Mn(C(G)) be its fundamental representa- tion (see [57]). Then we can explicitly write down the maps from Definition 3.1 for C(E+× G). Indeed, recall that by Theorem 2.11 we have that d (cid:4) (3.4) C(E+× (cid:125) G) ∼= (CC(G) ∗C . . . ∗C CC(G) (cid:123)(cid:122) (cid:124) d+1 )/I, where I is the ideal generated by 1 −(cid:80)d d i=0 ti . Here each ti denotes the inclusion of the half-open interval (0, 1] into C in a different copy of CC(G). Define the following G-equivariant ∗-homomorphisms ρi : C0((0, 1]) ⊗ C(G) → C(E+× t ⊗ ujk (cid:55)→(cid:2)√ i = 0, 1, . . . , d, ti ⊗ ujk (cid:3) , G) : √ d where [·] denotes the class of an element of CC(G)∗C d in (CC(G)∗C d)/I. We slightly abuse notation by denoting the generators in different copies of the iterated amal- gamated free product by the same ujk. One only needs to check the joint-unitality condition: 20 d(cid:88) i=0 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU u∗ k1uk1 i=0 ρi (cid:32) t ⊗ n(cid:88) d(cid:88) d(cid:88) n(cid:88) (cid:2)√ (cid:35) (cid:34) d(cid:88) ti ⊗ 1C(G) k=1 k=1 i=0 ti ⊗ u∗ = ρi t ⊗ u∗ (cid:33) (cid:17) (cid:16)√ d(cid:88) n(cid:88) (cid:34) d(cid:88) ti ⊗ n(cid:88) (cid:3)(cid:2)√ (cid:3) = (cid:3) = 1C(E+× =(cid:2)1C(G) ⊗ 1C(G) ti ⊗ uk1 G). k=1 k=1 i=0 i=0 k1 k1 d ρi(t ⊗ 1C(G)) = = = (cid:17) ρi t ⊗ uk1 (cid:16)√ (cid:35) u∗ k1uk1 i=0 G) → A. In other words, C(E+× (cid:5) Here we used the fact that (ujk) is a unitary matrix. Theorem 3.19. Let A be a G-C-algebra with a coaction δ : A → A⊗ C(G). Then dimLT(δ) ≤ n if and only if there exists a G-equivariant unital ∗-homomorphism C(E+× G) is an n-universal compact quantum princi- pal G-bundle. Proof. First suppose that there is a G-equviariant ∗-homomorphism C(E+× By (3.2) and Corollary 3.17, we have that dim G) → A. LT(A) ≤ n. There exist G-equivariant ∗-homomorphisms LT(A) ≤ dim Now suppose that dim LT(C(E+× G G)) ≤ n. G G n n n n n(cid:88) ρi : C0((0, 1]) ⊗ C(G) → A, i = 0, 1, . . . , n, with ρi(t ⊗ 1) = 1. i=0 The unitizations ρ+ product) give rise to a G-equivariant ∗-homomorphism i , i = 0, 1, . . . , n, of the above maps (by functoriality of the free ρ : CC(G) ∗C . . . ∗C CC(G) (cid:124) generated by the element (cid:80) Using Theorem 2.11, it suffices to check if ρ descends to the quotient by the ideal i ti − 1. This is however a consequence of the joint- (cid:4) → A. (cid:123)(cid:122) (cid:125) unitality condition. n+1 Theorem 3.19 motivates another definition, where the free noncommutative join is replaced by the equivariant noncommutative join. Definition 3.20. Let A a unital G-C-algebra with a coaction δ : A → A ⊗ C(G). Given a nonnegative integer d, we say that δ has the strong local-triviality dimension at most d, written dimSLT(δ) ≤ d, if there exists a G-equivariant ∗-homomorphism ρ : C(E∆ d We set dimSLT(δ) = ∞ if no such d exists. G) −→ A. Since C(E∆ 0 G) = C(G) and due to Proposition 3.5, for any coaction δ, we obtain dimWLT(δ) = 0 ⇐⇒ dimLT(δ) = 0 ⇐⇒ dimSLT(δ) = 0. (3.5) Using the same formulas as in the proof of Lemma 3.16, one can verify the fol- lowing result (see [13, Proposition 3.4]). THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 21 Proposition 3.21. Let A be a G-C-algebra with a coaction δ : A → A ⊗ C(G). δ(cid:126) C(G) with the Then, dim G)) ≤ n for all n. analog of the diagonal action of G. Consequently, dim G LT(C(E∆ n Using the above proposition and the inequality (3.2), for any coaction δ, we obtain δ(cid:126) C(G)) ≤ dim G LT(A) + 1, where we consider A G LT(A inequalities between all three local-triviality dimensions dimWLT(δ) ≤ dimLT(δ) ≤ dimSLT(δ). (3.6) Next example shows that the strong local-triviality dimension differs from the previously defined local-triviality dimensions. Example 3.22 (Antipodal action on odd free spheres). Consider the Z/2Z-action +)) ≥ 2. on C(S1 +)) ≤ 1, then we would obtain a Z/2Z- Indeed, if we would have that dim equivariant ∗-homomorphism +) described in Example 3.10. We will show that dim Z/2Z SLT (C(S1 Z/2Z SLT (C(S1 f : C(S1) ∼= C(E1Z/2Z) ∼= C(E∆ Z/2Z) −→ C(S1 +) −→ C(S1), 1 where the second arrow is simply the abelianization. Next, since K1(C(S1 the induced map +)) = 0, f∗ : K1(C(S1)) −→ K1(C(S1)) is trivial. This contradicts the classical Borsuk -- Ulam theorem (see [40, Theo- (cid:5) rem 3.1.1]). Similarly, if n is odd, we have that dim +)) ≥ n + 1. Z/2Z SLT (C(Sn 4. Locally trivial compact principal bundles In this section we prove that for unital commutative C-algebras our definition of the locally trivial compact quantum principal bundle recovers the notion of the locally trivial compact principal bundle in topology. First, let us recall the definition of a locally trivial principal G-bundle, which is the main motivation of this work. Definition 4.1. Let π : X → M be a fiber bundle of topological spaces and let X be equipped with a right action of a topological group G. The quadruple (X, M, π, G) is called a principal G-bundle if the following axioms hold: (1) For any x ∈ X and g ∈ G, we have π(xg) = π(x). (2) For each m ∈ M , there exists an open neighbourhood U of m in M and a fiber-preserving G-equivariant homeomorphism ϕ : π−1(U ) → U × G, which is called a trivialization of π over U with a typical fibre G. Point (2) above describes the local triviality of a principal G-bundle. From (1) and (2) one can prove that the action of G on X is free. On the other hand, by a result of Mostow [34, Theorem 3.1], free actions of compact Lie groups on regular topological spaces give rise to locally trivial principal bundles. We gave a standard definition of a principal G-bundle that can be found in text- books, e.g. [52, 54], and should be contrasted with the definition of a Cartan prin- cipal G-bundle [5, 11], where one assumes that the action of G is free and proper and that M ∼= X/G (in the original work of Cartan, it is phrased equivalently as the continuity of the translation map and a certain density condition) instead of (2). 22 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU For our purposes, we need a slight reformulation of the notion of local triviality. This can be achieved by means of certain invariants that are always finite for locally trivial compact principal G-bundles. Definition 4.2 ([50]). The Schwarz genus of a G-space X, denoted by gG(X), is the smallest number n such that X can be covered with open G-invariant subsets U0, U1, . . . , Un with the property that for every i = 0, 1, 2, . . . , n, there exists a G- equivariant map Ui → G. If no such n exists, we write gG(X) = ∞. Now let X and Y be two G-spaces and suppose that there exists a G-map X → Y . Then, there is an inequality gG(X) ≤ gG(Y ). (4.1) Note that if G is a compact Hausdorff group acting on a compact Hausdorff space X, then gG(X) < ∞ if and only if π : X → X/G is a locally trivial principal G- bundle. One can say even more, if gG(X) = n, this means that X/G can be covered with at most n + 1 trivializing open sets. Recall that gG(X) ≤ dim(X/G). (4.2) Indeed, dim(X/G) ≤ d if and only if any finite open cover of X/G can be refined by a finite open cover that splits into (d + 1) disjoint families of open sets. Note that open subsets of a trivializing open set are still trivializing, and thus if a finite open cover of X/G consists of trivializing open sets, then any finite open cover that refines it also consists of trivializing open sets. Combining this with the observation that a disjoint union of trivializing open sets is still a trivializing open set, we obtain the desired inequality. In [50, Theorem 9] Schwarz showed that for any topological group G there is an inequality gG(EnG) ≤ n. (4.3) We need another invariant which was introduced for purposes of the Borsuk -- Ulam-type theorems (e.g., see [29]). Definition 4.3. Let X be a G-space. We define the G-index of X by indG(X) := min{n : ∃ G-map X → EnG}. If there is no such G-map, we write indG(X) = ∞. As for the Schwarz genus, we have that if there is an G-equivariant map X → Y between two G-spaces, then indG(X) ≤ indG(Y ). (4.4) The following result shows that the above invariants are equal for compact Haus- dorff spaces. As we did not find it in the literature, we give its proof for reader's convenience. Proposition 4.4. Let G be a compact Hausdorff group acting continuously on a compact Hausdorff G-space X. Then, gG(X) = indG(X). THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 23 Proof. First assume that indG(X) = n. We have a G-map X → EnG. Using (4.1) and (4.3), we get that gG(X) ≤ gG(EnG) ≤ n = indG(X). φi : Ui = π−1(Vi) → G for some trivializing cover {Vi}n of unity subordinate to {Vi}n Now assume that gG(X) = n. Then, we know that there are G-equivariant maps i=0 be a partition i=0. Let {fi}n i=0. We define a G-map ψ : X → EnG , f (π(x))φi(x) , defined. Indeed, the condition (cid:80)n where we use the simplicial notation for the multi-join. The above G-map is well i=0 fi(m) = 1, for every m ∈ X/G, assures that the image of ψ lands in ∆n × Gn+1, where ∆n is the n-simplex. Equations (4.4) and (4.3), and the existence of ψ imply that indG(X) ≤ indG(EnG) ≤ n = gG(X). (cid:4) i=0 x (cid:55)→ n(cid:88) We emphasize that Proposition 4.4 shows that, for compact Hausdorff X and G, indG(X) < ∞ if and only if the G-bundle π : X → X/G is locally trivial. Thus, we obtain a different characterization of locally trivial compact principal G-bundles. Next, we show that for unital commutative C-algebras, i.e. algebras of complex- valued continuous functions on compact Hausdorff topological spaces, Definition 3.1 recovers the usual notion of local triviality. Theorem 4.5. Let G be a compact Hausdorff group, let X be a compact Hausdorff space, and let G act continuously on X. Denote by α : G → Aut(C(X)) the induced action. Then, (4.5) Proof. Let dimLT(α) = n. Take equivariant ∗-homomorphisms from Definition 3.1 and denote by indG(X) = dimLT(α). i=0 ρi(t ⊗ 1) = 1. We dualize the above to obtain i = 0, 1, . . . , n, ρi : (C0((0, 1]) ⊗ C(G))+ ∼= C(CG) → C(X), their unitizations. Note that (cid:80)n G-equivariant continuous maps ψ : X → EnG, ρi : X → CG, Next, we define a continuous G-map x (cid:55)→ n(cid:88) si ∈ [0, 1] and gi ∈ G. We only need to check if (cid:80)n n(cid:88) n(cid:88) n(cid:88) non-zero): i=0 (ρi(t ⊗ 1))(x) = 1 = i=0 Hence, indG(X) ≤ n = dimLT(α). i = 0, 1, . . . , n. ρi(x), where we used simplicial coordinates for EnG. Every element ρi(x) is a class in [0, 1] × G. Let us fix an arbitrary point x ∈ X and let ρi(x) := [(si, gi)] for some i=0 si = 1 (with each si being (t ⊗ 1)(ρi(x)) = (t ⊗ 1)([(si, gi)]) = si . i=0 i=0 i=0 n(cid:88) 24 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU Now suppose that indG(X) = n. We have a G-equivariant map ψ : X → EnG. Define G-equivariant continuous maps ψ→ EnG ρi : X ι (cid:44)→ (CG)n+1 pri+1(cid:16) CG . i=0 si = 1 and using the above calculation one can show that(cid:80)n (cid:80)n where i = 0, 1, . . . , n. Again each element ρi(x) is a class in [0, 1] × G. Let us denote it as previously by ρi(x) := [(si, gi)], where x ∈ X is arbitrary. By definition i=0 ρi(t ⊗ 1) = 1. This implies that dimLT(α) ≤ n = indG(X). (cid:4) Remark 4.6. Observe that for any compact Hausdorff group G and any compact Hausdorff G-space X we obtain dimG WLT(C(X)) = dimG LT(C(X)) = dimG SLT(C(X)). However, in the previous section we showed that all the dimensions can take different (cid:5) values in general. such that(cid:80) We end this section by some remarks of C-algebraic flavour. Let G be a compact Hausdorff group, X a compact Hausdorff space, and π : X → X/G be a compact principal G-bundle with the local-triviality dimension equal to n. Then we have G-equivariant ∗-homomorphisms ρi : C0((0, 1]) ⊗ C(G) → C(X), for i = 0, . . . , n, i ρi(t ⊗ 1) = 1. Now let pi := ρi(t ⊗ 1) ∈ C(X). Each pi is a positive Ai := {√ contractive element. We consider C-subalgebras of C(X) of the form pi f ∈ C(X/G)}cls ⊆ C(X/G), i = 0, ..., n. √ pif Since C(X) is commutative, each Ai is an ideal of functions supported on an open set Ui ⊆ X/G over which the principal bundle X → X/G is trivial. By the joint-unitality condition, we have that 1 ∈ A0 + A1 + . . . + An. Next, each ρi induces a G-equivariant unital ∗-homomorphism where M(Ai) is the multiplier algebra of Ai. ψi : C(G) → M(Ai), Remark 4.7. Note that one could consider the hereditary C-subalgebras Ai for actions of compact quantum groups on arbitrary unital C-algebras with finite local- triviality dimension. However, further investigations are needed to establish if they play a similar role to ideals of functions supported on the open sets constituting (cid:5) a trivializing cover. 5. Relations with piecewise triviality and the Rokhlin dimension In this section, we examine the connection between the local-triviality dimension and some other related notions, i.e. piecewise triviality and the Rokhlin dimension. THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 25 5.1. Piecewise triviality. We start by exploring the connection of the local-triviality dimension with piecewise triviality. We recall the definition of piecewise triviality in the classical context. Definition 5.1. A Cartan principal bundle (X, π, M, G) is called piecewise trivial, if there exist a covering of M by finitely many closed sets V0, . . . , Vn and fibre- preserving G-equivariant homomorphisms χi : π−1(Vi) → Vi × G, i = 0, 1, . . . , n. This definition was introduced in [5] along with an example (the bubble space) of a Cartan principal G-bundle that is piecewise trivial, but not locally trivial. For compact Hausdorff spaces, local triviality implies piecewise triviality. Indeed, for the cover {Vi} in the above definition, take the supports of functions of a partition of unity subordinate to the open trivializing cover given by local triviality. The concept of piecewise triviality admits a straightforward generalization to the realm of noncommutative geometry. Definition 5.2. Let A be a unital G-C-algebra, where G is a compact quantum group. An action of G on A is said to be piecewise trivial [24], if for some n ∈ N i=0 Ii = 0, and unital G-equivariant ∗-homomorphisms χi : C(G) → A/Ii. there exist G-invariant closed ideals I0, . . . , In of A, such that(cid:84)n Note that according to this definition, if a simple G-C-algebra is piecewise trivial it is in fact trivializable. In contrast with the classical case, local triviality in the sense of Definition 3.1 does not imply piecewise triviality, as the latter notion requires the existence of proper ideals, while the former may be even applied to simple algebras, as the next example shows. Example 5.3. (A Z/2Z-action on the irrational rotation algebra). Let θ ∈ (0, 1) be an irrational number. The irrational rotation algebra Aθ (or the noncommutative torus; see [17, 44, 49]) is the universal C-algebra generated by two unitaries U and V subject to the relation U V = e2πiθV U. This simple C-algebra plays a fundamental role in noncommutative geometry. We define an involutive automorphism of Aθ by mapping U (cid:55)→ −U and V (cid:55)→ V. This gives us an action of Z/2Z on Aθ. Note that the subalgebra C∗(U ) gener- ated by U is isomorphic to C(S1) and invariant under the above action, and that the restricted action amounts to the antipodal action on S1, whose local trivial- ity dimension is 1. Hence, applying inequality (3.2) to the equivariant embedding C∗(U ) ⊂ Aθ, we see that dim LT (Aθ) ≤ 1. Z/2Z However, this action cannot be piecewise trivial, because simplicity of Aθ would We show that the considered bundle is not trivial and thus dim force it to be trivial, which we will show is not possible. Z/2Z LT (Aθ) = 1. Indeed, if this were not the case, we would have a unital Z/2Z-equivariant ∗-homo- morphism ϕ : C(Z/2Z) → Aθ. Let p and q be the images under ϕ of the two minimal 26 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU projections generating C(Z/2Z). Thus p and q are orthogonal projections that add up to 1 and they are translates of each other under the Z/2Z action. The unique trace τ : Aθ → C given by (cid:32)(cid:88) m,n∈Z τ (cid:33) am,nU mV n = a0,0 is clearly invariant under Z/2Z-action on Aθ. We compute τ (p) = τ (q) = τ (p + q) 2 = . 1 2 This is impossible since we know the image of the homomorphism K0(Aθ) → R induced by τ is Z + θZ (this is in fact also injective; see [44]). Alternatively, we can prove this fact by noticing that the involution U (cid:55)→ −U and V (cid:55)→ V is homotopic to the identity via the homotopy defined by U (cid:55)→ eπitU and V (cid:55)→ V . This implies that p and q induce the same element in K0(Aθ) and 2[p] = [1], which contradicts with the fact that [1] is not divisible by 2 in K0(Aθ). We have proved that the noncommutative bundle under consideration is not triv- (cid:5) ial. Thus we conclude that dim Z/2Z LT (Aθ) = 1. 5.2. Rokhlin dimension. We proceed to the relation of the local-triviality dimen- sion with the Rokhlin dimension. Throughout the subsection all C-algebras are assumed to be separable and groups are assumed to be metrizable unless otherwise stated. Let us start by recalling the definitions of a sequence algebra and a com- pletely positive contractive order zero map, which are the basic ingredients of the definition of the Rokhlin dimension. Definition 5.4. Let A be a separable unital C-algebra, (cid:96)∞(N, A) denote the C- algebra of all bounded sequences with elements in A and c0(N, A) denote the ideal consisting of sequences converging to zero in norm. The sequence algebra is defined as the quotient The central sequence algebra is defined as the commutant A∞ ∩ A(cid:48). A∞ := (cid:96)∞(N, A)/c0(N, A). If G is a compact metrizable group acting on a separable unital C-algebra A, then there are actions of G on both A∞ and A∞∩A(cid:48). The continuity of those actions is a consequence of a result of Brown [8, Theorem 2]. Definition 5.5 ([56]). Let A and B be C-algebras. A completely positive contractive map ϕ : A → B is called order zero if and only if ϕ(a)ϕ(b) = 0, whenever ab = 0, for any a, b ∈ A. The theory of completely positive contractive order zero maps was developed by Winter and Zacharias and has played a fundamental role in the recent breakthrough in the classification theory of C-algebras. The following result, based on the Stine- spring theorem, will be crucial in exploring the connection of the Rokhlin dimension with local triviality. THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 27 Theorem 5.6 ([56]). Let A and B be unital C-algebras. Any completely positive contractive order zero map ϕ : A → B induces the ∗-homomorphism ρϕ : C0((0, 1]) ⊗ A → B determined by ρϕ(t ⊗ a) := ϕ(a) for all a ∈ A. Conversely, any ∗-homomorphism ρ : C0((0, 1]) ⊗ A → B induces the completely positive order zero map ϕρ : A → B given by ϕρ(a) := ρ(t ⊗ a) for all a ∈ A. The above is also true in the equivariant setting for actions of locally compact group on C-algebras (see [19, Corollary 2.10]). Note that to discuss order zero maps we need not restrict to separable C-algebras. Definition 3.1 was inspired by the following definition. Definition 5.7 ([19]). Let G be a compact metrizable group and let δ : G → Aut(A) be an action of G on a separable unital C-algebra A. We say that an action δ has the Rokhlin dimension n, written dimRok(δ) = n, if n is the minimal non-negative integer such that there exist G-equivariant completely positive contractive order zero maps ϕ0, . . . , ϕn : C(G) → A∞ ∩ A(cid:48) We set dimRok(δ) = ∞ if no such n exists. with n(cid:88) i=0 ϕi(1) = 1. Using Theorem 5.6, one can compare the local-triviality dimension and the Rokhlin dimension. Note that the local-triviality dimension is defined for actions of compact quantum groups to begin with, while the generalization of the Rokhlin dimension to actions of compact quantum groups is not straightforward and requires some reformulations [20]. Remark 5.8. Observe the original definition of the Rokhlin dimension [19, Defini- tion 3.2] is slightly more general: G is assumed to be compact and second countable, while A is assumed to be σ-unital. Furthermore, the (corrected) relative central se- quence algebra is used instead of A∞ ∩ A(cid:48). Note however that we do not use the (cid:5) invariant part A∞,δ of the sequence algebra due to [8, Theorem 2]. Next, using the notion of equivariant projectivity, we show that, for compact Lie group actions on unital commutative separable C-algebras, the notions of the local-triviality dimension and the Rokhlin dimension coincide. Let us first state the definition of the equivariant projectivity in the case of com- pact Hausdorff group actions. Definition 5.9 ([45, 46]). Let G be a compact Hausdorff group and let A be a G-C- algebra. We say that A is G-equivariantly projective if for any G-C-algebra B, a G-invariant closed ideal J ⊆ B, and an equivariant ∗-homomorphism σ : A → B/J, there is an equivariant ∗-homomorphism λ : A → B such that π ◦ λ = σ, where π : B → B/J is the quotient map. 28 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU Note that the above definition means projectivity in the category of general G-C-algebras and one can restrict this definition to subcategories of unital G- C-algebras, commutative G-C-algebras, etc. In the category of commutative C- algebras, an object C0(X) is G-equivariantly projective if and only if X is a G-AR (G-equivariant absolute retract [35, 36]). Remark 5.10. The notion of equivariant projectivity of [45, 46] should be distin- guished from the notion of equivariant projectivity in the category of projective (cid:5) modules that can be applied to C-algebras [9]. We state two propositions that establish a relation between the local-triviality dimension and the Rokhlin dimension. Proposition 5.11. Let A be a unital commutative separable C-algebra equipped with an action δ of a compact metrizable group G. Then, dimRok(δ) ≤ dimLT(δ). (cid:80)d Proof. Assume that dimLT(δ) = d. Due to Theorem 5.6, we have G-equivariant completely positive contractive order zero maps ϕ0, . . . , ϕd : C(G) → A such that i=0 ϕi(1) = 1. Using the unital inclusion ι : A → A∞ and the fact that A is commutative, we obtain G-equivariant completely positive contractive order zero Since ι is unital, we obtain(cid:80)d maps i=0(cid:101)ϕi(1) = 1. Therefore, dimRok(δ) ≤ d = dimLT(δ). (cid:101)ϕi := ι ◦ ϕi : C(G) → A∞ = A∞ ∩ A(cid:48), (cid:4) Proposition 5.12. Let G be a compact metrizable group such that C0((0, 1])⊗C(G) is G-equivariantly projective and let δ be an action of G on a unital separable C- algebra A. Then, i = 0, 1, 2, . . . , d. dimWLT(δ) ≤ dimRok(δ). Proof. Suppose that dimRok(δ) = d and we have equivariant completely positive contractive order zero maps ϕi : C(G) → A∞ ∩ A(cid:48) i = 0, 1, . . . , d. Using Theorem 5.6, we obtain equivariant ∗-homomorphisms ρi : C0((0, 1]) ⊗ C(G) → A∞ ∩ A(cid:48) (cid:44)→ A∞ . for Since C0((0, 1]) ⊗ C(G) is G-equivariantly projective, for each ρi, there exists a G-equivariant ∗-homomorphism λi : C0((0, 1]) ⊗ C(G) → (cid:96)∞(N, A) such that π ◦ λi = ρi, where π : (cid:96)∞(N, A) → A∞ is the quotient map. Define prn : (cid:96)∞(N, A) → A as a projection on the nth element of the sequence for some n ∈ N. Then, the maps (cid:101)ρn,i := prn ◦ λi : C0((0, 1]) ⊗ C(G) → A, i = 0, 1, 2, . . . , d, define G-equivariant ∗-homomorphisms. THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 29 i ρi(t ⊗ 1) = 1 and that ρi = π ◦ λi, we obtain From the fact that(cid:80) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)prn (cid:32) d(cid:88) d(cid:88) i=0 (cid:33) λi(t ⊗ 1) − 1 (cid:101)ρN,i(t ⊗ 1) = prN (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) → 0 as n → ∞. (cid:33) (cid:32) d(cid:88) λi(t ⊗ 1) Hence for large enough N , we can conclude that i=0 i=0 is invertible. (cid:4) If G is a compact Lie group, then the space (0, 1] × G is a G-AR (see [1, Corol- lary 2.3]). Hence, C0((0, 1]) ⊗ C(G) is G-equivariantly projective in the category In the general possibly noncommutative case, it is of commutative C-algebras. known that C0((0, 1]) ⊗ C(Z/2nZ) is equivariantly projective for any n ∈ N \ {0} [46, Proposition 2.10]. Combining Propositions 5.2 and 5.12, and taking advantage of the above remark about Lie groups, we arrive at Theorem 5.13. Let A be a unital commutative separable C-algebra equipped with an action δ of a compact Lie group G. Then, dimLT(δ) = dimRok(δ). Theorem 5.13 suggests that the Rokhlin dimension can be viewed as another noncommutative generalization of local triviality of compact principal G-bundles, where G is a compact Lie group. However, when we stay away from Lie groups, these notions differ even in the classical case as shown in the Theorem 5.17 below. First, we need to introduce a new characterization of the Rokhlin dimension for actions of compact metrizable groups on unital separable C-algebras. To this end, let us recall that every compact metrizable group G contains a decreasing sequence of normal subgroups (cid:92){Ni : i = 1, 2, . . .} = {e} , G = N1 ⊃ N2 ⊃ . . . with such that, for every i, Hi := G/Ni is a compact Lie group. Thus G = lim←− Hi. (See for example [48, Theorem 53]). Let δ denote the action of a compact metrizable group G on a unital C-algebra A. For any normal subgroup Ni ⊆ G as above, we can define the action δi of Hi on ANi by the formula δi [g]i (a) := δg(ιi(a)), g ∈ G, [g]i ∈ Hi, a ∈ ANi. Here ιi is the inclusion of ANi into A. Theorem 5.14. Let A be a unital separable C-algebra equipped with an action δ of a compact metrizable group G = lim←− Hi = lim←−(G/Ni) and let δi denote the action of Hi on ANi. Then, (cid:8)dimRok(δi)(cid:9) . dimRok(δ) = sup i 30 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU Proof. First note that A = lim−→ ANi and there is an action of G on each ANi defined by the quotient map G → Hi. Then, using [19, Theorem 3.8 (4)], we obtain (cid:8)dimRok(δi)(cid:9) ≤ sup (cid:8)dimRok(δi)(cid:9) . dimRok(δ) ≤ lim inf i i Next, suppose that dimRok(α) = d, namely that there are jointly-unital G-equi- variant completely positive contractive order zero maps ϕj : C(G) −→ A∞ ∩ A(cid:48), j = 0, 1, . . . , d. For every i, there is a G-equivariant ∗-homomorphism C(Hi) −→ C(G), where the action of G on Hi is again defined by the quotient map G → G/Ni. Com- posing the two maps above, for any i , we get jointly unital G-equivariant completely positive contractive order zero maps j : C(Hi) −→ A∞ ∩ A(cid:48), ψi j = 0, 1, . . . , d. Note that, as a G-space, Hi is Ni-invariant so that the image of the above map is contained in the fixed-point subalgebra under the Ni-action. Therefore, for all i, we obtain jointly-unital Hi-equivariant completely positive contractive order zero maps j : C(Hi) −→ (A∞ ∩ A(cid:48))Ni , ψi j = 0, 1, . . . , d. Observe that there is a ∗-homomorphism given by averaging each component over Ni , i.e., S : (A∞)Ni →(cid:0)ANi(cid:1) ∞ (cid:20)(cid:18)(cid:90) (cid:19) (cid:21) S : [(ai)i] (cid:55)→ αn(ai) dµ(n) , Ni i where µ denotes the normalized Haar measure on the compact group Ni. Moreover, it is straightforward to check that (cid:16) (A∞ ∩ A(cid:48))Ni(cid:17) ⊆ (ANi)∞ ∩ (ANi)(cid:48). S Hence, for every i, we obtain jointly unital Hi-equivariant completely positive con- tractive order zero maps j : C(Hi) −→(cid:0)ANi(cid:1) ∞ ∩(cid:0)ANi(cid:1)(cid:48) ψi , j = 0, 1, . . . , d. This implies that, for any i, Subsequently, sup i dimRok(δi) ≤ d = dimRok(δ). (cid:8)dimRok(δi)(cid:9) ≤ dimRok(δ). (cid:4) THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 31 Intuitively speaking, the Rokhlin dimension of an action does not see the small subgroups Ni. Let G = lim←− G/Ni = lim←− Hi be a compact metrizable group and let X be a compact metrizable G-space. For the purposes of the next corollary, Theorem 5.16, and Theorem 5.17, we introduce the notion of the homotopy quotient X×GHi . Let us assume that the G-action on X is on the right and we define a left action of G on Hi using the quotient map as before. Then, there is a diagonal G action on X × Hi and We equip X×GHi with a right Hi-action. Note that there is a isomorphism of Hi-spaces X×GHi := (X × Hi)/G. ∼= X/Ni . Hi X × G Note however that the homotopy quotient construction is much more functorial than the orbit space construction. Corollary 5.15. Let G = lim←− Hi be a compact metrizable group and let X be a compact metrizable G-space. Let δ denote the induced C(G)-coaction on C(X). The following equality holds: (cid:26) (cid:18) (cid:19)(cid:27) X × G Hi . dimRok(δ) = sup i indHi The following result gives a nice upper bound for the Rokhlin dimension in the commutative case. Theorem 5.16. Let X be a compact metrizable space equipped with an action δ of a compact metrizable group G and let dim(X/G) < ∞. Then, dimRok(δ) < ∞ if and only if δ is free. In fact, we have dimRok(δ) ≤ dim(X/G). Proof. We already know that dimRok(δ) < ∞ implies freeness [19, Theorem 4.1(1)]. Conversely, assume that the action of G on X is free. Then, for any Hi, the action of Hi on X×GHi is free as well. By Mostow's theorem [34, Theorem 3.1], we know that any free action of a compact Lie group is locally trivial, so we have indHi(X×GHi) < ∞ for any Hi. By Proposition 4.4 and the inequality (4.2), we have indHi(X×GHi) ≤ dim((X×GHi)/Hi), but the latter space is homeomorphic to X/G for any Hi. Therefore, Theorem 5.14 implies that dimRok(α) ≤ dim(X/G) < ∞. (cid:4) Next, we present the following striking dimension reduction phenomenon. Theorem 5.17. For any compact metrizable space X equipped with a continuous action by the p-adic group Zp. Let δ denote the induced C(G)-coaction on C(X). Then, if dimRok(δ) < ∞, then dimRok(δ) ≤ 3. Proof. For the sake of brevity, we write G := Zp, Hi := Z/piZ and Xi := X/Ni for all i. One can show that (cid:18) (cid:19) X × G Hn ∼= lim←− m≥n Xm × Hm Hn , 32 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU which implies that indHn (cid:19) (cid:26) (cid:18) (cid:19)(cid:27) X × G Hn ≤ min m≥n indHn Hn . Xm × Hm (cid:18) (cid:26) We compute the upper bound for the Rokhlin dimension: (cid:26) (cid:18) Xm × Hm (cid:19)(cid:27) Hn Hn X × (cid:18) (cid:19)(cid:27) (cid:26) k(cid:12)(cid:12) ∃ Xm × (cid:110) k(cid:12)(cid:12) ∃ Xm G Hm ≤ max n min m≥n Hn−→ EkHn Hn indHn (cid:27) (cid:111) Z/pm−→ EkHn . dimRok(δ) = max n indHn = max n min m≥n = max n min m≥n min min Here the last step uses the fact that if H is a quotient group of a compact group G, X is a G-space, and Y is an H-space (and hence also a G-space), then ∃ X G−→ Y if and only if ∃ X×GH H−→ Y . From Theorem 5.13, we know that indHm(Xm) = dimHm Rok(Xm) < ∞, therefore Hm−→ EdHm for some d ≥ 0. The conclusion follows from a result from [13] ∃ Xm which states that for any d ≥ 3 and n ≥ 1 there exists a Z/pn2d−3Z-equivariant map Ed(Z/pn2d−3Z) −→ E3(Z/pnZ). (cid:4) 6. The noncommutative Borsuk -- Ulam-type conjectures In this section, we prove a noncommutative Borsuk -- Ulam-type result for actions of compact quantum groups having a classical subgroup whose induced action has finite local-triviality dimension. This result is an easy consequence of the results presented in Sections 3.1 and 3.3. Let us emphasize that the noncommtuative Borsuk -- Ulam- type conjecture (see Conjecture 6.5 below) was one of the main reasons to introduce the local-triviality dimension, because in the commutative setting, this conjecture is known exactly in the locally trivial case [12]. The original Borsuk -- Ulam antipodal theorem [7, Satz II] can be equivalently formulated in the following way: there is no map Sn+1 → Sn intertwining the antipodal actions. This result was generalized to q-deformed spheres by Yamashita [58, Corollary 15] Matousek generalized the Borsuk -- Ulam theorem as follows: and θ-deformed spheres by Passer [41, Corollary 4.6]. let G be a finite group, then there is no G-equivariant map EnG → En−1G [29, Theorem 6.2.5]. If G = Z/2Z, we recover the usual Borsuk -- Ulam theorem. Chirvasitu, D (cid:44)abrowski and Hajac, based on the unpublished work of Bestvina and Edwards, extended the aforementioned non-existence result to all compact Hausdorff groups [12]. There is a natural question whether this result holds in the case of the n-fold equivariant noncommutative join of a given compact quantum group. THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 33 n G) → C(E∆ G) and dim G LT(C(E∆ n Conjecture 6.1. Let G be a compact quantum group. Then there does not exists a G-equivariant ∗-homomorphism C(E∆ G)) = n. An analogous conjecture for n-fold noncommutative free join of G would follow Indeed, this fact can be deduced from the following result from Conjecture 6.1. which is a direct consequence of the inequality (3.2) and Theorem 3.19. Proposition 6.2. Let G be a compact quantum group and suppose that there exists a single G-action with the local-triviality dimension equal to n. Then there is no G-equivariant ∗-homomorphism C(E+× G)) = n. LT(C(E+× G) and dim G G) → C(E+× n+1 n n+1 n Since dimG LT(C(EnG)) = n for any compact Hausdorff group G [12], we arrive at: Corollary 6.3. Let G be a compact Hausdorff group. Then there is no G-equivariant ∗-homomorphism C(E+× n G) → C(E+× n+1G) and dimG LT(C(E+× n G)) = n. Let us write the Borsuk -- Ulam theorem for free noncommutative spheres as a sep- arate result. Corollary 6.4. There is no ∗-homomorphism C(Sn antipodal actions and dim Z/2Z LT (C(Sn +)) = n. +) → C(Sn+1 + ) intertwining the Baum, D (cid:44)abrowski and Hajac postulated the following noncommutative Borsuk -- Ulam-type conjecture: Conjecture 6.5 ([3]). Let A be a unital C-algebra with a free action δ of a non-trivial compact quantum group G. There does not exist a G-equivariant ∗- homomorphism A → A δ(cid:126) C(G). Using the notion of the local-triviality dimension, we can obtain a new noncom- mutative Borsuk -- Ulam-type result. First, let us demonstrate that Conjecture 6.5 for locally trivial actions on non-simple C-algebras follows from Conjecture 6.1. Again, the notion of the n-universal compact quantum principal bundle is crucial for our arguments. Proposition 6.6. Let δ : A → A⊗C(G) be an action of a compact quantum group G LT(A) < ∞. Then on a unital C-algebra A such that A admits a character and dim Conjecture 6.1 implies Conjecture 6.5. G Proof. Suppose that we know that Conjecture 6.5 holds for A. Using universality of C(E+× G) (Theorem 3.19), we obtain a chain of G-∗-homomorphisms n C(E+× n G) → A → A δ(cid:126) C(G) → . . . → C(E∆ n+1 G). G LT(C(E∆ n G)) = n, which is the exact formulation of Conjecture 6.1. Hence, to prove Conjecture 6.5 for locally trivial actions it is enough to prove that (cid:4) dim Corollary 6.7. Let δ : A → A⊗ C(G) be an action of a compact Hausdorff group G LT(A) < ∞. There on a unital C-algebra A such that A admits a character and dimG does not exist a G-equivariant ∗-homomorphism A → A δ(cid:126) C(G). 34 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU Proof. The result is a consequence of Proposition 6.6, because Conjecture 6.1 is true (cid:4) for compact Hausdorff groups [12]. Finally, we can improve the result in Corollary 6.7 to arbitrary unital C-algebras equipped with an action of a compact quantum group admitting a classical subgroup whose induced action has finite local-triviality dimension: Theorem 6.8. Let G be a compact quantum group, and let A be unital C-algebra equipped with a coaction δ : A → A ⊗ C(G). Then, if G admits a non-trivial classical subgroup H whose induced action α satisfies dimWLT(α) < ∞, there is no G-equivariant ∗-homomorphism A → A δ(cid:126) C(G). Proof. We closely follow the reasoning presented in [42]. Suppose that there exists a G-equivariant ∗-homomorphism A → A δ(cid:126) C(G). Then by [16] there exists a H- equivariant ∗-homomorphism A → A (cid:126) C(H). Evaluation at e ∈ H in each point would produce a path of unital ∗-homomorphisms on A connecting a H-equivariant map to a one-dimensional representation. In what follows, we show that such a path cannot exists. Denote the aforementioned path of unital ∗-homomorphisms by ψt : A → A, where ψ1 is equivariant and ψ0 : A → C. Existence of ψ0 implies that the H-invariant ideal I = (cid:104)ab− ba : a, b ∈ A(cid:105) is proper and we can consider the abelianization Ac := A/I. By the Gelfand-Naimark theorem, Ac = C(X) for some compact Hausdorff space X. Since the action of H on A has finite weak local-triviality dimension, by inequality (3.3), the induced action on Ac has finite weak local-triviality dimension as well. In terms of spaces, this means that the principal H-bundle X → X/H is locally trivial. Now for every t ∈ [0, 1], ψt induces a map (cid:101)ψt : Ac → Ac. Such a path is dual to the homotopy of maps on the space X connecting an H-equivariant map with a constant map. This implies equivariant contractibility of X, which is equivalent with the existence of an H-map X ∗ H → X. However, this map cannot exists (cid:4) by [12]. Acknowledgement Authors are grateful to Ludwik D (cid:44)abrowski for valuable suggestions regarding the exposition of the paper and to G´abor Szab´o for discussions concerning c.p.c. order zero maps. M.T. would also like to thank Ben Passer and Alex Chirvasitu for many conversations about applications of the presented results to the Borsuk -- Ulam-type conjectures. This work is part of the project Quantum Dynamics supported by EU-grant RISE 691246 and Polish Government grant 317281. M.T. was partially supported by the project Diamentowy Grant No. DI2015 006945 financed by the Polish Ministry of Science and Higher Education. THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 35 References [1] S. Antonyan. Universal proper G-spaces. Topology Appl., 117, 1 (2002), 23 -- 43. [2] T. Bates, D. Pask, I. Raeburn, W. Szyma´nski, Wojciech. The C∗-algebras of row-finite graphs. New York J. Math., 6 (2000), 307 -- 324. [3] P. F. Baum, L. D (cid:44)abrowski, P. M. Hajac. Noncommutative Borsuk-Ulam-type conjectures. Banach Center Publications, 106 (2015), 9 -- 18. [4] P. F. Baum, K. de Commer, P. M. Hajac. Free actions of compact quantum groups on unital C∗-algebras. Doc. Math., 22 (2017), 825 -- 849. [5] P. F. Baum, P. M. Hajac, R. Matthes, W. Szyma´nski. Noncommutative geometry approach to principal and associated bundles. arXiv:math/0701033. [6] B. Blackadar. Operator Algebras. Theory of C-Algebras and von Neumann Algebras. Ency- clopaedia of Mathematical Sciences, 122. Operator Algebras and Non-commutative Geometry, III. Springer-Verlag, Berlin, 2006. [7] K. Borsuk. Drei Satze uber die n-dimensionale euklidische Sphare. Fund. Math., 20 (1933), 177 -- 190. [8] L. G. Brown. Continuity of actions of groups and semigroups on Banach spaces. J. London Math. Soc., (2) 62 (2000), no. 1, 107 -- 116. [9] T. Brzezi´nski, P. M. Hajac. Galois-Type Extensions and Equivariant Projectivity. arXiv:0901.0141. [10] R. J. Budzy´nski, W. Kondracki. Quantum Principal Fiber Bundles: Topological Aspects. Rep. Math. Phys., 37 (1996), no. 3, 365 -- 385. [11] H. Cartan, Seminaire Cartan. E.N.S. 1949/1950, reprinted by W.A. Benjamin, INC., New York, Amsterdam, 1967. [12] A. Chirvasitu, L. D (cid:44)abrowski, M. Tobolski. The weak Hilbert -- Smith conjecture from a Borsuk -- Ulam-type conjecture. arXiv:1612.09567 [13] A. Chirvasitu, B. Passer. Invariants in noncommutative dynamics. To appear in J. Funct. Anal. [14] A. Connes. Noncommutative Geometry. Academic Press, San Diego, CA, 1994. [15] L. D (cid:44)abrowski, T. Hadfield, P. M. Hajac. Equivariant Join and Fusion of Noncommutative Algebras. SIGMA, 11 (2015), 082. [16] L. D (cid:44)abrowski, P. M. Hajac, S. Neshveyev. Noncommutative Borsuk-Ulam-type conjectures revisited. arXiv:1611.04130. [17] E. Effros, F. Hahn. Locally compact transformation groups and C-algebras. Bull. Amer. Math. Soc., 73 (1967), 222 -- 226. [18] D. A. Ellwood. A new characterisation of principal actions. J. Funct. Anal., 173, 1 (2000), 49 -- 60. [19] E. Gardella. Rokhlin dimension for compact group actions. Indiana Univ. Math. J., 66, 2 (2017), 659 -- 703. [20] E. Gardella, M. Kalantar, M. Lupini. Rokhlin dimension for compact quantum group actions. arXiv:1709.00222. [21] I. Gelfand. Normierte Ringe. Mat. Sb., 9 (51), 1 (1941). [22] I. Gelfand, M. Naimark. On the imbedding of normed rings into the ring of operators in Hilbert space. Mat. Sb., 12 (54), 2 (1943), 197 -- 217. [23] P. M. Hajac, A. Kaygun, B. Zieli´nski. Finite closed coverings of compact quantum spaces. Banach Center Publ., 98 (2012), 215 -- 237. [24] P. M. Hajac, U. Krahmer, R. Matthes, B. Zieli´nski. Piecewise principal comodule algebras. J. Noncommut. Geom., 5, 4 (2011), 591 -- 614. [25] A. Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002. [26] I. Hirshberg, W. Winter, J. Zacharias. Rokhlin dimension and C-dynamics. Comm. Math. Phys., 335, 2 (2015), 637 -- 670. 36 E. GARDELLA, P. M. HAJAC, M. TOBOLSKI, AND J. WU [27] G. G. Kasparov. Equivariant KK-theory and the Novikov conjecture. Invent. math., 91 (1988), 147 -- 201. [28] G. Landi, W. van Suijlekom. Principal fibrations from noncommutative spheres. Comm. Math. Phys., 260 (2005), no. 1, 203 -- 225. [29] J. Matousek. Using Borsuk-Ulam theorem. Universitext, Lectures on topological methods in combinatorics and geometry, written in cooperation with Anders Bjorner and Gunter M. Ziegler, Springer-Verlag, Berlin, 2003. [30] K. Matsumoto. Noncommutative three-dimensional spheres. J apan. J. Math. (N.S.), 17, 2 (1991), 333 -- 356. [31] K. Matsumoto. Noncommutative three-dimensional spheres. II. Noncommutative Hopf fiber- ing. Yokohama Math. J., 38, 2 (1991), 103 -- 111. [32] E. Micheal. Local compactness and Cartesian products of quotient maps and k-spaces. Ann. Inst. Fourier, Grenoble, 18, 2 (1968), 281 -- 286. [33] J. Milnor. Construction of Universal Bundles, II. Ann. of Math., 68, 3 (1956), 430 -- 436. [34] G. D. Mostow. Equivariant Embeddings in Euclidean Space. Ann. of Math., 65, 3 (1957), 432 -- 446. [35] M. Murayama. On the G-Homotopy Types of G-ANR's. Publ. RIMS, Kyoto Univ., 18 (1982), 183 -- 189. [36] M. Murayama. On G-ANR's and their G-homotopy types. Osaka J. Math., 20 (1983), 479 -- 512. [37] G. Nagy. On the K-theory of the non-commutative circle. J. Operator Theory, 31, 2 (1994), 303 -- 309. [38] T. Natsume, C. L. Olsen. Toeplitz operators on noncommutative spheres and an index theo- rem. Indiana Univ. Math. J., 46 (1997), no. 4, 1055 -- 1112. [39] R. Nest, C. Voigt. Equivariant Poincar´e duality of quantum group actions. J. Funct. Anal., 258 (2010), 1466 -- 1503. [40] B. Passer. Noncommutative Borsuk-Ulam Theorems. Thesis (Ph.D.)-Washington University in St. Louis. ProQuest LLC, Ann Arbor, MI, 2016. 133 pp. [41] B. Passer. A noncommutative Borsuk -- Ulam theorem for Natsume -- Olsen spheres. J. Operator Theory, 75 (2016), no. 2, 337 -- 366. [42] B. Passer. Free Actions on C-algebra Suspensions and Joins by Finite Cyclic Groups. Indiana Univ. Math. J., 67 (2018), no. 1, 187 -- 203. [43] M. J. Pflaum. Quantum groups on fibre bundles. Comm. Math. Phys., 166 (1994), 279 -- 315. [44] M. Pimsner, D. Voiculescu. Imbedding the irrational rotation C-algebra into an AF-algebra. J. Operator Theory, 4, 2 (1980), 201 -- 210. [45] N. C. Phillips. Equivariant semiprojectivity. arXiv:1112.4584. [46] N. C. Phillips, A. P. W. Sørensen, H. Thiel. Semiprojectivity with and without a group action. J. Funct. Anal., 268, 4 (2015), 929 -- 973. [47] P. Podle´s. Quantum spheres. Lett. Math. Phys., 14, 3 (1987),193 -- 202. [48] L. Pontrjagin. Topological groups. (Translated from the Russian). Princeton University Press, Princeton, 1946. [49] M. A. Rieffel. C-algebras associated with irrational rotations. Pacific J. Math., 93, 2 (1981), 415 -- 429. [50] A. S. Schwarz. The genus of a fibre space. (Russian) Tr. Mosk. Mat. Obs., 10 (1961), 217 -- 272. [51] So(cid:32)ltan, Piotr M. On actions of compact quantum groups. Illinois J. Math., 55 (2011), no. 3, 953 -- 962. [52] N. Steenrod. The topology of fibre bundles. Princeton Landmarks in Mathematics, Reprint of the 1957 edition, Princeton Paperbacks, Princeton University Press, Princeton, 1999. [53] K. Thomsen. On the KK-theory and the E-theory of amalgamated free products of C*- algebras. J. Funct. Anal., 201 (2003), 30 -- 56. [54] T. tom Dieck. Algebraic topology. EMS Textbooks in Mathematics, European Mathematical Society (EMS), Zurich, 2008. THE LOCAL-TRIVIALITY DIMENSION OF ACTIONS OF COMPACT QUANTUM GROUPS 37 [55] J. H. C. Whitehead. Note on a theorem due to Borsuk. Bull. Amer. Math. Soc., 54 (1958), 1125 -- 1132. [56] W. Winter, J. Zacharias. Completely positive maps of order zero. Munster J. Math., 2 (2009), 311 -- 324. [57] S. L. Woronowicz. Compact matrix pseudogroups. Comm. Math. Phys., 111 (1987), 613 -- 665. [58] M. Yamashita. Equivariant comparison of quantum homogeneous spaces. Comm. Math. Phys., 317 (2013), no. 3, 593 -- 614. (E. Gardella) Mathematisches Institut, Fachbereich Mathematik und Informatik der Universitat Munster, Einsteinstrasse 62, 48149 Munster, Germany E-mail address: [email protected] (P. M. Hajac) Instytut Matematyczny, Polska Akademia Nauk, ul. ´Sniadeckich 8, Warszawa, 00-656 Poland E-mail address: [email protected] (M. Tobolski) Instytut Matematyczny, Polska Akademia Nauk, ul. ´Sniadeckich 8, Warszawa, 00-656 Poland E-mail address: [email protected] (J. Wu) Department of Mathematics, Penn State University, 109 McAllister Build- ing, University Park, PA 16802 USA E-mail address: [email protected]
1608.00178	2	1608	2017-01-28T16:00:23	Rigged modules I: modules over dual operator algebras and the Picard group	[ "math.OA", "math.FA" ]	In a previous paper we generalized the theory of W-modules to the setting of modules over nonselfadjoint dual operator algebras, obtaining the class of weak-rigged modules. At that time we promised a forthcoming paper devoted to other aspects of the theory. We fulfill this promise in the present work and its sequel "Rigged modules II", giving many new results about weak-rigged modules and their tensor products. We also discuss the Picard group of weak closed subalgebras of a commutative algebra. For example, we compute the weak Picard group of $H^{\infty}(D)$, and prove that for a weak* closed function algebra A, the weak Picard group of A is a semidirect product of the automorphism group of A, and the subgroup consisting of symmetric equivalence bimodules.	math.OA	math	RIGGED MODULES I: MODULES OVER DUAL OPERATOR ALGEBRAS AND THE PICARD GROUP DAVID P. BLECHER AND UPASANA KASHYAP Abstract. In a previous paper we generalized the theory of W ∗-modules to the setting of modules over nonselfadjoint dual operator algebras, obtaining the class of weak∗-rigged modules. At that time we promised a forthcoming paper devoted to other aspects of the theory. We fulfill this promise in the present work and its sequel "Rigged modules II", giving many new results about weak∗- rigged modules and their tensor products. We also discuss the Picard group of weak* closed subalgebras of a commutative algebra. For example, we compute the weak Picard group of H∞(D), and prove that for a weak* closed function algebra A, the weak Picard group is a semidirect product of the automorphism group of A, and the subgroup consisting of symmetric equivalence bimodules. 1. Introduction and notation The most important class of modules over a C ∗-algebra M are the Hilbert C ∗- modules: the modules possessing an M -valued inner product satisfying the natural list of axioms (see [19] or [9, Chapter 8]). A W ∗-module is a Hilbert C ∗-module over a von Neumann algebra which is selfdual (that is, it satisfies the module variant of the fact from Hilbert space theory that every continuous functional is given by inner product with a fixed vector, see e.g. [21, 4]), or equivalently which has a Banach space predual (see e.g. [10, Corollary 3.5]). A dual operator algebra is a unital weak* closed algebra of operators on a Hilbert space (which is not necessarily selfadjoint). One may think of a dual operator algebra as a nonselfadjoint analogue of a von Neumann algebra. The weak∗-rigged or w∗-rigged modules, introduced in [7] (see also [8, 17]), are a generalization of W ∗-modules to the setting of modules over a (nonselfadjoint) dual operator algebra. In [7] we generalized basic aspects of the theory of W ∗-modules, and this may be seen also as the weak* variant of the theory of rigged modules from [3] (see also [11]). In that paper we promised that some other aspects of the theory of weak∗-rigged modules would be presented elsewhere. Since that time is now long overdue, and since there has been some recent interest in rigged modules and related objects (see e.g. [20] or the survey [14] of Eleftherakis' work), we follow through on our promise here and in the sequel [5]. The present paper and its sequel consists of several topics and results about weak∗-rigged modules, mainly concerning their tensor products. For example fol- lowing the route in [1] we study the Picard group of weak* closed subalgebras of a commutative algebra. For a weak* closed function algebra A, the weak Picard group P icw(A) is a semidirect product of Aut(A), the automorphism group of A, Date: November 16, 2018. Key words and phrases. W ∗-algebra; Hilbert C ∗-module; Operator algebra; Dual operator algebra; W ∗-module; weak-rigged module; Morita equivalence; von Neumann algebra; Picard group. 1 2 DAVID P. BLECHER AND UPASANA KASHYAP and the subgroup of P icw(A) consisting of symmetric equivalence bimodules. In particular, we show that the weak Picard group of H ∞(D) is isomorphic to the group of conformal automorphisms of the disk. We will use the notation from [6, 7, 18], and perspectives from [2]. We will assume that the reader is familiar with basic notions from operator space theory which may be found in any current text on that subject (such as [13, 9]). The reader may consult [9] as a reference for any other unexplained terms here. We also assume that the reader is familiar with basic Banach space and operator space duality principles, e.g., the Krein-Smulian Theorem (see e.g., Section 1.4, 1.6, Ap- pendix A.2 in [9]). We often abbreviate 'weak∗' to 'w∗'. We use the letters H, K for Hilbert spaces. By a nonselfadjoint analogue of Sakai's theorem (see e.g. Section 2.7 in [9]), a dual operator algebra M is characterized as a unital operator algebra which is also a dual operator space. By a normal morphism we shall always mean a unital weak continuous completely contractive homomorphism on a dual oper- ator algebra. A concrete dual operator M -N -bimodule is a weak* closed subspace X of B(K, H) such that θ(M )Xπ(N ) ⊂ X, where θ and π are normal morphism representations of M and N on H and K respectively. An abstract dual operator M -N -bimodule is defined to be a nondegenerate operator M -N -bimodule X, which is also a dual operator space, such that the module actions are separately weak* continuous. Such spaces can be represented completely isometrically as concrete dual operator bimodules, and in fact this can be done under even weaker hypotheses (see e.g. [9, 10, 12]) and similarly for one-sided modules (the case M or N equals C). We use standard notation for module mapping spaces; e.g. CB(X, N )N (resp. CBσ(X, N )N ) are the completely bounded (resp. and weak* continuous) right N - module maps from X to N . We often use the normal module Haagerup tensor product ⊗σh M , and its universal property from [15], which loosely says that it 'lin- earizes' completely contractive M -balanced separately weak* continuous bilinear maps (balanced means that u(xa, y) = u(x, ay) for a ∈ M ). We assume that the reader is familiar with the notation and facts about this tensor product from [6, Section 2]. For any operator space X we write Cn(X) for the column space of n × 1 matrices with entries in X, with its canonical norm from operator space theory. Definition 1.1. [7] Suppose that Y is a dual operator space and a right mod- ule over a dual operator algebra M . Suppose that there exists a net of posi- tive integers (n(α)), and w∗-continuous completely contractive M -module maps φα : Y → Cn(α)(M ) and ψα : Cn(α)(M ) → Y , with ψα(φα(y)) → y in the weak* topology on Y , for all y ∈ Y . Then we say that Y is a right w∗-rigged module (or weak∗-rigged module) over M . As on p. 348 of [7], the operator space structure of a w∗-rigged module Y over M is determined by k[yij]kMn(Y ) = sup α k[φα(yij)]k, [yij] ∈ Mn(Y ). The following seems not to have been proved in the development in Section 2 and the start of Section 3 in [7] but seemingly assumed there. Lemma 1.2. A right w∗-rigged module over a dual operator algebra M is a dual operator M -module. Proof. Let Y be the w∗-rigged module, and let φα, ψα, nα be as in Definition 1.1. We need to show that the module action Y × M → Y is separately weak* continuous. RIGGED MODULES OVER DUAL OPERATOR ALGEBRAS 3 Given a bounded net mt → m weak* in M , and y ∈ Y , suppose that a subnet ymtν → y′ weak* in Y . Then φα(ψα(ymtν )) = φα(ψα(y)mtν ) weak* converges both to φα(ψα(y)m) = φα(ψα(ym)) and φα(ψα(y′)) with ν, for every α. Hence y′ = ym, so that by topology ymt → ym weak. So the map m 7→ ym is weak continuous by the Krein-Smulian theorem. That each m ∈ M acts weak* continuously on Y follows from e.g. [9, Corollary 4.7.7]. (cid:3) Every right w∗-rigged module Y over M gives rise to a canonical left w∗-rigged M -module Y , and a pairing (·, ·) : Y × Y → M (see [7]). Indeed Y turns out to be completely isometric to CBσ(Y, M )M as dual operator M -modules, together with its canonical pairing with Y . We also have eY = Y . The morphisms between w∗- rigged M -modules are the adjointable M -module maps; these are the M -module maps T : Y1 → Y2 for which there exists a S : Y2 → Y1 with (x, T (y)) = (S(x), y) for all x ∈ Y2, y ∈ Y1. These turn out to coincide with the weak* continuous completely bounded M -module maps (see [7, Proposition 3.4]). We often write B(Z, W ) for the weak* continuous completely bounded M -module maps from a w∗-rigged M - module Z into a dual operator M -module W , with as usual B(Z) = B(Z, Z). In [7], Section 4 we gave several equivalent definitions of w∗-rigged modules (an additional note that seems to be needed to connect the definitions may be found mentioned in the proof of [5, Theorem 2.3]). From some of these it is clear that every weak* Morita equivalence N -M -bimodule Y in the following sense, is a w∗- rigged right M -module and a w∗-rigged left N -module, and its 'dual module' Y and pairing (·, ·) : Y × Y → M may be taken to be the X below, and pairing corresponding to the first ∼= below. Definition 1.3. We say that two dual operator algebras M and N are weak* Morita equivalent, if there exist a dual operator M -N -bimodule X, and a dual operator N -M -bimodule Y , such that M ∼= X ⊗σh N Y as dual operator M -bimodules (that is, completely isometrically, w∗-homeomorphically, and also as M -bimodules), and similarly N ∼= Y ⊗σh M X as dual operator N -bimodules. We call (M, N, X, Y ) a weak* Morita context in this case. 2. Rigged and weak∗-rigged modules and their tensor product 2.1. Interior tensor product of weak* rigged modules. Suppose that Y is a right w∗-rigged module over a dual operator algebra M and, that Z is a right w∗-rigged module over a dual operator algebra N , and θ : M → B(Z) is a normal morphism. Because Z is a left operator module B(Z)-module (see p. 349 in [7]), Z becomes an essential left dual operator module over M under the action m · z = θ(m)z. In this case we say Z is a right M -N -correspondence. We form the normal module Haagerup tensor product Y ⊗σh M Z which we also write as Y ⊗θ Z. By 3.3 in [7] this a right w∗-rigged module over N . We call Y ⊗θ Z the interior tensor product of w∗-rigged modules. By 3.3 in [7] we have ^Y ⊗θ Z ∼= Z ⊗θ Y with N -valued pairing (w ⊗ x, y ⊗ z)N = (w, θ((x, y)M )z)N of Z ⊗θ Y with Y ⊗σh M Z. The w∗-rigged interior tensor product is associative. That is, if A, B, and C are dual operator algebras, if X is a right w∗-rigged module over A, Y is a right w∗-rigged module over B, and Z is a right w∗-rigged module over C, and if θ : A → B(Y ) and ρ : B → B(Z) are normal morphisms, then (X ⊗θ Y ) ⊗ρ Z ∼= 4 DAVID P. BLECHER AND UPASANA KASHYAP X⊗θ(Y ⊗ρZ) completely isometrically and w∗-homeomorphically. This follows from the associativity of the normal module Haagerup tensor product (see Proposition 2.9 in [6]). Let Y be a right w∗-rigged module over a dual operator algebra M . If N is a W ∗-algebra and if Z is a right W ∗-module over N and θ : M → B(Z) is a normal morphism, then as we said earlier, Y ⊗θ Z is a right w∗-rigged module over N . Since N is a W ∗-algebra it follows from Theorem 2.5 in [7] that Y ⊗θ Z is a right W ∗- module over N . An important special case of this is the W ∗-dilation. If Z = N is a von Neumann algebra containing (a weak* homeomorphic completely isometrically isomorphic copy of) M , with the same identity element, and θ : M → N is the inclusion, then Y ⊗θ N is called the W ∗-dilation of Y (see [18, 6, 7]). The following is an application of the W ∗-dilation to Morita equivalence. Theorem 2.1. Suppose that Y is a right w∗-rigged module over a dual operator algebra M . Suppose that N is a von Neumann algebra containing M as a weak* closed subalgebra (with the same identity), and that R is the weak* closed ideal in N generated by the range of the pairing Y × Y → M . Let θ : M → N be the inclusion, which also induces a map π : M → BR(R) ∼= R via the canonical left action. Then Z = Y ⊗θ N = Y ⊗π R is a von Neumann algebraic Morita equivalence bimodule (in the sense of Rieffel [22]), implementing a von Neumann algebraic Morita equivalence between R and B(Z) = Y ⊗π R ⊗π Y . In particular, if W ∗(M ) is a von Neumann algebra generated by (a weak* homeomorphic completely isometrically isomorphic copy of ) M , and if the range of the pairing Y × Y → M is weak* dense in M , then W ∗(M ) is Morita equivalent (in the sense of Rieffel) to the W ∗-algebra B(Y ⊗θ W ∗(M )) = Y ⊗θ W ∗(M ) ⊗θ Y . Proof. By the lines above the theorem, Z = Y ⊗θ N is a right W ∗-module over N . That Y ⊗θ N = Y ⊗π R may be seen for example by [7, Theorem 3.5] (see Lemma 2.5 below), since any weak* continuous left M -module map Y → N necessarily takes values in the ideal R since terms of form (x, y)x′ are weak* dense in Y , for x, x′ ∈ Y , y ∈ Y . By Corollary 8.5.5 in [9], B(Z) is a W ∗-algebra. Consider the canonical pairing Z × Z ∼= (N ⊗θ Y ) × (Y ⊗θ N ) → N. The w∗-closure of the range of this pairing is a weak* closed two sided ideal in N . It is easy to see that this ideal is R, since the terms of the form a(xy)b are contained in the range of the above pairing for all (x, y) ∈ Y × Y and a, b ∈ N . By [22] (or 8.5.14 in [9]), R and B(Z) are Morita equivalent in the sense of Rieffel. That B(Z) = Y ⊗σh M (R ⊗σh R R) ⊗σh M Y = Y ⊗θ R ⊗θ Y follows from the second paragraph on p. 357 of [7]. Since W ∗(M ) is a W ∗-algebra generated by M , and it is well known that the weak* closed ideals of W ∗-algebras are selfadjoint, these together imply that the weak* closure of the range of the above pairing is all of W ∗(M ). By 8.5.14 in [9], W ∗(M ) and B(Y ⊗θ W ∗(M )) are Morita equivalent in the sense of Rieffel. (cid:3) 2.2. Functorial properties. Note that Mm,n(B(Y1, Y2)) has two natural norms: the operator space one, coming from CB(Y1, Mm,n(Y2)), or the norm coming from CB(Cn(Y1), Cm(Y2)) via the identification of a matrix [fij ] ∈ Mm,n(CB(Y1, Y2)) with the map [yj] 7→ [Pj fij(yj)]. The next result asserts that these norms on Mm,n(B(Y1, Y2)) are the same. We will write this norm as k[fij]kcb. RIGGED MODULES OVER DUAL OPERATOR ALGEBRAS 5 Lemma 2.2. ([7, Corollary 3.6]) Suppose that Y1 and Y2 are right w∗-rigged mod- ules over a dual operator algebra M . For each m, n ∈ N we have Mm,n(B(Y, Z)) ∼= B(Cn(Y ), Cm(Z)) completely isometrically. The interior tensor product of w∗-rigged modules is functorial: Proposition 2.3. Suppose that Y1 and Y2 are right w∗-rigged modules over a dual operator algebra M , that Z1 and Z2 are right M -N -correspondences for a dual operator algebra N . Write the left action on Zk (abusively) as θ : M → B(Zk)N . If f = [fij ] ∈ Mm,n(B(Y1, Y2)M ), and if g = [gkl] ∈ Mp,q(B(Z1, Z2)N ) is a matrix of adjointables which are also left M -module maps, write f ⊗ g for [fij ⊗ gkl] ∈ Mmp,nq(B(Y1 ⊗θ Z1, Y2 ⊗θ Z2). Then kf ⊗ gkcb ≤ kf kcbkgkcb, where the 'subscript cb' refers to the norm discussed above the theorem. Further, ]f ⊗ g = g ⊗ f for any f ∈ B(Y1, Y2)M and g ∈ B(Z1, Z2)N . M Z1 → Y2 ⊗σh Proof. Suppose that f ∈ B(Y1, Y2)M and g ∈ B(Z1, Z2)N . Since g : Z1 → Z2 is a left M -module map, g is a right M -module map. Thus we can define f ⊗σh M g : Y1. By Corollary 2.4 in Y1 ⊗σh [6], f ⊗ g : Y1 ⊗θ Z1 → Y2 ⊗θ Z2 is a completely bounded weak* continuous right N -module map, with kf ⊗ gkcb ≤ kf kcbkgkcb. That f ⊗ g is adjointable follows from the fact that it is weak* continuous (by the remark a couple of paragraphs above Definition 1.3). Alternatively, for y ∈ Y1, z ∈ Z1, w ∈ Z2, x ∈ Y2 we have M Z2, and g ⊗σh M Y2 → Z1 ⊗σh f : Z2 ⊗σh M M (w ⊗ x, f (y) ⊗ g(z))N = (w, θ((x, f (y))M )g(z))N = (wθ(( f (x), y)M ), g(z))N = (g(wθ(( f (x), y)M )), z)N = (g(w), θ(( f (x), y)M )z)N = (g(w) ⊗ f (x), y ⊗ z)N , which also yields the last statement. Next, let f = [fij ], g = [gkl] be as in the statement. By a two step method we may assume that f = I or g = I. If g = I the norm inequality we want follows from the case in the last paragraph, Lemma 2.2, and because Mm,n(B(Y1 ⊗θ Z1, Y2 ⊗θ Z2)) ∼= B(Cn(Y1 ⊗θ Z1), Cm(Y2 ⊗θ Z2)) which may be viewed as B(Cn(Y1) ⊗θ Z1, Cm(Y2) ⊗θ Z2). Thus [fij ⊗ 1] may be regarded as the map h ⊗ I on Cn(Y1) ⊗θ Z1, where h is the map Cn(Y1) → Cn(Y2) associated with [fij] as in the discussion above Lemma 2.2. Thus k[fij ⊗ 1]kcb = kh ⊗ IkB(Cn(Y1)⊗θ Z1,Cm(Y2)⊗θ Z2) ≤ khkcb = kf kcb. If f = I, the norm inequality we want follows by a standard trick (which could also have been used to give an alternative proof of the previous computation). If y = [yij] ∈ Mp(Y ) and z = [zij] ∈ Mp(Z) then yir ⊗ gkl(zα k )(wα k , zrj) yir ⊗ gkl(zrj) = lim Xr k , z)] and ψα([bk]) = Pk zα α Xr,k where φα(z) = [(wα but for Z in place of Y . It follows that the norm of [Pr yir ⊗ gkl(zrj)] is dominated by supα k[gkl]kcbkykkzk. That kI ⊗ gkcb ≤ kgkcb follows easily from this if we use [6, Corollary 2.8]. (cid:3) k bk are the maps in Definition 1.1, 6 DAVID P. BLECHER AND UPASANA KASHYAP Remark. A similar result holds for rigged modules over approximately unital operator algebras. These matricial versions of the functoriality of the tensor product will be used in [5]. The w∗-rigged interior tensor product is projective. This is because of the fol- lowing: Proposition 2.4. The normal module Haagerup tensor product is projective: If u : Y1 → Y2 is a weak* continuous M -module complete quotient map between right dual operator M -modules, and v : Z1 → Y2 is a weak* continuous M -module complete quotient map between left dual operator M -modules, then u ⊗ v : Y1 ⊗σh M Z1 → Y2 ⊗σh M Z2 is a weak* continuous complete quotient map. Proof. By functoriality of ⊗σh M (see [6, Corollary 2.4]) we have that u ⊗ v is a weak* continuous complete contraction. If z ∈ Ball(Y2 ⊗σh M Z2), then by [6, Corollary 2.8] z is weak* approximable by a net zt ∈ Ball(Y2⊗hM Z2). We may assume that kztk < 1 for all t. By projectivity of ⊗hM (or of ⊗h) there exist wt ∈ Ball(Y1 ⊗hM Z1) with (u ⊗ v)(wt) = zt. Suppose that wtν → w ∈ Ball(Y1 ⊗σh M Z1), then clearly (u ⊗ v)(w) = z. So u ⊗ v is a quotient map. A similar argument works at the matrix levels using [6, Corollary 2.8]. (cid:3) Unlike the module Haagerup tensor product over a C ∗-algebra (see [9, Theorem 3.6.5 (2)]), the normal module Haagerup tensor product ⊗σh M need not be 'injective' for general dual operator modules if M is a W ∗-algebra. However by functoriality it is easy to see that we will have such injectivity for this tensor product for w∗- orthogonally complemented submodules (see the last section of [5]) of w∗-rigged modules, even if M is a dual operator algebra. For example if Y is a w∗-orthogonally complemented submodule of a w∗-rigged module W over M , and if i and P are the associated inclusion and projection maps, and if Z is a right M -N -correspondence, then P ⊗ IZ and i ⊗ IZ are completely contractive adjointable maps composing to the identity on Y ⊗θ Z. So Y ⊗θ Z is weak* homeomorphically completely isometrically M -module isomorphic to a w∗-orthogonally complemented submodule of W ⊗θ Z. A similar statement may be made for an appropriately complemented M -N -'subcorrespondence' of Z. This injectivity of the weak* interior tensor product will work for any weak* closed submodules of w∗-rigged modules over a W ∗-algebra (that is, W ∗-modules), because such are automatically w∗-orthogonally complemented [4]. It follows that the weak* interior tensor with a right correspondence is 'exact' on the category of W ∗-modules. Thus given an exact sequence of W ∗-modules over a W ∗-algebra M , 0 −→ D −→ E −→ F −→ 0 with the first of these adjointable morphisms completely isometric and the second a complete quotient map, and given a right M -N -correspondence Z, we get an exact sequence 0 −→ D ⊗θ Z −→ E ⊗θ Z −→ F ⊗θ Z −→ 0 of the same kind. Indeed this all follows from the 'commutation with direct sums' property at the end of [7, Section 3]. It might be interesting to investigate such exactness of the interior tensor with a correspondence on the category of w∗-rigged modules. RIGGED MODULES OVER DUAL OPERATOR ALGEBRAS 7 2.3. HOM-tensor relations. (See [4, Theorem 3.6] for the self-adjoint variant of these.) In our context, the HOM spaces will be the spaces B(−, −) of all weak* continuous completely bounded module maps. Lemma 2.5. If Y is a right w∗-rigged module over M , and Z is a left (resp. Y ∼= right) dual operator module over M , then Y ⊗σh B(Y, Z)M ) completely isometrically and weak* homeomorphically. M Z ∼= BM ( Y , Z) (resp. Z ⊗σh M Proof. The respectively case is [7, Theorem 3.5]. By the 'other-handed variant' of [7, Theorem 3.5] we have CBσ M Z = Y ⊗σh M Z. (cid:3) M ( Y , Z) ∼= eY ⊗σh Theorem 2.6. Let M and N be dual operator algebras. We have the following completely isometric identifications: N (1) B(Y ⊗θ Z, W )N ∼= B(Y, B(Z, W )N )M , where Y is a right w∗-rigged module over M , Z is a right M -N correspondence, and W is a right dual operator module over N . (2) B(Y, (Z ⊗σh N W ))M ∼= Z ⊗σh B(Y, W )M , where Y is a right w∗-rigged module over M , W is a dual operator N -M -bimodule and Z is a right dual operator N -module. (3) BN (B(Y, W )M , X) ∼= Y ⊗σh BN (W, X), where Y is a right w∗-rigged mod- ule over M , X is a dual left N -operator module, and W is a left N -M - correspondence. (4) BM (X, B(Z, W )N ) ∼= B(Z, BM (X, W ))N where X, Z are left and right w∗- rigged modules over M and N respectively, and W is a dual operator M - N -bimodule. M Proof. The proofs all follow from Lemma 2.5, and Corollary 3.3 in [7], and the associativity of the normal module Haagerup tensor product, and the fact that eY = Y for w∗-rigged modules. Since the proofs are all similar we just prove a couple of them. For (1) note that BN (Y ⊗θ Z, W ) ∼= W ⊗σh N ^Y ⊗θ Z ∼= W ⊗σh N ( Z ⊗σh M Y ) ∼= (W ⊗σh N Z) ⊗σh M Y which is isomorphic to B(Z, W )N ⊗σh M Y ∼= B(Y, B(Z, W )N )M . For (3), BN (B(Y, W )M , X) ∼= BN (W ⊗σh M Y ) ⊗σh N X ∼= (Y ⊗σh M which is isomorphic to Y ⊗σh M ( W ⊗σh BN (W, X). Y , X) ∼= ^ (W ⊗σh M N X) ∼= Y ⊗σh M W ) ⊗σh N X (cid:3) 3. The Picard group We now discuss the Picard group of a dual operator algebra, following the route in [1]. Throughout this section A will be a dual operator algebra. We define the weak Picard group of A, denoted by Picw(A), to be the collection of all A-A-bimodules implementing a weak* Morita equivalence of A with itself, with two such bimodules identified if they are completely isometrically isomorphic and weak* homeomorphic via an A-A-bimodule map. The multiplication on Picw(A) is given by the module normal Haagerup tensor product ⊗σh A . Any weak* continuous completely isometric automorphism θ of A defines a weak* Morita equivalence A-A-bimodule Aθ by 'change of rings' on the right. This is just A with the usual left module action, and with right module action x · a = xθ(a). 8 DAVID P. BLECHER AND UPASANA KASHYAP Lemma 3.1. The bimodule Aθ above is a weak* Morita equivalence bimodule for A, with 'inverse bimodule' Aθ−1 . Proof. This follows using Definition 1.3 and Lemma 3.3 below, but we will give a more explicit proof. If A is a dual operator algebra, then we prove that (A, A, Aθ, Aθ−1) is a weak* Morita context using Theorem 3.3 in [6]. Define a pairing (·) : Aθ × Aθ−1 → A taking (a, a′) 7→ aθ(a′). It is easy to check that (·) is a separately w∗-continuous completely contractive A-bimodule map which is balanced over A. Similarly, define another pairing [·] : Aθ−1 × Aθ → A taking [a, a′] 7→ aθ−1(a′). Again it is easy to check that [·] is a separately w∗-continuous completely contrac- tive A-module map which is balanced over A. It is simple algebra to check that (x, y)x′ = x[y, x′] and y′(x, y) = [y′, x]y. Checking the last assertion of Theorem 3.3 in [6] is also obvious. Hence (A, A, Aθ, Aθ−1) is a weak* Morita context. (cid:3) Let Aut(A) denote the group of weak* continuous completely isometric auto- morphisms of A. For α, β ∈Aut(A) let αAβ denote A viewed as an A-A-bimodule with the left action a · x = α(a)x and right action x · b = xβ(b). Lemma 3.2. If α, β, γ ∈Aut(A) then, αAβ ∼= γαAγβ completely A-A-isometrically. Proof. The map γ is the required isomorphism. (cid:3) Lemma 3.3. For θ1, θ2 ∈Aut(A), Aθ1 ⊗σh Proof. From Lemma 3.2, Aθ1 ⊗σh A Aθ2 A Aθ2 ∼= θ−1 1 ∼= Aθ1θ2 completely A-A-isometrically. A ⊗σh ∼= Aθ1θ2. Aθ2 A Aθ2 (cid:3) ∼= θ−1 1 Proposition 3.4. The collection {Aθ : θ ∈ Aut(A)} constitutes a subgroup of Picw(A), which is isomorphic to the group Aut(A) of weak* continuous completely isometric automorphisms of A. Proof. This follows from the above lemma. (cid:3) If X is a weak* Morita equivalence A-A-bimodule, and if θ ∈Aut(A), then let Xθ be X with the same left module action, but with right module action changed to x · a = xθ(a). Lemma 3.5. If X is a weak* Morita equivalence bimodule, then Xθ ∼= X ⊗σh completely A-A-isometrically. Also, Xθ is a weak* Morita equivalence bimodule. A Aθ Proof. The module action (x, a) 7→ xa is a completely contractive, separately weak* continuous balanced bilinear map, so by the universal property of the normal mod- ule Haagerup tensor product it induces a completely contractive weak* continuous linear map m : X ⊗σh A Aθ → Xθ. There is a completely contractive inverse map x 7→ x ⊗ 1, so that m is a surjective complete isometry, and it is easily seen to be an A-A-bimodule map. If (A, A, X, Y ) is a weak Morita context as in Definition 1.3 then by Lemma 3.1, and properties of the normal module Haagerup tensor product, (A, A, Xθ,θ−1 Y ) is a Morita context. To see this, note that by associativity of the normal module Haagerup tensor product, and some of the lemmas above in the present section, (X ⊗σh A Y ) ∼= X ⊗σh A Y ∼= X ⊗σh A (Aθ−1 ⊗σh A Aθ) ⊗σh A Y ∼= A A A ⊗σh completely isometrically and weak-homeomorphically. Similarly (Aθ−1 ⊗σh A A ⊗σh completely isometrically and weak-homeomorphically. A Aθ) ∼= Aθ−1 ⊗σh A (X ⊗σh A Y ) ⊗σh A Aθ ∼= Aθ−1 ⊗σh A Aθ ∼= A (cid:3) RIGGED MODULES OVER DUAL OPERATOR ALGEBRAS 9 An A-A-bimodule X will be called 'symmetric' if ax = xa for all a ∈ A, x ∈ X. Proposition 3.6. For a commutative dual operator algebra A, Picw(A) is a semidi- rect product of Aut(A) and the subgroup of Picw(A) consisting of symmetric equiv- alence bimodules. Thus, every V ∈ Picw(A) equals Xθ, for a symmetric X ∈ Picw(A), and some θ ∈ Aut(A). Proof. Suppose that X is any weak* Morita equivalence A-A-bimodule. Then any w∗-continuous right A-module map T : X → X is simply left multiplication by a fixed element of A. In fact we have A ∼= CBσ(X)A, via a map L : A → CB(X) (see e.g. Theorem 3.6 in [6]). For fixed a ∈ A, the map x 7→ xa on X, is a w∗- continuous completely bounded A-module map with completely bounded norm = kak. Hence by the above identification, there exists a unique a′ ∈ A such that a′x = xa for all x ∈ X and ka′k = kak. Define θ(a) = a′, then we claim that θ is a weak* continuous completely isometric unital automorphism of A. To see that θ is a homomorphism, let a1, a2 ∈ A and let T, S and U be maps from X to X simply given by right multiplication with a1, a2 and a1a2 respectively. Let θ(a1) = a′ 1, θ(a2) = a′ 2 and θ(a1a2) = a3. Since U = ST , we have L(U ) = L(T )L(S) (recall L is an anti-homomorphism, see e.g. Theorem 3.6 in [6]). This implies a3 = a′ 1a′ 2, that is, θ(a1a2) = θ(a1)θ(a2). Note that k[θ(aij)xkl]k = k[xklaij ]k ≤ k[xkl]kk[aij]k, and so using the isomorphism A ∼= CBσ A(X) above we see that θ is completely contractive. That θ is completely isometric follows (e.g. by a similar argument for θ−1). For the weak* continuity, note that if we have a bounded net at xat deduce that θ(at) w∗ w∗ → xa for all x ∈ X. Since the map L above is a weak* homeomorphism, we → θ(a). That θ is surjective follows by symmetry. w∗ → a then Thus we have defined a surjective group homomorphism P icw(A) → Aut(A) taking X 7→ θ. To see that this does define a group homomorphism, let X, Y ∈ P icw(A). Let X ⊗σh A Y 7→ θ and X 7→ θ1, Y 7→ θ2 under the above identification. We need to show that θ = θ1θ2. Let a ∈ A and θ(a) = a′, θ2(a) = a2 and θ1(a2) = a1. We need to show that a′ = a1. Consider a rank one tensor x ⊗ y ∈ X ⊗σh A Y for x ∈ X and y ∈ Y . Then a′x ⊗ y = x ⊗ ya, a1x = xa2 and a2y = ya. We have a′(x ⊗ y) = a′x ⊗ y and a1(x ⊗ y) = a1x ⊗ y = xa2 ⊗ y = x ⊗ a2y = x ⊗ ya. Since finite rank tensors are weak* dense in X ⊗σh z ∈ X ⊗σh A Y . This implies a′ = a1 which proves the required assertion. A Y , we have a′z = a1z for all From Lemma 3.1, for θ ∈ Aut(A) we have Aθ ∈ P icw(A), hence the above ho- momorphism has a 1-sided inverse Aut(A) → P icw(A). The above homomorphism restricted to modules of the form Aθ for θ ∈ Aut(A) is the identity map. That is, for θ ∈ Aut(A) the above homomorphism takes the weak* Morita equivalence bimodule Aθ to θ. Moreover the kernel of the homomorphism equals the symmetric equivalence bimodules. This proves the 'semidirect product' statement. For the last assertion, note that X = (Xθ−1)θ. From the above, xθ−1(a) = θ(θ−1(a))x = ax, which proves that Xθ−1 is symmetric. (cid:3) Remark. Similar results will hold for strong Morita equivalence bimodules in the sense of [11] over a norm closed operator algebra A, and their associated Picard group. 10 DAVID P. BLECHER AND UPASANA KASHYAP Thus we may assume henceforth that X is symmetric, if A is a commutative dual operator algebra. Proposition 3.7. Suppose that we have a weak* Morita context (A, A, X, Y ), with A a weak* closed subalgebra of a commutative von Neumann algebra M . Suppose that A generates M as a von Neumann algebra. Then every symmetric weak equiv- alence A-A-bimodule is completely isometrically A-A-isomorphic to a weak* closed A-A-subbimodule of M . Proof. From Theorem 5.5 in [6], X dilates to a weak Morita equivalence M -M - bimodule W = M ⊗σh A X. Let (M, M, W, Z) be the corresponding W ∗-Morita context. From Theorem 3.5 in [18], W contains X completely isometrically as a weak* closed A-submodule; and indeed it is clear that this is as a sub-bimodule over A. It is helpful to consider the inclusion (cid:20) A X Y A (cid:21) ⊂(cid:20) M W Z M (cid:21) of linking algebras. If X is symmetric, then since W = M ⊗σh A X we have wa = aw for all w ∈ W, a ∈ A. Similarly for Z ∼= Y ⊗σh A M , we have za = az. Since Z = W ∗, we have wm∗ = (mw∗)∗ = (w∗m)∗ = m∗w. Therefore xw = wx for all w ∈ W , x ∈ M . Thus W is a symmetric element of Picw(M ). If M is a commutative von Neumann algebra, then it is well known that the Pi- card group of M is just Aut(M ), and M is the only symmetric element of Picw(M ). We include a quick proof of this for the reader's convenience. Indeed suppose that Z is a symmetric weak equivalence M -M -bimodule. Suppose that M is a von Neu- mann algebra in B(H), and let K = Z ⊗θ H, the induced representation. Then Z ⊂ B(H, K) is a WTRO, which is commutative in the sense of [9, Proposition 8.6.5], and hence (see e.g. the proof of the last cited result) Z contains a unitary u with uu∗ = IK and u∗u = IH . Also, the map R : z 7→ u∗z is a completely isometric right M -module map from Z onto M . That is, Z = uM . Note that if θ : M → B(Z) is the left action of M on Z, then since Z is symmetric we have θ(a)(ub) = uba = uau∗(ub), for a, b ∈ M . That is, θ(a) corresponds to a 7→ uau∗ ∈ B(K). Then R(θ(a)z) = u∗uau∗z = aR(z), so that R is a bimodule map. That is, Z ∼= M as equivalence M -M -bimodules. Hence the Picard group of M is just Aut(M ). Putting the last two paragraphs together, we have proved that every symmetric (cid:3) A-A-bimodule 'is' a weak* closed A-A-subbimodule of M . Corollary 3.8. The weak Picard group of H ∞(D) is isomorphic to the group of conformal automorphisms of the disk. Proof. Since the monomial z generates A = H ∞(D) as a dual algebra, any au- tomorphism θ of A defines a map τ on the disk by θ(f )(w) = f (τ (w)), so that τ = θ(z) ∈ H ∞. By looking at θ−1, it is easy to see that τ is a conformal map. Thus the group of conformal automorphisms of the disk is isomorphic to the group Aut(A). We will be done by Proposition 3.6, if we can show that every symmetric equivalence A-A-bimodule is A-A-isomorphic to A. Let M = L∞(T), and let X be a symmetric equivalence A-A-bimodule. By Proposition 3.7, X can be taken to be a weak* closed A-submodule of M . By Beurling's theorem it follows that X is singly generated. Indeed there is a function k ∈ X, which is either a projection or a unitary in M , with X equal to the weak* RIGGED MODULES OVER DUAL OPERATOR ALGEBRAS 11 closure of M k or Ak respectively [16]. If k were a nontrivial projection p ∈ M , then XM ⊂ kM , so that X does not generate Z as an M -module, which is a contradiction. So k is unitary, and hence, as in the proof of Proposition 3.7, X ∼= A as dual operator A-A-bimodules. (cid:3) We thank Paul Muhly for a discussion on the proof of the last result. References [1] D. P. Blecher and K. Jarosz, Isomorphisms of function modules and generalized approxima- tion in modulus, Trans. Amer. Math. Soc. 354 (2002), 3663-3701. [2] D. P. Blecher, A generalization of Hilbert modules, J. Funct. Anal. 136 (1996), 365 -- 421. [3] D. P. Blecher, A new approach to Hilbert C ∗-modules, Math. Ann. 307 (1997), 253 -- 290. [4] D. P. Blecher, On selfdual Hilbert modules, in "Operator algebras and their applications", pp. 65 -- 80, Fields Inst. Commun., 13, Amer. Math. Soc., Providence, RI, 1997. [5] D. P. Blecher, Rigged modules II: multipliers and duality, Preprint 2016, to appear Studia Math. [6] D. P. Blecher and U. Kashyap, Morita equivalence of dual operator algebras, J. Pure and Applied Algebra 212 (2008), 2401-2412. [7] D. P. Blecher and U. Kashyap, A characterization and a generalization of W ∗-modules, Trans. Amer. Soc. 363 (2011), 345-363. [8] D. P. Blecher and J. Kraus, On a generalization of W ∗-modules, Banach Center Publications 91 (2010), 77-86. [9] D. P. Blecher and C. Le Merdy, Operator algebras and their modules -- an operator space approach, London Mathematical Society Monographs, Oxford Univ. Press, Oxford, 2004. [10] D. P. Blecher and B. Magajna, Duality and operator algebras: automatic weak continuity and applications,J. Funct. Anal. 224 (2005), 386 -- 407. [11] D. P. Blecher, P. S. Muhly, and V. I. Paulsen, Categories of operator modules (Morita equiv- alence and projective modules), Mem. Amer. Math. Soc. 143 (2000), no. 681. [12] E. G. Effros and Z-J. Ruan, Representations of operator bimodules and their applications, J. Operator Theory 19 (1988), 137 -- 157. [13] E. G. Effros and Z-J. Ruan, Operator Spaces, London Mathematical Society Monographs, New Series, 23, The Clarendon Press, Oxford University Press, New York, 2000. [14] G. Eleftherakis, Some notes on Morita equivalence of operator algebras, Serdica Math. J. 41 (2015), 117 -- 128. [15] G. K. Eleftherakis and V. I. Paulsen, Stably isomorphic dual operator algebras, Math. Ann. 341 (2008), 99 -- 112. [16] T. W. Gamelin, Uniform algebras, Prentice Hall (1969). [17] U. Kashyap, Morita equivalence of dual operator algebras, Ph. D. thesis (University of Hous- ton), December 2008. [18] U. Kashyap, A Morita theorem for dual operator algebras, J. Funct. Anal. 256 (2009), 3545- 3567. [19] E. C. Lance, Hilbert C-modules. A toolkit for operator algebraists, London Mathematical Society Lecture Note Series, 210. Cambridge University Press, Cambridge, 1995. [20] B. Mesland, Unbounded bivariant K-theory and correspondences in noncommutative geome- try, J. Reine Angew. Math. 691 (2014), 101 -- 172. [21] W. L. Paschke, Inner product modules over B∗-algebras, Trans. Amer. Math. Soc. 182 (1973), 443 -- 468. [22] M. A. Rieffel, Morita equivalence for C ∗-algebras and W ∗-algebras, J. Pure Appl. Algebra 5 (1974), 51 -- 96. Department of Mathematics, University of Houston, Houston, TX 77204-3008 E-mail address: [email protected] Department of STEM, Regis College, Weston, MA 02493 E-mail address: [email protected]
1612.04012	1	1612	2016-12-13T03:28:05	Fubini theorem in noncommutative geometry	[ "math.OA" ]	We discuss the Fubini formula in Alain Connes' noncommutative geometry. We present a sufficient condition on spectral triples for which a Fubini formula holds true. The condition is natural and related to heat semigroup asymptotics. We provide examples of spectral triples for which the Fubini formula fails.	math.OA	math	Fubini theorem in noncommutative geometry School of Mathematics and Statistics, University of New South Wales, F. Sukochev Kensington, Australia D. Zanin School of Mathematics and Statistics, University of New South Wales, Kensington, Australia Abstract We discuss the Fubini formula in Alain Connes' noncommutative geometry. We present a sufficient condition on spectral triples for which a Fubini formula holds true. The condition is natural and related to heat semigroup asymptotics. We provide examples of spectral triples for which the Fubini formula fails. 1. Introduction Fix throughout a separable infinite dimensional Hilbert space H. We let L(H) denote the algebra of all bounded operators on H. For a compact operator T on H, let λ(k, T ) and µ(k, T ) denote its k−th eigenvalue1 and k−th singular value (these are the eigenvalues of T). We let L1,∞ denote the principal ideal in L(H) generated by the operator diag({ 1 k+1}k≥0). Equivalently, L1,∞ = {T ∈ L(H) : µ(k, T ) = O( 1 k + 1 )}. Note that our notation differs from the one used in [4]. The following result is proposed on p. 563 in [4]. 6 1 0 2 c e D 3 1 ] . A O h t a m [ 1 v 2 1 0 4 0 . 2 1 6 1 : v i X r a Proposition 1.1. Let (1 + D2 L(H). If one of the elements T1(1 + D2 1)−p1/2, (1 + D2 2)−p2/2 ∈ L1,∞ and let T1, T2 ∈ 2)−p2/2 is convergent, 1)−p1/2, T2(1 + D2 ✩ Research supported by the ARC. Email addresses: [email protected] (F. Sukochev), [email protected] (D. Zanin) 1The eigenvalues are counted with algebraic multiplicities and arranged so that their ab- solute values are non-increasing. Preprint submitted to Elsevier ✩ then Γ(1 + p1 + p2 2 p1 2 )Trω((T1 ⊗ T2)(1 + D2 )Trω(T1(1 + D2 = Γ(1 + 1)−p1/2) · Γ(1 + holds for some (Dixmier) trace Trω on L1,∞. 1 ⊗ 1 + 1 ⊗ D2 2)−(p1+p2)/2) = p2 2 )Trω(T2(1 + D2 2)−p2/2) (1.1) The wording on p. 563 in [4] is that "one of the two terms is convergent"is open for interpretation. One possible interpretation is that the operator T1(1 + D2 2)−p2/2) is Tauberian. Recall that an operator A ∈ L1,∞ is called Tauberian if there exists a limit 1)−p1/2 (or T2(1 + D2 1 log(n + 2) lim n→∞ n Xk=0 λ(k, A) = c or, in a form convenient for comparison further below, n Xk=0 λ(k, A) = c · log(n + 1) + o(log(n + 1)). We have therefore rephrased the proposition as one of the two operators is Tauberian. The functional T 7→ Γ(1 + p 2 )Trω(T (1 + D2)−p/2) is considered the p-dimensional integral in Connes' noncommutative geometry [4]. That T (1 + D2)−p/2 is Tauberian implies that the functional is independent of which Dixmier trace Trω is used to define the functional, a property called measurability. The result proposed on p. 563 in [4] is a Fubini formulation for noncommutative geometry, emulating the classical Fubini theorem where the integral on the product space is calculated from the product of the integrals provided one of the integral exists. In recent personal communication, Professor Connes has kindly explained to the authors that the convergence he had in mind is "the convergence in the theta function formula (which in [4] is 4 lines above 2.Example a)). The assumed theta convergence is clearly stronger than the convergence of its Cesaro means and all the counter examples of the paper are about this nuance. As shown in Lemma 1.9 and Theorem 1.10, the Fubini formula indeed holds under theta convergence, by paying attention to the choice of the limiting processes, this is due to Professor Connes and we are grateful for his permission to include his proof into the paper." Our aim is to study the Fubini formula in detail. We show that the proposal in Proposition 1.1 does not hold under the condition that one of the terms is Tauberian. It does not hold either with an amended condition that both terms 2 are Tauberian. It does not hold if we ask if one or both of the terms satisfy the stronger condition that n Xk=0 λ(k, T (1 + D2)−p/2) = c · log(n + 1) + O(1). However, we show that there are natural conditions on the terms such that the Fubini formula as stated does hold. As explained above, one of them (see Condition 1.8 below) is also due to Professor Connes. To state our results we need some definitions. The following terminology was recently introduced in [2]. Definition 1.2. An operator A ∈ L1,∞ is universally measurable if ϕ(A) does not depend on the normalised2 trace ϕ on L1,∞. Equivalently (see Theorem 2.3) n Xk=0 λ(k, A) = c · log(n + 1) + O(1). Clearly being universally measurable is stronger than being Tauberian. Propo- sition 1.1 is false if we show that the same proposition is false for universally measurable operators. Definition 1.3. We say that a (p,∞)−summable spectral triple3 (A, H, D) ad- mits a noncommutative integral if, for every T ∈ A, the operator T (1 + D2)−p/2 is universally measurable. In this case, we set (T ) = ϕ(T (1 + D2)−p/2), T ∈ A for every normalised trace ϕ on L1,∞. The following condition is an analogue of the heat semigroup asymptotics found in Lemma 1.9.2 in [10]. It is satisfied by all commutative spectral triples of Riemannian manifolds (see the proof of Proposition 3.23 and Theorem 3.24 in [17]). Noncommutative tori also satisfies this condition (see the proof of Corollary 1.6). Condition 1.4. (A, H, D) is a (p,∞)−summable spectral triple such that, for every T ∈ A, there exists ε > 0 such that c(T ) tp + O( Tr(T e−(tD)2 t → 0. 1 tp−ε ), (1.2) ) = Our main Fubini theorem can be stated as follows. 2A trace on L1,∞ is a unitarily invariant linear functional on L1,∞. It is normalised if ϕ(diag({1, 1 2 , 1 3 , · · · })) = 1. 3We refer to [4] for the definition of a spectral triple. 3 Theorem 1.5. Suppose the spectral triples (A1, H1, D1) and (A2, H2, D2) sat- isfy the Condition 1.4. Then (a) (A1, H1, D1) and (A2, H2, D2) admit a noncommutative integral. (b) (A1 ⊗ A2, H1 ⊗ H2, (D2 (c) For every T1 ∈ A1, T2 ∈ A2, and for every normalised trace ϕ on L1,∞, we 2)1/2) satisfies the Condition 1.4 and admits a noncommutative integral. 1 ⊗ 1 + 1 ⊗ D2 have p1 + p2 Γ(1 + )ϕ((T1 ⊗ T2)(1 + D2 1 ⊗ 1 + 1 ⊗ D2 2)− p1 +p2 2 ) = = Γ(1 + )ϕ(T1(1 + D2 1)− p1 2 ) · Γ(1 + p2 2 )ϕ(T2(1 + D2 2)− p2 2 ). 2 p1 2 In particular, a Fubini formula holds for noncommutative tori, for sphere S2 and for the quantum group SUq(2). The proofs for noncommutative tori extend the idea used in the proof of the main result of [15]. Corollary 1.6. Let (A1, H1, D1) and (A2, H2, D2) be noncommutative tori. For every T1 ∈ A1, T2 ∈ A2, and for every normalised trace ϕ on L1,∞, we have p1 + p2 Γ(1 + )ϕ((T1 ⊗ T2)(1 + D2 1 ⊗ 1 + 1 ⊗ D2 2)− p1+p2 2 ) = = Γ(1 + )ϕ(T1(1 + D2 1)− p1 2 ) · Γ(1 + p2 2 )ϕ(T2(1 + D2 2)− p2 2 ). 2 p1 2 2 p1 2 Corollary 1.7. Let (A1, H1, D1) and (A2, H2, D2) be spectral triples which cor- respond either to sphere S2 or to the quantum group SUq(2). For every T1 ∈ A1, T2 ∈ A2, and for every normalised trace ϕ on L1,∞, we have p1 + p2 Γ(1 + )ϕ((T1 ⊗ T2)(1 + D2 1 ⊗ 1 + 1 ⊗ D2 2)− p1+p2 2 ) = = Γ(1 + )ϕ(T1(1 + D2 1)− p1 2 ) · Γ(1 + p2 2 )ϕ(T2(1 + D2 2)− p2 2 ). Condition 1.8, Theorems 1.9 and 1.10 below were suggested by Professor Connes. We are grateful for this valuable addition to the paper. Condition 1.8. (A, H, D) is a (p,∞)−summable spectral triple such that, for every T ∈ A, we have tpTr(T e−(tD)2 ) → c(T ), (1.3) In what follows, we use a notation ωu = ω ◦ Pu, u > 0. We refer the reader to Section 2 for the definition of Dixmier traces. Theorem 1.9. Let ω = ω◦ M be a state on L∞(0,∞). Suppose that the spectral triple (A1, H1, D1) (or (A2, H2, D2)) satisfies Condition 1.8. For every T1 ∈ A1, T2 ∈ A2, we have t → 0. p1 + p2 Γ(1 + 1 ⊗ 1 + 1 ⊗ D2 )Trωp1+p2 ((T1 ⊗ T2)(1 + D2 p2 2 ) · Γ(1 + 2 1)− p1 )Trωp1 (T1(1 + D2 2 p1 2 )Trωp2 (T2(1 + D2 2)− p2 2 ). 2)− p1+p2 2 ) = = Γ(1 + 4 Theorem 1.9 allows us to state another version of Fubini theorem as follows. Theorem 1.10. Suppose that the spectral triple (A1, H1, D1) (or (A2, H2, D2)) satisfies the Condition 1.8. For every T1 ∈ A1, T2 ∈ A2, and for every Dixmier trace Trω ∈ M4 on L1,∞, we have Γ(1 + p1 + p2 2 )Trω((T1 ⊗ T2)(1 + D2 1 ⊗ 1 + 1 ⊗ D2 2)− p1+p2 2 ) = = Γ(1 + p1 2 )Trω(T1(1 + D2 1)− p1 2 ) · Γ(1 + p2 2 )Trω(T2(1 + D2 2)− p2 2 ). It is important to note the difference between Conditions 1.4 and 1.8 and the difference between the assertions of Theorems 1.5 and 1.10. Indeed, Theorem 1.5 holds for arbitrary traces on L1,∞, while Theorem 1.10 holds for a certain subclass M in the class of Dixmier traces. Theorem 1.10 does not hold for some Dixmier traces outside of the subclass M. Condition 1.4 is stronger than universal measurability. Our second result complements Theorem 1.5 by stating that universal measurability is not suf- ficient for a Fubini theorem. In fact, the counterexample involves the nicest possible situation where the noncommutative integral is a normal functional on the algebra A. Definition 1.11. Suppose that (A, H, D) admits a noncommutative integral. We say that the noncommutative integral is normal if the mapping T → (T ), T ∈ A, is continuous in the weak operator topology. Theorem 1.12. There exists a (1,∞)−summable spectral triple (A, H, D) such that (a) D has simple spectrum Z+. (b) A is generated by a unitary operator U and is finite dimensional. (c) (A, H, D) admits a normal noncommutative integral. (d) (A ⊗ A, H ⊗ H, (D2 ⊗ 1 + 1 ⊗ D2)1/2) admits a normal noncommutative (e) For every normalised trace ϕ on L1,∞, we have integral. ϕ((U⊗U−1)(1+D2⊗1+1⊗D2)−1) 6= 0, ϕ(U (1+D2)− 1 2 ) = ϕ(U−1(1+D2)− 1 2 ) = 0. Corollary 1.13. In the setting of Theorem 1.12, there exists a positive element T ∈ A such that ϕ((T ⊗ T )(1 + D2 ⊗ 1 + 1 ⊗ D2)−1) 6= for every normalised trace ϕ on L1,∞. (ϕ(T (1 + D2)− 1 2 ))2 π 4 4Here, M is a subclass of Dixmier traces specified in the next section. 5 Our second counterexample shows that volume in noncommutative geometry is not necessarily well behaved under the product operation on spectral triples, even with the strong condition of universal measurability. Theorem 1.14. There exists an operator D such that (a) (1 + D2)−1/2 ∈ L1,∞ is universally measurable. (b) (1 + D2 ⊗ 1 + D2 ⊗ 1)−1 ∈ L1,∞ is universally measurable. (c) For every normalised trace ϕ on L1,∞, we have ϕ((1 + D2 ⊗ 1 + 1 ⊗ D2)−1) > π 4 (ϕ((1 + D2)− 1 2 ))2. Our final counterexample (proved in Appendix Appendix B) shows that it does not suffice to impose Condition 1.4 only on one spectral triple. It also shows that the assertion of Theorem 1.10 fails for some Dixmier trace (outside of the class M). Theorem 1.15. There exist spectral triples (A1, l2, D) and (C, l2, D) such that (a) D has simple spectrum Z+. In particular, the spectral triple (C, l2, D) sat- isfies the Condition 1.4. (b) There exists an operator T1 ∈ A1 and a (Dixmier) trace ϕ on L1,∞ such that ϕ((T1 ⊗ 1)(1 + D2 ⊗ 1 + 1 ⊗ D2)−1) 6= π 4 ϕ(T1(1 + D2)−1/2). We are grateful to Professor Connes for his kind explanation of noncommu- tative Fubini formula. We are also grateful to A. Carey, Y. Kuznetsova, S. Lord and A. Rennie for useful comments on the earlier versions of this manuscript. The authors would also like to mention that the idea of the proof of Lemma 5.1 was conveyed to them by the late N. Kalton. 2. Preliminaries The standard trace on L(H) is denoted by Tr. Fix an orthonormal basis in H (the particular choice of basis is inessential). We identify the algebra l∞ of bounded sequences with the subalgebra of all diagonal operators with respect to the chosen basis. We set l1,∞ = L1,∞ ∩ l∞. For a given sequence x ∈ l∞, we denote the corresponding diagonal operator by diag(x). Definition 2.1. A trace on L1,∞ is a unitarily invariant linear functional ϕ : L1,∞ → C. Traces on L1,∞ satisfying the condition ϕ(T S) = ϕ(ST ), T ∈ L1,∞, S ∈ L(H). 6 The latter may be reinterpreted as the vanishing of the linear functional ϕ on the commutator subspace [L1,∞,L(H)] = span{ST − T S, T ∈ L1,∞, S ∈ L(H)}. An example of a trace on L1,∞ is a Dixmier trace that we now explain (we use the definition from [18], which, according to Theorem 17 in [18], produces exactly the same class of traces on L1,∞ as the one in [4]). Namely, for every ultrafilter ω, the functional Trω defined on the positive cone of L1,∞ by the formula µ(k, A), log(n + 2) 0 ≤ A ∈ L1,∞, Trω(A) = lim n→ω Xk=0 is additive and, therefore, extends to a positive unitarily invariant linear func- tional on L1,∞ called a Dixmier trace. In order to properly state Theorem 1.10, we need a smaller subclass M of Dixmier traces. Let ω be a state on the algebra L∞(0,∞) which satisfies the condition ω = ω◦M (see p.35 in [1]). Here, the linear operator M : L∞(0,∞) → L∞(0,∞) is given by the formula (2.1) 1 n (M x)(t) = 1 log(t) t 1 x(s) ds s , t > 0. The functional Trω is defined on the positive cone of L1,∞ by the formula Trω(A) = ω(cid:16)t → 1 log(1 + t) t 0 µ(s, A)ds(cid:17), 0 ≤ A ∈ L1,∞. This functional is additive and, therefore, extends to a positive unitarily invari- ant linear functional on L1,∞ (see e.g. [1]). Let the group (R+,·) act on L∞(0,∞) by the formula u → Pu, (Pux)(t) = x(tu), u, t > 0. Note that M ◦ Pu = Pu ◦ M (see a similar formula (3) in [19]). In particular, Trωu = Trω◦Pu is also a positive unitarily invariant linear functional on L1,∞. We set M = {Trω : ω = ω ◦ M, ω = ω ◦ Pu, u > 0}. It is important to note that ω in this paragraph can never be an ultrafilter. However, Trω is still a Dixmier trace according to the main result of [18]. The following assertion is Theorem 3 in [1]. Theorem 2.2. Let ω = ω ◦ M be a state on the algebra L∞(0,∞). If the triple (A, H, D) is (p,∞)−summable, then Γ(1 + )Trω◦Pp (T (1 + D2)− p p 2 2 ) = ω(cid:16)t → t−pTr(T e−t−2D2 )(cid:17), T ∈ A. The following theorem provides the convenient spectral description for uni- versally measurable operators referred to earlier. It was originally proved in [7] for normal operators and, then in [12] and [8] for arbitrary operators (see also [13]). For accessible proof, we refer the interested reader to Theorem 10.1.3 in [16] and its proof in Chapter 5 in [16]. 7 Theorem 2.3. For A ∈ L1,∞, the following conditions are equivalent. (a) We have ϕ(A) = c for every normalised trace ϕ on L1,∞. (b) We have n λ(m, A) = c · log(n + 1) + O(1), n ≥ 0. Xm=0 In particular, A ∈ [L1,∞,L(H)] if and only if n Xm=0 λ(m, A) = O(1), n ≥ 0. Every universally measurable operator is Tauberian (that is, Dixmier-measurable [18]). For various sorts of measurability results in noncommutative geometry, we refer the interested reader to papers [19, 20] and to the book [16]. 3. Proof of Theorems 1.5 and 1.10 Lemma 3.1. Let Φ : (0,∞) → (0, 1) be such that (i) Φ is convex, decreasing and positive. (ii) Φ(0) = 1 and ∞ 1 Φ(t) dt t < ∞, 1 0 ( 1 t − 1)(1 − Φ(t))dt < ∞. For every 0 ≤ V ∈ L1,∞, we have k min{V Φ((nV )−1), 1 n}k1 = O(1), n → ∞ )+(1 − Φ((nV )−1))k1 = O(1), Proof. It is easy using (i) to check that the functions k(V − 1 n n → ∞. x → xΦ(x−1), x → (x − 1)+(1 − Φ(x−1)) increase on (0,∞). For simplicity of computations, let kV k1,∞ = 1. Let W ∈ L1,∞ be an operator commuting with V such that 0 ≤ V ≤ W and such that µ(k, W ) = 1 k+1 , k ≥ 0. It follows that k min{V Φ((nV )−1), 1 k + 1 )}∞k=nk1 + k{ ≤ k{ k + 1 Φ( n 1 1 n}k1 ≤ k min{W Φ((nW )−1), n}k1 ≤ dt t k=0k1 ≤ ∞ 1 n}n−1 + 1 Φ( t n ) n (ii) = O(1) and, similarly, k(V − 1 n )+(1 − Φ((nV )−1))k1 ≤ k(W − 1 n )+(1 − Φ((nW )−1))k1 ≤ 8 ≤ k{( 1 k + 1 − 1 n )(1 − Φ( k + 1 n k=0k1 ≤ n ))}n−1 0 ( 1 t − 1 n )(1 − Φ( t n ))dt (ii) = O(1). The following lemma extends Proposition 6 in [2]. Lemma 3.2. Let 0 ≤ V ∈ L1,∞ and let A ∈ L(H). Let Φ be as in Lemma 3.1. The following conditions are equivalent (a) ϕ(AV ) = c for every normalised trace ϕ on L1,∞. (b) We have Tr(AV Φ((nV )−1)) = c log(n) + O(1), n → ∞. Proof. It is clear that AV Φ((nV )−1) − AV EV [ 1 n ,∞) = = AV Φ((nV )−1)EV [0, 1 n ) + AV (Φ((nV )−1) − 1)EV [ 1 n ,∞). Therefore, 1 ,∞)) ≤ kAV Φ((nV )−1)− AV EV [ n 1 )k1 + kV (Φ((nV )−1) − 1)EV [ n 1 n 1 n ,∞)k1 ≤ ,∞)k1(cid:17) ≤ ,∞)k1+ 1 n Tr(AV Φ((nV )−1))− Tr(AV EV [ ≤ kAk∞(cid:16)kV Φ((nV )−1)EV [0, ≤ kAk∞(cid:16)kV Φ((nV )−1)EV [0, 1 n ≤ kAk∞(cid:16)k min{V Φ((nV )−1), It follows from Lemma 3.1 that + 1 )k1 + k(V − n 1 nk(Φ((nV )−1) − 1)EV [ 1 n 1 n}k1+k(V − )(Φ((nV )−1) − 1)EV [ 1 ,∞)k1(cid:17) ≤ n )+(Φ((nV )−1)−1)k1+ 1 nkEV [ 1 n ,∞)k1(cid:17). Tr(AV Φ((nV )−1)) − Tr(AV EV [ 1 n ,∞)) = O(1), n → ∞. It follows now from Lemma 8 in [2] that Tr(AV Φ((nV )−1)) − n Xk=0 λ(k, AV ) = O(1). The assertion follows now from Theorem 2.3. Lemma 3.3. If a spectral triple (A, H, D) satisfies the Condition 1.4, then it ad- mits a noncommutative integral. More precisely, if (A, H, D) is (p,∞)−summable, then c(T ) = Γ(1 + )ϕ(T (1 + D2)− p 2 ), p 2 T ∈ A, for every normalised trace ϕ on L1,∞. Here, c(T ) is the number which appears in (1.2). 9 Proof. It follows from (1.2) that Tr(T e−t2(1+D2)) = c(T ) tp + O( 1 tp−ε ), t → 0. Substituting t 1 p instead of t, we infer that Tr(T e−(t(1+D2) p 2 ) 2 p ) = c(T ) t + O( 1 t1−ε ), t → 0. (3.1) Set We have Φ(s) = 1 2 ) ∞ s Γ(1 + p e−t 2 p dt, s > 0. 1 2 p 2 ) s e−(t(1+D2) p dt = (1 + D2)−p/2(cid:16)Φ(s(1 + D2)p/2) − Φ((1 + D2)p/2)(cid:17), where the integral is understood in the Bochner sense in L1. In particular, we have Tr(T e−(t(1+D2) p 2 ) 2 p )dt = 1 s = Tr(T (1 + D2)− p 2 Φ((1 + D2) Integrating both sides in (3.1) over t ∈ [s, 1] and dividing by Γ(1 + p that 2 )) − Tr(T (1 + D2)− p 2 Φ(s(1 + D2) p p 2 )). 2 ) we infer Tr(T (1 + D2)− p 2 Φ(s(1 + D2) p 2 )) = c(T ) Γ(1 + p 2 ) log( 1 s ) + O(1), s → 0. Observe, that Φ satisfies the conditions of Lemma 3.1. The assertion follows now from Lemma 3.2 (as applied to V = (1 + D2)− p 2 and c = c(T ) Γ(1+ p 2 ) ). Proof of Theorem 1.5. Recall an abstract equality (which holds for all bounded operators T1, T2) (T1 ⊗ T2)e−t2(D2 1⊗1+1⊗D2 2)) = T1e−t2D2 1 ⊗ T2e−t2D2 2 . Take now T1 ∈ A1 and T2 ∈ A2. By Condition 1.4, we have c(T2) Tr(T1e−t2D2 tp2 tp1−ε ), Tr(T2e−t2D2 c(T1) tp1 +O( 2 ) = 1 ) = 1 +O( 1 tp2−ε ), t → 0. It follows that Tr((T1 ⊗ T2)e−t2(D2 1⊗1+1⊗D2 2))) = c(T1)c(T2) tp1+p2 + O( as t → 0. Thus, the spectral triple (A1 ⊗ A2, H1 ⊗ H2, (D2 satisfies the Condition 1.4 and 1 tp1+p2−ε ) 1 ⊗ 1 + 1 ⊗ D2 2)1/2) The assertion follows now from Lemma 3.3. c(T1 ⊗ T2) = c(T1)c(T2). 10 Proof of Theorem 1.9. Recall an abstract equality (which holds for all bounded operators T1, T2) (T1 ⊗ T2)e−t2(D2 1⊗1+1⊗D2 2)) = T1e−t2D2 1 ⊗ T2e−t2D2 2 . Take now T1 ∈ A1 and T2 ∈ A2. By Condition 1.8, we have c(T1) tp1 Tr(T1e−t2D2 1 ) = t → 0. Taking the trace and replacing t with t−1, we obtain that t−(p1+p2)Tr((T1⊗T2)e−t−2(D2 1⊗1+1⊗D2 + o( ), 1 tp1 2))) = (c(T1)+o(1))·t−p2 Tr(T2e−t−2D2 2 ), t → ∞. In particular, applying ω to the both sides of the equality, we arrive at ω(cid:16)t → t−(p1+p2)Tr((T1⊗T2)e−t−2(D2 1⊗1+1⊗D2 2 )(cid:17). 2 )))(cid:17) = c(T1)ω(cid:16)t → t−p2 Tr(T2e−t−2D2 It follows from Theorem 2.2 (applied to both sides of the equality) that Γ(1 + p1 + p2 2 )Trωp1+p2 ((T1 ⊗ T2)(1 + D2 1 ⊗ 1 + 1 ⊗ D2 2)− p1+p2 2 ) = = c(T1) · Γ(1 + p2 2 )Trωp2 (T2(1 + D2 2)− p2 2 ). Again using Theorem 2.2 (applied to the spectral triple (A1, H1, D1)), we infer that c(T1) = Γ(1 + This concludes the proof. p1 2 )Trωp1 (T1(1 + D2 1)− p1 2 ). Proof of Theorem 1.10. If ω = ω ◦ Pu, u > 0, then = Trω◦Pp1+p2 = Trω◦Pp2 Trω◦Pp1 = Trω. The assertion follows now from Lemma 1.9. 4. Physically relevant examples We supply 3 examples which satisfy the Condition 1.4. The first example is a sphere -- the simplest possible non-flat manifold. The second example is a noncommutative torus. The third and the most technically involved example is the quantum group SUq(2). The following elementary lemma is needed in all 3 examples. We incorporate the proof for convenience of the reader. Lemma 4.1. We have 11 (a) (b) (c) le−l2t2 Xl∈Z = t−2 + O(t−1), t → 0. e−l2t2 = 1 2 π t Xl∈Z + O(1), t → 0. l2e−l2t2 = 1 2 π 2t3 + O(t−2), t → 0. Xl∈Z Proof. Though the second and third equalities can be derived from the Poisson summation formula, this method gives nothing good for the first equality. We provide an elementary proof of the first equality. The proofs of the second and third are similar. The function s → se−s2t2 the function increases on the interval (0, 1 t√2 ( 1 t√2 ,∞). It follows that admits its maximum at the point s = 1 t√2 . Thus, ) and decreases on the interval ⌊ 1 t√2⌋−1 Xl=1 l l−1 se−s2t2 ds ≤ ⌊ 1 t√2⌋−1 Xl=0 le−l2t2 ≤ ⌊ 1 t√2⌋−1 Xl=0 l+1 l se−s2t2 ds. Thus, ⌊ 1 t√2⌋−1 Xl=0 le−l2t2 t√2⌋ − ⌊ 1 0 se−s2t2 (cid:12)(cid:12)(cid:12) Similarly, (cid:12)(cid:12)(cid:12) It follows that ∞ Xl=⌈ 1 t√2⌉+1 t√2 ⌋ ⌊ 1 t√2 ⌋−1 ds(cid:12)(cid:12)(cid:12) ≤ ⌊ 1 − ∞ ⌈ 1 t√2 ⌉ le−l2t2 se−s2t2 se−s2t2 ds ≤ sup s>0 se−s2t2 = 2 e− 1 t√2 . 2 e− 1 t√2 . ds(cid:12)(cid:12)(cid:12) ≤ =(cid:16) le−l2t2 ∞ Xl=0 =(cid:16) ⌊ 1 = ∞ 0 0 t√2⌋ se−s2t2 se−s2t2 ⌊ 1 t√2⌋−1 ∞ t√2⌉+1 le−l2t2(cid:17) +(cid:16) Xl=0 Xl=⌈ 1 ds + O(t−1)(cid:17) +(cid:16) ∞ ⌈ 1 t√2⌉ ds + O(t−1) = t−2 ∞ 0 le−l2t2(cid:17) +(cid:16) ⌈ 1 t√2⌉ Xl=⌊ 1 t√2⌋ le−l2t2(cid:17) = se−s2t2 ds + O(t−1)(cid:17) + O(t−1) = 1 2t2 + O(t−1). se−s2 ds + O(t−1) = 12 It follows that 0 Xl=−∞ le−l2t2 = ∞ Xl=0 le−l2t2 = 1 2t2 + O(t−1). Adding the last 2 formulae, we conclude the proof. 4.1. Example: sphere S2 We briefly recall the construction of a spectral triple on sphere S2. Interested reader is referred to [11] for details. Let s = (s1, s2, s3) be the point on the sphere S2 expressed in Cartesian coor- dinates. Define stereographic coordinates by the formula (x, y) = ( s1 ). 1−s3 Denote z = x + iy and q(x, y) = 1 + x2 + y2. The image of Lebesgue measure on sphere under stereographic projection is 4q−2(x, y)dxdy. Define unbounded operators D1 and D2 on (the subspace of all Schwartz functions in) the Hilbert space L2(R2, 4q−2(x, y)dxdy) by setting D1 = 1 ∂ ∂y . Consider now i a couple (A, A∗) of formally adjoint unbounded operators defined on (the sub- space of all Schwartz functions in) the Hilbert space L2(R2, 4q−2(x, y)dxdy) by the formula ∂x , D2 = 1 ∂ , s2 1−s3 i A = 1 2 Mq(D1 − iD2) + i 2 M¯z, A∗ = 1 2 Mq(D1 + iD2) + i 2 Mz. Our Hilbert space is C2⊗L2(R2, 4q−2(x, y)dxdy). Our von Neumann algebra is L∞(S2) with a smooth subalgebra C∞(S2). Its representation is given by the formula π(f ) = 1 ⊗ Mf◦Stereo−1, where Stereo denotes stereographic projection. Our Dirac operator is then defined by the formula (see equation (9.52) in [11]) D = e12 ⊗ A + e21 ⊗ A∗, where e12, e21 ∈ M2(C) are matrix units. It is established in Corollary 9.26 and Proposition 9.28 in [11] that D admits an orthonormal eigenbasis. In particular, D is self-adjoint. By Corollary 9.29 in [11], the constructed spectral triple (π(L∞(S2)), C2 ⊗ L2(R2, 4q−2(x, y)dxdy), D) In the following lemma, we show that it satisfies the is (2,∞)−summable. Condition 1.4. Lemma 4.2. For every f ∈ L∞(S2), we have Tr(π(f )e−t2D2 ) = t−2 · 1 4π S2 f (s)ds + O(t−1), t → 0. Proof. Recall how the group SU(2) acts on extended complex plane. g =(cid:18) a −¯b b ¯a(cid:19) ∈ SU (2), g(z) = az + b −¯bz + ¯a , z ∈ C. (4.1) 13 This action results in the unitary representation τ of the group SU(2) on the 1 Hilbert space C2 ⊗ L2(S2) by the formula 2 (cid:18)ψ1(g−1(z)) ψ2(cid:19)(cid:17)(z) =(cid:16) b¯z + ¯a (cid:16)τ (g)(cid:18)ψ1 ψ2(g−1(z))(cid:19) , ψ1, ψ2 ∈ L2(R2, 4q−2(x, y)dxdy). ¯bz + a(cid:17) The key fact (Proposition 9.27 in [11]) is that Dirac operator D commutes with the τ (g) for every g ∈ SU (2). projection of D. Since P commutes with τ (g), it follows that Let f ∈ L∞(S2) and denote for brevity F = f ◦ Stereo−1. Let P be a spectral Tr((1 ⊗ MF )P ) = Tr((1 ⊗ MF )τ (g) · τ (g−1)P ) = Tr((1 ⊗ MF )τ (g) · P τ (g−1)) = Note that Thus, = Tr(τ (g−1)(1 ⊗ MF )τ (g) · P ). τ (g−1)(1 ⊗ MF )τ (g) = MF◦g. Tr(π(f )P ) = Tr((1 ⊗ MF )P ) = Tr((1 ⊗ Mf◦g)P ), g ∈ SU (2). Since the latter equality holds for every g ∈ SU (2), it follows that Tr(π(f )P ) = SU(2) Tr((1 ⊗ MF◦g)P )dg = Tr((1 ⊗SU(2) MF◦gdg)P ), where dg is the Haar measure on SU(2). The action of SU(2) as given in (4.1) is conjugated (by means of stereographic projection, see Section 1.4 in [9]) to the action of SU(2) on sphere S2 by rotations. It follows that and, therefore, SU(2) (F ◦ g)dg = 1 4π S2 f (s)ds Tr(π(f )P ) = 1 4π S2 f (s)ds · Tr(P ). According to Corollary 9.29 in [11], spectrum of D is Z\{0} and, for every Tr(π(f )e−t2D2 l ∈ Z\{0}, Tr(ED{l}) = l. It follows that ) = Xl∈Z\{0} Tr(π(f )ED{l}) = e−t2l2 = Xl∈Z\{0} The assertion follows now from Lemma 4.1 (a). 14 Tr(π(f )e−t2D2 ED{l}) = 1 4π S2 f (s)ds ·Xl∈Z le−t2l2 . 4.2. Example: noncommutative torus We briefly recall a spectral triple for the noncommutative torus (originally introduced in Section II.2.β in [4]). After that, we show that the triple satisfies the Condition 1.4. Let Θ ∈ Mp(R), 1 < p ∈ N, be an anti-symmetric matrix. Let AΘ be the k=1 satisfying the conditions universal ∗−algebra generated by unitaries {Uk}p Uk2Uk1 = eiΘk1 ,k2 Uk1Uk2 , 1 ≤ k1, k2 ≤ p. Define a linear functional τ : AΘ → C by setting τ (U n1 1 ··· U np p ) = 0 unless (n1,··· , np) = 0. It can be demonstrated that τ is positive, that is τ (x∗x) ≥ 0 for x ∈ AΘ. We equip linear space AΘ with an inner product defined by the formula hx, yi = τ (xy∗), x, y ∈ AΘ. Natural action λ of AΘ on pre-Hilbert space (AΘ,h·,·i) by left multiplications extends to the action on the completed Hilbert space. The weak∗ closure of λ(AΘ) is denoted by L∞(Tp Θ) and τ extends to a faithful normal tracial state on L∞(Tp Θ) is naturally identified with L2(Tp Θ). The Hilbert space where L∞(Tp A natural spectral triple for the noncommutative torus is given as follows.5 Θ) and take λ(AΘ) to be the subalgebra of smooth elements. Let Θ). Define Set A = L∞(Tp m(p) = 2⌊ p self-adjoint operators Dk, 1 ≤ k ≤ p, on the Hilbert space L2(Tp Θ) and π(x) = 1⊗ Mx, x ∈ L∞(Tp 2 ⌋. Set H = Cm(p) ⊗ L2(Tp Θ) by setting Θ, τ ). Dk : U n1 1 ··· U np p → nkU n1 1 ··· U np p , (n1,··· , np) ∈ Zp. Those operators commute. Dirac operator D acts on the Hilbert space H by the setting p D = γk ⊗ Dk, Xk=1 where γk ∈ Mm(p)(C), 1 ≤ k ≤ p, are Pauli matrices. Lemma 4.3. For every x ∈ L∞(Tp θ), we have Proof. Let {uk}k∈Z be the standard basis in L2(Tp Tr(π(x)e−t2D2 ) = uk = U k1 1 U k2 2 ··· U kp p π tp 2 m(p) τ (x) + O( 1 tp−1 ). Θ, τ ), that is p , k = (k1,··· , kp) ∈ Zp. 5The C∗−algebra λ(AΘ) k·k∞ is isomorphic to a universal C∗−algebra constructed by Davidson (see pp.166-170 in [6]). 15 If {em}m(p) m=1 is the standard unit basis in Cm(p), then the elements em ⊗ uk, 1 ≤ m ≤ m(p), k ∈ Zp form an orthonormal basis in Cm(p) ⊗ L2(Tp Θ). We have (D2)(em ⊗ uk) = k2em ⊗ uk and hπ(x)(em ⊗ uk), em ⊗ uki = τ (x) for every x ∈ L∞(Tp θ ) and for every 1 ≤ m ≤ p, k ∈ Zp. Hence, Tr(π(x)e−t2D2 ) = m(p) Xm=1 Xk∈Zphπ(x)e−t2D2 (em ⊗ uk), em ⊗ uki = = m(p)τ (x) Xk∈Zp e−t2k2 = m(p)τ (x)(Xk∈Z e−t2k2 )p. The assertion follows now from Lemma 4.1 (b). Proof of Corollary 1.6. By Lemma 4.3, the noncommutative torus satisfies the Condition 1.4. The assertion follows now from Theorem 1.5. 4.3. Example: quantum group SUq(2) In what follows, O(SUq(2)) is the algebraic linear span of all words in a, c, a∗, c∗ with the following cancellation rules6 a∗a + c∗c = 1, aa∗ + q2cc∗ = 1, ac = qca, ac∗ = qc∗a, cc∗ = c∗c. Here, the parameter q takes value from the interval [−1, 1]. For q = 1, the algebra O(SUq(2)) is commutative and (its von Neumann envelope) equals to L∞(SU (2)). In what follows, we assume q ∈ (−1, 1). It follows from Proposition IV.4 in [14] that the elements {ancm(c∗)r, cm(c∗)r(a∗)n+1}m,n,r∈Z+ form a Hamel basis in O(SUq(2)). Define a linear functional7 τ on O(SUq(2)) by setting τ (ancm(c∗)r) = τ (cm(c∗)r(a∗)n) = 0, (m, n, r) 6= (0, 0, 0), τ (1) = 1. The algebra O(SUq(2)) acts on the Hilbert space H which is the completion of O(SUq(2)) with respect to the inner product (x, y) → h(xy∗). Here, h is the Haar state on SUq(2) (defined in Theorem IV.14 in [14]). Let l, m, n ∈ 1 2 Z+ be such that l ≥ 0, m,n ≤ l and l − m, l − n ∈ Z. Let tl m,n be the orthonormal basis in O(SUq(2)) constructed in Theorem IV.13 in [14]. By construction of the Hilbert space H, these elements also form an 6Here, we are using the notations from [14]. The definition appears on p.102 in [14]. The same definition is used by Connes (see formula (18) in [5]) and Chakraborty-Pal (see p.2 in [3]). However, those authors use the notation α for a and β for c. 7This is a trace on O(SUq(2)), not the Haar state. However, we don't need the tracial property of τ. Connes (see Theorem 4 in [5]) considered this functional on a larger algebra. 16 orthonormal basis in H. Connes defined Dirac operator D on the Hilbert space H in [5] (see formula (22) on p.8 there) by the formula Dtl m,n = 2l(2δ0(l − m) − 1)tl m,n. In this text, we are not interested in the sign of D, but only in its absolute value given by the formula m,n. Lemma 4.4. For every x ∈ O(SUq(2)), we have m,n = 2ltl Dtl Tr(Mxe−(tD)2 ) = 1 2 π 4t3 τ (x) + O(t−2), t → 0. Proof. Step 1: Let j ∈ 1 2 Z+ and let (r, s) 6= (0, 0). We claim that Tr(Mtj r,s e−(tD)2 ) = 0. Let A[·,·] be the linear subspace in O(SUq(2)) defined on p.105 in [14]. By m,n ∈ A[−2m,−2n]. m,n ∈ A[−2r − 2m,−2s− r,stl m,n is orthogonal Lemma IV.11 in [14], we have tj r,s ∈ A[−2r,−2s] and tl Using formula (26) on p.105 in [14], we infer that tj 2n]. It follows now from formulae (47) and (56) in [14] that tj to tl r,stl m,n (because (r, s) 6= (0, 0)). Thus, ) = Xl,m,n Step 2: Let j ∈ Z+. We claim that e−(tD)2 Tr(Mtj r,s e−4t2l2 r,stl htj m,n, tl m,ni = 0. Tr(M(cc∗)j e−(tD)2 ) = O(t−2). In what follows, pl = ED{2l}. Using formulae (2.1) -- (2.9) in [3] (or formulae (19) -- (21) in [5]), we infer that (plMcc∗pl)tl m,n = c(m, n, l)tl m,n, where c(m, n, l) = O(qm+l) + O(qn+l). Taking into account that q < 1, we obtain l Xm,n=−l Tr(plMcc∗pl) = It follows that c(m, n, l) = l Xm,n=−l(cid:16)O(qm+l) + O(qn+l)(cid:17) = O(l). Tr(M(cc∗)j pl) = Tr(plM(cc∗)j pl) ≤ kbk2j−2 ∞ Tr(plMcc∗pl) = O(l). 17 Hence, Tr(M(cc∗)j e−(tD)2 e−4t2l2 ) = Xl∈ 1 2 Z+ Tr(M(cc∗)j pl) = Xl∈ 1 2 Z+ e−4t2l2 · O(l). Thus, Tr(M(cc∗)j e−(tD)2 ) ≤ O(1) ·(cid:16) Xl∈ 1 2 Z+ le−4l2t2(cid:17) = O(1) ·(cid:16) Xl∈Z+ le−l2t2(cid:17). Replacing the sum with an integral, we conclude the proof in Step 2. Step 3: Let r, s ∈ 1 (47) in [14], we have 2 Z be such that (r, s) 6= (0, 0). By Step 1 and formula Tr(Mxe−(tD)2 ) = 0, x ∈ A[−2r,−2s]. By Step 2 and Proposition 10 (i) in [14], we have Tr(Mxe−(tD)2 ) = O(t−2), x ∈ A[0, 0], τ (x) = 0. A trivial computation shows that Tr(e−(tD)2 ) = Xl∈ 1 2 Z+ (2l + 1)2e−4l2t2 = O(t−2) + l2e−l2t2 . 1 2Xl∈Z It follows from Lemma 4.1 (c) that Tr(e−(tD)2 ) = 1 2 π 4t3 + O(t−2). Thus, for every m, n ∈ Z, we have π 4t3 τ (x) + O(t−2), The assertion follows now from formula (32) in [14]. Tr(Mxe−(tD)2 ) = 1 2 x ∈ A[m, n]. Proof of Corollary 1.7. By Lemma 4.2 and Lemma 4.4, spectral triples corre- sponding to the sphere S2 and to the quantum group SUq(2) satisfy the Condi- tion 1.4. The assertion follows now from Theorem 1.5. 5. Proof of Theorem 1.12 The following lemma provides a convenient formula for the sum of the first n eigenvalues of diag(x) ∈ L1,∞. Lemma 5.1. If x ∈ l∞ is such that x(k) ≤ 1 k+1 , k ≥ 0, then x(k) + O(1). λ(m, x) = n Xm=0 n Xk=0 18 Proof. Suppose first that x ≥ 0. It is clear that n Xk=0 x(k) ≤ n Xk=0 µ(k, x). On the other hand, there exists a set An ⊂ Z+ such that A = n + 1 and such that n Xk=0 µ(k, x) = Xk∈An Xk=0 x(k) + Xk∈An,k≥n n ≤ x(k) = Xk∈An,k≤n k + 1 ≤ 1 n x(k) + Xk=0 x(k) + Xk∈An,k≥n Xk=n+1 k + 1 ≤ 2n+1 1 x(k) ≤ x(k) + 1. n Xk=0 A combination of the latter estimates yields the assertion under the additional assumption that x ≥ 0. For an arbitrary x ∈ l1,∞, there exist 0 ≤ xp ∈ l1,∞, 1 ≤ p ≤ 4, such that x = x1 + ix2 + i2x3 + i3x4. If x(k) ≤ 1 5.7.5 in [16] that k+1 , k ≥ 0, then also xp(k) ≤ 1 k+1 , k ≥ 0. It follows from Lemma λ(m, x) = n Xm=0 ip−1 4 Xp=1 n Xm=0 λ(m, xp) + O(1). Applying the assertion for positive operators xp, 1 ≤ p ≤ 4, we infer that λ(m, x) = n Xm=0 ip−1 4 Xp=1 n Xk=0 xp(k) + O(1) = x(k) + O(1). n Xk=0 This concludes the proof. Lemma 5.2. If x ∈ l∞(Z2) is such that x(k, l) ≤ 1 1+k2+l2 , k, l ≥ 0, then n ⌊n1/2⌋ λ(m, x) = Xm=0 Xk,l=0 Proof. Define a bijection α2 : Z+ → Z2 z ∈ l∞ by setting z = x ◦ α2. It follows from Lemma Appendix A.2 that x(k, l) + O(1). + as in Lemma Appendix A.2. Define z(m) ≤ 1 1 + α2(m)2 ≤ const m + 1 , m ≥ 0. Therefore, λ(m, x) = n Xm=0 n Xm=0 λ(m, z) L.5.1 = n Xm=0 z(m) + O(1) = n Xm=0 x(α2(m)) + O(1). 19 Note that n Xm=0 (cid:12)(cid:12)(cid:12) x(α2(m)) − Xk∈Z2 + k≤α2(n) x(k)(cid:12)(cid:12)(cid:12) ≤ Xk∈Z2 k=α2(n) + x(k) ≤ Xk∈Z2 + k=α2(n) 1 1 + k2 = 1 = 1 + α2(n)2 Xk∈Z2 + k=α2(n) 1 ≤ 1 1 + α2(n)2 α2(n) Xk=0 1 ≤ 1. It follows from Lemma Appendix A.2 that (cid:12)(cid:12)(cid:12) Xk∈Z2 k≤α2(n) + x(k)− Xk∈Z2 + k2≤n x(k)(cid:12)(cid:12)(cid:12) ≤ Xk∈Z2 + k2∈[α2(n)2,n] x(k) ≤ Xk∈Z2 + k2∈[c1n,c2n] 1 1 + k2 = O(1). We also have (cid:12)(cid:12)(cid:12) Xk2 1+k2 2≤n x(k1, k2) − X0≤k1,k2≤n1/2 ≤ X0≤k1,k2≤n1/2 k2 1+k2 2>n 1 1 + k2 1 + k2 2 ≤ x(k1, k2) ≤ x(k1, k2)(cid:12)(cid:12)(cid:12) ≤ X0≤k1,k2≤n1/2 1 + n X0≤k1,k2≤n1/2 k2 1+k2 2>n 1 1 = O(1). A combination of the latter estimates yields the assertion. For a given θ ∈ R, we define xθ ∈ l∞ by setting xθ(0) = 1 and xθ(k) = einθ, k ∈ [2n, 2n+1), n ≥ 0. Let D = diag({k}k≥0), Uθ = diag({xθ(k)}k≥0). Clearly, (1 + D2)−1/2 ∈ L1,∞. Lemma 5.3. For every θ /∈ 2πZ, we have UθD−1 ∈ [L1,∞,L(H)]. Proof. We have 2n+1−1 Xk=1 xθ(k) k = eimθ n Xm=0 2m+1−1 Xk=2m eimθ(log(2) + O(2−m)) = = O(1) + log(2) = O(1). Hence, for every m ≥ 1, we have Xk=0 m n 1 k = Xm=0 ei(n+1)θ − eiθ eiθ − 1 xθ(k) (1 + k2)1/2 = O(1). 20 By Lemma 5.1, we have λ(Uθ(1 + D2)−1/2) = O(1). m Xk=0 The assertion follows from Theorem 2.3. Lemma 5.4. For every m1, m2 ∈ Z+, we have 2m1+1−1 Xk=2m1 2m2 +1−1 Xl=2m2 Here, 1 k2 + l2 = Ξ(m1 − m2) + O(min{2−m1, 2−m2}). Ξ(m) = 2 1 2m+1 2m dtds t2 + s2 , m ∈ Z. (5.1) Proof. It is clear that 2m2 +1 2m2 Thus, It follows that 2m2 +1−1 Xl=2m2 k2 + l2 = 2m2+1 1 2m2 dt t2 + k2 ≤ 2m2+1−1 Xl=2m2 1 k2 + l2 ≤ 2m2 +1 2m2 dt t2 + k2 + 1 k2 + 22m2 . dt t2 + k2 + O(1) k2 + 22m2 . 2m1 +1−1 Xk=2m1 2m2+1−1 Xl=2m2 1 k2 + l2 = 2m2+1 2m2 2m1 +1−1 Xk=2m1 dt t2 + k2 + O(1) · 2m1+1−1 Xk=2m1 1 k2 + 22m2 . Repeating the argument, we obtain that 2m1+1−1 Xk=2m1 +O(1) · 2m2+1 2m2 2m2 +1−1 Xl=2m2 dt t2 + 22m1 1 2m2 k2 + l2 = 2m2 +1 + O(1) · 2m1 +1 2m1 2m1+1 2m1 dtds t2 + s2 + ds s2 + 22m2 + O(1) 22m1 + 22m2 . Clearly, the second and third integrals above can be estimated as O(1) · 2m2 22m1 + 22m2 , O(1) · 2m1 22m1 + 22m2 . The reference to (5.1) concludes the proof. It is obvious that 0 ≤ Ξ(m) ≤ 2−m, m ∈ Z+. (5.2) 21 Lemma 5.5. For every θ ∈ R and for every p ∈ Z, we have θ(k)x−p xp k2 + l2 = F (pθ) θ (l) log(M ) log(2) + O(1), M ∈ N. M Xk,l=1 Here, F (θ) = Xm∈Z Ξ(m)eimθ, (5.3) where Ξ is given in (5.1). Proof. Since xp θ = xpθ, it follows that we may consider only the case p = 1. Firstly, we establish the assertion for M = 2n+1 − 1, n ∈ Z+. It follows from Lemma 5.4 that 2n+1−1 Xk,l=1 xθ(k)x−1 θ (l) k2 + l2 = n = Xm1,m2=0 n 2m1 +1−1 2m2+1−1 ei(m1−m2)θ Xk=2m1 Xm1,m2=0 Xl=2m2 Ξ(m1 − m2)ei(m1−m2)θ + O(1). 1 k2 + l2 = Rearranging the summands, we obtain that n Ξ(m1 − m2)ei(m1−m2)θ = Xm1,m2=0 Xm=1 (n + 1 − m)eimθΞ(m) + n n (n + 1 − m)e−imθΞ(m). Xm=1 = (n + 1)Ξ(0) + It follows from (5.2) that Ξ(m)eimθ(cid:17)+ O(1) = nF (θ)+ O(1). n n Ξ(m1− m2)ei(m1−m2) = n(cid:16) Xm1,m2=0 This proves the assertion for M = 2n+1 − 1, n ∈ Z+. Xm=−n Now, for an arbitrary M ∈ [2n, 2n+1), we have 2n+1−1 2n+1−1 Xk,l=1 Xk,l=1 xθ(k)x−1 k2 + l2 k2 + l2 − Xk,l=1 xθ(k)x−1 θ (l) θ (l) ≤ M 1 k2 + l2 − 2n−1 Xk,l=1 1 k2 + l2 ≤ 2n+1−1 2n+1−1 ≤ 2 Xk=2n Xl=1 This concludes the proof. 1 k2 + l2 ≤ 2 2n+1−1 Xk=2n 2n+1−1 Xl=1 2−2n ≤ 4. 22 The proof of the following lemma is parallel (though, not identical) to that of Lemma 5.5. Lemma 5.6. For every θ ∈ R and for every p, q ∈ Z such that (p + q)θ /∈ 2πZ, we have xp θ(k)xq k2 + l2 = O(1), M ∈ N. θ(l) M Xk,l=1 Proof. Firstly, we establish the assertion for M = 2n+1 − 1, n ∈ Z+. It follows from Lemma 5.4 that 2n+1−1 Xk,l=1 xp θ(k)xq k2 + l2 = θ(l) n 2m1+1−1 2m2 +1−1 ei(pm1+qm2)θ Xm1,m2=0 Ξ(m1 − m2)eip(m1−m2)θei(p+q)m2θ + O(1). Xk=2m1 Xl=2m2 1 k2 + l2 = = n Xm1,m2=0 Rearranging the summands, we obtain that n Ξ(m1 − m2)eip(m1−m2)θei(p+q)m2θ = Ξ(0)ei(p+q)mθ+ n Xm=0 Xm1,m2=0 Xm=1 n + Ξ(m)eipmθ ei(p+q)m2θ + n−m Xm2=0 Ξ(m)e−ipmθ n Xm=1 ei(p+q)m2θ. n Xm2=m The assumption (p + q)θ /∈ 2πZ guarantees that ei(p+q)m2θ = O(1), n−m Xm2=0 n Xm2=m ei(p+q)m2θ = O(1), m, n ∈ Z+. Therefore, appealing to (5.2), we obtain n Xm1,m2=0 Ξ(m1 − m2)eip(m1−m2)θei(p+q)m2θ ≤ n Xm=0 Ξ(0)ei(p+q)mθ+ + n Xm=1 Ξ(m) · O(1) + In other words, we have n Xm=1 Ξ(m) · O(1) = O(1). 2n+1−1 Xk,l=1 xp θ(k)xq k2 + l2 = O(1). θ(l) This proves the assertion for M = 2n+1 − 1, n ∈ Z+. 23 Now, for an arbitrary M ∈ [2n, 2n+1), we have 2n+1−1 2n+1−1 Xk,l=1 Xk,l=1 xp θ(k)xq k2 + l2 − θ(k)xq xp k2 + l2 Xk,l=1 ≤ θ(l) θ(l) M 1 k2 + l2 − 2n−1 Xk,l=1 1 k2 + l2 ≤ 2n+1−1 2n+1−1 ≤ 2 Xk=2n Xl=1 This concludes the proof. 1 k2 + l2 ≤ 2 2n+1−1 Xk=2n 2n+1−1 Xl=1 2−2n ≤ 4. Proof of Theorem 1.12. Let F be as in (5.3). Fourier coefficients of F are given by a non-zero sequence {Ξ(m)}m∈Z and, therefore F 6= 0. It follows from (5.2) that Fourier series for F converges uniformly and, therefore, F is continuous. It follows from the continuity of F that one can choose θ such that 2π ∈ Q, θ /∈ 2πZ and such that F (θ) 6= 0. Let Aθ be the von Neumann subalgebra in L(l2) generated by Uθ. θ = 1 and, therefore, Aθ is finite dimensional. Every linear functional on a finite dimensional subalgebra in L(H) is automatically normal. It follows that the mapping Since θ ∈ 2πQ, it follows that there exists 0 6= r ∈ Z such that U r θ T → ϕ(T (1 + D2)−1/2), T ∈ Aθ is normal for every linear functional on L1,∞ (in particular, for every trace on L1,∞). It follows from Lemma 5.3 that, for every p ∈ Z and for every normalised trace ϕ on L1,∞, we have ϕ(U p θ (1 + D2)−1/2) =(1, 0, pθ ∈ 2πZ pθ /∈ 2πZ. Hence, T (1 + D2)−1/2 is universally measurable for every T ∈ Aθ. This proves (c). Since Aθ ⊗ Aθ is also finite dimensional, it follows that the mapping T → ϕ(T (1 + D2 ⊗ 1 + 1 ⊗ D2)−1/2), T ∈ Aθ ⊗ Aθ is automatically normal for every linear functional on L1,∞ (in particular, for every trace on L1,∞). It follows from Lemma 5.5 that, for every θ ∈ R and for every p ∈ Z, we have xp θ(k)x−p 1 + k2 + l2 = F (pθ) θ (l) log(M + 1) log(2) + O(1), M ∈ Z+. M Xk,l=0 This equality combined with Lemma 5.2 provides that N Xm=0 λ(m, (U p θ ⊗U−p θ )(1+1⊗D2+D2⊗1)−1) = F (pθ) log(N + 1) 2 log(2) +O(1), N ∈ Z. 24 By Theorem 2.3, we have that ϕ((U p θ ⊗ U−p θ )(1 + 1 ⊗ D2 + D2 ⊗ 1)−1) = 1 2 log(2) F (pθ) (5.4) It follows from Lemma 5.6 that, for every θ ∈ R and for every p, q ∈ Z such for every normalised trace ϕ on L1,∞. that (p + q)θ /∈ 2πZ, we have xp θ(k)xq Xk,l=0 1 + k2 + l2 = O(1), M ∈ Z+. θ(l) M By Lemma 5.2, we have that N Xm=0 λ(m, (U p θ ⊗ U q θ )(1 + 1 ⊗ D2 + D2 ⊗ 1)−1) = O(1), N ∈ Z for every p, q ∈ Z with (p + q)θ /∈ 2πZ. By Theorem 2.3, we have ϕ((U p θ ⊗ U q θ )(1 + 1 ⊗ D2 + D2 ⊗ 1)−1) = 0 (5.5) for every normalised trace ϕ on L1,∞. Combining (5.4) and (5.5), we conclude that elements of the form T (1 + 1 ⊗ D2 + D2 ⊗ 1)−1, are universally measurable. This proves (d). T ∈ Aθ ⊗ Aθ, Finally, the first assertion in (e) follows from (5.4) (for p = 1) and the second assertion in (e) follows from Lemma 5.3. Proof of Corollary (1.13). Set T = U + U−1 + 2 ≥ 0. In the course of the proof of Theorem 1.12, we established a formula (5.5), which implies ϕ((U ⊗U )(1+1⊗D2+D2⊗1)−1) = ϕ((U−1⊗U−1)(1+1⊗D2 +D2⊗1)−1) = 0. It follows from Theorem 1.12 (e) that ϕ((T ⊗ T )(1 + 1 ⊗ D2 + D2 ⊗ 1)−1) = 4ϕ((1 + 1 ⊗ D2 + D2 ⊗ 1)−1)+ +2ϕ((U ⊗ U−1)(1 + 1 ⊗ D2 + D2 ⊗ 1)−1) 6= 4ϕ((1 + 1 ⊗ D2 + D2 ⊗ 1)−1). On the other hand, it follows from Lemma Appendix A.3 and Theorem 1.12 (e) that 4ϕ((1 + 1 ⊗ D2 + D2 ⊗ 1)−1) = π = π 4 (ϕ(T (1 + D2)−1/2))2. Remark 5.7. Neither Theorem 1.12 nor its proof specifies the dimension of the algebra A. However, if we replace 2 with 27 in the definition of xθ and set θ = π, then the algebra A becomes 2−dimensional. That F (π) 6= 0 can be showed as in the proof of Theorem 1.14 below. 25 6. Proof of Theorem 1.14 Define the sequence d by setting d(0) = 0, d(k) =(k, k ∈ [27n, 27(n+1)), n = 0 mod 2 27k, k ∈ [27n, 27(n+1)), n = 1 mod 2 and set D = diag({d(k)}k≥0). Set Ξ0(m) = 27 1 27(m+1) 27m dtds t2 + s2 . The proof of the following lemma is identical to that of Lemma 5.4 and is, therefore, omitted. Lemma 6.1. We have 27(n1+1)−1 27(n2 +1)−1 1 Xk1=27n1 = O(min{2−7n1, 2−7n2})+  Note that = d2(k1) + d2(k2) Xk2=27n2 Ξ0(n2 − n1), n1 = 0 mod 2, n2 = 0 mod 2 2−7Ξ0(n2 − n1 − 1), n1 = 1 mod 2, n2 = 0 mod 2 2−7Ξ0(n2 − n1 + 1), n1 = 0 mod 2, n2 = 1 mod 2 2−14Ξ0(n2 − n1), n1 = 1 mod 2, n2 = 1 mod 2 Ξ0(m) = Ξ0(m) ≤ 27 1 27(m+1) 27m dtds 214m ≤ (27 − 1)2 · 2−7m. In particular, we have Lemma 6.2. We have Ξ0(m) < ∞. Xm∈Z M Xk1,k2=1 1 d2(k1) + d2(k2) = log(M ) 14 log(2) (1 + 2−7)2 Xm∈Z Ξ0(2m) + O(1). 1 Proof. Suppose first that M = 27(n+1) − 1. It follows from Lemma 6.1 that 27(n+1)−1 Xk1,k2=1 +2−7 X0≤n1,n2≤n Ξ0(n2 − n1 − 1) + 2−7 X0≤n1,n2≤n Ξ0(n2−n1)+2−14 X0≤n1,n2≤n = X0≤n1,n2≤n Ξ0(n2 − n1 + 1) + O(1). d2(k1) + d2(k2) n1,n2=1 mod 2 n1,n2=0 mod 2 Ξ0(n2−n1)+ n1=1 mod 2 n2=0 mod 2 n1=0 mod 2 n2=1 mod 2 26 Making the substitution (n1, n2) =  we have that (m1, m2), n1 = 0 mod 2, n2 = 0 mod 2 (m1 − 1, m2), n1 = 1 mod 2, n2 = 0 mod 2 (m1, m2 − 1), n1 = 0 mod 2, n2 = 1 mod 2 (m1 − 1, m2 − 1), n1 = 1 mod 2, n2 = 1 mod 2 27(n+1)−1 Xk1,k2=1 1 d2(k1) + d2(k2) = (1 + 2−7)2 X0≤m1≤n 1≤m2≤n m1,m2=0 mod 2 Ξ0(m2 − m1) + O(1). Rearranging the summands as in Lemma 5.5, we infer that X0≤m1≤n 1≤m2≤n m1,m2=0 mod 2 Ξ0(m2 − m1) = n 2 Xm∈Z Ξ0(2m) + O(1). Passing from M = 27(n+1) − 1 to generic M as in Lemma 5.5, we conclude the proof. Lemma 6.3. For every normalised trace ϕ on L1,∞, we have )2 Xm∈Z ϕ((1 + D2 ⊗ 1 + 1 ⊗ D2)−1) = 1 + 2−7 7 log(2) 1 2 ( Ξ0(2m), ϕ((1 + D2)−1/2) = 1 + 2−7 2 . Proof. It follows from Lemma 6.1 that M Xk1,k2=0 1 1 + d2(k1) + d2(k2) = log(M ) 14 log(2) (1 + 2−7)2 Xm∈Z Ξ0(2m) + O(1). It follows now from Lemma 5.2 that M Xk=0 λ(k, (1 + D2 ⊗ 1 + 1 ⊗ D2)−1) = log(M ) 28 log(2) (1 + 2−7)2 Xm∈Z Ξ0(2m) + O(1). The first assertion follows now from Theorem 2.3. The second assertion follows from the equality 1 (1 + d2(k))1/2 = O(k−2) +(k−1, k ∈ [27n, 27(n+1)), n = 0 mod 2 2−7k−1, k ∈ [27n, 27(n+1)), n = 1 mod 2 ·(1, k ∈ [27n, 27(n+1)), n = 0 mod 2 1 − 2−7 −1, k ∈ [27n, 27(n+1)), n = 1 mod 2. = O(k−2) + 1 + 2−7 2k 2k + = 27 Proof of Theorem 1.14. According to the Lemma 6.3, it suffices to show that Ξ0(2m) > 7π 4 log(2). Xm∈Z In fact, we have Ξ0(0) = 27 1 27 1 dtds t2 + s2 ≥ 21≤s≤t≤27 dtds 2t2 = dtds t2 + s2 = 21≤s≤t≤27 = 27 (t − 1)dt t2 1 > 7 log(2) − 1. Therefore, Ξ0(2m) > Ξ0(0) > 7 log(2) − 1 > 7π 4 log(2). Xm∈Z Appendix A. Number-theoretic estimates The following lemmas are standard in number theory. Lemma Appendix A.1. For every p ∈ N, we have 1 = p 2 2−pπ Γ(1 + p 2 ) mp + O(mp−1), Xk∈Zp k≤m 1 = p 2 π Γ(1 + p 2 ) mp + O(mp−1). Xk∈Zp + k≤m Proof. We prove the first assertion by induction on p. Let K = [0, 1]p be the unit cube. For brevity, we denote p−tuple (1,··· , 1) = 1. We have Xk∈Np k≤m 1 = Xk∈Np k≤m K+k−1 dt = Cm where dt, Xk∈Zp + k≤m 1 = Xk∈Zp + k≤m K+k dt = Bm dt, Cm = [k∈Np k≤m (K + k − 1) ⊂ {t ∈ Rp + : t ≤ m} ⊂ [k∈Zp + k≤m (K + k) = Bm. It follows immediately that Xk∈Np k≤m 1 ≤ t∈Rp + t≤m 28 dt ≤ Xk∈Zp + k≤m 1. (A.1) It is clear that Xk∈Zp + k≤m 1 − Xk∈Np k≤m 1 ≤ p Xk∈Zp−1 + k≤m 1 = O(mp−1), (A.2) where we used induction with respect to p in the last equality. Combining (A.1) and (A.2), we infer that 1 = t∈Rp + t≤m dt + O(mp−1) = p 2 2−pπ Γ(1 + p 2 ) mp + O(mp−1). Xk∈Zp + k≤m This concludes the proof of the first equality. To see the second equality, note that 2p Xk∈Np k≤m 1 ≤ Xk∈Zp k≤m 1 ≤ 2p Xk∈Zp + k≤m 1. The second equality follows now from (A.2). Lemma Appendix A.2. Let αp : Z+ → Zp m → αp(m) increases, then + be a bijection. If the mapping p 2 p 2 αp(m)p = 2pπ−p/2Γ(1 + )m + O(m p−1 p ). Let αp : Z+ → Zp be a bijection. If the mapping m → αp(m) increases, then αp(m)p = π−p/2Γ(1 + )m + O(m p−1 p ). Proof. It follows from Lemma Appendix A.1 that m ≤ Xαp(k)≤αp(m) 1 = Xk∈Zp + k≤αp(m) 1 ≤ Xk∈Zp + k≤⌊αp(m)⌋+1 and = p 2 2−pπ Γ(1 + p m ≥ Xαp(k)<αp(m) 2 )αp(m)p + O(αp(m)p−1) 1 ≤ Xk∈Zp 1 = Xk∈Zp + + k<αp(m) k≤⌊αp(m)⌋−1 1 = 1 = = p 2 2−pπ Γ(1 + p 2 )αp(m)p + O(αp(m)p−1). A combination of these estimates yields the first assertion and the proof of the second one is identical. 29 Lemma Appendix A.3. For every p ∈ N, we have (1− ∆p)−p/2 ∈ L1,∞. For every normalised trace ϕ on L1,∞, we have Γ(1 + p 2 )ϕ((1 − ∆p)−p/2) = πp/2. Proof. It follows from Lemma Appendix A.2 that µ(m, (1 − ∆p)−p/2) = πp/2 Γ(1 + p 2 ) 1 m + 1 + O((m + 1)−1− 1 p ). Therefore, n Xm=0 µ(m, (1 − ∆p)−p/2) = πp/2 Γ(1 + p 2 ) log(n + 1) + O(1). The assertion follows from Theorem 2.3. Appendix B. An easy counter-example to formula 1.1 In Theorem 1.12, we required that both operators T1(1 + D2)−1/2 and T2(1 + D2)−1/2 are universally measurable. In this appendix, we show that a simpler counter-example with T2 = 1 does exist if the requirement of univer- sal measurability of T1(1 + D1)−p/2 is omitted. This gives a counter-example to the formula because one does not take into account the correction of the limiting process by powers (cf as in Lemma 1.9). Lemma Appendix B.1. There exist a 0 ≤ T1 ∈ L(l2), a universally measur- able operator (1 + D2)−1/2 ∈ L1,∞ and a Dixmier trace8 Trω such that Trω(T1(1 + D2)−1/2). Trω((T1 ⊗ 1)(1 + D2 ⊗ 1 + 1 ⊗ D2)−1) 6= π 4 Proof. Set D = diag({k}k≥0) and T1 = diag({x(k)}k≥0) with x = χ∪m[n2m,n2m+1), where log(nm) = o(log(nm+1)) as m → ∞. Suppose that for every Dixmier trace Trω we have Trω(cid:16)diag(cid:16)n x(k) 1 + k2 + l2ok,l≥0(cid:17)(cid:17) = π 4 Trω(cid:16)diag(cid:16)n x(k) k + 1ok≥0(cid:17)(cid:17). (B.1) In what follows, we omit diag to lighten the notations. Using definition (2.1) of Dixmier traces, we can equivalently rewrite (B.1) as lim n→ω 1 log(n + 2)(cid:16) n Xi=0 µ(cid:16)i,n x(k) 1 + k2 + l2ok,l≥0(cid:17) − π 4 n Xi=0 µ(cid:16)i,n x(k) k + 1ok≥0(cid:17)(cid:17) = 0 8See (2.1) for the definition of Dixmier trace. 30 for every ultrafilter ω. Equivalently, we have n Xi=0 µ(cid:16)i,n x(k) 1 + k2 + l2ok,l≥0(cid:17) − π 4 n Xi=0 µ(cid:16)i,n x(k) k + 1ok≥0(cid:17) = o(log(n)), n → ∞. Lemma 5.2 states that n Xi=0 µ(cid:16)i,n x(k) 1 + k2 + l2ok,l≥0(cid:17) − ⌊n1/2⌋ Xk,l=0 x(k) 1 + k2 + l2 = O(1), while Lemma 5.1 states that n Xi=0 µ(cid:16)i,n x(k) k + 1ok≥0(cid:17) − x(k) k + 1 n Xk=0 = O(1). We have Thus, π 4 n Xk=0 ⌊n1/2⌋ ⌊n1/2⌋ 1 ⌊n1/2⌋ x(k) x(k) 1 + k2 + l2 = Xl=0 1 + k2 ) + n1/2 1 + t2 + k2(cid:17) = dt 1 0 = ⌊n1/2⌋ Xk=0 1 + k2 + l2 = Xk,l=0 x(k)(cid:16)O( Xk=0 Xk=0 (k2 + 1)1/2 tan−1(( ⌊n1/2⌋ x(k) = n k2 + 1 1 2 ) + O(1). ) x(k) k + 1 − ⌊n1/2⌋ Xk=0 x(k) k + 1 tan−1(( n k2 + 1 ) 1 2 ) = o(log(n)), n → ∞. (B.2) We now show that (B.2) actually fails. For n = n2 2m, we have x(k) k + 1 ≥ n Xk=0 n2 2m Xk=n2m 1 k + 1 = log(n2m) + O(1) = 1 2 log(n) + O(1) and, taking into account that x vanishes on the interval [n2m−1, n2m), we have ⌊n1/2⌋ n2m−1 x(k) k + 1 tan−1(( n k2 + 1 ) 1 2 ) ≤ π 2 Xk=0 Hence, (B.2) fails for such x as n = n2 log(n2m−1)+O(1) = o(log(n)). 1 π 2 = k + 1 Xk=0 2m and m → ∞. 31 References [1] Benameur M., Fack T. Type II non-commutative geometry. I. Dixmier trace in von Neumann algebras. Adv. Math. 199 (2006), no. 1, 29 -- 87. [2] Carey A., Rennie A., Sukochev F., Zanin D. Universal measurability and the Hochschild class of the Chern character. J. Spectr. Theory 6 (2016), 1 -- 41. [3] Chakraborty P., Pal A. Equivariant spectral triples on the quantum SU(2) group. K-Theory 28 (2003), no. 2, 107 -- 126. [4] Connes A. Noncommutative Geometry. Academic Press, San Diego, 1994. [5] Connes A. Cyclic cohomology, quantum group symmetries and the local index formula for SUq(2). J. Inst. Math. Jussieu 3 (2004), no. 1, 17 -- 68. [6] Davidson K. C∗−algebras by example. Fields Institute Monographs, 6. American Mathematical Society, Providence, RI, 1996. [7] Dykema K., Figiel T., Weiss G., Wodzicki M. Commutator structure of operator ideals. Adv. Math. 185 (2004), no. 1, 1 -- 79. [8] Dykema K., Kalton N. Spectral characterization of sums of commutators. II. J. Reine Angew. Math. 504 (1998), 127 -- 137. [9] Gelfand I., Minlos R., Shapiro Z. Representations of the rotation and Lorentz groups and their applications. Oxford-London-New York-Paris: Pergamon Press. xviii, 366 pp. (1963); Moskva: Gosudarstv. Izdat. Fiz.- Mat. Lit., 368 pp. (1958). [10] Gilkey P. Invariance theory, the heat equation, and the Atiyah-Singer index theorem. Mathematics Lecture Series, 11. Publish or Perish, Inc., Wilm- ington, DE, 1984. [11] Gracia-Bondia J., Varilly J., Figueroa H. Elements of noncommutative geometry. Birkhauser Advanced Texts: Basler Lehrbucher. Birkhauser Boston, Inc., Boston, MA, 2001. [12] Kalton N. Spectral characterization of sums of commutators. I. J. Reine Angew. Math. 504 (1998), 115 -- 125. [13] Kalton N., Lord S., Potapov D., Sukochev F. Traces of compact operators and the noncommutative residue. Adv. Math. 235 (2013), 1 -- 55. [14] Klimyk A., Schmudgen K. Quantum groups and their representations. Texts and Monographs in Physics. Springer-Verlag, Berlin, 1997. [15] Lord S., Potapov D., Sukochev F. Measures from Dixmier traces and zeta functions. J. Funct. Anal. 259 (2010), no. 8, 1915 -- 1949. 32 [16] Lord S., Sukochev F., Zanin D. Singular Traces: Theory and Applications. volume 46 of Studies in Mathematics. De Gruyter, 2013. [17] Rosenberg S. The Laplacian on a Riemannian manifold. An introduction to analysis on manifolds. London Mathematical Society Student Texts, 31. Cambridge University Press, Cambridge, 1997. [18] Sedaev A., Sukochev F. Dixmier measurability in Marcinkiewicz spaces and applications. J. Funct. Anal. 265 (2013), no. 12, 3053 -- 3066. [19] Sukochev F., Usachev A., Zanin D. Generalized limits with additional in- variance properties and their applications to noncommutative geometry. Adv. Math. 239 (2013), 164 -- 189. [20] Sukochev F., Usachev A., Zanin D. Dixmier traces generated by exponen- tiation invariant generalised limits. J. Noncommut. Geom. 8 (2014), no. 2, 321 -- 336. [21] Sukochev F., Zanin D. ζ−function and heat kernel formulae. J. Funct. Anal. 260 (2011), no. 8, 2451 -- 2482. 33
1504.04816	1	1504	2015-04-19T09:56:34	A Note On Inner Quasidiagonal C*-Algebras	[ "math.OA" ]	In the paper, we give two new characterizations of separable inner quasidiagonal C-algebras. Base on these characterizations, we show that a unital full free product of two inner quasidiagonal C-algebras is inner quasidiagonal again. As an application, we show that a unital full free product of two inner quasidiagoanl C-algebras with amalgmation over a full matrix algebra is inner quasidiagonal. Meanwhile, we conclude that a unital full free product of two AF algebras with amalgamation over a finite-dimensional C-algebra is inner quasidiagonal if there are faithful tracial states on each of these two AF algebras such that the restrictions on the common subalgebra agree.	math.OA	math	A NOTE ON INNER QUASIDIAGONAL C-ALGEBRAS QIHUI LI Abstract. In the paper, we give two new characterizations of separable inner quasidiagonal C-algebras. Base on these characterizations, we show that a unital full free product of two inner quasidiagonal C-algebras is inner quasidi- agonal again. As an application, we show that a unital full free product of two inner quasidiagoanl C-algebras with amalgmation over a full matrix algebra is inner quasidiagonal. Meanwhile, we conclude that a unital full free product of two AF algebras with amalgamation over a finite-dimensional C-algebra is inner quasidiagonal if there are faithful tracial states on each of these two AF algebras such that the restrictions on the common subalgebra agree. 1. Introduction Quasidiagonal (QD) C-algebras have now been studied for more than 30 years. Voiculescu [17] give a characterization of quasidiagonal C-algebras as following: Definition 1. A C-algebra A is quasidiagonal if, for every x1, · · · , xn ∈ A and ε > 0, there is a representation π of A on a Hilbert space H, and a finite-rank projection P ∈ B (H) such that kP π (xi) − π (xi) P k < ε, kP π (xi) P k > kxik − ε for 1 ≤ i ≤ n. Voiculescu showed that A is QD if and only if π (A) is a quasidiagonal set of operators for a faithful essential representation π of A. In [4], we know that all separable QD C-algebras are Blackadar and Kirchberg's MF algebras. It is well known that the reduced free group C-algebra C∗ r (F2) is not QD. Haagerup and Thorbjφrnsen showed that C∗ r (F2) is MF [11]. This implies that the family of all separable QD C-algebras are strictly contained in the set of MF C-algebras. The concept of MF algebras was first introduced by Blackadar and Kirchberg in [4]. Many properties of MF algebras were discussed in [4]. In the same article, Blackadar and Kirchberg study NF algebras and strong NF algebras as well. A separable C-algebra is a strong NF algebra if it can be written as a generalized inductive limit of a sequential inductive system of finite-dimensional C-algebras in which the connecting maps are complete order embedding and asymptotically multiplicative in the sense of [4]. An NF algebra is a C-algebra which can be written as the generalized inductive limit of such system, where the connecting maps are only required to be completely positive contractions. It was shown that a separable C-algebra is an NF algebra if and only if it is nuclear and quasidiagonal. Whether the class of NF algebra is distinct from the class of strong NF algebras? 2000 Mathematics Subject Classification. 46L09, 46L35. Key words and phrases. Inner quasidiagonal C-algebras; Unital full free products of C- algebras; Unital full amalgamated free products of C-algebras. The research of the first author is partially supported by National Natural Science Foundation of China. 1 2 QIHUI LI For solving this question, Blackadar and Kirchberg introduce the concept of inner quasidiagonal by slightly modifying Voiculescu's characterization of quasidiagonal C-algebras: Definition 2. ([5]) A C-algebra A is inner quasidiagonal if, for every x1, · · · , xn ∈ A and ε > 0, there is a representation π of A on a Hilbert space H, and a finite-rank projection P ∈ π (A)′′ such that kP π (xi) − π (xi) P k < ε, kP π (xi) P k > kxik − ε for 1 ≤ i ≤ n. It was shown that a separable C-algebra is a strong NF algebra if and only if it is nuclear and inner quasidiagonal [5]. Blackadar and Kirchberg also gave examples of separable nuclear C-algebras which are quasidiagonal but not inner quasidiagonal, hence of NF algebras which are not strong NF. Therefore, the preceding question has been solved. In this note, we are interested in the question of whether the unital full free products of inner QD C-algebras are inner QD again. Note that every RFD C- algebra is inner QD [5]. We have known that a unital full free products of two RFD C-algebras is RFD [16]. Similar result holds for unital QD C-algebras [2]. Based on these results and the relationship among RFD C-algebras, inner QD C-algebras and QD C-algebras, it is natural to ask whether the same things will happen when we consider inner QD C-algebras. In this note we will show that a unital full free product of two unital inner QD C-algebra is inner again. As an application, we will consider the unital full free products of two inner QD C-algebras with amalgamation over finite-dimensional C-algebras. All C-algebras in this note are unital and separable. A brief overview of this paper is as follows. In Section 2, we fix some notation and give two new character- izations of inner QD C-algebras. Section 3 is devoted to results on the unital full free products of two unital inner QD C-algebras. We first consider unital full free products of unital inner QD C-algebras. As an application, we show that a unital full free product of two inner quasidiagonal C-algebras with amalgamation over a full matrix algebra is inner quasidiagonal. Meanwhile, we conclude that a unital full free product of two AF algebras with amalgamation over a finite-dimensional C-algebra is inner quasidiagonal if there are faithful tracial states on each of these two AF algebras such that the restrictions on the common subalgebra agree. 2. Inner Quasidiagonal C-algebras We denote the set of all bounded operators on H by B (H). Suppose {Mkn(C)}∞ n=1 is a sequence of complex matrix algebras. We introduce m=1 Mkm(C) of {Mkn (C)}∞ n=1 as follows: Mkn(C) = {(Yn)∞ n=1 ∀ n ≥ 1, Yn ∈ Mkn (C) and k(Yn)∞ the C-direct productQ∞ ∞Yn=1 Furthermore, we can introduce a norm-closed two sided ideal inQ∞ kYnk = 0) . Mkn (C) =((Yn)∞ ∞Xn=1 Mkn (C) : lim n→∞ follows: n=1 ∈ ∞Yn=1 n=1k = sup n≥1 kYnk < ∞}. n=1 Mkn(C) as A NOTE ON INNER QUASIDIAGONAL C-ALGEBRAS 3 Let π be the quotient map from Q∞ Then n=1 Mkn (C)/ Mkn (C). ∞Pn=1 n=1 Mkn(C) to Q∞ ∞Xn=1 Mkn (C) Mkn (C)/ ∞Yn=1 is a unital C-algebra. If we denote π ((Yn)∞ n=1) by [(Yn)n], then k[(Yn)n]k = lim sup n→∞ kYnk ≤ sup n kYnk = k(Yn)nk ∈ Mkn (C) ∞Yn=1 Recall that a C-algebra is residually finite-dimensional (RFD) if it has a sep- arating family of finite-dimensional representations. If a separable C- algebra A can be embedded intoQk Mnk (C) /Pk Mnk (C) for a sequence of positive integers {nk}∞ k=1 , then A is called an MF algebra. Many properties of MF algebras were discussed in [4]. Note that the family of all RFD C-algebras is strictly contained in the family of all inner QD C-algebras, and all QD C-algebras are MF C-algebras. Continuing the study of generalized inductive limits of finite-dimensional C- algebras, Blackadar and Kirchberg define a refined notion of quasidiagonality for C-algebras, called inner quasidiagonality. A cleaner alternative definition of inner quasidiagonality can be given using the socle of the bidual. Definition 3. If B is a C-algebra, then a projection p ∈ B is in the socle if pBp is finite-dimensional. Denote the set of the socle in B by socle (B) Theorem 1. ([8]) A separable C-algebra A is inner QD if there are projections pn ∈ A∗∗ such that (1) k[pn, a]k −→ 0 for all a ∈ A ⊆ A∗∗, (2) kak = lim kpnapnk for all a ∈ A and (3) pn ∈ socle (A∗∗) for every n. Theorem 2. ([5], Proposition 3.7.) Let A be a separable C-algebra. Then A is inner QD if and only if there is a sequence of irreducible representation {πn} of A on Hilbert space Hn,and finite-rank projection pn ∈ B (Hn), such that k[pn, πn (x)]k −→ 0 and lim sup kpnπn (x) pnk = kxk for all x ∈ A. The principal shortcoming of the definition of inner QD C-algebra is that it is often difficult to determine directly whether a C-algebra is inner QD, the following result for separable case is much easier for checking. Theorem 3. ([6]) A separable C-algebra is inner QD if and only if it has a separating family of quasidiagonal irreducible representations. Let π : B (H) → Q (H) be the canonical mapping onto the Calkin algebra and A is a unital C-algebra. Suppose ϕ : A → B (H) is a unital completely positive map then we say that ϕ is a representation modulo the compacts if π◦ϕ : A → Q (H) is a -homomorphism. If π ◦ ϕ is injective then we say that ϕ is a faithful representation modulo the compacts. for a sequence of positive integers {nk}∞ For an MF C-algebra, we are able to embed it into Qk it intoQk k=1 . For an RFD C-algebra, we can embed Mnk (C). Meanwhile, for a QD C-algebra, we can not only embed it into Mnk (C) /Pk Mnk (C) 4 QIHUI LI Mnk (C) /Pk Mnk (C) , but also lift this embedding to a faithful representation Qk into YMkm (C) modulo the compacts. Whether there is a similar characteriza- tion for the inner QD C-algebras? We will answer this question in the following theorem. The following lemma is a well-known result about completely positive map. We use c.p. to abbreviate "completely positive", u.c.p. for "unital completely positive" and c.c.p. for "contractive completely positive". Lemma 1. (Stinespring) Let A be a unital C-algebra and ϕ : A −→ B (H) be a know that, under this minimality condition, a Stinespring dilation is unique up to for every a ∈ A. In particular, kϕk = kV ∗V k = kϕ (1)k . ϕ (a) = V ∗πϕ (a) V When ϕ is unital, V ∗V = ϕ (I) = I, and hence V is an isometry. So in this case we may assume that V is a projection P and ϕ (a) = P πϕ (a) H. In general there could be many different Stinespring dilations, but we may always assume that a c.p.. map. Then, there exist a Hilbert space bH, a -representation πϕ : A −→ B (H) and an operator V : H −→bH such that We call the triplet (cid:16)πϕ, bH, V(cid:17) in preceding lemma a Stinespring dilation of ϕ. dilation (cid:16)πϕ, bH, V(cid:17) is minimal in the sense that πϕ (A) V H is dense in bH. We unitary equivalence. Note that if(cid:16)πϕ, bH, V(cid:17) is minimal Stinespring dilation of ϕ : A −→ B (H), then there exists a -homomorphism ρ : ϕ (A)′ −→ πϕ (A)′ ⊆ B(cid:16)bH(cid:17) the commutant ϕ (A)′ ⊆ B (H) also lifts to B(cid:16)bH(cid:17) . u.c.p. map. Suppose(cid:16)πϕ, bH, P(cid:17) is a minimal Stinespring dilation of ϕ where P is a projection in B(cid:16)bH(cid:17) . Then the -homomorphism ρ : ϕ (A)′ −→ πϕ (A)′ ⊆ B(cid:16)bH(cid:17) such that ϕ (a) x = V ∗πϕ (a) ρ (x) V for every a ∈ A and x ∈ ϕ (A)′ , it implies that Lemma 2. Let A be a unital C-algebra and ϕ : A −→ Mn (C) be a surjective is unital, i.e. ρ (CI) = CI ⊆ πϕ (A)′ Proof. Since ϕ is surjective, ϕ (A)′ = CI. Note ρ is a -homomorphism, it is easy to check that (I − P ) ρ (αI) P = P ρ (αI) (I − P ) = 0. and (I − P ) ρ (I) (I − P ) is a projection. We know(cid:16)πϕ, bH, P(cid:17) is a minimal Stine- spring dilation, then πϕ (A)′ has no proper projection bigger than P. It implies (I − P ) ρ (I) (I − P ) = 0, i.e. ρ (I) = I. Hence ρ (CI) = CI ⊆ πϕ (A)′ . (cid:3) Now, we are ready to give a new characterization of inner QD C-algebras. Theorem 4. Suppose A is a unital C-algebra. Then A is inner QD if and only if there is a faithful representation modulo compacts Φ : A −→ ΠMkn (C) for a se- quence {kn} of integers such that the u.c.p. maps ϕn : A −→ Mkn (C) is surjective for every n and the -homomorphism ρ : ϕn (A)′ −→ πϕn (A)′ ⊆ B(cid:16)bHn(cid:17) A NOTE ON INNER QUASIDIAGONAL C-ALGEBRAS 5 is surjective where(cid:16)πϕn , bHn, pn(cid:17) is a minimal Stinespring dilation of ϕn. Proof. (=⇒) Suppose A is inner QD. Then, by applying Theorem 2, we can find sequences of irreducible representations {πn} and finite projections {pn} where pn ∈ πn (A)′′ such that Φ : A −→ Πpnπn (A) pn is a faithful representation modulo compacts. Meanwhile, we have pnπn (A) pn ∼= Mkn (C) for some integer kn and (πn (A))′ = CI since πn is irreducible. Define ϕn : A −→pnπn (A) pn ∼= Mkn (C) by ϕn (a) = pnπn (a) pn. Then ϕn is u.c.p. and surjective for every n. Note (πn, Hn, pn) is a minimal Stinespring dilation of ϕn since πn is irreducible. There- fore the -homomorphism ρ : ϕn (A)′ −→ πn (A)′ ⊆ B (Hn) is surjective by Lemma 2 and the fact that π (A)′ = CI. (⇐=) Suppose there is a faithful representation modulo compacts Φ : A −→ ΠMkn (C) for a sequence {kn} such that the u.c.p. maps ϕn : A −→ Mkn (C) is surjective and the -homomorphism is surjective where(cid:16)πϕn , bHn, pn(cid:17) is a minimal Stinespring dilation of ϕn. Then ρ : ϕn (A)′ −→ πϕn (A)′ ⊆ B(cid:16)bHn(cid:17) ρ(cid:0)ϕn (A)′(cid:1) = ρ (CI) = CI It implies that πϕn (A)′ = CI since ρ is surjective. Hence πϕn by Lemma 2. is irreducible and pn ∈ πϕn (A)′′ . So, for these irreducible representation {πϕn } of A on Hilbert space bHn and finite-rank projection pn ∈ πϕn (A)′′, we have k[pn, πϕn (x)]k −→ 0 and lim sup kpnπϕn (x) pnk = kxk for all x ∈ A. It implies that A is inner QD by Theorem 2. (cid:3) Suppose A is a unital C-algebra and p ∈ socle (A∗∗) . Define Then we have the following few lemmas. Ap = {a ∈ A : [a, p] = 0} . Lemma 3. ([5], Corollary 3.5.) Let p ∈ socle (A∗∗) . Then d (a, Ap) = k[a, p]kfor all a ∈ A. Lemma 4. ([5], Proposition 3.4.) Let A be a C-algebra, and p ∈ socle (A∗∗) . Then (1) pAp = pApp = pA∗∗p = pAp (2) The weak closure of Ap in A∗∗ is pA∗∗p + (1 − p) A∗∗(1 − p). Lemma 5. Let A be a C-algebra, p1, · · · , pk ∈ socle (A∗∗) with p1 ≤ · · · ≤ pk.Then i=0A∗∗ pk(cid:0)∩k i=0Api(cid:1) = pk(cid:0)∩k p2(cid:0)A∗∗ pi(cid:1) p2(cid:1) = p1Ap1 + (p2 − p1) A (p2 − p1) p1 ∩ A∗∗ = p0Ap0 ⊕ (p1 − p0) A (p1 − p0) ⊕ · · · ⊕ (pk − pk−1) A (pk − pk−1) Proof. We only prove the case when k = 2. Since p1 ≤ p2 ∈ socle (A∗∗) , we have 6 QIHUI LI by Lemma 4. So it is obvious that p2 (Ap1 ∩ Ap2 ) ⊆ p1Ap1 + (p2 − p1) A (p2 − p1) . Meanwhile, p2Ap2 = p2Ap2 = p2A∗∗p2 ⊇ p1Ap1 + (p2 − p1) A (p2 − p1) by Lemma 4 and the fact that p1 ≤ p2 ∈ socle (A∗∗) . Then for every b1 ∈ p1Ap1 and b2 ∈ (p2 − p1) A (p2 − p1) ,there is a ∈ Ap2 such that p2a = b1 + b2. Hence a ∈ Ap1 ∩ Ap2 . It implies that p2 (Ap1 ∩ Ap2 ) ⊇ p1Ap1 + (p2 − p1) A (p2 − p1) . This completes the proof. (cid:3) Lemma 6. Let A be a C-algebra, {pn} be a sequence of projection in socle (A∗∗) with p0 ≤ p1 ≤ · · · and pi s.o.t−→ I (strong operator topology). Then i=0Api = p0Ap0 ⊕ ⊕∞ ∩∞ i=1 [(pi − pi−1) A (pi − pi−1)] ⊆ A Proof. By Lemma 5 and the fact that p1 ≤ p2 ≤ · · · with pi ∈ socle (A∗∗) , we have pk (∩∞ i=0Api ) = pk(cid:0)∩k i=0Api(cid:1) = pk(cid:0)∩k i=0A∗∗ pi(cid:1) = p0Ap0 ⊕ (p1 − p0) A (p1 − p0) ⊕ · · · ⊕ (pk − pk−1) A (pk − pk−1) for every k. Therefore i=0Api = p0 (∩∞ ∩∞ = p0 (∩∞ = p0Ap0 ⊕ ⊕∞ i=0Api) ⊕ ⊕∞ i=0Api) ⊕ ⊕∞ i=1 (pi − pi−1) (∩∞ k=0Apk ) i=1 (pi − pi−1)(cid:0)∩i k=0Apk(cid:1) i=1 [(pi − pi−1) A (pi − pi−1)] (cid:3) Remark 1. Suppose A is a unital inner QD C-algebras, then there is sequence {pn} of projections in socle (A∗∗) such that k[pn, a]k −→ 0 for all a ∈ A ⊆ A∗∗ and kak = lim kpnapnk for all a ∈ A by Theorem 1. Therefore we can define a sequence of u.c.p maps ϕn : A −→pnA∗∗pn by compression. It is obvious that Apn = Mϕn where Mϕn is the multiplicative domain of ϕn and kak = lim kϕn (a)k, d (a, Mϕn) −→ 0 for all a ∈ A by Lemma 3. Actually, this is a sufficient condition for a given C-algebra to be an inner QD C-algebra. Theorem 5. ([8]) A is inner QD if and only if there is a sequence of c.c.p. maps ϕn : A −→ Mkn (C) such that kak = lim kϕn (a)k and d (a, Mϕn) −→ 0 for all a ∈ A, where Mϕn is the multiplicative domain of ϕn. Now, we are ready to give another characterization of unital inner QD C- algebras. Theorem 6. Suppose A is a unital separable C-algebra. Then A is inner QD if and only if there is a sequence of unital RFD C-subalgebra {An}∞ n=1 of A such that ∪∞ n=1An is norm dense in A. A NOTE ON INNER QUASIDIAGONAL C-ALGEBRAS 7 Proof. (=⇒) Suppose F ⊆ A is a finite subset and ε > 0. Let {1} ∪ F ⊆ F1 ⊆ F2⊆ · · · be the sequence of finite subsets of A such that ∪iFi = A. Then, from Remark 1 and Theorem 5, we can find such that P0 ≤ P1 ≤ P2 ≤ · · · with Pi ∈ socle (A∗∗) and Pi s.o.t−→ P ∈ A∗∗ (i −→ ∞) d(a, APi ) = k[a, Pi]k = k(1 − Pi) aPi + Pia (1 − Pi)k < ε 2 · 2i+1 and kPiaPik > kak − ε i −→ ∞) and P ≥ Pi, we have 2i+1 for every a ∈ Fi (i ∈ N). Since Pi s.o.t−→ P ∈ A∗∗ (as kP aP k ≥ kPiaPik ≥ kak − ε 2i+1 for ∀a ∈ ∪iFi and i. It implies that kP aP k = kak for ∀a ∈ ∪iFi = A, therefore P = I. Now let Aε = ∩∞ i=0APi = P0AP0 ⊕ (P1 − P0) A (P1 − P0) ⊕ · · · by Lemma 6. So, for any a ∈ F , let x = P0aP0 + (P1 − P0) a (P1 − P0) + · · · ∈ Aε, we have d(a, Aε) ≤ ka − xk = P0a (P1 − P0) + P1a (P2 − P1) + · · · + (P1 − P0) aP0 + (P2 − P1) aP1 + · · · ≤ kP0a (1 − P0)k kP1 − P0k + · · · + kP1 − P0k k(1 − P0) aP0k < ∞Xi=0 ε 2 · 2i+1 + ε 2 · 2i+1 = ε. ∞Xi=0 Note that Aε is an RFD C-subalgebras of A, hence we can find a sequence of unital RFD C-subalgebra {An}∞ n=1An is norm dense in A. n=1 of A such that ∪∞ (⇐=) Suppose {An}∞ that ∪∞ C-subalgebra An and ba ∈ An such that ka − bak < ε that n=1 is a sequence of unital RFD C-subalgebra in A such = A, F ⊆ A is a finite subset and ε > 0. Then there is an RFD 3 for every a ∈ F . It follows n=1An k·k kak − ≤ kbak . ε 3 Since An is RFD, we can find a projection P such that ΦP : An −→ P AnP ⊆ Mt (C) is a -homorphism for some t ∈ C and kΦP (ba)k ≥ kbak − ε 3 . Extending and ΦP to a u.c.p. map fΦP : A −→ Mt (C) with An ⊆ MgΦP (cid:13)(cid:13)(cid:13)fΦP (ba)(cid:13)(cid:13)(cid:13) ≥ kbak − is the multiplicative domain of fΦP . Then (cid:13)(cid:13)(cid:13)fΦP (ba)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)fΦP (ba − a) + fΦP (a)(cid:13)(cid:13)(cid:13) ≤ where MgΦP ε 3 ε 3 +(cid:13)(cid:13)(cid:13)fΦP (a)(cid:13)(cid:13)(cid:13) for every a ∈ F . 8 QIHUI LI So from above inequalities, we have (cid:13)(cid:13)(cid:13)fΦP (a)(cid:13)(cid:13)(cid:13) ≥(cid:13)(cid:13)(cid:13)fΦP (ba)(cid:13)(cid:13)(cid:13) − ≥ kbak − 2ε 3 ε 3 ≥ kak − ε for every a ∈ F By the preceding discussion, for a finite subset F and ε > 0, there is a u.c.p map fΦP : A −→ Mt (C) for some t ∈ N such that d(cid:16)a, MgΦP(cid:17) < ε and (cid:13)(cid:13)(cid:13)fΦP (a)(cid:13)(cid:13)(cid:13) ≥ kak − ε for every a ∈ F . So by Theorem 5, A is inner. (cid:3) 3. Unital Full Free Products of Two Inner QD Algebras. In this section we will consider the question of whether the unital full free prod- ucts of inner QD C-algebras are inner QD again. First, we need a lemma for showing the main result in this section. Lemma 7. (Theorem 3.2, [16]) Suppose A1 and A2 are unital C-algebras. Then the unital full free product A = A1∗CA2 is RFD if and only if A1 and A2 are both RFD. Now, we are ready to give the main result of this section. Theorem 7. If A1 and A2 are both unital inner quasidiagonal C-algebras. Then A1 ∗C A2 is inner QD. Proof. Suppose τ is a fixed state on A1 ∗C A2 and F is a finite subset of A1 ∗C A2. Let A0 j = {a ∈ Aj : τ (a) = 0} , j = 1, 2. No loss of generality, we may assume that every b ∈ F can be decomposed into a finite sum with respect to τ, that is ai1 ai2 · · · ain α0 ∈ C, aij ∈ A0 ij , i1 6= i2 6= · · · 6= in b = α0I + Xi16=i26=···6=in 1 or A0 ij = A0 0 , j = 1, 2, the set of such elements of A0 where A0 j which appear in the decomposition of elements from F . Then we can find an RFD C-subalgebra Aj ε of Aj for j = 1, 2 such that 2. Denote by F j (1) d(a, Aj ε) < ε for ∀a ∈ F j 0 , j = 1, 2. Let b be an element in F . No loss of generality, we may assume that b can be decomposed into the form 2,1ai 1,2ai 2,2ai 1,3 · · · ai 1,ni , where and Then, for any ai that 1,1ai ai αI + 1,1, ai lXi=1 (cid:8)ai (cid:8)ai j,k, we can find gai (cid:12)(cid:12)(cid:12)τ(cid:16)ai j,k(cid:17)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)τ(cid:0)ai j,k −gai 2,1, ai j,k ∈ Aj 1,2, · · · , ai 2,2, · · · , ai 0 0 . 1,ni(cid:9) ⊆ F 1 2,n(cid:9) ⊆ F 2 ε such that(cid:13)(cid:13)(cid:13)ai j,k(cid:13)(cid:13)(cid:13) < ε by (1). Note j,k −gai j,k(cid:17)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)τ(cid:16)gai j,k(cid:17)(cid:12)(cid:12)(cid:12) < ε, j,k(cid:1) − τ(cid:16)gai A NOTE ON INNER QUASIDIAGONAL C-ALGEBRAS 9 j,k(cid:17) +(cid:16)gai j,k − τ(cid:16)gai There is an integer Mb > 0 such that j,k = τ(cid:16)gai j,k(cid:13)(cid:13)(cid:13) < 2ε. j,k(cid:17)(cid:17) and (cid:13)(cid:13)(cid:13)(cid:16)gai j,k(cid:17)(cid:17) − ai j,k − τ(cid:16)gai then gai with(cid:13)(cid:13)(cid:13)ai 1,k(cid:13)(cid:13)(cid:13) < ε(cid:1)0 1,k ∈(cid:0)A1 1,k −gai Therefore, no loss of generality, we may assume thatgai 2,k(cid:13)(cid:13)(cid:13) < 2ε with (cid:13)(cid:13)(cid:13)ai ε(cid:1)0 2,k ∈ (cid:0)A2 2,k −gai 2ε where k = 1, · · · , ni, i = 1, · · · , l. And gai where k = 1, · · · , ni, i = 1, · · · , l. Leteb = αI +Pl 2,1gai 1,1gai i=1gai (cid:13)(cid:13)(cid:13)b −eb(cid:13)(cid:13)(cid:13) 1,ni!(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1,ni − αI + lXi=1gai 2,1gai 1,1gai 1,2 · · · ]ai Since A1 that A1 ∗C A2 is inner QD. ε ∗C A2 ε is an RFD C-algebra by Lemma 7, then by Theorem 6 we have (cid:3) lXi=1 2,2ai 1,3 · · · ai ai 1,1ai 2,1ai 1,2ai ≤ Mbε. ∈ A1 ε ∗C A2 ε. αI + 1,2 · · · ] ai 1,ni Since every strong NF algebra is inner, then from [4], we know that every AF algebra and AH algebra are inner. Hence we have the following two corollaries. Corollary 1. Suppose A and B are both AF algebras, then A ∗C B is an inner QD algebra. Corollary 2. Suppose A and B are both AH algebras, then A ∗C B is an inner QD algebra. What will happen when the above amalgamation is over some other C-algebras instead of CI? In [13], it has been shown that a full amalgamated free product of two QD algebras may not be MF again, even for a unital full free product of two full matrix algebras with amalgamation over a two dimensional C-algebra which is -isomorphic to C ⊕ C. Therefore, a unital full amalgamated free product of two unital inner QD C-algebras may not be inner again. But we can give the affirmative answers for some specific cases. The following result can be found in [15] or [13]. Lemma 8. Suppose that A, B and D are unital C-algebras. Then (A⊗maxD) ∗ D (B⊗maxD) ∼=(cid:16)A∗ C B(cid:17) ⊗max D. Lemma 9. ([5])Let A be a C-algebra. Then, for any k, A is inner QD if and only if Mk (A) = A ⊗ Mk (C) is inner QD. Proposition 1. Let A and B be unital C-algebras. If D can be embedded as an unital C-subalgebra of A and B respectively, and D is -isomorphic to a full matrix algebra Mn (C) for some integer n, then the unital full amalgamated free product A∗ D B is inner QD if A and B are both inner QD. Proof. Since D is -isomorphic to a full matrix algebra, from Lemma 6.6.3 in [12], it follows that A ∼= A′ ⊗ D and B ∼= B′ ⊗ D. Then A′ and B′ are inner QD by Lemma 9. So the desired conclusion follows from Theorem 7, Lemma 8 and Lemma 9. (cid:3) Next, we will consider the case when the free products are amalgamated over some finite-dimensional C-algebras. 10 QIHUI LI Lemma 10. ([6]) An arbitrary inductive limit (with injective connecting maps) of inner quasidiagonal C-algebras is inner quasidiagonal. Lemma 11. ([15], Theorem 4.2) Assume that we have embeddings of C-algebras C ⊆ A1 ⊆ A2 and C ⊆ B1 ⊆ B2, then the natural morphism σ : A1∗CB1 −→ A2∗C B2 is injective. Lemma 12. ([15] corollary 4.13) If (An) and (Bn) are increasing sequences of C- algebras, all of which contain a common C-subalgebra C, then there is a natural isomorphism (An ∗C Bn) = lim −→ where lim−→ denotes the ordinary direct limit. lim −→ An ∗C lim −→ Bn The following lemma is a well-known property of AF algebras, we can find it in [9] Lemma 13. A C-algebra A is AF if and only if it is separable and (∗) for all ε > 0 and A1, · · · , An in A, there exists a finite dimensional C- subalgebra B of A such that dist (Ai, B) < ε for 1 ≤ i ≤ n. Moreover, if A1 is a finite-dimensional subalgebra of A, then we may choose B so that it contains A1. Lemma 14. ([1], Theorem 4.2) Consider unital inclusions of C-algebras A ⊇ C ⊆ B with A and B finite dimensional. Let A∗CB be the corresponding full amalgamated free product. Then A∗DB is RFD if and only there are faithful tracial states τA on A and τB on B with τA(x) = τB(x), ∀ x ∈ C. Corollary 3. Suppose A and B are AF algebras and A ⊇ C ⊆ B with C finite- dimensional. If there are faithful tracial states τA and τB on A and B respectively, such that τA(x) = τB(x), ∀ x ∈ C, then A ∗C B is inner QD.. Proof. Since C is a finite-dimensional C-subalgebra, then we can find a sequence of finite-dimensional C-subalgebras {An}∞ n=1 such that C ⊆ A1 ⊆ A2 ⊆ · · · with ∪An = A and C ⊆ B1 ⊆ B2 ⊆ · · · with ∪Bn = B by Lemma 13. Note that An ∗C Bn is RFD by Lemma 14, then A∗CB =lim (An ∗C Bn) is inner by −→ Lemma 12 and Lemma 10. (cid:3) n=1 and {Bn}∞ References [1] S. Armstrong, K. Dykema, R. Exel and H. Li, On embeddings of full amalgamated free product C-algebras, Proc. Amer. Math. Soc. 132 (2004), 2019-2030. [2] F. Boca, A note on full free product C-algebras, lifting and quasidiagonality, Oper- ator Theory, Operator Algebras and Related Topics (Proc. of the 16th Op.) [3] N. P. Brown and K. J. Dykema, Popa algebras in free group factors, J. reine angew. Math 573 (2004), 157-180. [4] B. Blackadar and E. Kirchberg, Generalized inductive limits of finite dimensional C-algebras, Math. Ann. 307 (1997), 343-380. [5] B. Blackadar and E. Kirchberg, Inner quasidiagonality and strong NF algebras, Pacific Journal of mathematics 198 (2001), 307 -- 329. [6] B. Blackadar and E. Kirchberg, Irreducible representations of inner quasidiagonal C-algebras, Canadian Mathematical Bulletin, 54(2011), 385-395. A NOTE ON INNER QUASIDIAGONAL C-ALGEBRAS 11 [7] N.P. Brown, On quasidiagonal C-algebras, Operator algebras and applications, 19 -- 64, Adv. Stud. Pure Math., 38, Math. Soc. Japan, Tokyo, 2004. [8] Nathanial P. Brown and Narutaka Ozawa, Algebras and Finite-Dimensional Approx- imations, American Mathematical Society (March 12, 2008). [9] K. Davidson, C-algebras by Example, Amer Mathematical Society (June 1996). [10] D. Hadwin, Q. Li and J. Shen, Topological free entropy dimensions in Nuclear C- algebras and in Full Free Products of Unital C-algebras, Canadian Journal of Math- ematics, 63(2011), 551-590. [11] U. Haagerup, S. Thorbjφrnsen, A new application of random matrices: Ext(C ∗ red(F2)) is not a group, Ann. of Math. (2) 162 (2005), no. 2, 711 -- 775. [12] R. Kadison, J. Ringrose, Fundamentals of the Operator Algebras (Academic, Orlando, FL), (1983, 1986)Vols. 1 and 2. [13] Qihui Li and Junhao Shen, On RFD property of unital full amalgamated free products of RFD C-algebras, Illinois J. Math, Summer 2012, 56(2),647-659. [14] Qihui Li and Junhao Shen, Full Amalgamated Free Products of MF Algebras, Oper- ators and Matrices,Vol.7(2013), 333-356. [15] Gert K. Pedersen, Pullback and Pushout Constructions in C-algebra Theory, Journal Functional Analysis 167(1999), 243-344. [16] R. Exel and T. Loring, Finite-dimensional representations of free product C -algebras, Internat. J. Math. 3 (1992), no. 4, 469 -- 476. [17] D. Voiculescu, A note on quasi-diagonal C-algebras and homotopy, Duke Mathemat- ical Jouranl Vol.62 (1991), No. 2, 267-271. Department of Mathematics, East China University of Science and Technology, Mei- long Road 130, 200237 Shanghai, P.R. China. E-mail address: qihui [email protected]
1508.00284	2	1508	2016-03-29T20:34:21	Exact large ideals of B(G) are downward directed	[ "math.OA" ]	We prove that if E and F are large ideals of B(G) for which the associated coaction functors are exact, then the same is true for the intersection of E and F. We also give an example of a coaction functor whose restriction to the maximal coactions does not come from any large ideal.	math.OA	math	EXACT LARGE IDEALS OF B(G) ARE DOWNWARD DIRECTED S. KALISZEWSKI, MAGNUS B. LANDSTAD, AND JOHN QUIGG Abstract. We prove that if E and F are large ideals of B(G) for which the associated coaction functors are exact, then the same is true for E ∩ F . We also give an example of a coaction functor whose restriction to the maximal coactions does not come from any large ideal. 1. Introduction In [BGW] Baum, Guentner, and Willett, striving to make advances in the Baum-Connes conjecture, studied crossed-product functors σ that take an action (A, α) of a locally compact group G to a C ∗-algebra A ⋊α,σ G lying between the full and reduced crossed products. It is particularly important to know when σ is exact in the sense that it preserves short exact sequences. Motivated by this, in [KLQa] we in- troduced coaction functors, a certain type of functor on the category of coactions of G. Every coaction functor gives rise to a crossed-product functor by composing with the full-crossed-product functor. Among other things, we showed that if the coaction functor is exact then so is the associated crossed-product functor. We paid particular attention to the coaction functors τE coming from large ideals E of the Fourier- Stieltjes algebra B(G). An obvious question is, "For which large ideals E is the coaction functor τE exact?" In the current paper we will call E exact if τE is exact; for example, B(G) is exact, but the reduced Fourier-Stieltjes algebra Br(G) is exact if and only if G is an exact group. In [KLQa, Remark 6.23] we asked whether the intersection of two exact large ideals is exact, and we mentioned that we had an idea of how to proceed, and promised to address the question in future work. In the current paper we fulfill that promise in Theorem 3.2. In [KLQa] we speculated that the proof would require a "somewhat more elaborate version of Morita compatibility", and that it would Date: December 16, 2015. 2000 Mathematics Subject Classification. Primary 46L55; Secondary 46M15. Key words and phrases. crossed product, action, coaction, Fourier-Stieltjes algebra, exact sequence, Morita compatible. 1 2 KALISZEWSKI, LANDSTAD, AND QUIGG "perhaps resemble the property that Buss, Echterhoff, and Willett call correspondence functoriality (see [BEWa, Theorem 4.9])". It transpires that we ended up doing something slightly different: rather than change our definition of Morita compatibility, we instead combine it with an- other concept from [BEWa], namely the ideal property. We also answer another question left open in [KLQa, Question 6.20]: there we asked whether every coaction functor, when restricted to the maximal coactions, is naturally isomorphic to one coming from a large ideal. In Example 3.16 we give a counterexample, stealing a trick from [BEWa]. We wish to thank the referee for suggestions that improved this pa- per. 2. Preliminaries We briefly recall a few definitions from [KLQa]. In the classical cat- egory C∗ of C ∗-algebras, the morphisms are homomorphisms between the C ∗-algebras themselves, not involving multipliers, and in the clas- sical category Coact of coactions the morphisms are morphisms in C∗ that are equivariant for the coactions. Since we are interested in the classical category instead of the nondegenerate one (involving nonde- generate homomorphisms into multiplier algebras), we regard maxi- malization (A, δ) 7→ (Am, δm) and normalization (A, δ) 7→ (An, δn) as functors on Coact (and we use the notation φm and φn for the respec- tive images of a morphism φ). We assume that we have fixed once and for all a maximalization func- tor (A, δ) 7→ (Am, δm) and a normalization functor (A, δ) 7→ (An, δn) on the classical category of coactions, with canonical equivariant sur- jections qm A : Am → A and ΛA : A → An. Recall from [KLQa, Defini- tion 4.1] that a coaction functor is a functor τ on the classical category of coactions, together with a natural transformation qτ from maxi- malization to τ such that for each coaction (A, δ) the homomorphism qτ A : Am → Aτ is surjective and has kernel contained in the kernel of the canonical map ΛAm : Am → An (which is both a normalization of (Am, δm) and a maximalization of (An, δn)). Maximalization, nor- malization, and the identity functor are all coaction functors. There are other known coaction functors, determined by large ideals of the Fourier-Stieltjes algebra B(G) (see [KLQa, Section 6]). Recall from [KLQb, Definition 3.1] that we say an ideal E of B(G) is large if it is weak* closed, G-invariant, and nonzero (in which case it will neces- sarily contain Br(G), by [KLQ13, Lemma 3.14]). In Example 3.16 we EXACT LARGE IDEALS OF B(G) ARE DOWNWARD DIRECTED 3 adapt a construction from [BEWa] (who studied crossed-product func- tors defined on a category of actions) to define new coaction functors not of the preceding types. In [KLQa, Definition 4.10] we defined a coaction functor to be exact if it preserves short exact sequences. Let (A, δ) and (B, ǫ) be coactions, and let (X, ζ) be an (A, δ) − (B, ǫ) imprimitivity-bimodule coaction. [KLQa, Lemma 4.15] gives an (Am, δm) − (Bm, ǫm) imprimitivity-bimodule coaction (X m, ζ m) such that X m-Ind ker qm B = ker qm A (see [KLQa, Lemma 4.21] for the latter). In [KLQa, Definition 4.16] we defined a coaction functor τ to be Morita compatible if for every (X, ζ) as above we also have X m-Ind ker qτ B = ker qτ A. Trivially, maximalization is Morita compatible, and by [KLQa, Propo- sition 6.10] every coaction functor coming from a large ideal is Morita compatible. Recall that in [KLQc, Definition 7.2] we called a coaction (A, δ) of G w-proper (and in [KLQb, Definition 5.1] we used the term "slice proper") if (ω ⊗ id) ◦ δ(A) ⊂ C ∗(G) for all ω ∈ A∗. If (A, δ) is a coaction, we call an ideal I of A strongly invariant (see, e.g.,[KLQa, Definition 3.16] if span{δ(I)(1 ⊗ C ∗(G))} = I ⊗ C ∗(G). Note that this is precisely what is needed for the restriction of δ to I to be a coaction. 3. Main result We recall a few definitions from [KLQa, Section 6]: given any coac- tion (A, δ) and any large ideal E of B(G), we define an ideal AE = {a ∈ A : E · a = {0}}, and we write AE = A/AE for the quotient C ∗-algebra. The quotient map QE A : A → AE is equivariant for δ and a coaction δE on AE, and (A, δ) 7→ (AE, δE) is a coaction functor that we denote by τE. Definition 3.1. We call a large ideal E of B(G) exact if the associated coaction functor τE is exact. We will prove that the set of exact large ideals of B(G) is downward directed by showing that it is in fact closed under finite intersections. It remains an open By induction, Theorem 3.2 below does the job. 4 KALISZEWSKI, LANDSTAD, AND QUIGG question whether the intersection of all exact large ideals of B(G) is exact. Theorem 3.2. The intersection of two exact large ideals of B(G) is exact. The key idea of our proof is the following: for two large ideals E and F of B(G), we compare the intersection E ∩ F to the product. The following definition makes this precise: Definition 3.3. For two large ideals E, F ⊂ B(G) we write hEF i for the weak-closed linear span of the set EF of products. Remark 3.4. It is somewhat frustrating that we do not know of any examples of exact large ideals other than B(G) (and, when G is ex- act, Br(G)). Perhaps other examples could be found using techniques similar to those of [BGW, Section 5]. Note that hEF i is a large ideal of B(G) contained in the intersection E ∩ F . In [KLQa, Corollary 6.9] we showed that if E or F is exact then hEF i = E ∩ F . On the other hand, in [KLQb, proof of Propo- sition 8.4] we observed that it follows from work of [Oka14] that if G is a noncommutative free group and Ep is the weak-closure in B(G) of span{P (G) ∩ Lp(G)}, where P (G) denotes the set of positive type functions on G, then for for every p > 2 we have hE2 pi ⊂ Ep/2 ( Ep. Note that in [KLQb, Section 8], Ep was defined using B(G) ∩ Lp(G); it now seems clear that this should be changed to span{P (G) ∩ Lp(G)} -- see [BEWb, Proposition 2.13]. We are grateful to Buss, Echterhoff, and Willett for pointing this out to us. Another key idea in our strategy is to first do it for w-proper coac- tions. Although w-properness is a quite strong hypothesis, in some sense it is not: Lemma 3.5. Every coaction is Morita equivalent to a w-proper one. Proof. Let (A, δ) be a coaction, with maximalization (Am, δm). Since (Am, δm) is maximal, the double crossed product gives a coaction (B, ǫ), an Am −B imprimitivity bimodule X, and a δm −ǫ compatible coaction ζ on X. By [KLQc, Corollary 7.8], (B, ǫ) is w-proper since it is a dual coaction. Let I be the kernel of the maximalization map qm A : Am → A, and let J be the ideal of B induced via the imprimitivity bimodule X. Since the imprimitivity bimodule X is equivariant, there is a coaction eǫ on the quotient B/J such that the given coaction (A, δ) is Morita EXACT LARGE IDEALS OF B(G) ARE DOWNWARD DIRECTED 5 equivalent to (B/J,eǫ). By [KLQb, Proposition 5.3], the coaction eǫ is w-proper. (cid:3) Lemma 3.6. If E and F are large ideals of B(G), then for every w- proper coaction (A, δ) there is a unique isomorphism θA making the diagram QE A A QhEF i A AhEF i commute. /❴❴❴❴❴ ≃ θA AE QF AE (AE)F . Proof. We will show that ker QF AhEF i, this will imply that AE = AhEF i/AE. Since ker QhEF i A = ker QhEF i A = ker QF AE ◦ QE A, and the result will follow. For all a + AE ∈ AE = A/AE we have AE = (AE)F if and only if F · (a + AE) = {AE}, a + AE ∈ ker QF equivalently F · a ⊂ AE, equivalently EF · a = {0}. By definition, hEF i is the weak-closed span of EF . Since δ is w-proper, the map f 7→ f · a : B(G) → A is weak-to-weakly continuous, so EF ·a = {0} if and only if hEF i·a = {0}, i.e., a ∈ AhEF i. Thus ker QF AE = (AE)F = AhEF i/AE. (cid:3) The following result almost shows that the θA of Lemma 3.6 gives a natural isomorphism between the coaction functors τhEF i and τF ◦ τE: Lemma 3.7. Let E and F be large ideals of B(G). Let (A, δ) and (B, ǫ) be w-proper coactions, and let ψ : A → B be a δ − ǫ equivariant homomorphism. Then the diagram (3.1) AhEF i θA (AE)F ψhEF i (ψE )F BhEF i θB / (BE)F commutes equivariantly for the appropriate coactions. / / / / / / 6 KALISZEWSKI, LANDSTAD, AND QUIGG Proof. Equation (3.1) is the outer square of the following diagram: AhEF i gPPPPPPPPPPPPPPP QhEF i A A θA QE A AE AE QF w♦♦♦♦♦♦♦♦♦♦♦♦♦ (AE)F ψhEF i ψ ψE (ψE )F BhEF i 6♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥ QhEF i B B QE B θB BE QF 'PPPPPPPPPPPP BE / (BE)F . The left and right quadrilaterals commute by Lemma 3.6. The top, middle, and bottom quadrilaterals commute by functoriality. Since QhEF i is surjective, it follows that the outer square commutes. Since all maps except possibly for θA and θB are equivariant for appropriate coactions, the isomorphisms θA and θB are also equivariant. (cid:3) A Definition 3.8. Let τ be a coaction functor. We say that a coac- tion (A, δ) is τ -exact if for every strongly δ-invariant ideal I of A the sequence 0 / I τ / Aτ / (A/I)τ / 0 is exact. Thus, a coaction functor τ is exact if and only if every coaction is τ -exact. Lemma 3.9. If E and F are exact large ideals of B(G), then every w-proper coaction is τhEF i-exact. Proof. Let (A, δ) be a w-proper coaction, and let I be a strongly δ- invariant ideal of A. Then we have an equivariant short exact sequence 0 / I φ / A ψ / B / 0, where φ is the inclusion, B = A/I, and ψ is the quotient map. We must show that the sequence 0 / I hEF i φhEF i / AhEF i ψhEF i / BhEF i / 0 (3.2) is exact. Since E is exact, the sequence / I E φE 0 / AE ψE / BE / 0 / / / / g 6 / / w ' / / / / / / / / / / / / / / / / / EXACT LARGE IDEALS OF B(G) ARE DOWNWARD DIRECTED 7 is exact. Then since F is exact, the sequence / (I E)F (φE )F / (AE)F (ψE )F / (BE)F 0 is exact. By Lemma 3.7, we have an isomorphism 0 0 / I hEF i θI ≃ / (I E)F ψhEF i φhEF i AhEF i θA ≃ BhEF i θB≃ / (AE)F (φE )F (ψE )F / (BE)F / 0 0 / 0 of sequences, so the top sequence is exact since the bottom one is. (cid:3) The following is adapted from [BEWa, Definition 3.1]. Definition 3.10. We say a coaction functor τ has the ideal property if for every coaction (A, δ) and every strongly δ-invariant ideal I of A, letting ι : I ֒→ A denote the inclusion map, the induced map is injective. ιτ : I τ → Aτ Note that in the above definition, if τ has the ideal property then the image of I τ in Aτ will be a strongly δτ -invariant ideal, and we will identify I τ with this image, regarding it as an ideal of Aτ . [BEWa, Remark 3.4] says that the ideal property holds for every crossed-product functor coming from a large ideal. This also follows from the following lemma. Lemma 3.11. For every large ideal E of B(G) the coaction functor τE has the ideal property. Proof. This follows from [KLQa, Proof of Proposition 6.7], where it is shown that [KLQa, Equation (6.4)] holds automatically. (cid:3) Remark 3.12. Every exact coaction functor has the ideal property, but normalization is a coaction functor that is not exact but nevertheless has the ideal property. We do not know an example of a decreasing coaction functor that is Morita compatible and does not have the ideal property. Proposition 3.13. Let τ be a Morita compatible coaction functor with the ideal property, and let (A, δ) and (B, ǫ) be Morita equivalent coac- tions. Then (A, δ) is τ -exact if and only if (B, ǫ) is. / / / / / / / / / / / / / / / 8 KALISZEWSKI, LANDSTAD, AND QUIGG Proof. Let X be an equivariant A−B imprimitivity bimodule. Without loss of generality assume that (B, ǫ) is τ -exact, let I be an invariant ideal of A, let J be the ideal of B corresponding to I via X, and let ψA : A → A/I ψB : B → B/J be the quotient maps. We know that ψτ and has kernel J τ , because we are assuming that (B, ǫ) is τ -exact. B : Bτ → (B/J)τ is surjective Since maximalization is an exact coaction functor [KLQa, Theo- rem 4.11], the homomorphism ψm A : Am → (A/I)m is surjective. Since qτ is a natural transformation from maximalization to τ , the diagram Am qτ A Aτ ψm A (A/I)m qτ A/I / (A/I)τ ψτ A commutes. Thus ψτ A is surjective. Since τ has the ideal property, I τ is an ideal of Aτ . For τ -exactness of (A, δ), it remains to show that I τ = ker ψτ A. B compatible imprimitivity-bimodule homomorphism qτ Since τ is Morita compatible, by [KLQa, Lemma 4.19] we have an equivariant Aτ − Bτ imprimitivity bimodule X τ and a surjective qτ A −qτ X : X m → X τ , where X m is the equivariant Am − Bm imprimitivity bimodule of [KLQa, Lemma 4.14]. (Note that [KLQa, Lemma 4.19] did not explic- itly mention surjectivity of qτ X, but this surjectivity follows from that of qτ A and qτ B.) We visualize this using the diagram Am X m Bm qτ A qτ X qτ B Aτ X τ Bτ Similarly for qτ X/Y : (X/Y )m → (X/Y )τ . / / / EXACT LARGE IDEALS OF B(G) ARE DOWNWARD DIRECTED 9 Consider the diagram (3.3) Am X m PPPPPPPPPPPPPP Bm ❘❘❘❘❘❘❘❘❘❘❘❘❘❘❘❘ (PPPPPPPPPPPP ❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚ )❘❘❘❘❘❘❘❘❘❘❘❘❘ ψm X ψm A ψm B qτ A qτ X qτ B (A/I)m )❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚ (X/Y )m (B/J)m qτ A/I qτ X/Y qτ B/J Aτ X τ ❘ PPPPPPPPPPPPPP ❘ Bτ ❘ ❘ ❘ ❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚ ❘ ❘ ❘ ψτ X ❘ ❘ ❘ ❘ ❘ ψτ A (PPPPPPPPPPPP (A/I)τ ψτ B ❘ )❘ )❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚❚ (X/Y )τ (B/J)τ . Claim: there is an imprimitivity-bimodule homomorphism ψτ indicates, with coefficient homomorphisms ψτ map ψτ X such that the diagram A and ψτ X as (3.3) B. To get a linear ψm X X m (X/Y )m qτ X qτ X/Y X τ /❴❴❴ ψτ X (X/Y )τ X . Since ker qτ commutes, it suffices to show that ker qτ ker qτ theorem we can factor x = x′ · b, where b ∈ ker qτ qτ X/Y ◦ ψm X = X m · ker qτ X ⊂ ker qτ X (x) τ . Suppose x ∈ B, by the Cohen-Hewitt factorization X/Y ◦ ψm B. Then X (x′ · b) X/Y ◦ ψm A/I (x′) · qτ (since qτ A/I (x′) · ψτ B/J ◦ ψm B (b) X/Y and ψm B(b) B ◦ qτ = qτ = qτ = qτ = 0, X are imprimitivity-bimodule homomorphisms) (by naturality of qτ ) as desired. The computations required to verify that the linear map ψτ X is an imprimitivity-bimodule homomorphism are routine: for the ( ) ) ( ) ) / / / 10 KALISZEWSKI, LANDSTAD, AND QUIGG right-module structures, let x ∈ X τ and b ∈ Bτ . By surjectivity we can write x = qτ B(b′) with x′ ∈ X m and b′ ∈ Bm, and then X (x′) and b = qτ X (x · b) = ψτ ψτ = ψτ = qτ ∗= qτ = ψτ = ψτ X (x′) · qτ X(x′ · b′) B(b′)(cid:1) X (x′ · b′) X (x′) · qτ X(cid:0)qτ X ◦ qτ X/Y ◦ ψm X/Y ◦ ψm X ◦ qτ X (x) · ψτ X(x′) · ψτ B(b), B (b′) B/J ◦ ψm B(b′) B ◦ qτ where the equality at ∗ follows since qτ X are imprimitivity- bimodule homomorphisms. Similarly for the left-module structures. For the right-hand inner products, let x, y ∈ X τ . Factor x = qτ X (x′) and y = qτ X/Y and ψm X(y′) with x′, y′ ∈ X m. Then A(cid:0)Aτ hx, yi(cid:1) = ψτ ψτ = ψτ = qτ A(cid:0)Aτ hqτ A ◦ qτ A/I ◦ ψm = (A/I)τ(cid:10)qτ = (A/I)τ(cid:10)ψτ = (A/I)τ(cid:10)ψτ X(y′)i(cid:1) X(x′), qτ A(cid:0)Amhx′, y′i(cid:1) A(cid:0)Amhx′, y′i(cid:1) X (x′), qτ X/Y ◦ ψm X ◦ qτ X (x), ψτ X(x′), ψτ X(y)(cid:11), X (y′)(cid:11) X/Y ◦ ψm X (y′)(cid:11) X ◦ qτ and similarly for the right-hand inner products. Thus ψτ bimodule homomorphism with coefficient homomorphisms ψτ proving the claim. X is an imprimitivity- A and ψτ B, It now follows from [EKQR06, Lemma 1.20] that ker ψτ A = X τ − Ind ker ψτ B = X τ − Ind J τ . Since we also have the imprimitivity-bimodule homomorphism ψm coefficient homomorphisms ψm exact coaction functor, X with B , and since maximalization is an A and ψm I m = ker ψm A = X m − Ind ker ψm B = X m − Ind J m. Now, δ restricts to a coaction (I, qI), and by surjectivity of qτ we have I τ = qτ B(I m), EXACT LARGE IDEALS OF B(G) ARE DOWNWARD DIRECTED 11 B(J m). Combining, we get and similarly J τ = qτ I τ = qτ = qτ = X τ − Ind qτ A(I m) A(X m − Ind J m) B(J m) (since qτ X is an imprimitivity-bimodule homomorphism) = X τ − Ind J τ = X τ − Ind ker ψτ B = ker ψτ A, finishing the proof. (cid:3) Proof of Theorem 3.2. Let E and F be exact large ideals. By [KLQa, Corollary 6.9] we have E ∩ F = hEF i, and by Lemma 3.11 the coac- tion functor τhEF i has the ideal property. Further, by [KLQa, Proposi- tion 6.10] τhEF i is Morita compatible. The conclusion now follows from Lemma 3.9, Proposition 3.13, and Lemma 3.5. (cid:3) Remark 3.14. The technique of proof of [KLQa, Theorem 4.22] shows that the greatest lower bound of any collection of exact coaction func- tors is exact. Thus, it might seem that Theorem 3.2 above implies that the intersection E of all exact large ideals of B(G) is exact. However, it is not clear to us how to show that τE coincides with the greatest lower bound of {τF : F is a exact large ideal}; it is certainly no larger than this greatest lower bound, but that is all we can prove at this point. To see what the problem is, let {Ei} be the set of exact large ideals of in the ideal AE. This is true for (C ∗(G), δG) since then AE = ⊥E B(G), so that E = Ti Ei. The issue is whether, for a given coaction (A, δ), the union Si AEi of the upward-direct family of ideals is dense and Si ⊥Ei)⊥ = Ti(⊥Ei)⊥ = E. In natural way). Obviously EA∗ ⊂ Ti span{EiA∗}, but we cannot see a the general case, we have (AE)⊥ = span{EA∗} (the weak-closure of the linear span of products, where E acts on the dual space A∗ in the ⊥Ei is dense in ⊥E because (Si reason to expect span{EA∗} to be weak-dense in this intersection. Remark 3.15. [BEWa, Subsection 9.2 Question (1)] asks whether, for every exact group G and all p ∈ [2, ∞), the crossed-product functor ⋊Ep is exact, where Ep is the weak-closure of B(G) ∩ Lp(G) (which should be changed to span{P (G) ∩ Lp(G)}, as in the discussion preceding Lemma 3.5 of the current paper and in [BEWb, Proposition 2.13]). We know that if G is a free group Fn with n > 1, then for 2 ≤ p < ∞ the coaction functor τEp is not exact. Of course, Fn is exact. We think we might be able to deduce that ⋊Ep is not exact. Note that this is 12 KALISZEWSKI, LANDSTAD, AND QUIGG if we compose a coaction functor τ with the full-crossed- nontrivial: product functor CP that takes an action (B, α) to the dual coaction (B ⋊α G,bα), we get a crossed-product functor µτ := τ ◦ CP that takes (B, α) to the coaction (cid:0)(B ⋊α G)τ , (bα)τ(cid:1). By [KLQa, Proposition 4.24], if τ is exact or Morita compatible then so is µτ . But to give a negative answer to the [BEWa] question we would be trying to draw a conclusion that goes in the "wrong direction". Example 3.16. [KLQa, Question 6.20] asks whether for every coaction functor τ there necessarily exists a large ideal E of B(G) such that the restrictions of τ and τE to the subcategory of maximal coactions (but still taking values in the ambient category of coactions) are naturally isomorphic. Borrowing a trick from Buss, Echterhoff, and Willett, we give a negative answer. We adapt a construction from [BEWa, Section 2.5 and Example 3.5]. Let R be a collection of coactions. For each coaction (A, δ) let RA,δ be the collection of all triples (B, ǫ, φ), where either (B, ǫ) ∈ R and φ : (A, δ) → (B, ǫ) is a morphism in Coact or (B, ǫ) = (An, δn) and ǫ : (A, δ) → (An, δn) is the normalization surjection. Then let   M (B,ǫ,φ)∈RA,δ (B, ǫ), M (B,ǫ,φ)∈RA,δ ǫ  be the direct-sum coaction. We can form the direct sum QR A := M (B,ǫ,φ)∈RA,δ φ : A → M   M (B,ǫ,φ)∈RA,δ B  , which is a nondegenerate δ − M (B,ǫ,φ)∈RA,δ ǫ equivariant homomorphism. Let AR be the image of A under this direct-sum homomorphism QR A . Then by the elementary Lemma 3.17 below there is a unique coaction δR of G on AR such that QR A is δ − δR equivariant. EXACT LARGE IDEALS OF B(G) ARE DOWNWARD DIRECTED 13 Claim: for every morphism φ : (A, δ) → (B, ǫ) in Coact, there is a unique morphism φR in Coact making the diagram (A, δ) QR A φ (B, ǫ) QR B (AR, δR) /❴❴❴❴❴ φR (BR, ǫR) commute. We have ker QR A = \ (C,η,ψ)∈RA,δ ker ψ, ker ψ ◦ φ. while Since we have ker QR B ◦ φ = \ (C,η,ψ)∈RB,ǫ {ψ ◦ φ : (C, η, ψ) ∈ RB,ǫ} ⊂ RA,δ, ker QR A ⊂ ker QR B ◦ φ, and the claim follows. Uniqueness of the maps φR and surjectivity of the maps QR that there is a unique decreasing coaction functor τR such that A implies (AτR, δτR) = (AR, δR) and φτR = φR (see [KLQa, Definition 5.1 and Lemma 5.2]). We will show that, whenever G is nonamenable, there is a suitable choice of R for which the coaction functor τR is not Morita compatible, and therefore its restriction to the maximal coactions is not naturally isomorphic to τE for any large ideal E of B(G). Let (A, δ) = (cid:0)C[0, 1) ⊗ C ∗(G), id ⊗ δG(cid:1). We let R = {(A, δ)}. The coactions (A, δ) and (K ⊗ A, id ⊗ δ) are Morita equivalent. We claim that QR K⊗A is not. Since the coaction functor τR is decreasing, it will follow that τR is not Morita compatible. A is faithful but QR The triple (A, δ, id) is in the collection RA,δ, which implies that QR A is faithful. On the other hand, we claim that the only morphism in the collection RK⊗A,id⊗δ is the normalization idK⊗C[0,1) ⊗ λ : K ⊗ C[0, 1) ⊗ C ∗(G) → K ⊗ C[0, 1) ⊗ C ∗ r (G). / / / 14 KALISZEWSKI, LANDSTAD, AND QUIGG Since G is nonamenable, this normalization is not faithful. To verify the claim, it will suffice to show that there are no nonzero homomorphisms from K ⊗ A to A. Any such homomorphism would be of the form ψ1 × ψ2, where ψ1 and ψ2 are commuting homomorphisms from K and A, respectively, to A, i.e., (ψ1 × ψ2)(k ⊗ a) = ψ1(k)ψ2(a). Since A has no nonzero projections, the homomorphism ψ1 must be 0, and so ψ1 × ψ2 = 0. In Example 3.16, we used the following lemma, which is presumably folklore. Since we could not find it in the literature we include the proof. Lemma 3.17. Let (A, δ) and (B, ǫ) be coactions, and let φ : A → M(B) be a δ − ǫ equivariant homomorphism. Let C = φ(A) ⊂ M(B). Then there is a unique coaction η of G on C such that φ : A → C is δ − η equivariant. Proof. By [Qui94, Corollary 1.7], it suffices to show that C is a nonde- generate A(G)-submodule of M(B). Now, [KLQa, Proposition A.1] says that a homomorphism from A to B is δ − ǫ equivariant if and only if it is a B(G)-module map. We need a slight extension of this, namely the case of homomorphisms φ : A → M(B). The argument of [KLQa, Proposition A.1] carries over, with the minor adjustment that in the second line of the multiline displayed computation the map φ⊗id must be replaced by the canonical extension φ ⊗ id : fM (A ⊗ C ∗(G)) → M(B ⊗ C ∗(G)), which exists by [EKQR06, Proposition A.6]. Thus, since we are assum- ing δ − ǫ equivariance, we can conclude that A(G) · φ(A) = φ(cid:0)A(G) · A(cid:1). Since A is a nondegenerate A(G)-module, we are done. (cid:3) References [BGW] [BEWa] P. Baum, E. Guentner, and R. Willett, Expanders, exact crossed prod- ucts, and the Baum-Connes conjecture, arXiv:1409.4332 [math.OA]. A. Buss, S. Echterhoff, and R. Willett, Exotic crossed products and the Baum-Connes conjecture, arXiv:1409.4332 [math.OA]. [BEWb] , Exotic Crossed Products, arXiv:1510.02556 [math.OA]. EXACT LARGE IDEALS OF B(G) ARE DOWNWARD DIRECTED 15 [EKQR06] S. Echterhoff, S. Kaliszewski, J. Quigg, and I. Raeburn, A Categorical Approach to Imprimitivity Theorems for C-Dynamical Systems, vol. 180, Mem. Amer. Math. Soc., no. 850, American Mathematical Society, Providence, RI, 2006. S. Kaliszewski, M. B. Landstad, and J. Quigg, Coaction functors, arXiv:1505.03487 [math.OA]. [KLQa] [KLQb] [KLQc] , Exotic coactions, Proc. Edinburgh Math. Soc., to appear, arXiv:1305.5489 [math.OA]. , Properness conditions for actions and coactions, arXiv:1504.03394 [math.OA]. [KLQ13] , Exotic group C ∗-algebras in noncommutative duality, New York [Oka14] [Qui94] J. Math. 10 (2013), 1 -- 23. R. Okayasu, Free group C ∗-algebras associated with ℓ Math. 25 (2014), no. 7, 1450065, 12pp. J. C. Quigg, Full and reduced C ∗-coactions, Math. Proc. Cambridge Philos. Soc. 116 (1994), 435 -- 450. p, Internat. J. School of Mathematical and Statistical Sciences, Arizona State University, Tempe, Arizona 85287 E-mail address: [email protected] Department of Mathematical Sciences, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway E-mail address: [email protected] School of Mathematical and Statistical Sciences, Arizona State University, Tempe, Arizona 85287 E-mail address: [email protected]
1611.10218	1	1611	2016-11-30T15:26:51	Low rank compact operators and Tingley's problem	[ "math.OA", "math.FA" ]	Let $E$ and $B$ be arbitrary weakly compact JB$^$-triples whose unit spheres are denoted by $S(E)$ and $S(B)$, respectively. We prove that every surjective isometry $f: S(E) \to S(B)$ admits an extension to a surjective real linear isometry $T: E\to B$. This is a complete solution to Tingley's problem in the setting of weakly compact JB$^$-triples. Among the consequences, we show that if $K(H,K)$ denotes the space of compact operators between arbitrary complex Hilbert spaces $H$ and $K$, then every surjective isometry $f: S(K(H,K)) \to S(K(H,K))$ admits an extension to a surjective real linear isometry $T: K(H,K)\to K(H,K)$.	math.OA	math	LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM FRANCISCO J. FERN ´ANDEZ-POLO AND ANTONIO M. PERALTA Abstract. Let E and B be arbitrary weakly compact JB∗-triples whose unit spheres are denoted by S(E) and S(B), respectively. We prove that every surjective isometry f : S(E) → S(B) admits an extension to a surjective real linear isometry T : E → B. This is a complete solution to Tingley's prob- lem in the setting of weakly compact JB∗-triples. Among the consequences, we show that if K(H, K) denotes the space of compact operators between arbitrary complex Hilbert spaces H and K, then every surjective isometry f : S(K(H, K)) → S(K(H, K)) admits an extension to a surjective real linear isometry T : K(H, K) → K(H, K). 1. Introduction It does not seem an easy task to write an introductory paragraph for a problem which has been open since 1987. As in most important problems, the precise question is easy to pose and reads as follows: Let X and Y be normed spaces, whose unit spheres are denoted by S(X) and S(Y ), respectively. Suppose f : S(X) → S(Y ) is a surjective isometry. The so-called Tingley's problem asks wether f can be extended to a real linear (bijective) isometry T : X → Y between the corresponding spaces (see [35]). The problem was named after D. Tingley proved in [35, THEOREM, page 377] that every surjective isometry f : X → Y between the unit spheres of two finite dimensional spaces satisfies f (−x) = −f (x) for every x ∈ S(X). Readers interested in a classic motivation, can sail back to the celebrated Mazur- Ulam theorem asserting that every surjective isometry between two normed spaces over R is a real affine function. In a subsequent paper, P. Mankiewicz established in [27] that, given two convex bodies V ⊂ X and W ⊂ Y , every surjective isometry g from V onto W can be uniquely extended to an affine isometry from X onto Y . Consequently, every surjective isometry between the closed unit balls of two Banach spaces X and Y extends uniquely to a real linear isometric isomorphism from X into Y . With this historical background in mind, Tingley's problem asks whether the conclusion in Mankiewicz's theorem remains true when we deal with the unit sphere whose interior is empty. In other words, when an isometric identification of the unit spheres of two normed spaces can produce an isometric (linear) identification of the spaces. 2010 Mathematics Subject Classification. Primary 47B49, Secondary 46A22, 46B20, 46B04, 46A16, 46E40, . Key words and phrases. Tingley's problem; extension of isometries; JB∗-triples; compact operators. 2 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA As long as we know, Tingley's problem remains open even for surjective isome- tries between the unit spheres of a pair of 2 dimensional Banach spaces. However, positive answers haven been established in a wide range of classical Banach spaces. In an interesting series of papers, G.G. Ding proved that Tingley's problem ad- mits a positive answer for every surjective isometry f : S(ℓp(Γ)) → S(ℓp(∆)) with 1 ≤ p ≤ ∞ (see [10, 11, 12] and [13]). More recently, D. Tan showed that the same conclusion remains true for every surjective isometry f : S(Lp(Ω, Σ, µ)) → S(Y ), where (Ω, Σ, µ) is a σ-finite measure space, 1 ≤ p ≤ ∞, and Y is a Banach space (compare [30, 31] and [32]). A result of R.S. Wang in [36] proves that for each pair of locally compact Hausdorff spaces L1 and L2, every surjective isometry f : S(C0(L1)) → S(C0(L1)) admits an extension to a surjective real linear isometry from C0(L1) onto C0(L1). V. Kadets and M. Mart´ın gave another positive answer to Tingley's problem in the case of finite dimensional polyhedral Banach spaces (see [22]). The surveys [14] and [37] contain a detailed revision of these results and additional references. In the setting of C∗-algebras, R. Tanaka recently establishes in [33] that every surjective isometry from the unit sphere of a finite dimensional C∗-algebra N into the unit sphere of another C∗-algebra M admits a unique extension to a surjective real linear isometry from N onto N . More recently, Tanaka also proves in [34] that the same conclusion holds when N and M are finite von Neumann algebras. As a result of a recent collaboration between the second author of this note and R. Tanaka (see [29]), new positive answers to Tingley´s problem have been revealed for spaces of compact operators. Concretely, denoting by K(H) the C∗-algebra of all compact operators on a complex Hilbert space H, it is shown that for every pair of complex Hilbert spaces H and H′, every surjective isometry f : S(K(H)) → S(K(H′)) admits a unique extension to a real linear isometry T from K(H) onto K(H′); and the same conclusion also holds when K(H) and K(H′) are replaced by arbitrary compact C∗-algebras (compare [29, Theorem 3.14]). The novelties in the just quoted note rely on the introduction of Jordan techniques to tackle Tingley's problem. From the wider point of view of weakly compact JB∗-triples, it is established that if E and B are weakly compact JB∗-triples not containing direct summands of rank smaller than or equal to 3, and with rank greater than or equal to 5, then surjective isometry f : S(E) → S(B) admits a unique extension to a surjective real linear isometry from E onto B (compare [29, Theorem 3.13] and see below for the concrete definitions of weakly compact JB∗-triples). The case of surjective isometries between the unit spheres of two weakly com- pact JB∗-triples of rank between 2 and 4 was a left as an open problem in [29]. This problem affects particularly interesting cases including spin factors, finite di- mensional Cartan factors, and space of the form K(H, H′) where H and H′ are complex Hilbert spaces with dim(H) = ∞ and 2 ≤dim(H′) ≤ 4. In this paper we complete the study left open in [29], solving Tingley's problem in the remaining cases of weakly compact JB∗-triples of low rank and providing a complete solu- tion of Tingley's problem for surjective isometries between the unit spheres of two weakly compact JB∗-triples. Besides Jordan techniques, the arguments applied in this note are strongly based on new geometric properties for Cartan factors, and results in Functional Analysis and Operator algebras. In order to have a precise idea of the results explored in this paper, it seems necessary to arrange the definition of those complex Banach spaces called Cartan LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 3 factors. All Banach spaces considered in this note are complex. Let H and H′ be complex Hilbert spaces, and let j : H → H be a conjugation (i.e., a conjugate linear isometry of order two) on H. A Banach space is called a Cartan factor of type 1 if it coincides with the space L(H, H′) of all bounded linear operators from H into H′. To understand the next Cartan factors let t : L(H) → L(H) be the complex linear involution on L(H) defined by xt = jx∗j (x ∈ L(H)). The spaces C2 = {x ∈ L(H) : xt = −x} and C3 = {x ∈ L(H) : xt = x} are closed subspaces of L(H) called Cartan factors of type 2 and 3, respectively. A Banach space X is called a Cartan factor of type 4 or spin if X admits a complete inner product (..) and a conjugation x 7→ x, for which the norm of X is given by kxk2 = (xx) +p(xx)2 − (xx)2. Cartan factors of types 5 and 6 (also called exceptional Cartan factors) are all finite dimensional. We refer to [6] for additional details. According to the terminology employed by L. Bunce and Ch.-H. Chu in [3], as- sociated with each Cartan factor we find an elementary JB∗-triple. The elementary JB∗-triples of types 1 to 6 (K1, . . . , K6) are defined as follows: types 4, 5 and 6 are precisely the Cartan factors of the same type, i.e. K4 = C4, K5 = C5, and K6 = C6. For the other types, we introduce some additional notation. The symbol K(H, H′) will denote the compact linear operators from H into H′. The elemen- tary JB∗-triples of types 1, 2 and 3 are K1 = K(H, H′), K2 = C2 ∩ K(H), and K3 = C3 ∩ K(H), respectively. Elementary JB∗-triples are the building blocks of weakly compact JB∗-triple, more precisely, every weakly compact JB∗-triple can be written as a c0-sum of a family of elementary JB∗-triples (see [3, Lemma 3.3 and Theorem 3.4]). The main result in this note (Theorem 4.1) establishes that every surjective isom- etry f : S(E) → S(B) between the unit spheres of two arbitrary weakly compact JB∗-triples admits a unique extension to a surjective real linear isometry T : E → B. The proof, which requires a substantial technical novelty, is divided into different partial results. In Section 4 we explore Tingley's problem when the domain is an elementary JB∗-triple of type 1, 2, 3, spin or finite dimensional (see Theorems 4.4, 4.5, 4.6, 4.7, and 4.9). Previously, in Section 3, we establish additional properties satisfied by every surjective isometry between the unit spheres of two elementary JB∗-triples. This geometric study complete the result obtained in [29] for elemen- tary JB∗-triples with rank greater than or equal to 5. For completeness reasons, the structure of the paper is supplemented with a subsection with basic notions, results and references applied in this note. It seems natural to conclude this introduction with a desiderata; once Tingley's problem has been solved for compact operators, compact C∗-algebras and weakly compact JB∗'triples, the unexplored frontier appears now in the case of B(H), general von Neumann algebras and C∗-algebras, and other operator spaces. 1.1. Basic background on JB∗-triples. In this subsection we revisit the basic notions and result in JB∗-triples. The reader who is familiar with these concepts can simply skip this part. Throughout this note the symbol BX will stand for the closed unit ball of a Banach space X. 4 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA Let H and H′ a pair of complex Hilbert spaces. In most of cases, the space L(H, H′) is not a C∗-algebra. An example of this claim appears when H is finite dimensional and H′ is infinite dimensional. However, the operator norm and the triple product defined by (1) {a, b, c} = 1 2 (ab∗c + cb∗a) (a, b, c ∈ L(H, H′)) equip L(H, H′) with a structure of JB∗-triple in the sense introduced by W. Kaup in [24]. A JB∗-triple is a complex Banach space E admitting a continuous triple product {a, b, c} which is conjugate linear in b and linear and symmetric in a and c, and satisfies the following axioms: (JB∗1) L(a, b)L(c, d) − L(c, d)L(a, b) = L(L(a, b)(c), d) − L(c, L(b, a)(d)), for every a, b, c, d in E, where L(a, b) is the operator on E defined by L(a, b)(x) = {a, b, x}; (JB∗2) L(a, a) is an hermitian operator on E with non-negative spectrum; (JB∗3) k{a, a, a}k = kak3, for every a ∈ E. Examples of JB∗-triples include the spaces L(H, H′), the spaces K(H, H′) of all compact operators between two complex Hilbert spaces, complex Hilbert spaces, and all C∗-algebras equipped with the same product defined in (1). JB∗-triples con- stitute a category which produces a Jordan model valid to generalize C∗-algebras. Every JB∗-algebra is a JB∗-triple under the triple product {a, b, c} = (a ◦ b∗) ◦ c + (c ◦ b∗) ◦ a − (a ◦ c) ◦ b∗. Cartan factors and elementary JB∗-triples of types 1, 2 and 3 are JB∗-triples with respect to the product given in (1). Spin factors are JB∗-triples when equipped with the product {x, y, z} = (xy)z + (zy)x − (xz)y. The exceptional Cartan factor C6 is a JB∗-algebra and C5 is a JB∗-subtriple of C6. A closed subtriple I of a JB∗-triple E is called an ideal in E if {I, E, E} + {E, I, E} ⊆ I holds. A JB∗-triple which cannot be decomposed into the direct sum of two non-trivial closed ideals is called a JB∗-triple factor. Those JB∗-triples which are also dual Banach spaces are called JBW∗-triples. Therefore, von Neumann algebras are examples of JBW∗-triples. Analogously as in the case of von Neumann algebras, JBW∗-triples admit a unique (isometric) predual and their triple product is separately weak∗-continuous (see [1]). The bidual, E∗∗, of a JB∗-triple, E, is a JBW∗-triple with a triple product extending the triple product of E (cf. [9]). Cartan factors are all JBW∗-triple factors. Partial isometries in a C∗-algebra can be generalized to tripotents in a JB∗-triple. More concretely, an element u in a JB∗-triple E is a tripotent if {u, u, u} = u. The set of all tripotents in a JB∗-triple E will be denoted by the symbol U(E). For each tripotent u in E, the mappings Pi(u) : E → Ei(u), (i = 0, 1, 2), defined by (2) P2(u) = L(u, u)(2L(u, u) − idE), P1(u) = 4L(u, u)(idE − L(u, u)), and P0(u) = (idE − L(u, u))(idE − 2L(u, u)), are contractive linear projections. Each Pj (u) is known as the Peirce-j projection induced by u, and corresponds to the projection of E onto the eigenspace Ej(u) of L(u, u) corresponding to the eigenvalue j 2 . The Peirce decomposition of E relative to u writes E in the form E = E2(u) ⊕ E1(u) ⊕ E0(u). LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 5 The following Peirce rules are satisfied, (3) {E2(u), E0(u), E} = {E0(u), E2(u), E} = 0, {Ei(u), Ej(u), Ek(u)} ⊆ Ei−j+k(u), (4) where Ei−j+k(u) = 0 whenever i − j + k /∈ {0, 1, 2} (compare [19]). It follows from (2) and the separate weak∗-continuity of the triple product that Peirce projections associated with a tripotent in a JBW∗-triple are weak∗ continuous. It is further known that E2(u) is a JB∗-algebra with product and involution determined by a ◦u b = {a, u, b} and a♯u = {u, a, u}, respectively. A non-zero tripotent u is called minimal if E2(u) = Cu, complete if the Peirce subspace E0(u) vanishes, and unitary if E = E2(e). Two tripotents u, e in E are said to be orthogonal (written e ⊥ u) if {e, e, u} = 0 (or equivalently, {u, u, e} = 0). Actually, the relation ⊥ can be considered on the whole E by defining a ⊥ b if {a, a, b} = 0 (see [5, §1] for additional details). It is known that orthogonal elements in E are geometrically M -orthogonal (see [19, Lemma 1.3] and [5, Lemma 1.1]), i.e. a ⊥ b ⇒ ka ± bk = max{kak,kbk}. (5) The relation of orthogonality is adequate to define a partial order in U(E) given by u ≤ e if e − u is a tripotent orthogonal to e (see, for example, [19] or [20]). The set U(E) is a lattice with respect to the partial order ≤. A tripotent u in a JBW∗-triple W is minimal if and only if it is a minimal element in the lattice (U(E),≤). A subset S in a JB∗-triple E is called orthogonal if 0 /∈ S and x ⊥ y for every x 6= y in S. The rank of E (denoted by rank(E) or by r(E)) is the minimal cardinal r satisfying that ♯S ≤ r for every orthogonal subset of E (compare [25]). The rank of a tripotent e in E is defined as the rank of the JB∗-triple E2(e). JB∗-triples of finite rank are all reflexive, they are described, for example in [2]. It is known that in a Cartan factor C, the rank of C is precisely the minimal cardinal r satisfying that ♯S ≤ r for every orthogonal subset of minimal tripotents in E (see [25, §3 or Theorem 5.8] and [2]). Let u, v be tripotents in a JB∗-triple E. We say that u and v are collinear (written u⊤v) if u ∈ E1(v) and v ∈ E1(u). Given u, v ∈ U(E) with u ∈ El(v) for some l = 0, 1, 2 then (6) Pk(u)Pj(v) = Pj(v)Pk(u), for all j, k ∈ {0, 1, 2} (compare [20, (1.9) and (1.10)]). There are JB∗-triples E for which U(E) is empty. However, since in a JB∗-triple E the extreme points of its closed unit ball are precisely the complete tripotents in E (see [26, Proposition 3.3]), we can obviously conclude, via Krein-Milman theorem, that every JBW∗-triple contains a huge set of tripotents. For each normal functional ϕ in the predual of a JBW∗-triple W , there exists a unique tripotent u ∈ W (called the support tripotent of ϕ in W ) satisfying ϕ = ϕP2(u) and ϕM2(u) is a faithful positive normal functional on the JBW∗-algebra W2(u) (see [19, Proposition 2]). The support tripotent of a normal functional ϕ will be denoted by e(ϕ). Suppose u is another tripotent in W such that ϕ(u) = 1 = kϕk. It follows from the arguments in [19, Propositions 1 and 2] and their proofs that 6 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA u = s(ϕ) + P0(s(ϕ))(u) (i.e. e ≤ u). Actually, the same arguments show the following: (7) x ∈ W, with kxk = 1 = ϕ(x) = kϕk ⇒ x = s(ϕ) + P0(s(ϕ))(x). By Proposition 4 in [19] we also know that minimal tripotents in a JBW∗-triple W are precisely the support tripotents of the extreme points of the closed unit ball of its predual. Suppose x is an element in a JB∗-triple E. The symbol Ex will denote the JB∗- subtriple generated by x, that is, the closed subspace generated by all odd powers of the form x[1] := x, x[3] := {x, x, x}, and x[2n+1] := (cid:8)x, x, x[2n−1](cid:9) , (n ∈ N). It is known that Ex is JB∗-triple isomorphic (and hence isometric) to a commutative C∗-algebra in which x is a positive generator (cf. [24, Corollary 1.15]). Suppose x is a norm-one element. The sequence (x[2n−1]) converges in the weak∗-topology of E∗∗ to a tripotent (called the support tripotent of x) u(x) in E∗∗ (see [16, Lemma 3.3] or [15, page 130]). C.M. Edwards and G.T. Ruttimann developed in [17] an analogue to the notion of compact projections in the bidual of a C∗-algebra in the more general setting of JB∗-triples. A tripotent e in the second dual, E∗∗, of a JB∗-triple E is said to be compact-Gδ if there exists a norm one element a in E satisfying u(a) = e. A tripotent e in E∗∗ is called compact if e = 0 or it is the infimum of a decreasing net of compact-Gδ tripotents in E∗∗. The symbol Uc(E∗∗) will stand for the lattice consisting of those compact tripotents in E∗∗ equipped with the order ≤. Theorem 3.4 in [4] shows that minimal tripotents in E∗∗ are compact, and con- sequently the relation min U(E∗∗) ⊆ min Uc(E∗∗) holds, where "min" denotes the minimal elements in the corresponding lattice. It is also observed in [4, page 47, comments after Theorem 3.4] that (8) min U(E∗∗) = min Uc(E∗∗). The weak∗-closed ideal J(v) generated by a minimal tripotent v in a JBW∗-triple W , is precisely a Cartan factor. It is further know that W writes as the direct orthogonal sum of J(v) and another weak∗-closed of W (see [8, MAIN THEOREM in page 302]). 2. Tingley's problem for weakly compact JB∗-triples which are not factors In order to resume the study of surjective isometries between the unit spheres of two weakly compact JB∗-triples, we start by recalling a result in [29]. Proposition 2.1. [29, Propositions 3.2 and 3.9] Let E and B be weakly compact JB∗-triples, and suppose that f : S(E) → S(B) is a surjective isometry. Then the following statements hold: (a) For each finite rank tripotent e in E there exists a unique finite rank tripotent u in B such that f ((e + BE ∗∗ 0 (e)) ∩ E) = (u + BB∗∗ 0 (u)) ∩ B; (b) The restriction of f to each norm closed face of BE is an affine function; (c) For each finite rank tripotent e in E there exists a unique finite rank tripotent u in B and a surjective real linear isometry Te : E0(e) → B0(u) such that f (e + x) = u + Te(x), for every x ∈ BE0(e), and Te(x) = f (x) for all x ∈ S(E0(e)); LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 7 (d) For each finite rank tripotent e in E there exists a unique finite rank tripotent (cid:3) u in B such that f (e) = u. One of the obstacles to the questions left open in [29] is mainly due to the fact that the arguments in [29, Lemma 3.4] are not valid for arbitrary weakly compact JB∗-triples. We begin the new contributions with a generalization of [29, Lemma 3.4] for general JB∗-triples. 2 has norm-one in E. Let u = s( e−x Proposition 2.2. Let e and x be norm-one elements in a JB∗-triple E. Suppose that e is a minimal tripotent and ke − xk = 2. Then x = −e + P0(e)(x). Proof. The element e−x 2 ) denote the support tripotent of e−x in the JBW∗-triple E∗∗. It is known that u is a compact(-Gδ) 2 tripotent in E∗∗ and e−x 2 ) (see [17, §4]). Since the lattice Uc(E∗∗) of those compact tripotents in E∗∗ is atomic (see [17, Theorem 4.5(ii)]), we can find at least a minimal element v ∈ min Uc(E∗∗) = min U(E∗∗) satisfying v ≤ u. Therefore, we can write 2 = u + P0(u)( e−x ) in E∗∗, and hence e − x 2 = v + P0(v)( e − x 2 1 2 P2(v)(e) + 1 2 P2(v)(−x) = P2(v)( e − x 2 ) = v. Since v is the unit of E∗∗2 (v) = Cv, and hence it is an extreme point of the closed unit of the latter space, we deduce that P2(v)(e) = P2(v)(−x) = v. The minimality of e in E (and in E∗∗) shows that e = v (compare [19, Corollary 1.7]). Having in mind that x is a norm-one element in E, it follows from [19, Lemma 1.6] that x = −v + P0(v)(x) = −e + P0(e)(x) in E. (cid:3) Many important consequences can be now derived from the above strengthened version of [29, Lemma 3.4]. For example, when in the proof of [29, Theorem 3.6] we replace [29, Lemma 3.4] with Proposition 2.2, the arguments remain valid, word- by-word, to prove the following: Theorem 2.3. Let f : S(E) → S(B) be a surjective isometry between the unit spheres of two weakly compact JB∗-triples. Suppose e is a finite rank tripotent in E. Then f (−e) = −f (e). Furthermore, if e1, . . . , em are mutually orthogonal minimal tripotents in E, then f (e1), . . . , f (em) are mutually orthogonal minimal tripotents and f (e1 + . . . + em) = f (e1) + . . . + f (em). Consequently, f maps tripotents of rank k to tripotents of rank k, f is additive on mutually orthogonal finite rank tripotents, and the ranks of E and B coincide. (cid:3) We can also remove now part of the hypothesis in [29, Theorem 3.12]. Theorem 2.4. Let f : S(E) → S(B) be a surjective isometry between the unit spheres of two weakly compact JB∗-triples. Suppose that E = ⊕c0 j∈J Kj, where ♯J ≥ 2, and every Kj is an elementary JB∗-triple. Then there exists a surjective real linear isometry T : E → B satisfying TS(E) = f . Proof. If we replace [29, Theorem 3.6] with Theorem 2.3 above, the arguments in the proof of [29, Theorem 3.12] also hold in this setting. (cid:3) Let f : S(E) → S(B) be a surjective isometry between the unit spheres of two weakly compact JB∗-triples. Suppose that E = ⊕c0 K′k, where j∈J the Kj and the K′k are elementary JB∗-triples. If ♯J ≥ 2 or ♯K ≥ 2 we deduce from Kj and B = ⊕c0 k∈K 8 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA Theorem 2.4 above that f can be extended to a real linear (bijective) isometry T : E → F . We can therefore restrict our attention to the case of a surjective isometry between the unit spheres of two elementary JB∗-triples. Let f : S(K) → S(K′) be a surjective isometry between the unit spheres of two elementary JB∗-triples. Theorem 2.3 implies that K and K′ both have the If rank(K) = 1 or rank(K) ≥ 5, Corollary 3.15 and Theorem 3.13 same rank. in [29] prove the existence of a surjective real linear isometry T : K → K′ which coincides with f on the unit sphere of K. We can actually restrict our attention to the cases in which 2 ≤rank(K) ≤ 4. Theorem 4.1 below assures (even under the weaker hypothesis 2 ≤rank(K)) the existence of a surjective complex linear or conjugate linear isometry T : K → K′ satisfying T (x) = f (x) for every x ∈ S(K). These arguments offer a complete answer to Tingley's problem for weakly compact JB∗-triples. Theorem 2.5. Let f : S(E) → S(B) be a surjective isometry between the unit spheres of two weakly compact JB∗-triples. Then there exists a surjective real linear isometry T : E → B satisfying TS(E) = f . (cid:3) 3. Geometric properties of the subtriple generated by two minimal tripotents Our first result can be easily derived from the Triple System Analyzer [8, Propo- sition 2.1]. Lemma 3.1. [8, Proposition 2.1 (i)] Let e1 and e2 be tripotents in a JB∗-triple E. The following statements hold: (a) If e1⊤e2 then e1 is minimal if and only if e2 is; (b) If e1 and e2 are minimal and e2 ∈ E1(e1) then e1⊤e2. (cid:3) Our next result is also essentially contained in [8]. We include an explicit state- ment with a justification for completeness. Lemma 3.2. Let C be a Cartan factor of rank greater or equal than 2. Let e1 and e2 be minimal tripotent in C with e1⊤e2. Then there exists a minimal tripotent u in C such that e1 ⊥ u, and u⊤e2. Proof. By [8, Corollary 2.2] the JBW∗-triple C1(e2) has rank one or two. Suppose first that C1(e2) has rank one. Since C1(e2) ∋ e1⊤e2, we deduce from [8, PROPOSITION in page 305] that C is isometric to a Hilbert space, and hence C must have rank one, which is impossible. Therefore C1(e2) has rank two, and we can thus find a tripotent u in C1(e2) such that u ⊥ e1 and u is minimal in C1(e2). Clearly, e1 + u is a tripotent in C1(e2) which is not minimal in the latter JB∗- triple. By the Triple System Analyzer [8, Proposition 2.1 (iii)], there exist two minimal tripotents u1, u2 in C satisfying e1 + u = u1 + u2, in other words, e1 + u is a rank two tripotent in C, and hence u must be minimal in C too. (cid:3) For later purposes, we shall make use of another decomposition which is also associated with a tripotent u in a JB∗-triple E. Since the mapping Q(u)E2(u) is the involution of the JB∗-algebra E2(u), it follows that E2(u) = E1(u) ⊕ E−1(u), where Ek(u) := {x ∈ E : Q(u)(x) := {u, x, u} = kx}. Clearly, E−1(u) = iE1(u). LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 9 By Peirce rules, E0(u) ⊕ E1(u) coincides with the kernel ker(Q(u)) = E0(u). The identity holds whenever ijk 6= 0. The spaces Ek(u) induce the following decomposition {Ei(u), Ej(u), Ek(u)} ⊆ Eijk(u), E = E−1(u) ⊕ E0(u) ⊕ E1(u). Let u and v be a couple of arbitrary minimal tripotents in a JB∗-triple E. The JB∗-subtriple J of E generated by u and v was totally described by Y. Friedman and B. Russo [19, Lemma 2.3 and Proposition 5]. It follows from the just quoted results that J is linearly spanned by u, v, P1(u)(v), and P1(v)(u). It is further known that J is isometrically (triple-)isomorphic to one in the following list: C, C ⊕∞ C, M1,2(C), M2(C) and S2(C), (9) where Mk,n(C) is the C∗-algebra of all k × n matrices with complex entries, and S2(C) denotes the symmetric 2 × 2 complex matrices. Given a minimal tripotent u in a JB∗-triple E, since E2(u) = Cu, there exists a unique norm-one functional φu : E → Cu satisfying P2(u)(x) = φu(x)u, for each x in E. Clearly, P 1(u)(x) = ℜeφu(x)u for all x in E. By an slight abuse of notation, when no confusion arises, we shall also write P 1(u)(x) for the real number φu(x). In our next result we shall establish a formula to compute the distance between two minimal tripotents in a JB∗-triple. Proposition 3.3. Let u and v be minimal tripotents in a JB∗-triple E. Then the following formula holds (10) ku − vk2 = (1 − ℜeφu(v)) +p(1 − ℜeφu(v))2 − kP0(u)(v)k2 = (1 − P 1(u)(v)) +p(1 − P 1(u)(v))2 − kP0(u)(v)k2. Proof. Let J denote the JB∗-subtriple generated by u and v. Suppose J = C⊕∞ C. The minimality of u and v implies that u ⊥ v and hence the conclusion in (10) is a consequence of [19, Lemma 1.3 (a)]. If J = C the statement is clear. We assume now that J = M1,2(C). There is no loss of generality in assuming that u = (1, 0) and v = (λ1, λ2) with λ12 + λ22 = 1. Therefore, ku − vk2 = k(1 − λ1,−λ2)k2 = 1 − λ12 + λ22 = 2 − 2ℜe(λ1) = 2(1 − ℜeφu(v)), which proves (10) because P0(u)(v) = 0. ξ2η1 We deal now with the case J = M2(C) with the spectral norm. We may assume ξ2η2 (cid:19) = (η1, η2) ⊗ (ξ1, ξ2), with ξj, ηk ∈ C that u =(cid:18) 1 0 0 0 (cid:19) and v =(cid:18) ξ1η1 satisfying ξ12 + ξ22 = 1 and η12 + η22 = 1. To simplify the notation we write v = (cid:18) α β δ (cid:19), with αδ = βγ and α2 + δ2 + β2 + γ2 = 1. According to this terminology, v − u =(cid:18) α − 1 β ξ1η2 γ γ kv − uk2 = k(v − u)(v − u)∗k =(cid:13)(cid:13)(cid:13)(cid:13) δ (cid:19) . By the Gelfand-Naimark axiom we have δ (cid:19)∗(cid:13)(cid:13)(cid:13)(cid:13) . δ (cid:19)(cid:18) α − 1 β (cid:18) α − 1 β γ γ 10 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA (α − 1)γ + βδ Since (v−u)(v−u)∗ =(cid:18) α − 12 + β2 (α − 1)γ + βδ γ2 + δ2 (cid:19) , its characteristic poly- nomial is precisely p(λ) = λ2−(α−12+β2+γ2+δ2)λ+(α−12+β2)(γ2+δ2)−(α−1)γ +βδ2. The conditions α2 + δ2 + β2 + γ2 = 1 and αδ = βγ imply (α − 12 + β2 + γ2 + δ2) = 2 − 2ℜe(α), and (α − 12 + β2)(γ2 + δ2) − (α − 1)γ + βδ2 = α−12γ2+α−12δ2+β2γ2+β2δ2−α−12γ2−β2δ2−2ℜe((α−1)γβδ) = α − 12δ2 + β2γ2 − 2ℜe((α − 1)γβδ) = (α − 1)δ − βγ2 = δ2. Therefore, and hence p(λ) = λ2 − 2(1 − ℜe(α))λ + δ2, ku − vk2 = (1 − ℜe(α)) +p(1 − ℜe(α))2 − δ2 = (1 − ℜeφu(v)) +p(1 − ℜeφu(v))2 − kP0(u)(v)k2. The case J = S2(C) follows by the same arguments. (cid:3) Corollary 3.4. Let u and v be minimal tripotents in a JB∗-triple E. Then the condition ku ± vk = √2 implies v ∈ E−1(u) ⊕ E1(u). Proof. Applying Proposition 3.3 we know that 2 = ku ∓ vk2 = (1 ∓ ℜeφu(v)) +p(1 ∓ ℜeφu(v))2 − kP0(u)(v)k2, and hence (1 ± ℜeφu(v)) =p(1 ∓ ℜeφu(v))2 − kP0(u)(v)k2, (1 ± ℜeφu(v))2 = (1 ∓ ℜeφu(v))2 − kP0(u)(v)k2, ±4ℜeφu(v) = −kP0(u)(v)k2, which proves ℜeφu(v) = P0(u)(v) = 0, and gives the desired conclusion. (cid:3) We can now prove that surjective isometries between the unit spheres of elemen- tary JB∗-triples of rank ≥ 2 preserve collinearity between minimal tripotents. Proposition 3.5. Let f : S(C) → S(C′) be a surjective isometry between elemen- tary JB∗-triples with rank(C) ≥ 2. Suppose e1 and e2 are minimal tripotents in C with e1⊤e2. Then f (e1)⊤f (e2). Proof. Since rank(C) ≥ 2 by Lemma 3.2 there exists a minimal tripotent u in C satisfying e1 ⊥ u and e2⊤u. By Proposition 2.1 f (e1), f (±e2), and f (u) are minimal tripotents with f (e1) ⊥ f (u). By hypothesis and Theorem 2.3 we have 2 = ke1 ± e2k2 = ku ± e2k2 = kf (e1) ± f (e2)k2 = kf (u) ± f (e2)k2. Thus, Corollary 3.4 shows that f (e2) = P −1(f (u))(f (e2)) + P1(f (u))(f (e2)) and f (e2) = P −1(f (e1))(f (e2)) + P1(f (e1))(f (e2)). Having in mind that f (e1) ⊥ f (u), and hence (C′)−1(f (e1)) ⊆ (C′)0(f (u)), we obtain f (e2) = P1(f (u))(f (e2)). Lemma 3.1(b) gives f (e1)⊤f (e2). (cid:3) The next result determines the behavior of a surjective isometry on the space C−1(e) = iRe associated with a minimal tripotent e. LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 11 Lemma 3.6. Let f : S(C) → S(C′) be a surjective isometry between elementary JB∗-triples with rank(C) ≥ 2. Suppose e is a minimal tripotent in C. Then f (ie) = ±if (e). Consequently, f (λe) = λf (e) or f (λe) = λf (e) for every λ ∈ C. Proof. The elements f (e) and f (ie) are minimal tripotents in C′ (see Proposition 2.1). Since, by Theorem 2.3, we have kf (e) ± f (ie)k = kf (e) − f (∓ie)k = ke ± iek = √2, we deduce from Corollary 3.4 that f (ie) ∈ C−1(f (e)) ⊕ C1(f (e)) = iRf (e) ⊕ C1(f (e)). If we consider the JB∗-subtriple J generated by f (e) and f (ie), we know that J is isomorphic to one of C, C ⊕∞ C, M1,2(C), M2(C) and S2(C) (see [19, Proposition 5]). Arguing case by case, we can find a minimal tripotent v12 ∈ C′, a real t and a complex λ12 satisfying t2+λ122 = 1, f (e)⊤v12, and f (ie) = ite+λ12v12. Applying Lemma 3.2 we find a minimal tripotent v22 in C′ such that f (e) ⊥ v22 and v12⊤v22. Theorem 2.3 applied to f−1 implies that e ⊥ f−1(v22) and f−1(v22) is a minimal tripotent in C. Therefore ie ⊥ f−1(v22). A new application of Theorem 2.3, shows that it + λ12v12 = f (ie) ⊥ v22, which implies λ12 = 0, and consequently, f (ie) = ±if (e). To prove the last statement, let us take u in C with e ⊥ u and let Tu : C0(u) → C′0(f (e)) be the surjective real linear isometry given by Proposition 2.1. For each complex number λ, we have f (λe) = Tu(λu) = ℜe(λ)Tu(e) + ℑm(λ)Tu(ie) = ℜe(λ)f (e) + ℑm(λ)f (ie) = ℜe(λ)f (e) ± iℑm(λ)f (e), that is f (λe) = λf (e) or f (λe) = λf (e), for every λ ∈ C. (cid:3) Following the notation in [8] (see also [28]), we recall that an ordered quadruple (u1, u2, u3, u4) of tripotents in a JB∗-triple E is called a quadrangle if u1⊥u3, u2⊥u4, u1⊤u2 ⊤u3⊤u4 ⊤u1 and u4 = 2 {u1, u2, u3} (the axiom (JB∗1) implies that the last equality holds if the indices are permutated cyclically, e.g. u2 = 2{u3, u4, u1}). Let u and v be tripotents in E. We say that u governs v, u ⊢ v, whenever v ∈ U2(u) and u ∈ U1(v). An ordered triplet (v, u, v) of tripotents in E, is called a trangle if v⊥v, u ⊢ v, u ⊢ v and v = Q(u)v. Proposition 3.7. Let f : S(C) → S(C′) be a surjective isometry between elemen- tary JB∗-triples with rank(C) ≥ 2, and let e be a minimal tripotent in C. The following statements hold: (a) If (u1, u2, u3, u4) is a quadrangle of minimal tripotents in C, then the quadruple (f (u1), f (u2), f (u3), f (u4)) is a quadrangle of minimal tripotents in C′; (b) Suppose (u1, u2, u3, u4) is a quadrangle of minimal tripotents in C and f (iu1) = if (u1) (respectively, f (iu1) = −if (u1)), then f (iuj) = if (uj) (respectively, f (iuj) = −if (uj)), for every j = 2, 3, 4; (c) Suppose f (ie) = if (e) (respectively, f (ie) = −if (e)), then f (iv) = if (v) (re- spectively, f (iv) = −if (v)) for every minimal tripotent v ∈ C with v⊤e; (d) If (v, u, v) is a trangle in C with v, v minimal, then (f (v), f (u), f (v)) is a trangle in C′, with f (v) and f (v) minimal; (e) Suppose (v, u, v) is a trangle in C with v, v minimal. If f (iv) = if (v) (respec- tively, f (iv) = −if (v)), then f (iv) = if (v) and f (iu) = if (u) (respectively, f (iv) = −if (v) and f (iu) = −if (u)); 12 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA (f ) Suppose (v, u, v) is a trangle in C with v, v minimal. If f (iu) = if (u) (respec- tively, f (iu) = −if (u)), then f (iv) = if (v) and f (iv) = if (v) (respectively, f (iv) = −if (v) and f (iv) = −if (v)). Proof. (a) We know from Theorem 2.3 and Proposition 3.5 that f (u1), f (u2), f (u3), and f (u4) are minimal tripotents in C′ with f (u1)⊤f (u2)⊤f (u3)⊤f (u4)⊤f (u1), f (u1) ⊥ f (u3) and f (u2) ⊥ f (u4). We only have to show that 2{f (u1), f (u2), f (u3)} = f (u4) to conclude the proof. To this end we observe that w = 1 w = 1 implies that 2 (u1 + u2 + u3 + u4) and 2 (u1 − u2 + u3 − u4) are minimal tripotents in C with w ⊥ w. Theorem 2.3 f (w) + f ( w) = f (w + w) = f (u1 + u3) = f (u1) + f (u3), (11) which implies that f (w) and f ( w) are orthogonal projections in the JBW∗-algebra C′2(f (u1)+f (u3)). To see this, we simply observe that {f (w), f (u1)+f (u3), f (w)} = {f (w), f (w), f (w)} = f (w) = {f (u1) + f (u3), f (w), f (u1) + f (u3)}, and the same for f ( w). In particular, (12) is a symmetry in the JBW∗-algebra C′2(f (u1) + f (u3)) and hence f (w) − f ( w) = f (u2 + u4) = f (u2) + f (u4) f (u2) + f (u4) = {f (u1) + f (u3), f (u2) + f (u4), f (u1) + f (u3)}, which via Peirce arithmetic, shows that f (u2) + f (u4) = 2{f (u1), f (u2), f (u3)} + 2{f (u1), f (u4), f (u3)}, and hence 2{f (u1), f (u2), f (u3)} = f (u4). (b) Let us assume that f (iu1) = if (u1). By (a), (f (u1), f (u2), f (u3), f (u4)) is a quadrangle of minimal tripotents in C′, and for the minimal tripotents w = 2 (u1 + u2 + u3 + u4) and w = 1 1 2 (u1 − u2 + u3 − u4), by (11) and (12), we have f (w) = 1 2 (f (u1)−f (u2)+f (u3)−f (u4)). By Lemma 3.6 it follows that f (i(w + w)) ∈ {±if (w ± w)}. Therefore, applying Theorem 2.3, we obtain 2 (f (u1)+f (u2)+f (u3)+f (u4)) and f ( w) = 1 if (u1) + f (iu3) = f (iu1) + f (iu3) = f (iu1 + iu3) = f (i(w + w)) ∈ {±if (w ± w)} = {±if (u1 + u3),±if (u2 + u4)} = {±i(f (u1) + f (u3)),±i(f (u2) + f (u4))}, and since f (iu3) ∈ {±if (u3)}, the unique possible choice for f (iu3) is if (u3). By or- thogonality f (iw)+f (i w) = f (iw +i w) = f (iu1 +iu3) = f (iu1)+f (iu3) = if (u1)+ if (u3) = i(f (w) + f ( w)), and since f (iw) ∈ {±if (w)} and f (i w) ∈ {±if ( w)}, we obtain f (iw) = if (w) and f (i w) = if ( w). Consequently, f (iu2 + iu4) = f (iw − i w) = f (iw) − f (i w) = if (w) − if ( w) = if (u2) + if (u4), which implies f (iu2) = if (u2) and f (iu4) = if (u4). (c) Let v be a minimal tripotent in C with e⊤v. By Lemma 3.2 there exists a minimal tripotent u ∈ C such that e ⊥ u and v⊤u. By [8, Proposition 1.7], the element v = 2{e, v, u} is a minimal tripotent in C and (e, v, u, v) is a quadrangle in C. Since f (ie) = if (e), it follows from (b) that f (iv) = if (v). (d) Let (v, u, v) be a trangle in C with v, v minimal. It is known that w = 2 (v + v−u) are minimal tripotents in C with w ⊥ w (compare, 1 2 (v + v +u) and w = 1 LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 13 for example, [23, Theorem 4.10] or [18, Corollary 2.8 and Remark 2.6]). Therefore u = w − w is a rank 2 tripotent in C. It follows from Theorem 2.3 that f (u), f (v), f (w), f ( w), and f (v) are tripotents in C′, where f (v), f (w), f ( w), and f (v) are minimal, f (u) has rank 2, f (v) ⊥ f (v), f (w) ⊥ f ( w), and f (u) = f (w) − f ( w). (1 + P 1(f (v))(f (w))) +p(1 + P 1(f (v))(f (w)))2 − kP0(f (v))(f (w))k2 By the hypothesis on f , Theorem 2.3 and Proposition 3.3, we have = kf (v) + f (w)k2 = kv + wk2 = + √2 3 2 and (1 − P 1(f (v))(f (w))) +p(1 − P 1(f (v))(f (w)))2 − kP0(f (v))(f (w))k2 = kf (v) − f (w)k2 = kv − wk2 = 1 2 , where we have identified P 1(f (v))(f (w))) with the real number ℜe(φf (v)(f (w))). It is not hard to see that the unique solution of the above system gives P 1(f (v))(f (w))) = By replacing v with v we get P 1(f (v))(f (w))) = 1 2 1 2 f (v), and kP0(f (v))(f (w))k = f (v), and kP0(f (v))(f (w))k = 1 2 . 1 2 . We also know that v + v = w + w and hence f (v) + f (v) = f (w) + f ( w), and hence f (w), f ( w) ∈ C′2(f (v) + f (v)). Therefore, P0(f (v))(f (w)) = P2(f (v))(f (w)). If we observe that (C′)1(f (v)) ⊆ C′2(f (v)) = Cf (v), we obtain P2(f (v))(f (w)) = P 1(f (v))(f (w))) = 1 2 f (v). It is also clear from the above that P1(f (v))(f (w)) = P1(f (v))(f (w)). We have therefore shown that 2 f (v) + P1(f (v))(f (w)) + 1 2 f (v), and thus f (w) = 1 f (w) = f (v) + f (v) + P1(f (v))(f (w)). 1 2 1 2 Similar arguments applied to f ( w) prove f ( w) = 1 2 f (v) + 1 2 f (v) + P1(f (v))(f ( w)). The equality f (v) + f (v) = f (w) + f ( w) implies that P1(f (v))(f ( w)) = −P1(f (v))(f (w)), and hence f (u) = f (w) − f ( w) = 2P1(f (v))(f (w)) = 2P1(f (v))(f (w)) ∈ C′1(f (v)) ∩ C′1(f (v)). The identity {f (u), f (v) + f (v), f (u)} = {f (w) − f ( w), f (w) + f ( w), f (w) − f ( w)} = {f (w), f (w), f (w)} + {f ( w), f ( w), f ( w)} = f (w) + f ( w) = f (v) + f (v), proves that f (u) actually is a symmetry in the JB∗-algebra C′2(f (v) + f (v)), and then C′2(f (v) + f (v)) = C′2(f (u)), which gives f (v), f (v) ∈ C′2(f (u)). Finally, since f (u) ∈ C′1(f (v)) ∩ C′1(f (v)), we deduce from Peirce rules that {f (u), f (v), f (u)} ∈ C′0(f (v)) ∩ C′2(f (v) + f (v)) = Cf (v), {f (u), f (v), f (u)} ∈ C′0(f (v)) ∩ C′2(f (v) + f (v)) = Cf (v), 14 and F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA {f (u), f (v) + f (v), f (u)} = f (v) + f (v), and hence {f (u), f (v), f (u)} = f (v) and {f (u), f (v), f (u)} = f (v), which finishes the proof of (d). (e) Let (v, u, v) be a trangle in C with v, v minimal, and f (iv) = if (v). By (d) the 3-tuple (f (iv), f (iu), f (iv)) is a trangle in C′ with f (iv), f (iv) minimal. Following the arguments in the proof of (d), iw = 1 2 i(v + v− u), f (iw) and f (i w) are minimal tripotents with iw ⊥ i w, f (iw) ⊥ f (i w), f (u) = f (w) − f ( w), f (v) + f (v) = f (w) + f ( w), 2 i(v + v + u), i w = 1 and f (iu) = f (iw) − f (i w) ∈ C′1(f (iv)) ∩ C′1(f (iv)), f (iv) + f (iv) = f (iw) + f (i w). Since f (iv) = if (v) and f (iz) ∈ {±if (z)}, for every z = w, w, v, we deduce that f (iv) = if (v) and f (iu) = if (u). (f ) With the notation employed in the proofs of (d) and (e), if f (iu) = if (u), we have i(f (w) − f ( w)) = if (w − w) = if (u) = f (iu) = f (iw − i w) = f (iw) − f (i w), and hence, by orthogonality relations, f (iw) = if (w), f (i w) = if ( w). We similarly get f (iv) = if (v) and f (iv) = if (v). (cid:3) We have developed enough tools to establish that surjective isometries between the unit spheres of elementary JB∗-triples of rank greater or equal than 2 are ℓ2- additive on collinear minimal tripotents. Proposition 3.8. Let f : S(C) → S(C′) be a surjective isometry between elemen- tary JB∗-triples with rank(C) ≥ 2. Suppose e1 and e2 are minimal tripotent in C with e1⊤e2. Then, the following statements hold: (a) If f (ie1) = if (e1), then f (αe1 + βe2) = αf (e1) + βf (e2), for all α, β ∈ C with α2 + β2 = 1; (b) If f (ie1) = −if (e1), then f (αe1 + βe2) = αf (e1) + βf (e2), for all α, β ∈ C with α2 + β2 = 1. Proof. Let us fix α, β ∈ C with α2+β2 = 1. We can assume α, β 6= 0. Proposition 2.1 assures that f (e1), f (e2), and f (αe1 + βe2) are minimal tripotents in C′. Let J denote the JB∗-subtriple of C′ generated by f (e1) and f (αe1 + βe2). We have already commented that J identifies with one of the following C, C⊕∞ C, M1,2(C), M2(C) and S2(C) (see [19, Proposition 5]). If J = C, we have f (αe1 + βe2) = λf (e1) for a suitable complex λ with λ = 1. By Lemma 3.6 we have λf (e1) = f (λe1) or λf (e1) = f (λe1). Therefore, αe1+βe2 = λe1 or αe1 + βe2 = λe1, and both equalities are impossible. If J = C ⊕∞ C, we can assume f (e1) = (1, 0) and f (αe1 + βe2) = (λ, 0) or f (αe1+βe2) = (0, λ) with λ = 1. In the first case, Lemma 3.6 gives αe1+βe2 = λe1 or αe1 + βe2 = λe1, which is impossible, while in the second case, by Theorem 2.4, we have αe1 + βe2 ⊥ e1, which is impossible too. LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 15 In the remaining cases, we can assume J ⊆ M2(C), f (e1) = (cid:18) 1 0 f (αe1 + βe2) = (cid:18) α′ β′ 0 0 (cid:19) and δ′ (cid:19) with α′2 + β′2 + γ′2 + δ′2 = 1 and α′δ′ = β′γ′. γ′ Applying Proposition 3.3 and the properties of f we get: 2(1 ± ℜe(α)) = ke1 ± (αe1 + βe2)k2 = kf (e1) ± f (αe1 + βe2)k2 = (1 ± ℜe(α′)) +p(1 ± ℜe(α′))2 − δ′2; ±4(1 ± ℜe(α))(ℜe(α′) − ℜe(α)) = −δ′2. 4(1 ± ℜe(α))2 + (1 ± ℜe(α′))2 − 4(1 ± ℜe(α))(1 ± ℜe(α′)) = (1 ± ℜe(α′))2 − δ′2; Therefore, δ′ = 0, and since α 6= 1 we also deduce that ℜe(α′) = ℜe(α). Therefore, P0(f (e1))(f (αe1 + βe2)) = 0 and P 1(f (e1))(f (αe1 + βe2)) = ℜe(α)f (e1). By Lemma 3.6 we have f (ie1) = if (e1) or f (ie1) = −if (e1). We shall distinguish these two cases. Case a) f (ie1) = if (e1). By Proposition 3.3, Theorem 2.3 and the hypothesis we get 2(1±ℜe(iα)) = ke1±i(αe1+βe2)k2 = kie1∓(αe1+βe2)k2 = kf (ie1)∓f (αe1+βe2)k2 = 2(1 ± ℜe(iα′)), iβ′ 0 2 0 0 (cid:19)(cid:13)(cid:13)(cid:13)(cid:13) 0 (cid:19) ∓(cid:18) α′ β′ γ′ =(cid:13)(cid:13)(cid:13)(cid:13) (cid:18) 1 0 0 0 (cid:19) ±(cid:18) iα′ iγ′ which shows that α = α′, and hence =(cid:13)(cid:13)(cid:13)(cid:13) (cid:18) i 2 0 (cid:19)(cid:13)(cid:13)(cid:13)(cid:13) (13) f (αe1 + βe2) = αf (e1) + P1(f (e1))(f (αe1 + βe2)). Since f (ie1) = if (e1) and e1⊤e2, Proposition 3.7(c) implies f (ie2) = if (e2). Thus, repeating the above arguments with e2 in the role of e1 we get (14) f (αe1 + βe2) = βf (e2) + P1(f (e2))(f (αe1 + βe2)). By combining (13) and (14) we get f (αe1 + βe2) = αf (e1) + βf (e2) + P1(f (e2))P1(f (e1))(f (αe1 + βe2)). Since f (e1) and f (e2) are collinear minimal tripotents (compare Proposition 3.5), αf (e1) + βf (e2) is a minimal tripotent in C′ (compare [8, LEMMA in page 306]). The element f (αe1 + βe2) lies in the unit sphere of C′ and, by Peirce arithmetic P1(f (e2))P1(f (e1))(f (αe1 + βe2)) ∈ C′1(αf (e1) + βf (e2)), and thus P2(αf (e1) + βf (e2))(f (αe1 + βe2)) = αf (e1) + βf (e2). Lemma 1.6 in [19] proves f (αe1 + βe2) = αf (e1) + βf (e2). In the case b) f (ie1) = −if (e1), the above arguments prove f (αe1 + βe2) = (cid:3) αf (e1) + βf (e2). Let C be a Cartan factor with rank greater or equal than 2. Let e1 be a minimal tripotent in C. The Peirce subspace C1(e1) cannot be zero, otherwise C = C2(e1) ⊕⊥ C0(e1) would be the direct sum of two orthogonal weak∗-closed triple ideals, which is impossible. Applying [8, Corollary 2.2 and Proposition 2.1] one of the following statements holds: (i) There exists a minimal tripotent v in C satisfying e1⊤v; (ii) There exist a rank 2 tripotent u and a minimal tripotent e1 in C such that (e1, u, e1) is a trangle; 16 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA (iii) There exist minimal tripotents v2, v3, v4 in C such that (e1, v2, v3, v4) is a quadrangle. In case (i), we can repeat the argument in the proof of Proposition 3.7(c) to deduce, via Lemma 3.2, the existence of minimal tripotents v2, v3, v4 in C such that (e1, v2, v3, v4) is a quadrangle. Therefore, for each minimal tripotent e1 in C one of the following holds: (X.1) There exist a rank 2 tripotent u and a minimal tripotent e1 in C such that (e1, u, e1) is a trangle; (X.2) There exist minimal tripotents v2, v3, v4 in C such that (e1, v2, v3, v4) is a quadrangle. Let e2 be a minimal tripotent with e1 ⊥ e2. In each one of the previous cases, by [25, Proposition 5.8], there exists a complex linear, isometric, JB∗-triple isomor- phism T : C → C such that (b.1) T (e1) = e1 and T (e1) = e2; (b.2) T (e1) = e1 and T (v3) = e2. Since T preserves quadrangles and trangles of the previous form, we can always conclude that one of the following statements is true: (c.1) There exist a rank 2 tripotent u in C such that (e1, u, e2) is a trangle; (c.2) There exist minimal tripotents v2, v4 in C such that (e1, v2, e2, v4) is a quad- rangle. The following corollary is therefore a consequence of the previous arguments (c.1) and (c.2) and Proposition 3.7(b) and (e). Corollary 3.9. Let f : S(C) → S(C′) be a surjective isometry between elementary JB∗-triples with rank(C) ≥ 2, and let e1 and e2 be minimal tripotents in C with e1 ⊥ e2. Suppose f (ie1) = if (e1) (respectively, f (ie1) = −if (e1)), then f (ie2) = if (e2) (respectively, f (ie2) = −if (e2)). (cid:3) Our next lemma will also follow from the comments prior to Corollary 3.9 and [19, Proposition 5]. Lemma 3.10. Let e and v be two minimal tripotents in a Cartan factor of rank greater or equal than two. Then one of the following statements holds: (a) There exist minimal tripotents v2, v3, v4 in C, and complex numbers α, β, γ, δ such that (e, v2, v3, v4) is a quadrangle, α2 + β2 + γ2 + δ2 = 1, αδ = βγ, and v = αe + βv2 + γv4 + δv3; (b) There exist a minimal tripotent e ∈ C, a rank two tripotent u ∈ C, and complex numbers α, β, δ such that (e, u, e) is a trangle, α2 + 2β2 + δ2 = 1, αδ = β2, and v = αe + βu + δe. Proof. Let J denote the JB∗-subtriple of C generated by e and v. We have re- peatedly applied that J identifies isomorphically with one of the following list: C, C ⊕∞ C, M1,2(C), M2(C) and S2(C) (see [19, Proposition 5]). Suppose J = C. Clearly v = λe for a suitable complex number λ with λ = 1. By (X.1) and (X.2), or there exist a rank 2 tripotent u and a minimal tripotent e in C such that (e, u, e) is a trangle, or there exist minimal tripotents v2, v3, v4 in C such that (e, v2, v3, v4) is a quadrangle. So, the desired conclusion holds with α = λ, β = γ = δ = 0 (where γ = β in the case of a trangle). LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 17 In the case J = C ⊕∞ C, we can assume that e = (1, 0) and v = (λ, 0) or v = (0, λ) with λ = 1. The case v = (λ, 0) was treated in the previous paragraph. For the second choice, we observe that v ⊥ e, and hence the statement follows from (c.1) and (c.2) with δ = 1, e1 = e, e2 = v and α = β = γ = 0 (where γ = β in the case of a trangle). We deal now with the remaining cases. There is no loss of generality in assuming If J = M1,2(C). We can obviously find a minimal tripotent v2 ∈ C such that e⊤v2 and complex numbers α, β satisfying v = αe + βv2 and α2 + β2 = 1. Let us take, via Lemma 3.2, a minimal tripotent v3 in C such that v3 ⊥ e and v2⊤v3. Setting v4 = 2{e, v2, v3} we define a quadrangle (e, v2, v3, v4) (see [8, Proposition 1.7]). The statement (b) holds with α, β, γ = δ = 0. J ⊆ M2(C), e =(cid:18) 1 δ (cid:19) with α2 +β2 +γ2 +δ2 = 1 and αδ = βγ (with β = γ in case J = S2(C)). We conclude by taking v2 =(cid:18) 0 1 0 0 (cid:19) , v3 = (cid:18) 0 0 (cid:19) and e = (cid:18) 0 0 0 1 (cid:19) , 0 (cid:19) and v =(cid:18) α β 1 (cid:19) , and v4 = (cid:18) 0 0 1 0 (cid:19) or u = (cid:18) 0 γ 0 0 1 1 0 0 respectively. (cid:3) We recall that a spin factor is a complex Hilbert space X, with inner product (..), provided with a conjugation (i.e. a conjugate linear isometry of period 2 for the Hilbertian norm given by kxk2 2 = (xx) (x ∈ X)) x 7→ x, where triple product and norm are given by (15) {x, y, z} = (xy)z + (zy)x − (xz)y, and kxk2 = (xx) +p(xx)2 − (xx)2, respectively. Let X1 = {x ∈ X : x = x} and X2 = {x ∈ X : x = −x}. It is not hard to see that X1 and X2 are real subspaces of X, X2 = iX1, and X = X1 ⊕ X2. Since . is a conjugation we can easily see that (xy) = (yx) for all x, y ∈ X. Therefore, if x1, y1 ∈ X1 and x2, y2 ∈ X2 we have (x1x2) = −(x2x1) = −(x1x2), (x1y1) = (y1x1) = (x1y1), and (x2y2) = (y2x2) = (x2y2). Therefore, (XjXj) ⊆ R and (x1x2) ∈ iR. The underlying real Banach space XR is a real Hilbert space with respect to the inner product hxyi := ℜe(xy). Clearly, the real subspaces X1 and X2 are orthogonal with respect to the inner product h..i, that is, hX1X2i = 0, and hxjyji = (xjyj), for every j = 1, 2. For x1 ∈ X1 and x2 ∈ X2, we have x1 + x2 = x1 − x2 and if (x1x2) = 0 we also have (16) kx1 + x2k = kx1k + kx2k. It is known that every spin factor X has rank two. We further known the precise form of minimal and rank two tripotents in X, more concretely, and min U(X) =n 1 2 (x1 + x2) : xi ∈ S(Xi), (x1x2) = 0o max U(X) =nλx1 : x1 ∈ S(X1), λ ∈ S(C)o. 18 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA Every maximal or complete tripotent in X is unitary. Given a minimal tripotent e = 1 2 (x1 + x2) ∈ min U(X), its Peirce-0 subspace (17) is one-dimensional. X0(e) = Ce = {x ∈ X : x ⊥ e} Let v = 1 2 (x1 + ix2) (xi ∈ S(X1), (x1x2) = 0) be a minimal tripotent in X. It is easy to check that X2(v) = Cv, X0(v) = Cv, and X1(v) = {x ∈ X : (xx1) = (xx2) = 0} = {x1, x2}⊥X . We further know that (18) and P2(v)(x) = 2(xv)v = ((xx1) − i(xx2))v, P0(v)(x) = 2(xv)v = ((xx1) + i(xx2))v, P1(v)(x) = x − 2(xv)v − 2(xv)v = x − (xx1)x1 − (xx2)x2 (x ∈ X). The projection P1(v) also coincides with the orthogonal projection of X onto {x1, x2}⊥X in the Hilbert space (X, (..)). Lemma 3.11. Let (v, u, v) be a trangle of tripotents in a Cartan factor C, where v and v are minimal. Let w = 1 2 (v − u + v), bu = v − v. Suppose α, β, δ are complex numbers with α2 + 2β2 + δ2 = 1, and αδ = β2. Let x be an element in C such that kxk ≤ 1, 2 (v + u + v), w = 1 P2(v)(x) = αv, P2(v)(x) = δv, P2(w)(x) = α + 2β + δ w, 2 and P2( w)(x) = α − 2β + δ 2 w. Then, for the minimal tripotent e = αv + βu + δv, we have x = e + P0(e)(x). Proof. By [23, Theorem 4·10] (see also [18, Lemma 2.7]), C2(v + v) is (isometrically isomorphic to) a spin factor. Let X denote this spin factor C2(v + v) = C2(v) ⊕ C2(v) ⊕ C1(v) ∩ C1(v). Let us observe that (w,bu, w) is a trangle in C with w, w minimal. Let · and (..) denote the involution and the inner product of X, respectively. Since v, v, u ∈ C2(v + v) = X we can assume that v = 1 2 (x1 − ix2) = v, and u = ix3, where xi ∈ S(X1), (x1x2) = 0, (x1x3) = 0, and (x2x3) = 0. We shall keep the notation given before this lemma. 2 (x1 + ix2), v = 1 Let y = P2(v + v)(x). Clearly kyk ≤ 1. By hypothesis P2(v)(y) = αv, P2(v)(y) = δv, P2(w)(y) = α + 2β + δ w, 2 and P2( w)(y) = α − 2β + δ 2 w. Applying the identities in (18), we deduce from the last four equalities that (19) (yx1) = α + δ 2 , (yx2) = i α − δ 2 , and (yx3) = iβ. LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 19 Let H be the (complex) subspace of X generated by x1, x2 and x3. And let P : X → H be the orthogonal projection of the Hilbert space (X, (..) onto H. Since H = H, it follows from [21, Remark 7] that max{kP (z)k,k(I − P )(z)k} ≤ kzk, for every z ∈ X. Moreover, kzk = kP (z)k if and only if z = P (z). We have shown in (19) that P (y) = α + δ 2 x1 + i α − δ 2 x2 + iβx3 = αv + βu + δv. Since, by the hypothesis on α, β, δ we have 1 ≥ kyk ≥ kP (y)k = kαv + βu + δvk = 1, we conclude that P (y) = y, and hence P2(v + v)(x) = y = αv + βu + δv. Finally, the element e = αv + βu + δv is a minimal tripotent in the spin factor X = C2(v + v) with P2(e)(x) = P2(e)(y) = e. The conditions 1 ≥ kxk, P2(e)(x) = e imply, via [19, Lemma 1.6], that P1(e)(x) = 0, and hence x = e + P0(e)(x). (cid:3) Our next theorem contains a key technical theorem needed for the main results of this note. Theorem 3.12. Let f : S(C) → S(C′) be a surjective isometry between elementary JB∗-triples with rank(C) ≥ 2. The following statements hold: (a) If (v1, v2, v3, v4) is a quadrangle of minimal tripotents in C and f (iv1) = if (v1) (respectively, f (iv1) = −if (v1)), then f (αv1 + βv2 + γv4 + δv3) = αf (v1) + βf (v2) + γf (v4) + δf (v3) (respectively, f (αv1 + βv2 + γv4 + δv3) = αf (v1) + βf (v2) + γf (v4) + δf (v3)), (b) If (v, u, v) is a trangle, with v, v ∈ C minimal tripotents, u ∈ C a rank two for all α, β, γ, δ ∈ C with α2 + β2 + γ2 + δ2 = 1, αδ = βγ; tripotent, and f (iv) = if (v) (respectively, f (iv) = −if (v)), then f (αv + βu + δv) = αf (v) + βf (u) + δf (v) (respectively, f (αv + βu + δv) = αf (v) + βf (u) + δf (v)), for all α, β, δ ∈ C with α2 + 2β2 + δ2 = 1, αδ = β2. Proof. (a) We assume f (iv1) = if (v1) (the case f (iv1) = −if (v1) follows similarly). Let e = αv1 + βv2 + γv4 + δv3. By Theorem 2.3, f (e) is a minimal tripotent and f (−e) = −f (e). By Lemma 3.10 one of the following statements holds: (1) There exist minimal tripotents w2, w3, w4 in C′, and complex numbers α′, β′, γ′, δ′ such that (f (v1), w2, w3, w4) is a quadrangle, α′2 +β′2 +γ′2 +δ′2 = 1, α′δ′ = β′γ′, and f (e) = α′f (v1) + β′w2 + γ′w4 + δ′w3; (2) There exist a minimal tripotent v ∈ C′, a rank two tripotent u ∈ C′, and complex numbers α′, β′, δ′ such that (f (v1), u, v) is a trangle, α′2 + 2β′2 + δ′2 = 1, α′δ′ = (β′)2, and f (e) = α′f (v1) + β′u + δ′v. 20 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA We shall first deal with case (1). By hypothesis kv1 ± ek = kf (v1) ± f (e)k. Applying Proposition 3.3 we obtain: (20) and that is, (21) and (22) (1 − ℜeα) +p(1 − ℜeα)2 − δ2 = kv1 − ek2 (1 + ℜeα) +p(1 + ℜeα)2 − δ2 = kv1 + ek2 = kf (v1) − f (e)k2 = (1 − ℜeα′) +p(1 − ℜeα′)2 − δ′2 = kf (v1) + f (e)k2 = (1 + ℜeα′) +p(1 + ℜeα′)2 − δ′2, (ℜeα′ − ℜeα) +p(1 − ℜeα)2 − δ2 =p(1 − ℜeα′)2 − δ′2, (ℜeα − ℜeα′) +p(1 + ℜeα)2 − δ2 =p(1 + ℜeα′)2 − δ′2. It is not hard to check that the unique solution to the system formed by (21) and (22) is δ′ = δ = 0, or ℜeα = ℜeα′ and δ′ = δ. In the case δ′ = δ = 0, it follows from (20) that ℜeα = ℜeα′. We have therefore shown that ℜeα = ℜeα′ and δ′ = δ. Now, Proposition 3.3 and the hypothesis give (1 ± ℑmα) +p(1 ± ℑmα)2 − δ2 = kv1 ± iek2 = ki(v1 ± ie)k2 = kiv1 ∓ ek2 = kf (iv1) ∓ f (e)k2 = kif (v1) ∓ f (e)k2 = k − f (v1) ∓ if (e)k2 = kf (v1) ± if (e)k2 = (1 ± ℑmα′) +p(1 ± ℑmα′)2 − δ′2. Arguing as above, we get ℑmα′ = ℑmα, and hence α = α′. We have therefore proved that α = α′, and δ = δ′, and thus (23) with kP0(f (v1))(f (e))k = kP0(v1)(e)k = δ. f (e) = αf (v1) + P1(f (v1))(f (e)) + P0(f (v1))(f (e)), We consider now case (2). The same arguments given in case (1) lead us to (23). Since f (iv1) = if (v1), Proposition 3.7(b) gives f (ivj) = if (vj), for every j ∈ {2, 3, 4}. When in previous arguments we replace v1 with v2, v4 and v3 we obtain (24) f (e) = βf (v2) + P1(f (v2))(f (e)) + P0(f (v2))(f (e)), (25) and (26) f (e) = δf (v3) + P1(f (v3))(f (e)) + P0(f (v3))(f (e)), f (e) = γf (v4) + P1(f (v4))(f (e)) + P0(f (v4))(f (e)). Since (f (v1), f (v2), f (v3), f (v4)) is a quadrangle of minimal tripotents in C′, and hence αf (v1) + βf (v2) + γf (v4) + δf (v3) is a minimal tripotent in C′, we deduce from (23), (24), (25) and (26) that P2(αf (v1) + βf (v2) + γf (v4) + δf (v3))(f (e)) = αf (v1) + βf (v2) + γf (v4) + δf (v3), and since f (e) is a minimal tripotent in C′, Lemma 1.6 in [19] implies that f (e) = αf (v1) + βf (v2) + γf (v4) + δf (v3), which concludes the proof of (a). LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 21 (b) Let us assume that f (iv) = if (v) (the case f (iv1) = −if (v1) follows sim- ilarly). By Proposition 3.7(e) we have f (iv) = if (v) and f (iu) = if (u). Let e denote αv + βu + δv. As before, f (e) is a minimal tripotent and f (−e) = −f (e) (compare Theorem 2.3), and by Lemma 3.10 one of the following statements holds: (1) There exist minimal tripotents w2, w3, w4 in C′, and complex numbers α′, β′, γ′, δ′ such that (f (v), w2, w3, w4) is a quadrangle, α′2 +β′2 +γ′2 +δ′2 = 1, α′δ′ = β′γ′, and f (e) = α′f (v) + β′w2 + γ′w4 + δ′w3; (2) There exist a minimal tripotent w ∈ C′, a rank two tripotent u ∈ C′, and complex numbers α′, β′, δ′ such that (f (v), u, w) is a trangle, α′2 + 2β′2 + δ′2 = 1, α′δ′ = (β′)2, and f (e) = α′f (v) + β′u + δ′ w. In case (1), arguing as above we get (1 ∓ ℜeα) +p(1 ∓ ℜeα)2 − δ2 = kv ∓ ek2 = kf (v) ∓ f (e)k2 = (1 ∓ ℜeα′) +p(1 ∓ ℜeα′)2 − δ′2, from which we obtain ℜeα′ = ℜeα. Repeating previous arguments, we also have (1 ± ℑmα) +p(1 ± ℑmα)2 − δ2 = kv ± iek2 = ki(v ± ie)k2 = kiv ∓ ek2 = kf (iv) ∓ f (e)k2 = kif (v) ∓ f (e)k2 = k − f (v) ∓ if (e)k2 = kf (v) ± if (e)k2 = (1 ± ℑmα′) +p(1 ± ℑmα′)2 − δ′2, and consequently ℑmα′ = ℑmα, and α = α′. Therefore (27) P2(f (v))(f (e)) = αf (v). In case (2) we also arrive to (27) with similar arguments to those given above. This discussion remains valid when v is replaced by v, and we therefore have (28) P2(f (v))(f (e)) = δf (v). Now, we set w = 1 2 (v + u + v), w = 1 (w,bu, w) is a trangle in C and e = bαw + bβbu +bδ w, where bα = α+2β+δ and bδ = α−2β+δ . By the arguments given above we have 2 (v − u + v) and bu = v − v. The triplet , bβ = α−δ 2 , 2 2 (29) and (30) P2(f (w))(f (e)) = bαf (w) = P2(f ( w))(f (e)) =bδf ( w) = α + 2β + δ f (w), 2 α − 2β + δ 2 f ( w). Having in mind (27), (28), (29) and (30), and applying Lemma 3.11 to the element f (e) and the triplet (f (v), f (u), f (v)) (compare Proposition 3.7(d)), we get f (e) = αf (v) + δf (v) + βf (u) + P0(αf (v) + δf (v) + βf (u))(f (e)), and, by the minimality of f (e), we deduce that f (e) = αf (v) + δf (v) + βf (u), as desired. (cid:3) Corollary 3.13. Let f : S(C) → S(C′) be a surjective isometry between elementary JB∗-triples with rank(C) ≥ 2. Then either f (iu) = if (u) for every finite rank tripotent u in C, or f (iu) = −if (u) for every finite rank tripotent u in C. Proof. Suppose there exists a minimal tripotent e ∈ C such that f (ie) = if (e), and let v be any other minimal tripotent in C. By Lemma 3.10 one of the following statements holds: 22 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA (a) There exist minimal tripotents v2, v3, v4 in C, and complex numbers α, β, γ, δ such that (e, v2, v3, v4) is a quadrangle, α2 + β2 + γ2 + δ2 = 1, αδ = βγ, and v = αe + βv2 + γv4 + δv3, and hence iv = iαe + iβv2 + iγv4 + iδv3; (b) There exist a minimal tripotent e ∈ C, a rank two tripotent u ∈ C, and complex numbers α, β, δ such that (e, u, e) is a trangle, α2 + 2β2 + δ2 = 1, αδ = β2, v = αe + βu + δe and iv = iαe + iβu + iδe. Both cases will be treated independently. Proposition 3.7 assures that (a) f (ivj) = if (vj), for every j = 2, 3, 4; (b) f (iu) = if (u) and f (ie) = if (e). An application of Theorem 3.12 proves (a) f (iv) = iαf (v1) + iβf (v2) + iγf (v4) + iδf (v3) = if (v); (b) f (iv) = iαf (e) + iβf (u) + iδf (e) = if (v). The final statement is a consequence of the first conclusion and Theorem 2.3. (cid:3) Before finishing this section, we shall present another refinement of the Triple System Analyzer [8, Proposition 2.1 (iii)]. Lemma 3.14. Let f : S(C) → S(C′) be a surjective isometry between elemen- tary JB∗-triples with rank(C) ≥ 2, and let e be a minimal tripotent in C. Then f (S(C1(e))) = S(C′1(f (e))). Proof. Let e be a minimal tripotent in C, and let us pick a minimal tripotent u in C1(e). By the Triple System Analyzer (see [8, Proposition 2.1 (iii)]) either u is minimal in C and e⊤u or u is not minimal in C, u ⊢ e and the triplet (e, u, e = Q(u)(e)) is a trangle with e and e minimal in C. Since in the first case, we can always find minimal tripotents e3 and e4 in C such that (e, u, e3, e4) is a quadrangle (compare the arguments in the proof of Proposition 3.7(c)), we deduce, applying Proposition 3.7, that f (u) ∈ C′1(f (e)). the unit sphere of C1(e), one of the following holds (a) x is a minimal tripotent in C (this happens when C1(e) has rank one); (b) We can find two orthogonal minimal tripotents u1, u2 ∈ C1(e) and λ ∈ R such In case (a), by the arguments the first paragraph, we have f (x) ∈ C′1(f (e)). In case (b) we observe that, by the Triple System Analyzer, u1 and u2 are finite rank tripotents in C, and thus, Proposition 2.1 it follows that that x = u1 + λu2 and λ ≤ 1 (compare [3, Remark 4.6]). Since C1(e) has rank one or two (see [8, Corollary 2.2]), given an element x in f (x) = f (u1 + λu2) = f (u1) + Tu1(λu2) = f (u1) + λf (u2) ∈ C′1(f (e)). (cid:3) 4. Synthesis of a real linear isometry In a tour the force, T. Dang and Y. Friedman [8] and E. Neher [28] developed, independently, a complete theory on coordinatization theorems for the Jordan triple systems "covered" by a "grid". A grid in a JB∗-triple E is a family of minimal and rank two tripotents in E built up of quadrangles of minimal tripotents or trangles of the form (v, u, v) with v and v minimal, where all the non-vanishing triple products among the elements of the grid are those associated to these types of trangles and quadrangles. A typical grid in the Cartan factor Mn,m(C) is given by the family of all matrix units. LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 23 The results in [8] and [28] prove, among other classification theorems, that every Cartan factor C admits a (rectangular, symplectic, hermitian, spin, or exceptional) grid G such that the elementary JB∗-triple K associated with C is precisely the norm closed linear span of the grid G, and C being the weak∗-closure of K is nothing but the weak∗-closure of the linear span of G (compare [28, Structure Theorem IV.3.14] or [8, §2]). A more detailed description of the grids will be given in subsequent results. We can now state the main result of this section. Theorem 4.1. Let f : S(C) → S(C′) be a surjective isometry between the unit spheres of two elementary JB∗-triples with rank greater or equal than two. Then there exists a surjective complex linear or conjugate linear isometry T : C → C′ satisfying TS(C) = f . The proof will follow from Theorems 4.4, 4.5, 4.6, 4.7, and 4.9 below. These results will be obtained by an individualized approach on each elementary JB∗- triple. Remark 4.2. Let f : S(C) → S(C′) be a surjective isometry between the unit spheres of two elementary JB∗-triples with rank greater or equal than two. By Corollary 3.13 we know that f (ie) = if (e) or f (ie) = −if (e), for every finite rank tripotent e ∈ C. In the second case, we can always replace C′ with the complex JB∗-triple C′′ obtained from C′ by keeping the original norm, triple product, and sum of vectors but replacing the product by scalars with the product given by λ · x = λx (λ ∈ C, x ∈ C′). Then the mapping f : S(C) → S(C′′), x 7→ f (x) is a surjective isometry with f (ie) = if (e) for every minimal tripotent e ∈ C. If there exists a surjective complex linear isometry T : C → C′′ extending the mapping f , then we can easily find a conjugate linear isometry T : C → C′, T (x) = T (x), whose restriction to S(C) is precisely f . Remark 4.3. Suppose f : S(C) → S(C′) is a surjective isometry between the unit spheres of two elementary JB∗-triples with rank(C) ≥ 2. Let e be a minimal tripotent in C and let Te : C0(e) → C′0(f (e)) be the surjective real linear isometry given by Proposition 2.1(c). If rank(C0(e)) ≥ 2, it follows from [7, Proposition 2.6] that Te either is complex linear if f (ie) = if (e) or conjugate linear if f (ie) = −if (e) (compare Corollary 3.13). When C0(e) has rank 1 (and hence it is a complex Hilbert space regarded as a type 1 Cartan factor), every element in S(C0(e)) is a minimal tripotent in C0(e). Therefore, it follows from Corollary 3.13 that Te either is complex linear if f (ie) = if (e) or conjugate linear if f (ie) = −if (e). Actually, if F is a JB∗-subtriple of C, having in mind that a JB∗-triple of a weak- lyc ompact JB∗-triple is weakly compact, every element in F can be approximated in norm by a finite linear combination of mutually orthogonal minimal tripotents in F (see [3, Remark 4.6]). Moreover, every minimal tripotent in F is a finite rank tripotent in C. Therefore, given a bounded real linear operator T1 : F → C′ such that T1(x) = f (x) for every x ∈ S(F ), we deduce from Corollary 3.13 that T1 either is complex linear if f (ie) = if (e) or conjugate linear if f (ie) = −if (e), for any minimal tripotent e ∈ C. 4.1. Elementary JB∗-triples of type 1. We begin our particular study for an elementary JB∗-triple C of type 1 and rank between 2 and 4. We are mainly interested in the case C = L(H, H′), where H and H′ are complex Hilbert spaces 24 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA with 2 ≤ min{dim(H′), dim(H)} ≤ 4 (see Section 2), however the next result is established under more general hypothesis. Theorem 4.4. Let C = K(H, H′), where H and H′ are complex Hilbert spaces with 2 ≤ min{dim(H′), dim(H)}, and let C′ be an elementary JB∗-triple. Suppose f : S(C) → S(C′) is a surjective isometry. Then there exists a surjective complex linear or conjugate linear isometry T : C → C′ satisfying TS(C) = f . Proof. Let us first assume that f (ie) = if (e), for every minimal tripotent e ∈ C (compare Corollary 3.13). We deduce from Remark 4.3 that the operator Te given by Proposition 2.1(c) is complex linear. Let {ξi : i ∈ I} and {ηj : j ∈ J} be orthonormal basis of H and H′, respectively. We set uij := ηj ⊗ ξi, (i, j) ∈ I × J. Then, the family {uij : (i, j) ∈ I × J} is a rectangular grid in C (compare [8], [28]), however we will not make an explicit use of the properties of the grid in this case. To simplify the notation, we assume that 1, 2 ∈ I, J. Let us consider the mini- mal tripotents u11, u12, u21, u22, and for each one of them the surjective real linear isometry Tuij : C0(uij) → C′0(f (uij )). We can decompose C in the form C = Cu11 ⊕ (C0(u21) ∩ C1(u11)) ⊕ (C0(u12) ∩ C1(u11)) ⊕ C0(u11). Let P10 = P1(u11)P0(u21) = P0(u21)P1(u11), P01 = P1(u11)P0(u12). The unique- ness of the above decomposition shows that the mapping T : C → C′ given by T (x) = P2(f (u11))(x) + Tu21 (P10(x)) + Tu12(P01(x)) + Tu11(P0(u11)(x)) is a well defined bounded real linear operator. Let u = η ⊗ ξ be a minimal tripotent in C with kηk = 1 = kξk. A concrete decomposition similar to that given by Lemma 3.10 can be materialized as follows: let us write η = λ1η1 + λ2η0 and ξ = µ1ξ1 + µ2ξ0, where kη0k = 1 = kξ0k, η1 ⊥ η0, ξ0 ⊥ ξ1 (in the Hilbertian sense), λ1, λ2, µ1, µ2 ∈ C with λ12 + λ22 = 1, and µ12 + µ22 = 1. Thus, we have u = λ1µ1u11 + λ1µ2w12 + λ2µ1w21 + λ2µ2w0, where w12 = η1 ⊗ ξ0, w21 = η0 ⊗ ξ1, w0 = η0 ⊗ ξ0, where (u11, w12, w0, w21) is a quadrangle of minimal tripotents, w12 ∈ C1(u11)∩C0(u21), w21 ∈ C1(u11)∩C0(u12), w0 ∈ C0(u11). We are in position to apply Theorem 3.12(a) and the complex linearity of T (u11), T (u12), and T (u21) to deduce that f (u) = λ1µ1f (u11) + λ1µ2f (w12) + λ2µ1f (w21) + λ2µ2f (w0) = λ1µ1T (u11) + λ1µ2Tu21 (w12) + λ2µ1Tu12(w21) + λ2µ2Tu11 (w0) = T (λ1µ1u11) + Tu21(λ1µ2w12) + Tu12(λ2µ1w21) + Tu11 (λ2µ2w0) = T (λ1µ1u11) + T (λ1µ2w12) + T (λ2µ1w21) + T (λ2µ2w0) = T (u). We observe that T is actually complex linear. We have therefore shown that f (u) = T (u), for every minimal tripotent u ∈ C. Proposition 3.9 in [29] concludes that TS(C) = f . Finally, if f (ie) = −if (e), for every minimal tripotent e ∈ C. Let f , and C′′ be the mapping and the elementary JB∗-triple defined in Remark 4.2. The arguments above show that we can find a surjective real linear isometry T : C → C′′ such LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 25 (cid:3) that TS(C) = f . The arguments in the just quoted Remark show the existence of a surjective conjugate linear isometry T : C → C′ satisfying TS(C) = f . 4.2. Elementary JB∗-triples of types 2 and 3. In the next results we deal with elementary JB∗-triples of type 2 and 3 with rank greater or equal than 2. For this reason we fix a complex Hilbert space H, a conjugation j : H → H, and the complex linear involution on L(H) defined by xt = jx∗j (x ∈ L(H)). Theorem 4.5. Let C = {x ∈ K(H) : xt = −x} with rank(C) ≥ 2 (i.e. dim(H) ≥ 4), and let C′ be an elementary JB∗-triple. Suppose f : S(C) → S(C′) is a sur- jective isometry. Then there exists a surjective complex linear or conjugate linear isometry T : C → C′ satisfying TS(C) = f . Proof. As in the proof of the previous theorem, we first assume that f (ie) = if (e), for every minimal tripotent e ∈ C (compare Corollary 3.13). The case f (ie) = −if (e), for every minimal tripotent e ∈ C follows by similar arguments. Let {ξi : i ∈ I} be an orthonormal basis of H. Defining uij = j(ξi)⊗ξj−j(ξj)⊗ξi (i, j ∈ I), the set {uij : i 6= j in I} is a sympletic grid in the sense of [8, page 317]. The element u12 is a minimal tripotent in C and f (u12) satisfies the same prop- erty (see Proposition 2.1). By Proposition 2.1(c), there exists a surjective real linear isometry Tu12 : C0(u12) → C′0(f (u12)) satisfying f (u12 + x) = f (u12) + Tu12(x), for every x ∈ BC0(u12), and Tu12(x) = f (x) for all x ∈ S(C0(u12)). By Lemma 3.14, fS(C1(u12)) : S(C1(u12)) → S(C′1(f (u12))) is a surjective isom- etry. Let p = ξ1 ⊗ ξ1 + ξ2 ⊗ ξ2, p = j(ξ1) ⊗ j(ξ1) + j(ξ2) ⊗ j(ξ2). Then p and p are rank-2 projections in B(H). The space C1(u12) is isometrically isomorphic to B(p(H), (1 − p)(H)) via the mapping a 7→ (1 − p)ap. Applying Theorem 4.4 we deduce the existence of a surjective real linear isometry T1 : C1(u12) → C′1(f (u12)) satisfying T1(x) = f (x) for every x ∈ S(C1(u12)). The uniqueness of the Peirce decomposition C = Cu12 ⊕ C1(u12) ⊕ C1(u12) and the real linearity of the mappings Tu12 and T1 guarantee that the mapping T : C → C′, defined by T (x) = T (λu12 + P1(u12)(x) + P0(u12)(x)) = λf (u12) + T1(P1(u12)(x)) + Tu12 (P1(u12)(x)) is a well-defined continuous linear operator with kTk ≤ 3. Let e be any minimal tripotent in C. By Lemma 3.10 one of the following statements holds: (a) There exist minimal tripotents v2, v3, v4 in C, and complex numbers α, β, γ, δ such that (u12, v2, v3, v4) is a quadrangle, α2 +β2 +γ2 + δ2 = 1, αδ = βγ, and e = αu12 + βv2 + γv4 + δv3; (b) There exist a minimal tripotent u12 ∈ C, a rank two tripotent u ∈ C, and complex numbers α, β, δ such that (u12, u, u12) is a trangle, α2+2β2+δ2 = 1, αδ = β2, and e = αu12 + βu + δu12. Applying Theorem 3.12 and the definition of T we get (a) f (e) = f (αu12 + βv2 + γv4 + δv3) = αf (u12) + βf (v2) + γf (v4) + δf (v3) 26 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA = αf (u12) + βT1(v2) + γT1(v4) + δTu12(v3) = (by Remark 4.3) = f (αu12) + T1(βv2) + T1(γv4) + Tu12 (δv3) = T (αu12) + T (βv2) + T (γv4) + T (δv3) = T (e), or, (b) f (e) = f (αu12 + βu + δu12) = αf (u12) + βf (u) + δf (u12) = αT (u12) + βT1(u) + δTu12 (u12) = (by Remark 4.3) = T (αu12) + T1(βu) + Tu12 (δu12) = T (e), respectively. We have therefore shown that f (e) = T (e) for every minimal tripotent e ∈ C. Proposition 3.9 in [29] proves that T (x) = f (x) for every x ∈ S(C), and hence T is surjective and isometric. We note that T actually is complex linear. (cid:3) We shall deal next with a Cartan factor of type 3. Theorem 4.6. Let C = {x ∈ K(H) : xt = x}, where dim(H) ≥ 2, and let C′ be an elementary JB∗-triple. Suppose f : S(C) → S(C′) is a surjective isometry. Then there exists a surjective complex linear or conjugate linear isometry T : C → C′ satisfying TS(C) = f . Proof. The proof follows similar guidelines to the proof of Theorem 4.5. We may assume thanks to Corollary 3.13, that f (ie) = if (e), for every minimal tripotent e ∈ C. The case f (ie) = −if (e), for every minimal tripotent e ∈ C is very similar. Let {ξi : i ∈ I} be an orthonormal basis of H. Defining uij = (j(ξi)⊗ ξj + j(ξj)⊗ ξi) (i 6= j ∈ I), and uii = (j(ξi) ⊗ ξi + j(ξi) ⊗ ξi) (i ∈ I), the set {uij : i, j ∈ I} is a hermitian grid in the sense of [8, page 308]. Let Tu11 : C0(u11) → C′0(f (u11)) be the surjective real linear isometry satisfying f (u11 + x) = f (u11) + Tu11(x), for every x ∈ BC0(u11), and Tu11 (x) = f (x) for all x ∈ S(C0(u11)), whose existence is guaranteed by Proposition 2.1(c). Lemma 3.14 implies that fS(C1(u11)) : S(C1(u11)) → S(C′1(f (u11))) is a sur- jective isometry. The elementary JB∗-triple C1(u11) has rank one, and hence, it is isometrically isomorphic to a Hilbert space. It follows from [10, Theorem 2.1] (see also [29, Corollary 3.15]) that there exists a surjective real linear isometry T1 : C1(u11) → C′1(f (u11)) satisfying T1(x) = f (x) for every x ∈ S(C1(u11)). Remark 4.3 guarantees that T1 and Tu12 are complex linear because we have assumed that f (ie) = if (e), for every minimal tripotent e ∈ C. that the mapping T : C → C′, defined by Repeating the arguments in the final part of the proof of Theorem 4.5 we deduce T (x) = T (λu12 + P1(u12)(x) + P0(u12)(x)) = λf (u12) + T1(P1(u12)(x)) + Tu12(P1(u12)(x)), is a well-defined continuous linear operator, and T (x) = f (x) for every x ∈ C, which concludes the proof. (cid:3) LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 27 4.3. Finite dimensional elementary JB∗-triples. Surjective isometries between finite-dimensional Cartan factors can be treated by a unified approach. We recall that Cartan factors of types 5 and 6 are finite dimensional. Theorem 4.7. Let C, C′ be elementary JB∗-triples with dim(C) < ∞. Suppose f : S(C) → S(C′) is a surjective isometry. Then there exists a surjective real isometry T : C → C′ satisfying TS(C) = f . Furthermore, if rank(C) ≥ 2 then T is complex linear or conjugate linear. Proof. We shall argue by induction on the dimension of C. conclusion follows from [10, Theorem 2.1] (see also [29, Corollary 3.15]). dim(C) = 1 the statement has been proved. If rank(C) = 1 the If the Henceforth, we assume rank(C) ≥ 2. Let us assume that our statement is true for any Cartan factor C with dimension ≤ n, and dim(C) = n + 1. Let us pick a minimal tripotent e ∈ C. We can assume, via Corollary 3.13, that f (ie) = if (e) (the case f (ie) = −if (e) follows with similar techniques). By Lemma 3.14 fS(C1(e)) : S(C1(e)) → S(C′1(f (e))) is a surjective isometry. Since dim(C1(e)) ≤ n, and C1(e) being generated by a standard grid (see [28, §IV.3]) is another Cartan factor, we conclude from the induction hypothesis that there exists a surjective real linear isometry T1 : C1(e) → C1(e) satisfying T1(x) = f (x) for all x ∈ S(C1(e)). By Proposition 2.1(c) we can find a surjective real linear isometry Te : C0(e) → C0(e) such satisfying f (x) = Te(x) for all x ∈ S(C0(e)). We define a bounded linear mapping T : C → C′ given by T (x) = T (λe + P1(e)(x) + P0(e)(x)) := λf (e) + T1(P1(e)(x)) + Te(P0(e)(x)). The arguments given in the last three paragraphs of the proof of Theorem 4.5 can be now repeated to show that T (u) = f (u) for every minimal tripotent u ∈ C. By [29, Proposition 3.9] we have f (x) = T (x) for every x ∈ S(C), which concludes the induction argument. The final observation follows as a consequence of Remark 4.3. (cid:3) 4.4. Spin factors. In this section we explore Tingley's problem for surjective isometries from the unit sphere of a spin factor into the unit sphere of an ele- mentary JB∗-triple. The starting lemma shows that we cannot find a surjective isometry from the unit sphere of an infinite dimensional spin factor onto the unit sphere of a type 1 Cartan factor. Lemma 4.8. Let X be a spin factor, and let C = L(H, H′), where H and H′ are complex Hilbert spaces with dim(H′) = 2, and dim(H) ≥ 3. Then there is no surjective isometry f : S(X) → S(C). Proof. Suppose we can find a surjective isometry g : S(C) → S(X). By hypothesis, we can find at least three minimal tripotents e11, e12 and e3 in C satisfying e11⊤e12, e11 and e12 generate a 2-dimensional complex Hilbert space, and e11, e12 ⊥ e3 (take for example e11 = η1 ⊗ ξ1, e12 = η1 ⊗ ξ2, and e3 = η2 ⊗ ξ3, where {η1, η2} and {ξ1, ξ2, ξ3} are orthonormal systems in H′ and H, respectively). Let us consider the surjective real linear isometry Te3 : C0(e3) → X0(f (e3)) given by Proposition 2.1(c). The elements e11, e12, and 1√2 (e11 + e12) are minimal tripotents in C0(e3). 28 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA Since f (e3) is a minimal tripotent in a spin factor X, its orthogonal comple- ment X0(f (e3)) = Cf (e3) is a one dimensional complex space, where · denotes the conjugation on X (compare (17)). The minimal tripotents f (e11), f (e12), and f ( 1√2 (e11 ± e12)) belong to X0(f (e3)). So, there exist λ1 and λ2 in the unit sphere of C such that f (e11) = λ1f (e3) and f (e12) = λ2f (e3). The elements f(cid:16) 1 √2 (e11 ± e12)(cid:17) = Te3(cid:16) 1 √2 (e11 ± e12))(cid:17) = 1 √2 √2(cid:16)Te3 (e11) ± Te3(e12)(cid:17) (λ1 ± λ2)f (e3) √2(cid:16)f (e11) ± f (e12)(cid:17) = = 1 1 must be norm-one, which implies λ2 = (−1)kiλ1 for some natural k. We observe that, by Remark 4.3, Te3 is complex linear or conjugate linear. Arguing as above, we prove the existence of a natural m such that f(cid:16) 1 √2 1 (e11 ± ie12)(cid:17) = Te3(cid:16) 1 √2 (e11 ± ie12))(cid:17) = √2(cid:16)f (e11) ± (−1)mif (e12)(cid:17) = 1 √2 1 ± (−1)k+1+m √2 = = λ1f (e3) 1 √2(cid:16)Te3(e11) ± (−1)miTe3(e12)(cid:17) (λ1 ± (−1)k+1+mλ1)f (e3) must be norm-one too, which is impossible. (cid:3) Theorem 4.9. Let X be a spin factor, let C′ be an elementary JB∗-triple, and let f : S(X) → S(C′) be a surjective isometry. Then there exists a surjective real linear isometry T : X → C′ satisfying TS(X) = f . Furthermore, the operator T can be chosen to be complex linear or conjugate linear. Proof. If X is finite dimensional the conclusion follows from Theorem 4.7. We can therefore assume that X is infinite dimensional. Theorem 2.3 implies that X and C′ are infinite dimensional rank 2 Cartan factors. Therefore C′ is either a spin factor or a type 1 Cartan factor of the form B(H, H′), where H, H′ are complex Hilbert spaces, dim(H′) = 2, and dim(H) = ∞. Lemma 4.8 shows that C′ = Y must be an infinite dimensional spin factor. As in previous results, we can assume that f (ie) = if (e) for every minimal tripotent e in X (compare Corollary 3.13 and Remark 4.3). Applying that X has rank two and Theorem 2.3 we deduce that f (iu) = if (u) for every tripotent u in X. By a little abuse of notation, the involutions on X and on Y will be both denoted · , similarly, we shall indistinctly write (..), k.k2, and k.k for the by the symbol inner products, the Hilbertian norms, and the spin norms on X and Y . The symbols X1, X2, Y1 and Y2 will have the usual meanings commented above, that is, X1 = {x ∈ X : x = x}, Y1 = {y ∈ Y : y = x}, X2 = iX1, and Y2 = iY1. By the arguments in the first paragraph, we can assume that X (and hence X1) is infinite dimensional. In a first step we shall show that given x1, x2 in S(X1) with (x1x2) = 0 we have f (x1) = µ1y1, f (x2) = µ2y2, for suitable y1, y2 ∈ S(Y1), µ1, µ2 ∈ C, with µj = 1, µ2 ∈ {±µ1}, and (y1y2) = 0. Indeed, it is known that, for each j = 1, 2, xj is a rank two tripotent which is LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 29 The elements e = x1+ix2 also unitary in X (see the remarks before Lemma 4.8). By Theorem 2.3 f (xj ) ∈ max U(Y ), and hence there exists a (unique) norm one element yj in Y1 and µj ∈ C with µj = 1 such that f (xj) = µjyj. are minimal tripotents in X with e ⊥ e (compare the comments before Lemma 4.8). Since f (e) and f (e) are orthogonal minimal tripotents in Y , µ1y1 = f (x1) = f (e) + f (e) and, since we have assumed that f (iu) = if (u) for every tripotent u in X, we have iµ2y2 = f (ix2) = f (e1) − f (e1) (see Theorem 2.3). Therefore , e = x1−ix2 2 2 f (e) = µ1y1 + iµ2y2 2 , and f (e) = µ1y1 − iµ2y2 2 . Having in mind that f (e) ⊥ f (e) in Y , the identity proves that 1 (31) f (e) = {f (e), f (e), f (e)} = 2(f (e)f (e))f (e) − (f (e)f (e))f (e) 2 = (f (e)f (e)) and 1 0 = (f (e)f (e)) = 4 2 + iµ1µ2(y1y2) + iµ2µ1(y2y1)). 1 − µ2 (µ2 Replacing f (e) with f (e) we also get 1 4 (µ2 1 − µ2 0 = (f (e)f (e)) = (32) Combining (31), and (32) we get R ∋ (y1y2) = 0 and µ2 µ0 = 1 satisfying 2, and hence µ2 = ±µ1. In the second step we shall prove the existence of a complex number µ0 with 2 − iµ1µ2(y1y2) − iµ2µ1(y2y1)). 1 = µ2 f (x1) ∈ µ0S(Y1), for every x1 ∈ S(X1). For this purpose, pick a norm one element x0 in X1. As before, there exist µ0 ∈ C and y0 ∈ S(Y1) such that f (x0) = µ0y0. Let x1 be any element in S(X1). We can find a third element x2 ∈ S(X1) satisfying (x1x2) = (x0x2) = 0. Applying the first step to the pairs (x0, x2) and (x1, x2) we get f (x2) = µ0z2, and f (x1) = µ0z1, for suitable z2, z1 ∈ S(Y1). This finishes the proof of the second step. The sets X1 and µ0Y1 are real linear closed subspaces of X and Y , respectively. Furthermore, the spin norms on X1 and on µ0Y1 are precisely the Hilbertian norms associated to the inner products (..), in other words, (X1,k.k) = (X1,k.k2) and (µ0Y1,k.k) = (µ0Y1,k.k2). We have proved in the second step that fS(X1) : S(X1) → S(µ0Y1) is a surjective isometry between the unit spheres of two Hilbert spaces. We deduce from [10] (see also Corollary 3.15 in [29]) the existence of a surjective real linear isometry F : X1 → µ0Y1 satisfying F (x1) = f (x1) for every x1 ∈ S(X1). We define a (complex) linear operator T : X → Y given by T (x1 + iz1) := F (x1) + iF (z1) (x1 + iz1 ∈ X = X1 ⊕ iX1). Every minimal tripotent in X is of the form e = x1+iz1 , where x1, z1 ∈ S(X1) and (x1z1) = 0. The elements u = x1 and w = iz1 are complete tripotents in X. For each t ∈ [0, 1] we have 2 ktu + (1 − t)wk2 2 = t2 + (1 − t)2, (tu + (1 − t)w tu + (1 − t)w)2 = (t2 − (1 − t)2)2, 30 F.J. FERN ´ANDEZ-POLO AND A.M. PERALTA and hence ktu + (1 − t)wk2 = t2 + (1 − t)2 +p(t2 + (1 − t)2)2 − (t2 − (1 − t)2)2 = 1. This shows that [u, w] = {tu + (1 − t)w : t ∈ [0, 1]} is a convex subset in S(X), and by Zorn's lemma, it is contained in some maximal proper norm closed face M of BX. We conclude, by Proposition 2.1(b), that 1 f (u) + 2 f (e) = f(cid:18) 1 w(cid:19) = 1 2 u + f (w) = f (x1) + f (iz1) 1 2 1 2 2 1 2 = 1 2 f (x1) + i 1 2 f (z1) = (by definition of T ) = 1 2 T (u) + 1 2 T (w) = T (e). Therefore, f (e) = T (e) for every minimal tripotent e ∈ X. An application of Proposition 3.9 in [29] proves that f (x) = T (x) for every x ∈ S(X), and hence T is a surjective linear isometry. We finally observe that if we assume that f (ie) = −if (e) for every minimal tripotent e in X, then the above arguments show the existence of a conjugate linear isometry T : X → Y satisfying f (x) = T (x) for every x ∈ S(X). (cid:3) Acknowledgements Authors partially supported by the Spanish Ministry of Economy and Competitiveness and European Regional Development Fund project no. MTM2014-58984-P and Junta de Andaluc´ıa grant FQM375. References [1] T. Barton abd R. M. Timoney, Weak∗-continuity of Jordan triple products and applications, Math. Scand. 59 (1986), 177 -- 191. [2] J. Becerra Guerrero, G. L´opez P´erez, A. M. Peralta, A. Rodr´ıguez-Palacios, A. Relatively weakly open sets in closed balls of Banach spaces, and real JB∗-triples of finite rank, Math. Ann. 330 (2004), no. 1, 45 -- 58. [3] L.J. Bunce and C.-H. Chu, Compact operations, multipliers and Radon-Nikodym property in J B∗-triples, Pacific J. Math. 153 (1992), 249 -- 265. [4] L.J. Bunce, F.J. Fern´andez-Polo, J. Mart´ınez Moreno, A.M. Peralta, A Saito-Tomita-Lusin theorem for JB∗-triples and applications, Quart. J. Math. Oxford, 57 (2006), 37-48. [5] M. Burgos, F.J. Fern´andez-Polo, J. Garc´es, J. Mart´ınez, A.M. Peralta, Orthogonality pre- servers in C∗-algebras, JB∗-algebras and JB∗-triples, J. Math. Anal. Appl. 348, 220-233 (2008). [6] Ch.-H. Chu. Jordan structures in geometry and analysis., Cambridge, Cambridge University Press, 2012. [7] T. Dang, Real isometries between JB∗-triples, Proc. Amer. Math. Soc. 114, 971-980 (1992). [8] T. Dang, Y. Friedman, Classification of JBW∗-triple factors and applications, Math. Scand. 61, no. 2, 292-330 (1987). [9] S. Dineen, The second dual of a JB∗-triple system, In: Complex analysis, functional analysis and approximation theory (ed. by J. M´ugica), 67-69, (North-Holland Math. Stud. 125), North- Holland, Amsterdam-New York, (1986). [10] G. Ding, The 1-Lipschitz mapping between the unit spheres of two Hilbert spaces can be extended to a real linear isometry of the whole space, Sci. China Ser. A 45 (2002), no. 4, 479-483. [11] G. G. Ding, The isometric extension problem in the spheres of lp(Γ) (p > 1) type spaces, Sci. China Ser. A 46 (2003), 333 -- 338. [12] G. G. Ding, The representation theorem of onto isometric mappings between two unit spheres of l∞-type spaces and the application on isometric extension problem, Sci. China Ser. A 47 (2004), 722 -- 729. [13] G. G. Ding, The representation theorem of onto isometric mappings between two unit spheres of l1(Γ) type spaces and the application to the isometric extension problem, Acta. Math. Sin. (Engl. Ser.) 20 (2004), 1089 -- 1094. LOW RANK COMPACT OPERATORS AND TINGLEY'S PROBLEM 31 [14] G. Ding, On isometric extension problem between two unit spheres, Sci. China Ser. A 52 (2009) 2069-2083. [15] C.M. Edwards, F.J. Fern´andez-Polo, C.S. Hoskin, A.M. Peralta, On the facial structure of the unit ball in a JB∗-triple, J. Reine Angew. Math. 641 (2010) 123-144. [16] C.M. Edwards, G.T. Ruttimann, On the facial structure of the unit balls in a JBW∗-triple and its predual, J. Lond. Math. Soc. 38, 317-322 (1988). [17] C.M. Edwards, G.T. Ruttimann, Compact tripotents in bi-dual JB∗-triples, Math. Proc. Camb. Phil. Soc. 120, 155-173 (1996). [18] F.J. Fern´andez-Polo, J. Mart´ınez, A.M. Peralta, Surjective isometries between real JB∗- triples, Math. Proc. Cambridge Phil. Soc., 137 709-723 (2004). [19] Y. Friedman and B. Russo. Structure of the predual of a J BW ∗-triple. J. Reine Angew. Math., 356:67 -- 89, 1985. [20] G. Horn, Characterization of the predual and ideal structure of a JBW∗-triple, Math. Scand. 61, no. 1, 117-133 (1987). [21] F.B. Jamjoom, A.M. Peralta, A.A. Siddiqui, H.M. Tahlawi, Cebysev subspaces of JBW∗- triples, J. Inequal. Appl. (2015) 2015:288. DOI 10.1186/s13660-015-0813-2. [22] V. Kadets and M. Mart´ın, Extension of isometries between unit spheres of infite-dimensional polyhedral Banach spaces, J. Math. Anal. Appl., 396 (2012), 441-447. [23] W. Kaup, Uber die Klassifikation der symmetrischen hermiteschen Mannigfaltigkeiten un- endlicher Dimension I, Math. Ann. 257, 463-486 (1981). [24] W. Kaup, A Riemann Mapping Theorem for bounded symmentric domains in complex Ba- nach spaces, Math. Z. 183, 503-529 (1983). [25] W. Kaup, On real Cartan factors, Manuscripta Math. 92, 191-222 (1997). [26] W. Kaup and H. Upmeier, Jordan algebras and symmetric Siegel domains in Banach spaces, Math. Z. 157, 179-200 (1977). [27] P. Mankiewicz, On extension of isometries in normed linear spaces, Bull. Acad. Pol. Sci., S´er. Sci. Math. Astron. Phys. 20 (1972) 367-371. [28] E. Neher, Jordan triple systems by the grid approach, Lecture Notes in Mathematics, 1280. Springer-Verlag, Berlin, 1987. [29] A.M. Peralta, R. Tanaka, A solution to Tingley's problem for isometries between the unit spheres of compact C∗-algebras and JB∗-triples, preprint 2016. arXiv:1608.06327v1. [30] D. Tan, Extension of isometries on unit sphere of L∞, Taiwanese J. Math. 15 (2011), 819 -- 827. [31] D. Tan, On extension of isometries on the unit spheres of Lp-spaces for 0 < p ≤ 1, Nonlinear Anal. 74 (2011), 6981-6987. [32] D. Tan, Extension of isometries on the unit sphere of Lp-spaces, Acta. Math. Sin. (Engl. Ser.) 28 (2012), 1197 -- 1208. [33] R. Tanaka, Spherical isometries of finite dimensional C ∗-algebras, J. Math. Anal. Appl. 445 (2017), no. 1, 337-341. [34] R. Tanaka, Tingley's problem on finite von Neumann algebras, preprint 2016. [35] D. Tingley, Isometries of the unit sphere, Geom. Dedicata 22 (1987) 371-378. [36] R.S. Wang, Isometries between the unit spheres of C0(Ω) type spaces, Acta Math. Sci. (Eng- lish Ed.) 14 (1994), no. 1, 82-89. [37] X. Yang, X. Zhao, On the extension problems of isometric and nonexpansive mappings. In: Mathematics without boundaries. Edited by Themistocles M. Rassias and Panos M. Pardalos. 725-748, Springer, New York, 2014. Departamento de An´alisis Matem´atico, Facultad de Ciencias, Universidad de Granada, 18071 Granada, Spain. E-mail address: [email protected] E-mail address: [email protected]
1803.01911	5	1803	2019-04-26T14:24:04	Dagger and Dilation in the Category of Von Neumann algebras	[ "math.OA", "cs.LO", "math.FA", "math.QA" ]	This doctoral thesis is a mathematical study of quantum computing, concentrating on two related, but independent topics. First up are dilations, covered in chapter 2. In chapter 3 "diamond, andthen, dagger" we turn to the second topic: effectus theory. Both chapters, or rather parts, can be read separately and feature a comprehensive introduction of their own.	math.OA	math	Dagger and Dilation in the Category of Von Neumann Algebras . A O h t a m [ 5 v 1 1 9 1 0 . 3 0 8 1 : v i X r a P ter aan op volgens in op dinsdag om 10.30 door Bastiaan Eelco Westerbaan geboren te 1 Third corrected version (april 26th, 2019) Identifiers hdl: 2066/201785 arXiv: 1803.01911 isbn: 978-94-6332-485-4 Persistent links https://arxiv.org/abs/1803.01911 https://doi.org/2066/201785 https://hdl.handle.net/2066/201785 Source code LATEX https://github.com/westerbaan/theses https://github.com/westerbaan/ndpt cover Printed by GVO drukkers & vormgevers B.V., Ede, https://proefschriften.nl. Where applicable, © 2019 B.E. Westerbaan, c b available under cc by, [1]. Dagger and Dilation in the Category of Von Neumann Algebras Proefschrift ter verkrijging van de graad van doctor aan de Radboud Universiteit Nijmegen op gezag van de rector magnificus prof. dr. J.H.J.M. van Krieken, volgens besluit van het college van decanen in het openbaar te verdedigen op dinsdag 14 mei 2019 om 12.30 uur precies door Bastiaan Eelco Westerbaan geboren op 30 augustus 1988 te Nijmegen 3 Promotor: Prof. dr. B.P.F. Jacobs Manuscriptcommissie: Prof. dr. J.D.M. Maassen Prof. dr. P. Panangaden (McGill University, Canada) Prof. dr. P. Selinger (Dalhousie University, Canada) Dr. C.J.M. Heunen (University of Edinburgh, Verenigd Koninkrijk) Dr. S. Staton (University of Oxford, Verenigd Koninkrijk) Dagger and Dilation in the Category of Von Neumann Algebras Doctoral Thesis to obtain the degree of doctor from Radboud University Nijmegen on the authority of the Rector Magnificus prof. dr. J.H.J.M. van Krieken, according to the decision of the Council of Deans to be defended in public on Tuesday, May 14, 2019 at 12.30 hours by Bastiaan Eelco Westerbaan born on August 30, 1988 in Nijmegen (the Netherlands) 5 Supervisor: Prof. dr. B.P.F. Jacobs Doctoral Thesis Committee: Prof. dr. J.D.M. Maassen Prof. dr. P. Panangaden (McGill University, Canada) Prof. dr. P. Selinger (Dalhousie University, Canada) Dr. C.J.M. Heunen (University of Edinburgh, United Kingdom) Dr. S. Staton (University of Oxford, United Kingdom) 7 1 Introduction 134 2 Dilations The basics of Hilbert C∗-modules 135 Stinespring's theorem . . . . . . . . . . . . . . . . . . . . . . . . 136 2.1 2.2 Hilbert C∗-modules . . . . . . . . . . . . . . . . . . . . . . . . . 141 2.2.1 . . . . . . . . . . . . . 142 2.2.2 Ultranorm uniformity . . . . . . . . . . . . . . . . . . . . 146 2.2.3 Self-dual completion . . . . . . . . . . . . . . . . . . . . . 150 2.3 Paschke dilations . . . . . . . . . . . . . . . . . . . . . . . . . . . 154 2.4 Kaplansky density theorem for Hilbert C∗-modules . . . . . . . 158 2.5 More on self-dual Hilbert C∗-modules . . . . . . . . . . . . . . . 159 Exterior tensor product . . . . . . . . . . . . . . . . . . . 164 2.6 Pure maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 2.5.1 3 Diamond, andthen, dagger 3.1.1 3.1.2 3.2.1 3.2.2 3.2.3 3.2.4 173 3.1 Effect algebras and related structures . . . . . . . . . . . . . . . 174 Effect monoids . . . . . . . . . . . . . . . . . . . . . . . . 178 Effect modules . . . . . . . . . . . . . . . . . . . . . . . . 179 3.2 Effectuses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 . . . . . . . . . . . . . . . . . . . . 181 From partial to total . . . . . . . . . . . . . . . . . . . . 182 From total to partial Predicates, states and scalars . . . . . . . . . . . . . . . . 190 Effectus of abstract M -convex sets . . . . . . . . . . . . . 193 3.3 Effectuses with quotients . . . . . . . . . . . . . . . . . . . . . . 197 3.4 Effectuses with comprehension and images . . . . . . . . . . . . 199 3.4.1 Sharp predicates . . . . . . . . . . . . . . . . . . . . . . . 203 ⋄-effectuses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 †-effectuses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214 3.7.1 Dilations . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 3.8 Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224 . . . . . . . . . . . . . . . . . . 224 . . . . . . . . . . . . . . . . . . . . . 225 . . . . . . . . . . . . . . . . . . . 226 3.8.1 Dagger kernel categories 3.8.2 3.8.3 Homological categories 3.5 3.6 &-effectuses 3.7 Sequential product 4 Outlook 229 Chapter 1 Introduction This thesis is a mathematical study of quantum computing, concentrating on two related, but independent topics. First up are dilations, covered in chapter 2. In chapter 3 "diamond, andthen, dagger" we turn to the second topic: effectus theory. Both chapters, or rather parts, can be read separately and feature a comprehensive introduction of their own. For now, we focus on what they have in common. Both parts study the category of von Neumann algebras (with ncpsu-maps∗ between them). A von Neumann algebra is a model for a (possibly infinite- dimensional) quantum datatype, for instance: 134 II a qubit a qutrit a qubit or a qutrit M2 M3 M2 ⊕ M3 The ncpsu-maps between these algebras represent physical (but not necessarily computable) operations in the opposite direction (as in the 'Heisenberg picture'): an (ordered) pair of qutrits a bit quantum integers M3 ⊗ M3 C2 B(ℓ2) 1. Measure a qubit in the computational basis (λ, µ) 7→ λ0ih0 + µ1ih1, C2 → M2, (qubit → bit) using braket notation† 2. Discard the second qubit (qubit × qubit → qubit) M2 → M2 ⊗ M2, a 7→ a ⊗ 1 3. Apply a Hadamard-gate to a qubit a 7→ H∗aH, where H = 1√2(cid:0) 1 1 1 −1(cid:1) (qubit → qubit) M2 → M2, ∗An abbreviation for normal completely positive contractive (i.e. subunital) linear maps. †See e.g. [91, fig. 2.1], so here 0i ≡ (cid:0) 1 0 (cid:1), h0 ≡ ( 1 0 ), 1i ≡ (cid:0) 0 1 (cid:1) and 0ih0 = (cid:0) 1 0 0 0 (cid:1). 134.. 9 4. Initialize a qutrit as 0 (empty → qutrit) M3 → C, a 7→ h0 a0i III The first part of this thesis starts with a concrete question: does the famous Stinespring dilation theorem [106], a cornerstone of quantum theory, extend to every ncpsu-map between von Neumann algebras? Roughly, this means that every ncpsu-map splits in two maps of a particularly nice form. We will see that indeed it does. The insights attained pondering this question lead naturally, as they so often do in mathematics, to other new results: for instance, we learn more about the tensor product of dilations (167 I) and the pure maps associated to them (171 VII). But perhaps the mathematically most interesting by-catch of pondering the question are the contributions (see 135 VI) to the theory of self-dual Hilbert C∗-modules over a von Neumann algebra, which underlie these dilations. In contrast to the first part of this thesis, the second part starts with the abstract (or rather: vague) question: what characterizes the category of von Neumann algebras? The goal is to find a set of reasonable axioms which pick out the category of von Neumann algebras among all other categories. In other words: a reconstruction of quantum theory with categorical axioms (presupposing von Neumann algebras are the right model of quantum theory). Unfortunately, we do not reach this goal here.‡ However, by studying von Neumann algebras with this self-imposed restriction, one is led to, as again is often the case in mathematics, upon several new notions and results, which are of independent interest, such as, for instance, ⋄-adjointness, pure maps and the existence of their dagger. IV V Twin theses A substantial part of the research presented here has been per- formed in close collaboration with my twin brother Abraham Westerbaan. Our results are interwoven to such a degree that separating them is difficult and, more importantly, detrimental to the clarity of presentation. Therefore we de- cided to write our theses as two consecutive volumes (with consecutive number- ing, more on that later). Abraham's thesis [124] is the first volume and covers the preliminaries on von Neumann algebras and, among other results, an ax- iomatization of the sequential product b 7→ √ab√a. The present thesis is the second volume and builds upon the results of Abraham's thesis, in its study of dilations and effectus theory. Because of this clear division, a significant amount of Abraham's original work appears in my thesis and, vice versa, my work ap- pears in his. However, the large majority of the new results in each of our theses is our own. VI Preliminaries For the first part of this thesis, only knowledge is presumed of, what is considered to be, the basic theory of von Neumann algebras. Unfor- ‡Building on this thesis, van de Wetering recently found a reconstruction of finite- dimensional quantum theory [117, 118]. tunately, this 'basic theory' takes most introductory texts close to a thousand pages to develop. This situation has led my twin brother to ambitiously rede- velop the basic theory of von Neumann algebras tailored to the needs of our theses (and hopefully others in the field). In his thesis, Abraham presents [124] a clear, comprehensive and self-contained development of the basic theory of von Neumann algebras in less than 200 pages. For a traditional treatment, I can wholeheartedly recommend the books of Kadison and Ringrose [73]. For the second part of this thesis, only elementary category theory is required as covered in any introductory text, like e.g. [9]. Writing style This thesis is divided into points, each a few paragraphs long. Points are numbered with Roman numerals in the margin: this is point VII. Points are organized into groups, which are assigned Arabic numerals, that ap- pear at the start of the group in the margin. The group of this point is 134. Group numbers below 134 appear in my twin brother's thesis [124]. Page num- bers have been replaced by the group numbers that appear on the spread.§ This unusual system allows for very short references: instead of writing 'see Theorem 2.3.4 on page 123', we simply write 'see 154 III'. Even better: 'see the middle of the proof of Theorem 2.3.4 on page 123' becomes the more precise 'see 154 VIII'. Somewhat unconventionally, this thesis includes exercises. Most exercises replace straight-forward, but repetitive, proofs. Other exercises, marked with an asterisk , are harder and explore tangents. In both cases, these exercises are not meant to force the reader into the role of a pupil, but rather to aid her, by reducing clutter. Solutions to these exercises (page 197) and errata to the printed version of this text (page 239), can be found at the end of this PDF. Possibly further errata will be made available on the arXiv under 1803.01911. VII Advertisements This thesis is based on the publications [22,69,126,127]. During my doctoral research, I also published the following papers, which although somewhat related, would've made this thesis too broad in scope if included. VIII 1. In A Kochen -- Specker System has at least 22 Vectors [116], Uijlen and I justify the title building upon the work of Arends, Ouaknine and Wampler. 2. In Unordered Tuples in Quantum Computation [35] Furber and I argue that M3 ⊕ C is the right algebra modeling an unordered pair of qubits. Then we compute the algebra for an unordered n-tuple of d-level systems. 3. In Picture-perfect QKD [76] Kissinger, Tull and I show how to graphically derive a security bound on quantum key distribution using the continuity of Stinespring's dilation. 4. In [100], Schwabe and I work out the details of a quantum circuit to break binary MQ using Grover's algorithm. §A spread are two pages that are visible at the same time. ..134.. 11 In the second part of this thesis, we use the notion of effectus (introduced by Jacobs [57]), solely for studying quantum theory. Effectus theory, however, has more breadth to it, see e.g. [3, 20, 21, 58 -- 64, 67, 68, 70]. For a long, but comprehensive introduction, see [23]. IX Next, I would like to mention some other recent approaches to the mathemat- ical study of quantum theory which are related in spirit and of which I am aware. First, there are colleagues who also study particular categories related to quantum theory. Spearheaded by Coecke and Abramsky [2], dozens of au- thors around Oxford have developed insightful graphical calculi by studying the category of Hilbert spaces categorically [25]. Not explicitly categorically (but certainly in spirit) is the work around operational probabilistic theories, see e.g. [27]. Besides these approaches (and that of effectus theory), there are many other varied categorical forages into quantum theory, e.g. [34, 78, 96, 105]. Then there are authors who pick out different structures from quantum theory to study. Most impressive is the work of Alfsen and Shultz, who studied and axiomatized the state spaces of many operator algebras, including von Neumann algebras [4]. Recently, the poset of commutative subalgebras of a fixed opera- tor algebra has received attention, see e.g. [51, 52, 85]. Finally, tracing back to von Neumann himself, the lattice of projections (representing sharp predicates) and the poset of effects (fuzzy predicates), has have been studied by many, see e.g. [29]. IXa About the cover One of the most familiar structures in operator algebras (the mathematical field of this thesis) is perhaps the infinite-dimensional Hilbert space ℓ2(N). What would its unit sphere look like? The cover depicts an ap- proximation: a 12-dimensional spherical shell split into two. On the outside the pieces are completely reflective -- on the inside they are dark blue and only mostly reflective. The pieces are sat between two (11-dimensional) hyperplanes with a somewhat reflective checkerboard pattern. The planes are cropped at a certain radius from the origin turning them into hyperdisks. The picture is ren- dered using a homebrew¶ path tracer, which simulates how rays of light would bounce to end up in the camera. The calculations involved (such as the one to determine how a ray bounces from an n-dimensional sphere) are all conve- niently expressed and performed using inner products. The primary rays cast by the camera have the biggest presence in the first three dimensions, only ten percent in the fourth, five percent in the fifth et cetera. This causes more higher- dimensional oddities to appear in the image where rays have travelled further or bounced farther in the higher dimensions. X Funding My research has been supported by the European Research Council under grant agreement № 320571. XI Acknowledgments Progress in theoretical fields is often dramatized as the fruit ¶Source code is available at https://github.com/westerbaan/ndpt. of personal struggles of einzelgangers grown in deep focus and isolation. But what is missing in this depiction is that often the seeds were sown in casual con- versation among colleagues. I would like to acknowledge two seeds I'm aware of. First, it was Sam Staton who suggested looking at a universal property for instruments, which led to quotients and comprehension. Secondly, Chris Heunen insisted, against my initial skepticism, that Stinespring's dilation must have a universal property. I've also had the immense pleasure of discussing re- search with Robin Adams, Guillaume Allais, Henk Barendregt, Kenta Cho, Joan Daemen, Robert Furber, Bart Jacobs, Bas Joosten, Martti Karvonen, Aleks Kissinger, Mike Koeman, Hans Maassen, Joshua Moerman, Teun van Nuland, Dusko Pavlovic, Arnoud van Rooij, Frank Roumen, Marc Schoolderman, Peter Schwabe, Peter Selinger, Alexandra Silva, Michael Skeide, Pawel Sobocinski, Sean Tull, Sander Uijlen, Marloes Venema, Abraham Westerbaan, John van de Wetering and Fabio Zanasi. I'm particularly grateful to Arnoud, Abraham and John for proofreading this thesis. I'm honored to have been received by Dusko for a two-month research visit, on which he, Jason, Jon, Mariam, Liang-Ting, Marine, Muzamil, Pawel and Peter-Michael made me feel at home on the other side of the planet. On this side of the planet, it is hard to imagine a friendlier group of colleagues. I'm much obliged to Marko van Eekelen, Wouter Geraedts, Engelbert Hubbers, Carlo Meijer and Joeri de Ruiter with whom I've performed security audits as a side-gig, which allowed for an extension of my position. I'm grateful to my supervisor, Bart Jacobs, who showed me a realistic optimism, which is in such short supply these days. I'm touched by the patience of my friends, (extended) family and Annelies, during the final stretch of writing this thesis. Last, but not least, I would like to share my deep appreciation for the (un- der)graduate education I've been fortunate enough to receive at the Radboud Universiteit. Only now, a few years later, I can see how rare it is, not only for its uncompromised treatment of the topics at hand, but also for the personal attention to each student. Of all the great professors, Henk Barendregt, Mai Gehrke, Arnoud van Rooij and Wim Veldman have been of particular personal influence to me -- I owe them a lot. ..134 13 Chapter 2 Dilations In this chapter we will study dilations. The common theme is that a complicated map is actually the composition of a simpler map after a representation into a larger algebra. The famous Gel'fand -- Naimark -- Segal construction (see 30 VI in [124]) is a dilation theorem in disguise: it shows that every state is a vector state on a larger algebra: Theorem (GNS') For each pu-map∗ ω : A → C from a C∗-algebra A , there is a Hilbert space H , an miu-map : A → B(H ) and a vector x ∈ H such that ω = h◦ where h : B(H ) → C is given by h(T ) ≡ hx, T xi. Probably the most famous dilation theorem is that of Stinespring [106, thm. 1]. (We will see a detailed proof later, in 137 II.) Theorem (Stinespring) For every cp-map ϕ : A → B(H ) between a C∗- algebra A and the C∗-algebra of bounded operators on a Hilbert space H , there is a Hilbert space K , an miu-map : A → B(K ) and a bounded opera- tor V : H → K such that ϕ = adV ◦ , where adV : B(K ) → B(H ) is given by adV (T ) ≡ V ∗T V . Furthermore 1. if ϕ is unital, then V is an isometry and 2. if A is a von Neumann algebra and ϕ is normal, then is normal as well. 135 II III IV Stinespring's theorem is fundamental in the study of quantum information and quantum computing: it is used to prove entropy inequalities (e.g. [84]), bounds on optimal cloners (e.g. [123]), full completeness of quantum programming lan- guages (e.g. [105]), security of quantum key distribution (e.g. [76,80]), to analyze V ∗pu-map is an abbreviation for positive unital linear map. See 10 II and 44 XV for other abbreviations, like nmiu-map and ncpsu-map. 135.. 15 quantum alternation (e.g. [13]), to categorify quantum processes (e.g. [102]) and as an axiom to single out quantum theory among information processing the- ories [18]. These are only a few examples of the many thousands building on Stinespring's discovery. Stinespring's theorem only applies to maps of the form A → B(H ) and so do most of its useful consequences. One wonders: is there an extension of Stine- spring's theorem to arbitrary ncp-maps A → B? A different and perhaps less frequently asked question is whether Stinespring's dilation is categorical in some way? That is: does it have a defining universal property? Both questions turn out to have an affirmative answer: Paschke's generalization of GNS for Hilbert C∗-modules [93] turns out to have the same universal property as Stinespring's dilation and so it extends Stinespring to arbitrary ncp-maps. Dilation (normal) GNS Maps (n)p-maps LHS (n)miu [36], 30 VI, II A ϕ −→ C RHS ncp vector state −−−−−→ C −→ B(K ) B(K ) hx,( )xi A (Hilbert space K ; von Neumann algebra A ; x ∈ K ) (n)cp-maps pure ncp-map (n)miu ϕ A adV−−→ B(H ) −→ B(H ) A −→ B(K ) B(K ) ϕ A (Hilbert spaces H , K ; vN alg. A ; V ∈ B(H , K )) ncp-maps −→ B (Von Neumann algebras A , B, P) cp-maps pure ncp-map h−→ B −→ P nmiu P A ϕ miu −→ Ba(E) A cp-map −→ Ba(F ) Ba(F ) A (C∗-algs A , B; Hilb. B-modules E, F ; T ∈ Ba(F, E)) adT−−→ Ba(E) (normal) Stinespring [106], IV Paschke [93, 127], 154 III KSGNS [75], cf. [81] Table 2.1: Various dilation theorems. Maps ϕ of the given type split as ϕ = h◦ , where is as in LHS and h is as in RHS. The first three appear prominently in this thesis; the last, KSGNS, is only briefly touched upon in 155. VI We start this chapter with a detailed proof of Stinespring's theorem. Then we show that it obeys a universal property. Before we move on to Paschke's GNS, we need to develop Paschke's theory of self-dual Hilbert C∗-modules. Paschke's work builds on Sakai's characterization of von Neumann algebras, which would take considerable effort to develop in detail. Thus to be as self-contained as possible, we avoid Sakai's characterization (in contrast to our article [127]) and give new proofs of Paschke's results where required. One major difference is that we will use a completion (see 150 II) of a uniform space instead of considering the dual space of a Hilbert C∗-module. This uniformity (which we call the ultranorm uniformity) plays a central role in our development of the theory and many of the new results. We finish the first part of this chapter by constructing Paschke's dilation and establishing that it indeed extends Stinespring's dilation. We start the second part of this chapter by characterizing when the representation in the Paschke dilation of an ncp-map is injective (see 156 II), which is a generalization of our answer [125] to the same question about the Stinespring dilation. Then we allow for an intermezzo to prove several new results on (self-dual) Hilbert C∗-modules, notably: 1. we generalize the Kaplansky density theorem to Hilbert C∗-modules, see 158 II; 2. we prove that every self-dual Hilbert C∗-module over a von Neumann algebra factor is of a particularly nice form, see 162 IV, and 3. we greatly simplify the construction of the exterior tensor product for Hilbert C∗-modules over von Neumann algebras in 164 II and show that its self-dual completion is particularly well-behaved, see e.g. 165 VI. We use this new insight into the exterior tensor product to compute the tensor product of dilations, see 167 I. In the final sections of this chapter we turn our attention to the maps that occur on the right-hand side of the Paschke dilation. In Stinespring's dilation these maps are always of the form adV , but in Paschke's dilation different maps may appear -- in any case, they are pure in the sense of 100 I. The main result of these last sections is an equivalence between this and another seemingly unrelated notion of purity: we show that an ncp-map ϕ is pure in the sense of 100 I if and only if the map on the left-hand side of the Paschke dilation of ϕ is surjective. This is a bridge to the next chapter in which we prove in greater generality that these pure maps have a †-structure, see 215 III. 2.1 Stinespring's theorem In the proof of Stinespring's theorem (137 II), it will be necessary to "com- plete"† a (complex) vector space with inner product into a Hilbert space. The details of this classical result were only sketched 30 I, where it's used in the GNS-construction (30 VI). Here we will work through the details as the corre- sponding completion (150 II) required in Paschke's dilation is more complex and its exposition will benefit from this familiar analog. 136 †Note that we do not require an inner product to be definite. The inner product on a Hilbert space is definite. Thus the completion will quotient out those vectors with zero norm with respect to the inner product. ..135 -- 136.. 17 II III Proposition Let V be a (complex) vector space with inner product [· , · ]. There is a Hilbert space H together with a bounded linear map η : V → H such that 1. [v, w] = hη(v), η(w)i for all v, w ∈ V and 2. the image of η is dense in H . Proof We will form H from the set of Cauchy sequences from V with a little twist. Recall that two Cauchy sequences (vn)n and (wn)n in V are said to be equivalent if for every ε > 0 there is a n0 such that kvn− wnk 6 ε for all n > n0, where as usual kvk ≡p[v, v]. Call a Cauchy sequence (vn)n fast if for each n0 we have kvn − vmk 6 1 2n0 for all n, m > n0. Clearly every Cauchy sequence has a fast subsequence which is (as all subsequences are) equivalent with it. IV Define H to be the set of fast Cauchy sequences modulo equivalence.‡ For brevity we will denote an element of H , which is an equivalence class of fast Cauchy sequences, simply by a single representative. Also, we often tersely write v for the Cauchy sequence (vn)n. The set H is a metric space with the standard distance d(v, w) ≡ limn→∞ kvn − wnk. It is independent of represen- tatives as by the reverse and regular triangle inequality combined, we get (cid:12)(cid:12)kvn − wnk − kv′n − w′nk(cid:12)(cid:12) 6 kvn − v′nk + kwn − w′nk → 0 for v, w equivalent with a Cauchy sequence v′, respectively w′. To show H is a complete metric space, assume v1, v2, . . . is a fast Cauchy sequence of fast Cauchy sequences. We will show (vn n)n is a Cauchy sequence. Here the restric- tion to fast sequences bears fruit: without it, the limit sequence is (notationally) harder to construct. Assume n, m, k > N for some N . We have n − vm k − vm k k + k − vm k k + kvm k − vm mk 6 kvn 2 kk + kvn kvn mk 6 kvn n − vn 2N . k k = d(vn, vm) 6 2−N we can find a k > N such Because limk→∞ kvn k − vm k k 6 2 n − vm 2N and so kvn that kvn n)n is a Cauchy n)n with respect to d, sequence. with one caveat: the sequence (vn n)n might not be fast and so, might not be in H . Like any other Cauchy sequence, (vn n)n has a fast subsequence. The equivalence class of this subsequence is the limit of v1, v2, . . . in H . We have shown H is a complete metric space. It's easily checked v1, v2, . . . converges to (vn 2N . Thus (vn mk 6 4 k − vm V Define η : V → H to be the map which sends v to the equivalence class of the constant sequence (v)n. By construction of η and H , the image of η is dense in H . Let f : V → X be any uniformly continuous map to a complete metric space X. We will show there is a unique continuous g : H → X such that g ◦ η = f . From the uniform continuity it easily follows that f maps Cauchy sequences to Cauchy sequences and preserves equivalence between them. Together with the completeness of X there is a unique g : H → X fixed ‡We use fast Cauchy sequences in the definition of H only to shorten the text: any Cauchy sequence is equivalent to a fast one anyway. by g((vn)n) = limn→∞ f (vn). Clearly g ◦ η = f . Finally, uniqueness of g follows readily from the fact g is fixed on the image of η, which is dense in H . Thus we directly get a scalar multiplication on H by extending η ◦ rz : V → H where rz(v) = zv is scalar multiplication by z ∈ C, which is uniformly continuous. In fact z(vn)n ≡ (zvn)n. Extending addition is more involved, but straightforward: given representatives v, w ∈ H we know (vn + wn)n is Cauchy in V by uniform continuity of addition and so picking a fast subsequence v + w of (vn + wn)n fixes an addition on H . With this expression for addition, and the similar one for scalar multiplication, it is easy to see they turn H into a complex vector space with zero η(0) for which η is linear. Extending the inner product is bit trickier. To define the inner product on H , first note that for Cauchy sequences v and w in V we have (using Cauchy -- Schwarz, 32 VI, on the second line): VI VII (cid:12)(cid:12)[vn, wn] − [vm, wm](cid:12)(cid:12) = (cid:12)(cid:12)[vn, wn − wm] + [vn − vm, wm](cid:12)(cid:12) 6 kvnkkwn − wmk + kvn − vmkkwmk. similar fashion we see(cid:12)(cid:12)[vn, wn]−[v′n, w′n](cid:12)(cid:12) 6 kvnkkwn−w′nk+kvn−v′nkkw′nk → 0 Thus as (kvnk)n and (kwnk)n are bounded we see ([vn, wn])n is Cauchy. In a as n → ∞ for v′ and w′ Cauchy sequences equivalent to v respectively w. Thus ([vn, wn])n is equivalent to ([v′n, w′n])n and so we define hv, wi ≡ limn→∞[vn, wn] on H . With vector space structure from VI we easily see this is an inner product H . The metric induced by the inner product coincides with d: hv − w, v − wi = lim n [vn − wn, vn − wn] = lim n kvn − wnk2 = d(v, w)2. And so H is a Hilbert space. Finally, hη(v), η(w)i = [v, w] is direct. (cid:3) We are ready to prove Stinespring's dilation theorem. Proof of Stinespring's theorem, 135 IV Let ϕ : A → B(H ) be any cp-map. Write A ⊙ H for the tensor product of A and H as vector spaces. We will define K as the Hilbert space completion of A ⊙ H with respect to the inner product [· , · ] fixed by 137 II [a ⊗ x, b ⊗ y] ≡ hx, ϕ(a∗b) yiH . (2.1) To proceed, we have to show that [· , · ] is indeed an inner product. By linear extension, (2.1) fixes a sesquilinear form [· , · ]. As ϕ preserves involution as a positive map, see 10 IV, the sesquilinear form [· , · ] is also symmetric, i.e. [t, s] = i=1 ai ⊗ xi is an arbitrary element of A ⊙ H . To see [· , · ] is positive, we need to show ai ⊗ xi,Xj aj ⊗ xj(cid:3) ≡ Xi,j ..136 -- 137.. hxi, ϕ(a∗i aj)xji . 19 [s, t]. AssumePn 0 6 (cid:2)Xi The trick is to consider the matrix algebra MnB(H ) acting on H ⊕n in the (In fact, this yields an isomor- obvious way: (A(x1, . . . , xn)t)i ≡ Pj Aij xj. phism MnB(H ) ∼= B(H ⊕n).) Writing x for (x1, . . . , xn)t in H ⊕n, we get Xi,j hxi, ϕ(a∗i aj) xji = (cid:10)x, (Mnϕ)(cid:0) (a∗i aj)ij(cid:1) x(cid:11)H ⊕n . (2.2) The matrix (a∗i aj)ij is positive as (C∗C)ij = a∗i aj with Cij ≡ 1√n aj. By com- plete positivity Mnϕ( (a∗i aj)ij ) is positive and so is (2.2), hence [· , · ] is an inner product. Write η : A ⊙ H → K for the Hilbert space completion of A ⊙ H with the inner product described in 136 II. III We need a lemma before we continue. Let T : A ⊙ H → A ⊙ H be a bounded linear map. We show there is an extension T : K → K . The operator T is uniformly continuous and so is η ◦ T : A ⊙ H → K . Thus by 136 V there is a unique continuous extension T : K → K with T (η(t)) = η(T (t)) for all t ∈ A ⊙ H . Clearly T is linear on the image of η, which is dense and so T is linear everywhere. As T is linear and continuous, it must be bounded, so T ∈ B(K ). It is easy to see \T + S = T + S and cλT = λ T for operators S, T on A ⊙ H and λ ∈ C. Also cT S = T S: indeed T Sη(t) = T η(St) = η(T St) and so by uniqueness T S = cT S. IV Assume b ∈ A . Let 0(b) be the operator on A ⊙ H fixed by 0(b) a ⊗ x ≡ (ba) ⊗ x. Clearly 0 is linear, unital and multiplicative. We want to show 0(b) is bounded for fixed b ∈ A , so that we can define (b) ≡ [0(b) ∈ B(K ) using III. To show 0(b) is bounded, we claim that in MnB(H ) we have (a∗i b∗baj)ij 6 kbk2(a∗i aj)ij . (2.3) Indeed, as b∗b 6 kbk2, there is some c with c∗c = kbk2−b∗b. Define Cij ≡ 1√n caj. We compute: (C∗C)ij = a∗i c∗caj = a∗i (kbk2 − b∗b)aj and so (2.3) holds. Hence and so 0(b) is bounded. Define (b) : A → B(H ) using III by (b) ≡ [0(b). (cid:13)(cid:13)0(b)Xi ai ⊗ xi(cid:13)(cid:13)2 = (cid:13)(cid:13)Xi = (cid:10)x, (Mnϕ)( (a∗i b∗baj)ij ) x(cid:11) 6 kbk2(cid:10)x, (Mnϕ)( (a∗i aj)ij ) x(cid:11) = kbk2(cid:13)(cid:13)Xi (bai) ⊗ xi(cid:13)(cid:13)2 ai ⊗ xi(cid:13)(cid:13)2 by (2.3) We already know is a mu-map. We want to show it preserves involution: (c∗) = (c)∗ for all c ∈ A . Indeed, for every a, b ∈ A and x, y ∈ H we have [0(c∗) a ⊗ x, b ⊗ y] = [(c∗a) ⊗ x, b ⊗ y] = hx, ϕ(a∗cb)yi = [a⊗ x, 0(c) b ⊗ y]. Hence h(c∗)η(a ⊗ x), η(b ⊗ y)i = hη(a ⊗ x), (c)η(b ⊗ y)i. As the linear span of vectors of the form η(a ⊗ x) is dense in K , we see (c∗) is adjoint to (c) and so (c∗) = (c)∗, as desired. Let V0 : H → A ⊙ H be given by V0x = 1 ⊗ x. This operator V0 is bounded: 1 Define V ≡ η ◦ V0. For all a ∈ A and x, y ∈ H , we have kV0xk = hx, ϕ(1∗1)xi 2 = kpϕ(1)xk 6 kpϕ(1)kkxk. [a ⊗ x, V0y] = hx, ϕ(a∗)yi = hx, ϕ(a)∗yi = hϕ(a)x, yi and so V ∗ satisfies V ∗η(a ⊗ x) = ϕ(a)x. Hence for all a ∈ A and x ∈ H V ∗(a)V x = V ∗(a)η(1 ⊗ x) = V ∗η(a ⊗ x) = ϕ(a)x. We have shown adV ◦ = ϕ. If ϕ is unital, then V ∗V x = V ∗(1⊗x) = ϕ(1)x = x for all x ∈ H and so V is an isometry. (Or equivalently adV is unital.) Assume A is a von Neumann algebra. It remains to be shown that is normal when ϕ is. So assume ϕ is normal. To start, note that the vector function- als ha ⊗ x, (· )a ⊗ xi are together a faithful set of np-functionals. Thus, by 48 II, it is sufficient to show that b 7→ ha ⊗ x, (b) a ⊗ xi ≡ hx, ϕ(a∗ba) xi is normal for every a ∈ A and x ∈ H . This is indeed the case, as normal maps are closed under composition and hx, (· )xi is normal by 37 IX, a∗(· )a by 44 VIII and ϕ by assumption. (cid:3) Remark Stinespring's theorem generalizes GNS. It seems tempting to directly prove Stinespring instead of GNS -- however, for our version with ncp-maps, we used GNS itself in the proof of Stinespring's theorem to show that is normal. In the following series of exercises, the reader will develop a few consequences of Stinespring's theorem which are in some fields better-known than Stinespring's theorem itself. As preparation we study the nmiu-maps between type I von Neumann algebras. The following result can be derived from the more gen- eral [110, thm. 5.5], but we will give a direct proof, which seems to be new. [130] Proposition Let : B(H ) → B(K ) be any non-zero§ nmiu-map for Hilbert spaces H and K . There is a Hilbert space K ′ and a unitary U : K → H ⊗K ′ (see 34a IV) such that (a) = U∗(a ⊗ 1)U for all a ∈ B(H ). § If K is 0-dimensional, then there zero-map B(H ) → B(K ) is an nmiu-map. ..137 -- 138.. 21 V VI VII VIII 138 II III Proof To start, we will show that is injective by proving ker = {0}. Clearly ker is a two-sided ideal -- indeed, if (a) = 0 for some a ∈ B(H ), then (ab) = (a)(b) = 0 and (ba) = 0 for any b ∈ B(H ). As is normal, we have (sup D) = supd∈D (d) = 0 for any bounded directed D ⊆ ker of self-adjoint elements. So by 69 II, we know ker = zB(H ) for some central projection z ∈ B(H ). Either z = 0 or z = 1. Clearly z 6= 1, for other- wise 0 = 6= 0, which is absurd. So z = 0 and ker = {0}, thus is injective. IV Pick an orthonormal basis (ei)i∈I of H . Write pi for the projection onto Cei, i.e. pi ≡ eiihei. Choose any i0 ∈ I and write K ′ ≡ (pi0 )K . Pick any orthonormal basis (dj)j∈J of K ′ and write qj for the projection onto Cdj, i.e. qj ≡ djihdj. Let ui denote the unique partial isometry fixed by uiei = ei0 and uiej = 0 for j 6= i -- viz. ui ≡ ei0ihei. Define rij ≡ (u∗i )dj . We will show (rij )(i,j)∈I×J is an orthonormal basis of K . For this, it is sufficient to as supremum over the partial sums), using that then rijihrij are pairwise orthogonal projections by 55 XIII. Clearly these rij are unit vectors, as by uiu∗i = pi0 and (pi0 )dj = dj, we have krijk2 = h(uiu∗i )dj, dji = 1. To show 1 = show rij are unit vectors and Pi,j rijihrij = 1 (where the sum is defined Pi,j rijihrij, we compute pi(cid:1) as ei is an orthn. basis as is unital 1 = (1) as is normal (pi) (u∗i pi0 ui) = (cid:0)Xi = Xi = Xi = Xi = Xi = Xi,j = Xi,j by dfn. ui as is mi-map as dj is an orthn. basis of K ′ by 44 VIII (ui)∗(pi0 )(ui) (ui)∗(cid:0)Xj qj(cid:1) (ui) (ui)∗ qj (ui) rijihrij . V Let U : K → H ⊗ K ′ denote the unitary fixed by U rij = ei ⊗ dj. For any i ∈ I, j ∈ J and a ∈ B(H ), it is straightforward to show that aeiihei0 = Pkhek, aeiiekihei0, where the sum converges ultrastrongly and consequently by dfn. rij and ui (a) rij = (a)(eiihei0) dj = (aeiihei0) dj = (cid:0)Xk = (cid:0)Xk = Xk = Xk = U∗(cid:0)Xk hek, aeiiekihei0(cid:1) dj hek, aeii (ekihei0)(cid:1) dj hek, aeii (ekihei0) dj hek, aeii U∗ ek ⊗ dj ekhek, aeii(cid:1) ⊗ dj = U∗ (aei) ⊗ dj = U∗ (a ⊗ 1) U rij as is us-cont. by 44 XV by dfn. U and uk as ek is an orthn. basis by dfn. rij . We have shown (a) = U∗(a ⊗ 1)U . (cid:3) Corollary The nmiu-isomorphisms B(H ) → B(K ) are precisely of the form adU for some unitary U : K → H . Exercise Use II and 135 IV to show that for every ncp-map ϕ : B(H ) → B(K ) there are a Hilbert space K ′ and a bounded operator V : K → H ⊗ K ′ such that ϕ(a) = V ∗(a ⊗ 1)V . Conclude that any quantum channel Φ from H to itself (i.e. completely positive trace-preserving linear map between trace-class operators over H , see e.g. [112, §2.6.2]) is of the form Φ() = trK ′ [U∗⊗v0ihv0 U ] for some unitary U . Exercise* (Kraus' decomposition) Let ϕ : B(H ) → B(K ) be any ncp-map. By VII there is a Hilbert space K ′ and bounded operator V : K → H ⊗ K ′ such that ϕ(a) = V ∗(a ⊗ 1)V . Choose an orthonormal basis (ei)i∈I of K ′. Show ϕ(a) =Pi∈I V ∗(a⊗eiihei)V , where the sum converges ultraweakly. De- duce there are projections Pi : H ⊗ K ′ → H with ϕ(a) =Pi∈I V ∗P ∗i aPiV . Conclude all ncp-maps ϕ : B(H ) → B(K ) are of the form ϕ(a) =Pi∈I V ∗i aVi for some bounded operators Vi : K → H , for which the partial sums ofPi V ∗i Vi are bounded. These Vi are called Kraus operators for ϕ. (For a different ap- proach, see e.g. [28, thm. 2.3].) VI VII VIII Now assume H and K are finite dimensional.¶ Show that in this case the total number of Kraus operators can be chosen to be less than or equal to (dim H ) · (dim K ). (There is a similar fact for infinite-dimensional H and K , where the product of numbers is replaced by a maximum of cardinals.) Definition Let ϕ : A → B(H ) be any ncp-map for some Hilbert space H ¶For arbitrary H and K , the cardinality of the set of Kraus operators can be chosen to be less than or equal to (dim H ) · (dim K ), with dimension understood as the cardinality of an orthonormal basis and the usual product on cardinals. 139 ..138 -- 139.. 23 and von Neumann algebra A . A normal Stinespring dilation (cf. [94, ch. 4]) is a triple (K , , V ) with K a Hilbert space, : A → B(K ) an nmiu-map (i.e. normal representation) and V : H → K a bounded operator such that ϕ decomposes as ϕ = adV ◦ . We just saw every ϕ has (at least one) normal Stinespring dilation. A (normal) Stinespring dilation (K , , V ) is said to be minimal if the linear span of (A )V H ≡ { (a)V x; a ∈ A , x ∈ H } is dense in K . By construction the (normal) Stinespring dilation from 137 II is minimal. Every (normal) Stinespring dilation (K , , V ) of ϕ restricts to a minimal dilation (K ′, , V ) for ϕ, where K ′ ⊆ K is the norm-closure of the linear span of (A )V H . It is a well-known fact that all minimal normal Stinespring dilations for a fixed ϕ are unitarily equivalent, see for instance [94, prop. 4.2]. We will adapt its proof to show that a minimal normal Stinespring dilation admits a universal property. Later we will use this to prove a second universal property 140 III that allows us to generalize Stinespring's dilation to arbitrary ncp-maps. The modification of the familiar argument is mostly straightforward, except for the following lemma, which we published earlier as [127, lem. 11]. II III A ❆❆❆❆❆❆❆❆ ′ B σ C Lemma Let : A → B and ′ : A → C be nmiu-maps be- tween von Neumann algebras, and let σ : C → B be an ncp- map such that σ ◦ ′ = . Then σ(cid:0)′(a1) c ′(a2)(cid:1) = (a1) σ(c) (a2) for all a1, a2 ∈ A and c ∈ C . IV Proof By Theorem 3.1 of [24] (see 34 XVIII), we know that for all c, d ∈ C : σ(d∗d) = σ(d)∗σ(d) =⇒ σ(cd) = σ(c)σ(d). We can apply this to ′(a)∗′(a) with a ∈ A -- indeed we have σ(′(a)∗′(a)) = σ(′(a∗a)) = (a∗a) = (a)∗(a) = σ(′(a))∗σ(′(a)) (2.4) and so by (2.4), we have σ(c′(a)) = σ(c)σ(′(a)) = σ(c)(a) for all c ∈ C . Thus also (taking adjoints): σ(′(a)c) = (a)σ(c) for all c ∈ C . Hence σ(′(a1)c′(a2)) = (a1)σ(c′(a2)) = (a1)σ(c)(a2) for all a1, a2 ∈ A and c ∈ C as desired. (cid:3) V Proposition Assume ϕ : A → B(H ) is any ncp-map with normal Stinespring dilations (K , , V ) and (K ′, ′, V ′). If (K , , V ) is minimal, then there is a unique isometry S : K → K ′ such that SV = V ′ and = adS ◦ ′. / / O O Proof (This is the same proof we gave in [127].) First we will deal with a pathological case. If V = 0, then ϕ = 0, V ′ = 0, K = {0} and = 0. Thus the unique linear map S : {0} → K ′ satisfies the requirements. Assume V 6= 0. (Uniqueness) Let S1, S2 : K → K ′ be any isometries for which SkV = V ′ and adSk ◦ ′ = for k = 1, 2. We want to show S1 = S2. First we will prove that adS1 = adS2. For k = 1, 2, n, m ∈ N, a1, . . . , an, α1, . . . , αm ∈ A , x1, . . . , xn, y1, . . . , ym ∈ H and c ∈ B(K ′), we have (cid:10)adSk (c)Xi (ai)V xi,Xj III (αj )V yj(cid:11) = Xi,j(cid:10)V ∗(α∗j ) adSk (c) (ai)V xi, yi(cid:11) = Xi,j(cid:10)V ∗adSk(cid:0)′(α∗j ) c ′(ai)(cid:1)V xi, yj(cid:11) = Xi,j(cid:10)(V ′)∗′(α∗j ) c ′(ai)V ′xi, yj(cid:11). As the linear span of (A )V H is dense in K , we see adS1(c) = adS2(c) for all c ∈ B(K ′). So adS1 = adS2 and thus λS1 = S2 for some λ ∈ C, cf. [126, lem. 9]. As V 6= 0, there is an x ∈ H with V x 6= 0. Also S1V x 6= 0 as S1 is an isometry and thus injective. From this and S1V x = V ′x = S2V x = λS1V x, we get λ = 1. Hence S1 = S2. (Existence) For any n ∈ N, a1, . . . , an ∈ A and x1, . . . , xn ∈ H , we have VI VII VIII It follows (again using denseness of the linear span of (A )V H ), that there is a unique isometry S : K → K ′ such that S(a)V x = ′(a)V ′x for all a ∈ A and x ∈ H . As SV x = S(1)V x = ′(1)V ′x = V ′x for all x ∈ H , we have SV = V ′. Furthermore we have (cid:13)(cid:13)Xi (ai)V xi(cid:13)(cid:13)2 = Xi,j(cid:10)V ∗(a∗j ai)V xi, xj(cid:11) = Xi,j(cid:10)ϕ(a∗j ai)xi, xj(cid:11) = (cid:13)(cid:13)Xi ′(ai)V ′xi(cid:13)(cid:13)2 . S(a)Xi S(aai)V xi ′(aai)V ′xi (ai)V xi = Xi = Xi = ′(a)Xi = ′(a)SXi ′(ai)V ′xi (ai)V xi. ..139.. 25 for all n ∈ N, a, a1, . . . , an ∈ A and x1, . . . , xn ∈ H . So S(a) = ′(a)S. Thus S∗′(a)S = S∗S(a) = (a) for all a ∈ A , whence adS ◦ ′ = . (cid:3) IX Exercise* The statement of the universal property V for Stinespring was pro- In this note he shows it is posed by Chris Heunen in an unpublished note. equivalent to the existence of a left-adjoint, as we will see in this exercise. Let Repcp denote the category with as objects ncpu-maps ϕ : A → B(H ) and as arrows between ϕ : A → B(H ) and ϕ′ : A ′ → B(H ′) pairs (m, S) where m : A → A ′ is an nmiu-map and S : H → H ′ is an isometry such that ϕ = adS ◦ ϕ′ ◦ m. Write Rep for the full subcategory of Repcp consisting of objects ϕ : A → B(H ) that are nmiu. Show using V that the inclusion functor U : Rep → Repcp has a left adjoint. IXa Remarks Recently there have been two interesting categorical discoveries con- cerning Stinespring's dilation theorem, which deserve a mention. Parzygnat found a rather different characterization of Stinespring's dilation theorem as an adjunction [92]. Secondly, Huot and Staton discovered that the category of completely positive maps between matrix algebras is the universal monoidal category with terminal unit with a functor from the category of matrix algebras with isometries between them. [54] This fact is surprisingly closely related to Stinespring's dilation theorem for matrix algebras. In the special case of ncp-maps of the form ϕ : B(H ) → B(K ), there is a different and noteworthy property concerning dilations 'of the same dimension' called essential uniqueness of purification. This property is the cornerstone of several publications, notably [18], which popularized it. X XI Exercise* (Essential uniqueness of purification) Let ϕ : B(H ) → B(K ) be any ncp-map. By 138 VII here is a Hilbert space K ′ and bounded opera- tor V : K → H ⊗K ′ such that V ∗(a⊗1)V = ϕ(a). Assume there is a bounded operator W : K → H ⊗ K ′ with W ∗(a ⊗ 1)W = ϕ(a) as well. We will show there is a unitary U : K ′ → K ′ such that V = (1 ⊗ U )W . First show that in the special case the dilations are minimal (i.e. the linear span of {(a ⊗ 1)V x; a ∈ B(H ), x ∈ K } is dense in H ⊗ K ′), there is a (unique) unitary U0 on H ⊗ K ′ with V = U0W and (a ⊗ 1)U0 = U0(a ⊗ 1) for all a ∈ B(H ). Derive from the latter that for each y ∈ K ′ and unit- vector e ∈ H , there is a y′ ∈ K ′ with U0(e⊗ y) = e⊗ y′. Conclude U0 = 1⊗ U for some unitary U on K ′. Show that the general case follows from the special case in which the dilations are minimal. 140 We are ready to state the more general universal property that Stinespring's dilation admits, which will allow us to generalize it to arbitrary ncp-maps. Definition Let ϕ : A → B be any ncp-map between von Neumann algebras. A Paschke dilation of ϕ is a triple (P, , h) with a von Neumann algebra P; an nmiu- map : A → P and ncp-map h : P → B with ϕ = h◦ and the following universal property holds. <③③③③ h A "❊❊❊❊ " ′ ϕ P σO P′ B II h′ For every triple (P′, ′, h′) with a von Neumann algebra P′; an nmiu-map ′ : A → P′ and ncp-map h′ : P′ → B with ϕ = h′ ◦ ′, there is a unique ncp-map σ : P′ → P (the "mediating map") with σ ◦ ′ = and h◦ σ = h′. Theorem Let ϕ : A → B(H ) be any ncp-map together with a minimal normal Stinespring dilation (K , , V ). Then (B(K ), , adV ) is a Paschke dilation of ϕ. Proof (This is essentially the same proof we gave in [127].) III IV ′ ad V A ϕ adV / B(H ) "❊❊❊❊❊❊❊❊❊ Let P′ be any von Neumann algebra to- gether with nmiu-map ′ : A → P′ and ncp- map h′ : P′ → B(H ) for which it holds that h′ ◦ ′ = ϕ. We must show that there is a unique ncp-map σ : P′ → B(K ) such that σ ◦ ′ = and adV ◦ σ = h′. First we will prove uniqueness of this map σ. Then we will show that such a σ exists by appealing to the universal property 139 V. (Uniqueness) Let σ1, σ2 : P′ → B(H ) be two such ncp-maps. Similarly to the uniqueness proof in 139 V, we compute that for all n, m ∈ N, a1, . . . , an, α1, . . . , αm ∈ A , x1, . . . , xn, y1 . . . , ym ∈ H and c ∈ P′, we have C✞✞✞✞✞✞✞✞ :ttttttttt d■■■■ ✞✞✞✞✞✞✞✞ / B( K ) adS■■■ B(K ) h′ P′ (cid:10)σk(c)Xi (ai)V xi,Xj (aj)V xj(cid:11) = Xi,j(cid:10)V ∗(α∗j )σk(c)(ai)V xi, yi(cid:11) 139 III= Xi,j(cid:10)V ∗σk(cid:0)′(α∗j )c′(ai)(cid:1)V xi, yj(cid:11) = Xi,j(cid:10)h′(′(α∗j )c′(ai))xi, yj(cid:11). formPi (ai)V xi is norm dense in K and so σ1(c) = σ2(c) as desired. (Existence) Let ( K , , V ) be a minimal normal Stinespring dilation of h′. The triple ( K , ◦ ′, V ) is a normal Stinespring dilation of ϕ. Thus by 139 V there is a (unique) isometry S : K → K with SV = V and adS ◦ ◦ ′ = . De- fine σ = adS ◦ . Clearly σ ◦ ′ = adS ◦ ◦ ′ = and adV ◦ σ = adV ◦ adS ◦ = adSV ◦ = ad V ◦ = h′ as desired. (cid:3) By definition of minimality of a Stinespring dilation, the set of vectors of the V VI ..139 -- 140.. 27 / / ) ) < S S O " / : C / O O d VII The converse holds as well: by VIII every Paschke dilation of a map ϕ : A → B(H ) is up to a unique nmiu-isomorphism a minimal Stinespring dilation. Later we will show that every ncp-map ϕ : A → B has a Paschke dilation: it turns out that Paschke's generalization of GNS to Hilbert C∗-modules fits the bill. Before we show this, we look at some other examples and basic properties of Paschke dilations. First, as Paschke dilations are defined by a simple universal property, they are unique up to a unique nmiu-isomorphism: VIII Lemma If (P1, 1, h1) and (P2, 2, h2) are two Paschke dilations for the same ncp-map ϕ : A → B, then there is a unique (nmiu) isomorphism ϑ : P1 → P2 with ϑ◦ 1 = 2 and h2 ◦ ϑ = h1. IX Proof There are unique mediating maps σ1 : P1 → P2 and σ2 : P2 → P1. It is easy to see σ1 ◦ σ2 satisfies the same defining property as the unique mediating map id : P2 → P2 and so σ1 ◦ σ2 = id. Similarly σ2 ◦ σ1 = id. Define ϑ = σ1. We just saw ϑ is an ncp-isomorphism. Note ϑ(1) = ϑ(1(1)) = 2(1) = 1 and so ϑ is unital. But then ϑ is an nmiu-isomorphism by 99 IX. (cid:3) X Exercise The following can be shown using only the defining universal property. 1. Show (A , , id) is a Paschke dilation of an nmiu-map : A → B. 2. Let (P, , h) be any Paschke dilation. Prove (P, id, h) is a Paschke dila- tion for h. (Hint: use id : P → P is the unique ncp-map with id◦ = and h◦ id = h.) 3. Let ( ϕ1 is a Paschke dilation of ( ϕ1 for i = 1, 2. ϕ2 ) : A → B1 ⊕ B2 be an ncp-map. Show (P1 ⊕ P2, ( 1 2 ) , h1 ⊕ h2) ϕ2 ) if (Pi, i, hi) is a Paschke dilation of ϕi 4. Let ϕ : A → B be any ncp-map with Paschke dilation (P, , h) and λ ∈ R with λ > 0. Prove (P, , λh) is a Paschke dilation of λϕ. XI Examples Later we will be able to compute more dilations. 1. In 167 I, we will see that if (Pi, i, hi) is a Paschke dilation of an ncp- map ϕi : Ai → Bi for i = 1, 2, then (P1 ⊗ P2, 1 ⊗ 2, h1 ⊗ h2) is a Paschke dilation of ϕ1 ⊗ ϕ2. 2. In 171 II, we will show that (⌈⌈p⌉⌉ A , h⌈⌈p⌉⌉, hp) is a Paschke dilation of the ncp-map h : A → pA p given by h(a) = pap for some projection p ∈ A . 3. In 171 VII, it's proven that a map ϕ with Paschke dilation (P, , h) is pure (in the sense of 100 I) if and only if is surjective. 4. In 169 XI the Paschke dilations of ϕ and c◦ ϕ for a filter c are related. 2.2 Hilbert C∗-modules 141 Recall that by GNS every state (pu-map) ω : A → C on a von Neumann alge- bra A splits as h◦ for some nmiu-map : A → B(H ), Hilbert space H and vector state h(T ) = hx, T xi for some x ∈ H . It is easy to see (H , , h) is a Paschke dilation of ω. Paschke showed that GNS can be generalized by replacing the scalars C with an arbitrary von Neumann algebra B: every 'state' ncp-map ϕ : A → B splits as h◦ for some nmiu-map : A → Ba(X), and 'vector state' h(T ) = hx, T xi, where X is a self-dual Hilbert C∗-module over B, x ∈ X and Ba(X) are the B- linear bounded operators on X. It turns out (Ba(X), , h) is a Paschke dilation of ϕ. Before we will prove this, we first develop the theory of (self-dual) Hilbert C∗-modules. We already encountered Hilbert C∗-modules in the development of the basics of completely positive maps in my twin brother's thesis [124]. (They appear in 32 I.) As Hilbert C∗-modules are central to the present topic, we will prove some basic results again, which were already covered by my twin. Definition Let B be a C∗-algebra and X be a right B-module. A B-valued inner product on X is a map h· , ·i : X × X → B such that II 1. hx, ·i : X → B is B-linear for all x ∈ X -- viz. hx, ybi = hx, yi b; 2. hx, yi∗ = hy, xi for all x, y ∈ X and 3. hx, xi > 0 for all x ∈ X. A B-valued inner product is called definite provided x = 0 whenever hx, xi = 0. A pre-Hilbert B-module X is a right B-module X together with a definite B- valued inner product. 1 It will turn out kxk = k hx, xik 2 defines a norm on a pre-Hilbert B-module X. A Hilbert B-module (also called a Hilbert C∗-module over B) is a pre- Hilbert B-module which is complete in this norm. A pre-Hilbert B-module is called self dual if for each bounded B-linear τ : X → B, there is a t ∈ X such that τ (x) = ht, xi for all x ∈ X. Beware In his original definition [93, §2.1], Paschke chose to make his B-valued inner product B-linear on the left instead of on the right. He also assumes his inner products to be definite. Examples Several examples [93] of (pre-)Hilbert C∗-modules follow, to each of which we will return in the remainder. • Hilbert C-modules are the same as Hilbert spaces. They are all self dual. ..140 -- 141.. 29 IIa IIb III • Any C∗-algebra B is a self-dual Hilbert C∗-module over itself with B- valued inner product ha, bi = a∗b. By the C∗-identity, the Hilbert B- module norm on B coincides with the norm of the C∗-algebra. • More generally: if J ⊆ B is any right ideal of B, then J is a pre-Hilbert B- module with ha, bi = a∗b. If J is closed with respect to the norm, then J is a Hilbert B-module. It might not be self dual. • Principal right-ideals are self dual: the right ideal eB for a projection e ∈ B is a self-dual Hilbert B-module with hea, ebi = a∗eb. Later we will see that every self-dual Hilbert B-module over a von Neumann algebra B is built up from such eB. • If X and Y are (pre-)Hilbert B-modules, then their direct sum X ⊕ Y (as right B-modules) is a (pre-)Hilbert B-module with the familiar inner product h(x, y), (x′, y′)i = hx, x′i + hy, y′i. If X and Y are self dual, then so is X ⊕ Y . 2.2.1 The basics of Hilbert C∗-modules 142 Perhaps the most important stones in the foundation of the theory of operator algebras are the Cauchy -- Schwarz inequality hx, yi2 6 kxk2kyk2 and the polar- 4P3 ization identity hx, yi = 1 k=0 ikkikx + yk2, see 4 XV. Also in the development of Hilbert C∗-modules, their generalizations support the bulk of the structure. II Definition Let X be a right B-module with B-valued inner product h· , ·i for some C∗-algebra B. For any positive functional f : B → C, write hx, yif = f . This h · , ·if is a complex-valued inner product f (hx, yi) and kxkf = hx, xi on X. Hence, by 4 XV, we know that k · kf is a seminorm (i.e. a not-necessarily- positive-definite norm.) 1/2 IIa We start with the Cauchy -- Schwarz inequality for B-valued inner products. The proof is the same as Paschke gives in [93, prop. 2.3 (ii)], but we use it to prove a slightly more general result. III Proposition (Cauchy -- Schwarz) Let X be a right B-module with B-valued in- ner product h· , ·i for some C∗-algebra B. Then hx, yihy, xi 6 k hy, yik hx, xi. IV Proof Let f : B → C be any state (pu-map). Since the states on B are order separating, see 22 VIII, it suffices to show f (hx, yihy, xi) 6 k hy, yikf (hx, xi). If f (hx, yihy, xi) = 0 the inequality holds trivially, so assume f (hx, yihy, xi) 6= 0. Using Cauchy -- Schwarz for h· , ·if , we derive f (hx, yihy, xi)2 = hx, y hy, xii2 f f 6 kxk2 = kxk2 6 kxk2 f ky hy, xik2 f f (hx, yihy, yihy, xi) f k hy, yik f (hx, yihy, xi), which yields the inequality by dividing by f (hx, yihy, xi). (cid:3) Exercise Let X be a right B-module with B-valued inner product h· , ·i for some C∗-algebra B. First use Cauchy -- Schwarz (III) to show k hx, yik 6 kxkkyk. 2 is a seminorm on X with kx · bk 6 kxkkbk With this, derive kxk = k hx, xik for all b ∈ B. (See [93, prop. 2.3].) From this, it follows easily that that the addition and the module action are jointly uniformly continuous and that the inner product is uniformly continuous in each argument separately. (The inner product is also jointly continuous, but not uniformly.) 1 Definition Let V be a right B-module for some C∗-algebra B. A B-sesquilinear form on V is a map B : V × V → B such that for each y ∈ V the maps x 7→ B(y, x) and x 7→ B(x, y)∗ are B-linear -- that is: V VI VII β∗ hx, yi b = h xβ , yb i for all x, y ∈ X and b, β ∈ B. Remark It seems that B-sesquilinear forms do not explicitly appear in the established literature, even though its associated polarization identity is often used implicitly. Example Let X be a pre-Hilbert B-module. For every B-linear T : X → X the map h(· ), T (· )i is a B-sesquilinear form. In 152 V we will see that on a self-dual Hilbert C∗-module all bounded B-sesquilinear forms are of this form. VIIa VIII Exercise Let B be a B-sesquilinear form. Prove the polarization identity: IX B(x, y) = 1 4 3Xk=0 ikB(ikx + y, ikx + y). Definition Let X, Y be pre-Hilbert B-modules for some C∗-algebra B. A (C- )linear map T : X → Y is said to be adjointable if there is a linear map S : Y → X such that hy, T xiY = hSy, xiX for all x ∈ X and y ∈ Y . Adjoints are unique: if S and S′ are both adjoints of T an easy computation shows h(S − S′)y, (S − S)′yi = 0 for all y ∈ Y . We write T ∗ for the adjoint of T (if it exists) and Ba(X) for the subset of (norm-)bounded operators on X which are adjointable. (See [93, §2].) 143 ..141 -- 143.. 31 Ia The second part of the following lemma is well-known, but is proven in a new II III way using the equivalence which is inspired by [93, thm. 2.8 (ii)]. Lemma For a linear map T : X → Y between pre-Hilbert B-modules X, Y and real number B > 0, the following are equivalent. 1. kT xk 6 Bkxk for all x ∈ X. 2. k hy, T xik 6 Bkykkxk for all x ∈ X and y ∈ Y . If T is bounded and adjointable, then kT ∗k = kTk and kT ∗Tk = kTk2. Proof Assume kT xk 6 Bkxk for all x ∈ X. Then by 142 V we find that for all y ∈ X we have k hy, T xik 6 kykkT xk 6 Bkykkxk, as desired. For the converse, pick x ∈ X and assume k hy, T xik 6 Bkykkxk for all y ∈ Y . Then kT xk2 = k hT x, T xik 6 BkT xkkxk. So kT xk 6 Bkxk by dividing kT xk if kT xk 6= 0 and trivially otherwise. Now assume T is adjointable and bounded. Then k hx, T ∗yik = k hy, T xik 6 kTkkykkxk for all x ∈ X, y ∈ Y and so by the previous kT ∗k 6 kTk. As adjoints are unique and T is an adjoint of T ∗ we get T ∗∗ = T and so kTk = kT ∗∗k 6 kT ∗k 6 kTk, as desired. For the final identity, note that for any x ∈ X we have kT xk2 = k hT x, T xik = k hx, T ∗T xik 6 kxkkT ∗T xk 6 kxk2kT ∗Tk and so kTk 6 kT ∗Tk 2 . Hence kTk2 6 kT ∗Tk 6 kT ∗kkTk = kTk2. 1 (cid:3) V IV Proposition Let X be a Hilbert B-module for some C∗-algebra B. With composition as multiplication, adjoint as involution and operator norm, the set of adjointable operators Ba(X) is a C∗-algebra. [93, §2] It is easy to see T ∗ + S∗ is an adjoint of T + S for T, S ∈ Ba(X) Proof and so T + S is adjointable with (T + S)∗ = T ∗ + S∗. Similarly it is easy to see Ba(X) is closed under scalar multiplication, composition and involution with (λT )∗ = λT ∗, (T S)∗ = S∗T ∗ and T ∗∗ = T . Also the identity 1 is clearly bounded and self-adjoint. So Ba(X) is a unital ∗-algebra. By II the C∗-identity It only remains to be shown that Ba(X) is complete. Let T1, T2, . . . holds. be a Cauchy sequence in Ba(X). As X is a Banach space B(X) is complete and so Tn → T for some T ∈ B(X). We have to show T is adjointable. For any n, m ∈ N we have kT ∗n−T ∗mk = k(Tn−Tm)∗k = kTn−Tmk and so T ∗1 , T ∗2 , . . . is also Cauchy and converges to, say, S ∈ B(X). So n hT ∗nx, yi = lim n hx, Tnyi = hx, T yi , T ∗n x, yi = lim hSx, yi = hlim n for any x ∈ X and y ∈ Y . Thus S = T ∗ and T ∈ Ba(X) as desired. (cid:3) 144 Proposition Let X be a Hilbert B-module for some C∗-algebra B. The vector states on Ba(X) are order separating -- that is: T > 0 if and only if hx, T xi > 0 for all x ∈ X. II III IV V VI 145 II − > 0 so T 3 Proof (Proof from [81, lem. 4.1].) Assume T ∈ Ba(X) with T > 0. Then T = S∗S for some S ∈ Ba(X) and so hx, T xi = hSx, Sxi > 0 for all x ∈ X. For the converse, assume hx, T xi > 0 for all x ∈ X. We claim T is self- k=0 ik(cid:10)ikx + y, T (ikx + y)(cid:11), 4P3 adjoint. By the polarization identity hx, T yi = 1 see 142 IX, we have hx, T yi = hT x, yi for all x, y ∈ X and so T ∗ = T . There are positive T+, T− ∈ Ba(X) with T = T+ − T− and T+T− = 0, see 24 II. By assumption 0 6 hT−x, T T−xi from which −(cid:10)x, T 3 −x(cid:11) > 0, but T 3 − = 0. Thus T− = 0 (by the functional calculus, 28 II) and so T > 0. (cid:3) Lemma Let T : X → Y be an adjointable linear map between pre-Hilbert B-modules X, Y . Then T is B-linear. [93, §2] Proof We have hy, (T x)bi = hT ∗y, xi b = hy, T (xb)i for any x ∈ X, y ∈ Y and b ∈ B. In particular we get h(T x)b − T (xb), (T x)b − T (xb)i = 0 taking y = (T x)b − T (xb) and so T (x)b = T (xb). (cid:3) Proposition Let X and Y be right B-modules with B-valued inner products for some C∗-algebra B. If T : X → Y is a bounded B-linear map, then we have the inequality hT x, T xi 6 kTk2 hx, xi for any x ∈ X. (cf. [93, rem. 2.9].) Proof (The proof is a slight variation on the first part of [93, thm. 2.8].) Pick ε > 0. By 17 V we know hx, xi + ε is invertible. Define h = (hx, xi + ε)− 1 2 . Clearly 0 6 hxh, xhi = h(hx, xi + ε)h − h2ε = 1 − h2ε 6 1. Thus kxhk 6 1 and so khhT x, T xi hk = kT (xh)k2 6 kTk2. From this and 0 6 hhT x, T xi h we get hhT x, T xi h 6 kTk2. Hence, by dividing by h on both sides, hT x, T xi 6 kTkh−2 = kTk2(hx, xi + ε) for all ε > 0 and so hT x, T xi 6 kTk2 hx, xi as desired. (cid:3) Proposition Let B be a von Neumann algebra, X be a Hilbert B-module and x ∈ X. Then the map h : Ba(X) → B defined by h(T ) = hx, T xi is completely positive. Proof Pick any n ∈ N, T1, . . . , Tn ∈ Ba(X) and b1, . . . , bn ∈ B and compute Xi,j b∗i h(T ∗i Tj)bj = Xi,j = Xi,j = (cid:10)Xi b∗i hx, T ∗i Tjxi bj hTixbi, Tjxbji Tixbi,Xi Tixbi(cid:11) > 0, which shows h is indeed completely positive. ..143 -- 145 (cid:3) 33 2.2.2 Ultranorm uniformity 146 In 136 II we saw how to complete a complex vector space with inner product to a Hilbert space. Now we will see how to complete a right B-module with B-valued inner product to a self-dual Hilbert B-module under the assumption B is a von Neumann algebra. We will use a different construction than Paschke used [93]. Paschke showed that for a pre-Hilbert B-module X the set of functionals X′ (i.e. B-module homomorphisms into B) turns out to be a self-dual Hilbert B- module, which he uses as the completion of X. A considerable part of his paper [93] is devoted to this construction. It requires Sakai's characterization of von Neumann algebras, which we have not covered. To avoid developing Sakai's theory, we give a different construction. Instead of embedding X into a dual space, we will stay closer to the similar fact for Hilbert spaces and use a topological completion. A simple metric completion will not do, we will need to complete X as a uniform space. Uniformities are structures that sit between topological spaces and metric spaces: every metric space is a uniformity (see V) and every uniformity is a topological space. We will refresh the basics of uniformities -- for a thorough treatment, see for instance [133, ch. 9]. II Definition A uniform space is a set X together with a family of relations Φ ⊆ ℘(X × X) called entourages satisfying the following conditions. 1. The set of entourages Φ is a filter. That is: (a) if ε, δ ∈ Φ, then ε ∩ δ ∈ Φ and (b) if ε ⊆ δ and ε ∈ Φ, then δ ∈ Φ. 2. For every ε ∈ Φ and x ∈ X we have x ε x. 3. For each ε ∈ Φ, there is a δ ∈ Φ such that δ2 ⊆ ε. So, if x δ y and y δ z then x ε z for any x, y, z ∈ X. 4. For every ε ∈ Φ, there is a δ ∈ Φ with δ−1 ⊆ ε. So, if x δ y then y ε x for any x, y ∈ X. A uniform space is Hausdorff whenever x ε y for all ε ∈ Φ implies that x = y. III The elements of X are the points of the uniform space. The entourages ε ∈ Φ are generalized distances: one reads x ε y as 'the point x is at most ε far from to y'. With this in mind, the second axiom states every point is arbitrarily close to itself. The third axiom requires that for every entourage ε there is an entourage which acts like ε/2. There might be several ε/2 which fit the bill. Whenever we write ε/2 we implicitly pick some entourage that satisfies (ε/2)2 ⊆ ε. Also we will use the obvious shorthand ε/4 = (ε/2)/2. IIIa Definition Let X be a set together with a family of relations B ⊆ ℘(X × X). The set B is said to be a subbase (for a uniformity on X) if it satisfies axioms 2, 3 and 4 of II. (Cf. [133, dfn. 5.5].) Exercise Let X be a set together with a subbase B. Write Φ for the filter generated by B (that is: δ ∈ Φ iff ε1 ∩ . . . ∩ εn ⊆ δ for some ε1, . . . , εn ∈ B). Show that (X, Φ) is a uniform space. Examples Using the definition of a subbase, it is easy to describe the entourages of some common uniformities. IV V 1. Let (X, d) be a metric space. Define ε ≡ {(x, y); d(x, y) 6 ε} for any ε > 0. The set B ≡ {ε; ε > 0} is a subbase and so fixes a uniformity Φ on X. In this sense every metric space is a uniformity. 2. Let X be a set together with a (infinite) family (dα)α∈I of pseudometrics.‖ Define Eα,ε = {(x, y); dα(x, y) 6 ε}. Then B ≡ {Eα,ε; ε > 0, α ∈ I} is a subbase and fixes a uniformity Φ on X. (Every uniform space is of this form, see e.g. [133, thm. 39.11].) 3. Let V be a vector space with a family of seminorms (k kα)α∈I . As a special case of the previous, V has a uniformity fixed by pseudometrics dα(x, y) = kx − ykα. (If additionally x = 0 provided kxkα = 0 for all α, then V is called a locally convex space.) 4. Assume B is a von Neumann algebra. There are several uniformities on B -- two of which are of particular interest to us. The ultrastrong uniformity 2 where f : B → C is an on B is fixed by the seminorms kbkf = f (b∗b) np-map. The ultraweak uniformity on B is given by the seminorms f (b) for np-maps f : B → C. 1 We are ready to define a uniformity for Hilbert B-modules. Later, we will complete X with respect to this uniformity. 1 Definition Let B be a von Neumann algebra. Assume X is a right B-module with B-valued inner product [· , · ]. We call the uniformity on X given by 2 for np-maps f : B → C the ultranorm uniformity. seminorms kxkf = f ([x, x]) The ultranorm uniformity will play a very similar role to the norm for Hilbert spaces. If B = C, then the ultranorm uniformity is the same as the uniformity induced by the norm. If X = B with [a, b] = a∗b, then the ultranorm uniformity coincides with the ultrastrong uniformity. Beware The ultranorm uniformity is (in general) not given by a single norm. Furthermore the ultranorm uniformity is weaker than the norm uniformity (that is: norm convergence implies ultranorm convergence, but not necessarily the other way around), even though on von Neumann algebras the ultraweak uni- formity is stronger than the weak uniformity. VI VII VIII IX ‖Like a metric, a pseudometric d is symmetric, obeys the triangle inequality and d(x, x) = 0, but unlike a metric d(x, y) = 0 is not required to entail x = y. 146.. 35 X Remarks The ultranorm uniformity has appeared earlier in the literature under various different names; for instance as the τ1-topology in [88, thm. 3.5.1] and as the s-topology in [38, 2.9]. In previous work the ultranorm uniformity only plays a minor role, often amongst other topologies. In this thesis we put the ultranorm uniformity center-stage as the right generalization of the norm and give it the 'ultranorm' name to match this view. 147 Definition Let X be a uniform space with entourages Φ. It is easy to translate the familiar notions for metric spaces to uniform spaces. 1. A net (xα)α is said to converge to x if for each ε ∈ Φ there is an α0 such that for all α > α0 we have x ε xα. Note that a net can converge to two different points. Indeed, this is the case if we start of with a pseudometric that is not a metric. 2. A net (xα)α is called Cauchy if for each ε ∈ Φ there is an α0 such that for all α, β > α0 we have xα ε xβ. The uniform space X is complete when every Cauchy net converges. 3. Let (Y, Ψ) be a second uniform space. A map f : X → Y is said to be uniformly continuous if for each ε ∈ Ψ there is a δ ∈ Φ such that for all x δ y we have f (x) ε f (y). The map f is merely continuous if for each x and ε ∈ Ψ, there is a δ ∈ Φ such that for all x δ y we know f (x) ε f (y). 4. We say two Cauchy nets (xα)α∈I and (yβ)β∈J are equivalent, in symbols: (xα)α ∼ (yβ)β, when for every ε ∈ Φ there are α0 ∈ I and β0 ∈ J such that for all α > α0 and β > β0 we have xα ε xβ . 5. A subset D ⊆ X is said to be dense if for each ε ∈ Φ and x ∈ X, there is a y ∈ D with x ε y. II Exercise In the same setting as the previous definition. 1. Show that equivalence of Cauchy nets is an equivalence relation. Show that if (xα)α is a subnet of a Cauchy net (yα)α, that then (xα)α is equivalent to (yα)α. 2. Prove that if (xα)α and (yα)α are equivalent Cauchy nets and xα → x, then also yα → x. 3. Assume (xα)α is a net with xα → x and xα → y. Prove x = y whenever X is Hausdorff. 4. Show that if f : X → Y is a continuous map between uniform spaces and we have xα → x in X, then f (xα) → f (x) in Y . 5. Assume f is a uniformly continuous map between uniform spaces. Show that f maps Cauchy nets to Cauchy nets and furthermore that f maps equivalent Cauchy nets to equivalent Cauchy nets. 6. Suppose D ⊆ X is a dense subset. Show that for each x ∈ X there is a Cauchy net (dα)α∈Φ in D with dα → x in X. 7. Assume f, g : X → Y are continuous maps between uniform spaces where Y is Hausdorff. Conclude from 4 and 6 that f and g are equal whenever they agree on a dense subset of X. Exercise Let (Xi)i∈I be a family of sets with uniformities (Φi)i∈I . For each i0 ∈ I and ε ∈ Φi0 , define a relation on Πi∈I Xi by (xi)i∈I ε (yi)i∈I ⇐⇒ xi0 ε yi0 . Show {ε; ε ∈ Φi, i ∈ I} is a subbase for Πi∈I Xi; that the projec- tions πi : Πi∈I Xi → Xi are uniformly continuous with respect to them and that they make Πi∈I Xi into the product of (Xi)i∈I in the category of uniform spaces with uniformly continuous maps. III Proposition Let X and Y be right B-modules with B-valued inner products. A bounded B-linear map T : X → Y is uniformly ultranorm continuous. Proof Let f : B → C be an np-map and ε > 0. Assume kx − ykf 6 ε kTk 144 V we have hT (x − y), T (x − y)i 6 kTk2 hx − y, x − yi and so . By 148 II kT x − T ykf = f (hT (x − y), T (x − y)i) 1 2 6 kTkkx − ykf 6 ε. Thus T is uniformly continuous w.r.t. the ultranorm uniformity. (cid:3) Corollary Let B be a von Neumann algebra and X a right B-module with B- valued inner product. For x0 ∈ X the maps III (x, y) 7→ x + y X × X → X x 7→ hx0, xi X → B b 7→ x0b B → X are all uniformly continuous w.r.t. the ultranorm uniformity. Exercise Let B be a von Neumann algebra and X a right B-module with B- valued inner product. Show that x 7→ xb is ultranorm continuous for any b ∈ B. Proposition Let X be a right B-module with B-valued inner product [· , · ] for some C∗-algebra B. If xα → x and yα → y in the ultranorm uniformity, for some nets (xα)α∈I , (yα)α∈I , then [xα, yα] → [x, y] ultraweakly. Proof Let ε > 0 and np-map f : B → C be given. Note that kyαkf → kykf ; indeed by the reverse triangle inequality kyαkf −kykf 6 kyα − ykf → 0. Thus we can find α1 such that kyαkf 6 kykf + 1 for α > α1. IV V VI ..146 -- 148.. 37 We can find α2 such that for all α > α2 we have kx− xαkf 6 1 2 ε(kykf + 1)−1 2 ε(kxkf + 1)−1. Pick α0 > α1, α2. Then for α > α0 we have and ky − yαkf 6 1 f ([xα, yα] − [x, y]) = [xα − x, yα]f + [x, yα − y]f 6 kxα − xkfkyαkf + kxkfkyα − ykf 6 kxα − xkf (kykf + 1) + (kxkf + 1)kyα − ykf 6 ε. by 142 III (cid:3) Thus [xα, yα] → [x, y] ultraweakly. VII Corollary Let X be a Hilbert B-module together with an ultranorm-dense subset D ⊆ X. The vector states from D are order separating: for each T ∈ Ba(X) we have T > 0 if and only if hx, T xi > 0 for all x ∈ D. Cf. 144 I. 149 Definition Let X be pre-Hilbert B-module for a von Neumann algebra B. Let E ⊆ X be some subset. 1. E is orthogonal if he, di = 0 for e, d ∈ E with e 6= d. 2. E is orthonormal if additionally he, ei is a non-zero projection for e ∈ E. 3. A family (be)e∈E in B is said to be ℓ2-summable if the partial sums ofPe∈E b∗ebe are bounded. 4. E is an (orthonormal) basis if it is orthonormal and additionally (a) for each x ∈ X we have x = Xe∈E e he, xi , where the sum converges in the ultranorm uniformity and (b) the sumPe∈E ebe converges in the ultranorm uniformity for any ℓ2- summable (be)e∈E in B. II Notation As with Hilbert spaces, it will sometimes be more convenient to view an orthonormal basis as a sequence/family (ei)i∈I , instead of a subset. This should cause no confusion. IIa Remark The first assumption '(a)' in the definition of an orthonormal basis E re- quires that every element x can be reconstructed from its coefficients (he, xi)e∈E. This assumption fails, for instance, if there is a non-zero element orthogonal to all basis elements. The second assumption '(b)' requires that behind any a rea- sonable family of coefficients, there is a corresponding element. This second assumption fails if the module is not ultranorm complete, see V. Beware In the literature, there are several different definitions of an orthonormal basis for a Hilbert B-module X. IIb such that x =Pe∈E e he, xi for all x ∈ X where the sum converges in the 1. Some [82,88] define an orthonormal basis as an orthonormal subset E ⊆ X norm. This notion of basis is independent of ours: there is a basis in our sense that is not a basis in their sense ({0ih0 ,1ih1 , . . .} in B(ℓ2)) and vice versa (the standard Hilbert module HB(ℓ2) defined in [82, §3] has a basis in their sense, but not a basis in ours as it is not self dual, cf. V.) 2. Others such as [12, lemma 8.5.23] define an orthonormal basis only when X is a self-dual (V) Hilbert C∗-module over a von Neumann algebra. In this case an orthonormal basis is defined as an orthonormal subset E ⊆ X such topology of X (with respect to the unique Banach space predual of X for which the inner product is separately w∗-continuous, see [12, lemma 8.5.4].) This notion is equivalent to ours, but only defined for self-dual modules over von Neumann algebras. that x =Pe∈E e he, xi for all x ∈ X where the sum converges in the w∗- Exercise Let X be a pre-Hilbert B-module for a C∗-algebra B. Show that we have ehe, ei = e for any e ∈ X where he, ei is a projection. (Hint: consider ke(1 − he, ei)k2; cf. [93, thm. 3.12].) Exercise Let X be a pre-Hilbert B-module for a von Neumann algebra B with an orthonormal basis E ⊆ X. Show Parseval's identity holds, that is: hx, xi = Xe∈E hx, eihe, xi, III IV where the sum converges ultraweakly. [93] (Hint: use 148 V.) Theorem Assume B is a von Neumann algebra. For any pre-Hilbert B- module X, the following conditions are equivalent. V 1. X is self dual. 2. X is ultranorm complete. 3. Every norm-bounded ultranorm-Cauchy net in X converges. 4. X has an orthonormal basis. Proof We will prove 1 ⇒ 3 ⇒ 4 ⇒ 2 ⇒ 4 ⇒ 1. The equivalence 1 ⇔ 4 was already shown by Paschke in [93, thm. 3.12], but we give a different proof. The equivalence 1 ⇔ 3 turns out proven by Frank already [32, thm. 3.2] in a different way. The final equivalence 1 ⇔ 2 is new, as far as we know. VI ..148 -- 149.. 39 VII (1 ⇒ 3, self-dual X are bounded complete) Let X be a self-dual pre-Hilbert B- module. To show norm-bounded ultranorm completeness, assume (xα)α is an ultranorm-Cauchy net in X with kxαk 6 B for some B > 0. For the moment pick an y ∈ X. By 142 III we have f (hxα, yihy, xαi) 6 kyk2f (hxα, xαi) for any np-map f : B → C and so hy, xαi is an ultrastrong-Cauchy net in B. Von Neumann algebras are ultrastrongly complete, see 77 I, so we can define τ (y) = (uslimα hy, xαi)∗, where uslim denotes the ultrastrong limit. As addition and multiplication by a fixed element are ultrastrongly continuous (see 45 IV) τ is B-linear. We will show τ is bounded. For each α we have hxα, yihy, xαi 6 kyk2 hxα, xαi 6 kyk2B2 and so τ (y)τ (y)∗ 46 II= uwlim α hxα, yihy, xαi 6 kyk2B2, where uwlim denotes the ultraweak limit. Thus kτ (y)k2 = kτ (y)τ (y)∗k 6 kyk2B2. So τ is bounded. By self-duality, there is a t ∈ X such that τ (y) = ht, yi for all y ∈ X. And so for each np-map f : B → C we have k hy, t − xαikf = k hy, ti − hy, xαikf = kτ (y)∗ − hy, xαikf → 0. (2.5) In 142 II we introduced the inner product h· , ·if defined by hx, yif ≡ f (hx, yi). From Kadison's inequality 30 IV it follows f (a)2 6 f (a∗a)f (1) and so hy, t − xαif 2 6 f (ht − xα, yihy, t − xαi)f (1) = k hy, t − xαik2 Combining (2.5) and (2.6), we see hy, t − xαif → 0. Let ε > 0 be given. As kxαkf is norm bounded we can find B > 0 such that kxαkf 6 B for all α. As (xα)α is ultranorm Cauchy, we can pick α such that kxβ − xαkf 6 (ktkf + B)−1 ε 2 . Putting it all together: 2 for all β > α. Find β such that ht − xα, t − xβif 6 ε f f (1). (2.6) ht − xα, t − xαif 6 ht − xα, t − xβif + ht − xα, xβ − xαif 6 6 ε 2 ε 2 + kt − xαkfkxβ − xαkf + (ktkf + B)kxβ − xαkf 6 ε. VIII Thus xα converges ultranorm to t. So X is ultranorm bounded complete. (3 ⇒ 4, bounded complete X has an orthonormal basis) Assume norm-bounded ultranorm-Cauchy nets converge in X. By Zorn's lemma, there is a maximal orthonormal subset E ⊆ X. We will show E is a basis. Let (be)e∈E be an E-tuple in B such that the partial sums ofPe∈E b∗ebe are bounded. We want to show Pe∈E ebe converges in the ultranorm uniformity. As the summands ofPe∈E b∗ebe are positive, the partial sums form a directed bounded net that converges ultrastrongly to its supremumPe∈E b∗ebe. By III, we have e he, ei = e for each e ∈ E and so hx, eihe, ei = hx, ei for every x ∈ X. Thus for any finite subset S ⊆ E DXe∈S ebe,Xe∈S ebeE = Xe∈S b∗e he, ei be 6 Xe∈S b∗ebe. To showPe ebe converges ultranorm, pick any np-map f : B → C. AsPe∈E b∗ebe converges ultraweakly, we have Pe∈E f (b∗ebe) < ∞. Thus the tails of the se- riesPe∈E−S f (b∗ebe) tend to zero, hence for finite S, T ⊆ E we know (cid:13)(cid:13)Xe∈S ebe −Xe∈T f = (cid:13)(cid:13) Xe∈S∆T ebe(cid:13)(cid:13)2 f ebe(cid:13)(cid:13)2 6 Xe∈S∆T in the ultranorm uniformity. The latter vanishes as S ∩ T grows. Thus the partial sums of Pe∈E ebe are ultranorm Cauchy and norm-bounded by kPe∈E b∗ebek, soPe∈E ebe converges Pick x ∈ X. We have to show Pe∈E xhe, xi converges in the ultranorm 0 6 Dx −Xe∈S e he, xiE = hx, xi −Xe∈S uniformity to x. For any finite subset S ⊆ E we have e he, xi , x −Xe∈S hx, eihe, xi . f (b∗ebe). Rearranging we find Bessel's inequality: Xe∈S hx, eihe, xi 6 hx, xi . can show x′ = 0. Hence the E-tuple (he, xi)e is ℓ2-summable and so Pe∈E e he, xi converges in the ultranorm uniformity. Consider x′ = x −Pe∈E e he, xi. We are done if we To see x′ = 0, we first show that X has polar decomposition: for each y ∈ X, 2 . By 80 IV the 2 has an approximate pseudoinverse h1, h2, . . . -- that is: there is an u ∈ X with hu, ui = ⌈hy, yi⌉ and y = u hy, yi positive element hy, yi hn hy, yi 2 = hy, yi 2 =Pn ⌈hn⌉ = ⌈hy, yi⌉. Note 2 hn = ⌈hn⌉ andPn hn hy, yi 1 1 1 1 1 1 1 hyhn, yhmi = (hn hy, yi 2 hm) = ⌈hn⌉⌈hm⌉ and so using that ⌈h1⌉ ,⌈h2⌉ , . . . are pairwise orthogonal, we see 2 )(hy, yi yhn, NXn=1 NXn=1 yhn(cid:11) = (cid:10) NXn=1 Thus (PN assumption this bounded net converges: write u ≡Pn yhn. (So in the special n=1 yhn)N forms an ultranorm-Cauchy net norm bounded by 1. By ⌈hn⌉ 6 ⌈hy, yi⌉ . ⌈hn⌉⌈hm⌉ = NXm=1 NXn=1 ..149.. 41 case X = B, we have u = y/hy, yi 1 2 , see 81 III.) We derive u hy, yi yhn hy, yi y ⌈hn⌉ 1 2 = Xn = Xn = yXn ⌈hn⌉ = y ⌈hy, yi⌉ = y, 1 2 by dfn. u and 148 IV by dfn. hn by 148 III by dfn. hn where the last equality follows from y(1 − ⌈hy, yi⌉) = 0, which is justified by This finishes the proof of polar decomposition. (cid:10)y(1 − ⌈hy, yi⌉), y(1 − ⌈hy, yi⌉)(cid:11) = (1 − ⌈hy, yi⌉)hy, yi (1 − ⌈hy, yi⌉) = 0. Recall we want to show x′ ≡ x −Pe∈E e he, xi = 0. Reasoning towards contradiction, assume x′ 2 for some u ∈ X with hu, ui = ⌈hx′, x′i⌉. Note hu, ui 6= 0 for otherwise x′ = 0, quod non. For any e0 ∈ E, we have 6= 0. By polar decomposition x′ = u hx′, x′i 1 he0, x′i = (cid:10)e0, x −Xe∈E e he, xi(cid:11) 148 III= he0, xi −Xe∈E he0, eihe, xi = 0 and so 0 = he, x′i = he, uihx′, x′i 1 2 for any e ∈ E. Thus he, ui = he, uihu, ui = he, ui⌈hx′, x′i⌉ 60 VIII= 0. Thus E ∪ {u} contradicts the maximality of E. Hence x′ = 0 and so x = Pe∈E e he, xi as desired. IX (4 ⇒ 2, basis implies completeness) Assume X has an orthonormal basis E. Let (xα)α be an ultranorm-Cauchy net. Pick e ∈ E. By 148 III (he, xαi)α is an ultrastrong-Cauchy net. As von Neumann algebras are ultrastrongly complete (see 77 I) there is a be ∈ B to which he, xαi converges ultrastrongly. We will show xα converges in the ultranorm uniformity toPe∈E ebe. We will first show that Pe∈E ebe converges. As before, it is sufficient to it suffices to show Pe∈E f (b∗ebe) is show Pe∈E b∗ebe is bounded. For this, bounded for each np-map f : B → C, as ultraweakly-bounded nets are norm bounded, see 87 VIII. For any e ∈ E we have kbekf = limα k he, xαikf and so f (b∗ebe) = limα f (hxα, eihe, xαi). Pick a finite subset S ⊆ E. Then we have f(cid:0)Xe∈S b∗ebe(cid:1) = Xe∈S = lim α lim α f (hxα, eihe, xαi) f(cid:0)Xe∈S hxα, eihe, xαi(cid:1) 6 f (hxα, xαi) by Bessel's inequality. As xα is ultranorm Cauchy, the net f (hxα, xαi) must be Cauchy and so bounded. HencePe∈E ebe converges in the ultranorm uniformity. To prove xα converges toPe∈E ebe, pick any ε > 0 and np-map f : B → C. We want to find α0 such that kxα −Pe∈E ebekf 6 ε for all α > α0. As (xα)α is ultranorm Cauchy, we can pick an α0 such that kxα − xβkf 6 1 2√2 all α, β > α0. As xα =Pe∈E e he, xαi we find with Parseval's identity (IV) that ε for 2 ε2. If we can 2 ε2, then we are done. For any β > α we f 6 1 f 6 1 k he, xαi − bek2 f . f = Xe∈E (cid:13)(cid:13)xα −Xe∈E e(he, xαi − be)(cid:13)(cid:13)2 have using the triangle inequality of k · kf on X f = (cid:13)(cid:13)Xe∈E ebe(cid:13)(cid:13)2 Take a finite subset S ⊆ E such thatPe∈E−S k he, xαi − bek2 also show Pe∈S k he, xαi − bek2 (cid:16)Xe∈S = (cid:13)(cid:13)Xe∈S 6 (cid:13)(cid:13)Xe∈S = (cid:16)Xe∈S e he, xαi − ebe(cid:13)(cid:13)f e he, xα − xβi(cid:13)(cid:13)f +(cid:13)(cid:13)Xe∈S e he, xβi − ebe(cid:13)(cid:13)f f(cid:17) 1 +(cid:16)Xe∈S k he, xα − xβik2 f(cid:17) 1 k he, xαi − bek2 . f(cid:17) 1 k he, xβi − bek2 2 2 2 2√2 ε(cid:1)2 f 6 kxα − xβk2 f 6 (cid:0) 1 We want to show the previous is bounded by 1√2 ε. We bound the two terms separately. For the first term using Bessel's inequality and our choice of α0 . As k he, xβi − bekf vanishes for each e ∈ S and S is finite, we can find sufficiently large β such that the right term is also bounded by 1 ε. Thus xα converges in the ultranorm 2√2 we have Pe∈S k he, xα − xβik2 uniformity toPe∈E ebe, as desired. (2 ⇒ 4, completeness implies basis) Follows from 3 ⇒ 4 as 2 ⇒ 3 is trivial. (4 ⇒ 1, basis implies self-duality) Assume X has an orthonormal basis E. Suppose τ : X → B is a bounded B-linear map. We want to show there is some t ∈ X such that τ (x) = ht, xi for all x ∈ X. Pick any x ∈ X. By assump- tionPe∈E e he, xi converges in the ultranorm uniformity to x. As by 148 I τ is ..149.. 43 X XI ultranorm continuous we see τ (x) = Xe∈E τ (e)he, xi = Xe∈E heτ (e)∗, xi . assumption this is the case if (τ (e)∗)e∈E is ℓ2-summable -- that is: the partial Thus if we can show Pe eτ (e)∗ converges in the ultranorm, we are done. By sums ofPe∈E τ (e)τ (e)∗ must be bounded. We will show that for finite S ⊆ E we have kPe∈S eτ (e)∗k 6 kτk, which is sufficient as then kPe∈S τ (e)τ (e)∗k = kPe∈S eτ (e)∗k2 6 kτk2. For the moment pick an arbitrary x ∈ X. By Bessel's inequality we have kPe∈S e he, xi k2 =Pe∈S k hx, ei he, xik 6 kxk2. Thus eτ (e)∗, x(cid:11)(cid:13)(cid:13) = (cid:13)(cid:13)τ(cid:0)Xe∈S SubstitutingPe∈S eτ (e)∗ for x, we find eτ (e)∗,Xe∈S (cid:13)(cid:13)(cid:10)Xe∈S eτ (e)∗k2 = (cid:13)(cid:13)(cid:10)Xe∈S kXe∈S e he, xi(cid:1)(cid:13)(cid:13) 6 kτkkxk. eτ (e)∗(cid:11)(cid:13)(cid:13) 6 kτk(cid:13)(cid:13)Xe∈S If kPe∈S eτ (e)∗k = 0, then we are done. Otherwise, divide both sides of the previous equation by kPe∈S eτ (e)∗k to find kPe∈S eτ (e)∗k 6 kτk as desired. (cid:3) eτ (e)∗(cid:13)(cid:13). 2.2.3 Self-dual completion 150 In 136 II we saw that every vector space with inner product can be completed into a Hilbert space. We continue with the generalization to Hilbert C∗-modules. Paschke proves a similar result [93, thm. 3.2] in a completely different way. II Theorem Let B be a von Neumann algebra and V a right B-module with B- valued inner product [· , · ]. There is a self-dual Hilbert B-module X together with a bounded B-linear η : V → X such that 1. [v, w] = hη(v), η(w)i and 2. the image of η is ultranorm dense in X. Proof Before we start the proof proper, we give an overview. (Overview) By 149 V and 148 V one would expect that one can simply extend the inner product of V to its ultranorm completion V by III IV [(xα)α, (yα)α] = uwlim [xα, yα], α (2.7) which will turn V into a self-dual Hilbert B-module. It will turn out that V is indeed a self-dual Hilbert B-module with this inner product, but it isn't clear at all how to show directly that the ultraweak limit in (2.7) converges. We will sketch an indirect construction, before delving into the details. As with a metric completion, V consists of equivalence classes of ultranorm Cauchy nets in V . Let V0 ⊆ V denote the equivalence classes of constant Cauchy nets. V0 is a right B-module with B-valued inner product using the operations of V . Let V1 ⊆ V denote the limits of norm-bounded ultranorm Cauchy nets over V0. Using the norm boundedness we can extend the module structure and inner product of V0 to V1. If all norm-bounded ultranorm Cauchy-nets over V1 would converge in V1, then V1 would be a self-dual Hilbert B-module and V1 would even be ultranorm complete. Later, when we have established that V is a Hilbert B-module, we will find out that in fact V ⊆ V1 due to 158 II -- i.e. V1 is already ultranorm complete. However, in a most frustrating state of affairs, we do not know how to prove this directly without first proving that V is a Hilbert B-module. Instead, we repeat: let V2 ⊆ V denote the limits of norm- bounded ultranorm Cauchy nets over V1. We have to go further: for any n ∈ N define Vn+1 = σ(Vn) where σ(U ) ⊆ V denotes the limits of norm-bounded ultranorm Cauchy over U . V0 ⊆ V1 ⊆ V2 ⊆ V3 ⊆ ··· ⊆ [n∈N Vn. Will we be able to show that Sn Vn is finally ultranorm complete? Not yet. We have to go even further. Define Vω = σ(Sn∈N Vn), Vω+1 = σ(Vω) and so for any ordinal number α > 0, define Vα = σ(Sβ<α Vβ ). Now Vα is an on: ascending chain of subsets of V . Thus Vα0 = Vα0+1 for some α0 6 2V (for otherwise 2V 6 V ). Thus norm-bounded ultranorm Cauchy nets in Vα0 must converge. We will see that at each step we can extend the module structure and inner product and with those Vα0 is self-dual. Thus Vα0 is ultranorm complete and actually Vα0 = V . V0 ⊆ V1 ⊆ ··· ⊆ Vω ⊆ Vω+1 ⊆ ··· ⊆ V2ω ⊆ ··· ··· ⊆ Vα0 = V To avoid requiring familiarity with transfinite induction, we will phrase the proof using Zorn's lemma instead. (V : fast nets) As we need some details in its construction, we will explicitly define V , the completion of V in the ultranorm uniformity using Cauchy nets. There are other ways to construct a completion of a uniform space, see for instance [133, thm. 39.12]. Let Φ denote the set of entourages of the ultranorm uniformity on V ; for its definition see 146 VII. Φ is a filter and thus can be used as index set for a net using reverse inclusion as order. (Thus ε > δ ⇔ ε ⊆ δ.) We say a net (xα)α∈Φ indexed by entourages is fast if for every ε ∈ Φ and α, β > ε we have xα ε2 xβ. Every Cauchy net is equivalent to a fast one, but this is not as evident as in the metric case. Let (xα)α∈I be an arbitrary Cauchy net in V . By definition V ..149 -- 150.. 45 we can find αε ∈ I such that for each ε ∈ Φ and α, β > αε we have xα ε xβ. Unfortunately α( · ) need not be order preserving and so (xαε )ε∈Φ need not be a subnet of (xα)α∈I . However, we claim the net (xαε )ε∈Φ is Cauchy, fast and equivalent to (xα)α∈I . To show it's Cauchy and fast, pick any ε ∈ Φ and ζ, ξ > ε (that is: ζ, ξ ⊆ ε). As I is directed, we can find β ∈ I with β > αζ, αξ. By definition of α( · ) we have xαζ ζ xβ ξ xαξ and so xαζ ε2 xαξ , as desired. To show equivalence, assume ε ∈ Φ is given. Assume δ > ε/2 and β > αε/2. There is a γ ∈ I with γ > αδ, β. Then xβ ε/2 xγ δ xαδ and so as δ ⊆ ε/2 we get xβ ε xαδ . We have shown equivalence. if fast Cauchy nets (xα)α∈Φ and (yα)α∈Φ are equivalent, then we can find for every ε ∈ Φ some β ∈ Φ such that for all γ > β, we have xγ ε yγ. (V : the uniform space N ) Write N for the set of fast Cauchy nets over V . Later we will define V as N modulo equivalence. Because of a subtlety with the definition of the uniformity on V later, it is helpful to consider N separately. Let ε ∈ Φ be given. For nets (xα)α and (yα)α in N , define A different fact about fast nets will be useful later on: (xα)α ε (yα)α ⇔ ∃β ∈ Φ∀γ > β. xγ ε yγ. If ε ⊆ δ, then ε ⊆ δ and ε1 ◦ ε2 ⊆ \ε1 ◦ ε2. Socε/2◦cε/2 ⊆ \ε/2◦ ε/2 ⊆bε, which is one requirement for {ε; ε ∈ Φ} to be a subbase for N . The others are easy as well. Also \ε1 ∩ ε2 = ε1 ∩ ε2 and so each entourage of N has some ε as subset. For the following, it is helpful to remark now that Cauchy nets (xα)α and (yα)α over V are equivalent if and only if (xα)α ε (yα)α for all ε ∈ Φ. (V : N is complete) As every Cauchy net is equivalent to a fast one, it is sufficient to show convergence of fast Cauchy nets. Thus let ((xγ α)α)γ be a fast Cauchy net in N . First we will show (x α α)α is a Cauchy net. It might not be fast, so formally ((xγ α)α)γ cannot converge to it, but it will converge to an equivalent fast Cauchy net. As ((xγ α)α)γ is fast, we know that for all γ1, γ2 > ε we have (xγ1 α )α. So there is a ζγ1,γ2 ∈ Φ such that for α > ζγ1,γ2 we have xγ1 α . Suppose α, β > ε. Pick ξ > α, β, ζ α, β. Then we have x α α ε2 x α α)α is Cauchy. Let (yα)α be a fast Cauchy net equivalent to (x α α)α → (yα)α. As (yα)α is equivalent α whenever α > α0. As ((xγ to (x α α)α)γ is fast Cauchy, we know that (x α such that x α β α )α ε2 (xγ2 α ε2 xγ2 β β ξ ε2 x ξ ε2 x β and so (x α α)α. We want to prove (xγ α)α we can find α0 such that yα ε/4 x α ε/4 xγ β)β if α, γ >cε/8. Thus there is some β0 β for β > β0 and α, γ >cε/8. Thus for β > β0∩ ε/8, α > α0∩ ε/8 β )β cε/4 (xγ α. Thus (yα)α ε (xγ ε/4 xγ β ε/4 x α β ε/4 xγ α)α and γ > cε/8 we get yα ε/4 x α whenever γ >cε/8, so (xγ α α)α → (yα)α. We have shown N is complete. (V : uniformity) Let V denote the set of fast Cauchy nets over V modulo equiv- alence. Write η : V → V for the map that sends x ∈ V to the equivalence class of the constant Cauchy net (x)α∈Φ. For brevity, we may write x = η(x). As VI VII VIII announced before, define V0 ≡ η(V ). What uniformity to put on V ? Unfortunately the relations ε do not neces- sarily preserve equivalence (∼) of Cauchy nets: there might be (xα)α ∼ (x′α)α and (yα)α ∼ (y′α)α in N such that (xα)α ε (yα)α, but not (x′α)α ε (y′α)α for some ε ∈ Φ. Instead we define for any ε ∈ Φ the following entourages on V : for all (x′α)α ∼ (xα)α and (y′α)α ∼ (yα)α. (xα)α ε (yα)α ⇔ (x′α)α ε (y′α)α By definition ε respects equivalence and can be considered as a relation on V . It is not hard to verify {ε; ε ∈ Φ} is a subbase for V and ^ε1 ∩ ε2 = ε1 ∩ ε2. It is easy to see V is a Hausdorff uniform space. Write Φ for the generated set of entourages on V . The entourages of V are exactly those relations which have some ε as a subset. Furthermore, if (xα)α ε (yα)α, then (xα)α ε3 (yα)α. Thus if a net (xγ α)α converges to (yα)α in N , then so do their equivalence classes in V . Hence V is complete. (The inclusion ε ⊆ ε3 ⊆ ε3 implies that the ε and ε generate the same uniformity on N .) (V is a right B-module) First we will define an addition on V . The addition on V is uniformly continuous (as the uniformity on V is given by seminorms, see 147 III for the uniformity on V 2.) Thus for any (xα)α and (yα)α in N , the net (xα +yα)α is again Cauchy, see 147 II. Also because of the uniform continuity of addition: if (xα)α ∼ (x′α)α and (yα)α ∼ (y′α)α, then (xα + yα)α ∼ (x′α + y′α)α. There is a net in N equivalent to (xα + yα)α. This fixes an addition on V , which turns it into an Abelian group. By construction η(x + y) = η(x) + η(y). 1 2 = f (b∗[x, x]b) 1 Pick any b ∈ B. We show the map rb : V → V given by rb(x) = xb is also uniformly continuous, which requires us to unfold the definitions further than with addition. Assume we are given an entourage in V , that is: np- maps f1, . . . , fn : B → C and ε > 0. For any np-map f : B → C, the map b ∗ f given by (b ∗ f )(x) ≡ f (b∗xb) is also an np-map, see 72 III. Clearly kxbkf = 2 = kxkb∗f for any x ∈ V . Thus if kx − ykb∗fi 6 ε f ([xb, xb]) for 1 6 i 6 n, then kxb − ybkfi = kx − ykb∗fi 6 ε as well. Hence rb is uniformly continuous. As before, this uniform continuity allows us to define a right B-action on V by sending the equivalence class of (xα)α to (xαb)α. By definition η(x)b = η(xb). It is straightforward to check this turns V into a right B-module. (Extending k · kf to V ) Let (xα)α be a Cauchy net over V and f : B → C be any np-map. From the reverse triangle inequality kxαkf − kxβkf 6 kxα − xβkf , it follows (kxαkf )α is Cauchy. Define k(xα)αkf = limα kxαkf . Again using the reverse triangle inequality kxαkf − kx′αkf 6 kxα − x′αkf we see k(xα)αkf = k(x′α)αkf whenever (xα)α ∼ (x′α)α. Thus k · kf lifts to V . These extended norms k · kf also induce a uniformity on V . We will now show it is the same uniformity. Let ((xγ α)α)γ be a net in V . It is sufficient to show that (xγ α)αkf → 0 for all np-maps f : B → C. α)α → 0 in the original uniformity if and only if k(xγ ..150.. 47 IX X Assume (xγ α)αkf → 0 as desired. For the converse, assume k(xγ α)αkf 6 1 2 ε for γ > γ0. Hence kxγ αk 6 ε whenever γ > γ0 and so k(xγ α)α → 0 in V . That is: for each f : B → C and ε > 0 there is a γ0 αkf 6 ε when α > α0. Let α)αkf = α)αkf → 0 for all np-maps f : B → C. Let ε > 0 αkf = αkf 6 ε for sufficiently large α, which is to such that for each γ > γ0 there is an α0 such that kxγ any f : B → C and ε > 0 be given. Find such γ0 for f and ε. Then k(xγ limα kxγ and np-map f : B → C be given. There is some γ0 such that limα kxγ k(xγ say (xγ (Induction set-up) The majority of the remaining work is to show we can define an inner product on V . As sketched earlier, we will extend the inner product from V0 step-by-step. Call a subset W ⊆ V together with a B-valued inner product [· , · ] a compatible extension if α)α → 0 in V . 1. V0 ⊆ W ; 2. W is a B-submodule of V ; 3. the inner product turns W into a pre-Hilbert B-module and 4. kxkV f = kxkW 1 f ([x, x]W ) f for all x ∈ W and np-maps f : B → C, where kxkW f is the extension of k · kf on V to V as in X. 2 and k · kV f = As k · kV f also generate the uniformity on V , the last requirement implies that on a compatible extension W the ultranorm uniformity and the uniformity induced by V coincide. For compatible extensions W1, W2 ⊆ V we say W1 6 W2 iff W1 ⊆ W2 and [v, w]W1 = [v, w]W2 for all v, w ∈ W1. Later, we will apply Zorn's lemma to the compatible extensions ordered in this way. (Induction base case) The inner product of V lifts to V0. Indeed, η(x) = 0 if and only if kxk = 0. So by Cauchy -- Schwarz [x, y] = 0 whenever η(x) = 0. Thus [x, y] = [x′, y′] if η(x) = η(x′) and η(y) = η(y′). Define [η(x), η(y)] = [x, y]. It's easy to see V0 is a pre-Hilbert B-submodule of V . Finally, for any np- map f : B → C we have kη(x)kV (Induction step) Assume W ⊆ V with [· , · ] is a compatible extension. Let σ(W ) denote the limits of norm-bounded Cauchy nets over W . We will define a B- valued inner product on σ(W ) that turns it into a compatible extension. In fact W 6 σ(W ). By assumption the induced uniformity on W is the ultra- norm uniformity and so addition and the B-action are uniformly continuous. Hence σ(W ) is again a submodule. f = kη(x)kV0 f = kxkV f , as desired. To define the inner product, assume x, y ∈ σ(W ). There are Φ-indexed norm- bounded Cauchy nets (xα)α and (yα)α over W such that xα → x and yα → y. XI XII XIII For any np-map f : B → C we have by Cauchy -- Schwarz (cid:12)(cid:12)f ([xα, yα] − [xβ, yβ])(cid:12)(cid:12) = (cid:12)(cid:12)[xα, yα − yβ]f + [xα − xβ, yβ]f(cid:12)(cid:12) 6 kxαkfkyα − yβkf + kxα − xβkfkyβkf . (2.8) (If not otherwise specified k · kf = k · kW f .) As (xα)α is norm bounded f = f ([xα, xα]) 6 kfkkxαk2, we see (kxαkf )α is bounded. Simi- and kxαk2 larly (kyβkf )β is bounded. Thus from (2.8) it follows ([xα, yα])α is a Cauchy net in the ultraweak uniformity of B, see 146 V. As k[xα, yα]k 6 kxαkkyαk, see 142 III, it is also a norm-bounded net. By 77 I norm-bounded Cauchy nets in the ultraweak uniformity converge, so uwlimα[xα, yα] exists. Assume we are given any Φ-indexed norm-bounded Cauchy nets (x′)α and (y′)α over W with (x′α)α → x and (y′α)α → y. Then, like (2.8): (cid:12)(cid:12)f ([xα, yα] − [x′α, y′α])(cid:12)(cid:12) = (cid:12)(cid:12)[xα, yα − y′α]f + [xα − x′α, y′α]f(cid:12)(cid:12) 6 kxαkfkyα − y′αkf + kxα − x′αkfky′αkf . From this it follows uwlimα[xα, yα] = uwlimα[x′α, y′α]. We are justified to define [x, y] ≡ uwlim α [xα, yα]. We will show [· , · ] is a B-valued inner product. Pick any x, y, z ∈ σ(W ) and Φ-indexed norm-bounded Cauchy nets (xα)α, (yα)α and (zα)α over W with xα → x, yα → y and zα → z in V . So xα + yα → x + y. With ultraweak continuity of addition [x + y, z] = uwlim [xα + yα, zα] = uwlim [xα, zα] + uwlim [yα, zα] = [x, z] + [y, z]. α α α In a similar fashion one proves the other axioms of a B-valued inner product: [x, y] = [y, x]∗ follows from the ultraweak continuity of (· )∗; B-homogeneity follows from ultraweak continuity of c 7→ cb (see 45 IV) and [x, x] > 0 follows from ultraweak-closedness of the positive cone of B. To prove definiteness of [· , · ], assume [x, x] = 0 for some x ∈ σ(W ). Pick some Φ-indexed norm-bounded Cauchy net (xα)α over W with xα → x. For any np-map f : B → C we have limα f ([xα, xα]) = f (uwlimα[xα, xα]) = f ([x, x]) = 0. Thus xα converges in the ultranorm uniformity on W to 0. Hence x = 0. Thus σ(W ) is a pre-Hilbert B-module. It remains to be shown that k · kσ(W ) f agree. Pick any x ∈ σ(W ) and f : B → C. By definition there is a and k · kV norm-bounded Cauchy net xα over W for which xα → x in V . Then f kxkσ(W ) f = f (uwlim α kxαkV as desired. We have shown that σ(W ) is a compatible extension. α kxαkW f = lim α [xα, xα]) 2 = lim 1 f = kxkV f , ..150.. 49 XIV (Induction limit step) Assume V is a non-empty chain of compatible extensions. See XI for the order on compatible extensions. Write W = S V . We will turn W into a compatible extension. Clearly W is a submodule of V . As V is not empty, V0 ⊆ W . For any v, w ∈ W there is some W ′ ∈ V with v, w ∈ W ′. Define [v, w]W = [v, w]W ′ . This turns W into a pre-Hilbert B-module as all its axioms only involve finitely many elements, which are contained in some single compatible extension far enough up in the chain. Finally, to see the seminorms agree, pick any np-map f : B → C and x ∈ W . Then x ∈ W ′ for some W ′ ∈ V . Consequently kxkV f = kxkW ′ 1 2 = kxkW f , f = f ([x, x]W ′ ) 2 = f ([x, x]W ) 1 as desired. We have shown W is a compatible extension. XV (Self-duality) We have seen that every non-empty chain of compatible exten- sions has an upper bound. Thus by Zorn's lemma there is a maximal compati- ble extension W ⊆ V . By maximality σ(W ) = W and so in W norm-bounded ultranorm-Cauchy nets converge. Thus by 149 V W must be self-dual and ultra- norm complete. By completeness W = V . Define X = V and hx, yi = [x, y]W . Clearly [v, w] = hη(v), η(w)i and the image of η is norm dense. (cid:3) 151 The completion 150 II we just constructed has a universal property, which can be seen as a generalization of [93, prop. 3.6]. Ia II Lemma Assume B is a von Neumann algebra and V is a right B-module with B-valued inner product. Let X be any self-dual Hilbert B-module together with inner product preserving B-linear map η : V → X such that the image of η is ultranorm dense in X. (E.g. the completion into a self-dual Hilbert B-module from 150 II.) Then η has the following universal property. Let Y be a self-dual Hilbert B-module. For each bounded B-linear map T : V → Y , there is a unique bounded B-linear map T : X → Y such that T ◦ η = T . Proof By 148 I bounded B-linear maps are ultranorm uniformly continuous. As η(V ) is ultranorm dense in X, the extension T is unique if it exists. Pick any x ∈ X. There is a net xα in V such that η(xα) → x ultranorm. So η(xα) is ultranorm Cauchy. As η preserves inner product, xα is ultranorm Cauchy as well. As T is ultranorm uniformly continuous T (xα) is ultranorm Cauchy. With the same motions we see T (xα) ∼ T (x′α) when xα ∼ x′α. Using this and that Y is ultranorm complete, see 149 V, we may define T x = unlimα xα, where unlim denotes the limit in the ultranorm uniformity. Clearly T ◦ η = T . It is straightforward to check that B-linearity of T follows from B-linearity of T and ultranorm continuity of addition and the B-module action. It remains to be shown T is bounded. By 144 V we have hT xα, T xαi 6 kTk2[xα, xα] = kTk2 hη(xα), η(xα)i and so using 148 V we see that hT xα, T xαi 6 kTk2 uwlim α α hη(xα), η(xα)i = kTk2 hx, xi , h T x, T xi = uwlim which implies that k Tk 6 kTk as desired. (In fact k Tk = kTk.) Before we show that every ncp-map has a Paschke dilation, we study Ba(X) a bit more for self-dual X. We start with B-sesquilinear forms. Definition Let V be a normed right B-module. A sesquilinear form B : V ×V → B is said to be bounded if there is an r > 0 such that kB(x, y)k 6 rkxkkyk for all x, y ∈ V . Example Let X be a pre-Hilbert B-module. For every T ∈ Ba(X) the map h(· ), T (· )i is a bounded B-sesquilinear form. For self-dual X every bounded B-sesquilinear form arises in this way. (cid:3) Proposition Let X be a self-dual Hilbert B-module. For every bounded B- sesquilinear form B on X, there is a unique T ∈ Ba(X) with B(x, y) = hx, T yi for all x, y ∈ X. Proof For each y ∈ X the map B(· , y)∗ is B-linear and bounded. Thus by self- duality of X there is a unique ty for each y ∈ X such that hty, xi = B(x, y)∗ for all x ∈ X. It is easy to see hty+y′, xi = hty + ty′ , xi and htyb, xi = htyb, xi, and so by uniqueness y 7→ ty is B-linear. Define T y = ty. Clearly hx, T yi = hty, xi∗ = B(x, y) and T must be the unique such B-linear map. Reasoning in the same way, we see there is a unique S with B(y, x)∗ = hx, Syi. Hence hx, T yi = B(x, y) = hy, Sxi∗ = hSx, yi, so T is adjointable. It remains to be shown T is bounded. There is an r > 0 such that for all x ∈ X: kT xk2 = k hT x, T xik = kB(T x, x)k 6 rkT xkkxk. 152 II III IV V VI So r is a bound by dividing out kT xk if kT xk 6= 0 and trivially otherwise. (cid:3) Remark In 36 V a more general result was shown. Exercise Suppose T : X → Y is a bounded B-linear map between Hilbert B- modules. Show that if X is self dual, then T is adjointable. [93, prop. 3.4] Exercise Let V be a right B-module with B-valued inner product for some von Neumann algebra B and η : V → X be the ultranorm completion of V from 150 II. Show that VII VIII IX { hx, (· ) xi : X → B ; x ∈ V } (where x ≡ η(x)) is an order separating set of ncp-maps, see 21 II. (That is: T > 0 iff hx, T xi > 0 for all x ∈ V . Consequently S = T iff hx, T xi = hx, S xi for all x ∈ V .) ..150 -- 152.. 51 X Theorem Suppose X is a self-dual Hilbert B-module for a von Neumann alge- bra B. Then Ba(X) is a von Neumann algebra. XII XI Proof The theorem is due to Paschke [93, prop. 3.10]; we give a new proof, which also appeared as 49 II. We already know that Ba(X) is a C∗-algebra, see 143 IV. (bounded order completeness) Let (Tα)α be a norm-bounded net of self-adjoint elements of Ba(X). We have to show it has a supremum. Pick x ∈ X and r ∈ R, r > 0 such that kTαk 6 r for all α. By 144 I (hx, Tαxi)α is a norm-bounded net of self-adjoint elements of B and so it has a supremum to which it converges ultrastrongly by 44 XIV. In particular 1 4 3Xk=0 ik(cid:10)ikx + y, Tα(ikx + y)(cid:11) = hx, Tαyi converges ultrastrongly for all x, y ∈ X. Define B(x, y) = uslimα hx, Tαyi. As addition and multiplication by a fixed element are ultrastrongly continuous (see 45 IV), we see that B is a B-sesquilinear form. By 142 III and 144 V we have hTαy, xihx, Tαyi 6 r2kxk2 hy, yi for each α, hence kB(x, y)k2 = kB(x, y)∗B(x, y)k 46 II= k uwlim 6 r2kxk2kyk2. α hTαy, xihx, Tαyik XIII So B is a bounded B-sesquilinear form. By V there is a unique T ∈ Ba(X) with hx, T yi = B(x, y) = uslimα hx, Tαyi. Noting B(x, y) = uwlimαhx, Tαyi, it is easy to see T is self-adjoint. Using 144 I and the fact that ultraweak limits respect the order (44 XI), we see hx, Tαxi 6 hx, T xi for all x ∈ X and so Tα 6 T . Now suppose we are given a self-adjoint S ∈ Ba(X) with Tα 6 S for all α. With a similar argument we see T 6 S and so S is the supremum of (Tα)α. (separating normal states) By 144 I, the states hx, (· )xi are separating. We are done if we can show that hx, (· )xi is normal, i.e. preserves suprema of bounded directed sets of self-adjoint elements. To this end, pick x ∈ X and let (Tα)α be a net with supremum T . As before (hx, Tαxi)α) is a norm bounded net which converges ultrastrongly to its supremum. Now, we just saw α hx, Tαxi hx, T xi = uwlim hx, Tαxi = sup α and so indeed hx, (· )xi is normal; the normal states are separating and conse- quently Ba(X) is a von Neumann algebra. (cid:3) 153 Proposition Assume T : X → Y is an adjointable bounded module map between Hilbert B-modules. Define adT : Ba(Y ) → Ba(X) by adT (B) = T ∗BT . The map adT is completely positive. If X and Y are self-dual, then adT is normal. II III IV 154 II III Proof For any n ∈ N , B1, . . . , Bn ∈ Ba(X) and A1, . . . , An ∈ Ba(Y ) we have Xi,j B∗j T ∗A∗j AiT Bi = (cid:0)Xi AiT Bi(cid:1)∗(cid:0)Xj AjT Bj(cid:1) > 0 and so adT is completely positive. Now we show adS is normal. By 48 II and 144 I, it suffices to show that for every x ∈ X, the map T 7→ hx, adS(T )xi = hSx, T Sxi is normal, which it indeed is by 152 XIII. (cid:3) Exercise Let A be a C∗-algebra. Assume n ∈ N and a1, . . . , an ∈ A . Use I to show ϕ : A → MnA given by ϕ(d) = (a∗i daj)ij is an ncp-map. 2.3 Paschke dilations We are ready to show that every ncp-map has a Paschke dilation. That is: for every ncp-map ϕ there is a triple (P, , h) with ϕ = h◦ via P for some ncp-map h, nmiu-map and von Neumann algebra P, which is minimal in the sense that for any such triple (P′, ′, h′) there is a unique ncp-map σ : P′ → P with h′ = h◦ σ and = σ ◦ ′. See 140 II. Definition Let ϕ : A → B be any ncp-map between von Neumann algebras. A complex bilinear map B : A ×B → X, where X is a self-dual Hilbert B-module is called ϕ-compatible if there is an r > 0 such that for all n ∈ N, a1, . . . , an ∈ A and b1, . . . , bn ∈ B we have (cid:13)(cid:13)Xi B(ai, bi)(cid:13)(cid:13)2 6 r ·(cid:13)(cid:13)Xi,j b∗i ϕ(a∗i aj)bj(cid:13)(cid:13) (2.9) and B(a, b1)b2 = B(a, b1b2) for all a ∈ A and b1, b2 ∈ B. Theorem Let ϕ : A → B be any ncp-map between von Neumann algebras. 1. There is a self-dual Hilbert B-module A ⊗ϕ B and ϕ-compatible bilinear ⊗ : A × B → A ⊗ϕ B such that for every ϕ-compatible bilinear map B : A × B → Y there is a unique bounded module map T : A ⊗ϕB → Y such that T (a⊗b) = B(a, b) for all a ∈ A and b ∈ B. 2. For a0 ∈ A , there is a unique (a0) ∈ Ba(A ⊗ϕ B) fixed by (a ∈ A , b ∈ B) (a0)(a ⊗ b) = (a0a) ⊗ b. Furthermore a 7→ (a) yields a nmiu-map : A → Ba(A ⊗ϕ B). ..152 -- 154.. 53 3. The map h : Ba(A ⊗ϕ B) → B given by h(T ) = h1 ⊗ 1, T (1 ⊗ 1)i is ncp. 4. The map : A → Ba(A ⊗ϕ B) together with 1 ⊗ 1 ∈ A ⊗ϕ B has the following universal property. Let ′ : A → Ba(X) be an nmiu-map for a self-dual Hilbert B- module X together with an element e ∈ X such that ϕ = h′ ◦ ′ for h′(T ) ≡ he, T ei. Then: there is a unique inner product preserving B-linear map S : A ⊗ϕB → X such that adS ◦ ′ = and S(1 ⊗ 1) = e. 5. (Ba(A ⊗ϕ B), , h) is a Paschke dilation of ϕ, see 140 II. IV Proof (This is a simplified version of the proof we published earlier in [127]. The construction of A ⊗ϕ B is essentially due to Paschke [93, thm. 5.2]: the self-dual completion of X in [93, thm. 5.2] is isomorphic to A ⊗ϕ B.) (1: A ⊗ϕ B) The algebraic tensor product A ⊙ B is a right B-module with the action (Pi ai ⊗ bi)β =Pi ai ⊗ (biβ). On A ⊙ B, define (cid:2)Xi ai ⊗ bi,Xj αj ⊗ βj(cid:3) ≡ Xi,j b∗i ϕ(a∗i αj)βj. V VI By complete positivity of ϕ this is a B-valued inner product on A ⊙ B. By 150 II, there is a self-dual Hilbert B-module A ⊗ϕB and B-linear inner product- preserving η : A ⊙ B → A ⊗ϕ B with ultranorm-dense range. Define a bilinear map ⊗ : A × B → A ⊗ϕ B by a ⊗ b = η(a ⊗ b). By definition we have (cid:13)(cid:13)Xi ai ⊗ bi(cid:13)(cid:13)2 = (cid:13)(cid:13)(cid:2)Xi ai ⊗ bi,Xj aj ⊗ bj(cid:3)(cid:13)(cid:13) = (cid:13)(cid:13)Xi,j b∗i ϕ(a∗i aj)bj(cid:13)(cid:13) and so ⊗ is a ϕ-compatible bilinear map. Let B : A × B → Y be a ϕ-compatible bilinear map to some self-dual Hilbert B-module Y . We must show that there is a unique bounded module map T : A ⊗ϕ B → Y such that T (a⊗ b) = B(a, b) for all a ∈ A and b ∈ B. By the defining property of the algebraic tensor product, there is a unique linear map T0 : A ⊙ B → Y such that T0(a⊗ b) = B(a, b) for all a ∈ A and b ∈ B. By definition of ϕ-compatibility and the inner product on A ⊙ B, the map T0 is in fact bounded and B-linear. By 151 Ia this map extends uniquely to a bounded module map T : A ⊗ϕ B → Y with T (a ⊗ b) = B(a, b), as desired. (2: : A → Ba(A ⊗ϕ B)) Let a0 ∈ A be given. To show the module map (a0) exists, it suffices to show that the bilinear map B : A ×B → A ⊗ϕ B given by B(a, b) = (a0a) ⊗ b is ϕ-compatible. Clearly B(a, bβ) = B(a, b)β. let n ∈ N, a1, . . . , an ∈ A and b1, . . . , bn ∈ B be given. To prove (2.9), The row vector (a1 ··· an) is an A -linear map a : A ⊕n → A in the usual way: a(α1, . . . , αn) = a1α1 +···+anαn. Similarly the column vector (b1 . . . bn)t is a B-linear map b : B → B⊕n with b(β) = (b1β, . . . , bnβ). We compute (cid:13)(cid:13)Xi B(ai, bi)(cid:13)(cid:13)2 = (cid:13)(cid:13)Xi = (cid:13)(cid:13)Xi,j = ka0k2(cid:13)(cid:13)Xi,j (a0ai) ⊗ bi(cid:13)(cid:13)2 b∗i ϕ(a∗i a∗0a0aj)bj(cid:13)(cid:13) b∗i ϕ(a∗i aj)bj(cid:13)(cid:13). = kb∗(Mnϕ)(a∗a∗0a0a)bk 6 ka∗0a0kkb∗(Mnϕ)(a∗a)bk Thus B is ϕ-compatible and so there is a unique B-linear bounded module map (a0) : A ⊗ϕ B → A ⊗ϕ B with (a0)(a ⊗ b) = (a0a) ⊗ b. Clearly is a unital, multiplicative and involution preserving map. It remains to be shown is normal. By 48 II and 152 IX, it suffices to show that d 7→ hx, (d)xi is normal for every x ∈ A ⊙ B. Find n ∈ N, row vector a ≡ (a1 ··· an) and column vector b ≡ (b1 ··· bn) such that x =Pn hx, (d)xi = Xi,j i=1 ai ⊗ bi. We compute b∗i ϕ(a∗i (d)aj )bj = b∗(Mnϕ)(a∗ d a)b. So d 7→ hx, (d)xi is normal by 153 I and 49 IV, whence is normal. (3: h : Ba(A ⊗ϕ B) → B) Define h(T ) ≡ h1 ⊗ 1, T 1 ⊗ 1i. It is completely positive by 145 I and normal by 152 XIII. (Uniqueness σ) Before we continue with 4, we will already prove the uniqueness property for point 5. Note that (h◦ )(a) = h1 ⊗ 1, a ⊗ 1i = ϕ(a) for all a ∈ A and so ϕ = h◦ . Assume ϕ = h′ ◦ ′ for some nmiu-map ′ : A → P′, ncp- map h : P′ → B and von Neumann algebra P′. For point 5 we must show there is a unique ncp-map σ : P′ → Ba(A ⊗ϕ B) such that h◦ σ = h′ and σ ◦ ′ = . Let σ1, σ2 : P → Ba(A ⊗ϕ B) be ncp-maps with h◦ σk = h′ and σk ◦ ′ = , k = 1, 2. We must show σ1 = σ2. Let c ∈ P′ and x ∈ A ⊙B be given. By 152 IX it suffices to prove that hx, σ1(c)xi = hx, σ2(c)xi. Find n ∈ N, a1, . . . , an ∈ A for any k = 1, 2, we find and b1, . . . , bn ∈ B such that x =Pi ai ⊗ bi. Note ai ⊗ bi = (ai)(1 ⊗ 1)bi. So VII VIII hx, σk(c)xi = Xi,j = Xi,j = Xi,j b∗i h( (a∗i )σk(c)(aj ) )bj b∗i h(σk( ′(a∗i )c′(aj) ))bj by 139 III b∗i h′(′(a∗i )c′(aj))bj. ..154.. (2.10) 55 Thus σ1 = σ2 as desired. One can show such σ exists by using (2.10) as defining formula. We will use, however, an indirect but shorter approach by first considering the 'spatial case', which was suggested by Michael Skeide. (4: spatial case) Let X be a self-dual Hilbert B-module together with e ∈ X and nmiu-map A → Ba(X) such that ϕ = h′ ◦ ′ with h′(T ) = he, T ei. To prove uniqueness, assume that for k = 1, 2, we have B-linear inner product preserv- ing Sk : A ⊗ϕ B → X with adSk ◦ ′ = and Sk 1 ⊗ 1 = e. Then h◦ adSk = h′ and so (as adSk is an ncp-map by 153 I), we know adS1 = adS2 by VIII. Hence there is a λ ∈ C, λ 6= 0 with S1 = λS2, cf. [126, lemma 9]. If e 6= 0, then from e = S1 1 ⊗ 1 = λS2 1 ⊗ 1 = λe it follows λ = 1 and so S1 = S2. In the other case, if e = 0, then h′ = 0 and so ϕ = 0, whence A ⊗ϕ B = {0} and indeed S1 = S2 = 0 trivially. We continue with existence. There is a unique linear S0 : A ⊙ B → X fixed by S0(a⊗ b) = ′(a)eb. Clearly S0 is also B-linear. For any x, y ∈ A ⊙ B, say x =Pi ai ⊗ bi and y =Pj αj ⊗ βj, we have hS0x, S0yi = Xi,j = Xi,j = Xi,j = [x, y]. h′(ai)ebi, ′(αj )eβji b∗i he, ′(a∗i αj)ei βj b∗i ϕ(a∗i αj)βj IX X Thus S0 preserves the inner product. In particular S0 is bounded and so there is a unique bounded B-linear S : A ⊗ϕ B → X with S ◦ η = S0. For all x ∈ A ⊙B we have hx, S∗S xi = hx, xi and so S∗S = 1 by 152 IX. Thus S preserves the inner product. By definition S(1 ⊗ 1) = ′(1)e1 = e. Pick any a ∈ A . A straightforward computation shows S(a) = ′(a)S hence S∗′(a) = (a)S∗ and S∗′(a)S = (a)S∗S = (a). We have shown adS ◦ ′ = adS. (5, σ existence) Assume ϕ = h′ ◦ ′ for some nmiu-map ′ : A → P′, ncp- map h : P′ → B and von Neumann algebra P′. It remains to be shown there is a σ : P′ → Ba(A ⊗ϕ B) with σ ◦ ′ = and h◦ σ = h′. To apply the previous point, we perform the whole construction for h′ instead of ϕ yielding h′ = hh′ ◦ h′ with h′ : P′ → Ba(P′ ⊗h′ B) and hh′ : Ba(P′ ⊗h′ B) → B. By IX there is a unique B-linear map S : A ⊗ϕ B → P′ ⊗h′ B with S∗S = 1, S1 ⊗ 1 = 1 ⊗ 1 and adS ◦ h′ ◦ ′ = . Define σ = adS ◦ h′ . This σ fits the bill: σ ◦ ′ = adS ◦ h′ ◦ ′ = and h◦ σ = h◦ adS ◦ h′ = hh′ ◦ h′ = h′. (cid:3) 155 We have shown that any ncp-map ϕ : A → B admits a dilation using Paschke's generalization of GNS to Hilbert C∗-modules. There is also a generalization of Stinespring's theorem to Hilbert C∗-modules due to Kasparov [75]: Theorem (KSGNS) Let ϕ : A → Ba(X) be a cp-map for some C∗-algebras A , B and Hilbert B-module X. There exists a Hilbert B-module Y , miu-map : A → Ba(Y ) and adjointable B-linear T : Y → X with ϕ = adT ◦ . For our purposes (the dilation of an arbitrary ncp-map A → B), Paschke's GNS construction sufficed. We leave open whether KSGNS admits a universal property like 139 V and whether it is a Paschke dilation. II III We have all the tools to characterize the ncp-maps ϕ for which the Paschke representation is injective. This is a generalization of our answer [125] to the same question for the Stinespring embedding. Theorem Let ϕ : A → B be any ncp-map with Paschke dilation (P, , h). Then ⌈⌉ = ⌈⌈ϕ⌉⌉. (⌈⌈ϕ⌉⌉ is the central carrier of ϕ, see 69 I.) Thus is injective (⌈⌉ = 1) if and only if ϕ maps no non-zero central projection to zero (⌈⌈ϕ⌉⌉ = 1). Proof (This is a simplified version of the proof we published earlier in [93, thm. 30].) By 140 VIII, it is sufficient to prove the equivalence for the dilation constructed in 154 III. Let p ∈ A be any projection. We will show p 6 ⌈⌉⊥ iff p 6 ⌈⌈ϕ⌉⌉⊥. Assume p 6 ⌈⌉, viz. (p) = 0. Pick any a ∈ A and b ∈ B. Then (p) a ⊗ b = 0. By definition of and the inner product, this is the case iff b∗ϕ(a∗p∗pa)b = 0. In particular ϕ(a∗pa) = 0. In other words a∗pa 6 ⌈ϕ⌉⊥. This happens if and only if a⌈ϕ⌉ a∗ 6 p⊥ by 55 V. Equivalently ⌈a⌈ϕ⌉ a∗⌉ 6 p⊥. As a was arbitrary, we see (using 68 I), that ⌈⌈ϕ⌉⌉ =Sa∈A ⌈a⌈ϕ⌉ a∗⌉ 6 p⊥, as desired. For the converse, note that the two implications we just used are, in if ϕ(a∗pa) = 0, then b∗ϕ(a∗pa)b = 0 for any b ∈ B and fact, equivalences: if (p) a ⊗ b = 0 for all a ∈ A and b ∈ B, then (p) = 0, by 152 IX. (cid:3) We will now prove a useful connection between the ncp-maps ncp-below ϕ (as defined in a moment) and the commutant of the image of a Paschke representa- tion ("") of ϕ. We will use it to give an alternative proof of the fact that pure maps and nmiu-maps are extreme among all ncpu-maps. Definition For linear maps ϕ, ψ : A → B, we say ϕ is ncp-below ψ (in sym- bols: ϕ 6ncp ψ) whenever ψ − ϕ is an ncp-map. Furthermore, write [0, ϕ]ncp = {ψ; ψ : A → B; 0 6ncp ψ 6ncp ϕ}. Definition Let ϕ : A → B be any ncp-map with Paschke dilation (P, , h). For t ∈ (A )(cid:3), the commutant of (A ), define ϕt = h(t(a)). The following is a generalization [93, prop. 5.4] to arbitrary Paschke dilations. Theorem Assume ϕ : A → B is an ncp-map with Paschke dilation (P, , h). The map t 7→ ϕt is a linear order isomorphism [0, 1](A )(cid:3) → [0, ϕ]ncp. ..154 -- 157.. 57 156 II III 157 II III IIIa IV V VI VII VIII Proof First we show the correspondence holds for the Paschke dilation con- structed in 154 III. This proof is a slight variation on [93, prop. 5.4]. Then we show the correspondence carries over to arbitrary Paschke dilations. (Set-up) Let (Ba(A ⊗ϕB), , h) denote the Paschke dilation constructed in 154 III. Clearly T 7→ ϕT is linear for T ∈ (A )(cid:3). Pick T ∈ (A )(cid:3). As √T ∈ (A )(cid:3) we have ϕT (a) = h(√T (a)√T ). Thus ϕT is ncp. In particular, if T 6 S for some S ∈ (A )(cid:3), then ϕS−T = ϕS − ϕT is ncp and so ϕT 6ncp ϕS. Conse- quently, if T 6 1, then ϕT 6 ϕ1 = ϕ. (Order embedding) Let T ∈ (A )(cid:3) be given such that ϕT is ncp. We must show T > 0. Pick any x ∈ A ⊙ B. By 152 IX and the construction of A ⊗ϕ B, it is sufficient to show hx, T xi > 0. Say x ≡Pi ai ⊗ bi. Then hx, T xi = Xi,j b∗i h1 ⊗ 1, T (a∗i aj)1 ⊗ 1i bj = Xi,j b∗i ϕT (a∗i aj)bj > 0. Thus T 7→ ϕT is an order embedding and in particular an injection. (Surjectivity) Pick any ψ ∈ [0, ϕ]ncp. We will show there is a T ∈ (A )(cid:3) with ϕT = ψ and 0 6 T 6 1. Let (Ba(A ⊗ψ B), ψ, hψ) denote the Paschke dilation of ψ constructed in 154 III. Define B : A × B → A ⊗ψ B by B(a, b) = a ⊗ b. This B is ϕ-compatible -- indeed, for any n ∈ N, a1, . . . , an ∈ A and b1, . . . , bn ∈ B, it follows from ψ 6ncp ϕ that (cid:13)(cid:13)Xi B(ai, bi)(cid:13)(cid:13)2 = (cid:13)(cid:13)Xi,j b∗i ψ(a∗i aj)bj(cid:13)(cid:13) 6 (cid:13)(cid:13)Xi,j b∗i ϕ(a∗i aj)bj(cid:13)(cid:13). So by 154 III, there is a unique bounded module map W : A ⊗ϕ B → A ⊗ψ B fixed by W a⊗b = a⊗b. As A ⊗ψ B is self-dual, the map W has an adjoint W ∗ by 152 VIII. For any a ∈ A , we have W (a) = ψ(a)W and so (a)W ∗ = W ∗ψ(a). Thus W ∗W (a) = W ∗ψ(a)W = (a)W ∗W . Apparently W ∗W ∈ (A )(cid:3). the inequality hx, xiψ 6 hx, xiϕ and so Define T ≡ W ∗W . Clearly 0 6 T . As ψ 6ncp ϕ, we have for each x ≡Pi ai⊗bi, hx, T xiϕ = hW x, W xiψ = hx, xiψ 6 hx, xiϕ. Thus T 6 1 by 152 IX. For any a ∈ A we have ϕT (a) = h1 ⊗ 1, W ∗W a ⊗ 1iϕ = hW 1 ⊗ 1, W a ⊗ 1iψ = h1 ⊗ 1, a ⊗ 1iψ = ψ(a), as desired. We have shown the correspondence holds for the Paschke dila- tion (Ba(A ⊗ϕ B), , h) constructed in 154 III. (Arbitrary dilation) Let (P′, ′, h′) be any Paschke dilation of ϕ. For brevity write P ≡ A ⊗ϕ B and for clarity write ϕP to distinguish between ϕT and ϕt for T ∈ (A )(cid:3) and t ∈ ′(A )(cid:3). By 140 VIII there is a unique nmiu- isomorphism ϑ : P′ → P such that ϑ◦ ′ = and h◦ ϑ = h′. It is easy to see ϑ restricts to a linear order isomorphism [0, 1]′(A )(cid:3) → [0, 1](A )(cid:3). For any t ∈ ′(A )(cid:3) and a ∈ A we have T and ϕP′ t ϕP′ t (a) = h′(t′(a)) = h(ϑ(t′(a))) = h(ϑ(t)(a)) = ϕP ϑ(t)(a) t and so t 7→ ϕP′ linear order isomorphisms ϑ−1 and t 7→ ϕP t . We return to t 7→ ϕt in 172, where we connect it to extremeness. is a linear order isomorphism as it is the composition of the (cid:3) IX X 2.4 Kaplansky density theorem for Hilbert C∗-modules 158 Before we continue our study of Paschke dilations, we need to develop some more theory on (self-dual) Hilbert C∗-modules. First we will prove a generalization of the Kaplansky density theorem for Hilbert C∗-modules, which we will need to compute the tensor product of Paschke dilations. The original density theorem (74 IV) is an important foundational result in the theory of operator algebras; quoting Pedersen: "The density theorem is Kaplansky's great gift to mankind. It can be used every day, and twice on Sundays." [95, §2.3.4]. We used the original theorem already several times; notably to show that a von Neumann algebra is ultrastrongly complete (77 I via 75 VIII). This ultrastrong completeness was used to define operator divi- sion (81 I) (and in turn filters 96 I), to give the spatial tensor product a universal property (112 XI) and in this thesis to show ultranorm completeness of self-dual Hilbert C∗-modules (149 V). So what does is this density theorem again? In essence, it is a remedy for the fact that strongly converging nets might not be bounded. The theorem is usually stated as follows: the unit ball of the SOT-closure of a self-adjoint algebra of operators A is contained in the SOT-closure of the unit ball of A . For our generalization, it is more convenient to consider the following variation. Kaplansky density theorem Let B be a von Neumann algebra with ultrastrongly- dense C∗-subalgebra A ⊆ B. Then: for every element a ∈ A , there is a net bα in B with bα → a ultranorm and kbαk 6 kbk for all α. Now we are ready to state and prove our generalization, which is inspired by the proof of the regular Kaplansky density theorem given in [7, thm. 1.2.2]. Cf. 74 IV. Ia Ib ..157 -- 158.. 59 II Kaplansky density theorem for Hilbert C∗-modules Let X be a Hilbert B- module for a von Neumann algebra B with an ultranorm-dense A -submodule D ⊆ X, where A ⊆ B is some C∗-subalgebra with hy, yi ∈ A for all y ∈ D. and kxαk 6 kxk for all α. Proof Let x ∈ X be given. The case x = 0 is trivial. Assume x 6= 0. Without loss of generality, we may assume kxk = 1. For this x and any y ∈ X, define Then: for every element x ∈ X, there is a net xα in D with xα → x ultranorm III h(y) ≡ y · 2 1 + hy, yi g(x) ≡ x · 1 . 1 +p1 − hx, xi We claim h is ultranorm continuous, h(g(x)) = x and kh(y)k 6 1 for all y ∈ X. From this, the promised result follows. Indeed, let xα be a net in D that converges ultranorm to g(x). Then h(xα) is still in D and furthermore h(xα) → h(g(x)) = x ultranorm and kh(xα)k 6 1 = kxk, as desired. IV We start with kh(y)k 6 1. The real map hr : λ 7→ 4λ(1 + λ)−2 is bounded by 1 and so by the functional calculus we see kh(y)k2 = k4hy, yi(1 + hy, yi)−2k = khr(hy, yi)k 6 1. We continue with the proof of h(g(x)) = x. Note hg(x), g(x)i = (1 +p1 − hx, xi)−2hx, xi and so with basic algebra 1 h(g(x)) = x · 1 +p1 − hx, xi = x · 2 (cid:0)1 +p1 − hx, xi(cid:1)2 1 +p1 − hx, xi = x. 1 +(cid:0)1 +p1 − hx, xi(cid:1)−2 + hx, xi 2 !−1 hx, xi V (Ultranorm continuity h) The real work is in the proof of the ultranorm continuity of h. Suppose yα → y ultranorm in X. Define 1 1 + hyα, yαi hy, yi 1 + hy, yi hyα, yαi 1 + hyα, yαi 1 + hy, yi 1 1 A1 ≡ A′1 ≡ A2 ≡ A′2 ≡ hy, yi hyα, yαi (1 + hy, yi)2 − (1 + hyα, yαi)2 − 1 + hy, yihyα − y, yαi 1 + hyα, yαihy − yα, yi 1 1 1 + hyα, yαi . 1 1 + hy, yi A straightforward computations shows hh(y) − h(yα), h(y) − h(yα)i = A1 + A′1 + A2 + A′2. (2.11) We will show that each of these terms converges ultraweakly to 0 as α → ∞. We start with A1. We need two facts. First, for any positive b ∈ B, we have 0 6 1 1+b 6 1 and 0 6 b 1+b 6 1. Secondly Putting these facts and the definition of A1 together, we get 1 1 + hy, yi − 1 1 + hyα, yαi = = 1 1 + hyα, yαi . 1 1 1 + hy, yi(cid:0)(1 + hyα, yαi) − (1 + hy, yi)(cid:1) 1 + hy, yi(cid:0)hyα, yαi − hy, yi(cid:1) 1 + hyα, yαi(cid:19) hy, yi 1 + hy, yi 1 + hyα, yαi 1 1 A1 = (cid:18) 1 = = 1 1 + hy, yi − 1 + hy, yi(cid:0)hyα, yαi − hy, yi(cid:1) 1 + hy, yi(cid:0)hyα − y, yαi + hy, yα − yi(cid:1) 1 1 1 + hyα, yαi hy, yi 1 + hy, yi 1 1 + hyα, yαi hy, yi 1 + hy, yi . As multiplying with constants is ultraweakly continuous (see 45 IV), it is suffi- cient to show that ultraweakly, as α → ∞, we have hy, yα − yi hyα − y, yαi (2.12) 1 1 1 + hyα, yαi → 0. 1 + hyα, yαi → 0 For any np-map f : B → C, we have (cid:12)(cid:12)(cid:12)f(cid:16)hyα − y, yαi 1 1 + hyα, yαi(cid:17)(cid:12)(cid:12)(cid:12) 2 = hyα − y, yα(1 + hyα, yαi)−1if2 f · kyα(1 + hyα, yαi)−1k2 6 kyα − yk2 f · kyα(1 + hyα, yαi)−1k2 · kfk 6 kyα − yk2 6 kyα − yk2 f · kfk → 0. f We have shown the LHS of (2.12). The proof for the RHS is different, but sim- pler. So A1 vanishes ultraweakly. In a similar way one sees A′1 → 0 ultraweakly. The proofs for A2, A′2 → 0 are very similar. (cid:3) 2.5 More on self-dual Hilbert C∗-modules We have a second look at self-dual Hilbert C∗-modules over von Neumann al- gebras and their orthonormal bases. We start with some good news. In B(H ) the linear span of eiihej is ultraweakly dense for any orthonormal basis (ei)i∈I of H . For self-dual Hilbert B-modules, we have a similar result. 159 ..158 -- 159.. 61 II Definition Let X be a Hilbert B-module. For any x, y ∈ X, define the bounded operator xihy ∈ Ba(X) by xihyz = xhy, zi. instance [88, §2.2]. IIa Remark The operator xihy is usually denoted by θx,y in the literature, see for III The following rules are easy to check. For x, y, v, w ∈ X and b ∈ B we have xihyvihw = xhy, viihw . xbihy = xihyb∗ xihy∗ = yihx If he, ei is a projection, then eihe is a projection (using 149 III). For any T ∈ Ba(X) we have T xihy = T xihy and xihy T ∗ = xihT y. IV Proposition Let X be a self-dual Hilbert B-module for a von Neumann al- gebra B. If (ei)i∈I is an orthonormal basis of X, then the linear span of the operators {eibihej ; i, j ∈ I, b ∈ B} is ultraweakly dense in Ba(X). Proof We start with some preparation. VI For a finite subset S ⊆ I, write pS =Pi∈S eiihei. The (eiihei)i∈I are pairwise orthogonal projections. Thus pS is a projection. Also pS 6 pS′ when S ⊆ S′. V Pick any x ∈ X. As pSx =Pi∈S ei hei, xi, we have hx, pSxi = hpSx, pSxi = Xi,j∈S hx, eii hei, ejihej, xi = Xi∈S hx, eiihei, xi , and so by Parseval supS hx, pSxi = hx, xi. Hence supS pS = 1 by 22 VIII and so pS converges ultraweakly to 1 by 44 XIV. VII Pick any T ∈ Ba(X). For i, j ∈ I we have eiihei T ejihej = eiiheiT ejihej = ei hei, T ejiihej . and so pST pS is in the linear span of {eibihej}. Thus it is sufficient to prove pST pS converges ultraweakly to T . VIII Pick any np-map f : B → C. We want to show f (T − pST pS) → 0. Clearly f (T − pST pS) 6 f ( (1 − pS)T ) + f ( pST (1 − pS) ). Using Cauchy -- Schwarz and VI we see the second term vanishes f (pST (1 − pS))2 6 f (pST T ∗pS)f (1 − pS) 6 kTk2f (1)f (1 − pS) → 0. Similarly the first term vanishes and so indeed pST pS → T ultraweakly. (cid:3) IX Proposition Let X be a self-dual Hilbert B-module for a von Neumann alge- bra B. If xα → x ultranorm for a norm-bounded net xα in X, then xαihy → xihy ultraweakly for any y ∈ Y . Proof We need some preparation. Define X Ω ≡ {f (hx, (· )xi); f : B → C np-map, x ∈ X}. From 144 I it follows Ω is order separating. We will use this fact twice. First we will prove k zihyk 6 kzkkyk. For any ω ∈ Ω, say ω ≡ f (hx, (· )xi), we have ω(yihy) = f (hx, yihy, xi) 6 f (hx, xi)kyk2 = kωkkyk2 and so from 21 VII it follows k yihyk 6 kyk2. Whence k zihyk2 = k yihzzihyk 1 = k yhz, zi 6 kyhz, zi 2 ihyhz, zi 2 k2. 1 1 2k 1 2 k 6 kykkzk, as promised. Hence k zihyk 6 kyhz, zi Now we start the proof proper. As ·ihy is linear and xα − x is norm-bounded, we may assume without loss of generality that xα → 0 ultranorm. We have to prove xαihy → 0 ultraweakly. Let g : Ba(X) → C be an arbitrary np-map. Using the order separation of Ω a second time and the fact that ω(T ∗(· )T ) ∈ Ω for all T ∈ Ba(X) and ω ∈ Ω, we see that the linear span of Ω is operator norm dense among all np-maps Ba(X) → C by 90 II. So there are n ∈ N i=1 ωi. By unfolding definitions and 148 V, it follows easily that ωi(·ihy) is ultranorm continuous. So g′(·ihy) is ultranorm continuous as well. Pick any ε > 0. By the previous, there is an α0 such that for all α > α0 we have g′(xαihy) 6 ε, hence and ω1, . . . , ωn ∈ Ω with kg − g′k 6 ε, where g′ ≡ Pn XI g(xαihy) 6 (g − g′)(xαihy) + g′(xαihy) 6 kg − g′k · k xαihyk + ε 6 ε (kxαk · kyk + 1). As xα is norm-bounded and ε was arbitrary, we conclude g(xαihy) → 0. Thus xαihy → 0 ultraweakly, as desired. (cid:3) For a Hilbert space H with linear subspace V ⊆ H we have V ⊆ V ⊥⊥ = V and H ∼= V ⊥⊥ ⊕ V ⊥. We generalize this to self-dual Hilbert C∗-modules over von Neumann algebras. Let's introduce the cast: 160 Exercise Let B be a von Neumann algebra. Assume X and Y are self-dual Hilbert B-modules with orthonormal bases E ⊆ X and F ⊆ Y . Write κ1 : X → X⊕Y and κ2 : Y → X⊕Y for the maps κ1(x) = (x, 0) and κ2(y) = (0, y). Show that κ1(E) ∪ κ2(F ) is an orthonormal basis of X ⊕ Y . Conclude that X ⊕ Y is self-dual. Definition Let X be a Hilbert C∗-module with subset V ⊆ X. Write V ⊥ for the orthocomplement of V , defined by II III V ⊥ = {x; x ∈ X; hx, vi = 0 for all v ∈ V }. ..159 -- 160.. 63 IV Proposition Assume X is a self-dual Hilbert B-module for a von Neumann algebra B. Let V ⊆ X be any subset. Then 1. V ⊥ is a ultranorm-closed Hilbert B-submodule of X (and so is V ⊥⊥); 2. V ⊥⊥ is the ultranorm closure of the B-linear span of V and 3. V ⊥⊥ ⊕ V ⊥ ∼= X as Hilbert C∗-modules via (x, y) 7→ x + y. V Proof It is easy to see V ⊥ is a submodule of X and V ⊆ V ⊥⊥. To show V ⊥ is ultranorm closed in X, assume xα is a net in V ⊥ converging ultranorm to x ∈ X. For each v ∈ V we have hv, unlimα xαi 148 III= uslimα hv, xαi = 0 and so x ∈ V ⊥ as well. Thus V ⊥ is ultranorm complete (and so self dual by 149 V). VI Write W for the ultranorm closure of the B-linear span of V in X. It follows from 148 III, that W is a submodule. Clearly the induced uniformity of X on W is the same as its own ultranorm uniformity: thus W is ultranorm complete and W ⊆ V ⊥⊥. By 149 V there is an orthonormal basis (ei)i∈I of W . Going back to the construction of (ei)i∈I (in 149 VIII), we see that we can extend it to a maximal orthonormal subset of X, which must be a basis for the whole of X. Pick any (dj )j∈J in X such that {ei}i∈I ∪{dj}j∈J is an orthonormal basis of X. VII By construction dj ∈ W ⊥ ⊆ V ⊥ for every j ∈ J. For any x ∈ X we have x = each j ∈ J. Thus V ⊥⊥ ⊆ W , so V ⊥⊥ = W . In particular, as V ⊥ is already an ultranorm closed submodule, we find W ⊥ = (V ⊥)⊥⊥ = V ⊥. VIII So (ei)i is a basis for W = V ⊥⊥ and (dj)j is a basis for W ⊥ = V ⊥. For brevity, let ϑ : V ⊥⊥⊕V ⊥ → X denote the map (y, y′) 7→ y +y′. Clearly ϑ is bounded B- linear. As hy, y′i = 0 for (y, y′) ∈ V ⊥⊥ ⊕ V ⊥ we easily see ϑ is inner product preserving and thus injective. As {ei}i∈I ∪ {dj}j∈J is an orthonormal basis of X, ϑ is surjective. As ϑ is inner product preserving, its inverse is bounded -- hence ϑ is an isomorphism of Hilbert C∗-modules, as promised. (cid:3) Pi ei hei, xi +Pj dj hdj, xi. If x ∈ V ⊥⊥ then x =Pi ei hei, xi as hx, dji = 0 for IX Exercise Assume X is a self-dual Hilbert B-module for a von Neumann alge- bra B. Suppose E ⊆ X is an orthonormal set. Show 1. E is a basis of E⊥⊥ and X 2. for any x ∈ X, we have x ∈ E⊥⊥ if and only if hx, xi =Pe∈E hx, eihe, xi. Exercise Let X be a self-dual Hilbert B-module for some von Neumann alge- bra B. Show that for any x1, . . . , xn ∈ X, there is a finite orthonormal basis of {x1, . . . , xn}⊥⊥ consisting of at most n elements. (Use the orthonormalization in the last part of 149 VIII.) Every Hilbert space H is isomorphic to some ℓ2(I). In fact, the cardinality of I is the only thing that matters, in the following sense: if ℓ2(I) ∼= ℓ2(J), then I and J have the same cardinality. What about Hilbert C∗-modules? Exercise In this exercise we prove [93, thm. 3.12]. Let B be a von Neu- mann algebra and (pi)i∈I a family of projections from B. Write ℓ2((pi)i∈I ) for the set of I-tuples (bi)i∈I from B that are ℓ2-summable (149 I) and further- more ⌈bib∗i ⌉ 6 pi for every i ∈ I. Show that ℓ2((pi)i∈I ) is a right B-module with coordinatewise operations. Prove that for any (bi)i and (ci)i in ℓ2((pi)i∈I ) the sum Pi∈I b∗i ci =: h(bi)i, (ci)ii converges ultraweakly and turns ℓ2((pi)i∈I ) into a pre-Hilbert B-module. Conclude ℓ2((pi)i∈I ) is self-dual and, in fact, that for every self-dual Hilbert B-module X with any orthonormal basis (ei)i∈I , we have X ∼= ℓ2( (hei, eii)i∈I ). 161 II Let B be a von Neumann algebra and (pi)i∈I , (qj)j∈J families of projections from B. What can we say if ℓ2((pi)i∈I ) ∼= ℓ2((qi)j∈J )? (ℓ2 as defined in II.) Let's start with some counterexamples. III 1. Do all orthonormal bases have the same cardinality? B is a self-dual Hilbert B-module over itself. Clearly, the element 1 by itself is an or- thonormal basis of B. However, for every projection p ∈ B the pair {p, 1− p} is also an orthonormal basis. Worse still, for B ≡ B(ℓ2(N)), the set {0ih0 , 1ih1 , 2ih2 , . . .}. is an orthonormal basis. 2. Perhaps any two orthonormal bases have a common coarsening in some sense? Take B = M3. It might not be obvious at this moment, but both { 0ih0 + 2ih2 ,1ih+} and { 0ih0 ,2ih2 + 1ih+} are orthonormal bases of M3 over itself. (Combine IV with V.) There does not seem to be an obvious way in which these two bases have a common coarsening of sorts. 3. Maybe Pi∈I hei, eii = Pj∈J hdj, dji for finite orthonormal bases (ei)i∈I and (dj )j∈J ? For commutative B this indeed holds using Parseval's iden- tity. However, in general it fails: both {0ih0 ,1ih1} and {0ih0 ,1ih0} are orthonormal bases of M2 over itself, but 0ih0 + 1ih1 6= 2 0ih0. Exercise Let (ei)i∈I be some orthonormal basis of a Hilbert B-module X. Show that if (ui)i∈I is a family of partial isometries from B (see 79 IV) such that uiu∗i = hei, eii for all i ∈ I, then (eiui)i∈I is an orthonormal basis of X. IV ..160 -- 161.. 65 Use this and II to conclude that ℓ2((pi)i∈I ) ∼= ℓ2((qi)i∈I ) for any projec- tions (pi)i∈I , (qi)i∈I from B with pi ∼ qi for i ∈ I. (Here, ∼ denotes the Murray -- von Neumann equivalence. That is: p ∼ q iff there is a partial isome- try u with u∗u = p and uu∗ = q. Eg. 0ih0 ∼ +ih+ via 0ih+.) V Exercise Let (ei)i∈I be some orthonormal basis of a Hilbert B-module X with distinguished 1, 2 ∈ I. Show that if he1, e1i + he2, e2i 6 1, then we can make a new orthonormal basis of X by removing e1 and e2 and inserting e1 + e2. Conclude pB ⊕ qB ∼= (p + q)B for projections p, q ∈ B with p + q 6 1. 162 Before we can prove our normal form (in IV), we need some more comparison theory of projections, which was touched upon in 83 II. Recall that for projec- tions p, q in a von Neumann algebra A , we write p . q if there is a partial isometry u with u∗u = p and uu∗ 6 q. In 83 IV we saw . preorders the projec- tions. II III Proposition Let B be a von Neumann algebra that is a factor (i.e. with cen- ter C1). For any two projections p, q ∈ B, either p . q or q . p (see 83 II). (For a more traditional proof, see e.g. [73, prop. 6.2.4].) Let p, q ∈ Proof B be any projections. By Zorn's lemma, there is a maximal set U ⊆ B of partial isometries with Pu∈U u∗u 6 p and Pu∈U uu∗ 6 q. Define p0 ≡ p − Pu∈U u∗u and q0 ≡ q−Pu∈U uu∗. Note that if p0 = 0, then p . q viaPu∈U u, p0 ⌈⌈q0⌉⌉ p0 = p0(Sa∈B ⌈a∗q0a⌉)p0 = Sa∈B ⌈p0a∗q0ap0⌉ by 68 I and 60 IX. Pick c.f. 89 III. Similarly, if q0 = 0, then q . p. So, reasoning towards contradiction, suppose p0 6= 0 and q0 6= 0. As B is a factor ⌈⌈q0⌉⌉ = 1 and so 0 < p0 = an a ∈ B with ⌈p0a∗q0ap0⌉ 6= 0. Then u0 ≡ [q0ap0], see 82 I, is a non-zero partial isometry satisfying u∗0u0 = ⌈p0a∗q0ap0⌉ 6 p0 and u0u∗0 = ⌈q0ap0a∗q0⌉ 6 q0, contradicting maximality of U . (cid:3) IV Theorem Let B be a von Neumann algebra that is a factor (i.e. with center C1). Suppose X is a self-dual Hilbert B-module. Either there is an infinite cardinal κ such that X ∼= ℓ2((1)α∈κ), with ℓ2 as in 161 II, or there is a natural number n ∈ N and a projection p ∈ B such that X ∼= ℓ2((1, . . . , 1, p)), where 1 occurs n times. Proof The case X = {0} is covered by n = 0 and p = 0. Assume X 6= {0}. V VI As a first step we will prove that either 1. X has an orthonormal basis E such that he, ei = 1 for some e ∈ E or 2. there is a single vector e ∈ X such that {e} is an orthonormal basis. Pick any orthonormal basis E0 of X. Let P denote the poset of subsets U ⊆ E0 × B satisfying 1. u is a partial isometry with uu∗ = he, ei for every (e, u) ∈ U and 2. the domains of these u are orthogonal: P(e,u)∈U u∗u 6 1. For any non-empty chain C ⊆ P it is easy to see S C ∈ P and so by Zorn's lemma, there is a maximal element U0 ∈ P . Define e0 = X(e,u)∈U0 eu E1 = E0\{e; (e, u) ∈ U0}. The set {e0}∪ E1 is an orthonormal basis of X -- indeed, it is clearly orthonor- mal and if x =Pe∈E0 e he, xi, then x = e0(cid:16) X(e,u)∈U0 u∗ he, xi(cid:17) + Xe∈E1 e he, xi . If E1 = ∅, then {e0} is an orthonormal basis of X and we have shown point 2. For the other case, assume E1 6= ∅. Pick e1 ∈ E1. Write p0 = he0, e0i and p1 = he1, e1i. Suppose p1 . 1 − p0. Then vv∗ = p1 and v∗v 6 1 − p0 for some v ∈ B. Hence v∗v + p0 6 1. So U0 ∪ {(e1, v)} ∈ P contradicting maximality of U0. Apparently p1 6. 1 − p0. So by II, we must have 1 − p0 . p1. Let v ∈ B be such that vv∗ = 1 − p0 and v∗v 6 p1. Write p′0 = 1 − v∗v. As 1 − p′0 6 p1 = 1 − (1 − p1) we have (1 − p′0) + (1 − p1) 6 1. Define q = 1 − ((1 − p′0) + (1 − p1)). We now have the following splittings of 1. p′0 1 − p1 q v∗v p1 p0 vv∗ Note q = p′0 + p1 − 1 and p′0, p1 > q. Define D = {d0, d1} where d0 = e1q d1 = e0 + e1v∗. From p1q = q it follows D is orthogonal. Clearly hd0, d0i = q and hd1, d1i = he0, e0i + vp1v∗ = p0 + vv∗ = 1. So D is an orthonormal set. Using 0 = (1 − p0)p0 = vv∗p0 and q + v∗v = p1, we see that for any x ≡ e0b0 + e1b1 with b0 ∈ p0B and b1 ∈ p1B, we have d0hd0, xi + d1hd1, xi = d0b1 + d1(vb1 + b0) = e1qb1 + e0vb1 + e1v∗vb1 + e0b0 + e1v∗b0 = e1qb1 + e0p0vv∗vb1 + e1v∗vb1 + e0b0 + e1v∗vv∗p0b0 = e1qb1 + e1v∗vb1 + e0b0 = e1(q + v∗v)b1 + e0b0 = e1b1 + e0b0 = x. ..161 -- 162.. 67 Thus D is an orthonormal basis of {e0, e1}⊥⊥. Hence E ≡ D ∪ (E1 − {e1}) is an orthonormal basis of X with d1 ∈ E such that hd1, d1i = 1. VII For brevity, call X 1-dim if there is a one-element orthonormal basis of X. In VI we saw how to create an orthonormal basis D of a non-1-dim X with If {e0}⊥ is not 1-dim, we can apply VI an e0 ∈ D such that he0, e0i = 1. on {e0}⊥ (instead of X) to find an orthonormal basis D′ of {e0}⊥ with e1 ∈ D′ such that he1, e1i = 1. This procedure can be continued using Zorn's lemma as follows. Let P denote the poset of orthogonal subsets E ⊆ X satisfying he, ei = 1 P . By Zorn's lemma there is a maximal E ∈ P . Suppose E⊥ 6= {0} and E⊥ is not 1-dim. Then by VI we can find an orthonormal basis E′ of E⊥ together with e ∈ E′ such that he, ei = 1. Now E′ ∪ {e} ∈ P , contradicting maximality of E. Apparently E⊥ = {0} or E⊥ is 1-dim. If E⊥ = {0} then we are done taking for κ the cardinality of E. For the other case, assume E⊥ = {e}⊥⊥ for some e ∈ X with he, ei a non-zero projection. If E is finite, then we are done as well. So, assume E is infinite. Pick a sequence e1, e2, . . . ∈ E of distinct elements. Write p = he, ei and E0 = {en; n ∈ N}. Define for all e ∈ E. Clearly ∅ ∈ P and for any non-empty chain C ⊆ P we haveS C ∈ E1 = {e + e1(1 − p), e1p + e2(1 − p), e2p + e3(1 − p), . . .}. It is easy to see E1 is an orthogonal set with hd, di = 1 for all d ∈ E1. Further- more E1 is an orthonormal basis for E0 ∪{e}, hence (E − E0)∪ E1 is the desired orthonormal basis of X. (cid:3) VIII Does this settle the issue raised in 161 I? One might hope that ℓ2((1B)α∈κ) ∼= ℓ2((1B)β∈λ) implies κ = λ for all cardinals κ,λ. This is not the case. Indeed, if B is a type III factor of operators on a separable Hilbert space, then for any non-zero projection p, we have p ∼ 1 − p (by combining [73, dfn. 6.5.1] and [73, cor. 6.3.5]) and so B ∼= pB⊕ (1− p)B ∼= B⊕ B as Hilbert C∗-modules. 163 Before we continue to the next topic, the exterior tensor product of Hilbert C∗-modules, we will show that the ultranorm completion 150 II is determined by its universal property. We will need this fact in our treatment of the exterior tensor product. II Proposition Let B be a von Neumann algebra and V a right B-module with B- valued inner product [· , · ]. There is an up-to-isomorphism unique self-dual Hilbert B-module X together with inner product preserving B-linear η : V → X with the following universal property. For every bounded B-linear map T : V → Y to some self-dual Hilbert B-module Y , there is a unique bounded B-linear map T : X → Y with T ◦ η = T . Moreover, for such η : V → X, the image of V under η is ultranorm dense. Proof We already know such a X and η : V → X exist by 150 II and 151 Ia. For this one η(V ) is ultranorm dense in X. Assume there is another self-dual Hilbert B-module X2 with inner product preserving B-linear η2 : V → X2 satisfying the universal property. By the universal property of η applied to η2, there must exist a unique U : X → X2 with U ◦ η = η2. Similarly, there is a V : X2 → X with V ◦ η2 = η. Clearly idX ◦ η = V ◦ η2 = V ◦ U ◦ η and so by the uniqueness id = V ◦ U . Similarly U ◦ V = id. It remains to be shown U preserves the inner product. As III hη(x), η(y)i = [x, y] = hη2(x), η2(y)i = hU η(x), U η(y)i for all x, y ∈ X, we know η(X) is ultranorm dense. So by ultranorm continuity and bijectivity of U , we know U (η(X)) = η2(X) is ultranorm dense in X2. Hence U preserves the inner product by 148 V (cid:3) 2.5.1 Exterior tensor product Suppose X is a Hilbert A -module and Y is a Hilbert B-module for some C∗- algebras A and B. The algebraic tensor product X⊙Y is a right A ⊙B-module via (x ⊗ y) · (a ⊗ b) = (xa) ⊗ (yb) and has an A ⊙ B-valued inner product fixed by [x ⊗ y, x′ ⊗ y′] = hx, x′i ⊗ hy, y′i. Write A ⊗σ B for the spatial C∗-tensor product [73, §11.3] of A and B. The inner product on X ⊙ Y can be extended to an A ⊗σ B-valued inner product on the norm completion of X ⊙ Y on which it is definite [81]. This is a Hilbert C∗-module, which is called the exterior tensor product of X and Y . It is quite difficult to perform this construction: Lance discusses the difficulties in [81, ch. 4]. But if we assume A and B are von Neumann algebras and that X and Y are self-dual, it is rather easier. In fact, we will directly construct the ultranorm completion of the exterior tensor product, which is a self-dual Hilbert A ⊗ B-module, which we call the self-dual exterior tensor product. This self-dual exterior tensor product behaves much better than the plain exterior tensor product. As a first example, we will see (in 165 VI) that Ba(X) ⊗ Ba(Y ) ∼= Ba(X ⊗ Y ) for the self-dual exterior tensor product, which is false in general for the plain exterior tensor product. The self-dual exterior tensor product also appears naturally when computing the Paschke dilation of a tensor product ϕ ⊗ ψ, see 167 I. Theorem Suppose A and B are von Neumann algebras. For any self-dual Hilbert A -module X and self-dual Hilbert B-module Y , there is an up-to- isomorphism unique self-dual Hilbert A ⊗B-module X ⊗ Y , called the (self-dual exterior) tensor product, together with a A ⊙ B-linear injective inner product preserving map η : X ⊙ Y → X ⊗ Y with the following universal property. For every self-dual Hilbert A ⊗ B-module Z with bounded A ⊙ B- linear T : X ⊙ Y → Z, there is a unique bounded A ⊗ B-linear ..162 -- 164.. 69 164 II map T : X ⊗ Y → Z with T ◦ η = T , where we norm X ⊙ Y with ktk = k[t, t]k such X ⊗ Y we have the following. 1 2 A ⊗B. Write x ⊗ y = η(x ⊗ y). For any 1. The image of X ⊙ Y under η is ultranorm dense in X ⊗ Y . 2. If (ei)i∈I is an orthonormal basis of X and (dj )j∈J of Y , then (a) (ei ⊗ dj)i,j∈I×J is an orthonormal basis of X ⊗ Y and (b) the linear span of {(cid:12)(cid:12) (eia) ⊗ (djb)(cid:11)h ek ⊗ dl ; a ∈ A , b ∈ B, i, k ∈ I, j, l ∈ J } is ultraweakly dense in Ba(X ⊗ Y ). Proof Pick orthonormal bases (ei)i∈I and (dj)j∈J of X and Y respectively. Write pij ≡ hei, eii⊗hdj, dji ∈ A ⊗ B. Define X ⊗ Y ≡ ℓ2((pij )i,j∈I×J ), with ℓ2 as in 161 II. This definition of X ⊗ Y seems to depend on the choice of bases (ei)i and (dj )h. We show it does not (up-to-isomorphism), in IX. (Definition η) Pick x ∈ X and y ∈ Y . Write xi ≡ hei, xi and yj ≡ hdj , yi. Then by Parseval hx, xi ⊗ hy, yi = (cid:0)Xi = Xi,j = Xi,j x∗i xi(cid:1) ⊗(cid:0)Xj y∗j yj(cid:1) x∗i xi ⊗ y∗j yj (xi ⊗ yj)∗(xi ⊗ yj). Hence (xi ⊗ yj)i,j is ℓ2-summable, and so there is a unique A ⊙ B-linear map η : X ⊙ Y → X ⊗ Y fixed by η(x ⊗ y) = (xi ⊗ yj)i,j. (η preserves inner product) Pick additional x′ ∈ X and y′ ∈ Y . Writing x′i = hei, x′i and y′j = hdj, y′i, we have (xi ⊗ yj)∗ ⊗ (x′i ⊗ y′j) = Xi,j x∗i x′i ⊗ y∗j y′j and hη(x ⊗ y), η(x′ ⊗ y′)i = Xi,j (cid:0)Xi x∗i x′i(cid:1) ⊗(cid:0)Xj y∗j y′j(cid:1) = hx, x′i ⊗ hy, y′i = hx ⊗ y, x′ ⊗ y′i . Thus it is sufficient to showPi,j x∗i x′i ⊗ y∗j y′j =(cid:0)Pi x∗i x′i(cid:1) ⊗(cid:0)Pj y∗j y′j(cid:1), which holds as the product np-functionals are separating by definition, see 108 II. Thus indeed η preserves the inner product. III IV V (η is injective) Let n ∈ N, x1, . . . , xn ∈ X and y1, . . . , yn ∈ Y be given such that η(Pl xl ⊗ yl) = 0 By 160 X there are orthonormal e′1, . . . , e′m ∈ X such that xl = Pi e′ihe′i, xli. Similarly yl = Pj d′jhd′j, yli for some orthonor- mal d′1, . . . , d′m′ ∈ Y . As η preserves the inner product, the elements e′i ⊗ d′j are again orthonormal and so from VI 0 = Xl xl ⊗ yl = Xi,j (e′i ⊗ d′j)(cid:0)Xl he′i, xli ⊗ hd′j, yli(cid:1) it followsPlhe′i, xli ⊗ hd′j , yli = 0 for all i, j. Consequently Xl xl ⊗ yl = Xi,j (e′i ⊗ d′j)(cid:0)Xl he′i, xli ⊗ hd′j, yli(cid:1) = 0, which shows η is injective (and that the inner product on X ⊙ Y is definite). (Image η ultranorm dense) For brevity, write E ≡ {ei ⊗ dj ; i, j ∈ I × J}. By 108 II (and 73 VIII) A ⊙ B is ultrastrongly dense in A ⊗ B. So E(A ⊙ B) is ultranorm dense in E(A ⊗B), see 148 III. In turn the linear span of E(A ⊗B) is (by construction) ultranorm dense in X ⊗ Y . Thus the linear span of E(A ⊙ B) is ultranorm dense in X ⊗ Y . Hence the image of η, which contains the linear span of E(A ⊙ B), is ultranorm dense in X ⊗ Y . (Universal property) Let Z be a self-dual Hilbert A ⊗ B-module with some bounded A ⊙ B-linear T : X ⊙ Y → Z. If X ⊙ Y were an A ⊗ B-module and both η and T were A ⊗ B-linear, we could simply apply 151 Ia to get the desired T . Instead, we will retrace the steps of its proof and modify it for the present situation. Uniqueness of T is the same: T is fixed by the ultranorm- dense image of η as it is bounded A ⊗ B-linear and so ultranorm continuous by 148 I. The argument for existence of T is more subtle. We will show VII VIII hT t, T ti 6 kTk2[t, t] for all t ∈ X ⊙ Y (2.13) 2 as a replacement for the ultranorm continuity of T . We cannot apply 144 V directly: as A ⊙ B is not a C∗-algebra, we cannot construct (ht, ti + ε)− 1 required for its proof. To prove (2.13), we will work in the norm completions. Write A ⊙ B for the norm-closure of A ⊙ B in A ⊗ B. Clearly A ⊙ B is a C∗- algebra (in fact it is the spatial C∗-tensor product of A and B). Write X ⊙ Y for the norm-closure of the image of η in X ⊗ Y . By norm continuity (see 142 V) the operations of X ⊗ Y restrict to turn X ⊙ Y into a Hilbert A ⊙ B- module. As T is uniformly continuous, there is a unique bounded A ⊙ B-linear T : X ⊙ Y → Z with T ◦ η = T and kTk = kTk. Now we can apply the proof of 144 V to T to find hT s, T si 6 kTk2 hs, si for s ∈ X ⊙ Y . Substituting s = η(t), we get (2.13). To define T , pick any t ∈ X ⊗ Y . As the image of η is ultranorm dense, there ..164.. 71 is a net (η(tα))α with η(tα) → t ultranorm. For any np-map f : A ⊗ B → C kT (tα) − T (tβ)k2 f = f (hT (tα − tβ), T (tα − tβ)i) 6 kTk2f ([tα − tβ, tα − tβ]) = kTk2f (hη(tα) − η(tβ), η(tα) − η(tβ)i) → 0. So (T tα)α is ultranorm Cauchy in Z. With a similar argument we see (T tα)α is equivalent to (T t′α)α whenever (ηtα)α is equivalent to (ηt′α)α. So we may define T t = unlim α T tα. Clearly T ◦ η = T . With the same argument as for 151 Ia, we see T is bounded and additive. To show T is A ⊗ B-linear, pick any b ∈ A ⊗ B. There is a net bβ in A ⊙ B with bβ → b ultrastrongly. If T is ultranorm continuous, we are done: ( T t)b = unlim β (unlim α T tα)bβ = unlim β unlim α T (tαbβ) = unlim β T (tbβ) = T (tb). To show T is ultranorm continuous, assume tα → 0 ultranorm in X ⊗ Y . It is sufficient to show T tα → 0 ultranorm. Let f : A ⊗ B → C be any np-map. Write Φ for the set of entourages for ultranorm uniformity on X ⊗ Y . As the image of η is ultranorm dense, there is a Φ-indexed Cauchy net (tα E)E∈Φ in η(X ⊙ Y ) with tα E. As tα → 0 there is some α0 such that for all α > α0 we have ktαkf 6 ε. We can find E0 > Ef,ε such that for all E > E0 we have k T tα0 − T tα0 E kf 6 ε. For E > E0 we have E > Ef,ε and so E E tα for each E ∈ Φ. Thus T tα = unlimE T tα kT tα0 E0kf Consequently (2.13) 6 kTkktα0 E0kf 6 kTk(ktα0 E0 − tα0kf + ktα0kf ) 6 2kTkε. k T tα0kf 6 k T tα0 − T tα0 E0kf + kT tα0 E0kf 6 ε + 2kTkε. IX So T tα → 0 ultranorm, as desired. (Uniqueness) Assume there is a self-dual Hilbert A ⊗B-module X⊗2Y together with a injective inner product-preserving A ⊙ B-linear map η2 : X ⊙ Y → X ⊗2 Y also satisfying the universal property. With the same reasoning as in 163 II there is an invertible bounded A ⊗ B-linear U : X ⊗ Y → X ⊗2 Y with U ◦ η = η2. Clearly hU x, U yi = hx, yi for x, y ∈ E(A ⊙ B), where we defined E = {ei ⊗ dj; i, j ∈ I × J}. As we saw the linear span of E(A ⊙ B) is ultranorm dense in X ⊗ Y , we see U must preserve the inner product by 148 V. X Property 1 from the statement of the Theorem follows immediately from the fact that the exterior tensor product is unique up to an isomorphism which respects the embeddings. Assume (e′i)i∈I ′ is an orthonormal basis of X and (d′j)j∈J ′ is an orthonormal basis of Y . We will show E2 = {e′i ⊗ d′j; i, j ∈ I′ × J′} is an orthonormal basis of X ⊗ Y . Clearly E2 is orthonormal. By 160 IX and 160 IV it is sufficient to show ei0 ⊗dj0 ∈ E⊥⊥2 for all i0 ∈ I and j0 ∈ J. Write ai = he′i, ei0i and bj =(cid:10)d′j , dj0(cid:11). By 160 IX and V it suffices to show the latter equality in a∗i ai ⊗ b∗j bj, b∗j bj(cid:1) = hei0 ⊗ dj0 , ei0 ⊗ dj0i = Xi,j a∗i ai(cid:1) ⊗(cid:0)Xj (cid:0)Xi which holds as product np-functionals are separating by definition, see 108 II. Finally we prove that the linear span of XI D ≡ {(cid:12)(cid:12)(e′ia) ⊗ (d′jb)(cid:11)hek ⊗ dl ; a ∈ A , b ∈ B, i, k ∈ I′, j, l ∈ J′} is ultraweakly dense in Ba(X ⊗ Y ). By 159 IV, the linear span of (the set of in Ba(X⊗Y ). We will show that t ∈ A ⊙B suffices. By 108 II, 73 VIII and 74 IV, each t ∈ A ⊗ B is the ultrastrong limit of a norm bounded net tα in A ⊙ B. operators of the form)(cid:12)(cid:12)(e′i ⊗ d′j)t(cid:11)(cid:10)e′k So by 159 IX (and 148 III), we see (cid:12)(cid:12)(e′i ⊗ d′j)tα(cid:11)(cid:10)e′k to (cid:12)(cid:12)(e′i ⊗ d′j)t(cid:11)(cid:10)e′k ⊗ d′l(cid:12)(cid:12) with t ∈ A ⊗ B is ultraweakly dense ⊗ d′l(cid:12)(cid:12) converges ultraweakly ⊗ d′l(cid:12)(cid:12). So indeed, the linear span of D is ultraweakly dense in Ba(X ⊗ Y ). (cid:3) Examples We give a few examples of the self-dual exterior tensor product and relate it to other notions of tensor product. To avoid confusion, we will denote the self-dual exterior tensor product by ⊗ext (instead instead of the plain ⊗.) 1. Any Hilbert space is a self-dual Hilbert C module. For Hilbert spaces H and K we have H ⊗Hilb K ∼= H ⊗ext K , where ⊗Hilb denotes the regular tensor product of Hilbert spaces. XII 2. Every von Neumann algebra is a self-dual Hilbert C∗-module over itself. For any von Neumann algebras A and B we have A ⊗vN B ∼= A ⊗ext B as Hilbert A ⊗vN B-modules, where A ⊗vN B denotes the (spatial) tensor product of von Neumann algebras. 3. Let X be a self-dual Hilbert A -module and Y be a self-dual Hilbert B- module for von Neumann algebras A , B. The norm closure of X ⊙ Y in X ⊗ext Y is the plain exterior tensor product of X and Y . In turn, the ultranorm completion of the exterior tensor product of X and Y is isomorphic to X ⊗ext Y , the self-dual exterior tensor product. ..164.. 73 4. In 167 I it will be shown that (A1 ⊗ϕ1 B1) ⊗ext (A2 ⊗ϕ2 B2) ∼= (A1 ⊗ A2) ⊗ϕ1⊗ϕ2 (B1 ⊗ B2) for all ncp-maps ϕi : Ai → Bi (i = 1, 2) between von Neumann algebras. 165 For Hilbert spaces H and K , we have B(H ⊗K ) ∼= B(H )⊗B(K ). However, for Hilbert modules X and Y and the regular (not necessarily self-dual) exterior tensor product it is not true in general that Ba(X) ⊗ Ba(Y ) ∼= Ba(X ⊗ Y ). However, if X and Y are self-dual Hilbert modules over von Neumann algebras, then we do have the familiar formula for the self-dual exterior tensor product. We need some preparation. II Setting Let X be a self-dual Hilbert A -module and Y be a self-dual Hilbert B- module for von Neumann algebras A and B. III Proposition For any operators S ∈ Ba(X) and T ∈ Ba(Y ), there is a unique operator S ⊗ T ∈ Ba(X⊗Y ) fixed by (S⊗T )(x⊗y) = (Sx)⊗(T y) for any x ∈ X and y ∈ Y . IV Proof We will define S ⊗ T using the universal property proven in 164 II. Write Θ : X⊙Y → X⊗Y for the map Θ(x⊗y) = (Sx)⊗(T y). It is easy to see Θ is A ⊙B-linear. The hard part is to show that Θ is bounded. Pick any t ∈ X⊙Y i=1 xi ⊗ yi. To start, note (hSxi, Sxji)ij is positive in MnA -- -- say t ≡ Pn indeed, for any a1, . . . , an ∈ A we have Pi,j a∗i hSxi, Sxji aj = hSx, Sxi > 0 where x ≡Pi xiai and so it is positive by 33 II. By reasoning in the same way forpkSk2 − S∗S instead of S, it follows (hSxi, Syii)ij 6 (kSk2 hxi, xji)ij. Sim- It is easy to see sAs∗1 = Pi,j Aij for ilarly 0 6 (hT yi, T yji)ij 6 (kTk2 hyi, yji)ij in MnB. Write s : (A ⊗ B)⊕n → A ⊗ B for the A ⊗ B-linear map given by s(t1, . . . , tn) = t1 + ··· + tn. (s is the row vector filled with 1s). any matrix A over A ⊗ B. As a final ingredient, recall that the bilinear map Mn(⊗) : MnA × MnB → Mn(A ⊗ B) defined in the obvious way is positive by 113 IV and 108 II. Putting everything together: hSxi, Sxji ⊗ hT yi, T yji h(Sxi) ⊗ (T yi), (Sxj ) ⊗ (T yj)i hΘt, Θti = Xi,j = Xi,j = s (hSxi, Sxji ⊗ hT yi, T yji)ij s∗1 = s Mn(⊗)(cid:0) (hSxi, Sxji)ij , (hT yi, T yji)ij(cid:1) s∗1 6 s Mn(⊗)(cid:0) (kSk2hxi, xji)ij , (kTk2hyi, yji)ij(cid:1) s∗1 = kSk2kTk2 ht, ti . We have proven that Θ is bounded (by kSkkTk). Thus by 164 II, there exists a unique map S⊗T : X⊗Y → X⊗Y with (S⊗T )(x⊗y) = Θ(x⊗y) ≡ (Sx)⊗(T y), which is what we desired to prove. (cid:3) Exercise Prove that for the operation ⊗ defined in III, we have V 1. x1ihx2 ⊗ y1ihy2 = x1 ⊗ y1ihx2 ⊗ y2, 2. 1 ⊗ 1 = 1; 3. (S ⊗ T )(S′ ⊗ T ′) = (SS′) ⊗ (T T ′) and 4. (S ⊗ T )∗ = S∗ ⊗ T ∗. (Hint: use that by the reasoning of 164 VII the linear span of {(ea) ⊗ (db); (a, b, e, d) ∈ A × B × E × D} is ultranorm dense in X ⊗ Y for any bases E of X and D of Y .) Theorem There is an nmiu-isomorphism ϑ : Ba(X) ⊗ Ba(Y ) ∼= Ba(X ⊗ Y ) fixed by ϑ(S ⊗ T ) = S ⊗ T . Proof We will show that the bilinear map Θ : Ba(X) × Ba(Y ) → Ba(X ⊗ Y ) given by Θ(S, T ) = S ⊗ T is a tensor product in the sense of 108 II using 116 VII. From this it follows by 114 II, that there is a nmiu-isomorphism VI VII ϑ : Ba(X) ⊗ Ba(Y ) ∼= Ba(X ⊗ Y ) with ϑ(S ⊗ T ) = Θ(S, T ) = S ⊗ T , as promised. From V, it follows Θ is an miu-bilinear map. By 164 II the linear span of op- erators of the form eiaihek ⊗ djbihdl = (eia) ⊗ (djb)ihek ⊗ dl is ultraweakly dense in Ba(X ⊗ Y ) for an orthonormal basis ei ⊗ dj of X ⊗ Y . So the image of Θ generates Ba(X ⊗ Y ). It remains to be shown that there are sufficiently many product functionals (see 108 II for the definition of product functional and their sufficiency). Write VIII IX ΩX ≡ {f (hx, (· )xi); f : A → C np-map, x ∈ X} ΩY ≡ {g(hy, (· )yi); g : B → C np-map, y ∈ Y }. From 144 I it follows ΩX and ΩY are order separating. For σ ∈ ΩX and τ ∈ ΩY -- say σ ≡ f (hx, (· )xi), τ ≡ g(hy, (· )yi) -- define σ⊗τ ≡ (f ⊗g)(x⊗y, (· )x⊗y) and Ω ≡ {σ ⊗ τ ; σ ∈ ΩX , τ ∈ ΩY }. We will show that Ω is faithful and that (σ ⊗ τ )(Θ(S, T )) = σ(S)τ (T ) for all S ∈ Ba(X) and T ∈ Ba(Y ), which Indeed for any S ∈ Ba(X) and is sufficient to show Θ is a tensor product. ..164 -- 165.. 75 T ∈ Ba(Y ), we have (σ ⊗ τ )(Θ(S, T )) = (f ⊗ g)(hx ⊗ y, (S ⊗ T )(x ⊗ y)i ) = (f ⊗ g)(hx ⊗ y, (Sx) ⊗ (T y)i ) = (f ⊗ g)(hx, Sxi ⊗ hy, T yii ) = f (hx, Sxi) · g(hy, T yii) = σ(S)τ (T ). X Finally, to show that Ω is faithful, assume A ∈ Ba(X ⊗ Y ), A > 0 and (σ ⊗ τ )(A) = 0 for all σ ⊗ τ ∈ Ω. Then for all x ∈ X and y ∈ Y , we have hx ⊗ y, A(x⊗ y)i = 0, thus k√A(x⊗ y)k2 = 0, hence √A(x⊗ y) = 0. So √A vanishes on X ⊙ Y , which is ultranorm dense in X ⊗ Y , whence √A = 0. So A = 0. (cid:3) 166 To compute the tensor product of Paschke dilations, we will need to know a bit more about the ultranorm continuity of the exterior tensor product. II III Lemma Let X be a self-dual Hilbert A -module and Y be a self-dual Hilbert B- module for von Neumann algebras A and B. If xα → x and yα → y ultranorm for norm-bounded nets xα in X and yα in Y , then xα ⊗ yα → x ⊗ y ultranorm. Proof As xα ⊗ yα − x⊗ y = (xα − x)⊗ yα + x⊗ (yα − y), it is sufficient to show that both (xα − x) ⊗ yα → 0 and x ⊗ (yα − y) → 0 ultranorm. Clearly h(xα − x) ⊗ yα, (xα − x) ⊗ yαi = hxα − x, xα − xi ⊗ hyα, yαi = (hxα − x, xα − xi ⊗ 1) · (1 ⊗ hyα, yαi). The maps b 7→ 1⊗ b and a 7→ a⊗ 1 are ncp so both 1⊗hyα, yαi and hxα− x, xα− xi⊗1 are bounded nets of positive elements. Also hxα−x, xα−xi⊗1 → 0 as xα → x ultranorm. Thus 44 III applies and we see (hxα−x, xα−xi⊗1)·(1⊗hyα, yαi) → 0 ultraweakly, hence (xα − x) ⊗ yα → 0 ultranorm. With a similar argument one shows x ⊗ (yα − y) → 0 ultranorm. (cid:3) Lemma Let X be a self-dual Hilbert A -module and Y be a self-dual Hilbert B- module for von Neumann algebras A and B. Suppose U ⊆ X and V ⊆ Y are ultranorm-dense submodules. Then the linear span of U ⊗ V ≡ {u ⊗ v; u ∈ U, v ∈ V } is ultranorm dense in X ⊗ Y . It is sufficient to show that every x ⊗ y ∈ X ⊗ Y is the ultranorm limit Proof of elements from U ⊗ V as the linear span of elements of the form x ⊗ y is ultranorm dense in X ⊗ Y . So pick any element of the form x ⊗ y ∈ X ⊗ Y . By the Kaplansky density theorem for Hilbert C∗-modules, see 158 II, there are norm-bounded nets xα in U and yα in V with xα → x and yα → y ultranorm. So by II, we see xα ⊗ yα → x ⊗ y. (cid:3) IV V Lemma Let ϕ : A → B be an ncp-map between von Neumann algebras together with ultrastrongly dense ∗-subalgebras A ′ ⊆ A and B′ ⊆ B. Then: the linear span of T ≡ {a ⊗ b; a ∈ A ′, b ∈ B′} is ultranorm dense in A ⊗ϕ B, which was defined in 154 III. It is sufficient to show that any a ⊗ b ∈ A ⊗ϕ B is the ultranorm limit Proof of elements from T . Pick any a ⊗ b ∈ A ⊗ϕ B. By 74 VI there are norm- bounded nets aα in A ′ and bα in B′ with aα → a and bα → b ultrastrongly. Say kbαk,kaαk 6 B for some B > 0. We will show aα⊗bα−a⊗b → 0 ultranorm. As usual, we split it into two: VI VII aα ⊗ bα − a ⊗ b = aα ⊗ (bα − b) + (aα − a) ⊗ b. For the first term, note haα ⊗ (bα − b), aα ⊗ (bα − b)i = (bα − b)∗ ϕ(a∗αaα) (bα − b) 6 B2kϕk (bα − b)∗(bα − b), which converges to 0 ultraweakly as bα → b ultrastrongly. The second h(aα − a) ⊗ b(aα − a) ⊗ bi = b∗ ϕ((aα − a)∗(aα − a)) b, converges ultraweakly to 0 as and b∗ϕ(· )b is normal and (aα − a)∗(aα − a) converges ultraweakly to 0 as aα → a ultrastrongly. (cid:3) Theorem Suppose ϕi : Ai → Bi is an ncp-map between von Neumann algebras with Paschke dilation (Pi, i, hi) for i = 1, 2. Then (P1 ⊗ P2, 1 ⊗ 2, h1 ⊗ h2) is a Paschke dilation of ϕ1 ⊗ ϕ2. Furthermore (A1 ⊗ϕ1 B1) ⊗ (A2 ⊗ϕ2 B2) ∼= (A1 ⊗ A2) ⊗ϕ1⊗ϕ2 (B1 ⊗ B2). (2.14) 167 B1) ⊗ (A2 ⊗ϕ2 Proof We will prove (2.14) first. We proceed as follows. To start, we prove that (A1 ⊙ B1) ⊙ (A2 ⊙ B2) (with the obvious inclusion) is ultranorm dense B2). Then we show that (A1 ⊙ A2) ⊙ (B1 ⊙ B2) is in (A1 ⊗ϕ1 ultranorm dense in (A1 ⊗ A2) ⊗ϕ1⊗ϕ2 (B1 ⊗ B2). We continue and show that the natural bijection U0 : (A1 ⊙ B1) ⊙ (A2 ⊙ B2) ∼= (A1 ⊙ A2) ⊙ (B1 ⊙ B2) preserves the induced inner products and so extends uniquely to an isomor- phism (2.14). Using this isomorphism, we prove that the tensor product of the standard Paschke dilations is a dilation of the tensor product. From this, we finally derive the stated result. (Density) By construction, A1 ⊙ B1 is an ultranorm-dense B1-submodule of A1 ⊗ϕ1 B2. Thus by 166 IV (A1⊙ B1)⊙ (A2⊙ B2) is ultranorm dense in (A1 ⊗ϕ1 B2). By 108 II (with 73 VIII and 74 VI), we know A1 ⊙ A2 is an ultrastrongly-dense B1. Similarly A2 ⊙ B2 is ultranorm dense in A2 ⊗ϕ2 B1)⊗ (A2 ⊗ϕ2 ..165 -- 167.. 77 II III subalgebra of A1 ⊗ A2. Similarly B1 ⊙ B2 is ultrastrongly dense in B1 ⊗ B2. Hence (A1⊙ A2)⊙ (B1⊙ B2) is ultranorm dense in (A1⊗ A2)⊗ϕ1⊗ϕ2 (B1⊗ B2) by 166 VI, IV (Inner product preservation) Clearly (cid:10) (a1 ⊗ b1) ⊗ (a2 ⊗ b2) , (α1 ⊗ β1) ⊗ (α2 ⊗ β2)(cid:11) = ha1 ⊗ b1, α1 ⊗ β1i ⊗ ha2 ⊗ b2, α2 ⊗ β2i = (b∗1ϕ1(a∗1α1)β1) ⊗ (b∗2ϕ2(a∗2α2)β2) = (b1 ⊗ b2)∗ (ϕ1 ⊗ ϕ2)(cid:0)(a1 ⊗ a2)∗(α1 ⊗ α2)(cid:1) (β1 ⊗ β2) = (cid:10) (a1 ⊗ a2) ⊗ (b1 ⊗ b2) , (α1 ⊗ α2) ⊗ (β1 ⊗ β2)(cid:11); so U0 : (A1 ⊙ B1) ⊙ (A2 ⊙ B2) ∼= (A1 ⊙ A2) ⊙ (B1 ⊙ B2) preserves the inner product by linearity. V As U0 preserves the inner product, it is ultranorm continuous. So with the usual reasoning, there is a unique bounded B1 ⊙ B2-linear extension U1 : (A1 ⊗ϕ1 B1) ⊙ (A2 ⊗ϕ2 B2) → (A1 ⊗ A2) ⊗ϕ1⊗ϕ2 (B1 ⊗ B2). In turn, by the universal property of the self-dual exterior tensor product, there exists a unique bounded B1 ⊗ B2-linear extension U : (A1 ⊗ϕ1 B1) ⊗ (A2 ⊗ϕ2 B2) → (A1 ⊗ A2) ⊗ϕ1⊗ϕ2 (B1 ⊗ B2). By the ultranorm density III and 148 V we know U also preserves the inner product. Also from III, it follows that U is surjective. Thus U is an isomorphism with U−1 = U∗. We have shown (2.14). By 165 VI there is a unique nmiu- isomorphism B1) ⊗ Ba(A2 ⊗ϕ2 ϑ : Ba(A1 ⊗ϕ1 fixed by ϑ(S ⊗ T ) = S ⊗ T . Let (Ba((A1 ⊗ A2) ⊗ϕ1⊗ϕ2 (B1 ⊗ B2)), , h) and (Ba(Ai⊗ϕi Bi), i, hi) (for i = 1, 2) denote the Paschke dilations of ϕ1⊗ ϕ2 and ϕi, respectively, constructed in 154 III. B2) → Ba((A1 ⊗ϕ1 B1) ⊗ (A2 ⊗ϕ2 B2)) VI Our next goal is to show adU ∗ ◦ ϑ◦(1 ⊗ 2) = which is sufficient to show that (Ba(A1⊗ϕ1 h◦ adU ∗ ◦ ϑ = h1 ⊗ h2, B1)⊗Ba(A2⊗ϕ2 B2), 1⊗2, h1⊗h2) (2.15) is a Paschke dilation of ϕ1 ⊗ ϕ2. We start with the equation on the left. (cid:0)(adU ∗ ◦ ϑ◦(1 ⊗ 2))(α1 ⊗ α2)(cid:1) (a1 ⊗ a2) ⊗ (b1 ⊗ b2) = (cid:0)(adU ∗ ◦ ϑ)(1(α1) ⊗ 2(α2))(cid:1) (a1 ⊗ a2) ⊗ (b1 ⊗ b2) = U (1(α1) ⊗ 2(α2)) U∗ (a1 ⊗ a2) ⊗ (b1 ⊗ b2) = U (1(α1) ⊗ 2(α2)) (a1 ⊗ b1) ⊗ (a2 ⊗ b2) = U ((α1a1) ⊗ b1) ⊗ ((α2a2) ⊗ b2) = ((α1 ⊗ α2) · (a1 ⊗ a2)) ⊗ (b1 ⊗ b2) = (α1 ⊗ α2) (a1 ⊗ a2) ⊗ (b1 ⊗ b2). As the linear span of elements of the form (a1⊗a2)⊗(b1⊗b2) is ultranorm dense in (A1⊗A2)⊗ϕ1⊗ϕ2 (B1⊗B2), we see (adU ∗ ◦ ϑ◦(1⊗2))(α1⊗α2) = (α1⊗α2) for all α1 ∈ A1 and α2 ∈ A2. In turn, as the linear span of elements of the form α1 ⊗ α2 is ultrastrongly dense in A1 ⊗ A2 and ncp-maps are ultrastrongly continuous, we see adU ∗ ◦ ϑ◦(1 ⊗ 2) = . We continue with the equation on the right of (2.15). (h◦ adU ∗ ◦ ϑ)(T ⊗ S) = h(1 ⊗ 1) ⊗ (1 ⊗ 1), U (T ⊗ S) U∗(1 ⊗ 1) ⊗ (1 ⊗ 1)i = hU∗ (1 ⊗ 1) ⊗ (1 ⊗ 1), (T ⊗ S) U∗(1 ⊗ 1) ⊗ (1 ⊗ 1)i = h(1 ⊗ 1) ⊗ (1 ⊗ 1), (T ⊗ S) (1 ⊗ 1) ⊗ (1 ⊗ 1)i = h1 ⊗ 1, T 1 ⊗ 1i ⊗ h1 ⊗ 1, S 1 ⊗ 1i = h1(T ) ⊗ h2(S) = (h1 ⊗ h2)(T ⊗ S) From ultrastrong density, again, it follows h◦ adU ∗ ◦ ϑ = h1 ⊗ h2. (cid:3) 2.6 Pure maps 168 Schrodinger's equation is invariant under the reversal of time and so the isolated quantum mechanical processes (ncp-maps) described by it are invertible. In stark contrast, the ncp-maps corresponding to measurement and discarding are rarely invertible. There are, however, many processes in between which might not be invertible, but are still pure enough to be 'reversed' in some consistent and useful fashion. For instance, the ncp-map adV has the obvious reversal (adV )† = adV ∗ (which is in general not the inverse.) Stinespring's theorem (135 IV) implies that every ncp-map into B(H ) splits as a reversible adV after a (possibly) non- reversible nmiu-map . Using Paschke's dilation, we see that every ncp-map factors as an nmiu-map before some particular ncp-map h. Is this ncp-map h reversible in some sense? In 171 VII we will see that these h are pure (in the ..167 -- 168.. 79 sense of 100 I) and in 215 III we will see that these pure maps are (in a sense) reversible. II The reversibility of pure maps (in a much more general setting) is the main result of the next chapter. In the remainder of this chapter, we study purity of ncp-maps a bit more. We consider a seemingly unrelated alternative notion of purity, which is based on Paschke's dilation, and show (in 171 VII) that this alternative notion is in fact equivalent to the original. We end this chapter by giving a different proof of the ncp-extremeness of pure and nmiu-maps. III Before we continue, let's rule out generalizations of more familiar notions of purity to arbitrary ncp-maps. 1. A state ϕ : A → C is called pure if it is an extreme point among all states. It is unreasonable to define an arbitrary ncp-map to be pure if it is extreme, as every nmiu-map (including von Neumann measurements and discard- ing) is extreme among the unital ncp-maps. This follows from 102 III and 102 V, but will be proven again in 172 VIII. In fact, in 172 XII we show that every ncp-map is the composition of two extreme maps. 2. Inspired by the GNS-correspondence between pure states and irreducible representations, Størmer defines a map ϕ : A → B(H ) to be pure if the only maps below ϕ in the completely positive ordering are scalar multiples of ϕ. For every central element z of A , the map a 7→ za is below idA and so the identity on non-factors is not pure in the sense of Størmer, even though the identity map is clearly reversible. IV Let's recall how we came to the notion of purity in 100 I, by generalizing adV . Pick V : H → K . By polar decomposition, see 82 I, we know that V = U A for A ≡ √V ∗V : H → ⌈A⌉ H and some isometry U : ⌈A⌉ H → K . Thus we have adV = adA ◦ adU . These two maps admit dual universal properties: adU is a corner and adA is a filter. An ncp-map ϕ is pure if it is the composition of a filter after a corner. It turns out that ϕ is pure in this sense if and only if the map on the left-hand side of its Paschke dilation is surjective. 169 We recall the definition of corner from [126, dfn. 2] and 95 I. II Definition An ncp-map h : A → B is a corner for a ∈ [0, 1]A if h(a) = h(1) and for every ncp-map f : A → C with f (1) = f (a), there is a unique ncp-map f′ : B → C with f = f′ ◦ h. III Remark In the next chapter we will study maps with a similar universal property as corners in a more general setting, but with all arrows in the reverse direction. To help alleviate some confusion caused by the change of direction, we call those maps comprehensions, see 199 II. See also [23]. Example The map ha : A → ⌊a⌋ A ⌊a⌋ given by b 7→ ⌊a⌋ b ⌊a⌋ is a corner of a ∈ [0, 1]A . We call ha the standard corner of a. See 98 I, 95 II or [126, prop. 5]. Lemma If (P, , h) is a Paschke dilation of a unital ncp-map, then h is a corner. Proof (This is a different proof of our similar result [127, cor. 27].) Let ϕ : A → B be any unital ncp-map. It is sufficient to prove h is a corner for the standard dilation (P, , h) constructed in 154 III, so P ≡ Ba(A ⊗ϕ B) and h(T ) = he, T ei, where e ≡ 1 ⊗ 1. Write p ≡ eihe. From he, ei = ϕ(1) = 1 it follows p is a projection. Define ϑ : B → pPp by ϑ(b) = e · bihe. Some straightforward computations (using 159 III) show ϑ is an miu-map with inverse pT p 7→ h(pT p). So ϑ is an miu-isomorphism and thus also normal. Furthermore h = ϑ−1(p(· )p) and so h is indeed a corner (of p). (cid:3) IV V VI We recall the definition of filter, see 96 I. Definition An ncp-map c : A → B is a filter for b ∈ B, b > 0 if c(1) 6 b and VII VIII for every ncp-map f : C → B with f (1) 6 b, there is a unique ncp-map f′ : C → A with f = c◦ f′. Remark The direction-reversed counterpart of filters are called quotients, see 197 II and [23] . Filters were introduced by us in [126, dfn. 2] under the name com- pressions. To avoid confusion with [4] and be closer to e.g. [132], we changed the name from compression to filter. Filters are named after polarization filters: the action of a polarization filter on the polarization of a photon is the same as a filter for the projection of the space of polarizations the polarization filter allows to pass without disturbance. Combining two polarization filters corresponds to composing the corresponding filters. This composed filter is in general a filter for an effect instead of a projection. This corresponds with the fact that there might not be a polarization of the photon anymore, with which it can pass both filters undisturbed. Example The map cb : ⌈b⌉ B ⌈b⌉ → B given by a 7→ √ba√b is a filter of b ∈ B, b > 0. We call cb the standard filter of b. See 96 V, 98 I or [126, prop. 6]. Exercise Let ϕ : A → B be any ncp-map between von Neumann algebras. IX X XI 1. Let c : B → C be any filter. Show that if (P, , h) is a Paschke dilation of ϕ, that then (P, , c◦ h) is a Paschke dilation of c◦ ϕ. 2. Assume c′ : C ′ → B is a filter of ϕ(1). Prove that there is a unique unital ncp-map ϕ′ with ϕ = c◦ ϕ′. Conclude that if (P, , h) is a Paschke dilation of ϕ′, that then (P, , c′ ◦ h) is a Paschke dilation of ϕ. (See 221 IV and 197 VII for the proofs in the more general case.) Exercise Derive (perhaps from X and 60 VIII), that filters are injective. XII ..168 -- 169 81 170 Definition Filters, corners and their compositions are called pure. Ia Remark This definition is equivalent to our previous definition of purity given in [127, dfn. 21]; see 100 III. II Examples Isomorphisms are pure. Less trivial examples follow. 1. The pure maps B(H ) → B(K ) are precisely of the form adT for some bounded operator T : K → H . 2. Let (P, , h) be a Paschke dilation of some ncp-map ϕ. In 169 V we saw that if ϕ is unital, that then h is a corner (and thus pure). In general, h = c◦ h′ for some filter c and corner h′ by 169 XI, and so h is pure. III Remark By definition, the set of pure maps is closed under composition. In the next chapter we will see (215 III) that there is a unique dagger † which turns the subcategory of von Neumann algebras with pure maps into a †-category such that f† is contraposed to f ; ⋄-positive maps are †-positive and †-positive maps have unique roots. For this unique †, we have (adV )† = adV ∗ . IV Exercise Show that any surjective nmiu-map between von Neumann algebras is a corner of a central projection, hence pure -- and conversely, that any corner of a central projection is a surjective nmiu-map. 171 We will prove that a map is pure if and only if the left-hand side map of its Paschke dilation is surjective. We need the following result. III II Theorem Let A be a von Neumann algebra together with a projection p ∈ A . Then a Paschke dilation of the standard corner hp : A → pA p, a 7→ pap is given by (⌈⌈p⌉⌉ A , h⌈⌈p⌉⌉, h′p), where ⌈⌈p⌉⌉ is the central carrier of p, see 68 I; h⌈⌈p⌉⌉ is the standard corner for ⌈⌈p⌉⌉ and h′p : ⌈⌈p⌉⌉ A → pA p is the restriction of hp. Proof (This is a simplified proof of the result we published earlier. [127, thm. 28]) Let (A ⊗hp pA p, , h) denote the Paschke dilation constructed in 154 III. We proceed in three steps: first we show A p is a self-dual Hilbert pA p-module, then we prove A ⊗hp pA p ∼= A p and finally we show Ba(A p) ∼= ⌈⌈p⌉⌉ A . IV Note A p is a pre-Hilbert pA p-module with scalar multiplication (αp) · (pap) = αpap and inner product hαp, api = pα∗ap. Since by the C∗-identity the norm on A p as a Hilbert module coincides with that as a subset of pA p; A p is norm closed in A and A is norm complete, we see A p is a Hilbert pA p-module. Recall A itself is a self-dual Hilbert A -module. The uniformity induced on A p by the ultranorm uniformity of A coincides with the ultranorm uni- formity of A p itself. The module A p is ultranorm closed in A as xα → x ultranorm implies xαp → xp ultranorm. As self-duality is equivalent to ultra- norm completeness, we see A p is self dual as well. Concerning A ⊗hp pA p ∼= A p, define B : A × pA p → A p by B(α, pap) = αpap. A straightforward computation shows B is a hp-compatible complex bilinear map (see 154 II) and so by 154 III, there is a unique bounded module map U : A ⊗hp pA p → A p fixed by U (α ⊗ pap) = B(α, pap) ≡ αpap. We will show U is an isomorphism. Clearly U is surjective. An easy computation shows a ⊗ pαp − apαp ⊗ p = 0 for all a, α ∈ A . So the set {a ⊗ p; a ∈ A } is ultranorm dense in A ⊗hp pA p. From this and hα ⊗ p, a ⊗ pi = pα∗ap = hU α ⊗ p, U a ⊗ pi we see U preserves the inner product. Hence U is an isomorphism with U∗ = U−1 given by U∗a = a ⊗ p. To show Ba(A p) ∼= ⌈⌈p⌉⌉ A , we will first find a convenient orthonormal ba- sis of A p. By 83 V, there are partial isometries (ui)i with Pi uiu∗i = ⌈⌈p⌉⌉ and u∗i ui 6 p. Write E = {ui; i ∈ I}. As ui = uip and ui = uiu∗i ui we see E is an orthonormal subset of A p. For any ap ∈ A p, we have ap = a⌈⌈p⌉⌉ p = ⌈⌈p⌉⌉ ap = (cid:0)Xi uiu∗i(cid:1)ap = Xi uiu∗i ap = Xi ui hui, api , where the sums converge ultrastrongly in A . Thus the last sum converges ultranorm in A p as well. Hence E is an orthonormal basis of A p. V VI Define 0 : A → Ba(A p) by 0 = adU ∗ ◦ . Note 0 is an nmiu-map and 0(α)ap = αap. The map 0 is surjective: indeed for any T ∈ Ba(A p) T (ap) = Xi T (ui)u∗i ap = (cid:0)Xi T (ui)u∗i(cid:1)ap = 0(cid:0)Xi T (ui)u∗i(cid:1)ap. Next, we show ⌈0⌉ = ⌈⌈p⌉⌉. It suffices to prove for all α ∈ A ,α > 0, that αap = 0 for all a ∈ A ⇐⇒ α⌈⌈p⌉⌉ = 0. Clearly, α⌈⌈p⌉⌉ = 0 implies αap = α⌈⌈p⌉⌉ ap = 0. For the converse, assume that we have αap = 0 for all a ∈ A . Then in particular αuip = 0, hence α⌈⌈p⌉⌉ = By 170 IV there is an nmiu-isomorphism ϕ : ⌈⌈p⌉⌉ A → Ba(A p) with 0 = αPi uiu∗i =Pi αuipu∗i = 0. Indeed ⌈0⌉ = ⌈⌈p⌉⌉. ϕ◦ h⌈⌈p⌉⌉. Hence h⌈⌈p⌉⌉ = (ϕ−1 ◦ adU ∗ )◦ . We compute h′p ◦ h⌈⌈p⌉⌉ = hp = h◦ = h◦ adU ◦ 0 = h◦ adU ◦ ϕ◦ h⌈⌈p⌉⌉. Using surjectivity of h⌈⌈p⌉⌉ and rearranging, we find h′p ◦(ϕ−1 ◦ adU ∗ ) = h. Thus indeed (⌈⌈p⌉⌉ A , h⌈⌈p⌉⌉, h′p) is a Paschke dilation of hp. (cid:3) Theorem Let ϕ : A → B be an ncp-map between von Neumann algebras with Paschke dilation (P, , h). The map ϕ is pure if and only if is surjective. Proof Assume is surjective. The kernel of is an ultraweakly closed two-sided ideal and so ker = zA for some central projection z, see 69 II. The standard corner hz⊥ : A → z⊥A is the quotient-map of zA and so by the isomorphism VII VIII 170 -- 171.. 83 theorem is a corner as well. In 170 II we saw h is pure. So ϕ is the composition of pure maps, hence pure itself. For the converse, assume ϕ is pure. For brevity, write p ≡ ⌈ϕ⌉. There is a unique ncp-map c with ϕ = c◦ hp, where h⌈ϕ⌉ is the standard corner of p. The map c is a filter, see 98 IX. By II, the triple (⌈⌈p⌉⌉ A , h⌈⌈p⌉⌉, h′p) is a Paschke dilation of hp, where h′p denotes the restriction of hp to ⌈⌈p⌉⌉ A . So by 169 XI we see (⌈⌈p⌉⌉ A , h⌈⌈p⌉⌉, c◦ h′p) is a Paschke dilation of ϕ. Clearly h⌈⌈p⌉⌉ is surjective. If (P′, ′, h′) is any other Paschke dilation of ϕ, then ′ = ϑ◦ h⌈⌈p⌉⌉ for some isomorphism ϑ and so ′ is surjective as well. (cid:3) 172 The order correspondence 157 IV has several corollaries. II Definition Assume ϕ : A → B is an ncp-map between von Neumann algebras. We say ϕ is ncp-extreme if ϕ is an extreme point among the ncp-maps with the same value on 1 [108] -- that is: for any 0 < λ < 1, ncp-maps ϕ1, ϕ2 : A → B with ϕ1(1) = ϕ2(1) = ϕ(1) we have λϕ1 + (1 − λ)ϕ2 = ϕ =⇒ ϕ1 = ϕ2 = ϕ. III Theorem Assume ϕ : A → B is an ncp-map with Paschke dilation (P, , h). Then the following are equivalent. 1. The map h is injective on (A )(cid:3), the commutant of (A ). 2. The map h is injective on [0, 1](A )(cid:3). 3. The map ϕ is ncp-extreme. V VI IV Proof (Based on [8, prop. 1.4.6] and [93, thm. 5.4].) We prove 1 ⇒ 2 ⇒ 3 ⇒ 1. (1 ⇒ 2) Trivial. (2 ⇒ 3) Assume h is injective on [0, 1](A )(cid:3). Suppose ϕ = λψ + (1 − λ)ψ′ for some 0 < λ < 1 and ncp ψ, ψ′ : A → B with ψ(1) = ψ′(1) = ϕ(1). Note 0 6ncp λψ 6ncp ϕ hence λψ = ϕt for some t ∈ [0, 1](A )(cid:3) by 157 IV. So h(λ1) = ϕλ1(1) = λϕ(1) = λψ(1) = ϕt(1) = h(t). VII By the injectivity of h, we find t = λ1, thus λψ = λϕ and so ψ = ϕ. The proof of ψ′ = ϕ is the same. Hence ϕ is ncp-extreme. (3 ⇒ 1) Suppose ϕ is ncp-extreme. Pick any a ∈ (A )(cid:3) with h(a) = 0. We want to show a = 0. Clearly h(a)R = h(aR) = 0 and h(a)I = h(aI) = 0, so it is sufficient to prove aR = aI = 0. Thus we may assume without loss of generality that a is self adjoint. By 170 II h is pure and thus splits as h = c◦ hp for some filter c : pPp → B and the standard corner hp : P → pPp for p = ⌈h⌉. By 169 XII filters are 4 p 6 ptp = λp 6 3 injective and so from the assumption 0 = h(a) = c(pap) it follows pap = 0. Using 9 X and some easy algebra, we can find 0 < µ, λ such that 1 4 6 µa+λ 6 3 4 . Define t = µa + λ. For the moment, note ϕt is ncp. As pap = 0, we see ptp = λp and so 1 4 p. Thus either 0 < λ < 1 or p = 0. If p = 0 then ϕ = 0 hence P = {0} and a = 0 as desired. For the non-trivial case, assume 0 < λ < 1. Note ϕt(1) = h(t) = c(ptp) = λc(p1p) = λϕ(1) and similarly ϕ1−t(1) = (1 − λ)ϕ(1). Define ψ1 = λ−1ϕt and ψ2 = (1−λ)−1ϕ1−t. By the previous, ψ1 and ψ2 are ncp-maps with ψ1(1) = ψ2(1) = ϕ(1). Also λψ1+(1−λ)ψ2 = ϕ. As ϕ is ncp- extreme, we must have ψ1 = ψ2 = ϕ. So ϕt = λϕ = ϕλ1, thus λ = t ≡ µa + λ. Rearranging: µa = 0. As µ > 0, we find a = 0, as desired. (cid:3) Corollary Any nmiu-map : A → B is ncp-extreme. (For a different proof, see e.g. [108, thm. 3.5].) Proof Easy: (B, , id) is a Paschke dilation of and id is injective. Theorem Any pure ncp-map ϕ : A → B is ncp-extreme. Proof By 100 III ϕ = c◦ hp for a filter c : pA p → B and projection p ≡ ⌈ϕ⌉. Combining 169 XI and 171 II we see (⌈⌈p⌉⌉ A , h⌈⌈p⌉⌉, c◦ hp) is a Paschke dilation of ϕ. Appealing to III, we are done if we can show c◦ hp is injec- tive on [0, 1]h⌈⌈p⌉⌉(A )(cid:3) . As h⌈⌈p⌉⌉ is surjective, h⌈⌈p⌉⌉(A )(cid:3) = Z(⌈⌈p⌉⌉ A ). As- sume t ∈ [0, 1]Z(⌈⌈p⌉⌉A ) with c(ptp) = 0 We want to show t = 0. As c is in- jective by 169 XII, we must have ptp = 0. So p 6 1 − ⌈t⌉ and ⌈⌈p⌉⌉ 6 1 − ⌈t⌉. Hence ⌈⌈p⌉⌉ 6 ⌈⌈p⌉⌉ − ⌈t⌉ 6 ⌈⌈p⌉⌉. So t 6 ⌈t⌉ = 0, as desired. (cid:3) Corollary Any ncp-map is the composition of two ncp-extreme maps. (cid:3) VIII IX X XI XII ..171 -- 172 85 Chapter 3 Diamond, andthen, dagger In the previous chapter and in [124] we have studied categorical properties of von Neumann algebras. In this chapter we change pace: we study categories that resemble the category of von Neumann algebras. The goal of this line of study is to identify axioms which uniquely pick out the category of von Neumann algebras. We do not reach this ambitious goal -- instead we will build up to a characterization of the existence of a unique †-structure on the pure maps of a von Neumann algebras-like category. 173 As a basic axiom we will require the categories we consider to be effectuses (which will be defined in 180 I). For us the definition of an effectus will play a similar role to that of a topological space for a geometer. In many applications, general topological spaces on their own are of little interest: the axioms are so weak that there are many (for the application) pathological examples. These weak axioms, however, are very expressive in the sense that they allow for the definition of many useful notions. Herein lies the strength of topological spaces -- as a stepping stone: many important classes of mathematical spaces are just plain topological spaces with a few additional axioms. Similarly, there will be many pathological effectuses. Their use for us lies in their expressiveness. Before we dive into effectuses, we need to learn about effect algebras (and related structures), which were introduced by mathematical physicists as a generaliza- tion of Boolean algebra to study fuzzy predicates in quantum mechanics. For us, effectuses are the categorical counterpart of effect algebras.∗ After this, we con- tinue with a brief, but thorough development of the basic theory of effectuses. There is a lot more to say about effectuses (see [23]), which has been omitted to avoid straining the reader. We will indulge, however, in one tangent, which II III ∗Effect algebroids [97, 98] are a different categorification with applications in cohomology. 173.. 87 can be skipped: the study of abstract convex sets in §3.2.4. For the moment, think of an effectus as a generalization of the opposite category of von Neumann algebras: the objects represent the physical systems (or data types, if you like) and the maps represent the physical processes (or properly typed programs). IV The first two axioms we add on top of those of an effectus are the existence of quotients and comprehension. It's helpful to discuss the origin of these axioms. It started with the desire to axiomatize categorically the sequential product on a von Neumann algebra A , that is the operation a & b ≡ √ab√a, which rep- resents sequential measurement of first a andthen b. The sequential product is a quantum mechanical generalization of classical conjunction (logical 'and'). In contrast to the tame ∧ in a Boolean algebras, the sequential product does not obey a lot of insightful formulas. The sequential product as a binary operation has received quite some attention [6, 37, 44 -- 47, 71, 72, 83, 104, 111, 122]. We ap- proach the sequential product in a rather different way: we take a step back and consider the map asrta(b) ≡ a & b for fixed a. (asrta is named after the assert statement used in many programming languages, see 211 III.) This map asrta factors in two: A b ✤ π / ⌈a⌉ A ⌈a⌉ / ⌈a⌉ b ⌈a⌉ ✤ ξ / A / √ab√a. It turned out that both π and ξ have nice and dual (in a sense) defining universal properties which can be expressed in an effectus: π is a comprehension and ξ is a quotient.† V The existence of quotients and comprehension in some effectus C is interest- ing on its own, but with two additional axioms (the existence of images and preservation of images under orthocomplementation), we get a surprisingly firm handle on the possibilistic side of C, by means of the existence of a certain functor ( )⋄ : C → OMLat to the category of orthomodular lattices. In the guid- ing example of von Neumann algebras, we have f⋄ = g⋄ if and only there are no post-measurement a and initial state ω that can determine with certainty whether f or g has been performed.‡ With this functor we can introduce several possibilistic notions of which ⋄-adjoint and ⋄-positive are the most important. Because of the central role of ⋄, we call an effectus with these axioms a ⋄-effectus. VI The axioms of a ⋄-effectus do not force any coherence between the quotients and comprehension. This has been a roadblock to the axiomatization of the sequential product for quite a while. The key insight was the following: the map asrta (i.e. b 7→ a & b) is the unique ⋄-positive map on A with asrta(1) = a. We turn this theorem into an axiom: an &-effectus is a ⋄-effectus where there are †Beware: an effectus quotient is not the same thing as a von Neumann-algebra quotient. ‡In symbols: f⋄ = g⋄ ⇐⇒ (cid:0)∀a, ω. ω(f (a)) = 0 ⇔ ω(g(a)) = 0(cid:1). / / / / such unique ⋄-positive maps asrta and where an additional polar decomposition axiom holds. In an &-effectus, the predicates on an object carry a canonical binary operation & which is the intended sequential product in the case of von Neumann algebras. VII In an effectus, we call a map pure if it can be written as the composition of quotients and comprehensions. In 171 VII we saw that in the case of von Neumann algebras, a map is pure (in this sense) if and only if the corresponding Paschke embedding is surjective. In particular, the pure maps B(H ) → B(K ) are exactly of the form adv (for bounded operators v : K → H ) which carry a †- structure, namely (adv)† = adv∗ . This begs the question: can we define a † on all pure maps in an &-effectus? The main result of this chapter is Theorem 215 III, which gives necessary and sufficient conditions for an &-effectus to have a well- behaved †-structure on its pure maps. After this apogee, we tie up some lose ends and discuss how these structures compare to those already known in the literature. 3.1 Effect algebras and related structures Before we turn to effectuses, it is convenient to introduce some lesser-known algebraic structures of which the effect algebra is the most important. It will turn out that the set of predicates associated to an object in an effectus can be arranged into an effect algebra. In this way, effect algebras play the same role for effectuses as Heyting algebras for toposes. First, we will recall the definition of partial commutative monoid (PCM) -- as later in 180 X we will see that the partial maps between two objects in an effectus can be arranged into a PCM. Definition A partial commutative monoid (PCM) M is a set M together with distinguished element 0 ∈ M and a partial binary operation > such that for all a, b, c ∈ M -- writing a ⊥ b whenever a > b is defined -- we have 1. (partial commutativity) if a ⊥ b, then b ⊥ a and a > b = b > a; 2. (partial associativity) if a ⊥ b and a > b ⊥ c, then b ⊥ c, a ⊥ b > c and (a > b) > c = a > (b > c) and 174 II 3. (zero) 0 ⊥ a and 0 > a = a. A map f : M → N between PCMs M and N is called an PCM homomorphism if for all a, b ∈ E with a ⊥ b, we have f (a) ⊥ f (b) and f (a) > f (b) = f (a > b). Write PCM for the category of PCMs with these homomorphisms. (Cf. [65,121]) For a, b in a PCM M , we say a 6 b iff a > c = b for some c ∈ M . Exercise Show a PCM is preordered by 6. ..173 -- 174.. III 89 IV The partial commutativity and associativity of a PCM ensure that a sum only depends on which elements occur (and how often), instead of their order. For- mally: if (··· (x1 > x2) >··· ) > xn exists some x1, . . . , xn in some PCM, then so does (··· (xπ(1) > xπ(2)) > ··· ) > xπ(n) for any permutation π of {1, . . . , n} and (··· (x1 > x2) > ··· ) > xn = (··· (xπ(1) > xπ(2)) > ··· ) > xπ(n). Thus parantheses are superfluous and we will often leave them out: n > i=1 xi ≡ x1 > ··· > xn ≡ (··· (x1 > x2) > ··· ) > xn. 175 Definition A PCM E together with distinguished element 1 ∈ E is called an effect algebra (EA) [31] provided that 1. (orthocomplement) for every a there is a unique a⊥ with a > a⊥ = 1 and 2. (zero -- one) if a ⊥ 1, then a = 0. A map f : E → F between effect algebras E and F is called an effect algebra homomorphism if it is a PCM homomorphism that preserves 1; concretely: 1. (additive) a ⊥ b implies f (a) ⊥ f (b) and f (a) > f (b) = f (a > b) and 2. (unital) f (1) = 1. (It follows that f (0) = 0 and f (a⊥) = f (a)⊥. See 176 V.) Write EA for the category of effect algebras with these homomorphisms. A subset D ⊆ E is a sub-effect algebra of E if 0, 1 ∈ D and for any a, b ∈ D with a ⊥ b, we have a > b ∈ D and a⊥ ∈ D. Examples There are many examples of effect algebras -- we only give a few. 1. The unit interval [0, 1] with partial addition is an effect algebra -- i.e.: x ⊥ y whenever x + y 6 1 and then x > y = x + y, x⊥ = 1 − x. II 2. Generalizing the previous: if G is an ordered group with distinguished element 1, then the order interval [0, 1]G ≡ {x; x ∈ G; 0 6 x 6 1} is an effect algebra with x ⊥ y whenever x+y 6 1; x>y = x+y and x⊥ = 1−x. if A is any von Neumann algebra, then the set of ef- fects [0, 1]A forms an effect algebra with a ⊥ b whenever a + b 6 1; a > b = a + b and a⊥ = 1− a. The 'effect' in effect algebra originates from this example. 3. In particular, 4. Any orthomodular lattice L (defined in 177 IV) is an effect algebra with the same orthocomplement, x ⊥ y whenever x 6 y⊥ and x > y = x ∨ y. (See e.g. [129, prop. 27].) 5. In particular, any Boolean algebra L is an effect algebra with complement as orthocomplement, x ⊥ y whenever x ∧ y = 0 and x > y = x ∨ y. 6. The one-element Boolean algebra 1 ≡ {0 = 1} is the final object in EA and the two-element Boolean algebra 2 ≡ {0, 1} is the initial object in EA. Exercise Let E and F be effect algebras. First show that the cartesian prod- uct E × F is an effect algebra with componentwise operations. Show that this is in fact the categorical product of E and F in EA. (The category EA is in fact complete and cocomplete; for this and more categorical properties, see [65].) Exercise There is a small redundancy in our definition of effect algebra: show that the zero axiom (x > 0 = x) follows from the remaining axioms (partial commutativity, partial associativity, orthocomplement and zero -- one.) Proposition In any effect algebra E with a, b, c ∈ E, we have 1. (involution) a⊥⊥ = a; 2. 1⊥ = 0 and 0⊥ = 1; 3. (positivity) if a > b = 0, then a = b = 0; 4. (cancellation) if a > c = b > c, then a = b; 5. the relation 6 (from 174 II) partially orders E; 6. a 6 b if and only if b⊥ 6 a⊥; 7. if a 6 b and b ⊥ c, then a ⊥ c and a > c 6 b > c and 8. a ⊥ b if and only if a 6 b⊥. Proof (These proofs are well-known, see for instance [29].) By partial commu- tativity and definition of orthocomplement both a⊥ > a = 1 and a⊥ > a⊥⊥ = 1. So by uniqueness of orthocomplement, we must have a = a⊥⊥, which is point 1. Clearly 0 > 1 = 1, so 1⊥ = 0 and 0⊥ = 1, which is point 2. For point 3, assume a > b = 0. Then a > b ⊥ 1 and so by partial associativity b ⊥ 1. By zero -- one, we get b = 0. Similarly a = 0, which shows point 3. For point 4, assume a > c = b > c. From partial associativity and commutativity, we get ((a > b)⊥ > a) > c = ((a > b)⊥ > a) > b = 1 and so by uniqueness of or- thocomplement c = ((a > b)⊥ > a)⊥ = b, which is point 4. By 174 III, we only need to show that 6 is antisymmetric for point 5. So assume a 6 b and b 6 a for some a, b ∈ E. Pick c, d ∈ E with a > c = b and b > d = a. Then a = (a > c) > d = a > (c > d). By cancellation c > d = 0. So by pos- itivity c = d = 0. Hence a = b. For point 6, assume a 6 b. Pick c ∈ E ..174 -- 175.. 91 III IV V VI with a > c = b. Clearly a > a⊥ = 1 = b > b⊥ = a > c > b⊥, so by cancel- lation a⊥ = c > b⊥, which is to say b⊥ 6 a⊥. For point 7, assume a 6 b and b ⊥ c. Pick d with a > d = b. By partial associativity and commutativity we have b > c = (a > d) > c = (a > c) > d, so a ⊥ c and a > c 6 b > c. For point 8, first assume a ⊥ b. Then a > b > (a > b)⊥ = 1 = b > b⊥. So by can- cellation b⊥ = a > (a > b)⊥, hence a 6 b⊥. For the converse, assume a 6 b⊥. Then a > c = b⊥ for some c. Hence a > c ⊥ b and so by partial associativity and commutativity, we get a ⊥ b, as desired. (cid:3) 176 Definition Suppose E is an effect algebra. Write b ⊖ a for the (by cancellation) II unique element (if it exists) with a > (b ⊖ a) = b. Exercise* Show that for any effect algebra E, we have (D1) a ⊖ b is defined if and only if b 6 a; (D2) a ⊖ b 6 a (if defined); (D3) a ⊖ (a ⊖ b) = b (if defined) and (D4) if a 6 b 6 c, then c ⊖ b 6 c ⊖ a and (c ⊖ a) ⊖ (c ⊖ b) = b ⊖ a. III Let E be a poset with maximum element 1 and partial binary operation ⊖ satisfying (D1) -- (D4). Define a > b = c ⇔ c ⊖ b = a and a⊥ = 1 ⊖ a. Show that this turns E into an effect algebra with compatible order and ⊖. IV Remark Such a structure (E,⊖, 1) is called a difference-poset (D-poset) [77] and is, as we have just seen, an alternative way to axiomatize effect algebras. V Exercise Show that for an effect algebra homomorphism f : E → F , we have 1. 2. 3. (preserves zero) f (0) = 0; (order preserving) if a 6 b, then f (a) 6 f (b); if a ⊖ b is defined, then f (a ⊖ b) = f (a) ⊖ f (b) and 4. consequently f (a⊥) = f (a)⊥. 177 For real numbers we have a + b = min{a, b} + max{a, b}. The following is a generalization to effect algebras. Ia Proposition Suppose E is an effect algebra. If the infimum a ∧ b exists for some a, b ∈ E with a ⊥ b, then their supremum a ∨ b exists as well and a > b = (a ∧ b) > (a ∨ b). Proof (This result appeared in my master's thesis [129, prop. 15]. It turns out that it was already proven before (in D-poset form) [29, prop. 1.8.2].) The result if (x ⊖ c) ∧ (x ⊖ d) exists, then c ∨ d exists as well follows from this lemma: and (x ⊖ c) ∧ (x ⊖ d) = x ⊖ (c ∨ d). Indeed: a ∧ b = ((a > b) ⊖ a) ∧ ((a > b) ⊖ b) = (a > b) ⊖ (a ∨ b) and so (a ∧ b) > (a ∨ b) = a > b, as desired. Now, we prove the lemma. Assume (x⊖c)∧(x⊖d) exists. Note x⊖c = (x⊥ >c)⊥. As z 7→ z⊥ is an order anti-isomorphism, we see (x⊥ > c) ∨ (x⊥ > d) exists and (x⊥ > c)∨ (x⊥ > d) = ((x⊖ c)∧ (x⊖ d))⊥ . Clearly (x⊥ > c)∨ (x⊥ > d) > x⊥. Write r := ((x⊥ > c) ∨ (x⊥ > d)) ⊖ x⊥. We will show r = c ∨ d. It is easy to see r > c and r > d. Assume s is any element with s > c and s > d. Note x⊥ > s > x⊥ > c and x⊥ > s > x⊥ > d, hence x⊥ > s > (x⊥ > c) ∨ (x⊥ > d) and so s > r. We have shown r = c ∨ d. Consequently (x ⊖ c) ∧ (x ⊖ d) = ((x⊥ > c) ∨ (x⊥ > d))⊥ = (r > x⊥)⊥ = x ⊖ (c ∨ d), as promised. (cid:3) Definition An ortholattice L is a bounded lattice with orthocomplement -- that is: it has a minimum 0, a maximum 1 and a unary operation ( )⊥ satisfying 1. a ∧ a⊥ = 0; 2. a ∨ a⊥ = 1; 3. a 6 b ⇒ b⊥ 6 a⊥ and 4. a⊥⊥ = a. An ortholattice is orthomodular provided a 6 b =⇒ a ∨ (a⊥ ∧ b) = b. See, for instance [11, 29, 74]. Example The lattice of projections in a von Neumann algebra is an orthomod- ular lattice with orthocomplement p⊥ ≡ 1 − p. Proposition An effect algebra E that is an ortholattice is also orthomodular. Proof (For a different proof, see [29, prop. 1.5.8].) Assume a 6 b. We have to show a ∨ (a⊥ ∧ b) = b. Note a ∧ (a⊥ ∧ b) 6 a ∧ a⊥ = 0 and so by Ia we have a ∨ (a⊥ ∧ b) = a > (a⊥ ∧ b). Thus it is sufficient to prove a⊥ ∧ b = b ⊖ a. Note a⊥ = (b ⊖ a) > b⊥ and b = (b ⊖ a) > a. Similar to II one can show that (b ⊖ a) > (b⊥ ∧ a) = ((b ⊖ a) > b⊥) ∧ ((b ⊖ a) > a) ≡ a⊥ > b, but b⊥ ∧ a 6 b⊥ ∧ b = 0 and so a⊥ > b = b ⊖ a, as desired. 3.1.1 Effect monoids (cid:3) II III IV V VI VII In 190 II we will see that the scalars of an effectus form an effect monoid: 178 ..175 -- 178.. 93 II Definition An effect monoid M [56] is an effect algebra together with a binary operation ⊙ such that for all a, b, c, d ∈ M , we have 1. (unit) 1 ⊙ a = a ⊙ 1 = a; 2. (associativity) (a ⊙ b) ⊙ c = a ⊙ (b ⊙ c) and 3. (distributivity) if a ⊥ b and c ⊥ d, then the following sum exists and furthermore (a ⊙ c) > (b ⊙ c) > (a ⊙ d) > (b ⊙ d) = (a > b) ⊙ (c > d). (Phrased categorically: an effect monoid is a monoid in EA with the obvious tensor product that relates bimorphisms to morphisms, see [56, 65].) An effect monoid M is said to be commutative if we have a ⊙ b = b ⊙ a for all a, b ∈ M . A map f : M → N between effect monoids is called an effect monoid homomorphism if it is an effect algebra homomorphism and furthermore f (a ⊙ b) = f (a) ⊙ f (b) for all a, b ∈ M . Examples Effect monoids are less abundant than effect algebras. III 1. The effect algebra [0, 1] is a commutative effect monoid with the usual product. (This is the only way to turn [0, 1] into an effect monoid [129, prop. 41].) 2. We saw earlier that every Boolean algebra is an effect algebra with x > y = x ∨ y defined iff x ∧ y = 0. The Boolean algebra is turned into an effect monoid with x⊙ y ≡ x∧ y. In fact, every finite (not necessarily commuta- tive) effect monoid is of this form [129, prop. 40] and thus commutative. 3. In particular: the two-element Boolean algebra 2 = 0, 1 from 175 II is an effect monoid with x ⊙ y = x ∧ y. 4. There is a non-commutative effect monoid based on the lexicographically ordered vector space R5, see [129, cor. 51]. 5. Let M be an effect monoid. Write M op for the effect monoid M with the opposite multiplication -- that is: a ⊙M b = b ⊙M op a. IV V IIIa Exercise Show that a ⊙ 0 = a = 0 ⊙ a for any a in an effect monoid M . Later, in a tangent, we will need the following specific fact about effect monoids. Exercise Assume M is an effect monoid with a1, . . . , an, b1, . . . , bn ∈ M such that >i ai = 1 and >i ai ⊙ bi = 1. Prove ai ⊙ bi = ai for every 1 6 i 6 n. 3.1.2 Effect modules 179 We will see that in an effectus, the effect monoid of scalars will act on every effect algebra of predicates, turning them into effect modules (see 190 II): Definition Suppose M is an effect monoid. An effect module E over M [65] is an effect algebra together with an operation M × E → E denoted by (λ, a) 7→ λ · a such that for all a, b ∈ E and λ, µ ∈ M , we have II 1. (λ ⊙ µ) · a = λ · (µ · a); 2. if a ⊥ b, then λ ⊙ a ⊥ λ ⊙ b and (λ ⊙ a) > (λ ⊙ b) = λ ⊙ (a > b); 3. if λ ⊥ µ, then λ ⊙ a ⊥ µ ⊙ a and (λ ⊙ a) > (µ ⊙ a) = (λ > µ) ⊙ a and 4. 1 ⊙ a = a. (Categorically: an effect module over M is an M -action.) An effect algebra homomorphism f : E → F between effect modules over M is an M -effect mod- ule homomorphism provided λ · f (a) = f (λ · a) for all λ ∈ M and a ∈ E. Write EModM for the category of effect modules over M with effect module homomorphisms between them. Examples There are many effect modules. III 1. Every effect algebra is an effect module over the two-element effect monoid 2. (In fact EA ∼= EMod2.) The only effect module up-to-isomorphism over the one-element effect monoid 1 is the one-element effect algebra 1 itself. 2. Effect modules over [0, 1] are the same thing as convex effect algebras, see [43,48]. If V is an ordered real vector space with order unit u, then [0, u] is an effect module over [0, 1]. In fact, every effect module over [0, 1] is of this form [48]. See also [66, thm. 3] for the stronger categorical equivalence. 3.2 Effectuses An effectus comes in two guises: axiomatizing either a category of total maps or a category of partial maps. We will start off with the total form as it has the simplest axioms. Later we will prefer to work with the partial form. Definition A category C is said to be an effectus in total form [23, 57, 69] if 180 1. C has finite coproducts (hence an initial object 0) and a final object 1; ..178 -- 180.. 95 2. all diagrams of the following form§ are pullbacks X + Y id+! X + 1 !+id !+id ! X κ1 1 κ1 (3.1) 1 + Y / 1 + 1 id+! X + Y / 1 + 1 !+! 3. and the following two arrows are jointly monic. 1 + 1 + 1 [κ1,κ2,κ2] [κ2,κ1,κ2] 2 1 + 1 III IV An arrow f : X → Y + 1 is called a partial map and written f : X ⇀ Y . II One with an interest in physics might think of the objects of an effectus as physical systems and its arrows as the physical operations between them. The final object 1 is the physical system with a single state. The coproduct X + Y is the system that can be prepared as either X or as Y . An arrow 1 → X corresponds to a physical preparation of the system X and an arrow X → 1 + 1 is a yes -- no measurement. In this sense an effectus can be seen as a generalized probabilistic theory -- not unlike the operational-probabilistic theories (OPT) of Chiribella et al [17]. There are differences: for instance, OPTs are equipped with a parallel compo- sition of systems, effectuses are not and OPTs always have the unit interval as scalars, effectuses might not. In [113] Tull compares effectuses and OPTs. There are other approaches to categorical probabilistic theories; see, for in- stance [40, 131]. Studying programming languages, one would do better thinking of the objects of an effectus as data types and its arrows as the allowed operations between them (semantics of programs). The final object 1 is the unit data type. The coproduct X + Y is the union data type of X and Y . An arrow 1 → X is a value of X and an arrow X → 1 + 1 is a predicate on X. V Our main example of an effectus in total form is the category vNop of von Neumann algebras with completely positive normal unital maps in the opposite direction. (To see vNop is an effectus in total form, adapt the proof of 191 II.) The partial maps correspond to ncp-maps f with f (1) 6 1. (Equivalently: contractive ncp-maps.) Some other examples appear later on in 191 II, 192 III, 196 II and 189a I. For a comprehensive list of examples, see [23]. §We write κi for coproduct coprojections; square brackets [f, g] for coproduct cotupling; h + k = [κ1 ◦ h, κ2 ◦ k] and ! for the unique maps associated to either the final object 1 or initial object 0. / / / / / / , , 2 Let C be an effectus in total form. Given two arrows f : X → Y + 1 and g : Y → Z + 1 (i.e. partial maps X ⇀ Y and Y ⇀ Z) their composition as partial maps is defined as g ◦ f ≡ [g, κ2]◦ f . Write Par C for the category of partial maps, which has the same objects as C, but as arrows X → Y in Par C we take arrows of the form X → Y + 1 in C, which we compose using ◦ and with identity on X in Par C given by κ1 : X → X + 1.¶ The category Par C is not an effectus in total form -- instead it is an effectus in partial form. VI Definition A category C is called an effectus in partial form [19, 23] if VII 1. C is a finPAC‖ [19] (cf. [5]) -- that is: (a) C has finite coproducts (+, 0); (b) C is PCM-enriched -- that is: i. every homset C(X, Y ) has a partial binary operation > and dis- tinguished element 0 ∈ C(X, Y ) that turns C(X, Y ) into a partial commutative monoid, see 174 II; ii. if f ⊥ g then both (h◦ f ) ⊥ (h◦ g) and (f ◦ k) ⊥ (g ◦ k) and h◦(f > g) = (h◦ f ) > (h > g) for any f, g : X → Y , h : Y → Y ′ and k : X′ → X and iii. (zero) 0 ◦ f = 0 and f ◦ 0 = 0 for any f : X → Y and (f > g)◦ k = (f ◦ k) > (g ◦ k) (c) (compatible sum) for any b : X → Y + Y we have ⊲1 ◦ b ⊥ ⊲2 ◦ b, where ⊲i : Y + Y → Y are partial projectors∗∗ defined by ⊲1 ≡ [id, 0] and ⊲2 ≡ [0, id] and (d) (untying) if f ⊥ g, then κ1 ◦ f ⊥ κ2 ◦ g, where κ1 and κ2 are copro- jections on the same coproduct and 2. it has effects -- that is: there is a distinguished object I such that (a) for each object X, the PCM C(X, I) ≡ Pred X is an effect algebra, see 175 I -- in particular C(X, I) has a maximum element 1; (b) if 1 ◦ f ⊥ 1 ◦ g, then f ⊥ g and (c) if 1 ◦ f = 0, then f = 0. In this setting, a map f : X → Y is called total if 1 ◦ f = 1. Remark The untying and zero axioms are redundant: they follows from the others. We include them, as they are part of the definition of finPAC. VIII ¶In categorical parlance: Par C is the Kleisli category of the monad ( ) + 1 : C → C. ‖Finitarily partially additive category. ∗∗Later, in 181 VII, we will use the more slightly more general ⊲1 : X1 + X2 → X1 and ⊲2 : X1 + X2 → X2 defined by ⊲1 = [id, 0] and ⊲2 = [0, id]. ..180.. 97 IX At first glance an effectus in partial form seems to have a much richer structure than an effectus in total form. This is not the case -- effectuses in total and partial form are two views on the same thing [19, 23]: X Theorem (Cho) Let C be an effectus in total form and D be an effectus in partial form. 1. The category Par C is an effectus in partial form with I = 1. 2. The total maps of D form an effectus in total form Tot D. 3. Nothing is lost: Par(Tot D) ∼= D and Tot(Par C) ∼= C. XI (The result can be rephrased categorically as a 2-equivalence of the 2-category of effectuses in partial form and the 2-category of effectuses in total form. For the details, see [19, §5].) To prove the Theorem (in 188 II), we need some prepa- ration. 3.2.1 From partial to total 181 We will first show that the subcategory of total maps of an effectus in partial form is an effectus in total form. This proof and especially the demonstration that the squares in (3.1) are pullbacks, will elucidate the axioms of an effectus in total form and will make the proof in the opposite direction more palatable. II III Lemma In an effectus in partial form, coprojections are total. Proof (For the original and different proof see [19, lem. 4.7(5)].) Let κ1 : X → X + Y be any coprojection. We have to show 1 ◦ κ1 = 1. Note that the first 1 denotes the maximum of Pred X + Y and the second the maximum of Pred X. Hence 1X = 1X ◦ idX = (1X ◦[idX , 0])◦ κ1 6 1X+Y ◦ κ1 6 1X . Indeed 1 ◦ κ1 = 1. (cid:3) IV Proposition In an effectus in partial form, the cotupling bijection (f, g) 7→ [f, g] is a PCM-isomorphism [23] -- that is: 1. [f, g] ⊥ [f′, g′] if and only if f ⊥ f′ and g ⊥ g′; 2. if [f, g] ⊥ [f′, g′], then [f, g] > [f′, g′] = [f > f′, g > g′] and 3. [0, 0] = 0. Furthermore [1, 1] = 1 for maps into I, so the cotupling map is an effect algebra isomorphism Pred(X) × Pred(Y ) ∼= Pred(X + Y ). Proof First we show [h, l] = [h, 0] > [0, l]. By the compatible sum axiom V [h, 0] = ⊲1 ◦(h + l) ⊥ ⊲2 ◦(h + l) = [0, l]. By PCM-enrichment ([h, 0] > [0, l])◦ κ1 = ([h, 0]◦ κ1) > ([0, l]◦ κ1) = h. Simi- larly ([h, 0] > [0, l])◦ κ2 = l. Thus indeed [h, l] = [h, 0] > [0, l]. Assume [f, g] ⊥ [f′, g′]. By PCM-enrichment we have f = [f, g]◦ κ1 ⊥ [f′, g′]◦ κ1 = f′. Similarly g ⊥ g′. Conversely, assume f ⊥ f′ and g ⊥ g′. Again, by PCM-enrichment [f, 0] = f ◦ ⊲1 ⊥ f′ ◦ ⊲1 = [f′, 0] and [f > f′, 0] = [f, 0] > [f′, 0]. Similarly [0, g > g′] = [0, g] > [0, g′]. Putting it all together: [f, g] > [f′, g′] = [f, 0] > [0, g] > [f′, 0] > [0, g′] = [f > f′, 0] > [0, g > g′] = [f > f′, g > g′]. To show cotupling is a PCM-isomorphism, it only remains to be shown that [0, 0] = 0. As 0 ◦ κ1 = 0 and 0 ◦ κ2 = 0, we indeed have 0 = [0, 0]. Similarly by II we have 1 ◦ κ1 = 1 and 1 ◦ κ2 = 1, so 1 = [1, 1]. (cid:3) The coproduct in an effectus in partial form is almost a (bi)product: Proposition In an effectus in partial form, we have a bijective correspondence VI VII h : Z → X + Y f : Z → X g : Z → Y 1 ◦ f ⊥ 1 ◦ g as follows [19, lem. 4.8]: for every f : Z → X and g : Z → Y with 1 ◦ f ⊥ 1 ◦ g, there is a unique map hf, gi : Z → X + Y such that ⊲1 ◦hf, gi = f and ⊲2 ◦hf, gi = g, where ⊲1 = [id, 0] and ⊲2 = [0, id]. In fact hf, gi = (κ1 ◦ f ) > (κ2 ◦ g). Conversely h = h⊲1 ◦ h, ⊲2 ◦ hi with 1 ◦ ⊲1 ◦ h ⊥ 1 ◦ ⊲2 ◦ h for h : Z → X + Y . Proof Let f : Z → X and g : Z → Y be given such that 1 ◦ f ⊥ 1 ◦ g. By II, we have 1 ◦ κ1 ◦ f = 1 ◦ f ⊥ 1 ◦ g = 1 ◦ κ2 ◦ g. Thus κ1 ◦ f ⊥ κ2 ◦ g. Define hf, gi = (κ1 ◦ f ) > (κ2 ◦ g). By the PCM-enrichedness, we have VIII ⊲1 ◦hf, gi = [id, 0]◦((κ1 ◦ f ) > (κ2 ◦ g)) = ([id, 0]◦ κ1 ◦ f ) > ([id, 0]◦ κ2 ◦ g) = f > (0 ◦ g) = f. Similarly ⊲2 ◦hf, gi = g. To show uniqueness, assume f = ⊲1 ◦ h and g = ⊲2 ◦ h for some h : Z → X + Y . Note κ1 ◦ ⊲1 = [κ1, 0] and κ2 ◦ ⊲2 = [0, κ2], so by IV we have (κ1 ◦ ⊲1) > (κ2 ◦ ⊲2) = [κ1, κ2] = id, hence hf, gi = h⊲1 ◦ h, ⊲2 ◦ hi = (κ1 ◦ ⊲1 ◦ h) > (κ2 ◦ ⊲2 ◦ h) = ((κ1 ◦ ⊲1) > (κ2 ◦ ⊲2))◦ h = id◦ h = h, ..180 -- 181.. 99 which demonstrates uniqueness. Finally, let h : Z → X+Y be any map. We must show that h = h⊲1 ◦ h, ⊲2 ◦ hi. Note 1 ◦ ⊲1 = [1, 0] and 1 ◦ ⊲2 = [0, 1]. So by PCM-enrichment 1 ◦ ⊲1 ◦ h ⊥ 1 ◦ ⊲2 ◦ h and so ⊲1 ◦ h ⊥ ⊲2 ◦ h. By the previous h⊲1 ◦ h, ⊲2 ◦ hi = (κ1 ◦ ⊲1 ◦ h)> (κ2 ◦ ⊲2 ◦ h) = h as desired. (cid:3) IX Exercise Show that in an effectus in partial form, we have 1. [a, b]◦hf, gi = (a◦ f ) > (b ◦ g); 2. 1 ◦hf, gi = (1 ◦ f ) > (1 ◦ g); 3. (k + l)◦hf, gi = hk ◦ f, l ◦ gi and hf, gi◦ k = hf ◦ k, g ◦ ki 4. X assuming 1 ◦ f ⊥ 1 ◦ g. [23] Remark In an effectus in partial form, there is a straightforward generalization of the bijective correspondence VII to h : Z → X1 + ··· + Xn f1 : Z → X1 ··· fn : Z → Xn ⊥n 1 ◦ fn with fi = ⊲i ◦ h and h = hf1, . . . , fni ≡ (κ1 ◦ f1) > ··· > (κn ◦ fn), where ⊲i is defined in the obvious way. The expected generalizations of the rules in IX also hold. XI Theorem Let C be an effectus in partial form. The category of total maps of C is an effectus in total form. [19, thm. 4.10] XII Proof The total maps indeed form a subcategory: 1 ◦ id = 1 and 1 ◦ f ◦ g = 1 ◦ g = 1 for composable total f, g. In II we saw coprojections are total. By IV we have 1 ◦[f, g] = [1 ◦ f, 1 ◦ g] = [1, 1] = 1 for total f : X → Z and g : Y → Z, so Tot C has binary coproducts. The unique map ! : 0 → X must be total as 1 ◦! = 1 is the unique map 0 → I, so Tot C has initial object 0, hence all finite coproducts. XIII To show I is the final object of Tot C, we need idI = 1. As C(I, I) is an effect algebra 1 = id > id⊥ for some id⊥. So by PCM-enrichment 1 = 1 ◦ 1 = (1 ◦ id)>(1 ◦ id⊥) = 1>(1 ◦ id⊥). By the zero -- one axiom 1 ◦ id⊥ = 0. So id⊥ = 0 and indeed id = 1. To show I is final in Tot C, pick any object X in Tot C. We claim 1 : X → I is the unique total map. Indeed, by the previous 1 = idI ◦ 1 = 1 ◦ 1, so 1 is total and if h : X → I is total, then 1 = 1 ◦ h = idI ◦ h = h. To show the square on the left of (3.1) is a pullback in Tot C, let f : Z → X + I and g : Z → I +Y be total maps with (id+1)◦ g = (1+id)◦ f . By VII, f = hα, ai and g = hb, βi for some maps α : Z → X, β : Z → Y and a, b : Z → I. XIV Z hα,ai≡f hα,βi ( X + Y id+1 g≡hb,βi 1+id I + Y id+1 X + I 1+id / I + I By IX, we have 1 = 1 ◦ f = 1 ◦hα, ai = (1 ◦ α) > a and so a⊥ = 1 ◦ α. Simi- larly b⊥ = 1 ◦ β. Again, using IX, we see that hb, 1 ◦ βi = (id + 1)◦hb, βi = (id + 1)◦ g = (1 + id)◦ f = h1 ◦ α, ai, so 1 ◦ β = a = (1 ◦ α)⊥, hence 1 ◦ β ⊥ 1 ◦ α, so hα, βi : Z → X + Y exists and is total as 1 ◦hα, βi = (1 ◦ α) > (1 ◦ β) = 1. We compute (id + 1)◦hα, βi = hα, 1 ◦ βi = hα, ai = f . Similarly (1 + id)◦hα, βi = g. Assume h : Z → X + Y is any map with (1 + id)◦ h = g and (id + 1)◦ h = f . Say h = hh1, h2i. Then hα, ai = f = (id + 1)◦ h = hh1, 1 ◦ h2i, so α = h1. Similarly β = h2. Thus h = hα, βi, which shows our square is indeed a pullback. To show the square on the right of (3.1) is a pullback in Tot C, assume (using VII) hα, βi : Z → X + Y is some total map with (1 + 1)◦hα, βi = κ1 ◦ 1. XV Z α hα,βi 1 ( X κ1 X + Y 1 1+1 I κ1 / I + I With IX, we see h1 ◦ α, 1 ◦ βi = (1 + 1)◦hα, βi = κ1 ◦ 1 = h1, 0i. So α is total and β = 0. Hence hα, βi = hα, 0i = κ1 ◦ α as desired. Finally, to show m1 ≡ [κ1, κ2, κ2], m2 ≡ [κ2, κ1, κ2] : I + I + I → I + I are jointly monic, let f1 ≡ ha1, b1, c1i, f2 ≡ ha2, b2, c2i : X → I + I + I be any total maps with m1 ◦ f1 = m1 ◦ f2 and m2 ◦ f1 = m2 ◦ f2. Then XVI a1 = [id, 0, 0]◦ f1 = ⊲1 ◦ m1 ◦ f1 = ⊲1 ◦ m1 ◦ f2 = a2. and similarly from the equality involving m2, we get b1 = b2. As f1 is total, we have 1 = 1 ◦ f1 = a1 > b1 > c1, so c1 = (a1 > b1)⊥. With the same reasoning c2 = (a2 > b2)⊥. Thus c1 = c2 and so f1 = f2, as desired. (cid:3) ..181 101 ' ' ) ) ( / / / & & ) ) ( / / / 3.2.2 From total to partial 182 Let C be an effectus in total form. In this section we will show that Par C is an effectus in partial form. Before we get to work, it is helpful to discuss the axioms of an effectus in total form, now we have some experience with an effectus in partial form. 1. In 181 VII we saw that the coproduct in an effectus (in partial form) is almost a biproduct. This structure is hidden (for the most part) in the left pullback square of (3.1), which allows the formation of hα, βi given partial maps α, β. 2. The right pullback square of (3.1) is used to extract a total map in C from a partial map f in Par C that is total (i.e. 1 ◦ f = 1). See 186 VIII. 3. The joint-monicity of [κ1, κ2, κ2] and [κ2, κ1, κ2] will imply the joint monic- ity of ⊲1 and ⊲2, which is required for uniqueness of the partial sum of partial maps. See 186 X. To prove the theorem, we need to study pullbacks: first in any category, then in an effectus C in total form and finally in Par C. Our proof is different from the original proof of Cho by using several more general facts about pullbacks, which we will prove next. 183 We start with two classic facts about pullbacks. II Exercise Show that if we have any pullback square -- say P m2 A ❴ m1 g B f / X , then m1 and m2 are jointly monic. III Exercise Prove the pullback lemma -- that is: if we have a commuting diagram A k X f f ′ B l / Y g g′ C m , / Z then we have the following two implications. 1. If the left and right inner squares are pullbacks, then so is the outer square. ✤ / / / / / / / / / 2. If the outer square is a pullback and l and g are jointly monic, then the left inner square is a pullback. Remark In the literature (e.g. [87, III.5 exc. 8]), one often finds a weaker second implication, which assumes that the right square is a pullback. IV It is well-known that monos are stable under pullbacks -- that is: 184 if n P ❴ g / m / A / X f / B , then / n P ❴ g / m / A / X f / B , where we used / for jointly monic maps. to denote a monic map. We will need an analogous result Lemma If the pairs (m1, m2), (n1, g1), (n2, g2) and (h1, h2) in the following com- muting diagram are jointly monic (for instance: if they span pullback squares), then (n1 ◦ h1, n2 ◦ h2) is jointly monic as well. II n1 m1 X1 f1 B1 P h2 P2 n2 X2 h1 g2 f2 P1 g1 A m2 / B2 Proof To show joint monicity of n1 ◦ h1 and n2 ◦ h2, assume α1, α2 : Z → P are maps with n1 ◦ h1 ◦ α1 = n1 ◦ h1 ◦ α2 and n2 ◦ h2 ◦ α1 = n2 ◦ h2 ◦ α2. To start m1 ◦ g1 ◦ h1 ◦ α2 = f1 ◦ n1 ◦ h1 ◦ α2 = f1 ◦ n1 ◦ h1 ◦ α1 = m1 ◦ g1 ◦ h1 ◦ α1. Reasoning on the other side of the diagram, we find III m2 ◦ g1 ◦ h1 ◦ α2 = m2 ◦ g2 ◦ h2 ◦ α2 = f2 ◦ n2 ◦ h2 ◦ α2 = f2 ◦ n2 ◦ h2 ◦ α1 = m2 ◦ g2 ◦ h2 ◦ α1 = m2 ◦ g1 ◦ h1 ◦ α1. So by joint monicity of m1 and m2, we conclude g1 ◦ h1 ◦ α2 = g1 ◦ h1 ◦ α1. So by the joint monicity of n1 and g1 (and by assumption n1 ◦ h1 ◦ α2 = n1 ◦ h1 ◦ α1), 182 -- 184.. 103 ✤ / / ✤ / / / / / / / / / / / / / / / we conclude h1 ◦ α1 = h1 ◦ α2. Reasoning in the same way mirrored over the diagonal, we find h2 ◦ α1 = h2 ◦ α2. Thus, by joint monicity of h1 and h2, we conclude that α1 = α2, as desired. (cid:3) 185 Proposition In an effectus in total form, any square of one of the two following forms is a pullback. X + A ❴ g+id id+f X + B g+id f X ❴ κ1 / Y κ1 Y + A / Y + B id+f X + A / Y + B f +g II Proof To start with the left square, consider the following commuting diagram. X + A id+f X + B id+! / X + 1 g+id (1) g+id (2) g+id Y + A !+id 1 + A id+f (3) id+f Y + B id+! Y + 1 !+id (4) !+id / 1 + B / 1 + 1 id+! We want to show (1) is a pullback. The inner square (4) and right rectangle (2,4) are pullbacks by axiom. Thus the inner square (2) is also a pullback by the pullback lemma, see 183 III. With the same reasoning, we see square (3) is a pullback. Thus by the pullback lemma, it is sufficient to show that left rectangle (1,3) is a pullback. The left rectangle is indeed a pullback as both the outer square (1,2,3,4) and the right rectangle (2,4) are pullbacks. For the right square, we consider the following diagram. f X κ1 ! Y κ1 1 κ1 X + A / Y + B / 1 + 1 !+! f +g The inner right and outer square are pullbacks by axiom. So the inner left square is also a pullback by the pullback lemma, as desired. (cid:3) 186 Definition Assume C is an effectus in total form. For a map f : X → Y , write f = κ1 ◦ f : X → Y + 1. (This is the Kleisli embedding C → Par C.) [19] ✤ / / / ✤ / / / / / / / / / / / / / / / / / (Recall that we use f : X ⇀ Y to denote a map f : X → Y + 1, see 180 I.) Exercise Assume C is an effectus in total form. Show that if Ia II is a coproduct in C, then κ1 : X → X + Y ← Y : κ2 κ1 : X ⇀ X + Y ↼ Y : κ2 is a coproduct in Par C. Show 0 is also the initial object of Par C. Beware While cotupling in C and Par C coincide: [f, g]C = [f, g]Par C for any (properly typed) partial maps f, g and \f +C g = f +Par C g, however, in general III [κ1 ◦ f, κ2 ◦ g] = f +C g 6= f +Par C g = [[κ1, κ3]◦ f, [κ2, κ3]◦ g]. Proposition Let C be an effectus in total form. Squares of the following form are pullbacks in Par C. IV X + A ❴ g+id id+ f X + B g+id A + X ❴ ⊲1 f +id B + X Y + A / Y + B id+ f A f ⊲1 / B Proof (A different proof for the pullback on the right is given in [19, lem. 4.1(5)].) It is sufficient to show that the following squares are pullbacks in C. V X + (A + 1) X + (B + 1) id+(f +id) A + (X + 1) f +id B + (X + 1) g+id g+id id+! id+! Y + (A + 1) / Y + (B + 1) id+(f +id) A + 1 / B + 1 f +id These are indeed pullbacks by 185 I. (cid:3) Definition Assume C is an effectus in total form. For arbitrary objects X, Y in C, write 0 ≡ κ2 ◦! : X → Y + 1 and 1 ≡ κ1 ◦! ≡ ! : X → 1 + 1. [19] Exercise Show 0 is a zero object in Par C with unique zero map as in VI. Proposition Assume C is an effectus in total form and f is a map in Par C. VI VII VIII 1. If 1 ◦ f = 1, then f = g for a unique g in C. 2. If 1 ◦ f = 0, then f = 0. ..184 -- 186.. 105 ✤ / ✤ / / / / / / / IX Proof (This proof is essentially the same as [19, lem. 4.7(1) & prop. 4.6].) Both points follow from the right pullback square of (3.1) as follows. X g ! f Y ❴ κ1 Y + 1 1 κ1 / 1 + 1 ! !+! X ! f ! 1 ❴ κ2 Y + 1 1 κ2 / 1 + 1 ! !+! If 1 ◦ f = 1, then (!+!)◦ f = κ1 ◦! and so there is a unique g with f = κ1 ◦ g as shown above on the left. For the other point, if 1 ◦ f = 0, then (!+!)◦ f = κ2 ◦! and so f = κ2 ◦! as shown above on the right. (cid:3) X Proposition Assume C is an effectus in total form. The partial projectors ⊲1 ≡ [id, 0] and ⊲2 ≡ [0, id] are jointly monic in Par C. [19, lem. 4.1(4)] XI Proof (This is a new proof.) Consider the following diagram in Par C. id+! ⊲1 ⊲1 X + 1 ❴ !+id 1 + 1 ⊲1 id+! ⊲2 / 1 ! X + Y ❴ !+id 1 + Y ❴ ⊲2 Y X ! 1 ⊲2 This diagram commutes and each of the inner squares is a pullback by IV. The maps ⊲1, ⊲2 : 1 + 1 ⇀ 1 are jointly monic, which is a reformulation of the joint monicity axiom of an effectus in total form. Thus by 184 II the outer ⊲1 and ⊲2 are jointly monic. (cid:3) 187 Theorem If C is an effectus in total form, then Par C is an effectus in partial II III In 186 II, we saw that Par C has finite coproducts. form. [19, thm. 4.2] Proof (PCM-enrichment, I) Assume f, g : X ⇀ Y . We define that f ⊥ g iff there is a bound b : X ⇀ Y + Y with ⊲1 ◦ b = f and ⊲2 ◦ b = g. (By 186 X this b is unique if it exists.) In this case, define f >g = ∇ ◦ b, where ∇ ≡ [id, id]. We will show > with 0 as defined in 186 VI forms a PCM. Partial commutativity is obvious. The map κ1 ◦ f : X ⇀ Y +Y is a bound for f ⊥ 0 and so f >0 = ∇ ◦ κ1 ◦ f = f , which & & # # & & ✤ / / / & & # # ' ' ✤ / / / ✤ & & & & / ✤ / ✤ / / shows 0 is indeed a zero for >. Only partial associativity remains. Assume f ⊥ g via bound b and f > g ⊥ h via bound c. Note ∇ ◦ b = f > g = ⊲1 ◦ c and thus by the right pullback square of 186 IV (see diagram below), there is a unique d : X ⇀ (Y + Y ) + Y with ⊲1 ◦ d = b and (∇ + id) ◦ d = c. X c d b (Y + Y ) + Y ❴ ⊲1 Y + Y ∇+id ∇ Y + Y ⊲1 / Y The map (⊲2 + id) ◦ d is a bound for g ⊥ h, indeed: ⊲1 ◦(⊲2 + id) ◦ d = ⊲2 ◦ ⊲1 ◦ d ⊲2 ◦(⊲2 + id) ◦ d = ⊲2 ◦ d = ⊲2 ◦ b = g = ⊲2 ◦(∇ + id) ◦ d = ⊲2 ◦ c = h. We need one more: [id, κ2] ◦ d is a bound for f ⊥ g > h, indeed ⊲1 ◦[id, κ2] ◦ d = ⊲1 ◦ ⊲1 ◦ d = ⊲1 ◦ b = f. ⊲2 ◦[id, κ2] ◦ d = ∇ ◦(⊲2 + id) ◦ d = g > h Finally, we compute f > (g > h) = ∇ ◦[id, κ2] ◦ d = ∇ ◦(∇ + id) ◦ d = ∇ ◦ c = (f > g) > h, which shows the partial associativity. Homsets of Par C are indeed PCMs. (PCM-enrichment, II) Assume f ⊥ g with bound b. It is easy to see b ◦ k is a bound for f ◦ k ⊥ g ◦ k and consequently (f ◦ k)>(g ◦ k) = ∇ ◦ b ◦ k = (f >g) ◦ k. For the other side: (h + h) ◦ b is a bound for h ◦ f ⊥ h ◦ g, indeed ⊲1 ◦(h + h) ◦ b = h ◦ ⊲1 ◦ b = h ◦ f ⊲2 ◦(h + h) ◦ b = h ◦ ⊲2 ◦ b = h ◦ g. And so (h ◦ f ) > (h ◦ g) = ∇ ◦(h + h) ◦ b = h ◦∇ ◦ b = h ◦(f > g). Finally, the zero axiom holds by 186 VII. (FinPAC) We just saw Par C is PCM-enriched. We already know Par C has coproducts. The compatible sum axiom holds by definition of ⊥. To show Par C is a finPAC, only the untying axiom remains to be proven. If f ⊥ g with bound b, then (κ1 + κ2) ◦ b is a bound for κ1 ◦ f ⊥ κ2 ◦ g, which proves the untying axiom. ..186 -- 187.. 107 IV V ( ' ' ( ( ✤ / VI (Effect algebra of predicates) Pick any predicate p : X → 1 + 1. We define p⊥ = [κ2, κ1]◦ p. (p⊥ is p with swapped outcomes.) We compute ⊲1 ◦ p = [id, κ2]◦ κ1 ◦ p = p ⊲2 ◦ p = [[κ2, κ1], κ2]◦ κ1 ◦ p = [κ2, κ1]◦ p ∇ ◦ p = [[κ1, κ1], κ2]◦ κ1 ◦ p = κ1 ◦! = 1. So p is a bound for p ⊥ p⊥ and p > p⊥ = 1. To show p⊥ is the unique orthocomplement, assume p > q = 1 for some q via a bound b : X ⇀ 1 + 1. Note 1 ◦ b = ∇ ◦ b = 1 and so by 186 VIII we know b = c for some c : X → 1 + 1. And so p = ⊲1 ◦ b = [id, κ2]◦ κ1 ◦ c = c. Hence q = ⊲2 ◦ b = [[κ2, κ1], κ2]◦ κ1 ◦ c = [κ2, κ1]◦ p = p⊥. To show Pred X is an effect algebra, it only remains to be proven that the zero -- one axiom holds. So assume 1 ⊥ p via some bound b. We must show p = 0. By the following instance of the right pullback square of (3.1), we see b = κ1 ◦ κ1 ◦!. X ! b ! 1 ❴ κ1 ◦ κ1 (1 + 1) + 1 [id,κ2] 1 κ1 / 1 + 1 VII Hence p = ⊲2 ◦ b = [[κ2, κ1], κ1]◦ κ1 ◦ κ1 ◦! = κ2 ◦! = 0. In 186 VIII we proved that f = 0 whenever 1 ◦ f = 0. It only (Final axioms) remains to be shown that f ⊥ g provided 1 ◦ f ⊥ 1 ◦ g. Let b be a bound for 1 ◦ f ⊥ 1 ◦ g. We apply the right pullback square of 186 IV twice in succession as follows (using ⊲2 ◦ c = ⊲2 ◦(! + id) ◦ c = ⊲2 ◦ b = 1 ◦ g for the right one.) X c f b X c Y + 1 ❴ ⊲1 Y !+id ! 1 + 1 ⊲1 / 1 d g Y + Y ❴ ⊲2 Y id+! ! Y + 1 ⊲2 / 1 Clearly ⊲2 ◦ d = g and ⊲1 ◦ d = ⊲1 ◦(id + !) ◦ d = ⊲1 ◦ c = f , so d is a bound for f ⊥ g, as desired. (cid:3) 188 We are ready to show the equivalence of effectuses in partial and total form. ( ( % % & & ✤ / & $ $ ) ) ✤ / & $ $ ) ) ✤ / Proof of 180 X Let C and D be effectuses in total and respectively partial form. The first point, that Par C is an effectus in partial form, is shown in 187 I. The second point, that Tot D is an effectus in total form, is shown in 181 XI. It remains to be shown that nothing is lost. (Par Tot D ∼= D) Let D be an effectus in partial form. A map f : X → Y in Par Tot D is by definition a map f : X → Y + 1 in D with 1 ◦ f = 1. Let P : Par Tot D → D denote the identity-on-objects map f 7→ ⊲1 ◦ f . This is a functor: clearly P id = P κ1 = id and if g : Y → Z in Par Tot D, then ⊲1 ◦(g ◦ f ) = ⊲1 ◦[g, κ2]◦ f = [⊲1 ◦ g, ⊲1 ◦ κ2]◦ f = [⊲1 ◦ g, 0]◦h⊲1 ◦ f, ⊲2 ◦ fi = ⊲1 ◦ g ◦ ⊲1 ◦ f. The functor P is an isomorphism with inverse P ′f ≡ hf, (1 ◦ f )⊥i. Indeed, we have P P ′f = ⊲1 ◦hf, (1 ◦ f )⊥i = f and P ′P f = h⊲1 ◦ f, (1 ◦ ⊲1 ◦ f )⊥i = (κ1 ◦ ⊲1 ◦ f ) > (κ2 ◦ ⊲1 ◦ f )⊥ = f. (Tot Par C ∼= C) Assume C is an effectus in total form. Recall 1 ◦ g = 1 and df ◦ g = f ◦ g, so Q : C → Tot Par C given by Qg = g is an identity-on- objects functor. It is an isomorphism by the first part of 186 VIII: for every f in Tot Par C, there is a unique g in C with f = g. (cid:3) II III IV Exercise In this exercise we will distinguish effectuses in total and partial form. 189 1. To start, show that the initial object of an effectus in total form is strict -- that is: show that any map into 0 is an isomorphism. (Hint: use the right pullback square of 185 I with X = Y = 0.) 2. Conclude that if C is both an effectus in total and partial form, then every object in C is isomorphic to 0. Convention For the remainder of this text, we will work with effectuses in partial form and write 1 instead of I. To spare ink, a phrase like "let C be an effectus" should be read as "let C be an effectus in partial form" and simply "C is an effectus" means either "C is an effectus in total form" or "C is an effectus in partial form". (In non-trivial cases, this is unambiguous, as seen in I.) As before, when we took the effectus in partial form as base category, we will denote partial maps by f : X → Y (instead of f : X ⇀ Y ) and their composition as f ◦ g (instead of f ◦ g). II ..187 -- 189 109 189a Examples Our main example of an effectus (in total form) is the category vNop of von Neumann algebras with ncpu-maps between them in the opposite direction, see 180 V. Its full subcategory CvNop of commutative von Neumann algebras is an effectus as well. In [23] several other examples of effectuses are discussed in detail: II 1. The category OUSop of order unit spaces (i.e. ordered real vector spaces with distinguished order unit) with positive unit-preserving linear maps in the opposite direction is an effectus (in total form). 2. A related example is the category OUGop of order unit groups (i.e. ordered abelian groups with distinguished order unit) with positive unit-preserving homomorphisms in the opposite direction, which is also an effectus. 3. Every extensive category [15] with final object is an effectus in total form. Examples of extensive categories with final object include (a) Set, the category of sets with maps between them; (b) CRngop, the category of commutative unit rings with unit-preserving homomorphisms between them in the opposite direction; (c) CH, the category of compact Hausdorff spaces with continuous maps between them; and III Together with van de Wetering, we recently investigated [128] the category EJAop of Euclidean Jordan algebras with positive unit-preserving linear maps in the opposite direction, which is also an effectus (in total form). 3.2.3 Predicates, states and scalars 190 We turn to the internal structures of an effectus. [23] II Definition Let C be an effectus (in partial form). 1. As alluded to in the definition of an effectus in partial form, a predicate on X is a map X → 1. The set of predicates Pred X on X is (by definition of an effectus in partial form) an effect algebra. 2. A scalar is a predicate on 1; that is, a map 1 → 1. We write Scal C ≡ M ≡ Pred 1 for the set of scalars. We multiply two scalars λ, µ : 1 → 1 simply by composing λ ⊙ µ ≡ λ◦ µ. As 1M = idM (see 181 XIII) and C is PCM-enriched, the set of scalars M with ⊙ is an effect monoid, see 178 II. 3. A real effectus is an effectus where the set of scalars M is isomorphic (as effect monoid) to [0, 1]. 4. For a scalar λ : 1 → 1 and a predicate p : X → 1, we write λ · p ≡ λ◦ p. Again due to PCM-enrichment and 1M = idM , this scalar multiplication turns Pred X into an M -effect module, see 179 II. 5. For f : X → Y in C, define Pred(f ) : Pred Y → Pred X by Pred(f )(p) = p◦ f . It is easy to see Pred(f ) is an M -effect module homomorphism if f is total and that, in fact, Pred: Tot C → EModop M is a functor. 6. A substate on X is a map ω : 1 → X. A state is a total substate. We denote the set of states on X by Stat X. 7. An effectus C has separating predicates if Pred X (as a set of maps X → 1) is jointly monic for every X in C. Similarly, an effectus has separating states if Stat X (as a set of maps 1 → X) is jointly epic for every X in C. Examples The category vNop (see 180 V) is a real effectus (M ∼= [0, 1]) with separating states and predicates. [23] The predicates on a von Neumann alge- bra A correspond to the set of effects [0, 1]A and Stat A can be identified with the convex set of normal states. We briefly cover the examples from [23] mentioned in 189a I. III IV 1. The category OUSop (see 189a I) is a real effectus with separating predi- cates, but without seperating states. (The states on a single order unit space are separating if and only if the order unit space is archimedean.) A predicate on an order unit space X corresponds to a point x ∈ X with 0 6 x 6 1, where 1 is the distinguished order unit. The states are exactly what are called states for order unit spaces in the literature. 2. The category OUGop (see 189a I) has the two-element effect monoid 2 as scalars and seperating predicates. The predicates of an order unit group G correspond to the elements x ∈ G with 0 6 x 6 1, where 1 is the distin- guished order unit. States on G correspond to unit-preserving positive homomorphisms G → Z, which are not separating. 3. Any extensive category with final object (such as Set, CRngop and CH) has as scalars the two-element effect monoid 2. (a) In Set the predicates on a set X correspond to subsets U ⊆ X and are thus separating. States on X correspond t elements x ∈ X, which are also separating. (b) In CRngop the predicates on R correspond to idempotents of R, which are not separating. States correspond to unit-preserving homomor- phisms R → Z, which are not separating either. 189a -- 190.. 111 (c) In CH the predicates on X correspond to clopen subsets U ⊆ X, which are not separating. States correspond on X correspond to its points x ∈ X, which are separating. V The category EJAop is a real effectus with separating states and predicates. The predicates on an Euclidean Jordan algebra (EJA) E correspond to its effects; i.e. the elements 0 6 a 6 1. The states are exactly what are usually considered states for EJAs. 191 Does every effect monoid M occur as the effect monoid of scalars of some effec- tus? Does every M -effect module occur as effect module of predicates? We will show they do, in the following strong sense. II Theorem Let M be any effect monoid. The category EModop III IV V M is an effectus, we have to show the dual axioms for EModM . M is an effectus in total form with scalars M and separating predicates. [23] In fact: every effectus C in total form with scalars M and separating predicates is equivalent to a subcategory of EModop M . Proof To show EModop (Finite products) For every M -effect module E, there is a unique M -effect module map ! : M → E, which is given by λ 7→ λ · 1. So M is the initial object of EModM . The one-element M -effect module {0 = 1} is the final object. If E and F are M -effect modules, then the set of pairs E × F with compo- nentwise operations is again an M -effect module. The effect module E × F with π2−→ F is the categorical product of E and F obvious projections E in EModM (as is the case with effect algebras, see 175 III). (Pushout diagrams) To show the pushout diagrams corresponding to (3.1) hold, assume we are given M -effect modules E, F, G with module maps α, β, δ that make the outer squares of the following diagrams commute. π1←− E × F G α G ! β E × F !×id M × F id×! id×! E × M !×id M × M δ E π1 ! M π1 E × F M × M !×! We have to show that there are unique dashed arrows (as shown) that make these diagrams commute. We start with the left diagram. By assumption α◦(! × id) = β ◦(id×!), so in particular α(1, 0) = β(1, 0). For any (x, y) ∈ E × F , we have α(x, 0) 6 α(1, 0) = β(1, 0) ⊥ β(0, 1) > β(0, y), so α(x, 0) ⊥ β(0, y). Define f : E × F → G by f (x, y) = α(x, 0) > β(0, y). It is easy to see f is (partially) additive and f (1, 1) = α(1, 0) > β(0, 1) = β(1, 1) = 1, so f is a c c o o o o O O Q Q o o O O c c o o o o O O Q Q o o O O module map. We compute f (x, λ · 1) = α(x, 0) > β(0, λ · 1) = α(x, 0) > λ · α(0, 1) = α(x, λ) and so f ◦(id×!) = α. Similarly β = f ◦(! × id). It is easy to see f is the unique map with f ◦(id×!) = α and β = f ◦(! × id) and so the left square of (3.1) is a pullback in EModop M . Concerning the right diagram: as δ ◦(!×!) =!◦ π1, we have δ(λ·1, µ·1) = λ·1 and so δ(0, y) 6 δ(0, 1) = 0. Hence δ(x, y) = δ(x, 0). Define g : E → G by g(x) = δ(x, 0). Clearly g is additive, g(1) = δ(1, 0) = δ(1, 1) = 1 and δ = g ◦ π1 by definition. Obviously g is the unique such map. Thus the right square of (3.1) is a pullback in EModop M . (Joint epicity) To show EModop M is an effectus, it only remains to be shown that hπ1, π2, π2i,hπ2, π1, π2i : M × M → M × M × M are jointly epic. So as- sume f, g : M×M×M → E are two M -effect module maps with f ◦hπ1, π2, π2i = g ◦hπ1, π2, π2i and f ◦hπ2, π1, π2i = g ◦hπ2, π1, π2i. From the first equality, it follows that f (1, 0, 0) = g(1, 0, 0). The other one implies f (0, 1, 0) = g(0, 1, 0). Thus f (0, 0, 1) = f (1, 1, 0)⊥ = g(1, 1, 0)⊥ = g(0, 0, 1). As for (λ1, λ2, λ3) ∈ M 3, we have (λ1, λ2, λ3) = λ1 · (1, 0, 0) > λ2 · (0, 1, 0) > λ3 · (0, 0, 1) we get f = g. (Representation) Let C be an effectus in total form. Clearly Pred f = Pred g (for f, g : X → Y in C) if and only if p◦ f = (Pred f )(p) = (Pred g)(p) = p◦ g for every p ∈ Pred X. Thus the functor Pred : C → EModop M is faithful if and only if C has separating predicates. So if C has separating predicates, C is equivalent to the subcategory Pred C of EModop M . (cid:3) VI VII Exercise Show that the category Rngop of unit rings with unit-preserving ho- momorphisms in the opposite direction is an effectus in total form. VIII 1. Show that the predicates on a ring R correspond to its idempotents; that p ⊥ q iff pq = qp = 0, p⊥ = 1 − p and p > q = p + q. (So 2 is its effect monoid of scalars.) Conclude Rngop does not have separating predicates. 2. Show that there is no unit-preserving ring-homomorphism Z2 → Z. Con- clude Rngop does not have separating states. We studied the structure of predicates in an effectus. What about the states? They will turn out to form an abstract M -convex set. Before we define these, we introduce a generalized distribution monad. Definition Let M be an effect monoid and X be any set. A formal M - convex combination over X is a function p : X → M with finite support such that >x∈X p(x) = 1. We use λ1 x1i > ··· > λn xni as a shorthand for the for- mal M -convex combination p : X → M given by p(x) = >xi=x λi. (So p(xi) = λi if the xi are distinct.) We write DM X for the set of all formal M -convex combinations over X. [56] 192 II ..190 -- 192.. 113 III Exercise* In this exercise you study DM as a monad. See [56, 129]. 1. For f : X → Y , define DM f : DM X → DM Y by (DM f )(p)(y) = > x;f (x)=y p(x). (That is: (DM f )(λ1 x1i > ··· > λn xni) = λ1 f (x1)i > ··· > λn f (xn)i.) Show that this turns DM into a functor Set → Set. 2. Define η : X → DM X by η(x) = 1 xi and µ : DMDM X → DM X by µ(Φ)(x) = > ϕ Φ(ϕ) ⊙ ϕ(x). (That is: µ(>i σi >j λij xijii) = >i,j(σi ⊙ λij )xiji.) Show that (DM , η, µ) is a monad. 3. Show that Kℓ DM , the Kleisli category of DM (see e.g. [129, §2.6]), is an effectus with M as scalars. IV Definition Let X be a set together with a map h : DM X → X. We say (X, h) is an abstract M -convex set provided h(xi) = x and h(cid:16)> i,j σi ⊙ λij xiji(cid:17) = h(cid:16)> i σi h(cid:0) > j λij xiji(cid:1)i(cid:17) (3.2) for every >i σi >j λij xijii in DMDM X. (Equivalently: h◦ µ = h◦DM h and h◦ η = id.) We call h the convex structure on X. A map f : X → Y between abstract M -convex sets is M -affine if f(cid:0)hX(cid:0)> i λi xii(cid:1)(cid:1) = hY(cid:0) > i λi f (xi)i(cid:1) for every >i λi xii in DM X. (Equivalently: hX ◦DM f = f ◦ hY .) Write AConvM for the category of abstract M -convex sets with M -affine maps between them. (Equivalently: AConvM is the Eilenberg -- Moore category of the monad DM .) We call an abstract M -convex set (X, h) cancellative provided h(λyi > λ⊥ x1i) = h(λyi > λ⊥ x2i) implies x1 = x2 for any x1, x2, y ∈ X and λ ∈ M, λ 6= 1. IVa Remark For an abstract [0, 1]-convex set (X, h), binary convex combinations are sufficient: for instance, for any λ > µ 6= 1 and x, y, z ∈ X we have h(cid:0)λxi > µyi > (λ > µ)⊥ zi(cid:1) = h(cid:16)λxi > λ⊥ h(cid:0)µ/λ⊥ yi > (λ>µ)⊥/λ⊥ zii(cid:1)(cid:17). This reduction to binary convex combinations uses the division on [0, 1]. Some effect monoids have a suitable division as well, but many do not as we will see later on in 195 II and 195 VI. It seems plausible that it is impossible to write an arbitrary three-element M -convex combination over DM{x, y, z} using only binary convex combinations in the case that M is the effect monoid on the unit interval of C[0, 1]. I do not have a proof to back this claim. Examples We give some examples of abstract convex sets. V 1. Every convex subset X ⊆ V of some real vector space V is a cancellative abstract [0, 1]-convex set with h(λ1 x1i>···>λn xni) = λ1x1 +···+λnxn. 2. Not every abstract [0, 1]-convex set is a convex subset of a real vector space. Consider T ⊆ [0, 1]2 defined by T ≡ {(x, y); x + y 6 1}. Say (x, y) ∼ (x′, y′) whenever x = x′ and either x 6= 0 or y = y′. This equivalence relation ∼ is a congruence (we will cover these in 193 II) for the convex structure on T (inherited by R2) and so T /∼ is an abstract [0, 1]-convex set. That is: we start off with a filled triangle T ≡ , identify all vertical lines except for the y-axis and are left with T /∼ ≡ . The abstract [0, 1]- convex set T /∼ is not cancellative: h(cid:16) 1 2 i > 1 2 i(cid:17) = but 6= . = h(cid:16) 1 2 i > 1 2 i(cid:17), 3. Semilattices are exactly abstract 2-convex sets. (2 is the two-element Effect Monoid, see 178 III.) Furthermore, every semilattice (L,∨) is an abstract [0, 1]-convex set with h(cid:0)λ1 x1i > ··· > λn xni(cid:1) = _i;λi6=0 xi. A semilattice L is cancellative as abstract [0, 1]-convex set if and only if x = y for all x, y ∈ L. See [90]. 4. Every cancellative [0, 1]-convex set is isomorphic to a convex subset of a real vector space. See e.g. [69, thm. 8]. Remarks The study of convex subsets (of real vector spaces) as algebraic struc- tures goes back a long time, see e.g. [30, 33, 42, 90, 107]. The more pathological abstract [0, 1]-convex sets have been studied before under different names: they are semiconvex sets in [30, 109] and convex spaces in [33]. The description of convex sets using Eilenberg -- Moore algebras is probably due to ´Swirszcz [109]. For the even more exotic abstract M -convex sets, see [23] or [56]. VI Proposition Let C be an effectus with scalars M . For any object X, the set of states Stat X is an abstract M op-convex set with h : DM op Stat X → Stat X VII ..192.. 115 defined by h(cid:0)λ1 ϕ1i > ··· > λn ϕni(cid:1) = [ϕ1, . . . , ϕn]◦hλ1, . . . , λni. Furthermore, for any total map f : X → Y , the map (Stat f )(ϕ) = f ◦ ϕ is an M op-affine map Stat X → Stat Y yielding a functor Stat : Tot C → AConvM op . [23] VIII Proof Clearly h(ϕi) = ϕ. To show the other axiom (3.2), assume we have any >n j=1 λij ϕijii in DM opDM op Stat X. Write λi ≡ hλi1, . . . , λimii and ϕi ≡ [ϕi1, . . . , ϕimi ]. Then λij ϕiji(cid:1)i(cid:17) = [ϕ1 ◦ λ1, . . . , ϕn ◦ λn]◦hσ1, . . . , σni i=1 σi >mi h(cid:16)> σi h(cid:0) > j i = > i = > = > i,j = > i,j ϕi ◦ λi ◦ σi i (cid:0)> j ϕij ◦ λij(cid:1)◦ σi ϕij ◦(λij ◦ σi) ϕij ◦(σi ⊙M op λij ) = h(cid:16)> i,j σi ⊙M op λij ϕiji(cid:17). So Stat X is indeed an abstract M op-convex set. Let f : X → Y be any total map. To show Stat f is affine, it is sufficient to show that for every >i λi ϕii in DM op , we have f ◦ >i ϕi ◦ λi = >i f ◦ ϕi ◦ λi, which is clearly true. Functo- riality of Stat is obvious. (cid:3) 3.2.4 Effectus of abstract M-convex sets 193 II For our next project, we indulge ourselves in a tangent: we investigate whether the category AConvM is an effectus. We can show it is, if M has a certain partial division. Unfortunately it will remain unclear if AConvM is an effectus in general. The case M = [0, 1] is much easier and has already been dealt with in [69]. The first step is to study the coproduct in AConvM . To construct the coproduct, we will need to know about quotients of abstract M -convex sets. Exercise In this exercise you construct the quotient of an abstract M -convex set by a congruence. Assume (X, h) is an abstract M -convex set and ∼ ⊆ X 2 is an equivalence relation. Write X/∼ for the set of equivalence classes of ∼ and q : X → X/∼ for q(x) = [x]∼, where [x]∼ is the equivalence class of x. For ϕ, ψ ∈ DM X, we write ϕ ∼ ψ provided (DM q)(ϕ) = (DM q)(ψ). We say ∼ is a congruence for (X, h) if ϕ ∼ ψ implies h(ϕ) ∼ h(ψ). 1. Prove that q, DM q and DMDM q are all surjective. 2. Show that the following are equivalent. (a) ∼ is a congruence. (b) There is a map h∼ : DM X/∼ → X/∼ such that h∼ ◦DM q = q ◦ h. By the previous point h∼ is unique, if it exists. 3. Assume ∼ is a congruence. Use the previous two points to show h∼ ◦DM h∼ ◦DMDM q = h∼ ◦ µ◦DMDM q. Conclude that (X/∼, h∼) is an abstract M -convex set and that the qou- tient map q : X → X/∼ is M -affine. Exercise Assume f : X → Y is an affine map between abstract M -convex sets. Show that the kernel {(x, y); f (x) = f (y)} of f is a congruence. Exercise Assume (X, h) is an abstract M -convex set and R ⊆ X 2 is some relation. It is easy to see there is a least congruence containing R. We need to know a bit more than mere existence. Write R∗ for the reflexive symmetric transitive closure of R (i.e. the least equivalence relation containing R). For ψ1, ψ2 ∈ DM X, we write ψ1 ≈ ψ2 if there is a derivation: a tu- j=1 λij xiji) such that ϕ1 = ψ1, ϕn = ψ2 ple ϕ1, . . . , ϕn ∈ DM X (say ϕi ≡ >ni and for each 1 6 i < n one of the following two conditions holds. III IV 1. ni = ni+1 and for all 1 6 j 6 ni we have xij R∗ x(i+1)j and λij = λ(i+1)j . 2. h(ϕi) = h(ϕi+1). We will show that ∼ given by x ∼ y ⇐⇒ η(x) ≈ η(y) is the least congruence containing R: 1. To start, show that η(h(ψ)) ≈ ψ and so ϕ ≈ ψ implies h(ϕ) ∼ h(ψ). 2. Show that if ϕ ≈ ψ, then µ(cid:16)λ0 ψi > n > j=1 λj χji(cid:17) ≈ µ(cid:16)λ0 ϕi > n > j=1 λj χji(cid:17). for any >j λj = 1 in M and χ1, . . . , χn ∈ DM X. 3. Use the previous point to show that ϕ ∼ ψ implies ϕ ≈ ψ. Conclude ∼ is the smallest congruence containing R. ..192 -- 193.. 117 V Proposition Assume (X, hX ) and (Y, hY ) are abstract M -convex sets. We will construct their coproduct. Note that µ turns DM (X + Y ) into an abstract M - convex set. Let ∼ denote the least congruence on DM (X + Y ) with (DM κ1)(ϕ) ∼ η(κ1(hX (ϕ))) and (DM κ2)(ψ) ∼ η(κ2(hY (ψ))) (3.3) for all ϕ ∈ DM X and ψ ∈ DM Y . (Alternatively: ∼ is the least congruence that makes η ◦ κ1 and η ◦ κ2 affine.) Write q : DM (X + Y ) → C for the quo- tient of DM (X + Y ) by ∼, see II. Denote the convex structure on C by h. Define c1 : X → C and c2 : Y → C by c1 ≡ q ◦ η ◦ κ1 and c2 ≡ q ◦ η ◦ κ2. Then: c1 : X → C ← Y : c2 is a coproduct of X and Y in AConvM . VI Proof We start with the affinity of the coprojections: c1 ◦ hX = q ◦ η ◦ κ1 ◦ hX = q ◦ DM κ1 = q ◦ µ◦DM η ◦DM κ1. = µ◦DM q ◦ DM η ◦DM κ1. = µ◦DM c1, by dfn. of ∼ so c1 is affine. With similar reasoning, we see c2 is affine. VII Assume f : X → Z and g : Y → Z are two affine maps to some abstract M - convex set (Z, hZ ). We want to show hZ ◦DM [f, g] lifts to C. With an easy calculation, one sees hZ ◦DM[f, g] is affine. We compute hZ ◦DM [f, g]◦DM κ1 = hZ ◦DM f = f ◦ hX = hZ ◦ η ◦[f, g]◦ κ1 ◦ hX = hZ ◦DM [f, g]◦ η ◦ κ1 ◦ hX . So ( (DM κ1)(ϕ), η(κ1(hX (ϕ))) ) is in the kernel (see III) of hZ ◦DM [f, g] for ϕ ∈ DM . Similarly ( (DM κ2)(ψ), η(κ2(hY (ψ))) ) is also in the kernel of hZ ◦ DM [f, g] for all ψ ∈ DM . As kernels of affine maps are congruences and ∼ is the least congruence that contains these pairs, we conclude ∼ is a subset of the kernel of hZ ◦ DM [f, g]. Thus there is a unique affine k : C → Z fixed by k ◦ q = hZ ◦DM [f, g]. Now k ◦ c1 = k ◦ q ◦ η ◦ κ1 = hZ ◦DM [f, g]◦ η ◦ κ1 = hZ ◦ η ◦[f, g]◦ κ1 = f and similarly k ◦ c2 = g. Only uniqueness of k remains. So assume k′ : C → Z is an affine map such that k′ ◦ c1 = f and k′ ◦ c2 = g. We turn the wheel: k′ ◦ q = k′ ◦ q ◦ µ◦DM η = k′ ◦ h◦ DM (q ◦ η) = hZ ◦ DM (k′ ◦ q ◦ η) = hZ ◦ DM [k′ ◦ q ◦ η ◦ κ1, k′ ◦ q ◦ η ◦ κ2] = hZ ◦ DM [k′ ◦ c1, k′ ◦ c2] = hZ ◦ DM [f, g] = k ◦ q. So k = k′ by surjectivity of q. (cid:3) Remark The existence of coproducts of abstract M -convex sets already follows from the more general theorem of Linton [86]. (Cf. [10].) However, for the results in the remainder of this section we need to know more about the coproduct than Linton's construction provides. VIII IX By our construction, we know that every z ∈ X + Y is of the form z = h(cid:16) n > i=1 λi c1(xi)i > m > i=1 σi c2(yi)i(cid:17) for some λi, σi ∈ M , xi ∈ X and yi ∈ Y . Indeed using IV, we see that In general these are not unique. h(cid:16) n > i=1 λi c1(xi)i > m > i=1 σi c2(yi)i(cid:17) = h(cid:16) n′ > i=1 λ′i c1(x′i)i > m′ > i=1 σ′i c2(y′i)i(cid:17) iff there is a derivation Φ1, . . . , Φl ∈ D2 M (X + Y ), say Φi ≡ >j ζij ϕiji, with Φ1 = n > i=1 λi κ1(xi)i > m > i=1 and for each 1 6 i < l either 1. µ(Φi) = µ(Φi+1) or σi κ2(yi)ii , Φl = n′ > i=1 λ′i κ1(xi)i > m′ > i=1 σ′i κ2(yi)ii 2. for each j we have ζij = ζ(i+1)j and one of the following. (a) ϕij = ϕ(i+1)j . (b) ϕij = (DM κ1)(χ) and ϕ(i+1)j = (DM κ1)(χ′) for some χ, χ′ ∈ DM X (c) ϕij = (DM κ2)(χ) and ϕ(i+1)j = (DM κ2)(χ′) for some χ, χ′ ∈ DM Y with hX (χ) = hX (χ′). with hY (χ) = hY (χ′). ..193.. 119 X Exercise Show that the one-element set is an abstract M -convex set and, in fact, the final object of AConvM . Moreover, show that in AConvM n · 1 ≡ 1 + ··· + 1 n times ∼= DM{1, . . . , n}. {z } (You might want to use that DM , as a left-adjoint, preserves coproducts.) 194 Proposition The category AConvM obeys all the axioms of an effectus in total II form except perhaps for the left pullback square of (3.1). Proof In 193 V we proved AConvM has binary coproducts and in 193 X that the one-element set 1 is its final object. As DM∅ = ∅, the empty set is trivially also an abstract M -convex set and in fact the initial object of AConvM . III We continue with joint monicity of [κ1, κ2, κ2], [κ2, κ1, κ2] : 1 + 1 + 1 → 1 + 1. We can represent the convex set 1 + 1 + 1 as the set of triples (a, b, c) ∈ M 3 with a > b > c = 1 with the obvious convex structure, cf. 193 X. Similarly, we identify 1 + 1 with pairs (a, a⊥) ∈ M 2. The maps at hand are given by (a, b, c) 7→ (a, b > c) and (a, b, c) 7→ (b, a > c). (3.4) Assume (a, b > c) = (a′, b′ > c′) and (b, a > c) = (b′, a′ > c′) for a, a′, b, b′, c, c′ ∈ M with a > b > c = 1 and a′ > b′ > c′ = 1. Then clearly a = a′, b = b′ and so c = (a > b)⊥ = (a′ > b′)⊥ = c′. Thus the maps (3.4) are jointly injective, hence jointly monic. IV At last, we are ready to prove that the square on the right of (3.1) is a pull- back diagram in AConvM . So assume α : Z → X + Y is a map in AConvM with (!+!)◦ α = κ1 ◦!. Z γ α ! X κ1 ! 1 κ1 X + Y / 1 + 1 !+! We have to show α = κ1 ◦ γ, for a unique γ : Z → X + Y in AConvM . For the moment, assume κ1 is injective. Then γ, if it exists, is unique. Pick z ∈ Z. We know α(z) = h(cid:16) n > i=1 λi κ1(xi)i > m > i=1 σi κ2(yi)i(cid:17) for some λi, σi ∈ M , xi ∈ X and yi ∈ Y , where h is the convex structure on X + Y . As (!+!)◦ α =!◦ κ1, we must have >i σi = 0 and so α(z) = κ1(xz) " " $ $ # # / / / V for some xz ∈ X. This xz is unique by supposed injectivity of κ1. De- fine γ(z) = xz. By definition κ1 ◦ γ = α. To see γ is affine, pick any ϕ ∈ DM (Z). We have κ1(γ(hZ(ϕ))) = α(hZ (ϕ)) = h(DM α(ϕ)) = h(DM κ1(DM γ(ϕ))) = κ1(hX (DM γ(ϕ))) and so indeed γ(hZ(ϕ)) = hX (DM γ(ϕ)) -- i.e. γ is affine. Finally, we will show that κ1 is indeed injective. Assume κ1(x0) = κ1(x′0) for some x0, x′0 ∈ X. Let Φ1, . . . , Φl ∈ D2 M (X + Y ) be a derivation of κ1(x0) = κ1(x′0) as in 193 IX. The remainder of the proof is, in essence, a simple induction over 1 6 i < l, which has been complicated by a technicality. Define IH(i) ≡ " > y µ(Φi)(κ2(y)) = 0 and hX(cid:16)> x µ(Φi)(κ1(x))xi(cid:17) = x0". Clearly IH(1) holds and if IH(l) should hold, then x0 = x′0 as desired. So, to prove the inductive step, assume IH(i) holds. There are two possible derivation steps, see 193 IX. In the first case (i.e. µ(Φi) = µ(Φi+1)) it is obvious IH(i + 1) holds. So we continue with the other case, where (among others) we have ζij = ζ(i+1)j , where Φi ≡ >j ζij ϕiji. Without loss of generality, we may assume ζij 6= 0. For each j there are three possibilities, see 193 IX. The third does not occur: if ϕij0 = (DM κ2)(χ) for some χ ∈ DM Y and j0, then we reach a contradiction: 0 = > y µ(Φi)(κ2(y)) = > y,j ζij ⊙ ϕij (κ2(y)) > > y ζij0 ⊙ χ(y) 6= 0. We will show the first part of IH(i + 1). Pick any y ∈ Y . We are left with two possibilities -- to distinguish those, write j ∈ J if ϕij = ϕ(i+1)j and j /∈ J if ϕij = (DM κ1)(χ) and ϕ(i+1)j = (DM κ1)(χ′) for some χ, χ′ ∈ DM X with hX (χ) = hX (χ′). If j ∈ J, then ζ(i+1)j ⊙ ϕ(i+1)j (κ2(y)) = ζij ⊙ ϕij (κ2(y)) = 0. In the other case, if j /∈ J, then clearly ζ(i+1)j ⊙ ϕ(i+1)j (κ2(y)) = 0 as ϕ(i+1)j = (DM κ1)(χ′). So, we have >y µ(Φi+1)(κ2(y)) = >y,j ζ(i+1)j ⊙ ϕ(i+1)j (κ2(y)) = 0. Write rij = (>x ϕij (κ1(x)))⊥. The remainder of the proof is complicated by the fact that in general rij 6= 0. We do have >j ζij ⊙ r⊥ij = 1 and >j ζij = 1 so by 178 V we see ζij ⊙ r⊥ij = ζij, which forces ζij ⊙ rij = 0. We compute hX(cid:16)> x x,j µ(Φi)(κ1(x))xi(cid:17) = hX(cid:16)> = hX(cid:16)> = hX(cid:16)µ(cid:16)> = hX(cid:16)> j j j ζij ⊙ ϕij (κ1(x))xi(cid:17) ζij ⊙ rij x0i > > x ζij rij x0i > > ζij hX(cid:16)rij x0i > > x x ζij ⊙ ϕij (κ1(x))xi(cid:17) ϕij (κ1(x))xii(cid:17)(cid:17) ϕij (κ1(x))xi(cid:17)i(cid:17). ..193 -- 194.. 121 So, if we show that for each j, we have hX(cid:16)rij x0i > > x ϕij (κ1(x))xi(cid:17) then x0 = hX(cid:16)> x = hX(cid:16)r(i+1)j x0i > > µ(Φi)(κ1(x))xi(cid:17) = hX(cid:16)> x x ϕ(i+1)j(κ1(x))xi(cid:17), µ(Φi+1)(κ1(x))xi(cid:17), (3.5) as desired. If j ∈ J, then ϕij = ϕ(i+1)j and so rij = r(i+1)j and obviously (3.5). So assume j /∈ J. Then clearly rij = 0 = r(i+1)j as ϕij = (DM κ1)(χ)) and ϕ(i+1)j = (DM κ1)(χ′). Now (3.5) follows from hX (χ) = hX (χ′). (cid:3) 195 It is unclear whether the left square of (3.1) is a pullback in AConvM . We will see that it is, if we assume that M has a partial division in the following sense. II Definition An effect divisoid is an effect monoid M with partial binary opera- tion (a, b) 7→ a/b that is defined iff a 6 b and satisfies the following axioms. 1. a/b is the unique element of M with a/b 6 b/b and b ⊙ a/b = a 2. a 6 a/a 3. / ( a/a ) ( a/a ) = a/a IIa Beware The first axiom might be read as " a/b is defined to be the unique element such that (...)" in which case the definition is cyclical and thus problematic. What is meant instead is that is part of the structure and satisfies a/b • a/b 6 b/b • if c 6 b/b and b ⊙ a/b = a and and b ⊙ c = a, then c = a/b for all a, b, c ∈ M with a 6 b. In particular, it is not ruled out that there is an effect monoid which is an effect divisoid in two different ways. Remark The element a/a behaves like a support-projection for a. III IV Exercise Show that for an effect divisoid M , the following holds. 1. 0/0 = 0, 1/1 = 1, a/1 = a, a/a ⊙ a/a = a/a and a ⊙ b /a = a/a ⊙ b. 2. For any a 6 b 6 c, we have b/c ⊙ a/b = a/c . Examples Almost all effect monoids we encountered are effect divisoids. V 1. The effect monoid [0, 1] is an effect divisoid with a/b = a b if b 6= 0 and 0/0 = 0. With the same partial division, we see the two-element effect monoid 2 is also an effect divisoid. 2. More interestingly, if M1 and M2 are effect divisoids, then M1 × M2 is an ). In particular [0, 1]n is an effect divisoid with effect divisoid with component-wise partial division. / (a1, a2) (b1, b2) a1/b1 a2/b2 = ( , 3. Later, in 213 I, we will see that if C is an &-effectus, then (Scal C)op (the scalars with multiplication in the opposite direction) is an effect divisoid. 4. If M is a division effect monoid as in [19, dfn. 6.3], then M op is an effect divisoid in the obvious way. 5. The effect monoid on the unit interval of C[0, 1] is not an effect divisoid, but the effect monoid on the unit interval of L∞[0, 1] is, see VI. Exercise* Let X be a compact Hausdorff space. In this exercise you will show that the unit interval of C(X) is an effect divisoid if and only if X is if supp f is open for every f ∈ basically disconnected [39, 1H] -- that is: C(X). Equivalently: C(X) is σ-Dedekind complete [39, 3N.5], which is in turn equivalent to bounded ω-completeness (i.e. C(X) has suprema, resp. infima, of bounded ascending, resp. descending, sequences). VI 1. Suppose the unit interval of C(X) is an effect divisoid. Let f ∈ C(X) with 0 6 f 6 1. If supp f = X, then you're done. Otherwise pick any y /∈ supp f . Show, using Urysohn's lemma, that there is a g ∈ C(X), 0 6 g 6 f/f with g(y) = 0 and g(x) = 1 for all x ∈ supp f . Show that this implies ( is the characteristic function of supp f . Conclude X is basically disconnected. )(y) = 0 and so f/f f/f 2. Assume X is basically disconnected. Let f, g ∈ C(X) with 0 6 f 6 g 6 1. n} has open closure and so For n > 0, show that Un ≡ {x; g(x) > 1 hn ≡ ( f (x) g(x) 0 x ∈ Un otherwise. is continuous. Note h1 6 h2 6 . . . 6 1 and define is the characteristic function of supp f and partial division turns the unit interval of C(X) into an effect divisoid. f/g = supn hn. Show f/f f/g 6 g/g . Prove that this ..194 -- 195.. 123 /c /c ⊙ c ⊙ a = c ⊙ b⊥ / c ⊙ a c ⊙ b⊥ Hence ( c/c ⊙ b)⊥ > c/c ⊙ b⊥ > c ⊙ a lemma, we get /c = c/c ⊙ b⊥ ⊙ / c ⊙ a c ⊙ b⊥ = c/c ⊙ a. Indeed: c/c ⊙ a ⊥ c/c ⊙ b. 6 c/c ⊙ b⊥. VII Proposition If a ⊥ b and a > b 6 c in an effect divisoid, then a > b = a/c > b/c . VIII Proof We will first show that if c ⊙ a ⊥ c ⊙ b, then c/c ⊙ a ⊥ c/c ⊙ b. To start, /c note c ⊙ b⊥ 6 c ⊙ b⊥ > c⊥ = (c ⊙ b)⊥ 6 c ⊙ a and so IX Assume a ⊥ b and a > b 6 c. Clearly c ⊙ a/c = a ⊥ b = c ⊙ b/c . By our initial a/c = c/c ⊙ a/c ⊥ c/c ⊙ b/c = b/c . Clearly c ⊙ ( a/c > b/c ) = a > b, so a > b /c = c ⊙ ( a/c > b/c ) /c = c/c ⊙ ( a/c > b/c ) = a/c > b/c , as desired. (cid:3) 196 The following is a generalization of [19, prop. C.3]. See also [69, prop. 15]. II Theorem If M is an effect divisoid, then AConvM is an effectus. III Proof We only have to show that the square on the left of (3.1) is a pullback in AConvM -- the other axioms were proven in 194 I. Z β γ α X + Y id+! X + 1 !+id !+id 1 + Y / 1 + 1 id+! To this end, assume α : Z → X + 1 and β : Z → Y + 1 are maps in AConvM with (! + id)◦ β = (id+!)◦ α. We have to show there is a unique γ : Z → X + Y in AConvM with (id+!)◦ γ = β and (! + id)◦ γ = α. Pick any z ∈ Z. Write · for the unique element of 1. By construction of the coproduct, see 193 V, we have α(z) = h1(cid:16)λ0 κ1(·)i > β(z) = h2(cid:16)σ0 κ2(·)i > n > i=1 m > i=1 λi κ2(yi)i(cid:17) σi κ1(xi)i(cid:17) for some λi, σi ∈ M , xi ∈ X and yi ∈ Y , where h1 and h2 are the convex structures on 1 + Y and X + 1, respectively. Unfolding definitions (and identi- fying 1 + 1 ∼= M via κ1(·) = 1), it is easy to see that by assumption λi(cid:17)⊥. σi = (! + id)(β(z)) = (id+!)(α(z)) = λ0 = (cid:16) n m > i=1 > i=1 (3.6) " " # # # # / / / This suggests the following definition γ(z) = h3(cid:16) n > i=1 λi κ2(yi)i > m > i=1 σi κ1(xi)i(cid:17), (3.7) where h3 is the convex structure on X + Y . To show this a proper definition, let n′, m′ ∈ N, λ′i, σ′i ∈ M , x′i ∈ X and y′i ∈ Y be any elements such that α(z) = h1(cid:16)λ′0 κ1(·)i > β(z) = h2(cid:16)σ′0 κ2(·)i > n′ > i=1 m′ > i=1 λ′i κ2(y′i)i(cid:17) σ′i κ1(x′i)i(cid:17). We have to show h3(cid:16) n′ > i=1 λ′i κ2(y′i)i > IV m′ > i=1 σ′i κ1(x′i)i(cid:17) = h3(cid:16) n > i=1 λi κ2(yi)i > m > i=1 σi κ1(xi)i(cid:17). (3.8) Without loss of generality, we may assume n, n′, m, m′ > 1 (as we allow, for instance, λ1 = 0). Due to (3.6), it is clear λ0 = λ′0 and σ0 = σ′0. By 193 IX we know there are derivations Φ1, . . . , Φl ∈ D2 M (X + 1), say Φi ≡ >j ζij ϕiji, Ψi ≡ >j ξij ψiji with M (1 + Y ) and Ψ1, . . . , Ψk ∈ D2 Φ1 = λ0 κ1(·)i > Ψ1 = σ0 κ2(·)i > n > i=1 m > i=1 λi κ2(yi)ii Φl = λ0 κ1(·)i > σi κ1(xi)i i Ψk = σ0 κ2(·)i > n′ > i=1 m′ > i=1 λ′i κ2(y′i)ii σ′i κ1(x′i)ii . We are going to combine the Φi and Ψi into a single derivation of (3.8) by first 'applying' the Φi and then the Ψi. Define Ω1, . . . , Ωl+k ∈ D2 M (X + Y ) by Ωi = > j ζij ωiji Ωi+l = > j ξij ω(i+l)ji for 1 6 i 6 l for 1 6 i 6 k, ..195 -- 196.. 125 where ωij ∈ DM (X + Y ) are given by ωij(κ2(y)) = ϕij (κ2(y)) ωij(κ1(x)) = ϕij (κ1(·)) ⊙ ( ω(i+l)j(κ1(x)) = ψij (κ1(x)) ω(i+l)j(κ2(y)) = ψij (κ2(·)) ⊙ ( /λ0 ψ11(κ1(x)) > r(κ1(x))) /σ0 ϕl1(κ2(y)) > r(κ2(y))) r(z) =  λ0/λ0 ( σ0/σ0 ( 0 )⊥ z = κ1(x1) )⊥ z = κ2(y1) otherwise, where on the first two lines 1 6 i 6 l and on the second two 1 6 i 6 k. Before we continue, we check whether the ωij are distributions, as claimed. For 1 6 i 6 l this amounts to >x ωij(κ1(x)) = ϕij (κ1(·)). Indeed > x ωij(κ1(x)) = ϕij (κ1(·)) ⊙ r(κ1(x1)) > > x ϕij (κ1(·)) ⊙ ψ11(κ1(x)) /λ0 = ϕij (κ1(·)) ⊙ ( = ϕij (κ1(·)) ⊙ ( = ϕij (κ1(·)). λ0/λ0 λ0/λ0 )⊥ > ϕij(κ1(·)) ⊙ >x /λ0 )⊥ > ϕij(κ1(·)) ⊙ λ0/λ0 ψ11(κ1(x)) With an analogous argument, one checks ω(i+l)j is a distribution for 1 6 i 6 k. To show Ωi is a derivation (in the sense of 193 IX), first pick any 1 6 i < l. We distinguish between the two allowable derivation steps. If µ(Φi) = µ(Φi+1), then µ(Ωi) = µ(Ωi+1) by a straightforward computation and so Ωi has a valid step. In the other case, we have ζij = ζ(i+1)j and for each j we have three possibilities. In the first case, if ϕij = ϕ(i+1)j , then clearly ωij = ω(i+1)j as well. For the second case, assume ϕij = (DM κ2)(χ) and ϕ(i+1)j = (DM κ2)(χ′) for some χ, χ′ ∈ DM Y with hY (χ) = hY (χ′). Clearly ϕij(κ1(·)) = 0 and so ωij = (DM κ2)(χ). Similarly ω(i+1)j = (DM κ2)(χ′). The third case is trivial. So Ωi also makes a valid step of the second kind. With the same kind of argument, we cover Ωi+l for 1 6 i < k. The only step left to check is the one from Ωl to Ωl+1. Note ζl1 = 1, ϕl1(κ1(·)) = λ0 and λ0 ⊙ ( )⊥ = 0, so with some easy manipulation, we find λ0/λ0 Ωl = n′ > i=1 λ′i κ2(y′i)i > m > i=1 σi κ1(xi)ii . and in a similar fashion Ωl+1 = n′ > i=1 λ′i κ2(y′i)i > Ω1 = Ωl+k = n > i=1 n′ > i=1 λi κ2(yi)i > λ′i κ2(y′i)i > m > i=1 m > i=1 m′ > i=1 σi κ1(xi)ii σi κ1(xi)ii σ′i κ1(x′i)ii . So Ωi is a valid derivation of (3.8). We are in the home stretch now. Define γ(z) as in (3.7). It is easy to see γ is the unique map with (id+!)◦ γ = β and (! + id)◦ γ = α. It remains to be shown γ is affine. Write q3 : DM (X + 1) → X + 1 q1 : DM (X + Y ) → X + Y for the quotient maps used in the construction of the coproducts. We worked hard to show that for all z ∈ Z, ϕ ∈ DM Y , ψ ∈ DM X, and χ ∈ DM (X + Y ) with χ(κ1(y)) = ϕ(κ1(y)) and χ(κ2(x)) = ψ(κ2(x)), we have q2 : DM (1 + Y ) → 1 + Y V α(z) = q1(ϕ) β(z) = q2(ψ) (cid:21) =⇒ γ(z) = q3(χ). Assume z = hZ (>i λi zii). We want to show γ(z) = h3(>i λi γ(zi)i). To this end, pick ϕi, ψi, χi with α(zi) = q1(ϕi), β(zi) = q2(ψi), χi(κ1(y)) = ϕi(κ1(y)) and χi(κ2(x)) = ψi(κ2(x)). So γ(zi) = q3(χi). Define ϕ, ψ, χ as follows. ϕ ≡ µ(cid:0)> i λi ϕii(cid:1) ψ ≡ µ(cid:0)> i λi ψii(cid:1) χ ≡ µ(cid:0)> i λi χii(cid:1) Note q1(ϕ) = h1(>i λi q1(ϕi)i) = h1(>i λi α(zi)i) = α(>i λi zii) = α(z). Similarly q2(ψ) = β(z). For any y ∈ Y , we have χ(κ1(y)) = >i λi⊙ χi(κ1(y)) = >i λi ⊙ ϕi(κ1(y)) = ϕ(κ1(y)). Similarly χ(κ2(x)) = ψ(κ2(x)) for any x ∈ X. So γ(z) = q3(χ) and consequently γ(z) = q3(χ) = h3(cid:0)> i as desired. λi q3(χi)i(cid:1) = h3(cid:0)> i λi γ(zi)i(cid:1), (cid:3) 3.3 Effectuses with quotients We return to the main program of this chapter: the attempted axiomatisation of the effectus vNop. 197 ..196 -- 197.. 127 II Definition Let C be an effectus. We say C is an effectus with quotients if for each predicate p : X → 1, there exists a map ξp : X → X/p with 1 ◦ ξp 6 p⊥ satisfying the following universal property. For any map f : X → Y with 1 ◦ f 6 p⊥, there is a unique map f′ : X/p → Y such that f′ ◦ ξp = f . Any map with this universal property is called a quotient for p. IIa Remarks The universal property of quotients is essentially the same as the universal property of (contractive) compressions, which we proposed in [126]. Effectuses with quotients also appear in [23]. III Notation Unless otherwise specified ξp will denote some quotient for p. IV Examples In vNop quotients are exactly the same thing as contractive filters, see 169 VIII, 96 I and 96 V. An example of a quotient map for 1− a ∈ A is given by ξ : ⌈a⌉ A ⌈a⌉ → A with ξ(b) = √ab√a. IVa The effectus EJAop has quotients, which are in essence very similar to those of vNop, when ignoring the technical complications Jordan algebras impose. [128, prop. 25] IVb The effectus OUSop has quotients, which are rather different from the previous two examples. An example of a quotient for a predicate v on an order unit space V is given by the inclusion of the order ideal generated by 1 − v h1 − vi ≡ {a; a ∈ V ; ∃n. − n(1 − v) 6 a 6 n(1 − v)} ⊆ V, where the order ideal is considered as an order unit space with order-unit 1− v. The effectus OUGop has very similar quotients. See [23]. V Exercise Verify the following basic properties of quotients. (See [23].) 1. If ξ : X → Y is a quotient for p and ϑ : Y → Z is an isomorphism, then ϑ◦ ξ is a quotient for p as well. 2. Conversely, if ξ1 and ξ2 are both quotients for p, then there is a unique isomorphism ϑ with ξ1 = ϑ◦ ξ2. 3. Isomorphisms are quotients (for 0). 4. Maps into 0 are quotients (for 1). 5. If ξ is a quotient for p, then 1 ◦ ξ = p⊥ (Hint: apply the universal property to p⊥). 6. Quotients are epic. VI The following proposition is easy to prove, but shows an important property of quotients: any map f factors as a total map after a quotient for (1 ◦ f )⊥. Proposition Assume ξp⊥ : X → X/p⊥ is a quotient for p⊥. For any f : X → Z with 1 ◦ f = p, there is a unique total g : X/p⊥ → Z with f = g ◦ ξp⊥ . Proof (This is a simplified version of our previous proof in [23].) By definition of quotient, there is a unique g : X/p⊥ → Z with g ◦ ξp⊥ = f . Note 1 ◦ g ◦ ξp⊥ = 1 ◦ f = p = 1 ◦ ξp⊥. Thus, as ξp⊥ is epi (V), we conclude 1 ◦ g = 1. That is: g is total. (cid:3) Proposition In an effectus with quotients, quotients are closed under composi- tion. Proof (This is essentially the same proof as we gave in [23].) Assume ξ1 : X → Y is a quotient for p⊥ and ξ2 : Y → Z is a quotient for q⊥. We will prove ξ2 ◦ ξ1 is a quotient for (q ◦ ξ1)⊥. As our effectus has quotients, we can pick a quo- tient ξ : X → X/(q ◦ ξ1)⊥ of (q ◦ ξ1)⊥. First, some preparation. Note 1 ◦ ξ = q ◦ ξ1 6 1 ◦ ξ1 = p. Thus by the universal property of ξ1, there is a unique map h1 : Y → X/(q ◦ ξ1)⊥ with h1 ◦ ξ1 = ξ. As 1 ◦ h1 ◦ ξ1 = 1 ◦ ξ = q ◦ ξ1 and ξ1 is epi, we see 1 ◦ h1 = q. Thus by VII, there is a unique total map h2 : Z → X/(q ◦ ξ1)⊥ with h2 ◦ ξ2 = h1. Let g : X/(q ◦ ξ1)⊥ → Z be the unique map such that g ◦ ξ = ξ2 ◦ ξ1. We are in the following situation. VII VIII IX X ξ1 X Y ξ ξ2 h1 Z g h2 X/(q ◦ ξ1)⊥ We claim g and h2 are each other's inverse. Indeed: from g ◦ h2 ◦ ξ2 ◦ ξ1 = g ◦ ξ = ξ2 ◦ ξ1 we get g ◦ h2 = id and from h2 ◦ g ◦ ξ = h2 ◦ ξ2 ◦ ξ1 = ξ we find h2 ◦ g = id. Thus ξ2 ◦ ξ1 = g ◦ ξ for an isomorphism g, which shows ξ2 ◦ ξ1 is a quotient, see V. (cid:3) Exercise By VII we know that each map f in an effectus with quotients factors as t◦ ξ for some total map t and quotient ξ. Show that this forms an orthogonal factorization system (cf. [79]) -- that is: prove that if t′ ◦ ξ′ = t◦ ξ for some quotient ξ′ and total map t′, then there is a unique isomorphism ϑ with ξ′ = ϑ◦ ξ and t = t′ ◦ ϑ. XI In 197 XI we saw that every map f factors uniquely via a quotient for (1 ◦ f )⊥. In this sense quotients are the universal way to restrict to the support on the domain -- so why do we call them quotients instead of, say, initial-support- maps? We answer this question in IV, but need some preparation first. Definition For an effectus C, writeR Pred(cid:3) for the category with 1. as objects pairs (X, p), where X is an object of C and p ∈ Pred X and 198 II ..197 -- 198.. 129 / / , , / / ' ' H H III IV 2. an morphism between (X, p) and (Y, q) corresponds to a partial map f : X → Y in C for which we have p 6 (q⊥ ◦ f )⊥. and U f = f for arrows f . The functor U has left- and right-adjoint 0 ⊣ U ⊣ 1 Write U : R Pred(cid:3) → C for the forgetful functor -- that is: U (X, p) = X with 0, 1 : C →R Pred(cid:3) given by 0X = (X, 0), 1X = (X, 1) and 0f = 1f = f . Exercise* Show that for an effectus C, the following are equivalent. 1. C has quotients. 2. The functor 0 : C →R Pred(cid:3), has a left-adjoint Q : R Pred(cid:3) → C. If the existence of a left adjoint Q to 0 gives quotients, what does the existence of a right adjoint K to 1 amount to? It will turn out (199 VI) this is equivalent to the existence of so-called comprehensions, the topic of the next section. With both quotients and comprehension, we have a chain of four adjunctions: R Pred(cid:3) ⊣ vC ⊣Q 0 ⊣ ⊣ K 1 There are numerous other examples of categories which also have such a chain of four adjunctions between it and a natural category of object-with-predicates. See [22] for several examples. In most of these examples, the predicates corre- spond to subspaces and the left-most functor to quotienting by this subspace. This is the reason for the name 'effectus with quotients'. V Originally, we considered the universal properties of quotients and comprehen- sions separately and under different names. It was Jacobs who recognized that both universal properties appear together in a chain of adjunctions. 3.4 Effectuses with comprehension and images 199 We continue with the formal introduction of our second axiom: the existence of comprehension. II Definition Let C be an effectus. We say C has comprehension if for each predicate p : X → 1, there exists a map πp : {Xp} → X with p◦ πp = 1 ◦ πp satisfying the following universal property. For any map g : Z → X with p◦ g = 1 ◦ g, there is a unique map g′ : Z → {Xp} such that πp ◦ g′ = g. ( ( v C C [ [ Any map with this universal property is called a comprehension for p. Beware We do not assume comprehensions are total (in contrast to [23].) In 202 VIII we will see that in an effectus with quotients, comprehensions must be total. Remarks The universal property of comprehensions is essentially the same as the universal property of (contractive) corners, which we proposed in [126]. Effectuses with compression also appear in [23]. In partial form, comprehensions are essentially the same as categorical kernels, see 200 V later on. Notation Unless otherwise specified πp will denote some comprehension of p. Examples In vNop comprehensions are exactly the same thing as corners, see 169 II or 95 I. An example of a comprehension for a ∈ A is given by π : A → ⌊a⌋ A ⌊a⌋ with π(b) = ⌊a⌋ b ⌊a⌋ as proven in 95 II. The effectus EJAop also has comprehension, which is very similar to those of vNop. [128, prop. 24] The comprehension maps of OUSop are again quite different from those of vNop and EJAop. Confusingly, a comprehension for a predicate v of an or- der unit space V is given by the vector-space-quotient map q : V → V /h1−vi, where h1 − vi is the order ideal generated by 1 − v defined in 197 IVb. The effectus OUGop has similar comprehension. See [23]. III IIIa IV V Va Any extensive category with final object has comprehension, which includes Set, CH and CRng, see [23]. Exercise* Show that an effectus C has comprehension if and only if the func- tor 1 : C →R Pred(cid:3) from 198 II has a right adjoint K. Exercise Show the following basic properties of comprehensions. (See [23].) VI VII 1. If π : X → Y is a comprehension for p and ϑ : Z → X is an isomorphism, then π ◦ ϑ is a comprehension for p as well. 2. Conversely, if π1 and π2 are both comprehensions for p, then there is a unique isomorphism ϑ with π1 = π2 ◦ ϑ. 3. Isomorphisms are comprehensions (for 1). 4. Zero maps are comprehensions (for 0). 5. Comprehensions are monic. 6. p⊥ ◦ π = 0 if π is a comprehension for p. Comprehensions and categorical kernels are very similar. 200 ..198 -- 200.. 131 II Definition Let C be a category with a zero object. A (categorical) kernel of an arrow f (in symbols: ker f ) is an equalizer of f with the parallel zero map. (That is: f ◦(ker f ) = 0 and for every g with f ◦ g = 0, there exists a unique g′ with (ker f )◦ g′ = g.) We will use cok f to denote a cokernel of f -- that is: a kernel of f in C op. III Proposition An effectus with comprehension has all kernels. The kernel of a map f is given by a comprehension π(1 ◦ f )⊥ for (1 ◦ f )⊥. IV Proof Clearly 1 ◦ f ◦ π(1 ◦ f )⊥ = 0 and so f ◦ π(1 ◦ f )⊥ = 0. Assume g is any map with f ◦ g = 0. Then 1 ◦ f ◦ g = 0 and so (1 ◦ f )⊥ ◦ g = 1 ◦ g. Thus there exists a unique g′ with π(1 ◦ f )⊥ ◦ g′ = g. (cid:3) Exercise Show that in an effectus, a map f is a comprehension for p if and only if it is a kernel of p⊥. V 201 We are ready to define purity in effectuses. II Definition In an effectus a map f is called pure if f = π ◦ ξ for some compre- hension π and quotient ξ. III Example Due to 100 III pure maps in vNop are precisely the pure maps as defined in 170 I and 100 I. The pure maps B(H ) → B(K ) are exactly the maps of the form adT where T is a contractive map K → H . IV Remark In the previous chapter we discussed (in 168 III) why some more famil- iar notions of purity for states, like extremality, do not generalize properly to arbitrary maps. Recently, three other definitions of purity have been proposed in different contexts. [16, 26, 101] All three are inspired by the essential unique- ness of purification (see 139 XI) and require a monoidal structure to state. The identity map might not be pure in the sense of [16], which is too restrictive for our taste. Hazarding a guess, it seems likely that [26] considers maps pure which we don't and that [101] does not contain all our pure maps. The exact relation between our definition of pure and [26, 101] is left open. V We do not have a good handle on the structure of pure maps in an arbitrary effectus. We will need a few additional assumptions. 202 Definition Let C be an effectus. 1. We say C has images if for each map f : X → Y , there is a least predi- cate im f on Y with the property (im f )◦ f = 1 ◦ f -- that is, for every predicate p on Y with p◦ f = 1 ◦ f , we must have im f 6 p. For brevity, write im⊥ f ≡ (im f )⊥. Note im⊥ f is the greatest predicate with the property (im⊥ f )◦ f = 0. 2. We say a map f : X → Y is faithful if im f = 1. That is: f is faithful if and only if p◦ f = 0 implies p = 0 for every predicate p. Notation The expression im f ◦ g is read as im(f ◦ g). Remarks The predicate im f will play for comprehension the same role as 1 ◦ f plays for quotients. Similarly faithful is the analogue of total. Images in effec- tuses are also studied in [23]. Examples In vNop the image of a map f : A → B is given by the least projec- tion e ∈ A with the property f (1 − e) = 0. The map f is faithful if and only if f (a∗a) = 0 implies a∗a = 0 for all a ∈ A . The effectus EJAop also has all images, which are defined similarly, see [128, prop. 29]. In contrast, the effectuses OUGop and OUSop do not have all images, see [23]. Exercise Show im f ◦ g 6 im f . Conclude im f ◦ α = im f for any iso α. Exercise Show that in an effectus, any quotient is faithful. In contrast to quotients, it is not clear at all whether (without additional as- sumptions) comprehensions are total; they are part of a factorization system or are closed under composition. Lemma In an effectus with quotients, comprehensions are total. Proof Assume π is some comprehension for p. Let ξ be a quotient for (1 ◦ π)⊥. There is a total πt with π = πt ◦ ξ. As 1 ◦ πt ◦ ξ = 1 ◦ π = p◦ π = p◦ πt ◦ ξ and ξ is epi, we see 1 ◦ πt = p◦ πt. Thus, as π is a comprehension for p, there exists an f with πt = π ◦ f . Now π = πt ◦ ξ = π ◦ f ◦ ξ and so id = f ◦ ξ, since π is mono. Hence 1 = 1 ◦ id = 1 ◦ f ◦ ξ 6 1 ◦ ξ = 1 ◦ π and so π is total. (cid:3) II III IV IVa V VI VII VIII IX 3.4.1 Sharp predicates Definition Let C be an effectus with comprehension and images. 203 1. We say a predicate p is (image) sharp if p = im f for some map f . Write SPred X for the set of all sharp predicates on X. 2. We define ⌊p⌋ = im πp, where πp is some comprehension for p. (Com- prehensions for the same predicate have the same image by 202 V.) Also write ⌈p⌉ =(cid:4)p⊥(cid:5)⊥. ..200 -- 203.. 133 II III IV Beware In [23] p is called sharp whenever p ∧ p⊥ = 0. In general this is weaker than image-sharpness, which we use in this text. In IV we will see ⌊p⌋ is sharp. It is unclear whether ⌈p⌉ is sharp (without additional assumptions). Example In vNop the sharp predicates are the projections. For a predicate a ∈ A , the sharp predicate ⌈a⌉ is the least projection above a. Lemma In an effectus with comprehension and images, we have 1. ⌊p⌋ 6 p 2. πp = π⌊p⌋ ◦ α for some iso α; 3. ⌊⌊p⌋⌋ = ⌊p⌋; 4. p 6 q =⇒ ⌊p⌋ 6 ⌊q⌋; 5. ⌈p⌉◦ f 6 ⌈p◦ f⌉ and 6. ⌈p⌉◦ f = 0 iff p◦ f = 0, V VI VII VIII IX X XI for any map f : X → Y , predicates p, q on X and πp and π⌊p⌋ comprehensions for p and ⌊p⌋ respectively. Proof We will prove the statements in listed order. (Ad 1) Let π be a comprehension for p. By definition p◦ π = 1 ◦ π. Thus ⌊p⌋ = im π 6 p, as desired. (Ad 2) It is sufficient to show πp is a comprehension for ⌊p⌋. First, note ⌊p⌋◦ πp = (im πp)◦ πp = 1 ◦ πp. To show the universal property, assume g : Z → X is some map with ⌊p⌋◦ g = 1 ◦ g. Then 1 ◦ g = ⌊p⌋◦ g 6 p◦ g 6 1 ◦ g and so 1 ◦ g = p◦ g. As πp is a comprehension for p, there is a unique g′ with πp ◦ g′ = g and so πp is indeed a comprehension for ⌊p⌋ as well. (Ad 3) Follows from the previous point and 202 V. (Ad 4) Pick a comprehension πp for p and πq for q. Note 1 ◦ πp = p◦ πp 6 q ◦ πp 6 1 ◦ πp so q ◦ πp = 1 ◦ πp and thus πp = πq ◦ f for some f . By 202 V we see ⌊p⌋ = im πp = im πq ◦ f 6 im πq = ⌊q⌋. (Ad 5) Clearly p◦ f ◦ π(p ◦ f )⊥ = 0. (Recall our convention 199 IV for π's.) Thus there is some h with f ◦ π(p ◦ f )⊥ = πp⊥ ◦ h. By point 2 there is some (isomorphism) α with πp⊥ = π⌊p⊥⌋ ◦ α. We compute ⌈p⌉◦ f ◦ π(p ◦ f )⊥ = ⌈p⌉◦ πp⊥ ◦ h = ⌈p⌉◦ π⌊p⊥⌋ ◦ α ◦ h = ⌈p⌉◦ π Thus ⌈p⌉◦ f 6 im⊥ π(p ◦ f )⊥ = ⌊(p◦ f )⊥⌋⊥ = ⌈p◦ f⌉, as promised. (Ad 6) Assume ⌈p⌉◦ f = 0. From p 6 ⌈p⌉ it follows p◦ f 6 ⌈p⌉◦ f = 0. Thus p◦ f = 0. For the converse, assume p◦ f = 0. Then ⌈p◦ f⌉ = ⌈0⌉ = (im id)⊥ = 0 as id is a comprehension for 1. Thus ⌈p⌉◦ f 6 ⌈p◦ f⌉ = 0. (cid:3) ⌈p⌉⊥ ◦ α◦ h = 0. XII Exercise Let C be an effectus with comprehension and images. Show that p is sharp if and only if ⌊p⌋ = p. Conclude im πs = s for sharp s. XIII Exercise Show that in an effectus with comprehension and images we have the equality ⌈⌈p⌉◦ f⌉ = ⌈p◦ f⌉ for any map f : X → Y and predicate p on X. Exercise Show that in an effectus with images im hf, gi = [im f, im g]. Conclude that a predicate [p, q] is sharp if and only if p and q are sharp. XIV Lemma In an effectus with comprehension and images we have 204 s 6 t ⇐⇒ πs = πt ◦ h for some h, for all sharp predicates s, t on the same object. Proof (This simpler proof than ours in [23].) Assume πs = πt ◦ h. Then 6 im πt = ⌊t⌋ 203 XII= t. s 203 XII= ⌊s⌋ = im πs = im πt ◦ h 202 V as desired. Conversely assume s 6 t. Then t⊥ ◦ πs 6 s⊥ ◦ πs = 0 and so πs = πt ◦ h for some h by the universal property of πt. (cid:3) Lemma In an effectus with images, we have im[f, g] = (im f ) ∨ (im g). Proof We get im[f, g] > im f and im[f, g] > im g from II III IV [1 ◦ f, 1 ◦ g] = 1 ◦[f, g] = (im[f, g])◦[f, g] = [(im[f, g])◦ f, (im[f, g])◦ g]. To show im[f, g] is the least upper-bound, assume p > im f and p > im g. Then 1 ◦[f, g] > p◦[f, g] = [p◦ f, p◦ g] > [(im f )◦ f, (im g)◦ g] = 1 ◦[f, g], hence 1 ◦[f, g] = p◦[f, g] and so p > im[f, g], as desired. (cid:3) Corollary In an effectus with comprehension and images, we have for any sharp s, t a supremum (among all predicates) given by s ∨ t = im[πs, πt]. In particular: s ∨ t is sharp. V In 200 V we saw that comprehensions are precisely kernels of predicates. What about cokernels and quotients? 205 Proposition An effectus with quotients and images has all cokernels. A cokernel of a map f is given by a quotient ξim f of im f . Proof As 0 = (im⊥ f )◦ f = 1 ◦ ξim f ◦ f , we have ξim f ◦ f = 0. Assume g is a map with g ◦ f = 0. Then 1 ◦ g 6 im⊥ f and so g = g′ ◦ ξim f for a unique g′. Thus indeed, ξim f is a cokernel of f . (cid:3) Exercise Show that in an effectus with comprehension and images, a map f is a quotient of sharp s if and only if f is a cokernel of a comprehension πs of s. II III IV ..203 -- 205 135 3.5 ⋄-effectuses 206 In quantum physics it is not uncommon to restrict one's attention to sharp predicates (i.e. projections). Does this restriction hurt the expressivity? It does not: every (normal) state on a von Neumann algebra is determined by its values on projections. In fact, on B(H ) with dim H > 3, Gleason's famous theorem states that every measure on the projections extends uniquely to a state. We take the idea of restricting oneself to projections one step further: we also want to restrict to projections for our outcomes. It is rare for quantum processes to send projections to projections (see 210 III), so instead we consider the least projection above the outcome. The simple idea of taking the restriction to sharp predicates seriously leads to a host of interesting new notions. We will give these right off the bat and study their relevance later on. II Definition A ⋄-effectus ("diamond effectus") is an effectus with quotients, comprehension and images such that s⊥ is sharp for every sharp predicate s. In a ⋄-effectus, define for f : X → Y the following restrictions to sharp predicates. SPred X - SPred Y f⋄ f ⋄ by f⋄(s) = ⌈s◦ f⌉ and f⋄(s) = im f ◦ πs 1. We say maps f : X ⇆ Y : g are ⋄-adjoint if f⋄ = g⋄. 2. An endomap f : X → X is ⋄-self-adjoint if f is ⋄-adjoint to itself. 3. Two maps f, g : X → Y are ⋄-equivalent if f⋄ = g⋄ (or equivalently when- ever f⋄ = g⋄, see 207 VIIa.) 4. A pure endomap f is ⋄-positive if f = g ◦ g for some ⋄-self-adjoint g. For brevity, write f (s) = f⋄(s⊥)⊥ and f (s) = f⋄(s⊥)⊥ Examples The categories vNop, CvNop, EJAop and Set are all ⋄-effectuses. Let's investigate the basic properties of ( )⋄, ( )⋄ and ( ) . Exercise Show that in a ⋄-effectus both f⋄ and f are order preserving maps. III 207 II III Proposition For f : X → Y in a ⋄-effectus we have f⋄(s) 6 t⊥ ⇐⇒ f⋄(t) 6 s⊥ (3.9) for all sharp s, t. In other words: f⋄ is the left order-adjoint of f . IV Proof To start, let's prove the order-adjunction reformulation f⋄(s) 6 t (3.9) ⇐⇒ f⋄(t⊥) ≡ f (t)⊥ 6 s⊥ ⇐⇒ s 6 f (t). - m m To prove (3.9), first assume f⋄(s) 6 t⊥. Then s◦ f 6 ⌈s◦ f⌉ = f⋄(s) 6 t⊥ = im⊥ πt, where the last equality is due to 203 XII. Thus s◦ f ◦ πt = 0 which is to say s 6 im⊥ f ◦ πt, so f⋄(t) = im f ◦ πt 6 s⊥. For the converse, assume f⋄(t) 6 s⊥. Then as before (but in the other direction) we find s◦ f ◦ πt = 0 and so s◦ f 6 t⊥. Hence f⋄(s) = ⌈s◦ f⌉ 6 ⌈t⊥⌉ = ⌊t⌋⊥ = t⊥, as desired. (cid:3) Exercise Use the fact that there is an order adjunction between f⋄ and f to show that in a ⋄-effectus 1. f⋄ is order preserving; 2. f⋄ preserves suprema; 3. f preserves infima; 4. f⋄ preserves suprema; 5. f⋄ ◦ f ◦ f⋄ = f⋄ and 6. f ◦ f⋄ ◦ f = f . Lemma In a ⋄-effectus ( )⋄, ( ) and ( )⋄ are functorial -- that is 1. (id)⋄ = id, 2. (f ◦ g)⋄ = g⋄ ◦ f⋄, 3. (id) = id, 4. (f ◦ g) = g ◦ f , 5. (id)⋄ = id and 6. (f ◦ g)⋄ = f⋄ ◦ g⋄. V VI Proof We get (id)⋄ = id directly from 203 XII. For 2 we only need a single line: VII (f ◦ g)⋄(s) = ⌈s◦ f ◦ g⌉ 203 XIII= ⌈⌈s◦ f⌉◦ g⌉ = g⋄(f⋄(s)). Note that we used that ceilings are sharp. Point 3 and 4 follow easily from 1 and 2 respectively. The identity (id)⋄ = id is again 203 XII. We claim f⋄ ◦ g⋄ is left order-adjoint to (f ◦ g) , indeed f⋄(g⋄(s)) 6 t ⇐⇒ g⋄(s) 6 f (t) ⇐⇒ s 6 g (f (t)) = (f ◦ g) (t). Thus by uniqueness of order adjoints, we find (f ◦ g)⋄ = f⋄ ◦ g⋄. Exercise Derive from III that f⋄ = g⋄ if and only if f⋄ = g⋄. Lemma In a ⋄-effectus, sharp predicates are order sharp -- that is: for any predicate p and sharp predicate s with p 6 s and p 6 s⊥, we must have p = 0. Proof Note ⌈p⌉ 6 ⌈s⌉ = s and ⌈p⌉ 6 ⌈s⊥⌉ = s⊥. So by 204 I, there is an h with π⌈p⌉ = πs ◦ h. We compute (cid:3) VIIa 208 II 1 ◦ π⌈p⌉ = ⌈p⌉◦ π⌈p⌉ = ⌈p⌉◦ πs ◦ h 6 s⊥ ◦ πs ◦ h = 0. Thus π⌈p⌉ = 0 and so p 6 ⌈p⌉ = im π⌈p⌉ = 0, as desired. 206 -- 208.. (cid:3) 137 III Proposition (Cho) In a ⋄-effectus, the poset SPred X of sharp predicates on any object X, is a sub-effect algebra of Pred X and an orthomodular lattice. V IV Proof Let C be a ⋄-effectus with some object X. We will show SPred X is a sub-effect algebra of Pred X, which is additionally an ortholattice. By 177 VI this is sufficient to show SPred X is orthomodular. (Ortholattice) In 204 V we already saw that sharp s, t have a sharp supre- mum s ∨ t = im[πs, πt] in Pred X. So s ∨ t is also the supremum of s and t in SPred X. As ( )⊥ is an order anti-automorphism of both Pred X and SPred X, we know (s⊥ ∨ t⊥)⊥ is the infimum of s and t in Pred X and SPred X. As any sharp predicate s is order sharp by I, we find s ∧ s⊥ = 0 (and conse- quently s⊥ ∨ s = 1). We have shown SPred X is an ortholattice. (Sub-EA) Clearly 0, 1 are sharp and s⊥ is sharp for sharp s by definition of ⋄- effectus. To prove SPred X is a sub-effect algebra of Pred X, it only remains to be shown s > t is sharp for sharp and summable s, t. So, assume s, t are sharp and s ⊥ t. Note s ∧ t 6 s ∧ s⊥ = 0 as SPred X is an ortholattice and so by 177 Ia we find s > t = s ∨ t, which is indeed sharp. (cid:3) VII Corollary In an ⋄-effectus C the assignment X 7→ SPred X, f 7→ (f⋄, f ) yields a functor from C to OMLatGal, the category of orthomodular lattices with Galois connection between them, as defined in [55]. VI VIII There is a rather different formula for the infima of sharp predicates, which will IX X be useful later on. Lemma In a ⋄-effectus, we have s ∧ t = (πs)⋄(πs (t)) for sharp predicates s, t. Proof It is easy to see (πs)⋄(πs (t)) is a lower bound: indeed (πs)⋄(πs (t)) 6 t and (πs)⋄(πs (t)) 6 (πs)⋄(1) = im πs = s. We have to show (πs)⋄(πs (t)) is the greatest lower bound. Let r be any sharp predicate with r 6 s and r 6 t. By 204 I we have πr = πs ◦ h for some h. Thus (πr)⋄ = (πs)⋄ ◦ h⋄ = (πs)⋄ ◦ πs ◦(πs)⋄ ◦ h⋄ = (πs)⋄ ◦ πs ◦(πr)⋄. Hence r = (πr)⋄(1) = ((πs)⋄ ◦ πs ◦(πr)⋄)(1) = (πs)⋄(πs (r)) 6 (πs)⋄(πs (t)). (cid:3) XI Remark In [23] there appears a similar result due to Jacobs, which is based on subtly different assumptions. XII Exercise Show that in a ⋄-effectus, we have (ξ ◦ ξ⋄)(t) = s ∨ t for sharp s, t and quotient ξ of s. (Hint: mimic IX.) 209 We turn to ⋄-adjointness. II Exercise Show the following basic properties of ⋄-adjointness 1. f⋄ = g⋄ (f is ⋄-adjoint to g) if and only if f⋄ = g⋄. 2. If f and g are ⋄-adjoint, then im f = ⌈1 ◦ g⌉. Exercise Show in order: 1. If f is ⋄-self-adjoint, then f ◦ f is ⋄-self-adjoint. 2. If f is ⋄-positive, then f is ⋄-self-adjoint. 3. If f is ⋄-positive and f ◦ f is pure, then f ◦ f is ⋄-positive. Lemma Let α be an isomorphism in a ⋄-effectus. Then 1. s◦ α is sharp for sharp predicates s and 2. α⋄(s) = s◦ α and α⋄(s) = s◦ α−1 (so α and α−1 are ⋄-adjoint). Proof Let s be a sharp predicate. Then s = im πs. Note im α−1 ◦ πs = s◦ α -- indeed, s◦ α ◦ α−1 ◦ πs = 1 and when p◦ α−1 ◦ πs = 1, we must have p◦ α−1 > s, which gives p > s◦ α as desired. So s◦ α is indeed sharp. So α⋄(s) = ⌈s◦ α⌉ = s◦ α and α⋄(s) = im α ◦ πs = s◦ α−1 as promised. (cid:3) Definition A map f in a ⋄-effectus is a sharp map provided s◦ f is sharp for all sharp predicates s. Exercise Show that the following are equivalent. 1. f is a sharp map. 2. ⌈p◦ f⌉ = ⌈p⌉◦ f for every predicate p. III IV V 210 II Example In vNop the sharp maps are exactly the mni-maps (i.e. the normal ∗-homomorphisms). See 99 XII. III 3.6 &-effectuses In a ⋄-effectus quotient and comprehension are not tied together by its axioms. With two additional axioms, we will see quotient and comprehension become tightly interwoven. Definition An &-effectus ("andthen effectus") is a ⋄-effectus such that 1. for each object X and each predicate p on X, there is a unique ⋄-positive map asrtp : X → X ("assert p") with 1 ◦ asrtp = p (see 206 II) and 211 II 2. for every quotient ξ : Y → Z and comprehension π : X → Y the compos- ite ξ ◦ π is pure. ..208 -- 211.. 139 III In an &-effectus, we define p & q ≡ q ◦ asrtp, pronounced "p andthen q". For brevity, we will write p2 ≡ p & p. Remarks The asrtp maps are named after the assert statement found in many programming languages. The assert statement checks whether the provided Boolean expression (i.e. predicate) is true (at the time of execution) -- and if it is, it will continue the program without further action; if it isn't, it will halt execution immediately. Our asrtp maps can be thought of in the same way. Alternatively, asrtp can be viewed as a filter that blocks states for which p⊥ holds. With this intuition, the predicate p & q corresponds to whether p and q are true, by first checking p and then checking q. The asrtp maps can be used to construct more complicated measurements. For instance, let p, q, r be three predicates (measurement outcomes) on an ob- ject X with p > q > r = 1. Then hasrtp, asrtq, asrtri : X → X + X + X models measuring whether p, q, r hold without discarding the state of the (possibly af- fected) system after measurement (in contrast to hp, q, ri : X → 1 + 1 + 1). If we discard the measurement outcome, we find the map asrtp>asrtq >asrtr : X → X, which is the side-effect of measuring p, q, r. IIIa The second axiom is closely related to polar decomposition: in the case of vNop (flipping direction of arrows now), the unique isomorphism ϕ such that h◦ c = c′ ◦ ϕ◦ h′ for the suitable standard filter c′ and corner h′ is given by adu where u stricted to a unitary. is the isomtry from the polar decomposition up⌈h⌉ c(1)⌈h⌉ ≡ pc(1)⌈h⌉ re- IV Examples The category vNop is an &-effectuses with asrta : b 7→ √ab√a. The first axiom is proven in 105 V and the second in 100 III. The full subcate- gory CvNop of commutative von Neumann algebras is a &-effectus as well. The only other known example of an &-effectus is the category EJAop of Euclidean Jordan algebras with positive unital maps in the opposite direction, see [128]. V Proposition For a predicate p in an &-effectus the following are equivalent. 1. p is sharp, 2. p & p = p and 3. asrtp ◦ asrtp = asrtp. VI Proof First we prove that p is sharp if and only if p&p = p. So, assume p is sharp. As ⋄-positive maps are ⋄-self-adjoint we have im asrtp = ⌈1 ◦ asrtp⌉ = ⌈p⌉ = p. Thus p & p = p◦ asrtp = 1 ◦ asrtp = p. For the converse, assume p & p = p. From im asrtp = ⌈1 ◦ asrtp⌉ = ⌈p⌉, we get ⌈p⌉◦ asrtp = p & ⌈p⌉ = 1 ◦ asrtp = p. By assumption p&p = p. So p&(⌈p⌉⊖p) = 0. Hence ⌈p⌉⊖p 6 im⊥ asrtp = ⌈p⌉⊥. However ⌈p⌉ ⊖ p 6 ⌈p⌉. Thus by 208 I get ⌈p⌉ ⊖ p = 0. So p is indeed sharp. Clearly, if asrtp ◦ asrtp = asrtp, then p = 1 ◦ asrtp = 1 ◦ asrtp ◦ asrtp = p & p. It only remains to be shown asrtp ◦ asrtp = asrtp whenever p is sharp. So as- sume p is sharp. By definition asrtp is pure: asrtp = π ◦ ξ for some quotient ξ and comprehension π. By assumption ξ ◦ π is pure as well, so there is a quo- tient ξ′ and comprehension π′ with ξ ◦ π = π′ ◦ ξ′. Note 1 ◦ ξ = 1 ◦ asrtp = p and 1 ◦ asrtp ◦ asrtp = p & p = p = 1 ◦ ξ, hence 1 ◦ ξ = 1 ◦ asrtp ◦ asrtp = 1 ◦ π ◦ ξ ◦ π ◦ ξ = 1 ◦ π ◦ π′ ◦ ξ′ ◦ ξ = 1 ◦ ξ′ ◦ ξ, so 1 ◦ ξ′ = 1. Thus ξ′ is an iso. Also (im π′)◦ ξ ◦ π = (im π′)◦ π′ ◦ ξ′ = 1 ◦ ξ′ = 1 from which it follows p = im π 6 (im π′)◦ ξ 6 1 ◦ ξ = p. Thus (im π′)◦ ξ = p = 1 ◦ ξ. Hence im π′ = 1 and so π′ is an isomorphism. Now we know π ◦ π′ is a comprehension and ξ′ ◦ ξ is a quotient, we see asrtp ◦ asrtp is pure. As asrtp is ⋄-self-adjoint, we see asrtp ◦ asrtp is ⋄-positive. By uniqueness of positive maps asrtp ◦ asrtp = asrtp, as desired. (cid:3) Proposition Let C be an &-effectus and s be any sharp predicate. There exist comprehension πs of s and quotient ζs of s⊥ such that ζs ◦ πs = id and πs ◦ ζs = asrts. In fact, for every comprehension π of s, there is a quotient ξ of s⊥ with ξ ◦ π = id and π ◦ ξ = asrts and conversely for every quotient ξ of s⊥ there exists a comprehension π of s with ξ ◦ π = id and π ◦ ξ = asrts. Proof Let s be any sharp predicate. By definition ⋄-positive maps are pure and so there is a quotient ξ and comprehension π with π ◦ ξ = asrts. So VII VIII π ◦ ξ = asrts V = asrts ◦ asrts = π ◦ ξ ◦ π ◦ ξ. Thus ξ ◦ π = id. We compute s = 1 ◦ asrts = 1 ◦ ξ and so as ξ ◦ π = id by dfn. ξ and π by dfn. ( )⋄ by ⋄-s.a. asrts by dfn. ( )⋄ im π = im π ◦ ξ ◦ π = im asrts ◦ π = (asrts)⋄(im π) = (asrts)⋄(im π) = ⌈(im π)◦ π ◦ ξ⌉ = ⌈1 ◦ ξ⌉ = s. Thus π is a comprehension of s and ξ is a quotient of s⊥. We have proven the first part. For the second part, let π′ be any comprehension of s. Then π′ = π ◦ α for some iso α. Define ξ′ = α−1 ◦ ξ. It is easy to see π′ ◦ ξ′ = asrts and ξ′ ◦ π′ = id. The other statement is proven in a similar way. (cid:3) ..211.. 141 IX Notation In an &-effectus together with chosen comprehension πs for s, we will write ζs for the unique quotient for s⊥ satisfying ζs ◦ πs = id and πs ◦ ζs = the corresponding quotient of πs and vice versa πs the asrts. We call this ζs corresponding comprehension of ζs. X Warning ξs⊥ = ζs! XI Proposition In an &-effectus both comprehensions and pure maps are closed under composition. XII Proof We will first prove that comprehensions are closed under composition. Assume π1 : X → Y and π2 : Y → Z are comprehensions with s = im π1 and t = im π2. We will show π2 ◦ π1 is a comprehension for im π2 ◦ π1. To this end, let f : V → Z be any map with (im π2 ◦ π1)⊥ ◦ f = 0. As im π2 ◦ π1 6 im π2 = t we get t⊥ ◦ f = 0, so f = π2 ◦ g2 for a unique g2 : V → Y . Let ζ2 be a quotient for t⊥ such that ζ2 ◦ π2 = id, which exists by VII. Then s◦ ζ2 ◦ π2 ◦ π1 = s◦ π1 = 1 so s◦ ζ2 > im π2 ◦ π1. Consequently s◦ g2 = s◦ ζ2 ◦ π2 ◦ g2 = s◦ ζ2 ◦ f > (im π2 ◦ π1)◦ f = 1 ◦ f = 1 ◦ g2. Hence there exists a unique g1 : V → X with π1 ◦ g1 = g2 and so π2 ◦ π1 ◦ g1 = f . By monicity of π2 ◦ π1, this g1 is unique and so π2 ◦ π1 is indeed a comprehension. XIII Now we will prove that pure maps are closed under composition. Assume g : X → Y and f : Y → Z are pure maps. Say f = π1 ◦ ξ1 and g = π2 ◦ ξ2 for some com- prehensions π1, π2 and quotients ξ1, ξ2. By definition of &-effectus there is a comprehension π′ and quotient ξ′ such that ξ1 ◦ π2 = π′ ◦ ξ′. By the previ- ous point π1 ◦ π′ is a comprehension and ξ′ ◦ ξ2 is a quotient by 197 IX. Thus f ◦ g = π1 ◦ π′ ◦ ξ′ ◦ ξ2 is pure. (cid:3) XIV Exercise Show that in an &-effectus, we have asrtp ◦ asrtp = asrtp&p. XV Exercise Show that in an &-effectus, we have im f 6 s ⇐⇒ asrts ◦ f = f 1 ◦ f 6 t ⇐⇒ f ◦ asrtt = f for any sharp predicates s and t. XVI Definition Let C be an &-effectus. In XI we saw pure maps are closed under composition. Write Pure C for the subcategory of pure maps. 212 II Lemma In an &-effectus, ζ⌈p⌉ ◦ asrtp is a quotient for p⊥. Proof Recall im asrtp = ⌈1 ◦ asrtp⌉ = ⌈p⌉ and so im ζ⌈p⌉ ◦ asrtp = (ζ⌈p⌉)⋄(⌈p⌉) = im ζ⌈p⌉ ◦ π⌈p⌉ = im id = 1, where π⌈p⌉ is the comprehension corresponding to ζ⌈p⌉. By 211 XI ζ⌈p⌉ ◦ asrtp = π ◦ ξ for some comprehension π and quotient ξ. Putting it together: 1 = im ζ⌈p⌉ ◦ asrtp = im π ◦ ξ 6 im π, so im π = 1 and thus π is an isomorphism, hence ζ⌈p⌉ ◦ asrtp is a quotient for the orthocomplement of ⌈p⌉◦ asrtp = p. (cid:3) Proposition Every map f in an &-effectus factors as III f = πim f ◦ g ◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f (3.10) for a unique total and faithful map g. Furthermore the following holds. 1. If f is pure, then g is an isomorphism. 2. If f is pure and im f = 1, then f is a quotient. 3. If f is pure and 1 ◦ f = 1, then f is a comprehension. Proof By the universal property of πim f , there is a unique g′ with f = πim f ◦ g′. To show g′ is faithful, assume p◦ g′ = 0 for some predicate p. Then 0 = p◦ g′ = p◦ ζim f ◦ πim f ◦ g′ = p◦ ζim f ◦ f and so p◦ ζim f 6 im⊥ f , hence p = p◦ ζim f ◦ πim f 6 (im⊥ f )◦ πim f = 0, which shows g′ is indeed faithful. Note 1 ◦ g′ = 1 ◦ πim f ◦ g′ = 1 ◦ f . By 197 VII and I there is a unique total g with g′ = g ◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f . Clearly (3.10) holds and g is the unique map for which (3.10) holds as comprehensions are mono and quotients are epi. As 1 = im g′ = im g ◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f 6 im g, we see im g = 1 and so g is faithful. To prove point 1, assume f is pure. That is: f = π ◦ ξ for some compre- hension π and quotient ξ. As 1 ◦ π = 1 and im ξ = 1, we see im π = im f and 1 ◦ ξ = 1 ◦ f . Thus π = πim f ◦ α and ξ = β ◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f for some iso- morphisms α and β. Thus f = πim f ◦ α◦ β ◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f and so by unique- ness of g, we see g = α◦ β is an isomorphism. To prove point 2, additionally assume im f = 1. Then πim f is an iso and so using the previous, we see f is indeed a quotient. Point 3 is just as easy. (cid:3) λ/µ = λ′ ◦ ζ⌈µ⌉. Proposition If C is an &-effectus, then (Scal C)op is an effect divisoid, see 195 II. Proof Let λ, µ be scalars with λ 6 µ. Recall that the scalar 1 = id and so µ = 1 ◦ asrtµ = asrtµ. By 212 I, there is a unique λ′ with λ′ ◦ ζ⌈µ⌉ ◦ µ = λ. Define For the moment, assume λ = µ. Then λ′ ◦ ζ⌈µ⌉ ◦ µ = µ = 1 ◦ ζ⌈µ⌉ ◦ µ as im µ = im asrtµ = ⌈µ⌉. By epicity of ζ⌈µ⌉ ◦ µ, we see λ′ = 1, hence µ/µ = ⌈µ⌉. As we have µ 6 ⌈µ⌉ and ⌈⌈µ⌉⌉ = ⌈µ⌉, we see axioms 2 and 3 hold. λ/µ = λ′ ◦ ζ⌈µ⌉ 6 λ/µ = λ′ ◦ ζ⌈µ⌉ ◦ µ = λ as required. Assume σ is 1 ◦ ζ⌈µ⌉ = ⌈µ⌉ = µ/µ an arbitrary scalar with µ ⊙op σ = λ and σ 6 µ/µ ≡ ⌈µ⌉. Then σ = σ′ ◦ ζ⌈µ⌉ for a unique σ′. As λ′ ◦ ζ⌈µ⌉ ◦ µ = λ = σ′ ◦ ζ⌈µ⌉ ◦ µ we must have λ′ = σ′, whence (cid:3) We return to the general case to prove axiom 1. Clearly and µ ⊙op λ/µ = σ. ..211 -- 213.. 143 IV 213 II III Lemma In an &-effectus, if s ⊥ t for sharp s, t, then s & t = 0 = t & s. IV Proof Write r ≡ (s > t)⊥. As 1 = s > t > r and s & s = s, we have s = s & 1 = s & s > s & t > s & r = s > s & t > s & r and so s & t 6 s & t > s & r = 0. (cid:3) V Exercise Show that in an &-effectus, we have p 6 s ⇐⇒ s & p = p for sharp s and any predicate p. VI Exercise Show that in a &-effectus, a map hf, gi is sharp if and only if both f and g are sharp. (Hint: use 208 III and 203 XIV.) 3.7 †-effectuses 214 Definition A †-category ("dagger category") is a category C together with an involutive identity-on-objects functor ( )† : C → C op -- that is, for all objects X and maps f, g in C, we have 1. (f ◦ g)† = g† ◦ f† 2. id† = id 3. f†† = f and 4. X† = X. Cf. [14, 49, 50, 102]. In any †-category we may define the following. 1. An endomap f is called †-self-adjoint if f† = f . 2. An endomap f is †-positive if f = g† ◦ g for some other map g. 3. An isomorphism α is called †-unitary whenever α−1 = α†. II Example The category Hilb of Hilbert spaces with bounded linear maps is a †-category with the familiar adjoint as †. 215 Definition We call an &-effectus C a †-effectus ("dagger effectus") provided 1. Pure C is a †-category satisfying asrt†p = asrtp and f is ⋄-adjoint to f†; 2. for every †-positive f , there is a unique †-positive g with g ◦ g = f and 3. ⋄-positive maps are †-positive. II III IV V VI VIa VIIb Examples In VI we will see that the category vNop is a †-effectus. Recently we have shown with van de Wetering that the category EJAop of Euclidean Jordan algebras with positive unital maps in the opposite direction, is a †-effectus as well. [128] Theorem An &-effectus is a †-effectus if and only if 1. for every predicate p, there is a unique predicate q with q & q = p; 2. asrt2 p&q = asrtp ◦ asrt2 q ◦ asrtp for all predicates p, q and 3. a quotient for a sharp predicate (e.g. ζs) is a sharp map, see 210 I. Proof Necessity will be proven in 216 XI and sufficiency in 220 II. (cid:3) Especially the sufficiency requires quite some preparation. For convenience, call C a †′-effectus if C is an &-effectus satisfying axioms 1, 2 and 3 from the Theorem above. Corollary The category vNop of von Neumann algebras with ncpu-maps in the opposite direction, which is an &-effectus by 211 IV, is also a †-effectus. Remarks The dagger on the pure maps of vN (that exists by the previous corollary) is fixed by the following two rules. and pure maps are rigid, see 102 IX. 1. For any (nmiu-)isomorphism ϑ we have ϑ† = ϑ−1 as ϑ is ⋄-adjoint to ϑ−1 2. The standard filter c : A → ⌈b⌉ A ⌈b⌉ for b ∈ A (ie. c(a) = √ba√b, see 169 VIII) has as dagger c† : ⌈b⌉ A ⌈b⌉ → A given by c†(a) = √ba√b. (This follows from asrt†p = asrtp and 216 VII.) More concretely, for any pure ncp-map ϕ : A → B we can find a unique iso- morphism ϑ : ⌈ϕ⌉ A ⌈ϕ⌉ → ⌈ϕ(1)⌉ B ⌈ϕ(1)⌉ such that In special cases, we can give a simpler definition of the dagger. For instance, if the pure map is of the form ϕ : B(H ) → B(K ), then ϕ = adT for some operator T : K → H and in that case ϕ† = adT ∗ . Another example is a pure map ϕ : A → B between finite-dimensional von Neumann algebras. The Hilbert -- Schmidt inner product [a, b] ≡ tr a∗b turns a finite-dimensional von Neumann algebra into a Hilbert space. The adjoint of ϕ with respect to this inner product equals ϕ†. This can be seen by con- sidering the special cases of nmiu-isomorphisms and standard filters (noting nmiu-isomorphisms are trace-preserving). ..213 -- 215.. 145 (this follows from 100 III and 98 IX) and so by the previous two rules we get ϕ(a) = pϕ(1) ϑ(cid:0)⌈ϕ⌉ a⌈ϕ⌉(cid:1)pϕ(1) ϕ†(b) = ⌈ϕ⌉ ϑ−1(cid:16)pϕ(1) bpϕ(1)(cid:17) ⌈ϕ⌉ (for all a ∈ A ) (for all b ∈ B). VII Remarks In an &-effectus with separating predicates (see 190 II), the second axiom of a †′-effectus is equivalent to (p & q)2 & r = p & (q2 & (p & r)) for all predicates p, q, r. This is essentially the fundamental formula of quadratic Jordan algebras [89, §4.2] (with Uxy ≡ x2 & y). by Selinger [103] who calls it the unique square root axiom. The second axiom of a †-effectus has been considered before in †-categories 216 II Lemma In a †-effectus, a map is †-positive if and only if it is ⋄-positive. Proof Assume f is †-positive. By assumption 2, there is a (unique) †-positive g with f = g ◦ g. By †-positivity of g, there is an h with g = h† ◦ h. Using the fact h is ⋄-adjoint to h†, we see g is ⋄-self-adjoint: g⋄ = (h† ◦ h)⋄ = h⋄ ◦(h†)⋄ = (h†)⋄ ◦ h⋄ = (h† ◦ h)⋄ = g⋄. Thus f is the square of the ⋄-self-adjoint map g, hence f is ⋄-positive. (cid:3) III Lemma Predicates in a †-effectus have a unique square root: for every predi- cate p, there is a unique predicate q with q & q = p. IV Proof Let p : X → 1 be any predicate. By assumption 3, the ⋄-positive map asrtp is also †-positive. So by assumption 2, there is (a unique) †-positive map f with f ◦ f = asrtp. Define q = 1 ◦ f . By I f is ⋄-positive and so by uniqueness of ⋄-positive maps, we get f = asrt1 ◦ f ≡ asrtq. We compute q & q = q ◦ asrtq ≡ 1 ◦ f ◦ asrt1 ◦ f = 1 ◦ f ◦ f = 1 ◦ asrtp = p, which shows p has as square root q. To show uniqueness, assume p = r & r for some predicate r. Note asrtp = asrtr&r 211 XIV= asrtr ◦ asrtr. As asrtr is ⋄-positive, it is also †-positive by the third axiom. So by the second axiom asrtr = asrtq. Thus r = q, which shows uniqueness of the square root. (cid:3) V Proposition In an &-effectus with square roots (e.g. †- or †′-effectus) we have for every isomorphism α and predicate p. asrtp ◦ α = α◦ asrtp ◦ α VI Proof There is some q with q & q = p. The map α−1 ◦ asrtq ◦ α is ⋄-self-adjoint: (α−1 ◦ asrtq ◦ α)⋄ = α−1 ⋄ ◦(asrtq)⋄ ◦ α⋄ 209 IV= α⋄ ◦(asrtq)⋄ ◦(α−1)⋄ = (α−1 ◦ asrtq ◦ α)⋄. Thus α−1 ◦ asrtq ◦ α◦ α−1 ◦ asrtq ◦ α = α−1 ◦ asrtp ◦ α is ⋄-positive. By unique- ness of ⋄-positive maps, we get α−1 ◦ asrtp ◦ α = asrt1 ◦ α−1 ◦ asrtp ◦ α = asrtp ◦ α. Postcomposing α, we find asrtp ◦ α = α◦ asrtp ◦ α, as desired. (cid:3) Proposition In a †-effectus: ζ†s = πs (recall convention 211 IX for πs and ζs). Proof As πs is ⋄-adjoint to π†s, we have VII VIII im π†s = (π†s)⋄(1) = (πs)⋄(1) = ⌈1 ◦ πs⌉ = 1 and similarly (cid:6)1 ◦ π†s(cid:7) = im πs = s. As 1 ◦ π†s 6 (cid:6)1 ◦ π†s(cid:7) = s, there is some pure h with π†s = h◦ ζs. By 212 III, there is also some pure and faithful g with ζ†s = πs ◦ g. Using ζs ◦ πs = id twice we find id = id† = π†s ◦ ζ†s = h◦ ζs ◦ πs ◦ g = h◦ g. (3.11) Now 1 = 1 ◦ h◦ g 6 1 ◦ g and so g is total. Clearly ζ†s ◦ ζs is †-positive, hence ⋄- positive and so by uniqueness of ⋄-positive maps: ζ†s ◦ ζs = asrt1 ◦ ζ† s ◦ ζs = asrt1 ◦ πs ◦ g ◦ ζs = asrts = πs ◦ ζs. (cid:3) So by epicity of ζs, we find ζ†s = πs, as desired. Corollary In a †-effectus, πs is ⋄-adjoint to ζs. Also α† = α−1 for any iso α. Exercise Show that in an &-effectus where every predicate has a square root and where πs is ⋄-adjoint to ζs (e.g. a †-effectus) we have asrtp ◦ ζs = ζs ◦ asrtp ◦ ζs . (Hint: mimic the proof of V.) Theorem A †-effectus is a †′-effectus. Proof Assume C is a †-effectus. Axiom 1 is already proven in III. (Ax. 2) Let p, q be predicates. To start, note 1 ◦ asrtq ◦ asrtp = p & q and IX X XI XII XIII im asrtq ◦ asrtp = (asrtq)⋄(p) = (asrtq)⋄(p) = ⌈q & p⌉ . So using 212 III we know asrtq ◦ asrtp = π⌈q&p⌉ ◦ α◦ ζ⌈p&q⌉ ◦ asrtp&q for some iso α. Applying the dagger to both sides we get, using VII and IX: asrtp ◦ asrtq = (asrtq ◦ asrtp)† ⌈q&p⌉ = asrtp&q ◦ ζ† = asrtp&q ◦ π⌈p&q⌉ ◦ α−1 ◦ ζ⌈q&p⌉. ⌈p&q⌉ ◦ α† ◦ π† ..215 -- 216.. 147 Combining both: asrtp ◦ asrt2 q ◦ asrtp = asrtp&q ◦ π⌈p&q⌉ ◦ α−1 ◦ ζ⌈q&p⌉ ◦ π⌈q&p⌉ ◦ α◦ ζ⌈p&q⌉ ◦ asrtp&q = asrtp&q ◦ π⌈p&q⌉ ◦ ζ⌈p&q⌉ ◦ asrtp&q = asrtp&q ◦ asrt⌈p&q⌉ ◦ asrtp&q 211 XV= asrt2 p&q, as desired. XIV (Ax. 3) Pick any sharp predicates s, t. We want to show t◦ ζs is sharp. To this end, we will show t◦ ζs is the image of πs ◦ πt. Clearly t◦ ζs ◦ πs ◦ πt = 1. Let p be any sharp predicate with p◦ πs ◦ πt = 1. Then p◦ πs > im πt = t. we get t◦ ζs 6 ⌈t◦ ζs⌉ 6 p, which shows t◦ ζs is the image of πs ◦ πt and consequently sharp. (cid:3) So p⊥ ◦ πs 6 t⊥ and hence (cid:6)p⊥ ◦ πs(cid:7) 6 t⊥. As πs is ⋄-adjoint to ζs by IX, 217 Let f be a pure map in a †′-effectus. We will work towards the definition of f†. By 212 III we know there is an iso α with f = πim f ◦ α ◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f . In a †-effectus we have asrt†p = asrtp, ζ†s = πs, α† = α−1 and π†s = ζs (for corresponding ζs and πs) -- so we are forced to define f† = asrt1 ◦ f ◦ π⌈1 ◦ f⌉ ◦ α−1 ◦ ζim f , (3.12) where ζim f is the unique corresponding quotient of πim f and π⌈1 ◦ f⌉ the unique corresponding comprehension of and ζ⌈1 ◦ f⌉, see 211 IX. Before we declare (3.12) a definition, we have to check whether it is independent of choice of π (and corresponding ζ). So suppose f = π′ ◦ α′ ◦ ζ′ ◦ asrt1 ◦ f for some iso α′, com- prehension π′ of im f and quotient ζ′ of ⌈1 ◦ f⌉⊥. There are isos β and γ such that π′ = πim f ◦ β and ζ′ = γ ◦ ζ⌈1 ◦ f⌉. We will take a moment to relate α and α′: as πim f is mono and ζ1 ◦ f ◦ asrt1 ◦ f is epic, we have β ◦ α′ ◦ γ = α and so (α′)−1 = γ ◦ α−1 ◦ β. To continue, it is easy to see β−1 ◦ ζim f is the unique corresponding quotient to π′ and π⌈1 ◦ f⌉ ◦ γ−1 is the unique corresponding com- prehension to ζ′. So with this choice of quotient and comprehension, we are forced to define f† = asrt1 ◦ f ◦ π⌈1 ◦ f⌉ ◦ γ−1 ◦ α′−1 ◦ β−1 ◦ ζim f = asrt1 ◦ f ◦ π⌈1 ◦ f⌉ ◦ γ−1 ◦ γ ◦ α−1 ◦ β ◦ β−1 ◦ ζim f = asrt1 ◦ f ◦ π⌈1 ◦ f⌉ ◦ α−1 ◦ ζim f , which is indeed consistent with (3.12). So we are justified to declare: Definition In a †′-effectus, for a pure map f , define f† = asrt1 ◦ f ◦ π⌈1 ◦ f⌉ ◦ α−1 ◦ ζim f , where α is the unique iso such that f = πim f ◦ α◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f . Exercise Show that in a †′-effectus, we have asrt†p = asrtp π†s = ζs ζ†s = πs α† = α−1 for a quotient ζs corresponding to πs and iso α. To compute f††, we need to put f† in the standard form of 212 III. To do this, we need to pull asrtp from one side to the other. In this section we will work towards a general result for this. Lemma In a †′-effectus, πs is ⋄-adjoint to ζs. Proof Pick any sharp t, u. We have to show t◦ ζs 6 u⊥ if and only if u ◦ πs 6 t⊥. So assume t◦ ζs 6 u⊥. Then t = t◦ ζs ◦ πs 6 u⊥ ◦ πs = (u ◦ πs)⊥ so u ◦ πs 6 t⊥. For the converse, assume u ◦ πs 6 t⊥. Then u ◦ asrts = u ◦ πs ◦ ζs 6 t⊥ ◦ ζs. From this, the ⋄-self-adjointness of asrts and the fact that t⊥ ◦ ζs is sharp, we find (t⊥ ◦ ζs)⊥ ◦ πs ◦ ζs 6 u⊥ and so II III 218 II III t◦ ζs = (t⊥ ◦ id)⊥ ◦ ζs = (t⊥ ◦ ζs ◦ πs)⊥ ◦ ζs = (t⊥ ◦ ζs)⊥ ◦ πs ◦ ζs 6 u⊥, as desired. (cid:3) Definition In an &-effectus, a map f is pristine if it is pure and 1 ◦ f is sharp. Remarks In general, pristine maps are not closed under composition: in vNop, the maps asrt0ih0 and asrt+ih+ on M2 are both pristine, but their composite is not. However, with a non-standard composition of pristine maps (namely f · g ≡ asrt⌊1 ◦ f ◦ g⌋ ◦ f ◦ g), the pristine maps are closed under composition and even form a †-category [23]. Essentially the same construction has been found independently by [53] for partial isometries between Hilbert spaces. IV V Exercise Show that in an &-effectus, every pristine map h is of the form VI for some iso α. h = πim h ◦ α◦ ζ1 ◦ h ..216 -- 218.. 149 VII Proposition In a †′-effectus, with pristine map h, we have for any predicate p with p 6 im h. asrtp ◦ h = h◦ asrtp ◦ h VIII Proof For brevity, write t = 1 ◦ h and s = im h. By VI there is some iso α such that h = πs ◦ α◦ ζt. Note im asrtp = ⌈p⌉ 6 ⌈s⌉ = s and so asrtp = asrts ◦ asrtp by the first rule of 211 XV. By the second rule of 211 XV and 1 ◦ p = p 6 s, we see p = p◦ asrts. Thus asrtp ◦ πs = asrts ◦ asrtp ◦ asrts ◦ πs = πs ◦ ζs ◦ asrtp ◦ πs ◦ ζs ◦ πs = πs ◦ asrtp ◦ πs ◦ ζs ◦ πs = πs ◦ asrtp ◦ πs. Putting everything together asrtp ◦ h = asrtp ◦ πs ◦ α◦ ζt = πs ◦ asrtp ◦ πs ◦ α◦ ζt = πs ◦ α◦ asrtp ◦ πs ◦ α ◦ ζt = πs ◦ α◦ ζt ◦ asrtp ◦ πs ◦ α ◦ ζt = h◦ asrtp ◦ h, by 216 X and II by 216 V by 216 X and II as desired. (cid:3) IX Exercise Working in a †′-effectus, show in order 1. if h ≡ πim h ◦ α◦ ζ1 ◦ h is some pristine map, then h† = π1 ◦ h ◦ α−1 ◦ ζim h; 2. h†† = h for any pristine h; 3. h† ◦ h = asrt1 ◦ h for any pristine map h; 4. p◦ h† 6 im h for any predicate p and pristine map h and 5. if p 6 1 ◦ h, then asrtp ◦ h† ◦ h = h◦ asrtp for any pristine map h. X Proposition In a †′-effectus, for every pure map f , there exists a unique pristine map h with 1 ◦ h = ⌈1 ◦ f⌉ and f = h◦ asrt1 ◦ f . Furthermore f† = asrt1 ◦ f ◦ h†. XI Proof By 212 III we know f = πim f ◦ α◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f for some iso α. De- fine h = πim f ◦ α ◦ ζ⌈1 ◦ f⌉. Clearly 1 ◦ h = ⌈1 ◦ f⌉ and f = h◦ asrt1 ◦ f . Also f† ≡ asrt1 ◦ f ◦ π⌈1 ◦ f⌉ ◦ α−1 ◦ ζim f = asrt1 ◦ f ◦ asrt⌈1 ◦ f⌉ ◦ π⌈1 ◦ f⌉ ◦ α−1 ◦ ζim f ≡ asrt1 ◦ f ◦ h†. by 211 XV Only uniqueness remains. Assume f = h′ ◦ asrt1 ◦ f for some pristine map h′ with ⌈1 ◦ h′⌉ = ⌈1 ◦ f⌉. By 212 III and 211 XV, there is an iso α′ with h′ = πim h ◦ α′ ◦ ζ⌈1 ◦ f⌉. As ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f is a quotient (by 212 I), quotients are faithful and f = h′ ◦ asrt1 ◦ f , we see im h′ = im f and so α = α′ by 212 III. (cid:3) Proposition In a †′-effectus, we have f†† = f for any pure map f . Proof By X we have f = h◦ asrt1 ◦ f and f† = asrt1 ◦ f ◦ h† for some pristine h with 1 ◦ h = ⌈1 ◦ f⌉. Clearly 1 ◦ f 6 ⌈1 ◦ f⌉ = 1 ◦ h = im h†, so by VII we get f† = asrt1 ◦ f ◦ h† = h† ◦ asrt1 ◦ f ◦ h† . Consequently XII XIII f†† = asrt1 ◦ f ◦ h† ◦ h†† IX= asrt1 ◦ f ◦ h† ◦ h IX= h◦ asrt1 ◦ f = f, as desired. (cid:3) Now we tackle the most tedious part of 215 III: we will show (f ◦ g)† = g† ◦ f† in a †′-effectus. To avoid too much repetition, let us fix the setting. Setting Let f, g be two composable pure maps in a †′-effectus. For brevity, write p = 1 ◦ f , q = 1 ◦ g, s = im f and t = im g. Let ϕ and ψ be the unique isomorphisms (see 212 III) such that 219 II f = πs ◦ ϕ◦ ζ⌈p⌉ ◦ asrtp and g = πt ◦ ψ ◦ ζ⌈q⌉ ◦ asrtq. Define h = πs ◦ ϕ◦ ζ⌈p⌉ and k = πt ◦ ψ ◦ ζ⌈q⌉. To compute (f ◦ g)†, we have to put f ◦ g in the standard form of 212 III. We will do this step-by-step, first we put ζ⌈p⌉ ◦ asrtp ◦ πt in standard form. 1 ◦ ζ⌈p⌉ ◦ asrtp ◦ πt = ⌈p⌉◦ asrtp ◦ πt 211 XV= p◦ πt im ζ⌈p⌉ ◦ asrtp ◦ πt = (ζ⌈p⌉ ◦ asrtp)⋄(t) 218 II= (cid:6)t◦ asrtp ◦ π⌈p⌉(cid:7) . Thus there is a unique isomorphism χ with ζ⌈p⌉ ◦ asrtp ◦ πt = π⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ χ◦ ζ⌈p ◦ πt⌉ ◦ asrtp ◦ πt. (3.13) Next, we consider asrtp ◦ k ◦ asrtq, clearly 1 ◦ asrtp ◦ k ◦ asrtq = p◦ k ◦ asrtq = p◦ g. Concerning the image, first note p◦ k = p◦ πt ◦ ψ ◦ ζ⌈q⌉ 6 1 ◦ ζ⌈q⌉ = ⌈q⌉ and so we must have im asrtp ◦ k 6 ⌈q⌉, which implies im asrtp ◦ k ◦ asrtq = (asrtp ◦ k)⋄(q) = ⌈⌈q⌉◦ asrtp ◦ k⌉ = ⌈p◦ k⌉ . So there is a unique isomorphism ω such that asrtp ◦ k ◦ asrtq = π⌈p ◦ k⌉ ◦ ω ◦ ζ⌈p ◦ g⌉ ◦ asrtp ◦ g. ..218 -- 219.. (3.14) 151 Next, we consider ζ⌈p ◦ πt⌉ ◦ ψ ◦ ζ⌈q⌉. Note ψ ◦ ζ⌈q⌉ is a quotient for a sharp predicate, hence sharp and so ⌈p◦ πt⌉◦ ψ ◦ ζ⌈q⌉ = ⌈p◦ πt ◦ ψ ◦ ζ⌈q⌉⌉ = ⌈p◦ k⌉ by 210 II. As quotients are closed under composition (197 IX), there is an iso β with (3.15) Finally, we deal with πs ◦ ϕ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉. By 211 XI this is again a compre- hension. We compute ζ⌈p ◦ πt⌉ ◦ ψ ◦ ζ⌈q⌉ = β ◦ ζ⌈p ◦ k⌉. im πs ◦ ϕ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ = (πs ◦ ϕ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉)⋄(1) 218 II= ζ⋄s (ϕ⋄(⌈t◦ asrtp ◦ π⌈p⌉⌉)) = ⌈t◦ asrtp ◦ π⌈p⌉⌉◦ ϕ−1 ◦ ζs 210 II= ⌈t◦ asrtp ◦ π⌈p⌉ ◦ ϕ−1 ◦ ζs⌉ = ⌈t◦ f†⌉ and so there must be a unique iso α with πs ◦ ϕ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ = π⌈t ◦ f †⌉ ◦ α. (3.16) III Lemma In setting II, we have f ◦ g = π⌈t ◦ f †⌉ ◦ α◦ χ◦ β ◦ ω ◦ ζ⌈p ◦ g⌉ ◦ asrtp ◦ g. IV Proof It's a long, but easy verification, either with a diagram • g ◦ p t r s a ◦ ⌉ g ◦ p ⌈ ζ ◦ ω • ❖❖❖❖❖❖ asrtq '❖❖❖❖❖❖ (3.14) asrtp ◦ k π⌈p ◦ k⌉ 7♦♦♦♦♦♦♦♦♦♦♦♦♦♦ ζ⌈p ◦ k⌉ g • • • ❖❖❖❖❖ ζ⌈p⌉ ◦ asrtp '❖❖❖❖❖ πt • • asrtp ◦ πt (3.13) • ζ⌈p ◦ πt⌉ / • f • χ ψ ◦ ζ⌈q⌉ 218 VII ψ ◦ ζ⌈q⌉ (3.15) β πs ◦ ϕ 7♦♦♦♦♦♦♦♦♦♦♦♦♦♦ _❅❅❅❅❅❅❅❅ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ (3.16) • α ◦ ⌉ † f ◦ t ⌈ π ❅❅❅❅❅❅❅❅ / • or with equational reasoning f ◦ g = πs ◦ ϕ◦ ζ⌈p⌉ ◦ asrtp ◦ πt ◦ ψ ◦ ζ⌈q⌉ ◦ asrtq (3.13)= πs ◦ ϕ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ χ◦ ζ⌈p ◦ πt⌉ ◦ asrtp ◦ πt ◦ ψ ◦ ζ⌈q⌉ ◦ asrtq 218 VII= πs ◦ ϕ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ χ◦ ζ⌈p ◦ πt⌉ ◦ ψ ◦ ζ⌈q⌉ ◦ asrtp ◦ k ◦ asrtq (3.15)= πs ◦ ϕ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ χ◦ β ◦ ζ⌈p ◦ k⌉ ◦ asrtp ◦ k ◦ asrtq = π⌈t ◦ f †⌉ ◦ α ◦ χ◦ β ◦ ζ⌈p ◦ k⌉ ◦ asrtp ◦ k ◦ asrtq (3.14)= π⌈t ◦ f †⌉ ◦ α ◦ χ◦ β ◦ ζ⌈p ◦ k⌉ ◦ π⌈p ◦ k⌉ ◦ ω ◦ ζ⌈p ◦ g⌉ ◦ asrtp ◦ g = π⌈t ◦ f †⌉ ◦ α ◦ χ◦ β ◦ ω ◦ ζ⌈p ◦ g⌉ ◦ asrtp ◦ g, (3.16) / / ' / / ' / / O O 7 / / 7 / / O O _ whichever the Reader might prefer. Corollary (f ◦ g)† = asrtp ◦ g ◦ π⌈p ◦ g⌉ ◦ ω−1 ◦ β−1 ◦ χ−1 ◦ α−1 ◦ ζ⌈t ◦ f †⌉ . To show (f ◦ g)† = g† ◦ f†, it is sufficient to proof that the 'daggered' version of each of the subdiagrams (3.16), (3.15), (3.13), (3.14) and 218 VII of the above diagram holds. We start with the simple ones. (cid:3) Lemma In setting II, the daggered version of (3.15) holds -- that is: V VI VII π⌈p ◦ k⌉ ◦ β−1 = π⌈q⌉ ◦ ψ−1 ◦ π⌈p ◦ πt⌉. Proof The map π⌈q⌉ ◦ ψ−1 ◦ π⌈p ◦ πt⌉ ◦ β is a comprehension for ⌈p◦ k⌉ -- indeed VIII (π⌈q⌉ ◦ ψ−1 ◦ π⌈p ◦ πt⌉ ◦ β)⋄(1) = (ζ⋄ = ⌈p◦ πt⌉◦ ζ⌈q⌉ ◦ ψ ⌈q⌉ ◦ ψ⋄)(⌈p◦ πt⌉) = (cid:6)p◦ πt ◦ ζ⌈q⌉ ◦ ψ(cid:7) = ⌈p◦ k⌉ . Furthermore ζ⌈p ◦ k⌉ ◦ π⌈q⌉ ◦ ψ−1 ◦ π⌈p ◦ πt⌉ ◦ β (3.15) = β−1 ◦ ζ⌈p ◦ πt⌉ ◦ ψ ◦ ζ⌈q⌉ ◦ π⌈q⌉ ◦ ψ−1 ◦ π⌈p ◦ πt⌉ ◦ β = id. So π⌈q⌉ ◦ ψ−1 ◦ π⌈p ◦ πt⌉ ◦ β is the unique comprehension corresponding to ζ⌈p ◦ k⌉ -- that is: π⌈q⌉ ◦ ψ−1 ◦ π⌈p ◦ πt⌉ ◦ β = π⌈p ◦ k⌉, as desired. (cid:3) Exercise Show that in the setting II, the daggered version of (3.16) holds -- i.e. α−1 ◦ ζ⌈t ◦ f †⌉ = ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs. (Hint: mimic the proof of VII.) Exercise Show that in the setting II, we have π⌈q⌉ ◦ ψ−1 ◦ asrtp ◦ πt = asrtp ◦ k ◦ π⌈q⌉ ◦ ψ−1. (This is the daggered version of the subdiagram marked 218 VII.) Proposition If in a †′-effectus, ν is the unique iso (cf. 216 XIII) such that asrta ◦ asrtb = π⌈a&b⌉ ◦ ν ◦ ζ⌈b&a⌉ ◦ asrtb&a, then asrtb ◦ asrta = asrtb&a ◦ π⌈b&a⌉ ◦ ν−1 ◦ ζ⌈a&b⌉. IX X XI ..219.. 153 XII Proof Let µ be the unique iso with asrtb ◦ asrta = π⌈b&a⌉ ◦ µ◦ ζ⌈a&b⌉ ◦ asrta&b. We will see µ = ν−1. For brevity, write a & b = (a & b)◦ π⌈a&b⌉ b & a = (b & a)◦ π⌈b&a⌉ By 218 IX and 211 XV, we have π⌈a&b⌉ ◦ asrta&b ◦ ζ⌈a&b⌉ = π⌈a&b⌉ ◦ ζ⌈a&b⌉ ◦ asrta&b = asrta&b. Now the second axiom of a †′-effectus comes into play π⌈a&b⌉ ◦ asrt2 a&b ◦ ζ⌈a&b⌉ a&b = π⌈a&b⌉ ◦ asrta&b ◦ ζ⌈a&b⌉ ◦ π⌈a&b⌉ ◦ asrta&b ◦ ζ⌈a&b⌉ = asrt2 = asrta ◦ asrt2 = π⌈a&b⌉ ◦ ν ◦ asrtb&a ◦ ζ⌈b&a⌉ ◦ π⌈b&a⌉ ◦ µ◦ asrta&b ◦ ζ⌈a&b⌉ = π⌈a&b⌉ ◦ ν ◦ asrtb&a ◦ µ◦ asrta&b ◦ ζ⌈a&b⌉. b ◦ asrta Thus as quotients are epis and comprehensions are monos: asrt2 a&b = ν ◦ asrtb&a ◦ µ◦ asrta&b. (3.17) We want to show asrta&b is an epi. First note (cid:6)a & b(cid:7)◦ ζ⌈a&b⌉ 210 II= (cid:6)a & b◦ ζ⌈a&b⌉(cid:7) 211 XV= ⌈a & b⌉ = 1 ◦ ζ⌈a&b⌉ and so im asrta&b =(cid:6)a & b(cid:7) = 1, which tells us asrta&b is a quotient and there- fore an epi. So, from (3.17) we get asrta&b = ν ◦ asrtb&a ◦ µ and so a & b = 1 ◦ asrta&b = 1 ◦ ν ◦ asrtb&a ◦ µ = b & a◦ µ. (3.18) Using this equation again in the previous, we find ν ◦ µ◦ asrta&b = ν ◦ µ◦ asrtb&a ◦ µ = ν ◦ asrtb&a ◦ µ = asrta&b. Thus ν ◦ µ = id and so µ = ν−1. Write l = π⌈b&a⌉ ◦ ν−1 ◦ ζ⌈a&b⌉. Then (a & b)◦ l† = (a & b)◦ π⌈a&b⌉ ◦ ν ◦ ζ⌈b&a⌉ = a & b ◦ ν ◦ ζ⌈b&a⌉ = b & a◦ ζ⌈b&a⌉ = b & a. by 218 IX by (3.18) and µ = ν−1 And so, keeping in mind a & b 6 ⌈a & b⌉ = 1 ◦ l, we have asrtb ◦ asrta = l ◦ asrta&b = asrt(a&b) ◦ l† ◦ l = asrtb&a ◦ l = asrtb&a ◦ π⌈b&a⌉ ◦ ν−1 ◦ ζ⌈a&b⌉, by 218 IX as promised. (cid:3) Corollary In setting II, the daggered version of (3.14) holds -- that is: XIII asrtq ◦ asrtp ◦ k = asrtp ◦ g ◦ π⌈p ◦ g⌉ ◦ ω−1 ◦ ζ⌈p ◦ k⌉. Lemma In setting II, the daggered version of (3.13) holds -- that is: ζt ◦ asrtp ◦ π⌈p⌉ = asrtp ◦ πt ◦ π⌈p ◦ πt⌉ ◦ χ−1 ◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉. Proof The heavy lifting has been done in XI already. To start, note XIV XV asrtp ◦ asrtt 211 XV= asrt⌈p⌉ ◦ asrtp ◦ asrtt = π⌈p⌉ ◦ ζ⌈p⌉ ◦ asrtp ◦ πt ◦ ζt (3.13)= π⌈p⌉ ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ χ◦ ζ⌈p ◦ πt⌉ ◦ asrtp ◦ πt ◦ ζt 218 VII= π⌈p⌉ ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ χ◦ ζ⌈p ◦ πt⌉ ◦ ζt ◦ asrtp ◦ πt ◦ ζt = π⌈p⌉ ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ χ◦ ζ⌈p ◦ πt⌉ ◦ ζt ◦ asrtt&p = π⌈p&t⌉ ◦ α2 ◦ χ◦ β2 ◦ ζ⌈t&p⌉ ◦ asrtt&p, where α2 and β2 are the unique isomorphisms such that π⌈p⌉ ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ = π⌈p&t⌉ ◦ α2 ζ⌈p ◦ πt⌉ ◦ ζt = β2 ◦ ζ⌈t&p⌉. With the same reasoning as in IX and VII, we see ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ζ⌈p⌉ = α−1 2 ◦ ζ⌈p&t⌉ πt ◦ π⌈p ◦ πt⌉ = π⌈t&p⌉ ◦ β−1 2 . Now we can apply XI: ζt ◦ asrtp ◦ π⌈p⌉ 2 ◦ χ−1 ◦ α−1 211 XV= ζt ◦ asrtt ◦ asrtp ◦ π⌈p⌉ XI= ζt ◦ asrtt&p ◦ π⌈t&p⌉ ◦ β−1 = ζt ◦ asrtt&p ◦ πt ◦ π⌈p ◦ πt⌉ ◦ χ−1 ◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ζ⌈p⌉ ◦ π⌈p⌉ = ζt ◦ asrtp ◦ πt ◦ ζt ◦ πt ◦ π⌈p ◦ πt⌉ ◦ χ−1 ◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ = asrtp ◦ πt ◦ π⌈p ◦ πt⌉ ◦ χ−1 ◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉, 2 ◦ ζ⌈p&t⌉ ◦ π⌈p⌉ as desired. ..219.. (cid:3) 155 XVI Proposition In a †′-effectus (f ◦ g)† = g† ◦ f† holds. XVII Proof We work in setting II. The equality (f ◦ g)† = g† ◦ f† follows from V and the commutativity of the following diagram • o ❖❖❖❖❖❖ asrtq ❖❖❖❖❖❖ g† • o ζt • o • π⌈q⌉ ◦ ψ−1 ❖❖❖❖❖ asrtp ◦ ζ⌈p⌉ ❖❖❖❖❖ f † • w XIII asrtp ◦ k X asrtp ◦ πt XIV ζ⌈p ◦ k⌉ ♦♦♦♦♦♦♦♦♦♦♦♦♦♦ π⌈p ◦ k⌉ π⌈q⌉ ◦ ψ−1 • o VII β−1 • o •O π⌈p ◦ πt⌉ • o 1 − ω ◦ ⌉ g ◦ p ⌈ π g ◦ p t r s a • w ϕ−1 ◦ ζs ♦♦♦♦♦♦♦♦♦♦♦♦♦♦ ❅❅❅❅❅❅❅❅ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ IX ❅❅❅❅❅❅❅❅ ⌉ † f ◦ t ⌈ ζ ◦ 1 − α • • χ−1 or alternatively by g† ◦ f† = asrtq ◦ π⌈q⌉ ◦ ψ−1 ◦ ζt ◦ asrtp ◦ π⌈p⌉ ◦ ϕ−1 ◦ ζs XIV= asrtq ◦ π⌈q⌉ ◦ ψ−1 ◦ asrtp ◦ πt ◦ π⌈p ◦ πt⌉ ◦ χ−1 ◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs X= asrtq ◦ asrtp ◦ k ◦ π⌈q⌉ ◦ ψ−1 ◦ π⌈p ◦ πt⌉ ◦ χ−1 ◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs VII= asrtq ◦ asrtp ◦ k ◦ π⌈p ◦ k⌉ ◦ β−1 ◦ χ−1 ◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs IX= asrtq ◦ asrtp ◦ k ◦ π⌈p ◦ k⌉ ◦ β−1 ◦ χ−1 ◦ α−1 ◦ ζ⌈t ◦ f †⌉ XIII= asrtp ◦ g ◦ π⌈p ◦ g⌉ ◦ ω−1 ◦ ζ⌈p ◦ k⌉ ◦ π⌈p ◦ k⌉ ◦ β−1 ◦ χ−1 ◦ α−1 ◦ ζ⌈t ◦ f †⌉ = asrtp ◦ g ◦ π⌈p ◦ g⌉ ◦ ω−1 ◦ β−1 ◦ χ−1 ◦ α−1 ◦ ζ⌈t ◦ f †⌉ V= (f ◦ g)†, whichever the Reader might prefer. (cid:3) 220 We are ready to finish the proof of 215 III. II Theorem A †′-effectus is a †-effectus with † as defined in 217 II. III IV Proof Let C be a †′-effectus. (Ax. 1) By 219 XVI, 218 XII and 217 III the † defined in 217 II turns Pure C into a †-category. Also by 217 III, we have asrt†p = asrtp for any predicate p. Pick any pure f . We have to show f is ⋄-adjoint to f†. By 212 III we have f = o g g O O o g g o O O O O w o O O O w o o πim f ◦ ϕ◦ ζ⌈1 ◦ f⌉ ◦ asrt1 ◦ f for some iso α. We compute f⋄ = (πim f )⋄ ◦ ϕ⋄ ◦(ζ⌈1 ◦ f⌉)⋄ ◦(asrt1 ◦ f )⋄ 218 II= (ζim f )⋄ ◦ ϕ⋄ ◦(π⌈1 ◦ f⌉)⋄ ◦(asrt1 ◦ f )⋄ 209 IV= (ζim f )⋄ ◦(ϕ−1)⋄ ◦(π⌈1 ◦ f⌉)⋄ ◦(asrt1 ◦ f )⋄ = (asrt1 ◦ f ◦ π⌈1 ◦ f⌉ ◦ ϕ−1 ◦ ζim f )⋄ = (f†)⋄, so f is indeed ⋄-adjoint to f†. (Ax. 2) Let f be a †-positive map. That is: f = h† ◦ h for some pure map h. By 212 III we have h = πim h ◦ α◦ ζ⌈1 ◦ h⌉ ◦ asrt1 ◦ h for some iso α. We compute V f = h† ◦ h = asrt1 ◦ h ◦ π⌈1 ◦ h⌉ ◦ α−1 ◦ ζim h ◦ πim h ◦ α◦ ζ⌈1 ◦ h⌉ ◦ asrt1 ◦ h = asrt1 ◦ h ◦ asrt⌈1 ◦ h⌉ ◦ asrt1 ◦ h 211 XV= asrt1 ◦ h ◦ asrt1 ◦ h. Form this it follows that †-positive maps are ⋄-positive. Let q be the predicate such that q & q = 1 ◦ h. Then asrt†q ◦ asrtq = asrtq ◦ asrtq 211 XIV= asrtq&q = asrt1 ◦ h. Thus asrt1 ◦ h is a †-positive with asrt1 ◦ h ◦ asrt1 ◦ h = f . We have to show asrt1 ◦ h is the unique map with this property. Let g be any †-positive map with g ◦ g = f . Recall both g and f are ⋄-positive, hence asrt1 ◦ f = f = g ◦ g = asrt1 ◦ g ◦ asrt1 ◦ g 211 XIV= asrt(1 ◦ g)&(1 ◦ g). So (1 ◦ g) & (1 ◦ g) = 1 ◦ f = (1 ◦ h) & (1 ◦ h). Hence 1 ◦ g = 1 ◦ h by uniqueness of the square root. So g = asrt1 ◦ g = asrt1 ◦ h = h, as desired. (Ax. 3) Let asrtp be any ⋄-positive map. Write q for the unique predicate with q & q = p. Then VI asrtp = asrtq&q 211 XIV= asrtq ◦ asrtq = asrt†q ◦ asrtq, so asrtp is †-positive, as desired. 3.7.1 Dilations (cid:3) Being the main topic of the first part of this thesis, we cannot finish our discus- sion of †-effectuses without mentioning dilations. The presented abstract theory of dilations in a †-effectus is preliminary and as of yet rather unimpressive. 221 ..219 -- 221.. 157 II Definition Let f : X → Y be any map in an effectus. A dilation of f is a triple (P, , h) of a sharp total map : P → Y and a pure map h : X → P such that ◦ h = f and the following universal property holds. For every triple (P ′, ′, h′) with ′ : P ′ → Y total sharp, h′ : X → P arbitrary and f = ′ ◦ h′, there is a unique σ : P → P ′ with σ ◦ h = h′ and ′ ◦ σ = . We say an effectus has dilations if there exists a dilation for every map. III Example The effectus vNop (180 V) has dilations, as shown in 154 III, but the full subcategory CvNop of commutative von Neumann algebras does not as shown in the next exercise. IIIa Exercise In this exercise you will show that CvNop does not have dilations. First, show that a corner between commutative von Neumann algebras is mul- tiplicative. Conclude that if CvNop were to have dilations, then any ncpu-map would be multiplicative, quod non. IV Proposition Let C be an effectus with dilations. 1. If (P1, 1, h1) and (P2, 2, h2) are both dilations of f , then there is a unique isomorphism α : P1 → P2 with α◦ h1 = h2 and 2 ◦ α = 1. 2. Conversely, if (P, , h) is a dilation of f and α : P → P ′ is some isomor- phism, then (P ′, ◦ α−1, α ◦ h) is also a dilation of f . 3. (X, , id) is the dilation of a sharp total map . 4. If (P, , h) is some dilation, then (P, id, h) is the dilation of h. 5. For any quotient ξ : X → Q and map f : Q → Y with dilation (P, , h), the triple (P, , h◦ ξ) is a dilation of f ◦ ξ. 6. Conversely, if (P, , h) is a dilation of f ◦ ξ for a map f : Q → Y and quotient ξ : X → Q, then (P, , h′′) is a dilation of f , where h′′ : Q → P is the unique map with h′′ ◦ ξ = h. 7. If for i = 1, 2 the triple (Pi, i, hi) is a dilation of fi : Xi → Y , then (P1 + P2, [1, 2], h1 + h2) is a dilation of [f1, f2] : X1 + X2 → Y . V Proof Point 1 has been proven in 140 VIII. Point 2 is easily verified. To prove point 3, assume is a sharp total map. Assume = ′ ◦ h′ for some h′ and sharp total ′. Then h′ is clearly the unique map satisfying h′ = h′ ◦ id (and = ′ ◦ h′), which demonstrates point 3. For point 4, assume (P, , h) is some dilation. Assume h = ′ ◦ h′ for some h′ and sharp total ′. There is a unique σ with h′ = σ ◦ h and ◦ ′ ◦ σ = . Now ′ ◦ σ ◦ h = ′ ◦ h′ = h. So ′ ◦ σ satisfies the properties of the unique mediating map id : P → P , hence ′ ◦ σ = id. Only uniqueness of σ remains, so assume σ′ is a ncp-map with h′ = σ′ ◦ h and ′ ◦ σ′ = id. Then ◦ ′ ◦ σ′ = and so σ = σ′. To demonstrate point 5, assume ′ ◦ h′ = f ◦ ξ for some h′ and total sharp ′. Then 1 ◦ h′ = 1 ◦ ′ ◦ h′ = 1 ◦ f ◦ ξ 6 1 ◦ ξ, so there is a unique h′′ with h′′ ◦ ξ = h′. By epicity of ξ, we get ′ ◦ h′′ = f from ′ ◦ h′′ ◦ ξ = ′ ◦ h′ = f ◦ ξ. Thus there is a unique σ with σ ◦ h = h′′ and ′ ◦ σ = and so σ ◦ h◦ ξ = h′′ ◦ ξ = h′. Only uniqueness of σ remains, so suppose σ′ is a map with σ′ ◦ h◦ ξ = h′ and ′ ◦ σ = . Then σ′ ◦ h = h′′ by epicity of ξ, hence σ′ = σ as desired. For point 6, assume f = ′ ◦ h′ for some h′ and sharp total ′. Then f ◦ ξ = ′ ◦ h′ ◦ ξ and so there is a unique σ with ◦ σ = ′ and σ ◦ h′ ◦ ξ = h ≡ h′′ ◦ ξ. Thus σ ◦ h′ = h′′ by epicity of ξ. To show uniqueness, assume σ′ is such that σ′ ◦ h′ = h′′ and ◦ σ′ = ′. Then σ′ ◦ h′ ◦ ξ = h′′ ◦ xi = h and in- deed σ′ = σ. To show point 7, assume [f1, f2] = ′ ◦ h′ for some h′ and total sharp ′. For any i = 1, 2 we have ′ ◦ h′ ◦ κi = fi, so there is a unique σi with σi ◦ hi = h′ ◦ κi and i = ′ ◦ σi. Define σ = [σ1, σ2]. Clearly σ ◦(h1 + h2) = [σ1 ◦ h1, σ2 ◦ h2] = [h′ ◦ κ1, h′ ◦ κ2] = h′ and ′ ◦ σ = [′ ◦ σ1, ′ ◦ σ2] = [1, 2]. Assume σ′ is any ncp-map with σ′ ◦(h1 + h2) = h′ and ′ ◦ σ′ = [1, 2]. For i = 1, 2 we have ′ ◦ σ′ ◦ κi = i and h′ ◦ κi = σ′ ◦(h1 + h2)◦ κi = σ′ ◦ hi, so σ′ ◦ κi = σi. Hence σ′ = [σ′ ◦ κ1, σ′ ◦ κ2] = [σ1, σ2] = σ. (cid:3) 2 , 1 With some additional assumptions, the existence of dilations forces the existence of some familiar quantum gates. Assume C is a †-effectus with dilation with a scalar λ such that λ⊥ = λ. (For instance: λ = 1 2 in vNop.) Let (P,hs, s⊥i, h) be a dilation of the map(cid:10)h, h⊥(cid:11) : 1 → 1 + 1. We think of P as an abstract qubit 2(cid:11) is of the form (M2,he, e⊥i, h) where h is -- indeed in vNop any dilation of(cid:10) 1 the vector state for some v ∈ C2 and e is a projector on wC, with hv, wi = 1√2 (for instance v = 0i and w = +i). By assumption hs, s⊥i◦ h = hλ, λi, so s◦ h = s⊥ ◦ h = λ. and hs⊥, si◦ h = hλ, λi. Hence by the definition of dilation, there is a unique map X : P → P with X ◦ h = h and hs⊥, si◦ X = hs, s⊥i, whence s◦ X = s⊥ ◦ X and s⊥ ◦ X = s◦ X. Clearly X ◦ X ◦ h = h and hs, s⊥i◦ X ◦ X = hs, s⊥i, so by uniqueness of the mediating map id : P → P , we must have X ◦ X = id. Thus the iso- morphism X behaves like the X-gate -- indeed, in vNop, we have X = adX if e = 0ih0. For the next gate, assume s is pure and im s = 1. Note s† is total. As asrtλ = λ, we have h† ◦ s† = λ† = λ and so (h†)⊥ ◦ s† = λ⊥ = λ. Combined: hh†, (h†)⊥i◦ s† = hλ, λi. Thus using the universal property of the dilation, there is a unique map H : P → P with H ◦ h = s† and hh†, (h†)⊥i◦ H = hs, s⊥i. So h† ◦ H = s. In general, we do not know a lot more about H. Assume H is an comprehen- 222 II III ..221 -- 222.. 159 sion (as it is in vNop in this situation). Then s† = H† ◦ h and s = h† ◦ H†, so s⊥ = (h†)⊥ ◦ H†, whence hs, s⊥i = hh†, (h†)⊥i◦ H†. Thus, by unique- ness of H, we must have H† = H. Thus both H† and H are total, which forces H ◦ H† = H† ◦ H = asrt1 = id, by uniqueness of ⋄-positive maps. Hence H ◦ H = id, s◦ H = h† and h† ◦ H = s. So H is similar to the Hadamard- gate -- indeed, in vNop, we have H = adH if e = 0ih0. IV Do our additional assumptions (purity of s, im s = 1, etc.) follow from more general principles? Can we define the other gates, like the CNOT? Do these gates interact in the usual way? Unfortunately, we do not have an answer to any of these questions. 223 In 157 IV we saw that for any ncp-map ϕ with Paschke dilation (P, , h), there is a linear order isomorphism [0, 1](A )(cid:3) ∼= [0, ϕ]ncp. To finish the discussion of dilations, we mention how to state this correspondence in an &-effectus. II Definition In an &-effectus, for any predicate p, define seffp ≡ asrtp > asrtp⊥ . This map models the side-effect of measuring p. For any map f : X → Y , define Inv f ≡ {p; p ∈ Pred X; f ◦ seffp = f}. III This is the set of predicates whose measurement does not disturb f . Lemma In vNop, for any nmiu-map : A → B, we have Inv = [0, 1](A )(cid:3) -- that is: Inv is the set of effects of the commutator of the image of . IV Proof Assume a ∈ [0, 1]f (A )(cid:3). Simply unfolding definitions, we find seff a((b)) = √a(b)√a + √1 − a(b)√1 − a = (b) for any b ∈ A , so a ∈ Inv . For the con- verse, assume a ∈ Inv . Recall (A , , id) is a Paschke dilation of . As = seff a ◦ = (asrta ◦ ) > (asrt1−a ◦ ) >ncp asrta ◦ , we know by 157 IV that there is a t ∈ [0, 1](A )(cid:3) with asrta((b)) ≡ √a(b)√a = t(b) for all b ∈ A , but then clearly a = √a(1)√a = t(1) = t. V Definition Let C be an &-effectus. For a map f : X → Y , write (cid:3) ↓ f ≡ {g; g : X → Y ; g 6 f}. We say a dilation (P, , h) of f has the order correspondence if there is an order isomorphism Θ : ↓ f → Inv such that for every g ∈↓ f , we have g = ◦ asrtΘ(g) ◦ h. VI Example By 157 IV every dilation in vNop has the order correspondence. 3.8 Comparisons 3.8.1 Dagger kernel categories To finish this chapter, we relate the structures in a †-effectus with existing structures in the literature. Just like we try to axiomatize vNop, Heunen in the third chapter of his Ph.D-thesis attempted to axiomatize Hilb, the category of Hilbert spaces with bounded operators. Heunen came very close: he proves that a †-category obeying certain additional axioms must embed into Hilb. [49, 3.7.18] We will discover that some of his axioms for Hilb hold in Pure(C) for a †-effectus C and others do not in general. (We do not cover all his axioms.) Definition Let C be a †-category. We say that an arrow f is †-mono iff f† ◦ f = id and dually f is †-epi iff f ◦ f† = id. An arrow f is a †-partial isometry if f = m◦ e for some †-mono m and †-epi e. A †-kernel of f is an equalizer of f with 0, which is †-mono. A †-kernel category is a †-category with zero object and where each arrow has a †-kernel [49, 3.2.20]. Proposition In Pure(C) for some †-effectus C, the following holds. 1. f is †-mono iff f is a comprehension. Dually: f is †-epi iff f is a quotient of a sharp predicate. 2. The †-partial isometries are exactly the pristine maps, see 218 IV. Furthermore: Pure C is a †-kernel category: the †-kernel of f is given by π(1 ◦ f )⊥. Proof Let f be any †-mono pure map, say f ≡ πs ◦ α◦ ζ⌈p⌉ ◦ asrtp for some iso α, s ≡ im f and p ≡ 1 ◦ f . As we saw before (using 211 XV), we see id = f† ◦ f = asrtp2 . Thus p & p = p2 = 1 = 1 & 1 and so by uniqueness of the square root we get p = 1. Hence f is a comprehension. Dualizing, we see †-epis are exactly quotients of sharp predicates. Now it is clear that the †-partial isometries are exactly the pristine maps. By the previous π(1 ◦ f )⊥ is †-mono. To see it is the †-kernel of f , we have to show it is the equalizer of f with 0. Clearly (1 ◦ f )⊥ ◦ π(1 ◦ f )⊥ = 1 and so 1 ◦ f ◦ π(1 ◦ f )⊥ = 0, whence f ◦ π(1 ◦ f )⊥ = 0. Assume f ◦ g = 0 for some pure map g. Then 1 ◦ f ◦ g = 0, so 1 ◦ f 6 im⊥ g, hence (1 ◦ f )⊥ > im g, so g = π(1 ◦ f )⊥ ◦ g′ for a unique g′. (cid:3) Exercise* Let C be a †-effectus. In this exercise, you will show that in general Pure C does not have finite coproducts. In particular, Pure C does not have †-biproducts, see [49, 3.2.15]. Consider C ≡ vNop. Reasoning to- π2−→ C of C wards contradiction, assume (Pure vNop)op has a product C π1←− A and C. First show that for any non-zero pure map f : A → C, there ex- ists a Hilbert space H , an element x ∈ H , a von Neumann algebra C and ..222 -- 224.. 161 224 II III IV V VI isomorphism ϕ : B(H ) ⊕ C → A with f (varphi(T, c)) = hx, T xi. Use this on π1, π2 and hidC, idCi to show there is an isomorphism ϕ : B(H ) ⊕ C ∼= A with π1 ◦ ϕ = hx, (· )xi and π2 ◦ ϕ = hy, (· )yi for some von Neumann alge- bra C , Hilbert space H and x, y ∈ H . Press on: show C is trivial. Then prove dim H 6 2. Almost there: show dim H > 2. Derive a contradiction. VII Exercise* In this exercise you will show that Pure(vNop) does not have all It is helpful to first consider two concrete coequalizers in Hilb coequalizers. and vNop. (These coequalizers capture unordered pairs of qubits.) Write σ : C2 ⊗ C2 → C2 ⊗ C2 for the unitary fixed by σ iji = jii. The equalizer of σ with id exists in Hilb and is given by eS , the inclusion of the subspace S spanned by {00i ,11i ,01i + 10i} into C2 ⊗ C2. (S is called the symmetric tensor of C2 with itself.) The coequalizer of σ with id is simply e†S . There is also an equalizer of adσ : M4 → M4 with id in vN, but curiously enough, it is given by e : M3 ⊕ C → M4, where e(a, λ) = ade† (λ) and A is the subspace spanned by 01i − 10i. For the proof and similar results, see [35]. Assume ξ : C → M4 is some filter (i.e. quotient in the opposite category) with adσ ◦ ξ = ξ. Show that for each b ∈ C , we have ξ(b)σ = σξ(b). From this, derive (perhaps in the same way as [35]), that ξ(b) = pA ξ(b)pA + pS ξ(b)pS for all b ∈ C , where pA = eA e†A and pS = eS e†S . Prove that either ξ(b) 6 pA for all b ∈ C or ξ(b) 6 pS for all b ∈ C . Conclude Pure(vNop) does not have all coequalizers. (a) + ade† S A VIII To summarize: the category Pure(C) of a †-effectus C is a †-category, has †- kernels and each †-mono is a †-kernel, but in general it does not have †-biproducts or †-equalizers, which shows it is not, in general, a pre-Hilbert category as in [49, 3.7.1]. VIIIa Recently Tull axiomatized [115] the †-category Hilb∼ of Hilbert spaces with bounded operators between them modulo global phase (i.e. T = S in Hilb∼ iff T = λS for some λ ∈ C,λ = 1.) Write Hilbc ∼ for the restruction of Hilb∼ to contractive operators. This category Hilbc ∼ is a full †-subcategory of Pure(vNop) via the functor H 7→ B(H ), (cid:0)T : H → K(cid:1) 7→ (cid:0)adT : B(K ) → B(H )(cid:1). The category Hilbc ∼ (and in fact Hilb∼) do not have all (co)products. Tull shows that Hilb∼ does have phased biproducts. [114, 115] This raises the question: do the phased biproducts of Hilb∼ extend to Pure(vNop). If this is the case, then a phased product of C with itself in Pure(vNop) is given by: 0ih0 ,1ih1 : C → M2. Unfortunately this leads to a contradiction. It hasn't been determined yet whether Pure(vNop) has phased products at all. 3.8.2 Sequential product The sequential product -- the operation a & b ≡ √ab√a on the effects of a von Neumann algebra -- has been studied before. In [37, 47] Gudder and coauthors establish some basic algebraic properties of & on B(H ). Most of them are surprisingly hard to prove -- for instance, the proof that a&b = b&a implies ab = ba, requires the Fuglede -- Rosenblum -- Putnam Theorem. Then, in [46], Gudder and Latr´emoli`ere (G&L) characterize & on B(H ) as the unique operation satisfying (for effects a, b and density matrix ) 225 II 1. tr[(a & )b] = tr[(a & b)] 3. a & (a & b) = (a & a) & b = a2 & b 2. a & 1 = 1 & a = a 4. a 7→ a & b is strongly continuous In 105 V we characterized asrta : b 7→ a& b as the unique ⋄-positive map with a& 1 = a. How do these characterizations compare? G&L's first axiom plays a very similar role as the ⋄-self adjointness of ⋄-positive maps for us. Their third axiom is somewhat related to ⋄-positivity. Their fourth axiom seems unrelated to our characterization and conversely the purity of our ⋄-positive maps has no counterpart in their axioms. A few years earlier, Gudder and Greechie started [47] the abstract study of some of the algebraic properties of the sequential product on arbitrary effect algebras, which has been picked up by several other authors [6, 44, 45, 71, 72, 83, 104, 111, 119, 120, 122]. III Definition A sequential effect algebra (SEA) with a binary operation & satisfying is an effect algebra E together IV (S1) if a ⊥ b, then c & a ⊥ c & b and (c & a) > (c & b) = c & (a > b); (S2) 1 & a = a; (S3) if a & b = 0, then a & b = b & a; (S4) if a & b = b & a, then a & b⊥ = b⊥ & a and (a & b) & c = a & (b & c) and (S5) if c & a = a & c, c & b = b & c and a ⊥ b, then c & (a & b) = (a & b) & c and c & (a > b) = (a > b) & c. Examples The effect algebra [0, 1]A of effects on a von Neumann algebra A is a sequential effect algebra with a& b = √ab√a. Any commutative effect monoid is a sequential effect algebra with a & b = a ⊙ b. Proposition In a †-effectus, the set of predicates Pred X on any object X with p & q = q ◦ asrtp satisfies axioms (S1), (S2) and (S3) of a SEA. V VI ..224 -- 225.. 163 VII Proof The proofs of (S1) and (S2) are obvious. To show (S3), assume p & q = 0 for some predicates p, q. Then 1 ◦ asrtq ◦ asrtp = p & q = 0 and so asrtq ◦ asrtp = 0 = asrt0. Applying the dagger, we find 0 = asrt†0 = asrt†p ◦ asrt†q = asrtp ◦ asrtq. Thus q & p = 0 = p & q, as desired. (cid:3) VIII Remark It's unclear whether the set of predicates in a †-effectus forms a sequen- tial effect algebra. A seemingly related open problem is whether p & q = q & p implies that asrtp ◦ asrtq = asrtq ◦ asrtp. 3.8.3 Homological categories II 226 Next we compare our effectuses to Grandis' homological categories [41], which generalize abelian categories used in homological algebra. First, we need a lemma. Lemma Suppose s, t are any sharp predicates on the same object in a †-effectus. If s⊥ 6 t, then s & t is sharp. Proof By 208 III, it is sufficient to show s & t⊥ is sharp as (s & t)⊥ = s & t⊥ > s⊥. Note s⊥ ⊥ t⊥, so by 213 III, we find t⊥ & s⊥ = 0, hence t⊥ & s = t⊥ is sharp. By the same reasoning as in the final part of the proof of 219 XI, there exists a (unique) pristine map l with asrts ◦ asrtt⊥ = l ◦ asrtt⊥&s, l ◦ l† = asrt⌈t⊥&s⌉ and l† ◦ l = asrt⌈s&t⊥⌉. It follows l ◦ asrtt⊥&s ◦ l† = asrts&t⊥ and so III asrt2 s&t⊥ = l ◦ asrtt⊥&s ◦ l† ◦ l ◦ asrtt⊥&s ◦ l† = l ◦ asrtt⊥&s ◦ asrt⌈s&t⊥⌉ ◦ asrtt⊥&s ◦ l† 211 XV= l ◦ asrtt⊥&s ◦ asrtt⊥&s ◦ l† 211 V= l ◦ asrtt⊥&s ◦ l† = asrts&t⊥, which shows s & t⊥ is sharp. (cid:3) IV Definition A category C is a pointed semiexact category [41, §1.1] if C has a zero object and all kernels and cokernels. In a pointed semiexact category: 1. We put a partial order on kernels in the usual way: we write n 6 m for kernels n : A → X and m : B → X, if there is an f : A → B with n = m◦ f . If both n 6 m and m 6 n, then we write n ≈ m. We denote the poset of kernels on A modulo ≈ by Nsb A. (Normal subobjects, [41, §1.5].) 2. For every map f there is a unique map g with f = (ker cok f )◦ g ◦(cok ker f ). If this g is an iso, then f is called exact. A pointed semiexact category C is a pointed homological category if 1. kernels are closed under composition; 2. cokernels are closed under composition and 3. (homology axiom) for any kernel m : M → A and cokernel q : A → Q with ker q 6 m, the composition q ◦ m is exact. Theorem Any †-effectus (in partial form) is a (pointed) homological category: V 1. a map is a kernel iff it is a comprehension; 2. a map is a cokernel iff it is a quotient of a sharp predicate and 3. a map is exact iff it is pristine. Proof Let C be a †-effectus. By 200 III and 205 II, we know C has all kernels and cokernels: a kernel of f is exactly a comprehension of (1 ◦ f )⊥ and a cokernel of f is exactly a quotient of im f . By 211 XI, comprehensions are closed under composition, and so kernels are closed under composition as well. As for the cokernels, assume ζ1, ζ2 are two composable cokernels. Both ζ1 and ζ2 are quotients of a sharp predicate. By 197 IX the composition ζ1 ◦ ζ2 is a quotient as well. The map ζ2 is sharp by 197 IX and so the predicate 1 ◦ ζ1 ◦ ζ2 is sharp. Hence ζ1 ◦ ζ2 is a cokernel too. Unfolding definitions, it is easy to see exact maps correspond precisely to pristine maps. (Homology axiom) We have to show q ◦ m is exact for any cokernel q and kernel m with ker q 6 m. As q ◦ m is pure, it is sufficient to show 1 ◦ q ◦ m is sharp. By 204 I, the assumption ker q 6 m is equivalent to (1 ◦ q)⊥ ≡ im ker q 6 im m and so im⊥ m 6 1 ◦ q. Our lemma II shows (im m) & (1 ◦ q) ≡ 1 ◦ q ◦ m◦ m† is sharp and so VI VII ⌈1 ◦ q ◦ m⌉ = ⌈1 ◦ q ◦ m⌉◦ m† ◦ m = (cid:6)1 ◦ q ◦ m◦ m†(cid:7)◦ m = 1 ◦ q ◦ m◦ m† ◦ m = 1 ◦ q ◦ m, as comprehensions are †-mono as m† is sharp which show q ◦ m is sharp, as desired. In any homological category, a generalization of the famous Snake Lemma holds. Before we can discuss it, we need some more homological category theory. (cid:3) Definition Let a (pointed) semiexact category be given. 227 II 1. We say the diagram A −→ C is exact at B, if ker cok(f ) ≈ ker(g). f1−→ A2 → ··· → An is a (long) exact sequence if it is exact −→ B f g We say A1 at A2, . . . , An−1. ..225 -- 227.. 165 2. The poset Nsb A of kernels modulo ≈ is a bounded lattice with minimum 0 and maximum 1 ≡ id. [41, §1.5] 3. For any f : A → B, define f∗ : Nsb A ⇆ Nsb B : f∗ by f∗(k) = ker cok(f ◦ k) and f∗(k) = ker((cok k)◦ f ). 4. We say f is left-modular at k if f∗(f∗(k)) = k ∨ f∗(0) and f is right- modular at k if f∗(f∗(k)) = k ∧ f∗(1). We call f simply left-modular (resp. right-modular) if it is left-modular (resp. left-modular) at every k. We say f is modular if it is both left- and right-modular. III Example In a †-effectus, the previous concepts are related to our familiar notions as follows. g f −→ B 1. A diagram A −→ C is exact at B if and only if im⊥ f = ⌈1 ◦ f⌉. 2. The lattice Nsb A is isomorphic to the lattice SPred A via k 7→ im k. 3. For any f , we have im f∗(k) = f⋄(im k) and im f∗(k) = f (im k). 4. A map f is left-modular at k iff f (f⋄(im k)) = (im k) ∨ ⌈1 ◦ f⌉ and right- modular at k iff f⋄(f (im k)) = (im k) ∧ im f . V IV Notation As Nsb A and SPred A are isomorphic, we will abbreviate "f is left- modular at πs" by "f is left-modular at s" and similarly for right-modularity. Lemma In a †-effectus, we have π ◦ π⋄ = id and ζ⋄ ◦ ζ = id for any compre- hension π and sharp quotient ζ. VI Proof We start with π. The trick is to use π† is sharp: π ◦ π⋄(s) = π ◦(π†)⋄(s) = ⌈⌈s◦ π†⌉⊥ ◦ π⌉⊥ = ⌈(s◦ π†)⊥ ◦ π⌉⊥ = ⌈(s◦ π† ◦ π)⊥⌉⊥ = ⌈s⊥⌉⊥ = s. As for the other equality: ζ⋄ ◦ ζ (s) = (ζ†)⋄ ◦ ζ (s) = ⌈⌈s⊥ ◦ ζ⌉⊥ ◦ ζ†⌉ = ⌈(⌈s⊥ ◦ ζ⌉◦ ζ†)⊥⌉ = ⌈(⌈s⊥⌉◦ ζ ◦ ζ†)⊥⌉ = ⌈⌈s⊥⌉⊥⌉ = s. (cid:3) 228 The following is a translation of (a part of) Grandis' Snake Lemma [41, §3.4] into †-effectus jargon. We include a proof as it highlights how the reasoning in a homological category is subtly different from the reasoning we already saw in a †-effectus: maps l appear, where we would have expected l⋄. Snake Lemma Suppose we have a commuting diagram in †-effectus as follows. II A a 0 0 / A′ f h / B b / B′ g k / C 0 / 0 c / C′ Furthermore, assume / / / / / / 1. the diagram has exact rows; 2. b is left-modular over im f ; 3. b is right-modular over im h; 4. f is right-modular over ⌈1 ◦ b⌉⊥ and 5. k is left-modular over im b. These additional assumptions are equivalent to the following conditions. 1. b ◦ b⋄(im f ) = ⌈1 ◦ b⌉⊥ ∨ im f , 2. b⋄ ◦ b (im h) = (im h) ∧ im b, 3. k ◦ k⋄(im b) = (im h) ∨ im b, 4. f⋄ ◦ f ◦ b (0) = ⌈1 ◦ b⌉⊥ ∧ im f , 5. im⊥ f = ⌈1 ◦ g⌉, 6. im⊥ h = ⌈1 ◦ k⌉, 7. g is a sharp quotient and 8. h is a comprehension. Completing the diagram, write aπ, bπ, cπ for kernels of a, b and c, respectively; aζ, bζ, cζ for the cokernels and f , g, h, k for the induced maps between the kernels and cokernels. It is easy to see g = c†π ◦ g ◦ bπ k = cζ ◦ k ◦ b†ζ. f = b†π ◦ f ◦ aπ h = bζ ◦ h◦ a†ζ Then: there is a map d : ker c → cok a that turns†† ker a f / ker b g / ker c d / cok a h / cok b k / cok c (3.19) into a (long) exact sequence. This reveals the snake: d 0 0 ker a aπ A a A′ aζ cok a f f h h ker b bπ B b B′ bζ cok b g g k k ker c cπ C c C′ cζ cok c 0 0 ††Beware: we use ker a (and cok) inconsistently -- earlier in the text ker f refers to a kernel map -- here it refers to a kernel object instead. ..227 -- 228.. 167 / / / / / III Proof Before we start, we give an overview of the proof. We will first show there are comprehensions m, h′ and sharp quotients g′, v such that the left and right faces of the following cube commute. m =③③③③③③③③ B g b A′- h <②②②②②②②②② B′ v Q Z. g′ =④④④④④④④④ cπ C d ker c / cok a aζ <②②②②②②②② h′ (These two faces are, what Grandis calls, subquotients.) Then we will prove using exactness of the rows that there is a unique lifting of b to b′ : Z → A′ along the comprehensions m, h. In turn, d is defined as the unique lifting of b′ along the sharp quotients h′, aζ. (d is the map regularly induced by b along the two subquotients.) Before demonstrating the exactness of the snake, we need some non-trivial identities for d⋄ and d . (These also follow from Grandis' calculus of direct and inverse images along an induced morphism.) IV Define v ≡ ξh⋄(im a). As (v ◦ h)⋄(1) = h⋄(h⋄(im a)) = im a = 1 ◦ aζ, there must be a total pure map h′ such that v ◦ h = h′ ◦ aζ. Note ⌈1 ◦ v⌉⊥ = h⋄(im a) 6 h⋄(1) = im h. So by the homology axiom, we know v ◦ h is pris- tine and so h′ must be a comprehension. Moving to the other side, define m ≡ πg (⌈1 ◦ c⌉⊥). Reasoning in a similar way, we see there is a unique sharp quo- tient g′ with g ◦ m = cπ ◦ g′. hensions) to show (b ◦ m)⋄(1) ≡ im b ◦ m 6 im h ≡ h⋄(1). by dfn. m V To show b lifts along m and h, it suffices (by the universal property of compre- (b⋄ ◦ m⋄)(1) = (b⋄ ◦ g ◦ c )(0) = (b⋄ ◦ b ◦ k )(0) = (b⋄ ◦ b ◦ h⋄)(1) 6 h⋄(1) by exactness at B′ as b⋄ ⊣ b . So there is a unique b′ : Z → A′ with h◦ b′ = b ◦ m. Before we continue, we note b′ = h† ◦ b ◦ m (as h† ◦ h = id) and so surprisingly: m ◦ b ◦ h = b′ 227 V= b′ ◦ h ◦ h⋄ = m ◦ b ◦ h⋄. (3.20) / / = < . = / - < Next, to show b′ lifts along g′ and aζ, it is sufficient to prove (b′ ◦ aζ )(0) = ⌈1 ◦ b′ ◦ aζ⌉⊥ > ⌈1 ◦ g′⌉⊥ = g′ (0). (b′ ◦ aζ)(0) = (b′ ◦ a⋄)(1) = (m ◦ b ◦ h⋄ ◦ a⋄)(1) = (m ◦ b ◦ b⋄ ◦ f⋄)(1) > (m ◦ f⋄)(1) = (m ◦ g )(0) = (g′ ◦(cπ) )(0) > g′ (0). by dfn. aζ by (3.20) as b⋄ ⊣ b by exactness at B by dfn. g′ Thus there is a unique d : ker c → cok a with d◦ g′ = aζ ◦ b′. More concretely, using g′ ◦ g′† = id, we see d = aζ ◦ h† ◦ b ◦ m◦ g′†. Our next goal is to show VI d⋄ = (aζ )⋄ ◦ h ◦ b⋄ ◦ m⋄ ◦ g′ = h′ ◦ v⋄ ◦ b⋄ ◦ m⋄ ◦ g′ = (aζ )⋄ ◦ h ◦ b⋄ ◦ g ◦(cπ)⋄ = h′ ◦ v⋄ ◦ b⋄ ◦ g ◦(cπ)⋄. (3.21) As a first step, note im m = (g ◦ c⋄)(1) > (g ◦(cπ)⋄)(s) and so (g ◦(cπ)⋄)(s) = (g ◦(cπ)⋄(s)) ∧ (im m) = (m⋄ ◦ m ◦ g ◦(cπ)⋄)(s) = (m⋄ ◦ g′ ◦ cπ ◦(cπ)⋄)(s) = (m⋄ ◦ g′ )(s) by 208 IX by dfn. g′ by 227 V. and so we only have to show the first two equalities of (3.21). The first is easy: d⋄ = d⋄ ◦ g′ ⋄ ◦ g′ = (aζ )⋄ ◦ b′ ⋄ ◦ g′ = (aζ )⋄ ◦ h ◦ h⋄ ◦ b′ = (aζ )⋄ ◦ h ◦ b⋄ ◦ m⋄ ◦ g′ ⋄ ◦ g′ by 227 V by dfn. d by 227 V by dfn. b′. The second is trickier. We need some preparation. Recall v (0) 6 h⋄(1) and so h⋄(1) = v (0) ∨ h⋄(1)208 XII= (v ◦ v⋄ ◦ h⋄)(1) = (v ◦ h′ ⋄)(1). ..228.. (3.22) 169 As h⋄ ◦ aζ is injective, the last equality follows from (h⋄ ◦ aζ ◦(aζ )⋄ ◦ h ◦ b⋄ ◦ m⋄ ◦ g′ )(s) = (h⋄(h ◦ b⋄ ◦ m⋄ ◦ g′ )(s) ∨ aζ(0)) = (h⋄ ◦ h ◦ b⋄ ◦ m⋄ ◦ g′ )(s) ∨ (h⋄ ◦ aζ)(0) = (h⋄ ◦ h ◦ b⋄ ◦ m⋄ ◦ g′ )(s) ∨ v (0) = ( (b⋄ ◦ m⋄ ◦ g′ )(s) ∧ h⋄(1) ) ∨ v (0) = (b⋄ ◦ m⋄ ◦ g′ )(s) ∨ v (0) = ( (b⋄ ◦ m⋄ ◦ g′ )(s) ∨ v (0) ) ∧ h⋄(1) = (v ◦ v⋄ ◦ b⋄ ◦ m⋄ ◦ g′ )(s) ∧ h⋄(1) = v ◦ v⋄ ◦ b⋄ ◦ m⋄ ◦ g′ (s) ∧ v ◦ h′ ⋄(1) = v ( (v⋄ ◦ b⋄ ◦ m⋄ ◦ g′ )(s) ∧ h′ ⋄(1) ) = (v ◦ h′ ⋄ ◦ h′ ◦ v⋄ ◦ b⋄ ◦ m⋄ ◦ g′ )(s) = (h⋄ ◦ aζ ◦ h′ ◦ v⋄ ◦ b⋄ ◦ m⋄ ◦ g′ )(s), by 208 XII as h⋄ ⊣ h by dfns. v,aζ by 208 IX as b⋄ ◦ m⋄(1) 6 h⋄(1) as v (0) 6 h⋄(1) by 208 XII by (3.22) as v ⊢ v⋄ by 208 IX where h⋄ ◦ aζ ◦ = v ◦ h′ ⋄ is proven in the same way as g ◦(cπ)⋄ = m⋄ ◦ g′ . VII We are ready to show exactness of (3.19). We will first derive exactness in cok a from the right-modularity of b over im h. d⋄(1) = ((aζ)⋄ ◦ h ◦ b⋄ ◦ g ◦(cπ)⋄)(1) = ((aζ)⋄ ◦ h ◦ b⋄ ◦ g ◦ c )(0) = ((aζ)⋄ ◦ h ◦ b⋄ ◦ b ◦ k )(0) = ((aζ)⋄ ◦ h ◦ b⋄ ◦ b ◦ h⋄)(1) = ((aζ)⋄ ◦ h )(b⋄(1) ∧ h⋄(1)) = ((aζ)⋄ ◦ h ◦ h⋄ ◦ h ◦ b⋄)(1) = ((aζ)⋄ ◦ h ◦ b⋄)(1) = ((aζ)⋄ ◦ h ◦ bζ)(0) = ((aζ)⋄ ◦ aζ ◦ h )(0) = h (0) by (3.21) by dfn. cπ by exactness in B′ by right-modularity by 208 IX as h ⊢ h⋄ by dfn. bζ by dfn. h by 227 V. By a dual argument one derives exactness of (3.19) in ker c from the left- modularity of b over im f . To show exactness of (3.19) in cok b, we will use left-modularity of k in im b: VIII h⋄(1) = (h⋄ ◦(aζ)⋄ ◦ aζ)(1) = (h⋄ ◦(aζ)⋄)(1) = ((bζ )⋄ ◦ h⋄)(1) = ((bζ )⋄ ◦ bζ ◦(bζ)⋄ ◦ h⋄)(1) = (bζ)⋄(h⋄(1) ∨ bζ(0)) = (bζ)⋄(h⋄(1) ∨ b⋄(1)) = (bζ)⋄(k (0) ∨ b⋄(1)) = ((bζ )⋄ ◦ k ◦ k⋄ ◦ b⋄)(1) = ((bζ )⋄ ◦ k ◦ c⋄ ◦ g⋄)(1) = ((bζ )⋄ ◦ k ◦ c⋄)(1) = ((bζ )⋄ ◦ k ◦ cζ)(0) = ((bζ )⋄ ◦ bζ ◦ k )(0) = k (0) by 227 V as aζ ⊣ (aζ )⋄ by dfn. h as (bζ)⋄ ⊣ bζ by 208 XII by dfn. bζ exactness in B′ by left-modularity as quotients are faithful by dfn. cζ by dfn. k by 227 V. Finally, the exactness of (3.19) in ker b is shown with a dual argument using the right-modularity of f over ⌈1 ◦ b⌉⊥. (cid:3) Remark As vNop is a †-effectus, Grandis' Snake Lemma also holds for von Neumann algebras. This is somewhat puzzling for the following reason. To study spaces, one associates in algebraic geometry to each space an object (or even a diagram of objects) of an abelian category representing some algebraic invariant. The key point is that it is easier to compute and understand these algebraic invariants than it is to work with the spaces themselves. One normally thinks of a category of spaces as being quite different from a category of algebraic invariants, hence it's surprising that vNop, which is a rather complicated category of (non-commutative) spaces, behaves in this way, similar to an abelian category. IX ..228 171 Chapter 4 Outlook For my closing remarks, I would like to speculate on possible applications and suggest possible directions for future research. 229 We started this thesis by asking the question whether the incredibly useful Stinespring dilation theorem extends to any ncp-map between von Neumann algebras. Now that we have seen it does, the next question is obvious: do the applications that make the Stinespring dilation so useful also extend to the Paschke dilation? I want to draw attention to one application in particular: proving security bounds on quantum protocols. Here one uses a continuous version of the Stinespring dilation theorem [80]. Is there also such a continuous version for Paschke dilation? Next, we have seen that self-dual Hilbert C∗-modules over von Neumann alge- bras are very well-behaved compared to arbitrary Hilbert C∗-modules. Frankly, I'm surprised they haven't been studied more extensively before: the way re- sults about Hilbert spaces generalize elegantly to self-dual Hilbert C∗-modules over von Neumann algebras seems hard to ignore. Only one immediate open question remains here: does the normal form of 162 IV extend in some way for to arbitrary von Neumann algebras? As announced, in the second part of this thesis, we did not reach our goal of ax- iomatizing the category of von Neumann algebras categorically. In our attempt, we did find several new concepts such as ⋄-adjointness, ⋄-positivity, purity de- fined with quotient and comprehension and the † on these pure maps. Building on this work, van de Wetering recently announced [117, 118] a reconstruction of finite-dimensional quantum theory. To discuss it, we need a definition: call an effectus operational if its scalars are isomorphic to the real interval; both the states and the predicates are order separating; all predicate effect modules are embeddable in finite-dimensional order unit spaces and each state space is a closed subset of the base norm space of all unital positive functionals on the II III 229.. 173 corresponding order unit space of predicates. It follows from [118] that the state space of any operational &-effectus is a spectral convex set in the sense of Alfsen and Shultz and that any operational †-effectus is equivalent to a subcategory of the category of Euclidean Jordan Algebras with positive maps between them in the opposite direction. Are there infinite-dimensional generalizations of these results? Are the state spaces of a real &-effectuses perhaps spectral convex sets? Is any real †-effectus equivalent to a subcategory of, say, the category of JBW-algebras with positive maps between them in the opposite direction? Index &, andthen, 211 II [0, 1]A , 175 II {Xp}, 199 II !, initial/final map, 180 I X/p, 197 II †, dagger -category, 214 I -effectus, 215 I -epi, 224 II in a †′-effectus, 217 II -kernel, 224 II category, 224 II -mono, 224 II -partial isometry, 224 II -positive, 214 I -self-adjoint, 214 I -unitary, 214 I ⋄, diamond -adjoint, 206 II -effectus, 206 II -equivalent, 206 II -positive, 206 II -self-adjoint, 206 II ε, 150 VI ε, 150 VIII ⌊ ⌋, floor, 203 I ⌈ ⌉, ceiling, 203 I f in an effectus, 186 I xihy for Hilbert B-modules, 159 II λ · p, 190 II h· ,·i in an effectus, 181 VII B-valued, 141 II 6 among kernels, 226 IV in PCM/EA, 174 II ., 162 I ∇, 187 III ⊙ effect monoid, 178 II vector spaces, 137 II ⊖, 176 I ( )⋄, ( )⋄, 206 II ( ) , ( ) , 206 II ( )∗, ( )∗, 227 II ( )⊥ element effect algebra, 175 I subset Hilbert B-module, 160 III πp, 199 IV, 211 IX ⊲i, 180 VII ⇀, 180 I ⊗ Hilbert B-modules, 164 II module maps, 165 III ⊗ϕ, 154 III ξp, 197 III ζs, 211 IX h· , ·if [· , · ] B-module, 142 II cotupling, 180 I inner product, 136 II, 137 II B-valued, 146 VII, 148 V, 150 II, 150 XI, 150 XIII, 163 II [0, ϕ]ncp, 157 II ℓ2((pi)i∈I ), 161 II Ba(X), 143 I 6ncp, 157 II, 198 II ϕt, 157 III adV Hilbert B-modules, 153 I A AConvM , 192 IV adV Hilbert spaces, 135 IV adjointable, 143 I 175 M -affine, 192 IV asrtp, assert, 211 II B basically disconnected, 195 VI (orthonormal) basis, 149 I Bessel's inequality, 149 VIII bound, 187 III C cancellative M -convex set, 192 IV Cauchy net, 147 I equivalent, 147 I fast, 150 V Cauchy sequence fast, 136 III CH, 189a II ϕ-compatible, 154 II comprehension, 199 II corresponding, 211 IX congruence M -convex set, 193 II (abstract) M -convex set, 192 IV (formal) M -convex combination, 192 II corner, 169 II standard, 169 IV CRng, 189a II CvN, 189a I D DM , 192 II dilation in an effectus, 221 II D-poset, 176 IV E EA, 175 I effect algebra, 175 I convex, 179 III homomorphism, 175 I sequential-, 225 IV sub-, 175 I effect divisoid, 195 II effect module, 179 II homomorphism, 179 II effect monoid, 178 II commutative, 178 II homomorphism, 178 II effects, 175 II effectus, 189 II †′-, 215 V &-, 211 II ⋄-, 206 II, 215 I in partial form, 180 VII in total form, 180 I operational, 229 III real, 190 II with comprehension, 199 II with dilations, 221 II and order correspondence, 223 V with images, 202 I with quotients, 197 II EModM , 179 II entourages, 146 II exact, 226 IV at, 227 II sequence, 227 II F cok f , 200 II faithful, 202 I filter, 169 VIII standard, 169 X H Hilbert B-module, 141 II pre-, 141 II self-dual, 141 IIa homological category pointed, 226 IV I im, im⊥, 202 I inner product B-valued, 141 II definite, 141 II Inv f , 223 II K κi, coprojection, 180 I ker f , 200 II kernel †−, 224 II category theoretic, 200 II Kraus operators, 138 VIII M M , 190 II map ncp(u)-, 135 II nmiu-, 135 II modular, 227 II N ncp-extreme, 172 II Nsb A, 226 IV O orthocomplement in an effect algebra, 175 I w.r.t. a Hilbert B-module, 160 III orthogonal, 149 I ortholattice, 177 IV orthomodular lattice, 177 IV orthonormal, 149 I OUG, 189a II OUS, 189a II P ℘, power set, 146 II Par C, 180 VI Parseval's identity, 149 IV partial map, 180 I partial projectors, 180 VII Paschke dilation, 140 II PCM, 174 II homomorphism, 174 II polar decomposition, 149 VIII polarization identity, 142 IX Pred(f ), 190 II predicate, 190 II Pred X, 180 VII pristine, 218 IV pullback lemma, 183 III pure in an effectus, 201 II ncp-map, 170 I Pure C, 211 XVI Q quotient, 197 II corresponding, 211 IX R Rng, 191 VIII S scalar, 190 II Scal C, 190 II SEA, 225 IV seff p, 223 II semiexact category pointed, 226 IV separating predicates, 190 II states, 190 II sesquilinear form B-valued, 142 VII bounded, 152 II Set, 189a II sharp, 203 I map, 210 I order-, 208 I SPred X, 203 I state, 190 II sub-, 190 II Stat X, 190 II Stinespring dilation, 139 I minimal, 139 I T tensor product of Hilbert B-modules, 164 II 177 of module maps, 165 III total map, 180 VII U ultranorm, 146 VII ultrastrong, 146 V ultraweak, 146 V uniform space, 146 II complete, 147 I continuous, 147 I convergence, 147 I dense, 147 I Hausdorff, 146 II uniformly continuous, 147 I unlim, 151 II uslim, 149 VII uwlim, 149 VII V vN, 180 V Bibliography [1] Creative commons public https://creativecommons.org/licenses/by/4.0/legalcode. international attribution 4.0 license. [2] Samson Abramsky and Bob Coecke. A categorical semantics of quantum protocols. In LICS 2004, pages 415 -- 425. IEEE. arXiv:quant-ph/0402130. [3] Robin Adams and Bart Jacobs. A type theory for probabilistic and Bayesian reasoning. 2015. arXiv:1511.09230. [4] Erik M. Alfsen and Frederic W. Shultz. State spaces of operator algebras. 2001. doi:10.1007/978-1-4612-0147-2. [5] Michael A. Arbib and Ernest G. Manes. Partially additive categories and flow-diagram semantics. Journal of Algebra, 62(1):203 -- 227, 1980. doi:10.1016/0021-8693(80)90212-4. [6] Alvaro Arias and Stanley P. Gudder. Almost sharp quantum ef- Journal of Mathematical Physics, 45(11):4196 -- 4206, 2004. fects. doi:10.1063/1.1806532. [7] William Arveson. An invitation to C-algebras, volume 39. Springer, 2012. doi:10.1007/978-1-4612-6371-5. [8] William B. Arveson. Subalgebras of C-algebras. Acta Mathematica, 123(1):141 -- 224, 1969. doi:10.1007/BF02392388. [9] Steve Awodey. Category theory. Oxford University Press, 2010. [10] Michael Barr and Charles Wells. Toposes, triples and theories, volume 278. Springer-Verlag New York, 1985. [11] Garrett Birkhoff and John von Neumann. The logic of quantum mechan- ics. Annals of Mathematics, pages 823 -- 843, 1936. doi:10.2307/1968621. [12] David P Blecher, Christian Le Merdy, et al. Operator algebras and their modules: an operator space approach. Number 30. Oxford University Press, 2004. [13] Costin Badescu and Prakash Panangaden. Prospects and problems. 33 -- 42. Open Publishing Association, 2015. doi:10.4204/EPTCS.195.3. Quantum alternation: In QPL 2015, volume 195 of EPTCS, pages 179 [14] Mark S. Burgin. Categories with involution, and correspondences in γ- categories. Trudy Moskovskogo Matematicheskogo Obshchestva, 22:161 -- 228, 1970. [15] Aurelio Carboni, Stephen Lack, and Robert F.C. Walters. Introduction to extensive and distributive categories. Journal of Pure and Applied Algebra, 84(2):145 -- 158, 1993. doi:10.1016/0022-4049(93)90035-R. [16] Giulio Chiribella. Distinguishability and copiability of programs in general process theories. Int J. Software Informatics, 1(2), 2014. arXiv:1411.3035. [17] Giulio Chiribella, Giacomo Mauro D'Ariano, and Paolo Perinotti. Proba- bilistic theories with purification. Physical Review A, 81(6):062348, 2010. doi:10.1103/PhysRevA.81.062348. [18] Giulio Chiribella, Giacomo Mauro D'Ariano, and Paolo Perinotti. Infor- mational derivation of quantum theory. Physical Review A, 84(1):012311, 2011. doi:10.1103/PhysRevA.84.012311. [19] Kenta Cho. Total and partial computation in categorical quantum foun- In QPL 2015, volume 195 of EPTCS, pages 116 -- 135, 2015. dations. doi:10.4204/EPTCS.195.9 arXiv:1511.01569v1. [20] Kenta Cho and Bart P.F. Jacobs. The EfProb library for probabilistic In CALCO 2017, volume 72 of LIPIcs, pages 25:1 -- 25:8, calculations. 2017. doi:10.4230/LIPIcs.CALCO.2017.25. [21] Kenta Cho and Bart P.F. Jacobs. Disintegration and Bayesian inver- sion via string diagrams. MSCS, 2019. doi:10.1017/S0960129518000488 arXiv:1709.00322v3. [22] Kenta Cho, Bart P.F. Jacobs, Abraham A. Westerbaan, and Bas E. Westerbaan. Quotient -- comprehension chains. In QPL 2015, vol- ume 195 of EPTCS, pages 136 -- 147, 2015. doi:10.4204/EPTCS.195.10 arXiv:1511.01570v1. [23] Kenta Cho, Bart P.F. Jacobs, Bas E. Westerbaan, and Abraham A. West- erbaan. An introduction to effectus theory. arXiv:1512.05813v1, 2015. [24] Man-Duen Choi. A Schwarz inequality for positive linear maps on Illinois Journal of Mathematics, 18(4):565 -- 574, 1974. C-algebras. doi:10.1215/ijm/1256051007. [25] Bob Coecke and Aleks R. Kissinger. Picturing quantum processes. 2017. doi:10.1017/9781316219317. [26] Oscar Cunningham and Chris J.M. Heunen. Purity through factori- In QPL 2017, volume 266 of EPTCS, pages 315 -- 328, 2017. sation. doi:10.4204/EPTCS.266.20 arXiv:1705.07652. [27] Giacomo Mauro D'Ariano, Giulio Chiribella, and Paolo Perinotti. Quan- tum Theory from First Principles. Cambridge University Press, 2016. doi:10.1017/9781107338340. [28] Edward Brian Davies. Quantum theory of open systems. Academic Press London, 1976. [29] Anatolij Dvurecenskij and Sylvia Pulmannov´a. New trends in quantum structures, volume 516. Springer, 2013. doi:10.1007/978-94-017-2422-7. [30] Joe Flood. Semiconvex geometry. Journal of the Australian Mathematical Society, 30(4):496 -- 510, 1981. doi:10.1017/S1446788700017973. [31] David J. Foulis and Mary K. Bennett. Effect algebras and un- sharp quantum logics. Foundations of Physics, 24(10):1331 -- 1352, 1994. doi:10.1007/BF02283036. [32] Michael Frank. Self-duality and C-reflexivity of Hilbert C-moduli. fur Analysis und ihre Anwendungen, 9(2):165 -- 176, 1990. Zeitschrift doi:10.4171/ZAA/390. [33] Tobias Fritz. Convex spaces I: Definition and examples. 2009. arXiv:0903.5522. [34] Robert W.J. Furber and Bart P.F. Jacobs. From Kleisli cat- egories to commutative C∗-algebras: probabilistic Gelfand duality. In CALCO 2013, volume 8089 of LNCS, pages 141 -- 157, 2013. doi:10.1007/978-3-642-40206-7 12. [35] Robert W.J. Furber and Bas E. Westerbaan. Unordered tuples in In QPL 2015, EPTCS, pages 196 -- 207, 2015. quantum computation. doi:10.4204/EPTCS.195.15. [36] Israel Gelfand and Mark Neumark. On the imbedding of normed rings into the ring of operators in Hilbert space. Rec. Math. [Mat. Sbornik] N.S., 12(2):197 -- 217, 1943. doi:10.1090/conm/167/16. [37] Aurelian Gheondea and Stanley P. Gudder. Sequential product of quantum effects. Proceedings of the American Mathematical Society, 132(2):503 -- 512, 2004. doi:10.2307/1194082. [38] Paul Ghez, Ricardo Lima, and John Roberts. W∗-categories. Pacific Journal of Mathematics, 120(1):79 -- 109, 1985. 181 [39] Leonard Gillman and Meyer Jerison. Rings of continuous functions. Springer, 2013. doi:10.1007/978-1-4615-7819-2. [40] Stefano Gogioso and Carlo Maria Scandolo. Categorical probabilistic In QPL 2017, volume 266 of EPTCS, pages 367 -- 385, 2017. theories. doi:10.4204/EPTCS.266.23 arXiv:1701.08075. [41] Marco Grandis. On the categorical foundations of homological and homotopical algebra. Cahiers de Topologie et G´eom´etrie Diff´erentielle Cat´egoriques, 33(2):135 -- 175, 1992. [42] Stanley P. Gudder. A general theory of convexity. Milan Journal of Mathematics, 49(1):89 -- 96, 1979. doi:10.1007/BF02925185. [43] Stanley P. Gudder. In- ternational Journal of Theoretical Physics, 38(12):3179 -- 3187, 1999. doi:10.1023/A:1026678114856. Convex structures and effect algebras. [44] Stanley P. Gudder. Open problems for sequential effect algebras. International Journal of Theoretical Physics, 44(12):2199 -- 2206, 2005. doi:10.1007/s10773-005-8015-1. [45] Stanley P. Gudder and Richard J. Greechie. Uniqueness and order in sequential effect algebras. International Journal of Theoretical Physics, 44(7):755 -- 770, 2005. doi:10.1007/s10773-005-7054-y. [46] Stanley P. Gudder and Fr´ed´eric Latr´emoli`ere. Characterization of the sequential product on quantum effects. Journal of Mathematical Physics, 49(5):052106, 2008. doi:10.1063/1.2904475. [47] Stanley P. Gudder and Gabriel Nagy. Sequential quantum mea- surements. Journal of Mathematical Physics, 42(11):5212 -- 5222, 2001. doi:10.1063/1.1407837. [48] Stanley P. Gudder and Sylvia Pulmannov´a. Representation theorem for convex effect algebras. Commentationes Mathematicae Universitatis Car- olinae, 39(4):645 -- 660, 1998. [49] Chris J.M. Heunen. logics. and https://www.math.ru.nl/~landsman/PhDChris.pdf. PhD thesis, Categorical Radboud quantum University, models 2009. [50] Chris J.M. Heunen and Martti Karvonen. Limits in dagger categories. arXiv:1803.06651. [51] Chris J.M. Heunen, Nicolaas P. Landsman, and Bas Spitters. Bohrification of operator algebras and quantum logic. Synthese, 186(3):719 -- 752, 2012. doi:10.1007/s11229-011-9918-4. [52] Chris J.M. Heunen and Bert J. Lindenhovius. Domains of com- In LICS 2015, pages 450 -- 461, 2015. mutative C∗-subalgebras. doi:10.1109/LICS.2015.49. [53] Peter Hines and Samuel L. Braunstein. The structure of partial isometries. Semantic Techniques in Quantum Computation, pages 361 -- 389, 2010. [54] Mathieu Huot and Sam Staton. Universal properties in quantum the- In QPL 2018, volume 287 of EPTCS, pages 213 -- 223, 2019. ory. doi:10.4204/EPTCS.287.12. [55] Bart P.F. Jacobs. Orthomodular lattices, Foulis semigroups and dagger kernel categories. LMCS, 6, June 2010. doi:10.2168/LMCS-6(2:1)2010 arXiv:0912.0931. [56] Bart P.F. Jacobs. Probabilities, distribution monads, and convex categories. Theoretical Computer Science, 412(28):3323 -- 3336, 2011. doi:10.1016/j.tcs.2011.04.005. [57] Bart P.F. Jacobs. New directions in categorical logic, probabilistic and quantum logic. sical, doi:10.2168/LMCS-11(3:24)2015 arXiv:1205.3940v5. LMCS, 11(3), for clas- 2015. [58] Bart P.F. Jacobs. Affine monads and of LNCS, pages side-effect-freeness. 2016. 53 -- 72, In CMCS 2016, doi:10.1007/978-3-319-40370-0 5. volume 9608 [59] Bart P.F. Jacobs. Effectuses from monads. ENTCS, 325:169 -- 183, 2016. doi:10.1016/j.entcs.2016.09.037. [60] Bart P.F. Jacobs. discrete for doi:10.23638/LMCS-13(3:17)2017 arXiv:1607.02790v3. probability LMCS, Hyper distributions. normalisation and conditioning 2017. 13(3), [61] Bart P.F. Jacobs. Quantum effect logic in cognition. Journal of Mathe- matical Psychology, 81:1 -- 10, 2017. doi:10.1016/j.jmp.2017.08.004. [62] Bart P.F. Jacobs. A recipe for state-and-effect triangles. LMCS, 13(2), 2017. doi:10.23638/LMCS-13(2:6)2017 arXiv:1703.09034v3. [63] Bart P.F. Jacobs. A channel-based perspective on conjugate priors. arXiv:1707.00269v2, 2018. [64] Bart P.F. Jacobs. From probability monads to commutative effectuses. Journal of Logical and Algebraic Methods in Programming, 94:200 -- 237, 2018. doi:10.1016/j.jlamp.2016.11.006. 183 [65] Bart P.F. Jacobs and Jorik Mandemaker. Coreflections in alge- Foundations of physics, 42(7):932 -- 958, 2012. braic quantum logic. doi:10.1007/s10701-012-9654-8. [66] Bart P.F. Jacobs, Jorik Mandemaker, and Robert W.J. Furber. The ex- pectation monad in quantum foundations. Information and Computation, 250:87 -- 114, 2016. doi:10.1016/j.ic.2016.02.009. [67] Bart P.F. Jacobs and Abraham A. Westerbaan. An effect- theoretic account of lebesgue integration. ENTCS, 319:239 -- 253, 2015. doi:10.1016/j.entcs.2015.12.015. [68] Bart P.F. Jacobs and Abraham A. Westerbaan. Distances between states and between predicates. arXiv:1711.09740v2, 2018. [69] Bart P.F. Jacobs, Bas E. Westerbaan, and Abraham A. Westerbaan. In FoSSaCS 2015, volume 9034 of LNCS, pages States of convex sets. 87 -- 101, 2015. doi:10.1007/978-3-662-46678-0 6. [70] Bart P.F. Jacobs and Fabio Zanasi. A formal semantics of influence In MFCS 2017, volume 83 of LIPIcs, 2017. in Bayesian reasoning. doi:10.4230/LIPIcs.MFCS.2017.21. [71] Wang Jia-Mei, Wu Jun-De, and Cho Minhyung. Entropy of partitions on sequential effect algebras. Communications in Theoretical Physics, 53(3):399, 2010. doi:10.1088/0253-6102/53/3/01. [72] Shen Jun and Junde Wu. gebras. doi:10.1016/S0034-4877(09)90015-5. Reports on Mathematical Physics, Remarks on the sequential effect al- 63(3):441 -- 446, 2009. [73] Richard V. Kadison and John R. Ringrose. Fundamentals of the The- doi:10.1090/gsm/015 & 1983. ory of Operator Algebras, I & II. doi:10.1090/gsm/016. [74] Gudrun Kalmbach. Orthomodular lattices, volume 18. Academic Pr, 1983. doi:10.1016/S0304-0208(08)73817-9. [75] Gennadi G. Kasparov. Hilbert C-modules: theorems of Stinespring and Voiculescu. Journal of Operator Theory, pages 133 -- 150, 1980. [76] Aleks R. Kissinger, Sean Tull, and Bas E. Westerbaan. Picture-perfect quantum key distribution. 2017. arXiv:1704.08668. [77] Frantisek Kopka and Ferdinand Chovanec. D-posets. Mathematica Slo- vaca, 44(1):21 -- 34, 1994. [78] Andre Kornell. Quantum collections. Mathematics, 28(12):1750085, 2017. arXiv:1202.2994v2. International Journal of doi:10.1142/S0129167X17500859 [79] Mareli Korostenski and Walter Tholen. Factorization systems as eilenberg- moore algebras. Journal of Pure and Applied Algebra, 85(1):57 -- 72, 1993. doi:10.1016/0022-4049(93)90171-O. [80] Dennis Kretschmann, Dirk Schlingemann, and Reinhard F. Werner. The information-disturbance tradeoff and the continuity of Stinespring's repre- sentation. Information Theory, IEEE Transactions on, 54(4):1708 -- 1717, 2008. doi:10.1109/TIT.2008.917696. [81] E. Christopher Lance. ator algebraists, volume 210. doi:10.1017/CBO9780511526206. Hilbert C-modules: for oper- Cambridge University Press, 1995. a toolkit [82] Giovanni Landi and Alexander Pavlov. On orthogonal systems in Hilbert C-modules. Journal of Operator Theory, pages 487 -- 500, 2012. [83] Yuan Li and Xiu-Hong Sun. Sequential product and Jordan product of quantum effects. International Journal of Theoretical Physics, 50(4):1206 -- 1213, 2011. doi:10.1007/s10773-010-0615-8. [84] Goran Lindblad. Completely positive maps and entropy inequalities. Com- munications in Mathematical Physics, 40(2):147 -- 151, 1975. [85] Bert J. Lindenhovius. C(A). PhD thesis, Radboud University, 2016. doi:2066/158429. [86] Fred E.J. Linton. Coequalizers in categories of algebras. In Seminar on triples and categorical homology theory, volume 80, pages 75 -- 90. Springer, 1969. doi:10.1007/BFb0083082. [87] Saunders Mac Lane. Categories for the working mathematician. 1971. doi:10.1007/978-1-4757-4721-8. [88] VM Manuilov and EV Troitsky. Hilbert C-and W-modules and their morphisms. Journal of Mathematical Sciences, 98(2):137 -- 201, 2000. [89] Kevin McCrimmon. A taste of Jordan algebras. Springer, 2006. doi:10.1007/b97489. [90] Walter D. Neumann. On the quasivariety of convex subsets of affine spaces. Archiv der Mathematik, 21(1):11 -- 16, 1970. doi:10.1007/BF01220869. [91] Michael A. Nielsen and Isaac L. Chuang. Quantum computation and quantum information, 2010. doi:10.1017/CBO9780511976667. 185 [92] Arthur J. Parzygnat. Stinespring's construction as an adjunction. 2018. arXiv:arXiv:1807.02533. [93] William L. Paschke. Inner product modules over B-algebras. Trans- the American Mathematical Society, 182:443 -- 468, 1973. actions of doi:10.1090/S0002-9947-1973-0355613-0. [94] Vern I. Paulsen. Completely bounded maps and operator algebras, vol- ume 78 of Cambridge Studies in Advanced Mathematics. Cambridge Uni- versity Press, Cambridge, 2002. [95] Gert K. Pedersen. C-algebras and their automorphism groups. London Mathematical Society Monographs, 1979. [96] Mathys P.A. Rennela, Sam Staton, and Robert W.J. Furber. Infinite- dimensionality in quantum foundations: W-algebras as presheaves over matrix algebras. In QPL 2016, volume 236 of EPTCS, 2017. doi:10.4204/EPTCS.236.11 arXiv:1701.00662v1. [97] Frank Roumen. Cohomology of effect algebras. In QPL 2016, volume 236 of EPTCS, pages 174 -- 202, 2016. doi: arXiv:1602.00567. [98] Frank Roumen. Effect algebroids. PhD thesis, Radboud Universiteit Ni- jmegen, 2017. doi:2066/181438. [99] Shoichiro Sakai. C∗-algebras and W-algebras. 1998. doi:10.1007/978-3-642-61993-9. [100] Peter Schwabe and Bas E. Westerbaan. Solving binary-MQ with Grover's algorithm. In International Conference on Security, Privacy, and Applied Cryptography Engineering, pages 303 -- 322. Springer, 2016. doi:10.1007/978-3-319-49445-6 17 eprint:2019/151. [101] John Selby and Bob Coecke. Leaks: quantum, classical, intermediate and more. Entropy, 19(4):174, 2017. doi:10.3390/e19040174. [102] Peter Selinger. Dagger compact closed categories and completely positive maps. ENTCS, 170:139 -- 163, 2007. doi:10.1016/j.entcs.2006.12.018. [103] Peter Selinger. Idempotents in dagger categories. ENTCS, 210:107 -- 122, 2008. doi:10.1016/j.entcs.2008.04.021. [104] Jun Shen and Junde Wu. sharply dominating. doi:10.1016/j.physleta.2009.02.073. Not each sequential effect algebra is Physics Letters A, 373(20):1708 -- 1712, 2009. [105] Sam Staton. Algebraic effects, linearity, and quantum programming lan- In ACM SIGPLAN Notices, volume 50, pages 395 -- 406. ACM, guages. 2015. doi:10.1145/2676726.2676999. [106] W. Forrest Stinespring. Pro- the American Mathematical Society, 6(2):211 -- 216, 1955. Positive functions on C-algebras. ceedings of doi:10.2307/2032342. [107] Marshall H. Stone. Postulates for the barycentric Annali doi:10.1007/BF02413910. di Matematica Pura ed Applicata, 29(1):25 -- 30, calculus. 1949. [108] Erling Størmer. Positive linear maps of operator algebras. Springer, 2012. doi:10.1007/978-3-642-34369-8. [109] Tadeusz ´Swirszcz. Monadic functors and categories of convex sets. Polish Academy of Sciences. Institute of Mathematics, 1975. [110] Masamichi Takesaki. Theory of operator algebras I. 1979. doi:10.1007/978-1-4612-6188-9. [111] Josef Tkadlec. Atomic sequential effect algebras. International Journal of Theoretical Physics, 47(1):185 -- 192, 2008. doi:10.1007/s10773-007-9492-1. [112] Marco Tomamichel. Resources: Mathematical Foundations, volume 5. doi:10.1007/978-3-319-21891-5. Quantum Information Processing with Finite Springer, 2015. [113] Sean Tull. Operational theories of physics as categories. 2016. arXiv:1602.06284. [114] Sean Tull. Quotient categories and phases. 2018. arXiv:1801.09532. [115] Sean Tull. Categorical Operational Physics. PhD thesis, University of Oxford, 2019. arXiv:1902.00343. [116] Sander Uijlen and Bas E. Westerbaan. A Kochen-Specker system has at least 22 vectors. New Generation Computing, 34(1-2):3 -- 23, 2016. arXiv:1412.8544 doi:10.1007/s00354-016-0202-5. [117] John van de Wetering. An effect-theoretic reconstruction of quantum theory. arXiv:1801.05798. [118] John van de Wetering. Reconstruction of quantum theory from universal filters. arXiv:1801.05798. [119] John van de Wetering. Sequential product spaces are jordan algebras. arXiv:1803.11139. 187 [120] John van de Wetering. Three characterisations of the sequential prod- uct. Journal of Mathematical Physics, 59(8), 2018. doi:10.1063/1.5031089 arXiv:1803.08453. [121] Friedrich Wehrung. Partial Commutative Monoids, pages 23 -- 69. Springer International Publishing, Cham, 2017. doi:10.1007/978-3-319-61599-8 2. [122] Liu Weihua and Wu Junde. A uniqueness problem of the sequence prod- uct on operator effect algebra. Journal of Physics A: Mathematical and Theoretical, 42(18):185206, 2009. doi:10.1088/1751-8113/42/18/185206. [123] Reinhard F. Werner. Optimal cloning of pure states. Physical Review A, 58(3):1827, 1998. doi:10.1103/PhysRevA.58.1827. [124] Abraham A. Westerbaan. The Category of Von Neumann Algebras. PhD thesis, Radboud University, 2019. arXiv:1804.02203 doi:2066/201611. [125] Abraham A. Westerbaan and Bas E. Westerbaan. When does Stinespring dilation yield a faithful representation? Mathematics Stack Exchange. http://math.stackexchange.com/q/1701574 (version: 2016-03-17). [126] Abraham A. Westerbaan and Bas E. Westerbaan. A universal property for sequential measurement. Journal of Mathematical Physics, 57(9):092203, 2016. doi:10.1063/1.4961526 arXiv:1603.00410v1. [127] Abraham A. Westerbaan and Bas E. Westerbaan. Paschke dilations. In QPL 2016, volume 236, pages 229 -- 244, 2017. doi:10.4204/EPTCS.236.15 arXiv:1603.04353v2. [128] Abraham A. Westerbaan, Bas E. Westerbaan, and John van de Wetering. In Peter Selinger and Pure maps between Euclidean Jordan algebras. Giulio Chiribella, editors, QPL 2018, volume 287 of EPTCS, pages 345 -- 364. Open Publishing Association, 2019. doi:10.4204/EPTCS.287.19. [129] Bas E. Westerbaan. Sequential ics. https://www.ru.nl/publish/pages/813276/masterscriptie_bas_westerbaan.pdf. Master's thesis, 2013. product on Radboud effect log- University. [130] Bas E. Westerbaan, Abraham A. and Westerbaan. Reference request uni- tal normal -homomorphisms B(H) → B(K). Mathematics Stack Ex- change. https://math.stackexchange.com/q/3140021 (version: 2019- 03-11). [131] Alexander Wilce. A shortcut from categorical quantum theory to convex operational theories. In QPL 2017, volume 266 of EPTCS, pages 222 -- 236, 2017. doi:10.4204/EPTCS.266 arXiv:1802.09737. [132] Alexander Wilce. A royal road to quantum theory (or thereabouts). En- tropy, 20(4):227, 2018. doi:10.3390/e20040227. [133] Stephen Willard. General topology. Addison -- Wesley, 1970. 189 Lekensamenvatting De volgende twee technische vraagstukken in de wiskundige theorie van quantum computers liggen ten grondslag aan dit proefschrift. 1. Breidt de stelling van Stinespring uit naar alle von Neumann algebra's? 2. Is de categorie van von Neumann algebra's te axiomatiseren? Het eerste vraagstuk wordt beantwoord: ja, de stelling breidt uit. Het tweede vraagstuk blijft helaas voor een groot deel onbeantwoord. De belangrijkste bijdragen van dit proefschrift aan de wetenschap zijn niet per se deze directe resultaten, maar de bijvangst die het onderzoek erna heeft opgeleverd: er worden verscheidene nieuwe stellingen bewezen en begrippen ingevoerd die toepassing hebben buiten deze vraagstukken (zoals bijvoorbeeld een uitbreiding van de stelling van Kaplansky en het begrip ⋄-geadjungeerdheid). Maar waar gaan deze twee originele vraagstukken eigenlijk over? Wat is die stelling van Stinespring? Wat zijn von Neumann algebra's? En wat is een quantum computer uberhaupt? Quantum computers Een quantum computer gebruikt bepaalde quantum mechanische eigenaardighe- den om sommige berekeningen veel efficienter uit te voeren dan een traditionele computer. Een zorgelijk voorbeeld is dat quantum computer erg goed is in het ontbinden van getallen in priemfactoren. Het grootste deel van de moderne cryptografie die ons beveiligt is gestoeld op de aanname dat het vinden van zulke ontbindingen lastig is en deze wordt dan ook makkelijk door een voldoende grote quantum computer gebroken. Gelukkig zijn de huidige quantum computers nog erg klein en hebben wij de tijd om onze cryptografie aan te passen. Aan de positieve kant zijn quantum computer erg geschikt voor het simuleren van het gedrag van moleculen, waaronder het lokale effect van potentiele medicijnen. In functie is een quantum computer te vergelijken met een grafische kaart van een computer: voor de meeste berekeningen is een normale CPU het best, 191 maar voor sommige berekeningen is een grafische kaart veel efficienter. Dit komt doordat een normale CPU gemaakt is om zo snel mogelijk losse rekenstappen op afzonderlijke getallen achter elkaar uit te voeren, terwijl een grafische kaart rekenstappen uitvoert op miljoenen getallen tegelijkertijd. Niet elke berekening heeft baat bij dat grote parallelisme. Voor diegene die dat wel hebben (zoals het weergeven van een 3D-wereld in een computerspel) is het lastig om het originele algoritme aan te passen om optimaal gebruik te maken van de grafische kaart. Programmeren voor een grafische kaart vraagt om een heel andere mindset. Ook de theoretische analyse van algoritmes voor grafische kaarten is van een andere aard dan die voor een traditionele computer. De situatie bij een quantum computer is vergelijkbaar, maar extremer: de efficientiewinst tussen een quantum computer en PC is groter dan die tussen een CPU en grafische kaart. Helaas is het vinden en analyseren van algoritmes voor quantum computers ook vele malen ingewikkelder. Er is ook geen methode of zelfs vuistregel bekend om te bepalen of een berekening efficienter uit te voeren is op een quantum computer of niet. Von Neumann algebra's Het gedrag van een algoritme voor een traditionele computer, zoals diegene die het kwadraat van een telgetal berekent, wordt vaak beschreven met wiskundige functie, zoals in dit geval de functie f : N → N met f (n) = n2. De no- tatie f : N → N betekent dat de functie f als invoer een telgetal N ≡ {0, 1, 2, . . .} neemt en ook een telgetal als uitvoer geeft. Een echt programma draait op een computer dat maar een beperkte hoeveelheid geheugen heeft: er is dus een grootste telgetal dat in de computer past. Dat we het algoritme beschrijven met willekeurig grote telgetallen is een idealisering van de situatie. Het zou echter onhandig zijn om alle generiek toepasbare algoritme te beschrijven als wiskundige functies tussen eindige verzamelingen (zoals {0, 1, 2, 3}) in plaats van tussen oneindige verzamelingen zoals N. Dit is helaas precies de situatie voor de meeste beschrijvingen van algo- ritmes voor quantum computers. Het probleem is dat het analoog van N en de wiskundige functies daartussen vele malen complexer van aard zijn dan de wiskundige functies die het gedrag beschrijven voor het eindige geval. Ook al dit eindige geval op zichzelf is stukken ingewikkelder dan de simpele functies tussen eindige verzamelingen voor de normale algoritmes. Von Neumann algebra's zijn de wiskundige structuren die gebruikt worden om oneindige quantumdatatypes te beschrijven zoals het analoog van N. Zoge- noemde ncp-functies tussen von Neumann algebras zijn het analoog van de normale wiskundige functies tussen verzamelingen. Het zijn deze von Neumann algebra's met bijbehorende ncp-functies, die in dit proefschrift onderzocht wor- den om ze beter te begrijpen en eenvoudiger toe te kunnen passen. De stelling van Stinespring Het is lastig om grip te krijgen op een willekeurige ncp-functie tussen von Neu- mann algebra's. De stelling van Stinespring helpt hierbij: deze zegt dat elke ncp-functie f : A → BI tussen von Neumann algebra's A en BI (waar BI van een speciale soort is) in feite de samenstelling is van twee veel simpelere ncp- functies: eerst f1 : A → P en dan f2 : P → BI . Deze bijzondere opsplitsing wordt ook wel een dilatie genoemd en wordt gebruikt als cruciale stap in het bewijs van vele stellingen, niet alleen over ncp-functies zelf, maar ook over bi- jvoorneeld quantum protocollen en quantum informatie. Het probleem is dat de stelling van Stinespring alleen toepasbaar is als BI van de speciale soort type I is. Daarmee zijn de stellingen die ermee worden bewezen vaak ook maar toepas- baar op zulke speciale von Neumann algebra's. In het eerste hoofdstuk van het proefschrift wordt een generalisatie van de stelling van Stinespring bewezen, die toepasbaar is op alle ncp-functies. Axiomatisatie Waarom zijn juist von Neumann algebra's en ncp-functies geschikt voor oneindige quantum datatypes? In feite is het een wiskundige gok: von Neumann algebra's en ncp-functies lijken de eenvoudigste structuren die de voorbeelden die we kennen goed beschrijven. Eenvoud hier is relatief: de ingewikkelde definitie van een von Neumann algebra staat ver van de intuıtie van een informaticus en natuurkundige. Het valt ook te betwijfelen of elke von Neumann algebra een realistisch quantum mechanisch systeem beschrijft, laat staan een quantum datatype. Een mogelijke manier om deze kwestie te beslechten is de volgende: wij zoeken een lijst van kenmerkende (en voor onze toepassing relevante) eigen- schappen waaraan von Neumann algebra's samen met ncp-functies aan voldoen. Als er bewezen kan worden dat alleen von Neumann algebra's met ncp-functies aan deze lijst van eigenschappen kan voldoen (en geen ander soort wiskundige structuur) dan hebben wij met deze lijst van eigenschappen (nu verheven tot axioma's) een nieuw inzicht in de aard van von Neumann algebra's. Een dergelijke aanpak is eerder al succesvol gebruikt om de regels van quan- tum theorie voor eindige systemen beter te begrijpen. Een voorbeeld hiervan zijn de axioma's van Chiribella et al, die quantum theorie karakteriseren door de wijze waarop informatie verwerkt wordt. In het tweede hoofdstuk van dit proefschrift wordt gezocht naar een dergeli- jke axiomatisatie. Een van de kernaxioma's is het bestaan van een soort omker- ing, een zogenoemde dagger, op de speciale ncp-functies die aan de rechterkant voorkomen van een dilatie, wat de twee hoofdstuken met elkaar verbindt. Helaas wordt een volledige axiomatisatie niet bereikt. Het verhaal eindigt hier niet: 193 mijn college van de Wetering heeft recent bewezen dat met een paar toevoeg- ingen, de axioma's uit dit proefschrift zogeheten Euclidische Jordan algebra's (EJA's) karakteriseren. Hiermee is de kous nog niet af: EJA's zijn niet geschikt voor oneindige quantum datatypes. Dit resultaat suggereert wel de mogeli- jkheid dat JBW-algebra's, welke een kruising zijn van von Neumann en Jordan algebra's, uiteindelijk de juiste algebra's zijn voor onze toepassing. Hoewel ons tweede vraagstuk niet beantwoord is, heeft de zoektocht naar geschikte axioma's een aantal nieuwe begrippen aan het licht gebracht (zoals ⋄- geadjungeerdheid, hoeken en filters), die een nieuw inzicht verschaffen in von Neumann algebra's en daarmee quantum computers in het algemeen. About the author Bas Westerbaan, born in 1988, enrolled to the Radboud Universiteit as a physics student in 2007. He received his bachelor's degree in mathematics with a thesis on computability theory supervised by dr. Wim Veldman. He continued his study of logic and foundational computer science during his master's eduction and developed a keen interest for, what he thought to be, an unrelated math- ematical field: functional analysis. These interests however, are conveniently combined in the present thesis which continues the research of his master's the- sis, which was supervised by prof. Bart Jacobs as well. During his undergraduate studies, Bas also served as secretary of the board of the 150 member student club Karpe Noktem and he was elected to the student board of the faculty of science. His master's degree in mathematics was awarded cum laude in 2013. While perfoming the research as a promovendus at the Digital Security group that led to this thesis, Bas coauthered several security audits including those of the DigiD- and BerichtenBox -app; the Dutch digital identity and mail sys- tem. In the year between the submission and defense of his thesis, Bas was employed as a post-doctoral researcher where he assisted with the development of a cryptographic system for polymorphic pseudonymisation and encryption. During that same year he also taught a quarter-year course on computer net- working and was kindly invited by dr. Heunen to speak about his research at the University of Edinburgh. 195 Solutions to exercises Solution to 138 VII. To prove the first statement, let ϕ : B(H ) → B(K ) If ϕ = 0 then K ′ = 0 and V = 0 does the job, so for be any ncp-map. the other case, assume ϕ 6= 0. By 135 IV there is a Hilbert space H ′, a bounded operator W : H → H ′ and an nmiu-map : B(H ) → B(H ′) such that ϕ = adW ◦ . Clearly 6= 0. By 138 II there is a Hilbert space K ′ and a unitary U : H ′ → H ⊗ K ′ with (A) = U∗(A ⊗ 1)U for all A ∈ B(H ). Define V ≡ U W . Then ϕ(A) = W ∗(A)W = W ∗U∗(A⊗ 1)U W = V ∗(A⊗ 1)V , as desired. Before we can continue with the second statement, we need to understand the relationship between quantum channels and ncpu-maps. This relationship is best understood with predual characterization of von Neumann algebras due to Sakai [99], which we have been avoiding. The characterization is as thus: a C∗-algebra A is a von Neumann algebra if and only if it is isomorphic to the dual of a Banach space. Then this Banach space is unique up-to-isomoprhism as it must be isomorphic to the space of normal functionals on A (denoted by A∗) and is appropriately called the predual of A . Any normal linear map ϕ : A → B between von Neumann algebras A and B yields a linear map ϕ∗ : B∗ → A∗ via ϕ∗(ω) = ω ◦ ϕ. In the other direction, any linear map ϕ∗ : B∗ → A∗ gives rise to a normal linear map ϕ : A → B by defining ϕ(a)(ω) = ϕ∗(ω)(a) where we identified A ≡ (A∗)∗. Clearly ϕ is positive precisely if ϕ∗ maps positive functionals to positive functionals. A normal state ω : B(H ) → C is precisely of the form ω(a) = tr[ρa] for some density matrix ρ over H . Thus the predual of B(H ) can be identified with the set of trace-class operators over H . Let ϕ∗ be a linear map from the density operators on H to those on K . The map ϕ∗ is completely positive in its usual sense if the corresponding map ϕ is completely positive. Furthermore ϕ is unital if and only if ϕ∗ is trace-preserving. To prove the second statement, let Φ be any quantum channel mapping density matrices over H to those of H again. (Note: in the printed version of the thesis the exercise incorrectly assumes Φ to map density matrices over H to those over some other Hilbert space K .) It follows from the previous, that 197 there is a unique ncpu-map ϕ : B(H ) → B(H ) with tr[Φ(ρ)A] = tr[ρϕ(A)] for any density matrix ρ ∈ B(H ). See also [112] for a more direct approach. By the first part if the exercise, we know that there is a Hilbert space K ′ and a bounded operator V : H → H ⊗ K ′ with ϕ(A) = V ∗(A ⊗ 1)V . Tracing back the definition of V , we see that V is an isometry because ϕ is unital. Pick any orthonormal bases E and F of H and K ′ respectively. We may assume, without loss of generality, that K ′ is not zero-dimensional by setting K ′ = C and V = 0 in the case that Φ = 0. Pick any f0 ∈ F and any unitary U : H ⊗ K ′ → H ⊗ K ′ with U∗x⊗ f0 = V x, which exists as V is an isometry. Now we compute tr[ϕ(A)ρ] = tr[ρV ∗(A ⊗ 1)V ] hV ρe, (A ⊗ 1)V ei = Xe∈E = Xe∈E(cid:10)U∗(ρ ⊗ 1) e ⊗ f0, (A ⊗ 1)U∗ e ⊗ f0(cid:11). Inserting f0ihf0 in the previous, we may sum over all f ∈ F and get tr[ϕ(A)ρ] = Xe∈E f∈F(cid:10)U∗(ρ ⊗ f0ihf0) e ⊗ f, (A ⊗ 1)U∗ e ⊗ f(cid:11) = Xe∈E f∈F(cid:10)(U∗(e ⊗ f )), U∗(ρ ⊗ f0ihf0)U (A ⊗ 1) (U (e ⊗ f ))(cid:11) = tr(cid:2)U∗(ρ ⊗ f0ihf0)U (A ⊗ 1)(cid:3) = tr(cid:2)A trK ′ [U∗(ρ ⊗ f0ihf0)U ](cid:3). This show that indeed Φ(ρ) = trK ′ [U∗(ρ ⊗ v0ihv0)U ] as desired with v0 ≡ f0. Solution to 138 VIII. Let ϕ : B(H ) → B(K ) be any ncp-map. By 138 VII there is a Hilbert space K ′ and a bounded operator V : K → H ⊗ K ′ with ϕ(A) = where the sum converges ultraweakly and so by ultraweak continuity of adV (44 VIII) and B 7→ A ⊗ B (116 III), we see V ∗(A ⊗ 1)V . Let E be any orthonormal basis of K ′. Then 1 = Pe∈E eihe ϕ(A) = V ∗(cid:16)A ⊗Xe∈E eihe(cid:17)V ) = Xe∈E V ∗(A ⊗ eihe)V. (4.1) For e ∈ E, define Pe : H ⊗ K ′ → H by Pe ≡ 1 ⊗ he, i.e. Pe(x ⊗ y) = xhe, yi. Define Ve ≡ PeV . Note that P ∗e APe = A ⊗ eihe and so by (4.1) ϕ(A) = Xe∈E = Xe∈E = Xe∈E V ∗(A ⊗ eihe)V V ∗P ∗e APeV V ∗e AVe, as desired. From the special case A = 1, we see thatPe∈E V ∗e Ve = ϕ(1) and so the partial sums ofPe∈E V ∗e Ve are bounded. For the final part, assume H and K are finite dimensional. Recall that the standard Stinespring dilation space (say K ′′) for ϕ is constructed using a completion and quotient of B(H )⊙ K . As B(H )⊙ K is finite dimensional it is already complete. Hence K ′′ has dimension at most (dim H )2(dim K ). By construction H ⊗ K ′ ∼= K ′′, hence dim K ′ 6 (dim H )(dim K ). Recall E is a basis of K ′ and so there are indeed at most (dim H )(dim K ) Kraus operators. Solution to 139 IX. We will show that U : Rep → Repcp has a left adjoint by demonstrating the universal mapping property. Let ϕ : A → B(H ) be any object of Repcp. Pick any minimal Stinespring dilation (K , , V ) of ϕ. The map : A → B(K ) is an object of Rep. Clearly adV ◦ ◦ id = ϕ and so ηϕ ≡ (id, V ) : ϕ → U is a morphism in Repcp. We will show that for each f : ϕ → U ′ in Repcp, there is a unique f′ : → ′ in Rep with U f′ ◦ ηϕ = f . This is sufficient to show that U has a left adjoint. So let f : ϕ → U ′ be any morphism in Repcp. Say ′ : A ′ → B(K ′). Then f ≡ (m′, V ′) consists of a nmiu-map m′ : A → A ′ and bounded opera- tor V ′ : H → K ′ with adV ′ ◦ ′ ◦ m′ = ϕ. By 139 V there is a unique bounded operator S : K → K ′ with SV = V ′ and = adS ◦ ′ ◦ m′. This turns f′ ≡ (m′, S) into a morphism → ′ in Rep. Furthermore U f′ ◦ ηϕ = (m′ ◦ id, SV ) = (m′, V ′) = f . To show uniqueness, assume there is some f′′ : → ′ in Rep with U f′′ ◦ ηϕ = f . Say f′′ = (m′′, S′′). Then (m′, V ′) = f = U f′ ◦ ηϕ = (m′′, S′′V ). So m′′ = m′ and V ′ = S′′V . The fact that f′′ is a morphism in Rep is equivalent to adS′′ ◦ ′ ◦ m′′ = . Thus adS′′ ◦ ′ ◦ m′ = . By uniqueness of S, we get S′′ = S. Hence f′′ = (m′′, S′′) = (m′, S) = f′, as desired. Solution to 139 XI. Let ϕ : B(H ) → B(K ) be any ncp-map. As in the de- scription of the exercise, let K be a Hilbert space and V, W : K → H ⊗ K ′ be bounded operators with V ∗(a⊗ 1)V = ϕ(a) = W ∗(a⊗ 1)W . Write V for the closed linear span of {(a ⊗ 1)V x; a ∈ B(H ), x ∈ K } in H ⊗ K ′ and simi- larly W for that of {(a⊗ 1)W x; a ∈ B(H ), x ∈ K }. Note that for any n ∈ N, 199 (cid:13)(cid:13)Xi (ai ⊗ 1)V xi(cid:13)(cid:13)2 hxi, V ∗((a∗i aj) ⊗ 1)V xji hxi, W ∗((a∗i aj) ⊗ 1)W xji = Xi,j = Xi,j = (cid:13)(cid:13)Xi (ai ⊗ 1)W xi(cid:13)(cid:13)2 . x1, . . . , xn ∈ K and a1, . . . , an ∈ B(H ) we have Thus there is a unique unitary U0 : W → V fixed by U0(a⊗1)W x = U0(a⊗1)V x. We see U0W = V by setting a = 1. Furthermore (α ⊗ 1)U0(a ⊗ 1)W x = ((αa) ⊗ 1)V x = U0(α ⊗ 1)(a ⊗ 1)W x for any α, a ∈ B(H ) and x ∈ K , hence (α ⊗ 1)U0 = U0(α ⊗ 1). For any a ∈ B(H ), the operator a ⊗ 1 ∈ B(H ⊗ K ′) restricts to B(W ). Pick an orthonormal basis E of H and some e0 ∈ E. Note that (e0ihe0 ⊗ 1)W = e0 ⊗ W ′ for some closed subspace W ′ ⊆ K ′. In fact, for any e ∈ E we have e⊗ W ′ = (eihe0⊗ 1)(e0 ⊗ W ′) = (eihe0⊗ 1)(1⊗e0ihe0)W = (eihe T ⊗ 1)W = (eihe ⊗ 1)W , where T is the unitary on H that only swaps e and e0. Hence W = H ⊗W ′. Similarly V = H ⊗V ′ for some closed subspace V ′ ⊆ K ′. For any non-zero w ∈ W ′ and unit-vector x ∈ H , we have U0(x ⊗ w) = U0(xihx⊗1)(x⊗w) = (xihx⊗1)U0(x⊗w) so U0(x⊗w) = x⊗y for some y ∈ V ′. Clearly kwk = kx ⊗ wk = kU0(x ⊗ w)k = kx ⊗ yk = kyk, so there is a unique unitary U1 : W ′ → V ′ with U0(x ⊗ w) = x ⊗ U1w. It follows U0 = 1 ⊗ U1. As V and W are isomorphic, they have the same dimension and so do V ⊥ and W . Consequently, there is an unitary U : K ′ → K ′ extending U1. We have V = (1 ⊗ U )W = (1 ⊗ U1)W = U0W = V as desired. Solution to 140 X. We cover the points in order. 1. Let : A → B be a mniu-map. Assume there are nmiu ′ : A → P′ and ncp h′ : P′ → B with h′ ◦ ′ = . We have to show there is a unique map σ : P → B with id◦ σ = h′ and h′ ◦ ′ = . Clearly σ ≡ h′ fits the bill. 2. Let (P, , h) be any Paschke dilation. We will show that (P, id, h) is a Paschke dilation of h. To this end, let ′ : A → P′ be any nmiu- map and h′ : P′ → B be any ncp-map with h′ ◦ ′ = h. Consider ′ ◦ and h′. By the universal property of the original dilation, there is a unique ncp-map σ : P′ → P with σ ◦ ′ ◦ = and h◦ σ = h′. Further- more id : P → P is the unique ncp-map with id◦ = and h◦ id = h. Now (σ ◦ ′)◦ = and h◦(σ ◦ ′) = h′ ◦ ′ = h, so σ ◦ ′ = id. We are are halfway demonstrating that σ is also the mediating map for our dilation of h. It remains to be shown that σ is the unique ncp-map with σ ◦ ′ = id and h◦ σ = h′. So assume there is a ncp-map σ′ : P′ → P with h◦ σ′ = h′ and σ′ ◦ ′ = id. Clearly σ′ ◦ ′ ◦ = and so by uniqueness of σ as the mediating map for the orignal dilation, we see σ′ = σ, as desired. 3. Let ( ϕ1 2 ) = ( h1 ◦ 1 ϕ2 ) : A → B1⊕B2 be any ncp-map. Pick Paschke dilations (Pi, i, hi) of ϕi for i = 1, 2. We will show that (P1 ⊕ P2, ( 1 ϕ2 ). Clearly h1 ⊕ h2 ◦ ( 1 Paschke dilation of ( ϕ1 Let ′ : A → P′ be any nmiu-map and (cid:0) h1 ncp-map with (cid:0) h1 ϕ2 ). Note hi ◦ ′ = ϕi (for i = 1, 2) and so there is a unique σi : P′ → Pi with σi ◦ ′ = i and hi ◦ σi = h′i. We will show that ( σ1 σ2 ) : P′ → P1 ⊕ P2 is the unique mediating map. Clearly ( σ1 h2(cid:1)◦ ′ = ( ϕ1 σ2(cid:1) = (cid:0) h′ 2 ) and (h1 ⊕ h2)◦(cid:0) σ1 σ2 ◦ ′(cid:1) = ( 1 (cid:0) h1 h2(cid:1). To show uniqueness of ( σ1 σ2 ), assume there is ncp-map(cid:0) σ′ 2(cid:1) =(cid:0) h1 h2(cid:1). Then hi ◦ σ′i = 2(cid:1)◦ ′ =(cid:0) 1 2(cid:1) and (h1⊕h2)◦(cid:0) σ′ P1⊕P2 such that(cid:0) σ′ σ2 ) =(cid:0) σ′ 2(cid:1). hi and σ′i ◦ ′ = i for i = 1, 2 and so σi = σ′i by the uniqueness of the seperate σi. Thus indeed ( σ1 2 ) , h1 ⊕ h2) is a h2 ◦ 2 ) = ( ϕ1 ϕ2 ). h2(cid:1) : P′ → B1 ⊕ B2 any 2 ◦ σ2(cid:1) = 2(cid:1) : P′ → σ2 ) ◦ ′ = (cid:0) σ1 ◦ ′ 1 ◦ σ1 h′ 1 σ′ 1 σ′ 1 σ′ 1 σ′ 4. Let ϕ : A → B be any ncp-map with Paschke dilation (P, , h). As- sume λ ∈ R, λ > 0. We will show (P, , λh) is a Paschke dilation of λϕ. Clearly λh◦ = λϕ. To this end, assume ′ : A → P′ is a nmiu-map and h′ : P′ → B is an ncp-map with h′ ◦ ′ = λϕ. Then λ−1h′ ◦ ′ = ϕ. Thus there is a unique σ : P′ → P with σ ◦ ′ = and h◦ σ = λ−1h′. Clearly λh◦ σ = h′ and so σ also serves as the unique mediating map for the dilation of λϕ. Solution to 142 V. Let X be a right B-module with B-valued inner prod- uct h· , ·i for some C∗-algebra B. Using the C∗-identity, 142 III and the def- inition of k · k on X, we get khx, yik2 = khx, yi∗hx, yik = khy, xihx, yik 6 2 is a seminorm on X. Clearly kxk > 0 for any x ∈ X. For any λ ∈ C and x ∈ X we have hλx, λxi = λhx, xiλ = λ2hx, xi, hence kλxk = kλ2hx, xik (cid:13)(cid:13)khx, xikhy, yi(cid:13)(cid:13) = kxk2kyk2, so indeed khx, yik 6 kxkkyk. 2 = λkxk. Next, for any x, y ∈ X we have Now we will show kxk ≡ khx, xik 1 1 kx + yk2 = khx + y, x + yik 6 khx, xik + khy, yik + khx, yik + khx, yi∗k = kxk2 + kyk2 + 2khx, yik 6 kxk2 + kyk2 + 2kxkkyk = (kxk + kyk)2 201 and thus k · k is indeed a seminorm. Finally, we will show kx · bk 6 kxkkbk for any b ∈ B and x ∈ X. As hx, xi is positive, we have hx, xi 6 khx, xik = kxk2. Also recall a 7→ b∗ab is clearly desired. positive. Thus kx · bk2 ≡ khxb, xbik = kb∗hx, xibk 6(cid:13)(cid:13)kxk2b∗b(cid:13)(cid:13) = kbk2kxk2 as module X for some C∗-algebra B. Distributing B inP3 Solution to 142 IX. Let B be any B-sesquilinear form on a pre-Hilbert B- k=0 ikB(ikx + y, ikx + y) we get 16 terms consisting of four of each B(x, x), B(y, y), B(x, y) and B(y, x) with the following coefficients. k = 0 k = 1 k = 2 k = 3 1 i · (−i) · i = i (−1)3 = −1 (−i) · i · (−i) = −i B(x, x) B(y, y) 1 i −1 −i B(x, y) 1 i · (−i) = 1 (−1)2 = 1 (−i) · i = 1 B(y, x) 1 i2 = −1 (−1)2 = 1 (−i)2 = −1 Note that the coefficients in every column sum to 0, except for the coefficients for B(x, y) which sum to 4. HenceP3 Solution to 146 IV. Let X be a set together with a subbase B. Write Φ for the filter generated by B. Note B ⊆ Φ. We will show (X, Φ) is a uniform space, by proving its axioms in order. k=0 ikB(ikx + y, ikx + y) = 4B(x, y). 1. By construction Φ is a filter. 2. Pick any ε ∈ Φ. We have to show ∆ ≡ {(x, x); x ∈ X} ⊆ ε. By definition of Φ, there are ε1, . . . , εn ∈ B with ε1 ∩ . . . ∩ εn ⊆ ε. By definition of a base, we have ∆ ⊆ εi for each 1 6 i 6 n and so ∆ ⊆ ε1 ∩ . . . ∩ εn ⊆ ε as well. 3. Pick any ε ∈ Φ. By definition of Φ, there are ε1, . . . , εn ∈ B with ε1∩ . . .∩ εn ⊆ ε. As B is a base, there are δ1, . . . , δn with δi ◦ δi ⊆ εi for 1 6 i 6 n. 4. Pick any ε ∈ Φ. By definition of Φ, there are ε1, . . . , εn ∈ B with ε1∩ . . .∩ i ⊆ εi. Define δ = Define δ = δ1 ∩ . . . ∩ δn. Then δ ◦ δ ⊆Ti,j δi ◦ δj ⊆Ti δi ⊆Ti εi ⊆ ε. εn ⊆ ε. As B is a base, there are δ1, . . . , δn with δ−1 δ1 ∩ . . . ∩ δn. Then δ−1 =Ti δ−1 i ⊆Ti εi ⊆ ε. Thus indeed, (X, Φ) is a uniform space. Solution to 147 II. We prove the subexercises in order. 1. First we will show that equivalence of Cauchy nets is an equivalence rela- tion. As for every entourage ε we have x ε x, we see that every Cauchy net is equivalent to itself. Assume (xα)α ∼ (yβ)β. We will show (yβ)β ∼ (yα)α. Let ε be some entourage. There is some entourage δ with δ−1 ⊆ ε. By as- sumption there are α0 and β0 such that xα δ yβ for all α > α0 and β > β0. But then yβ ε xα for α > α0 and β > β0. Hence (yβ)β ∼ (xα)α. To prove transitivity, assume we are given Cauchy nets (xα)α ∼ (yβ)β ∼ (zγ)γ. Let ε be some entourage. There is an entourage δ with δ ◦ δ ⊆ ε. There are α0, β0, γ0 such that xα δ yβ and yβ δ zγ for α > α0, β > β0 and γ > γ0. Hence xα ε zγ for such α and γ, which shows (xα)α ∼ (zγ)γ. Next, assume that (xα)α is a subnet of a Cauchy net (yα)α. To show (xα)α ∼ (yα)α, assume ε is some entourage. By the definition of Cauchy net, there is some α0 such that xα ε xβ for all α, β > α0. In particular xα ε yβ for all α, β > α0 which shows (xα)α ∼ (xβ )β. 2. Assume (xα)α ∼ (yβ)β and xα → x. To prove yα → x, pick any en- tourage ε. Pick δ such that δ2 ⊆ ε. There are α0 and β0 such that yβ δ xα and xα δ x for all α > α0 and β > β0. Then yβ ε x for all β > β0, whence yβ → x. 3. Assume (xα)α → x and (xα)α → y in some Hausdorff uniform space. Let ε be any entourage. Pick δ and δ′ with δ2 ⊆ ε and δ′ ⊆ δ−1. There is an α0 such that xα δ x and xα δ′ y for all α > α0. Thus x ε y. As our space is Hausdorff the previous implies x = y. 4. Let f : X → Y be a continuous map between uniform spaces. Assume xα → x in X. Let ε be any entourage of Y . By continuity there is a δ such that x δ y implies f (x) ε f (y) for any y ∈ Y . There is an α0 such that x δ xα for all α > α0. For those α we also have f (x) ε f (xα), which shows f (xα) → f (x). 5. Let f : X → Y be a uniformly continuous map between uniform spaces. Let (xα)α and (yβ)β be nets of X such that for each entourage ε of X there are α0, β0 with xα ε yβ for all α > α0 and β > β0. The map f preserves this relation between the nets in the following way. Let ε be any entourage of Y By uniform continuity there is a δ such that x δ y implies f (x) ε f (y). There are α0 and β0 with xα δ yβ for all α > α0 and β > β0. Hence f (xα) ε f (yβ) for such α, β. From the previous it follows that f preserves Cauchy nets (by setting xα = yα) and that it preserves equivalence between Cauchy nets. 6. Let D ⊆ X be a dense subset of a uniform space X. Let x ∈ X be any point. Pick for every ε ∈ Φ an element dε ∈ D with x ε dε. Clearly (dε)ε∈Φ is a net with inverse inclusion. We have dε → x as dδ ε x whenever δ ⊆ ε. 7. Assume f, g : X → Y are continuous maps between uniform spaces that agree on a dense subset D ⊆ X. Let x ∈ X be any point. Pick a net xα from D with xα → x. Then f (x) = f (limα xα) = limα f (xα) = limα g(xα) = g(limα xα) = g(x). Hence f = g. 203 so B also satisfies condition 4 of 146 II. Assume i0 ∈ I and ε is an entourage of Xi0 . Define δ ≡ ε. Let (xi)i and (yi)i uniformly continuous. To show (πi)i is a categorical product, assume we are given a uniform space Y together with uniformly continuous maps fi : Y → Xi for each i ∈ I. We with πi ◦ f = fi for all i ∈ I. Define f by (f (y))i ≡ fi(y). Clearly πi ◦ f = fi. a δ ∈ Φi with δ2 ⊆ ε. Then δ2 = bδ2 ⊆ ε and so B satisfies condition 3 of 146 II. Now let δ denote an entourage of Xi with δ−1 ⊆ ε. Then δ−1 = dδ−1 ⊆ ε and Next we show that the projectors πi : Qi Xi → Xi are uniformly continuous. from Qi Xi be given with (xi)i δ (yi)i. Then xi0 ε yi0. Thus πi0 is indeed have to show that there is a unique uniformly continuous map f : Y →Qi Xi To show f is uniformly continuous, pick any entourage ε of Qi Xi. By definition, there are i1, . . . , in and ε1, . . . , εn with εj ∈ Φij andTj bεj ⊆ ε. For Define δ ≡Tj δj Assume x δ y. Then for each 1 6 j 6 n we have fij (x) εj fij (y) and so f (x) bεj f (y), hence f (x) ε f (y). Thus f is uniformly continuous. Qi Xi with πi ◦ f′ = fi. Then (f′(y))i = fi(y) = (f (y))i for every y ∈ Y and each 1 6 j 6 n pick an entourage δj of Y such that x δj y implies fij (x) εj fij (y). Solution to 147 III. Write B ≡ {ε; ε ∈ Φi; i ∈ I}. First we show B is a subbase, i.e. that is satisfies conditions 2, 3 and 4 of 146 II. Let ε be an arbitrary element of B and i ∈ I denote the index element with ε ∈ Φi. Assume (xi)i∈I ∈ ΠiXi. Clearly xi ε xi and so (xi)i ε (xi)i. Thus B satisfies condition 2 of 146 II. Pick To show uniqueness of f , assume there is a uniformly continuous map f′ : Y → so f′ = f . Solution to 148 IV. Let B be a von Neumann algebra and X a right B-module with B-valued inner product. We will show that x 7→ xb is ultranorm continuous for any b ∈ B. It is sufficient to show it is ultranorm continuous at 0, so assume xα → 0 ultranorm for some net xα in X. Let f : B → C be any f = f ([xαb, xαb]) = f (b∗[xα, xα]b) 6 kb∗bkf ([xα, xα]) = np-map. Then kxαbk2 kbk2kxαk2 Solution to 149 III. For brevity, write p ≡ he, ei. Note kep− ek2 = ke(1− p)k2 = he(1 − p), e(1 − p)i = (1 − p)he, ei(1 − p) = (1 − p)p(1 − p) = 0. Thus ep− e = 0 and so ep = e as desired. f → 0 thus xαb → 0 ultranorm as well. Solution to 149 IV. Let X be a pre-Hilbert B-module for a von Neumann algebra B with orthonormal basis E ⊆ X. Assume x ∈ X. By definition of or- thonormal basis, we know x =Pe∈E ehe, xi where the sum converges ultranorm. That is: Pe∈S ehe, xi → x as S ranges over the finite subsets of E. By 148 V we see(cid:10)Pe∈S ehe, xi,Pe∈S ehe, xi(cid:11) → hx, xi ultraweakly. ThusPe∈Shx, eihe, xi = Pe,d∈Shx, eihe, dihd, xi =(cid:10)Pe∈S ehe, xi,Pe∈S ehe, xi(cid:11) → hx, xi ultraweakly, as desired. Solution to 152 VIII. Let T : X → Y be a bounded B-linear map between Hilbert B-modules. Assume X is self dual. For any y ∈ Y , the map x 7→ hy, T xi is B-linear and bounded and so by self-duality of X, there is a ty ∈ X with hty, xi = hy, T xi for all x ∈ X. For any z, y ∈ Y and x ∈ X, we have htz + ty, xi = htz, xi + hty, xi = hz, T xi + hy, T xi = hz + y, T xi = htz+y, xi. Thus tz + ty = tz+y For any λ ∈ C, y ∈ Y and x ∈ X we have hλty, xi = hλy, T xi = htλy, xi and so λty = tλy. Hence T ∗y ≡ ty defines a linear map from Y to X with hy, T xi = hty, xi ≡ hT ∗y, xi for all x ∈ X and y ∈ Y . So T ∗ is the adjoint of T . Solution to 152 IX. Let V be a right B-module with B-valued inner product for some von Neumann algebra B. Write η : V → X for the ultranorm completion of V from 150 II. Let T ∈ Ba(X) be given with hx, T xi > 0 for all x ∈ V . We have to show T > 0. Let x ∈ X be an arbitrary vector. As all vector states on Ba(X) are order separating by 148 VII, it is sufficient to show hx, T xi > 0. As the image of V under η is ultranorm dense in X, we can find find a net xα indeed T > 0, as desired. with cxα → x. Then by 148 V we get hx, T xi = uwlimαhcxα, Tcxαi > 0. So Solution to 153 IV. Let A be a C∗-algebra with a1, . . . , an ∈ A . Define ϕ : A → MnA by ϕ(d) ≡ (a∗i daj)ij . We have to show ϕ is an ncp-map. Recall A is a self- dual Hilbert A -module with ha, bi ≡ a∗b and so its n-fold direct product A n It is also bounded: (i.e. T is the row-vector (ai)i.) Clearly T is A -linear. is also self dual (see also 160 II.) Define T : A n → A by T ((bi)i) ≡ Pi biai kT (bi)ik2 =Pi kbiaik2 6Pi Akbik2 = Ak(bi)ik2, where A ≡ maxi kaik2. It's then adT (d) ((bi)i) = T ∗dT (bi)i = T ∗dPi aibi = (Pi(a∗j dai)bi)j = ϕ(d) ((bi)i). easy that T ∗(b) ≡ (a∗i b)i is the adjoint of T . We may identify Ba(A n) = Mn and Thus adT = ϕ. By 153 I the map adT and thus ϕ is as well. Solution to 160 II. Assume X and Y are self-dual Hilbert B-modules over a von Neumann algebra B with orthonormal bases E ⊆ X and F ⊆ Y . Write G ≡ {(e, 0); e ∈ E} ∪ {(0, d); d ∈ D}. Clearly G is orthonormal. To show G is an orthonormal basis, two conditions remain. For the first, let (x, y) ∈ X ⊕ Y be given. As E and F are orthonormal bases we know x =Pe∈E ehe, xi and y = Pf∈F fhf, xi, where the sums converge ultranorm. The inclusions x 7→ (x, 0) 205 and y 7→ (0, y) are bounded B-linear and thus ultranorm continuous, hence Xg∈G ghg, (x, y)i = (cid:16)Xe∈E = (cid:16)Xe∈E = (cid:16)Xe∈E (e, 0)h(e, 0), (x, y)i(cid:17) + (cid:16)Xf∈F (e, 0)he, xi(cid:17) + (cid:16)Xf∈F (0, f )hf, yi(cid:17) fhf, yi(cid:17) ehe, xi,Xf∈F = (x, y), (0, f )h(0, f ), (x, y)i(cid:17) mal basis of E⊥⊥. let (bg)g∈G be which proves the first condition. For the second condition, some ℓ2-summable family from B. The subfamilies (be0 )e∈E and (b0,f )f∈F are ℓ2-summable as well. Hence the sums Pe∈E eb(e,0) andPf∈F f b(0,f ) con- verge ultranorm. Thus Pe∈E(e, 0)b(e,0) +Pf∈F (0, f )b(0,f ) = Pg∈G gbg con- verges ultranorm as well. We have shown G is an orthonormal basis of X ⊕ Y . Consequently X ⊕ Y is self dual by 149 V. Solution to 160 IX. Assume X is a self-dual Hilbert B-module for a von Neu- mann algebra B. Suppose E ⊆ X is an orthonormal set. We will show E is an orthonormal basis of E⊥⊥. Clearly E is orthonormal. Because of this, and For the second part, assume x ∈ X. By Parseval's identity (see 149 IV) ultranorm in E⊥⊥ for any ℓ2-family (be)e. Assume x ∈ E⊥⊥. To show E is an the fact that E⊥⊥ is ultranorm closed (by 160 IV) we know Pe bbe converges orthonromal base, it only remains to be shown that x =Pe ehe, xi. Define x′ ≡ x −Pe ehe, xi. By 160 IV we know E⊥⊥ is ultranorm closed and so x′ ∈ E⊥⊥. For any e0 ∈ E we also have he0, x′i = he0, xi −Pe∈Ehe0, eihe, xi = 0 and so x′ ∈ E⊥. Hence hx′, x′i = 0, so indeed x =Pe ehe, xi and E is an orthonor- we have hx, xi = Pehx, eihe, xi for any x ∈ E⊥⊥. To prove the converse, assume hx, xi =Pehx, eihe, xi. By 160 IV we know x = x′ + x′′ for x′ ∈ E⊥⊥ for E⊥⊥ we see hx′, x′i = Pehx, eihe, xi = hx, xi. Now, using hx′′, x′i = 0 and x′′ ∈ E⊥. Note he, x′i = he, xi for any e ∈ E and so by Parseval's identity we see hx, xi = hx′′, x′′i + hx′, x′i = hx′′, x′′i + hx, xi and so hx′′, x′′i = 0, whence x ∈ E⊥⊥. Solution to 160 X. Let X be a self-dual Hilbert B-module for some von Neu- mann algebra B. Assume x1, . . . , xn ∈ X. We will show that there is a finite or- thonormal set E of n or fewer elements such that E is a basis of {x1, . . . , xn}⊥⊥. We do this by induction over n. For n = 0, the set E = ∅ suffices. For the in- duction step, assume n > 0 and E′ is an orthonormal basis of {x1, . . . , xn−1}⊥⊥. case, assume x′ 6= 0. By polar decomposition (see the end of 149 VIII), there is 2 and hu, ui = ⌈hx′, x′i⌉. Define E ≡ E′ ∪ {u}. an u ∈ X with x′ = uhx′, x′i Write x′ ≡ xn −Pe∈E′ ehe, xni. If x′ = 0, then E ≡ E′ suffices. For the other 1 f + 2kaikfkbikf kaikfkbikf . f(cid:16)Xi∈S 1 2 = 0 and so x′ ∈ E⊥⊥. ClearlyPe∈E′ ehe, xni ∈ E′⊥⊥ ⊆ E⊥⊥ and so xn = x′ +Pe∈E′ ehe, xni ∈ E⊥⊥. Together with the pre- Clearly E is an othonormal set of n or fewer elements. By the induction assump- tion and 160 IX, we know xi ∈ E′⊥⊥ ⊆ E⊥⊥ for i 6 1 6= n − 1. For any d ∈ E⊥ we have hd, x′i = hd, uihx′, x′i vious, we see {x1, . . . , xn}⊥⊥ ⊆ E⊥⊥⊥⊥ = E⊥⊥ ⊆ {x1, . . . , xn}⊥⊥, thus E⊥⊥ = {x1, . . . , xn}⊥⊥. By 160 IX we know E is an orthonormal basis of E⊥⊥, which completes the proof by induction. Solution to 161 II. For brevity write ℓ2 ≡ ℓ2((pi)i∈I ). We will first proof that ℓ2 is a right B-module. Assume (ai)i, (bi)i ∈ ℓ2. We want to show (ai + bi)i ∈ ℓ2. First, we to show Pi(ai + bi)∗(ai + bi) is bounded. Pick A, B ∈ R+ withPi a∗i ai 6 A andPi b∗i bi 6 B which exist as (ai)i and (bi)i are ℓ2. Let f be any normal state on B and S ⊆ I some finite subset. Then f kai + bik2 kaik2 (ai + bi)∗(ai + bi)(cid:17) = Xi∈S 6 Xi∈S f + kbik2 6 A + B + 2Xi∈S f(cid:1) 1 By Cauchy -- SchwarzPi∈S kaikfkbikf 6(cid:0)Pi∈S kaik2 f(cid:1) 1 2(cid:0)Pi∈S kbik2 converging sum of positive elementsPi(ai +bi)∗(ai +bi) 6 A+B +(AB) Suppose b ∈ B and (ai)i ∈ ℓ2. Then b∗(cid:0)Pi a∗i ai(cid:1)b =Pi(aib)∗aib as x 7→ Next we will show h(ai)i, (bi)ii ≡Pi a∗i bi defines a B-valued inner product on B. Pick A, B ∈ R+ with Pi a∗i ai 6 A and Pi b∗i bi 6 B. For any finite (cid:12)(cid:12)(cid:12)f(cid:16)Xi∈S 2 6 (AB) As normal states are order separating, we see that we have a bounded and thus 2 . Sup- pose i ∈ I. It remains to be shown ⌈(ai + bi)(ai + bi)∗⌉ 6 pi. Recall from 59 VI that ⌈xx∗⌉ 6 pi if and only if pix = x. Clearly pi(ai + bi) = piai + pibi = ai + bi as (ai)i, (bi)i ∈ ℓ2 and so indeed (ai + bi)i ∈ ℓ2. b∗xb is normal by 44 VIII and so (aib)i is ℓ2. Furthermore piaib = aib for any i ∈ I, so (aib)i ∈ ℓ2. We have shown ℓ2 is a right B-module with coordinatewise operations. on ℓ2. First we have to show the sum converges. Suppose f is any normal state a∗i bi(cid:17)(cid:12)(cid:12)(cid:12) 6 Xi∈S [ai, bi]f 6 Xi∈S where we used Cauchy -- Schwarz for B-valued inner products in the second in- equality and classic Cauchy -- Schwarz in the final inequality. We have shown thatPi∈S a∗i bi is a norm-bounded net in S. We claim it is ultraweakly Cauchy as well. For now, pick any finite sets S, T ⊆ I. Assume f is any normal state subset S ⊆ I we have kaikfkbikf 6 (AB) 1 2 . 1 1 2 , 207 on B. We want to show that the following quantity vanishes for sufficiently large S ∩ T . (4.2) (4.3) a∗i bi −Xi∈T (cid:12)(cid:12)(cid:12)f(cid:16)Xi∈S f (a∗i bi)(cid:12)(cid:12)(cid:12). Note that [(ai)i, (bi)i] ≡Pi∈S−T f (a∗i bi) is an inner product and so f (b∗i bi)(cid:17) 1 f (a∗i bi)(cid:12)(cid:12)(cid:12) + (cid:12)(cid:12)(cid:12) Xi∈T−S f (a∗i ai)(cid:17) 1 2(cid:16) Xi∈S−T a∗i bi(cid:17)(cid:12)(cid:12)(cid:12) 6 (cid:12)(cid:12)(cid:12) Xi∈S−T f (a∗i bi)(cid:12)(cid:12)(cid:12) 6 (cid:16) Xi∈S−T (cid:12)(cid:12)(cid:12) Xi∈S−T 2 . small by picking sufficiently large S ∩ T . And so (with a similar argument for the other factor), we see that (4.3) vanishes, which is the left term of (4.2). The The sumPi f (a∗i ai) converges and soPi∈S−T f (a∗i ai) can be made arbitrarily argument for the other term of (4.2) is the same. ThusPi∈S a∗i bi is ultraweakly Pi a∗i bi is an inner product on ℓ2. We claim this inner product is definite. Assume 0 = h(ai)i, (ai)ii ≡Pi a∗i ai. Then a∗i ai = 0 for each i ∈ I and so ai = 0, Cauchy and converges by 76 III. From the fact that a 7→ a∗, a 7→ ab and (a, b) 7→ a + b are all ultraweakly continuous, it follows readily that h(ai)i,i(bi)ii ≡ which is to say (ai)i = 0. Indeed our inner product is definite. Thus ℓ2 is a pre-Hilbert B-module. Write E ≡ {δi; i ∈ I}, where (δi)j = 0 for i 6= j and (δi)i = pi. Clearly E is an orthonormal set. We claim it's an orthonormal basis of ℓ2. Assume (ai)i ∈ ℓ2. It is easy to see Pi∈I δiai = (ai)i and so Pδi∈E δihδi, (ai)ii = Pi∈I δiai = (ai)i. It remains to be shown thatPδi∈E δibδi converges ultranorm for any ℓ2 summable (bi)δi∈E, which indeed it does to (bδi)i∈I as we already saw. Thus ℓ2 is self dual. For the final part of the exercise, assume X is a self-dual Hilbert B-module over some von Neumanna algebra B with orthonormal basis E ⊆ X. De- fine ϑ : X → ℓ2((he, ei)e∈E) by ϑ(x) ≡ (he, xi)e. Clearly ϑ is B-linear. It also preserves the inner product: hx, yi = hx,Pe ehe, yii = Pehx, ehe, yii = Pehx, eihe, yi = hϑ(x), ϑ(y)i. which entails it's injective. To show ϑ is sur- jective, let (xe)e ∈ ℓ2((he, ei)e∈E) be given. The family (xe)e is ℓ2 so Pe ebe converges ultranorm. Clearly ϑ(Pe ebe) = (he,Pe ebei)e = (be)e, so ϑ is indeed surjective. It follows ϑ is an isomorphism X ∼= ℓ2((he, ei)e∈E). Solution to 161 IV. Let X be a self-dual Hilbert B-module X for some von Neumann algebra B with orthonormal basis (ei)i∈I . Assume (ui)i∈I is a family partial isometries with uiu∗i = hei, eii. We will show (eiui)i∈I is an orthonormal basis of X. For brevity, write di ≡ eiui. To start, note hdi, dji = u∗i hei, ejiuj, If i = j, then hdi, dii = u∗i hei, eiiui = u∗i ui which which is zero if i 6= j. is a projection. So (di)i is orthonormal. Hence Pi dibi converges for any ℓ2- family (di)i. It remains to be shown that x =Pi dihdi, xi. NotePi dihdi, xi = Pi eiheiuiu∗i , xi = Pi eiheihei, eii, xi = Pi eihei, xi = x, so indeed (di)i is an orthonormal basis of X. For the next part, assume (pi)i∈I and (qi)i∈I are projections with pi ∼ qi. Let pi denote the partial projection with u∗i ui = qi and uiu∗i = pi. Con- sider ℓ2((pi)i∈I ). Define ((δi)i) ∈ ℓ2 by (δi)j = 0 if i 6= j and (δi)i = pi. In 161 II we saw (δi)i is an orthonormal basis of ℓ2((pi)i∈I ). Note uiu∗i = pi = hδi, δii. By the previous δiui is another orthonormal basis of ℓ2((pi)i∈I ). By the sec- ond part of 161 II we see ℓ2((pi)i∈I ) ∼= ℓ2((hδiui, δiuii)i∈I ) = ℓ2((u∗i piui)i∈I ) = ℓ2((qi)i∈I ), as promised. Solution to 161 V. Let X be a self-dual Hilbert B-module. Assume E ⊆ X and e1, e2 ∈ X such that E ∪ {e1, e2} is an orthonormal basis and he1, e1i + he2, e2i 6 1 We will show that E′ ≡ E ∪ {e1 + e2} is an orthonormal basis as well. For brevity, write p1 ≡ he1, e1i and p2 ≡ he2, e2i. Clearly E itself is an orthonormal set. For any e ∈ E we have he, e1 + e2i = he, e1i + he, e2i = 0 and so E′ is an orthogonal set. By assumption p1 and p2 are projections with p1 + p2 6 1 and so by 55 XIII they are orthogonal and thus in particular p1 + p2 is again a projection. Hence he1 + e2, e1 + e2i = p1 + p2 is a projection. Thus E′ is an orthonormal set. Let x ∈ X be given. By 149 III we have e2 = e2p2 and so e2he1 + e2, xi = e2h(e1 + e2)p2, xi = e2he2, xi. Similarly e1he1 + e2, xi = e1he1, xi. Thus x = shows the first condition on an orthonormal set to be an orthonormal basis. The second (and final) condition holds automatically as E′ is an orthonormal set and X is ultranorm complete. Thus E′ is indeed an orthonormal basis. e1he1, xi + e2he2, xi +Pe∈E ehe, xi = (e1 + e2)he1 + e2, xi +Pe∈E ehe, xi, which The the last part of the exercise, assume p, q ∈ B are projections with p+q 6 1. Clearly {p + q} is an orthonormal basis of (p + q)B and so is {p, q} by the previous. Hence by 161 II we see (p + q)B ∼= ℓ2({p, q}) = pB⊕ qB, as promised. Solution to 165 V. Assume X is a self-dual Hilbert A -module and Y is a self- dual Hilbert B-module for von Neumann algebras A and B. 1. Assume x1, x2, x ∈ X and y1, y2, y ∈ Y . Then x1ihx2 ⊗ y1ihy2 x ⊗ y = (x1ihx2 x) ⊗ (y1ihy2 y) = (x1hx2, xi) ⊗ (y1hy2, yi) = (x1 ⊗ y1)(hx2, xi ⊗ hy2, yi) = (x1 ⊗ y1)hx2 ⊗ y2, x ⊗ yi = x1 ⊗ y1ihx2 ⊗ y2 x ⊗ y. This is sufficient to show that x1ihx2 ⊗ y1ihy2 = x1 ⊗ y1ihx2 ⊗ y2, either by appealing to the defining universal property of X ⊗ Y or by the related property that the A ⊙ B-linear span of {x ⊗ y; x ∈ X, y ∈ Y } is 209 ultranorm dense in X ⊗ Y . (In fact the C-linear span is already ultranorm dense.) 2. Clearly, for any x ∈ X and y ∈ Y we have (1 ⊗ 1) (x ⊗ y) = 1 x⊗ y and so by same token (as the conclusion of the previous point) we see 1 = 1 ⊗ 1. 3. Assume x ∈ X, y ∈ Y , S, S′ ∈ Ba(X) and T, T ′ ∈ Ba(Y ). Then (S ⊗ T )(S′ ⊗ T ′) x ⊗ y = (SS′x) ⊗ (T T ′y) = (SS′) ⊗ (T T ′) x ⊗ y, which is sufficient to show (SS′) ⊗ (T T ′) = (S ⊗ T ) ⊗ (S′ ⊗ T ′) (see the conclusion of the first point.) 4. Assume S ∈ Ba(X), T ∈ Ba(Y ), x1, x2 ∈ X and y1, y2 ∈ Y . Then h(S ⊗ T )∗ x1 ⊗ y1, x2 ⊗ y2i = hx1 ⊗ y1, (S ⊗ T ) x2 ⊗ y2i = hx1 ⊗ y1, (Sx2) ⊗ (T y2)i = hx1, Sx2i ⊗ hy1, T y2i = hS∗x1, x2i ⊗ hT ∗y1, y2i = h(S∗ ⊗ T ∗) x1 ⊗ y1, x2 ⊗ y2i. This is sufficient (see the conclusion of the first point) to show that the vector functionals for (S⊗T )∗x1⊗y1 and (S∗⊗T ∗)x1⊗y1 agree. Hence (S⊗ T )∗x1⊗y1 = (S∗⊗T ∗)x1⊗y1. In turn this is sufficient to show (S⊗T )∗ = S∗ ⊗ T ∗, as desired. Solution to 169 XI. Assume ϕ : A → B is an ncp-map between von Neumann algebras. 1. Assume c : B → C is a filter and that (P, , h) is a Paschke dilation of ϕ. We will show that (P, , c◦ h) is a Paschke dilation of c◦ ϕ. To this end, assume : A → P′ is some nmiu-map and h′ : P′ → C is any ncp-map with h′ ◦ ′ = c◦ ϕ. Assume ϕ(1) 6 1. (We will reduce the general to this one later on.) Note h′(1) = h′(′(1)) = c(ϕ(1)) 6 c(1) and so by the universal property of c, there is a unique ncp-map map h′′ : P′ → B with c◦ h′′ = h′. Then c◦ ϕ = h′ ◦ ′ = c◦ h′′ ◦ ′. And so ϕ = h′′ ◦ ′ as c is injective, see 169 XII. There is a unique ncp-map σ : P′ → P with h◦ σ = h′′ and σ ◦ ′ = . Then h′ = c◦ h′′ = c◦ h◦ σ. As- sume σ′ : P′ → P is any ncp-map with σ′ ◦ = and c◦ h◦ σ′ = h′. Then c◦ h◦ σ′ = h′ = c◦ h′′ and so h◦ σ′ = h′′. Hence σ = σ′ by the uniqueness of σ. Thus (P, , c◦ h) is a Paschke dilation of c◦ ϕ. in the general case, assume ϕ(1) 66 1. Then ϕ(1) 6= 0. De- Now, fine ϕ′ ≡ kϕ(1)k−1ϕ. Then ϕ′(1) = kϕ(1)k−1ϕ(1) 6 1. By 140 X we know (P, ,kϕ(1)k−1h) is a Paschke dilation of ϕ′. Now we are in our previous case and so (P, ,kϕ(1)k−1c◦ h) is a Paschke dilation of c◦ ϕ′. Applying 140 X once again, we conclude (P, , c◦ h) is a Paschke dilation of c◦ ϕ as desired. 2. Assume c′ : C ′ → B is a filter of ϕ(1). By the universal property of c′, there is a unique ncp-map ϕ′ : A → C ′ with c′ ◦ ϕ′ = ϕ. Then c′(ϕ′(1)) = ϕ(1) = c′(1) and so by injectivity of c′ (see 169 XII) we get ϕ′(1) = 1 -- that is to say: ϕ′ is unital. For the final part of this exercise, assume (P, , h) is a Paschke dilation of ϕ′. By the previous point, we know (P, , c′ ◦ h) is a Paschke dilation of c′ ◦ ϕ′ = ϕ, as desired. Solution to 169 XII. Let c : C → B be some filter for b ∈ B. By 169 X we know cb : ⌈b⌉ B ⌈b⌉ → B given by a 7→ √ba√b is also a filter for b. By the universal property of c, there is a unique ncp-map ϑ1 with c◦ ϑ1 = cb. In the other direction, by the universal property of cb, there is a unique ncp-map ϑ2 with cb ◦ ϑ2 = c. Hence c◦(ϑ1 ◦ ϑ2) = cb ◦ ϑ2 = c◦ id. Clearly c(1) 6 c(1) and so by the universal property of c the map id is the unique ncp-map with c◦ id = c, hence ϑ1 ◦ ϑ2 = id. Reasoning similarly on the other side, we see ϑ2 ◦ ϑ1 = id and so ϑ2 is an isomorphism. Recall c = cb ◦ ϑ2 and so it is sufficient to show cb is injective. To this end, let a1, a2 ∈ ⌈b⌉ B ⌈b⌉ be given with cb(a1) = cb(a2); i.e. √ba1√b = √ba2√b. Then a1 = a2 by 60 VIII and so cb is indeed injective. Solution to 170 IV. Assume : A → B is a surjective nmiu-map between von Neumann algebras. We already saw in 138 III that the kernel of an nmiu-map is an ultraweakly-closed two-sided ideal and hence a principal ideal of a central projection by 69 II. Write z for the central projection of A with ker = (1−z)A ; hz : A → zA for the map hz(a) = za and ′ : zA → B for the restriction of to zA . Clearly ′ ◦ hz = and so ′ is surjective. Furthermore ′ is injective, because for any a ∈ zA with (a) = 0 we have (a) = 0 and so a ∈ It is easy to see that ′ as a bijective miu- ker = (1 − z)A , hence a = 0. map has an miu inverse and so it is an miu-isomorphism. Consequently, it is an nmiu-isomorphism (as miu-maps are order-preserving) and in particular an ncp-isomorphism. The map hz is a standard corner for z (see 169 IV) and so = ′ ◦ hz is also a corner for z as ′ is an ncp-isomoprhism. Conversely, assume h : A → C is a corner for a central projection z ∈ A . The map hz : A → zA given by a 7→ za is another corner for z, see 169 IV. By the universal property of h, there is a unique ncp-map ϑ1 : zA → C with h◦ ϑ1 = ϑ2. On the other side, by the universal property of hz, there is a unique ncp-map ϑ2 : C → zA with hz ◦ ϑ2 = ϑ1. Again, by the uni- versal property of h, the map id : C → C is the unique ncp-map C → C with h = h◦ id. Note h = hz ◦ ϑ2 = h◦(ϑ1 ◦ ϑ2) and so id = ϑ1 ◦ ϑ2. Simi- larly ϑ2 ◦ ϑ1 = id. Thus ϑ1 is an ncp-isomorphism and consequently an nmiu- 211 isomoprhism by 99 IX. Clearly hz is a surjective nmiu-map and thus h = hz ◦ ϑ2 is a surjective nmiu-map as well. Solution to 174 III. Reflexivity is easy: a > 0 = 0 > a = a by the partial commutativity and zero axiom, so a 6 a. For transitivity, assume a 6 b and b 6 c. Pick d, e ∈ M with a > d = b and b > e = c. Then by partial associativity c = b > e = (a > d) > e = a > (d > e) so a 6 c. Solution to 175 III. Assume E and F are effect algebras. We define a > b for a, b ∈ E × F whenever both a1 ⊥ b1 and a2 ⊥ b2 and in that case a > b ≡ (a1 > b1, a2 > b2). Furthermore, define 1 ≡ (1, 1) and 0 ≡ (0, 0). We claim that these operations turn E × F into an effect algebra. We start with partial commutativity. Assume a, b ∈ E × F with a ⊥ b. Then a1 ⊥ b1 and a2 ⊥ b2, hence b1 ⊥ a1, b2 ⊥ a2 and a >b = (a1 >b1, a2 >b2) = (b1 > a1, b2 > a2) = b > a. We continue with partial associativity. Assume a, b, c ∈ E × F with a ⊥ b and a > b ⊥ c. Then ai ⊥ bi, ai > bi ⊥ ci for i = 1, 2 and so bi ⊥ ci and (ai > bi) > ci = ai > (bi > ci). Consequently (a > b) > c = a > (b > c). Next the zero axiom: for any a ∈ E × F we have a1 ⊥ 0, a2 ⊥ 0, 0 > a1 = a1 and 0 > a2 = a2. Thus 0 ≡ (0, 0) ⊥ (a1, a2) and 0 > a = a. To prove the orthocomplement law, assume a ∈ E×F . Then for a ≡ (a⊥1 , a⊥2 ) we have a ⊥ a⊥ and a > a⊥ = (1, 1) ≡ 1. Furthermore, if a ∨ b = 1 for some b ∈ E × F , then b1 = a⊥1 and b2 = a⊥2 , hence b = a⊥. Only the zero -- one law remains. Asumme a ⊥ 1 for some a ∈ E × F . Then a1 ⊥ 1 and a2 ⊥ 1, hence a1 = 0 and a2 = 0. Thus a = 0, as desired. We have shown E × F is an effect algebra with componentwise operations. Write π1 : E×F → E and π2 : E×F → F for the maps given by π1(a1, a2) = a1 and π2(a1, a2) = a2. Clearly π1 and π2 are additive and unital, hence effect algebra homomorphisms. We will show that E × F with projections π1 and π2 forms a categorical product of E and F in EA. To this end, assume f1 : G → E and f2 : G → F are effect algebra homomorphisms for some effect algebra G. Define g : G → E × F by g(a) = (f1(a), f2(a)). Clearly g is the unique map with πi ◦ g = fi for i = 1, 2. It is also an effect algebra homomorphism and so E × F is indeed the product of E and F . Solution to 175 IV. Let E be an effect algebra with any a ∈ E. We will first show that 1⊥ = 0 and a⊥⊥ = a. By the orthocomplement law 1 > 1⊥ = 1, so 1⊥ ⊥ 1, hence 1⊥ = 0 by the zero -- one law. By the orthocomplement law a > a⊥ = 1. Thus a⊥ > a = a > a⊥ = 1 by partial commutativity. Hence a = a⊥⊥ by uniqueness of the orthocomplement. Now, to prove the zero law, note a⊥ > a⊥⊥ = 1 = (a⊥ > a⊥⊥) > (a⊥ > a⊥⊥)⊥ = (a⊥ > a⊥⊥) > 0 = a⊥ > (a⊥⊥ > 0) by the orthocomplement law and partial associativity. Thus by uniqueness of the orthocomplement a⊥⊥ = a⊥⊥ > 0, hence 0 > a = a by partial commutativity and a⊥⊥ = a. Solution to 176 II. We start with (D1). Assume b 6 a. Then b > c = a for some c. By definition c = a ⊖ b. Conversely, assume a ⊖ b is defined. Then b > (b ⊖ a) = a, hence b 6 a. To prove (D2), assume a ⊖ b is defined. Then a ⊖ b 6 b > (a ⊖ b) = a as desired. For (D3) assume a ⊖ b and a ⊖ (a ⊖ b) are defined. Note (a ⊖ b) > b = b > (a ⊖ b) = a and so b = a ⊖ (a ⊖ b). (D4) is next. Assume a 6 b 6 c. Then (c ⊖ b) > (b ⊖ a) > a = c = (c ⊖ a) > a and so by cancellation c⊖ b 6 (c⊖ b) > (b⊖ a) = c⊖ a. Note ((c⊖ a)⊖ (c⊖ b)) > (c⊖b) >a = c = (b⊖a) >a >(c⊖b) and so by cancellation (c⊖a)⊖(c⊖b) = b⊖a. For the final part of the exercise, assume E is a poset with maximum ele- ment 1 and a partial operation ⊖ satisfying (D1), (D2), (D3) and (D4). We will show E is an effect algebra with a > b = c ⇔ a = c ⊖ b and 0 = 1 ⊖ 1. First we have to show that if c1⊖b = c2⊖b for some b 6 c1, c2 in E, then c1 = c2 so that > is at most single-valued. By (D4) we have (1⊖ b)⊖ (1⊖ c1) = c1⊖ b and so (1 ⊖ b) ⊖ a = (1 ⊖ b) ⊖ (c1 ⊖ b) = (1 ⊖ b) ⊖ ((1 ⊖ b) ⊖ (1 ⊖ c1)) = 1 ⊖ c1 by (D3). Similarly 1 ⊖ c2 = (1 ⊖ b) ⊖ a, so with another application of (D3) we see c1 = 1 ⊖ (1 ⊖ c1) = 1 ⊖ (1 ⊖ c2) = c2. The first axiom we will prove is partial commutativity. Assume a, b ∈ E with a ⊥ b. By definition, there is some c > b with c ⊖ b = a. By (D3) we have c⊖ a = c⊖ (c⊖ b) = b and so b ⊥ a with b > a = c = a > b as desired. Note that we have also shown that a 6 a > b and (a > b) ⊖ a = b for a ⊥ b. Next, to prove partial associativity, assume a, b, c ∈ E with a ⊥ b and a >b ⊥ c. Then a 6 a > b 6 (a > b) > c and thus b = (a > b) ⊖ a (D4) = (cid:0)((a > b) > c) ⊖ a(cid:1) ⊖ (cid:0)((a > b) > c) ⊖ (a > b)(cid:1) = (cid:0)((a > b) > c) ⊖ a(cid:1) ⊖ c, which shows b ⊥ c with b>c = ((a>b)>c)⊖a. From that it immediately follows that b > c ⊥ a with (b > c) > a = (a > b) > c and so a ⊥ b > c with (a > b) > c = a > (b > c) by partial commutativity. We continue with the zero axiom. Assume a ∈ E. Clearly a 6 1 6 1 and so (1 ⊖ a) ⊖ (1 ⊖ 1) = 1 ⊖ a by (D4), which shows 1 ⊖ a ⊥ 1 ⊖ 1 ≡ 0 with (1 ⊖ a) > 0 = 1 ⊖ a. The zero law follows by partial commutativity. Now we prove the orthocomplement law. Assume a ∈ E. By (D3) we have 1 ⊖ (1 ⊖ a) = a, which demonstrates a ⊥ 1 ⊖ a ≡ a⊥ with a > a⊥ = 1. Assume b ∈ E with a > b = 1. By definition, we have a = 1 ⊖ b and so by (D3) we see b = 1 ⊖ (1 ⊖ b) = 1 ⊖ a ≡ a⊥. To show the last axiom (zero -- one), assume a ∈ E with a ⊥ 1. Then 1 6 a > 1 6 1 and so a > 1 = 1. Hence a = (a > 1)⊖ 1 = 1⊖ 1 ≡ 0. We have shown E is an effect algebra. The conclude this exercise, we show that the original order and ⊖ coincide with those defined for E as an effect algebra. For clarity, write 6EA and ⊖EA for the latter. 213 To start, assume a 6 b for some a, b ∈ E (as a D-poset). Then b = a > (b⊖ a) as trivially b⊖a = b⊖a. Hence a 6EA b. To prove the converse, assume a 6EA b. Then a > c = b for some c ∈ E. Hence a 6 a > c 6 b. Indeed 6=6EA. Assume a, b ∈ E. The element a ⊖EA b is defined if and only if b 6 a, which is precisely when a ⊖ b is defined. Furthermore b > (a ⊖EA b) = a and so a ⊖EA b = (b > (a ⊖EA b)) ⊖ b = a ⊖ b. Solution to 176 V. Assume f : E → F is an homomorphism between effect algebras. 1. We have 0 ⊥ 1 and so f (0) ⊥ f (1) = 1, hence f (0) = 0 by the zero -- one law. 2. Assume a 6 b in E. Then a > c = b for some c ∈ E. Hence f (b) = f (a > c) = f (a) > f (c), which shows f (a) 6 f (b). 3. Assume a ⊖ b is defined (i.e. a > b). Recall (a ⊖ b) > b = a and so f (a ⊖ b) > f (b) = f (a). Hence f (a ⊖ b) = f (a) ⊖ f (b). 4. Assume a ∈ E. In 176 II we saw 1 ⊖ a = a⊥ and so by the previous point f (a⊥) = f (1 ⊖ a) = f (1) ⊖ f (a) = 1 ⊖ f (a) = f (a)⊥. Solution to 178 IIIa. Let M be an effect monoid with any a ∈ M . Then a⊙ 1 = a⊙(0>1) = (a⊙0)>(a⊙1). Hence by cancellation a⊙0 = 0. Similarly 0⊙a = 0. Solution to 178 V. Assume M is an effect monoid with a1, . . . , an, b1, . . . , bn ∈ M such that >i ai = 1 and >i ai ⊙ bi = 1. Note that for any 1 6 i 6 n we have ai ⊙ bi 6 ai as (ai ⊙ bi) > ai ⊙ b⊥i = ai ⊙ 1 = ai. Thus a⊥i = > j6=i aj > > j6=i aj ⊙ bj = (ai ⊙ bi)⊥ > a⊥i and so a⊥i = (ai ⊙ bi)⊥, hence ai = ai ⊙ bi. Solution to 181 IX. Let C be an effectus in partial form. Assume f : X → A and g : X → B are any arrows in C with 1 ◦ f ⊥ 1 ◦ g. 1. Assume [a, b] : A + B → Y is any map. Then 1 ◦ a◦ f 6 1 ◦ f ⊥ 1 ◦ g > 1 ◦ b ◦ g and so 1 ◦ a◦ f ⊥ 1 ◦ b ◦ g. Hence a◦ f ⊥ b ◦ g and so [a, b]◦hf, gi = [a, b]◦((κ1 ◦ f ) > (κ2 ◦ g)) = ([a, b]◦(κ1 ◦ f )) > ([a, b]◦(κ2 ◦ g)) = (a◦ f ) > (b ◦ g). 2. By the previous point and 181 IV, we have 1 ◦hf, gi = [1, 1]◦hf, gi = (1 ◦ f ) > (1 ◦ g). 3. Assume k : A → A′ and l : B → B′ are two maps in C. Then by the first point of this exercise, we have (k + l)◦hf, gi ≡ [κ1 ◦ k, κ2 ◦ l]◦hf, gi = (κ1 ◦ k ◦ f ) > (κ2 ◦ l ◦ g) ≡ hk ◦ f, l ◦ gi. 4. Assume k : X′ → X is some map in C. Then by PCM-enrichment of C, we get hf, gi◦ k ≡ ((κ1 ◦ f ) > (κ2 ◦ g))◦ k = (κ1 ◦ f ◦ k) > (κ2 ◦ g ◦ k) ≡ hf ◦ k, g ◦ ki. Solution to 183 II. Assume h1, h2 : Q → P are two arrows with m1 ◦ h1 = m1 ◦ h2 and m2 ◦ h1 = m2 ◦ h2. Note (f ◦ m1)◦ h1 = g ◦ m2 ◦ h1 = (g ◦ m2)◦ h2 and so by the universal property of a pullback there is a unique map h : Q → P with m1 ◦ h = m1 ◦ h1 and m2 ◦ h = m2 ◦ h2 and so h1 = h2 = h. Solution to 183 III. For the first point, assume the left and right inner squares are pullbacks. To prove the outer square is a pullback, assume α : S → C and β : S → X are any maps with g′ ◦ f′ ◦ β = m◦ α. As the right inner square is a pullback, there is a unique map γ : S → B with g ◦ γ = α and l ◦ γ = f′ ◦ β. By this last equality and the fact that the left inner square is a pullback, there is a unique map δ : S → A with f ◦ δ = γ and k ◦ δ = β. We will prove that this δ is also the mediating map for the outer square. Clearly g ◦ f ◦ δg ◦ γ = α. Assume δ′ : S → A is some other map with g ◦ f ◦ δ′ = α and k ◦ δ′ ◦ β. Note that g ◦ f ◦ δ′ = α = g ◦ γ and l ◦ f ◦ δ′ = f′ ◦ k ◦ δ′ = f′ ◦ β = l ◦ γ. Thus f ◦ δ′ = γ as g and l are jointly monic by 183 II. Consequently δ = δ′ and so the outer square is indeed a pullback. For the second point, assume the outer square is a pullback and that l and g are jointly monic. We will prove that the left inner square is a pullback. To this end, assume α : S → B and β : S → X are maps with l ◦ α = f′ ◦ β. Note m◦ g ◦ α = g′ ◦ l ◦ α = g′ ◦ f′ ◦ β and thus by the fact that the outer square is a pullback, there is a unique γ : S → A with g ◦ f ◦ γ = g ◦ α and k ◦ γ = β. This will also be the unique mediating map demonstrating that the left inner square is a pullback. We have l ◦ f ◦ γ = f′ ◦ k ◦ γ = f′ ◦ β = l ◦ α and so by the joint monicity of l and g, we conclude f ◦ γ = α. Assume there is some map γ′ : S → A with f ◦ γ′ = α and k ◦ γ′ = β. Then clearly g ◦ f ◦ γ′ = g ◦ α and so by the uniqueness of γ as a mediating map for the outer square, we find γ′ = γ. Thus the left inner square is a pullback as well. Solution to 186 II. Let f : X ⇀ Z and g : Y ⇀ Z be two partial maps in C; that is f : X → Z + 1 and g : Y → Z + 1. Then the cotuple [f, g] : X + Y → Z + 1 (with respect to the coproduct that we assumed to exist) is also a partial map X + Y ⇀ Z. Note [f, g] ◦ κ1 = [[f, g], κ2]◦ κ1 ◦ κ1 = f and [f, g] ◦ κ2 = [[f, g], κ2]◦ κ1 ◦ κ2 = g. Suppose h : X + Y ⇀ Z is a partial map with h ◦ κ1 = f and h ◦ κ2 = g. Then h◦ κ1 = h ◦ κ1 = f and h◦ κ2 = h ◦ κ2 = g, so h = [f, g]. This shows X + Y is also a coproduct in Par C with coprojectors κ1 and κ2. To show 0 is an initial object of Par C, assume X is some object of Par C, i.e. of C. As 0 is initial in C, there is a unique map ! : 0 → X + 1 in C and so there is a unique map ! : 0 ⇀ X in Par C. Solution to 186 VII. We already know that the initial object 0 of C is also initial 215 in Par C. We will show that it is also final in Par C. To this end, assume X is some object in Par C, i.e. of C. There is a unique map ! : X → 1 in C and so there is a unique map ! : X ⇀ 0 in Par C. Thus 0 is final in Par C. For any objects X and Y from Par C, the zero map 0 : X → Y in Par C is by definition equal to i ◦ f , where f : X ⇀ 0 is the unique final map and i : 0 →⇀ Y is the unique initial map. Unfolding definitions, we see 0 = i ◦ f = [i, κ2]◦ f = κ2 ◦! as desired. Solution to 189 I. For the first point, assume C is an effectus in total form together with a map f : A → 0. We will show that f is an isomorpism. By initiality of 0, we know f ◦!A =!0 = id. By 185 I, the following square is a pullback. ❴ 0 ! A 0 / 0 f As id0 ◦ f = f ◦ idA, there is a unique map h : A → 0 with id◦ h = f and !◦ h = id. So h = f and !◦ f =!◦ h = id. Thus f is indeed an isomorphism with inverse !. For the second point, assume C is an effectus in total and partial form. Let X and Y be any two objects. As 0 is final in an effectus in total form, there are maps !X : X → 0 and !Y : Y → 0. As 0 is a strict initial object in an effectus in partial form, both !X and !Y are isomorphisms. Hence !Y ◦!−1 X is an isomorphism between X and Y . Solution to 191 VIII. To show Rngop is an effectus in total form, we shown the dual axioms for Rng. Clearly, for any two unit rings R and S, their cartesian product R×S is a categorical product with projectors π1(r, s) = r and π2(r, s) = s. The integer ring Z is the initial object of Rng. The zero ring (the unique unit ring with a single element 0 = 1) is the final object of Rng. To show the pushout diagrams corresponding to (3.1) hold, assume we are given unit rings R, S, T with unit-preserving homomorphisms α, β, δ that make the outer squares of the following diagrams commute. T α T ! β R × S !×id Z × S id×! id×! R × Z !×id Z × Z δ R π1 ! Z π1 R × S !×! Z × Z We have to show that there are unique dashed arrows (as shown) that make ✤ / b b o o o o O O Q Q o o O O b b o o o o O O Q Q o o O O these diagrams commute. We start with the left diagram. Define f : R × S → T by f (r, s) ≡ α(r, 0) + β(0, s). Clearly f is additive and f (0, 0) = α(0, 0) + β(0, 0) = 0. By assumption α(n, m) = β(n, m) for any n, m ∈ Z and so α(1, 0) = β(1, 0) in particular, hence f (1, 1) = α(1, 0) + β(0, 1) = β(1, 1) = 1. It remains to be shown that f is multiplicative. First note that α(1, 0)β(0, 1) = β(1, 0)β(0, 1) = β(0, 0) = 0. Thus α(r, 0)β(0, s) = α(r, 0)α(1, 0)β(0, 1)β(0, s) = 0 for any r ∈ R and s ∈ S and similarly β(0, s)α(r, 0) = 0. Hence f (r, s)f (r′, s′) = α(r, 0)α(r′, 0) + β(0, s)β(0, s′) = α(rr′, 0) + β(0, ss′) = f (rr′, ss′). We have shown f is a unit-preserving ring homomorphism. Clearly f (r, m) = α(r, 0) + β(0, m) = α(r, 0) + α(0, m) = α(r, m) for any r ∈ R and m ∈ Z hence α = f ◦(id×!). Similarly β = f ◦(!◦ id). Assume f′ : R × S → T is any unit- preserving ring homomorphism with α = f′ ◦(id×!) and β = f′ ◦(!◦ id). Then clearly f′(r, 0) = α(r, 0) and f′(0, s) = β(0, s), so f′(r, s) = f′(r, 0) + f′(0, s) = α(r, 0) + β(0, s) = f (r, s). This shows the left square above is indeed a pushout. We continue with the diagram on the right. By assumption δ(n, m) = n for any n, m ∈ Z hence δ(0, s) = δ(0, s)δ(0, 1) = 0 for any s ∈ S. Thus δ(r, s) = δ(r, 0) + δ(0, s) = δ(r, 0) for any r ∈ R and s ∈ S. Define g : R → T by g(r) = δ(r, 0). Clearly g is additive, multiplicative and g(0, 0) = 0. Furthermore g(1) = δ(1, 0) = δ(1, 1) = 1, so it is a unit-preserving ring homomorphism. It is easy to see that g ◦ π1 = δ and that g is the unique such unit-preserving ring homo- morphism. We have shown the right diagram is a pushout too. To show Rngop is an effectus in total form, it only remains to be shown that hπ1, π2, π2i,hπ2, π1, π2i : Z × Z → Z × Z × Z are jointly epic. So as- sume f, g : Z×Z×Z → R are two unit-preserving ring homomorphisms for which we have f ◦hπ1, π2, π2i = g ◦hπ1, π2, π2i and f ◦hπ2, π1, π2i = g ◦hπ2, π1, π2i. By the first equality f (k, 0, 0) = g(k, 0, 0) and by the second f (0, l, 0) = g(0, l, 0) for any k, l ∈ Z. In particular f (0, 0, 1) = 1 − f (1, 0, 0)− f (0, 1, 0) = 1 − g(1, 0, 0)− g(0, 1, 0) = g(0, 0, 1) and so f (0, 0, m) = mf (0, 0, 1) = mg(0, 0, 1) = g(0, 0, m) for any m ∈ Z. Putting it all together: f (k, l, m) = f (k, 0, 0) = f (0, l, m) + f (0, 0, m) = g(k, 0, 0) = g(0, l, m) + g(0, 0, m) = g(k, l, m). So f = g and so we have shown the required joint epicity. We have shown Rngop is indeed an effectus in total form. We continue with the two additional points. Let p : Z × Z → R be any predicate on R. Then p(1, 0) ∈ R with p(1, 0)2 = p(1, 02) = p(1, 0) and so p(1, 0) is an idempotent. Furthermore p(0, 1) = p(1, 1) − p(1, 0) = 1 − p(1, 0), so p is fixed by the idempotent p(1, 0). Let e ∈ R be any idempotent. It is easy to see that p(n, m) ≡ ne + m(1 − e) is a unit-preserving ring homomorphism with p(1, 0) = e. Thus the predicates on R correspond to its idempotents. This also shows that the set of scalars of Rngop is the two-element effect monoid 2 as Z has exactly two idempotents: 0 and 1. Assume p ⊥ q for two predicates p, q : Z× Z → R on R. Then, by definition, p ⊥ q if there is a b : Z× Z× Z → R with b(n, m, m) = p(n, m) and b(m, n, m) = q(n, m) for all n, m ∈ Z In particular p(1, 0)q(1, 0) = b(1, 0, 0)b(0, 1, 0) = 0, 217 thus the idempotents corresponding to p and q are orthogonal. Conversely, if p(1, 0) and q(1, 0) are orthogonal, then b(k, l, m) = kp(1, 0) + lq(1, 0) + m(1 − p(1, 0) − q(1, 0)) defines a unit-presrving ring homomorphism that shows p ⊥ q with (p > q)(1, 0) = b(1, 1, 0) = p(1, 0) + q(1, 0). From this it is also clear that p⊥(1, 0) = 1 − p(1, 0). There are many examples to show unit-preserving ring homomorphisms need not be fixed on their value on idempotents (and thus that Rngop does not have separting predicates.) For instance, let R denote the unit ring of continuous real- valued functions on the unit interval [0, 1]. This ring has only two idempotents: the functions that are constant 0 and 1, which are also the zero and unit element (respectively) of the ring. For every x ∈ [0, 1] the map δx : R → R given by δx(f ) = f (x) is a unit-preserving ring homomorphism. Clearly δx(0) = 0 = δy(0) and δx(1) = 1 = δy(1) for any x, y ∈ [0, 1] with x 6= y, but not δx = δy. Finally, we treat the second and last point of the exercise. Reasoning to- wards contradiction, let f : Z2 → Z be any unit-preserving ring homomorphism. Then 0 = f (0) = f (1 + 1) = f (1) + f (1) = 1 + 1 = 2, which is absurd. Thus there is no such homomorphism. However, there are two unit-preserving ring homomorphisms Z × Z → Z2 corresponding to the idempotents 0 and 1 in Z2. Hence the states cannot be separating (for otherwise there could at most be a single unit-preserving ring homomorphism Z × Z → Z2.) Solution to 192 III. We will first show that DM is a functor. To start, clearly DM (id)(p)(x) ≡ > y;id(y)=x p(y) = p(x) and so DM (id) = id. Let f : X → Y and g : Y → Z be given. Then DM (g ◦ f )(p)(z) ≡ > x; g(f (x))=z p(x) = > > p(x) y; g(y)=z x; f (x)=y = > y; g(y)=z(cid:0)(DM f )(p)(cid:1)(y) = (cid:0) (DM g)(cid:0)(DM f )(p)(cid:1)(cid:1)(z) and so DM (g ◦ f ) = (DM g)◦(DM f ). Next, we show that η is a natural transformation. Let f : X → Y be any map. Then for any x0 ∈ X, we have ((DM f )◦ ηX )(x0) = > x; f (x)=y (ηX (x0))(x) = > 1 x; f (x)=y x=x0 = (ηY (f (x0))) and so DM f ◦ ηX = ηY ◦ f , i.e. η is a natural transformation id ⇒ DM . To show µ is a natural transformation, assume f : X → Y is some map, y ∈ Y and Φ ∈ DMDM X. Then (DM f )(µX (Φ))(y) = > x; f (x)=y µX (Φ)(x) = > x; f (x)=y > p Φ(p) ⊙ p(x) = > p > x; f (x)=y Φ(p) ⊙ p(x) = > p Φ(p) ⊙ > x; f (x)=y p(x) = > p Φ(p) ⊙ (DM f )(p)(y) > p; (DM f )(p)=q q (cid:16) > = > q = > = > q Φ(p) ⊙ q(y) Φ(p)(cid:17) ⊙ q(y) p; (DM f )(p)=q (DMDM f )(Φ)(q) ⊙ q(y) = µY ((DMDM f )(Φ))(y) and so (DM f )◦ µX = µY ◦(DMDM f ), hence µ is a natural transformation DMDM ⇒ DM . We continue with the monad laws. Assume X is a set, x ∈ X and ℵ ∈ 219 DMDMDM X. Then µX (µDM X (ℵ))(x) = > p (µDM X )(ℵ)(p) ⊙ p(x) = > p > Φ ℵ(Φ) ⊙ Φ(p) ⊙ p(x) = > Φ ℵ(Φ) ⊙ > p Φ(p) ⊙ p(x) = > Φ = > p = > p ℵ(Φ) ⊙ µX (Φ)(x) > Φ; µX (Φ)=p ℵ(Φ) ⊙ p(x) (DM µX )(ℵ)(p) ⊙ p(x) = µX ((DM µX )(ℵ))(x) and so µ◦ µDM = µ◦(DM ). For any x ∈ X and p ∈ DM X, we have ηDM X (p)(q) ⊙ q(x) = p(x) µX (ηDM X (p))(x) = > q and so µ◦ ηDM = id. For the final monad law, assume x ∈ X and p ∈ DM X. Then µX ((DM ηX )(p))(x) = > q (DM ηX )(p)(q) ⊙ q(x) = > q = > y > y; ηX (y)=q p(y) ⊙ q(x) p(y) ⊙ ηX (y)(x) = p(x) and so µ◦DM η = id. We have shown (DM , η, µ) is a monad. Our next project is to show that Kℓ DM is an effectus in total form. It is easy to see that the coproduct X + Y from Set is also a coproduct in DM with coprojectors η ◦ κi, where κi are the coprojectors for X + Y in Set. (In fact, for any category C and monad T , the inclusion functor K : C → Kℓ T given by KX = X and Kf = η ◦ f is a left adjoint and so K preserves colimits.) The empty set is also the initial object of Kℓ DM . As DM 1 ∼= 1, the category Kℓ DM has as final object the one-element set 1. Now we will show that a square such as that on the left of (3.1) is a pull- back. So assume α : Z → DM (X + 1) and β : Z → DM (1 + Y ) are maps with (!+ id) ◦ α = ( id+!) ◦ β, where ◦ denotes the composition in Kℓ DM and f ≡ η ◦ f the Kleisli embedding. Assume z ∈ Z. By assumption α(z)(κ2∗) = ((! + id) ◦ α)(z)(κ2∗) = (( id + !) ◦ β)(z)(κ2∗) = >y β(z)(κ2y), where ∗ is the unique element of 1. We want to define δ : Z → DM (X + Y ) by δ(z)(κ1x) = α(z)(κ1x) and δ(z)(κ2y) = β(z)(κ2y), but first need to check the image of δ(z) sums Indeed >w δ(z)(w) = >y β(z)(κ2y) > >x α(z)(κ1x) = α(z)(κ2∗) > to 1. >x α(z)(κ1x) = 1. Clearly α = ( id+!) ◦ δ and β = (!+ id) ◦ δ. Suppose δ′ : Z → DM (X+Y ) is some map with α = ( id+!) ◦ δ′ and β = (!+ id) ◦ δ′. Then δ′(κ1x) = (( id + !) ◦ δ′)(κ1x) = α(κ1x) for any x ∈ X and similarly δ′(κ2y) = β(κ2y) for any y ∈ Y , so δ′ = δ. We have shown that the square on the left of (3.1) is indeed a pullback in Kℓ DM . To show that the square on the right of (3.1) is also a pullback, assume α : Z → DM 1 and β : Z → DM (X + Y ) are maps with (! +!) ◦ β = κ1 ◦ α. Assume z ∈ Z and y ∈ Y . Then β(z)(κ2y) 6 ((! + !) ◦ β)(z)(κ2∗) = (κ1 ◦ α)(z)(κ2∗) = 0. This allows us to define δ : Z → DM X by δ(z)(x) = β(κ1x) as >x δ(z)(x) = >x β(κ1x) = >x β(κ1x) > >y β(κ2y) = 1. Clearly κ1 ◦ δ = β and ! ◦ δ = α. Suppose δ′ : Z → DM X is some map with κ1 ◦ δ′ = β and ! ◦ δ′ = α. Then δ′(z)(x) = (κ1 ◦ δ′)(κ1x) = β(z)(κ1x) = δ(z)(x) and so δ = δ′. We have shown that the square on the right of (3.1) is a pullback in Kℓ DM . It only remains to be shown that the maps [κ1, κ2, κ2], [κ2, κ1, κ2] : 1+1+1 → DM (1 + 1) are jointly monic in Kℓ DM . To this end, assume f1, f2 : Z → DM (1 + 1 + 1) are given with [κ1, κ2, κ2] ◦ f1 = [κ1, κ2, κ2] ◦ f2 and [κ2, κ1, κ2] ◦ f1 = [κ2, κ1, κ2] ◦ f2. Assume z ∈ Z. Clearly and f1(z)(κ1∗) = ([κ1, κ2, κ2] ◦ f1)(z)(κ1∗) = ([κ1, κ2, κ2] ◦ f2)(z)(κ1∗) = f2(z)(κ1∗) f1(z)(κ2∗) = ([κ2, κ1, κ2] ◦ f1)(z)(κ1∗) = ([κ2, κ1, κ2] ◦ f2)(z)(κ1∗) = f2(z)(κ2∗). 221 So f1(z)(κ3∗) = (cid:0)f1(z)(κ1∗) > f1(z)(κ2∗)(cid:1)⊥ = (cid:0)f2(z)(κ1∗) > f2(z)(κ2∗)(cid:1)⊥ = f2(z)(κ3∗), hence f1 = f2. We have shown that Kℓ DM is an effectus. A scalar in Kℓ DM corresponds to a map λ : 1 → DM (1 + 1), which corre- sponds in turn to an element of M . Unfolding definitions, it is straight-forward to see that the effect monoid structure defined on the scalars is the same as that of M . Solution to 193 II. We start with the surjectivity of q, DM q and DMDM q. By definition q is clearly surjective. Pick a section r : X/∼ → X of q; i.e. q ◦ r = id. Then DM r and DMDM r are sections of DM q and DMDM q respectively, hence DM q and DMDM q are surjective. To prove the second point, assume ∼ is a congruence. Assume ϕ, ψ ∈ DM X with (DM q)(ϕ) = (DM q)(ψ). By definition ϕ ∼ ψ and h(ϕ) ∼ h(ψ) by assumption that ∼ is a congruence. Thus q ◦ h(ϕ) = q ◦ h(ψ). Thus to- gether with the surjectivity of DM q, there is a unique map h∼ : DM X/∼ → X/∼ fixed by h∼ ◦(DM q) = q ◦ h. To prove the converse, assume there is a map h∼ : DM X/∼ → X/∼ with h∼ ◦(DM q) = q ◦ h and assume ϕ ∼ ψ for some ϕ, ψ ∈ DM X. Then q(h(ϕ)) = h∼(DM q(ϕ)) = h∼(DM q(ψ)) = q(h(ψ)) and so h(ϕ) ∼ h(ψ), which shows that ∼ is a congruence. For the third point, assume ∼ is a congruence. Then h∼ ◦DM h∼ ◦ DMDM h∼ = h∼ ◦DM (h∼ ◦DM q) = h∼ ◦DM (q ◦ h) = q ◦ h◦DM h = q ◦ h◦ µ = h∼ ◦DM q ◦ µ = h∼ ◦ µ◦DMDM q by the second point idem as (X, h) is a.conv. by the second point by naturality µ. Hence by surjectivity of DMDM q, we find h∼ ◦ DM h∼ = h∼ ◦ µ. Using natural- ity of η, the second point and the fact that (X, h) is an abstract convex set (in that order), we find h∼ ◦ η ◦ q = h∼ ◦DM q ◦ η = q ◦ h◦ η = q and so h∼ ◦ η = id by surjectivity of q. We have shown (X/∼, h∼) is an abstract M -convex set. The equality q ◦ h = h∼ ◦DM q from the second point shows that q is an M -affine map. Solution to 193 III. To show ∼ is a congruence, assume ϕ ∼ ψ for ϕ, ψ ∈ DM X. Write q : X → X/∼ for the quotient map for ∼ and f∼ : X/∼ → X for the unique map with f∼ ◦ q = f . Note that DM q(ϕ) = DM q(ψ) and so f (h(ϕ)) = h(DM f (ϕ)) = h(DM f∼(DM q(ϕ))) = h(DM f∼(DM q(ψ))) = f (h(ψ)) and so h(ϕ) ∼ h(ψ), which shows ∼ is a congruence. Solution to 193 IV. Concerning the first point: clearly h(η(h(ψ))) = h(ψ) and so ( η(h(ψ)), ψ ) is a derivation of η(h(ψ)) ≈ ψ. Assume ϕ ≈ ψ. Then η(h(ϕ)) ≈ ϕ ≈ ψ ≈ η(h(ψ)) and so h(ϕ) ∼ h(ψ). We continue with the second point. Suppose we are given χ1, . . . , χn, ϕ, ψ ∈ DM X and λ0, . . . , λn ∈ M with >i λi = 1. To cover the first base case in the definition of a derivation, assume h(ϕ) = h(ψ). Then (h◦ µ)(cid:16)λ0 ψi > n > j=1 λj χji(cid:17) = (h◦DM h)(cid:16)λ0 ψi > n > j=1 λj χji(cid:17) λj h(χj)i(cid:17) λj h(χj)i(cid:17) λj χji(cid:17) n > j=1 n > j=1 = h(cid:16)λ0 h(ψ)i > = h(cid:16)λ0 h(ϕ)i > = (h◦ µ)(cid:16)λ0 ϕi > n > j=1 and so µ(cid:0)λ0 ψi > >n base case in the definition of a derivation, assume ϕ ≡ >m >m j=1 λj χji(cid:1) ≈ µ(cid:0)λ0 ϕi > >n j=1 λj χji(cid:1). Next, for the other i µi xii and ψ ≡ i µi yii with xi R∗ yi for 1 6 i 6 m. Then > x n > j=1 n > j=1 λ0 ⊙ µi yii > m > i=1 m > i=1 n > j=1 n > j=1 µ(cid:16)λ0 ψi > µ(cid:16)λ0 ϕi > λj χji(cid:17) = λj χji(cid:17) = This shows that µ(cid:0)λ0 ψi > >n second base case of a derivation. Thus by induction µ(cid:0)λ0 ψi > >n µ(cid:0)λ0 ϕi > >n j=1 λj χji(cid:1) whenever merely ϕ ≈ ψ. j=1 λj χji(cid:1) ≈ µ(cid:0)λ0 ϕi > >n λ0 ⊙ µi xii > > x λj ⊙ χj(x)xi λj ⊙ χj(x)xi . For the third point, we need some preparation. Pick representatives R ⊆ X for every x ∈ X there is a unique rx ∈ R with rx ∼ x. By defini- of ∼: j=1 λj χji(cid:1) by the j=1 λj χji(cid:1) ≈ 223 tion η(rx) ≈ η(x). Let ϕ ≡ >n i=1 λi xii ∈ DM X given. By the previous point λi η(xi)i(cid:17) λi η(xi)i(cid:17) n > i=2 n > i=2 ϕ = µ(cid:16)λ1 η(x1)i > ≈ µ(cid:16)λ1 η(rx1 )i > = µ(cid:16)λ2 η(x2)i > λ1 rx1i > ≈ µ(cid:16)λ2 η(rx2 )i > λ1 rx1i > n > i=3 n > i=3 ... λi η(xi)i(cid:17) λi η(xi)i(cid:17) n > i=1 ≈ λi rxii r∈R(cid:16)> ≈ > x∼r ϕ(x)(cid:17) ri . Now, assume ϕ ∼ ψ. Then >x∼x0 ϕ(x) = >x∼x0 ψ(x) for any x0 ∈ X and so ϕ ≈ > r∈R(cid:16)> x∼r ϕ(x)(cid:17) ri = > r∈R(cid:16)> x∼r ψ(x)(cid:17) ri ≈ ψ. Clearly ∼ is an equivalence relation on X. To show it is a congruence, assume ϕ ∼ ψ. By the previous, we know ϕ ≈ ψ and so h(ϕ) ∼ h(ψ) by the first point, which shows ∼ is indeed a congruence. Furthermore R is contained in ∼; indeed, if x R y, then (η(x), η(y)) is a derivation of η(x) ≈ η(y) and so x ∼ y. To show ∼ is the smallest congruence containing R, assume S ⊆ X 2 is some congruence of X with R ⊆ S. It is sufficient to show that ϕ ≈ ψ implies h(ϕ) S h(ψ). Indeed if this is the case and x ∼ y, then by definition η(x) ≈ η(y) and so x = h(η(x)) S h(η(y)) = y. We will prove that ϕ ≈ ψ implies h(ϕ) S h(ψ) by induction over the defini- tion of ≈. For the first base case in the definition of ≈, assume ϕ ≡ >n i=1 λi xii and ψ ≡ >n i=1 λi yii for some (x1, y1), . . . , (xn, yn) ∈ R∗. Then xi S yi and so ϕ S ψ, hence h(ϕ) S h(ψ). For the other base case, suppose h(ϕ) = h(ψ) for some ϕ, ψ ∈ DM X. Then clearly h(ϕ) S h(ψ). Thus, by induction h(ϕ) S h(ψ) whenever ϕ ≈ ψ. Solution to 193 X. As the one element set is final in Set, there is a unique map ! : DM 1 → 1. This turns DM 1 into an abstract M -convex set: !◦DM ! = 0/0 , we have ! = !◦ µ and !◦ η = ! = id. Note that there is a single formal M -convex combination over 1 and so DM 1 ∼= 1. As AConvM is the Eilenberg -- Moore cat- egory of the monad DM , it follows that the functor F : Set → AConvM given by F (X) = (DM X, µ) and F f = DM f is a left adjoint and so preserves coprod- ucts. Hence DM{1, . . . , n} ∼= F (n · 1) ∼= n · F (1) ∼= n · 1. and 0 ⊙ 0 = 0. Thus, by uniqueness Solution to 195 IV. Clearly 0 6 0/0 a/1 = 1 ⊙ a/1 = a and so in of and a⊙ a/a ⊙ a/a = a⊙ a/a = a and a/a ⊙ a/a 6 a/a particular a/a ⊙ a/a = a/a . For the final equation of the first so by uniqueness of point, assume b ∈ M . As a⊙ b 6 a we know a ⊙ b a/a ⊙ b 6 a/a /a = a/a ⊙ b. and a ⊙ a/a ⊙ b = a ⊙ b. So by uniqueness of /a a ⊙ b For the second point, suppose a, b, c ∈ M with a 6 b 6 c. Note b/c ⊙ and c ⊙ b/c ⊙ a/b = b ⊙ a/b = a. So by uniqueness of a/b 6 b/c 6 c/c , we a/c b/c ⊙ a/b = a/c , as desired. get 0/0 = 0. For any a ∈ M , we have is defined. Note we see 1/1 = 1. Next, note a/a , we have a ⊙ b /a Solution to 195 VI. Let X be a compact Hausdorff space. Assume that the unit interval of C(X) is an effect divisoid. To show X is basically disconnected, assume f ∈ C(X). We have to show that supp f is open. Note suppf = supp f and so, without loss of generality, we may assume that f > 0. By compactness f is bounded. Pick B > 0 with B > f . Then supp B−1f = supp f and B−1f 6 1, so we may also assume without loss of generality, that 0 6 f 6 1. If supp f = X, then we are done, so assume supp f 6= X. Pick any y /∈ supp f . By Urysohn's lemma, there is a g ∈ C(X) with g(y) = 0 and g(x) = 1 for every x ∈ supp f . Define h ≡ f/f ∧ (0 ∨ g). Clearly 0 6 h 6 f/f 6 1 and h(y) 6 (0 ∨ g)(y) = 0. is zero -- one valued: a characteristic Note that function. Pick any x ∈ supp f . Then 0 < f (x) 6 f/f (x) = (x) ∧ g(x) = 1. By continuity 1 ∧ 1 = 1. Hence (f ⊙ h)(x) = f (x)h(x) = f (x) for x ∈ supp f and (f ⊙ h)(x) = f (x)h(x) = 0 = f (x) whenever x /∈ supp f . Thus f ⊙ h = f . By uniqueness of (y) = h(y) = 0. Recall that y f/f was an arbitrary element y /∈ supp f and thus is the characteristic function f/f of supp f , which is thus open. Hence X is basically disconnected. (x) = 1 for every x ∈ supp f and so h(x) = f/f f/f ≡ f/f ⊙ f/f = f/f (x) ∈ {0, 1}, hence f/f = h. Consequently , we have and so f/f f/f f/f f/f f/f To prove the converse, assume X is basically disconnected. Let f, g ∈ C(X) n}. Note Un = be given with 0 6 f 6 g 6 1. Define Un ≡ {x; g(x) > 1 supp((g − 1 n ) ∨ 0) and so Un is open as X is basically disconnected. Define hn ≡ ( f (x) g(x) 0 x ∈ Un otherwise. To show hn is continuous, assume there is a net (xα)α with xα → x for some x ∈ X. Suppose x ∈ Un. As Un is open, we know xα ∈ Un for sufficiently large α and so hn(xα) → hn(x) as f (x) n > 0. g(x) is continuous for x ∈ Un as then g(x) > 1 225 Clearly 0 6 h1 6 h2 6 . . . 6 1 and so we may define is ω-complete by basic disconnectedness of X. Note 0 6 f/g 6 1. For the other case, suppose x /∈ Un. The set X − Un is open and so xα /∈ Un for sufficiently large α and then hn(xα) = 0 = hn(x). Thus hn is continuous. f/g ≡ supn hn as C(X) Suppose f = g and x ∈ supp f . Then hn(x) = 1 for all n > f (x)−1 and (x) = 1 for all x ∈ supp f . Write χ for the characteristic f/f (x). As Un ⊆ supp f and hn 6 1, we f/f = χ, the characteristic function so function of supp f . We just saw χ 6 f/f have hn 6 χ for all n and so of supp f . In particular f 6 f/f f/f 6 χ. Thus (x) = 1. Thus f/f = f/f . and f/f Let f, g ∈ C(X) be given with 0 6 f 6 g 6 1 and h1 6 h2 6 . . . as above. We will show f/g 6 g/g . Assume x ∈ X. Suppose x ∈ supp g. Then hn(x) = f (x) g(x) 6 1 = g/g (x) for all n > g(x)−1 and so f/g (x) 6 g/g (x) for all x ∈ supp g. For the other case, assume x /∈ supp g. Then x /∈ Un and so hn(x) = 0 for all n ∈ N, whence (x) = 0 6 g/g (x). Thus indeed f/g 6 g/g . g(x) for x ∈ supp g. To show In a similar way, it is easy to see that (x) 6 g/g (x). In particular (x) 6 f (x) f/g f/g f/g f/f ⊙ f/f ≡ f/f equality, we define kn ≡ ( f (x) g(x) 1 x ∈ Un otherwise. f/g (x) 6 kn(x) = f (x) g(x) for n > g(x)−1 and so With similar reasoning as for hn, we see that these kn are continuous. Fur- f/g 6 kn for any n. Pick any x ∈ thermore hm 6 kn for all m and so supp g. Then g(x) for any x ∈ supp g. Thus (g ⊙ f/g )(x) = f (x) for x ∈ supp g by continuity. For the other case, assume x /∈ supp g. Then g⊙ f/g (x) 6 g/g (x) = 0 = g(0) > f (0). We have shown g⊙ f/g = f . remains. So assume h ∈ C(X) with 0 6 h 6 g/g and g ⊙ h = f . Assume x ∈ supp g. Then g(x)h(x) = f (x) and so h(x) = f (x) (x) for all x ∈ supp g. For the other g(x) = f/g case, assume x /∈ supp g. As both h, (x). We have shown h = f/g )(x) = f (x) for x ∈ supp g and so (g ⊙ f/g and thus the unit interval of C(X) is an effect divisoid. f/g 6 g/g , we have h(x) = 0 = f/g (x). By continuity h(x) = f/g Only uniquness of (x) = f (x) (x) 6 f/g f/g f/g Solution to 197 V. Assume C is an effectus. 1. Assume ξ : X → Y is a quotient for p and ϑ : Y → Z is an isomor- phism. Note 1 ◦ ϑ◦ ξ = 1 ◦ ξ 6 p⊥. To prove ϑ◦ ξ is a quotient for p, assume f : X → Y ′ is some map with 1 ◦ f 6 p⊥. As ξ is a quotient, there exists a unique map f′ : Y → Y ′ with f′ ◦ ξ = f . Clearly f = f′ ◦ ξ = (f′ ◦ ϑ−1)◦(ϑ◦ ξ). Assume h : Z → Y ′ is any map with f = h◦(ϑ◦ ξ). Then h◦ ϑ = f′ by uniqueness of f′ and so h = f′ ◦ ϑ−1. We have shown that ϑ◦ ξ is a quotient of p as well. 2. Assume ξ1 : X → Y1 and ξ2 : X → Y2 are quotients for p. By the universal property of ξ1, there is a unique map ϑ1 : Y1 → Y2 with ϑ1 ◦ ξ1 = ξ2. Similarly, there is a unique map ϑ2 : Y2 → Y1 with ϑ2 ◦ ξ2 = ξ1. Note that ξ1 = ϑ2 ◦ ξ2 = ϑ2 ◦ ϑ1 ◦ ξ1. By the universal property of ξ1 again, the map id : Y1 → Y1 is the unique map with id◦ ξ1 = ξ1 and so ϑ2 ◦ ϑ1 = id. Similarly ϑ1 ◦ ϑ2 = id. Thus ϑ2 is an isomorphism with ϑ2 ◦ ξ2 = ξ1. It is the unique such isomorphism by the universal property of ξ2. 3. Assume f : X → Y is some map. Clearly 1 ◦ f 6 1 = 0⊥. Furthermore f is itself the unique map f′ with f′ ◦ id = f . Thus id is a quotient for 0. Consequently any isomorphism is a quotient for 0 by the first point. 4. Assume X is any object. As 0 is a zero object, there is a unique map 0X : X → 0. Clearly 1 ◦ 0 = 0 = 1⊥. Assume f : X → Y is any map with 1 ◦ f 6 1⊥ = 0. Then f = 0 and so f = 0 = 0 ◦ 0X . If f = h◦ 0X for some h : 0 → Y , then h = 0 and so any map into zero is indeed a quotient for 1. 5. Assume ξ : X → Y is a quotient for p. Clearly 1 ◦ p⊥ = p⊥ 6 p⊥ and so there is a unique map f : Y : 1 with f ◦ ξ = p⊥. Then p⊥ = 1 ◦ p⊥ = 1 ◦ f ◦ ξ 6 1 ◦ ξ = p⊥ and so 1 ◦ ξ = p⊥. 6. Assume f1 ◦ ξ = f2 ◦ ξ for some quotient ξ of p. Then 1 ◦ f1 ◦ ξ 6 1 ◦ ξ = p⊥. Hence there is a unique f with f ◦ ξ = f1 ◦ ξ. Both f1 and f2 fit the bill, hence f1 = f = f2. Thus ξ is epic. Solution to 197 XI. Let C be an effectus with quotients. Assume t′ ◦ ξ′ = t◦ ξ for some quotient ξ, ξ′ and total maps t, t′. Then 1 ◦ ξ = 1 ◦ t◦ ξ = 1 ◦ t′ ◦ ξ′ = 1 ◦ ξ′. Thus by 197 V there is a unique isomorphism ϑ with ξ′ = ϑ◦ ξ. Note t◦ ξ = t′ ◦ ξ′ = t′ ◦ ϑ◦ ξ and so by epicity of ξ (see 197 V), we see t = t′ ◦ ϑ, as desired. Solution to 198 III. Assume C is an effectus with quotients. We will show that 0 has a left adjoint by demonstrating the universal mapping property. Pick for ev- the components of the unit of the adjunction. Let f : (X, p) → (Y, 0) be some erty of ξp, there exists a unique map f′ : X/p → Y with f = f′ ◦ ξp in C. Clearly as well. Then g ◦ ξp = f and so g′ = f′. This shows that 0 has a left adjoint. ery object (X, p) inR Pred(cid:3) a quotient ξp : X → X/p of p. Note (0⊥ ◦ ξp)⊥ = p and so ξp is a map (X, p) → (X/p, 0) inR Pred(cid:3). We will use these maps ξp as map inR Pred(cid:3). By definition of map 1 ◦ f 6 p⊥ and so by the universal prop- also (0f )◦ ξp = f inR Pred(cid:3). Let g : X/p → Y be some map with (0g)◦ ξp = f To prove the converse, assume C is an effectus where 0 : C →R Pred(cid:3) has a left adjoint Q : R Pred(cid:3) → C. Let X be some object in C with a predicate p. tion Q ⊣ 0. By definition of maps in R Pred(cid:3), we know 1 ◦ η 6 p⊥. We will with 1 ◦ f 6 p⊥. Then f : (X, p) → (Y, 0) in R Pred(cid:3). By the universal map- Write η : (X, p) → 0Q(X, p) for the (X, p) component of the unit of the adjunc- show that η is a quotientfor p. To this end, assume f : X → Y is some map in C ping property, there is a unique map f′ : Q(X, p) → Y in C with (0f′)◦ η = f , 227 i.e. f′ ◦ η = f . This shows that η is indeed a quotient of p. Hence C has quotients. Solution to 199 VI. Assume C is an effectus with comprehension. We will show that 1 has a right adjoint by demonstrating the dual of the univer- hension πp : {Xp} → X. Note that for any map f , we have 1 ◦ f = p◦ f if and only if p⊥ ◦ f = 0. Thus (p⊥ ◦ πp)⊥ = 1, which shows that πp is a for the counit of the adjunction. Let f : (Y, 1) → (X, p) be some map in C. By by the universal property of πp, there is a unique map f′ : Y → {Xp} in C sal mapping property. Pick for every object (X, p) in R Pred(cid:3) a compre- map ({Xp}, 1) → (X, p) in R Pred(cid:3). We will use these maps as components definition of maps inR Pred(cid:3), we have 1 6 (p⊥ ◦ f )⊥, viz. 1 ◦ f = p◦ f . Thus, with πp ◦ f′ = f . Hence 1f′ is the unique map in R Pred(cid:3) with πp ◦(1f′) = f inR Pred(cid:3). We have shown that 1 has a right adjoint. To prove the converse, assume C is an effectus where 1 : C → R Pred(cid:3) has a right adjoint K : R Pred(cid:3) → C. Let X be some object of C with a of the adjunction 1 ⊣ K. By definition of maps in R Pred(cid:3), we know 1 6 (p⊥ ◦ f )⊥ and so f is a map (Y, 1) → (X, p) in R Pred(cid:3). By the dual of the universal mapping property, there is a unique map f′ : Y → K(X, p) inR Pred(cid:3) predicate p. Write ε : 1K(X, p) → (X, p) for the (X, p) component of the counit (p⊥ ◦ ε)⊥, hence 1 ◦ ε = 1 ◦ ε. We will show that ε is a comprehension for p. To this end, assume f : Y → X is some map in C with 1 ◦ f = p◦ f . Then 1 = with ε◦(1f′) = f , i.e. ε◦ f′ = f . This shows that ε is indeed a comprehension for p. Hence C has comprehension. Solution to 199 VII. Let C be an effectus. 1. Assume π : X → Y is a comprehension for p and ϑ : Z → X is an iso- morphism. We will show π ◦ ϑ is a comprehension for p as well. To start, note 1 ◦ π ◦ ϑ = p◦ π ◦ ϑ. Assume f : X′ → Y is any map with 1 ◦ f = p◦ f . Then, by the universal property of π, there is a unique map f′ : X′ → X with π ◦ f′ = f . Then f = (π ◦ ϑ)◦(ϑ−1 ◦ f′). Assume g : X′ → X is a map with g = (π ◦ ϑ)◦ g as well. Then ϑ◦ g = f′ by uniqueness of f′ and so g = ϑ−1 ◦ f′. We have shown that π ◦ ϑ is a comprehension for p as well. 2. Assume π1 : X1 → Y and π2 : X2 → Y are comprehensions of p. By the universal property of π1, there is a unique map ϑ1 : X2 → X1 with π2 = π1 ◦ ϑ1. On the other side, by the universal property of π2, there is a unique map ϑ2 : X1 → X2 with π1 = π2 ◦ ϑ2. Then π1 = π2 ◦ ϑ2 = π1 ◦ ϑ1 ◦ ϑ2. By the universal property of π1 again, the map id : X1 → X1 is the unique map with π1 ◦ id = π1 and so ϑ1 ◦ ϑ2 = id. Similarly ϑ2 ◦ ϑ1 = id. Thus ϑ2 is an isomorphism with ϑ1 = π2 ◦ ϑ2. It is the unique such isomorphism by the universal property of π2. 3. Assume f : X → Y is some map. Trivially 1 ◦ f = 1 ◦ f and 1 ◦ id = 1 ◦ id. The map f itself is the unique map g with g ◦ id = f and so id is a comprehension for 1. Consequently any isomorphism is a comprehension for 1 by the first point. 4. Assume Y is any object. As 0 is a zero object, there is a unique map 0Y : 0 → Y . Clearly 1 ◦ 0Y = 0 = 0 ◦ 0Y . Assume f : X → Y is any map with 1 ◦ f = 0 ◦ f . Then 1 ◦ f = 0 ◦ f = 0 and so f = 0. The zero map is the unique map X → 0 and for that map, we have 0Y ◦ 0 = 0 = f and so 0Y is a comprehension for 0. 5. Assume π ◦ f1 = π ◦ f2 for some comprehension π of p. Then 1 ◦ π ◦ f1 = p◦ π ◦ f1 and so by the universal property of π, there is a unique map f with π ◦ f = π ◦ f1. Both f1 and f2 fit the bill, hence f1 = f = f2. Thus π is monic. 6. Assume π is a comprehension of p. Then (p⊥ ◦ π) >(p◦ π) = (p⊥ >p)◦ π = 1 ◦ π = p◦ π. Thus p⊥ ◦ π = 0 by cancellation. Solution to 200 V. Let C be any effectus. Assume π : X → Y is a comprehension of p. We will show π is a categorical kernel p⊥. To start, note p⊥ ◦ π = 0 by 199 VII. Now assume f : X′ → Y is some map with p⊥ ◦ f = 0. Then p◦ f = (p◦ f ) > (p⊥ ◦ f ) = 1 ◦ f and so by the universal property of π, there is a unique map f′ : X′ → X with f = π ◦ f′. Hence π is a kernel of p⊥. To prove the converse, assume k : X → Y is a kernel of p⊥ : Y → 1. We have 1 ◦ k = (p◦ p⊥)◦ k = (p◦ k) > (p⊥ ◦ k) = p◦ k. Assume f : X′ → Y is some map with p◦ f = 1 ◦ f . Then (p⊥ ◦ f ) > (p◦ f ) = 1 ◦ f = p◦ f and so by cancellation p⊥ ◦ f = 0. Hence by the universal property of k there is a unique map f′ : X′ → X with k ◦ f′ = f . This shows that k is a comprehension of p. Solution to 202 V. We have (im f )◦ f ◦ g = 1 ◦ f ◦ g 6 1 ◦ g and so im f ◦ g 6 im f . Consequently im f ◦ α 6 im f = im f ◦ α◦ α−1 6 im f ◦ α and so im f ◦ α = im f . Solution to 202 VI. Assume ξ : X → Y is some quotient and p is a predicate on Y with p◦ ξ = 0. Then p◦ ξ = 0 = 0 ◦ ξ and so p = 0 by epicity of ξ. Hence ξ is faithful. Solution to 203 XII. Assume ⌊p⌋ = p. Then p = ⌊p⌋ ≡ im π for some com- prehension π of p and so p is sharp. To prove the converse, assume p is sharp; that is: p = im f for some f . Pick any comprehension π of p. Clearly p◦ f = (im f )◦ f = 1 ◦ f and so there is a unique g with f = π ◦ g. By 202 V we find p = im f = im π ◦ g 6 im π ≡ ⌊p⌋ 6 p and so p = ⌊p⌋. Solution to 203 XIII. Note ⌈p⌉ 6 ⌈q⌉ whenever p 6 q and so ⌈p◦ f⌉ 6 ⌈⌈p⌉◦ f⌉ 6 ⌈⌈p◦ f⌉⌉ = ⌈p◦ f⌉ as p 6 ⌈p⌉ and ⌈p⌉◦ f 6 ⌈p◦ f⌉ by 203 XIII. Thus ⌈p◦ f⌉ = 229 ⌈⌈p⌉◦ f⌉. Solution to 203 XIV. To start, note [im f, im g]hf, gi = ((im f )◦ f )>((im g)◦ g) = (1 ◦ f ) > (1 ◦ g) = 1 ◦hf, gi by 181 IX. Assume p ≡ [p1, p2] is some predicate with p◦hf, gi = 1 ◦hf, gi. Then (1 ◦ f ) > (1 ◦ g) = 1 ◦hf, gi = p◦hf, gi = (p1 ◦ f ) >(p2 ◦ g) 6 (p1 ◦ f ) >(1 ◦ g) and so by cancellation 1 ◦ f 6 p1 ◦ f 6 1 ◦ f , hence 1 ◦ f = p1 ◦ f . Thus im f 6 p1. Similarly im g 6 p2. Thus [im f, im g] 6 [p1, p2] = p. We have shown imhf, gi = [im f, im g]. Assume [p, q] is sharp. That is, there is some f with [p, q] = im f . Then [p, q] = im f = imh⊲1 ◦ f, ⊲2 ◦ fi = [im ⊲1 ◦ f, im ⊲2 ◦ f ] and so p = im ⊲1 ◦ f and q = im ⊲2 ◦ f , which shows p and q are sharp. To prove the converse, assume p and q are sharp. Pick f and g with p = im f and g = im g. Note f + g = hf ◦ ⊲1, g ◦ ⊲2i. We have im f = im f ◦ ⊲1 ◦ κ1 6 im f ◦ ⊲1 6 im f and so im f ◦ ⊲1 = im f . Similarly im g ◦ ⊲2 = im g. Thus im f + g = imhf ◦ ⊲1, g ◦ ⊲2i = [im f ◦ ⊲1, im g ◦ ⊲2] = [im f, im g] = [p, q], which shows [p, q] is sharp. Solution to 205 IV. Let C be an effectus with comprehension and images. As- sume f : X → Y is a quotient of sharp s. Let πs denote a comprehension of s. Clearly 1 ◦ f ◦ πs = s⊥ ◦ πs = 0. Assume g : X → Y ′ is some map with g ◦ πs = 0. Then 1 ◦ g ◦ πs = 0 and so 1 ◦ g 6 im⊥ πs = s⊥. Thus by the universal property of f as a quotient, there is a unique map g′ : Y → Y ′ with g = g′ ◦ f . This shows f is a cokernel of πs. To prove the converse, assume f is a cokernel of πs. By assumption f ◦ πs = 0. Thus 1 ◦ f ◦ πs = 0. Hence 1 ◦ f 6 im⊥ πs = s⊥. Let g : X → Y ′ be a map with 1 ◦ g 6 s⊥. Then 1 ◦ g ◦ πs 6 s⊥ ◦ πs = 0 and so g ◦ πs = 0. Thus by the universal property of f as a cokernel, there is a unique map g′ : Y → Y ′ with g = g′ ◦ f . This shows f is a quotient of s. Solution to 207 II. Assume s 6 t. Then s◦ f 6 t◦ f , hence (t◦ f )⊥ 6 (s◦ f )⊥ f⋄(t). This shows f⋄ is order preserving. Furthermore t⊥ 6 s⊥, so f⋄(t⊥) 6 f⋄(s⊥), hence f (s) ≡ f⋄(s⊥)⊥ 6 f⋄(t⊥)⊥ ≡ f (t), which shows that f is order preserving. Solution to 207 V. Let f : X → Y be some map in a ⋄-effectus. so(cid:4)(t◦ f )⊥(cid:5) 6(cid:4)(s◦ f )⊥(cid:5). Thus f⋄(s) ≡ ⌈s◦ f⌉ ≡(cid:4)(s◦ f )⊥(cid:5)⊥ 6(cid:4)(t◦ f )⊥(cid:5)⊥ ≡ 1. Assume s 6 t. Clearly f⋄(t) 6 f⋄(t) so s 6 t 6 f (f⋄(t)), hence f⋄(s) 6 f⋄(t), which shows that f⋄ is order preserving. 2. Assume D ⊆ SPred X is a set of sharp predicates with a supremum sup D. We will show f⋄(sup D) is the supremum of f⋄(D). For any d ∈ D, we have d 6 sup D and so f⋄(d) 6 f⋄(sup D) by the previous point. Assume x is some sharp predicate of Y with f⋄(d) 6 x for all d ∈ D as well. Then d 6 f (x) for all d ∈ D and so sup D 6 f (x). Hence f⋄(sup D) 6 x, which shows that sup f⋄(D) = f⋄(sup D). 3. Assume D ⊆ SPred Y is a set of sharp predicates with an infimum inf D. We will show f (inf D) is the infimum of f (D). For any d ∈ D, we have inf D 6 d and so f (inf D) 6 f (d) by 207 II. Assume x is some sharp predicate of X with x 6 f (d) for all d ∈ D as well. Then f⋄(x) 6 d for all d ∈ D and so f⋄(x) 6 inf D, hence x 6 f (inf D). This shows that inf f (D) = f (inf D). 4. For for any D ⊆ SPred Y with supremum, we have f⋄(sup D) ≡ f ((sup D)⊥)⊥ (d⊥))⊥ f (d⊥)(cid:1)⊥ = (cid:0) inf = f ( inf d∈D d∈D = sup d∈D = sup d∈D f (d⊥)⊥ f⋄(d) and so f⋄ preserves suprema. 5. For any s ∈ SPred X we have f⋄(s) 6 f⋄(s) and so s 6 f (f⋄(s)). Simi- larly, for any t ∈ SPred Y we have f⋄(f (t)) 6 t. Hence f⋄(f (f⋄(s))) 6 f⋄(s) 6 f⋄(f (f⋄(s))), so f⋄ = f⋄ ◦ f ◦ f⋄. 6. We have f (f⋄(f (t))) 6 f (t) 6 f (f⋄(f (t))) and so f = f ◦ f⋄ ◦ f . Solution to 207 VIIa. Let f, g : X → Y be some maps in a ⋄-effectus. As- sume f⋄ = g⋄ and t ∈ SPred X. Trivially g⋄(t) 6 (g⋄(t)⊥)⊥ and so f⋄(g⋄(t)⊥) = g⋄(g⋄(t)⊥) 6 t⊥. Thus f⋄(t) 6 g⋄(t)⊥⊥ = g⋄(t). Reasoning in the same way with f and g swapped, we see g⋄(t) 6 f⋄(t) and so f⋄ = g⋄. To prove the converse, assume f⋄ = g⋄ and s ∈ SPred Y . Trivially g⋄(s) 6 (g⋄(s)⊥)⊥ and so f⋄(g⋄(s)⊥) = g⋄(g⋄(s)⊥) 6 t⊥. Thus f⋄(t) 6 g⋄(t)⊥⊥ = g⋄(t). Reasoning in the same way with f and g swapped, we see g⋄(t) 6 f⋄(t) and so f⋄ = g⋄. Solution to 208 XII. As ξ⋄ is left adjoint to ξ , we have t 6 ξ (ξ⋄(t)). Further- more s = (cid:6)s⊥(cid:7)⊥ = ⌈1 ◦ ξ⌉⊥ = ξ (0) 6 ξ (ξ⋄(t)). Thus ξ (ξ⋄(t)) is an upper bound of s and t. We have to show it is the least sharp upper bound. So as- sume r is any sharp predicate with r > s and r > t. Then 1 ◦ ξr = r⊥ 6 s⊥ and so there is a (unique) map h with ξr = h◦ ξ. We have (ξr) = ξ ◦ h = ξ ◦ ξ⋄ ◦ ξ ◦ h = ξ ◦ ξ⋄ ◦(ξr) and so r = (ξr) (0) = ξ (ξ⋄((ξr) (0))) = ξ (ξ⋄(r)) > ξ (ξ⋄(t)). 231 Solution to 209 II. For the first point, assume f⋄ = g⋄ for some f : X ⇆ Y : g. Trivially g⋄(s) 6 (g⋄(s)⊥)⊥ for any for any s ∈ SPred X and so f⋄(g⋄(s)⊥) = g⋄(g⋄(s)⊥) 6 s⊥. Hence f⋄(s) 6 g⋄(s)⊥⊥ = g⋄(s). With essentially the same argument (swapping both f and g and the up/down position of the diamonds), we see g⋄(s) 6 f⋄(s), hence g⋄ = f⋄. Swapping f and g again, gives the argument for the converse. For the second point, assume f and g are ⋄-adjoint. Then im f = f⋄(1) = g⋄(1) = ⌈1 ◦ g⌉, as desired. Solution to 209 III. For the first point, assume f is ⋄-self-adjoint. Then (f ◦ f )⋄ = f⋄ ◦ f⋄ = f⋄ ◦ f⋄ = (f ◦ f )⋄ and so f ◦ f is ⋄-self-adjoint as well. For the second point, assume f is ⋄-positive. Then f = g ◦ g for some ⋄-self- adjoint g. Hence f is ⋄-self-adjoint by the previous point. For the third point, assume f is ⋄-positive and f ◦ f is pure. By the previous point f is ⋄-self-adjoint and so f ◦ f is ⋄-self-adjoint by the first point. Thus f ◦ f is ⋄-positive. Solution to 210 II. Assume f : X → Y is a sharp map and p is a predicate on Y . Using 203 XIII and 203 XII we see ⌈p◦ f⌉ = ⌈⌈p⌉◦ f⌉ = ⌈p⌉◦ f . For the converse assume ⌈p⌉◦ f = ⌈p◦ f⌉ for every predicate p on Y . Assume s is any sharp predicate on Y . Then ⌈s◦ f⌉ = ⌈s⌉◦ f = s◦ f and so s◦ f is sharp. Thus f is sharp. Solution to 211 XIV. Note 1 ◦ asrtp ◦ asrtp = p◦ asrtp ≡ p & p. By 209 III the map asrtp ◦ asrtp is ⋄-positive and thus asrtp ◦ asrtp = asrtp&p by uniqueness of ⋄-positive maps. Solution to 211 XV. Assume f : X → Y is a map in a &-effectus with s ∈ SPred Y and t ∈ SPred X. Assume im f 6 s. Then 1 ◦ f = s◦ f . Thus there is a g with f = πs ◦ g. Then asrts ◦ f = πs ◦ ζs ◦ πs ◦ g = πs ◦ g = f . For the converse, as- sume asrts ◦ f = f . Then 1 ◦ f = 1 ◦ asrts ◦ f = s◦ f and so im f 6 s. We have shown the first equivalence. For the second equivalence, assume 1 ◦ f 6 t. Then f = h◦ ζt for a (unique) map h. So f ◦ asrtt = h◦ ζt ◦ πt ◦ ζt = h◦ ζt = f . For the converse, assume f = f ◦ asrtt. Then 1 ◦ f = 1 ◦ f ◦ asrtt 6 1 ◦ asrtt = t. Solution to 213 V. Clearly p 6 s if and only if 1 ◦ p 6 s. Even more trivially, s & p = p if and only if p◦ asrts ≡ s & p = p. Thus the result follows by applying the second equivalence of 211 XV with f = p and t = s. Solution to 213 VI. Let f : X → A and g : X → B be maps in a &-effectus. As- sume hf, gi is sharp. Pick any sharp predicate s on A. By 203 XIV we know [s, 0] is sharp and so [s, 0]◦hf, gi = (s◦ f ) > (0 ◦ g) = s◦ f is sharp as well. This shows f is sharp. With a similar argument we see g is sharp. For the converse, assume f and g are sharp. Assume [s, t] is some sharp predicate on A + B. By 203 XIV we know s and t are sharp. Thus s◦ f and t◦ g are sharp. Conse- quently [s, t]◦hf, gi = (s◦ f ) > (t◦ g) is sharp as well by 208 III. Thus hf, gi is sharp, as desired. Solution to 216 X. Assume C is a &-effectus where every predicate has a square root and where πs is ⋄-adjoint to ζs. Let p be a predicate and s be another predicate on the same object that is sharp. Pick a predicate q with q & q = p. The map f ≡ πs ◦ asrtq ◦ ζs is pure and ⋄-self-adjoint: f⋄ = (ζs)⋄ ◦(asrtq)⋄ ◦(πs)⋄ = (πs)⋄ ◦(asrtq)⋄ ◦(ζs)⋄ = f⋄. Thus f ◦ f is ⋄-positive. By 211 XIV, we get f ◦ f = πs ◦ asrtq ◦ ζs ◦ πs ◦ asrtq ◦ ζs = πs ◦ asrtq&q ◦ ζs = πs ◦ asrtp ◦ ζs and so 1 ◦ f ◦ f = 1 ◦ πs ◦ asrtp ◦ ζs = p◦ ζs. By uniqueness of ⋄-positive maps, we get πs ◦ asrtp ◦ ζs = f ◦ f = asrt1 ◦ f ◦ f = asrtp ◦ ζs . Thus asrtp ◦ ζs = ζs ◦ πs ◦ asrtp ◦ ζs = ζs ◦ asrtp ◦ ζs . Solution to 217 III. By 211 XV we have asrtp = asrtp ◦ asrt⌈p⌉ = π⌈p⌉ ◦ id◦ ζ⌈p⌉ ◦ asrtp and so asrt†p = asrtp ◦ π⌈p⌉ ◦ id−1 ◦ ζ⌈p⌉ = asrtp ◦ asrt⌈p⌉ = asrtp. Next, note ζ1 = π1 = asrt1 = id and so πs = πs ◦ id◦ ζ1 ◦ asrt1, hence π†s = asrt1 ◦ π1 ◦ id−1 ◦ ζs = ζs. Also ζs = π1 ◦ id◦ ζs ◦ asrt1 and so ζ†s = asrt1 ◦ πs ◦ id−1 ◦ ζ1 = πs. Finally, note α = asrt1 = π1 ◦ α ◦ ζ1 for any isomorphism α and so α† = asrt1 ◦ π1 ◦ α−1 ◦ ζ1 = α−1. Solution to 218 VI. Assume h is a pristine map in a &-effectus. By 212 III we know there is an isomorphism α with h = πim h ◦ α◦ ζ⌈1 ◦ h⌉ ◦ asrt1 ◦ h. As 1 ◦ h is sharp, we have ζ⌈1 ◦ h⌉ ◦ asrt1 ◦ h = ζ1 ◦ h ◦ π1 ◦ h ◦ ζ1 ◦ h = ζ1 ◦ h and so h = πim h ◦ α◦ ζ1 ◦ h. Solution to 218 IX. Assume h ≡ πim h ◦ α◦ ζ1 ◦ h is a pristine map in a †′-effectus, cf. 218 VI. 1. Clearly h = πim h ◦ α◦ ζ⌈1 ◦ h⌉ ◦ asrt1 ◦ h as 1 ◦ h is sharp and by 211 XV. Thus h† = asrt1 ◦ h ◦ π⌈1 ◦ h⌉ ◦ α−1 ◦ ζim h = π1 ◦ h ◦ α−1 ◦ ζim h using other equivalence of 211 XV. 233 2. Note that by the previous point 1 ◦ h† = im h and im h† = 1 ◦ h, hence by a second appeal to the previous point: h†† = π1 ◦ h† ◦(α−1)−1 ◦ ζim h† = πim h ◦ α◦ ζ1 ◦ h = h. 3. By the first point h† ◦ h = π1 ◦ h ◦ α−1 ◦ ζim h ◦ πim h ◦ α ◦ ζ1 ◦ h = π1 ◦ h ◦ ζ1 ◦ h = asrt1 ◦ h as promised. 4. Clearly p◦ h† 6 1 ◦ h† = im h. 5. Assume p 6 1 ◦ h. Then p◦ asrt1 ◦ h = p by 211 XV and so asrtp ◦ h† ◦ h218 VII= h◦ asrtp ◦ h† ◦ h pt. 3 = h◦ asrtp ◦ asrt1 ◦ h = h◦ asrtp, as desired. Solution to 219 IX. The map α◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs is a quotient for(cid:6)t◦ f†(cid:7): 1 ◦ α◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs = (cid:6)t◦ asrtp ◦ π⌈p⌉(cid:7)◦ ϕ−1 ◦ ζs = (cid:6)t◦ asrtp ◦ π⌈p⌉ ◦ ϕ−1 ◦ ζs(cid:7) = (cid:6)t◦ f†(cid:7) . Furthermore α◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs ◦ π⌈t ◦ f †⌉ (3.16) = α◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs ◦ πs ◦ ϕ◦ π⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ α−1 = id. Thus α◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs is the unique quotient corresponding to the comprehension π⌈t ◦ f †⌉. Hence α◦ ζ⌈t ◦ asrtp ◦ π⌈p⌉⌉ ◦ ϕ−1 ◦ ζs = ζ⌈t ◦ f †⌉, as de- sired. Solution to 219 X. Note that π⌈q⌉ ◦ ψ−1 = π⌈q⌉ ◦ ψ−1 ◦ ζ1 ◦ asrt1 and so by definition of the dagger, we have (π⌈q⌉ ◦ ψ−1)† = asrt1 ◦ π1 ◦(ψ−1)−1 ◦ ζ⌈q⌉ = ψ ◦ ζ⌈q⌉. Furthermore p◦ πt 6 1 = 1 ◦ π⌈q⌉ ◦ ψ−1. Thus we can apply the fourth point of 218 IX: π⌈q⌉ ◦ ψ−1 ◦ asrtp ◦ πt = asrtp ◦ πt ◦(π⌈q⌉ ◦ ψ−1)−1 ◦ π⌈q⌉ ◦ ψ−1 = asrtp ◦ πt ◦ ψ ◦ ζ⌈q⌉ ◦ π⌈q⌉ ◦ ψ−1 = asrtp ◦ k ◦ π⌈q⌉ ◦ ψ−1, which is the desired equality. Solution to 221 IIIa. Assume A is a commutative von Neumann algebra with effect a ∈ A . The standard corner ha : A → ⌈a⌉ A ⌈a⌉ for a (see 169 IV) is given by b 7→ ⌈a⌉ b ⌈a⌉. If A is commutative, then ha(bc) = ⌈a⌉ bc⌈a⌉ = ⌈a⌉ b ⌈a⌉⌈a⌉ c⌈a⌉ = ha(b)ha(c) and so ha is an nmiu-map. Every corner is a standard corner up-to an isomorphism, thus every corner on a commutative von Neumann algebra is an nmiu-map as well. If CvNop were to have dilations, then any ncpu-map would be the composition of a corner and an nmiu-map. We just saw that such corners are nmiu and so any ncp-map would be an nmiu-map as well, which is absurd. Thus CvNop does not have dilations in the sense of 221 II. Solution to 224 VI. We will first show that for any non-zero pure map f : A → C, there is a Hilbert space H , an element x ∈ H , von Neumann algebra C and isomorphism ϕ : B(H ) ⊕ C → A with f (ϕ(T, c)) = hx, T xi. By definition of purity, we know f = c◦ h for some corner h and filter c for f (1). Clearly c(1) = f (1) 6= 0 as f 6= 0 and so c : C → C is simply given by c(1)id, hence f = f (1)h for a corner h : A → C of p ∈ A . As pA p ∼= C both have only two projections and so p is a mimimal projection. In particular 0 is the only central element strictly below p, thus ⌈⌈p⌉⌉ A is a factor. Let : ⌈⌈p⌉⌉ A → B(H ), y ∈ H be a GNS-representation of the restriction of h to ⌈⌈p⌉⌉ A with y ∈ H such that h(⌈⌈p⌉⌉ a) = hy, (⌈⌈p⌉⌉ a)yi. This restriction of h is still a corner, hence pure and so by 171 VII we know is surjective. As ⌈⌈p⌉⌉ A is a factor, must be injective as well. Thus is an nmiu-isomorphism. Hence ϕ : B(H ) ⊕ B → A defined by ϕ(T, c) ≡ −1(T ) + c is an nmiu-isomorphism as well. Define x ≡pf (1). Then f (ϕ(T, c)) = f (1)h(−1(T ) + c) = f (1)h(−1(T )) = f (1)hy, (−1(T ))yi = f (1)hy, T yi = hx, T xi as desired. We continue with the main task of this exercise. Assume (Pure vNop)op has a π2−→ C of C and C. Note π1 6= 0 for otherwise idC = π1 ◦ h = 0, product C π1←− A where h : C → A is the unique map given by the universal property of the product for idC and idC. Similarly π2 6= 0. By assumption, there is a unqiue pure map h : C → A with π1 ◦ h = id and π2 ◦ h = id. With the reasoning at the start, we know there is a Hilbert space H , an element z ∈ H , a von Neumann algebras C and an isomor- phism ϕ : B(H ) ⊕ C with h†(ϕ(T, c)) = hz, T zi. Note that T 7→ hz, T zi is the dagger of the standard filter czihz. Write p : B(H ) ⊕ C → B(H ) for the regular coprojection (T, c) 7→ T , which is corner with p†(T ) = (T, 0). Then h† ◦ ϕ = c† zihz ◦ p and so h = ϕ◦ p† ◦ czihz. By definition of h, we have id = π1 ◦ h = π1 ◦ ϕ◦ p† ◦ czihz. (4.4) Reasoning the same way as we did for f at the start of this exercise, we know that ⌈π1 ◦ ϕ⌉ is a minimal projection. If ⌈π1 ◦ ϕ⌉ 6 (0, 1), then π1 ◦ ϕ◦ p† = 0 and so id = 0 by (4.4), which is absurd. Thus ⌈π1 ◦ ϕ⌉ = (p1, 0) for some minimal projection p1 of B(H ). Thus π1(ϕ(T, c)) = hx, T xi for some x ∈ H . Similarly π2(ϕ(T, c)) = hy, T yi for some y ∈ H . 235 Note that πi ◦ ϕ◦(id ⊕ ϑ) = πi ◦ ϕ for any map ϑ : C → C and i = 1, 2. Hence id◦ πi ◦ ϕ◦(id⊕ϑ)◦ ϕ−1 ◦ h for i = 1, 2 and so (id⊕ϑ)◦ ϕ−1 ◦ h = ϕ−1 ◦ h for every ϑ by uniqueness of h, which is absurd unless C = 0. Suppose dim H > 3. Then there is some non-zero z0 ∈ H with z0 ⊥ x and x0 ⊥ y. So πi ◦ ϕ◦ cz0ihz0 = 0 for i = 1, 2. By the presumed universal property of A , there is a unique map h : C → A with π1 ◦ h = π1 ◦ ϕ◦ cz0ihz0 and π2 ◦ h = π2 ◦ ϕ◦ cz0ihz0 Both ϕ◦ cz0ihz0 and 0 fit the bill, so ϕ◦ cz0ihz0 = 0, which is absured. Thus dim H 6 2. Clearly dim H 6= 0, for otherwise there would be a unqiue pure map C → C, which is absurd. If dim H = 1, then x = λy for some λ ∈ C, λ 6= 0 and so π1 = λπ2, hence for any two pure maps f1, f2 : B → C we would have f1 = λf2, which is absurd. Reducing clutter a bit, we may assume without loss of generality A = M2, π1(T ) = hx, T xi and π2(T ) = hy, T yi for some x, y ∈ C2. Write f1, f2 : C2 → C for the maps f1(λ, µ) = λ and f2(λ, µ) = µ. By the presumed universal prop- erty of M2 as product, there is a unique pure map f : C2 → A with π1 ◦ f = f1 and π2 ◦ f = f2. From the definition of purity, it follows that f = c◦ h for some corner h : C2 → ⌈h(1)⌉ M2 ⌈h(1)⌉ and the standard filter c : ⌈h(1)⌉ M2 ⌈h(1)⌉ → M2. Note ⌈h⌉ C2 ⌈h⌉ ∼= ⌈h(1)⌉ M2 ⌈h(1)⌉. The former is commutative, so ⌈h(1)⌉ 6= 1. If ⌈h(1)⌉ = 0, then h = 0 and so f1 = 0, quod non. Thus ⌈h(1)⌉ M2 ⌈h(1)⌉ ∼= C and so either ⌈h⌉ = (1, 0) or ⌈h⌉ = (0, 1). In the first case 1 = f2(0, 1) = π2(h(0, 1)) = π2(0) = 0, which is a contradiction. The second case leads to a similar contradiction. So there is no product of C and C in (Pure vNop)op. Solution to 224 VII. Assume ξ : C → M4 is some filter with adσ ◦ ξ = ξ. Note σ∗ = σ. Assume c ∈ C . Then ξ(c) = adσ(ξ(c)) = σξ(c)σ. Let b ∈ M4 be any element with σbσ = b. Then (pS + pA )b(pS + pA ) = b = σbσ = (σpS + σpA )b(pS σ + pA σ) = (pS − pA )b(pS − pA ) and so pA bpS + pS bpA = 0. As pA and pS have orthogonal ranges, we see pA bpS = 0 = pS bpA . Hence b = pS bpS + pA bpA . Furthermore bpS = pS bpS . If b is additionally self adjoint, then pS b = pS b∗pS = pS bpS = pS b. Thus pS ξ(c) = ξ(c)pS for any c ∈ C There is some isomorphism ϑ : C → ⌈ξ(1)⌉ M4 ⌈ξ(1)⌉ with ξ(c) = pξ(1)ϑ(c)pξ(1) and so pSpξ(1)bpξ(1) =pξ(1)bpξ(1)pS for any b ∈ M4. From this it follows (using the pseudoinverse ofpξ(1)) that pS ⌈ξ(1)⌉ b ⌈ξ(1)⌉ = ⌈ξ(1)⌉ b ⌈ξ(1)⌉ pS . Thus ⌈ξ(1)⌉ pS ⌈ξ(1)⌉ is central in ⌈ξ(1)⌉ M4 ⌈ξ(1)⌉. So ei- ther ⌈ξ(1)⌉ pS ⌈ξ(1)⌉ = 0 or ⌈ξ(1)⌉ pS ⌈ξ(1)⌉ = ⌈ξ(1)⌉. In the first case, we have ⌈ξ(1)⌉ pS = 0 hence ξ(c) = pA ξ(c)pA + pS ξ(c)⌈ξ(1)⌉ pS = pA ξ(c)pA 6 pA for all c ∈ C . With similar reasoning we see that ξ(c) 6 pS for all c ∈ C in the second case. Assume there is an equalizer e : E → M4 of idM4 and adσ in (Pure vNop)op. By definition of purity e = c◦ h for some filter c and comprehension h. As e is supposed to be an equalizer of id and adσ, we have adσ ◦ c◦ h = adσ ◦ e = e = c◦ h, hence adσ ◦ c = c by the surjectivity of the corner h. We have just shown that then either c(b) 6 pA for all b or or c(b) 6 pS for all b. Hence h(e) 6 pA for all e ∈ E or h(e) 6 pS for all e ∈ E. In the first : M3 → M4. Clearly adσ ◦ ade† case, consider the map ade† and so there must be a (unique) map h : M3 → E with e ◦ h = ade† . However pA > e(h(T )) = ade† = 0, quod non. We reach a contradiction in the other case in a similar fashion. Hence Pure vNop has no coequalizers. (T ) 6 pS for all T ∈ M3 and so ade† S = ade† S S S S S 237 Errata These are errata to the printed version of this thesis, which have been corrected in this PDF. Additional errata will be published on the arXiv under 1803.01911. Erratum to 138 VII. The second part of the exercise starts with "Conclude that any quantum channel Φ : B(K ) → B(H ) . . . ". This should have read "Conclude that any quantum channel Φ from H to itself . . . " Erratum to 139 XI. The third-to-last sentence reads "Derive from the latter that for each y ∈ K ′ and rank-one projector e ∈ H , there is a y′ ∈ K ′ with U0(e⊗ y) = e⊗ y′." This should have been "Derive from the latter that for each y ∈ K ′ and unit vector e ∈ H , there is a y′ ∈ K ′ with U0(e⊗y) = e⊗y′.". Erratum to 141 IIa. A pre-Hilbert B-module X is self dual if every bounded B- linear τ : X → B is of the form τ (x) = ht, xi for some t ∈ X. In print the condition that τ should be bounded was accidentally lost when it was moved from 149 V. Erratum to 147 II. The assumption that the net xα is Cauchy in point 3 is superfluous. Erratum to 152 VIII. The map T should be presumed to be bounded. Erratum to 152 IX. The C∗-algebra B isn't assumed to be a von Neumann algebra, which it should have been (for otherwise 150 II wouldn't be applicable.) Erratum to 159 VIII. The displayed inequality f (T − pST pS) 6 f ( (1 − pS)T + pST (1 − pS) ) should have read f (T − pST pS) 6 f ( (1 − pS)T ) + f ( pST (1 − pS) ). Erratum to 169 X. The standard filter cb for b ∈ B is a map "⌈b⌉ B ⌈b⌉ → B" instead of a map "B → ⌈b⌉ B ⌈b⌉". Erratum to 197 V. Not every zero map is a quotient; only those into 0. Erratum to 199 VII. Not every zero map is a comprehension; only those from 0. 239 Erratum to 218 IX. Also in the last two points of the exercise, the map h should be presumed to be pristine. Erratum to 224 VI. The condition "f ◦ ϕ = hx, (· )xi" is a typo. What was meant is "f (ϕ(T, c)) = hx, T xi for all c ∈ C and T ∈ B(H )". The same mistake was made writing "π1 ◦ ϕ = hx, (· )xi" and "π2 ◦ ϕ = hy, (· )yi".
1305.0111	1	1305	2013-05-01T07:28:44	Bures Distance For Completely Positive Maps	[ "math.OA", "math.FA" ]	D. Bures had defined a metric on the set of normal states on a von Neumann algebra using GNS representations of states. This notion has been extended to completely positive maps between $C^$-algebras by D. Kretschmann, D. Schlingemann and R. F. Werner. We present a Hilbert $C^$-module version of this theory. We show that we do get a metric when the completely positive maps under consideration map to a von Neumann algebra. Further, we include several examples and counter examples. We also prove a rigidity theorem, showing that representation modules of completely positive maps which are close to the identity map contain a copy of the original algebra.	math.OA	math	Bures Distance For Completely Positive Maps B.V. Rajarama Bhat and K. Sumesh Abstract D. Bures had defined a metric on the set of normal states on a von Neumann algebra using GNS representations of states. This notion has been extended to completely positive maps between C ∗-algebras by D. Kretschmann, D. Schlingemann and R. F. Werner. We present a Hilbert C ∗-module version of this theory. We show that we do get a metric when the completely positive maps under consideration map to a von Neumann algebra. Further, we include several examples and counter examples. We also prove a rigidity theorem, showing that representation modules of completely positive maps which are close to the identity map contain a copy of the original algebra. Keywords: Completely positive maps, Hilbert C ∗-modules, Bures distance. AMS Classification: 46L08, 46L30 1 Preliminaries 1.1 Introduction Given a state φ on a C∗-algebra B we have the familiar GNS-triple (H, π, x), where H is a Hilbert space, π : B → B(H) is a representation (i.e., ∗-homomorphism) and x ∈ H is a vector such that φ(·) = hx, π(·)xi. Now it is a natural question to ask: If two states φ1, φ2 are close in some metric, whether the associated triples are close in some sense? Keeping this idea in mind, D. Bures ([7]) defines a distance between two states φ1, φ2 on B, as β(φ1, φ2) := inf kx1 − x2k , where the infimum is taken over all GNS-triples with common representation spaces: (H, π, x1), (H, π, x2) of φ1, φ2. D. Bures showed that β is indeed a metric. The notion has found uses in many areas ([1, 3, 4, 8]). D. Kretschmann, D. Schlingemann and R. F. Werner ([13]) extended this notion at first to completely positive (CP) maps from a C∗-algebra A to B(G) for some Hilbert space G and then to more general range C∗-algebras using an alternative definition of the Bures distance. They use the Stinespring representation ([23]) for the initial definition, which in the usual formulation requires the range space to be the whole algebra B(G). Here we develop the theory using Hilbert module language, which allows the range algebra to be any C∗-algebra, and the definition of the metric is a very natural extension of the definition given by Bures for states. Working with modules has several advantages. The results we get are of course same as that of [13], when the range algebra is a von Neumann algebra or an injective C∗-algebra. However, we show that one may not even get a metric (triangle inequality may fail) when the range algebra is a general C∗-algebra. There have been several papers ([2, 9]) on different methods to make exact com- putations of the Bures metric for states. We provide several examples with explicit computations of the Bures distance for CP-maps. In particular, we show that the 1 infimum in the definition of Bures metric may not be attained in all common rep- resentation modules, answering a question raised in [12, 13]. It turns out that the example is quite simple involving CP-maps on 2 × 2 matrix algebra. In the last Section we prove a rigidity theorem, which says that on von Neumann algebras, if a CP-map is strictly within unit distance (in Bures metric) from the identity map, then the GNS-module of the CP-map contains a copy of the original von Neumann algebra as a direct summand. We consider this as the most impor- tant positive result of this paper and we expect that the result will have further applications in the study of CP-maps. 1.2 Hilbert C∗-modules Let B be a C∗-algebra. A complex vector space E is a Hilbert B-module if it is a right B-module with a B-valued inner product, which is complete with respect to the associated norm (see [14, 17, 21] for basic theory). We denote the space of all bounded and adjointable maps between two Hilbert B-modules E1 and E2 by Ba(E1, E2). In particular, if E1 = E2 = E, then Ba(E, E) = Ba(E), which forms a C∗-algebra with natural algebraic operations. Let π : B → B(G) be a non-degenerate (i.e., span π(B)G = G) representation of B on a Hilbert space G. Given a Hilbert B-module E, we define the Hilbert space H := E ⊙ G as the completion of the inner product space obtained from the algebraic tensor product E ⊗ G by dividing out the null space of the semi-inner product hx ⊗ g, x′ ⊗ g′i := hg, π(hx, x′i)g′i, where x, x′ ∈ E, g, g′ ∈ G. We denote the equivalence class containing x ⊗ g by x ⊙ g. To each x ∈ E we associate the linear map Lx : g 7→ x ⊙ g in B(G, H) with adjoint L∗x : y ⊙ g 7→ π(hx, yi)g. Clearly L∗xLy = π(hx, yi) and Lxb = Lxπ(b) for all x, y ∈ E, b ∈ B. Also kLxk2 = kπ(hx, xi)k = kxk2. By identifying B with π(B) and x with Lx we may assume that E ⊆ B(G, H). Note that a 7→ a⊙idG : Ba(E) → B(H) is a unital ∗-homomorphism, and hence an isometry. So we may consider Ba(E) ⊆ B(H). Suppose A is another C∗-algebra. A Hilbert B-module E is called a Hilbert A- B-module if there exists a representation τ : A → Ba(E) which is non-degenerate (equivalently, unital if A is unital). If E is a Hilbert A-B-module, we may consider A ⊆ Ba(E), and we denote τ (a) by a itself and thereby τ (a)x = ax for all x ∈ E, a ∈ A. The composition map A τ−→ Ba(E) → B(H) is denoted by ρ; i.e., ρ(a) = a⊙ idG. Note that Lax = ρ(a)Lx. Also B(G, H) forms a Hilbert A-B(G)-module with left action ax := ρ(a)x. If E1 and E2 are two Hilbert A-B-modules, then a linear map Φ : E1 → E2 is said to be A-B-linear (or bilinear) if Φ(axb) = aΦ(x)b for all a ∈ A, b ∈ B, x ∈ E. The space of all bounded, adjointable and bilinear maps from E1 to E2 is denoted by Ba,bil(E1, E2). If E is a Hilbert A-B-module, then Ba,bil(E) is the relative commutant of the image of A in Ba(E). Suppose B ⊆ B(G) is a von Neumann algebra and E is a Hilbert B-module. Then we say E is a von Neumann B-module if E is strongly closed in B(G, H) ⊆ B(G ⊕ H). Thus, if x is an element in the strong closure E of a Hilbert B-module SOT−−→ x. All von Neumann E, then there exists a net (xα) ⊆ E such that Lxα B-modules are self-dual (in the sense that all B-valued functionals are given by a B-valued inner product), and hence they are complemented in all Hilbert B-module which contains it as a B-submodule. In particular, strongly closed B-submodules are complemented in a von Neumann B-module. If we think Ba(E) ⊆ B(H), then Ba(E) is a von Neumann algebra acting non-degenerately on the Hilbert space H. s 2 If A is a C∗-algebra, then by a von Neumann A-B-module we mean a von Neumann B-module E with a non-degenerate representation τ : A → Ba(E). In addition, if A is a von Neumann algebra and a 7→ hx, axi : A → B is a normal mapping for all x ∈ E (equivalently, the representation ρ : A → B(H) is normal), then we call E a two-sided von Neumann A-B-module. For more details see [17, 21, 22]. It is well-known that if ϕ : A → B is a CP-map between unital C∗-algebras, then there exists a Hilbert A-B-module E and x ∈ E such that ϕ(a) = hx, axi for all a ∈ A. The construction of E is by starting with A ⊗ B and defining a B-valued semi-inner product on it as ha1 ⊗ b1, a2 ⊗ b2i := b∗1ϕ(a∗1a2)b2, and usual quotienting and completion procedure (see [11, 16, 17, 18, 23]). The comparison with GNS for states is obvious. The pair (E, x) is called a GNS-construction for ϕ and E is called a GNS-module for ϕ. If further, span AxB = E, then (E, x) is said to be a minimal GNS-construction, and is unique up to isomorphism. If (both A and) B is a von Neumann algebra, then (E, x) can be chosen such that E is a (two-sided) von Neumann A-B-module. Here the closure for minimality is taken under strong operator topology. Note that if B = B(G), then L∗xρ(a)Lx = hx, axi = ϕ(a) for all a ∈ A. Thus (ρ, Lx, H) is a Stinespring representation for the CP-map ϕ : A → B(G). 1.3 Bures distance Given two unital C∗-algebras A and B, we let CP (A,B) denote the set of all nonzero CP-maps from A into B. 1.3.1 Definition. A Hilbert A-B-module E is said to be a common representation module for ϕ1, ϕ2 ∈ CP (A,B) if both of them can be represented in E, that is, there exist xi ∈ E such that ϕi(a) = hxi, axii, i = 1, 2. Note that we are demanding no minimality for the common representation mod- ule. So we can always have such a module. For, if ( Ei, xi) is the minimal GNS- construction for ϕi, then take E = E1 ⊕ E2, x1 = x1 ⊕ 0 and x2 = 0 ⊕ x2. For a common representation module E, define S(E, ϕi) to be the set of all x ∈ E such that ϕi(a) = hx, axi for all a ∈ A. 1.3.2 Definition. Let E be a common representation module for ϕ1, ϕ2 ∈ CP (A,B). Define βE(ϕ1, ϕ2) := inf(cid:8)kx1 − x2k : xi ∈ S(E, ϕi), i = 1, 2(cid:9) and the Bures distance β(ϕ1, ϕ2) := inf E βE(ϕ1, ϕ2) where the infimum is taken over all common representation module E. We have called β as a 'distance' in anticipation. Later we will show that it is indeed a metric under most situations, for instance, when B is a von Neumann algebra. But surprisingly β is not a metric in general. Our first job is to show that the definition here matches with that of [13]. We see it as follows. Suppose B = B(G). If E is a common representation module and xi ∈ S(E, ϕi), then (ρ, Lxi, E⊙G) is a Stinespring representation for ϕi with kx1 − x2k = kLx1 − Lx2k. On the other way if (π′, Vi, H′) is a Stinespring representation for ϕi, then E := B(G, H′) is a Hilbert1 A-B(G)-module with inner product hx1, x2i := 1If A is a von Neumann algebra and π is normal, then a 7→ hx, ayi = hx, π(a)yi is normal map from A → B for all x, y ∈ E. Thus E can be made into a two-sided von Neumann A-B(G)-module. 3 x∗1x2, composition as the right module action and left action given by ax := π′(a)x for all a ∈ A, x ∈ E. Clearly (E, Vi) is a GNS-construction for ϕi. Note that span EG = H′. We have H := E⊙G is a Hilbert space with inner product hx⊙g, x′⊙ g′i = hg′, x∗x′g′i = hxg, x′g′i. Thus x⊙ g 7→ xg is a unitary from U : H → H′. Note that ULVi = Vi and Uρ(a)U∗ = π′(a) for all a ∈ A. Identifying H with H′ through U, we get π′ = ρ and LVi = Vi. Therefore (π, Vi, H′) = (ρ, LVi, H). Thus there exists a one-one correspondence between the GNS-constructions {(E, x1), (E, x2)} and the Stinespring representations {(π′, V1, H′), (π′, V2, H′)} such that kx1 − x2k = kV1 − V2k. Hence β(ϕ1, ϕ2) coincides with the definition given in [13]. In particular, if B = B(C) = C, then β(ϕ1, ϕ2) is the Bures distance given in [7]. The following proposition says that β(ϕ1, ϕ2) coincide with the alternative defi- nition, given in [13], of Bures distance for CP-maps between arbitrary C∗-algebras. Subsequently will not be needing this definition and we present it here for the sake of completeness. 1.3.3 Proposition. With notation as above, β(ϕ1, ϕ2) = inf ϕ kϕ11(1) + ϕ22(1) − ϕ12(1) − ϕ21(1)k 1 2 where the infimum is taken over all CP-extensions ϕ : A → M2(B) of the form ϕ21 ϕ22(cid:21) with completely bounded maps ϕij : A → B satisfying ϕii = ϕi. ϕ =(cid:20)ϕ11 ϕ12 Proof. Let E be a common representation module and xi ∈ S(E, ϕi). Define ϕ : A → M2(B) by a 7→ [ϕij(a)], where ϕij(a) := hxi, axji. Then ϕ is a CP-map with kx1 − x2k2 = khx1, x1i + hx2, x2i − hx1, x2i − hx2, x1ik = kϕ11(1) + ϕ22(1) − ϕ12(1) − ϕ21(1)k . 1 ϕ kϕ11(1) + ϕ22(1) − ϕ12(1) − ϕ21(1)k 2 . To get Since E is arbitrary β(ϕ1, ϕ2) ≥ inf the reverse inequality, assume that ϕ = [ϕij] : A → M2(B) is a CP-map with ϕii = ϕi. Let ( E, x) be a GNS-construction of ϕ. Note that E is a Hilbert A-M2(B)- module. Given b ∈ B, x ∈ E define xb := x(bI), where I ∈ M2(B) is the identity matrix. Under this action E becomes a right B-module. Now for x1, x2 ∈ E define hx1, x2i′ := Pi,j hx1, x2iij, where hx1, x2iij is the (i, j)th entry of hx1, x2i ∈ M2(B). Then h·,·i′ is a B-valued inner product on E. Denote the resulting inner product B-module by E0. The left action of A on E induce a non-degenerate left action of A on E0. Complete E0 to get the Hilbert A-B-module E. Set xi = xeii, where {eij}, 1 ≤ i, j ≤ 2 are matrix units of M2(B). Then xi ∈ S(E, ϕi) and kx1 − x2k2 = khx1 − x2, x1 − x2i′k = khx1, x1i′ + hx2, x2i′ − hx1, x2i′ − hx2, x1i′k = kϕ11(1) + ϕ22(1) − ϕ12(1) − ϕ21(1)k . Since ϕ is arbitrary β(ϕ1, ϕ2) ≤ inf ϕ kϕ11(1) + ϕ22(1) − ϕ12(1) − ϕ21(1)k 1 2 . The following proposition says that Bures distance is stable under taking ampli- ations. 4 1.3.4 Proposition. Let A and B be unital C∗-algebras. Then for ϕ, ψ ∈ CP (A,B), β(ϕ, ψ) = β(ϕn, ψn) where ϕn, ψn : Mn(A) → Mn(B) are the amplifications of ϕ, ψ respectively for n ≥ 1. Proof. Fix n ≥ 1. Suppose E is a common representation module for ϕ, ψ and x1 ∈ S(E, ϕ), x2 ∈ S(E, ψ). Then diag(x1,· · · , x1) ∈ S(Mn(E), ϕn) and diag(x2,· · · , x2) ∈ S(Mn(E), ψn), and hence β(ϕn, ψn) ≤ kdiag(x1 − x2,· · · , x1 − x2)k = kx1 − x2k . Since x1, x2 and E are arbitrary β(ϕn, ψn) ≤ β(ϕ, ψ). Conversely, suppose F is a common representation module for ϕn, ψn and y1 ∈ S(F, ϕn), y2 ∈ S(F, ψn). If {eij},{fij}, 1 ≤ i, j ≤ n are matrix units of Mn(A), Mn(B) respectively, then E := {e11F f11} is a common representation module for ϕ, ψ in the natural way and more over, e11y1f11 ∈ S(E, ϕ) and e11y2f11 ∈ S(E, ψ). Also, ke11y1f11 − e11y2f11k2 = kf11he11(y1 − y2), e11(y1 − y2)if11k ≤ ky1 − y2k2 . Therefore β(ϕ, ψ) ≤ β(ϕn, ψn). 1.3.5 Proposition. Let A,B and C be unital C∗-algebras. Then for ϕi ∈ CP (A,B) and ψi ∈ CP (B,C), i = 1, 2, β(ψ1 ◦ ϕ1, ψ2 ◦ ϕ2) ≤ kϕ1k 2 β(ψ1, ψ2) + kψ2k 1 1 2 β(ϕ1, ϕ2). In particular, β(ψ2 ◦ ϕ1, ψ2 ◦ ϕ2) ≤ kψ2k 1 2 β(ϕ1, ϕ2). Proof. Suppose E, F are common representation modules for ϕi, ψi respectively, and xi ∈ S(E, ϕi), yi ∈ S(F, ψi), i = 1, 2. Then xi ⊙ yi ∈ S(E ⊙ F, ψi ◦ ϕi), and hence β(ψ1 ◦ ϕ1, ψ2 ◦ ϕ2) ≤ kx1 ⊙ y1 − x2 ⊙ y2k ≤ kx1 ⊙ y1 − x1 ⊙ y2 + x1 ⊙ y2 − x2 ⊙ y2k ≤ kx1kky1 − y2k + kx1 − x2k ky2k ≤ kϕ1k 2 ky1 − y2k + kx1 − x2k kψ2k 2 . 1 1 Since xi, yi, E and F are arbitrary the results holds. 1 2 . 1.3.6 Proposition. Let ϕ1, ϕ2 ∈ CP (A,B). Then (i) β(ϕ1, ϕ1 + ϕ2) ≤ kϕ2k (ii) (cid:12)(cid:12)β(ϕ1, ϕ2) − β(ϕ1, ǫϕ1 + (1 − ǫ)ϕ2)(cid:12)(cid:12) ≤ ǫ (iii) If ϕi(1) ≤ 1, then(cid:12)(cid:12)kϕ1k − kϕ2k(cid:12)(cid:12) ≤ 2β(ϕ1, ϕ2). 2 (kϕ1k 1 1 1 2 + kϕ2k 2 ) for 0 ≤ ǫ ≤ 1. Proof. (i) Suppose (Ei, xi) is a GNS-construction for ϕi, i = 1, 2. Then z1 := x1 ⊕ 0 ∈ S(E1 ⊕ E2, ϕ1) and z2 := x1 ⊕ x2 ∈ S(E1 ⊕ E2, ϕ1 + ϕ2), and hence β(ϕ1, ϕ1 + ϕ2) ≤ kz1 − z2k = kx2k = kϕ2k 1 2 . 5 (ii) Using triangle inequality and part (i), (cid:12)(cid:12)β(ϕ1, ϕ2) − β(ϕ1, ǫϕ1 + (1 − ǫ)ϕ2)(cid:12)(cid:12) ≤ β(ϕ2, ǫϕ1 + (1 − ǫ)ϕ2) ≤ β(ϕ2, (1 − ǫ)ϕ2) + β((1 − ǫ)ϕ2, ǫϕ1 + (1 − ǫ)ϕ2) ≤ kǫϕ2k ≤ ǫ 2 + kǫϕ1k 2 + kϕ2k 2 (kϕ1k 2 ). 1 2 1 1 1 1 (iii) Let E be a common representation for ϕ1, ϕ2 and xi ∈ S(E, ϕi). Then (cid:12)(cid:12)kϕ1k − kϕ2k(cid:12)(cid:12) =(cid:12)(cid:12)kx1k2 − kx2k2(cid:12)(cid:12) =(cid:12)(cid:12)(kx1k + kx2k)(kx1k − kx2k)(cid:12)(cid:12) = (kx1k + kx2k)(cid:12)(cid:12)kx1k − kx2k(cid:12)(cid:12) ≤ 2 kx1 − x2k . Since x1, x2 and E are arbitrary the result follows. 2 Bures distance: von Neumann algebras As is well-known one of the problems in dealing with Hilbert C∗-modules in contrast to Hilbert spaces is that in general submodules are not complemented, that is, there is a problem in taking orthogonal complements and writing the whole space as a direct sum. This problem is not there for von Neumann modules. Here we generalize almost all the results of [13], where the results stated mainly for the case when the range algebra is the algebra of all bounded operators on a Hilbert space. The proofs are similar, though we have also taken some ideas from [7]. We also give several examples and answer a question of [13] in the negative. In this Section we assume that A is a unital C∗-algebra, B ⊆ B(G) is a von Neumann algebra and ϕ1, ϕ2 ∈ CP (A,B). 2.1 Metric property To begin with we have the following proposition. 2.1.1 Proposition. If B ⊆ B(G) is a von Neumann algebra, then β(ϕ1, ϕ2) = inf E βE(ϕ1, ϕ2) (2.1) where the infimum is taken over all common representation modules E which are von Neumann A-B-module. Proof. Since von Neumann B-modules are Hilbert B-modules we have β(ϕ1, ϕ2) ≤ inf βE(ϕ1, ϕ2). To get the reverse inequality, assume that E is a common repre- sentation module for ϕ1, ϕ2. Then E := E ⊆ B(G, E ⊙ G) forms a von Neumann A-B-module. Since E ⊆ E we have E is a common representation module for ϕ1, ϕ2, and hence inf βE(ϕ1, ϕ2) ≤ β(ϕ1, ϕ2). s As we have taken B as von Neumann algebra for this Section, we may use (2.1) as the definition of Bures distance. Also by a common representation module and GNS-module we will mean a von Neumann A-B-module. However, note that for all 6 the results here, the algebra A can be a general C∗-algebra and the left action by A need not be normal. So we do not need that ϕ1, ϕ2 to be normal. The following result shows the existence of a sort of universal module where we can take infimum to compute the Bures distance. 2.1.2 Proposition. There exists a von Neumann A-B-module E such that: (i) For all ϕ1, ϕ2 ∈ CP (A,B), β(ϕ1, ϕ2) = βE(ϕ1, ϕ2); (ii) For a fixed ϕ1 ∈ CP (A,B) there exists ξ1 ∈ S(E, ϕ1) such that β(ϕ1, ϕ2) = inf(cid:8)kξ1 − ξ2k : ξ2 ∈ S(E, ϕ2)(cid:9) for all ϕ2 ∈ CP (A,B). Proof. For each ϕ ∈ CP (A,B) fix a GNS-construction (Eϕ, xϕ). Set Hϕ = Eϕ ⊙ G and H = ⊕Hϕ. Then E0 := ⊕ s Eϕ ⊆ B(G, H) is a von Neumann A-B-module. Note that S(E0, ϕ) is nonempty for all ϕ ∈ CP (A,B). Take E = E0 ⊕ E0 which is a von Neumann A-B-module. (i) Suppose ϕ1, ϕ2 ∈ CP (A,B) and E is a common representation module. We will prove that βE(ϕ1, ϕ2) ≤ βE(ϕ1, ϕ2). For that, it is enough to show that for all xi ∈ S(E, ϕi) there exists ξi ∈ S(E, ϕi) such that kξ1 − ξ2k ≤ kx1 − x2k. Take ξ′1 ∈ S(E0, ϕ1). Let U : span s Aξ′1B → span s Ax1B be the bilinear unitary satisfying U(aξ′1b) = ax1b. Let P be the bilinear projection of E onto span s Ax1B. Set x′2 := P x2 ∈ span s Ax1B ⊆ E, x′′2 := (1 − P )x2 ∈ (span s Ax1B)⊥ ⊆ E, ϕ′2(·) := hx′2, (·)x′2i and ϕ′′2(·) := hx′′2, (·)x′′2i. Clearly ϕ2 = ϕ′2 + ϕ′′2. Let ξ′2 = U∗(x′2) ∈ span s Aξ′1B ⊆ E0. Then hξ′2, aξ′2i = hU∗x′2, aU∗x′2i = hU∗x′2, U∗(ax′2)i = hx′2, ax′2i = ϕ′2(a). Let ξ′′2 ∈ S(E0, ϕ′′2). Set ξ1 = ξ′1 ⊕ 0 and ξ2 = ξ′2 ⊕ ξ′′2 . Then ξi ∈ S(E, ϕi) with kξ1 − ξ2k2 = khξ1, ξ1i + hξ2, ξ2i − 2Re(hξ1, ξ2i)k = khξ′1, ξ′1i + hξ′2, ξ′2i + hξ′′2 , ξ′′2i − 2Re(hξ′1, ξ′2i)k = khξ′1 − ξ′2, ξ′1 − ξ′2i + hξ′′2 , ξ′′2ik = khU(ξ′1 − ξ′2), U(ξ′1 − ξ′2)i + hξ′′2 , ξ′′2ik = khx1 − x′2, x1 − x′2i + hx′′2, x′′2ik = khx1, x1i + hx′2, x′2i − 2Re(hx1, x′2i) + hx′′2, x′′2ik = khx1, x1i + hx2, P x2i − 2Re(hx1, x′2i) + hx2, (1 − P )x2ik = khx1, x1i + hx2, x2i − 2Re(hx1, x′2i)k = khx1 − x2, x1 − x2ik = kx1 − x2k2 . (x1 = x1 ⊕ 0, x2 = x′2 ⊕ x′′2 in E) Since x1, x2 and E are arbitrary βE(ϕ1, ϕ2) ≤ β(ϕ1, ϕ2). (ii) Note that ξ1 ∈ S(E, ϕ1) is independent of E and ϕ2. If we denote ξ2 obtained 7 in part(i) by ξ2(x1, x2), then βE(ϕ1, ϕ2) = inf(cid:8)kξ − ξ′k : ξ ∈ S(E, ϕ1), ξ′ ∈ S(E, ϕ2)(cid:9) ≤ inf(cid:8)kξ1 − ξ′k : ξ′ ∈ S(E, ϕ2)(cid:9) ≤ inf(cid:8)kξ1 − ξ2(x1, x2)k : xi ∈ S(E, ϕi)(cid:9) = inf(cid:8)kx1 − x2k : xi ∈ S(E, ϕi)(cid:9) = βE(ϕ1, ϕ2) Since this is true for all common representation module E, we get β(ϕ1, ϕ2) ≤ βE(ϕ1, ϕ2) ≤ inf(cid:8)kξ1 − ξ′k : ξ′ ∈ S(E, ϕ2)(cid:9) ≤ β(ϕ1, ϕ2). This completes the proof. 2.1.3 Theorem. β is a metric on CP (A,B). Proof. Positive definiteness: Let ϕ1, ϕ2 ∈ CP (A,B). Take E and ξ1 ∈ S(E, ϕ1) as in proposition 2.1.2(ii). By definition β(ϕ1, ϕ2) ≥ 0. Now if β(ϕ1, ϕ2) = 0, then inf(cid:8)kξ1 − ξ2k : ξ2 ∈ S(E, ϕ2)(cid:9) = 0. Since S(E, ϕ2) is a norm closed subset of E, above equality implies that ξ1 ∈ S(E, ϕ2). Therefore ϕ1 = ϕ2. Symmetry: Clear from the definition. Triangle inequality: Let ϕ1, ϕ2, ϕ3 ∈ CP (A,B). Suppose E and ξ1 ∈ S(E, ϕ1) are as in proposition 2.1.2(ii). Then β(ϕ2, ϕ3) = inf(cid:8)kξ2 − ξ3k : ξi ∈ S(E, ϕi), i = 2, 3(cid:9) ≤ inf(cid:8)kξ2 − ξ1k : ξ2 ∈ S(E, ϕ2)(cid:9) + inf(cid:8)kξ1 − ξ3k : ξ3 ∈ S(E, ϕ3)(cid:9) = β(ϕ2, ϕ1) + β(ϕ1, ϕ3). 2.2 Intertwiners and computation of Bures distance The definition of Bures distance is abstract and does not give us indications as to how to compute it for concrete examples. In this Section, motivated by the work of [13], we show that Bures distance can be computed using intertwiners between two (minimal) GNS-constructions of CP-maps. Suppose E is a common representation module for ϕi and xi ∈ S(E, ϕi), i = 1, 2. Then kx1 − x2k2 = khx1 − x2, x1 − x2ik = kϕ1(1) + ϕ2(1) − 2Re(hx1, x2i)k. Thus β(ϕ1, ϕ2) is completely determined by the subsets {hx1, x2i : xi ∈ S(E, ϕi)} ⊆ B. This observation leads to the following Definition. 2.2.1 Definition. Given a common representation module E for ϕ1 and ϕ2 define and NE(ϕ1, ϕ2) :=(cid:8)hx1, x2i : xi ∈ S(E, ϕi)(cid:9) N(ϕ1, ϕ2) := ∪E NE(ϕ1, ϕ2) where the union is taken over all common representation module E. 8 Note that N(ϕ1, ϕ2) ⊆ B is always nonempty. Also if E is a common represen- tation module for ϕ1 and ϕ2, then βE(ϕ1, ϕ2) = inf N∈NE (ϕ1,ϕ2) kϕ1(1) + ϕ2(1) − 2Re(N)k 1 2 (2.2) with ϕ1(1) + ϕ2(1) − 2Re(N) = hx1 − x2, x1 − x2i ≥ 0. 2.2.2 Definition. Let (Ei, xi) be a GNS-construction for ϕi, i = 1, 2. Then define M(ϕ1, ϕ2) :=(cid:8)hx1, Φx2i : Φ ∈ Ba,bil(E2, E1),kΦk ≤ 1(cid:9). 2.2.3 Lemma. The set M(ϕ1, ϕ2) ⊆ B depends only on the CP-maps ϕi and not on the GNS-constructions (Ei, xi). Proof. We show that M(ϕ1, ϕ2) defined via (Ei, xi) coincides with M(ϕ1, ϕ2) which is defined via the minimal GNS-construction ( Ei, xi). Let Ui : Ei → span s AxiB be the bilinear unitary satisfying Ui(a xib) = axib for all a ∈ A, b ∈ B. Since span s AxiB ⊆ Ei is a complemented B-submodule, Ui ∈ Ba,bil( Ei, Ei) is an ad- jointable isometry ([14], Theorem 3.6). Note that Ui( xi) = xi and U∗i (xi) = xi. Now suppose hx1, Φx2i ∈ M(ϕ1, ϕ2), where Φ ∈ Ba,bil(E2, E1) with kΦk ≤ 1. Set Φ = U∗1 ΦU2. Then Φ ∈ Ba,bil( E2, E1) with k Φk ≤ 1. Also hx1, Φx2i = hU1 x1, ΦU2 x2i = h x1, U∗1 ΦU2 x2i = h x1, Φ x2i ∈ M (ϕ1, ϕ2). Hence M(ϕ1, ϕ2) ⊆ M(ϕ1, ϕ2). To get the reverse inclusion start with a Φ ∈ Ba,bil( E2, E1) and set Φ = U1 ΦU∗2 ∈ Ba,bil(E2, E1). 2.2.4 Proposition. If (Ei, xi) is a GNS-construction for ϕi, i = 1, 2, then (i) M(ϕ1, ϕ2) = N(ϕ1, ϕ2) = NE1⊕E2(ϕ1, ϕ2) and (ii) β(ϕ1, ϕ2) = 2 . 1 inf M∈M (ϕ1,ϕ2) kϕ1(1) + ϕ2(1) − 2Re(M)k Proof. (i) Suppose E is a common representation module and hz1, z2i ∈ NE(ϕ1, ϕ2). Set E1 = E2 = E and Φ = idE. Then, from above Lemma, hz1, z2i = hz1, Φz2i ∈ M(ϕ1, ϕ2). Since z1, z2 and E are arbitrary N(ϕ1, ϕ2) ⊆ M(ϕ1, ϕ2). In particular, M(ϕ1, ϕ2) is nonempty. For the reverse inclusion, let hx1, Φx2i ∈ M(ϕ1, ϕ2). Set z1 = x1 ⊕ 0 and z2 = Φx2 ⊕ √idE − Φ∗Φx2 in E1 ⊕ E2. Then hz1, az1i = hx1, ax1i = ϕ1(a) and hz2, az2i = hΦx2 ⊕pidE2 − Φ∗Φx2, a(Φx2) ⊕ apidE2 − Φ∗Φx2i = hΦx2, Φ(ax2)i + hpidE2 − Φ∗Φx2,pidE2 − Φ∗Φax2i = hx2, Φ∗Φ(ax2)i + hx2, (idE2 − Φ∗Φ)ax2i = hx2, ax2i = ϕ2(a) for all a ∈ A. Thus (E1⊕E2, zi) is a GNS-construction for ϕi. Note that hx1, Φx2i = hz1, z2i ∈ NE1⊕E2(ϕ1, ϕ2). Hence M(ϕ1, ϕ2) ⊆ NE1⊕E2(ϕ1, ϕ2). Thus N(ϕ1, ϕ2) ⊆ M(ϕ1, ϕ2) ⊆ NE1⊕E2(ϕ1, ϕ2) ⊆ N(ϕ1, ϕ2). (ii) Follows from equation (2.2). 2.2.5 Corollary. If (Ei, xi) is GNS-construction for ϕi, i = 1, 2, then β(ϕ1, ϕ2) = βE1⊕E2(ϕ1, ϕ2) = inf(cid:8)kx1 ⊕ 0 − y1 ⊕ y2k : y1 ⊕ y2 ∈ S(E1 ⊕ E2, ϕ2)(cid:9). 9 Proof. Suppose hx1, Φx2i ∈ M(ϕ1, ϕ2). Then, from the proof of Proposition 2.2.4, we have hx1, Φx2i = hz1, z2i, where zi ∈ S(E1 ⊕ E2, ϕi) with z1 = x1 ⊕ 0. Denote the z2 obtained by z2(Φ). Then, from proposition 2.2.4(ii), β(ϕ1, ϕ2) = inf(cid:8)kϕ1(1) + ϕ2(1) − 2Re(hx1 ⊕ 0, z2(Φ)i)k ≥ inf(cid:8)kϕ1(1) + ϕ2(1) − 2Re(hx1 ⊕ 0, y1 ⊕ y2i)k = inf(cid:8)kx1 ⊕ 0 − y1 ⊕ y2k : y1 ⊕ y2 ∈ S(E1 ⊕ E2, ϕ2)(cid:9) ≥ βE1⊕E2(ϕ1, ϕ2). 1 1 2 : Φ ∈ Ba,bil(E2, E1),kΦk ≤ 1(cid:9) 2 : y1 ⊕ y2 ∈ S(E1 ⊕ E2, ϕ2)(cid:9) the states ϕi : A → C given by ϕi(f ) =R f dµi, where µ1 and µ2 are two equivalent 2.2.6 Example. Let (X, F, µ) be a measure space and let A = L∞(X, µ). Consider (i.e., absolutely continuous each other) probability measures on (X, F) such that µi << µ, i = 1, 2. Let h be a positive function (Radon Nikodym derivative) on X such that dµ1 = hdµ2. Clearly Ei = L2(X, µi) is a von Neumann A-C-module with left multiplication as the left action. Also (Ei, 1) is a GNS-construction for ϕi. Suppose g1 ⊕ g2 ∈ S(E1 ⊕ E2, ϕ2). Then Z f dµ2 = hg1 ⊕ g2, f (g1 ⊕ g2)i =Z g12f dµ1 +Z g22f dµ2 =Z (g12h + g22)f dµ2 for all f ∈ A, and hence g12h + g22 = 1 a.e., µ2. Therefore 1 β(ϕ1, ϕ2) = inf(cid:8)k1 ⊕ 0 − g1 ⊕ g2k : g1 ⊕ g2 ∈ S(E1 ⊕ E2, ϕ2)(cid:9) 2 : g12h ≤ 1 a.e., µ2(cid:9) 2 : g1 ≥ 0 and 0 ≤ g2 = inf(cid:8)(h1 − g1, 1 − g1i + hg2, g2i) = inf(cid:8)(2 − 2Re(Z g1dµ1)) = √2 inf(cid:8)(1 −Z g1hdµ2) = √2(1 −Z √hdµ2) 2 : g12h + g22 = 1 a.e., µ2(cid:9) 1h ≤ 1 a.e., µ2(cid:9) 2 . 1 1 1 In particular, if we take X = {1, 2, . . . , n}, µ the counting measure, µ1(i) = pi and µ2(i) = qi, where 0 < pi, qi < 1 such that P pi = P qi = 1, then β(ϕ1, ϕ2) = √2(1 −P√piqi) Here we compute the Bures distance for homomorphisms and for some other 2 . 1 special cases. 2.2.7 Corollary. Let ϕ1, ϕ2 : A → B be two unital ∗-homomorphisms. 2 : b ∈ B,kbk ≤ 1, ϕ1(a)b = bϕ2(a) ∀a ∈ (i) Then β(ϕ1, ϕ2) = √2 inf(cid:8)k1 − Re(b)k A(cid:9). 1 (ii) If A = B and ϕ2(a) = u∗ϕ1(a)u for some unitary u ∈ B, then β(ϕ1, ϕ2) = √2 inf(cid:8)k1 − Re(b′u)k 1 2 : b′ ∈ ϕ1(A)′,kb′k ≤ 1(cid:9). 10 (iii) If u ∈ Mn(C) is a unitary and ϕ : Mn(C) → Mn(C) is the ∗-homomorphism ϕ(a) = u∗au, then β(id., ϕ) = √2 inf(cid:8)k1 − Re(λu)k 1 2 : λ ∈ [−1, 1](cid:9). Proof. (i) Let Ei be the von Neumann A-B-module B with left action ax := ϕi(a)x for all a ∈ A, x ∈ Ei. Then (Ei, 1) is the minimal GNS-construction for ϕi. Suppose Φ ∈ Ba,bil(E2, E1). Then ϕ1(a)Φ(1) = aΦ(1) = Φ(a1) = Φ(ϕ2(a)) = Φ(1)ϕ2(a) for all a ∈ A. Clearly, for a fixed b0 ∈ B satisfying ϕ1(a)b0 = b0ϕ2(a), the map b 7→ b0b is an element of Ba,bil(E2, E1). Thus β(ϕ1, ϕ2) = inf(cid:8)kϕ1(1) + ϕ2(1) − 2Re(M)k 1 = inf(cid:8)k2 − 2Re(Φ(1))k = √2 inf(cid:8)k1 − Re(b)k 1 1 2 : M ∈ M(ϕ1, ϕ2)(cid:9) 2 : Φ ∈ Ba,bil(E2, E1),kΦk ≤ 1(cid:9) 2 : b ∈ B,kbk ≤ 1, ϕ1(a)b = bϕ2(a) ∀a ∈ A(cid:9). (ii) Suppose b ∈ B. Then ϕ1(a)b = bϕ2(a) for all a ∈ A implies that bu∗ ∈ ϕ1(A)′, and hence b = b′u for some b′ ∈ ϕ1(A)′ ⊆ B. (iii) This follows from (ii), since M′n = CI. In [13] it is shown that the Bures distance is comparable with completely bounded norm when B = B(G), and the following bounds were obtained. In fact, the lower bound holds even for an arbitrary unital C∗-algebra B. 2.2.8 Theorem ([13]). For ϕ1, ϕ2 ∈ CP (A, B(G)), kϕ1 − ϕ2kcb pkϕ1kcb +pkϕ2kcb ≤ β(ϕ1, ϕ2) ≤qkϕ1 − ϕ2kcb. Moreover, there exists a common representation module E and corresponding GNS- construction (E, xi) for ϕi such that β(ϕ1, ϕ2) = βE(ϕ1, ϕ2) = kx1 − x2k . 2.2.9 Example. In general, the upper bound given in Theorem 2.2.8 may fails to hold if the cb-norm is replaced by the operator norm. For example, consider the CP-maps ϕi : M2(C) → M2(C) given by Let E = M8×2(C) which is a von Neumann M2(C)-M2(C)-module with module actions given by a12 ϕ1((cid:2)aij(cid:3)) :=(cid:20)a11 + 2a22  axb := ax1b ax2b ax3b ax4b a21 a22 + 2a11(cid:21) ∀x = x1 x2 x3 x4 0 and ϕ2((cid:2)aij(cid:3)) :=(cid:20)2a22  ∈ E and a, b, xi ∈ M2(C). 11 0 2a11(cid:21) . Then E is a common representation module with and 0 −1√2 √3√2 √3√2 0 0 0 1 0 #t z1 : =" 1 0 0 0 z2 : =(cid:20) 0 0 0 0 0 1 0 −1 0 (cid:21)t 0 0 0 0 1 0 1 0 1√2 0 ∈ S(E, ϕ1) ∈ S(E, ϕ2), where 't' stands for transpose. Note that if x ⊕ y = [xij] ⊕ [yij] ∈ S(E ⊕ E, ϕ2), then by evaluating ϕ2 at matrix units, we see that xi1 = yi1 = 0 = xk2 = yk2, i = 1, 3, 5, 7, k = 2, 4, 6, 8 and Hence (xi22 + yi22).  (∗) β(ϕ1, ϕ2) = inf kz1 ⊕ 0 − x ⊕ yk (xi1xi−1,2 + yi1yi−1,2) = 0, Xi=2,4,6,8 Xi=2,4,6,8 (xi12 + yi12) = 2 = Xi=1,3,5,7 = inf (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≥ inf (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) −x12 − x41 0 2 1 −x12 − x41 "5 − Re(√6x61 − √2x81) "5 − Re(√6x61 − √2x81) 5 − Re(√6x52 + √2x72)#(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 5 − Re(√6x52 + √2x72)#(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) elementary calculus shows that β(ϕ1, ϕ2) ≥p5 − √2 − √6. Note that kz1 − z2k = p5 − √2 − √6, and hence β(ϕ1, ϕ2) = p5 − √2 − √6 > 1. But ϕ1 − ϕ2 is the where the infimums are taken over all x ⊕ y ∈ E1 ⊕ E2 satisfying (∗). Now some transpose map. Therefore 1 = kϕ1 − ϕ2k < β(ϕ1, ϕ2)2 < kϕ1 − ϕ2kcb = 2 (see [18] for the computation of cb-norm for transpose map). 0 1 2 Theorem 2.2.8 guarantees the existence of a common representation module, where Bures distance is attained. It is a natural question as to whether Bures distance is attained in every common representation module. This is true for states ([4]). The question in the general case was asked by [12, 13]. Here we resolve it in the negative through a simple counter example. 2.2.10 Example. Consider the (normal) CP-maps ϕi : M2(C) → M2(C) given by ϕi(a) := a∗i aai, where a1 =(cid:20)1 0 is the minimal GNS-construction for ϕi. Set x1 = x1 ⊕ 0 and x2 = 0 ⊕ x2. Then xi ∈ S( E1 ⊕ E2, ϕi) and 0 0(cid:21). Then ( Ei, xi) := (M2(C), ai) 0 0(cid:21) and a2 =(cid:20)0 1 β(ϕ1, ϕ2) = β E1⊕ E2(ϕ1, ϕ2) ≤ kx1 − x2k = kIk = 1. Clearly, E := M2(C) is a common representation module. It is not hard to see that S(E, ϕi) = {λai : λ ∈ C,λ = 1}. Now for any xi = λiai ∈ S(E, ϕi), kx1 − x2k2 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) " 1 −λ2λ1 −λ1λ2 1 #(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 12 = sup(λ : λ ∈ σ(cid:16)" 1 −λ2λ1 −λ1λ2 1 #(cid:17)) = 2. Hence βE(ϕ1, ϕ2) = √2 > 1 ≥ β(ϕ1, ϕ2). Note that here β(ϕ1, ϕ2) ≤ 1 = pkϕ1 − ϕ2k. Conjecture. If ϕ, ψ ∈ CP (A,B), then β(ϕ, ψ) = sup φ,n β(φ ◦ ϕn, φ ◦ ψn) where the supremum is taken over all states φ : Mn(B) → C, n ≥ 1. From Proposition 1.3.4 and 1.3.5 we have β(φ◦ ϕn, φ◦ ψn) ≤ β(ϕn, ψn) = β(ϕ, ψ) for all states φ : Mn(B) → C, n ≥ 1. If the conjecture can be proved directly, then using the upper bound for states [7, 13] we get an alternative proof of the upper bound for Bures metric: β(ϕ, ψ) = sup φ,n β(φ ◦ ϕn, φ ◦ ψn) ≤ sup φ,n pkφ ◦ ϕn − φ ◦ ψnk =qkϕ − ψkcb. 3 Bures distance: C∗-algebras This Section consists mostly of counter examples. But results similar to the last section do hold for injective C∗-algebras. 3.1 Counter examples We saw that if the range algebras are von Neumann algebras, then the Bures metric can be computed using intertwiners. It was crucial that the space of intertwiners was independent of the choice of GNS-constructions (Lemma 2.2.3 ). The first example here shows that this is no longer the case for some range C∗-algebras. We have another example to show that the upper bound computed for β in Theorem 2.2.8 may not hold for general range C∗-algebras. Finally, as a worst case scenario we have a tricky example to show that even the triangle inequality may fail to hold. 3.1.1 Example. If ϕ1 and ϕ2 are CP-maps between C∗-algebras, then M(ϕ1, ϕ2) may depends on the GNS-construction. For example, consider the CP-maps ϕi : C([0, 2π]) → C([0, 2π]) given by ϕi(f ) := gif , where gi(t) = sin(t)i for all t ∈ [0, 2π], i = 1, 2. Set xi = √gi and Ei = span{√gif : f ∈ C([0, 2π])} = {f ∈ C([0, 2π]) : f (0) = f (π) = f (2π) = 0}. Then ( Ei, xi) is the minimal GNS-construction for ϕi. Define the adjointable bilinear map Φ : E2 → E1 by Φ(f ) = gf , where g(t) =( 1 if 0 ≤ t < π, 1 if π ≤ t ≤ 2π. 2 Since Φ is a contraction h x1, Φ x2i ∈ M (ϕ1, ϕ2). We have (Ei, xi) := (C([0, 2π]), xi) is also a GNS-construction for ϕi. Now if Φ : E2 → E1 is an adjointable bilinear map, then Φ(f ) = Φ(1)f for all f ∈ C([0, 2π]). Thus Ba,bil(E2, E1) = {f 7→ hf : h ∈ C([0, 2π])}. Hence if h x1, g x2i = h x1, Φ x2i ∈ M(ϕ1, ϕ2), then h x1, g x2i = h x1, h x2i for some h ∈ C([0, 2π]); i.e., x1(t)g(t) x2(t) = x1(t)h(t) x2(t), ∀t ∈ [0, 2π] g(t) = h(t), ∀t ∈ [0, 2π] r {0, π, 2π} ⇒ 13 which is not possible since h is continuous on [0, 2π]. So h x1, Φ x2i /∈ M(ϕ1, ϕ2). 3.1.2 Example. Suppose H is an infinite dimensional Hilbert space and p ∈ B(H) is an orthogonal projection such that both p and q := (1 − p) have infinite rank. Let A = C∗{K(H) ∪ {I}} and let u = λp + λq, where λ = eiθ is a scalar with − π 2 < θ < π 2 . Note that u ∈ B(H) is a unitary. Define ∗-homomorphisms ϕi : A → A by ϕ1(a) := a and ϕ2(a) := u∗au. Now suppose E is a common representation module for ϕ1, ϕ2 and xi ∈ S(E, ϕi). Since kaxi − xiϕi(a)k = 0, we get axi = xiϕi(a) for all a ∈ A. Then ahx1, x2i = ϕ1(a)hx1, x2i = hx1, x2iϕ2(a) = hx1, x2iu∗au for all a ∈ A, and hence hx1, x2iu∗ ∈ A′. Therefore hx1, x2i = λ′u for some λ′ ∈ C. Since hx1, x2i ∈ A and u 6∈ A we have λ′ = 0, whence hx1, x2i = 0. Also since E and xi ∈ S(E, ϕi) are arbitrary β(ϕ1, ϕ2) = inf 2 = √2. E,xi kx1 − x2k = kϕ1(1) + ϕ2(1)k 1 Now we prove thatpkϕ1 − ϕ2kcb < β(ϕ1, ϕ2). For a = [aij] ∈ B(H) = B(Hp ⊕ H⊥p ), where Hp = ran(p), kϕ1(a) − ϕ2(a)k = ka − u∗auk 2 2 0 )a12 (1 − λ 0 (1 − λ2)a21 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:20)a11 a12 0 λ(cid:21)(cid:13)(cid:13)(cid:13)(cid:13) a21 a22(cid:21) −(cid:20)λ 0 0 λ(cid:21)∗(cid:20)a11 a12 a21 a22(cid:21)(cid:20)λ 0 (cid:21)(cid:13)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:20) )a12(cid:13)(cid:13),(cid:13)(cid:13)(1 − λ2)a21(cid:13)(cid:13)(cid:9) = max(cid:8)(cid:13)(cid:13)(1 − λ ≤(cid:12)(cid:12)1 − λ2(cid:12)(cid:12)kak so that kϕ1 − ϕ2k ≤ 1 − λ2. But a =(cid:20)0 I I 0(cid:21) is of norm one and k(ϕ1 − ϕ2)(a)k = 1 − λ2, whence kϕ1 − ϕ2k = 1 − λ2 =(cid:12)(cid:12)λ(λ − λ)(cid:12)(cid:12) =(cid:12)(cid:12)λ − λ(cid:12)(cid:12). Now for all n ≥ 1, A ∈ Mn(A). Then, as above, we get k(ϕ1 − ϕ2)nk =(cid:12)(cid:12)λ − λ(cid:12)(cid:12). Thus pkϕ1 − ϕ2k =qkϕ1 − ϕ2kcb =q(cid:12)(cid:12)λ − λ(cid:12)(cid:12) < √2 = β(ϕ1, ϕ2). Now if ϕi is considered as a map into B(H) denote it by eϕi. Then b ∈ eϕ1(A)′ ⊆ if we let Un, Pn and Qn denote the n × n diagonal matrix with diagonal u, p and q respectively, then Un = λPn + λQn and (ϕ1 − ϕ2)n(A) = A − U∗nAUn for all B(H) implies that ba = ab for all a ∈ K(H) ⊆ A, so that b = λbI for some λb ∈ C. From Corollary 2.2.7, 1 2 : λ′ ∈ C,λ′ ≤ 1(cid:9) β(eϕ1,eϕ2) = √2 inf(cid:8)k1 − Re(λ′u)k 2 1 √2 k1 − Re(u)k ≤ = √21 − Re(λ) < √2 = β(ϕ1, ϕ2). 1 2 14 3.1.3 Example. Let H be an infinite dimensional Hilbert space. Consider the unital C∗-subalgebra A : = C∗(K(H ⊕ H) ∪n(cid:20)I 0 =((cid:20)λ1I + a11 a21 0 0(cid:21) ,(cid:20)0 0 0 I(cid:21)o) a12 λ2I + a22(cid:21) : λi ∈ C, aij ∈ K(H)) of B(H ⊕ H), where K(·) is the set of compact operators. Suppose u ∈ B(H) is a unitary and 1 < r ∈ R. Set in B(H ⊕ H). Define CP-maps ϕi : A → A by ϕi(a) := z∗i azi, i = 1, 2, 3. Note that each ϕi has the form, ϕi(·) =(cid:20)0 0 I 0 0 rI(cid:21) z1 =(cid:20)0 u 0 rI(cid:21) and z3 =(cid:20)0 0 rI(cid:21) , z2 =(cid:20)0 0 ∗(cid:21). Let E12 =((cid:20)x11 λ1u + x12 x21 λ2I + x22(cid:21) : λi ∈ C, xij ∈ K(H)) which is a Hilbert A-A-module with a natural inner product and bimodule structure. Note that zi ∈ S(E12, ϕi), i = 1, 2, and hence β(ϕ1, ϕ2) ≤ kz1 − z2k = 1. Similarly E23 =((cid:20)x11 λ1I + x12 x21 λ2I + x22(cid:21) : λi ∈ C, xij ∈ K(H)) is a Hilbert A-A-module with zi ∈ S(E23, ϕi), i = 2, 3, and β(ϕ2, ϕ3) ≤ kz2 − z3k = 1. Now we will show that β(ϕ1, ϕ3) > 2 ≥ β(ϕ1, ϕ2) + β(ϕ2, ϕ3) so that β fails to satisfy triangle inequality. Suppose E is a common representation module for ϕ1, ϕ3. We prove that hx1, x3i = 0 for all xi ∈ S(E, ϕi). If we proved this, then E and xi ∈ S(E, ϕi) arbitrary implies that β(ϕ1, ϕ3) = inf E,xi kx1 − x3k = kϕ1(1) + ϕ3(1)k Suppose hx1, x3i = [aij]. Since 0 ≤(cid:20)hx1, x1i hx3, x1i have a11 = a12 = a21 = 0. Also for all a ∈ K(H), we get (cid:20)a 0 0 0(cid:21) x1 = x1(cid:20)0 0 u∗au(cid:21) 0 and (Simply look at the norm of the difference.) Hence 1 2 =p2(1 + r2) > 2. hx1, x3i 0 0 ∗ 0 a∗11 a∗21 a∗12 a∗22 hx3, x3i(cid:21) = (cid:20)a 0 0 0(cid:21) x3 = x3(cid:20)0 0 0 a(cid:21) . a11 a12 a21 a22 0 0 0 ∗  we which implies that u∗aua22 = a22a; i.e., aua22 = ua22a for all a ∈ K(H). Hence ua22 = λI for some λ ∈ C. Thus a22 = λu∗. Since a22 ∈ K(H) and u∗ /∈ K(H) we have λ = 0, and hence a22 = 0 and hx1, x3i = 0. 15 i.e., 0 (cid:20)0 0 u∗au(cid:21)hx1, x3i = hx1, x3i(cid:20)0 0 0 a(cid:21) ; (cid:20)0 0 u∗au(cid:21)(cid:20)0 0 a22(cid:21)(cid:20)0 0 0 a(cid:21) 0 a22(cid:21) =(cid:20)0 0 0 0 3.2 Injective C∗-algebras Recall that a C∗-algebra B is an injective C∗-algebra if, whenever C is a C∗-algebra, S an operator system contained in C, and ϕ : S → B is a completely positive contraction, then ϕ extends to a completely positive contraction ϕ : C → B. Further, this is equivalent to saying that there is a faithful representation π of B on a Hilbert space G, such that there is a conditional expectation from B(G) onto π(B). See [5, 18, 24] for details. 3.2.1 Proposition. Let A and B be unital C∗-algebras and ϕ1, ϕ2 ∈ CP (A,B). Suppose B is an injective unital C∗-algebra and π : B → B(G) is a faithful repre- sentation of B on G. Then β(ϕ1, ϕ2) = β(π ◦ ϕ1, π ◦ ϕ2). Proof. Since B is injective there exists a completely positive conditional expectation P : B(G) → π(B). Take ϕ = π−1◦ P : B(G) → A. Then ϕ is a contractive CP-map. Moreover, ϕ ◦ π ◦ ϕi = ϕi, i = 1, 2. Now by Proposition 1.3.5, β(ϕ1, ϕ2) = β(ϕ ◦ π ◦ ϕ1, ϕ ◦ π ◦ ϕ2) ≤ β(π ◦ ϕ1, π ◦ ϕ2) ≤ β(ϕ1, ϕ2). From Proposition 1.3.5, we know that β(π ◦ ϕ1, π ◦ ϕ2) ≤ β(ϕ1, ϕ2) even for an arbitrary C∗-algebra B. But, in general, equality may not holds. See example 3.1.2. The following bounds were first obtained in [13]. 3.2.2 Corollary. If B is an injective unital C∗-algebra, then β is a metric on CP (A,B) and kϕ1 − ϕ2kcb pkϕ1kcb +pkϕ2kcb ≤ β(ϕ1, ϕ2) ≤qkϕ1 − ϕ2kcb. Further, there exists common representation module E and corresponding GNS- construction (E, xi) for ϕi such that β(ϕ1, ϕ2) = βE(ϕ1, ϕ2) = kx1 − x2k . Proof. Suppose π : B → B(G) is a faithful representation of B. Now the first part follows from Theorem 2.1.3 and Proposition 3.2.1. Also from Theorem 2.2.8 and Proposition 3.2.1, we have β(ϕ1, ϕ2) = β(π ◦ ϕ1, π ◦ ϕ2) ≤qkπ ◦ ϕ1 − π ◦ ϕ2kcb =qkϕ1 − ϕ2kcb. Now, from Theorem 2.2.8, we know that there exists a von Neumann A-B(G)- module F with yi ∈ S(F, π ◦ ϕi) such that ky1 − y2k = β(π ◦ ϕ1, π ◦ ϕ2). Given b ∈ B, y ∈ F define yb := yπ(b). Under this action, F forms a right B-module, denoted by E0. Let P : B(G) → π(B) be a completely positive conditional expectation satisfying P (b1ab2) = b1P (a)b2 for all bi ∈ π(B), a ∈ B(G). Now define a B-valued semi-inner product on E0 by hx1, x2i′ := π−1P (hx1, x2i). Let E be the completion of the B-valued inner product space E0/N, where N := {x ∈ E0 : hx, xi′ = 0}. Then E is a Hilbert A-B-module with left action induced by that of A on F . Note that xi := yi + N ∈ S(E, ϕi), i = 1, 2 are such that βE(ϕ1, ϕ2) ≤ kx1 − x2k ≤ ky1 − y2k = β(π ◦ ϕ1, π ◦ ϕ2) = β(ϕ1, ϕ2). Thus β(ϕ1, ϕ2) = βE(ϕ1, ϕ2) = kx1 − x2k. 16 4 Bures distance and a rigidity theorem Observe that for the identity map on a unital C∗-algebra B the GNS-module is B itself. Here we show that if a CP-map on a von Neumann algebra B is close to the identity map in Bures distance then the GNS-module has a copy of B. Suppose B ⊂ B(G) is a von Neumann algebra and ϕ : B → B is a CP-map. 4.0.3 Proposition. If (E, x) is the minimal GNS-construction for ϕ, then the fol- lowing are equivalent: (i) The center CB(E) := {y ∈ E : by = yb ∀b ∈ B} contains a unit vector. (ii) E ∼= B ⊕ F for some von Neumann B-B-module F . (iii) There exists an element c ∈ B such that the two sided (strongly closed) ideal generated by c is B, and a CP-map ψ : B → B such that ϕ(b) = c∗bc + ψ(b) ∀b ∈ B. Proof. (i) ⇒ (ii): Let z ∈ CB(E) be a unit vector. The two sided B-B-module generated by z is naturally isomorphic to B by bz 7→ b, and let us denote it by Bz. Then E decomposes as Bz ⊕ (Bz)⊥. (ii) ⇒ (iii): Without loss of generality, we may take E = B ⊕ F . Then x ∈ E decomposes as x = c⊕ y with c ∈ B, y ∈ F . Clearly, ϕ(b) = hx, bxi = c∗bc +hy, byi, and we can take ψ(b) = hy, byi for all b ∈ B. Note that B is the two sided (strongly closed) ideal generated by c. (iii) ⇒ (i): Note that the CP-map b 7→ c∗bc is dominated by the CP-map ϕ, and hence there exists a vector z ∈ E (see [6, 17]) such that c∗bc = hz, bzi for all b ∈ B. Note that, for elements a, a′, b, d, d′ ∈ B, (acd)∗b(a′cd′) = d∗(c∗a∗ba′c)d = d∗hz, a∗ba′zid′ = hazd, ba′zd′i. It follows that for any element d in the (strongly closed) ideal generated by c, there exists an element zd ∈ E such that d∗bd = hzd, bzdi. Taking d = 1, we have an element w ∈ E such that b = hw, bwi for all b ∈ B. Observe that w is a unit vector. Direct computation yields hbw−wb, bw−wbi = 0, hence w is in the center CB(E). 4.0.4 Theorem. Let ϕ : B → B be a CP-map such that β(id., ϕ) < 1. Let (E, x) be a GNS-construction for ϕ. Then E ∼= B ⊕ F for some von Neumann B-B-module F . Proof. Without loss of generality, assume that (E, x) is the minimal GNS-construction for ϕ. Let ε > 0 be such that β(id., ϕ) + ε < 1. Since the identity map has (B, 1) as its GNS-construction, from Theorem 2.2.5, there exists z1 = 1⊕0, z2 = c⊕y in B⊕E such that kz1 − z2k ≤ β(id, ϕ) + ε < 1 and ϕ(b) = hz2, bz2i = c∗bc +hy, byi. Further, as k1 − ck ≤ kz1 − z2k < 1 we note that c is invertible. Therefore the ideal generated by c is whole of B. Now the result follows from the previous Proposition. Acknowledgment. The authors wish to thank G. Ramesh and K. B. Sinha for several useful discussions on Bures distance. References [1] Damian F. Abasto,P. Zanardi, Thermal states of the Kitaev honeycomb model: Bures metric analysis. Phys. Rev. A (3) 79 (2009), no. 1, 012321, 6 pp. 17 [2] S. J. Akhtarshenas, An explicit computation of the Bures metric over the space of N-dimensional density matrices. J. Phys. A 40 (2007), no. 37, 11333- 11341. [3] P.M. Alberti, G. Peltri, On Bures distance over standard form vN-algebras, math.OA/0008164 v3, September 2000. [4] H. Araki, Bures distance function and a generalization of Sakai's non- commutative Radon-Nikodym theorem, Publ. Res. Inst. Math. Sci. 8 (1972), 335-362 [5] W. B. Arveson, On subalgebras of C∗-algebras, Bull. Amer. Math. Soc. 75 (1969), 790-794. [6] B. V. Rajarama Bhat, M. Skeide, Tensor product systems of Hilbert mod- ules and dilations of completely positive semigroups, Infin. Dimens. Anal. Quan- tum Probab. Relat. Top. 3 (2000), no. 4, 519-575. [7] D. Bures, An extension of Kakutani's theorem on infinite product measures to the tensor product of semifinite W ∗-algebras, Trans. Amer. Math. Soc. 135 (1969), 199-212. [8] J. Dittmann, Yang-Mills equation and Bures metric. Lett. Math. Phys. 46 (1998), no. 4, 281-287. [9] J. Dittmann, Explicit formulae for the Bures metric J. Phys. A 32 (1999), no. 14, 2663-2670. [10] I. Kaplansky, Modules over operator algebras, Amer. J. Math. 75 (1953), 839-858. [11] G. G. Kasparov, Hilbert C∗-modules: theorems of Stinespring and Voiculescu, J. Operator Theory 4 (1980), no. 1, 133-150. [12] D. Kretschmann, D. Schlingemann, R. F. Werner, The information- disturbance tradeoff and the continuity of Stinespring's representation, quant- ph/0605009, May 2006. [13] D. Kretschmann, D. Schlingemann, R. F. Werner, A continuity theo- rem for Stinespring's dilation, J. Funct. Anal. 255 (2008), no. 8, 1889-1904 [14] E. C. Lance, Hilbert C∗-modules, Cambridge Univ. Press, Cambridge, 1995. [15] G. J. Murphy, C∗-algebras and operator theory, Academic Press, Boston, MA, 1990. [16] G. J. Murphy, Positive definite kernels and Hilbert C∗-modules, Proc. Edin- burgh Math. Soc. (2) 40 (1997), no. 2, 367-374. [17] W. L. Paschke, Inner product modules over B∗-algebras, Trans. Amer. Math. Soc. 182 (1973), 443-468. [18] V. Paulsen, Completely bounded maps and operator algebras, Cambridge Univ. Press, Cambridge, 2002. [19] M. A. Rieffel, Induced representations of C∗-algebras, Advances in Math. 13 (1974), 176-257. 18 [20] S. Sakai, C∗-Algebras and W ∗-Algebras, Springer, Berlin, 1971. [21] M. Skeide, Generalised matrix C∗-algebras and representations of Hilbert mod- ules, Math. Proc. R. Ir. Acad. 100A (2000), no. 1, 11-38. [22] M. Skeide, Von Neumann modules, intertwiners and self-duality, J. Operator Theory 54 (2005), no. 1, 119-124. [23] W. F. Stinespring, Positive functions on C∗-algebras, Proc. Amer. Math. Soc. 6 (1955), 211-216. [24] M. Takesaki, Theory of operator algebras. III, Springer, Berlin, 2003. Statistics and Mathematics Unit, Indian Statistical Institute, R. V. College Post, Bangalore 560059, India. E-mail: [email protected], [email protected] 19
1011.3160	1	1011	2010-11-13T20:45:14	A note on the von Neumann algebra of a Baumslag-Solitar group	[ "math.OA", "math.GR" ]	We study qualitative properties of the group von Neumann algebra of a Baumslag-Solitar group. Namely, we prove that, in the non-amenable and {ICC} case, the associated ${\rm II}_1$ factor is prime, not solid, and does not have any Cartan subalgebra.	math.OA	math	A NOTE ON THE VON NEUMANN ALGEBRA OF A BAUMSLAG-SOLITAR GROUP Pierre Fima(1,2) Abstract We study qualitative properties of the group von Neumann algebra of a Baumslag-Solitar group. Namely, we prove that, in the non-amenable and ICC case, the associated II1 factor is prime, not solid, and does not have any Cartan subalgebra. 1 Introduction In their breakthrough paper [OP07] Ozawa and Popa obtained new structural results for group (and group measure space) von Neumann algebras. for any diffuse amenable von Neumann subalgebra Q ⊂ L(Fn), the normalizer of Q generates an amenable von Neumann subalgebra. This strengthened two well-known results: Voiculescu's result [Vo96] showing that L(Fn) has no Cartan subalgebra and Ozawa's result [Oz04] showing that L(Fn) is solid (the relative commutant of any diffuse von Neumann subalgebra is amenable) which itself strengthened Ge's result [Ge98] showing that L(Fn) is prime (it is not decomposable into tensor product of II1 factors). In particular, they showed that L(Fn) is strongly solid, i.e. In this note we study these properties for Baumslag-Solitar groups factors. Let n, m ∈ Z − {0}. The Baumslag-Solitar group is defined by BS(n, m) := ha, b abna−1 = bmi. It was proved in [Mo91] that BS(n, m) ≃ BS(p, q) if and only if {n, m} = {ǫp, ǫq} for some ǫ ∈ {−1, 1}. Moreover, Γ is known to be non-amenable but inner-amenable and ICC whenever n, m ≥ 2 and n 6= m (see [St06]). Gal and Januszkiewicz [GJ03] proved that BS(n, m) has the Haagerup property. Note that their proof also implies that BS(n, m) has the complete approximation property (CMAP). Actually one just has to check that the automorphism group of a locally finite tree has the CMAP as a locally compact group (for the compact-open topology). Also, the first L2 Betti number of BS(n, m) is zero. Our results can be summarized as follows. Theorem 1.1. Let Γ = BS(n, m). Assume n, m ≥ 2 and n 6= m. The following holds. 1. L(Γ) is a prime II1 factor. 2. L(Γ) is not solid and does not have any Cartan subalgebra. We prove actually a general primeness result for groups acting on trees (see Corollary 3.3). We also prove a stronger property than the absence of Cartan subalgebra (see Theorem 4.2). Namely, we prove that L(Γ) is robust i.e, the relative commutant of any regular and amenable von Neumann subalgebra is non-amenable. 2 Preliminaries 2.1 Weakly compact actions Weakly compact actions were introduced by Ozawa-Popa [OP07]. The following theorem is similar to Theorem 4.9 in [OP07]. The main ingredients of the proof are contained in the proofs of Theorem 4.9 in [OP07] and Theorem B in [0P08] as explained in the proof of Theorem 3.5 in [HS09] (see also Theorem 3.3 in [Ho10]). This result is not stated explicitly in any of these papers but the proof is the same as the one of Theorem 3.5 in [HS09]. 1Partially supported by ERC Starting Grant VNALG-200749. 2Institut Math´ematiques de Jussieu, Universit´e Paris Diderot, 175 rue de Chevaleret, 75013 Paris. E-mail: [email protected] 1 Theorem 2.1. Let P be a tracial von Nemann algebra that admits the following deformation property: there exists a tracial von Neumann algebra eP , a trace preserving inclusion P ⊂ eP and a one-parameter group (αs)s∈R of trace-preserving automorphisms of eP such that • lims→0αs(x) − x2 = 0 for all x ∈ P . • P L2( eP ) ⊖ L2(P )P is weakly contained in P L2(P ) ⊗ L2(P )P . • There exists c > 0 such that α2s(x)−x2 ≤ cαs(x)−EP ◦αs(x)2, for all x ∈ P, s ∈ R (transversality). Let Q ⊂ P be a von Neumann subalgebra and G ⊂ NP (Q) be a subgroup such that the action G y Q is weakly compact. If, for all non zero projection z in Z(G′ ∩ P ), αs does not converge uniformly on (zQ)1, then G′′ is amenable. 2.2 HNN extensions of von Neumann algebras HNN extensions of general von Neumann algebras were introduced in [Ue04]. In this note we follow the approach of [FV10]. Let (M, τM ) be a tracial von Neumann algebra and N ⊂ M a von Neumann subalgebra. Let θ : N → M be trace-preserving embedding. Let P = HNN(M, N, θ) be the HNN extension. We recall that a tracial von Neumann algebra eP with a trace preserving inclusion P ⊂ eP and one parameter group of automorphisms (αs) of eP satisfying the first condition of Theorem 2.1 were constructed in [FV10]. It was also observed that eP = P ∗N (N ⊗ LZ). This implies that, when N is amenable, P L2( eP ) ⊖ L2(P )P is weakly contained in the coarse P -P -bimodule. A detailed argument can be found e.g. in [CH08, Proposition 3.1]. Also, an automorphism β ∈ Aut(P ) such that β ◦ αs = α−s ◦ β and β(x) = x for all x ∈ P was introduced in [FV10]. Such a deformation is s-malleable. As such, it automatically satisfies the following transversality property (see Lemma 2.1 in [Po08]): α2s(x) − x2 ≤ 2αs(x) − EM ◦ αs(x)2 for all x ∈ M, s ∈ R. Hence, if N is amenable, Theorem 2.1 applies to the HNN extension P = HNN(M, N, θ). 3 Primeness results for groups acting on trees We will use the following proposition. The proof is similar to the one of Theorem 5.2 in [CH08] (even easier because we state it in the finite case). Proposition 3.1. Let M1 and M2 be finite von Neumann algebras with a common von Neumann subalgebra B of type I. Let M = M1 ∗B M2. Let p ∈ M be a non zero projection. If pM p is a non amenable II1 factor then pM p is prime. The following result is a direct corollary of the preceding Proposition and Remark 4.6 in [FV10]. Corollary 3.2. Let Γ = HNN(H, Σ, θ) be a non-trivial HNN extension (i.e. Σ, θ(Σ) 6= H). Assume that the following conditions are satisfied. 1. Γ is non-amenable and ICC. 2. Σ is abelian or finite. Then, L(Γ) is a prime II1 factor. We now extend this result to groups acting on trees. Corollary 3.3. Let Γ by a group satisfying the following properties. 2 • Γ is non-amenable and ICC. • Γ admits an action Γ y T without inversion on a tree T such that there exists a finite subtree with a finite stabilizer and such that there exists an edge e ∈ E(T ) with the properties that Stab e is abelian or finite and that the smallest subtrees containing all vertices Γ · s(e), resp. Γ · r(e), are both equal to the whole of T . Then L(Γ) is a prime II1 factor. Proof. In Corollary 3.2 we have seen that the conclusion of the corollary holds for certain HNN extensions. In [CH08, Theorem 5.2] it was shown that the conclusion also holds for certain amalgamated free product groups. So, it suffices to prove that all groups satisfying the assumptions of the corollary fall into one of both families and this can be done as in the proof of Theorem 1.2 in [FV10]. 4 Robustness for certain HNN extensions We call a von Neumann algebra robust if the relative commutant of any regular and amenable von Neumann subalgebra is non-amenable. Clearly, robustness implies the absence of Cartan subalgebra. Let (P, τ ) be a tracial von Neumann algebra and A, B ⊂ M be possibly non-unital von Neumann subalgebras. Following [Po03, Section 2] we say that A embeds into B inside M , and we denote it by A ≺M B, if there exist n ∈ N, a non-zero projection q ∈ Bn = Mn(C) ⊗ B, a normal unital ∗-homomorphism ϕ : A → qBnq and a non-zero partial isometry v ∈ 1A(M1,n(C) ⊗ M ) satisfying av = vϕ(a) for all a ∈ A. Otherwise, we write A 6≺M B. Lemma 4.1. Let P = HNN(M, N, θ) be an HNN extension of finite von Neumann algebras and suppose that N is amenable and P has the CMAP. Let Q ⊂ P a unital von Neumann subalgebra. If Q is amenable and Q ⊀ P M then NP (Q)′′ is amenable. Proof. Let z ∈ NP (Q)′ ∩ P be a non-zero projection. Observe that z ∈ Q′ ∩ P . As Q ⊀ P M we get zQ ⊀ P M . By [FV10, Theorem 3.4] we get that the deformation (αs) does not converge uniformly on the unit ball of zQ, for all non-zero projection z ∈ NP (Q)′ ∩ P . We can apply [OP07, Theorem 3.5] and Theorem 2.1 to conclude that NP (Q)′′ is amenable. Theorem 4.2. Let Γ = HNN(H, Σ, θ). Suppose that the following conditions are satisfied. 1. H is abelian, 2 ≤ H/Σ < ∞ and 3 ≤ H/θ(Σ), 2. Γ has the CMAP. Then L(Γ) is robust. If moreover Σ is infinite, then L(Γ) is not solid. Proof. Let Γ = hH, t θ(σ) = tσt−1 ∀σ ∈ Σi. Define G = hH, t−1Hti ⊂ Γ and Σ′ = {g ∈ Γ gσ = σg for all σ ∈ Σ}. As H is abelian, we have H ⊂ Σ′. Moreover, for all σ ∈ Σ and h ∈ H, we have t−1htσ = t−1hθ(σ)t = t−1θ(σ)ht = σt−1ht. It follows that t−1Ht ⊂ Σ′. We conclude that G ⊂ Σ′. Let eH be a copy of H and view Σ as a subgroup of eH via the map θ. Define the following group homomor- phisms: the first one from H to G is the identity, the second one from eH to G maps h onto t−1ht. These groups homomorphisms agree on Σ (because we see Σ < eH via the map θ). We get a group homomorphism eH from H ∗ Σ eH is not eH to G which is clearly surjective. It is also injective because it maps each reduced word in H ∗ Σ onto a reduced word (in the HNN extension sense) in G. As H/Σ ≥ 2 and H/θ(Σ) ≥ 3, H ∗ Σ amenable. 3 Let Q ⊂ L(Γ) be an amenable regular subalgebra. By Lemma 4.1, Q ≺ L(Γ) L(H). Because Σ has finite index in H we obtain Q ≺ L(Γ) L(Σ). It follows that L(Σ)′ ∩ L(Γ) ≺ L(Γ) Q′ ∩ M . In particular, L(G) ≺ L(Γ) Q′ ∩ M . Because L(G) has no amenable direct summand, Q′ ∩ M is not amenable. If L(Σ) is infinite, L(Γ) is obviously not solid. Actually, L(Σ) is a diffuse amenable von Neumann subalgebra and its relative commutant is not amenable as it contains L(G). We obtain the following obvious corollary. Corollary 4.3. Let Γ = BS(m, n). If n, m ≥ 2 and n 6= m then L(Γ) is a non-solid II1 factor and does not have any Cartan subalgebra. Acknowledgements Many thanks are due to Stefaan Vaes for several helpful conversations. References [CH08] I. Chifan and C. Houdayer, Bass-Serre rigidity results in von Neumann algebras. To appear in Duke Math J. arXiv:0805.1566. [FV10] P. Fima and S. Vaes, HNN extensions and unique group measure space decomposition of II1 factors. To appear in Transactions of the American Mathematical Society. arXiv:1005.5002 [GJ03] S.R. Gal and T. Januszkiewicz, New a-T-menable HNN-extension. J. Lie Theory. 13(2) (2003), 383- 385. [Ge98] L. Ge, Applications of free entropy to finite von Neumann algebras, II. Ann. of Math. 147 (1998), 143-157. [Ho10] C. Houdayer, Strongly solid group factors which are not interpolated free group factors. Math. Ann. 346 (2010), 969-989. [HS09] C. Houdayer and D. Shlyakhtenko, Strongly solid II1 factors with an exotic MASA. arXiv:0904.1225. [Mo91] D.I. Moldavanskii, On the isomorphisms of Baumslag-Solitar groups. (Russian) Ukrain. Mat. Zh. 43 (1991), no. 12, 1684-1686. [Oz04] N. Ozawa, Solid von Neumann algebras. Acta Math. 192 (2004), 111-117. [OP07] N. Ozawa, S. Popa, On a class of II1 factors with at most one Cartan subalgebra. To appear in Annals of Mathematics arXiv: 0706.3623. [0P08] N. Ozawa, S. Popa, On a class of II1 factors with at most one Cartan subalgebra, II. Amer. Journal of Math. 132 (2010), no. 3, 841-866. [Po03] S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, I. Invent. Math. 165 (2006), 369-408. [Po08] S. Popa, On the superrigidity of malleable actions with spectral gap. J. Amer. Math. Soc. 21 (2008), 981-1000. [St06] Y. Stalder, Moyennabilit´e int´erieure et extensions HNN. Ann. Inst. Fourier. 56 (2006), no. 2, 309-323. [Ue04] Y. Ueda, HNN extensions of von Neumann algebras. J. Funct. Anal. 225 (2005), 383-426. [Vo96] D.-V. Voiculescu, The analogues of entropy and of Fishers information measure in free probability theory, III. GAFA, Geom. funct. anal. 6 (1996), 172-199. 4
1809.05772	1	1809	2018-09-15T21:37:03	Geometry of $C^*$-algebras, the bidual of their projective tensor product, and completely bounded module maps	[ "math.OA", "math.FA" ]	Let $\mathcal{A}$ be a $C^$-algebra, and consider the Banach algebra $\mathcal{A} \otimes_\gamma \mathcal{A}$, where $\otimes_\gamma$ denotes the projective Banach space tensor product; if $\mathcal{A}$ is commutative, this is the Varopoulos algebra $V_\mathcal{A}$. It has been an open problem for more than 35 years to determine precisely when $\mathcal{A} \otimes_\gamma \mathcal{A}$ is Arens regular. Even the situation for commutative $\mathcal{A}$, in particular the case $\mathcal{A} = \ell_\infty$, has remained unsolved. We solve this classical question for arbitrary $C^$-algebras by using von Neumann algebra and operator space methods, mainly relying on versions of the (commutative and non-commutative) Grothendieck Theorem, and the structure of completely bounded module maps. Establishing these links allows us to show that $\mathcal{A} \otimes_\gamma \mathcal{A}$ is Arens regular if and only if $\mathcal{A}$ has the Phillips property; equivalently, $\mathcal{A}$ is scattered and has the Dunford--Pettis Property. A further equivalent condition is that $\mathcal{A}^$ has the Schur property, or, again equivalently, the enveloping von Neumann algebra $\mathcal{A}^{}$ is finite atomic, i.e., a direct sum of matrix algebras. Hence, Arens regularity of $\mathcal{A} \otimes_\gamma \mathcal{A}$ is encoded in the geometry of the $C^$-algebra $\mathcal{A}$. In case $\mathcal{A}$ is a von Neumann algebra, we conclude that $\mathcal{A} \otimes_\gamma \mathcal{A}$ is Arens regular (if and) only if $\mathcal{A}$ is finite-dimensional. For commutative $C^$-algebras $\mathcal{A}$, we determine precisely the centre of the bidual, namely, $Z({V_\mathcal{A}}^{})$ is Banach algebra isomorphic to $\mathcal{A}^{} \otimes_{eh} \mathcal{A}^{*}$, where $\otimes_{eh}$ denotes the extended Haagerup tensor product.	math.OA	math	Geometry of C ∗-algebras, the bidual of their projective tensor product, and completely bounded module maps Matthias Neufang Abstract Let A be a C ∗-algebra, and consider the Banach algebra A⊗γ A, where ⊗γ denotes the projective Banach space tensor product; if A is commutative, this is the Varopoulos algebra VA. It has been an open problem for more than 35 years to determine precisely when A ⊗γ A is Arens regular; cf. [26], [40], [41]. Even the situation for commutative A, in particular the case A = ℓ∞, has remained unsolved. We solve this classical question for arbitrary C ∗-algebras by using von Neumann algebra and operator space methods, mainly relying on versions of the (commutative and non-commutative) Grothendieck Theorem, and the structure of completely bounded module maps. Establishing these links allows us to show that A ⊗γ A is Arens regular if and only if A has the Phillips property; equivalently, A is scattered and has the Dunford -- Pettis Property. A further equivalent condition is that A∗ has the Schur property, or, again equivalently, the enveloping von Neumann algebra A∗∗ is finite atomic, i.e., a direct sum of matrix algebras. Hence, Arens regularity of A ⊗γ A is encoded in the geometry of the C ∗-algebra A. In case A is a von Neumann algebra, we conclude that A ⊗γ A is Arens regular (if and) only if A is finite-dimensional. We also show that this does not generalize to the class of non-selfadjoint dual (even commutative) operator algebras. Specializing to commutative C ∗-algebras A, we obtain that VA is Arens regular if and only if A is scattered. In fact, we determine precisely the centre ∗∗) is Banach algebra isomorphic to A∗∗ ⊗eh A∗∗, where ⊗eh denotes of the bidual, namely, Z(VA the extended Haagerup tensor product. We deduce that VA is strongly Arens irregular (if and) only if A is finite-dimensional. Hence, VA is neither Arens regular nor strongly Arens irregular, if and only if A is non-scattered (as mentioned above, this is new even for the case A = ℓ∞). 1 Introduction Let A be a C ∗-algebra. We shall be concerned with the Banach algebra A ⊗γ A, where we write ⊗γ for the projective Banach space tensor product. In case A is commutative, this is the so-called Varopoulos algebra VA. As is well-known, any C ∗-algebra A is Arens regular, i.e., the left and right Arens product on the bidual A∗∗ coincide; A∗∗ with this product is the enveloping von Neumann al- gebra of A. The problem of characterizing those C ∗-algebras A such that A⊗γA is Arens regular, has to our knowledge first been systematically studied in the Ph.D. thesis [26] of M. Ljeskovac, written in 1981 under the supervision of A.M. Sinclair (Edinburgh); apparently independent from this work are the articles [40] and [41] by A. Ulger, and [25] by A.T.-M. Lau and A. Ulger. Results regarding Arens regularity of A⊗γA have been obtained by these authors in several cases. However, even the case of general commutative C ∗-algebras has thus far not been settled: it is known that c0⊗γc0 is Arens regular while C(G)⊗γ C(G) is not, where G is an infinite compact group; but the case of ℓ∞⊗γℓ∞, for instance, has been open. In summary, no characterization of commutative C ∗-algebras 1 with Arens regular projective tensor square is known, let alone the situation for non-commutative C ∗-algebras. In this paper, we shall use von Neumann algebra and operator space techniques to solve this question for arbitrary C ∗-algebras, and present further results on the structure of the bidual. Our approach may be described in essence as follows. Our starting point is to pass, based on (commutative and non-commutative) versions of the Grothendieck Theorem, from the projective Banach space tensor product to the Haagerup tensor product, and then to use the latter and the extended Haagerup tensor product, as well as their intimate link to completely bounded module maps on B(H), where H is a Hilbert space: in this fashion, the left Arens product translates into composition of operators on B(H), and results on automatic normality of these maps yield structural results on the topological centre of the bidual algebra in question. By establishing these connections, we are then able to completely characterize Arens regularity of the projective tensor square of an arbitrary C ∗-algebra in terms of intrinsic properties of the latter. As we shall prove, Arens regularity of A⊗γA is encoded in the geometry of the C ∗-algebra A. In section 1 we provide some preliminary results pertaining to C ∗- and operator space theory In section 3, for commutative C ∗-algebras A, we determine which will be needed in the sequel. precisely the centre of the bidual of the Varopoulos algebra VA, showing that it is Banach algebra isomorphic to A∗∗ ⊗eh A∗∗, where ⊗eh denotes the extended Haagerup tensor product. It follows that VA is strongly Arens irregular (if and) only if A is finite-dimensional. Section 4 gives our main result, stating that A ⊗γ A is Arens regular if and only if A has the Phillips property; equivalently, A is scattered and has the Dunford -- Pettis Property. Further equivalent statements are that A∗ has the Schur property, or, again equivalently, the enveloping von Neumann algebra A∗∗ is finite atomic, i.e., a direct sum of matrix algebras. In fact, we prove that if A has the Phillips property, then A ⊗γ B is Arens regular for any C ∗-algebra B. Moreover, we deduce that for a von Neumann algebra A, Arens regularity of A⊗γA is equivalent to A being finite- dimensional. We also show that this characterization does not extend, in general, to non-selfadjoint dual operator algebras, even commutative ones: indeed, given any p ∈ [1, ∞), the algebra ℓp (Oℓp) with pointwise product is (completely) isomorphic and w∗-homeomorphic to a dual operator algebra with Arens regular projective tensor square; here, Oℓp denotes Pisier's operator space structure on ℓp, obtained by complex interpolation. Considering commutative C ∗-algebras A, we obtain that VA is Arens regular if and only if A is scattered. Combined with our earlier result, this shows that the Varopoulos algebra VA is neither ∗∗) lies strictly between VA and its Arens regular nor strongly Arens irregular (i.e., the centre Z(VA bidual), if and only if A is non-scattered. This result is already new for the fundamental example ℓ∞. Indeed, already the question whether ℓ∞⊗γ ℓ∞ is Arens regular has been open since 1981. 2 Preliminaries We start by recalling the definitions of geometric properties for a C ∗-algebra A and a Banach space E which we shall encounter. • A is scattered if any positive functional on A is a finite or countable (pointwise) sum of pure functionals (cf. [19]). • E has the Dunford -- Pettis property (DPP) if any weakly compact operator from E into any 2 Banach space is completely continuous; equivalently, for any weakly null sequences xn in E and fn in E∗, the sequence hfn, xni tends to 0 (going back to Grothendieck's classical work [13]). • E has the Phillips property if the Dixmier projection E∗∗∗ → E∗ is sequentially w∗ -- norm continuous (cf. [10]). • E has the Schur property if in E any weakly convergent sequence is norm convergent. We now review some related structure results for C ∗-algebras; see [11, Corollary VII-10], [2, Theorem 3] and [19, Theorem 2.2] for the first, and [10, Lemma 3.1] together with [25, Theorems 3.4 and 3.6] for the second. Theorem 2.1. Let A be a C ∗-algebra. Then the following are equivalent: (i) A is scattered; (ii) A does not contain an isomorphic copy of ℓ1; (iii) A∗ has the Radon -- Nikodym Property (RNP); (iv) A∗∗ is atomic, i.e., a direct sum of B(Hi) for Hilbert spaces Hi. Theorem 2.2. Let A be a C ∗-algebra. Then the following are equivalent: (i) A has the Phillips property; (ii) A is scattered and has the DPP; (iii) A∗ has the Schur property; (iv) A∗∗ is finite atomic, i.e., a direct sum of matrix algebras. We now fix some terminology. By "w∗-continuous" we mean "w∗-w∗-continuous". We write ✷ for the left Arens product in the bidual of a Banach algebra A. Recall that, for X, Y ∈ A∗∗, we have where hX✷Y, f i := hX, Y ✷f i (f ∈ A∗) hY ✷f, ai := hY, f ✷ai (a ∈ A); of course, hf ✷a, bi := hf, abi for all b ∈ A. Note that, if xi and yj are nets in A converging w∗ to X and Y , respectively, we have X✷Y = w∗- lim i w∗- lim j xiyj. Obviously, the map A∗∗ ∋ X 7→ X✷Y is w∗-continuous. The set of all X ∈ A∗∗ such that the map A∗∗ ∋ Y 7→ X✷Y is w∗-continuous, is called the (left) topological centre, denoted by Zt(A∗∗). Now, A is said to be Arens regular if Zt(A∗∗) is maximal, i.e., equals A∗∗ (this is equivalent to saying that the left and right Arens products on A∗∗ coincide, but we shall only use the left one); also, A is said to be (left) strongly Arens irregular if Zt(A∗∗) is minimal, i.e., equals A. Topological centres have been studied intensely in abstract harmonic analysis in recent years; see, e.g., the articles [3], [4], 3 [8], [16], [24], [28] and [29]. Let us mention that, for any locally compact group G, the group algebra L1(G) and the measure algebra M (G) are strongly Arens irregular; these are the main theorems of [23] and [29], respectively, and we shall make use of the former result in section 4. We write Z(A) for the centre of an algebra A. Note that if A is a commutative Banach algebra, Zt(A∗∗) equals the centre Z(A∗∗). Even in this situation, determining Zt(A∗∗) can be very diffi- cult. For instance, denoting by A(G) the Fourier algebra of a locally compact group G, the centre Z(A(G)∗∗) is not known in general, even when G is discrete or compact. Indeed, it is open if A(G) can be Arens regular for discrete groups without infinite amenable subgroups nor copies of free groups, such as Olshanskii's 'Tarski monsters'; cf. [9], [28]. In the case of infinite compact groups, it is known that A(G) is not Arens regular; cf. [9]. As the projective tensor product A := C(G) ⊗γ C(G) contains a subalgebra isomorphic to A(G), by Varopoulos's classical work, one thus obtains that A is not Arens regular, as mentioned in section 1; cf. [41]. However, determining when A(G) is strongly Arens irregular, is more complicated: for instance, A(SU (3)) is not, but A(SU (3)ℵ0 ) and A(SU (3)ℵ1 ) are; cf. [27], [24], [8]. Given a set S ∈ B(H), where H is a Hilbert space, we write S ′ for its commutant. Let A ⊆ B(H) and B ⊆ B(K) be C ∗-algebras. We denote by Aop the opposite algebra. We denote by CBA,B(B(K, H)) the algebra of completely bounded operators on B(K, H) that are left A- and right B-module maps, i.e., Φ(aT b) = aΦ(T )b for all T ∈ B(K, H) and a ∈ A, b ∈ B. If H = K and A = B, we simply write CBA(B(H)). We denote by CBσ A,B(B(K, H)) the subalgebra of normal (i.e., w∗-continuous) such maps. Of particular importance in our approach is the work of Magajna [30], which contains a further development of the results obtained in [15] by Hofmeier -- Wittstock. We write ⊗γ for the projective Banach space tensor product, ⊗h for the Haagerup tensor product, ⊗eh for the extended Haagerup tensor product, and ⊗σh for the normal Haagerup tensor product. For the convenience of the reader, we shall briefly recall the definitions and some fundamental relations. For Banach spaces E and F , the projective tensor product E⊗γ F is the completion of the algebraic tensor product E ⊗ F with respect to the norm kukγ = inf( mXk=1 kxkkkykk u = xk ⊗ yk) . mXk=1 Now let E and F be operator spaces, n ∈ N, and u = [uij] ∈ Mn(E ⊗ F ). The Haagerup norm of u is defined by kukh = inf{kxkkyk}, with the infimum taken over all p ∈ N, and all representations k=1 xik ⊗ ykj]. The Haagerup tensor product E ⊗h F is the completion of E ⊗ F with respect to the Haagerup matrix norms. For C ∗-algebras A and B, the Haagerup norm on A ⊗ B can be expressed as u = x ⊙ y, where x ∈ Mn,p(E) and y ∈ Mp,n(F ); here, x ⊙ y = [Pp kukh = inf (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) akak nXk=1 bk nXk=1 ∗(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∗bk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1 2 u = nXk=1 ak ⊗ bk . Note that on E ⊗ F we have k · kh ≤ k · kγ. If A and B are commutative C ∗-algebras, the classical Grothendieck Inequality says that where K is the Grothendieck constant; cf., e.g., [21]. k · kγ ≤ Kk · kh, 4 The extended Haagerup tensor product E ⊗eh F is defined as the subspace of (E∗ ⊗h F ∗)∗ corresponding to the completely bounded bilinear forms E∗ × F ∗ → C which are separately w∗- continuous. Note that the w∗-Haagerup tensor product of dual operator spaces is defined by E∗ ⊗w∗h F ∗ = (E ⊗h F )∗; so we have E∗ ⊗eh F ∗ = E∗ ⊗w∗h F ∗. Finally, the normal Haagerup tensor product of dual operator spaces is defined as E∗ ⊗σh F ∗ = (E ⊗eh F )∗. Hence, we have (E ⊗h F )∗∗ = (E∗ ⊗eh F ∗)∗ = E∗∗ ⊗σh F ∗∗. More information on these tensor products can be found, e.g., in [1], [5], [37], and [6]. As noted above, the classical Grothendieck Inequality yields a canonical isomorphism between A⊗γA and A ⊗h A for commutative C ∗-algebras A, which plays a significant role in our approach; as do the consequences of the non-commutative Grothendieck Inequality -- see the seminal work of Pisier -- Shlyakhtenko [38] and Haagerup -- Musat [14] -- drawn in [20] regarding Arens regularity of the algebras A⊗γA and A ⊗h A, where A is an arbitrary C ∗-algebra. In the following we collect the results from [20] which will be of importance for us. (Note that the projective operator space tensor product is of course crucial in the work [20], but we shall not explicitly use it.) Theorem 2.3. Let A and B be C ∗-algebras. (i) If A ⊗γ B is Arens regular, then A ⊗h B is as well. (ii) If A ⊗h B and A ⊗h Bop are Arens regular, then so is A ⊗γ B. Proof. (i) follows from [20, Theorems 2.6, 2.7], and (ii) from [20, Theorem 2.9]. In the sequel, we shall use the following identification without explicit reference. Theorem 2.4. Let A and B be C ∗-algebras, and let R := A∗∗ ⊆ B(H), S := B∗∗ ⊆ B(K). Then via a completely isometric w∗-homeomorphism Θ such that (A ⊗h Bop)∗∗ = CBR′,S ′(B(K, H)) (Θ(a ⊗ b))(T ) = aT b for all a ∈ A, b ∈ B, and Θ(X✷Y ) = Θ(X)Θ(Y ) for all X, Y ∈ (A ⊗h B)∗∗. Moreover, by restricting Θ, one obtains the completely isometric multiplicative w∗-homeomorphic identification R ⊗eh S op = CBσ R′,S ′(B(K, H)). Proof. In view of the canonical identification (E ⊗h F )∗∗ = E∗∗ ⊗σh F ∗∗ for operator spaces E and F , combined with [30, Theorem 2.2], we only need to show that Θ is multiplicative. To see this, let X, Y ∈ (A ⊗h Bop)∗∗, and let xi, yj be bounded nets in A ⊗h Bop converging w∗ to X, respectively, Y . Note that Θ is multiplicative on A ⊗h Bop. We then have (with limits in the w∗-topology): Θ(X✷Y ) = lim i lim j Θ(xiyj) = lim i lim j Θ(xi)Θ(yj) = lim i Θ(xi)Θ(Y ) = Θ(X)Θ(Y ), 5 as claimed. For the second-last equation, note that every Θ(xi) ∈ CBσ(B(K, H)), and left multipli- cation in CB(B(K, H)) by such an element is w∗-continuous; for the last equation, note that right multiplication in CB(B(K, H)) by any element is w∗-continuous. Finally, note that the restriction of ✷ to R ⊗eh S op defines a product on this space since Θ is multiplicative, Θ(R ⊗eh S op) = CBσ R′,S ′(B(K, H)), and CBσ R′,S ′(B(K, H)) is an algebra. 3 The case of Varopoulos algebras Theorem 3.1. Let A be a commutative C ∗-algebra. If VA is Arens regular, then A is scattered. Proof. Since A is commutative, VA = A ⊗γ A is isomorphic to A ⊗h A by the classical Grothendieck Inequality, so the latter is Arens regular. Represent the von Neumann algebra R := A∗∗ in B(H) such that it is maximal commutative; cf. [39, Chapter 5, Proposition 6]. As A ⊗h A is Arens regular and commutative, (A ⊗h A)∗∗ = CBR′(B(H)) = CBR(B(H)) is commutative. Hence, we have CBR(B(H)) = CBR(B(H))c ∩ CBR(B(H)), where the commutant is taken in CB(B(H)). Since R is commutative, it has no direct summand of type IJ,n with J an infinite cardinal and n ∈ N. So, we have CBR(B(H))c = CBσ R′(B(H)), by [30, Theorem 2.4]. Thus, as R = R′, we obtain CBR(B(H)) = CBσ R(B(H)). By [30, Theorem 3.5], this implies that R is atomic. Owing to Theorem 2.1, this is equivalent to A being scattered. ∗∗ is isomorphic Theorem 3.2. Let A be a commutative C ∗-algebra. The (topological) centre of VA to A∗∗ ⊗eh A∗∗ (via ϕ∗∗ where ϕ : A ⊗h A → A⊗γA is the canonical isomorphism, based on the classical Grothendieck Inequality). Proof. As above, represent R := A∗∗ in B(H) such that it is maximal commutative. By [30, Theorem 2.4], we have CBR(B(H))c = CBσ R′(B(H)) = CBσ R(B(H)), where the commutant is taken in CB(B(H)). So we see that Z(CBR(B(H))) = CBR(B(H))c ∩ CBR(B(H)) = CBσ R(B(H)). Hence we obtain: Θ(R ⊗eh R) = CBσ R(B(H)) = Z(CBR(B(H))) = Z(Θ((A ⊗h A)∗∗)) = Θ(Z((A ⊗h A)∗∗)). So, Z((A ⊗h A)∗∗) = A∗∗ ⊗eh A∗∗. As VA is isomorphic to A ⊗h A (via ϕ), the assertion follows. Corollary 3.3. Let A be a commutative C ∗-algebra. The Varopoulos algebra VA is strongly Arens irregular (if and) only if A is finite-dimensional. 6 Proof. Assume that VA is strongly Arens irregular. Then we have by Theorem 3.2: ϕ∗∗(A∗∗ ⊗eh A∗∗) = Z(VA ∗∗) = VA = ϕ∗∗(A ⊗h A). Since A ⊗h A ⊆ A∗∗ ⊗h A∗∗ ⊆ A∗∗ ⊗eh A∗∗, we deduce that A∗∗ ⊗h A∗∗ = A∗∗ ⊗eh A∗∗ (as Banach spaces). By [18, Corollary 3.8], this implies that A is finite-dimensional. The result [18, Corollary 3.8] to which we have referred above, implies that for a C ∗-algebra A, the equality A ⊗h A = A ⊗eh A (at the Banach space level) holds if and only if A is finite- dimensional. We shall show in the following, in passing, that this equivalence fails for A = ℓ1 with pointwise product. Note that the latter, endowed with any operator space structure, is completely isomorphic to an operator algebra, as shown by Blecher -- Le Merdy; cf. [1, Proposition 5.3.3]. For our proof, recall that a Banach space E is said to have Pelczynski's property (V) if any set K ⊆ E∗ satisfying sup ψ∈K hψ, xni −−→ n 0 for every weakly unconditionally Cauchy (wuC) series P xn in E, is relatively weakly compact. Here,P xn being wuC means thatP hψ, xni < ∞ for all ψ ∈ E∗. Theorem 3.4. We have (the Banach space equality) ℓ1 ⊗h ℓ1 = ℓ1 ⊗eh ℓ1. Proof. Note that, as shown by Pelczynski [35], c0 has property (V); cf. also [36, p. 351]. Since c∗ 0 = ℓ1 has the Schur property, [7, Corollary 5] implies that c0 ⊗γ c0 has property (V). But by the classical Grothendieck Theorem, the Banach algebras c0 ⊗γ c0 and c0 ⊗h c0 are isomorphic. So, c0 ⊗h c0 has property (V). Hence, its dual ℓ1 ⊗eh ℓ1 is weakly sequentially complete, again by a well-known result of Pelczynski [35]; see also [36, p. 351]. Thus, ℓ1 ⊗h ℓ1, as a closed subspace of ℓ1 ⊗eh ℓ1, is also weakly sequentially complete, so it does not contain an isomorphic copy of c0. Now, by a result of Pisier [37, Exercise 5.1], we have as desired. ℓ1 ⊗h ℓ1 = (c0 ⊗h c0)∗ = ℓ1 ⊗eh ℓ1, Remark 3.5. The following is a quick geometric proof of the well-known result that c0 ⊗γ c0 is Arens regular. -- As noted above, since c0 has property (V) and its dual has the Schur property, c0 ⊗γ c0 has property (V). In view of [12, Corollary I.2], this implies Arens regularity. 4 The case of general C ∗-algebras Our goal is to prove the complete characterization given below of the Arens regularity of A⊗γA. Theorem 4.1. Let A be a C ∗-algebra. Then the following are equivalent: (i) A⊗γA is Arens regular; (ii) A has the Phillips property; equivalently, A is scattered and has the DPP. In fact, if (ii) holds, then A ⊗γ B is Arens regular for any C ∗-algebra B. 7 Corollary 4.2. Let A be a von Neumann algebra. Then the following are equivalent: (i) A⊗γA is Arens regular; (ii) A is finite-dimensional. Proof. This follows from our Theorem 4.1 combined with [10, Corollary 2.14], which states that a dual Banach space with the Phillips property is finite-dimensional. Remark 4.3. As a fundamental example, we obtain that ℓ∞ ⊗γ ℓ∞ is Arens irregular. This solves a problem which has been open since the initial work of [26] in 1981. We will now show that Corollary 4.2 does not extend to the setting of non-selfadjoint dual operator algebras, even commutative ones. Recall that a dual operator algebra is, by definition, a w∗-closed subalgebra of B(H), for some Hilbert space H. For p ∈ [1, ∞), we shall consider ℓp with pointwise multiplication. We note that, given any Banach A, the algebra ℓp⊗γA is Arens regular if and only if A is; cf. [40, Corollaries 4.7 and 4.12]. We denote by Oℓp Pisier's operator space structure on ℓp, obtained by complex interpolation. Proposition 4.4. Let p ∈ [1, ∞). Then the algebra ℓp (Oℓp) with pointwise product is (completely) isomorphic and w∗-homeomorphic to a commutative dual operator algebra with Arens regular pro- jective tensor square. Proof. First, note that Oℓp is completely isomorphic to an operator algebra, as shown by Blecher -- Le Merdy; cf. [1, Corollary 5.3.5]. Since Oℓp is a dual operator space, and the product is separately w∗- continuous, there exists a commutative dual operator algebra Ap and a Banach algebra isomorphism and w∗-homeomorphism Φ : Oℓp → Ap such that Φ and Φ−1 are completely bounded, by a result of Le Merdy's; cf. [1, Theorem 5.2.16]. Let us now prove (working in the Banach algebra category) that Ap⊗γAp is Arens regular. As mentioned above, the Arens regularity of ℓp passes to ℓp⊗γℓp. Since Φ ⊗ Φ : ℓp⊗γℓp → Ap⊗γAp is an algebra homomorphism, it suffices to show that it is surjective. As ℓp has the approximation property, we have the isometric identification ℓp⊗γℓp = N ((ℓp)∗, ℓp), where the latter space denotes the nuclear operators (we write (ℓp)∗ for the predual to avoid distinguishing the cases p = 1 and p > 1). Now let z ∈ Ap⊗γAp. Then we have a series representation z =P xn⊗yn withP kxnkkynk < ∞. Hence, z0 :=P Φ−1(xn) ⊗ Φ−1(yn) ∈ N ((ℓp)∗, ℓp), and (Φ ⊗ Φ)(z0) = z, as desired. We now return to the C ∗-case. Noting that commutative C ∗-algebras have the DPP, our Theorem 4.1 immediately yields the following. Corollary 4.5. Let A be a commutative C ∗-algebra. Then the Varopoulos algebra VA is Arens regular if and only if A is scattered. The following consequence of Corollaries 3.3 and 4.5 provides a natural class of Banach algebras that are neither Arens regular nor strongly Arens irregular. Corollary 4.6. Let A be a commutative C ∗-algebra. Then VA is neither Arens regular nor strongly Arens irregular, if and only if A is non-scattered. 8 Example 4.7. The case A = ℓ∞ provides a natural example for the above situation; note that ∗∗ is ℓ∞ = C(βN) and βN is non-scattered. By our Theorem 3.2, the (topological) centre of Vℓ∞ isomorphic to ℓ∞ ∗∗ ⊗eh ℓ∞ ∗∗. In order to prove Theorem 4.1, we start with the generalization of our Theorem 3.1 to arbitrary C ∗-algebras. Theorem 4.8. Let A be a C ∗-algebra. If A ⊗γ A is Arens regular, then A is scattered. Proof. Since A ⊗γ A is Arens regular, so is A ⊗h A, according to Theorem 2.3 (i). Now, let A0 be a commutative C ∗-subalgebra of A. Then A0 ⊗h A0 is Arens regular as a subalgebra of A⊗h A. By our Theorem 3.1, this entails that A0 is scattered. Hence, we see that every commutative C ∗-subalgebra of A is scattered, which implies that A is scattered, by [22, Lemma 2.2] (in fact, the last implication is an equivalence). We now complete the proof of the implication (i) =⇒ (ii) in Theorem 4.1. Theorem 4.9. Let A be a C ∗-algebra. If A ⊗γ A is Arens regular, then A has the DPP. Proof. By Theorem 4.8, A is scattered, whence it remains to show, in view of Theorems 2.1 and 2.2, that the atomic von Neumann algebra R := A∗∗ is finite. Atomicity means that R is a direct sum of B(Hi), and we need to show that all Hilbert spaces Hi are finite-dimensional. Assume towards a contradiction that some Hi0 =: H is infinite-dimensional. Let K be a Hilbert space such that R ⊆ B(K). Then we have: (A ⊗h A)∗∗ = CBR′(B(K)) ⊇ CBB(H)′ (B(K)). By [31, Theorem 3.2 and Corollary 3.3], the latter is isomorphic to CBB(H)′(B(H)) = CB(B(H)). As A ⊗γ A is Arens regular, so is A ⊗h A, by Theorem 2.3 (i). Hence, the above entails that multiplication in CB(B(H)) is separately w∗-continuous. Realize H as ℓ2(G) where G is a discrete group of cardinality dim(H), and note that ℓ1(G)∗∗, endowed with the left Arens product, embeds isometrically and w∗-continuously as a subalgebra in CB(B(ℓ2(G))); cf. [32, Satz 2.1.1], [34, Remark 3.7], [17, Proposition 6.5]. Indeed, the map Θ : ℓ∞(G)∗ → CB(B(ℓ2(G))) given by hΘ(m)T ξ, ηi = mt(hLtT Lt−1 ξ, ηi), where m ∈ ℓ∞(G)∗, T ∈ B(ℓ2(G)), and ξ, η ∈ ℓ2(G), yields such an embedding (here, Lt denotes left translation by t ∈ G). This entails that the (left) topological centre Zt(ℓ1(G)∗∗) equals ℓ1(G)∗∗. But Zt(ℓ1(G)∗∗) = ℓ1(G), by [23, Theorem 1]; cf. also [33, Theorem 1.1]. Thus we obtain that G is finite, a contradiction. The following completes the proof of Theorem 4.1. Theorem 4.10. Let A be a scattered C ∗-algebra with the DPP. Then A ⊗γ B is Arens regular for any C ∗-algebra B. Proof. By Theorem 2.2, R := A∗∗ is finite atomic. Let H be such that R is represented in standard form in B(H). Then R is isomorphic to (R′)op, so it follows that R′ is also finite atomic. Put S := B∗∗. 9 We first show that A ⊗h B is Arens regular. To this end, choose K such that S op ⊆ B(K). Then we obtain: (A ⊗h B)∗∗ = CBR′,(S op)′(B(K, H)) = CBσ R′,(S op)′(B(K, H)), using in the second step [30, Lemma 3.4]. Similarly, A ⊗h Bop is Arens regular. Indeed, choosing K such that S ⊆ B(K), we have: (A ⊗h Bop)∗∗ = CBR′,S ′(B(K, H)) = CBσ R′,S ′(B(K, H)), using again [30, Lemma 3.4]. As A ⊗h B and A ⊗h Bop are Arens regular, it follows that A ⊗γ B is too, by Theorem 2.3 (ii). References [1] D. Blecher and C. LeMerdy, Operator algebras and their modules -- an operator space approach, London Mathematical Society Monographs, New Series, 30, Oxford Science Publications, The Clarendon Press, Oxford University Press, Oxford, 2004. [2] C.-H. Chu, A note on scattered C ∗-algebras and the Radon-Nikod´ym property, J. London Math. Soc. (2) 24 (1981), no. 3, 533 -- 536. [3] H.G. Dales and A.T.-M. Lau, The second duals of Beurling algebras, Mem. Amer. Math. Soc. 177 (2005), no. 836 [4] H.G. Dales, A.T.-M. Lau and D. Strauss, Banach algebras on semigroups and on their com- pactifications, Mem. Amer. Math. Soc. 205 (2010), no. 966 [5] E.G. Effros and Z.-J. Ruan, Operator spaces, Clarendon Press, Oxford, New York, 2000. [6] E.G. Effros and Z.-J. Ruan, Operator space tensor products and Hopf convolution algebras, J. Operator Theory 50 (2003), no. 1, 131 -- 156. [7] G. Emmanuele and W. Hensgen, Property (V) of Pelczynski in projective tensor products, Proc. Roy. Irish Acad. Sect. A 95 (1995), no. 2, 227 -- 231. [8] M. Filali, M. Neufang and M.S. Monfared, Weak factorizations of operators in the group von Neumann algebras of certain amenable groups and applications, Math. Ann. 351 (2011), no. 4, 935 -- 961. [9] B. Forrest, Arens regularity and discrete groups, Pacific J. Math. 151 (1991), no. 2, 217 -- 227. [10] W. Freedman and A. Ulger, The Phillips properties, Proc. Amer. Math. Soc. 128 (2000), no. 7, 2137 -- 2145. [11] N. Ghoussoub, G. Godefroy, B. Maurey and W. Schachermayer, Some topological and geometric structures in Banach spaces, Mem. Amer. Math. Soc. 70 (1987), no. 378. [12] G. Godefroy and B. Iochum, Arens-regularity of Banach algebras and the geometry of Banach spaces, J. Funct. Anal. 80 (1988), no. 1, 47 -- 59. 10 [13] A. Grothendieck, Sur les applications lin´eaires faiblement compactes d'espaces du type C(K), Canad. J. Math. 5 (1953), 129 -- 173. [14] U. Haagerup and M. Musat, The Effros -- Ruan conjecture for bilinear forms on C ∗-algebras, Invent. Math. 174 (2008), 139 -- 163. [15] H. Hofmeier and G. Wittstock, A bicommutant theorem for completely bounded module homo- morphisms, Math. Ann. 308 (1997), no. 1, 141 -- 154. [16] Z. Hu, M. Neufang and Z.-J. Ruan, Multipliers on a new class of Banach algebras, locally compact quantum groups, and topological centres, Proc. London Math. Soc. (3) 100 (2010), no. 2, 429 -- 458. [17] Z. Hu, M. Neufang and Z.-J. Ruan, Completely bounded multipliers over locally compact quan- tum groups, Proc. London Math. Soc. (3) 103 (2011), no. 1, 1 -- 39. [18] T. Itoh and M. Nagisa, On the extended Haagerup tensor product in operator spaces, J. Korean Math. Soc. 51 (2014), no. 2, 345 -- 362. [19] H.E. Jensen, Scattered C ∗-algebras, Math. Scand. 41 (1977), no. 2, 308 -- 314. [20] A. Kumar and V. Rajpal, Arens regularity of projective tensor products, Arch. Math. (Basel) 107 (2016), no. 5, 531 -- 541. [21] A. Kumar and A.M. Sinclair, Equivalence of norms on operator space tensor products of C ∗- algebras, Trans. Amer. Math. Soc. 350 (1998), no. 5, 2033 -- 2048. [22] M. Kusuda, C ∗-algebras in which every C ∗-subalgebra is AF, Q. J. Math. 63 (2012), no. 3, 675 -- 680. [23] A.T.-M. Lau and V. Losert, On the second conjugate algebra of L1(G) of a locally compact group, J. London Math. Soc. (2) 37 (1988), no. 3, 464 -- 470. [24] A.T.-M. Lau and V. Losert, The centre of the second conjugate algebra of the Fourier algebra for infinite products of groups, Math. Proc. Cambridge Philos. Soc. 138 (2005), no. 1, 27 -- 39. [25] A.T.-M. Lau and A. Ulger, Some geometric properties on the Fourier and Fourier -- Stieltjes algebras of locally compact groups, Arens regularity and related problems, Trans. Amer. Math. Soc. 337 (1993), no. 1, 321 -- 359. [26] M. Ljeskovac, The projective tensor product of commutative Banach algebras, Ph.D. thesis, University of Edinburgh, 1981. [27] V. Losert, On the centre of the bidual of Fourier algebras (the compact case), Talk at Interna- tional Conference on Abstract Harmonic Analysis, Ko¸c University, Istanbul, 2004. [28] V. Losert, The centre of the bidual of Fourier algebras (discrete groups), Bull. London Math. Soc. 48 (2016), no. 6, 968 -- 976. [29] V. Losert, M. Neufang, J. Pachl and J. Stepr¯ans, Proof of the Ghahramani -- Lau conjecture, Adv. Math. 290 (2016), 709 -- 738. 11 [30] B. Magajna, On completely bounded bimodule maps over W ∗-algebras, Studia Math. 154 (2003), no. 2, 137 -- 164. [31] G. May, E. Neuhardt and G. Wittstock, The space of completely bounded module homomor- phisms, Arch. Math. (Basel) 53 (1989), no. 3, 283 -- 287. [32] M. Neufang, Abstrake Harmonische Analyse und Modulhomomorphismen uber von Neumann- Algebren, Ph.D. thesis, University of Saarland, 2000. [33] M. Neufang, A unified approach to the topological centre problem for certain Banach algebras arising in abstract harmonic analysis, Arch. Math. (Basel) 82 (2004), no. 2, 164 -- 171. [34] M. Neufang, Z.-J. Ruan and N. Spronk, Completely isometric representations of McbA(G) and U CB( G)∗, Trans. Amer. Math. Soc. 360 (2008), no. 3, 1133 -- 1161. [35] A. Pelczynski, Banach spaces on which every unconditionally converging operator is weakly compact, Bull. Acad. Polon. Sci. S´er. Sci. Math. Astronom. Phys. 10 (1962), 641 -- 648. [36] H. Pfitzner, Weak compactness in the dual of a C ∗-algebra is determined commutatively, Math. Ann. 298 (1994), no. 2, 349 -- 371. [37] G. Pisier, Introduction to operator space theory, London Mathematical Society Lecture Note Series, 294, Cambridge University Press, Cambridge, 2003. [38] G. Pisier and D. Shlyakhtenko, Grothendieck's theorem for operator spaces, Invent. Math. 150 (2002), 185 -- 217. [39] D.M. Topping, Lectures on von Neumann algebras, Van Nostrand Reinhold Mathematical Stud- ies, 36, 1971. 623 -- 639. [40] A. Ulger, Arens regularity of the algebra Ab⊗B, Trans. Amer. Math. Soc. 305 (1988), no. 2, [41] A. Ulger, Erratum to: "Arens regularity of the algebra Ab⊗B" [Trans. Amer. Math. Soc. 305 (1988), no. 2, 623 -- 639], Trans. Amer. Math. Soc. 355 (2003), no. 9, 3839. Author's affiliations: School of Mathematics and Statistics, Carleton University, 1125 Colonel By Drive, Ottawa, Ontario K1S 5B6, Canada Email: [email protected] and Laboratoire de Math´ematiques Paul Painlev´e (UMR CNRS 8524), Universit´e Lille 1 -- Sciences et Technologies, UFR de Math´ematiques, 59655 Villeneuve d'Ascq Cedex, France Email: [email protected] 12
1007.0058	3	1007	2011-11-23T00:06:41	Infinite divisibility and a non-commutative Boolean-to-free Bercovici-Pata bijection	[ "math.OA", "math.PR" ]	We use the theory of fully matricial, or non-commutative, functions to investigate infinite divisibility and limit theorems in operator-valued non-commutative probability. Our main result is an operator-valued analogue of the Bercovici-Pata bijection. An important tool is Voiculescu's subordination property for operator-valued free convolution.	math.OA	math	INFINITE DIVISIBILITY AND A NON-COMMUTATIVE BOOLEAN-TO-FREE BERCOVICI-PATA BIJECTION S. T. BELINSCHI, M. POPA, AND V. VINNIKOV Abstract. We use the theory of fully matricial, or noncommutative, functions to investigate infinite divisibility and limit theorems in operator-valued non- commutative probability. Our main result is an operator-valued analogue for the Bercovici-Pata bijection. An important tool is Voiculescu's subordination property for operator-valued free convolution. 1. introduction The study of problems in operator algebras from a probabilistic perspec- tive recorded numerous successes in the recent decades. Notions of independence specific to noncommutative probability setting - especially Voiculescu's free in- dependence - were shown to be important for encoding structural properties of certain operator algebras. Numerous analogues of classical probabilistic notions and phenomena were found to hold for free, Boolean or conditionally free indepen- dences. In particular, conditional expectations in the context of noncommutative algebras are old and well-known; however, noncommutative independences over subalgebras with respect to conditional expectations are new and not yet very well studied and understood: it was Voiculescu [25] who generalized his own free in- dependence to freeness (with amalgamation) over a noncommutative subalgebra, while for the monotone independence of Muraki [12], Boolean independence of Spe- icher and Woroudi [21] and the conditionally free independence and convolution of Bozejko, Leinert and Speicher [8], the generalization was achieved by Popa [15, 16] and [17]. Let us briefly introduce some of the more important notions from noncom- mutative probability that we will study in this paper. We shall call a noncommuta- tive probability space a pair (A, ϕ), where A is a unital ∗-algebra over the complex numbers and ϕ : A → C is a positive functional normalized so that it carries the unit 1 ∈ A of A into the complex number 1. The algebra A plays the role of the algebra of complex-valued measurable functions on a classical probability space and ϕ plays the role of integration with respect to the probability measure. By following this analogy, an operator-valued noncommutative probability space is naturally de- fined as a triple (A, EB,B), where A is again a unital ∗-algebra, B is a ∗-subalgebra of A containing the unit of A and EB : A → B is a conditional expectation, i.e. a positive linear B−B bimodule map. (In some contexts, this definition will turn out to be too restrictive or too broad; thus, we will usually specify our requirements, assumptions and notations on a case-by-case basis.) When B = C, we deal with an Date: November 7, 2018. Research of STB was supported by a Discovery grant from NSERC Canada, and a University of Saskatchewan start-up grant. 1 2 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV ordinary noncommutative probability space. Elements X ∈ A are called random variables or (in the second context) operator-valued (or B-valued) random variables. If (A, ϕ) is a non-commutative probability space, the distribution of a self- adjoint element X of A (non-commutative random variable) is a real measure µX described via Z tndµX (t) = ϕ(X n). As shown in [25], for the operator-valued non-commutative random variables, the appropriate analogue of the distribution is an element of ΣB, the set of positive conditional expectations from the ∗-algebra of non-commutative polynomials with coefficients in some C∗-algebra B to B. In this setting, the op-valued analogues of compactly supported real measures are positive conditional expectations from ΣB with moments not growing faster than exponentially (see below). In probability theory limit theorems play a central role. Historically, among the first results proved for each new type of noncommutative independence was a Central Limit Theorem: Voiculescu identified the Wigner law as the free cen- tral limit for scalar-valued free independence [23] and its operator-valued analogue [25], Muraki [12] showed that the arcsine distribution is the monotone central limit and Popa [15] identified an operator-valued analogue for the arcsine, Speicher and Woroudi [21] proved the Boolean central limit theorem, and Bozejko, Leinert and Speicher [8] described the pairs of measures that appear as conditionally free central limits. However, the "most general" limit theorems involve so-called infinitesimal arrays, and the limit distributions are usually shown to identify with infinitely di- visible distributions. Some descriptions/characterizations of infinite divisibility are known for all noncommutative scalar independences [5, 21, 12, 11, 28], but very little is known about operator-valued infinite divisibility; until recently, the only exception we know of was Speicher's work [20]. In a recent breakthrough, Popa and Vinnikov [18] gave a description of free, Boolean and conditionally free infinitely divisible distributions in terms of their linearizing transforms that parallels the re- sults of Bercovici and Voiculescu [5], Speicher and Woroudi [21] and Krysztek [11], respectively. In this paper we will use some of the results from [18], the subordina- tion result of Voiculescu [26] and the theory of noncommutative functions to prove a Hin¸cin type theorem for free and conditionally free convolutions and to identify operator-valued analogues of the Bercovici-Pata bijection [4]. In addition, we prove a generalization to conditionally free convolution of the result of Belinschi and Nica [3] which states that the Boolean Bercovici-Pata bijection is a homomorphism with respect to free multiplicative convolution. Our paper is organized as follows: in the second section we define the main notions and tools that we shall use, and prove a few preparatory results, in Section 3 we prove a Hincin type theorem for certain infinitesimal arrays of operator-valued random variables; using this result, we obtain a restricted Boolean-to-free Bercovici- Pata bijection, in Section 4 we prove our main result, the Boolean-to-conditionally free Bercovici-Pata bijection, and finally in Section 5 we show that the Boolean-to- conditionally free Bercovici-Pata bijection for scalar-valued random variables is a homomorphism with respect to multiplicative c-free convolution. INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 3 2. independences, transforms and subordination We start with a precise definition of an operator valued distribution. Gen- erally, we will assume that B is a unital C-algebra. We will denote by BhXi the ∗-algebra freely generated by B and the selfadjoint symbol X . We will also use the notation BhX1,X2, . . .i for the ∗-algebra freely generated by B and the non- commutating self-adjoint symbols X1,X2, . . . . The set of all positive conditional expectations from BhXi to B will be denoted by ΣB. For B ⊆ D a unital inclusion of C∗-algebras, we denote by ΣB:D the set of all unital, positive B-bimodule maps µ : BhXi −→ D with the property that for all positive integers n and all {fi(X )}n (1) i=1 ⊂ BhXi we have that: i,j=1 ≥ 0 in Mn(B). (cid:2)µ(fj(X )∗fi(X ))]n Remark that ΣB = ΣB:B, as an easy consequence of Exercise 3.18 from [14]. Using these notations, we define the distribution of an operator-valued random variable: Definition 2.1. If B ⊆ A, B ⊆ D are unital inclusions of ∗-algebras, respectively C∗-algebras, φ : A −→ D is a unital positive B-bimodule map and a is a selfadjoint element of A, we will denote by Bhai the ∗-algebra generated in A by B and a and by φa, "the D-distribution" of a, that is the positive B-bimodule map φa : BhXi −→ D defined by φa = φ ◦ τa where τa : BhXi −→ A is the unique homomorphism such that τa(X ) = a and τa(b) = b for all b ∈ B. We will denote by Σ0 B:D the set of elements from ΣB:D with the property that there exists some M > 0 such that, for all b1, . . . , bn ∈ B we have that (2) µ(X b1X b2 ···X bnX ) < M n+1b1···bn. B:D (see [18], Proposition 1.2). With the notations from Definition 2.1, an element µ ∈ ΣB:D can be realized as φa for some C∗-algebra A containing B and a a self-adjoint from A if and only if µ ∈ Σ0 2.1. Independences. There are several independences important for noncommu- tative probability. We shall start with the oldest and best-known, Voiculescu's free independence. We present it in a C∗-algebraic context. Definition 2.2. Let B be a C∗-algebra, B ⊆ A be a unital inclusion of ∗-algebras and φ : A → B be a positive conditional expectation. A family {Xi}i∈I of selfadjoint elements from A is said to be free with respect to φ if φ(A1A2 ··· An) = 0 whenever Aj ∈ BhXǫ(j)i with φ(Aj ) = 0, where ǫ(j) ∈ I, with ǫ(k) 6= ǫ(k + 1) for 1 ≤ k ≤ n − 1. Let now N ∈ N and {µj}N j=1 be a family of elements from ΣB. We define their additive free convolution the following way: Consider the symbols {Xj}N j=1; on the algebra BhX1, X2, . . . , XNi take the conditional expectation µ such that µ ◦ τXj = µj and the mixed moments of X1, . . . , Xn are computed via the rules from Definition 2.2. The free additive convolution of {µj}N j=1 is the conditional expectation ⊞N j=1µj = µ ◦ τX1+X2+···+XN : BhX1 + X2 + ··· + XNi ∼= BhXi −→ B. 4 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV We have that ⊞N j=1µj is also an element of ΣB: in [20] it is shown that µ, Secondly, we give the op-valued equivalent of Speicher and Woroudi's defined as above, is a positive conditional expectation, therefore so is µ◦τX1+X2+···+XN . Boolean independence [21] as it appears in [18]: Definition 2.3. Let B ⊆ A, B ⊆ D be unital inclusions of ∗-algebras, and φ : A → D a unital completely positive B-bimodule map. A family {Xi}i∈I of selfadjoint elements from A is said to be Boolean independent with respect to φ if φ(A1A2 ··· An) = φ(A1)φ(A2)··· φ(An) for all a ∈ Ωn, b ∈ Ωm. Notice that matrices over C act from the right and from the left on matrices over V by the standard rules of matrix multiplication. whenever Aj is in the nonunital ∗-subalgebra over B generated by Xj, where ǫ(j) ∈ I with ǫ(k) 6= ǫ(k + 1) for 1 ≤ k ≤ n − 1. The definition of Boolean convolutions of distributions from ΣB:D is done similarly to free convolutions of distributions from ΣB , as shown in [18], by simply replacing free with Boolean independence. The reader will observe that this defini- tion makes sense for B = D; the broader context that we provide adds in fact more depth to the theory. This will become clearer in the following definition, which essentially unites free and Boolean independence. We aim to extend the results that can be obtained for free independence to the case µ ∈ ΣB:D. In this setting, if B is simply replaced by D, the resulting relation does not uniquely determine the joint moments of X1, . . . , Xn. As shown in [7, 6], a more suitable approach is the c-freeness (see also [17] and [8]). Definition 2.4. Let B ⊆ A, B ⊆ D be unital inclusions of ∗-algebras, EB : A −→ B be a positive conditional expectation and θ : A −→ D be a unital B-bimodule map. The family {Xi}i∈I of selfadjoint elements from A is said to be c-free with respect to (θ, ϕ) if (i) the family {Xi}i∈I is free with respect to EB (ii) θ(A1A2 ··· An) = θ(A1)θ(A2)··· θ(An) for all Ai ∈ BhXǫ(i)i such that EB(Ai) = 0 and ǫ(k) 6= ǫ(k + 1). The reason for switching to the notation EB will be seen later. We will consider the above relations in the framework of B ⊂ D a unital inclusion of C∗- algebras and θ a completely positive map.As shown in [7], this setting (that includes ΣB:D is closed with respect to c-free convolution. Next, we describe one of our main tools in analyzing distributions of sums of independent (freely, Boolean or c-freely) operator-valued random variables, namely noncommutative sets and functions [10]. We will use the terminology from [10] (see also [22], [18], but translating the results in the different terminology from [27] is trivial.) For a vector space V over C, we let Mn×m(V), denote n× m matrices over V and write Mn(V) for Mn×n(V). We define the noncommutative space over V by Vnc = Mn(V). We call Ω ⊆ Vnc a noncommutative set if it is closed under direct sums. Explicitly, denoting Ωn = Ω ∩ Mn(V), we have ∞an=1 a ⊕ b =(cid:20)a 0 b(cid:21) ∈ Ωn+m 0 INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 5 Let V and W be vector spaces over C, and let Ω ⊆ Vnc be a noncommuta- tive set. A mapping f : Ω → Wnc with f (Ωn) ⊆ Mn(W) is called a noncommutative function if f satisfies the following two conditions: Ωn, then f (sas−1) = sf (a)s−1. • f respects direct sums: f (a ⊕ b) = f (a) ⊕ f (b) for all a, b ∈ Ω. • f respects similarities: if a ∈ Ωn and s ∈ Mn(C) is invertible with sas−1 ∈ We will denote f (n) = fΩn : Ωn → Mn(W). For convenience, we will refer some- times, when there is no risk of confusion, to f (n) and Ωn as the nth coordinate of the respective noncommutative function and set. A noncommutative set Ω ⊆ Vnc is called upper admissible if for all a ∈ Ωn, b ∈ Ωm and all c ∈ Mn×m(V), there exists λ ∈ C, λ 6= 0, such that (cid:20)a λc b(cid:21) ∈ Ωn+m. 0 (1) The set Nilp(A) =`∞ Non-commutative functions on upper-admissible sets admit a consistent differen- tial calculus, including Taylor expansions by defining the (right) noncommutative difference-differential operators by evaluating a noncommutative function on block upper triangular matrices (see again [10]). In this paper we will use only the fol- lowing three upper admissible noncommutative sets (A will denote a C∗-algebra): n=1 Nilp(A; n); here the set Nilp(A; n) consists of all a ∈ Mn(A) such that ar = 0 for some r, where we view a as a matrix over the tensor algebra T(A) of A over C (2) Noncommutative balls B(A, ρ) = {a ∈ Anc : kak < ρ} of radius ρ > 0 over A (A could have been replaced by any operator space with the corresponding operator space norm). (3) Noncommutative half-planes H+(Anc) = {a ∈ Anc : ℑa > 0} over A. Here ℑa = (a− a∗)/2i denotes the imaginary part of a. (We say that an element b in a unital C∗-algebra satisfies b > 0 if there exists ε ∈ (0, +∞) so that b ≥ ε1, where here 1 is the unit of A; of course, for an element a ∈ Anc to have imaginary part strictly greater than zero means simply that each "coordinate" has imaginary part strictly positive.) It has been first noted by Voiculescu [27] that H+(Anc) is indeed a noncommutative set. 2.2. Op-valued distributions and properties of their transforms. As for scalar-valued (non)commutative probability, there are transforms that linearize dif- ferent kinds of convolutions of operator-valued distributions. It turns out that these transforms can be described in terms of noncommutative functions defined on non- commutative spaces, associated to operator-valued distributions. First, we intro- duce some terminology and notations. If A ⊇ B is a unital inclusion of ∗-algebras and EB : A → B is a conditional expectation, then EMn(B) = EB ⊗ 1n : Mn(A) → Mn(B) is still a conditional expectation for any n ∈ N and any linear functional (in particular any trace τ ) on B extends to τ ⊗ trn : B ⊗ Mn(C) → C, where trn is the canonical normalized trace on Mn(C). Note also that if X, Y ∈ A are free, boolean independent, respectively c-free with respect to EB and φ : A → D for some algebra D containing B, then so are X ⊗ 1n and Y ⊗ 1n with respect to EMn(B) and φ⊗ 1n. We recall that a ∈ A is called selfadjoint if a = a∗. Any element a ∈ A in a ∗-algebra can be written uniquely a = ℜa + iℑa, where ℜa = (a + a∗)/2, ℑa = (a − a∗)/(2i) are selfadjoint. For any given ∗-algebra A on which a notion 6 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV of positivity coherent with the star operation has been defined, we shall denote the upper half-plane of A by H+(A) = {a ∈ A : ℑa > 0}, n=1 and H−(A) = −H+(A). (Thus, H+(Anc) =`∞ H+(Mn(A)); see also [27].) We note that a ∈ H+(A) =⇒ a∗ ∈ H−(A). A useful generalization, noted by Voiculescu, of the fact that the operation of taking inverse changes the imaginary part of a complex number from positive to negative and vice-versa is the following implication, which holds in any unital ∗-algebra in which analytic functional calculus is available: a ∈ H+(A) =⇒ a−1 ∈ H−(A). (3) (Note that the invertibility of a is part of the statement.) u = ℜa, v = ℑa, and a = u + iv, we have a = u + iv = u + i(√v)2 = √v[(√v)−1u(√v)−1 + i]√v. Indeed, by writing The ability to take square root is guaranteed by the analytic functional calculus and the fact that v > ε1A for some ε > 0. As (√v)−1u(√v)−1 is selfadjoint, it is clear that i does not belong to its spectrum, so (√v)−1u(√v)−1 + i is invertible in A. Invertibility of √v, again guaranteed by its strict positivity and the existence of analytic functional calculus, implies that a is itself invertible in A. Now writing its inverse gives = (√v)−1[(√v)−1u(√v)−1 + i]−1(√v)−1 a−1 = (cid:0)√v[(√v)−1u(√v)−1 + i]√v(cid:1)−1 = (√v)−1(cid:2)[(√v)−1u(√v)−1]2 + 1(cid:3)−1 = (√v)−1(cid:2)[(√v)−1u(√v)−1]2 + 1(cid:3)−1 {z +in−(√v)−1(cid:2)[(√v)−1u(√v)−1]2 + 1(cid:3)−1 {z =d =c [(√v)−1u(√v)−1 − i](√v)−1 [(√v)−1u(√v)−1](√v)−1 } . (√v)−1o } Analytic functional calculus guarantees that c = c∗, d = d∗ and −d > 0. The uniqueness of the expansion into real and imaginary part guarantees that c = ℜ(a−1), d = ℑ(a−1), so our claim is proved. We indicate next how an operator-valued distribution can be encoded by noncommutative functions. Note that if µ ∈ ΣB:D, then (µ ⊗ 1n) ∈ ΣMn(B):Mn(D). Moreover, (see [10], [27]) for b1 0 . . . 0 0 0 0 . . . 0 0 b = (µ ⊗ 1n+1)(cid:0)[X b]n) = 0 b2 . . . 0 0 . . . . . . . . . . . . . . . 0 0 . . . bn 0  ∈ Mn+1(B) 0 0 . . . 0 . . . . . . . . . . . . 0 0 . . . 0 µ(X b1X b2 . . .X bn) 0 . . . 0  , we have that (4) INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 7 so µ is completely characterized by the sequence(cid:8)(µ ⊗ Idn)(cid:0)[X · b]m(cid:1)(cid:9)m,n. Using this observation, we shall indicate how all the information describ- ing a distribution can be encapsulated in a fully matricial (or noncommutative) function. For a given µ ∈ ΣB:D, we define its moment-generating series as the non-commutative function of components M (n) µ (b) = ∞Xk=0 (µ ⊗ 1n)([X · b]k) = 1n + (µ ⊗ 1n)(X · b) +(µ ⊗ 1n)(X · b · X · b) + ··· . As shown in [18], Mµ is always well-defined on Nilp(B). Moreover, from Engel's Theorem, a ∈ Nilp(Mn(B)) if and only if T aT −1 is strictly upper triangular for some T ∈ GL(n), therefore T [Mµ(b) − 1]T −1 is also upper-triangular hence Mµ(b) − 1 ∈ Nilp(D). As in [18], we used the symbol 1 for Idn on each component from Mn(B). If µ ∈ Σ0 B:D, then Mµ is also well-defined on a small non-commutative ball from Bnc which is mapped in a non-commutative ball from Dnc. via the functional equations For ν ∈ ΣB and µ ∈ ΣB:D we define their R-, respectively B-transforms (5) (6) Mν(b) − 1 = Rν (b · Mν(b)) Mµ(b) − 1 = Bµ(b) · Mµ(b) In [18] is is shown that the R- and B-transforms are non-commutative functions well-defined on Nilp(B). If ν ∈ Σ0 B, respectively µ ∈ Σ0 B:D, then Rν and Bµ are also well-defined in some non-commutative balls from Bnc. their linearizing property: we have The main reason for which we have introduced the R and B-transforms is (7) (8) Rµ⊞ν(b) = Rµ(b) + Rν(b), µ, ν ∈ ΣB, Bµ⊎ν (b) = Bµ(b) + Bν(b), µ, ν ∈ ΣB:D. The first result is due to Voiculescu [25], and the second to Popa [16]. We warn the reader that the other version of the R-transform, defined below, namely the original one of Voiculescu, as well as the one used by Dykema (that we will call here R), is related to this version by a simple multiplication to the right with the variable b: Rµ(b) = Rµ(b)b. Depending on the functional context, sometimes it is convenient to use slight variations of these transforms, which benefit of similar properties. We start with the most straightforward: following [15], we introduce the shifted moment generating function Hµ of µ ∈ ΣB:D as the non-commutative function of components µ (b) = b · M (n) H(n) µ (b). When there is no risk of confusion, we will denote H (1) µ also by Hµ. For a better understanding of the way noncommutative functions gener- alize the classical functions associated to probability distributions, it will be conve- nient to express the transform Hµ in terms of the generalized resolvent or operator- valued Cauchy transform. Several properties of scalar-valued Cauchy transforms (easily proved, but available also in [1, Chapter III]) are preserved when we pass to the operator-valued context. We shall express the operator-valued Cauchy trans- form first in terms of random variables. Suppose that B ⊂ A, B ⊂ D are inclusions 8 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV of unital C∗-algebras and φ : A → D is a completely positive B-bimodule map. For any fixed X = X ∗ ∈ A we let GX = (G(n) X )n, G(n) X : H+(Mn(B)) → H−(Mn(D)), G(n) X (b) = φn[(b − X ⊗ Idn)−1], where φn is the extension of φ from Mn(A) to Mn(D). For n = 1 we shall denote G(1) X (b) simply by GX (b). Let us first remark that this expression makes sense: indeed, since ℑb > 0 and X = X ∗, it follows that ℑ(b − X ⊗ Idn) > 0, so, as noted in (3), b− X ⊗ Idn is invertible in A and its inverse has strictly negative imaginary part. Since φ is completely positive, it follows in addition that G(n) X takes values in the lower matricial half-plane. We observe that, whenever kb−1k < 1/kXk, we can write GX (b) =P∞ as a convergent series. Thus, it follows easily that for µ ∈ Σ0 B:D we shall write G(n) µ (b) = ∞Xn=0 (µ ⊗ Idn)(b−1(X · b−1)n) = (µ ⊗ Idn)[(b − X )−1]; n=0 b−1φ[(Xb−1)n] (of course, these equalities require that we consider an extension of µ to BhhXii, the algebra of formal power series generated freely by B and the selfadjoint symbol X ). This also indicates a very important equality, namely, for µ ∈ Σ0 (9) Moreover, Gµ(b∗) = [Gµ(b)]∗ extends Gµ to the lower half-planes, analytically through points b with inverse of small norm. H+(Mn(B)) and GX with the structures defined above are fully matricial sets and functions. It is easy to observe that the same is true for FX , the reciprocals of GX , namely It has been shown by Voiculescu [27] that H(Bnc) = `∞ Gµ(b−1) = Hµ(b), b ∈ H+(Mn(B)). n=1 B:D FX (b) = [GX (b)]−1, X (b) = [G(n) F (n) X (b)]−1. Remark 2.5. When X is a B-valued selfadjoint random variable, the transform F has many properties in common with its scalar-valued analogue. First of all, it follows straightforwardly from the similar property of G that F (n) X necessarily maps H+(Mn(B)) into itself. Moreover, under the condition that B has a rich enough collection of positive linear functionals (for example if it is a C∗-algebra) we always have ℑF (n) X (b) ≥ ℑb for all b ∈ H+(Mn(B)). Indeed, for n = 1, for any fixed positive linear functional f on B, let us define FX,f (z) = f (FX (ℜb + zℑb)). Clearly, since ℑb > 0, we have that FX,f : H+(C) → H+(C). In addition, (iy)n ((X − ℜb)(ℑb)−1)n(cid:21)#−1 Thus [1, Chapter III] there exists a positive compactly supported Borel measure ρ z−t dρ(t)i−1 Gρ(z) =hRR on the real line of mass 1/f ((ℑb)) so that FX,f (z) = 1 for all z in the upper half-plane. The Nevanlinna representation of FX,f implies that ℑFX,f (z) ≥ f (ℑb)ℑz for all z ∈ H+(C), and so, since this holds for all positive linear functionals f on B, ℑFX (b) ≥ ℑb for all b ∈ H+(B). Moreover, we note that (ℑb)−1φ(cid:20) 1 f" ∞Xn=0 = f (ℑb) > 0. FX,f (iy) y→+∞ y→+∞ lim lim iy = 1 INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 9 if equality holds for a given b0, then X − ℜb0 must be a multiple of the identity of B. The argument for a general component F (n) X is analogous. We would like also to mention the connection between F and B: (10) 1 − Fµ(b−1)b = Bµ(b), b−1 ∈ H+(Bnc). For classical measures it is known that weak convergence to a finite mea- sure is equivalent to uniform convergence on compact sets for the Cauchy transforms to the Cauchy transform of the limit, and, if all measures involved are compactly supported, these two statements are equivalent to the convergence of moments (we say σn → σ in moments if R tjdσn(t) converges to R tjdσ(t) for any j ∈ N.) We shall provide below two versions of this result for operator-valued distributions. First, let us define convergence in moments for an operator-valued distri- bution. Definition 2.6. Given a sequence of distributions µn ∈ ΣB:D, we say that (a) µn converges to µ ∈ ΣB:D pointwise in moments if for any ϕ ∈ D∗ we have lim n→∞ µn(X b1X b2 ··· bjX ) = µ(X b1X b2 ··· bjX ), for all b1, . . . , bj ∈ B; (b) µn norm-converges to µ ∈ ΣB:D in moments if lim n→∞ sup kbkk=1,1≤k≤j kµn(X b1X b2 ···X bjX ) − µ(X b1X b2 ···X bjX )k = 0, for all j ∈ N. Remark first that for finite dimensional algebra B the two notions are equivalent. Also, remark that condition [(b)] is equivalent to (b′) lim n→∞ Mµn(b) = Mµ(b), for all b ∈ Nilp(B). In this paper we will mainly be interested in norm-convergence of moments. We note next the following simple remark: Remark 2.7. Assume that B ⊆ D is a unital inclusion of von Neumann algebras and {µn}n∈N is sequence from Σ0 B:D Then µn converges pointwise in moments to µ ∈ Σ0 Proof. First, let us assume that µn converges to µ pointwise in moments. It suffices to prove that for any state ϕ on Mm(D) we have that B:D if and only if Gµn converges pointwise to Gµ on H+(Bnc). lim n−→∞ ϕ(cid:0)G(m) µn (b)(cid:1) = ϕ(cid:0)G(m) µ (b)(cid:1). Let ϕ as above and b ∈ H+(Mm(B)). The function H+(C) ∋ z 7→ ϕ(cid:0)G(m) zℑb)(cid:1) is the Cauchy transform of some positive real measure σn,ϕ,b. Indeed, and since ϕ, µn ⊗ 1m,ℑb and ℑz are all positive, it follows that ℑϕ(G(m) zℑb)) < 0. Moreover, µn (ℜb + zℑb)) = ϕ((µn ⊗ 1m)(cid:2)(ℜb + zℑb − X )−1(cid:3)), ϕ(G(m) µn (ℜb + µn (ℜb + lim z→∞ z(µn ⊗ 1m)(cid:2)(ℜb + zℑb − X )−1(cid:3) = (11) (ℑb)−1(µn ⊗ 1m)(cid:20) ((ℜb − X )(ℑb)−1)j zj (cid:21) lim z→∞ ∞Xj=0 = (ℑb)−1 10 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV (the limit is in the norm topology of Mm(D); these expressions make sense for z large enough because of the exponential growth condition) Hence applying ϕ and using that ℑb > 0 gives the result. Thus, according to [1], Chapter III, we have that ϕ(G(m) µn (ℜb + zℑb)) =ZR (z − t)−1 dσn,ϕ,b(t), with σn,ϕ,b(R) = ϕ((ℑb)−1). We observe in addition that ZR tj dσn,ϕ,b(t) = ϕ(cid:0)(µn ⊗ 1m)(cid:2)(ℑb)−1((ℜb − X )(ℑb)−1)j(cid:3)(cid:1) , so by our hypothesis and the continuity of the multiplication with the constant (ℑb)−1 we obtain that the moments of σn,ϕ,b(t) converge. Normality of the family {ϕ(G(m) µn (ℜb + zℑb))}n guarantees that a weak limit of this sequence of measures exists, and the limit has the prescribed moments. We conclude that ϕ(G(m) µn (ℜb + zℑb)) converges to ϕ(G(m) µ (ℜb + zℑb)) uniformly on compacts of H+(C) for any fixed b ∈ H+(Mm(B)). µ (b)) for any m and any b ∈ H+(Mm(B)). We define again σn,ϕ,b as above and observe that ϕ(G(m) µn (ℜb + zℑb)) is the Cauchy transform of this measure and it converges uniformly on compacts of the complex upper half-plane to the Cauchy transform of a limit measure σϕ,b. By applying ϕ to equation (11) we obtain that σϕ,b has moments Assume next that limn→∞ G(m) µn (b)) = G(m) ϕ(cid:0)(µ ⊗ 1m)(cid:2)(ℑb)−1((ℜb − X )(ℑb)−1)j(cid:3)(cid:1), so that ϕ(cid:0)(ℑb)−1(µn ⊗ 1m)(cid:2)(ℜb − X )(ℑb)−1(ℜb − X )··· (ℜb − X )(cid:3) (ℑb)−1(cid:1) = ϕ(cid:0)(ℑb)−1(µ ⊗ 1m)(cid:2)(ℜb − X )(ℑb)−1(ℜb − X )··· (ℜb − X )(cid:3) (ℑb)−1(cid:1) , from which we obtain pointwise convergence for all symmetric moments of the fully matricial extensions of µn to µ, hence for all moments of µn and µ. (cid:3) Definition 2.8. A sequence {µn}n from Σ0 B:D is said to be uniformly bounded (by M ) if there exists some constant M > 0 such that for all n, p and all b1, . . . , bp ∈ B we have that lim n→∞ kµn(X · b1 · X ··· bp · X )k < M p+1kb1k ···kbpk. Note that, according to [18], Proposition 1.2, the above condition holds true (with the same M ) for µn ⊗ 1m and b1, . . . , bp ∈ Mm(B). For the next Proposition we shall find useful the following (purely Banach space) results, which can be found in the first chapters of [9](Theorems 1.5 and 1.6). Let E and E1 be complex Banach spaces, D ⊂ E and D1 ⊂ E1 be bounded domains. Following [9], we denote by Hol(D, D1) the set of holomorphic mappings a (h, . . . , h) on a from D into D1, that is, functions f for which f (a + h) =P∞ neighborhood of a, for any a in D; where ∂n n=0 f (n f (n a (h1, . . . , hn) = 1 n! ∂t1 ··· ∂tn f (a + t1h1 + ··· + tnhn) is a continuous n-linear map from En to E1. We shall also denote B ⊂⊂ D if B is a subset of D with the additional property that the norm distance from B to ∂D is strictly positive. Theorem 2.9. Let (fj)j∈J be a net in Hol(D, D1) and f ∈ Hol(D, D1), and B ⊂⊂ D be a ball centered at a ∈ D. The following statements are equivalent: INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 11 (1) The net (fj)j∈J is uniformly convergent to f on B; (2) For all k ∈ N we have limj∈J kf (k j,a − f (k a k = 0. Theorem 2.10. Let (fj)j∈J be a net in Hol(D, D1). For any two balls B1, B2 ⊂⊂ D the following statements are equivalent: (1) fj → f relative to k · kB1 ; (2) fj → f relative to k · kB2 . where kfkB = supx∈B kf (x)k. We would like to emphasize that in Remark 2.7 we do NOT require that the sequence {µn} is uniformly bounded. However, in order to be able to prove the similar result for norm-convergence of moments in the most general C∗-algebraic context, we will have to require that. Proposition 2.11. Assume B ⊆ D is a unital inclusion of C∗-algebras and that the sequence {µn}n∈N ⊂ Σ0 B:D is uniformly bounded. The following statements are equivalent: (1) {µn}n∈N norm-converges in moments to some µ ∈ Σ0 (2) for all positive integers m, G(m) µn converges uniformly to G(m) H+(Mm(B)) which lay at positive distance from ∂H+(Mm(B)). B:D; µ on balls in Proof. Let us denote by M a common bound for {µn}n and µ as in Definition 2.8. (0), the (0), Mm(D)) is well-defined in B 1 2M µ ∈ Hol(B 1 For (2)⇒(1), note first that for all m, H 2M from Mm(B). Moreover, H ball of center zero and radius 1 and (m) µ (m) 2M (12) H (m)(k µ,0 (h1, . . . , hk) = 1 k! Xσ∈Sk (µ ⊗ 1m)(cid:0)hσ(1)X hσ(2) ···X hσ(k)(cid:1) where Sk denotes the symmetric group with k elements. Fix now b ∈ H−(Mm(B)) with kbk < 1 2M . Then there exist some small R > 0 such that BR(b−1) ⊂⊂ H+(Mm(B)). Since Gµ(h−1) = Hµ(h), it follows (0) and {Hµn}n converges that there exists some r > 0 such that Br(b) ⊂⊂ B 1 uniformly to Hµ on Br(b). Applying Theorem 2.10, we have that {Hµn}n converges uniformly to Hµ on B 1 (0), hence 2M 4M kH (m),(k µn,0 − H (m),(k µ,0 k −→ 0 as k-linear maps from Mm(cB)k to Mm(D). But (12) gives (h, . . . , h) = (µ ⊗ 1m) (hX h···X h) and, since m is arbitrary, equation (4) allows us to conclude. H (m),(k µ,0 For (1)⇒(2), we will use the result from [18], Proposition 1.2 namely that if µ ∈ Σ0 B:D then there exist a C∗-algebra A containing B, some selfadjoint X ∈ A and a unital completely positive B-bimodule map φ : A −→ D such that for all noncommutative polynomials f with coefficients in B we have that µ(f (X )) = φ(f (X)). 12 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV Fix now b0 ∈ H+(Mm(B)). Since ℑ(X − b0) = −ℑ(b0) < 0, we have that X − b0 is invertible in Mm(A) and (13) k(X − b0)−1k = k(iℑb0 + ℜb0 − X)−1k = k(ℑb0)−1/2(i + (ℑb0)−1/2(ℜb0 − X)(ℑb0)−1/2)−1(ℑb0)−1/2k ≤ k(ℑb0)−1kk(i + (ℑb0)−1/2(ℜb0 − X)(ℑb0)−1/2)−1k ≤ k(ℑb0)−1k. (We have used here the fact that Mm(B) is a C∗-algebra, the fact that ℑb0 is selfadjoint, as well as the fact that i + (ℑb0)−1/2(ℜb0 − X)(ℑb0)−1/2 is normal, so that one can apply continuous functional calculus to it.) Note that the above majorization is independent of X. Also, for khk < 1 −1k , we have kℑb0 (b0 + h − X)−1 = [h + (b0 − X)]−1 = (X − b0)−1[h(X − b0)−1 − 1m]−1 = (b0 − X)−1 . ∞Xn=0(cid:2)h(X − b0)−1(cid:3)n Also, G(m),(k µ,b0 = 1 k! Xσ∈Sk φm(cid:0)(b0 − X)−1hσ(1)(X − b0)−1hσ(2) ··· hσ(k)(X − b0)−1(cid:1) To simplify the notations, we will prove (2) for m = 1; for an arbitrary m By the above, it is easy to observe that each of these k-linear functionals is bounded in norm by (k + 1)!(k(ℑb)−1kk + k(ℑb)−1kkkℜbk). all the computations are similar, using the matricial extensions of µ and {µn}n. We shall prove that there exists a point, namely Q = (1 + 2M )i, around which there exists a ball of radius 1/2 on which Gµn converges to Gµ uniformly in norm. Then we shall use Theorem 2.10 to argue that this implies uniform convergence on any ball B ⊂⊂ H+(B) - the result will be proved by using Theorem Since φ is unital and completely positive, we have that kφkcb = kφ(1)k = 1, hence we can apply φ to the power series development of (b0 + h − X)−1. It follows that G(m) µ (b0 + h) = φm(cid:0)(b0 + h − X)−1(cid:1) = ∞Xn=0 φm(cid:0)(b0 − X)−1[h(X − b0)−1]n(cid:1) , hence kGµ(b)k = kφ[(b − X)−1]k ≤ ≤ ≤ ∞Xn=0(cid:13)(cid:13)(cid:13)φh(X − b0)−1(cid:2)(b − b0)(X − b0)−1(cid:3)ni(cid:13)(cid:13)(cid:13) ∞Xn=0 kφkcbk(X − b0)−1kn+1kb − b0kn k(ℑb0)−1k . 1 − k(ℑb0)−1kkb − b0k INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 13 2.9. Indeed, let us start by observing that Theorems 2.10 and 2.9 indeed apply to our functions whenever we restrict them to H+(B) + ic for any c > 0. For a fixed pozitive integer k we have that 1 1 1 = = X m1 Qm1 Q(cid:19)−1# µ(cid:2)(Q − X )−1h1(Q − X )−1 ··· hk(Q − X )−1(cid:3) = ··· hk(cid:18)1 − X h1(cid:18)1 − X Q(cid:19)−1 Qmk+1 h1X m2 Qm2 ··· hk X mk+1 Qk+1 µ"(cid:18)1 − X Q(cid:19)−1 Qk+1 µ ∞Xm1,...,mk+1=0 Qk+1 µ ∞Xm1,...,mk+1=0 = µ ∞Xq=0 Qk+1+q Xm1+···+mk+1=q ∞Xq=0 Qk+1+q Xm1+···+mk+1=q kµ [X m1 h1X m2 ··· hkX mk+1 ]k ≤ M m1+m2+···+mk+1kh1k ···khkk Qm1+m2+···+mk+1 X m1 h1X m2 ··· hkX mk+1 X m1 h1X m2 ··· hkX mk+1 µ [X m1 h1X m2 ··· hkX mk+1 ] . = 1 1 1 = M qkh1k ···khkk We note that the majorization guarantees that the last sum above is majorized in norm by M q(q + k)k; since Q = 1 + 2M, the convergence of this expression is not a problem. Note that all the above estimates hold true also for {µn}n. We claim that Indeed, let ε > 0 be fixed. By the choice of Q (and the norm convergence of the last series from above) it follows that there exists a positive integer q(ε, Q) not depending on n, so that lim n→∞(cid:13)(cid:13)µn(cid:2)(Q − X )−1h1(Q − X )−1 ··· hk(Q − X )−1(cid:3) −µ(cid:2)(Q − X )−1h1(Q − X )−1 ··· hk(Q − X )−1(cid:3)(cid:13)(cid:13) = 0. µn [X m1 h1X m2 ··· hkX mk+1 ](cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) < 1 Qk+1+q Xm1+···+mk+1=q for any r ≥ q(ε, Q). Fix r = 2 + q(ε, Q), and observe that by the norm-convergence of the moments of µn to the moments of µ, we can find Nε ∈ N so that for any n ≥ Nε we have that whenever kh1k, . . . ,khkk ≤ 1. kµn [X m1 h1X m2 ··· hkX mk+1] − µ [X m1 h1X m2 ··· hkX mk+1 ]k < ε 2 ∞Xq=r (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXq=0 Xm1+···+mk+1=q ε 8kh1k ···khkk This proves that limn→∞ kG(k µ (Q)k = 0, as k-linear operators from B to D. Since k is arbitrary, the second condition of Theorem 2.9 is satisfied, so Gµn converges locally uniformly in the norm topology of D to Gµ, as claimed. µn (Q) − G(k (cid:3) 14 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV The last two results have established the connection between transforms and distributions in the most general case that we will consider. Remark 2.12. (a) Due to the relevance of this particular case, we emphasize again that if B,D are finite dimensional, then the sequence {µn}n∈N from Proposition 2.11 needs not be uniformly bounded. Indeed, generally local compactness of finite dimensional spaces makes proofs considerably simpler. Unfortunately, infinite dimensional Ba- nach spaces are not locally compact in the norm topology, so in particular closed balls are not compact. However, some properties of analytic maps on finite di- mensional spaces remain true, including the continuity of composition operation [9, Theorem 1.10]. (b) Using Theorem 2.10 and [9, Theorem 1.10], we can replace the Cauchy transform G in Proposition 2.11 with any of the transforms F , B, M , H, or, if B = D, R. (c) Not the same thing can be said about the lemma preceding it; in that case, in the most general context G can only be replaced by H or M . However, there are many special cases in which G can be replaced by F or R. We introduce a few other new transforms and modifications of the trans- forms already introduced which will be of use to us. We recall the R-transform, which can be expressed as Rµ(b) = G−1 µ (b)− b−1 for b invertible and of small norm, and which linearizes free additive convolution. (We use the convention that the −1 as exponent on the letter which denotes a function means compositional inverse, while a −1 exponent on the function evaluated in a point means the multiplica- tive inverse of the value of the function in that point: thus, f (b)f (b)−1 = 1, while f (f −1(b)) = b.) From it we shall define the Voiculescu transform ϕµ of µ simply as ϕµ(b) = Rµ(b−1). This function is easily seen to be defined on an open set in B and Fµ(ϕµ(b) + b) = b, so ℑϕµ(b) ≤ 0 whenever ℑb > 0. Obviously, the Voiculescu transform also satisfies ϕµ(b) + ϕν (b) = ϕµ⊞ν (b). Finally, denoting hµ(b) = Fµ(b) − b, b ∈ H+(Bnc), we re-write (8) using hµ⊎ν(b) = hµ(b) + hν(b) b ∈ H+(Bnc). (10): (14) We shall deal next with linearizing transforms for conditionally free convo- lutions. Distributions (µ, ν) ∈ ΣB:D × ΣB can be associated a linearizing transform for c-free convolution, the cR-transform. It is defined by the functional equation (15) [Mµ(b) − 1] · Mν(b) = Mµ(b) · cRµ,ν(bMν(b)). (We remind the reader that the second coordinate is linearized by the R-transform.) Note that when B = D and EB = θ in Definition 2.4 we obtain µ = ν, so c-free convolution simply coincides with free convolution for both coordinates. In addition, if ν is the distribution of the zero random variable (corresponding to Mν(b) = 1), then we obtain Mµ(b)− 1 = Mµ(b)cRµ,ν(b), which is equivalent to the definition of the B-transform. Thus, conditionally free convolution can be viewed as interpolating between free and Boolean convolution, as in the scalar case [8]. We will often work in terms of selfadjoint random variables, and not ele- ments in ΣB:D or Σ0 B:D; of course, the two approaches are fully equivalent. INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 15 2.3. Voiculescu's subordination result and c-freeness. Next, we come to the problem of subordination (the reason why we chose in one of the above definitions to write conditional expectation from A to B as EB instead of ϕ). Generally we denote by EV the conditional expectation from the "large" algebra onto the subalgebra V. Voiculescu shows in [26, Theorem 3.8] that Theorem 2.13. Assume that the selfadjoint operator valued random variables X and Y are free with amalgamation over B. Then there exists a unique map ω(n) : H+(Mn(B)) → H+(Mn(B)) so that X+Y (b) = G(n) b ∈ H+(Mn(B)). (16) EMn(BhXi)[(b − (X + Y ) ⊗ 1n)−1] = [ω(n)(b) − X ⊗ 1n]−1, In particular, G(n) X (ω(n)(b)) for all n ∈ N, b ∈ H+(Mn(B)). In addition, ω = (ω(n))n is a non-commutative function and ℑω(b) ≥ ℑb for all b ∈ H+(Bnc). Now let us look at the subordination problem from two different perspec- tives. First, pick A,B,D, θ, EB as in Definition 2.4. Let X, Y be selfadjoint and c-free in A over the pair of algebras B,D. We call their (pair) distributions (µX , νX ) and (µY , νY ). We shall re-express the c-free convolution of those two in terms of the subordination functions for the second coordinates. We note first that relation (15) holds for b of small enough norm. Next, recall that if we require in addition that b is also invertible, then Mµ(b) = b−1Gµ(b−1). This holds in general, for distri- butions µ ∈ ΣB:D. In such a case, G(n) µX (b) = (θ ⊗ 1n)[(b − X)−1] for b with positive imaginary part, or with norm of its inverse small enough. Now we rewrite (15) in terms of G: (b−1Gµ(b−1) − 1)b−1Gν (b−1) = b−1Gµ(b−1)cRµ,ν(Gν (b−1)). Replace b−1 by b: (b · Gµ(b) − 1) · b · Gν (b) = b · Gµ(b)cRµ,ν (Gν (b)). Observe that here we can simplify a b to get Gµ(b) · b · Gν (b) − Gν(b) = Gµ(b)cRµ,ν (Gν (b)). Now take (µ, ν) = (µX , νX ), denote by ω1 the subordination function with respect to X for the second coordinate (so that EBhXi[(b − X − Y )−1] = [ω1(b) − X]−1) and replace b by ω1(b). Since ℑω1(b) ≥ ℑb, the equation will hold provided (ℑb)−1 is sufficiently small. We get GµX (ω1(b))ω1(b)GνX (ω1(b)) − GνX (ω1(b)) = GµX (ω1(b))cRµX ,νX (GνX (ω1(b))). Finally, we multiply to the left by FµX (ω1(b)): ω1(b)GνX (ω1(b)) − FµX (ω1(b))GνX (ω1(b)) = cRµX ,νX (GνX (ω1(b))). Repeat the same process with X replaced by Y and ω1 by ω2 to get ω2(b)GνY (ω2(b)) − FµY (ω2(b))GνY (ω2(b)) = cRµY ,νY (GνY (ω2(b))). Two remarks which are immediate: first, from Theorem 2.13, we have GνX (ω1(b)) = GνY (ω2(b)) = GνX+Y (b). Second, cR linearizes c-free convolution. Thus, replacing in the above relations and adding them gives us cRµX+Y ,νX+Y (GνY +X (b)) = [ω1(b) − FµX (ω1(b)) + ω2(b) − FµY (ω2(b))]GνX+Y (b) 16 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV But this means (if we express cR properly) that (b − FµX+Y (b))GνX+Y (b) = [ω1(b) − FµX (ω1(b)) + ω2(b) − FµY (ω2(b))]GνX+Y (b) Denoting h(b) = F (b) − b and simplifying the invertible GνX+Y (b) provides us with the operator-valued analogue of [2, Corollary 4]: b ∈ H+(B). hµX+Y (b) = hµX (ω1(b)) + hµY (ω2(b)), (17) If there is one drawback to this formula it is that we cannot state that ℑFµ(b) ≥ ℑb, as the target algebra for F is D (or, to be more precise, its upper half-plane.) Let us also note that BµX (b−1)b = b − FµX (b) = −hµX (b), so the above can be written (however in a less pleasant form) in terms of B. For our purposes (related to the infinite divisibility and the triangular arrays of identically distributed rows), we note that the n times c-free convolution is given as (18) hµX1+···+Xn (b) = nhµX1 (ωn(b)), (ωn(b)). where ωn is the subordination function with respect to n-fold free additive convo- lution: GνX1 +···+Xn (b) = GνX1 In addition to this result, we would like to emphasize the connection of c-free independence with Markovianity, as discussed by Voiculescu. Let us recall that µX : BhXi → D is given by µX (P (X )) = θ(P (X)) ∈ D. In particular, we can look at Voiculescu's subordination theorem and apply θ to its formula: b ∈ H+(B) θ(cid:0)EBhXi[(b − (X + Y ))−1](cid:1) = θ(cid:0)[ω1(b) − X]−1(cid:1) , it is clear from the definition that θ(cid:0)[ω1(b) − X]−1(cid:1) = Since ω1(b) ∈ H+(B), GµX (ω1(b)). In particular, if D happens to be any von Neumann algebra so that B ⊂ D ⊂ BhXi and θ itself is a conditional expectation, this simply indicates that c-freeness is in fact a different expression of Markovianity, or, differently said, c-freeness generalizes the Markov property for free algebras. We conclude this section with a short remark: it follows from (17) with µX = νX , µY = νY that hνX (ω1(b)) + hνY (ω2(b)) = hνX+Y (b), so b ∈ H+(Bnc). FνX+Y (b) = ω1(b) + ω2(b) − b (19) 3. A Hincin type theorem for free additive convolution and the restricted Boolean Bercovici-Pata bijection We shall first prove a restricted version of the Boolean Bercovici-Pata bijection, using methods inspired by [3]. There the bijection was conveniently expressed in terms of the subordination function for the power two free additive convolution. We shall obtain the same result here: it follows from the operator- valued R-transform property (7), Theorem 2.13 and analytic continuation that Proposition 3.1. For any B-valued distribution µ, we denote ω the subordination function for µ ⊞ µ. Then the following functional equations hold : (20) (21) ω(b) = 1 2 (b + Fµ⊞µ(b)) = 1 2 (b + Fµ(ω(b)), Fµ⊞µ(b) = Fµ(cid:18) 1 2 (b + Fµ⊞µ(b))(cid:19) , b ∈ H+(B). INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 17 Both these equations hold for the fully matricial extensions of the functions in- volved. We prove next a Hincin type theorem. Theorem 3.2. Assume that {Xjk}j∈N,1≤k≤kj is a triangular array of random vari- ables in (A, EB,B) of elements free over B so that {Xjk : 1 ≤ k ≤ kj} have the same distribution with respect to EB for each j ∈ N (i.e. rows are identically distributed). Assume in addition that j→∞ kXj1 + ··· + Xjkjk ≤ M lim sup for some M ≥ 0. If limj→∞ Xj1 + Xj2 + ··· + Xjkj exists in distribution as norm- limit of moments in ΣB, then the limit distribution is freely infinitely divisible over B. Remark/Definition 3.3. We shall call a triangular array of identically distributed rows satisfying limj→∞ kj = ∞ and lim supj→∞ kXj1 + ··· + Xjkjk ≤ M for some M ≥ 0 infinitesimal. We shall call a triangular array of distributions infinitesimal if they can be realized operatorially by an infinitesimal triangular array of random variables. It should be noted that in scalar-valued probability, being infinitesimal means that for any ǫ > 0, limj→∞R χ[−ǫ,ǫ](t) dµXj1 (t) = 1; thus, we require a stronger notion of infinitesimality for our theorem. It should be noted again, how- ever, that when B is finitely dimensional, the above theorem remains true even when the stronger requirement of infinitesimality is removed. Proof. We first observe that, due to the fact that GX is a noncommutative function and the upper half-plane an admissible noncommutative set, it is enough to consider in our proof the functions GX . Denote µj the distribution of X1j, i.e. Gµj = EB[(b − X1j)−1]. We shall use the subordination result: Gµj (ωj(b)) = G ⊞kj µ j (b), ωj(b) = 1 kj b +(cid:18)1 − 1 kj(cid:19) Fµj (ωj(b)), b ∈ H+(B). we observe that ωj is in fact the reciprocal of the Cauchy transform of a positive B-valued distribution; indeed, it is clear from (14) that ωj is indeed the reciprocal . Using the characterization in terms of of the Cauchy transform of (µ R-transform of the free infinite divisibility of Popa and Vinnikov [18, Theorem 5.9], as the operator-valued Voiculescu transform of this distribution is simply ϕ(w) = (kj − 1)(w − Fµj (w)), w ∈ H+(B), it follows that (µ is freely infinitely divisible. )⊎1−k−1 )⊎1−k−1 ⊞kj j ⊞kj j j j Now our proof is complete: the subordination relation tells us that ⊞kj j limj→∞ ωj(b) = limj→∞ F ⊞kj µ j (b) norm-uniformly on subsets D ⊂⊂ H+(B), so is freely limj→∞(µ infinitely divisible and the set of freely infinitely divisible operator-valued distri- butions is closed under taking norm-moment limits, by Proposition 2.11 we are done. (cid:3) . Since, as noted above, (µ j = limj→∞ µ )⊎1−k−1 )⊎1−k−1 ⊞kj j ⊞kj j j Theorem 3.4. Consider two infinitesimal triangular arrays {Xjk}j∈N,1≤k≤kj and {Yjk}j∈N,1≤k≤kj in (A, EB,B) so that Xjk are all free with amalgamation over B, Yjk are Boolean independent with amalgamation over B, and {Xjk : 1 ≤ k ≤ 18 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV kj} ∪ {Yjk : 1 ≤ k ≤ kj} have the same distribution with respect to EB for each j ∈ N (i.e. rows are identically distributed). The following are equivalent: vergence of moments); we call the limit distribution µX ; (a) limj→∞ Xj1 + Xj2 + ··· + Xjkj exists in distribution (as norm-limit con- (b) limj→∞ Yj1 + Yj2 + ··· + Yjkj exists in distribution (as norm-limit conver- gence of moments); we call the limit distribution µY . Moreover, the correspondence between the two limiting distributions is given ana- lytically by the fully matricial extension of the relation (22) FµX (b) = 1 2 (b + FµY (FµX (b))) b ∈ H+(B). The correspondence µX ↔ µY is the Boolean Bercovici-Pata bijection. It can be easily seen to be, as in the case of scalar-valued distributions, a morphism between ⊞ and ⊎. Proof. Let us assume that limj→∞ Xj1 +Xj2 +···+Xjkj exists. As seen in Theorem 3.2, µX is ⊞-infinitely divisible. In terms of the subordination function, this simply corresponds to the limit described in the proof of the previous theorem. On the other hand, if existing, FY (b) − b = limj→∞ kj (FYj1 (b) − b). Since we know that each of kj(b − FYj1 (b)) is itself a Voiculescu transform of a proba- kj +1 ), it is bility measure (namely of (µXj1+···+Xjkj +Xjkj +1) enough to show that the inverse of b + kj(FYj1 (b) − b) with respect to composition (which is ⊞-infinitely divisible), converges to FX uniformly on compacts. But this inverse is simply ωj from the previous theorem's proof. This completes the proof of one implication. ⊎1− 1 ) kj +1 = (µ ⊎1− 1 ⊞kj +1 Xj1 The converse is simpler. The statement that Yj1 +··· + Yjkj tends to Y in distribution as j → ∞ is equivalent to limj→∞ kj (FYj1 (b)−b) = FY (b)−b uniformly on compacts of H+(B). But then the expression of the Voiculescu transform of the distribution associated with ωj is simply ϕj(w) = (kj−1)(w−FYj1 (w)), w ∈ H+(B). Since convergence of Voiculescu transforms and convergence in distribution are equivalent, it follows that the distribution associated to ωj converges weakly. Since, as seen in the proof of the previous theorem, the distribution associated to ωj is simply the (1 − k−1 )th Boolean power of the distribution of Xj1 + ··· + Xjkj , it follows that Xj1 + ··· + Xjkj also converges in distribution to the same limit. This shows that limj→∞ Xj1 +··· + Xjkj exists in distribution and its reciprocal Cauchy transform's formula is the one indicated in (22). (cid:3) j As a by-product, we obtain the following corollary describing distributions which are the second power with respect to free additive convolution. Corollary 3.5. An operator-valued distribution µ is the nth power with respect to free additive convolution of an operator-valued distribution if and only if µ⊎1−1/n is freely infinitely divisible. n Proof. First implication has been noted in the proof of Theorem 3.2, where it is noted that (ν⊞n)⊎1−1/n must be freely infinitely divisible. The converse follows from if µ is freely infinitely divisible, then as µ⊎n/(n−1) satisfies Fµ⊎n/(n−1) (b) = [18]: n−1Fµ(b) − 1 n−1 b, we can use the definition of ϕµ to conclude Fµ⊎n/(n−1) (b + n−1 ϕµ(b). Applying F −1 ϕµ(b)) = b − 1 µ⊎n/(n−1) and using again the definition of ϕ gives ϕµ(b) = ϕµ⊎n/(n−1) (b − 1 n−1 ϕµ(b). Simple arithmetic gives n−1 ϕµ(b)) − 1 INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 19 1 n−1 ϕµ(b)) + b − 1 n ϕµ⊎n/(n−1) (b − 1 n−1 ϕµ(b) = b. All these relations are valid whenever ℑb has an inverse which is small enough. Analytic continuation gives F(µ⊎n/(n−1)) n−1 ϕµ(b). The description from [18] of freely infinitely divisible distributions in terms of their R-transform together with the relation be- tween Rµ and ϕµ guarantees that b − 1 n−1 ϕµ(b) is well-defined on all H+(Bnc) with imaginary part at least ℑb. (b) = b − 1 (cid:3) 1 n 4. C-free Bercovici-Pata bijection In this section we shall connect Boolean, free and conditionally free in- finitely divisible distributions both via their Hincin-type description as limits of triangular arrays and via explicit formulas linking their transforms. Our results will generalize the results of the previous section, but we will also use those results in our proofs. In some of the statements and proofs made below, we will use the non- commutative set of nilpotent elements of our C∗-algebras and noncommutative functions defined on it; we are aware that there exist many C∗-algebras that have no nilpotent elements except zero, however the fully matricial extensions of those elements (as noted in Section 2) are very rich. As we did before, we sometimes write the proofs using the first coordinate of our noncommutative maps and sets, but we do this only for the relative simplicity of notation: the noncommutative structure allows each proof to be re-written for any coordinate. Lemma 4.1. Let {kn}n be an increasing sequence of positive integers and {µn}n be a sequence of elements from ΣB:D. Then µ⊎kn norm-converges in moments to µ ∈ ΣB:D if and only if for all b ∈ Nilp(B) (23) n Bµ(b) = lim n→∞ kn · [Mµn (b) − 1]. Proof. Since for all b ∈ Nilp(B) we have that Mµ(b) − 1 ∈ Nilp(D) as noted in Section 2, it follows that Mµ(b) is invertible, so Bν(b) = [Mν(b) − 1] · Mν(b)−1 for any ν ∈ ΣB:D, hence the norm-convergence in moments of µ⊎kn (24) knBµn (b) = Bµ(b). for all b ∈ Nilp(B) lim n→∞ n is equivalent to Suppose first that (24) holds true. Then limn→∞ Bµn (b) = 1 kn Bµ(b) = 0, but [1 − Bµn (b)] · Mµn(b) = 1, therefore limn→∞ Mµn (b)−1 = 1, so lim n→∞ kn[Mµn (b) − 1] = lim n→∞ = lim n→∞ kn[Mµn (b) − 1]Mµn (b)−1 knBµn (b) = Bµ(b). limn→∞ For the converse, if (23) holds true, then we have limn→∞(Mµn (b) − 1) = Bµ(b) = 0, that is limn→∞ Mµn (b)−1 = 1, therefore 1 kn lim n→∞ kn[Mµn (b) − 1] = lim n→∞ = lim n→∞ kn[Mµn (b) − 1]Mµn (b)−1 knBµn (b). (cid:3) 20 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV The next lemma gives a similar characterization for the linearizing trans- forms of c-free convolution. Lemma 4.2. Let {kn}n be an increasing sequence of positive integers and (µ, ν) ∈ B:D × Σ0 Σ0 B, respectively {(µn, νn)}n, be an infinitesimal sequence from ΣB:D × ΣB. Then the following are equivalent: a. (µn, νn) c kn norm-converges in moments to (µ, ν) b. for all b ∈ Nilp(B) we have that Rν (b) = cRµ,ν (b) = lim n−→∞ lim n−→∞ kn · Mν(b) kn · Mµ(b). Proof. The implication"(b)⇒(a)" is a re-phrasing of Theorem 3.4 and the proof is identical. We give it here for the convenience of the reader. Recall the equation (b) − b = ϕν(b), for b ∈ H+(B) with kb−1k small. The convergence for the F −1 second coordinate comes to limn→∞ knϕνn (b) = ϕν(b). We shall (for now formally) replace b by Fνn (b). This gives us ν ϕν(Fνn (b)) = lim n→∞ knϕνn (Fνn (b)) = lim n→∞ kn(b − Fνn (b)) = lim n→∞ knBνn (b−1)b. We note that indeed we are indeed allowed to make the substitution and take limits by [9, Theorem 1.10], as Fνn (b) → b in norm, uniformly on closed balls inside any proper region of the upper half-plane, by Remark 2.12. same argument works for fully matricial extensions of the above maps. This together with the previous lemma and the equivalence of convergence on Nilp and H+ proves the first statement. The implication "(a)⇒(b)" is now direct, from the definition of cRµ,ν(b) as given in (15): (Mµ(b) − 1) · Mν(b) = Mµ(b)cRµ,ν(bMν(b)). We re-write this as kn · Bµn (b)b−1Hνn (b) = kn · cRµn,νn(Hνn (b)) and note again that Hνn (b) converges uniformly to b. Using the previous lemma, the linearizing property of cR and Remark 2.12, we conclude. (cid:3) As shown in [18] for each µ ∈ ΣB:D, there exist some selfadjoint γ ∈ B and some linear map σ : BhXi −→ D satisfying (1) such that (25) Bµ(b) =(cid:2)γ · 1 +eσ(cid:0)b(1 − X b)−1(cid:1)(cid:3) · b. whereeσ is the fully matricial extension of σ to BhhXii - the formal non-commutative power series with coefficients in B. Also, if (µ, ν) ∈ ΣB:D × ΣB is c -infinitely divisible, then there exist some selfadjoint γ0 ∈ B, γ1 ∈ D and some linear maps σ0 : BhXi −→ B, σ1 : BhXi −→ ΣB:D satisfying 1 such that (26) (27) Rν(b) =(cid:2)γ0 · 1 +fσ0(cid:0)b(1 − X b)−1(cid:1)(cid:3) · b cRµ,ν(b) =(cid:2)γ1 · 1 +fσ1(cid:0)b(1 − X b)−1(cid:1)(cid:3) · b B:D × Σ0 Moreover, if the moments of µ, ν do not grow faster than exponentially (that is (µ, ν) ∈ Σ0 Definition 4.3. We define the bijection BP from ΣB:D × ΣB to the set of all if Bµ, respectively Bν c -infinitely divisible elements of ΣB:D × ΣB as follows: B) then so do the moments of σ, σ0 and σ1 from above. INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 21 are determined as above by the pairs (γ, σ) and (γ0, σ0), then BP(µ, ν) is the c - infinitely divisible pair (µ′, ν′) such that cRµ′,ν ′ and Rν ′ are determined, as above, by (γ, σ), respectively (γ0, σ0). Remark 4.4. BP maps Σ0 Σ0 B:D × Σ0 BP(µ, ν), then B onto the c -infinitely divisible elements of B and µ′ is the first coordinate of B. Moreover, if (µ, ν) ∈ Σ0 B:D × Σ0 B:D × Σ0 hµ′ (b) = hµ(FBP (ν)(b)) (28) FBP(ν)(b) = 1 2 (b + Fν(FBP(ν)(b))), b ∈ H+(Bnc). Proof. The first assertion is an immediate consequence of Remark 5.5 from [18]; more precisely if the components of the R- and cR-transforms of ν and (µ, ν) do not grow faster than exponentially, then so do the moments of ν and µ. The second equation of the second assertion is simply equation (22); look- ing at RBP(ν)(b) = γ0 · 1 +fσ0(cid:0)(b−1 − X )−1(cid:1), ℑb < 0, we observe that ϕBP(ν)(b) = RBP(ν)(b−1) = γ0 · 1 +fσ0(cid:0)(b − X )−1(cid:1), ℑb > 0. We note that if ℑb−1 is small in norm, then ℑfσ0(cid:0)(b − X )−1(cid:1) is also small in norm; indeed, this has been noted in the proof of Proposition 2.11. Thus, when ℑb−1 is small enough, b + ϕBP (ν)(b) ∈ H+(Bnc). Using the expression of ϕ in terms of F , by replacing b with b + ϕBP(ν)(b) our equation is equivalent to which is trivially equivalent to 1 b = 2(cid:0)b + ϕBP (ν)(b) + Fν(b)(cid:1) , γ0 · 1 +fσ0(cid:0)(b − X )−1(cid:1) = −ϕBP(ν)(b) = Fν(b) − b = −Bν(b−1)b. Analytic continuation concludes the proof of the second equation. We note once again that this second equation together with Proposition 3.1 indicates that FBP(ν) is in fact simply the subordination function corresponding to the free additive convolution of ν with itself: Fν(FBP(ν)(b)) = Fν⊞ν(b). For the first equation let us simply rewrite the defining relation (15) for cR expressed in −hµ′ (b) = b − Fµ′(b) = γ1 · 1 +fσ1(cid:0)(FBP(ν)(b) − X )−1(cid:1) , terms of γ1 andfσ1 as As in the definition of the Bercovici-Pata bijection Bµ is given by γ1 and fσ1, this b ∈ H+(Bnc). is the claimed relation. We go now to the main result of this section. (cid:3) Theorem 4.5. Let (µn, νn)n be an infinitesimal sequence from ΣB:D × ΣB and {kn}n be an increasing sequence of positive integers. The following properties are equivalent: (1) (µn, νn)⊎kn norm-converges in moments to some (µ, ν) ∈ Σ0 (2) (µn, νn) c kn norm-converges in moments in Σ0 (3) (µn, νn) c kn norm-converges in moments to BP(µ, ν) for some (µ, ν) ∈ B:D × Σ0 B:D × Σ0 B B Σ0 B:D × Σ0 B 22 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV (4) There exist the pairs (γ, σ) and (γ0, σ0) determining Bµ and Bν as above such that for all m amd all b1, . . . , bm ∈ B: lim n−→∞ lim n−→∞ lim n−→∞ lim n−→∞ (µn)(X ) = γ (νn)(X ) = γ0 kn · µn(X b1X b2 ··· bmX ) = σ(b1X b2 ···X bm) kn · νn(X b1X b2 ··· bmX ) = σ0(b1X b2 ···X bm) Proof. Note that, by Theorem 3.4 ν⊎kn converges. The two limits will be called in this proof ν and ν⊞ respectively. We observe that by the previous remark and Theorem 3.4, ν⊞ = BP(ν). As shown in subsection 2.3, (µX1,n+···+Xkn ,n , νX1,n+···+Xkn ,n ) = (µn, νn) c kn can be expressed coordinatewise in terms of transforms: converges if and only if ν⊞kn n n hµX1,n +···+Xkn ,n (b) = knhµX1,n (ωn(b)) (c-free) ωn(b) = 1 kn b +(cid:16)1 − 1 kn(cid:17) FνX1,n +···+Xkn ,n (b) = 1 kn b +(cid:16)1 − 1 kn(cid:17)FνX1,n (ωn(b)) (free) and ν⊎1−k−1 n X1,n+···+Xkn ,n Recall from the proof of Theorem 3.4 that ωn = Fν converges to ν⊞. −1 n ⊎1−k X1,n +···+Xkn ,n (1) =⇒ (2): Assume first that (1) holds. Then, as seen above, ν⊞kn converges to µ, it was shown in Proposition 2.11 that (b) = hµ(b) uniformly on closed bounded subsets of the upper converges to ν⊞. Since µ⊎kn limn→∞ knhµX1,n half-plane which are at positive distance from ∂H+(B). Thus, n n lim n→∞ kn · hµX1,n (ωn(b)) = hµ(Fν⊞ (b)), b ∈ H+(B), and the limit is uniform on closed bounded subsets of the upper half-plane, as shown in [9, Theorem 1.10]. Same argument works for the fully matricial extensions, and we conclude that (2) holds. n X1,n+···+Xkn ,n (2) =⇒ (1): Next, assume that (2) holds. This clearly means that in relation (free) above we can take limits when n tends to infinity to obtain that both ν⊎1−k−1 and νX1,n+···+Xkn ,n converge to ν⊞. So ωn norm-converges to F ν⊞. We recall that ωn has as right inverse with respect to composition the (w), and by Theorem 3.4 and Proposition 2.11 function w 7→ knw + (1 − kn)FνX1,n this function converges uniformly on closed bounded sets of the upper half-plane H+(Bnc) to w − hν(w). So for ℑb sufficiently large in order for b − hν(b) to belong to H+(Bnc) + i1, lim n→∞ kn · hµX1,n (b) = lim n→∞ = lim n→∞ (ωn(knb + (1 − kn)FνX1,n kn · hµX1,n hµ(b − hν(b)) (b))) In the last equality we have used the assumption that kn · hµX1,n ◦ ωn converges to hµ uniformly on bounded closed sets of the upper half-plane and that the function (w) converges also uniformly to w 7→ w − hν(w). Since w 7→ knw + (1 − kn)FνX1,n we assumed b − hν(b) ∈ H+(Bnc) + i1, the equality follows. So we have proved (1) must hold. This shows the equivalence between (1) and (2). INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 23 The previous remark indicates that indeed the limit of (µn, νn) c kn as n tends to infinity, if existing, must be BP(µ, ν), thus establishing (2) ⇔ (3). Finally the equivalence of (4) to (1), (2) and (3) is a direct consequence of Lemma 4.2, the equivalence between norm-convergence on Nilp and H+ for the corresponding transforms, and the definition of the Bercovici-Pata bijection. (cid:3) Remark 4.6. We would like to come back once again to the issue of distributions in ΣB:D, or even in a larger set. As shown by Bercovici and Voiculescu [5], one can define free convolutions of probability measures with unbounded support, even when they have no moments whatsoever. In the same paper, they provide an operatorial representation, by showing that if A1,A2 are free in (A, ϕ), and Xj = X ∗ j are un- bounded operators affiliated with Aj and having distributions µXj , j = 1, 2, then ⊞µX2. X1+X2 is a selfadjoint unbounded operator affiliated to A and µX1+X2 = µX1 In the case of operator-valued distributions, to our best knowledge such a result does not exist as of now. However, it is very easy to observe that if (A, EB,B) is an op-valued noncommutative probability space and X = X ∗ is affiliated to A, then (b − X)−1 ∈ A for any b ∈ H+(B) (one only needs to use analytic functional calcu- lus), so EB[(b−X)−1] is well-defined. However, except when B is finite dimensional and A is a finite von Neumann algebra, we do not know whether any of the results of Voiculescu and Speicher (existence and good behavior of R-transform, subordi- nation etc) remains valid. When B is finite dimensional, one can make certain generalizations (for example one can talk of convergence of distributions in Σ0 B to a distribution in ΣB, or even one without moments - provided we define it appro- priately - and even obtain Theorem 3.2 for such a limit when the infinitesimality of an array is defined as convergence of GXij to b 7→ b−1). However, we repeat that our purpose in this paper was to provide results in maximum generality in terms of B. We shall postpone a detailed discussion of finite dimensional scalar algebras to future papers. 5. The homomorphism property of the Bercovici-Pata bijection In this section, we restrict ourselves to the scalar case, i.e. we consider B = D = C and in Definition 2.4 EB, θ to be a pair of states φ, ψ . Also, we will use the complex analytic R-, cR- and B-transforms1, not their non-commutative versions. For two pairs of distributions (φX , ψX ) and (φY , ψY ) given as in Definition 2.1 by the c-free random variables X and Y , we shall denote ψX ⊠ ψY for φXY and (φX , ψX ) ⊠c (φY , ψY ) for (φXY , ψXY ). In terms of transforms, if Tν(z) = z R−1(z) , cTµ,ν(z) = 1 R−1 ν (z) cRµ,ν(R−1 ν (z)) it has been shown (see [24], respectively [19]) that Tµ⊠ν (z) = Tµ(z)Tν(z) cT(µ1,ν1)⊠c(µ2,ν2)(z) = cT(µ1,ν1)(z) · cT(µ2,ν2)(z). 1The reader is warned that in scalar probability, the transform B is denoted by η, unlike in operator-valued probability; we have chosen not to change the notation because we use the operator-valued definitions and results already introduced. 24 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV (Recall the convention that f −1(z) denotes the compositional inverse of f evaluated 1 in z, while f (z)−1 means f (z) .) All these relations hold on a neighborhood of zero. It has been shown by Belinschi and Nica [3] that the Boolean Bercovici- Pata bijection is a homomorphism with respect to free multiplicative convolution, meaning that BP(µ ⊠ ν) = BP(µ) ⊠ BP(ν). The purpose of this section is to generalize this result to multiplicative c-free con- This observation turns out to be true also for the T -transform: volution. The main tool in [3] was the observation that SBP(µ)(z) = Sµ(cid:16) z 1−z(cid:17) . R t dν(t) 6= 0, we have Lemma 5.1. For any compactly supported pair of distributions (µ, ν) so that cTBP(µ,ν)(z) = cT(µ,ν)(cid:18) z 1 − z(cid:19) , for z in a neighborhood of zero. Proof. To simplify the notations, if σ is a compactly supported real measure, we will write Mσ(z) =Z tz 1 − tz dσ(z) = Mσ(z) − 1. z Bµ(z) = cR(µ,ν)(z+zMν (z)) Using equation (15) as re-written in the proof of Lemma 4.2, we obtain . Recall the defining relation for Rν as Mν(z) = ν (Mν(z)) = z + zMν(z), and that 1 Rν(z + zMν(z)); inverting Rν to the left gives R−1 inverting now Mν to the right yields z+zMν (z) R−1 ν (z) = M−1 ν (z)(1 + z), again for z in a neighborhood of zero. We compose to the right with M−1 a small enough neighborhood of the origin) in the relation that defined cR and use the above: assumption thatR t dν(t) 6= 0 implies that the compositional inverse does exist on (the ν (29) 1 M−1 ν (z) Bµ(M−1 ν (z)) = cR(µ,ν)(R−1 ν (z) R−1 ν (z)) = cT(µ,ν)(z). On the other hand, as shown in [3], RBP(ν)(z) = Bν (z), so cTBP(µ,ν)(z) = = 1 R−1 BP(µ)(z) 1 B−1 ν (z) cRBP(µ)(R−1 BP(ν)(z)) cRBP(µ)(B−1 ν (z)). Of course, by using the definition of B we obtain that B−1 our lemma is proved, since from Definition 4.3 we have that cRBP(µ) = Bµ. ν (z) = M−1 ν (cid:16) z 1−z(cid:17). Thus, (cid:3) Now the main result of the section follows: Proposition 5.2. For any compactly supported distributions (µ1, ν1), (µ2, ν2), we have BP((µ1, ν1) ⊠c (µ2, ν2)) = BP(µ1, ν1) ⊠c BP(µ2, ν2). INFINITE DIVISIBILITY, NON-COMMUTATIVE BERCOVICI-PATA BIJECTION 25 Proof. For the second coordinate, this has been proved in [3]. For the first coordi- nate, we use the previous lemma to write 1 − z(cid:19) cTBP((µ1,ν1)⊠c(µ2,ν2))(z) = cT(µ1,ν1)⊠c(µ2,ν2)(cid:18) z = cT(µ1,ν1)(cid:18) z 1 − z(cid:19) cT(µ2,ν2)(cid:18) z 1 − z(cid:19) = cTBP((µ1,ν1))(z)cTBP((µ2,ν2))(z) = cTBP(µ1,ν1)⊠cBP(µ2,ν2)(z). (cid:3) We would like to mention, however, that regrettably the c-free convolution of positive probability measures on the positive half-line is not necessarily well- defined, while the Boolean - or c-free - Bercovici-Pata bijection is not well defined for measures on the unit circle, as sums of unitaries are not unitaries. Thus, the above result must be viewed in terms of algebraic distributions. References [1] Akhieser, N. I.The classical moment problem and some related questions in analysis. Trans- lated by N. Kemmer. Hafner Publishing Co., New York, 1965. [2] Belinschi, S. T. C-free convolution of measures with unbounded support. C∗-algebras in Sibiu, 1 -- 7. Theta 2008. [3] Belinschi, S. T.; Nica, A. On a remarkable semigroup of homomorphisms with respect to free multiplicative convolution. Indiana Univ. Math. J. [4] Bercovici, H.; Pata, V. Stable laws and domains of attraction in free probability theory, With an appendix by Philippe Biane, Ann. of Math. (2), 149, (1999), no.3, 1023 -- 1060 [5] Bercovici, H.; Voiculescu, D. Free convolutions of measures with unbounded support. Indiana Univ. Math. J. 42 (1993), no. 3, 733 -- 773. [6] Blanchard, E.; Dykema, K. Embeddings of reduced free products of operator algebras Pacific Journal of Mathematics 199 (2001), 1-19 [7] Boca, Florin. Free products of completely positive maps and spectral sets, J. Functional Anal- ysis 97 (1991), 251-263 [8] Bozejko, Marek; Leinert, Michael; Speicher, Roland. Convolution and limit theorems for conditionally free random variables. Pacific J. Math. 175, (1996), no. 2, 357 -- 388. [9] Isidro, Jos´e M.; Stach´o, L´aszl´o. Holomorphic automorphism groups in Banach spaces: an elementary introduction. North-Holland Mathematics Studies 105, (1985). [10] Kaliuzhnyi-Verbovetskyi D. S. and Vinnikov, Victor. Foundations of noncommutative func- tion theory, preprint [11] Krysztek, Anna Dorota. Infinite divisibility for the conditionally free convolution. IDAQP Vol 10(4) (2007), 499 -- 522. [12] Muraki, Naofumi Monotonic independence, monotonic central limit theorem and monotonic law of small numbers, IDAQP Vol. 4(1) (2001) pp. 39-58 [13] Nica, Alexandru; Speicher, Roland. Lectures on the combinatorics of free probability. Cam- bridge University Press, 2006. [14] Paulsen, Vern. Completely bounded maps and operator algebras. Cambridge Studies in Ad- vanced Mathematics 78, Cambridge University Press, Cambridge, UK, 2002. [15] Popa, Mihai. A combinatorial approach to monotonic independence over a C-algebra, Pac J. of Math, Vol. 237 (2008), No. 2, 299-325 [16] Popa, Mihai. A new proof for the multiplicative property of the boolean cumulants with ap- plications to the operator-valued case Colloquium Mathematicum, Vol 117 (2009), No. 1 , 81-93 [17] Popa, Mihai. Multilinear function series in conditionally free probability with amalgamation Com. on Stochastic Anal., Vol 2, No 2 (Aug 2008) 26 S. T. BELINSCHI, M. POPA, AND V. VINNIKOV [18] Popa, Mihai; Vinnikov, Victor. Non-Commutative Functions and Non-Commutative Free Levy-Hincin Formula, preprint [19] Popa, Mihai; Wang, Jiun-Chau. On multiplicative conditionally free convolution, Trans. Amer. Math. Soc., to appear; available also on arXiv:0805.0257 [20] Speicher, Roland. Combinatorial theory of the free product with amalgamation and operator- valued free probability theory. Memoirs of the Amer. Math. Soc. 132 (1998), x+88 [21] Speicher, Roland; Woroudi, Reza. Boolean convolution. Fields Institute Communications, Vol. 12 (D. Voiculescu, ed), AMS, 1997, 267 -- 279. [22] J.L. Taylor, Functions of several noncommuting variables, Bull. Amer. Math. Soc. 79 (1973), 1 -- 34. [23] Voiculescu, D. V. Addition of certain non-commutative random variables, J. Funct. Anal. 66 (1986), 323 -- 346. [24] D. Voiculescu, Multiplication of certain noncommuting random variables, J. Operator Theory 18(1987), 223 -- 235. [25] Voiculescu, D. V. Operations on certain non-commutative operator-valued random variables, Ast´erisque (1995), no. 232, 243 -- 275. [26] Voiculescu, D. V. The coalgebra of the free difference quotient and free probability. Internat. Math. Res. Not. (2000) No. 2. [27] Voiculescu, D. V. Free Analysis Questions I: Duality Transform for the Coalgebra of ∂X :B Internat. Math. Res. Not. (2004), No. 16. [28] Wang, J.-C. Limit theorems for additive c-free convolution, preprint arXiv:0805.0607 Department of Mathematics and Statistics, University of Saskatchewan, 106 Wiggins Road, Saskatoon, SK, S7N 5E6, CANADA, and Institute of Mathematics "Simion Stoilow" of the Romanian Academy. E-mail address: [email protected] Center for Advanced Studies in Mathematics at the Ben Gurion University of Negev, P.O. B. 653, Be'er Sheva 84105, Israel, and Institute of Mathematics "Simion Stoilow" of the Romanian Academy, P.O. Box 1-764, Bucharest, RO-70700, Romania E-mail address: [email protected] Department of Mathematics, Ben Gurion University of Negev, Be'er Sheva 84105, Israel E-mail address: [email protected]
1608.03143	1	1608	2016-08-10T11:52:29	Some notes on Morita equivalence of Operator Algebras	[ "math.OA" ]	In this paper we present some key moments in the history of Morita equivalence for operator algebras.	math.OA	math	Some notes on Morita equivalence of operator algebras G.K. Eleftherakis Abstract In this paper we present some key moments in the history of Morita equivalence for operator algebras. We briefly describe some events in the history of Morita equivalence for operator algebras. We make no claim of completeness in this presentations. Also, we have modified some of the definitions so as to be compatible with our presentation. The reader can use the books [2, 13] for the notions of operator space theory used in this article. In what follows if H, K are Hilbert spaces, we denote by B(H, K) the space of bounded operators from H to K and we write B(H) for B(H, H). 1 Morita equivalence of C ∗-algebras M. A. Rieffel introduced the idea of Morita equivalence of C ∗ and W ∗ al- gebras. The definitions and results of this section can be deduced from [5, 15, 16, 17]. Let H, K be Hilbert spaces and T be a subspace of B(H, K). The space T is called a ternary ring of operators (TRO) if T T ∗T ⊂ T . Definition 1.1 Let A be a C ∗ algebra. We denote by AHM the category of representations of A : The objects of AHM are pairs (π, H) where H is a Hilbert space and π : A → B(H) 1 is a ∗−homomorphism such that π(A)(H) = H. If (πi, Hi) ∈ AHM, i = 1, 2 the corresponding space of category homorphisms is the following: BA(H1, H2) = {x ∈ B(H1, H2) : xπ1(a) = π2(a)x ∀ a ∈ A}. Definition 1.2 Let A, B be C ∗ algebras. These algebras are called strongly Morita equvalent in the sense of Rieffel (R-SME), if there exist 1-1 ∗−homomorphisms π : A → B(H), ρ : B → B(K) where H, K are Hilbert spaces and there exists a TRO T ⊂ B(H, K) such that π(A) = [T ∗T ]−k·k, ρ(B) = [T T ∗]−k·k. Theorem 1.1 Let A, B be R-SME C ∗ algebras. Then the categories AHM and BHM are equivalent. Definition 1.3 Let K be the space of compact operators acting on an in- finite dimensional separable Hilbert space. The C ∗ algebras A and B are called strongly stably isomorphic if the C ∗ algebras A ⊗ K and B ⊗ K are C ∗ isomorphic. (The tensor product ⊗ is the spatial tensor product, see [2]). We can easily see that if A and B are strongly, stably isomorphic then they are R-SME. The converse doesn't hold. But if A and B possess countable approximate identities the converse is true. In particular, if A and B are separable or unital, then they are R-SME if and only if they are strongly stably isomorphic. 2 Morita equivalence of W ∗-algebras Definition 2.1 Let M be a W ∗ algebra. We denote by M NHM the category of normal representations of M : The objects of M NHM are pairs (π, H) where H is a Hilbert space and π : M → B(H) is a weak* continuous ∗−homomorphism such that the identity operator be- longs to π(M). If (πi, Hi) ∈ M NHM, i = 1, 2 the corresponding space of category homorphisms is the following: BM (H1, H2) = {x ∈ B(H1, H2) : xπ1(a) = π2(a)x ∀ a ∈ M}. 2 Definition 2.2 Let M, N be W ∗ algebras. These algebras are called weakly Morita equvalent in the sense of Rieffel (R-WME), if there exist 1-1 weak* continuous ∗−homomorphisms π : M → B(H), ρ : N → B(K) where H, K are Hilbert spaces and there exists a TRO T ⊂ B(H, K) such that π(M) = [T ∗T ]−w∗ , ρ(N) = [T T ∗]−w∗ . Definition 2.3 Let M, N be a W ∗ algebras and F : M NHM → N NHM be a functor. This functor is called a ∗-functor if for every pair of objects H1, H2 ∈ M NHM and every x ∈ BM (H1, H2), the map F (x∗) is equal to F (x)∗. Theorem 2.1 Let M, N be W ∗ algebras. The following are equivalent: (i) The algebras M and N are R-WME. (ii) The categories M NHM, N NHM are equivalent through a ∗−functor. (iii) There exists a cardinal I such that the algebras MI(M), MI (N) of all I × I matrices whose finite submatrices have uniformly bounded norms are isomorphic as W ∗ algebras. (iv) There exist faithful normal representations π of M and ρ of N such that the commutants π(M)′ and ρ(N)′ are isomorphic. The equivalence of (i) and (ii) was proved in [14]. The others can be found in [2]. 3 Strong Morita equivalence of operator al- gebras. In this section we assume that all operator algebras (see [2, 13]) have contrac- tive approximate identities (cai). The definitions and results of this section can be deduced from [1, 3]. 3 Definition 3.1 Let A be an operator algebra and X be an operator space. We say that X is a left A-operator module if there exists a completely con- tractive bilinear map Similarly we can define right A-operator modules. A × X → X. Definition 3.2 Let A be an operator algebra. We denote by AOM the category of left operator modules: The objects of AOM are the left operator modules X for which eix → x ∀ x ∈ X, where (ei)i is a cai for the algebra A. If X1, X2 ∈ AOM, the corresponding space of category homorphisms is the following: CBA(X1, X2) = {u : X1 → X2 : u is a compl. bounded A-module map }. Definition 3.3 Let A, B be operator algebras. A functor F : AOM → BOM is called completely contractive if the map x → F (x) is completely contractive for all pairs X1, X2 ∈ AOM and all x ∈ CBA(X1, X2). Definition 3.4 Let A, B be operator algebras. These algebras are called strongly Morita equivalent, in the sense of Blecher Muhly and Paulsen, (BMP- SME) if there exist completely isometric homomorphisms π : A → B(H), ρ : B → B(K), where H, K are Hilbert spaces such that π(A)(H) = H, ρ(B)(K) = K and such that there exist a (ρ(B), π(A))-module U and a (π(A), ρ(B))-module V satisfying π(A) = [V U]−k·k, ρ(B) = [UV ]−k·k. We also require the existence of nets of positive integers (nt)t, (mi)i and nets (v1 t ) ⊂ Ball(Rnt(V )), (u1 t ) ⊂ Ball(Cnt(U)) i ) ⊂ Ball(Rmi(U)), (u2 t )t is an approximate identity for π(A) and (u2 i ) ⊂ Ball(Cmi(V )) (v2 such that (v1 proximate identity for ρ(B). t u1 i v2 i )i is an ap- 4 We recall (see [2, 13]) the Haagerup tensor product U ⊗hV of two operator spaces U and V. This space is an operator space and it has the property of linearizing completely bounded bilinear maps. Suppose U is a right A operator module and V is a left A operator module over an operator algebra A. We denote by Ω the space [u ⊗ (av) − (ua) ⊗ v : u ∈ U v ∈ V a ∈ A]−k·k and by U ⊗h A V the quotient (U ⊗h V )/Ω. This space has the property of linearizing A-balanced completely bounded bilinear maps. This means that if X is an operator space and φ : U × V → X is a completely bounded bilinear map satisfying φ(ua, v) = φ(u, av) ∀ u ∈ U, v ∈ V, a ∈ A then there exists a completely bounded linear map such that φ : U ⊗h A V → X φ(u ⊗A v) = φ(u, v). If B is an operator algebra and U (resp. V ) is a left (resp. right) B operator module then U ⊗h A V is a left (resp. right) B operator module. Theorem 3.1 Let A, B be operator algebras. The following are equivalent: (i) The algebras A and B are BMP-SME. (ii) The categories AOM, BOM are equivalent through a completely contractive functor. (iii) There exist a (B, A) operator module U and an (A, B) operator mod- B U and A) are equivalent as B A V and B (resp. V ⊗h ule V such that U ⊗h (resp. A) operator modules. Theorem 3.2 Two C ∗ algebras are R-SME if and only if they are BMP- SME. 5 4 Weak* Morita equivalence of dual operator algebras. In this section we assume that all dual operator algebras (see [2]) are unital. If M is a dual operator algebra, a normal representation of M is a map α : M → B(H), where H is a Hilbert space and α is a completely contractive weak* continuous homomorphism. The definitions and results of this section can be deduced from [4, 12]. Definition 4.1 Let M be a dual operator algebra and X be a dual operator space. We say that X is a left dual M-operator module if there exists a completely contractive weak* continuous bilinear map M × X → X. Similarly we can define right dual operator modules over M. Definition 4.2 Let M be a dual operator algebra. We denote by M NOM the category of left dual operator modules: The objects of M NOM are the = X. If X1, X2 ∈ M NOM left dual operator modules X for which MX the corresponding space CBσ M (X1, X2) of category morphisms is the space of all weak* continuous completely bounded M-module maps from X1 into X2. w∗ Definition 4.3 Let M, N be dual operator algebras. These algebras are called weakly* Morita equivalent in the sence of Blecher and Kashyap (BK- WME) if there exist completely isometric normal unital representations π : M → B(H), ρ : N → B(K), where H, K are Hilbert spaces such that there exist a (ρ(N), π(M)) bimodule U ⊂ B(H, K) and a (π(M), ρ(N)) bimodule V ⊂ B(K, H) satisfying π(M) = [V U]−w∗ , ρ(N) = [UV ]−w∗ . Also we require the existence of nets of positive integers nt, mi and nets (v1 t ) ⊂ Ball(Rnt(V )), (u1 t ) ⊂ Ball(Cnt(U)) (u2 i ) ⊂ Ball(Rmi(U)), i v2 t )t and (u2 such that the nets (v1 identity operators. t u1 (v2 i ) ⊂ Ball(Cmi(V )) i )i converge in the weak* topology to the 6 We recall (see [2]) the normal Haagerup tensor product U ⊗σh V of the dual operator spaces U and V. This space is a dual operator space and it has the property of linearizing weak* continuous completely bounded bilinear maps. Suppose U is a right M dual operator module and V is a left M dual operator module over a dual operator algebra M. We denote by Ωσ the space [u ⊗ (av) − (ua) ⊗ v : u ∈ U v ∈ V a ∈ A]−w∗ and by U ⊗σh M V the quotient (U ⊗σh V )/Ωσ. This space has the property of linearizing M-balanced weak* continuous completely bounded bilinear maps [11]. This means that if X is a dual operator space and φ : U × V → X is a weak* continuous completely bounded bilinear map satisfying φ(ua, v) = φ(u, av) for all u ∈ U, v ∈ V and a ∈ A then there exists a weak* continuous completely bounded linear map such that φ : U ⊗σh M V → X φ(u ⊗M v) = φ(u, v). If N is an operator algebra and the above U, (resp. V ) is a left (resp. right ) N operator module then U ⊗σh M V is a left (resp. right ) N operator module [11]. Definition 4.4 Let N, M be dual operator algebras. A functor F : M NOM → N NOM is called completely contractive and normal if the map F : CBσ M (X1, X2) → CBσ N (F (X1), F (X2)) is completely contractive and weak* continuous for all X1, X2 ∈ M NOM. 7 Theorem 4.1 Let M, N be dual operator algebras. The following are equiv- alent: (i) The algebras M and N are BK-WME. (ii) The categories M NOM, N NOM are equivalent through a completely contractive and normal functor. (iii) There exist a (N, M) dual operator module U and an (M, N) dual N U and M) are M V and N (resp. V ⊗σh operator module V such that U ⊗σh equivalent as N (resp. M) dual operator bimodules. Theorem 4.2 Two W ∗ algebras are R-WME if and only if they are BK- WME. Theorem 4.3 Let A, B be BMP-SME approximately unital operator alge- bras. Then their second duals A∗∗, B∗∗ are BK-WME algebras. Theorem 4.4 [10] Let N1, N2 be nests (see [6]) acting on the Hilbert spaces H1, H2 respectively, let M1, M2 be the corresponding nest algebras and let K(M1), K(M2) be the subalgebras of compact operators. The following are equivalent: (i) The algebras M1, M2 are BK-WME. (ii) The algebras K(M1), K(M2) are BMP-SME. (iii)The nests N1, N2 are isomorphic. (iv) There exist an (M1, M2) bimodule U ⊂ B(H2, H1) and an (M2, M1) bimodule V ⊂ B(H1, H2) such that M1 = [UV ]−w∗ , M2 = [V U]−w∗ . 5 Stable isomorphism of dual operator alge- bras. Definition 5.1 [9] Let M, N be weak* closed algebras acting on the Hilbert spaces H and K respectively. We call them TRO equivalent if there exists a TRO T ⊂ B(H, K) such that M = [T ∗NT ]−w∗ N = [T MT ∗]−w∗ . Theorem 5.1 [9] Let L1, L2 be reflexive lattices and let M = Alg(L1), N = Alg(L2) be the corresponding algebras of operators leaving invariant every 8 element of L1 and L2 (see [6]); also let ∆(M) = M ∩M ∗ and ∆(N) = N ∩N ∗ be their diagonals. The algebras M and N are TRO equivalent if and only if there exists a ∗-isomorphism θ : ∆(M)′ = L′′ 1 → ∆(N)′ = L′′ 2 such that θ(L1) = L2. Definition 5.2 Let M, N be dual operator algebras. We call them ∆-equivalent if there exist completely isometric normal representations α of M and β of N such that α(M) and β(N) are TRO equivalent. As in definition 2.1 if M is a unital dual operator algebra, we denote by M NHM the category of normal completetely contractive representations of M: If (Hi, αi), i = 1, 2 are objects of M NHM, the space of morphisms HomM (H1, H2) is the following: HomM (H1, H2) = {x ∈ B(H1, H2) : xα1(m) = α2(m)x ∀ m ∈ M}. Let ∆(M) = M ∩ M ∗ be the diagonal of M. Observe that αi∆(M ) is a ∗−homomorphism since αi is a contraction. We also define the category M DNHM which has the same objects as M NHM but for every pair of objects (Hi, αi), i = 1, 2 the space of morphisms is the following: HomD M (H1, H2) = {x ∈ B(H1, H2) : xα1(m) = α2(m)x ∀ m ∈ ∆(M)}. Observe that HomM (H1, H2) ⊂ HomD M (H1, H2). We say that F : M DNHM → M DNHM is a ∗-functor if for every pair H1, H2 ∈ M DNHM F (x∗) = F (x)∗ ∀ x ∈ HomD M (H1, H2). The main theorem of this section, which is a generalization of Theorem 2.1, is the following: Theorem 5.2 [7, 9, 11] Let M, N be unital dual operator algebras. The following are equivalent: (i) The algebras M and N are ∆−equivalent . 9 (ii) There exists a ∗-functor of equivalence such that F : M DNHM → N DNHM F ( M NHM) = M NHM. (iii) The algebras M and N are stably isomorphic in the following sense: there exists a cardinal I such that the algebras MI(M), MI (N) of I ×I matri- ces whose finite submatrices have uniformly bounded norms are isomorphic as dual operator algebras. (iv) There exist completely isometric normal representations α of M and β of N such that α(M) = Alg(L1), β(N) = Alg(L2) for reflexive lattices L1 and L2 and there exists a ∗−isomorphism θ : α(∆(M))′ → β(∆(N))′ mapping L1 onto L2. Theorem 5.3 [8] Let M, N be CSL algebras (see [6]). Then M and N are ∆-equivalent if and only if they are TRO equivalent. Theorem 5.4 [8] Let M, N be ∆-equivalent dual operator algebras. For ev- ery completely isometric normal representation α of M there exists a com- pletely isometric normal representation β of N such that the algebras α(M) and β(N) are TRO equivalent. Theorem 5.5 [7] Two W ∗ algebras are ∆-equivalent if and only if they are R-WME. Theorem 5.6 [4] If two dual operator algebras are ∆-equivalent then they are BK-WME. We are going to prove that ∆-equivalence is strictly stronger than BK- WME: Theorem 5.7 [7, 8, 10] Two BK-WME algebras are not necessarily stably isomorphic. 10 Proof Let {qn : n ∈ N} be an enumarations of the rationals. We define the measure µ = Pn δqn on the Borel σ-algebra of R, where δqn is the Dirac measure. Let H = L2(R, µ) and let Q+ t ) be the projection onto L2((−∞, t], µ) (resp. L2((−∞, t), µ) ). The set t (resp. Q− N1 = {Q+ t , Q− t : t ∈ R} is a totally atomic nest in H. Let λ be Lebesgue measure on R and let Nt be the projection onto L2((−∞, t], λ). We define the nest N2 = {Q+ t ⊕ Nt, Q− t ⊕ Nt : t ∈ R} t to Q− t ⊕ Nt and Q− acting on H ⊕ L2(R, λ). The map θ : N1 → N2 sending the projection Q+ t to Q+ t ⊕ Nt is a nest isomorphism [6]. Therefore, by Theorem 4.4, the nest algebras M1 = Alg(N1) and M2 = Alg(N2) are BK- WME. Suppose now that they are stably isomorphic. By Theorems 5.2 amd 5.3, M1 and M2 should be TRO equivalent. So by Theorem 5.1 there would exist a ∗-isomorphism σ : N ′′ 1 → N ′′ 2 such that σ(N1) = N2. But N ′′ algebra (masa) and N ′′ a contradiction. So M1 and M2 are not stably isomorphic. 1 is a totally atomic maximal abelian selfadjoint 2 is a masa with a nontrivial continuous part. This is (cid:3) References [1] D.P. Blecher, A Morita theorem for algebras of operators on Hilbert space, J. Pure Appl. Algebra 156 (2001), 153 -- 169. [2] Blecher, David P. ; Le Merdy, Christian. Operator algebras and their modules -- an operator space approach. London Mathematical Society Monographs. New Series, 30. Oxford Science Publications. The Claren- don Press, Oxford University Press, Oxford, 2004. x+387 pp. ISBN: 0-19-852659-8 [3] D.P. Blecher, P.S. Muhly, V.I. Paulsen, Categories of operator modules - Morita equivalence and projective modules, Memoirs of the A.M.S. 143 (2000) No 681. 11 [4] D.P. Blecher and U. Kashyap, Morita equivalence of dual operator al- gebras, J. Pure Appl. Algebra , 212 (2008), 2401 -- 2412. [5] L.G. Brown, P. Green and M. A. Rieffel, Stable isomorphism and strong Morita equivalence of C ∗-algebras , Pacific J. Math. 71 (1977), 349 -- 363. [6] K. R. Davidson, Nest Algebras, Longman Scientific & Technical, Harlow, 1988. [7] G.K. Eleftherakis, A Morita type equivalence for dual operator algebras, J. Pure Appl. Algebra, 212 (2008), no5, 1061 -- 1071. [8] G.K. Eleftherakis, Morita type equivalence and reflexive algebras, J. Operator Theory , 64 (2010), no 1, 3 -- 17. [9] G.K. Eleftherakis, TRO equivalent algebras, Houston J. of Mathematics, 38 (2012), no1, 153 -- 175. [10] G.K. Eleftherakis, Morita equivalence of nest algebras, Math. Scand. (to appear) Arxiv: 1002.2335. [11] G.K. Eleftherakis, V.I. Paulsen, Stably isomorphic dual operator alge- bras, Math. Ann. 341 (2008), no 1, 99-112. [12] U. Kashyap, A Morita theorem for dual operator algebras, J. Funct. Analysis, 256 (2009), 3545 -- 3567. [13] V.I. Paulsen, Completely Bounded Maps and Operator Algebras, Cam- bridge Studies in Advanced Math. 78, Cambridge University Press, Cambridge, 2002. [14] M. A. Rieffel, Morita equivalence for C ∗-algebras and W ∗-algebras, J. Pure Appl. Algebra 5 (1974), 51 -- 96. [15] M. A. Rieffel, Morita equivalence for operator algebras. Operator alge- bras and applications, Part I (Kingston, Ont., 1980), pp. 285 -- 298, Proc. Sympos. Pure Math., 38, Amer. Math. Soc., Providence, R.I., 1982. [16] M. A, Rieffel, Induced representations of C ∗-algebras , Adv. in Math. 13 (1974), 176 -- 257. 12 [17] M. A. Rieffel, Strong Morita equivalence of certain transformation group C ∗-algebras , Math. Ann. 222 (1976), 7 -- 22. G.K. ELEFTHERAKIS Department of Mathematics Faculty of sciences University of Patras 265 04 Patras, Greece email: [email protected] 13
1506.01272	1	1506	2015-06-03T15:05:27	Classification of C*-algebras generated by representations of the unitriangular group $UT(4,\mathbb{Z})$	[ "math.OA", "math.KT" ]	It was recently shown that each C-algebra generated by a faithful irreducible representation of a finitely generated, torsion free nilpotent group is classified by its ordered K-theory. For the three step nilpotent group $UT(4,\mathbb{Z})$ we calculate the ordered K-theory of each C-algebra generated by a faithful irreducible representation of $UT(4,\mathbb{Z})$ and see that they are all simple A$\mathbb{T}$ algebras. We also point out that there are many simple non A$\mathbb{T}$ algebras generated by irreducible representations of nilpotent groups.	math.OA	math	CLASSIFICATION OF C-ALGEBRAS GENERATED BY REPRESENTATIONS OF THE UNITRIANGULAR GROUP UT (4, Z) CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY Abstract. It was recently shown that each C-algebra generated by a faithful irreducible representation of a finitely generated, torsion free nilpotent group is classified by its ordered K-theory. For the three step nilpotent group U T (4, Z) we calculate the ordered K-theory of each C-algebra generated by a faithful irreducible representation of U T (4, Z) and see that they are all simple AT algebras. We also point out that there are many simple non AT algebras generated by irreducible representations of nilpotent groups. 1. Introduction The last few years witnessed several breakthroughs in the theory of simple, nuclear C-algebras. The hands of Rørdam, H. Lin, Z. Niu, Winter, Matui and Sato joined to show that if a C-algebra satisfies several abstract properties (see Theorem 2.1) then it necessarily has the concrete property of being an approximately subhomogenous algebra and is moreover classified1 by its Elliott invariant. In other words, if a C-algebra A satisfies Theorem 2.1 then it is an inductive limit of subhomogeneous C-algebras -- but knowledge of the Elliott invariant of A provides (in theory) an explicit decomposition of A as a limit of subhomogenous C-algebras and therefore a wealth of information about its structure. This provides an entire new avenue of study for those important classes of C-algebras that do not have an obvious inductive limit structure, e.g. C-algebras produced by dynamical systems or group representations. We travel this new avenue of study by calculating the Elliott invariant of C-algebras generated by faithful irreducible representations of the three step nilpotent group UT (4, Z). The odd-numbered authors showed that if Γ is a finitely generated torsion free nilpotent group and π is a faithful irreducible representation of Γ, then C∗(π(Γ)) satisfies Theorem 2.1. Therefore our present results combined with Theorem 2.3 de- termine the structure of C-algebras generated by faithful irreducible representations of UT (4, Z). Our current program was carried out (without the aid of Theorem 2.1) for many two-step nilpotent groups. Perhaps the best known example is Elliott and Evans's work on the irrational rotation algebras in [10]. They showed that the C-algebras generated by faithful irreducible representations of the discrete Heisenberg group H3 are AT algebras. The Elliott invariant had long been known for these algebras by the work of Rieffel and Pimsner and Voiculescu [22, 23, 25]. 1See Definition 2.2 for our working definition of "classified" 1 2 CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY It follows from the main theorem in Phillips's preprint [21] that the C-algebra generated by a faithful irreducible representation of a finitely generated two-step nilpotent group is an AT algebra. This is essentially the reason that we focus on the three step nilpotent group UT (4, Z) as it is the least complicated, most natural group covered by Theorem 2.3, but not by Phillips's theorem. We briefly explain the technical aspects of our calculations. The C-algebras gen- erated by faithful irreducible representations of UT (4, Z) are parameterized by the irrational numbers in (0, 1). For α ∈ (0, 1) and irrational let us denote by Bα the C-algebra generated by this representation. The Pimsner-Voiculescu six term exact sequence and a straightforward application of a theorem of Packer and Raeburn [19] show that Ki(Bα) ∼= Z10 for i = 0, 1. The bulk of our work is then devoted to divining the order structure on K0(Bα). We do this by locating a well-behaved (with respect to order structure) finite index subgroup G ≤ UT (4, Z) and applying Pimsner's [24] to show that the order on K0(Bα) is determined by the representation restricted to G. In particular we show the order on K0(Bα) is given by the hyperplane with normal vector (1, α, α2, 0, ..., 0). With the K-theory of each Bα in hand, in Section 6, we address when Bα ∼= Bβ. It is fairly easy to see that if α is transcendental or algebraic with minimal polynomial of degree greater than or equal to five, we have Bα ∼= Bβ if and only if α = ±β mod Z. On the other hand, if the degree of the minimal polynomial for α is less than or equal to four the situation is much more interesting (see Theorem 6.2 and following examples) and contrasts with the case of the irrational rotation algebras. An extremely crude summation of the works [22,23,25] is "two irrational rotation algebras are isomorphic if and only if they are obviously isomorphic." Theorem 6.2 shows cases with Bα ∼= Bβ that are not obviously isomorphic, i.e. the classification theorem (Theorem 2.1) is essential. As mentioned above, Phillips showed that every C-algebra generated by an ir- reducible representation of a finitely generated two-step nilpotent group is an AT algebra. All of the algebras considered here also turn out to be AT algebras. For the sake of completeness we finish the paper with Section 7 by pointing out that there are many C-algebras generated by faithful irreducible representations of three-step nilpotent groups that are not AT algebras. 2. Preliminaries In this section we define our objects of study and recall the necessary C-algebraic background. For information on the properties of AT algebras we refer the reader to Rørdam's monograph [27]. The following major theorem is crucial to our investiga- tions. Theorem 2.1 (Rørdam, H. Lin, Z. Niu, Winter, Matui and Sato [13 -- 16, 28, 33]). Let A and B be unital, separable, simple, nuclear, quasidiagonal C-algebras with unique tracial states and finite nuclear dimension that satisfy the universal coefficient theorem. Then A is an approximately subhomogeneous C-algebra. Moreover, if (K0(A), K0(A)+, [1A], K1(A)) ∼= (K0(B), K0(B)+, [1B], K1(B)), then A ∼= B. UNITRIANGULAR GROUPS 3 Definition 2.2. In this paper, when we say a C-algebra is classifiable we mean that it satisfies the hypotheses of Theorem 2.1. In reality, the term "classifiable" refers to a much larger class of C-algebras. We sacrificed generality for clarity in our definition. See the recent preprint [11] for a much more general definition of classifiable. Theorem 2.3 (See [5,6]). Let Γ be a torsion free finitely generated nilpotent group and π a faithful irreducible representation of Γ. Then C∗(π(Γ)) is classifiable (Definition 2.2). Definition 2.4. For a group Γ, let Z(Γ) ≤ Γ denote the center of Γ. For a, b ∈ Γ, set [a, b] = aba−1b−1. The unitriangular subgroup of GL(4, Z) is defined as UT (4, Z) = a = 1 1 a12 a13 a14 a23 a24 a34 1 1  : aij ∈ Z . The following subgroup of UT (4, Z) plays a key role in our calculations, H4 = {a ∈ UT (4, Z) : a23 = 0}. For each 1 < i < j ≤ 4, let eij ∈ M4(Z) be the (i, j)-matrix unit. One easily verifies the following commutation relations: (2.1) (2.2) Notice that (2.3) [1 + eij, 1 + ekℓ] = 1 + δjkeiℓ − δiℓekj. Z(UT (4, Z)) =  τθ(x) =(cid:26) e2πixθ 0 1 0 0 a 1 0 0 1 0 1  : a ∈ Z ∼= Z. if x ∈ Z(UT (4, Z)) if x 6∈ Z(UT (4, Z)) . 2.1. Representation-theoretic description. Definition 2.5. Let θ ∈ R. Define the trace on UT (4, Z) as follows: Let πθ denote the GNS representation of UT (4, Z) associated with τθ. Let π be an irreducible, faithful, unitary representation of UT (4, Z). It is well known (see [3, 12, 17] or the introduction of [5]) that there is an irrational θ ∈ R such that C∗(π(UT (4, Z))) ∼= C∗(πθ(UT (4, Z))). Definition 2.6. For each θ ∈ R we define Bθ = C∗(πθ(UT (4, Z))). 4 CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY 2.2. Crossed product construction. In order to describe the order structure on K0(Bθ) we describe Bθ as a crossed product. Recall the definition of H4 in Definition 2.4. Lemma 2.7. Let θ be irrational and Aθ denote the irrational rotation algebra asso- ciated with θ. Then C∗(πθ(H4)) ∼= Aθ ⊗ Aθ (πθ is the representation from Definition 2.5). Proof. By the relations in (2.2), one sees that C∗(πθ(1+e12), πθ(1+e24)) and C∗(πθ(1+ e13), πθ(1 + e34)) are two commuting copies of Aθ. The conclusion then follows by the simplicity of Aθ, the nuclearity of Aθ and Takesaki's theorem on the simplicity of simple minimal tensor products (see [32, Corollary IV.4.21]). (cid:3) We now describe the conjugation action of 1 + e23 on Aθ ⊗ Aθ. Define the automor- phism β : H4 → H4 by (2.4) β(x) = (1 + e23)x(1 + e23)−1 = [1 + e23, x]x. This combined with the relations in (2.2) produces β(1 + e12) = 1 + e12 − e13, β(1 + e13) = 1 + e13, β(1 + e34) = 1 + e24 + e34, β(1 + e24) = 1 + e24. Summarizing the above discussion we obtain Theorem 2.8. Let u, v be standard generators of Aθ. Then Bθ ∼= (Aθ ⊗ Aθ) ⋊β Z where β(u ⊗ 1) = u ⊗ u−1, β(v ⊗ 1) = v ⊗ 1, β(1 ⊗ u) = 1 ⊗ u, β(1 ⊗ v) = v ⊗ v. 2.3. Twisted group C-algebra. In order to calculate the K-groups of Bθ we de- scribe Bθ as a twisted group C-algebra. Everything in this section will be well- known to experts in twisted group C-algebras. We recall the necessary definitions (see for example Section 1 of [18]) for the non-experts. Let Γ be a discrete group and σ : Γ × Γ → T a 2-cocycle (also referred to as a multiplier ). The involutive Banach algebra ℓ1(Γ, σ) is formed with the following multiplication and involution f∗(s) = σ(s−1, s)f (s−1). f (t)g(t−1s)σ(t, t−1s), (2.5) f ∗ g(s) :=Xt∈Γ The left regular representation of ℓ1(Γ, σ) is defined on B(ℓ2(Γ)) as λ(f )(g)(t) =Xs∈Γ σ(s, s−1t)f (s)g(s−1t). The reduced twisted group C-algebra is defined as C∗r (Γ, σ) = C∗(λ(ℓ1(Γ, σ))). Let Z be the center of UT (4, Z) and C = {x ∈ UT (4, Z) : x14 = 0} ⊆ UT (4, Z). UNITRIANGULAR GROUPS 5 Notice that C is a complete choice of coset representatives for UT (4, Z)/Z. Let c : UT (4, Z)/Z → C be the unique lifting of the quotient map. Following [20] for each θ ∈ R, we define the 2-cocycle ωθ : UT (4, Z)/Z × UT (4, Z)/Z → T by ωθ(xZ, yZ) = τθ(c(xZ)c(yZ)c(xyZ)−1). Proposition 2.9. Let ωθ be the cocycle from above. Then Bθ is isomorphic to the reduced twisted group C-algebra C∗r (UT (4, Z)/Z, ωθ). Proof. Let L2(UT (4, Z), τθ) denote the Hilbert space associated to the GNS represen- tation of τθ. It was shown in [5, Lemma 2.4], that {δx : x ∈ C} is an orthonormal basis for L2(UT (4, Z), τθ) and W : L2(UT (4, Z), τθ) → ℓ2(UT (4, Z)/Z), given by W δx = δxZ, is unitary. Moreover [5, Lemma 2.4] shows that Bθ is generated by {πθ(x) : x ∈ C}. Very easy calculations and [5, Lemma 2.4] show that for each x ∈ C we have W πθ(x) = λ(xZ)W. It then follows that Bθ ∼= C∗r (UT (4, Z)/Z, ωθ). (cid:3) We now use the fact that each Bθ is a twisted group C-algebra to calculate their unordered K-groups. Throughout this section we set 3. Computation of K∗ Lemma 3.1. We have Γ = UT (4, Z)/Z(UT (4, Z)). K0(C∗(Γ)) = K1(C∗(Γ)) = Z10. Proof. Note that Γ ∼= Z4 ⋊α Z where he12, e13, e24, e34i/Z(UT (4, Z)) ∼= Z4 and the automorphism α is implemented by conjugation of e23 mod Z(UT (4, Z)). In other words, let x1, x2, x3, x4 be a free basis of Z4 and define α(x1) = x1 − x2, α(x2) = x2, α(x3) = x3, α(x4) = x3 + x4, (3.1) then Γ ∼= Z4 ⋊α Z. It is well known (see for example [8, 2.1] or [2]) that the graded ring K∗(C(T4)) = K0(C(T4)) ⊕ K1(C(T4)) can be identified with the exterior algebra of 4 generators e1, e2, e3, e4 over Z. Furthermore, K0 is identified with those terms of even degree and the standard coordinate functions in C(T4) correspond to e1, e2, e3, e4. The induced action of α on K∗(C(T4)) is a ring automorphism. This fact combined with (3.1) shows that α∗(x) = x for x ∈ {1, e2, e3, e1 ∧ e2, e2 ∧ e3, e3 ∧ e4, e1 ∧ e2 ∧ e3, e2 ∧ e3 ∧ e4, e1 ∧ e2 ∧ e3 ∧ e4}. Furthermore, α∗(e1) = e1 − e2, α∗(e1 ∧ e3 ∧ e4) = e1 ∧ e3 ∧ e4 − e2 ∧ e3 ∧ e4, α∗(e4) = e3 + e4, α∗(e1 ∧ e2 ∧ e4) = e1 ∧ e2 ∧ e3 + e1 ∧ e2 ∧ e4, which determines the homomorphism α∗ : K1(C(T4)) → K1(C(T4)). It follows that (3.2) K1(C(T4))/(id − α∗)(K1(C(T4))) ∼= Z4. 6 CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY Similarly, the following calculations determine the K0 counterpart: α∗(e1 ∧ e3) = e1 ∧ e3 − e2 ∧ e3, α∗(e2 ∧ e4) = e2 ∧ e3 + e2 ∧ e4, α∗(e1 ∧ e4) = e1 ∧ e3 + e1 ∧ e4 − e2 ∧ e3 − e2 ∧ e4. It follows that K0(C(T4))/(id − α∗)(K0(C(T4))) ∼= Z2. (3.3) As was implied above, we have K0(C(T4)) ∼= K1(C(T4)) ∼= Z8. This fact combines with (3.2), (3.3) and the Pimsner-Voiculescu six term exact sequence for crossed products [23, Theorem 2.4] to prove the claim. (cid:3) Corollary 3.2. Let θ ∈ R. Then Ki(Bθ) ∼= Z10 for i = 0, 1. Proof. By Proposition 2.9, we have Bθ ∼= C∗(Γ, ωθ). When θ = 0, Bθ is isomorphic to the group C-algebra C∗(Γ). For any two θ1, θ2, the cocycles ωθ1 and ωθ2 are homotopic. Indeed, define r(t) := tθ1 + (1 − t)θ2, and a homotopy of 2-cocycles ω : Γ × Γ × [0, 1] → T by ω(·,·, t) := ωr(t)(·,·). Since Γ is discrete in the simply connected (contractible actually) nilpotent Lie group UT (4, R)/Z(UT (4, R)), it follows from Theorem 4.2 of [19] that K∗(Bθ) ∼= K∗(C∗(Γ)). (cid:3) 4. Range of the trace Let A be a unital C-algebra and τ a tracial state on A. We also denote by τ the state on K0(A). In general it is difficult to calculate the range τ (K0(A)) ⊂ R. Pimsner showed in [24] that in the case of crossed products by free groups an examination of the de la Harpe-Skandalis determinant [4] can sometimes lead to a satisfying description of the range of the trace of the crossed product. Pimsner's ideas work particularly well for the cases at hand. We very briefly recall the necessary background from [4] and [24] (see also [2]) and refer the reader to [24] for more information and proofs of the claims made below. Let U(A) denote the unitary group of a C-algebra. Let U∞(A) be the inductive limit of U(Mn(A)) in the usual way. Let U∞(A)0 denote the connected component of the identity in U∞(A). For a piecewise differentiable path ξ : [0, 1] → U∞ one defines ∆τ (ξ) = 1 2πi 1Z0 τ (ξ′(t)ξ(t)−1) dt. The map ∆τ is constant on homotopy classes with fixed endpoints. Let ξ be a differentiable path of unitaries and n ≥ 1 an integer. From the following easily derivable formula, d dt ξn(t) = nξ′(t)ξn−1(t) − n−1Xi=1 [ξ′(t), ξn−i(t)]ξ(t)i−1, UNITRIANGULAR GROUPS 7 one deduces that (4.1) ∆τ (ξn) = n∆τ (ξ). Lemma 4.1. Let u, v ∈ Un(A) be unitaries. Then for any piecewise differentiable path ξ from 1 to uvu−1v−1 we have ∆τ (ξ) = 0. Proof. Let w(t) =(cid:18) cos(πt/2) sin(πt/2) − cos(πt/2) (cid:19) and define 1 (cid:19) w(t)(cid:18) v 1 (cid:19) w(t)(cid:18) u−1 ξ(t) =(cid:18) u sin(πt/2) 1 (cid:19) w(t)(cid:18) v−1 1 (cid:19) w(t). It is then a straightforward calculation to see that ∆τ (ξ) = 0. (cid:3) Under the Bott isomorphism K0(A) ∼= K1(SA), one sees that ∆τ restricted to K0(A) is just τ. Let q : R → R/τ (K0(A)) be the quotient map. Then for any unitary u ∈ U∞(A)0 and any piecewise differentiable path ξ from 1 to u the map is a well-defined group homomorphism. ∆τ (u) = q ◦ ∆τ (ξ) Suppose now that α is an automorphism of A and that τ = τ ◦ α. Define the group homomorphism ∆α τ : ker(K1(id) − K1(α)) → R/τ (K0(A)) by ∆α τ ([u]1) = ∆τ (uα−1(u−1)). Finally, we have Theorem 4.2 (Pimsner [24]). The following sequence is exact: 0 −→ τ (K0(A)) −→ τ (K0(A ⋊α Z)) −→ ∆α τ (ker(K1(id) − K1(α))) −→ 0, where the first map is the inclusion and the second is the quotient q : R → R/τ (K0(A)). It is well known that Aθ arises as a crossed product C(T) ⋊ Z, hence we may apply the Kunneth formula [30], which provides K0(Aθ ⊗ Aθ) ∼=(cid:16)K0(Aθ) ⊗ K0(Aθ)(cid:17) ⊕(cid:16)K1(Aθ) ⊗ K1(Aθ)(cid:17). Let τ be the unique trace on Aθ ⊗ Aθ. We also denote by τ the extension of this trace to Bθ. Pimsner and Voiculescu showed in [23] that the range of the trace on K0(Aθ) is Z + θZ. From this it follows that Moreover from the description of the map in the Kunneth formula one checks that Z + θZ + θ2Z ⊆ τ (K0(Aθ ⊗ Aθ)). hence (4.2) ker(τ ) ⊇ K1(Aθ) ⊗ K1(Aθ), τ (K0(Aθ ⊗ Aθ)) = Z + θZ + θ2Z. 8 CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY 4.1. A finite index subgroup. Eventually we will apply Pimsner's ideas of the previous section to compute the range of the trace for Bθ ∼= Aθ ⊗ Aθ ⋊β Z. As one might expect from Pimsner's Theorem 4.2, a good description of ker(K1(id)− K1(β)) is helpful. In our case a complication arises because the automorphism β of Aθ ⊗ Aθ mixes up the tensor factors making it difficult to describe ker(K1(id)−K1(β)). In this section we look at a subalgebra A ⊆ Bθ that is isomorphic to a crossed product of A2θ⊗A2θ by an automorphism that does factor as a tensor product of automorphisms of A2θ (providing an easy path to the range of the trace calculation for A). The algebra A is "big enough" to then yield the range of trace calculation for Bθ. Recall the Heisenberg group H4 defined in (2.1), and the automorphism β from (2.4). Then H4/Z(H4) ∼= Z4. Since β fixes the center, it drops to an automorphism of Z4 (which we still denote by β). The matrix for β y Z4 with respect to the basis {1 + e12, 1 + e13, 1 + e24, 1 + e34} mod Z(H4) is Then the following two subgroups of Z4 are invariant under β and β−1 : (4.3) eX = 0 −1 1 0 1 1 0 1 1 0 −1 1  . β = eY =   , + , Z4/heX,eY i ∼= Z/2Z × Z/2Z. 0 −1 −1 0 1 0 0 1  , 1 0 0 −1  + . map. Then 3π(e1) = π(e1) = π(e4) 6= π(e2) = π(e3) = 3π(e3), so (4.4) Let e1, e2, e3, e4 be the standard basis for Z4 and π : Z4 → Z4/heX,eY i the quotient Now set X (resp. Y ) equal to the inverse image of eX (resp. eY ) under the quotient map H4 → H4/Z(H4). Definition 4.3. Set u1 = πθ(1−e13 +e24), v1 = πθ(1+e12 +e34), u2 = πθ(1−e13−e24), v2 = πθ(1 + e12 − e34). Then C∗(u1, v1) = C∗(πθ(X)) ∼= A2θ ∼= C∗(πθ(Y )) = C∗(u2, v2) and by similar reasoning as in Lemma 2.7 we have C∗({ui, vi : i = 1, 2}) ∼= C∗(u1, v1)⊗ C∗(u2, v2) ∼= A2θ ⊗ A2θ. Furthermore we have we have βA2θ⊗A2θ = β1 ⊗ β2 where (4.5) βj(uj) = uj, and βj(vj) = exp((−1)j2πiθ)ujvj, for j = 1, 2. The above information combines with [22, Corollary 2.5] to provide Lemma 4.4. For j = 1, 2 we have K0(βj) = id and K1(βj)([uj]1) = [uj]1 and K1(βj)([vj]1) = [ujvj]1. Let G ≤ H4 be the subgroup generated by X and Y . Then H4/G = 4 by (4.4). Let e = x0, x1, x2, x3 ∈ H4 be H4/G coset representatives. As in Definition 4.3, we have A2θ ⊗ A2θ ∼= C∗(πθ(G)). UNITRIANGULAR GROUPS 9 Consider the unitary representation πθG of G. The unitary representation πθ of H4 is just the induced representation IndH4 G (πθG). It then follows from the general theory of induced representations (see for example, [1, Appendix E]) and C-algebras that there is an embedding σ : Aθ ⊗ Aθ → MH4/G(A2θ ⊗ A2θ) = M4(A2θ ⊗ A2θ) such that σ(πθ(t)) = πθ(t) πθ(x−1 1 tx1) πθ(x−1 2 tx2) πθ(x−1 3 tx3)  for all t ∈ G. Notice that each generator u1, u2, v1, v2 ∈ C∗(πθ(G)) is mapped to a scalar multiple of itself under the automorphisms Ad(πθ(xi) for i = 1, 2, 3. In par- ticular each of the automorphisms Ad(πθ(xi)) is homotopic to the identity. We have shown the following Proposition 4.5. Let ι : A2θ ⊗ A2θ → Aθ ⊗ Aθ be the inclusion map given by C∗(πθ(G)) ⊆ C∗(πθ(H4)). Let σ be as above. Then K∗(σ ◦ ι) = 4 · idK∗(A2θ⊗A2θ). By the functoriality of K∗, and the fact that all the K-groups of Aθ ⊗ Aθ and A2θ ⊗ A2θ are torsion free, we have K∗(ι) and K∗(σ) are both injective. 4.2. Range of trace. Let A and B be C-algebras. In [30], Schochet describes a Z/2Z graded pairing (4.6) α : Kp(A) ⊗ Kq(B) → Kp+q(A ⊗ B), p, q ∈ Z/2Z that is an isomorphism when A is in the "bootstrap class" and B has torsion free K-theory [30, Theorem 2.14]. Suppose now that A and B satisfy the conditions of the previous sentence. Let β = β1 ⊗ β2 be an automorphism of A ⊗ B. It is clear that K∗(β1) ⊗ K∗(β2) induces a graded automorphism of K∗(A) ⊗ K∗(B). Moreover, by the description of α, a straightforward verification reveals that the following diagram commutes: (4.7) K∗(A) ⊗ K∗(B) α K∗(A ⊗ B) . K∗(β1)⊗K∗(β2) K∗(β1⊗β2) K∗(A) ⊗ K∗(B) α / K∗(A ⊗ B) Since Aθ is isomorphic to a crossed product of the form C(T) ⋊ Z, it is in the boot- strap class by [30, Proposition 2.7]. Since Aθ has torsion free K-theory, we have an isomorphism in (4.6) when A = B = Aθ. Recall from [22] that K0(A2θ) ∼= K1(A2θ) ∼= Z2. Moreover for unitary generators w1 and w2, we have {[w1]1, [w2]1} a free basis for K1(A2θ). By [30] we have K1(A2θ ⊗ A2θ) ∼= [K0(A2θ) ⊗ K1(A2θ)] ⊕ [K1(A2θ) ⊗ K0(A2θ)] ∼= Z4 ⊕ Z4 ∼= Z8. / / / 10 CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY Under this identification and the ordered free bases {[uj]1, [vj]1} of K1(A2θ), by Lemma 4.4, we have K1(β1 ⊗ β2) =hidZ2 ⊗(cid:18) 1 1 0 1 (cid:19)i ⊕h(cid:18) 1 1 0 1 (cid:19) ⊗ idZ2i. We therefore have the following Lemma 4.6. The subgroup ker(K1(id) − K1(β1 ⊗ β2)) ≤ K1(A2θ⊗2θ) is generated by the set (4.8) {[p ⊗ u2 + (1 − p) ⊗ 1]1, [u1 ⊗ p + 1 ⊗ (1 − p)]1 : p ∈ A2θ is a projection}. Theorem 4.7. We have τ (K0(Bθ)) = τ (K0((Aθ ⊗ Aθ) ⋊β Z)) = Z + θZ + θ2Z. Proof. By (4.2) and Theorem 4.2 it suffices to show that for all x ∈ ker(K1(id)−K1(β)) we have ∆β τ (x) = 0. By Proposition 4.5, for any x ∈ ker(K1(idAθ⊗Aθ) − K1(β)) we have 4x ∈ ker(K1(idA2θ⊗A2θ ) − K1(β)). Therefore by (4.1) we need to show that for any [w]1 ∈ ker(K1(idA2θ⊗A2θ) − K1(β)) and differentiable path ξ from the identity to wβ−1(w) in (U∞)0 we have ∆τ (ξ) = 0. By Lemma 4.6 we only need to show this for those w of the form in (4.8). We will only show it for w of the form [p⊗u2+(1−p)⊗1]1 (one proves it for [u1⊗p+1⊗(1−p)]1 in exactly the same manner). 1 (p) are unitarily equivalent. Since β2(u2) = u2 it follows that p⊗ u2 + (1− p)⊗ 1 is unitarily equivalent to β(p ⊗ u2 + (1 − p) ⊗ 1). The conclusion now follows from Lemma 4.1. By a result of Rieffel [26, Corollary 2.5] the projections p and β−1 (cid:3) 5. Elliott invariants We now gather the information of the proceeding sections to describe the Elliott invariants of the algebras Bθ. Theorem 5.1. Let θ be irrational. Then Bθ is a simple AT algebra with unique trace and Ki(Bθ) ∼= Z10 for i = 0, 1. Let Θ = (1, θ, θ2, 0, ..., 0) ∈ R10. Then (5.1) K0(Bθ)+ = {0} ∪ {x ∈ Z10 : hx, Θi > 0}. Proof. The K-groups were calculated in Corollary 3.2. Since Bθ is the C-algebra generated by an irreducible representation of a finitely generated nilpotent group, it is simple with a unique trace (this is well known, see e.g. the introduction of [5]). It follows from [6, Theorems 2.9 & 4.4 ] that Bθ has strict comparison. Since Bθ has unique trace, this means the order structure on K0(Bθ) = Z10 is completely determined by the range of the trace, which is Z + θZ + θ2Z by Theorem 4.7. This shows (5.1). This description of (K0(Bθ), K0(Bθ)+) shows it is a Riesz group, and therefore a dimension group by the Effros-Handelman-Shen theorem [7]. Therefore Bθ has the same Elliott invariant of an AT algebra by [9]. Since Aθ is an AT algebra by [10], it follows that Aθ ⊗ Aθ is an AT2 algebra (it is actually an AT algebra see [27, Proposition 3.25]) and therefore satisfies the universal UNITRIANGULAR GROUPS 11 coefficient theorem by [29]. Also by [29], it follows that (Aθ ⊗ Aθ) ⋊β Z satisfies the universal coefficient theorem. By [5], Bθ is quasidiagonal and by [6] Bθ has finite nuclear dimension. By Theorem 2.1, Bθ is therefore isomorphic to an AT algebra. (cid:3) 6. An isomorphism criterion As mentioned in the introduction we intend to show there are irrational numbers θ, η such that Bθ ∼= Bη but Bθ and Bη are not "obviously" isomorphic (i.e. θ 6= −η mod Z). This shows that one must use the classification theorem (Theorem 2.1) to classify the algebras {Bθ : θ ∈ (0, 1) \ Q} amongst themselves. By the results of the preceding sections it follows that Bθ and Bη are isomorphic if and only if the ordered groups (Z3, (1, 0, 0), Pθ) and (Z3, (1, 0, 0), Pη) with distin- guished order unit are isomorphic, where Pθ = {x ∈ Z3 x1 + x2θ + x3θ2 > 0}∪{0}. An automorphism of Z3 to Z3 is implemented by some At ∈ GL(3, Z) ( t denotes transpose), and it is easy to see that At sends Pθ to Pη if and only if (6.1) 1 θ θ2 = η2 . A 1 η We will find below that for most pairs (θ, η), (6.1) holds if and only if η = ±θ (mod Z). This reflects the situation with the irrational rotation algebras Aθ. However, for certain θ, there are more possibilities. Fix an irrational θ. Our goal is to more easily describe the relation between θ and η given by (6.1). Equation (6.1) defines an equivalence relation θ ∼ η which extends the equivalence relation θ = ±η (mod Z), since if η = ±θ + k with k ∈ Z, then 1 0 k ±1 k2 ±2k 1  0 0 1 θ 1 η η2 . θ2 = We will describe the relation ∼ in several cases based on degθ (that is, the degree of the minimal polynomial of θ over Q). 6.1. deg θ > 2. Throughout this subsection fix a θ with degree strictly bigger than two. Suppose that θ ∼ η. The degree restriction implies that the first row of A must be (1 0 0). Also, since we are only counting the number of η's which are distinct mod Z and modulo the automorphism η 7→ −η, we may assume that a21 = 0 and det A = 1. Then we may write A in the form (6.2) By (6.1), A = 1 0 0 0 a b c d e . (aθ + bθ2)2 = η2 = c + dθ + eθ2, 12 hence (6.3) CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY b2θ4 + 2abθ3 + (a2 − e)θ2 − dθ − c = 0. We now split into three subcases. Case (a): θ has degree > 4. Then the coefficients of the polynomial in (6.3) must In particular, b = 0 and e = a2. Then A is lower triangular, and since be zero. det A = 1, it follows that a = e = 1. But then η = θ. Case (b): θ has degree 4. Let p(x) = x4 + λ3x3 + λ2x2 + λ1x + λ0, with λi ∈ Q, be the minimal polynomial for θ. Then it follows that 2ab = b2λ3, a2 − e = b2λ2, −d = b2λ1, −c = b2λ0. Moreover, ae − bd = det(A) = 1. Note that if b = 0, then just as in case (a), we have a = 1 and hence η = θ. Otherwise, we have a = bλ3, 1 2 1 4 3 − b2λ2, b2λ2 e = d = −b2λ1, c = −b2λ0, and hence µb3 = 1, where µ = 1 8 λ3 3 − 1 2 λ3λ2 + λ1. Case (c): θ has degree 3. Let p(x) = x3 +λ2x2 +λ1x+λ0 be the minimal polynomial for θ. Then, we have (2ab − b2λ2)θ3 + (a2 − e − b2λ1)θ2 − (d + b2λ0)θ − c = 0 in which case, a2 − e − b2λ1 = (2ab − b2λ2)λ2, −(d + b2λ0) = (2ab − b2λ2)λ1, −c = (2ab − b2λ2)λ0. The above equations allow us to express c, d and e in terms of a and b. Then, again using det A = ae − bd = 1, we have (6.4) The above equation is of the form F (a, b) = 0, where F (x, y) = x3+px2y+qxy2+ry3− 1 is a cubic polynomial with rational coefficients. Its projectivization is F (X, Y, Z) = X 3 + pX 2Y + qXY 2 + rY 3 − Z 3. 2 + λ1)ab2 + (λ0 − λ1λ2)b3 = det A = 1. a3 − 2λ2a2b + (λ2 UNITRIANGULAR GROUPS 13 Lemma 6.1. The projective curve defined by F (X, Y, Z) = 0 is nonsingular, that is, there are no points [X : Y : Z] ∈ P2 for which F, ∂F ∂Z all vanish simultaneously. ∂Y , and ∂F ∂X , ∂F Proof. Since ∂F satisfy ∂F ∂Z = 3Z 2, any nonsingular point must lie in R2. So suppose (x, y) ∈ R2 ∂x (x, y) = ∂F ∂y (x, y) = 0. Writing t = x + 1 3py, we have which implies t2 =(cid:0) 1 9p2 − 1 using that t2 = ky2, we get 1 3 ∂F ∂x p2(cid:19) y2 = 0 = 3t2 +(cid:18)q − 3q(cid:1) y2. Let k = 1 9p2 − 1 =(cid:18)2pk − 6k3/2 + 3r − 1 3 pq(cid:19) y2 = 0. ∂F ∂y 3q. Rewriting ∂F ∂y in terms of t, and If y = 0, then t = 0 and hence x = 0. But then F (x, y) = −1. Hence we may assume instead that Now, rewriting F in terms of t, and again using that t2 = ky2, we get α = 2pk − 6k3/2 + 3r − 1 3 pq = 0. F (x, y) =(cid:18)−2k3/2 − 1 27 p3 + pk + r(cid:19) y3 − 1. Writing β for the coefficient of y3 above, we see that 3β = −6k3/2 − 1 9 p3 + 3pk + 3r = α + pk − 1 9 p3 + 1 3 pq = α. Since α = 0, it follows that F (x, y) = −1, and hence (x, y) does not lie on the curve. (cid:3) Siegel's theorem ( [31]) implies that the curve F (x, y) = 0 can only have finitely many integral points. It follows that there are only finitely many choices for A, and hence for η. 6.2. deg θ = 2. Here we take a different strategy for determining how many η may satisfy θ ∼ η, by studying the groups Gη = Z + ηZ + η2Z. Note that if θ ∼ η, then Gθ = Gη. So suppose this is true for some η. Since θ has degree 2, we may write where a, b, c ∈ Z have no common factors, k ∈ Q, and k can be written r/s where r, s ∈ Z have no common factors and are each square free. Since Q + √kQ = Q + θQ + θ2Q = Q + ηQ + η2Q, it follows that deg η = 2 as well, and moreover x + y√k η = a + b√k θ = c z 14 CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY for some x, y, z ∈ Z with no common factors. Now, we have Gθ = Z + a + b√k c ! Z + a2 + b2k + 2ab√k c2 ! Z ⊆ 1 c2s(cid:16)Z + √kZ(cid:17) . Since η ∈ Gθ, it follows that (6.5) Now consider the group Q + Gθ. We clearly have z c2s. √k(cid:18)Z + 2a c Z(cid:19) = Q + √kZ b cd Q + Gθ = Q + b c where d is the denominator of 2a we must have c when written in lowest form. Since Q+Gθ = Q+Gη, (6.6) b cd = y zw where w is the denominator of 2x/z when written in lowest form. Now by (6.5), there can be only finitely many choices for z. Moreover, w must divide z, so there are only finitely many choices for w. By equation (6.6), once z and w are fixed, there can only be finitely many choices for y. Finally, we only have finitely many choices (mod Z) for x, and hence we can have only finitely many choices for η (mod Z). Summarizing the above, we have Theorem 6.2. Let θ be irrational. • If θ has degree > 4, then for all η, Bθ ∼= Bη if and only if θ = ±η (mod Z). • Suppose θ has degree 4, with minimal polynomial p(x) = x4 + λ3x3 + λ2x2 + λ1x + λ0. If, for some nonzero k ∈ Z, then Bθ ∼= Bη if and only if η = ±θ (mod Z) or η = ±ζ (mod Z), where 1 8 λ3 3 − 1 2 λ3λ2 + λ1 = 1 k3 , ζ = kλ3θ + kθ2. 1 2 Otherwise, Bθ ∼= Bη if and only if η = ±θ (mod Z). • If θ has degree 2 or 3, then there are finitely many distinct values of η (mod Z) such that Bθ ∼= Bη. It is already apparent from Theorem 6.2 that in the case where the degree of θ is less than or equal to four, that Bθ ∼= Bη will happen more often than Aθ ≃ Aη. For the cases deg θ = 2, 3, we provide examples to illustrate that possibly even more can happen. Example 6.3. Let θ be the real solution to θ3 = θ + 1. Then the equation (6.4) from case (d) in this case is a3 − ab2 − b3 = 1 UNITRIANGULAR GROUPS 15 whose integer solutions are (1, 0), (0,−1), (±1,−1), and (4, 3). These correspond to the values η = θ, −θ2, ±θ − θ2, and 4θ + 3θ2. Since deg θ > 2, these values are all distinct mod Z. Finally, we provide a similar example in the case where deg θ = 2. Example 6.4. Let θ = 1+√2 have 3 and η = 1+2√2 3 1 9 1 3 Gθ = Z + . Then (in the notation of case (d)) we √2Z = Gη. Therefore, Bθ ∼= Bη. On the other hand, θ is not equal to ±η + n for any integer n. 7. A C-algebra generated by a faithful irreducible representation of a finitely generated torsion free nilpotent group that is not an AT algebra So far all of the algebras we considered turned out to be AT algebras. By Chris Phillips' theorem [21] if G is any finitely generated two-step nilpotent group and π a faithful irreducible representation of G, then C∗(π(G)) is an AT algebra. Since a natural conjecture forms from the previous sentences we would like to point out that it is very easy to produce irreducible representations of finitely generated nilpotent groups (of nilpotency step necessarily larger than 2) that are not AT algebras. To this end, let θ be irrational and u and v be generators of Aθ satisfying the commutation relation. Consider the automorphism β of Aθ defined by β(u) = u and β(v) = u2v. It is then clear from the Pimsner-Voiculescu six-term exact sequence that K1(Aθ ⋊β Z) contains an element of order 2. Since all AT algebras have torsion free K1, it follows that Aθ ⋊β Z is not an AT algebra. We claim that Aθ ⋊β Z is isomorphic to the C-algebra generated by an irreducible representation of a 3-step nilpotent group. Indeed, let H3 ≤ GL(3, Z) be the Heisenberg group with generators a, b. Then β(a) = a and β(b) = a2b defines an automorphism of H3. Notice that β fixes Z(H3). Moreover the induced action of β on H3/Z(H3) ∼= Z2 is given by the unipotent matrix(cid:18) 1 2 0 1 (cid:19) , showing that H3 ⋊β Z is a three step nilpotent group. Let πθ be the representation of H3 ⋊β Z induced from θ ∈ T = \Z(H3). One checks fairly easily that C∗(πθ(H3 ⋊β Z)) ∼= Aθ ⋊β Z. Remark 7.1. The above example can clearly be generalized (with minimal effort) in a variety of ways to produce all kinds of finitely generated K1 groups. Acknowledgements C. E. thanks Marius Dadarlat and Andrew Toms for several informative conversa- tions about ordered K-theory and also to Marius for suggesting the method of proof for Lemma 3.1. 16 CALEB ECKHARDT, CRAIG KLESKI, AND PAUL MCKENNEY References [1] Bachir Bekka, Pierre de la Harpe, and Alain Valette. Kazhdan's property (T), volume 11 of New Mathematical Monographs. Cambridge University Press, Cambridge, 2008. [2] Bruce Blackadar. K-theory for operator algebras, volume 5 of Mathematical Sciences Research Institute Publications. Cambridge University Press, Cambridge, second edition, 1998. [3] A. L. Carey and W. Moran. Characters of nilpotent groups. Math. Proc. Cambridge Philos. Soc., 96(1):123 -- 137, 1984. [4] P. de la Harpe and G. Skandalis. D´eterminant associ´e `a une trace sur une alg´ebre de Banach. Ann. Inst. Fourier (Grenoble), 34(1):241 -- 260, 1984. [5] Caleb Eckhardt. Quasidiagonal representations of nilpotent groups. Adv. Math., 254:15 -- 32, 2014. [6] Caleb Eckhardt and Paul McKenney. Finitely generated nilpotent group C-algebras have finite nuclear dimension. arXiv:1409.4056 [math.OA], to appear in J. Reine Angew. Math., 2015. [7] Edward G. Effros, David E. Handelman, and Chao Liang Shen. Dimension groups and their affine representations. Amer. J. Math., 102(2):385 -- 407, 1980. [8] G. A. Elliott. On the K-theory of the C ∗-algebra generated by a projective representation of a torsion-free discrete abelian group. In Operator algebras and group representations, Vol. I (Neptun, 1980), volume 17 of Monogr. Stud. Math., pages 157 -- 184. Pitman, Boston, MA, 1984. [9] George A. Elliott. On the classification of C ∗-algebras of real rank zero. J. Reine Angew. Math., 443:179 -- 219, 1993. [10] George A. Elliott and David E. Evans. The structure of the irrational rotation C ∗-algebra. Ann. of Math. (2), 138(3):477 -- 501, 1993. [11] Guihua Gong, Huaxin Lin, and Zhuang Niu. Classification of finite simple Z-stable C-algebras. [12] Roger E. Howe. On representations of discrete, finitely generated, torsion-free, nilpotent groups. arXiv:1501.00135v3 [math.OA], 2015. Pacific J. Math., 73(2):281 -- 305, 1977. [13] Huaxin Lin. Classification of simple C ∗-algebras of tracial topological rank zero. Duke Math. J., 125(1):91 -- 119, 2004. [14] Huaxin Lin and Zhuang Niu. Lifting KK-elements, asymptotic unitary equivalence and classi- fication of simple C ∗-algebras. Adv. Math., 219(5):1729 -- 1769, 2008. [15] Hiroki Matui and Yasuhiko Sato. Strict comparison and Z-absorption of nuclear C ∗-algebras. [16] Hiroki Matui and Yasuhiko Sato. Decomposition rank of UHF-absorbing C-algebras. Duke Acta Math., 209(1):179 -- 196, 2012. Math. J., 163(14):2687 -- 2708, 2014. [17] Calvin C. Moore and Jonathan Rosenberg. Groups with T1 primitive ideal spaces. J. Functional Analysis, 22(3):204 -- 224, 1976. [18] Judith A. Packer. Twisted group C ∗-algebras corresponding to nilpotent discrete groups. Math. Scand., 64(1):109 -- 122, 1989. [19] Judith A. Packer and Iain Raeburn. Twisted crossed products of C ∗-algebras. II. Math. Ann., 287(4):595 -- 612, 1990. [20] Judith A. Packer and Iain Raeburn. On the structure of twisted group C ∗-algebras. Trans. Amer. Math. Soc., 334(2):685 -- 718, 1992. [21] N. Christopher Phillips. Every simple higher dimensional noncommutative torus is an AT al- gebra. arXiv:math/0609783 [math.OA], 2006. [22] M. Pimsner and D. Voiculescu. Exact sequences for K-groups and Ext-groups of certain cross- product C ∗-algebras. J. Operator Theory, 4(1):93 -- 118, 1980. [23] M. Pimsner and D. Voiculescu. Imbedding the irrational rotation C ∗-algebra into an AF-algebra. J. Operator Theory, 4(2):201 -- 210, 1980. UNITRIANGULAR GROUPS 17 [24] Mihai V. Pimsner. Ranges of traces on K0 of reduced crossed products by free groups. In Op- erator algebras and their connections with topology and ergodic theory (Bu¸steni, 1983), volume 1132 of Lecture Notes in Math., pages 374 -- 408. Springer, Berlin, 1985. [25] Marc A. Rieffel. C ∗-algebras associated with irrational rotations. Pacific J. Math., 93(2):415 -- 429, 1981. [26] Marc A. Rieffel. The cancellation theorem for projective modules over irrational rotation C ∗- algebras. Proc. London Math. Soc. (3), 47(2):285 -- 302, 1983. [27] M. Rørdam. Classification of nuclear, simple C ∗-algebras. In Classification of nuclear C ∗- algebras. Entropy in operator algebras, volume 126 of Encyclopaedia Math. Sci., pages 1 -- 145. Springer, Berlin, 2002. [28] Mikael Rørdam. The stable and the real rank of Z-absorbing C ∗-algebras. Internat. J. Math., [29] Jonathan Rosenberg and Claude Schochet. The Kunneth theorem and the universal coefficient 15(10):1065 -- 1084, 2004. theorem for Kasparov's generalized K-functor. Duke Math. J., 55(2):431 -- 474, 1987. [30] Claude Schochet. Topological methods for C ∗-algebras. II. Geometric resolutions and the Kunneth formula. Pacific J. Math., 98(2):443 -- 458, 1982. [31] Joseph H. Silverman and John Tate. Rational points on elliptic curves. Undergraduate Texts in Mathematics. Springer-Verlag, New York, 1992. [32] M. Takesaki. Theory of operator algebras. I, volume 124 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, Berlin, 2002. Reprint of the first (1979) edition, Operator Algebras and Non-commutative Geometry, 5. [33] Wilhelm Winter. Nuclear dimension and Z-stability of pure C∗-algebras. Invent. Math., 187(2):259 -- 342, 2012. Department of Mathematics, Miami University, Oxford, OH, 45056
0907.2618	3	0907	2010-01-26T18:12:53	Noncommutative Semialgebraic sets and Associated Lifting Problems	[ "math.OA" ]	We solve a class of lifting problems involving approximate polynomial relations (soft polynomial relations). Various associated C-algebras are therefore projective. The technical lemma we need is a new manifestation of Akemann and Pedersen's discovery of the norm adjusting power of quasi-central approximate units. A projective C-algebra is the analog of an absolute retract. Thus we can say that various noncommutative semialgebraic sets turn out to be absolute retracts. In particular we show a noncommutative absolute retract results from the intersection of the approximate locus of a homogeneous polynomial with the noncommutative unit ball. By unit ball we are referring the C-algebra of the universal row contraction. We show projectivity of alternative noncommutative unit balls. Sufficiently many C-algebras are now known to be projective that we are able to show that the cone over any separable C-algebra is the inductive limit of C-algebras that are projective.	math.OA	math	NONCOMMUTATIVE SEMIALGEBRAIC SETS AND ASSOCIATED LIFTING PROBLEMS TERRY A. LORING AND TATIANA SHULMAN Abstract. We solve a class of lifting problems involving approximate polynomial relations (soft polynomial relations). Various associated C ∗-algebras are therefore projective. The technical lemma we need is a new manifestation of Akemann and Pedersen's discovery of the norm adjusting power of quasi-central approximate units. A projective C ∗-algebra is the analog of an absolute retract. Thus we can say that various noncommutative semialgebraic sets turn out to be absolute retracts. In particular we show a noncommutative absolute retract results from the intersection of the approximate locus of a homogeneous polynomial with the noncommutative unit ball. By unit ball we are referring the C ∗-algebra of the universal row contraction. We show projectivity of alternative noncommutative unit balls. Sufficiently many C ∗-algebras are now known to be projective that we are able to show that the cone over any separable C ∗-algebra is the inductive limit of C ∗-algebras that are projective. 0 1 0 2 n a J 6 2 ] . A O h t a m [ 3 v 8 1 6 2 . 7 0 9 0 : v i X r a Contents Introduction 1. 2. Quasi-Central Approximate Units Fix Norms 3. Lifting Softened Homogeneous Relations 4. Soft Versions of Known Projectives 5. Fattened Curves in Various NC unit balls 6. Soft Cylinders 7. Cones are Limits of Projective C ∗-Algebras References 1 5 6 10 13 17 18 22 1. Introduction Lifting problems for relations in C ∗-algebras have tended to have ad hoc solutions. Olsen and Pedersen prove in [23] that a nilpotent aways has a nilpotent lift, specifically that given x in a C ∗-algebra quotient A/I with xn = 0 there is always X in A with π(X) = x and X n = 0. Their proof is rather different from the techniques Akemann and Pedersen used in [2] to show that for x in A/I with kxnk ≤ ǫ (ǫ > 0) 2000 Mathematics Subject Classification. 46L05, 47B99 . Key words and phrases. C ∗-algebra, relation, projectivity, row contraction, noncommutative star- polynomial, lifting. 1 Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 2 there is always a lift X with kX nk ≤ ǫ. Different still are the techniques used in [22] to show that the relations describing CMn = C0 ((0, 1], Mn) are liftable: given x1, . . . , xn in A/I satisfying all the relations (∀j) kxjk ≤ 1 x∗ j xk = 0 (j 6= k) j xj = x∗ x∗ x1x∗ 1 = x∗ kxk 1x1 (∀j, ∀k) there are lifts X1, . . . , Xn in A that also satisfy these relations. More recently, M-ideals showed up in [24] to settle the lifting problem for the relations kxk ≤ 1 xn = 0. The lifting results above (most of them, anyway) show various C ∗-algebras are projective. Projectivity was introduced by Effros and Kaminker, in [10]. A C ∗-algebra P is projective if the map ρ ◦ -- : hom(P, B) → hom(P, C) is onto whenever ρ : B → C is onto. Projectivity was shown by Blackadar in [5] to be the noncommutative analog of a space being an absolute retract (AR). The analog of absolute neighborhood retract is semiprojec- tivity, which we will not discuss in detail in this paper except in Section 6. Systematic investigations of projectivity exist, but only in the case of at-most one-dimensional spectrum. There was a study of C0(X) for X + a tree in [17]. Chigogidze and Dranishnikov solved the general question for C0(X) being projective, in [7]. The answer is that C0(X) is projective if and only if X + is a dedrite. The finite mapping telescopes associated to inclusions of finite-dimensional C ∗-algebras were shown to be projective, in [22]. In the later terminology of [13], this says we have projectivity for a large class of one-dimensional noncommutative CW complexes. "NC" will stand for "noncommutative." Thus noncommutative CW complex becomes NCCW. In the commutative case, very sweeping statements can be made about what spaces are AR or ANR. For example, every compact semialgebraic set in finite-dimensional Euclidean space is an absolute neighborhood retract. See [25, p. 79] and [16] for precise results and definitions. A subset of Euclidean space is said to be semialgebraic if it is the union of solution sets of polynomial equations and polynomial inequalities. As we are interested in closed and connected sets, it will suffice to have in mind sets of the form This general result about semialgebraic sets being ANR cannot translate directly to C ∗- algebras. We know that for the unit disk D, the C ∗-algebra C0 (D \ {0}) ∼= C ∗(cid:10)x(cid:12)(cid:12) x∗x = xx∗, kxk ≤ 1(cid:11) fails to be projective. Normals don't generally lift to normals. Some normals fail to have par- tial lifts, and are bounded away from other normals that have partial lifts. To get technical, C0 (D \ {0}) is not even weakly semiprojective ([12]). where the pj are polynomials. (cid:8)(x1, . . . , xn) ∈ R n(cid:12)(cid:12) pj (x1, . . . , xn) ≤ ǫj for j = 1 . . . J(cid:9) , x x x, y x, y kxk ≤ 1 xn = 0 kxk ≤ 1 kxnk ≤ C kxk ≤ 1 kyk ≤ 1 x∗x = y ∗y x∗y = y ∗x = 0 x2 = y2 = 0 −1 ≤ x ≤ 1 −1 ≤ y ≤ 1 xy = 0 x1, . . . , xn kxjk ≤ 1, (∀j) h, k, x x, y, z, w x1, . . . , xr 0 ≤ x ≤ 1 0 ≤ y ≤ 1 0 ≤ z ≤ 1 0 ≤ w ≤ 1 x∗ i xi = x∗ x∗ i xi = 0 (if i 6= j) j xj (∀i, ∀j) hk = 0 0 ≤(cid:20) 1 − h x∗ k (cid:21) ≤ 1 x xy = 0 zw = 0 (1 − x)z(1 − x) = 0 (1 − x)w(1 − x) = 0 xkx∗ ≤ 1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1 k(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Credit Shulman [24](Olsen, Pedersen, Loring for kxk < 1.) Akemann, Pedersen [1] C0 ((0, 1], M3) Loring, Pedersen [22] C0 ((0, 1], C 4) Loring [17] Usefully in the qC picture of K-theory. C0(X) where X + is a tree with four edges. Loring, Pedersen [22, Example 3.11] Loring [19] Loring [17] Folklore, functional calculus Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 3 Gen- erators Individual restrictions Other Relations Name or Comment Table 1. Some Known Projective C ∗-algebras / Liftable relations There is a way to avoid the difficulty posed by this nonliftable example other than keeping to small dimension. We will avoid exact relations. An important instance of Theorem 3.2 is the fact that for any positive ǫ, an element x in a C ∗-algebra quotient A/I with kxk ≤ 1 kx∗x − xx∗k ≤ ǫ has a lift to X, so π(X) = x, with kXk ≤ 1 and kX ∗X − XX ∗k ≤ ǫ. Put another way, we show is projective for all positive ǫ. Since Aǫ = C ∗(cid:10)x(cid:12)(cid:12) kx∗x − xx∗k ≤ ǫ, kxk ≤ 1(cid:11) C0(D \ {0}) ∼= lim A 1 k → Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 4 Gen- erators Individual restrictions h, k h, k −1 ≤ h ≤ 1 −1 ≤ k ≤ 1 −1 ≤ h ≤ 1 −1 ≤ k ≤ 1 Other Relations Name / Remark Reference k[h, k]k ≤ C Soft Square Theorem 3.1 k[h, k]k ≤ C kh + ikk ≤ 1 Soft Disk I Theorem 3.1 x kxk ≤ 1 k[x∗, x]k ≤ 2C Soft Disk I a second presentation −1 ≤ h ≤ 1 −1 ≤ k ≤ 1 k[h, k]k ≤ C kh2 + k2k ≤ 1 Soft Disk II Theorem 3.1 h, k a, k h, p h, k, x x, y, z, w x1, . . . , xr kak ≤ 1 0 ≤ k ≤ 1 0 ≤ h ≤ 1 0 ≤ p ≤ 1 0 ≤ x ≤ 1 0 ≤ y ≤ 1 0 ≤ z ≤ 1 0 ≤ w ≤ 1 Might be useful investigating commutators and square roots k[a, k]k ≤ C kh(p2 − p)k ≤ C khkk ≤ C 0 ≤(cid:20) 1 − h x∗ k (cid:21) ≤ 1 x kxyk ≤ C kzwk ≤ C k(1 − x)z(1 − x)k ≤ C k(1 − x)w(1 − x)k ≤ C (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1 q 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Theorem 3.1 Theorem 3.1 Theorem 3.1 Theorem 4.2 (xkx∗ k) ≤ 1 1 < q < ∞ Theorem 5.3 Table 2. Some New Projective C ∗-algebras / Liftable relations we have shown C0(D \ {0}) has a shape system (c.f. [5, 6, 10]) that is trivial in the sense that all the C ∗-algebras in the system are projective. It was previously unknown whether C0(D \ {0}) could be written as an inductive limit of semiprojective C ∗-algebras. An important special case that we study is the approximate zero locus of a homogeneous NC ∗-polynomial intersected with the NC unit ball. The homogeneity is imposed to give contractability, and so gives us an expectation of finding not only semiprojectivity, but projectivity. By approximate zero locus we mean the universal C ∗-algebra Aǫ = C ∗x1, . . . , xr(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) and the "row contraction" conditionP xjx∗ kp (x1, . . . , xr)k ≤ ǫ, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXj=1 xjx∗ j(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1+ , with the unit ball. A special case of Theorem 3.2 states that Aǫ is projective for all ǫ > 0. j ≤ 1 (c.f. [3, 9]) is implementing the intersection Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 5 In many cases the relations we can handle have as their universal C ∗-algebra something that is very unfamiliar. In these cases, it is perhaps best to see our results as lifting results for the relations rather than projectivity results for the C ∗-algebras. Table 1, lists some known liftable relations. It is somewhat representative. There are closure results, such as when A is projective also Mn(A) is projective, which lead to many more liftable sets of relations, so no table can be complete. Table 2, lists some of the new examples. It is not certain these are new projective C ∗- algebras, as projective C ∗-algebras are contractible and so invariants such as K-theory are of no avail. We can say with some certainty the relations were not known to lift. An useful result in topological shape theory is that every compact metric space is the projective limit of ANRs [6, IX.1.4]. Blackadar writes in [5, 4.4]: It is not clear that every C-algebra has a strong shape system [is an inductive limit of semiprojective C-algebras]. We show in Section 7 that every cone over a separable C ∗-algebra is the inductive limit of projective C ∗-algebras. 2. Quasi-Central Approximate Units Fix Norms Our key tool for lifting is Theorem 2.3. It was extracted from the difficult terrain that is page 127 of Akemann and Pedersen's paper [2]. Approximate units are assumed to satisfy 0 ≤ uλ ≤ 1. If I is an ideal in A we let π : A → A/I denote the quotient map. Lemma 2.1. Suppose I ⊳ A. For any approximate unit uλ of I, any h in A+, and any real 0 ≤ δ ≤ 1, lim sup λ (cid:13)(cid:13)(cid:13)(1 − uλ) 1 1 2 h (1 − uλ) 2 + (1 − δ)u 1 2 λ hu Proof. We can lift π(h) to k with 0 ≤ k ≤ kπ(h)k . Setting x = h − k we have x in I and lim sup λ (cid:13)(cid:13)(cid:13)(1 − uλ) = lim sup λ 1 1 2 h (1 − uλ) 2 + (1 − δ)u 1 2 λ hu 1 (cid:13)(cid:13)(cid:13)(1 − uλ) Now we use the order structure in A and find 1 2 λ(cid:13)(cid:13)(cid:13) ≤ max (kπ(h)k , (1 − δ) khk) . 2 1 λ(cid:13)(cid:13)(cid:13) 1 2 k (1 − uλ) 2 + (1 − δ)u 1 2 λ hu 2 1 λ(cid:13)(cid:13)(cid:13) . 1 1 1 1 (1 − uλ) 2 k (1 − uλ) 2 + (1 − δ)u 2 λ hu 2 λ ≤ kkk (1 − uλ) + (1 − δ) khk uλ ≤ max (kπ(h)k , (1 − δ) khk) . Lemma 2.2. Suppose I ⊳ A. For any approximate unit uλ of I quasicentral for A, any a in A, and any real 0 ≤ δ ≤ 1, (cid:3) lim sup λ (cid:13)(cid:13)(cid:13)a (1 − δuλ) 1 2(cid:13)(cid:13)(cid:13) ≤ max(cid:16)kπ(a)k , (1 − δ) 1 2 kak(cid:17) . Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 6 Proof. Using the quasicentral property and Lemma 2.1, lim sup λ 2 (cid:13)(cid:13)(cid:13)a (1 − δuλ) 1 2(cid:13)(cid:13)(cid:13) = lim sup ka∗a (1 − δuλ)k λ = lim sup ka∗a (1 − uλ) + (1 − δ)a∗auλk λ = lim sup λ (cid:13)(cid:13)(cid:13)(1 − uλ) ≤ max (kπ(a∗a)k , (1 − δ) ka∗ak) = (max (kπ(a)k , (1 − δ) kak))2 . 1 2 a∗a (1 − uλ) 1 2 + (1 − δ)u 1 2 λ a∗au 1 2 λ(cid:13)(cid:13)(cid:13) Theorem 2.3. Suppose I ⊳ A, that uλ is a approximate unit uλ for I quasicentral for A, and a is in A. If f is a continuous function of [0, 1] so that (cid:3) then 1 = f (0) ≥ f (t) ≥ f (1) ≥ 0 lim sup kaf (uλ)k ≤ max (kπ(a)k , f (1) kak) . λ Proof. Let δ = 1 − f (1)2 and This function is continuous and g(t) = δ−1(cid:0)1 − f (t)2(cid:1) . 0 = g(0) ≤ g(t) ≤ g(1) = 1 so g (uλ) is also a quasicentral approximate unit. By the Lemma 2.2, lim sup λ kaf (uλ)k = lim sup 1 λ (cid:13)(cid:13)(cid:13)a (1 − δg (uλ)) 2(cid:13)(cid:13)(cid:13) 2 kak(cid:17) ≤ max(cid:16)kπ(a)k , (1 − δ) = max (kπ(a)k , f (1) kak) . 1 3. Lifting Softened Homogeneous Relations We will consider ∗-polynomials in infinitely many variables that are homogeneous in some finite subset of the variables. These we take to be the first r-variables, which we label x1, . . . , xr, and the remaining variables we label y1, y2, . . . . We also use the n-tuple notation x = (x1, . . . , xr) and y = (y1, y2, . . .) and with a NC ∗-polynomial p we use the notation (cid:3) p(x, y) = p(x1, . . . , xr, y1, y2, . . .). For scalar t we use We will say p is d-homogeneous in the first r variables if tx = (tx1, . . . , txr). p(tx, y) = tdp(x, y) for all real scalars t. In other words, in each monomial the xj and x∗ times. As d is not necessarily the degree of p we call d the degree of homogeneity of p. j appear collectively d Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 7 We will rather quickly run out of letters if we insist on other symbols when evaluating p on specific elements of a C ∗-algebra A. Given x1, . . . , xr in A and m in A or eA we define (We use eA to denote the unitization of A.) If ϕ : A → B then ϕ(x) = (ϕ(x1), . . . , ϕ(xr)) mx = (mx1, . . . , mxr). and ϕ(y) = (ϕ(y1), ϕ(y2), . . .). If z1, . . . , zr are in A then z ≤ y shall mean zj ≤ xj for j = 1, . . . , r. If 0 ≤ x then x 1 2 shall 1 denote(cid:16)x 2 1 , . . . , x 2 1 r(cid:17) . For a1, . . . , an and b1, . . . , bn in A we use the notation a • b = a1b1 + · · · + anbn. Recall π is our generic notation for the quotient map A → A/I. Theorem 3.1. Suppose p1, . . . , pJ are NC ∗-polynomials in infinitely many variables that are homogeneous in the first r variables, with each degree of homogeneity dj at least one. Suppose Cj > 0 are constants. For every C ∗-algebra A and I ⊳ A an ideal, given x1, . . . , xr and y1, y2, . . . in A with 0 ≤ x and kpj (π(x), π(y))k ≤ Cj, there are z1, . . . zr in A with 0 ≤ z ≤ x and π(z) = π(x) and (3.1) kpj (z, y)k ≤ Cj. Proof. Our proof is modeled on that from [2]. We start by performing the easier lifting where Cj in (3.1) is replaced by (1 + ǫ1) Cj. We pick ǫ1 later, but it will be positive. Since Cj is not allowed to be zero, (1 + ǫ1) Cj will be strictly larger than Cj. Let uλ be any approximate unit uλ for I that is quasicentral for A. By quasicentrality and the homogeneity in x, we have 1 2 • (1 − uλ) x lim λ (cid:13)(cid:13)(cid:13)pj(cid:16)x 1 2 , y(cid:17)(cid:13)(cid:13)(cid:13) = lim λ (cid:13)(cid:13)(cid:13)pj (x, y) (1 − uλ)dj(cid:13)(cid:13)(cid:13) kpj (x, y) (1 − uλ)k ≤ lim λ = kπ (pj (x, y))k = kpj (π (x) , π (y))k ≤ Cj. We define z(1) k = x 1 1 2 k (1 − uλ1) x k 2 where λ1 is large enough to give us (cid:13)(cid:13)pj(cid:0)z(1), y(cid:1)(cid:13)(cid:13) ≤ (1 + ǫ1) Cj for j = 1, . . . , J. Clearly 0 ≤ z(1) ≤ x and π(cid:0)z(1)(cid:1) = π (x) . k (cid:17) 1 (1 − δ2uλ2)(cid:16)z(1) k (cid:17) 1 k =(cid:16)z(1) We will create ever better lifts by defining z(2) 2 2 , Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 8 and so forth. For consistency, we let δ1 = 1 and z(0) = x. We choose δ1 > δ2 > . . . all positive with δc < ∞. ∞Xc=1 We set ǫ1 and the rest of a sequence ǫc by requiring Notice the ǫc are positive and decreasing to zero. Assume we have found z(1) through z(c−1) with (1 − δc+1) (1 + ǫc) = 1. (3.2) (3.3) (3.4) and (3.5) (3.6) Moreover, assume the z(w) have been constructed via the formula Theorem 2.3 tells us lim sup 0 ≤ z(c−1) ≤ z(c−2) · · · ≤ z(1) ≤ x (w = 1, . . . , c − 1) . π(cid:0)z(c−1)(cid:1) = π(cid:0)z(c−2)(cid:1) = · · · = π(cid:0)z(1)(cid:1) = π (x) (cid:13)(cid:13)pj(cid:0)z(w), y(cid:1)(cid:13)(cid:13) ≤ (1 + ǫw) Cj, 2 (1 − δwuλw)(cid:0)z(w−1)(cid:1) 1 2 , y(cid:17)(cid:13)(cid:13)(cid:13) 2 • (1 − δcuλ)(cid:0)z(c−1)(cid:1) 1 (cid:13)(cid:13)(cid:13)pj(cid:0)z(c−1), y(cid:1) (1 − δcuλ)dj(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)pj(cid:0)z(c−1), y(cid:1) (1 − δcuλ)(cid:13)(cid:13) z(w) =(cid:0)z(w−1)(cid:1) 1 (cid:13)(cid:13)(cid:13)pj(cid:16)(cid:0)z(c−1)(cid:1) 1 ≤ lim sup = lim sup 2 , λ λ λ (w = 1, . . . , c − 1) . (cid:13)(cid:13)(cid:13)pj(cid:16)(cid:0)z(c−1)(cid:1) 1 ≤ max (Cj, (1 − δc) ((1 + ǫc−1) Cj)) = Cj ≤ max(cid:0)(cid:13)(cid:13)π(cid:0)pj(cid:0)z(c−1), y(cid:1)(cid:1)(cid:13)(cid:13) , (1 − δc)(cid:13)(cid:13)pj(cid:0)z(c−1), y(cid:1)(cid:13)(cid:13)(cid:1) = max(cid:0)kpj (π (x) , π (y))k , (1 − δc)(cid:13)(cid:13)pj(cid:0)z(c−1), y(cid:1)(cid:13)(cid:13)(cid:1) 2 , y(cid:17)(cid:13)(cid:13)(cid:13) ≤ (1 + ǫc) Cj. 2 , and we may because 2 • (1 − δcuλc)(cid:0)z(c−1)(cid:1) 1 z(c) =(cid:0)z(c−1)(cid:1) 1 (cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(cid:13)(cid:16)z(c−1) ≤(cid:13)(cid:13)(cid:13)z(c−1) 2 (1 − δcuλc)(cid:0)z(c−1)(cid:1) 1 (δcuλc)(cid:16)z(c−1) (cid:17) 1 (cid:13)(cid:13)(cid:13) kδcuλck ≤ 2 kxkk δc. k k k k 2 2(cid:13)(cid:13)(cid:13)(cid:13) (cid:17) 1 k − z(c−1) (cid:13)(cid:13)(cid:13)z(c) so we may choose λc with We set and the construction continues. We wish to set zk = lim c z(c) j Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 9 Equations (3.3), (3.4) and (3.5) give us 0 ≤ z ≤ x and π(z) = x and finally the norm conditions kpj (z, y)k = lim c (cid:13)(cid:13)pj(cid:0)z(c), y(cid:1)(cid:13)(cid:13) ≤ Cj. (cid:3) If we have soft ∗-polynomial relations involving self-adjoint variables we can replace each by two positive variables. A variable that is a contraction can be replaced by four positive variables. These replacements will preserve any homogeneity in a subset of the variables. Thus we can have a more flexible version of Theorem 3.1. As stated, Theorem 3.2 it is not a corollary as we are very specific in how the lifts are adjusted for the different types of variables. Theorem 3.2. Suppose p1, . . . , pJ are NC ∗-polynomials in infinitely many variables that are homogeneous in the first r variables, with each degree of homogeneity dj at least one. Suppose Cj > 0 are constants. Suppose S = {1, . . . , r} is partitioned as S = S+ ∪ Sh ∪ Sg, we have positive constants Cj, nonnegative constants Dk, Ek, Fk and Gk, and consider the relations (3.7) (3.8) (3.9) (3.10) 0 ≤ xk ≤ Dk (k ∈ S+) Ek ≤ xk ≤ Fk (k ∈ Sh) kxkk ≤ Gk (k ∈ Sg) kpj (x, y)k ≤ Cj. For every C ∗-algebra A and I ⊳ A an ideal, given x1, . . . , xr and y1, y2, . . . in A so that (x, y) satisfies (3.7-3.9) and (π (x) , π (y)) satisfies (3.10), there are elements z1, . . . , zr in A so that (z, y) satisfy (3.7-3.10) and π (z) = π (x) . Moreover, it is possible to do so with 1 2 m2 (xk) 1 2 , (k ∈ S+) zk = (xk) zk = mxkm, zk = xkm2, (k ∈ Sh) (k ∈ Sg) for some m in 1 + I with 0 ≤ m ≤ 1. Proof. Let ǫc, δc and uλ be as before. We modify the construction used for Theorem 3.1 by requiring m0 = 1 and mc = (1 − δcuλc) mc−1 (1 − δcuλc) and We want 1 1 2 2 z(c) k m2 k = x cx k z(c) k = mcxkmc z(c) k = xkm2 c (k ∈ S+) (k ∈ Sh) (k ∈ Sg). (cid:13)(cid:13)pj(cid:0)z(c), y(cid:1)(cid:13)(cid:13) ≤ (1 + ǫc) Cj Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 10 given that we already have defined mc−1 in 1 + I to established (3.11) To unify the initial step and subsequent steps, we take ǫ0 large enough to force (3.11) when c = 1. We need to find the right λc to define z(c) j = w(λc) j where (cid:13)(cid:13)pj(cid:0)z(c−1), y(cid:1)(cid:13)(cid:13) ≤ (1 + ǫc−1) Cj. 1 2 k ((1 − δcuλ) mc−1 (1 − δcuλ))2 x w(λ) k = x w(λ) k = ((1 − δcuλ) mc−1 (1 − δcuλ)) xk ((1 − δcuλ) mc−1 (1 − δcuλ)) w(λ) k = xk ((1 − δcuλ) mc−1 (1 − δcuλ))2 (k ∈ S+) (k ∈ Sg) k 1 2 (k ∈ Sh) and see Therefore lim sup λ lim λ (cid:13)(cid:13)(cid:13)w(λ) (cid:13)(cid:13)pj(cid:0)w(λ), y(cid:1)(cid:13)(cid:13) = lim sup = lim sup λ j − (1 − δcuλ)4 z(c−1) j (cid:13)(cid:13)(cid:13) = 0. (cid:13)(cid:13)pj(cid:0)(1 − δcuλ)4 z(c−1), y(cid:1)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)pj(cid:0)z(c−1), y(cid:1) (1 − δcuλ)4dj(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)pj(cid:0)z(c−1), y(cid:1) (1 − δcuλ)(cid:13)(cid:13) λ ≤ lim sup λ ≤ max(cid:0)kpj (π (x) , π (y))k , (1 − δc)(cid:13)(cid:13)pj(cid:0)z(c−1), y(cid:1)(cid:13)(cid:13)(cid:1) ≤ Cj and it is possible to chose the needed λc. It is clear that mc stays in i + I, so these are all lifts of the original x. What is left to check is that m = lim c mc exists. Indeed it does, as kmc − mc−1k = k(1 − δcuλc) mc−1 (1 − δcuλc) − mc−1k ≤ k(δcuλc) mc−1 (1 − δcuλc)k + kmc−1 (δcuλc)k ≤ 2 kδcuλcmc−1k ≤ 2δc. (cid:3) We get from Theorem 3.2 a myriad of projective C ∗-algebras, simply by adding relations such as −1 ≤ yj ≤ 1 that are liftable and that impose a norm restriction forcing the universal C ∗-algebra to exist. We generally add the relation kyjk = 0 to most of the yj so as to be working with a finitely generated projective C ∗-algebra. 4. Soft Versions of Known Projectives Consider C0(X) where X + is a (finite) tree. The presentation in [17] for the projective C ∗-algebra C0(X) was based on a partial order (cid:22) on {1, . . . , s}. This was not a general partial order, it had to be the partial order on the non-root vertices determined by paths away from the root. Let us call such a relation a tree order. Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 11 The presentation associated to the tree order (cid:22) had generators h1, . . . , hs and relations 0 ≤ hj ≤ 1, hihj = hj, hihj = 0, (j = 1, . . . , s) (if i ≺ j) (if i 6(cid:22) j and j 6(cid:22) i). The last two lines of relations are not generally homogeneous in any subset of the variables. Taking advantage of the precise way the liftings are modified in our main theorem, we can still lift the soft version of these. The concrete function in C0(X) that corresonds to the abstract hj only varies along one egdge, the edge incident to the jth vertex on a path from that vertex to the root. From this point of view, it makes more sense to index the generators by the edges (as in [17]) but in graph theory, tree orders are on the vertices. We do need to make two changes to the relations. Some are redundant, since for positive elements and h1h2 = h2 & h3h1 = 0 =⇒ h3h2 = 0 h1h2 = h2 & h3h1 = h1 =⇒ h3h2 = h2. Let us swich to the indexing being over the non-root vertices. We then can speak of i being a child of j, meaning j (cid:22) i and j (cid:22) k (cid:22) i =⇒ k = j or k = i. The only relations we need are those that ask that the parent act as a unit on the child and that two children of the same parent must be orthogonal. We call two children of the same vertex siblings, of course. Children of children, and so forth, we call descendents. A second change is we replace hihj = hj by Theorem 4.1. Suppose (cid:22) is a tree order on {1, . . . , s}. The relations (hi − 1) hj (hi − 1) = 0. 0 ≤ hj ≤ 1, k(hi − 1) hj (hi − 1)k ≤ ǫ, khihjk ≤ ǫ, (j = 1, . . . , s) (if j is a child of i) (if i and j are siblings) are liftable. The following, slightly stronger result is more easily proven. Theorem 4.2. Suppose (cid:22) is a tree order on {1, . . . , s} and ǫ > 0. For every C ∗-algebra A and I ⊳ A an ideal, given h1, . . . , hs in A with 0 ≤ h ≤ 1 and kπ ((hi − 1) hj (hi − 1))k ≤ ǫ, kπ (hi) π (hj)k ≤ ǫ, (if j is a child of i) (if i and j are siblings) there are k1, . . . ks in A with 0 ≤ k ≤ h and π(k) = π(h) and k(ki − 1) kj (ki − 1)k ≤ ǫ, kkikjk ≤ ǫ, (if j is a child of i) (if i and j are siblings). Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 12 Proof. The very trivial base case for our proof by induction is the case of zero generators. Re-indexing, we may assume the minimal elements (i.e. elders or vertices closest to the root) in this partial order are {1, . . . , r}. If m and n are descendents of different minimal elements i and j then there are no relations involving both hm and hn. The relations not involving the minimal elements are a disjoint union of relations of the type in the statement of the theorem. The relations we need that involve the minimal elements are 0 ≤ hj ≤ 1 for j = 1, . . . , r and k(hi − 1) hm (hi − 1)k ≤ ǫ, kkikjk ≤ ǫ, (m is the child of i) (1 ≤ i < j ≤ r). These are homogeneous in {h1, . . . , hr}. By Theorem 3.1 there are k1, . . . , kr in A with 0 ≤ kj ≤ hj and π(hj) = π(hj) for j ≤ r and k(ki − 1) hm (ki − 1)k ≤ ǫ, kkikjk ≤ ǫ, (m is the child of i) (1 ≤ i < j ≤ r). The induction hypothesis tells us there are kr+1, . . . , ks with 0 ≤ km ≤ hm and π(km) = π(hm) for m > r with all the relations not involving indices {1, . . . , r}. We might have lost the relations between some hi and hm with m a child of i, but we have not, since (hi − 1) km (hi − 1) ≤ (hi − 1) hm (hi − 1) . (cid:3) A rather different example, with a similar proof, is a soft version of the projective C ∗- algebra considered in [19]. (For a detailed explanation of how the second relation is valid, see [20].) Theorem 4.3. For any positive ǫ, the C ∗-algebra hk = 0 k (cid:21) ≤ 1 + k (cid:21) ≤ 1 + x x khkk ≤ ǫ 0 ≤(cid:20) 1 − h x∗ C ∗h, k, x(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) C ∗h, k, x(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 0 ≤(cid:20) 1 − π (h) π (x)∗ 0 ≤(cid:20) 1 − h x∗ kπ (h) π (k)k ≤ ǫ π (x) π (k) (cid:21) ≤ 1. Proof. Suppose h, k and x are in A, which we may assume is unital, are such that is projective. and We know positive contractions lift to positive contractions from M2(A/I) to M2(A) and so we can find h, k and x in A so that π(cid:16)h(cid:17) = π (h) , π(cid:16)k(cid:17) = π (k) , π (x) = π (x) and 0 ≤(cid:20) 1 − h x∗ k (cid:21) ≤ 1. x Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 13 The polynomial hk is homogeneous in k so Theorem 3.2 tells us there is 0 ≤ m ≤ 1 in 1 + I so that(cid:13)(cid:13)(cid:13)hmkm(cid:13)(cid:13)(cid:13) ≤ ǫ. Let ¯h = h, ¯x = mx and ¯k = mkm. These are still lifts of h, x and k, and now(cid:13)(cid:13)¯h¯k(cid:13)(cid:13) ≤ ǫ and 0 (cid:20) 1 − ¯h ¯x∗ ¯h ¯h (cid:21) =(cid:20) 1 0 ≤(cid:20) 1 − ¯h ¯x∗ 0 m (cid:21)(cid:20) 1 − h x∗ ¯h (cid:21) ≤ 1. ¯h x k (cid:21)(cid:20) 1 0 m (cid:21) 0 implies (cid:3) 5. Fattened Curves in Various NC unit balls Theorem 5.1. Suppose p1, . . . , pJ are NC ∗-polynomials in x1, . . . , xs. Suppose 1 ≤ r ≤ s and each pj is homogeneous in {x1, . . . , xr} with degree of homogeneity dj ≥ 1. For ǫ > 0, the C ∗-algebra is projective. Proof. This is immediate consequence of Theorem 3.2. For a single NC ∗-polynomial p, we can think of (5.1) (5.2) as a approximate zero locus of a NC curve intersected with the NC unit square. Likewise we can think of (cid:3) kxkk ≤ 1 kxkk ≤ 1, k = 1, . . . , s kpj (x1, . . . , xs)k ≤ ǫ, j = 1, . . . , J (cid:29) Aǫ = C ∗(cid:28)x1, . . . , xs(cid:12)(cid:12)(cid:12)(cid:12) C ∗(cid:28)x1, . . . , xs(cid:12)(cid:12)(cid:12)(cid:12) C ∗x1, . . . , xs(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) kxk(cid:13)(cid:13)(cid:13) ≤ 1 implies kxkk ≤ 1 so we can still apply Theorem 5.1. kp (x1, . . . , xs)k ≤ ǫ (cid:29) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kp (x1, . . . , xs)k ≤ ǫ + sXk=1 kxk(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ 1 x∗ as a approximate zero locus of a NC curve with the NC unit ball. Notice that the "row contraction" condition(cid:13)(cid:13)(cid:13)X x∗ For ǫ > 0, and with p homogeneous in x1, . . . , xs we find (5.1) and (5.2) define projective C ∗-agebras. We will see that it is possible to work with other unit balls, not just the ones corresponding to the ℓ2 and ℓ∞ norms. Lemma 5.2. Suppose 0 < α < ∞ is a scalar. For every ǫ > 0 there is a δ > 0 so that for any two positive contractions in any C ∗-algebra, khk − khk ≤ δ =⇒ (cid:13)(cid:13)(hkh)α − kαh2α(cid:13)(cid:13) ≤ ǫ. Proof. This can be rephrased so it becomes a special case of Lemma 10 of [19], but it is easier to just revise the proof. We know for nonnegative scalars (xyx)α = yαx2α so by spectral theory, hk = kh =⇒ (hkh)α = kαh2α. Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 14 If the lemma is false, there must be some ǫ0 and hn and kn in An with 0 ≤ hn ≤ 1 and 0 ≤ kn ≤ 1 and and This creates an element in with hk = kh and (hkh)α − kαh2α 6= 0, a contradiction. (cid:3) Theorem 5.3. Suppose r is a natural number. For 0 < p < ∞ define If 0 < p ≤ ∞ then Bp is projective. Proof. Suppose I ⊳ A with quasicentral approximate unit uλ. Suppose xk are in A with π (xk) in A/I satisfying ≤ 1+ 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) khnkn − knhnk ≤ 1 n p nh2α (xkx∗ k) rXk=1 n (cid:13)(cid:13) ≥ ǫ0. (cid:13)(cid:13)(hnknhn)α − kα Y An.M An (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Bp = C ∗x1, . . . , xr(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k (cid:17)∗(cid:17) p rXk=1(cid:16)z(1) p(cid:17) < ∞. ∞Xc=1(cid:16)1 − (1 − δc) k (cid:16)z(1) (π (xk) (π (xk))∗) 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p 1 ≤ 1. ≤ 1 + ǫ0. (1 − δc+1) (1 + ǫc) = 1. Let z(0) k = xk and let ǫ0 be sufficiently large so as to have Choose δc a positive sequence decreasing to zero with δ1 = 1 and Define ǫc ց 0 for c ≥ 1 by the formula Assume we have found z(c−1) with and Using Lemma 5.2 and Theorem 2.3 we find ≤ 1 + ǫc. k (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1(cid:16)z(c−1) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1(cid:16)(1 − δcuλ) = lim sup λ ≤ 1. k π(cid:0)z(c−1)(cid:1) = π (x) 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:17)∗(cid:17) p (cid:16)z(c−1) (cid:17)∗ (cid:16)z(c−1) (cid:16)z(c−1) rXk=1(cid:16)z(c−1) p z(c−1) k k k k 1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (1 − δcuλ) 2 (cid:17)∗(cid:17) p 1 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p(cid:17) p (1 − δcuλ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) lim sup λ Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 15 We can chose λc and set z(c) k = (1 − δcuλc) 1 p z(c−1) k where λc is large enough to ensure The z(c) j converge because ≤ 1 + ǫc. (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) k (cid:17)∗(cid:17) p k (cid:16)z(c) rXk=1(cid:16)z(c) (cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(1 − δcuλc) ≤(cid:13)(cid:13)(cid:13)(1 − δcuλc) ≤(cid:16)1 − (1 − δc) 1 1 k − z(c−1) k (cid:13)(cid:13)(cid:13)z(c) p z(c−1) k − z(c−1) k k p − 1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)z(c−1) p(cid:17) kxkk . 1 (cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) (cid:3) We still cannot tell if all the Bp are isomorphic. They do interact with homogeneous ∗- polynomials in about the same fashion as the usual unit ball. The set of NC ∗-polynomials that we know we can mix with the nonstandard unit ball condition depends on p. We have no idea if this is a limitation of our methods, or a real limitation. Theorem 5.4. Suppose r ≤ s and p1, . . . , pJ are NC ∗-polynomials in x1, . . . , xs, each homo- geneous in {x1, . . . , xr} with degree of homogeneity dj at least one. For Cj > 0 and 0 < q ≤ 2 the C ∗-algebra kpj (x1, . . . , xs)k ≤ Cj, j = 1, . . . , J + sXk=1 (xkx∗ k) ≤ 1 q 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Aǫ = C ∗x1, . . . , xs(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) sXk=1 is projective. and Proof. Suppose we are given π : A → A/I with x1, . . . , xs in A with We first apply Theorem 5.3 to find y1, . . . , ys in A with π(yk) = π(xk) and kpj (π (x1) , . . . , π (xs))k ≤ Cj. (π (xkx∗ k)) ≤ 1 q 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) q (yky ∗ k) ≤ 1. sXk=1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Theorem 3.2 gives us z1, . . . , zs in A with π(zj) = π(xj) and but also with zk = ykm for k ≤ r and zk = yk for k > r, where 0 ≤ m ≤ 1. Therefore kpj (z1, . . . , zs)k ≤ Cj, zkz∗ k = ykm2y ∗ k ≤ yky ∗ k Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 16 for 1 ≤ k ≤ r. Since for q ≤ 2 the function tq/2 is operator-monotone we get (zkz∗ k) q 2 ≤ sXk=1 sXk=1 (yky ∗ k) q 2 ≤ 1. (cid:3) Theorem 5.5. Suppose p1, . . . , pJ are homogeneous, degree-dj NC ∗-polynomials in x1, . . . , xr with dj ≥ 1. For Cj > 0 and 2 < q < ∞ the C ∗-algebra (xkx∗ k) ≤ 1 kpj (x1, . . . , xr)k ≤ Cj, j = 1, . . . , J (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1 q 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) + Aǫ = C ∗x1, . . . , xr(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1 is projective. and (π (xkx∗ k)) ≤ 1 q 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kpj (π (x1) , . . . , π (xr))k ≤ Cj. Proof. Suppose we are given π : A → A/I with x1, . . . , xr in A with Choose δc and ǫc as before, with the δc summable. Keeping with our earlier notation, we are going to define z(c) from z(c−1) by Lemma 5.2 and Theorem 2.3 give us k = (1 − δcuλc) z(c−1) z(c) k . lim sup λ (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) and k = lim sup (cid:16)z(c−1) rXk=1(cid:16)(1 − δcuλ) z(c−1) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1(cid:16)z(c−1) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) rXk=1(cid:16)z(c−1) ≤ lim sup = 1 λ λ lim sup λ (cid:13)(cid:13)pj(cid:0)(1 − δ1uλ) z(c−1)(cid:1)(cid:13)(cid:13) = lim sup λ ≤ lim sup λ = Cj. 2 2 k k k k k 2(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:17)∗ (1 − δcuλ)(cid:17) q (1 − δcuλ)q(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:17)∗(cid:17) q (cid:16)z(c−1) (1 − δcuλ)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:17)∗(cid:17) q (cid:16)z(c−1) (cid:13)(cid:13)(cid:13)pj(cid:0)z(c−1)(cid:1) (1 − δcuλ)dj(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)pj(cid:0)z(c−1)(cid:1) (1 − δcuλ)(cid:13)(cid:13) Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 17 The limit of the z(c) j will exist because k − z(c−1) k (cid:13)(cid:13)(cid:13)z(c) − z(c−1) k k (cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)(1 − δcuλc) z(c−1) ≤ k(1 − δcuλc) − 1k(cid:13)(cid:13)(cid:13)z(c−1) ≤ δc kxkk . k (cid:13)(cid:13)(cid:13) (cid:3) (cid:13)(cid:13)(cid:13) 6. Soft Cylinders When we stray from homogeneous relations, we come across K-theoretical obstructions to projectivity. To illustrate what properties can still hold, we offer the example of the "soft cylinder." The weaker properties are semiprojectivity (as in [5]) and the RFD property, meaning "residually finite dimensional." Projectivity implies semiprojectivity and also RFD ([22, §1]). For ǫ ≥ 0 we define the soft cylinder almost like Exel's soft torus ([14]): Notice we retained some homogeneity. Aǫ = C ∗ 1(cid:10)u, h(cid:12)(cid:12) u∗u = uu∗ = 1, −1 ≤ h ≤ 1, kuh − huk ≤ ǫ(cid:11) . Theorem 6.1. For positive ǫ, the soft cylinder Aǫ is semiprojective. Proof. Suppose B is a unital C ∗-algebra, with ideal I =S I n for some increasing sequence of ideals In. Suppose we are given u and h in B/I where u is unitary, −1 ≤ h ≤ 1 and kuh − huk ≤ ǫ. For some n it is possible to lift u to v in B/In that is a unitary ([5, Prop. 2.21]). Take any lift of h to −1 ≤ k ≤ 1 in B/In. Theorem 3.2 tells us there is k in A/I with −1 ≤ k ≤ 1 and (cid:13)(cid:13)(cid:13)vk − kv(cid:13)(cid:13)(cid:13) ≤ ǫ. (cid:3) Eilers and Exel ([11]) have shown that the soft torus is RFD. The same can be said, and proven much more easily, for the soft cylinder. Theorem 6.2. For positive ǫ, the soft cylinder Aǫ is RFD. Proof. Consider the surjection that sends the obvious unitary generator to u and the obvious positive, norm-one generator to h. By [15, Theorem 3.2] the free product is RFD. Theorem 3.2 tells us that ρ is split. Thus Aǫ can be embedded in an RFD C ∗-algebra and so is itself RFD. (cid:3) ρ : C(S1) ∗C C[0, 1] ։fAǫ Our lifting theorems can be used to determine many more C ∗-algebras are RFD. The study of weak projectivity ([21]) and RFD of the C ∗-algebras associated to rather general relations that have some homogeneity might lead to some interesting examples. These topics will be explored elsewhere. Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 18 7. Cones are Limits of Projective C ∗-Algebras We end with a tantalizingly result: every cone is the limit of projectives. As a C ∗-algebra with a projective cone must be semiprojective ([4, II.8.3.10]) it would seem that we are close to proving that every separable C ∗-algebra is a limit of semiprojective C ∗-algebras. We say definitively, projectivity is not "extremely rare" ([18, p. 73]). Lemma 7.1. Suppose A is the unital C ∗-algebra where the p1, p2 . . . are NC polynomials in the xk of degrees Dj with zero constant term. Then the cone CA has presentation A = C ∗ 1(cid:28)x1, x2, . . .(cid:12)(cid:12)(cid:12)(cid:12) CA = C ∗h, x1, x2, . . .(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) −Ck ≤ xk ≤ Ck (∀k) pj (x) = 0 (∀j) (cid:29) 0 ≤ h ≤ 1 hxk = xkh (∀k) −Ckh ≤ xk ≤ Ckh (∀k) qj (h, x) = 0 (∀j) + , where qj is the NC polynomial derived from the pj by padding monomials on the left with various powers of h so that qj is homogeneous with degree Dj. Proof. To illustrate the construction of the qj, if then p1 = x1 + 3x1x∗ 2x1 q1 = h2x1 + 3x1x∗ 2x1. In general, we can break up pj into homogeneous summands and then describe the qj as qj (h, x) = hDj −dpj,d (x) . pj = pj,d DjXd=1 DjXd=1 Let the universal C ∗-algebra for these relations be denoted U. This exists, as the relations satisfy the needed four axioms as in [20]. One of the axioms is that setting all variables to the zero elements in {0} leads to a representation of the relations, which is true because we require the constant terms to be zero. To define a ∗-homomorphism U → CA we define in CA = C0 ((0, 1], A) elements xk = txk and h = t, shorthand for x(t) = tx and so forth. It is obvious that 0 ≤ h ≤ 1 and that h commutes with each xk. Also −Ck ≤ xk ≤ Ck =⇒ −tCk ≤ txk ≤ tCk Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 19 so −Cjh ≤ xj ≤ Cjh. The last relation holds as well since (cid:16)qj(cid:16)h, x(cid:17)(cid:17) (t) = DjXd=1 = tDj tDj −dpj,d (tx) pj,d (x) DjXd=1 = tDj pj (x) = 0. Next we will show this map is onto. Basic algebra, and the usual isomorphism of CA with C0(0, 1] ⊗ A, tells us that functions of the form tmw generate the cone, where w ranges over words in the xk. Suppose w = w1w2 · · · wn. If m ≥ n then this is easily in the image, as tmw = tm−n (tw1) (tw2) . . . (twn) . If 1 ≤ m < n then the Stone-Weierstrass theorem tells us we can approximate in C0(0, 1] the function tm by a polynomial in tn, tn+1, . . . and so can approximate tmw by a polynomial in tnw, tn+1w, . . . and the map is indeed onto. We turn to proving it is one-to-one. Consider an irreducible representation in B() of the relations defining U by H and X1, X2, . . . . Since 0 ≤ H ≤ 1 and HXk = XkH and we find that H is central and so H = λI for some scalar λ with 0 ≤ λ ≤ 1. If λ = 0 then H = 0 and −Ckλ ≤ Xk ≤ Ckλ =⇒ Xk = 0. This is the zero representation, which is the pullback of the zero representation of CA. If λ is positive, then −Ckλ ≤ Xk ≤ Ckλ =⇒ −Ck ≤ λ−1Xk ≤ Ck and qj (H, X) = 0 implies pj(cid:0)λ−1X(cid:1) = DjXd=1 λ−dpj,d (X) = λ−Dj H Dj −dpj,d (X) DjXd=1 = λ−Dj qj (H, X) = 0. Thus the λ−1Xk form a representation of A on and so a representation of CA via the composition CA δλ / A B(). This sends h to λI = H and xk to Xk, finishing the proof. (cid:3) Theorem 7.2. If q1, q2 . . . are homogeneous NC polynomials, each of degree at least one, in noncommuting variables h, x1, x2, . . . then for positive constants C1, . . . , CJ and D1, . . . , DK, / / / Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 20 the C ∗-algebra is projective. C ∗h, x1, x2, . . .(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 0 ≤ h ≤ 1, −Ckh ≤ xk ≤ Ckh, kqj (h, x1, x2, . . .)k ≤ Dj (1 ≤ j ≤ J) + (1 ≤ k ≤ K) Proof. We can find some r so that xr+1, xr+2, . . . are not in any of the polynomials q1, . . . , qJ . If we relabel these y1, y2, . . . our lifting problem becomes 0 ≤ h ≤ 1, −Ckh ≤ xk ≤ Ckh, −C ′ kh, kh ≤ yk ≤ C ′ kqj (h, x)k ≤ Dj (1 ≤ j ≤ J) where now the qj are homogeneous in {h, x1, . . . , xr} . We are using x for (x1, . . . , xr) . Given h, xk and yk in A with 0 ≤ π (h) ≤ 1, −Ckπ (h) ≤ π (xk) ≤ Ckπ (h) , −C ′ kπ (h) , (1 ≤ j ≤ J) kπ (h) ≤ π (yk) ≤ C ′ kqj (π (h) , π (x))k ≤ Dj we first find a new lift h of π(h) with 0 ≤ h ≤ 1. Using Davidson's two-sided order lifting theorem ([8]) we find xk and yk with −Ckh ≤ xk ≤ Ckh, h −C ′ k h ≤ yk ≤ C ′ k and π (xk) = π (xj) and π (yk) = π (yj) . By Theorem 3.2 there is an m with 0 ≤ m ≤ 1 in 1 + I so that Our desired lifts are mhm, mxm and mym. Lemma 7.3. Let D be a separable C ∗-algebra. Then (cid:13)(cid:13)(cid:13)qj(cid:16)mhm, mxm(cid:17)(cid:13)(cid:13)(cid:13) ≤ Dj. (cid:3) (∀j) (∀k) (cid:29) for a countable collection of NC polynomials. D ∼= C ∗(cid:28)x1, x2, . . .(cid:12)(cid:12)(cid:12)(cid:12) −Cj ≤ xj ≤ Cj pk (x1, x2, . . .) = 0 Proof. Example 1.3(b) in [5] tells us that D has a presentation with countably many gener- ators, countably many relations in the form of a NC ∗-polynomial set to zero and countably many norm conditions. We will modify Blackadar's method a bit. Let F = + i , which is a countable dense subfield of C. Select a countable dense sequence in D and apply to this sequence all polynomials over F in countably many variables. This results in a countable, dense F-∗-subalgebra B of D. Enumerate B as x1, x2, . . . . The algebraic operations for B can be encoded in ∗-polynomial relations. For example, if αxj = xk for some α in F, then we use the relation αxj − xk = 0. If x∗ j = xk then we use the relation x∗ j − xk = 0, and so forth. This means B is the universal F-∗-algebra for generators x1, x2, . . . Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 21 and some countable set of ∗-polynomial relations pj(x1, x2, . . .) = 0. We now add to these relations the C ∗-relations kxkk ≤ Ck where Ck is the norm of the element xk in D. Then any function f : B → G, for G a C ∗-algebra, that satisfies these relations is first of all an F-linear ∗-algebra homomorphism. It is continuous with respect to the norm on D since xj − xk will equal some xℓ so we have the relation kf (xℓ)k ≤ kxℓk and so kf (xj) − f (xk)k = kf (xℓ)k ≤ kxℓk = kxj − xkk . It therefore extends to a continuous function ϕ : D → G. This extended function will be linear over C. To verify this, consider α = lim αn, a limit of scalars from F, and d = lim dj, a limit of elements in B. Then ϕ(αnd) = ϕ(lim j αndj) = lim j f (αndj) = αn lim j f (dj) = αnϕ(d) and ϕ(αd) = ϕ(lim n αnd) = lim n ϕ(αnd) = lim n αnϕ(d) = αϕ(d). Finally, continuity implies that ϕ Is a ∗-homomorphism. It is uniquely determined by f and so D is universal for these relations. We can eliminate many of the norm conditions. Suppose we keep only the norm restrictions kxkk ≤ Ck for those xk that are self-adjoint. Then the estimate that gave continuity changes a little. Any xj − xk will equal some xℓ and for some r and s we will have xr = 1 2x∗ ℓ and xs = −i ℓ . Of course these are the real and imaginary part of xℓ, and as they are self-adjoint we have the relations kf (xr)k ≤ kxrk and kf (xs)k ≤ kxsk . Therefore 2 xℓ − i 2 xℓ + 1 2x∗ kf (xj) − f (xk)k = kf (xr) + if (xs)k ≤ kf (xr)k + kf (xs)k ≤ kxrk + kxsk ≤ 2 kxℓk = 2 kxj − xkk . This still gives us continuity and so the rest of the proof goes through. We can toss the generators that are not self-adjoint if we modify each polynomial by the evaluating xk at xr + ixs whenever xr and xs are the real and imaginary parts of xk. Among the polynomial relations will be x∗ j − xj = 0 for the generators we are keeping. Given this, it is our option to use the relation kxjk ≤ Cj or −Cj ≤ xj ≤ Cj. (cid:3) Theorem 7.4. If A is a separable C ∗-algebra then its cone CA is isomorphic to the inductive limit of a countable system of projective C ∗-algebras with surjective bonding maps. Proof. We start with the case where A = eD for some separable, possibly unital C ∗-algebra. Lemma 7.3 tells us −Ck ≤ xk ≤ Ck pj (x1, x2, . . .) = 0 (∀j) (cid:29) (∀k) 0 ≤ h ≤ 1 hxk = xkh (∀k) −Ckh ≤ xk ≤ Ckh (∀k) qj (h, x1, x2, . . .) = 0 (∀j) + and then Lemma 7.1 tells us A ∼= C ∗ 1(cid:28)x1, x2, . . .(cid:12)(cid:12)(cid:12)(cid:12) CA ∼= C ∗h, x1, x2, . . .(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 22 where the qk are homogeneous. Clearly CA ∼= lim −→ Pn where Pn = C ∗*h, x1, x2, . . .(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 0 ≤ h ≤ 1 −Ckh ≤ xk ≤ Ckh, (∀k) khxk − xkhk ≤ 1 n , kqj (x1, x2, . . .)k ≤ 1 n (k = 1, . . . , n) (j = 1, . . . , n) + . Since the commutators are homogeneous NC polynomials, Theorem 7.2 applies and the Pn are projective. We are done for CA = C(cid:16)eD(cid:17) . What about CD? We have the exact sequence 0 / CD / CA / CC / 0. Of course CC equals C0(0, 1] and is projective. Let Qn be the kernel of the map of Pn onto C0(0, 1] that sends h to t 7→ t and xk to zero. Then we have 0 0 0 / Qn Pn CC / Qn+1 Pn+1 CC / CD / CA / CC / 0 / 0 / 0 with the rows exact. Also, CD is isomorphic to lim −→ Qn, which we can see as follows. There is a ∗-homomorphism ϕ : lim −→ Qn → CD induced by the maps Qn → CD. The maps Qn → Pn are inclusions and hence isometries. Pn is an also an isometry. Theorem 13.1.2.2 in [18] implies that the induced map lim From the commutative diagram Qn → lim −→ −→ lim −→ Qn lim −→ Pn ϕ CD ∼= / CA we conclude ϕ is injective. As to surjectivity, consider x in CD. This gets sent to 0 in CC. Any lift of x to y in P1 is also sent to zero in CC, and so is Qn. This shows ϕ is onto. By Theorem 5.3 of [22] the Qn are projective. (cid:3) References [1] Charles A. Akemann. Left ideal structure of C ∗-algebras. J. Functional Analysis, 6:305 -- 317, 1970. [2] Charles A. Akemann and Gert K. Pedersen. Ideal perturbations of elements in C ∗-algebras. Math. Scand., 41(1):117 -- 139, 1977. [3] William Arveson. Subalgebras of C ∗-algebras. III. Multivariable operator theory. Acta Math., 181(2):159 -- 228, 1998. / / / / / / / / / / / / / / / / / / / / / / / Noncommutative Semialgebraic sets and Associated Lifting Problems (version 13) 23 [4] B. Blackadar. Operator algebras, volume 122 of Encyclopaedia of Mathematical Sciences. Springer- Verlag, Berlin, 2006. Theory of C ∗-algebras and von Neumann algebras, Operator Algebras and Non- commutative Geometry, III. [5] Bruce Blackadar. Shape theory for C ∗-algebras. Math. Scand., 56(2):249 -- 275, 1985. [6] Karol Borsuk. Theory of shape. PWN -- Polish Scientific Publishers, Warsaw, 1975. Monografie Matem- atyczne, Tom 59. [7] Alex Chigogidze and Alexander N. Dranishnikov. Which compacta are noncommutative ARs? Topology Appl, to appear. [8] Kenneth R. Davidson. Lifting positive elements in C ∗-algebras. Integral Equations Operator Theory, 14(2):183 -- 191, 1991. [9] Kenneth R. Davidson and David R. Pitts. Nevanlinna-Pick interpolation for non-commutative analytic Toeplitz algebras. Integral Equations Operator Theory, 31(3):321 -- 337, 1998. [10] Edward G. Effros and Jerome A. Kaminker. Homotopy continuity and shape theory for C ∗-algebras. In Geometric methods in operator algebras (Kyoto, 1983), volume 123 of Pitman Res. Notes Math. Ser., pages 152 -- 180. Longman Sci. Tech., Harlow, 1986. [11] Søren Eilers and Ruy Exel. Finite-dimensional representations of the soft torus. Proc. Amer. Math. Soc., 130(3):727 -- 731 (electronic), 2002. [12] Søren Eilers and Terry A. Loring. Computing contingencies for stable relations. Internat. J. Math., 10(3):301 -- 326, 1999. [13] Søren Eilers, Terry A. Loring, and Gert K. Pedersen. Stability of anticommutation relations: an appli- cation of noncommutative CW complexes. J. Reine Angew. Math., 499:101 -- 143, 1998. [14] Ruy Exel. The soft torus and applications to almost commuting matrices. Pacific J. Math., 160(2):207 -- 217, 1993. [15] Ruy Exel and Terry A. Loring. Finite-dimensional representations of free product C ∗-algebras. Internat. J. Math., 3(4):469 -- 476, 1992. [16] Stanislaw Lojasiewicz. Triangulation of semi-analytic sets. Ann. Scuola Norm. Sup. Pisa (3), 18:449 -- 474, 1964. [17] Terry A. Loring. Projective C ∗-algebras. Math. Scand., 73(2):274 -- 280, 1993. [18] Terry A. Loring. Lifting solutions to perturbing problems in C ∗-algebras, volume 8 of Fields Institute Monographs. American Mathematical Society, Providence, RI, 1997. [19] Terry A. Loring. A projective C ∗-algebra related to K-theory. J. Funct. Anal., 254(12):3079 -- 3092, 2008. [20] Terry A. Loring. C ∗-algebra relations. Math. Scand., to appear. http://arxiv.org/abs/0807.4988. [21] Terry A. Loring. Weakly projective C ∗-algebras. Rocky Mountain J. Math., to appear. http://arxiv.org/abs/0905.1520. [22] Terry A. Loring and Gert K. Pedersen. Projectivity, transitivity and AF-telescopes. Trans. Amer. Math. Soc., 350(11):4313 -- 4339, 1998. [23] Catherine L. Olsen and Gert K. Pedersen. Corona C ∗-algebras and their applications to lifting problems. Math. Scand., 64(1):63 -- 86, 1989. [24] Tatiana Shulman. Lifting of nilpotent contractions. Bulletin of the London Mathematical Society, 40(6):1002, 2008. [25] Viktor A. Vassiliev. Applied Picard-Lefschetz theory, volume 97 of Mathematical Surveys and Mono- graphs. American Mathematical Society, Providence, RI, 2002. Department of Mathematics and Statistics, University of New Mexico, Albuquerque, NM 87131, USA. Department of Mathematics, University of Copenhagen, Universitetsparken 5, DK-2100 Copenhagen Ø, Denmark
1003.1650	2	1003	2010-06-06T01:51:06	A note on some group $C^*$-algebras which are quasi-directly finite	[ "math.OA", "math.FA" ]	An algebra is said to be quasi-directly finite when any left-invertible element in its unitization is automatically right-invertible. It is an old observation of Kaplansky that the von Neumann algebra of a discrete group has this property; in this note, we collate some analogous results for the group $C^*$-algebras of more general locally compact groups. Partial motivation comes from earlier work of the author on the phenomenon of empty residual spectrum for convolution operators.	math.OA	math	A note on some group C∗-algebras which are quasi-directly finite Yemon Choi October 29, 2018 Abstract An algebra is said to be quasi-directly finite when any left-invertible element in its uni- tization is automatically right-invertible. It is an old observation of Kaplansky that the von Neumann algebra of a discrete group has this property; in this note, we collate some analogous results for the group C∗-algebras of more general locally compact groups. Partial motivation comes from earlier work of the author on the phenomenon of empty residual spectrum for convolution operators. MSC 2010: Primary 22D25; Secondary 46L05 1 Introduction Following Munn [8], we say that an algebra R is quasi-directly finite if its unitization R♯ is directly finite; that is, if every left invertible element in R♯ is automatically right invertible. In the present article, we consider the question of when the reduced or full C∗-algebras of a locally compact group are quasi-directly finite. Apart from intrinsic interest, this question is motivated by arguments in the author's previous article [2], where attention was mostly confined to the case of discrete groups. In this case, the group von Neumann algebras are known to be directly finite; one can exploit this to show that, in a variety of settings, convolution operators associated to actions of discrete groups have empty residual spectrum. As a hint that the situation will be subtler than in the discrete case, consider the reduced group C∗-algebra and the group von Neumann algebra of SL(2, R), denoted by C∗ r(SL(2, R)) and VN(SL(2, R)) respectively. Both algebras have been studied in some detail, and it follows from standard facts about them, that while VN(SL(2, R)) is not directly finite, the unitization of C∗ r(G) is quasi-directly finite if G is unimodular. r(SL(2, R)) is. More generally, we shall see below (Theorem 3.2) that C∗ There seems to be no known characterization, in terms of G, of when C∗ r(G) is quasi-directly finite. One of our secondary aims is to collate some of the relevant partial results in one place for ease of reference, as a precursor to possible further work. Residual spectrum and directly finite algebras In order to give some background and motivation to what follows, we shall outline the link between the residual spectrum of an operator and direct finiteness of an algebra which contains it. Recall that if T is a bounded linear operator on a Banach space X, then the residual spectrum of X is the set of all λ ∈ C such that T − λI is injective with closed, proper range; put slightly differently, it is the set of all λ in the spectrum of T which are not approximate or genuine eigenvalues. Thus, knowing in advance that the residual spectrum of T is empty is of interest in working out the spectral theory of T . 1 Let us say that a subalgebra A ⊆ B(X) is surjunctive if every a ∈ A has empty residual spectrum (note that this is, a priori, not just a property of the algebra A as an abstract algebra, but a property of A and its realization inside B(X)). If A is surjunctive, then a short argument shows that the subalgebra of B(X) generated by A and the identity operator must be directly finite; the converse is in general false, as illustrated by a construction of G. A. Willis (see [2, Proposition 4.1] and the surrounding remarks for some discussion of this). On the other hand, under extra conditions, we do get a partial converse: Theorem 1.1. Let A ⊆ B(X) be a closed subalgebra. Suppose that A is bicontinuously isomorphic to a quasi-directly finite C∗-algebra. Then A is surjunctive. Proof sketch. This is essentially proved in [2], but the arguments are somewhat scattered and are moreover restricted to the case where A is unital. Thus, for sake of convenience, we sketch how the argument goes. Let a ∈ A, λ ∈ C, and suppose a − λI is injective but not invertible. It suffices to find a sequence (bn) in A such that k(a − λI)bnk → 0 while inf n kbnk > 0; for, given such a sequence, it follows that there is a sequence (xn) in X of approximate eigenvectors for a corresponding to λ, showing that λ is not in the residual spectrum. It will follow from Lemma 2.6 below that since A is quasi-directly finite, the subalgebra Aun ⊆ B(X) that is generated by A and I will be directly finite; moreover, Aun is bicontinu- ously isomorphic to a unital, directly finite C∗-algebra. Note also that left multiplication by a − λI must be an injective operator from A to itself. The existence of a sequence (bn) with the required properties now follows exactly as in the proof of [2, Theorem 3.2]. One way to obtain examples satisfying the conditions of Theorem 1.1 is as follows. Let M be a semifinite von Neumann algebra, equipped with a faithful, normal, semifinite trace τ on M, let 1 ≤ p ≤ ∞, and let Lp(M, τ ) be the noncommutative Lp-space associated to (M, τ ), as constructed in e.g. [3]; then Lp(M, τ ) has the structure of an M-bimodule in a natural way, and the corresponding homomorphism ıτ : M → B(Lp(M, τ )) is injective with closed range. Thus, if A is a quasi-directly finite C∗-subalgebra of M, Theorem 1.1 implies that ıτ (A) will be a surjunctive subalgebra of B(Lp(M, τ )). By a result of Dixmier [4], VN(G) is semifinite for every connected locally compact group G, and so the discussion of the previous example can be applied. It is worth noting that for any locally compact group G, there is a natural action of VN(G) on the Fourier algebra A(G) (see [6] for the definition). It is not hard to show that the corresponding homomorphism ıA : VN(G) → B(A(G)) is an isometry, giving us another possible source of examples. Remark 1.2. Given a von Neumann algebra M, its predual M∗ has a natural M-bimodule structure. Moreover: (i) if M has a faithful normal semifinite trace τ , then it can be shown that there is an isomor- phism of M-bimodules from L1(M, τ ) onto M∗; (ii) if M = VN(G) for some locally compact group G, then there is an isomorphism of M- bimodules from A(G) onto VN(G)∗. Thus if VN(G) is semifinite, with a trace τ , this gives us two ways to view the action of VN(G) on its predual. 2 2 Notation and preliminaries It is convenient, in phrasing some of our results, to use the notions of left and right quasi-inverses. Definition 2.1. Given a, b ∈ A, let a • b := a + b − ab. If a • b = 0 then we say that a is a left quasi-inverse for b and b is a right quasi-inverse for a. The idea, of course, is that if A has an identity element 1, then 1 − a • b = (1 − a)(1 − b). It is more intuitive to reason with left, right and two-sided invertible elements than with their "quasi-" counterparts; but since we will be working with rings that may or may not have identity elements, the language of quasi-inverses streamlines some of the statements. This is particularly true when we start to move between various ideals in non-unital rings, where adjoining an identity would destroy the ideal property and make the phrasing of various conditions slightly cumbersome. Basic properties These are surely well-known, but we collect them here for ease of reference. Note first of all that • is associative: although one could check this by a direct comparison of (a • b) • c with a • (b • c), it is more instructive to adjoin a formal identity 1 and observe that 1 − (a • b) • c = (1 − a • b)(1 − c) = (1 − a)(1 − b)(1 − c) = (1 − a)(1 − b • c) = 1 − a • (b • c) An easy yet fundamental fact about unital Banach algebras is that the group of invertible elements is open in the norm topology. This has an obvious analogue for quasi-inverses; we state a slightly more precise version below, for later reference. Lemma 2.2. Let A be a Banach algebra and let a ∈ A. Suppose a has a right quasi-inverse in A; then so does every a′ that is sufficiently close to a. Proof. Suppose there exists b ∈ A such that a • b = 0. Adjoining an identity element 1 to A, chosen to have norm 1, we thus have (1 − a)(1 − b) = 1. Put δ = (1 + kbk)−1 > 0. Then, given any a′ ∈ A such that a′ − a < δ, put u := (1 − a′)(1 − b) ∈ A♯ We have 1 − u = (a′ − a)(1 − b) which has norm < 1; thus u is invertible in the Banach algebra A♯. Moreover, the usual formula for the inverse shows that c := 1 − u−1 = −Xn≥1 (1 − u)n lies in A. Since 1 − u = a′ • b by construction, we see that (a′ • b) • c = 0. By associativity of •, it follows that b • c is a right quasi-inverse to a′. We will also use the following simple observation, whose proof is straightforward from the definitions. Lemma 2.3. Let A be a k-algebra and J a right ideal in A. If a ∈ J has a right quasi-inverse b ∈ A, then b ∈ J. 3 Quasi-directly finite algebras A ring R with identity 1 is said to be directly finite if any x, y ∈ R which satisfy xy = 1 necessarily satisfy yx = 1. Many of the examples considered in the present article are algebras without an identity element, and while we can always pass to the unitization, it is more convenient to be able to work within the original algebra: see Proposition 2.8 below for an example of this. Thus, following Munn [8], we make the following definition. Definition 2.4. Let R be a ring, not necessarily having an identity element. We say R is quasi- directly finite if every element which is left quasi-invertible is also right quasi-invertible. (By our earlier remarks, if an element has both a left quasi-inverse bL and a right quasi-inverse bR, then bL = bR.) An algebra is said to be quasi-directly finite if its underlying ring is. For sake of brevity, we shall henceforth abbreviate "quasi-directly finite" to "q.d.f". Remark 2.5. We gave our definition in terms of quasi-inverses, since this is the formalism we will use in following sections. One can rephrase the definition as follows: R is q.d.f. if and only if, for every a, b ∈ R satisfying a + b = ab, we have a + b = ba. Lemma 2.6. Let k be a field. (i) If A is a q.d.f. k-algebra, then its forced unitization A♯ is directly finite. (ii) If B is a k-algebra with an identity element, then B is q.d.f. if and only if it is directly finite. Proof sketch. If A is q.d.f. and a♯, b♯ ∈ A♯ satisfy a♯b♯ = 1, then we must have a♯ = λ(1 − a) and b♯ = λ−1(1 − b) for some a, b ∈ A and λ ∈ C; by rescaling if necessary, we may assume without loss of generality that λ = 1. But then we have a • b = 1 − a♯b♯ = 0; since A is q.d.f. this implies that 0 = b • a = 1 − b♯a♯, so that b♯a♯ = 1. This completes the proof of part (i). The proof of (ii) is similar, and we omit the details. Remark 2.7. Since a directly finite algebra is q.d.f, and since subalgebras of q.d.f. algebras are q.d.f, (ii) implies that we can reverse the implication in (i). Proposition 2.8. Let A be a Banach algebra and let J be a right ideal in A which is dense for the norm topology. Then J is q.d.f. if and only if A is. Proof. Sufficiency is obvious, so we need only prove necessity. Suppose J is q.d.f. Let a, b ∈ A satisfy a • b = 0. By Lemma 2.2 and density of J in A, we can find a′ ∈ J which is close to a and which has a right quasi-inverse in A, say b′; by Lemma 2.3, b′ ∈ J. Therefore, since J is assumed to be q.d.f, b′ • a′ = 0. Moreover, the proof of Lemma 2.3 shows that we can take b′ to be of the form b • c for some c ∈ A. Thus b • (c • a′) = 0, i.e. b has a right quasi-inverse in A. Since we initially assumed that a • b = 0, it follows that b • a = 0 as required. A sufficient criterion for a C∗-algebra to be q.d.f. Let H be an infinite-dimensional Hilbert space. While the algebra B(H) is evidently not directly finite, the closed subalgebra K(H) is q.d.f. We shall see in this section that this is a special case of a more general result for semifinite von Neumann algebras. Proposition 2.9. Let A be a C∗-algebra and J a dense right ideal in A. Suppose that there exists a linear functional τ : J → C with the following properties: (i) τ (ab) = τ (ba) for all a, b ∈ J; and (ii) if c ∈ J and τ (c∗c) = 0 then c = 0. Then A is q.d.f. 4 In the case where A has an identity element and J = A, this result is well-known (and is indeed the basis of Montgomery's proof in [7] that the complex group algebra of a group is always directly finite). However, since I am unaware of a convenient reference where this case is stated explicitly, and since we are aiming for something slightly more general, we shall give a complete proof. The key observation is the following standard result about C∗-algebras: If p is an idempotent in a unital C∗-algebra, there exists a hermitian idempotent e in that Fact. algebra which satisfies ep = p and pe = e. This result seems to be part of the C∗-algebraic folklore. In particular, in several sources it is merely observed, without further proof, that we can take e to be e = pp∗(1 + (p − p∗)(p∗ − p))−1 . (†) This formula makes it clear that e = f e, but it is not so transparent that ep = p, nor that e is an idempotent. Probably the easiest, if not the quickest, way to verify these properties is to regard p as a projection inside B(H), and hence as a 2 × 2 operator matrix (cid:18)I R 0(cid:19) with regard to the 0 decomposition of H as ran(p) ⊕ ran(p)⊥. One now checks that the formula on the right hand side of (†) comes out to equal (cid:18)I 0 0 0(cid:19), i.e. the orthogonal projection of H onto ran(p); it is then clear that ep = p and pe = e = e2 as claimed. Proof of Proposition 2.9. By Proposition 2.8, it suffices to show that J is q.d.f. Let a, b ∈ J be such that a • b = 0. Put p = b • a = ab − ba ∈ J; clearly τ (p) = 0, by the 'tracial' property of τ . Moreover, since 2p − p2 = p • p = b • a • b • a = b • 0 • a = p , we have p = p2; thus p is an idempotent element of J. By the remarks above, there is a hermitian idempotent e ∈ A♯ such that pe = e and ep = p; in particular, e ∈ J, since J is a right ideal in A and hence in A♯. Then, using the tracial property of τ , we have τ (e) = τ (pe) = τ (ep) = τ (p) = 0 . But since e ∈ J and e = e∗e has trace zero, the 'faithfulness' of τ forces e = 0, hence forces p = 0. Thus b • a = 0 = a • b and the proof is complete. Let A be a C∗-algebra and A+ its cone of positive elements. By a trace on A+, we mean a function τ : A+ → [0, ∞] which is R+-linear and satisfies τ (a∗a) = τ (aa∗) for all a ∈ A. We say that τ is faithful if τ (x) > 0 for all x ∈ A+ \ {0}. Given a trace on A+, standard procedures from C∗-algebra theory ensure that there is a 2- sided ∗-ideal mτ ⊂ A, whose positive cone coincides with mτ ∩ A+, and a linear tracial functional mτ → C which coincides with τ on mτ ∩ A+; by abuse of notation, we will denote this functional also by τ . Moreover, the set nτ := {x ∈ A : τ (xx∗) < ∞} has the same norm-closure in A as mτ does. See [5, Proposition 6.1.2] for the details. It is then clear that Proposition 2.9 has the following consequence. Corollary 2.10. Let A be a C∗-algebra and τ : A+ → [0, ∞] a faithful trace. Then the norm closure of nτ in A is q.d.f. 5 Remark 2.11. A corollary of Proposition 2.9 is that continuous-trace C∗-algebras ([5, Ch. 4, §5]) are q.d.f. In this context it is natural to wonder which Type I C∗-algebras are q.d.f. On the one hand, all CCR algebras are q.d.f (as remarked in [2]), but that there are natural examples of GCR algebras which are not q.d.f, such as the Toeplitz algebra T . 3 Applications to group C∗-algebras The three most obvious C∗-algebras associated to a locally compact group G are the reduced group C∗-algebra, the full group C∗-algebra, and the group von Neumann algebra VN(G). It is in fact well known when VN(G) is directly finite: we give the characterization for sake of completeness and for background context. We shall then turn to examples where C∗ r(G) is q.d.f: here our main result is Theorem 3.2, which can be thought of as an extension of the main result from [7]. Lastly we shall make some comments on C∗(G). The case of VN(G) Being unital, a von Neumann algebra is q.d.f. if and only if it is directly finite; and this in turn happens if and only if it is a finite von Neumann algebra, that is, one in which the identity is a finite projection. (Briefly: if M is directly finite, then every isometry in M is unitary; this forces 1M to be a finite projection. Conversely, if M is finite, then standard von Neumann algebra theory tells us that M supports a separating family of faithful finite tracial states; applying Proposition 2.9 we deduce that M is directly finite.) It is known that we can get summands of any type in the type decomposition of VN(G) by choosing appropriate locally compact groups G: see [11] for some of the possibilities. On the other hand, it has long been known exactly when VN(G) is a finite von Neumann algebra. To state this characterization we need the following terminology: a topological group G is said to have small invariant neighbourhoods, or to be a SIN group, if it has a neighbourhood base at the identity consisting of conjugation-invariant neighbourhoods. Any compact, abelian or discrete group has this property, for example. Theorem 3.1. Let G be a locally compact group. Then VN(G) is a finite von Neumann algebra if and only if G is a SIN group. This result is essentially [5, Proposition 13.10.5]. (To be precise: the result there is stated under the additional assumption that G is unimodular, but inspection of the argument shows this extra condition to be unnecessary. For a less compressed account, see e.g. Propositions 3.2 and 4.1 of [12] and the surrounding remarks.) In particular, since SL(2, R) is known to be non-SIN, we deduce that VN(SL(2, R)) is not q.d.f. (Alternatively, we could have appealed to known structure theory for VN(SL(2, R)), which tells us that it has a unital subalgebra isomorphic to B(ℓ2).) Results for C∗ r(G) The simplest case to consider is when G is unimodular (thus including all SIN groups, but also those connected Lie groups which are semisimple or nilpotent). We recall the definition of the Plancherel weight on VN(G). Let F be the family of all compact, 6 finite-measure neighbourhoods of the origin in G, and define τ : VN(G)+ → [0, ∞] by τ (a) := sup K∈F K−2ha(χK), χKi (1) Clearly τ is faithful, and it is finite-valued on Cc(G)+. Moreover, an easy calculation shows that if G is unimodular then τ is a trace on VN(G)+. Since Cc(G) ⊆ nτ and mτ shares the same norm closure as nτ , applying Corollary 2.10 yields the following result. Theorem 3.2. Let G be a unimodular, locally compact group. Then C∗ r(G) is q.d.f. By Theorem 1.1 and the remarks which follow it, this gives us several examples of natural group representations θ : G → B(X) where the algebra θ(L1(G)) is surjunctive. Remark 3.3. Since semisimple Lie groups are unimodular, Theorem 3.2 provides another proof of [2, Theorem 3.4]. Remark 3.4. It is unclear whether having a dense q.d.f. *-subalgebra is always sufficient for a C∗-algebra to be q.d.f. Although the Plancherel weight on VN(G)+ is tracial if and only if G is unimodular, there are non-unimodular groups G for which VN(G) is a semifinite von Neumann algebra (hence supports some other faithful normal semifinite trace). This happens, for instance, if G is connected and separable (see [4, 9] for the details, and also [11, pp. 248 -- 249] for a more general perspective). In such situations, it is tempting to hope that this trace τ on VN(G)+ can be chosen such that mτ ∩ C∗ r(G) is dense in C∗ r(G). However, this does not work: we can find rather basic examples of connected, solvable Lie groups whose (reduced) group C∗-algebras are not q.d.f. This seems to be well-known to specialists: a particularly simple and explicit example is discussed by Rosenberg in [10]. Since the emphasis in that paper is rather different from the subject of this article, we shall for sake of convenience outline the salient facts. Let G be the "complex ax + b group" G = (cid:26)(cid:18)a b 1(cid:19) : a ∈ C∗, b ∈ C(cid:27) . 0 Up to unitary equivalence, G has precisely one infinite-dimensional unitary representation, which we denote by σ; this representation is faithful in the sense that σ : C∗(G) → B(Hσ) is injective. (See the remarks on [10, p. 177].) We then have the following result, whose proof can be found in that of [10, Proposition 1]. Proposition 3.5. There exists ϕ ∈ L1(G) such that I + σ(ϕ) is injective with closed, proper range. (In fact it has a one-dimensional cokernel.) Corollary 3.6. C∗ r(G) is not q.d.f. Proof of the corollary. We start by noting that since G is amenable, we can work with the full C∗ -- algebra instead of the reduced one. By Proposition 3.5, −1 lies in the residual spectrum of σ(ϕ). Since σ is faithful, if C∗(G) were q.d.f. then σ(C∗(G)) would be surjunctive, by Theorem 1.1. But this contradicts our initial observation that σ(ϕ) has non-empty residual spectrum. 7 Rosenberg's calculations actually went much further, and determined (up to isomorphism) the reduced group C∗-algebras of certain relatives of the ax + b group. The methods, which use a mix of commutative Fourier analysis with BDF techniques, were extended by Wang in [13] to the class of groups G(p, q, α): here p, q are positive integers, and G(p, q, α) is the semidirect product arising from an action of R on Rp+q via t 7→ diag(eα1t, . . . , eαpt, e−αp+1t, . . . , e−αp+qt) where α1, . . . , αp+q > 0. En route to determining C∗(G(p, q, α)), the following result is proved. Theorem 3.7 ([13, Corollary 3.2]). There is an embedding of C∗(G(p, q, α)) as a subalgebra of C(Sp−1 × Sq−1, K(L2(R))). Corollary 3.8. C∗(G(p, q, α)) is q.d.f. Remark 3.9. An alternative proof that C∗(G(p.q, α)) is q.d.f. goes as follows. First, note that the isomorphism class of C∗(G(p, q, α)) depends only on p and q and not on the lengths of the 'roots' α1, . . . , αp+q. (See [13, pp. 12 -- 13] and [10, pp. 190 -- 191].) Secondly, note that if we choose the αi so that α1 + · · · + αp = αp+1 + · · · + αp+q, then for each x in the Lie algebra g of G(p, q, α), the operator adx : g → g has trace zero; and this is known (see the remarks on [10, p. 190]) to imply that the group is unimodular. Applying Theorem 3.2, we deduce that C∗(G(p, q, α)) is q.d.f for this -- and hence any -- choice of the αi. The full group C∗-algebra What can be said for the full group C∗-algebra of G? If G is amenable then C∗(G) = C∗ r(G) and we can apply the results of the previous section. If G is non-amenable, then the situation is unclear even for discrete groups. There exist nonamenable discrete groups Γ for which C∗(Γ) has a faithful tracial state, and hence is q.d.f. by Proposition 2.9 -- for instance, this happens if G is residually finite. On the other hand, in [1] Bekka and Louvet give examples of discrete groups Γ for which C∗(Γ) has no faithful trace; but I do not know if such examples fail to be q.d.f. For certain classes of connected groups, we can do better. It was observed in [2] that if G is a CCR group (for instance, a connected Lie group which is either semisimple, nilpotent, or real algebraic) then C∗(G) has directly finite unitization, and hence in the language of this article is q.d.f. 4 Final remarks Many questions remain. We shall close with three in particular. 1. Does there exist a discrete group Γ (necessarily non-amenable) such that C∗(Γ) is not q.d.f? 2. Let G be a connected Lie group with Lie algebra g. Are there "reasonable' characterizations r(G) is q.d.f? What if we restrict attention to solvable Lie groups? in terms of g for when C∗ 3. If G is not SIN, then as remarked above its group von Neumann algebra cannot be q.d.f. What about its measure algebra M (G)? 8 Acknowledgements My thanks go to C. Zwarich for pointing out to me the characterization of when the group von Neumann algebra is finite, R, Archbold for several comments and corrections, and M. Daws for useful conversations leading to a more streamlined presentation of Theorem 3.2. References [1] M. B. Bekka and N. Louvet, Some properties of C ∗-algebras associated to discrete linear groups, in C ∗-algebras (Munster, 1999), Springer, Berlin, 2000, pp. 1 -- 22. [2] Y. Choi, Group representations with empty residual spectrum, Int. Eq. Op. Th., 67 (2010), pp. 95 -- 107. [3] J. Dixmier, Formes lin´eaires sur un anneau d'op´erateurs, Bull. Soc. Math. France, 81 (1953), pp. 9 -- 39. [4] J. Dixmier, Sur la repr´esentation r´eguli`ere d'un groupe localement compact connexe, Ann. Sci. ´Ecole Norm. Sup. (4), 2 (1969), pp. 423 -- 436. [5] , C ∗-algebras, North-Holland Publishing Co., Amsterdam, 1977. Translated from the French by Francis Jellett, North-Holland Mathematical Library, Vol. 15. [6] P. Eymard, L'alg`ebre de Fourier d'un groupe localement compact, Bull. Soc. Math. France, 92 (1964), pp. 181 -- 236. [7] M. S. Montgomery, Left and right inverses in group algebras, Bull. Amer. Math. Soc., 75 (1969), pp. 539 -- 540. [8] W. D. Munn, Direct finiteness of certain monoid algebras. I, Proc. Edinburgh Math. Soc. (2), 39 (1996), pp. 365 -- 369. [9] L. Pukanszky, Action of algebraic groups of automorphisms on the dual of a class of type I groups, Ann. Sci. ´Ecole Norm. Sup. (4), 5 (1972), pp. 379 -- 395. [10] J. Rosenberg, The C ∗-algebras of some real and p-adic solvable groups, Pacific J. Math., 65 (1976), pp. 175 -- 192. [11] C. E. Sutherland, Type analysis of the regular representation of a nonunimodular group, Pacific J. Math., 79 (1978), pp. 225 -- 250. [12] K. F. Taylor, The type structure of the regular representation of a locally compact group, Math. Ann., 222 (1976), pp. 211 -- 224. [13] X. Wang, The C ∗-algebras of a class of solvable Lie groups, vol. 199 of Pitman Research Notes in Mathematics Series, Longman Scientific & Technical, Harlow, 1989. Y. Choi D´epartement de math´ematiques et de statistique, Pavillon Alexandre-Vachon Universit´e Laval Qu´ebec, QC Canada, G1V 0A6 Email: [email protected] 9
1306.6669	2	1306	2013-08-22T23:00:05	A class of C*-algebras that are prime but not primitive	[ "math.OA", "math.RA" ]	We establish necessary and sufficient conditions on a (not necessarily countable) graph E for the graph C-algebra C(E) to be primitive. Along with a known characterization of the graphs E for which C(E) is prime, our main result provides us with a systematic method for easily producing large classes of (necessarily nonseparable) C-algebras that are prime but not primitive. We also compare and contrast our results with similar results for Leavitt path algebras.	math.OA	math	A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE GENE ABRAMS AND MARK TOMFORDE Abstract. We establish necessary and sufficient conditions on a (not necessarily countable) graph E for the graph C ∗-algebra C ∗(E) to be primitive. Along with a known characterization of the graphs E for which C ∗(E) is prime, our main result provides us with a systematic method for easily producing large classes of (necessarily nonseparable) C ∗-algebras that are prime but not primitive. We also compare and contrast our results with similar results for Leavitt path algebras. . A O h t a m [ 2 v 9 6 6 6 . 6 0 3 1 : v i X r a 1. Introduction It is well known that any primitive C ∗-algebra must be a prime C ∗- algebra, and a partial converse was established by Dixmier in the late 1950's when he showed that every separable prime C ∗-algebra is primitive (see [11, Corollaire 1] or [24, Theorem A.49] for a proof). For over 40 years after Dixmier's result, it was an open question as to whether every prime C ∗- algebra is primitive. This was answered negatively in 2001 by Weaver, who produced the first example of a (necessarily nonseparable) C ∗-algebra that is prime but not primitive [28]. Additional ad hoc examples of C ∗-algebras that are prime but not primitive have been given in [9], [19, Proposition 31], and [21, Proposition 13.4], with this last example being constructed as a graph C ∗-algebra. In this paper we identify necessary and sufficient con- ditions on the graph E for the C ∗-algebra C ∗(E) to be primitive. Conse- quently, this will provide a systematic way for easily describing large classes of (necessarily nonseparable) C ∗-algebras that are prime but not primitive. In particular, we obtain infinite classes of (nonseparable) AF-algebras, as well as infinite classes of non-AF, real rank zero C ∗-algebras, that are prime but not primitive. Compellingly, but perhaps not surprisingly, the conditions on E for which C ∗(E) is primitive are identical to the conditions on E for which the Leavitt path algebra LK (E) is primitive for any field K [2, Theorem 5.7]. However, as is typical in this context, despite the similarity of the statements of the Date: July 26, 2021. 2010 Mathematics Subject Classification. 46L55. Key words and phrases. C ∗-algebras, graph C ∗-algebras, prime, primitive. This work was supported by Collaboration Grants from the Simons Foundation to each author (Simons Foundation Grant #20894 to Gene Abrams and Simons Foundation Grant #210035 to Mark Tomforde). 1 2 GENE ABRAMS AND MARK TOMFORDE results, the proofs for graph C ∗-algebras are dramatically different from the proofs for Leavitt path algebras, and neither result directly implies the other. Acknowledgement: We thank Takeshi Katsura for providing us with use- ful comments and some helpful observations. 2. Preliminaries on graph C ∗-algebras In this section we establish notation and recall some standard definitions. Definition 2.1. A graph (E0, E1, r, s) consists of a set E0 of vertices, a set E1 of edges, and maps r : E1 → E0 and s : E1 → E0 identifying the range and source of each edge. Definition 2.2. Let E := (E0, E1, r, s) be a graph. We say that a vertex v ∈ E0 is a sink if s−1(v) = ∅, and we say that a vertex v ∈ E0 is an infinite emitter if s−1(v) = ∞. A singular vertex is a vertex that is either a sink or an infinite emitter, and we denote the set of singular vertices by E0 sing, and refer to the elements of E0 reg as regular vertices; i.e., a vertex v ∈ E0 is a regular vertex if and only if 0 < s−1(v) < ∞. A graph is row-finite if it has no infinite emitters. A graph is finite if both sets E0 and E1 are finite. A graph is countable if both sets E0 and E1 are (at most) countable. sing. We also let E0 reg := E0 \ E0 Definition 2.3. If E is a graph, a path is a finite sequence α := e1e2 . . . en of edges with r(ei) = s(ei+1) for 1 ≤ i ≤ n − 1. We say the path α has length α := n, and we let En denote the set of paths of length n. We consider the vertices of E (i.e., the elements of E0) to be paths of length zero. We also let Path(E) :=Sn∈N∪{0} En denote the set of paths in E, and we extend the maps r and s to Path(E) as follows: for α = e1e2 . . . en ∈ En with n ≥ 1, we set r(α) = r(en) and s(α) = s(e1); for α = v ∈ E0, we set r(v) = v = s(v). Also, for α = e1e2 · · · en ∈ Path(E), we let α0 denote the set of vertices that appear in α; that is, α0 = {s(e1), r(e1), . . . , r(en)}. Definition 2.4. If E is a graph, the graph C ∗-algebra C ∗(E) is the universal C ∗-algebra generated by mutually orthogonal projections {pv : v ∈ E0} and partial isometries with mutually orthogonal ranges {se : e ∈ E1} satisfying (1) s∗ (2) ses∗ ese = pr(e) e ≤ ps(e) for all e ∈ E1 for all e ∈ E1 (3) pv =P{e∈E 1:s(e)=v} ses∗ e for all v ∈ E0 reg. Definition 2.5. We call Conditions (1) -- (3) in Definition 2.4 the Cuntz- Krieger relations. Any collection {Se, Pv : e ∈ E1, v ∈ E0} of elements of a C ∗-algebra A, where the Pv are mutually orthogonal projections, the Se are partial isometries with mutually orthogonal ranges, and the Cuntz-Krieger relations are satisfied is called a Cuntz-Krieger E-family in A. The universal A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 3 property of C ∗(E) says precisely that if {Se, Pv : e ∈ E1, v ∈ E0} is a Cuntz- Krieger E-family in a C ∗-algebra A, the there exists a ∗-homomorphism φ : C ∗(E) → A with φ(pv) = Pv for all v ∈ E0 and φ(se) = Se for all e ∈ E1. For a path α := e1 . . . en, we define Sα := Se1 · · · Sen; and when α = 0, we have α = v is a vertex and define Sα := Pv. Remark 2.6. Using the orthogonality of the projections {ses∗ see that if α, β ∈ Path(E), then sαs∗ β = αδ for some γ, δ ∈ Path(E); in the former case we get sαs∗ while in the latter we get sαs∗ sαs∗ αsβs∗ sαs∗ α. e : e ∈ E1}, we β is nonzero if and only if α = βγ or β = sαs∗ α, β. Specifically, if α = β , then β is nonzero precisely when α = β, in which case the product yields αsβs∗ αsβs∗ β = sβs∗ αsβs∗ Definition 2.7. A cycle is a path α = e1e2 . . . en with length α ≥ 1 and r(α) = s(α). If α = e1e2 . . . en is a cycle, an exit for α is an edge f ∈ E1 such that s(f ) = s(ei) and f 6= ei for some i. We say that a graph satisfies Condition (L) if every cycle in the graph has an exit. Definition 2.8. A simple cycle is a cycle α = e1e2 . . . en with r(ei) 6= s(e1) for all 1 ≤ i ≤ n−1. We say that a graph satisfies Condition (K) if no vertex in the graph is the source of exactly one simple cycle. (In other words, a graph satisfies Condition (K) if and only if every vertex in the graph is the source of no simple cycles or the source of at least two simple cycles.) Our main use of Condition (L) will be in applying the Cuntz-Krieger Uniqueness Theorem. The Cuntz-Krieger Uniqueness Theorem was proven for row-finite graphs in [6, Theorem 1], and for countably infinite graphs in [13, Corollary 2.12] and [23, Theorem 1.5]. The result for possibly un- countable graphs is a special case of the Cuntz-Krieger Uniqueness Theorem [20, Theorem 5.1] for topological graphs. Alternatively, one can obtain the result in the uncountable case by using the version for countable graphs and applying the direct limit techniques described in [23] and [14]. Theorem 2.9 (Cuntz-Krieger Uniqueness Theorem). If E is a graph that satisfies Condition (L) and φ : C ∗(E) → A is a ∗-homomorphism from C ∗(E) into a C ∗-algebra A with the property that φ(pv) 6= 0 for all v ∈ E0, then φ is injective. It is a consequence of the Cuntz-Krieger Uniqueness Theorem that if E is a graph satisfying Condition (L) and I is a nonzero ideal of C ∗(E), then there exists v ∈ E0 such that pv ∈ I. (To see this, consider the quotient map q : C ∗(E) → C ∗(E)/I.) Definition 2.10. If v, w ∈ E0 we write v ≥ w to mean that there exists a path α ∈ Path(E) with s(α) = v and r(α) = w. (Note that the path α could be a single vertex, so that in particular we have v ≥ v for every v ∈ E0.) 4 GENE ABRAMS AND MARK TOMFORDE If E is a graph, a subset H ⊆ E0 is hereditary if whenever e ∈ E1 and s(e) ∈ H, then r(e) ∈ H. Note that a short induction argument shows that whenever H is hereditary, v ∈ H, and v ≥ w, then w ∈ H. For v ∈ E0 we define H(v) := {w ∈ E0 v ≥ w}. It is clear that H(v) is hereditary for each v ∈ E0. A hereditary subset H reg : r(s−1(v)) ⊆ H} ⊆ H. For any hereditary is called saturated if {v ∈ E0 subset H, we let H :=\{K : K ⊆ H and K is a saturated hereditary subset} denote the smallest saturated hereditary subset containing H, and we call H the saturation of H. Note that if H is hereditary and we define H0 := H reg : r(s−1(v)) ⊆ Hn−1} for n ∈ N, then H0 ⊆ and Hn := Hn−1 ∪ {v ∈ E0 n=0 Hn. It is clear that both of the sets ∅ and H1 ⊆ H2 ⊆ . . . and H = S∞ E0 are hereditary subsets of E0. Lemma 2.11. Let E = (E0, E1, r, s) be a graph. If H ⊆ E0 and K ⊆ E0 are hereditary subsets with H ∩ K = ∅, then H ∩ K = ∅. Proof. Since H ∩ K = ∅, we have H0 ∩ K0 = ∅. A straightforward inductive argument shows that Hn ∩ Kn = ∅ for all n ∈ N. Since H0 ⊆ H1 ⊆ H2 ⊆ . . . n=0 Kn = ∅. Hence (cid:3) and K0 ⊆ K1 ⊆ K2 ⊆ . . ., it follows that S∞ n=0 Hn ∩S∞ H ∩ K = ∅. If H is a hereditary subset, we let IH := span(cid:0){sαs∗ β : α, β ∈ Path(E), r(α) = r(β) ∈ H}). It is straightforward to verify that if H is hereditary, then IH is a (closed, two-sided) ideal of C ∗(E) and IH = IH. In addition, the map H 7→ IH is an injective lattice homomorphism from the lattice of saturated hereditary subsets of E into the gauge-invariant ideals of C ∗(E). (In particular, if H and K are saturated hereditary subsets of E, then IH ∩ IK = IH∩K .) When E is row-finite, the lattice homomorphism H 7→ IH is surjective onto the gauge-invariant ideals of C ∗(E). Definition 2.12. A graph E is downward directed if for all u, v ∈ E0, there exists w ∈ E0 such that u ≥ w and v ≥ w. Lemma 2.13. Let E = (E0, E1, r, s) be a graph, and let v, w ∈ E0. If I is an ideal in C ∗(E), pv ∈ I, and v ≥ w, then pw ∈ I. Proof. Let α be a path with s(α) = v and r(α) = w. Since pv ∈ I, we have (cid:3) pw = pr(α) = s∗ α ∈ I. αsα = sαps(α)s∗ α = sαpvs∗ The following graph-theoretic notion will play a central role in this paper. A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 5 Definition 2.14. Let E = (E0, E1, r, s) be a graph. For w ∈ E0, we define U (w) := {v ∈ E0 : v ≥ w}; that is, U (w) is the set of vertices v for which there exists a path from v to w. We say E satisfies the Countable Separation Property if there exists a Remark 2.15. It is useful to note that E does not satisfy the Countable Separation Property if and only if for every countable subset X ⊆ E0 we countable set X ⊆ E0 for which E0 =Sx∈X U (x). have E0 \Sx∈X U (x) 6= ∅. 2.1. The notions of "prime" and "primitive" for algebras, and for C ∗-algebras. When we are working with a ring R, an ideal in R shall always mean a two-sided ideal. When working with a C ∗-algebra A, an ideal in A shall always mean a closed two-sided ideal. If we have a two-sided ideal in a C ∗-algebra that is not closed, we shall refer to it as an algebraic ideal. Many properties for rings are stated in terms of two-sided ideals. However, when working in the category of C ∗-algebras it is natural to consider the corresponding C ∗-algebraic properties stated in terms of closed two-sided ideals. Thus for C ∗-algebras, one may ask whether a given ring-theoretic property coincides with the corresponding C ∗-algebraic property. In the next few definitions we will consider the notions of prime and primitive, and explain how the ring versions of these properties coincide with the C ∗- algebraic versions. In particular, a C ∗-algebra is prime as a C ∗-algebra if and only if it is prime as a ring, and a C ∗-algebra is primitive as a C ∗-algebra if and only if it is primitive as a ring. This will allow us to unambiguously refer to C ∗-algebras as "prime" or "primitive". If I, J are two-sided ideals of a ring R, then the product IJ is defined to be the two-sided ideal IJ :=( n Xℓ=1 iℓjℓ n ∈ N, iℓ ∈ I, jℓ ∈ J) . Definition 2.16. A ring R is prime if whenever I and J are two-sided ideals of R and IJ = {0}, then either I = {0} or J = {0}. If I, J are closed two-sided ideals of a C ∗-algebra A, then the product IJ is defined to be the closed two-sided ideal IJ := IJ =( n Xℓ=1 iℓjℓ n ∈ N, iℓ ∈ I, jℓ ∈ J). Definition 2.17. A C ∗-algebra A is prime if whenever I and J are closed two-sided ideals of A and IJ = {0}, then either I = {0} or J = {0}. Remark 2.18. We note that if a ring R admits a topology in which multi- plication is continuous (e.g., if R is a C ∗-algebra), then it is straightforward to show that R is prime if and only if R has the property that whenever I and J are closed two-sided ideals of R and IJ = {0}, then either I = {0} 6 GENE ABRAMS AND MARK TOMFORDE or J = {0}. Thus a C ∗-algebra is prime as a ring (in the sense of Defini- tion 2.16) if and only if it is prime as a C ∗-algebra (in the sense of Defi- nition 2.17). Moreover, since any C ∗-algebra has an approximate identity, and any closed two-sided ideal of a C ∗-algebra is again a C ∗-algebra, we get that whenever I and J are closed two-sided ideals in a C ∗-algebra, then IJ = I ∩J. Thus a C ∗-algebra A is prime if and only if whenever I and J are closed two-sided ideals in A and I ∩J = {0}, then either I = {0} or J = {0}. In addition, the existence of an approximate identity in a C ∗-algebra implies that ideals of ideals are ideals. In other words, if A is a C ∗-algebra, I is a closed two-sided ideal of A, and J is a closed two-sided ideal of I, then J is a closed two-sided ideal of A. Consequently, if I is a closed two-sided ideal in a prime C ∗-algebra, then I is a prime C ∗-algebra. Definition 2.19. Recall that for a ring R a left R-module RM consists of an abelian group M and a ring homomorphism π : R → End(M ) for which π(R)(M ) = M , giving the module action r · m := π(r)(m). We say that RM is faithful if the homomorphism π is injective, and we say RM is simple if M 6= {0} and M has no nonzero proper R-submodules; i.e., there are no nonzero proper subgroups N ⊆ M with π(r)(n) ⊆ N for all r ∈ R and n ∈ N . We make similar definitions for right R-modules MR, faithful right R-modules, and simple right R-modules. Definition 2.20. Let R be a ring. We say that R is left primitive if there exists a faithful simple left R-module. We say that R is right primitive if there exists a faithful simple right R-module. There are rings that are primitive on one side but not on the other. The first example was constructed by Bergman [7, 8]. In his ring theory textbook [26, p.159] Rowen also describes another example found by Jategaonkar that displays this distinction. Definition 2.21. If A is a C ∗-algebra, a ∗-representation is a ∗-homomorphism π : A → B(H) from the C ∗-algebra A into the C ∗-algebra B(H) of bounded linear operators on some Hilbert space H. We say that a ∗-representation π : A → B(H) is faithful if it is injective. For any subset S ⊆ H, we define π(A)S := span{π(a)h : a ∈ A and h ∈ S}. If S = {h} is a singleton set, we often write π(A)h in place of π(A){h}. A subspace K ⊆ H is called a π-invariant subspace (or just an invariant subspace) if π(a)k ∈ K for all a ∈ A and for all k ∈ K. Observe that a closed subspace K ⊆ H is invariant if and only if π(A)K ⊆ K. Definition 2.22. Let A be a C ∗-algebra, and let π : A → B(H) be a ∗- representation. (1) We say π is a countably generated ∗-representation if there exists a countable subset S ⊆ H such that π(A)S = H. (2) We say π is a cyclic ∗-representation if there exists h ∈ H such that π(A)h = H. (Note that a cyclic ∗-representation could also be called a singly generated ∗-representation.) A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 7 (3) We say π is an irreducible ∗-representation if there are no closed invariant subspaces of H other than {0} and H. We note that π is irreducible if and only if π(A)h = H for all h ∈ H \ {0}. Remark 2.23. One can easily see that for any ∗-representation of a C ∗- algebra, the following implications hold: irreducible =⇒ cyclic =⇒ countably generated. In addition, if H is a separable Hilbert space (and hence H has a countable basis), then π is automatically countably generated. Definition 2.24. A C ∗-algebra A is primitive if there exists a faithful irre- ducible ∗-representation π : A → B(H) for some Hilbert space H. The following result is well known among C ∗-algebraists. Proposition 2.25. If A is a C ∗-algebra, then the following are equivalent. (1) A is a left primitive ring (in the sense of Definition 2.20). (2) A is a right primitive ring (in the sense of Definition 2.20). (3) A is a primitive C ∗-algebra (in the sense of Definition 2.24). Proof. If A is a C ∗-algebra and Aop denotes the opposite ring of A, then we see that a 7→ a∗ is a ring isomorphism from A onto Aop. Hence A − Mod and Aop − Mod ∼= Mod −A are equivalent categories. (We mention that, in general, a C ∗-algebra is not necessarily isomorphic as a C ∗-algebra to its opposite C ∗-algebra; see [22].) The equivalence of (1) and (3) is proven in [12, Theorem 2.9.7, p.57] and uses the results of [12, Corollary 2.9.6(i), p.57] and [12, Corollary 2.8.4, p.53], which show two things: (i) that any algebraically irreducible representation of A on a complex vector space is algebraically equivalent to a topologi- cally irreducible ∗-representation of A on a Hilbert space, and (ii) that any two topologically irreducible ∗-representations of A on a Hilbert space are (cid:3) algebraically isomorphic if and only if they are unitarily equivalent. Remark 2.26. In light of Proposition 2.25, a C ∗-algebra is primitive in the sense of Definition 2.24 if and only if it satisfies any (and hence all) of the three equivalent conditions stated in Proposition 2.25. 3. Prime and primitive graph C ∗-algebras In this section we establish graph conditions that characterize when the associated graph C ∗-algebra is prime and when the associated graph C ∗- algebra is primitive. The following proposition was established in [19, Theorem 10.3] in the context of topological graphs, but for the convenience of the reader we pro- vide a streamlined proof here for the context of graph C ∗-algebras. We also mention that this same result was obtained for countable graphs in [5, Proposition 4.2].) 8 GENE ABRAMS AND MARK TOMFORDE Proposition 3.1. Let E be any graph. Then C ∗(E) is prime if and only if the following two properties hold: (i) E satisfies Condition (L), and (ii) E is downward directed. Proof. First, let us suppose that E satisfies properties (i) and (ii) from above. If I and J are nonzero ideals in C ∗(E), then it follows from property (i) and the Cuntz-Krieger Uniqueness Theorem for graph C ∗-algebras that there exists u, v ∈ E0 such that pu ∈ I and pv ∈ J. By property (ii) there is a vertex w ∈ E0 such that u ≥ w and v ≥ w. It follows from Lemma 2.13 that pw ∈ I and pw ∈ J. Hence 0 6= pw = pwpw ∈ IJ, and so C ∗(E) is a prime C ∗-algebra. For the converse, let us suppose that C ∗(E) is prime and establish prop- erties (i) and (ii). Suppose C ∗(E) is prime, and E does not satisfy Con- dition (L). Then there exists a cycle α = e1 . . . en in E that has no ex- its. Since α has no exits, the set H := α0 = {s(e1), r(e1), . . . , r(en−1)} is a hereditary subset of E, and it follows from [5, Proposition 3.4] that the ideal IH = IH is Morita equivalent to the C ∗-algebra of the graph EH := (H, s−1(H), rH , sH). Since EH is the graph consisting of a single cycle, C ∗(EH ) ∼= Mn(C(T)) for some n ∈ N (see [16, Lemma 2.4]). There- fore, the ideal IH is Morita equivalent to C(T). However, since C ∗(E) is a prime C ∗-algebra, it follows that the ideal IH is a prime C ∗-algebra. Be- cause Morita equvalence preserves primality, and IH is Morita equivalent to C(T), it follows that C(T) is a prime C ∗-algebra. However, it is well known that C(T) is not prime: If C and D are proper closed subsets of T for which C ∪ D = T, and we set IC := {f ∈ C(T) : f C ≡ 0} and ID := {f ∈ C(T) : f D ≡ 0}, then IC and ID are ideals with IC ∩ ID = {0} but IC 6= {0} and ID 6= {0}. This provides a contradiction, and we may conclude that E satisfies Condition (L) and that property (i) holds. Next, suppose C ∗(E) is prime. Let u, v ∈ E0, and consider H(u) := {x ∈ E0 : u ≥ x} and H(v) := {x ∈ E0 : v ≥ x}. Then H(u) and H(v) are nonempty hereditary subsets, and the ideals IH(u) = IH(u) and IH(v) = IH(v) are each nonzero. Since C ∗(E) is a prime C ∗-algebra, it follows that IH(u)∩H(v) = IH(u) ∩ IH(v) 6= {0}. Thus H(u) ∩ H(v) 6= ∅, and Lemma 2.11 implies that H(u) ∩ H(v) 6= ∅. If we choose w ∈ H(u) ∩ H(v), then u ≥ w and v ≥ w. Hence E is downward directed and property (ii) (cid:3) holds. The following is well known, but we provide a proof for completeness. Lemma 3.2. Any primitive C ∗-algebra is prime. Proof. Let A be a primitive C ∗-algebra. Then there exists a faithful irre- ducible ∗-representation π : A → B(H). Let I and J be ideals of A, and suppose that IJ = {0}. If J 6= {0}, then the faithfulness of π implies π(J)H 6= {0}, and the fact that J is an ideal shows that π(J)H is a closed A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 9 invariant subspace for π. Since π is irreducible, it follows that π(J)H = H. Using the fact that IJ = {0}, it follows that {0} = π(IJ)H = π(I)π(J)H = π(I)H. Since π is faithful, this implies that I = {0}. Thus A is a prime C ∗-algebra. (cid:3) The following two lemmas are elementary, but very useful. Lemma 3.3. If X ⊆ (0, ∞) is an uncountable subset of positive real num- bers, then the sum Xx∈X x diverges to infinity. Proof. For each n ∈ N, let Xn := {x ∈ X : x ≥ 1 n=1 Xn, and since X is uncountable, there exists n0 ∈ N such that Xn0 is infinite. (cid:3) n }. Then X = S∞ 1 n = ∞. Thus Px∈X x ≥Px∈Xn0 x ≥P∞ i=1 Lemma 3.4. Let A be a C ∗-algebra, and let {pi i ∈ I} be a set of nonzero mutually orthogonal projections in A. If there exists a ∗-representation π : A → B(H) and a vector ξ ∈ H such that π(pi)ξ 6= 0 for all i ∈ I, then I is at most countable. images of the π(pi). Since the π(pi) are mutually orthogonal projections, Proof. Let P := Li∈I π(pi) be the projection onto the direct sum of the the Pythagorean theorem shows Pi∈I kπ(pi)ξk2 = kP ξk2 ≤ kξk2 < ∞. Since each kπ(pi)ξk2 term is nonzero, and since any uncountable sum of positive real numbers diverges to infinity (see Lemma 3.3), it follows that (cid:3) the index set I is at most countable. The following proposition provides a necessary condition for a graph C ∗- algebra to be primitive. Proposition 3.5. If E is a graph and C ∗(E) has a faithful countably gen- erated ∗-representation, then E satisfies the Countable Separation Property. Proof. By hypothesis there is a faithful countably generated ∗-representation π : C ∗(E) → B(H). Thus there exists a countable set of vectors S := {ξi}∞ i=1 ⊆ H with π(C ∗(E))S = H. For every n ∈ N ∪ {0} and for every i ∈ N, define Γn,i := {α ∈ Path(E) : α = n and π(sαs∗ α)ξi 6= 0}. (Recall that we view vertices as paths of length zero, and in this case sv = pv.) By Remark 2.6, for any n ∈ N ∪ {0} the set {sαs∗ α : α ∈ Path(E) and α = n} consists of mutually orthogonal projections, and hence Lemma 3.4 implies that for any n ∈ N ∪ {0} and for any i ∈ N, the set Γn,i is countable. Define Γ := ∞ [n=0 ∞ [i=1 Γn,i, 10 GENE ABRAMS AND MARK TOMFORDE which is countable since it is the countable union of countable sets. Also define U (r(α)). Θ := [α∈Γ Then Θ ⊆ E0 is a set of vertices, and we shall show that Θ = E0. We note that Θ consists precisely of the vertices v in E for which there is a path from v to w, where w = r(α) for a path α having the property that π(SαS∗ α)ξi is nonzero for some i. Let v ∈ E0, and let I denote the closed two-sided ideal of C ∗(E) generated by pv. Since I is a nonzero ideal and π is faithful, it follows that π(I)H 6= {0}. Thus π(I)S = π(IC ∗(E))S = π(I)π(C ∗(E))S = π(I)H 6= {0} and hence there exists a ∈ I and ξi ∈ S such that π(a)ξi 6= 0. If H(v) := {w ∈ E0 : v ≥ w} is the hereditary subset of E0 generated by v (as given in Definition 2.10), then it follows from [5, §3] that I = IH(v) = span{sαs∗ β : α, β ∈ Path(E) and r(α) = r(β) ∈ H(v)}. Since a ∈ I and π(a)ξi 6= 0, it follows from the linearity and continuity of π that there exists α, β ∈ Path(E) with r(α) = r(β) ∈ H(v) and π(sαs∗ β)ξi 6= 0. Hence π(sαs∗ β)π(sβs∗ 6= 0. and thus π(sβs∗ β)ξi β)ξi = π(sαs∗ βsβs∗ β)ξi = π(sαs∗ β)ξi 6= 0, If we let n := β, then β ∈ Γn,i ⊆ Γ. Since r(β) ∈ H(v), it follows that v ≥ r(β) and v ∈ U (r(β)) ⊆Sα∈Γ U (r(α)) = Θ. Thus E0 = Θ := Sα∈Γ U (r(α)), and since Γ is countable, E satisfies the Countable Separation Property. (cid:3) Corollary 3.6. If E is a graph and C ∗(E) has a faithful cyclic ∗-representation, then E satisfies the Countable Separation Property. Corollary 3.7. If E is a graph and C ∗(E) is primitive, then E satisfies the Countable Separation Property. Our main objective in this article is the following result, in which we provide necessary and sufficient conditions on a graph E for the C ∗-algebra C ∗(E) to be primitive. In particular, we identify the precise additional condition on E that guarantees a prime graph C ∗-algebra C ∗(E) is primitive. With the previously mentioned result of Dixmier [11, Corollaire 1] as context (i.e., that any prime separable C ∗-algebra is primitive), we note that our additional condition is not a separability hypothesis on C ∗(E), but rather the Countable Separation Property of E. Theorem 3.8. Let E be any graph. Then C ∗(E) is primitive if and only if the following three properties hold: (i) E satisfies Condition (L), (ii) E is downward directed, and A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 11 (iii) E satisfies the Countable Separation Property. In other words, by Proposition 3.1, C ∗(E) is primitive if and only if C ∗(E) is prime and E satisfies the Countable Separation Property. Proof. To prove this result we establish both the sufficiency and the necessity of properties (i), (ii), and (iii) for C ∗(E) to be primitive. Proof of Sufficiency: Since E satisfies the Countable Separation Prop- erty by (iii), there exists a countable set X ⊆ E0 such that E0 =Sx∈X U (x). Since X is countable, we may list the elements of X as X := {v1, v2, . . .}, where this list is either finite or countably infinite. For convenience of nota- tion, let us write X = {vi}i∈I where the indexing set I either has the form I = {1, . . . , n} or I = N. We inductively define a sequence of paths {λi}i∈I ⊆ Path(E) satisfying the following two properties: (a) For each i ∈ I we have vi ≥ r(λi). (b) For each i ∈ I with i ≥ 2 there exists µi ∈ Path(E) such that λi = λi−1µi. To do so, define λ1 := v1 and note that for i = 1 Property (a) is satisfied trivially and Property (b) is satisfied vacuously. Next suppose that λ1, . . . λn have been defined so that Property (a) and Property (b) are satisfied for 1 ≤ i ≤ n. By hypothesis (ii) E is downward directed, and hence there exists a vertex un+1 in E such that r(λn) ≥ un+1 and vn+1 ≥ un+1. Let µn+1 be a path from r(λn) to un+1, and define λn+1 := λnµn+1. Then the paths λ1, . . . , λn+1 satisfy Property (a) and Property (b) for all 1 ≤ i ≤ n. Continuing in this manner, we either exhaust the elements of I or inductively create a sequence of paths {λi}i∈I ⊆ Path(E) satisfying Property (a) and Property (b) for all i ∈ I. Note that, in particular, for each i < n, we have λn = λiµi+1 . . . µn, and hence by Remark 2.6 for i ≤ n we have sλis∗ λisλns∗ λn = sλns∗ λn. We now establish for future use that every nonzero closed two-sided ideal J of C ∗(E) contains sλns∗ λn for some n ∈ I. Using hypothesis (i), our graph satisfies Condition (L), and the Cuntz-Krieger Uniqueness Theorem (Theorem 2.9) implies that there exists w ∈ E0 such that pw ∈ J. By the Countable Separation Property there exists vn ∈ X such that w ≥ vn. In addition, by Property (a) above vn ≥ r(λn), so that there is a path γ in E for which s(γ) = w and r(γ) = r(λn). Since pw ∈ J, we have by Lemma 2.13 that pr(λn) ∈ J, so that sλns∗ λn = sλnpr(λn)s∗ λn ∈ J as desired. Define L :=( n Xi=1 (xi − xisλis∗ λi) : n ∈ I and x1, . . . , xn ∈ C ∗(E)) . Recall that λ1 = v1, and by our convention (see Definition 2.5) sλ1 = pv1 and sλ1s∗ λ1 = pv1. Clearly L is an algebraic (i.e., not necessarily closed) left 12 GENE ABRAMS AND MARK TOMFORDE ideal of C ∗(E). We claim that pv1 /∈ L. For otherwise there would exist n ∈ I and x1, ..., xn ∈ C ∗(E) with n (xi − xisλis∗ λi) = pv1. Xi=1 But then multiplying both sides of this equation by sλns∗ λn on the right gives (xisλns∗ λn − xisλis∗ λisλns∗ λn) = sλns∗ λn. n Xi=1 n Xi=1 Using the previously displayed observation, the above equation becomes (xisλns∗ λn − xisλns∗ λn) = sλns∗ λn, which gives 0 = sλns∗ pv1 /∈ L, and L is a proper left ideal of C ∗(E). λn, a contradiction. Hence we may conclude that By the definition of L, we have a−apv1 ∈ L for each a ∈ C ∗(E) and hence L is a modular left ideal of C ∗(E), a property which necessarily passes to any left ideal of C ∗(E) containing L. (See [10] for definitions of appropriate terms. Also, cf. [25, Chapter 2, Theorem 2.1.1].) Let T denote the set of (necessarily modular) left ideals of C ∗(E) that contain L but do not contain pv1. Since L ∈ T , we have T 6= ∅. By a Zorn's Lemma argument there exists a maximal element in T , which we denote M . We claim that M is a maximal left ideal of C ∗(E). For suppose that M ′ is a left ideal of C ∗(E) having M $ M ′. Then by the maximality of M in T we have pv1 ∈ M ′, so that xpv1 ∈ M ′ for each x ∈ C ∗(E), which gives that x = (x − xpv1) + xpv1 ∈ L + M ′ = M ′ for each x ∈ C ∗(E), so that M ′ = C ∗(E). Thus M is a maximal left ideal, and since M is also modular, M is a maximal modular left ideal. Thus by [10, VII.2, Exercise 6] (cf. [25, Chapter 2, Corollary 2.1.4]) M is closed. Since M is closed we may form the regular ∗-representation of C ∗(E) into C ∗(E)/M ; i.e., the homomorphism π : C ∗(E) → End(C ∗(E)/M ) given by π(a)(b+M ) := ab+M . In this way C ∗(E)/M becomes a left C ∗(E)-module. The submodules of C ∗(E)/M correspond to left ideals of C ∗(E) containing M , and by the maximality of M the only submodules of C ∗(E)/M are {0} and M . Hence C ∗(E)/M is a simple module. We claim that ker π = {0}. If a ∈ ker π, then for all b ∈ C ∗(E) we have ab + M = π(a)(b + M ) = 0 + M , so that ab ∈ M for all b ∈ C ∗(E). If {eλ}λ∈Λ is an approximate unit for C ∗(E), then the previous sentences combined with the fact that M is closed implies that a = limλ aeλ ∈ M . Hence we have established that ker π ⊆ M . Furthermore, since ker π is a closed two-sided ideal of C ∗(E), it follows from above that if ker π is nonzero, then sλns∗ λn ∈ ker π for some n ∈ I. But then M contains asλns∗ In addition, λn ∈ L ⊆ M for all a ∈ C ∗(E), it follows that a ∈ M for since a − asλns∗ all a ∈ C ∗(E), and C ∗(E) ⊆ M , which is a contradiction. We conclude that ker π = {0}. Hence C ∗(E)/M is a faithful left C ∗(E)-module, which λn for every a ∈ C ∗(E). A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 13 with the simplicity of C ∗(E)/M yields that C ∗(E) is primitive as a ring. It follows (see Remark 2.25) that C ∗(E) is primitive as a C ∗-algebra. Proof of Necessity: If C ∗(E) is primitive, then C ∗(E) is necessarily prime by Lemma 3.2, so by Proposition 3.1 we get that E satisfies Condi- tion (L) and is downward directed. In addition, Corollary 3.7 shows that if C ∗(E) is primitive, then E satisfies the Countable Separation Property, (cid:3) thus completing the proof. Corollary 3.9. Let E be a graph. Then the following are equivalent: (i) C ∗(E) is primitive. (ii) C ∗(E) is prime and E satisfies the Countable Separation Property. (iii) C ∗(E) is prime and C ∗(E) has a faithful cyclic ∗-representation. (iv) C ∗(E) is prime and C ∗(E) has a faithful countably generated ∗- representation. Proof. The equivalence of (i) and (ii) follows from Theorem 3.8 and Proposi- tion 3.1. The implications (i) =⇒ (iii) =⇒ (iv) are trivial. The implication (iv) =⇒ (ii) follows from Corollary 3.7. (cid:3) Remark 3.10. In [21, Problem 13.6] Katsura asks whether a prime C ∗- algebra is primitive if it has a faithful cyclic ∗-representation. Corollary 3.9 shows that the answer to Katsura's question is affirmative in the class of graph C ∗-algebras. Moreover, Corollary 3.9 prompts us to ask the following more general question: Question: If a C ∗-algebra is prime and has a faithful countably generated representation, then is that C ∗-algebra primitive? Again, Corollary 3.9 provides us with an affirmative answer to this question for the class of graph C ∗-algebras. In addition, an affirmative answer to this question in general implies an affirmative answer to Katsura's question in [21, Problem 13.6]. Remark 3.11. In the introduction of [28] Weaver notes that his example of a prime and not primitive C ∗-algebra ". . . places competing demands on the set of partial isometries: it must be sufficiently abundant . . . and [simulta- neously] sufficiently sparse . . ." Effectively, Theorem 3.8 identifies precisely how these two competing demands play out in the context of a graph C ∗- algebra C ∗(E): primeness (abundance of partial isometries) corresponds to E satisfying Condition (L) and being downward directed, while nonprimi- tivity (sparseness of partial isometries) corresponds to E not satisfying the Countable Separation Property. Remark 3.12. It is shown in [5, Proposition 4.2] that for a graph E having both E0 and E1 countable, C ∗(E) is primitive if and only if E is downward directed and satisfies Condition (L). When E0 is countable, then E trivially satisfies the Countable Separation Property, with E0 itself providing the 14 GENE ABRAMS AND MARK TOMFORDE requisite countable set of vertices. Thus Theorem 3.8 provides both a gen- eralization of (since the countability of E1 is not required), and an alternate proof for, the result in [5, Proposition 4.2]. 4. Examples of prime but not primitive C ∗-algebras We now offer a number of examples of prime nonprimitive C ∗-algebras that arise from the characterizations presented in Proposition 3.1 and The- orem 3.8. These examples are similar in flavor to the classes of examples that played an important role in [2]. Definition 4.1. Let X be any nonempty set, and let F(X) denote the set of finite nonempty subsets of X. We define three graphs EA(X), EL(X), and EK(X) as follows. and EA(X)1 = {eA,A′ : A, A′ ∈ F(X) and A $ A′}, (1) The graph EA(X) is defined by EA(X)0 = F(X) with s(eA,A′) = A and r(eA,A′) = A′ for each eA,A′ ∈ EA(X)1. (2) The graph EL(X) is defined by EL(X)0 = F(X) with s(eA,A′) = A and r(eA,A′) = A′ for each eA,A′ ∈ EL(X)1. (3) The graph EK(X) is defined by and EL(X)1 = {eA,A′ : A, A′ ∈ F(X) and A j A′}, EK(X)0 = F(X) EK(X)1 = {eA,A′ : A, A′ ∈ F(X) and A j A′} ∪ {fA : A ∈ F(X)}, and with s(eA,A′) = A, and r(eA,A′) = A′ for each eA,A′ ∈ EK(X)1, and with s(fA) = r(fA) = A for all fA. Remark 4.2. Observe that for any nonempty set X, the graph EA(X) is acyclic, the graph EL(X) has as its only simple cycles the single loop at each vertex (so that EL(X) satisfies Condition (L), but not Condition (K)), and the graph EK (X) has as its only simple cycles the two loops at each vertex (so that EK (X) satisfies Condition (K)). In this section all direct limits that we discuss will be direct limits in the category of C ∗-algebras, so that the direct limit algebras discussed are C ∗-algebras. An AF-algebra is typically defined to be a C ∗-algebra that is the direct limit of a sequence of finite-dimensional algebras. Since it is a sequential direct limit, an AF-algebra is necessarily separable. Some authors, such as Katsura in [19], have considered arbitrary direct limits of finite-dimensional C ∗-algebras. Following Katsura in [19], we shall define an AF-algebra to be a C ∗-algebra that is the direct limit of finite-dimensional C ∗-algebras. (Equivalently, an AF-algebra is a C ∗-algebra A with a directed family of finite dimensional C ∗-subalgebras whose union is dense in A.) An A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 15 AF-algebra in our sense is a sequential direct limit (i.e., an AF-algebra in the traditional sense) if and only if it is separable. Lemma 4.3. Let EA(X), EL(X), and EK(X) be the graphs presented in Definition 4.1. (1) Each of the graphs EA(X), EL(X), and EK(X) is downward directed. (2) Each of the graphs EA(X), EL(X), and EK(X) satisfies the Count- able Separation Property if and only if X is countable. Proof. To establish (1), we observe that in each of the graphs EA(X), EL(X), and EK (X), each pair of vertices A, A′ corresponds to a pair of finite subsets of X, so that if B denotes the finite set A ∪ A′ then there is an edge eA,B from A to B and an edge eA′,B from A′ to B. Downward directedness of each of the three graphs follows. For (2), we note that if X is countable then F(X) is countable, so that in this case each of the three graphs has countably many vertices, and thus trivially satisfies the Countable Separation Property. On the other hand, if X is uncountable, then any countable union of elements of F(X) includes only countably many elements of X, so that there exists some vertex (in- deed, uncountably many vertices) which does not connect to the vertices (cid:3) represented by such a countable union. Proposition 4.4. Let X be an infinite set, and let EA(X), EL(X), and EK(X) be the graphs presented in Definition 4.1. Then (1) C ∗(EA(X)) is a prime AF-algebra for any set X. Furthermore, C ∗(EA(X)) is primitive if and only if X is countable. In addition, C ∗(EA(X)) is a separable AF-algebra if and only if X is countable. (2) C ∗(EL(X)) is a prime C ∗-algebra that is not AF for any set X. Fur- thermore, C ∗(EL(X)) is primitive if and only if X is countable. In addition, C ∗(EL(X)) contains an ideal that is not gauge invariant. (3) C ∗(EK (X)) is a prime C ∗-algebra of real rank zero that is not AF for any set X. Furthermore, C ∗(EK (X)) is primitive if and only if X is countable. In addition, every ideal of C ∗(EK(X)) is gauge invariant. Proof. The indicated primeness and primitivity properties of each of the three algebras C ∗(EA(X)), C ∗(EL(X)), and C ∗(EK (X)) follow directly from Proposition 3.1, Theorem 3.8, and Lemma 4.3. We now take up the discussion of the additional properties. (1) Since EA(X) is a countable graph if and only if X is countable, it follows that C ∗(EA(X)) is separable if and only if X is countable. (2) We see that EL(X) has exactly one loop at each vertex, and that these are the only simple cycles in EL(X). Thus every vertex of EL(X) is the base of exactly one simple cycle, so that EL(X) does not satisfy Condition (K). In addition, if α is the loop based at the vertex A, then since A is finite and X is infinite there exists an element x ∈ X \ A, and the edge from A to A ∪ {x} provides an exit for α. Hence every cycle in EL(X) has an exit and EL(X) 16 GENE ABRAMS AND MARK TOMFORDE satisfies Condition (L). Since EL(X) contains a cycle, C ∗(EL(X)) is not AF. Moreover, since EL(X) does not satisfy Condition (K), it follows that C ∗(EL(X)) contains an ideal that is not gauge invariant. (This is established for row-finite countable graphs in [27, Theorem 2.1.19], although the same argument works for non-row-finite or uncountable graphs. Alternatively, the result for uncountable graphs may also be obtained as a special case of [21, Theorem 7.6].) (3) As EK (X) has two loops at each vertex, every vertex in EK(X) is the base point of two distinct simple cycles, so that EK(X) satisfies Condi- tion (K). Since EK (X) contains a cycle, C ∗(EK (X)) is not AF. In addition, since EK (X) satisfies Condition (K), C ∗(EK (X)) has real rank zero. (This was established for C ∗-algebras of locally-finite countable graphs in [18, The- orem 4.1] and for C ∗-algebras of countable graphs in [15, Theorem 2.5], and can be extended to uncountable graphs using the approximation methods of [23, Lemma 1.2].) Moreover, since EK (X) satisfies Condition (K), all ideals of C ∗(EK (X)) are gauge invariant. (cid:3) We can now produce infinite classes of C ∗-algebras that are prime and not primitive. In fact, we are able to produce three such classes of C ∗-algebras: one in which all the C ∗-algebras are AF-algebras, one in which all the C ∗- algebras are non-AF and contain ideals that are not gauge invariant, and one in which all the C ∗-algebras are non-AF, have all of their ideals gauge invariant, and are real rank zero. Proposition 4.5. For a set X we let X denote the cardinality of X. (1) If C := {Xi : i ∈ I} is a collection of sets with Xi > ℵ0 for all i ∈ I, and with Xi 6= Xj for all i, j ∈ I with i 6= j, then {C ∗(EA(Xi)) : i ∈ I} is a collection of AF-algebras that are prime and not primitive. Moreover C ∗(EA(Xi)) is not Morita equivalent to C ∗(EA(Xj )) for all i, j ∈ I with i 6= j. (2) If C := {Xi : i ∈ I} is a collection of sets with Xi ≥ 2ℵ0 for all i ∈ I, and with Xi 6= Xj for all i, j ∈ I with i 6= j, then {C ∗(EL(Xi)) : i ∈ I} is a collection of non-AF C ∗-algebras each of which is prime and not primitive, and each of which has the property that it contains ideals that are not gauge invariant. Moreover, C ∗(EL(Xi)) is not Morita equivalent to C ∗(EL(Xj)) for all i, j ∈ I with i 6= j. (3) If C := {Xi : i ∈ I} is a collection of sets with Xi > ℵ0 for all i ∈ I, and with Xi 6= Xj for all i, j ∈ I with i 6= j, then {C ∗(EK(Xi)) : i ∈ I} is a collection of non-AF C ∗-algebras of real rank zero each of which is prime and not primitive, and each of which has the property that A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 17 all of its ideals are gauge invariant. Moreover, C ∗(EK (Xi)) is not Morita equivalent to C ∗(EK(Xj )) for all i, j ∈ I with i 6= j. Proof. For (1), the fact that the C ∗(EA(Xi)) are AF-algebras that are prime and not primitive follows from Proposition 4.4(1). It remains to show the C ∗(EA(Xi)) are mutually non-Morita equivalent. For each i ∈ I, the graph EA(Xi) satisfies Condition (K), and hence the ideals in C ∗(EA(Xi)) are in one-to-one correspondence with admissible pairs (H, S), where H is a saturated hereditary subset of EA(Xi)0 and S is a subset of breaking vertices for H. For each x ∈ Xi define Hx := EA(Xi)0 \ {x}. Because {x} is a source and an infinite emitter in EA(Xi) that only emits edges into Hx, the set Hx is saturated and hereditary, and Hx has no breaking vertices. In addition, any saturated hereditary subset of EA(Xi)0 that contains Hx is either equal to Hx or equal to EA(Xi)0. Thus I(Hx,∅) is a maximal ideal in C ∗(EA(Xi)). Conversely, any maximal ideal must have the form I(H,S) for an admissible pair (H, S), and in order to be a proper ideal there exists x ∈ EA(Xi)0 \ H. Because H ⊆ Hx, the maximality of I(H,S) implies H = Hx and S = ∅. Thus we may conclude that the map x 7→ I(Hx,∅) is a bijection from X onto the set of maximal ideals of C ∗(EA(Xi)). Hence the cardinality of the set of maximal ideals of C ∗(EA(Xi)) is equal to Xi. Since any two Morita equivalent C ∗-algebras have isomorphic lattices of ideals, any two Morita equivalent C ∗-algebras have sets of maximal ideals with the same cardinality. Thus, when i, j ∈ I with i 6= j, the fact that Xi 6= Xj implies that C ∗(EA(Xi)) is not Morita equivalent to C ∗(EA(Xj)). For (2), the fact that the C ∗-algebras C ∗(EL(Xi)) are non-AF C ∗-algebras that are prime and not primitive and that each contains ideals that are not gauge invariant follows from Proposition 4.4(2). It remains to show the C ∗(EL(Xi)) are mutually non-Morita equivalent. Fix i ∈ I, and let J ⊳ C ∗(E) be a maximal ideal in C ∗(EL(Xi)). Then there exists exactly one x ∈ Xi such that IHx ⊆ J, where Hx := EL(Xi)0 \ {x}. (If there did not exist such an x, then J would be all of C ∗(EL(Xi)), and if there existed more than one x, then J would not be maximal.) Since IHx ⊆ J, it follows that J corresponds to a maximal ideal of the quotient C ∗(EL(Xj))/IHx . Because the graph EL(Xi) \ Hx is a single vertex with a single loop, we see that ∼= C(T). For this maximal ideal there is a unique z ∈ T C ∗(EL(Xj))/IHx such that the ideal is equal to {f ∈ C(T) : f (z) = 0}. This line of reasoning shows that the map J 7→ (x, z) is a bijection from the set of maximal ideals of C ∗(EL(Xi)) onto the set X × T. Since Xi ≥ 2ℵ0 and T = 2ℵ0, we may conclude that Xi ×T = Xi. Thus the set of maximal ideals of C ∗(EL(Xi)) has cardinality equal to Xi. Since any two Morita equivalent C ∗-algebras have isomorphic lattices of ideals, any two Morita equivalent C ∗-algebras have sets of maximal ideals with the same cardinality. Thus, when i, j ∈ I with i 6= j, the fact that Xi 6= Xj implies that C ∗(EL(Xi)) is not Morita equivalent to C ∗(EL(Xj)). 18 GENE ABRAMS AND MARK TOMFORDE For (3), the fact that the C ∗(EK (Xi)) are non-AF C ∗-algebras of real rank zero that are prime and not primitive and that all ideals are gauge invariant It remains to show the C ∗(EK (Xi)) are follows from Proposition 4.4(3). mutually non-Morita equivalent. The proof follows much like the proof of part (1): For each i ∈ I, the graph EK(Xi) satisfies Condition (K), and hence the ideals in C ∗(EK(Xi)) are in one-to-one correspondence with admissible pairs (H, S), where H is a saturated hereditary subset EK (Xi)0 and S is a subset of breaking vertices for H. For each x ∈ Xi define Hx := EK(Xi)0 \ {x}. Because {x} is a source and an infinite emitter in EA(Xi), the set Hx is saturated and hereditary. In addition, because there is a loop at {x}, it is a (unique) breaking vertex for Hx. In addition, any saturated hereditary subset of EK(Xi)0 that contains Hx is either equal to Hx or equal to EA(Xi)0. Thus I(Hx,{x}) is a maximal ideal in C ∗(EK (Xi)). Conversely, any maximal ideal must have the form I(H,S) for an admissible pair (H, S), and in order to be a proper ideal there exists x ∈ EK (Xi)0 \ H. Because H ⊆ Hx, the maximality of I(H,S) implies H = Hx and S = {x}. Thus we may conclude that the map x 7→ I(Hx,{x}) is a bijection from X onto the set of maximal ideals of C ∗(EK (Xi)). Thus the cardinality of the set of maximal ideals of C ∗(EK (Xi)) is equal to Xi, and as argued in (1) this implies that when i, j ∈ I with i 6= j, the C ∗-algebra C ∗(EK (Xi)) is not Morita equivalent to C ∗-algebra C ∗(EK (Xj)). (cid:3) Remark 4.6. Note that in each of parts (1) -- (3) of Proposition 4.5 we are constructing a prime, nonprimitive C ∗-algebra for each set in the collection C. We mention that for any cardinal number κ, there exists a collection of κ sets of differing cardinalities all greater than or equal to 2ℵ0. (This fact is well known; see e.g. [17, Lemma 7.7].) Hence in each of parts (1) -- (3) of Proposition 4.5 one can choose the collection C to be of any desired cardinality κ. Example 4.7. There are, of course, many examples of uncountable graphs whose associated C ∗-algebras are primitive. For instance, let X be any uncountable set, let P(X) denote the set of all subsets of X, and let EP (X) be the graph having E0 P(X) = P(X) and E1 P(X) = {eA,A′ A, A′ ∈ P(X) and A $ A′}, with s(eA,A′) = A, and r(eA,A′) = A′ for each eA,A′ ∈ E1 P(X). Then EP(X) is not a countable graph, and C ∗(EP(X)) is not a separable C ∗-algebra. However, EP(X) satisfies the three conditions of Theorem 3.8, and hence C ∗(EP(X)) is a primitive C ∗-algebra. (In particular, we observe that any vertex in EP(X) emits an edge pointing to {X} ∈ E0 P(X), so EP(X) triv- ially satisfies the Countable Separation Property.) The graph EP(X) has no cycles, so that C ∗(EP(X)) is an AF-algebra. In a like manner, we could construct additional examples of uncountable graphs having primitive graph C ∗-algebras, and one could easily produce non-AF examples by adding one or two loops at every vertex of EP(X). A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 19 The following definition provides a second graph-theoretic construction which produces examples of graphs whose corresponding graph C ∗-algebras are prime but not primitive. Definition 4.8. Let κ > 0 be any ordinal. We define the graph Eκ as follows: E0 κ = {α α < κ}, E1 κ = {eα,β α, β < κ, and α < β}, s(eα,β) = α, and r(eα,β) = β for each eα,β ∈ E1 κ. (cid:3) Recall that an ordinal κ is said to have countable cofinality in case κ is the limit of a countable sequence of ordinals strictly less than κ. For example, any countable ordinal has countable cofinality. The ordinal ω1 does not have countable cofinality, while the ordinal ωω does have this property. With this definition in mind, it is clear that E0 κ has the Countable Separation Property if and only if κ has countable cofinality. Thus by Theorem 3.8 we get Proposition 4.9. Let {κα α ∈ A} denote a set of distinct ordinals, each without countable cofinality. Then the collection {C ∗(Eκα) α ∈ A} is a set of nonisomorphic AF-algebras, each of which is prime but not primitive. 5. A comparison of the conditions for a graph C ∗-algebra to be simple, to be primitive, and to be prime In Proposition 3.1 and Theorem 3.8, we obtained conditions for a graph C ∗-algebra to be prime and primitive, respectively. In this section we com- pare these conditions with the conditions for a graph C ∗-algebra to be simple. Recall that an infinite path p in E is a nonterminating sequence p = e1e2e3 . . . of edges in E, for which r(ei) = s(ei+1) for all i ≥ 1. (Note that this notation, while standard, can be misleading; an infinite path in E is not an element of Path(E), as the elements of Path(E) are, by definition, finite sequences of edges in E.) We denote the set of infinite paths in E by E∞. Proposition 5.1. Let E be a graph. The graph C ∗-algebra C ∗(E) is simple if and only if the following two con- ditions are satisfied (1) E satisfies Condition (L), and (2) E is cofinal (i.e., if v ∈ E0 and α ∈ E∞ ∪ E0 sing, then v ≥ α0). The graph C ∗-algebra C ∗(E) is primitive if and only if the following three conditions are satisfied (1) E satisfies Condition (L), (2) E is downward directed (i.e., for all v, w ∈ E0 there exists x ∈ E0 such that v ≥ x and w ≥ x), and (3) E satisfies the Countable Separation Property (i.e., there exists a countable set X ⊆ E0 such that E0 ≥ X). The graph C ∗-algebra C ∗(E) is prime if and only if the following two con- ditions are satisfied 20 GENE ABRAMS AND MARK TOMFORDE (1) E satisfies Condition (L), and (2) E is downward directed (i.e., for all v, w ∈ E0 there exists x ∈ E0 such that v ≥ x and w ≥ x). Proof. Proposition 3.1 and Theorem 3.8 give the stated conditions for a graph C ∗-algebra to be prime and primitive, respectively. The conditions for simplicity are established in [27, Theorem 2.12]. (Although all of [27] is done under the implicit assumption the graphs are countable, the proof of [27, Theorem 2.12] and the proofs of the results on which it relies all go (cid:3) through verbatim for uncountable graphs.) Every C ∗-algebra has a nonzero irreducible representation. (This follows from the GNS construction, which shows that GNS-representations con- structed from pure states are irreducible [24, Lemma A.12], and the Krein- Milman Theorem, which asserts that pure states exist for any C ∗-algebra [24, Lemma A.13].) Thus any simple C ∗-algebra has a faithful irreducible ∗-representation, and any simple C ∗-algebra is necessarily primitive. More- over, as was shown in Lemma 3.2, any primitive C ∗-algebra is necessarily prime. Thus we have C ∗(E) is simple =⇒ C ∗(E) is primitive =⇒ C ∗(E) is prime. We observe that the form of each of the three results presented in Proposi- tion 5.1, in which simplicity, primeness, and primitivity of C ∗(E) are given in graph-theoretic terms, may be described as "Condition (L) plus some- thing extra". This having been said, our goal for this section is solely graph-theoretic: we show that these three "extra" conditions may be seen as arising in a common context, by considering subsets of E0 of the form H(v). Proposition 5.2. Let E be a graph. Then the following equivalences hold. (1) E is cofinal if and only if for all v ∈ E0 one has H(v) = E0. (2) E is downward directed if and only if for all v, w ∈ E0 one has H(v) ∩ H(w) 6= ∅. (3) E satisfies the Countable Separation Property if and only if there exists a countable collection of subsets of vertices {Si : i ∈ I} (so, I is countable and Si ⊆ E0 for all i ∈ I) with E0 = Si∈I Si and with Tv∈Si H(v) 6= ∅ for all i ∈ I. Proof. For (1), suppose first that for all v ∈ E0 we have H(v) = E0. Choose v ∈ E0 and α ∈ E∞ ∪ E0 sing, then α ∈ H(v) = E0. Since every element of H(v) \ H(v) is a regular vertex, it follows that α ∈ H(v) If instead α ∈ E∞, then we may write α = e1e2e3 . . . for and v ≥ α. ei ∈ E1 with r(ei) = s(ei+1) for all i ∈ N. Since s(e1) ∈ H(v) = E0, it follows that s(ei) ∈ H(v)n for some n ∈ N (recall the notation of Lemma 2.11). Thus s(e2) = r(e1) ∈ H(v)n−1, and continuing recursively we have s(en) = r(en−1) ∈ H(v)1, and s(en+1) = r(en) ∈ H(v)0 = H(v). Hence sing. If α ∈ E0 A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 21 v ≥ s(en+1), and v ≥ α0. Thus E is cofinal. Conversely, if there exists v ∈ E0 with H(v) 6= E0, then there is a vertex w ∈ E0 \ H(v). Since H(v) is saturated, either w ∈ E0 sing or there exists an edge e1 ∈ E1 with s(e1) = w and r(e1) /∈ H(v). Using r(e1) and continuing inductively, we either produce a singular vertex z ∈ E0 \ H(v) or an infinite path α := e1e2e3 . . . with s(ei) ∈ E0 \ H(v). Since H(v) ⊆ H(v), it follows that either there is a singular vertex z with v 6≥ z or there is an infinite path α with v 6≥ α0. Hence E is not cofinal. write H(v) = S∞ For (2), suppose first that E is downward directed. If v, w ∈ E0, then the definition of downward directed implies that H(v)∩H(w) 6= ∅. Since H(v) ⊆ H(v) and H(w) ⊆ H(w), it follows that H(v) ∩ H(w) 6= ∅. Conversely, suppose that for all v, w ∈ E0 one has H(v)∩H(w) 6= ∅. Then for any v, w ∈ E0, we may choose x ∈ H(v) ∩ H(w). Using the notation of Definition 2.10 n=0 H(w)n. Choose the smallest n1 ∈ N∪{0} such that x ∈ H(v)n1, and choose the smallest n2 ∈ N∪{0} such that x ∈ H(w)n2. If we let n := max{n1, n2}, then x ∈ H(v)n ∩ H(w)n. If n = 0, then x ∈ H(v) ∩ H(w) and we have that v ≥ x and w ≥ x. If n ≥ 1, then there exists an edge e1 ∈ E1 such that s(e1) = x and r(e1) ∈ H(v)n−1 ∩ H(w)n−1. Using r(e1) next, and continuing recursively, we produce a finite path α := e1 . . . en with r(en) ∈ H(v) ∩ H(w). Hence v ≥ r(en) and w ≥ r(en). Thus E is downward directed. n=0 H(v)n and H(w) = S∞ For (3), suppose first that E satisfies the Countable Separation Property. x ∈ X. Thus the condition in (3) holds with I := X and Si := U (i) for all i ∈ I. Conversely, suppose that there exists a countable collection of subsets Then there is a countable nonempty set X ⊆ E0 such that E0 =Sx∈X U (x). In addition, x ∈ Tv∈U (x) H(v) for all x ∈ X, so Tv∈U (x) H(v) 6= ∅ for all of vertices {Si : i ∈ I} with E0 = Si∈I Si and with Tv∈Si H(v) 6= ∅ for all i ∈ I. For each i ∈ I, choose a vertex vi ∈Tv∈Si H(v) and define Cvi := {r(α) : α ∈ Path(E), s(α) = v, and s(αi) ∈ E0 reg for all 1 ≤ i ≤ α}. (Note that if vi ∈ E0 sing, then Cvi := {vi}.) Since there are only a finite num- ber of edges emitted from any regular vertex, we see that Cvi is a countable countable union of countable sets, X is countable. If w ∈ E0 is any vertex, By the definition of vi we then have that vi ∈ H(w). Hence there exists a path α ∈ Path(E) such that s(α) = vi, r(α) ∈ H(w), and s(αi) ∈ E0 reg for set for all i ∈ I. We define X := Si∈I Cvi, and observe that since X is a then by the hypothesis that E0 =Si∈I Si there exists i ∈ I such that w ∈ Si. all 1 ≤ i ≤ α. Thus r(α) ∈ Cvi, and w ≥ r(α), so that w ∈ Sx∈X U (x). We have therefore shown that E0 = Sx∈X U (x), and hence E satisfies the Countable Separation Property. (cid:3) It is clear that Property (1) of Proposition 5.2 implies both Property (2) and Property (3) of that Proposition. (Note that we may use the singleton 22 GENE ABRAMS AND MARK TOMFORDE set S = E0 to establish Property (3) from Property (1).) Thus, as promised, using Proposition 5.1, we have established a natural connection between simplicity, primitivity, and primeness for graph C ∗-algebras from a graph- theoretic point of view. We summarize this observation as the following result. Corollary 5.3. Let E be a graph. The graph C ∗-algebra C ∗(E) is simple if and only if the following two con- ditions are satisfied (1) E satisfies Condition (L) (2) If v ∈ E0, then H(v) = E0. The graph C ∗-algebra C ∗(E) is primitive if and only if the following three conditions are satisfied (1) E satisfies Condition (L) (2) If v, w ∈ E0, then H(v) ∩ H(w) 6= ∅. (3) There exists a countable collection of subsets of vertices {Si : i ∈ I} (so, I is countable and Si ⊆ E0 for all i ∈ I) such that E0 =Si∈I Si and Tv∈Si H(v) 6= ∅ for all i ∈ I. The graph C ∗-algebra C ∗(E) is prime if and only if the following two con- ditions are satisfied (1) E satisfies Condition (L) (2) If v, w ∈ E0, then H(v) ∩ H(w) 6= ∅. We conclude this graph-theoretic section with the following observation. Since the simplicity of C ∗(E) clearly implies its primeness, it is perhaps not surprising that there should be a direct graph-theoretic connection be- Indeed, one tween the germane properties appearing in Proposition 5.1. can easily see that cofinal implies downward directed, as follows. If E is cofinal, and v, w ∈ E0, then one may inductively create a sequence of edges α := e1e2e3 . . . with s(e1) = w and s(ei) = r(ei−1) for all i ≥ 2, and such that this sequence either ends at a sink or goes on forever to produce an infinite path. Hence either v ≥ r(α) (if α ends at a sink) or v ≥ α0 (if α is an infinite path), and E is downward directed. From this point of view, the difference between the notion of cofinal and the notion of downward directed can be viewed as follows: E is cofinal if and only if "for all v, w ∈ E0 and for all α ∈ Path(E) with s(α) = w there exists x ∈ E0 such that v ≥ x and r(α) ≥ x", while E is downward directed if and only if "for all v, w ∈ E0 and for some α ∈ Path(E) with s(α) = w there exists x ∈ E0 such that v ≥ x and r(α) ≥ x." Specifically, the cofinality property allows for the path from one of the vertices to start along any specified initial segment α, while the downward directedness property contains no such requirement. A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 23 6. Primality and primitivity of graph C ∗-algebras compared with primality and primitivity of Leavitt path algebras In this final section we compare the notions of primeness and primitivity for graph C ∗-algebras C ∗(E) with primeness and primitivity for Leavitt path algebras. Briefly, for any graph E and any field K, one may define the K-algebra LK(E), the Leavitt path algebra of E with coefficients in K. When K = C, then LC(E) may be viewed as a dense ∗-subalgebra of C ∗(E). For reasons which remain not well understood, many structural properties are simultaneously shared by both LC(E) and C ∗(E). We show in this section that the primitivity property may be added to this list. Additional information about Leavitt path algebras may be found in [1] or [2]. The map Pn i is a K-algebra isomorphism from LK(E) onto its opposite algebra LK(E)op. Hence there is a natural cor- respondence between left LK(E)-modules and right LK(E)-modules, which yields 7→ Pn i=1 λiαiβ∗ i i=1 λiβiα∗ Proposition 6.1. [2, Proposition 2.2] If E is a graph and K is a field, then the algebra LK (E) is left primitive if and only if it is right primitive. Definition 6.2. In light of Proposition 6.1, we shall say a Leavitt path algebra LK(E) is primitive if it is left primitive (which occurs if and only if LK(E) is also right primitive). Remark 6.3. When we say LK (E) is primitive, we mean that LK(E) is primitive as a ring. The astute reader may notice that it seems more natural to consider primitivity of LK(E) as an algebra; that is, to reformulate the definition of primitive as having a simple faithful left K-algebra module (not merely a simple faithful left ring module). However, since LK (E) has local units, any ring module also carries a natural structure as a K-algebra module. Hence any Leavitt path algebra is primitive as a ring if and only if it is primitive as a K-algebra. Likewise, again using that Leavitt path algebras have local units, any ring ideal of LK(E) is closed under scalar multiplication by K, and hence the ring ideals of LK(E) are precisely the K-algebra ideals of LK(E). Consequently, a Leavitt path algebra is prime as a ring if and only if it is prime as a K-algebra, and a Leavitt path algebra is simple as a ring if and only if it is simple as a K-algebra. Thus for Leavitt path algebras the ring-theoretic notions of primitive, prime, and simple coincide with the corresponding K-algebra-theoretic notions. Theorem 6.4. Let E be a graph. Then the following are equivalent. (i) C ∗(E) is primitive. (ii) LK (E) is primitive for some field K. (iii) LK (E) is primitive for every field K. (iv) E satisfies Condition (L), is downward directed, and satisfies the Countable Separation Property. 24 GENE ABRAMS AND MARK TOMFORDE Proof. The equivalence of (i) and (iv) is precisely Theorem 3.8. The equiv- (cid:3) alence of (ii), (iii), and (iv), is shown in [2, Theorem 5.7]. Theorem 6.4 provides yet another example of a situation in which the same ring-theoretic property holds for both of the algebras C ∗(E) and LC(E) (in- deed, LK(E) for any field K), but for which the proof that the pertinent property holds in each case is wildly different. In particular, no "direct" con- nection between C ∗(E) and LC(E) is established. We note that the proof of the sufficiency direction of Theorem 3.8 looks, on the surface, nearly identi- cal to the proof that LC(E) is primitive whenever E satisfies Condition (L), is downward directed, and satisfies the Countable Separation Property [2, Theorem 3.5 with Proposition 4.8]. However, in the proof of the result herein we invoke the Cuntz-Krieger Uniqueness Theorem, whose justification is sig- nificantly different than that of the correspondingly invoked algebraic result [4, Corollary 3.3]. Furthermore, the proof of the necessity direction of The- orem 3.8 is significantly different than the proof of the analogous result for Leavitt path algebras [2, Proposition 5.6]. In this regard, it is worth noting that for Leavitt path algebras, in contrast to Lemma 3.4 for C ∗-algebras, it is perfectly possible to have a graph E and left LK (E)-module M con- taining an element m for which there exists an uncountable set of nonzero orthogonal projections in LK(E) which do not annihilate m. For example, let U be an uncountable set, and let EU denote the graph having one vertex v, and uncountably many loops {ei i ∈ U } at v. Let R = LK(EU ), and let M = RR. Then for m = 1R ∈ M , {eie∗ i i ∈ U } is such a set. In contrast to the result presented in Theorem 6.4, the class of graphs which produce prime Leavitt path algebras is not the same as the class of graphs which produce prime graph C ∗-algebras. For example, if we let E be the graph with one vertex and one edge • then for any field K, the Leavitt path algebra LK(E) is isomorphic to K[x, x−1], the algebra of Laurent polynomials with coefficients in K, which is prime. (Indeed, K[x, x−1] is a commutative integral domain.) However, the graph C ∗-algebra C ∗(E) is isomorphic to C(T), the C ∗-algebra of con- tinuous functions on the circle, which is not prime. Thus "primeness" yields one of the relatively uncommon contexts in which an algebraic property of LK(E) does not coincide with the corresponding C ∗-algebraic property of C ∗(E). Hence the conditions on E for LK(E) to be prime are different than the conditions on E for C ∗(E) to be prime. Necessary and sufficient conditions for a Leavitt path algebra to be prime are given in [3, Corollary 3.10] (see also [2, Theorem 2.4]), which we state here. Proposition 6.5. Let E be a graph. Then the following are equivalent A CLASS OF C ∗-ALGEBRAS THAT ARE PRIME BUT NOT PRIMITIVE 25 (i) LK (E) is prime for some field K. (ii) LK (E) is prime for every field K. (iii) E is downward directed. We conclude this article with the following summary of comparisons of germane properties between Leavitt path algebras and graph C ∗-algebras. A proof that the indicated conditions on E which yield the simplicity of LK(E) for any field K is given in [1]. The remaining comparisons follow from Proposition 6.5 with Proposition 3.1 and Theorem 6.4. LK(E) is simple ⇐⇒ C ∗(E) is simple ⇐⇒ E is cofinal, and E satisfies Condition (L) LK (E) prime ⇐⇒ E is downward directed C ∗(E) prime ⇐⇒ E is downward directed, and E satisfies Condition (L) LK(E) is primitive ⇐⇒ C ∗(E) is primitive ⇐⇒ E is downward directed, E satisfies Condition (L), and E has the Countable Separation Property In particular, we note that C ∗(E) is prime if and only if LK(E) is prime and E satisfies Condition (L). Specifically, C ∗(E) prime implies LK(E) prime for every field K, but not conversely. References [1] G. Abrams, P. Ara, and M. Siles Molina, Leavitt path algebras. Lecture Notes in Mathematics, Springer Verlag, to appear. [2] G. Abrams, J. Bell, and K.M. Rangaswamy, On prime non-primitive von Neumann regular algebras, Trans. Amer. Math. Soc., to appear. [3] G. Aranda Pino, E. Pardo, and M. Siles Molina, Exchange Leavitt path algebras and stable rank, J. Algebra 305 (2006), 912 -- 936. [4] G. Aranda Pino, D. Mart´ın Barquero, C. Mart´ın Gonz´alez, and M. Siles Molina, The socle of a Leavitt path algebra, J. Pure App. Alg. 212 (2008), 500 -- 509. [5] T. Bates, J. Hong, I. Raeburn, and W. Szyma´nski, The ideal structure of the C ∗- algebras of infinite graphs, Illinois J. Math. 46(4) (2002), 1159 -- 1176. [6] T. Bates, D. Pask, I. Raeburn, and W. Szyma´nski, The C ∗-algebras of row-finite graphs, New York J. Math. 6 (2000), 307 -- 324. [7] G.M. Bergman, A ring primitive on the right but not on the left, Proc. Amer. Math. Soc. 15 (1964) 473 -- 475. [8] G.M. Bergman, Errata: A ring primitive on the right but not on the left', Proc. Amer. Math. Soc. 15 (1964), 1000. [9] M.J. Crabb, A new prime C ∗-algebra that is not primitive, J. Funct. Analysis 236 (2006), 630 -- 633. 26 GENE ABRAMS AND MARK TOMFORDE [10] J.B. Conway, A Course in Functional Analysis. Graduate Texts in Mathematics 96, Springer-Verlag, New York, 1985. ISBN 0-387-96042-2. [11] J. Dixmier, Sur les C ∗-algebres, Bull. Soc. math France 88 (1960), 95 -- 112. [12] J. Dixmier, Les C ∗-algebres et leurs representations, Gauthier-Villars, 1969. (English translation: C ∗-algebras, North-Holland, 1977.) [13] D. Drinen and M. Tomforde, The C ∗-algebras of arbitrary graphs, Rocky Mountain J. Math. 35 (2005), 105 -- 135. [14] K.R. Goodearl, Leavitt path algebras and direct limits, Rings, Modules, and Repre- sentations, 165 -- 187, Contemp. Math., 480, Amer. Math. Soc., Providence, RI, 2009. [15] J.H. Hong and W. Szyma´nski, Purely infinite Cuntz-Krieger algebras of directed graphs, Bull. London Math. Soc. 35 (2003), 689 -- 696. [16] A. an Huef and I. Raeburn, The ideal structure of Cuntz-Krieger algebras, Ergodic Theory Dynam. Systems 17 (1997), 611 -- 624. [17] T. Jech, Set theory. The third millennium edition, revised and expanded. Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2003. xiv+769 pp. [18] J. A Jeong and G.H. Park, Graph C ∗-algebras with real rank zero, J. Funct. Anal. 188 (2002), 216 -- 226. [19] T. Katsura, Non-separable AF-algebras, in Operator Algebras: The Abel Symposium 2004. Proceedings of the First Abel Symposium, Oslo, September 3 -- 5, 2004, Springer- Verlag, Berlin Heidelberg, 2006, pp.165 -- 173. ISBN: 3-540-34196-X. [20] T. Katsura, A class of C ∗-algebras generalizing both graph algebras and homeomor- phism C ∗-algebras I, Fundamental results, Trans. Amer. Math. Soc. 356 (2004), 4287 -- 4322. [21] T. Katsura, A class of C ∗-algebras generalizing both graph algebras and homeomor- phism C ∗-algebras III, Ideal structures, Ergodic Theory Dynam. Systems 26 (2006), 1805 -- 1854. [22] N.C. Phillips and M.G. Viola, A simple separable exact C ∗-algebra not anti- isomorphic to itself, Math. Ann. 355 (2013), 783 -- 799. [23] I. Raeburn and W. Szyma´nski, Cuntz-Krieger algebras of infinite graphs and matrices, Trans. Amer. Math. Soc. 356 (2004), 39 -- 59. [24] I. Raeburn and D.P. Williams, Morita Equivalence and Continuous-Trace C ∗- algebras, Math. Surveys & Monographs, vol. 60, Amer. Math. Soc., Providence, 1998. [25] C.E. Rickart, General theory of Banach algebras, The University Series in Higher Mathematics, D. van Nostrand Co., Inc., Princeton, N.J.-Toronto-London-New York 1960 xi+394 pp. [26] L.H. Rowen, Ring theory. Vol. I. Pure and Applied Mathematics, 127. Academic Press, Inc., Boston, MA, 1988. xxiv+538 pp. [27] M. Tomforde, Structure of graph C*-algebras and their generalizations, Chapter in the book "Graph Algebras: Bridging the gap between analysis and algebra", Eds. Gonzalo Aranda Pino, Francesc Perera Dom`enech, and Mercedes Siles Molina, Servicio de Publicaciones de la Universidad de M´alaga, M´alaga, Spain, 2006. [28] N. Weaver, A prime C ∗-algebra that is not primitive, J. Funct. Analysis 203 (2003), 356 -- 361. Department of Mathematics, University of Colorado, Colorado Springs, CO 80918, USA E-mail address: [email protected] Department of Mathematics, University of Houston, Houston, TX 77204- 3008, USA E-mail address: [email protected]
1501.01038	9	1501	2016-10-02T15:39:52	The Cuntz semigroup of the tensor product C*-algebras	[ "math.OA" ]	We calculate the Cuntz semigroup of the tensor product A with A. We restrict our attention to C*-algebras A which are unital, simple, nuclear, stably finite, have stable rank one, absorbs the Jiang-Su algebra tensorially and satisfy the UCT.	math.OA	math	THE CUNTZ SEMIGROUP OF THE TENSOR PRODUCT OF C-ALGEBRAS CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y Abstract. We calculate the Cuntz semigroup of a tensor prod- uct C∗-algebra A ⊗ A. We restrict our attention to C∗-algebras which are unital, simple, separable, nuclear, stably finite, Z-stable, satisfy the UCT, with finitely generated K0(A) and have trivial K1(A). On calcule le semigroupe de Cuntz d'une C∗-alg`ebre produit tensoriel A ⊗ A. On consid`ere seulement les C-alg`ebres simples, s´eparable, nucl´eaires, `a ´el´ement unit´e, stablement finies, Z-stables, satisfaisant au UCT, dont le groupe K0(A) est de type fini, et dont le groupe K1(A) est trivial. Key words and phrases. C∗-algebra, Cuntz semigroup, K0-group, tensor product. Date: Received December 8, 2014. 2000 Mathematics Subject Classification. Primary 46L35, 46L06. 1 2 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y 1. Introduction The Cuntz semigroup has been studied since the late seventies but only recently has it proven to be an important invariant for C ∗-algebras. First, in the early 2000s, M. Rordam and A. Toms constructed exam- ples of C ∗-algebras that appeared to be counterexamples to the Elliott conjecture. Shortly afterwards, Toms realized that the Cuntz semi- group distinguishes some of the newly constructed algebras; hence, the Cuntz semigroup could be added to the Elliott invariant. Toms's dis- covery obviously prompted major questions, such as: "What is the range of the Cuntz semigroup?" or "What is the relation between the Cuntz semigroup and the Elliott invariant?" or "What are the proper- ties of the Cuntz semigroup?" In this paper we propose to study one property of the Cuntz semi- group, namely, how the Cuntz semigroup of the tensor product, A ⊗ A, of two identical copies of the C-algebra A relates to the Cuntz semi- group of A. It is well known that the tensor product of two positive elements is still a positive element. This property allows us to define a natural tensor product map from A+⊗A+ to (A⊗A)+. The usual inter- pretation of A+ ⊗ A+ is as a subset of the usual tensor product of (nu- clear) C ∗-algebras. However, defining maps at the level of Cuntz semi- groups requires defining tensor products of Cuntz semigroups, which are a priori tensor products of abelian semigroups. Hence we must consider semigroup tensor products, discussed below. Our approach to tensor products of Cuntz semigroups is to first take an algebraic tensor product of abelian semigroups and then to take a completion with respect to a suitable topology. See [19, para. 6.L] for more infor- mation on topological completions. The basic reason for introducing completions is that if we use only algebraic tensor products we can obtain surjectivity results only in very limited situations, such as the finite-dimensional case. In the first three sections of this paper we work with the algebraic tensor product, and we use the term "dense range" for results from which we later obtain surjectivity as a corollary, after THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 3 taking a completion. We consider completions in the last section of the paper. As defined by Grillet [15], the tensor product of two abelian semi- groups is constructed by forming a free abelian semigroup and passing to the quotient by the relations (a + a′) ⊗ b = (a ⊗ b) + (a′ ⊗ b) and (a ⊗ b′) + (a ⊗ b) = a ⊗ (b + b′). This definition is equivalent [15] to the definition by a universal property. Stating the universal product defi- nition for a family (Ai)i∈I of semigroups, we first say that a mapping s of the Cartesian product of semigroups Q Ai into a semigroup C is I-linear if the mapping is a semigroup homomorphism in each variable separately. Then, if an I-linear mapping t of Q Ai into a semigroup T has the property that, for any I-linear mapping s of Q Ai into some semigroup C, there exists a unique homomorphism u of T into C such that s = u ◦ t, then we call the pair (t, T ), and also the semigroup T, a tensor product of the family (Ai)i∈I. It is well known that not every positive element of a tensor product can be written as a tensor product of positive elements, even after allowing sums. Thus, the naive tensor product map from A+ ⊗ A+ to (A ⊗ A)+ is in general not surjective. It seems interesting that, as we shall see, in some cases this map becomes surjective if we pass to Cuntz equivalence classes. In a recent paper [2], the question of determining surjectivity, at the level of Cuntz semigroups, of the natural tensor product map is posed; and left as an open problem. In that paper, the authors state that surjectivity does hold in the cases of AF algebras and O∞-stable algebras. This is not surjectivity at the level of algebraic tensor prod- ucts, rather it is with respect to a particular choice of Cuntz semigroup tensor product introduced in [2], called the Cuntz category tensor prod- uct. We will consider the case of simple, separable, unital, stably finite, nuclear, Z-stable C∗-algebras, with finitely generated K0 group, triv- ial K1-group and satisfying the UCT, and we show that the image of the natural tensor product map on algebraic elements is dense in the 4 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y sense that it becomes surjective after passing to a completion. We con- sider completions with respect to several different possible topologies, the coarsest of these being given by pointwise suprema, as will be ex- plained in the last section of the paper. We use [23] to deduce that algebras in the abovementioned class have stable rank one. The stable rank one property and its consequences are used several times in our proofs. Brown, Perera and Toms, [9], showed an important representation result for the original version of the Cuntz semigroup. This result was extended to the nonunital case, using the stabilized version of the Cuntz semigroup, by Elliott, Robert and Santiago, [14], and with more ab- stract hypotheses by Tikuisis and Toms, [28]. Their results (see Theo- rems 2.1 and 2.1) imply that for certain simple exact C∗-algebras, a part of the Cuntz semigroup is order isomorphic to an ordered semigroup of lower semicontinuous functions defined on a compact Hausdorff space. 2. The Cuntz semigroup Let A be a separable C-algebra. For positive elements a, b ∈ A ⊗ K, we say that a is Cuntz subequivalent to b, and write a (cid:22) b, if vnbv∗ n → a in the norm topology, for some sequence (vn) in A ⊗ K. We say that a is Cuntz equivalent to b and write a ∼ b if a (cid:22) b and b (cid:22) a. Denote by Cu(A) the set of Cuntz equivalence classes of the positive cone of A ⊗ K, i.e. Cu(A) = (A ⊗ K)+/∼. The order relation a (cid:22) b defined for the positive elements of A ⊗ K induces an order relation on Cu(A): [a] ≤ [b] if a (cid:22) b, where [a] denotes the Cuntz equivalence class of the positive element a. Note (cf. [24, page 151]) that this order relation does not need to be the algebraic order with respect to the addition operation defined by setting [a] + [b] := [a′ + b′], where a′ and b′ are orthogonal positive elements. It turns out that in a stabilization we can always find such orthogonal representatives, i.e., in (A ⊗ K)+ we have a ∼ a′, b ∼ b′ with a′b′ = 0. Moreover, the choice of the orthogonal representatives does not affect the Cuntz class of their sum. So the ordered set Cu(A) becomes an abelian semigroup, under an addition THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 5 operation that is sometimes called Brown-Douglas-Fillmore addition [8]. If A is unital, we denote by T (A) the simplex of tracial states. By V (A) we denote the projection semigroup defined by the Murray von Neumann equivalence classes of projections in A ⊗ K. The order structure on V (A) is defined through Murray-von Neumann comparison of projections. 2.1. Representations of the Cuntz semigroup. Brown, Perera and Toms's representation result [9] for the Cuntz semigroup is as fol- lows: Theorem 2.1. Let A be a simple, separable, unital, exact, stably finite Z-stable C∗-algebra. Then there is an order preserving isomorphism of ordered semigroups, W (A) ∼= V (A) ⊔ Lsc(T (A), (0, ∞)). In the statement of the above theorem, W (A) is the original defini- tion of the Cuntz semigroup, i.e., W (A) = M∞(A)+/∼, and Lsc(T (A), (0, ∞)) denotes the set of lower semicontinuous, affine, strictly positive func- tions on the tracial state space of the unital C-algebra A. Addition within Lsc(T (A), (0, ∞)) is done pointwise and order is defined through pointwise comparison, as is usual for functions. For [p] ∈ V (A) and f ∈ Lsc(T (A), (0, ∞]), addition is defined by [p] + f := [p] + f ∈ Lsc(T (A), (0, ∞) ), where [a](τ ) = lim n→∞ a projection. The order relation is given by: τ (a1/n), τ ∈ T (A), which reduces to τ (a) when a is [p] ≤ f if [p](τ ) < f (τ ) for all τ ∈ T (A), f ≤ [p] if f (τ ) ≤ [p](τ ) for all τ ∈ T (A). Elliott, Robert and Santiago's representation result [14] is very sim- ilar, and uses the stabilized Cuntz semigroup. In this result, the func- tions that appear may take infinite values and the algebras are not necessarily unital. Since we restrict our attention to the case of unital 6 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y algebras, the domain, T (A), can be taken to be a compact simplex, which in turn gives a simplified version of their result: Theorem 2.2. Let A be a simple, separable, unital, exact, stably finite Z-stable C-algebra. Then there is an order preserving isomorphism of ordered semigroups, Cu(A) ∼= V (A) ⊔ Lsc(T (A), (0, ∞]), where Lsc(T (A), (0, ∞]) will denote the set of lower semicontinuous, possibly infinite, affine, strictly positive functions on the tracial state space of a unital C-algebra A. Within Lsc(T (A), (0, ∞]) addition is pointwise and pointwise comparison is used. For [p] ∈ V (A) and f ∈ Lsc(T (A), (0, ∞]), addition is given by [p] + f := [p] + f ∈ Lsc(T (A), (0, ∞]), where [a](τ ) = lim n→∞ a projection. The order relation is given by: τ (a1/n), τ ∈ T (A), which reduces to τ (a) when a is [p] ≤ f if [p](τ ) < f (τ ) for all τ ∈ T (A), f ≤ [p] if f (τ ) ≤ [p](τ ) for all τ ∈ T (A). In the proof of the above theorems, a semigroup map i : Cu(A) −→ Lsc(T (A), (0, ∞]) is defined, i([a])(τ ) = dτ (a), with dτ to be explained later. These theorems show that the Cuntz semigroup, say Cu(A), is the disjoint union of the semigroup of positive elements coming from pro- jections in (A ⊗ K)+, denoted V (A), and the set of lower semicontinu- ous, affine, strictly positive, functions on the tracial state space of A, denoted by Lsc(T (A), (0, ∞]). In [9], the elements of the Cuntz semi- group that correspond to lower semicontinuous, affine, strictly positive, functions on the tracial state space are termed purely positive elements. In general, the set of purely positive elements does not form an object in the Cuntz category. To see this, consider an element x with spectrum THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 7 [ǫ, 1] in a C-algebra of stable rank 1. The increasing sequence (x− 1 n)+ is at first purely positive, but has a supremum that is projection-class. Theorem 2.2 implies that the subsemigroup of purely positive ele- ments of the Cuntz semigroup, for certain C-algebras A, is isomorphic to the semigroup Lsc(T (A), (0, ∞]). The convex structure of the space of tracial states, T (A), makes it a Choquet simplex when A is unital [13], metrizable when A is separable. We will need a result about lower semicontinuous functions on metriz- able Choquet simplices. Proposition 2.3. Let S be a compact metrizable Choquet simplex. Then every positive element of Lsc(S, (0, ∞]), bounded or not, is the pointwise supremum of some pointwise nondecreasing sequence of con- tinuous positive functions on S. Proof. Let S be a compact metrizable Choquet simplex. Using Ed- ward's separation theorem inductively shows, see Lemma 6.1 in [1], that every lower semicontinuous positive affine function on the simplex, possibly with infinite values, is the pointwise supremum of a strictly in- creasing sequence of affine continuous functions without infinite values. Compactness lets us arrange that the functions are everywhere posi- tive, for example, we may replace each fn by the pointwise supremum of {fn, ǫ1} for a suitably small ǫ. (cid:3) The dual of the Cuntz semigroup is denoted by D(A), and is referred to as the set of all dimension functions. It is the set of all additive, suprema-preserving, and order preserving maps d : Cu(A) → (0, ∞] such that, in the unital case, d([1]) = 1. If the map on A+ given by x 7→ d([x]) is lower semicontinuous, we say that the dimension function is lower semicontinuous. The lower semicontinuous dimension functions correspond to the 2-quasitraces, by Proposition 2.24 of [6]; for more information, see [14]. In the general case, once Theorem 2.2 is no longer applicable, the dual space of the Cuntz semigroup is strictly larger than the set of traces, T (A). See [5, page 307] for an example of a dimension function that is not lower semicontinuous and thus does 8 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y not come from either a trace or quasitrace. This example arises from a nonsimple and nonunital C-algebra. We don't know if there is an example coming from a simple and nuclear C-algebra. We also note that nonsimple purely infinite algebras may have a nontrivial Cuntz semigroup, but do not have traces. Thus, their dimension functions do not come from traces. Any (necessarily not exact; see [18]) C-algebra for which the quasi- traces are not all traces will be an example where the states on the Cuntz semigroup do not correspond to the traces. Reviewing the liter- ature dealing with the Cuntz semigroup, the algebraic structure of the Cuntz semigroup is generally the main topic of interest, and the topo- logical structure is hardly ever explicitly mentioned. We mention here some minor but apparently new observations about the topology of the Cuntz semigroup. We have Hausdorff metrics: D(A) is metrizable when A is separable, the metric being given by P d1(xk) − d2(xk)2−k, where xk is a dense subsequence of the positive part of the unit ball of A. Similarly, the Cuntz semigroup itself has at least a pseudomet- ric, in the presence of separability, of the form P dk(x1) − dk(x2)2−k, where dk is a dense subsequence of D(A). We note that in general there may exist projection-class elements that are equal, under the dimension functions, to purely positive elements. In the stable rank 1 case, it is possible to discriminate between projection-class elements and purely positive elements on the basis of a spectral criterion. 2.2. Dimension functions and a conjecture of Blackadar and Handelman. We have seen, as in Theorem 2.2, that the map i is useful in describing the order on the Cuntz semigroup. The map i is i(a) = dτ (a), where we define dτ (a) to be an extended version of the τ (a1/n), where τ is a tracial state. This map, dτ , rank of a: dτ (a) = lim n→∞ also called a dimension function, is lower semicontinuous as a map from A+ to [0, ∞], possibly taking infinite values, and defines a state on the Cuntz semigroup. In 1982, Blackadar and Handelman conjectured, see [5], that the set of lower semicontinuous dimension functions that come from traces is weakly dense in the set of dimension functions (or states THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 9 on the Cuntz semigroup). The conjecture is known to be true for a large class of C-algebras, see [29, page 426], that includes the algebras that we propose to study in this paper, namely: simple, unital, stably finite, nuclear, Z-stable C-algebras, with stable rank one, trivial K1- group and with the UCT property. We note that by [23], the stable rank one property follows from the other properties. From now on, we thus assume that the Blackadar-Handelman conjecture holds. Consider the map t : A+ × A+ → Cu(A ⊗ A) defined by t(a, b) = [a ⊗ b]. Let us check that the above map t respects Cuntz equivalence. Lemma 2.4. Let A be a σ-unital C-algebra. Given positive elements a, a′, b in A such that a′ (cid:22) a we have a ⊗ b (cid:22) a′ ⊗ b. Proof. Let en be a countable approximate unit. Since a′ (cid:22) a, choose an en such that enae∗ n → a′. We have (en ⊗ en)(a ⊗ b)(en ⊗ en)∗ − a′ ⊗ b = enae∗ n ⊗ enbe∗ n − a′ ⊗ b, and so (en ⊗ en)(a ⊗ b)(en ⊗ en)∗ − a′ ⊗ b → 0 (cid:3) If a ∼ a′ then a ⊗ b ∼ a′ ⊗ b by applying the lemma twice, and thus we obtain the Corollary: Corollary 2.5. Consider the map t : A+ × A+ → Cu(A ⊗ A) defined by t(a, b) = [a ⊗ b]. If a and a′ are positive elements of A that are Cuntz equivalent, then t(a, b) = t(a′, b). 3. Main result We begin with a technical lemma that is used in proving our main results. Lemma 3.1. Let S1 and S2 be compact metrizable Choquet simplices. Let F be a positive, (bi)affine, continuous finite-valued function on S1 × S2. The function F can be approximated uniformly from below by 10 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y a finite sum Pi,j aijf (1) positive functions on Sk and the aij are positive scalars. (y) where the f (k) i (x)f (2) j are continuous affine i Proof. Suppose first that S1 = S2 = S. The affine continuous functions A(S) on the compact Choquet simplex S happen to be a Banach space whose dual is an L1-space, see, e.g., [21, pg.181]. The space A(S) is separable because S is metrizable. We can therefore apply Theorem 3.2 in [21] to obtain an inductive limit decomposition of A(S) in the form A(S) = [ En, ∞ has a basis, {fj}mi where the En are finite-dimensional l∞-spaces and the connecting maps are inclusion maps. We denote the dimension of En by mn. Each subspace Ei = ℓmi j=1, of elements of A(Smi)+ with Pj fj = 1 in that subspace. Since the connecting maps in the above inductive limit are inclusion maps, we can inductively choose the bases in such a way that the set of basis functions for ℓmi ∞ is contained in the set of basis functions for ℓmi+1 ∞ . The union of these sets of basis func- tions gives an infinite sequence (fi) ∈ A(S)+. Choose a dual sequence (xi) ∈ S; thus, fi(xj) is 1 if i = j, and is zero otherwise. Passing to duals we also obtain a projective limit decomposition of S in the form Sm1 ✛ Sm2 ✛ Sm3 ✛ · · · where the Smi are mi-dimensional simplexes and the maps are affine surjective maps. Consider the (bi)affine function Fn(x, y) := Pmn j,k ajkfj(x)fk(y) where ajk is the given function F (x, y) evaluated at (xj, xk). The Fn are an increasing sequence of positive functions that converges pointwise, on a compact space, and Dini's theorem implies uniform convergence. De- noting the (bi)affine function obtained in the limit by G, we note that G is equal to the given function F at points of the form (xi, xj). Be- cause both G and F are already known to be continuous and positive, to prove equality it suffices to show that the (xi) are (affine) linearly dense in a suitable sense. We recall that the affine span of a subset of an affine space is the set of all finite linear combinations of the points of THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 11 the subset, with coefficients adding up to 1. When we apply a projec- tion map pn : S → Smn, a dimension argument using the basis property shows that the simplex generated by (pn(xi)) has maximal dimension; or in other words, the affine span of the (pn(xi)) includes all of Smn. It follows from the definition of the topology on a projective limit that the affine span of (xi) is dense in S. It then follows that G equals F on a dense set, and thus everywhere. We thus obtain finite sums, Pi,j aijfi(x)fj(y), that approximate the given function F as required, where the aij are positive scalars, however, the non-negative functions fk may have zeros. We can arrange that there are no zeros by first approximating F − ε for sufficiently small ε > 0, and then adding small positive multiples of 1 to the resulting functions fk. The case S1 6= S2 is a straightforward generalization. (cid:3) The next theorem appears to be of interest in the setting of nonuni- tal stably projectionless C-algebras. In the nonunital case, one has to find a compact simplex as base for the tracial cone: see [28, Prop. 3.4]. We will subsequently build upon the proof of the next theorem to accomodate projection-class elements. We denote the algebraic tensor product of abelian semigroups [15] by ⊗alg. Under the hypotheses of the theorem, the Cuntz semigroups are sets of semi-continuous func- tions, and dense range means that the image is norm-dense among the continuous functions (which are dense in a pointwise sense among the lower-semi-continuous functions in question, by Proposition 2.3). See Section 4 for more information on the topology that is implied by the density condition that we are using. The next lemma extends, so to speak, the familiar properties of tensor products of continuous functions to the case of lower semi- continuous functions, by combining Proposition 2.3 and Lemma 3.1. Lemma 3.2. Assume A and B are separable C-algebras that are such that Cu(A) ∼= Lsc(T (A), (0, ∞]), Cu(B) ∼= Lsc(T (B), (0, ∞]), and Cu(A ⊗ B) ∼= Lsc(T (A) × T (B), (0, ∞]). Then the tensor product map t from Cu(A) ⊗alg Cu(B) to Cu(A ⊗ B) has dense range, in the 12 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y sense that every element of Cu(A⊗B) can be obtained as the pointwise supremum of some sequence of elements of the range of the map t. Proof. When A and B are non-unital, we use [28, Prop. 3.4] to find a compact simplex, denoted T (A) or respectively T (B), as a base for the tracial cone. Define Φ : Lsc(T (A), (0, ∞]) ⊗alg Lsc(T (B), (0, ∞]) → Lsc(T (A) × T (B), (0, ∞]), a semigroup morphism that acts on the el- ements (f ⊗ g) of Lsc(T (A), (0, ∞]) ⊗alg Lsc(T (B), (0, ∞]) as Φ(f ⊗ g)(x, y) = f (x)g(y), where the right hand side is the pointwise product, f (x)g(y), inter- preted as a function on a Cartesian product. We note that the semi- group tensor product on the left is defined [15, page 270] by forming a free abelian semigroup and imposing all possible bilinearity relations: (a + a′) ⊗alg b = a ⊗alg b + a′ ⊗alg b and a ⊗alg b′ + a ⊗alg b = a ⊗alg (b + b′). Evidently these relations do not alter the function obtained on the right hand side. Note that the product of non-negative lower semicontinous functions is a non-negative lower semicontinuous function. The given representa- tions of the Cuntz semigroup(s) evidently intertwine Φ with the given tensor product map t. We now consider the range of the map Φ. Let G be a continuous element of Cu(A ⊗ B) ∼= Lsc(T (A) × T (B), (0, ∞]). We recall that for a unital C-algebra, the tracial states, T (A), are a Choquet simplex (see [13]). In the nonunital case, T (A) is understood to be a compact and simplicial base of the tracial cone (as already mentioned above, see also [28]). Since A is separable, the simplex is metrizable. By Lemma 3.1 we have an element Fn of the form PN i fi(x)gi(y) that is bounded above by G and approximates the continuous function G within 1 n . Since the functions fi and gi are positive, the element Fn is in the image of the map Φ. The more general case of G being lower semicontinuous follows by taking suprema, using Proposition 2.3. (cid:3) 3.1. Beyond the purely positive case. The next step is to assume our algebra A has non-trivial projections, so that V (A), the projection THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 13 monoid, is non-trivial. We will make use of the fact that the Cuntz semigroup is the disjoint union of projection elements from the Cuntz semigroup, denoted V (A), and purely positive elements, denoted Lsc(T (A), (0, ∞]). Ara, Perera and Toms, [3, Proposition 2.23], prove that that for algebras of stable rank one, the Cuntz class of a positive element is given by a projection if and only if {0} is not in the spectrum or if it is an isolated point of the spectrum. If a ∈ M∞(A)+, b ∈ M∞(B)+ are positive elements then it follows that a ⊗ b is a positive element in M∞(A ⊗ B)+. This induces a bilinear morphism from Cu(A) × Cu(B) to Cu(A ⊗ B) which in turn induces a natural Cuntz semigroup map t : Cu(A) ⊗alg Cu(B) → Cu(A ⊗ B) t([a] ⊗ [b]) = [a ⊗ b]. In the case that A = B and A is a simple, separable, unital, stably finite, nuclear, Z-stable C-algebra, satisfies the UCT, with stable rank one and vanishing K1 group, we now show that the map t has dense range. We begin with a technical lemma on enveloping groups: Lemma 3.3. Let V be a semigroup, and let G denote the formation of the enveloping group. We have G(V ⊗ V ) = G(V ) ⊗Z G(V ), where ⊗ denotes the tensor product of semigroups, and ⊗Z denotes the tensor product of abelian groups. Proof. Let us denote by C(S) the greatest homomorphic image of a commutative semigroup S. This is just the quotient of S by an equiv- alence relation: x ∼ y if x + b = y + b for some element b. Follow- ing [16], we define a tensor product of commutative semigroups by S1 ⊗C S2 := C(S1 ⊗ S2). It is shown on page 201 of [16] that this is an associative and commutative tensor product of commutative semi- groups. By Proposition 3 of [16], we have that the enveloping group of 14 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y V is given by G(V ) = V ⊗C Z, where Z denotes the group of integers. But then, by the associative and commutative properties of the tensor product, G(V ) ⊗C G(V ) = (V ⊗C V ) ⊗C Z. It then follows that G(V ) ⊗C G(V ) = C(V ⊗ V ) ⊗C Z = G(C(V ⊗ V )). The definition of an enveloping group is such that the enveloping group of C(V ⊗ V ) is the same as the enveloping group of V ⊗ V. We have therefore shown that G(V ⊗ V ) = G(V ) ⊗C G(V ). However, by Proposition 1.4 of [15], the semigroup tensor product of abelian groups coincides with the usual tensor product of abelian groups, and since groups are already cancellative, there is then no difference between the tensor products ⊗, ⊗Z and ⊗C when the factors are both abelian groups. It follows that G(V ⊗ V ) = G(V ) ⊗Z G(V ), as claimed. (cid:3) Lemma 3.4. If a C-algebra A is simple, separable, unital, nuclear, Z-stable, and finitely generated K0(A), K1(A) = {0}, and satisfies the UCT, then the natural map Cu(A) ⊗alg Cu(A) → Cu(A ⊗ A), given by ([a] ⊗ [b]) 7→ [a ⊗ b] is an isomorphism from V (A) ⊗alg V (A) to V (A ⊗ A), i.e. V (A) ⊗alg V (A) ∼= V (A ⊗ A). Proof. We remark that the hypotheses on A imply stable rank 1 by [23]. Since the tensor product, A ⊗ A, also is simple, separable, unital, nuclear, and Z-stable, it follows that the tensor product, A ⊗ A, also has stable rank one. Our algebra A is assumed to satisfy the UCT, THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 15 so then A will satisfy the Kunneth formula for tensor products in K- theory [27], see also the partial counterexample due to Elliott in [25] which means that we must assume finitely generated K0-group: 0 → K0(A) ⊗ K0(A) ⊕ K1(A) ⊗ K1(A) → K0(A ⊗ A) → T or(K0(A), K1(A)) ⊕ T or(K1(A), K0(A)) → 0. It follows from the above exact sequence that K0(A) ⊗ K0(A) → K0(A ⊗ A) is an injective map. Since K1(A) = {0} it follows that we have an order-preserving isomorphism t : K0(A) ⊗ K0(A) → K0(A ⊗ A). We will show that when restricted to the positive cones, this isomor- phism becomes equivalent to the given map. Since A has stable rank 1, there is an injective map i : V (A) → K0(A) and V (A) has the cancella- tion property. Taking algebraic tensor products of semigroups, we con- sider V (A) ⊗alg V (A). Taking the (semigroup) tensor product of maps, we obtain a map i ⊗alg i : V (A) ⊗alg V (A) → K0(A) ⊗alg K0(A) where K0(A) ⊗alg K0(A) is a semigroup tensor product of abelian groups. Moreover, the semigroup tensor product of abelian groups coincides with the usual tensor product of abelian groups (by Proposition 1.4 in [15] and the remarks after that Proposition). We note that the map i ⊗alg i is an injective map (using Lemma 3.3). Since A is stably finite and the positive cone of a tensor product of finitely generated ordered abelian groups is the tensor product of the positive cones of the ordered abelian groups, it follows that the range of the map i ⊗alg i is exactly the positive cone of K0(A) ⊗ K0(A). Composing with the above injective map t, we obtain an injective map from V (A)⊗alg V (A) to K0(A⊗A), which takes an element p⊗alg q to p ⊗ q. Since t is an order isomorphism, the range of t ◦ (i ⊗alg i) is exactly the positive cone of K0(A⊗A). We now observe that the map we have obtained is in fact equal to the given map, because, as A ⊗ A has 16 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y stable rank 1 and is stably finite, it follows that V (A ⊗ A) is embedded in K0(A ⊗ A) as the positive cone. It then follows that t ◦ (i ⊗alg i), which acts on elements by taking p ⊗alg q to p ⊗ q, is in fact an injection of V (A) ⊗alg V (A) onto V (A ⊗ A). Evidently, this map coincides with the given map. (cid:3) Remark 3.5. The argument of the above lemma can be adapted to pro- vide a class of counter-examples to the possible surjectivity, even after closure, of the tensor product map t : Cu(A) ⊗alg Cu(B) → Cu(A⊗B). If an algebra is in the UCT class, and the K-theory groups are such that the last term in the Kunneth sequence of [27] does not vanish, then we see that the first map in the short exact sequence 0 → K0(A) ⊗ K0(A) ⊕ K1(A) ⊗ K1(A) → K0(A ⊗ A) → T or(K0(A), K1(A)) ⊕ T or(K1(A), K0(A)) → 0 will not be surjective. But then, in particular, the tensor product map from K0(A) ⊗ K0(A) to K0(A ⊗ A) will not be surjective. Hence the tensor product map at the level of projection semigroups will not be surjective either, and this is an obstacle to the surjectivity of the tensor product map at the level of Cuntz semigroups (using the result that the tensor product of elements that are equivalent to a projection is an element that is equivalent to a projection). In the next theorem, the range is dense in a sense discussed in Section 4. Theorem 3.6. If a C-algebra A is simple, separable, unital, sta- bly finite, nuclear, Z-stable, satisfies the UCT, with finitely generated K0(A), and has K1(A) = {0}, then the natural semigroup map t : Cu(A) ⊗alg Cu(A) → Cu(A ⊗ A) given by t([a]⊗[b]) = [a⊗b] has dense range (under pointwise suprema). Proof. Under these hypotheses, any element x in Cu(A ⊗ A) is either in V (A ⊗ A) or in Lsc(T (A ⊗ A), (0, ∞]), by, for example, Theorem THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 17 In the case that x is in V (A ⊗ A), Lemma 3.4 shows surjec- 2.1. tivity of V (A) ⊗alg V (A) onto V (A ⊗ A). The traces are a metriz- able simplex [13, 7], so in the case that x is a continuous function in Lsc(T (A ⊗ A), (0, ∞]), by Theorem 3.2, there is an element of Lsc(T (A), (0, ∞])⊗algLsc(T (A), (0, ∞]) that (uniformly) approximates x. Since any lower semicontinuous function is a supremum of continu- ous functions, this completes the proof. (cid:3) Remark 3.7. We mention that the above result applies in the finite- dimensional case, in other words, the case of matrix algebras, and in this case, dense range is equivalent to surjectivity. 4. Deducing surjectivity results To obtain surjectivity results in greater generality, some form of com- pletion operation must be introduced, and we do this next. The natural topology on the projection-class elements of the Cuntz semigroup is the discrete topology. Thus, we were able to handle projection-class elements without explicit reference to a topology. How- ever, when considering the tensor product of purely positive elements, it appears inevitable that we must consider topologies, which means that we must consider completions of the algebraic tensor product with respect to a specified topology. Our approach is based on a very general construction of the so-called inductive (or injective) tensor product of topological vector spaces, due to Grothendieck, see [17, d´efinition 3]. Thus, we form the algebraic ten- sor product of abelian semigroups, see [15], and then view elements of Cu(A)⊗alg Cu(B) as functions on D(A)×D(B), where D(A) and D(B) denote the dimension functions on the Cuntz semigroup(s). The induc- tive topology is the (initial) topology induced by this embedding. We then take the topological completion of Cu(A) ⊗alg Cu(B) with respect to this topology (see [19, ex. 6.L, and pp. 195-6] for information on completions). As previously mentioned, we only need to perform the above construction on the set of those elements whose image under the 18 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y tensor product map is purely positive: this set will be further char- acterized later. For brevity we refer to these elements as the purely positive elements. We have that for the purely positive elements of Cu(A) ⊗alg Cu(B), an increasing sequence xn converges (pointwise) to an element of the completion if and only if (d1, d2)(xn) converges for all d1 ∈ D(A) and d2 ∈ D(B). The limit of the sequence exists in the completion, and we define the inductive tensor product of Cuntz semi- groups, denoted Cu(A) ⊗ Cu(B), to be Cu(A) ⊗alg Cu(B) augumented by the set of all such limits. We also have the minimal embedding tensor product, given as fol- lows. Consider the natural tensor product map t : Cu(A)⊗alg Cu(B) → Cu(A ⊗min B). This map induces a uniformizable topology on its do- main (sometimes called the inital topology). Taking, then, the comple- tion of the domain with respect to this topology, and proceeding as in the previous paragraph, we obtain the minimal embedding tensor prod- uct, Cu(A) ⊗min Cu(B). If, in the above, we replace Cu(A ⊗min B) by Cu(A⊗max B), then we obtain the maximal embedding tensor product, denoted Cu(A) ⊗max Cu(B). Proposition 4.1. If the dimension functions are determined by traces, and if the normalized traces T (A ⊗min B) are isomorphic to T (A) × T (B), then the inductive tensor product Cu(A) ⊗ Cu(B) coincides with the minimal embedding tensor product Cu(A) ⊗min Cu(B). Proof. The inductive tensor product is given by adding to the algebraic tensor product the limits of increasing sequences (xn) that are such that (d1 × d2)(xn) converges for each d1 ∈ D(A) and d2 ∈ D(B). (We need only consider purely positive elements.) On the other hand, the minimal embedding tensor product is given by adding to the algebraic tensor product the limits of increasing sequences (xn) that are such that d(t(xn)) converges for each d ∈ D(A ⊗min B), again for purely positive elements. The hypothesis implies that D(A ⊗min B) is isomorphic to D(A) × D(B), but then the two constructions considered above will coincide. (cid:3) THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 19 The results of the previous sections are valid within the class of algebra where the purely positive elements of Cu(A⊗A) are isomorphic to the set of lower semicontinuous, affine, strictly positive functions on the tracial state space of the unital C-algebra A. In general, of course, since we regard Cu(A) as being topologically a disjoint union of two sets, namely the set of elements having the same class as a projection, and the set of purely positive elements, the tensor product of Cuntz semigroups will a priori be a union of four sets. As explained above, the topological issues only arise when considering the component Cu(A)pure ⊗ Cu(B)pure. In general, when considering the tensor product map t : Cu(A) ⊗ Cu(A) → Cu(A ⊗ A) the following three cases appear: Case 1: projection elements tensored with projection elements, Case 2: purely positive tensored with purely positive elements, and Case 3: projection elements tensored with purely positive elements. We now summarize our results on algebraic tensor products, for each of these three cases, in the setting of completed tensor products. Theorem 4.2. Suppose that A is a C-algebra which is unital, sim- ple, separable, stably finite, nuclear, Z-stable, satisfies the UCT, with finitely generated K0(A), has K1(A) = {0} and satisfies the Blackadar -- Handelman conjecture. Then each of the maps t : Cu(A)pure ⊗ Cu(A)pure → Cu(A ⊗ A) t : Cu(A)pure ⊗ Cu(A)proj → Cu(A ⊗ A) t : Cu(A)proj ⊗ Cu(A)pure → Cu(A ⊗ A) t : Cu(A)proj ⊗ Cu(A)proj → Cu(A ⊗ A) is injective. The first three of these maps have range contained in the purely positive elements of Cu(A ⊗ A), the last map has range contained in the projection-class elements of Cu(A ⊗ A). The domains are injective tensor products of Cuntz semigroups. The first and last of the above maps are surjective onto their range. 20 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y Proof. We first show that the behaviour with respect to purely positive and projection-class elements is as claimed. A positive element has the class of a projection if and only if the spectrum of the element has a spectral gap at zero, and the spectrum of a ⊗ b is given by the set of all pairwise products {λµ λ ∈ Sp(a), µ ∈ Sp(b)}. It can thus be seen that a ⊗ b has the class of a projection if and only if both a and b have the class of a projection. We now consider surjectivity. Our Lemma 3.4 gives surjectivity onto the projection elements of Cu(A ⊗ A), and Lemma 3.1 provides an ap- proximation (from below) of purely positive elements of Cu(A)⊗Cu(A) that are continuous when viewed as functions on D(A) × D(A). More precisely, the image of Cu(A) ⊗alg Cu(A) is dense, with respect to the supremum norm over the (compact) space D(A) × D(A), within the continuous positive affine functions on D(A) × D(A). It follows that after taking completions, the range of the tensor product map t : Cu(A) ⊗ Cu(A) → Cu(A ⊗ A) contains at least the purely posi- tive elements corresponding to continuous functions. Since any lower semicontinuous function is a supremum of continuous functions, this implies surjectivity. We now consider the injectivity of these maps. Suppose that x, x′ ∈ Cu(A)⊗Cu(A) are such that t(x) = t(x′). Consider first the case where t(x) and t(x′) belong to the purely positive part of Cu(A) ⊗ Cu(A). We have that d ◦ t(x) = d ◦ t(x′) for all dimension functions d on A ⊗ A. Choose a dimension function d on A ⊗ A that comes from a tensor product τ1 ⊗ τ2 of traces on A. Thus, if t is the map taking [a] ⊗ [b] to [a ⊗ b], and d(t(x)) = d(t(x′)), we deduce that (d1 ⊗ d2)(x) = (d1 ⊗ d2)(x′) where d1 and d2 are the dimension states on Cu(A) that come from τ1 and τ2 respectively. Since A satisfies the Blackadar- Handelman conjecture, this means that d1 and d2 can be chosen to be equal to any two dimension states of A. But this means that x and x′ are equal in the tensor product Cu(A) ⊗ Cu(A). The two mixed cases are similar. The case of projections follows from the fact that the THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 21 natural map from V (A) ⊗ V (A) to V (A ⊗ A) is an injection, by Lemma 3.4. (cid:3) Combining the above with Theorem 3.6, we have the corollary: Corollary 4.3. If an unital C-algebra A is simple, separable, sta- bly finite, nuclear, Z-stable, satisfies the UCT, with finitely generated K0(A), has K1(A) = {0} then the natural tensor product map t : Cu(A) ⊗ Cu(A) → Cu(A ⊗ A) given by t([a], [b]) = [a ⊗ b] is surjective, and becomes an isomorphism when restricted to Cu(A)pure ⊗ Cu(A)pure ⊔ Cu(A)proj ⊗ Cu(A)proj. Thus the mixed elements of the tensor product are evidently an obstacle to injectivity of the unrestricted tensor product map. This is because, by Theorem 3.6, every purely positive element of Cu(A ⊗ A) is the image of some element of Cu(A)pure ⊗ Cu(A)pure, but on the other hand, as shown in the proof of Theorem 4.2, the image of a purely positive element tensored with a projection-class element is a purely positive element. It thus follows that in this case the tensor product map is not injective. Therefore, it is quite rare for the tensor product map to be an isomorphism. We note the following further corollary: Corollary 4.4. If an unital C-algebra A is simple, stably projection- less, stably finite, nuclear, Z-stable, satisfies the UCT, with finitely generated K0(A), and has K1(A) = {0} then the natural tensor prod- uct map t : Cu(A) ⊗ Cu(A) → Cu(A ⊗ A) given by t([a] ⊗ [b]) = [a ⊗ b] is an isomorphism. A class of algebras to which our results can be applied is the class of simple AT-algebras with trivial K1-group (or, equivalently, AI alge- bras). 22 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y Corollary 4.5. If A is a simple AI-algebra with finitely generated K- theory, then the natural tensor product map t : Cu(A) ⊗ Cu(A) → Cu(A ⊗ A) given by t([a], [b]) = [a ⊗ b] is surjective, and becomes an isomorphism when restricted to Cu(A)pure ⊗ Cu(A)pure ⊔ Cu(A)proj ⊗ Cu(A)proj. Dini's theorem shows that over a compact base space, pointwise con- vergence of increasing sequences of continuous functions implies uni- form convergence. It follows that in the simple special case of C- algebras having Cuntz semigroups isomorphic to the positive lower semicontinuous functions on a compact space, the rather coarse topol- ogy provided by the definition of the injective tensor product is equiv- alent to a much finer topology. Thus, for the class of algebras under consideration, we expect that most of the ways to construct the tensor product will coincide. The tensor product of [2] involves both a joint continuity condition and a separate continuity condition. (See page 7 of [2] for the precise statement of both conditions.) In the closely related setting of tensor products of topological vector spaces, a tensor product involving both a joint continuity condition and a separate continuity condition would lie strictly between the inductive tensor product and the projective tensor product. We have already shown (see Proposition 4.1) that in our setting, the equivalent of the inductive tensor prod- uct and the projective tensor product happen to coincide. At present, however, we do not have a good topological description of the tensor product of [2]. Remark 4.6. The tensor product of two abelian semigroups is con- structed by forming a free abelian semigroup and passing to the quo- tient in which (a + a′) ⊗ b = (a ⊗ b) + (a′ ⊗ b) and (a ⊗ b′) + (a ⊗ b) = a ⊗ (b + b′). One can equivalently define the tensor product as an abstract abelian semigroup on which additive maps are given by bi- additive maps on the original semigroups. However, Grillet in [15], THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 23 Theorem 2.1, shows that, as for any universal property definition, the tensor product of abelian semigroups is unique up to semigroup iso- morphism. We recall that the stable rank one property was used in a funda- mental way in our proofs, to, for example, simplify the compact con- tainment property. (We further used the stable rank one property to obtain cancellation in the projection semigroup, and to ensure that projection class and purely positive elements can be distinguished by a spectral criterion.) Investigating the exact relationship between the several possible tensor products suggests venturing beyond the stable rank one case; in which case additional issues arise that are beyond the scope of this paper, but seem to be a reasonable question for future investigation. References [1] R. Antoine, J. Bosa and F. Perera, Completions of monoids with applications to the Cuntz semigroup. Internat. J. Math. 22 (2011), no. 6, pp. 837-861. [2] R. Antoine, F. Perera, and H. Thiel, Tensor products and regularity properties of Cuntz semigroups, preprint (arXiv: 1410.0483v1 [math.OA], 2 Oct 2014, 183 pages.) [3] P. Ara, F. Perera, and A. S. Toms, K-theory for operator algebras. Classifica- tion of C-algebras, Contemporary Mathematics, 534 (2011) [4] R. J. Archbold, On the centre of a tensor product of C-algebras. J. London Math. Soc. (2) 10 (1975), 257 -- 262 [5] B. Blackadar, and D. Handelman, Dimension functions and traces on C- algebras, J. Funct. Anal. 45 (1982), 297 -- 340. [6] ´E. Blanchard, E. Kirchberg, Non-simple purely infinite C-algebras: the Haus- dorff case, J. Funct. Anal. 207 (2004), 461 -- 513. [7] L. Brown, Stable isomorphism of hereditary subalgebras of C-algebras, Pacific J. Math 71 (1977), 335 -- 384 [8] L. G. Brown, R. G. Douglas, and P. A. Fillmore, Unitary equivalence modulo the compact operators and extensions of C-algebras. Proc. Conf. on Operator Theory, Lecture Notes in Math, 345, pp. 58 -- 128, Springer Verlag, Berlin, 1973. [9] N. P. Brown, F. Perera, and A. S. Toms, The Cuntz semigroup, the Elliott conjecture, and dimension functions on C-algebras, J. Reine Angew. Math. 621 (2008), 193 -- 211. 24 CRISTIAN IVANESCU AND DAN KU CEROVSK ´Y [10] K. T. Coward, G. A. Elliott, and C. Ivanescu, The Cuntz semigroup as an invariant for C-algebras, J. Reine Angew. Math. 621 (2008), 161 -- 193. [11] A. Ciuperca, L. Roberts, and L. Santiago, Cuntz semigroup of ideals and quo- tients and a generalized Kasparov stabilization theorem, J. Operator Theory Vol 64 (2010), 155 -- 169. [12] M. Dadarlat, G. Nagy, A. N´emethi, C. Pasnicu, Reduction of topological stable rank in inductive limits of C-algebras, Pacific J. Math. 153 (1992), 267 -- 276. [13] E. G. Effros, F. Hahn, Locally compact transformation groups and C-algebras, Memoirs of the Amer. Math. Soc., No. 75, Amer. Math. Soc., Providence, R.I. (1967). [14] G. A. Elliott, L. Robert, and L. Santiago, The cone of lower semicontinuous traces on a C-algebra, Amer. J. Math. 133 (2011), 969 -- 1005. [15] P. A. I. Grillet, The tensor product of semigroups, Trans. Amer. Math. Soc., 138 (1968), 267 -- 281. [16] P. A. I. Grillet, The tensor product of cancellative semigroups, J. Natur. Sci. and Math. 10 (1970), 199-208. [17] A. Grothendieck, Produits tensoriels topologiques et espaces nucl´eaires, Mem- oirs Amer. Math. Soc. 16 (1955), 1 -- 191. [18] U. Haagerup, Quasitraces on exact C∗-algebras are traces, C. R. Math. Acad. Sci. Soc. R. Can., 36 (2014), 67 -- 92. [19] J. L. Kelley, General Topology, Van Nostrand (1955), reprinted by Springer [20] S-H. Kye and H. Osaka, Classification of bi-qutrit positive partial transpose entangled edge states by their ranks, J. Math. Phys. 53 (2012), 052201. [21] A. J. Lazar, J. Lindenstrauss, Banach spaces whose duals are L1 spaces and their representing matrices, Acta Math. 126 (1971), 165 -- 193. [22] H. Osaka, Non-commutative dimension for C-algebras, Interdiscip. Inform. Sci. 9 (2003), no. 2, 209-220. [23] M. Rørdam, The stable and the real rank of Z-absorbing C-algebras, Interna- tional J. Math., 15, No. 10 (2004), 1065 -- 1084 [24] M. Rørdam, W. Winter, The Jiang-Su algebra revisited, J. Reine Angew. Math. 642 (2010), 129 -- 155. [25] J. Rosenberg, C. Schochet, The Kunneth theorem and the universal coefficient theorem for Kasparov's generalized K-functor, Duke Math. J. 55 (1987), 431 -- 474 [26] R. Ryan, Introduction to tensor products of Banach spaces, Springer-Verlag, London, 2010. [27] C. Schochet, Topological methods for C-algebras, II: Geometry resolutions and the Kunneth formula, Pacific J. Math. 98 (1982), 443 -- 458. THE TENSOR PRODUCT OF CUNTZ SEMIGROUPS 25 [28] A. Tikuisis and A. S. Toms, On the structure of the Cuntz semigroup in (pos- sibly) nonunital C*-algebras, Canad. Math. Bull. 58 (2015), 402 -- 414. [29] A. S. Toms, Comparison theory and smooth minimal C-∗-dynamics, Comm. Math. Phys. 289 (2009), no.2, 401 -- 433. (1) Department of Mathematics and Statistics, Grant MacEwan Uni- versity, Edmonton, Alberta, T5J 4S2, e-mails: [email protected], [email protected] (2) Department of Mathematics and Statistics, University of New Brunswick, Fredericton, New Brunswick, E3B 5A3, e-mail: [email protected]
1512.00276	4	1512	2018-10-31T18:57:54	K-theory of cluster C*-algebras	[ "math.OA", "math.GT", "math.RT" ]	It is proved that the K_0-group of a cluster C-algebra is isomorphic to the corresponding cluster algebra. As a corollary, one gets a shorter proof of the positivity conjecture for cluster algebras. As an example, we consider a cluster C-algebra A(1,1) coming from triangulation of an annulus with one marked point on each boundary component.	math.OA	math	K-theory of cluster C∗-algebras Igor Nikolaev Abstract It is proved that the K0-group of a cluster C ∗-algebra is isomorphic to the corresponding cluster algebra. As a corollary, one gets a shorter proof of the positivity conjecture for cluster algebras. As an example, we consider a cluster C ∗-algebra A(1, 1) coming from triangulation of an annulus with one marked point on each boundary component. Key words and phrases: cluster C ∗-algebras, K-theory MSC: 13F60 (cluster algebras); 46L85 (noncommutative topology); 1 Introduction Cluster algebras are a class of commutative rings introduced by [Fomin & Zelevinsky 2002] [4] having deep roots in hyperbolic geometry and Te- ichmuller theory [Williams 2014] [15]. Namely, the cluster algebra A(x, B) of rank n is a subring of the field of rational functions in n variables de- pending on a cluster of variables x = (x1, . . . , xn) and a skew-symmetric matrix B = (bij) ∈ Mn(Z); the pair (x, B) is called a seed. A new cluster x′ = (x1, . . . , x′ ij) is obtained from (x, B) by the exchange relations: k, . . . , xn) and a new skew-symmetric matrix B′ = (b′ xkx′ k = xmax(−bik,0) , i if i = k or j = k otherwise. (1) n n i + xmax(bik ,0) Yi=1 Yi=1 bij + bikbkj+bikbkj 2 b′ ij = (−bij The seed (x′, B′) is said to be a mutation of (x, B) in direction k, where 1 ≤ k ≤ n; the algebra A(x, B) is generated by cluster variables {xi}∞ i=1 1 obtained from the initial seed (x, B) by the iteration of mutations in all possible directions k. The Laurent phenomenon proved by [Fomin & Zelevinsky 2002] [4] says that A(x, B) ⊂ Z[x±1], where Z[x±1] is the ring of the Laurent polynomials in variables x = (x1, . . . , xn) depending on an initial seed (x, B); in other words, each generator xi of algebra A(x, B) can be written as a Laurent polynomial in n variables with integer coefficients. The famous Positivity Conjecture says that coefficients of the Laurent polynomials corresponding to variables xi are always non-negative integers, see [Fomin & Zelevinsky 2002] [4]. A general form of the Positivity Conjecture was proved by [Lee & Schiffler 2015] [7] using a clever combinatorial formula for the variables xi. Cluster C ∗-algebras A(x, B) are a class of non-commutative rings intro- duced in [11]. The A(x, B) is an AF -algebra given by the Bratteli diagram [Bratteli 1972] [1]; such a diagram is obtained from a mutation tree of the initial seed (x, B) modulo an equivalence relation between the seeds lying at the same level, see Section 2.2. (We refer the reader to Figures 1 and 2 for an immediate example of such algebras.) The aim of our note is the K-theory of the AF -algebra A(x, B). Namely, the ordered abelian group is a pair (G, G+) consisting of an abelian group G and a semigroup G+ ⊂ G of positive elements of G; the order ≤ on G is defined by the positive cone G+, i.e. a ≤ b if and only if b − a ∈ G+. An order-isomorphism ∼= between (G, G+) and (H, H +) is an isomorphism ϕ : G → H, such that ϕ(G+) = H +. Denote by K0(A(x, B)) the K0-group of the AF -algebra A(x, B) and by K + 0 (A(x, B)) ⊂ K0(A(x, B)) its Grothendieck semigroup [Effros 1981, Chapter 8] [2]. The pair (K0(A(x, B)), K + 0 (A(x, B))) is an invariant of Morita equivalence of the AF -algebra A(x, B) [Elliott 1976] [3]. In view of the Laurent phenomenon, let Aadd(x, B) be an additive group add(x, B) be a semigroup inside the of the cluster algebra A(x, B); Aadd(x, B) consisting of the Laurent polynomials with positive coefficients. The pair (Aadd(x, B),A+ add(x, B)) is an abelian group with order. The order a > b is defined between two elements a, b ∈ Aadd(x, B) if and only if a− b ∈ A+ add(x, B) . Our main result can be formulated as follows. 0 (A(x, B)) ∼= (Aadd(x, B), A+ Theorem 1 (K0(A(x, B)), K + let A+ add(x, B)). An application of theorem 1 is as follows. Recall that the dimension group is a triple (G, G+, Γ) consisting of an abelian group G, a semigroup of positive elements G+ ⊂ G and a scale Γ ⊆ G+, i.e. a generating, hereditary and directed subset of G+ [Effros 1981, Chapter 7] [2]. For instance, the Γ ∼= G+ 2 is a scale called stable; thus the pair (G, G+) is a special case of the dimension group. An order-isomorphism ∼= between dimension groups (G, G+, Γ) and (H, H +, Γ′) is an isomorphism ϕ : G → H, such that ϕ(G+) = H + and ϕ(Γ) = Γ′. Denote by Γ ⊂ K + 0 (A(x, B)) the set of the Murray-von Neumann equivalence classes of projections in the algebra A(x, B). It is known, that the triple (K0(A(x, B)), K + 0 (A(x, B)), Γ) is an invariant of the isomorphism class of the AF -algebra A(x, B) [Elliott 1976] [3]. It is not hard to observe, that the set X = {xi}∞ i=1 of all variables xi in the cluster algebra Aadd(x, B) is a scale, since it is a generating, hereditary and directed subset of A+ add(x, B). Notice that choosing a different initial seed (x, B) for the Laurent expansion add(x, B), X) ∼= of variables xi yields a new scale X ′, such that (Aadd(x, B),A+ (Aadd(x, B),A+ add(x, B) for any dimension group; therefore theorem 1 implies a new proof of the Positivity Conjecture for the cluster algebras. add(x, B), X ′). But X ⊆ A+ Corollary 1 The coefficients of the Laurent polynomials corresponding to the cluster variables xi are non-negative integers. The article is organized as follows. The preliminary facts are introduced in Section 2. Theorem 1 and corollary 1 are proved in Section 3. In Section 4 we consider an example of the cluster C ∗-algebra A(1, 1) coming from trian- gulation of an annulus with one marked point on each boundary component. 2 Preliminaries This section is a brief review of the AF -algebras, cluster C ∗-algebras and Mundici dimension groups. For a general review of C ∗-algebras we refer the reader to [Murphy 1990] [10]. The AF -algebras were introduced in [Bratteli 1972] [1]. The general K-theory of C ∗-algebras is covered in [Rørdam, Larsen & Laustsen 2000] [13] and K-theory of the AF -algebras in [Effros 1981] [2]. Cluster C ∗-algebras were the subject of [11]. Mundici dimension groups were introduced by [Mundici 1988] [8]. 2.1 AF -algebras and dimension groups A C ∗-algebra is an algebra A over C with a norm a 7→ a and an involution a 7→ a∗ such that it is complete with respect to the norm and ab ≤ a b and a∗a = a2 for all a, b ∈ A. Any commutative C ∗-algebra is 3 isomorphic to the algebra C0(X) of continuous complex-valued functions on some locally compact Hausdorff space X; otherwise, A can be thought of as a noncommutative topological space. An AF -algebra (Approximately Finite C ∗-algebra) is defined to be the norm closure of a dimension-increasing sequence of finite dimensional C ∗- algebras Mn, where Mn is the C ∗-algebra of the n × n matrices with entries in C. Here the index n = (n1, . . . , nk) represents the semi-simple matrix algebra Mn = Mn1 ⊕ . . . ⊕ Mnk . The ascending sequence mentioned above can be written as (2) M1 ϕ1−→ M2 ϕ2−→ . . . , where Mi are the finite dimensional C ∗-algebras and ϕi the homomorphisms between such algebras. The homomorphisms ϕi can be arranged into a graph be the as follows. Let Mi = Mi1 ⊕ . . . ⊕ Mik and Mi′ = Mi′ semi-simple C ∗-algebras and ϕi : Mi → Mi′ the homomorphism. (To keep it simple, one can assume that i′ = i + 1.) One has two sets of vertices joined by brs edges whenever the summand Mi′ Vi1, . . . , Vik and Vi′ contains brs copies of the summand Mis under the embedding ϕi. As i varies, one obtains an infinite graph called the Bratteli diagram of the AF -algebra. The matrix B = (brs) is known as a partial multiplicity matrix; an infinite sequence of Bi defines a unique AF -algebra. 1 ⊕ . . . ⊕ Mi′ k 1, . . . , Vi′ k r For a unital C ∗-algebra A, let V (A) be the union (over n) of projections in the n× n matrix C ∗-algebra with entries in A; projections p, q ∈ V (A) are equivalent if there exists a partial isometry u such that p = u∗u and q = uu∗. The equivalence class of projection p is denoted by [p]; the equivalence classes of orthogonal projections can be made to a semigroup by putting [p] + [q] = [p + q]. The Grothendieck completion of this semigroup to an abelian group is called the K0-group of the algebra A. The functor A → K0(A) maps the category of unital C ∗-algebras into the category of abelian groups, so that projections in the algebra A correspond to a positive cone K + 0 ⊂ K0(A) and the unit element 1 ∈ A corresponds to an order unit u ∈ K0(A). The ordered abelian group (K0, K + 0 , u) with an order unit is called a dimension group; an order-isomorphism class of the latter we denote by (G, G+). If A is an AF -algebra, then its dimension group (K0(A), K + 0 (A), u) is a complete isomorphism invariant of algebra A [Elliott 1976] [3]. The order- isomorphism class (K0(A), K + 0 (A)) is an invariant of Morita equivalence of the algebra A, i.e. an isomorphism class in the category of finitely generated projective modules over A. 4 The scale Γ is a subset of K + 0 (A) which is generating, hereditary and directed, i.e. (i) for each a ∈ K + 0 (A) there exist a1, . . . , ar ∈ Γ(A), such that a = a1 + . . . + ar; (ii) if 0 ≤ a ≤ b ∈ Γ, then a ∈ Γ; (iii) given a, b ∈ Γ there exists c ∈ Γ, such that a, b ≤ c. If u is an order unit, then the set Γ := {a ∈ K + 0 (A) 0 ≤ a ≤ u} is a scale; thus the dimension group of algebra A can be written in the form (K0(A), K + 0 (A), Γ). 2.2 Cluster C∗-algebras Let Tn be an oriented tree whose vertices correspond to the seeds (x, B) and outgoing edges correspond to mutations in direction 1 ≤ k ≤ n. Notice that the the tree Tn of a cluster algebra A(x, B) has a grading by levels, i.e. minimal distance from the root of Tn. We shall say that a pair of clusters x and x′ with exchange matrices B and B′ are ℓ-equivalent, if: (i) x and x′ lie at the same level; (ii) x and x′ coincide modulo a cyclic permutation of variables xi; (iii) B = B′. It is not hard to see that ℓ is an equivalence relation on the set of vertices of graph Tn. Definition 1 By a cluster C ∗-algebra A(x, B) one understands an AF -algebra given by the Bratteli diagram B(x, B) of the form: B(x, B) := Tn mod ℓ. (3) Remark 1 Notice that the graph B(x, B) is no longer a tree; the cycles of B(x, B) appear after gluing together vertices lying at the same level of the tree according to the relation ℓ. The B(x, B) is not a regular graph, since the valency of vertices can vary. However, the B(x, B) is always a Bratteli diagram, since it is obtained from a regular tree by an addition of extra edges and subsequent contraction of the respective edges. Notice also that the B(x, B) is a finite graph if and only if A(x, B) is a finite cluster algebra. Example 1 Let x = (x1, x2) and B = (cid:18) 0 −2 2 0(cid:19) . 5 (4) ✑ ❜ ◗ ◗s❄ ◗ ✑ ✑✰ ❜ ❜ ❜ ❜ ❜ ❜ ✠ ❏❏❫❄✠ ❅❘❄ ✠ ❅❘❄ ✠ ❅❘❄✠ ❅❘❄ ✠❅❘❄ ✠❅❘❄ ❜ ✑✑✰ ◗◗s❄ ✑✑✰ ◗◗s❄ ✑✑✰ ◗◗s❄ ❜ ❜ ❜ Figure 1: The Bratteli diagram of Markov's cluster C ∗-algebra. The cluster algebra A(x, B) is associated to an ideal triangulation of an annulus with one marked point on each boundary component, see [Fomin, Shapiro & Thurston 2008, Example 4.4] [5]. The exchange relations (1) in this case can be written as xi−1xi+1 = 1+x2 i and B′ = −B. It is easy to verify using definition 1, that the Bratteli diagram T2 mod ℓ of the corresponding cluster C ∗-algebra A(x, B) is given by Figure 3. We refer the reader to Section 4 for an extended discussion of the properties of such an algebra. Example 2 ([11]) Let x = (x1, x2, x3) and B =   0 2 −2 −2 0 2 0 2 −2 .   (5) The cluster algebra A(x, B) is called Markov's; it is associated to an ideal triangulation of the hyperbolic torus with a cusp, see e.g. [Williams 2014] [15]. The Bratteli diagram T3 mod ℓ of the cluster C ∗-algebra A(x, B) is shown in Figure 1. (The corresponding mutation tree T3 and the equivalence classes of relation ℓ are given in full detail in [11], Figure 4.) The algebra A(x, B) has a non-trivial primitive spectrum being isomorphic to an AF - algebra M1 introduced by [Mundici 1988] [8]; for a general theory we refer the reader to the monograph [Mundici 2011] [9]. 2.3 Mundici dimension groups A broad class of dimension groups has been introduced by [Mundici 1988] [8]. We shall use such groups as the main technical tool in proof of theorem 6 1. We refer the reader to [Mundici 1988] [8] and [Mundici 2011] [9] for a detailed account. A lattice-ordered (or ℓ-group) is a structure (G, +,−, 0,∨,∧) such that (G, +,−, 0) is an abelian group, (G,∨,∧) is a lattice, and x + (y ∨ z) = (x + y)∨ (x + z) for all x, y, z ∈ G. An order unit in a partially ordered group G is an element u ≥ 0 such that for each x ∈ G there is an integer n ≥ 0 with x < nu. A unital ℓ-group is an ℓ-group with distinguished order unit. The function f : [0, 1]n → R is called a McNaughton function over [0, 1]n iff f is continuous and there are a finite number of linear functions: α1 = b1 + a11x1 + . . . + a1nxn α2 = b2 + a21x1 + . . . + a2nxn ... αm = b1 + am1x1 + . . . + amnxn, (6)   where all aij and bi are integers, such that for every (x1, . . . , xn) ∈ [0, 1]n there is i ∈ {1, . . . , m} with f (x1, . . . , xn) = αi(x1, . . . , xn), see [Mundici 1988] [8] and [Mundici 2011] [9]. In other words, the McNaughton function is a piecewise linear function with integer coefficients. It is easy to see, that the set of all McNaughton functions over [0, 1] is an ℓ-group with the pointwise operations +,−,∨,∧ of R and with the constant function 1 as the distinguished order unit. The Mundici dimension group Mn is an ℓ-group defined by the McNaughton functions over [0, 1]n. Theorem 2 ([8], [9]) (K0(A(x, B)), K + is defined by a subset of all McNaughton functions over [0, 1]n. 0 (A(x, B)), u) ∼= (Mn, 1), where Mn Remark 2 Theorem 2 for n = 1 was proved in [Mundici 1988] [8]. In par- ticular, the Markov cluster C ∗-algebra A(x, B) in Figure 1 has the dimension group M1. By an extension of the argument of [Mundici 2011] [9], one can prove Theorem 2 for n ≥ 1. 3 Proofs 3.1 Proof of theorem 1 We shall split the proof into a series of lemmas. 7 add(x, B). ai ≤ c ≤ bj. add(x, B)) is a dimen- Lemma 1 The ordered abelian group (Aadd(x, B), A+ sion group with the stable scale Γ ∼= A+ Proof. Recall that an ordered abelian group (G, G+) satisfies the Riesz in- terpolation property, if given {ai, bj ∈ G ai ≤ bj for i, j = 1, 2} there exists c ∈ G, such that (7) Let us show that the ordered group (Aadd(x, B), A+ Riesz interpolation property. Indeed, if ai = P Aixi and bj = P Bixi are the Laurent polynomials of ai, bj ∈ Aadd(x, B), then one can choose c = P Cixi such that Ci = Ai if Ai 6= 0 and 0 < Ci < Bi if Ai = 0. Clearly, the condition (7) is satisfied. By the Effros-Handelman-Shen Theorem, a countable ordered abelian group is a dimension group if and only if it satisfies the Riesz interpolation property, see [Effros 1981, Theorem 3.1] [2]. Thus (Aadd(x, B), A+ add(x, B)) is a dimension group with the stable scale. Lemma 1 is proved. add(x, B)) satisfies the Lemma 2 The exists a canonical isomorphism ϕ between the abelian groups K0(A(x, B)) and Aadd(x, B). Proof. The idea is to construct an isomorphism ϕ : Mn → Aadd(x, B), where Mn is the Mundici dimension group. The rest of the proof will follow from Theorem 2. We assume without loss of generality, that the linear functions αi in the set (6) constitute the Schauder-type basis for [0, 1]n. 1 We shall assign to each αi = bi + ai1x1 + . . . + ainxn a Laurent monomial, which is a gener- ator of the group Aadd(x, B). Roughly speaking, this can be done by an "exponentiation" of the variables xi. Indeed, consider a map ϕ acting by the formula: bi + ai1x1 + ai2x2 + . . . + ainxn 7→ bixai1 1 xai2 2 . . . xain n , 1 ≤ i ≤ m, (8) where aij ∈ Z and bi ∈ Z. polynomial according to the formula: The map ϕ sends the McNaughton function f : [0, 1]n → R to a Laurent f 7→ m Xi=1 bixai1 1 xai2 2 . . . xain n ∈ Z[x±1]. (9) 1Such a choice of αi provides injectivity of our construction. I am grateful to the referee for pointing out this fact to me. 8 It is verified directly, that pointwise addition of the McNaughton func- tions maps to an addition of the Laurent polynomials. One can see that our construction provides an injective mapping into the ring of Laurent poly- nomials. Working backwards our construction, it can be proved that the mapping is also surjective. Thus one gets an isomorphism of the abelian groups: (10) On the other hand, it follows from Theorem 2 that Mn ∼= K0(A(x, B)), where the set of the McNaughton functions over [0, 1]n is defined by the algebra A(x, B). Thus one obtains an isomorphism of the abelian groups: ϕ : Mn → Aadd(x, B). ϕ : K0(A(x, B)) → Aadd(x, B). (11) Lemma 2 is proved. Remark 3 Using the McNaughton functions over [0, 1]n, one can see that for the finite cluster algebras the group K0(A(x, B)) is isomorphic to a direct sum of finitely many copies of Z. This fact implies that the corresponding tracial simplex is spanned by n extremal traces. Remark 4 Lemma 2 implies that the group K0(A(x, B)) has the natural structure of a commutative ring, since K0(A(x, B)) ⊂ Z[x±1]. It is an in- teresting question to find an interpretation of the product in terms of the K-theory. Lemma 3 The isomorphism ϕ is order-preserving, i.e. 0 (A(x, B)) ∼= (Aadd(x, B), A+ Proof. In view of Theorem 2, it is sufficient to show that (K0(A(x, B)), K + add(x, B)). (Mn, 1) ∼= (Aadd(x, B), A+ add(x, B)) (12) are isomorphic dimension groups. The semi-group M+ n of positive elements of the Mundici dimension group (Mn, 1) consists of all piecewise linear functions with bi > 0. Likewise, the semigroup A+ On the other hand, formula (8) says that ϕ sends the coefficient bi into the coefficient bi of the Laurent monomial. Thus one gets the equality: add(x, B) consists of the Laurent polynomials with bi > 0. add(x, B). ϕ(M+ n ) = A+ 9 (13) In other words, the isomorphism ϕ : Mm → Aadd(x, B) preserves the semi- group of positive elements of the respective dimension groups. Lemma 3 follows from (13) and Theorem 2. Theorem 1 follows from lemma 3. Remark 5 ([11]) Theorem 1 implies that the category of cluster algebras can be embedded into the category of dimension groups (G, G+) with the stable scale. The following (partial) characterization of cluster algebras in terms of the dimension groups is true: The cluster algebras correspond to the dimension groups with a non-trivial spectrum P rim (G, G+) ∼= {Rn n ≥ 1}, where P rim (G, G+) is the space of primitive ideals of (G, G+) endowed with the Jacobson topology. 3.2 Proof of corollary 1 Let ψ be an inverse of the map ϕ constructed in lemma 2. We shall fix an isomorphism class of the AF -algebra A(x, B) and consider the corresponding dimension group (K0(A(x, B)), K + 0 (A(x, B)), Γ). In view of theorem 1, we have: ( Aadd(x, B) = ψ(K0(A(x, B))) A+ add(x, B) = ψ(K + 0 (A(x, B))). 0 (A(x, B)), one gets a scale ψ(Γ) ⊆ A+ (14) add(x, B) in the cluster On the other hand, it is verified directly that the set X = {xi}∞ Since Γ ⊆ K + algebra Aadd(x, B). i=1 of all cluster variables xi is a scale, since it is a generating, hereditary and directed subset of Aadd(x, B). But given isomorphism class of algebra A(x, B) can define only one scale on the cluster algebra Aadd(x, B); thus X ∼= ψ(Γ). It remains to recall that ψ(Γ) ⊆ A+ add(x, B). In other words, the coefficients of the Laurent polynomials corresponding to the cluster variables xi are non-negative integers. Corollary 1 is proved. add(x, B) and therefore X ⊆ A+ 4 An example To illustrate theorem 1, we shall consider a cluster C ∗-algebra A(1, 1) as- sociated to a triangulation of an annulus with one marked point on each boundary component, see [Fomin, Shapiro & Thurston 2008, Example 4.4] [5]; we shall keep the original notation of cited paper. 10 ❜ ✑✑✰ ◗◗s ❜ ✑✑✰ ◗◗s ✑✑✰ ◗◗s ❜ ❜ ✑✑✰ ◗◗s ✑✑✰ ◗◗s ✑✑✰ ◗◗s ❜ ❜ ❜ . . . ❜ . . . ❜ . . . ❜ . . . Figure 2: Bratteli diagram of algebra A(1, 1). 4.1 Cluster C∗-algebra A(1, 1) Let A = {z = x + iy ∈ C r ≤ z ≤ R} be an annulus in the complex plane such that r < R. Recall that the Riemann surfaces A and A′ are isomorphic if and only if R/r = R′/r′; the real number t = R/r is called a modulus of A. By TA = {t ∈ R t > 1} we understand the Teichmuller space of A, i.e. the space of all Riemann surfaces A endowed with a natural topology. The cluster algebra A(x, BT ) of rank two given by a matrix: 2 BT = (cid:18) 0 −2 0(cid:19) (15) is the coordinate ring of TA [Fomin, Shapiro & Thurston 2008, Example 4.4] [5]; the A(x, BT ) is related to the Penner coordinates on the space TA corresponding to an ideal triangulation T of A with one marked point on each boundary component of A [Williams 2014, Section 3] [15]. By A(1, 1) := A(x, BT ) we shall understand a cluster C ∗-algebra given by matrix BT ; the reader is encouraged to verify using formulas (1) that the Bratteli diagram of A(1, 1) has the form of a Pascal triangle shown in Figure 3. (The A(1, 1) is the so-called GICAR algebra [Effros 1980, p.13(e)] [2]; such an algebra has a rich set of ideals [Bratteli 1972, Section 5.5] [1].) By {σt : A(1, 1) → A(1, 1) t ∈ R} we denote a group of modular automor- phisms constructed in [11], Section 4; the σt is generated by the geodesic flow T t on the space TA. The A(1, 1) embeds into an UHF-algebra: M2(C). (16) M2∞ := ∞ Oi=1 11 0 i=1 exp(√−1(cid:18) 1 The M2∞ is known as a CAR algebra; unlike the A(1, 1), it is a simple AF - algebra with the Bratteli diagram shown in [Effros 1980, p.13(c1)] [2]. The 0 λ(cid:19)) 0 < λ < 1} defines a group of Powers product {⊗∞ modular automorphisms {σt : M2∞ → M2∞ t ∈ R}; it is not hard to observe, that σt ≡ σt on A(1, 1). Recall that if eij are the matrix units in M2(C), one can define a projection 1+t (e11 ⊗ e11 + te22 ⊗ e22 + √t(e12 ⊗ e ∈ M2(C) ⊗ M2(C) by the formula e = 1 e21 + e21 ⊗ e12)), where t ∈ R is a parameter. If θ is the shift automorphism of the UHF -algebra M2∞, then projections ei := θi(e) ∈ M2i satisfy the following relations: ( eiej eiei±1ei = = ejei, i − j ≥ 2 if (1+t)2 ei t and the Powers state ϕt : M2∞ → C satisfies the Jones equality: ϕt(wen+1) = t (1 + t)2 ϕt(w), ∀w ∈ M2n+1, (17) (18) see [Jones 1991, Section 5.6] [6] for the details. The ei generate the alge- bra M2∞ and taking new generators si such that σt(si) = tei − (1 − ei) one gets a representation of the braid group Bn = {s1, . . . , sn sisi+1si = si+1sisi+1, sisj = sjsi if i − j ≥ 2} in the algebra M2n. 4.2 Jones Index Theorem As an application of theorem 1, one gets a short proof of the Jones Index Theorem in terms of the cluster algebras. Corollary 2 Relations (17) define a C ∗-algebra if and only if the values of index (1+t)2 belong to the set: t [4,∞) [ {4 cos2(cid:18)π n(cid:19) n ≥ 3}. (19) Proof. To find admissible values of parameter t, we shall use a simple analysis of the cluster algebra A(1, 1) ∼= K0(A(1, 1)). Recall that algebra A(1, 1) has a unique canonical basis B consisting of the positive elements of A(1, 1), i.e. the Laurent polynomials with positive integer coefficients; the elements of B 12 generate the whole algebra A(1, 1). An explicit construction of B was given by [Sherman & Zelevinsky 2004, Theorem 2.8] [14]. Namely, i xq B = {xp i+1 p, q ≥ 0} [ {Tn(x1x4 − x2x3) n ≥ 3}, where Tn(x) are the Chebyshev polynomials of the first kind. Since T0 = 1 and Tn(cid:20) 1 2 (t + t−1)(cid:21) = 1 2 (tn + t−n), (20) (21) we shall look for a modulus t such that 1 2 (t + t−1) = x1x4 − x2x3. This is always possible since the Penner coordinates on the Teichmuller space TA are given by the cluster (x1, x2), where each xi is a function of modulus t [Williams 2014, Section 3] [15]. (i) Since t > 1, it is easy to see by direct substitution that the values of index belong to the interval: (4,∞). (22) (ii) To get discrete values, we shall assume that A(1, 1) is a finite cluster algebra, i.e. the number of xi is finite. It is immediate that B < ∞ and from the second series in (20) one obtains: Tn(x1x4 − x2x3) = T0 = 1 (23) for some integer n ≥ 1. But x1x4 − x2x3 = 1 for the Chebyshev polynomials, one gets an equation: 2(t + t−1) and using formula (21) tn + t−n = 2 (24) for (possibly complex) values of modulus t. Since (24) is equivalent to the equation t2n − 2tn + 1 = (tn − 1)2 = 0, one gets the n-th root of unity: However, the index t ∈ {e 2πi n n ≥ 1}. (1 + t)2 t = t−1 + 2 + t = 2(cid:20)cos (cid:18)2π n (cid:19) + 1(cid:21) = 4 cos2(cid:18)π n(cid:19) (25) (26) is a real number. Thus relations (17) define a C ∗-algebra. (We must exclude the case n = 2 corresponding to t = −1, because otherwise one gets a division by zero in (17).) Bringing together (22) and (26) one gets the conclusion of corollary 2. 13 Remark 6 The finite cluster algebras corresponding to the discrete moduli come from a triangulation of the n-gons or the n-gons with one puncture, see [Fomin, Shapiro & Thurston 2008, Table 1] [5]; such algebras are classified by their Coxeter-Dynkin diagrams of type An−3 and Dn, respectively. As explained, the A(1, 1) is a finite-dimensional C ∗-algebra having the Bratteli diagram similar to one shown in [Jones 1991, pp. 37-38] [6]. 4.3 Dimension group of the GICAR algebra We shall use theorem 1 to calculate a dimension group of the algebra A(1, 1). Corollary 3 (K0(A(1, 1)), K + is the semigroup of all positive-definite polynomials on the interval (0, 1). 0 (A(1, 1)), u) ∼= (Z[x], P +(0, 1), u), where P +(0, 1) Proof. It is known that the Chebyshev polynomials of the first kind Tn(x) lie in a basis B of the cluster algebra A(1, 1) [Sherman & Zelevinsky 2004, Theorem 2.8] [14]. For each 0 ≤ k ≤ n, we shall introduce a new basis B′ in A(1, 1) comprising the elements: 1 (x)(T0(x) − T1(x))n−k = xk(1 − x)n−k. T k (27) On the other hand, the Bratteli diagram in Figure 3 says that the group K0(A(1, 1)) is generated by the (equivalence classes of) projections [en k ] sub- ject to the relations: [en k ] = [en+1 ] + [en+1 k+1]. (28) Take a representation ρ of K0(A(1, 1)) in the cluster algebra A(1, 1) given by the formula: (29) ρ([en k k ]) = xk(1 − x)n−k, 0 ≤ k ≤ n. The reader can verify that relations (28) are satisfied. It is easy to see, that xk(1 − x)n−k are generators of the polynomial ring Z[x] and the rest of the proof repeats the argument in [Renault 1980, Appendix] [12]. Corollary 3 follows. Remark 7 Corollary 3 was first proved by [Renault 1980, Appendix] [12]; the GICAR algebra involved in the original proof is isomorphic to the cluster C ∗-algebra A(1, 1), see Figure 3. 14 − λ 2 λ 2 λ ≥ 2 −1 λ3 1 Figure 3: The fundamental domain of the Hecke group. 4.4 The Hecke group The space TA can be parameterized by a Fuchsian group of the second type. Namely, the Hecke group H(λ) is a subgroup of SL2(R) generated by matrices: (cid:18) 0 −1 0 (cid:19) and (cid:18) 1 λ 1(cid:19) , 1 0 (30) where λ ∈ R. The group H(λ) acts on the Lobachevsky plane H = {z = x + iy y > 0} by the linear fractional transformations. The Hecke theorem says that the action is discrete if and only if: λ ∈ [2,∞) ∪ {λn := 2 cos(cid:16) π n(cid:17) n ≥ 3}. The fundamental domain {z ∈ H ℜ(z) < λ z > 1} of H(λ) is 2 , sketched in Figure 4. It is easy to see, that H/H(λ) is a disk if λ ≥ 2 and there exits an Ahlfors map f : A → H/H(λ) of degree 2, such that: n(cid:19) n ≥ 3}}. TA ∼= {t = λ2 t ∈ [4,∞) ∪ {4 cos2(cid:18)π (31) (In fact, the Hecke theorem follows from (31) and analysis of algebra A(1, 1) given in Section 4.2; the A(1, 1) is a coordinate ring of TA.) Remark 8 Formula (31) relates the Hecke group H(λ) to the algebra A(1, 1) and henceforth to the index of subfactors by taking the weak closure of a representation of A(1, 1) ⊂ M2∞; such a question was raised in [Jones 1991, Section 3.1] [6]. I am grateful to Ralf Schiffler for an invitation to his Acknowledgments. Cluster Algebra Seminar at the University of Connecticut and many helpful discussions. I thank the anonymous referee for the interest, careful reading and useful suggestions regarding this paper. 15 References [1] O. Bratteli, Inductive limits of finite dimensional C ∗-algebras, Trans. Amer. Math. Soc. 171 (1972), 195-234. [2] E. G. Effros, Dimensions and C ∗-Algebras, in: Conf. Board of the Math. Sciences No.46, AMS (1981). [3] G. A. Elliott, On the classification of inductive limits of sequences of semisimple finite-dimensional algebras, J. Algebra 38 (1976), 29-44. [4] S. Fomin and A. Zelevinsky, Cluster algebras I: Foundations, J. Amer. Math. Soc. 15 (2002), 497-529. [5] S. Fomin, M. Shapiro and D. Thurston, Cluster algebras and trian- gulated surfaces, Part I: Cluster complexes, Acta Math. 201 (2008), 83-146. [6] V. F. R. Jones, Subfactors and Knots, CBMS Series 80, AMS, 1991. [7] K. Lee and R. Schiffler, Positivity for cluster algebras, Annals of Math. 182 (2015), 73-125. [8] D. Mundici, Farey stellar subdivisions, ultrasimplicial groups, and K0 of AF C ∗-algebras, Adv. in Math. 68 (1988), 23-39. [9] D. Mundici, Advanced Lukasiewicz Calculus and MV -algebras, Springer, 2011. [10] G. J. Murphy, C ∗-Algebras and Operator Theory, Academic Press, 1990. [11] I. Nikolaev, On cluster C ∗-algebras, J. Funct. Spaces 2016, Article ID 9639875, 8p. (2016) [12] J. Renault, A Groupoid Approach to C ∗-Algebras, Lecture Notes in Mathematics 793, Springer-Verlag 1980. [13] M. Rørdam, F. Larsen and N. Laustsen, An introduction to K-theory for C ∗-algebras. London Mathematical Society Student Texts, 49. Cam- bridge University Press, Cambridge, 2000. 16 [14] P. Sherman and A. Zelevinsky, Positivity and canonical bases in rank 2 cluster algebras of finite and affine types, Moscow Math. J. 4 (2004), 947-974. [15] L. K. Williams, Cluster algebras: An introduction, Bull. Amer. Math. Soc. 51 (2014), 1-26. Department of Mathematics and Computer Science, St. John's University, 8000 Utopia Parkway, New York, NY 11439, United States; E-mail: [email protected] 17
1510.05472	2	1510	2017-03-29T20:44:15	Purely Infinite Totally Disconnected Topological Graph Algebras	[ "math.OA" ]	We give a sufficient condition on totally disconnected topological graphs such that their associated topological graph algebras are purely infinite.	math.OA	math	PURELY INFINITE TOTALLY DISCONNECTED TOPOLOGICAL GRAPH ALGEBRAS HUI LI Abstract. We give a sufficient condition on totally disconnected topological graphs such that their associated topological graph algebras are purely infinite. 1. Introduction The Elliott program developed rapidly over the last twenty years with the goal of classi- fying all simple separable nuclear C ∗-algebras by the so-called Elliott's invariants (Notice that different types of classification might have different invariants). In the purely infinite case, Kirchberg in [9] and Phillips in [14] separately showed that all simple separable nuclear purely infinite C ∗-algebras in the UCT class can be classified by their K-theoretic data. Katsura in [6] gave a sufficient condition on simple topological graph algebras such that they are purely infinite, and constructed all simple separable nuclear purely infinite C ∗-algebras in the UCT class. Graph algebras were firstly defined by Kumjian, Pask, Raeburn, and Renault in [12] using the groupoid C ∗-algebras method. Topological graph algebras studied by Katsura (see [4]), which are seen to be a generalization of graph algebras, were defined by using a modified version of Pimsner's construction (see [15]). A few people then tried to realize topological graph algebras as groupoid C ∗-algebras, and these results can naturally be regarded as a generalization of Kumjian, Pask, Raeburn, and Renault's approach to graph algebras in [12]. For example, Katsura in [7] showed that the topological graph algebra of a compact topological graph with a surjective range map is isomorphic to a Renault- Deaconu groupoid C ∗-algebra. Yeend in [22] proved that every topological graph algebra is indeed an ´etale groupoid C ∗-algebra. Kumjian and Li in [10] strengthened Yeend's result by showing that every topological graph algebra is indeed a Renault-Deaconu groupoid C ∗-algebra (Yeend's result, and Kumjian and Li's result both cover a result of Brownlowe, Carlsen, and Whittaker from [3, Proposition 2.2]). In this article, we give a sufficient condition on totally disconnected topological graphs such that their associated topological graph algebras are purely infinite. Our approach is different than Katsura's. Our strategy is that we deal with topological graph algebras under the groupoid model, and we apply Anantharaman-Delaroche's criterion in [1], which yields purely infinite ´etale groupoid C ∗-algebras, to our settings. Date: 24 Feb 2017. 2010 Mathematics Subject Classification. 46L05. Key words and phrases. C ∗-algebra; topological graph; Renault-Deaconu groupoid; groupoid C ∗- algebra; pure infiniteness. This research was supported by Research Center for Operator Algebras of East China Normal Univer- sity and supported by Science and Technology Commission of Shanghai Municipality (STCSM), grant No. 13dz2260400. 1 2 HUI LI This paper is organized as follows. In Section 2, we give a background review on topological graph algebras and state the main theorem of [10]. In Section 3, we prove our main theorem, that is Theorem 3.11, which is a sufficient condition giving rise to purely infinite topological graph algebras. In Section 4, we give some remarks on our main theorem. 2. Preliminaries Throughout this paper, all the topological spaces are assumed to be second countable; and all the topological groupoids are assumed to be second countable. Our work in this article utilizes the Hilbert module, the C ∗-correspondence, and the Cuntz-Pimsner algebra machinery. These materials can be referred to [8, 13, 15, 16], etc. This paper also involves groupoids, groupoid C ∗-algebras, which can be found in [17]. 2.1. Topological Graph Algebras. In this subsection, we recap some background about topological graphs and topological graph algebras from [4, 10]. Definition 2.1. Let T be a locally compact Hausdorff space. Then T is said to be totally disconnected if T has an open base consisting of compact open subsets of T . Definition 2.2 ([4, Definition 2.1]). A quadruple E = (E0, E1, r, s) is called a topological graph if E0, E1 are locally compact Hausdorff spaces, r : E1 → E0 is a continuous map, and s : E1 → E0 is a local homeomorphism. In addition, E is said to be totally disconnected if E0, E1 are both totally disconnected. Definition 2.3 ([4]). Let E be a topological graph. For x, y ∈ Cc(E1), f ∈ C0(E0), e ∈ E1, and for v ∈ E0, define (x · f )(e) := x(e)f (s(e)); (f · x)(e) := f (r(e))x(e); and hx, yiC0(E 0)(v) := Xs(e)=v x(e)y(e). Then Cc(E1) is a right inner product C0(E0)-module with an adjointable left C0(E0)- action. Its completion X(E) under the k · kC0(E 0)-norm is called the graph correspondence associated to E. The Cuntz-Pimsner algebra of X(E), which is denote by O(E), is called the topological graph algebra of E. A subset N of E1 is called an s-section if sN : N → s(N) is a homeomorphism with respect to the subspace topologies. Define some useful subsets of E0 as follows. Define (1) E0 (2) E0 sce := E0 \ r(E1). fin := {v ∈ E0 : there exists an open neighborhood N of v such that r−1(N) is compact }. rg := E0 fin \ E0 sce. sg := E0 \ E0 rg. (3) E0 (4) E0 For n ≥ 2, define En :=nµ = (µ1, . . . , µn) ∈ nYi=1 E1 : s(µi) = r(µi+1), i = 1, . . . , n − 1o PURELY INFINITE TOTALLY DISCONNECTED TOPOLOGICAL GRAPH ALGEBRAS 3 regarded as a subspace of the product spaceQn `∞ n=0 En with the disjoint union topology. Define the infinite-path space i=1 E1. Define the finite-path space E∗ := E∞ :=nµ ∈ ∞Yi=1 E1 : s(µi) = r(µi+1), i = 1, 2, . . .o. Denote the length of a path µ ∈ E∗ ∐ E∞ by µ. A finite path µ ∈ E∗ \ E0 is called a cycle if r(µ) = s(µ). The vertex r(µ) is called the base point of µ. The cycle µ is said to be without entrances if r−1(r(µi)) = {µi}, for i = 1, . . . , µ. On the other hand, the cycle µ is said to have entrances if there exists 1 ≤ i ≤ µ such that r−1(r(µi)) 6= {µi}. Definition 2.4 ([4, Definition 5.4]). Let E be a topological graph. Then E is said to be topologically free if the set of base points of cycles without entrances has empty interior. Definition 2.5 ([10, Definitions 4.1, 4.7]). Let E be a topological graph. Define the boundary path space to be ∂E := E∞ ∐ {µ ∈ E∗ : s(µ) ∈ E0 sg}. For a subset S ⊂ E∗, define the cylinder set by Z(S) := {µ ∈ ∂E : there exists α ∈ S, such that µ = αβ}. Define a locally compact Hausdorff topology on ∂E to be generated by the basic open sets Z(U) ∩ Z(K)c, where U is an open set of E∗ and K is a compact set of E∗. By [10, Lemma 7.1], the one-sided shift map σ : ∂E \ E0 sg → ∂E is a local homeomor- phism. Definition 2.6. Let E be a topological graph and let v ∈ E0. Define the positive orbit space of v (see [5, Definition 4.1]) by Orb+(v) := {w ∈ E0 : there exists ν ∈ E∗, such that r(ν) = w, s(ν) = v}. E is said to be cofinal if for any µ ∈ ∂E, Orb+(r(µ))S(cid:16)Sµ E0. i=1 Orb+(s(µi))(cid:17) is dense in Remark 2.7. The cofinality for topological graphs is a generalization of the cofinality for directed graphs (see [11, 12]). Let E be a directed graph, the cofinality of E means that for any v ∈ E0, and for any µ ∈ ∂E, there exists ν ∈ E∗, such that r(ν) = v, s(ν) ∈ {r(µ)}S(cid:16)Sµ i=1{s(µi)}(cid:17). Theorem 2.8 ([5, Proposition 8.9, Theorem 8.12]). Let E be a topological graph. Then O(E) is simple if and only if E is topologically free and cofinal. 2.2. Groupoid C ∗-algebras. In this subsection, we state one of the main theorems of [10]. A topological groupoid is called a locally compact groupoid if its topology is locally compact Hausdorff. A locally compact groupoid is said to be ´etale if its range map is a local homeomorphism. Let Γ be a locally compact groupoid and let N ⊂ Γ. Then N is called an s-section if sN : N → s(N) is a homeomorphism with respect to the subspace topologies; N is called an r-section if rN : N → r(N) is a homeomorphism with respect to the subspace topologies; and N is called a bisection if sN , rN are both homeomorphisms with respect to the subspace topologies. 4 HUI LI Definition 2.9 ([18, Definition 2.4]). Let T be a locally compact Hausdorff space and let σ : dom(σ) → ran(σ) be a partial local homeomorphism. Define the Renault-Deaconu groupoid Γ(T, σ) as follows: Γ(T, σ) := {(t1, k1 − k2, t2) ∈ T × Z × T :k1, k2 ≥ 0, t1 ∈ dom(σk1), t2 ∈ dom(σk2), σk1(t1) = σk2(t2)}. Define the unit space Γ0 := {(t, 0, t) : t ∈ T }. For (t1, n, t2), (t2, m, t3) ∈ Γ(T, σ), define the multiplication, the inverse, the source and the range map by (t1, n, t2)(t2, m, t3) := (t1, n + m, t3); (t1, n, t2)−1 := (t2, −n, t1); r(t1, n, t2) := (t1, 0, t1); s(t1, n, t2) := (t2, 0, t2). Define the topology on Γ(T, σ) to be generated by the basic open set U(U, V, k1, k2) := {(t1, k1 − k2, t2) : t1 ∈ U, t2 ∈ V, σk1(t1) = σk2(t2)}, where U ⊂ dom(σk1), V ⊂ dom(σk2) are open in T, σk1 is injective on U, and σk2 is injective on V . Definition 2.10 ([10, Definition 7.5]). Let E be a topological graph. Define the boundary path groupoid to be the Renault-Deaconu groupoid Γ(∂E, σ) := {(µ, k − l, ν) ∈ ∂E × Z × ∂E : µ ∈ dom(σk), ν ∈ dom(σl), σk(µ) = σl(ν)}. The following theorem is a special case of [10, Theorem 7.6]. Theorem 2.11. Let E be a topological graph. Then O(E) is isomorphic to the groupoid C ∗-algebra C ∗(Γ(∂E, σ)). 3. Sufficient Conditions of Purely Infinite Topological Graph Algebras Definition 3.1 ([20]). Let A be a C ∗-algebra. Then A is said to be purely infinite if every nonzero hereditary C ∗-subalgebra of A contains an infinite projection. We firstly prove the following two technical lemmas. set K ⊂ E∗ satisfying Z(U) ∩ Z(K)c 6= ∅. Write K = Sk Lemma 3.2. Let E be a topological graph. Fix an open set U ⊂ E∗, and fix a compact i=0(K ∩ Ei) for some k ≥ 0. Then for any µ ∈ Z(U) ∩ Z(K)c with µ ≥ k, there exists an open subset V of E∗ such that µ ∈ Z(V ) ⊂ Z(U) ∩ Z(K)c. In particular, if µ = k, then V can be chosen to be an open neighborhood of µ in Eµ. Proof. We prove the first statement. Write µ = αβ where α ∈ U. Case 1. K = ∅. Let V := U. Then we are done. Case 2. K 6= ∅ and K ⊂ E0. Then r(µ) /∈ K. Take an open neighborhood N of r(µ) which does not intersect with K. Let V := (rα)−1(N) ∩ U. We have µ ∈ Z(V ) ⊂ Z(U) ∩ Z(K)c. Case 3. K 6= ∅ and K 6⊂ E0. Then k ≥ 1. Since µ ∈ Z(K)c, then for 1 ≤ i ≤ k there exists an open neighborhood Ni ⊂ Ei of µ1 . . . µi such that Ni does not intersect with PURELY INFINITE TOTALLY DISCONNECTED TOPOLOGICAL GRAPH ALGEBRAS 5 K ∩ Ei and r(Ni) does not intersect with K ∩ E0. Let if k > α > 0 if α > k if α = k if α = 0. V :=  i=1 (Ni × Ek−i)(cid:17) ∩ Nk ∩ ((U ∩ Eα) × Ek−α) (cid:16)Tk−1 (cid:16)Tk i=1(Ni × Eα−i)(cid:17) ∩ (U ∩ Eα) i=1 (Ni × Ek−i)(cid:17) ∩ Nk ∩ (U ∩ Eα) (cid:16)Tk−1 i=1 (Ni × Ek−i)(cid:17) ∩ Nk ∩ (rk)−1(U ∩ Eα) (cid:16)Tk−1 V := (U ∩ Eα) × Eβ (rk)−1(U) U ∩ Ek if α > 0, β > 0 if α = 0, β > 0 if β = 0. Then V is an open subset of Emax{α,k} and µ ∈ Z(V ) ⊂ Z(U) ∩ Z(K)c. We prove the second statement. If K = ∅, let Then V is an open neighborhood of µ in Eµ, and Z(V ) ⊂ Z(U) ∩ Z(K)c. If K 6= ∅, then it follows directly from the above construction. (cid:3) Lemma 3.3. Let E be a topological graph. Suppose that E0 sce = ∅. Then for any open set U ⊂ E∗ and any compact set K ⊂ E∗ satisfying Z(U) ∩ Z(K)c 6= ∅, there exists an open subset V of E∗ such that ∅ 6= Z(V ) ⊂ Z(U) ∩ Z(K)c. sg = E0 \ E0 sce = ∅ gives E0 Proof. The assumption E0 fin)c. Fix an open set U ⊂ E∗ and fix a compact set K ⊂ E∗ satisfying Z(U) ∩ Z(K)c 6= ∅. Write K = i=0(K ∩ Ei), for some k ≥ 0. Fix µ ∈ Z(U) ∩ Z(K)c. If µ ≥ k then by Lemma 3.2 we are done. So we may assume that µ < k. By Lemma 3.2, there exists an open i=0(K ∩ Ei))c. Then s(V ) is fin, we have s(µ) ∈ E0 \E0 fin. Sk neighborhood V ⊂ Eµ of µ such that Z(V ) ⊂ Z(U) ∩ Z(Sµ an open neighborhood of s(µ). Since µ ∈ ∂E and E0 Consider the set sg = E0 \E0 fin E0 = E0 rg ∪ (E0 F := {αµ+1 ∈ E1 : α ∈ K, α ≥ µ + 1}, which is a compact subset of E1. Then there exists e ∈ r−1(s(V )) such that e /∈ F . Take an open neighborhood W ⊂ E1 of e which does not intersect with F . Let O :=((V × W ) ∩ Eµ+1 W if µ > 0 if µ = 0. So ∅ 6= Z(O) ⊂ Z(U) ∩ Z(K)c. (cid:3) Definition 3.4 ([1, Page 202]). Let Γ be an ´etale groupoid. Then Γ is said to be essentially free if the set of elements in Γ0 whose isotropy group are trivial form a dense subset of Γ0. Lemma 3.5. Let T be a locally compact Hausdorff space and let σ : dom(σ) → ran(σ) be a partial local homeomorphism. Then the Renault-Deaconu groupoid Γ(T, σ) is essentially free if and only if the set {t ∈ T : t, σ(t), . . . are distinct } is dense in T . Proof. It is straightforward to see. (cid:3) Definition 3.6. Let E be a topological graph. Then E is said to be essentially free if the boundary path groupoid Γ(∂E, σ) is essentially free. 6 HUI LI The following proposition is a generalization of [11, Lemma 3.4]. Proposition 3.7. Let E be a topological graph. Then E is topologically free if and only if E is essentially free. Proof. First of all, suppose that E is not essentially free. We aim to show that E is not topologically free. Since E is not essentially free, there exists a nonempty open set N ⊂ ∂E which does not intersect with {µ ∈ ∂E : µ, σ(µ), . . . are distinct } due to Lemma 3.5. For each 0 ≤ p < q, define a closed subset of ∂E to be Bp,q := {µ ∈ ∂E : σp(µ) = σq(ν)}. For each 0 ≤ p < q, let Ap,q := Bp,q ∩ N, then Ap,q is a closed subset of the subspace N. Notice that N = S0≤p<q Ap,q because N does not intersect with {µ ∈ ∂E : µ, σ(µ), . . . are distinct }. By the Baire's category theorem, there exists a nonempty open subset O of N (O is also open in ∂E) contained in Ap0,q0 for some 1 ≤ p0 < q0. By Definition 2.5, there exist an open set U ⊂ E∗ and a compact set K ⊂ E∗ such that ∅ 6= Z(U) ∩ Z(K)c ⊂ O. By Lemma 3.2, there exist n ≥ 0 and an open subset V ⊂ En such that ∅ 6= Z(V ) ⊂ O. We may assume that n 6= 0. Let W := (V × (rp)−1(s(V ))) ∩ En+p0. Then Z(W ) ⊂ Z(V ). So Z(W ) ⊂ Ap0,q0. We deduce that every µ ∈ Z(W ) satisfies that σn+p0(µ) = σn+q0(µ) since µ ∈ Ap0,q0. [4, Proposition 2.8] assures that for each α ∈ W there exists µ ∈ Z(W ) such that µ = αβ. We then conclude that s(W ) is an open subset of E0 consisting of base points of cycles. We claim that s(W ) consists of base points of cycles without entrances. Suppose not, for a contradiction. We obtain two infinite paths αβ, αβ′ ∈ Z(W ), where α ∈ W, β1 6= β2. Since ν := β1 . . . βq0−p0, ν′ := β′ q0−p0 are cycles, ανν′νν′ · · · ∈ Z(W ) /∈ Ap0,q0, which is a contradiction. Hence s(W ) is an open subset of E0 consisting of base points of cycles. Therefore E is not topologically free. 1 . . . β′ Conversely suppose that E is essentially free. Suppose that E is not topologically free, for a contradiction. By Lemma 3.5, the set {µ ∈ ∂E : µ, σ(µ), . . . are distinct } is dense in ∂E. By [5, Proposition 6.12], there exist a nonempty open set V ⊂ E0 consisting of base points of cycles without entrances, and a homeomorphism σ on V such that σ = r ◦ (sr−1(V ))−1. By [4, Proposition 2.8], V ⊂ E0 rg. Fix a vertex v ∈ V . Let ν be the unique simple cycle such that r(ν) = v. By the assumption, there is a convergent sequence (ν(n))∞ n=1 ⊂ {µ ∈ ∂E : µ, σ(µ), . . . are distinct } with the limit ν. By [10, Lemma 4.8], r(ν(n)) → r(ν) = v, and so there exists N ≥ 1 such that r(ν(N )) ∈ V . Let α be the unique simple cycle with r(α) = r(ν(N )). Since V consists of base points of cycles without entrances and V ⊂ E0 rg, we deduce that ν(N ) = αα · · · , which is a contradiction. So E is topologically free. (cid:3) Definition 3.8 ([1, Definition 2.1]). Let Γ be an ´etale groupoid. Then Γ is said to be locally contracting if for any nonempty open set U ⊂ Γ0, there exist an open subset V ⊂ U and an open bisection N ⊂ Γ such that V ⊂ s(N) and s ◦ r−1 N −1(V ) $ V . Definition 3.9. Let E be a topological graph. Then E is said to be locally contracting if the boundary path groupoid Γ(∂E, σ) is locally contracting. Definition 3.10. Let E be a topological graph and let v ∈ E0. Then v is said to connect to a cycle if there exists u ∈ E0 such that v ∈ Orb+(u) and u is the base point of a cycle. The following theorem is a generalization of partial results from [11, Theorem 3.9]. PURELY INFINITE TOTALLY DISCONNECTED TOPOLOGICAL GRAPH ALGEBRAS 7 Theorem 3.11. Let E be a topologically free totally disconnected topological graph. Sup- pose that the following subset of E0 is dense in E0. B := {v : v connects to a cycle µ with entrances such that any open neighborhood N of µ contains an open neighborhood U of µ with rµ(U) ⊂ sµ(U)}. Then E is essentially free and locally contracting. Hence O(E) is purely infinite. Proof. Since every cycle of E has entrances, E is topologically free. So E is essentially free by Proposition 3.7. We claim that E0 sce = ∅. Suppose that E0 sce 6= ∅, for a contradiction. Fix v ∈ E0 sce. Then there exists an open neighborhood V of v not intersecting with r(E1). Since B is dense in E0, V ∩ B 6= ∅ which is a contradiction. So E0 sce = ∅ and we finish proving the claim. Now we prove that E is locally contracting. Fix a nonempty open set N ⊂ ∂E. By Lemma 3.3 there exists a nonempty open subset U of E∗ such that ∅ 6= Z(U) ⊂ N. Then s(U) is a nonempty open subset of E0. Since B is dense in E0, there exist µνeβ, µναeβ ∈ ∂E such that µ ∈ U, α is a cycle, and e 6= α1. We may assume that µν > 0 (the case µν = 0 would follow a similar argument). Take a compact open sµν-sections O1 such that µν ∈ O1 and Z(O1) ⊂ Z(U); and take compact open sα-sections O2 such that α ∈ O2, rα(O2) ⊂ sµν(O1) ∩ sα(O2), e /∈ {α′ 1 : α′ ∈ O2}. Define a compact open bisection of the boundary path groupoid Γ(∂E, σ) by S := U(Z((O1×O2)∩Eµνα), Z(O1), µνα, µν). Define a nonempty compact open subset of N by W := s(S). Then W ⊂ s(S). For any µ′ν′, µ′′ν′′ ∈ O1, α′ ∈ O2, β ∈ ∂E such that (µ′ν′α′β, µνα − µν, µ′′ν′′β) ∈ S, there exist unique µ′′′, ν′′′ ∈ O1, α′′′ ∈ O2 such that µ′′′ν′′′α′′′α′β ∈ Z(Z((O1 × O2) ∩ Eµνα)). So µ′ν′α′β = s(µ′′′ν′′′α′′′α′β, µνα − µν, µ′ν′α′β) ∈ W . Hence s ◦ r−1 S−1(W ) ⊂ W . Pick up an arbitrary γ ∈ s(e)∂E (see [4, Proposition 2.8]). Then µνeγ ∈ W but µνeγ /∈ s ◦ r−1 S−1(W ) because e /∈ {α′ S−1(W ) $ W . By Definition 3.9, Γ(∂E, σ) is locally contracting. 1 : α′ ∈ O2}. Therefore s ◦ r−1 By Theorem 2.11, O(E) is isomorphic with C ∗(Γ(∂E, σ)). Since E is essentially free (cid:3) and locally contracting, [1, Proposition 2.4] gives O(E) is purely infinite. Corollary 3.12. Let E be a totally disconnected topological graph such that O(E) is simple. Suppose that there exists a cycle µ with entrances such that any open neighborhood N of µ contains an open neighborhood U of µ with rµ(U) ⊂ sµ(U). Then E is essentially free and locally contracting. Hence O(E) is purely infinite. Proof. Since O(E) is simple, by Theorem 2.8 E is topologically free. By Proposition 3.7 E is essentially free. Fix v ∈ E0 and fix an open neighborhood V of v. Since O(E) i=1 Orb+(s(µi)) is dense in E0. Hence is simple, by Theorem 2.8 E is cofinal. So Sµ V ∩(cid:16)SN i=1 Orb+(s(µi))(cid:17) 6= ∅. Thus V ∩ B 6= ∅ (see Theorem 3.11). Theorem 3.11 implies that E is locally contracting. Therefore [1, Proposition 2.4] yields that O(E) is purely infinite. (cid:3) 4. Concluding Remarks Katsura in [6] defined a concept called contracting topological graphs which can pro- vide purely infinite topological graph algebras. We recall the definition of contracting topological graphs and state Katsura's result. 8 HUI LI Definition 4.1 ([6, Definition 2.3]). Let E be a topological graph. A nonempty precom- pact open set V ⊂ E0 is said to be contracting if there exists a finite family of nonempty open sets {Ui ⊂ Eni : ni ≥ 1}k (1) r(Ui) ⊂ V, i = 1, . . . , k; (2) for i 6= j, ni ≤ nj, we have {(µ1, . . . , µni) : (µ1, . . . , µni) ∈ Ui, µ ∈ Enj } = ∅; and i=1 satisfying the following. (3) V $Sk i=1 s(Ui). Moreover, E is said to be contracting at a vertex v ∈ E0 if Orb+(v) = E0 and every open neighborhood of v contains a contracting open set. Furthermore, E is said to be contracting if E is contracting at some vertex in E0. Theorem 4.2 ([6, Theorem A]). Let E be a topological graph. Suppose that O(E) is simple. If E is contracting then O(E) is purely infinite. We provide two examples which indicate that both of pure infiniteness conditions for simple totally disconnected topological graph algebras in Corollary 3.12 and Theorem 4.2 are not comparable. Example 4.3. Define E0 := {v}; E1 := {e0, e1}. Then O(E) ∼= O2 which is simple and purely infinite. Notice that e0 is a cycle with entrances and {e0} is an open neighborhood of e0 with r({e0}) ⊂ s({e0}). So the assumption of Corollary 3.12 is satisfied. However, it is easy to see that this graph is not contracting. Therefore the assumption of Theorem 4.2 is not satisfied. Furthermore, Katsura in [6, Remark 2.8] asked that whether the converse of Theorem 4.2 is true and this is a counterexample of the question. n=1{0, 1},Q∞ Example 4.4. Define E0 := {v}; E1 := {e0, e1}. This time we consider the dual graph n=1{0, 1}, id, σ) where σ is the one-sided shift. By [10, Theo- has no entrances, the assumption of Corollary 3.12 is not satisfied. On the other hand, ) of bE := (Q∞ rem 7.5], O(bE) ∼= O(E) so O(bE) is simple and purely infinite. Since every cycle of bE Orb+(000 . . . ) = bE0 and Z(0 · · · 0 ) ⊂ bE0 is contracting for all n ≥ 1. To see that Z(0 · · · 0 {z}n {z}n 1) ⊂ bE1. It is straightforward ) ⊂ bE1, U2 := Z(0 · · · 0 {z} {z}n to check that {U1, U2} satisfies Conditions (1) -- (3) of Definition 4.1. So Z(0 · · · 0 ) is con- tracting. Hence E is contracting and the assumption of Theorem 4.2 is satisfied. is contracting, let U1 := Z(0 · · · 0 n+1 Yeend in [22] showed that topological 1-graph C ∗-algebras coincide with topological graph algebras (see also [10]). Later, Renault, Sims, Williams, and Yeend in [19] pro- vided a sufficient condition for simple compactly-aligned topological higher-rank graph C ∗-algebras to be purely infinite. In the following we interpret their condition in the topological graph setting and present their result. Definition 4.5 ([19, Definition 5.7]). Let E be a topological graph. A nonempty pre- compact open set U ⊂ E0 is said to be contracting if there exist 0 ≤ n < m, a nonempty precompact open sn-section Yn, and a nonempty precompact open sm-section Ym, such that (1) sm(Ym) = sn(Yn); (2) r(Ym) ⊂ r(Yn) = U; (3) for µ ∈ Ym, ν ∈ Yn with rm(µ) = rn(ν), there exists γ ∈ Em−n such that µ = νγ; {z}n PURELY INFINITE TOTALLY DISCONNECTED TOPOLOGICAL GRAPH ALGEBRAS 9 (4) there exists a nonempty open subset W of YnE∗ such that {ν1 . . . νn : ν ∈ W } = Yn and that for µ ∈ Ym, ν ∈ W there is no γ ∈ E∗ satisfying µ = νγ or ν = µγ. Theorem 4.6 ([19, Proposition 5.8]). Let E be a topological graph. Suppose that O(E) is simple and that for any v ∈ E0, there exist n ≥ 0 and an open subset U of En satisfying that v ∈ r(U) and s(U) is contracting in the sense of Definition 4.5. Then O(E) is purely infinite. The disadvantage of this theorem is that in order to prove the pure infiniteness of a topological graph algebra, one has to check the contracting condition for every vertex. Overall, both criteria of Katsura and Renault-Sims-Williams-Yeend for topological graph algebras being purely infinite relies on the assumption of the given topological graph algebras being simple. Our approach does not but pay the price that we have to assume that the starting topological graphs are totally disconnected. Finally, we consider a compact totally disconnected topological graph E such that the range, source map are surjective and O(E) is simple. Schafhauser recently proved that O(E) is finite if and only if the source map is injective (see [21, Theorem 6.7]). On the other hand, it is hard to get a completely graphic characterization for O(E) to be purely infinite. However, by combining the recent work of Brown, Clark, and Sierakowski in [2] and the work of Kumjian and Li [10], we are able to reduce the problem by only considering the projections on the vertex space. More precisely, O(E) is purely infinite if and only if every nonzero projection on C(E0) is infinite. References [1] C. Anantharaman-Delaroche, Purely infinite C ∗-algebras arising from dynamical systems, Bull. Soc. Math. France 125 (1997), 199 -- 225. [2] J. Brown, L.O. Clark, and A. Sierakowski, Purely infinite C ∗-algebras associated to ´etale groupoids, Ergodic Theory Dynam. Systems 35 (2015), 2397 -- 2411. [3] N. Brownlowe, T.M. Carlsen, and M.F. Whittaker, Graph algebras and orbit equivalence, preprint, arXiv:1410.2308. [4] T. Katsura, A class of C ∗-algebras generalizing both graph algebras and homeomorphism C ∗-algebras I. Fundamental results, Trans. Amer. Math. Soc. 356 (2004), 4287 -- 4322. [5] T. Katsura, A class of C ∗-algebras generalizing both graph algebras and homeomorphism C ∗-algebras III. Ideal structures, Ergodic Theory Dynam. Systems 26 (2006), 1805 -- 1854. [6] T. Katsura, A class of C ∗-algebras generalizing both graph algebras and homeomorphism C ∗-algebras. IV. Pure infiniteness, J. Funct. Anal. 254 (2008), 1161 -- 1187. [7] T. Katsura, Cuntz-Krieger algebras and C ∗-algebras of topological graphs, Acta Appl. Math. 108 (2009), 617 -- 624. [8] T. Katsura, On C ∗-algebras associated with C ∗-correspondences, J. Funct. Anal. 217 (2004), 366 -- 401. [9] E. Kirchberg, The classification of purely infinite C ∗-algebras using Kasparov's theory, preprint, 1994. [10] A. Kumjian and H. Li, Twisted topological graph algebras are twisted groupoid C ∗-algebras, J. Oper- ator Theory, to appear, arXiv:1507.04449. [11] A. Kumjian, D. Pask, and I. Raeburn, Cuntz-Krieger algebras of directed graphs, Pacific J. Math. 184 (1998), 161 -- 174. [12] A. Kumjian, D. Pask, I. Raeburn, and J. Renault, Graphs, groupoids, and Cuntz-Krieger algebras, J. Funct. Anal. 144 (1997), 505 -- 541. [13] E.C. Lance, Hilbert C ∗-modules, A toolkit for operator algebraists, Cambridge University Press, Cambridge, 1995, x+130. [14] N.C. Phillips, A classification theorem for nuclear purely infinite simple C ∗-algebras, Doc. Math. 5 (2000), 49 -- 114. 10 HUI LI [15] M.V. Pimsner, A class of C ∗-algebras generalizing both Cuntz-Krieger algebras and crossed products by Z, Fields Inst. Commun., 12, Free probability theory (Waterloo, ON, 1995), 189 -- 212, Amer. Math. Soc., Providence, RI, 1997. [16] I. Raeburn and D.P. Williams, Morita equivalence and continuous-trace C ∗-algebras, American Mathematical Society, Providence, RI, 1998, xiv+327. [17] J. Renault, A groupoid approach to C ∗-algebras, Springer, Berlin, 1980, ii+160. [18] J. Renault, Cuntz-like algebras, Operator theoretical methods (Timi¸soara, 1998), 371 -- 386, Theta Found., Bucharest, 2000. [19] J. Renault, A. Sims, D.P. Williams, and T.Yeend, Uniqueness theorems for topological higher-rank graph C ∗-algebras, preprint, arXiv:0906.0829. [20] M. Rørdam and E. Størmer, Classification of nuclear C ∗-algebras Entropy in operator algebras, Operator Algebras and Non-commutative Geometry, 7, Springer-Verlag, Berlin, 2002, x+198. [21] C.P. Schafhauser, Finiteness properties of certain topological graph algebras, Bull. Lond. Math. Soc. 47 (2015), 443 -- 454. [22] T. Yeend,Topological higher-rank graphs and the C ∗-algebras of topological 1-graphs, Contemp. Math., 414, Operator theory, operator algebras, and applications, 231 -- 244, Amer. Math. Soc., Providence, RI, 2006. E-mail address: [email protected] Research Center for Operator Algebras and Shanghai Key Laboratory of Pure Math- ematics and Mathematical Practice, Department of Mathematics, East China Normal University, 3663 Zhongshan North Road, Putuo District, Shanghai 200062, China
1303.1842	1	1303	2013-03-07T22:48:49	On the problem of classifying simple compact quantum groups	[ "math.OA" ]	We review the notion of simple compact quantum groups and examples, and discuss the problem of construction and classification of simple compact quantum groups. Several new quantum groups constructed by Banica, Curran and Speicher since the author's first paper on simple quantum groups are shown to be simple using results of Raum and Weber.	math.OA	math	ON THE PROBLEM OF CLASSIFYING SIMPLE COMPACT QUANTUM GROUPS SHUZHOU WANG Dedicated to Professor S. L. Woronowicz on the Occasion of his 70th Birthday Abstract. We review the notion of simple compact quantum groups and examples, and discuss the problem of construction and classification of simple compact quantum groups. Several new quantum groups constructed by Banica, Curran and Speicher since the author's first paper on simple quantum groups are shown to be simple using results of Raum and Weber. 1. Introduction There are two main areas in the operator algebraic approach to quantum groups: com- pact quantum groups and locally compact quantum groups. The former has a satisfactory axiomatic theory due to Woronowicz [83, 85]. While the latter has witnessed great progress due to concerted efforts of several generations of mathematicians (cf. an incomplete list including [47, 48, 41, 42, 1, 50, 54]), it still does not have an axiomatic framework that contains the non-compact Drinfeld-Jimbo quantum groups [40] as examples except for special cases such as SLq(2, C) (cf. e.g. [55]), nor has the existence of Haar weight been established in general. This is in stark contrast with the fact that compact real forms of the Drinfeld-Jimbo quantum groups are special examples in the theory of compact quantum groups (see [59, 60, 62, 61, 51]), and Haar measure, Peter-Weyl theorem and Tannaka-Krein duality can be established from a very simple set of axioms [83, 84, 85]. Within this well established framework of compact quantum groups, recent work on com- pact quantum groups has been primarily on the construction, classification, structure and other (operator) algebraic properties of specific classes of compact quantum groups. The modern impetus in the theory of quantum groups came as a result of the discov- ery of new examples of Hopf algebras by Drinfeld-Jimbo on the algebraic side [40] and 2010 Mathematics Subject Classification. Primary 46L65; Secondary 17B37, 20G42, 58B32, 46L87, 46L55, 46L60, 46L89, 16T05, 81R50, 81R60. Key words and phrases. simple quantum groups, quantum permutation groups, quantum automor- phism groups, compact quantum groups, Woronowicz C ∗-algebras, Hopf Algebras. 1 2 SHUZHOU WANG Woronowicz on the analytic side [82]. These are deformation quantizations of the classi- cal Lie algebras and Lie groups, and much of the literature on quantum groups had been devoted to this approach to quantum groups. Starting in his thesis [70], the author took a different direction than the traditional deformation quantization method by viewing quantum groups as intrinsic objects and found in a series of papers (including [69] in collaboration with Van Daele) several classes of universal compact quantum groups that can not be obtained as deformations of Lie groups or Lie algebras, the most important of these are the universal compact quantum groups of Kac type Au(n) and their orthogonal counterpart Ao(n) [71], the more general universal compact quantum groups Au(Q) and their self-conjugate counterpart Bu(Q) [69, 74], where Q ∈ GL(n, C), and the quantum automorphism groups Aaut(B, tr) [75] of finite dimensional C ∗-algebras B endowed with a tracial functional tr, including the quantum permutation groups Aaut(Xn) on the space Xn of n points. These objects have been an international focus of study in the subject of compact quantum groups and interest in them continues unabated (cf. [2]-[35], [44], [49], [63]-[68]). The quantum groups Au(Q) have the remarkable universal property that can be used to give following alternative and concrete definition of compact matrix quantum groups that was originally defined by Woronowicz [83] more abstractly: a compact matrix quantum group is a quotient Au(Q)/I, where I is a Woronowicz C ∗-ideal in Au(Q). This means in geometric language that every compact matrix quantum group (including compact Lie group) is a quantum subgroup of Au(Q) for an appropriate choice of Q. In contrast, Drinfeld-Jimbo quantum groups and other deformations of Lie groups do not enjoy this property, but are quantum subgroups of Au(Q). Similarly, the quantum groups Bu(Q) have the universal property that every compact matrix quantum group with self-conjugate fundamental representation is of the form Bu(Q)/I, where I is a Woronowicz C ∗-ideal in Bu(Q). Without universal compact quantum groups, the Drinfeld-Jimbo quantum groups and other quantum groups obtained by deformation would be the end of the story. However, the outpouring of papers on the universal quantum groups in the last few years (e.g. [2]- [35], [44], [49], [63]-[68]) demonstrates depth of the subject: despite much work achieved so far, we have only seen the tip of the iceberg and the story is far from the end. Although compact quantum groups have a satisfactory axiomatic framework and Drinfeld- Jimbo quantum groups are their special examples, after the discovery of these new classes of universal compact quantum groups, it is a natural program to classify simple compact CLASSIFYING SIMPLE QUANTUM GROUPS 3 quantum groups. This program was initiated in [79], where it was shown that all compact quantum groups mentioned above are simple in generic cases. The main goals of the program on simple compact quantum groups are: (1) construct and classify simple compact quantum groups and their irreducible representations, (2) understand the structure of simple compact quantum groups and structure of compact quantum groups in terms of the simple ones, and (3) develop new applications of simple compact quantum groups in other areas of mathematics and physics, such as quantum symmetries in noncommutative geometry and algebraic quantum field theory. For goal (1), one would like to develop a theory of simple compact quantum groups that parallels the Killing-Cartan theory and the Cartan-Weyl theory for simple compact Lie groups. For this purpose, one must first construct all simple compact quantum groups. Though the work so far provides several infinite classes of examples of these, it should be pointed out that the construction of simple compact quantum groups is only at the beginning stage for this task at the moment, as all the simple compact quantum groups known so far are almost classical in the sense that their representation rings are isomorphic to those of ordinary compact groups and in particular are commutative. The first examples of simple compact quantum groups that are not almost classical should be directly re- lated to the universal quantum groups Au(Q) (see the footnote after Problem 4.1), where Q ∈ GL(n, C) are positive, n ≥ 2, though these quantum groups are not simple them- selves. The representation ring of Au(Q) is highly noncommutative, being roughly the free product of two copies of the ring of integers, according to Banica [3]. To construct other simple compact quantum groups, the most natural idea is to study quantum automor- phism groups of appropriate quantum spaces, such as those in the author's papers [75, 76], the papers of Banica, Bichon, Goswami and their collaborators [5]-[11], [32, 33], [44], [24]- [31]. In retrospect, both simple Lie groups and finite simple groups are automorphism groups. This suggests viability of this approach to the program. Natural mathematical and physical structures that have compact quantum automorphic group symmetries are compact commutative and noncommutative Riemannian manifolds in the sense of Connes [36, 38]. Such symmetries should be investigated first. The following heuristic may indicate the depth of the problem on classification problem of simple compact quantum groups. The finite dimensional factors are classified by the discrete set of natural numbers while the classification of infinite dimensional von Neu- mann factors involves continuous parameters. Similarly, simple compact Lie groups are classified by a discrete set of Cartan matrices, but the classification of simple compact quantum groups involves continuous parameters. However, since the algebraic structures 4 SHUZHOU WANG of compact quantum groups are richer and more rigid than those of von Neumann al- gebras, the classification of simple compact quantum groups might be more accessible than the classification of infinite dimensional factors. Even if a classification of simple compact quantum groups up to isomorphism is unattainable, just as von Neumann fac- tors are far from being classified up isomorphism, experience has demonstrated that the study of universal quantum groups and quantum automorphism groups is fruitful (cf. [71, 69, 74, 76, 78, 79], [2]-[35], [44], [49], [63]-[68] and references therein), and other types of classification theories may also be considered, such as the classification of easy quantum groups [21, 18, 19, 81] and the classification of restricted classes of quantum automorphism groups [8]-[12] by Banica et al.. An outline of the paper is as follows. In §2, we review the notion of compact quantum groups and simple compact quantum groups. In §3, we review examples of simple compact quantum groups constructed so far in [79]. In §4, we discuss problem of the fine structure of Au(Q) and simple quantum quotient groups from Au(Q), as well as quantum subgroups from free products of compact quantum groups. In §5, we give a list of problems related to almost classical compact quantum groups and compact quantum groups with property F . In §6 and §7, we discuss the problem of constructing simple compact quantum groups from quantum automorphisms of finite graphs and other quantum subgroups of the quantum permutation groups, easy quantum groups and quantum isometry groups. In §6, several new quantum groups constructed by Banica, Curran and Speicher [17] since [79] are shown to be simple using results in Raum [56] and Weber [81] along with results in [79]. 2. Compact quantum groups and simple compact quantum groups We first recall the definition of compact (matrix) quantum groups and then the notion of simple compact quantum groups. There are several equivalent definitions of compact (matrix) quantum groups, each has its own advantages over the others. We briefly describe below another equivalent defini- tion, which has the advantage of letting the reader to "visualize" all compact quantum groups more concretely. This equivalent definition is essentially in the literature, but not in the explicit form we describe below. Notation: For elements uij (i, j = 1, ...n) of a C ∗-algebra A, we define the following elements in the n × n matrix algebra Mn(A) over A: u := (uij)n i,j=1, ut := (uji)n i,j=1 and u∗ := ¯ut, i.e. u∗ = (u∗ i,j=1, ¯u := (u∗ ji)n i,j=1. ij)n CLASSIFYING SIMPLE QUANTUM GROUPS 5 Definition 2.1. (cf. [70, 71, 69, 74, 78]) The universal compact matrix quantum groups are defined to be the family of pairs (Au(Q), ∆u), where Q ∈ GL(n, C), Q > 0, and uij (i, j = 1, ...n) are generators of the universal C ∗-algebra Au(Q) that satisfies the following sets of relations: u∗u = In = uu∗, utQ¯uQ−1 = In = Q¯uQ−1ut, and ∆u : Au(Q) → Au(Q) ⊗ Au(Q) is the uniquely defined morphism such that ∆u(uij) = uik ⊗ ukj. n Xk=1 Note that instead of restricting Q to positive matrices, one can still define (Au(Q), ∆u) for any invertible Q. For such Q, one has the following free product decomposition [78], Au(Q) ∼= Au(P1) ∗ Au(P2) ∗ · · · ∗ Au(Pk) for appropriate positive matrices P1, P2, ..., Pk with compatible coproducts as in [71]. Definition 2.2. (cf. where π : Au(Q) → A and ∆ : A → A ⊗ A are C ∗-morphisms such that [83]) A compact matrix quantum group is a triple (A, ∆, π), (1) π is surjective, and (2) ∆π = (π ⊗ π)∆u. Note that using [71], conditions (1) and (2) above is equivalent to (1) plus the following condition: (2)′ ∆u(ker(π)) ⊂ ker(π ⊗ π). Hence a compact matrix quantum group can be defined even more simply as a pair (A, π) satisfying (1) and (2)′. The C ∗-algebra A in the definition of a compact matrix quantum group (A, ∆, π) is called a finitely generated Woronowicz C ∗-algebra. The morphism ∆ is called the coproduct of A. It can be shown that there is a Hopf ∗-algebra structure (Au(Q), ∆, ε, S) on the dense ∗-subalgebra Au(Q) of Au(Q) such that S(uij) = u∗ ji, ε(uij) = δij, i, j = 1, 2, · · · n. This ∗-Hopf algebra structure induces Hopf ∗-algebra structure on the dense ∗-subalgebra A of A in the above definition of compact matrix quantum group, and on the dense ∗-subalgebra A of A in the definition the compact quantum group below. 6 SHUZHOU WANG As in 2.3 of [71], one can define morphism between compact matrix quantum groups as opposite of morphisms of finitely generated Woronowicz C ∗-algebras. Under these morphisms, compact matrix quantum groups form a category, though this category is not closed under inverse limits, which leads to the following notion of compact quantum groups using 3.1 of [71], Definition 2.3. (cf. [1, 85]) A compact quantum group (A, ∆) is an inverse limit of compact matrix quantum groups (Aλ, ∆λ, πλλ′). We will see in the next few paragraphs that the above notion of compact quantum groups is equivalent to the elegant and abstract one in [85]. As an inductive limit (instead of inverse limit) of finitely generated Woronowicz C ∗- algebras Aλ, the C ∗-algebra A in the definition above is called a Woronowicz C ∗- algebra. Kernels of morphisms between Woronowicz C ∗-algebras are called Woronowicz C ∗-ideals, which can also be intrinsically defined as in 2.3 of [71]. Remark: Intuitively, we think of A = C(G), where G is a compact quantum group, even there might be no points in "G" other than the identity. We also use the notations AG = C(G) and C(GA) = A. As in the literature, by abuse of terminology, A is also called a compact (matrix) quantum group besides being called a (finitely generated) Woronowicz C ∗-algebra. As usual, one defines the Haar state/measure. One then establishes the existence and uniqueness of the Haar state/measure on any compact matrix quantum group just as in [83]. Using this and 3.3 of [71], one establishes the existence and uniqueness of the Haar state/measure on any compact quantum group G. In the following, the Haar state on A = C(G) is denoted by hG or simply h if no confusion arises. Using the Haar state/measure h, the Peter-Weyl theory for all compact quantum groups can be developed as in [85] or [53]. As a result, we see that the above definition of compact quantum groups is equivalent to the one in [85]. We use AG to denote the dense ∗-subalgebra of AG consisting of coefficients of finite dimensional representations of G. As a consequence of the Peter-Weyl theory for compact quantum groups, AG is a Hopf ∗-algebra. We need to recall the notion of normal quantum subgroups [71, 80] to define simple compact quantum groups. Let (N, π) be a quantum subgroup of a compact quantum group G, which, as defined in [70, 71], means that π : C(G) −→ C(N) is a surjection of C ∗-algebras such that (π ⊗ π)∆G = ∆N π, where ∆G, ∆N are coproducts of C(G) and C(N), respectively. CLASSIFYING SIMPLE QUANTUM GROUPS 7 Define C(G/N) := {a ∈ C(G)(id ⊗ π)∆(a) = a ⊗ 1N }, C(N\G) := {a ∈ C(G)(π ⊗ id)∆(a) = 1N ⊗ a}, ∆ being coproduct on C(G), 1N the unit of C(N). Definition 2.4. (cf. [71, 79]) We say N is normal in G if it satisfies one of the equivalent conditions in the proposition below. Proposition 2.5. (cf. [79]) Let N be a quantum subgroup of a compact quantum group G. The following conditions are equivalent: (1) C(N\G) is a Woronowicz C ∗-subalgebra of C(G). (2) C(G/N) is a Woronowicz C ∗-subalgebra of C(G). (3) C(G/N) = C(N\G). (4) For every irreducible representation uλ of G, either hN π(uλ) = Idλ or hN π(uλ) = 0, where hN is the Haar measure on N, dλ = dim(uλ) and Idλ the dλ × dλ identity matrix. Among the above four equivalent formulations of the notion of normal quantum sub- groups, condition (4) is the most convenient for our purposes. If G is a compact group, then the above definition coincides with the usual notion of closed normal subgroups. To avoid complications with classification of finite quantum groups, we want to restrict the notion of simple quantum groups to quantum groups that are connected. Definition 2.6. (cf. [79]) A compact quantum group G is called connected if for each non-trivial irreducible representation uα ∈ G, the C ∗-algebra C ∗(uα ij) generated by the coefficients of uα is of infinite dimension. A compact matrix quantum group G is called simple (resp. absolutely simple) if it is connected and has no non-trivial connected normal quantum subgroups (resp. non-trivial normal quantum subgroups) and no non-trivial representations of dimension one. Note that compact Lie groups are (commutative) examples of compact quantum groups. It is easy to show that a compact Lie group is simple (resp. connected) in the usual sense if and only if it is simple (resp. connected) in the sense above. Problem 2.7. In the definition of simple compact quantum groups, if one replaces the condition "it has no non-trivial representations of dimension one" with the apparently less stringent condition "C(G) 6= C ∗(Γ) where Γ is a discrete group", do we get an equivalent definition? 8 SHUZHOU WANG Note according to [82], the condition C(G) 6= C ∗(Γ) above means that G is a nonabelian compact quantum group. For a compact Lie group G, this condition simply means G is a nonabelian Lie group, and the answer to the above question is affirmative by Weyl's dimension formula. Just as the notion of simple compact Lie groups excludes the torus groups, the definition of simple quantum groups above (including the alternative one formulated in Problem 2.7) excludes the compact quantum groups coming from group C ∗-algebras C ∗(Γ) of discrete groups Γ (i.e., abelian quantum groups). This is important because the classification of discrete groups is out of reach. 3. Examples of simple compact quantum groups that are almost classical and have property F The simple compact quantum groups that have been constructed so far share many properties common to compact Lie groups. We recall two of these properties [79]. Just as for compact groups, the representation ring (also called the fusion ring) R(G) of a compact quantum group G is a partially ordered algebra over the integers Z generated by the irreducible characters of G. The set of characters of G is a semi-ring that defines the order of R(G). [79]) A compact quantum group G is said to have property F if Definition 3.1. (cf. each Woronowicz C ∗-subalgebra of C(G) is of the form C(G/N) for some normal quantum subgroup N of G. A compact quantum group is called almost classical if its representation ring R(G) is order isomorphic to the representation ring of a compact group. In plain language, a compact quantum G is said to have property F if its quantum function algebra C(G) behaves exactly as the function algebras of compact groups with respect to normal subgroups. Note that compact quantum groups C ∗(Γ) from dual Γ of a discrete group Γ do not have this property unless the discrete group Γ is abelian, in which case the group C ∗-algebra C ∗(Γ) is a genuine function algebra over the Pontryagin dual Γ of the discrete abelian group Γ. The notion dual to property F is given by following definition, which captures the prop- erty that compact quantum group C ∗(Γ) has with respect to normal quantum subgroups: Definition 3.2. (cf. [79]) A compact quantum G is said to have property F D if each of quantum subgroup of G is normal. CLASSIFYING SIMPLE QUANTUM GROUPS 9 Proofs of assertions in [79] concerning properties F and F D, along with other related properties of compact quantum groups, can be found in [80]. Quantum groups Bu(Q) (cf. [70, 71, 69, 74, 78]) Keeping the notation for definition of Au(Q) in §2. Let Q ∈ GL(n, C) be such that Q ¯Q = ±In, n ≥ 2. Bu(Q) is defined to be the universal C ∗-algebra with generators uij (i, j = 1, 2, · · · , n) that satisfy that following sets of relations: u∗u = In = uu∗, utQuQ−1 = In = QuQ−1ut. It can be shown that there is a well-defined morphism ∆ : Bu(Q) → Bu(Q) ⊗ Bu(Q) such that n ∆(uij) = uik ⊗ ukj. Xk=1 and that (Bu(Q), ∆) is a compact matrix quantum group. The quantum groups Bu(Q) have the universal property that every compact matrix quantum group with self-conjugate fundamental representation is of the form Bu(Q)/I, where I is a Woronowicz C ∗-ideal in Bu(Q). Note: Bu(Q) is also denoted by Ao(Q∗) by Banica et al. When Q = In, Bu(Q) is just Ao(n), the universal orthogonal quantum group of Kac type introduced in 4.5 of [71]. Banica et al. also invented the notation O+(n) to signify Ao(n) = C(O+(n)). In addition, when Q = " 0 −In 0 #, Bu(Q) is a quantum symplectic group. This is In one of the reasons that we use the notation Bu(Q) instead of the notation Ao(Q), as the latter only captures the special case Q = In. Following the notation O+(n) of Banica et al., we denote the above universal symplectic quantum group by Sp+(n). The quantum symplectic group has also appeared in recent work of Bhowmick, D'Andrea, Das and D¸abrowski on quantum gauge symmetries (see [26]). For any invertible Q ∈ GL(n, C) that does not satisfy Q ¯Q = ±In, Bu(Q) can be defined by the same relations as above, but Bu(Q) ∼= Au(P1) ∗ Au(P2) ∗ · · · ∗ Au(Pk)∗ ∗ Bu(Q1) ∗ Bu(Q2) ∗ · · · ∗ Bu(Ql). 10 SHUZHOU WANG for certain Pi > 0 and Qj such that Qj ¯Qj's are scalars (cf. [78]). We note that both Au(Q) and Bu(Q) can be alternatively described as quantum auto- morphism groups of appropriate spaces (cf. [76, 26]). Quantum automorphism groups Aaut(B, τ ) (cf. [75]) Let B be a of a finite dimensional C ∗-algebra and φ a functional on B. In general the quantum automorphism group Aaut(B, φ) that preserves the system (B, φ) exists, which is the universal object in the category of compact quantum groups acting on the system (B, φ). However, as shown in Banica [4], only when φ is the canonical trace τ does the quantum group Aaut(B, τ ) have a representation theory that is relatively easy to describe, where the trace τ on B is called canonical if it coincides with the restriction to B of the unique tracial state on the algebra L(B) of operators with B acting by the GNS representation associated with the trace τ . If one identifies B with Lm m m τ ( bk) = T r(bk), k=1 Mnk(C), then Xk=1 n2 k n Xk=1 where bk ∈ Mnk , n is the dimension of B, and T r is ordinary trace on Mnk , i.e. T r(bk) is the sum of the diagonal entries of the matrix bk. In either case, explicit description of Aaut(B, φ) or Aaut(B, τ ) in terms of generators and relation is complicated for a general finite dimensional C ∗-algebra B. However, when B = C(Xn) is the commutative C ∗-algebra of functions on the space where Xn is the space of n points, the quantum permutation group Aaut(Xn) := Aaut(C(Xn)) has a surprisingly simple description in terms of generators and relations: The C ∗-algebra Aaut(Xn) is generated by self-adjoint projections aij such that each row and column of the matrix (aij)n i,j=1 adds up to 1, i.e., a2 ij = aij = a∗ ij, n i, j = 1, · · · , n, aij = 1, i = 1, · · · , n, aij = 1, j = 1, · · · , n. n Xj=1 Xi=1 Banica et al. invented the very convenient geometric notation S+ n so that Aaut(Xn) = C(S+ n ). In [34], Bichon generalizes the quantum permutation groups to purely algebraic context and establishes the universal property of the quantum permutation groups in complete generality. CLASSIFYING SIMPLE QUANTUM GROUPS 11 Theorem 3.3. (cf. 4.1 and 4.7 in [79]) (a) For Q ∈ GL(n, C) such that Q ¯Q = ±In and n ≥ 2, Bu(Q) is almost classical simple compact quantum group with property F . (b) Let B be a finite dimensional C ∗-algebra endowed with its canonical trace τ and dim(B) ≥ 4, then quantum group Aaut(B, τ ) is an almost classical absolutely simple com- pact quantum groups with property F . As special cases of Bu(Q) and Aaut(B, τ ), we have Corollary 3.4. (a) The universal orthogonal quantum groups O+(n) (for n ≥ 2) and the universal symplectic quantum groups Sp+(n) (for n ≥ 1) are simple. (b) The quantum permutation groups S+ n are simple for n ≥ 4. The main ideas used in the proof of Theorem 3.3 include (1) Banica's fundamental work of on the structure of fusion rings of these quantum groups (cf. Th´eor`eme 1 and Theorem 4.1 in [2, 4] respectively); (2) Correspondence between Hopf ∗-ideals and Woronowicz C ∗-ideals (cf. 4.2 - 4.3 in [79]); (3) Reconstruction of a normal quantum group from the identity in the quotient quan- tum group (cf. 4.4 in [79]). In addition to the fusion rings of compact quantum groups such as those considered in [2, 4], the matters in (2) and (3) above are of interest in their own right and worth further investigation. The correspondence between Hopf ∗-ideals and Woronowicz C ∗-ideals and related matters in (2) relate algebraic and analytical aspects of compact quantum groups. In purely Hopf algebras context, reconstruction of a normal quantum group from the quotient quantum group in (3) has been an issue since 1970's and is related to several other important and old open questions [80]. Quantum groups Kq, K u q and KJ (cf. [62, 61, 52, 58, 73] and [59, 60]) A unified study of the compact quantum groups Kq, K u q is due to Soibelman & Vaks- man, Levendorskii [62, 61, 52]. The ∗-Hopf algebras AKq are algebras of "representative functions" of Drinfeld-Jimbo quantum groups Uq(g) and define in a sense their "compact real form". The quantum group Kq is a deformation of the Poisson Lie group K(1, 0) (cf. q are twisting of the ∗-Hopf algebras AKq by an element [52]). The ∗-Hopf algebras AK u u ∈ ∧2hR. The quantum group K u q is a deformation of the Poisson Lie group K(1, u) (cf. 12 SHUZHOU WANG [52]). As shown in [73], K u dimensional vector space as conjectured by Rieffel (cf. [57, 58] for the background). q is an example of Rieffel's deformation from action of finite Rieffel's quantum group deformation KJ [58] depends on J = S ⊕ (−S), where S is a skew symmetric operator on the Lie algebra (viewed as Rn) of a torus subgroup of the compact Lie group K. For appropriate choice of S, KJ is a deformation of Poisson Lie group K(0, u) [52, 58]. An action of Rd := Rn × Rn on A = C(K) can be constructed and Rieffel's theory of deformation for action of Rd [57] can be applied to obtain AJ [58], also denoted C(KJ). Precise description of Kq, K u q and KJ require more space than appropriate in this paper. For our purposes, these quantum groups can be roughly described as follows: (1) The associated dense Hopf ∗-algebras AKq, AK u q and AKJ are the same vector space as the un-deformed/un-twisted ones AK, AKq and AK respectively, but the algebras AKq, AK u q and AKJ have deformed products; (2) The Hopf ∗-algebras AKq, AK u q and AKJ have the same coproduct as the un- deformed/un-twisted ones AK, AKq and AK respectively; (3) Representation theories of the deformed/twisted quantum groups Kq, K u q and KJ are the same as the un-deformed/un-twisted ones K, Kq and K respectively. Theorem 3.5. cf. (5.1, 5.4 and 5.6 in [79]) If K is a simple compact Lie group, then Kq, K u q , KJ are almost classical simple compact quantum groups with property F . The main ideas used in the proof of Theorem 3.5 include (1) Representation theory of these quantum groups; (2) Correspondence between Hopf ∗-ideals and Woronowicz C ∗-ideals, as in the proof of Theorem 3.3; (3) Reconstruction of a normal quantum group from the identity in the quotient quan- tum group, as in the proof of Theorem 3.3; (4) The normal subgroups of the undeformed Lie group remain to be normal subgroups of the deformed quantum groups and explicit identification of normal quantum subgroups of the deformed quantum groups. Other deformations of compact Lie groups, though constructed not as systematic as the ones considered by Drinfeld-Jimbo, Soibelman et al. and Rieffel, are scattered in the literature. We believe the general ideas used in the proof of Theorem 3.5 can also be applied to such deformations. CLASSIFYING SIMPLE QUANTUM GROUPS 13 4. Quotient quantum groups from Au(Q) and free products In [78], the quantum groups Au(Q) are classified up to isomorphism for positive matrices Q > 0 and the quantum groups Bu(Q) are classified up to isomorphism for matrices Q with Q ¯Q = ±In. It is shown that the corresponding Au(Q) (resp. Bu(Q)) is not a free product, or a tensor product, or a crossed product. However, for general non-singular matrices Q, we have the decomposition theorem expressing Au(Q) and Bu(Q) in terms of free product of the forgoing quantum groups (cf. Theorem 3.1 and 3.3 in [78]). In the light of these results, the following problem seems to be fundamental: Problem 4.1. Study further the fine structure of Au(Q) for positive matrices Q ∈ GL(n, C) and n ≥ 2; Determine their simple quotient quantum groups. A solution of this problem will also provide the first examples1 of simple compact quantum groups that are not almost classical (see §3). Note that the Au(Q)'s have the 1 dimensional diagonal torus T as their (connected) normal quantum subgroup, as observed by Bichon (private communication, cf. 4.5 in [79]), so they are not simple. However, they are very close to being simple. For example, they have no non-trivial irreducible representations of dimension one [3, 78]. It is worth noting that in Problem 4.1, simple quotient quantum groups of Au(Q) should be easier to determine than simple quantum subgroups of Au(Q), since the latter is tantamount to finding all simple quantum groups due to the universal property of Au(Q), which include all simple compact quantum groups in [79] as reviewed in §3, simple compact Lie groups, as well as all the other unknown simple compact quantum groups. By investigating the fine structure of concrete quantum groups such as Au(Q), one can expect to gain insights into the structure of general compact quantum groups and simple quantum groups. Sections §6 and §7 below contain more directions of research on this approach to quantum groups. Suitable modifications of the method for the proofs of the main results in section 4 of the paper [79] should yield a solution to Problem 4.1. The extra work needed for this kl and u∗ 1After this paper was accepted for publication, Alexandru Chirvasitu informed the author that in the preprint "Free unitary quantum groups are (almost) simple", he showed that the quantum group generated by uij u∗ kluij is a simple compact quantum group with noncommutative representation ring and without property F , where uij are the generators of Au(Q) with Q > 0. In the same preprint, he also showed that the quotient quantum group of Au(Q) by its central subgroup T in the proof of 4.5 in [79] has no normal quantum subgroups but is not finitely generated, and all normal subgroups are subgroups of T, giving a complete classification of simple quotient groups of Au(Q). See also footnote before Problem 5.4 below. 14 SHUZHOU WANG problem that do not appear in [79] is that there are more Woronowicz subalgebras in Au(Q) to consider than therein. Some preliminary computations of these subalgebras give optimism to a positive solution of the problem. One of the main ingredients in this calculation is Banica's fundamental result on fusion rules of the irreducible representations of the quantum group Au(Q) [3]. The general Au(Q) (resp. Bu(Q)) for arbitrary Q ∈ GL(n, C) is not simple if C ∗(Z) appears in its free product decompositions as described in [78]. This is because of the fact that Au(Q′) = C ∗(Z) = C(T) (resp. Bu(Q′) = C ∗(Z/2Z)) for Q′ ∈ GL(1, C) and the following result ([80]): Proposition 4.2. Let G1, G2 be compact quantum groups. Let G = G1∗G2 be the free product compact quantum group [71] underlying AG1 ∗AG2. Let π1 be the natural embedding of G1 into G defined by the surjection π1 : AG1 ∗ AG2 → AG1, π1 = id1 ∗ ǫ2. If G1 has at least one irreducible representation of dimension greater than one, then (G1, π1) is not a normal quantum subgroup of G1∗G2. Otherwise, (G1, π1) is normal in G1∗G2. The hat in the symbol ∗ above signifies the "Fourier transform" of free product ∗ remiscent of classical case in which Gk = Γk, where Γk are discrete abelian groups. Note that a compact Lie group being simple means roughly that it is not a direct product of proper connected subgroups. A similar result also holds for quantum groups: If GA is a simple compact quantum group, then AG is not a tensor product (i.e. G is not a direct product of its non-trivial quantum subgroups) [72, 80]. However, the proposition above says that the evident quantum subgroups (G1, π1) and (G1, π2) of G1∗G2 are not normal in G1∗G2 when G1 and G2 have no non-trivial represen- tations of dimension one. Along with other results of the author, this may suggest that the following problem on the structure of simple quantum groups has a positive solution. Problem 4.3. Let G1 and G2 be simple compact quantum groups. Is G1∗G2 also simple? The results in author's paper [71] should be useful for a solution of this problem. In particular, Theorem 1.1 there should play a role, as it did in [2, 3, 33, 78]. Note that the formula for the Haar measure on G1∗G2 and the classification its irreducible representations are given in explicit formulas in Theorem 1.1 in [71]. According to the postulates in the definition of a normal quantum subgroup (Definition 2.4), these are important ingredients in determining whether G1∗G2 has normal quantum subgroups. CLASSIFYING SIMPLE QUANTUM GROUPS 15 The results and methods of the paper [79] (especially section 4 therein) should also be useful for this problem. A positive solution to Problem 4.3 would have the following implication: G1∗G2 would be a simple compact quantum group when both G1 and G2 are merely simple compact Lie groups. The following easier variation of Problem 4.3 is also of interest and is related to the problem of determining whether several families of easy quantum groups are simple (cf. §6 below). Problem 4.4. Let G1 be a simple compact quantum group and G2 a finite quantum group. Is G1∗G2 also simple? Note that in view of the problems in §6 below, it would be interesting to solve Prob- lem 4.4 for G2 a finite group. Also related problems can be formulated for Bichon's free wreath product of compact quantum groups [33] (cf. §6 below): Problem 4.5. Let G1 be a simple compact quantum group and G2 = S+ permutation group. Is the free wreath product G1∗S+ n also simple? n the quantum 5. Problems related to almost classical compact quantum groups and property F As reviewed in §3, the simple compact quantum groups known so far are almost classical and have property F . Such quantum groups seems to be most accessible at the moment. The following problem evidently is less difficult than the general problem of classifying simple compact quantum groups and should be attempted first: Problem 5.1. Classify simple compact quantum groups that are almost classical and have property F . The following closely related problems should be considered also: Problem 5.2. (a) Classify almost classical simple compact quantum groups. (b) Classify simple compact quantum groups with property F . Problem 5.3. Does simple compact quantum groups with property F D exist? If so, construct and classify them. 16 SHUZHOU WANG In Problem 5.1 and Problem 5.2, it would be interesting enough to restrict consideration to almost classical simple compact quantum groups that have the same representation rings as simple compact Lie groups. In another direction, for the apparently more difficult problem of classifying simple compact quantum groups that are not almost classical or without property F , we do not have a single example of them. Therefore the following is a basic problem2: Problem 5.4. (a) Construct an example of simple compact quantum group that is not almost classical. (b) Construct an example of simple compact quantum group that does not have property F . A more concrete problem than Problem 5.4 is the following Problem 5.5. Construct simple compact quantum groups with noncommutative repre- sentation ring. For the problems in this section, results on general structure of compact quantum groups such as those in the previous sections should also be useful. In this direction, we have the following result (cf. [80]). Theorem 5.6. Let G be a compact quantum group with property F . Then its quantum subgroups and quotient groups G/N by normal quantum subgroups N also have property F . It would be of interest to develop other general results on the structure of compact quantum groups. 6. Quantum automorphism groups of finite graphs and easy quantum groups How and where do we find quantum groups satisfying the properties in the problems in §5 above? The most natural approach, in our opinion, is by considering quantum automorphism groups of appropriate commutative and noncommutative spaces and their quantum subgroups. Much of the recent work on compact quantum groups falls into this category. We would like to mention three classes of these quantum groups: quantum automorphism groups of finite graphs and other quantum subgroups of the quantum 2Alexandru Chirvasitu informed the author that he has since solved Problem 5.4 and Problem 5.5 - see footnote after Problem 4.1. CLASSIFYING SIMPLE QUANTUM GROUPS 17 permutation groups, easy quantum groups and quantum isometry groups. The last of these three is discussed in the next section. We briefly look at the first two in this section. In part to understand quantum subgroups of the quantum permutation groups [75], Bichon [32] constructed the quantum automorphism groups of finite graphs, which are quantum subgroups of the former preserving the edges of the graphs. To further under- standing this new class of quantum groups, Bichon [33] also defined free wreath product of compact quantum groups using the quantum permutation groups and proved the beau- tiful formula stating that the free wreath product of the quantum automorphism group of a graph by the quantum permutation group Aaut(Xn) is the quantum automorphism group of n disjoint copies of the graph. Banica also independently studied quantum sub- groups of the quantum permutation groups [5, 6]. These works lead to a great deal of further studies of quantum automorphism of finite graphs and quantum subgroups of the quantum permutation groups, cf. [5] - [12] and references therein. It is instructive to see an immediate application of the quantum automorphism groups of finite graphs to related problems in the last section. Clearly, a quantum quotient group G/N of an almost classical quantum group G is also almost classical. However a crucial observation is that a quantum subgroup of an almost classical quantum group needs not be almost classical. For example, the quantum permutation groups Aaut(Xn) are almost classical (cf. [4, 75, 79] and §3), but according to of Bichon [33], their quantum subgroups A2(Z/mZ), as the quantum authomorphism group of certain graphs, are not almost classical if m ≥ 3 (see Corollary 2.7 and the paragraph following Corollary 4.3 of [33]). Though the quantum groups A2(Z/mZ) are not simple (see proposition 2.6 of [33]), and noting that the quantum automorphism group of the trivial graph (i.e. the quantum permutation group) is simple, it is natural to expect that it is possible to obtain simple compact quantum groups that are not almost classical by considering quantum automorphism groups of other appropriate finite graphs, including other free wreath products, thus solving Problem 5.4. Note that since the free wreath is constructed from the free product, such problems are related to see Problem 4.1 and the discussions following it. In another important and related new direction that has origins in works on free wreath product and quantum subgroups of the quantum permutation group discussed above, Banica and Speicher [21] initiated the study of easy quantum groups and found several interesting families of new compact quantum groups. A compact matrix quantum group G with fundamental representation u is called easy if (1) G lies between Sn and O+(n) (cf. §3 on Bu(Q) with Q = In); and 18 SHUZHOU WANG (2) For any k, l ≥ 0, Hom(u⊗k, u⊗l) is linearly generated by operators Tp canonically associated with partitions p of k + l. See [21] or [17] or [81] for a description of Tp. In addition to the six families of free easy quantum groups (also called orthogonal + of free easy quantum groups) in [21, 17], Weber also found another new family B′ n quantum groups in [81], where we follow his different notation from [21]. In [56], Raum computed the fusion rings of several easy quantum groups in [21] using free products. In works of tour de force, Banica and Vergnioux computed the fusion rings of the quantum reflection group H s n in [22] and the half-librated orthogonal quantum group O∗(n) in [23]. It would be interesting to see if any of these quantum groups provide solutions to some of the problems in this section. As the fusion rings of H s n and O∗(n) are non-commutative, the following problems seem to be most appealing: Problem 6.1. (a) Is the quantum reflection group H s n simple? (b) Is the half-librated orthogonal quantum group O∗(n) simple? A positive answer to either (a) or (b) would provide the first simple compact quantum group that is not almost classical and has noncommutative representation ring. In the light of Theorem 4.1 in Raum [56] and 3.1 and 3.2 in Weber [81] whose notation we follow, the relevant quantum groups considered in that theorem are either not simple or rely on solution of problems in §4. For instance, using their results above along with our Theorem 3.3, we see that n is simple for n ≥ 3; + (n ≥ 3) and S ′ n + (n ≥ 4) are simple modulo two components (i.e. discon- (1) B+ (2) B′ n nected); (3) B#+ n is not simple because of Proposition 4.2, but it seems not far from being simple (a concept to be precised) because of 4.1.(3) in [56] and 3.2.(a) in [81]. It is not clear if the quantum groups H + free wreath product, the quantum group H + n, H (s) n and H [s] n , H ∗ n are simple. Note that as a n is related to the general question Problem 4.5. 7. Quantum isometry groups of commutative and noncommutative Riemannian spaces After Banica's initial investigation of quantum isometry groups of finite spaces [5, 6], a recent conceptual breakthrough in compact quantum groups is Debashish Goswami's [44] theory of quantum isometry groups of spectral triples `a la Connes [36], where the universal quantum groups Au(Q) plays an essential role in the proof of existence. Using CLASSIFYING SIMPLE QUANTUM GROUPS 19 this notion he computed with Bhowmick [27, 24] several examples of quantum isometry groups and found that they are either isomorphic to classical isometry groups or are among the examples studied earlier by Rieffel and the author [58, 73]. Using the same circle of ideas they subsequently developed [28] an improved notion of quantum isometry group without relying on existence of a good Laplacian as required in [44]. The quantum isometry groups computed so far are either classical groups, or known quantum groups, or combinations of both based on free product or tensor product. Be- cause of this, Goswami made a rigidity conjecture earlier stating to the effect that con- nected spaces do not admit non-trivial quantum symmetries. On the other hand, Huichi Huang [46] has shown that connected non-smooth metric spaces admit faithful action even by quantum permutation groups, disproving this earlier conjecture. Most striking of all is that Goswami, along with his collaborators, have recently shown in a series of pa- pers [45, 39] a modified rigidity result stating that a connected and oriented Riemannian manifold does not have quantum symmetries other than the classical ones. As a fundamental new concept, one would naturally wonder if fundamentally new com- pact quantum groups can be constructed using quantum isometry groups. In the light of results of Huang and Goswami et al. above, one should look at quantum isometry groups of non-smooth metric spaces or disconnected Riemannian manifolds that are non-classical (i.e. not compact groups). It would be interesting to find if any such quantum groups are simple: Problem 7.1. Construct examples of simple quantum isometry groups of non-smooth metric spaces or disconnected Riemannian manifolds. It is conceivable that quantum isometry groups of disconnected Riemannian manifolds will be related to quantum subgroups of the quantum permutation group as considered in §6, since quantum permutation group can permute the connected components in quantum manner just as it does on the finite space. Another direction of research in the theory of quantum isometry group is the following. In the newly developed notion of quantum isometry group in [28], the quantum groups in the categories Q′ R and Q′ that are used to define the quantum isometry group there does not carry a C ∗-algebraic action on the spectral triple, and as a result, the universal object (i.e, the quantum isometry group) does not carry a C ∗-algebraic action in general. See [30] for an example of this situation. (We refer the reader to the above cited papers for detailed description of Q′ R and Q′ due to space limitation.) 20 SHUZHOU WANG To address this problem, we believe the categories Q′ R and Q′ are too large, and propose the following alternative for the notion of quantum isometry groups that will always carry C ∗-algebraic action. First, by a compact quantum transformation group (A, α, u) of a compact type [75]) spectral triple (B, H, D) we mean a compact quantum transformation group (cf. (A, α) of B that satisfies (QT1) There is a unitary representation u ∈ L(H ⊗ AG) of GA on H such that (ρ, u) is α-covariant: (ρ ⊗ 1)α(b) = u(ρ(b) ⊗ 1)u∗, b ∈ B; (QT2) D is an intertwiner of u with itself: (D ⊗ 1)u = u(D ⊗ 1). A morphism from another quantum transformation group ( A, α, u) to (A, α, u) is defined to be a morphism π from ( A, α) to (A, α) (cf. [75]) that satisfies u = (idH ⊗ π)u. The above defines category C of compact quantum transformation groups of (B, H, D) with objects {(A, α, u)} and morphisms {π}. Similar to [75], the quantum isometry group of (B, H, D) is defined to be a universal object of category C if it exists. As in [75], universal object does not always exist in C in general. Therefore, as in [75] and [28], consider measured spectral triple (B, H, D, φ) with a (usually positive) functional φ on B and consider the category Cφ of compact quantum transformation groups that satisfies (QT1)-(QT2) and (QT3) (φ ⊗ id)(α(b)) = φ(b)1A for b ∈ B The quantum isometry group of (B, H, D, φ) is defined to be the universal object of category Cφ if it exists. The following can be proved and justifies in part the above notion quantum isometry group. Proposition 7.2. (1) Let (B, H, D) be the spectral triple associated with a compact Rie- mannian manifold M. Then the universal object in the category of compact transformation groups of (B, H, D) in the sense above is the isometry group of M. (2) For an arbitrary spectral triple (B, H, D), the universal object in the category of compact transformation groups of (B, H, D) in the sense above is the compact group of isometries of (B, H, D) in the sense of Connes (see p6200 of [37]). Since composition of two continuous maps is continuous, the quantum isometry group defined above is evidently contained in the quantum isometry groups defined by Goswami [44] and Bhowmick and Goswami [28]. Because of this, and using ideas in [75] and [28], the following seems to be quite plausible: Problem 7.3. For a measured spectral triple (B, H, D, φ), show that universal object exists in the category Cφ. CLASSIFYING SIMPLE QUANTUM GROUPS 21 Many other problems naturally arise: (i) Calculate the quantum isometry groups in the sense above for the classical spaces such as the flat spheres, tori, and other Riemannian spaces; (ii) Calculate the quantum isometry groups in the sense above for the noncommutative tori; (iii) Study the properties of quantum isometry groups in general and apply them to study other properties of noncommutative spaces; (iv) Construct simple compact quantum groups through the study of quantum isometry groups. As in the definitions of quantum isometry groups by Goswami and Bhowmick, and because a quantum isometry group in our sense is contained in theirs, examples of simple quantum isometry groups in our sense might need to be constructed out of non-smooth metric spaces or disconnected Riemannian manifolds, as in Problem 7.1. Acknowledgments. The author would like to thank Hanfeng Li and Yi-Jun Yao for invitations to Chongqing University and Fudan University respectively during the summer of 2012, where the author had the opportunity to work on this paper. The author would also like to thank Hanfeng Li and Dechao Zheng at Chongqing University, and Yi-Jun Yao and Guoliang Yu at Fudan University for their hospitality and for creating a congenial atmosphere. He also wish to record his thanks to Huichi Huang, Pan Ma and Qinggang Ren for making his visit enjoyable. The author is indebted to Professor Marek Bozejko for drawing his attention to easy quantum groups during the conference in September 2011 celebrating Professor Woronowicz's seventieth birthday. References [1] Baaj, S. and Skandalis, G.: Unitaires multiplicatifs et dualit´e pour les produits crois´es de C ∗-alg`ebres, Ann. Sci. Ec. Norm. Sup. 26 (1993), 425-488. [2] Banica, T.: Th´eorie des repr´esentations du groupe quantique compact libre O(n), C. R. Acad. Sci. Paris t. 322, Serie I (1996), 241-244. [3] Banica, T.: Le groupe quantique compact libre U (n), Commun. Math. Phys. 190 (1997), 143-172. [4] Banica, T.: Symmetries of a generic coaction. Math. Ann. 314 (1999), 763-780. [5] Banica, T.: Quantum automorphism groups of homogeneous graphs. J. Funct. Anal. 224 (2005), no. 2, 243 -- 280. [6] Banica, T.: Quantum automorphism groups of small metric spaces. Pacific J. Math. 219 (2005), no. 1, 27 -- 51. [7] Banica, T., Bichon, J.: Free product formulae for quantum permutation groups. J. Inst. Math. Jussieu 6 (2007), no. 3, 381 -- 414. 22 SHUZHOU WANG [8] Banica, T., Bichon, J.: Quantum automorphism groups of vertex-transitive graphs of order ≤ 11. J. Algebraic Combin. 26 (2007), no. 1, 83 -- 105. [9] Banica, T., Bichon, J.: Quantum groups acting on 4 points, J. Reine Angew. Math., 626 (2009), 74-114. [10] Banica, T., Bichon, J., Chenevier, G.: Graphs having no quantum symmetry. Ann. Inst. Fourier (Grenoble) 57 (2007), no. 3, 955 -- 971. [11] Banica, T., Bichon, J., Collins, B.: The hyperoctahedral quantum group. J. Ramanujan Math. Soc. 22 (2007), no. 4, 345 -- 384. [12] Banica, T., Bichon, J., Collins, B.: Quantum permutation groups: a survey, Banach Center Publi- cations 78, 13-34, 2007. [13] Banica, T., Bichon, J., Natale, S.: Finite quantum groups and quantum permutation groups. Adv. Math. 229 (2012), 3320-3338 [14] Banica, T., Collins, B.: Integration over quantum permutation groups. J. Funct. Anal. 242 (2007), no. 2, 641 -- 657. [15] Banica, T., Collins, B.: Integration over compact quantum groups. Publ. Res. Inst. Math. Sci. 43 (2007), no. 2, 277 -- 302. [16] Banica, T., Collins, B.: Integration over the Pauli quantum group. J. Geom. Phys. 58 (2008), 942-961. [17] Banica, T., Curran, S., Speicher, R.: Classification results for easy quantum groups. Pacific J. Math. 247 (2010), no. 1, 1-26. [18] Banica, T., Curran, S., Speicher, R.: Stochastic aspects of easy quantum groups. Probab. Theory Related Fields 149 (2011), 435-462. [19] Banica, T., Curran, S., Speicher, R.: De Finetti theorems for easy quantum groups. Ann. Probab. 40 (2012), 401-435. [20] Banica, T., Moroianu, S.: On the structure of quantum permutation groups. Proc. Amer. Math. Soc. 135 (2007), no. 1, 21 -- 29. [21] Banica, T., Speicher, R.: Liberation of orthogonal Lie groups. Adv. Math. 222 (2009), 1461-1501. [22] Banica, T. and Vergnioux, R.: Fusion rules for quantum reflection groups. J. Noncommut. Geom. 3 (2009), 327-359. [23] Banica, T. and Vergnioux, R.: Invariants of the half-liberated orthogonal group. Ann. Inst. Fourier 60 (2010), 2137-2164. [24] Bhowmick, J.: Quantum Isometry Group of the n tori. Proc. Amer. Math. Soc. 137 (2009), no. 9, 3155-3161. [25] Bhowmick, J., D'Andrea, F., D¸abrowski, L.: Quantum isometries of the finite noncommutative geometry of the standard model. Comm. Math. Phys. 307 (2011), no. 1, 101-131 [26] Bhowmick, J., D'Andrea, F., Das, B., D¸abrowski, L.: Quantum gauge symmetries in Noncommuta- tive Geometry. arXiv:1112.3622 [27] Bhowmick, J. and Goswami, D.: Quantum Isometry Groups: Examples and Computations, Com- mun.Math.Phys. 285 (2009), 421-444, [28] Bhowmick, J. and Goswami, D.: Quantum group of orientation-preserving Riemannian isometries. J. Funct. Anal. 257 (2009), no. 8, 2530-2572. CLASSIFYING SIMPLE QUANTUM GROUPS 23 [29] Bhowmick, J. and Goswami, D.: Quantum isometry groups of the Podles spheres. J. Funct. Anal. 258 (2010), no. 9, 2937-2960. [30] Bhowmick, J., Goswami, D.: Some counterexamples in the theory of quantum isometry groups. Lett. Math. Phys. 93 (2010), no. 3, 279-293. [31] Bhowmick, J., Goswami, D., Skalski, A.: Quantum Isometry Groups of 0- Dimensional Manifolds. Trans. Amer. Math. Soc. 363 (2011), no. 2, 901-921. [32] Bichon, J.: Quantum automorphism groups of finite graphs. Proc. Amer. Math. Soc. 131 (2003), no. 3, 665 -- 673 [33] Bichon, J.: Free wreath product by the quantum permutation group. Algebr. Represent. Theory 7 (2004), 343 -- 362. [34] Bichon, J.: Algebraic quantum permutation groups. Asian-Eur. J. Math. 1 (2008), no. 1, 1 -- 13. [35] Bichon, J., De Rijdt, A., Vaes, S.: Ergodic coactions with large multiplicity and monoidal equivalence of quantum groups. Comm. Math. Phys. 262 (2006), 703 -- 728. [36] Connes, A.: Noncommutative Geometry, Academic Press, 1994. [37] Connes, A.: Noncommutative geometry and reality, J. Math. Phys. 36:11 (1995), 6194-6231 [38] Connes, A.: Gravity coupled with matter and the foundation of non commutative geometry, Com- mun. Math. Phys. 182 (1996), 155-176. [39] Das, B. and Goswami, D and Joardar, S.: Rigidity of action of compact quantum groups II. arXiv:1206.1718 [40] Drinfeld, V.G.: Quantum groups, in Proc. of the ICM-1986, Berkeley, Vol I, Amer. Math. Soc., Providence, R.I., 1987, pp798 -- 820. [41] Enock, M. et Schwartz, J.-M.: Une dualit´e dans les alg`ebres de von Neumann, Bull. Soc. Math. France. Suppl. M´emoire 44 (1975), 1-144. [42] Enock, M. et Schwartz, J.-M.: Kac algebras and duality of locally compact groups. (English summary) With a preface by Alain Connes. With a postface by Adrian Ocneanu. Springer-Verlag, Berlin, 1992. x+257 pp. ISBN: 3-540-54745-2 [43] Fredenhagen, K. and Rehren, K.H. and Schroer, B.: Superselection sectors with braid group statistics and exchange algebras I, Comm. Math. Phys. 125 (1989), 201-226. [44] Goswami, D.: Quantum Group of Isometries in Classical and Noncommutative Geometry, Comm. Math. Phys. 285 (2009), no. 1, 141-160. [45] Goswami, D.: Rigidity of action of compact quantum groups I, III arXiv:1106.5107, arXiv:1207.6470 [46] Huang, Huichi: Faithful compact quantum group actions on connected compact metrizable spaces, arXiv:1202.1175 [47] Kac, G.: Ring groups and the duality principle, I, II Proc. Moscow Math. Soc. 12 (1963), 259-303; ibid. 13 (1965), 84-113. [48] Kac, G. and Vainerman, L. I.: Nonunimodular ring groups and Hopf Von Neumann algebras, Math. USSR Sb. 23 (1974), 185-214. [49] Kostler, C., Speicher, R.: A noncommutative de Finetti theorem: Invariance under quantum permu- tations is equivalent to freeness with amalgamation. Comm. Math. Phys. 291 (2009), no. 2, 473-490. [50] Kustermans, Johan and Vaes, Stefaan Locally compact quantum groups. Ann. Sci. ´Eole Norm. Sup. (4) 33 (2000), no. 6, 837-934. 24 SHUZHOU WANG [51] Levendorskii, S.: Twisted algebra of functions on compact quantum group and their representations, Algebra i analiz 3:2 (1991), 180-198. St. Petersburg Math. J. 3:2 (1992), 405-423. [52] Levendorskii, S. and Soibelman, Y.: Algebra of functions on compact quantum groups, Schubert cells, and quantum tori, Commun. Math. Phys. 139 (1991), 141-170. [53] Maes, Annand Van Daele, Alfons: Notes on compact quantum groups. Nieuw Arch. Wisk. (4) 16 (1998), no. 1-2, 73-112. [54] Masuda, T.; Nakagami, Y.; Woronowicz, S. L. A C ∗-algebraic framework for quantum groups. Internat. J. Math. 14 (2003), no. 9, 903-1001 [55] Podles, P. and Woronowicz S. L.: Quantum deformation of Lorentz group, Commun. Math. Phys. 130 (1990), 381-431. [56] Raum, S.; Isomorphisms and fusion rules of orthogonal free quantum groups and their free complex- ifications. Proc. Amer. Math. Soc. 140 (2012), no. 9, 3207-3218. [57] Rieffel, M.: Deformation quantization for actions of Rn, Memoirs A.M.S. no. 506, 1993. [58] Rieffel, M.: Compact quantum groups associated with toral subgroups, Contemp. Math. 145 (1993), 465-491. [59] Rosso, M.: Comparaison des groupes SU (2) quantiques de Drinfeld et Woronowicz, C. R. Acad. Sci. Paris 304 (1987), 323-326. [60] Rosso, M.: Alg`ebres enveloppantes quantifi´ees, groupes quantiques compacts de matrices et calcul differentiel non-commutatif, Duke Math. J. 61 (1990), 11-40. [61] Soibelman, Y.: Algebra of functions on compact quantum group and its representations, Algebra i analiz 2:1 (1990), 190-212. Leningrad Math. J. 2:1 (1991), 161-178. [62] Soibelman, Y. and Vaksman, L.: The algebra of functions on quantum SU (2), Funct. Anal. ego Pril. 22:3 (1988), 1-14. Funct. Anal. Appl. 22:3 (1988), 170-181. [63] Soltan, P.M.: Quantum families of maps and quantum semigroups on finite quantum spaces. J. Geom. Phys. 59 (2009), no. 3, 354-368. [64] Vaes, S. and Vennet, N.V.: Identification of the Poisson and Martin boundaries of orthogonal discrete quantum groups. Journal of the Institute of Mathematics of Jussieu 7 (2008), 391-412. [65] Vaes, S. and Vennet, N.V.: Poisson boundary of the discrete quantum group \Au(F ). Compos. Math. 146 (2010), no. 4, 1073 -- 1095. [66] Vaes, S. and Vergnioux, R.: The boundary of universal discrete quantum groups, exactness and factoriality. Duke Mathematical Journal 140 (2007), 35 -- 84. [67] Vergnioux, R., Voigt , C.: The K-theory of free quantum groups. arXiv:1112.3291 [68] Voigt, C.: The Baum-Connes conjecture for free orthogonal quantum groups. Adv. Math. 227 (2011), 1873 - 1913. [69] Van Daele, A. and Wang, S.Z.: Universal quantum groups, International J. of Math. 7 no2 (1996), 255-264. [70] Wang, S. Z.: General Constructions of Compact Quantum Groups, Ph.D. Thesis, University of California at Berkeley, March, 1993. [71] Wang, S.Z.: Free products of compact quantum groups, Commun. Math. Phys. 167 (1995), 671-692. [72] Wang, S.Z.: Tensor products and crossed products of compact quantum groups, Proc. London Math. Soc. 71 (1995), 695-720. CLASSIFYING SIMPLE QUANTUM GROUPS 25 [73] Wang, S.Z.: Deformations of compact quantum groups via Rieffel quantization, Commun. Math. Phys. 178:3 (1996), 747-764. [74] Wang, S.Z.: Problems in the theory of quantum groups, in Quantum Groups and Quantum Spaces, Banach Center Publication 40 (1997), Inst. of Math., Polish Acad. Sci., Editors: R. Budzynski, W. Pusz, and S. Zakrzewski. pp67-78 [75] Wang, S.Z.: Quantum symmetry groups of finite spaces, Commun. Math. Phys. 195:1 (1998), 195- 211. [76] Wang, S.Z.: Ergodic actions of universal quantum groups on operator algebras, Commun. Math. Phys. 203 (1999), 481-498. [77] Wang, S.Z.: Classification of quantum groups SUq(n), J. London Math. Soc. 59:2 (1999), 669-680 [78] Wang, S.Z.: Structure and isomorphism classification of quantum groups Au(Q) and Bu(Q), Journal of Operator Theory 48(3) (2002), 573-583 [79] Wang, S.Z.: Simple compact quantum groups I, J. Funct. Anal. 256 (2009), no. 10, 3313-3341. [80] Wang, S.Z.: Equivalent notions of normal quantum subgroups, compact quantum groups with prop- erties F and F D, and other applications, preprint. [81] Weber, M.: On the classification of easy quantum groups - The nonhyperoctahedral and the half- liberated case. arXiv:1201.4723 [82] Woronowicz, S. L.: Twisted SU (2) group. An example of noncommutative differential calculus, Publ. RIMS, Kyoto Univ. 23 (1987), 117-181. [83] Woronowicz, S. L.: Compact matrix pseudogroups, Commun. Math. Phys. 111 (1987), 613-665. [84] Woronowicz, S. L.: Tannaka-Krein duality for compact matrix pseudogroups. Twisted SU (N ) groups, Invent. Math. 93 (1988), 35-76. [85] Woronowicz, S. L.: Compact quantum groups, in Quantum Symmetries, Les Houches Summer School-1995, Session LXIV, A. Connes, K. Gawedzki, J. Zinn-Justin, ed., Elsevier Science, Ams- terdam, 1998; pp845-884. Department of Mathematics, University of Georgia, Athens, GA 30602, USA Fax: 706-542-5907, Tel: 706-542-0884 E-mail address: [email protected]
1006.1021	1	1006	2010-06-05T05:33:25	A Gruss inequality for n-positive linear maps	[ "math.OA" ]	Let $\mathscr{A}$ be a unital $C^$-algebra and let $\Phi: \mathscr{A} \to {\mathbb B}({\mathscr H})$ be a unital $n$-positive linear map between $C^$-algebras for some $n \geq 3$. We show that $$\\|\Phi(AB)-\Phi(A)\Phi(B)\\| \leq \Delta(A,\|\|\cdot\|\|)\,\Delta(B,\|\|\cdot\|\|)$$ for all operators $A, B \in \mathscr{A}$, where $\Delta(C,\\|\cdot\\|)$ denotes the operator norm distance of $C$ from the scalar operators.	math.OA	math	A GR USS INEQUALITY FOR n-POSITIVE LINEAR MAPS MOHAMMAD SAL MOSLEHIAN1 AND RAJNA RAJI ´C2 Abstract. Let A be a unital C ∗-algebra and let Φ : A → B(H ) be a unital n-positive linear map between C ∗-algebras for some n ≥ 3. We show that kΦ(AB) − Φ(A)Φ(B)k ≤ ∆(A, · ) ∆(B, · ) for all operators A, B ∈ A , where ∆(C, k · k) denotes the operator norm distance of C from the scalar operators. 0 1 0 2 n u J 5 ] . A O h t a m [ 1 v 1 2 0 1 . 6 0 0 1 : v i X r a 1. Introduction Let B(H ) stand for the algebra of all bounded linear operators on a complex Hilbert space (H , h·, ·i), let k · k denote the operator norm and let I be the identity operator. For self-adjoint operators A, B the order relation A ≤ B means that hAξ, ξi ≤ hBξ, ξi (ξ ∈ H ). In particular, if 0 ≤ A, then A is called positive. If dimH = k, we identify B(H ) with the algebra Mk of all k × k matrices with entries in C. Let ∆(C, k · k) = inf λ∈C kC − λIk be the k · k-distance of C from the scalar operators. It is known that ∆(C, k · k) ≤ kCk and ∆(C, k · k) = c(C) for any normal operator C, where c(C) denotes the radius of the smallest disk in the complex plane containing the spectrum σ(C) of C; see [11]. A linear map Φ : A → B between C ∗-algebras is said to be positive if Φ(A) ≥ 0 whenever A ≥ 0. Every positive linear map Φ satisfies Φ(A∗) = Φ(A)∗ for all A. We say that Φ is unital if A , B are unital C ∗-algebras and Φ preserves the identity. A linear map Φ is called n- positive if the map Φn : Mn(A ) → Mn(B) defined by Φn([aij]n×n) = [Φ(aij)]n×n is positive, where Mn(A ) stands for the C ∗-algebra of n × n matrices with entries in A . It is known that kΦk = 1 for any unital n-positive linear map Φ. Φ is said to be completely positive if it is n-positive for every n ∈ N. For a comprehensive account on completely positive maps see [9]. The Gruss inequality [6], as a complement of Chebyshevs inequality, states that if f and g are integrable real functions on [a, b] and there exist real constants ϕ, φ, γ, Γ such that 2010 Mathematics Subject Classification. Primary 46L07; secondary 47A12, 47L25. Key words and phrases. Gruss inequality, operator inequality, positive operator, C ∗-algebra, completely positive map, n-positive map. 1 2 M.S. MOSLEHIAN, R. RAJI ´C ϕ ≤ f (x) ≤ φ and γ ≤ g(x) ≤ Γ hold for all x ∈ [a, b], then 1 b − aZ b a (cid:12)(cid:12)(cid:12)(cid:12) g(x)dx(cid:12)(cid:12)(cid:12)(cid:12) f (x)g(x)dx − 1 (b − a)2 Z b a f (x)dxZ b a ≤ 1 4 (φ − ϕ)(Γ − γ) . This inequality has been investigated, applied and generalized by many mathematicians in different areas of mathematics; see [5] and references therein. Peri´c and Raji´c proved a Gruss type inequality for unital completely bounded maps [10]. In what follows A will stand for a unital C ∗-algebra. In this paper we prove that if Φ : A → B(H ) is a unital n-positive linear map between C ∗-algebras for some n ≥ 3, then kΦ(AB) − Φ(A)Φ(B)k ≤ ∆(A, k · k) ∆(B, k · k) (1.1) for all operators A, B ∈ A . 2. Main result To achieve our main result we need three lemmas. The first lemma can be deduced from [8, Theorem 2] by adding a necessary assumption Φ(A∗) = Φ(A)∗. We state it for the sake of convenience. Lemma 2.1. Let Φ : A → B(H ) be a unital ∗-n-positive (not necessarily linear ) map for some n ≥ 3. Then kΦ(AB) − Φ(A)Φ(B)k2 ≤ kΦ(A∗2) − Φ(A∗)2k kΦ(B2) − Φ(B)2k (2.1) for all operators A, B ∈ A . Proof. Let A and B be two operators in A . We have 0 ≤ A∗ B ∗ I 0 ... 0     h A B I 0 · · · 0 i =   A∗A A∗B A∗ 0 · · · B ∗A B ∗B B ∗ 0 · · · 0 · · · ... . . . 0 · · · B ... 0 A ... 0 I ... 0 0 0 0 ... 0 .   Due to Φ is n-positive, we get Φ(A) Φ(A∗A) Φ(A∗B) Φ(A∗) 0 · · · Φ(B ∗A) Φ(B ∗B) Φ(B ∗) 0 · · · 0 · · · ... . . . 0 · · · Φ(B) Φ(I) ... 0 ... 0 ... 0 0 ≤   , 0 0 0 ... 0   A GR USS INEQUALITY FOR n-POSITIVE LINEAR MAPS 3 whence 0 ≤ . (2.2) Φ(A∗A) Φ(A∗B) Φ(B ∗A) Φ(B ∗B) Φ(A)∗ Φ(B)∗ Φ(A) Φ(B) I     It is known that the matrix " R T T ∗ S # is positive if and only if R, S are positive and R ≥ T S −1T ∗, where S −1 denotes the (generalized) inverse of S. Using this fact and noting to (2.2) we get " Φ(A∗A) Φ(A∗B) Φ(B ∗A) Φ(B ∗B) # ≥" Φ(A)∗ Φ(B)∗ # I −1 h Φ(A) Φ(B) i =" Φ(A)∗Φ(A) Φ(A)∗Φ(B) Φ(B)∗Φ(A) Φ(B)∗Φ(B) # or equivalently " Φ(A∗A) − Φ(A)∗Φ(A) Φ(A∗B) − Φ(A)∗Φ(B) Φ(B ∗A) − Φ(B)∗Φ(A) Φ(B ∗B) − Φ(B)∗Φ(B) # ≥ 0 As noted in [2], the inequality (2.3) implies that kΦ(A∗B) − Φ(A)∗Φ(B)k2 ≤ kΦ(A2) − Φ(A)2k kΦ(B2) − Φ(B)2k Replacing A by A∗ in (2.4) we obtain (2.1). (2.3) (2.4) (cid:3) Remark 2.2. The inequality (2.3) is known as the operator covariance-variance inequality; see [1] and references therein for more information. The second lemma includes our main idea. Lemma 2.3. Let Φ : A → B(H ) be a unital positive linear map between C ∗-algebras. Then kΦ(A∗A) − Φ(A)∗Φ(A)k ≤ ∆(A, k · k)2 (2.5) for every normal operator A ∈ A . Proof. Let A ∈ A be a normal operator, and let C denote the C ∗-algebra generated by A, A∗ and I. Then C is commutative, so that the restriction of Φ to C is completely positive by a known fact due to Choi; see [4, 9]. The Stinespring dilation theorem states that for any unital completely positive map Φ : C → B(H ) there exist a Hilbert space K , an isometry 4 M.S. MOSLEHIAN, R. RAJI ´C V : H → K and a unital ∗-homomorphism π : C → B(K ) such that Φ(A) = V ∗π(A)V . Now we have kΦ(A∗A) − Φ(A)∗Φ(A)k = kΦ(cid:0)(A − λI)∗(A − µI)(cid:1) − (Φ(A − λI))∗Φ(A − µI)k = kV ∗π(cid:0)(A − λI)∗(A − µI)(cid:1)V − V ∗(π(A − λI))∗V V ∗π(A − µI)V k = kV ∗(π(A − λI))∗(I − V V ∗)π(A − µI)V k (π is ∗-homomorphism) ≤ kπ(A − λI)k kπ(A − µI)k (I − V V ∗ is a projection) ≤ kA − λIk kA − µIk (π is unital and norm decreasing) for all λ, µ ∈ C. From this it follows that kΦ(A∗A) − Φ(A)∗Φ(A)k ≤ inf λ∈C kA − λIk inf µ∈C kA − µIk = ∆(A, k · k)2. (cid:3) Remark 2.4. Passing the proof of previous lemma, one can easily deduce that in the case of a unital completely positive map Φ : A → B(H ) the statement of Lemma 2.3 is valid for every A ∈ A . The next lemma is well known. Lemma 2.5. [7, Theorem 1] Let A ∈ A and kAk < 1 − (2/m) for some integer m greater than 2. Then there are m unitary elements U1, . . . , Um ∈ A such that mA = U1 + · · · + Um. We are ready to present our main result. Theorem 2.6. Let Φ : A → B(H ) be a unital n-positive linear map between C ∗-algebras for some n ≥ 3. Then kΦ(AB) − Φ(A)Φ(B)k ≤ ∆(A, k · k) ∆(B, k · k) (2.6) for all operators A, B ∈ A . Proof. Using Lemmas 2.1 and 2.3 we obtain (2.6) for normal operators A, B ∈ A . Let 0 6= A ∈ A . Let m > 2 be an integer and M = m2+2 m2−2m kAk. Then kA/Mk = m2 − 2m m2 + 2 < 1 − 2 m . A GR USS INEQUALITY FOR n-POSITIVE LINEAR MAPS 5 By Lemma 2.5, there are unitaries U1, . . . , Um ∈ A such that A = M any normal operator B ∈ A we have j=1 Uj. Hence for m Pm kΦ(AB) − Φ(A)Φ(B)k = ≤ ≤ ≤ Uj! Φ(B)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kΦ(UjB) − Φ(Uj)Φ(B)k (by (2.6) for normal operators) M m (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Φ m Xj=1 m2 + 2 m3 − 2m2 kAk m UjB! − Φ m Xj=1 Xj=1 Xj=1 kUjkkBk m m2 + 2 m3 − 2m2 kAk m2 + 2 m2 − 2m kAk kBk. Letting m → ∞, we infer that kΦ(AB) − Φ(A)Φ(B)k ≤ kAk kBk (2.7) for arbitrary A and normal B. By repeating the same argument for arbitrary B and by using (2.7), we deduce that kΦ(AB) − Φ(A)Φ(B)k ≤ kAk kBk for all A, B ∈ A . Next we observe that kΦ(AB) − Φ(A)Φ(B)k = kΦ(cid:0)(A − λI)(B − µI)(cid:1) − Φ(A − λI)Φ(B − µI)k ≤ kA − λIk kB − µIk for all λ, µ ∈ C. Thus, kΦ(AB) − Φ(A)Φ(B)k ≤ inf λ∈C kA − λIk inf µ∈C kB − µIk = ∆(A, k · k) ∆(B, k · k) for all A, B ∈ A . This proves the theorem. (cid:3) Remark 2.7. Let us remark here that the inequality (2.6) does not have to hold if Φ is assumed only to be a unital positive linear map. To see this, let us choose Φ : M2 → M2 to be the transpose map. It is known (see e.g. [9]) that such a map is positive, but not 2-positive. Hence, Φ is not a 3-positive map. Then, for matrices A =" 1 2 0 4 # 2 4 # , B =" 1 0 6 we have M.S. MOSLEHIAN, R. RAJI ´C kΦ(AB) − Φ(A)Φ(B)k = 6. Since A and B are positive matrices such that σ(A) = {0, 5} and σ(B) = {1, 4}, we conclude that ∆(A, k · k) = c(A) = 5 2 and ∆(B, k · k) = c(B) = 3 2. Therefore, kΦ(AB) − Φ(A)Φ(B)k = 6 > 15 4 = ∆(A, k · k) ∆(B, k · k). We do not know whether (2.6) holds if a map Φ is assumed to be 2-positive. Theorem 2.6 extends the result obtained in [10, Corollary 1] to the case of n-positive linear maps, where n ≥ 3. Some applications of Theorem 2.6 on completely positive maps were also given in [10]. The first example of n-positive map (n ≥ 2), which is not completely positive, was obtained by Choi in [3]. He showed that the map Φ : Mk → Mk defined by Φ(T ) = (k − 1)tr(T )I − T (T ∈ Mk) is (k − 1)-positive, but not k-positive. (Here 'tr' denotes the trace.) Later, Takasaki and Tomiyama [12] introduced a way to construct new examples of (k − 1)-positive linear maps from Mk to Mk which are not k-positive. In the next corollary we apply Theorem 2.6 on Choi's (k − 1)-positive linear map. Corollary 2.8. Let A, B ∈ Mk, where k ≥ 4. Then k(k2 − k − 1)tr(AB)I − kAB − (k − 1)tr(A)tr(B)I + tr(B)A + tr(A)Bk ≤ (k2 − k − 1)2 k − 1 ∆(A, k · k) ∆(B, k · k). Proof. Define Φ(T ) = k − 1 k2 − k − 1 tr(T )I − 1 k2 − k − 1 T (T ∈ Mk). By Theorem 2.6 we have kΦ(AB) − Φ(A)Φ(B)k ≤ ∆(A, k · k) ∆(B, k · k) since Φ is a unital (k − 1)-positive linear map. An easy computation shows that Φ(AB) − Φ(A)Φ(B) = k − 1 (k2 − k − 1)2 [(k2 − k − 1)tr(AB)I − kAB − (k − 1)tr(A)tr(B)I + tr(B)A + tr(A)B], from which the result immediately follows. (cid:3) A GR USS INEQUALITY FOR n-POSITIVE LINEAR MAPS 7 References [1] Lj. Arambasi´c, D. Baki´c, M.S. Moslehian, Gram matrix in inner product modules over C ∗-algebras, arXiv:0905.3509v2. [2] R. Bhatia, C. Davis, More operator versions of the Schwarz inequality, Comm. Math. Phys. 215 (2000), no. 2, 239 -- 244. [3] M.D. Choi, Positive linear maps on C ∗-algebras, Can. J. Math. 24 (1972), 520 -- 529. [4] M.D. Choi, A Schwarz inequality for positive linear maps on C ∗-algebras, Illinois J. Math. 18 (1974), 565 -- 574. [5] S.S. Dragomir, Advances in inequalities of the Schwarz, Gruss and Bessel type in inner product spaces, Nova Science Publishers, Inc., Hauppauge, NY, 2005. [6] G. Gruss, Uber das Maximum des absoluten Betrages von a f (x)g(x)dx − 1 b−aR b 1 (b−a)2 R b a f (x)dxR b a g(x)dx, Math. Z. 39 (1935), 215 -- 226. [7] R.V. Kadison, G.K. Pedersen, Means and convex combinations of unitary operators, Math. Scan. 57 (1985), 249 -- 266. [8] R. Mathias, A note on: "More operator versions of the Schwarz inequality" [Comm. Math. Phys. 215 (2000 ), no. 2, 239 -- 244; by R. Bhatia and C. Davis], Positivity 8 (2004), no. 1, 85 -- 87. [9] V. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Advanced Mathe- matics, 78. Cambridge University Press, Cambridge, 2002. [10] I. Peri´c, R. Raji´c, Gruss inequality for completely bounded maps, Linear Algebra Appl. 390 (2004), 287 -- 292. [11] J.G. Stampfli, The norm of a derivation, Pacific J. Math. 33 (1970), 737 -- 747. [12] T. Takasaki, J. Tomiyama, On the geometry of positive maps in matrix algebras, Math. Z. 184 (1983), 101 -- 108. 1 Department of Pure Mathematics, Center of Excellence in Analysis on Algebraic Structures (CEAAS), Ferdowsi University of Mashhad, P.O. Box 1159, Mashhad 91775, Iran. E-mail address: [email protected] and [email protected] 2 Faculty of Mining, Geology and Petroleum Engineering, University of Zagreb, Pierot- tijeva 6, 10000 Zagreb, Croatia E-mail address: [email protected]
1710.02409	3	1710	2019-05-30T16:18:30	Recovery map stability for the Data Processing Inequality	[ "math.OA", "quant-ph" ]	The Data Processing Inequality (DPI) says that the Umegaki relative entropy $S(\rho\|\|\sigma) := {\rm Tr}[\rho(\log \rho - \log \sigma)]$ is non-increasing under the action of completely positive trace preserving (CPTP) maps. Let ${\mathcal M}$ be a finite dimensional von Neumann algebra and ${\mathcal N}$ a von Neumann subalgebra if it. Let ${\mathcal E}_\tau$ be the tracial conditional expectation from ${\mathcal M}$ onto ${\mathcal N}$. For density matrices $\rho$ and $\sigma$ in ${\mathcal N}$, let $\rho_{\mathcal N} := {\mathcal E}_\tau \rho$ and $\sigma_{\mathcal N} := {\mathcal E}_\tau \sigma$. Since ${\mathcal E}_\tau$ is CPTP, the DPI says that $S(\rho\|\|\sigma) \geq S(\rho_{\mathcal N}\|\|\sigma_{\mathcal N})$, and the general case is readily deduced from this. A theorem of Petz says that there is equality if and only if $\sigma = {\mathcal R}_\rho(\sigma_{\mathcal N} )$, where ${\mathcal R}_\rho$ is the Petz recovery map, which is dual to the Accardi-Cecchini coarse graining operator ${\mathcal A}_\rho$ from ${\mathcal M} $ to ${\mathcal N} $. In it simplest form, our bound is $$S(\rho\|\|\sigma) - S(\rho_{\mathcal N} \|\|\sigma_{\mathcal N} ) \geq \left(\frac{1}{8\pi}\right)^{4} \\|\Delta_{\sigma,\rho}\\|^{-2} \\| {\mathcal R}_{\rho_{\mathcal N}} -\sigma\\|_1^4 $$ where $\Delta_{\sigma,\rho}$ is the relative modular operator. We also prove related results for various quasi-relative entropies. Explicitly describing the solutions set of the Petz equation $\sigma = {\mathcal R}_\rho(\sigma_{\mathcal N} )$ amounts to determining the set of fixed points of the Accardi-Cecchini coarse graining map. Building on previous work, we provide a throughly detailed description of the set of solutions of the Petz equation, and obtain all of our results in a simple self, contained manner.	math.OA	math	Recovery map stability for the Data Processing Inequality Eric A. Carlen1 and Anna Vershynina2 1Department of Mathematics, Hill Center, Rutgers University, 110 Frelinghuysen Road, Piscataway, NJ 2Department of Mathematics, Philip Guthrie Hoffman Hall, University of Houston, 3551 Cullen Blvd., 08854-8019, USA Houston, TX 77204-3008, USA May 31, 2019 Abstract The Data Processing Inequality (DPI) says that the Umegaki relative entropy S(ρσ) := Tr[ρ(log ρ− log σ)] is non-increasing under the action of completely positive trace preserving (CPTP) maps. Let M be a finite dimensional von Neumann algebra and N a von Neumann subalgebra of it. Let Eτ be the tracial conditional expectation from M onto N . For density matrices ρ and σ in M, let ρN := Eτ ρ and σN := Eτ σ. Since Eτ is CPTP, the DPI says that S(ρσ) ≥ S(ρNσN ), and the general case is readily deduced from this. A theorem of Petz says that there is equality if and only if σ = Rρ(σN ), where Rρ is the Petz recovery map, which is dual to the Accardi-Cecchini coarse graining operator Aρ from M to N . We prove a quantitative version of Peta's theorem. In it simplest form, our bound is 8(cid:17)4 S(ρσ) − S(ρNσN ) ≥(cid:16) π k∆σ,ρk−2kRρ(σN ) − σk4 1 . where ∆σ,ρ is the relative modular operator. Since k∆σ,ρk ≤ kρ−1k, this bound implies a bound with a constant that is independent of σ. We also prove an analogous result with a more complicated constant in which the roles of ρ and σ are interchanged on the right. Explicitly describing the solutions set of the Petz equation σ = Rρ(σN ) amounts to determining the set of fixed points of the Accardi-Cecchini coarse graining map. Building on previous work, we provide a throughly detailed description of the set of solutions of the Petz equation Rρ(Eτ γ) = γ, and obtain all of our results in a simple, self contained manner. Finally, we prove a theorem characterizing state ρ for which the orthogonal projection from M onto N in the GNS inner product is a conditional expectation. 1 2 1 Introduction 1.1 The Data Processing Inequality Let M be a finite dimensional von Neumann algebra, which we may regard as a subalgebra of Mn(C), the n× n complex matrices. The Hilbert-Schmidt inner product h·,·iHS on Mn(C) is given in terms of the trace by hX, Y iHS = Tr[X ∗Y ]. Let 1 denote the identity. A state on M is a linear functional ϕ on M such that ϕ(A∗A) ≥ 0 for A ∈ M and such that ϕ(1) = 1. A state ϕ is faithful in case ϕ(A∗A) > 0 whenever A 6= 0, and is tracial in case ϕ(AB) = ϕ(BA) for all A, B ∈ M. Every state on M is of the form X 7→ Tr[ρX], where ρ is a density matrix in M; i.e., a non-negative element ρ of M such that Tr[ρ] = 1. This state is faithful if and only if ρ is invertible. It will be convenient to write ρ(X) = Tr[ρX] to denote the state corresponding to a density matrix ρ. Given a faithful state ρ, the corresponding Gelfand- Naimark-Segal (GNS) inner product is given by hX, Y iGN S,ρ := ρ(X ∗Y ). one whose density matrix is n−11. The symbol τ is reserved throughout for this tracial state. In this finite dimensional setting, there is always a faithful tracial state τ on M, namely the Let N be a von Neumann subalgebra of M. Let E be any norm-contractive projection from M onto N . (Norm contractive means that kE (X)k ≤ kXk for all X ∈ M. Throughout the paper, k · k without any subscript denotes the operator norm.) By a theorem of Tomiyama [37], E preserves positivity, E (1) = 1, and E (AXB) = AE (X)B for all A, B ∈ N , X ∈ M . (1.1) Moreover, as Tomiyama noted, it follows from (1.1) and the positivity preserving property of E that E (X)∗E (X) ≤ E (X ∗X) for all X ∈ M , (1.2) In fact, more is true. As is well known, every norm contractive projection is completely positive. A conditional expectation from M onto N , in the sense of Umegaki [41, 42, 43], is a unital projection from M onto N that is order preserving and such that (1.1) and (1.2) are satisfied. Since every conditional expectation E is a unital completely positive map, its adjoint with respect to the Hilbert-Schmidt inner product, E †, is a completely positive trace preserving (CPTP) map, also known as a quantum operation. (Throughout this paper, a dagger † always denotes the adjoint with respect to the Hilbert-Schmidt inner product.) Let Eτ denote the orthogonal projection from M onto N with respect to the GNS inner product determined by τ . It is easy to see, using the tracial nature of τ , that Eτ is in fact a conditional expectation, and since Eτ = E † τ , Eτ is a quantum operation. 1.1 Definition. For any state ρ on M, ρN denotes the state on N given by ρN := Eτ (ρ) where, as always Eτ denotes the tracial conditional expectation onto N . The restriction of a state ρ on M to N is of course a state on N , and as such, it is represented by a unique density matrix belonging to N , which is precisely ρN . 3 1.2 Example. Let H = H1 ⊗ H2 be the tensor product of two finite dimensional Hilbert spaces. Let M = B(H) be the algebra of all linear transformations on H, and let N be the subalgebra 1H1 ⊗ B(H2) consisting of all operators in M of the form 1H1 ⊗ A, A ∈ B(H2). Then for the normalized trace τ , Eτ (X) = d−1 1 1H1 ⊗ Tr1X for all X ∈ M where d1 is the dimension of H1 and where Tr1 denotes the partial trace over H1. Given two states ρ and σ on M, the Umegaki relative entropy of ρ with respect to σ is defined [44] by (1.3) Lindblad's inequality [19] states that with Eτ being the tracial conditional expectation onto N , (1.4) S(ρσ) := Tr[ρ(log ρ − log σ)] . S(ρσ) ≥ S(Eτ (ρ)Eτ (σ)) . Lindblad showed that the monotonicity (1.4) is equivalent to the joint convexity of the relative entropy (ρ, σ) 7→ S(ρσ), and this in turn is an immediate consequence of Lieb's Concavity Theorem [16]. In the case that M = B(H1 ⊗ H2) and N = {1H1 ⊗ A, A ∈ B(H2)}, (1.4) was proved by Lieb and Ruskai [18], who showed it to be equivalent to the Strong Subadditivity (SSA) of the von Neumann entropy; more information on SSA is contained in Section 5. Using the fact that Stinesping's Dilation Theorem [33] relates general CPTP maps to tracial expectation, Lindblad [20] was able to prove, using (1.4) that for any CPTP map P, S(ρσ) ≥ S(P(ρ)P(σ)) . (1.5) This is known as the Data Processing Inequality (DPI). Because of the simple relation between (1.4) and (1.5) the problem of determining the cases of equality in the Data Processing Inequality largely comes down the problem of determining the cases of equality in (1.4), which was solved by Petz [27, 28]. His necessary and sufficient condition for equality in (1.4) is closely connected with the problem of quantum coarse graining, and in particular a quantum coarse graining operation introduced by Acardi and Cecchini [2], whose dual is now now known as the Petz recovery channel, the CPTP map Rρ given by Rρ(γ) = ρ1/2(ρ−1/2 (1.6) N γρ−1/2 N )ρ1/2 . It is obvious that Rρ(ρN ) = ρ, so that Rρ "recovers" ρ from ρN . Petz proved [27, 28] that there is equality in (1.4) if and only if and that this is true if and only if Rρ(σN ) = σ Rσ(ρN ) = ρ . (1.7) (1.8) There has been much recent work on stability for for the DPI: [13, 17, 32, 34, 35, 46]. Suppose that ρ and σ are such that there is approximate equality in (1.4). To what extent do ρ and σ provide approximate solutions to Pets's equation (1.7) and (1.8)? 4 1.2 Main results In this paper we further develop an approach that we introduced in [9] for proving stability for analogs of the DPI for R´enyi relative entropies. The R´enyi relative entropies include the Umegaki relative entropy (1.3) as a limiting case, but taking advantage of the special structure on the Umegaki relative entropy, we are able to sharpen the stability bounds obtained in [9] for this case. Our results in this direction are given in Theorem 1.5 and Corollary 1.7. These results are proved in Section 2. Moreover, we show how the approach introduced in [9] yields a simple specification of the structure of the pairs of densities ρ and σ that satisfy Petz's equations (1.7) and (1.8). This is a question that was first dealt with by Hayden, Jozsa, Petz and Winter in [14]. The question arose there because, although it is known (and explained in the appendix) that the DPI is equivalent to the quantum Strong Subadditivity of the Entropy (SSA) [18], it is not so simple to translate Petz's condition for equality in the DPI into the condition for equality in SSA. This was done in [14], but Petz himself, together with Mosnyi, returned to the problem of determining the structure of solutions of (1.7) and (1.8) in [23]. Here we give simple specification of the structure of solutions providing some new information. Our main results in this direction are given in Theorem 1.9 and Corollary 1.10. These results are proved in Section 3. As explained in the next two subsections, the questions considered here are closely related to several questions concerning quantum conditional expectations and quantum "coarse graining". in Section 4, we prove a theorem specifying conditions under which the orthogonal projection of M onto N is a conditional expectations, and hence, when expectation preserving conditional expectations exist. The full statement is given in Theorem 4.1. Before giving precise statements, it will be useful to recall the relatively simple situation regarding the classical DPI. 1.3 The classical DPI It will be useful to recall some aspects of the classical analog of the DPI. Let Ω be a finite set. Let F be a non-trivial partition of Ω. Let M denote the functions on Ω, and let N the functions on Ω that are constant on each set of the partition F . Then M and N are commutative von Neumann algebras, and N is a subalgebra of M. Let X be a function on Ω such that X(ω) = X(ω′) if and only ω and ω′ belong to the same set in F . Then X generates F in the sense that the sets constituting F are precisely the non-empty sets of the form {ω X(ω) = x}. Let ρ and σ be two strictly positive probability densities on the set Ω. Let τ denote the uniform probability density on Ω; i.e., τ (ω) = Ω−1 for all ω, where Ω is the cardinality of Ω. As above, let Eτ denote the orthogonal projection of M onto N , which is nothing other than the conditional expectation with respect the random variable Y and the probability measure τ . As above, let ρN = Eτ ρ and σN = Eτ σ. It is clear the ρN is a "coarse grained" version of ρ, obtained by averaging ρ on the sets of the partition F , making it constant on these. Let f (ωy) be the conditional density under ρ for ω given Y (ω), and likewise let g(ωy) be the conditional density under σ for ω given Y (ω). That is f (ωx) = ρ(ω) ρN (x) and g(ωx) = σ(ω) ρN (x) , 5 (1.9) which, for each x in the range of X are both probability densities on the set {ω X(ω) = x}. Then we have ρ(ω) = ρN (X(ω))f (ωX(ω)) and σ(ω) = σN (X(ω))g(ωX(ω)), and hence ρ(ω)([log ρN (ω) − log σN (ω)] + [log f (ωX(ω)) − log g(ωX(ω))]) ρn(X(ω)) [f (ωX(ω)) log f (ωX(ω)) − log g(ωX(ω))] (1.10) For each x in the range of X, it follows from Jensen's inequality that ρ(ω)(log ρ(ω) − log σ(ω)) S(ρσ) = Xω∈Ω = Xω∈Ω = S(ρNσN ) +Xω∈Ω X{ω : X(ω)=x} f (ωx) [log f (ωx) − log g(ωx)] ≥ 0 , (1.11) and there is equality if and only if f (ωx) = g(ωx) everywhere on {ω : X(ω) = x}. It follows that S(ρσ) ≥ S(ρNσN ) with equality if and only if for each x in the range of X, f (ωx) = g(ωx) everywhere on {ω : X(ω) = x}. In this case, X is called a sufficient statistic for the pair {ρ, σ}. Suppose we are given an independent identically distributed sequence of points {ωj}, drawn according to one of the two probability densities ρ or σ, and we want to determine which it is. If we know the function f (ωx), it suffices to observe the sequence {X(ωj)}, and to determine which of ρN (x) or σN (x) is governing its distribution. Indeed, we can define a classical recovery map as follows: For any probability density γ ∈ N , regarded as a probability density on the range of X, define Rργ to be the probability density in M given by Rργ(ω) = γ(X(ω))f (ωX(ω)) . Therefore, we can express the condition for equality in the classical DPI as RρσN = σ, and evidently this is true if and only if RσρN = ρ. This is the classical analog of Petz's result. Moreover, in this notation, we have that Xω∈Ω ρn(X(ω)) [f (ωX(ω)) log f (ωX(ω)) − log g(ωX(ω))] = S(ρRσρ) , so that (1.10) becomes Then by the classical Pinsker inequality, S(ρσ) − S(ρNσN ) ≥ S(ρRσρ) . S(ρσ) − S(ρNσN ) ≥ 1 2 Xω∈Ω ρ(ω) − Rσρ(ω)!2 . (1.12) (1.13) It remains an open problem to prove quantum analogs of (1.12) or (1.13), even with worse constants on the right. Here we prove a quantum analog of (1.13) with a worse constant, and with the power raised from 2 to 4 on the right. The line of argument has to be entirely different from the one we have just employed in the classical case because there is no effective quantum replacement for "conditioning on the observable X". 6 It is therefore useful to find a way of describing the classical recovery map that doe not refer explicitly to conditioning on the random variable X. Define Eσ to be the orthogonal projection of M onto N in L2(σ). Then for any random variable Y (i.e., any function on Ω), EσY is the condition expectation of Y given the sigma-algebra M. The operation Y 7→ EσY yields a "coarse grained version" of Y that is constant on the sets in F : With X and g as in (1.9), EσY (ω) = Xω′: X(ω′)=X(ω)} g(ω′X(ω)Y (ω′) . It is clear from this formula that Eσ preserves positivity, and preserves expectations with respect to σ, That is, σ(Y ) = σ(EσY ) . (1.14) Now let E † σ be the dual operation taking states on M (probability densities on the range of X) to states on M (probability densities on Ω). It is easily seen that this is nothing other than Rσ. That is, the classical recovery map Rσ is nothing other than the dual of the conditional expectation Eσ, which is nothing other than the orthogonal projection of M onto N in L2(σ). This analytic specification of Rσ, making no explicit mention of conditioning on X, provides a starting point for the construction of a quantum recovery map. 1.4 Quantum conditional expectations and quantum coarse graining The discussion of the classical DPI brings us to the question as to whether for any faithful state ρ on M there exists a conditional expectation E from M onto N that preserves expectations with respect to ρ, i.e. such that ρ(X) = ρ(E (X)) for all X ∈ M . (1.15) The property (1.15) says that "the expectation of a conditional expectation of an observable equals the expectation of the observable". If such a conditional expectation exists, then it is unique: Any such conditional expectation must be the orthogonal projection of M on N with respect to the GN S inner product for the state ρ. To see this, note that for all X ∈ M and all A ∈ N , using (1.1), hA, XiGN S,ρ = ρ(A∗X) = ρ(E (A∗X)) = ρ(A∗E (X)) = hA, E (X)iGN S,ρ . (1.16) Suppose that E is a conditional expectation satisfying (1.15). Then since E is a unital com- pletely positive map, E † is a CPTP map; i.e., a quantum channel. For any state γ on N , and any A ∈ M, we then have E †(A) = γ(E (A)) . Then when (1.15) is satisfied, taking γ = ρN , we have E †ρN (A) = ρN (E (A)) = ρ(A) , and this means that E † is a quantum channel that "recovers" ρ from ρN . As we have already noted, Eτ is a conditional expectation with the property (1.15). However, for non-tracial states ρ, a conditional expectation satisfying (1.15) need not exist. 7 A theorem of Takesaki [36] says, in our finite dimensional context, that for a faithful state ρ, there exists a conditional expectation E from M onto N if and only if ρAρ−1 ∈ N for all A ∈ N , and in general this is not the case. We give a short proof of this and somewhat more in Section 4: In Theorem 4.1, we prove that Eρ, the orthogonal projection from M onto N in the GNS inner product with respect to ρ, is real (that is, it preserves self-adjointness) if and only if ρAρ−1 ∈ N for all A ∈ N . Since every order preserving linear transformation is real, this precludes the general existence of conditional expectations satisfying (1.15) whenever N is not invariant under X 7→ ρXρ−1, thus implying Takesaki's Theorem (in this finite dimensional setting). There is another inner product on M that is naturally induced by a faithful state ρ, namely the Kubo-Martin-Schwinger (KMS) inner product. It is defined by hX, Y iKM S,ρ = Tr[ρ1/2X ∗ρ1/2Y ] = Tr[(ρ1/4Xρ1/4)∗(ρ1/4Y ρ1/4)] . (1.17) Evidently, for any X ∈ M and Y ∈ N , hX, Y iKM S,ρ = Tr[ρ1/2X ∗ρ1/2Y ] = Tr[(ρ−1/4 N ρ1/2Xρ1/2ρ−1/4 ≤ k(ρ−1/4 N ρ1/2Xρ1/2ρ−1/4 N Y ρ1/4 N )∗(ρ1/4 N )] N )kHSkY kKM S,ρN . Hence Y 7→ hX, Y iKM S,ρ is a bounded linear functional on (N ,h· ,·iKM S,ρN ), and then there is a uniquely determined Aρ(X) ∈ N such that for all X ∈ M and all Y ∈ N , The map Aρ was introduced by Accardi and Cecchini [2], building on previous work by Accardi hX, Y iKM S,ρ = hAρ(X), Y iKM S,ρN . (1.18) [1]. 1.3 Definition. Let ρ be a faithful state on M. The Accardi-Cecchini coarse graining operator Aρ from M to N is defined by (1.18). The map Aρ was introduced by Accardi and Cecchini [2], building on previous work by Accardi [1]. It is a "coarse graining" operation in that to each observable X in the larger algebra M, it associates an observable Aρ(X), in the smaller algebra N , and measurement of Aρ(X) will yields coarser information than a measurement of X itself. The same, of course, is true for conditional expectations, Since 1 ∈ N by definition, for all X ∈ M, h1, XiKM S,ρ = h1, Aρ(X)iKM S,ρ, and for all X ∈ M, h1, XiKM S,ρ = Tr[σ1/21σ1/2X] = ρ(X). Therefore ρ(Aρ(X)) = ρ(X) . (1.19) Thus, unlike conditional expectations in general, the Accardi-Cecchini coarse-graining operator always preserves expectations with respect to ρ. In the matricial setting, it is a particularly simple matter to derive an explicit expression for Aρ. By definition, for all X ∈ M and all Y ∈ N , 8 Aρ(X)] = Tr[ρ1/2Y ρ1/2X] . (1.20) N Y ρ1/2 N Make the change of variables Z = ρ1/2 ranges over N . Hence Tr[ρ1/2 N Y ρ1/2 N . Since ρ1/2 N is invertible and Y ranges over N , Z Tr[ZAρ(X)] = Tr[(ρ1/2ρ−1/2 N Zρ−1/2 N ρ1/2)X] = Tr[Z(ρ−1/2 N ρ1/2Xρ1/2ρ−1/2 N )] . Since the above holds for all Z ∈ N , it follows that Aρ(X) = ρ−1/2 N Eτ (ρ1/2Xρ1/2)ρ−1/2 N . (1.21) (1.22) It is evident from this formula that Aρ is a completely positive unital map from M to N , and therefore it is actually a contraction from M to N . By Tomiyama's Theorem, it cannot in general be a projection of M onto N . That is, if X ∈ N , it is not necessarily the case that Aρ(X) = X. The set of X ∈ N for which this is true turns out to be a subalgebra of N , as was shown by Accardi and Cecchini [2]. This subalgebra will be of interest in what follows. 1.4 Definition. The Petz recovery map Rρ is the Hilbert-Schmidt adjoint of Aρ [28]. That is, Rρ = A † ρ , or equivalently, for all density matrices γ ∈ N Tr[γAρ(X)] = Tr[Rρ(γ)X] . A dagger † always denotes the adjoint with respect to the Hilbert-Schmidt inner product. As the dual of a unital completely positive map, Rρ is a CPTP map. Moreover, it follows immediately from the definition and (1.22) that for all density matrices γ ∈ N , Rρ(γ) = ρ1/2(ρ−1/2 N γρ−1/2 N )ρ1/2 . (1.23) It is evident from this formula not only that Rρ is a CPTP map, but that Rρ(ρN ) = ρ; i.e., Rρ recovers ρ from ρN . Now suppose that σ is another density matrix in M and that Rρ(σN ) = σ . (1.24) Then by the Data Processing Inequality and (1.24), S(ρσ) ≤ S(Rρ(ρN )Rρ(σN )) = S(ρσ). Hence when Rρ(σN ) = σ, there is equality in (1.4). The deeper result of Petz [27, 28] is that there is equality in (1.4) only in this case. Our goal is to prove a stability bound for Petz's theorem on the cases of equality in (1.4). Our result involves the relative modular operator ∆ρ,σ on M defined by ∆σ,ρ(X) = σXρ−1 (1.25) for all X ∈ M. This is the matricial version of an operator introduced in a more general von Neumann algebra context by Araki [3]. Our main result is: 1.5 Theorem. Let ρ and σ be two states on M Let Eτ be the tracial conditional expectation onto a von Neumann subalgebra N , and let ρN = Eτ ρ and σN = Eτ σ. Then, with k · k2 denoting the Hilbert-Schmidt norm, 9 4(cid:17)4 S(ρσ) − S(ρNσN ) ≥(cid:16) π k∆σ,ρk−2kσ1/2 N ρ−1/2 N ρ1/2 − σ1/2k4 2 . (1.26) The quantity on the right hand side may be estimated in terms of the Petz recovery map. In Section 2 we prove: 1.6 Lemma. Let ρ, σ and σN be specified as in Theorem 1.5. Then, with k·k1 denoting the trace norm, 1 2kRρ(σN ) − σk1 . As an immediate Corollary of Theorem 1.5 and Corollary 1.6, we obtain k(σN )1/2(ρN )−1/2ρ1/2 − σ1/2k2 ≥ 1.7 Corollary. Let ρ and σ be two states on M. Let Eτ be the tracial conditional expectation onto a von Neumann subalgebra N , and let ρN = Eτ ρ and σN = Eτ σ. Then, with k · k1 denoting the trace norm, 8(cid:17)4 S(ρσ) − S(ρNσN ) ≥(cid:16) π k∆σ,ρk−2kRρ(σN ) − σk4 1 . (1.27) Note how the right hand side of (1.27) differs from the right hand side of (1.13): Apart from the constant and the power 4 in place of 2, the most striking difference is that the roles of ρ and σ are reversed. The expected result, with a worse constant, is obtained in Corollary 1.8. Recall that the modular operator is the right multiplication by ρ−1 and left multiplication by σ, so k∆σ,ρk ≤ kρ−1k, since kσk ≤ 1. While kρ−1k might be considerably larger than k∆σ,ρk, a bound in terms of kρ−1k has the merit that it is independent of σ: 8(cid:17)4 S(ρσ) − S(ρNσN ) ≥(cid:16) π kρ−1k−2kRρ(σN ) − σk4 1 . (1.28) Corollary 1.7 yields a result of Petz: With M, N , ρ and σ as above, S(ρσ) = S(ρNσN ) if and only if σ satisfies the Petz equation σ = Rρ(σN ) . (1.29) Theorem 1.5 gives what appears to be a stronger condition on relating ρ, σ, ρN and σN , namely that ρ−1/2 N ρ1/2 = σ−1/2 N σ1/2 . (1.30) While validity of (1.30) immediately implies that σ satisfies the Petz equation (1.29), the converse is also true: By what we have noted above, when (1.29) is satisfied, S(ρσ) = S(ρNσN ), and then by Theorem 1.5, (1.30) is satisfied. This may be made quantitative as follows: Letting LA denote the operator of left multiplication by A, Lρ 1/2 N Lσ −1/2 N (σ1/2 N ρ−1/2 N ρ1/2 − σ1/2) = (ρ1/2 − ρ1/2 N σ−1/2 N σ1/2) , and hence kρ1/2 − ρ1/2 N σ−1/2 N k = kρNk1/2 and kLσ N σ1/2k2 ≤ kL k = kσ−1 −1/2 N 1/2 ρ 1/2 N kkL Since kLρ kkσ1/2 N ρ−1/2 N ρ1/2 − σ1/2k2 . 10 (1.31) −1/2 N σ N k1/2, we may combine (1.31) with (1.26) to obtain 4(cid:17)4 S(ρσ) − S(ρNσN ) ≥(cid:16) π k∆σ,ρk−2kρNk−2kσ−1 N k−2kρ1/2 N σ−1/2 N σ1/2 − ρ1/2k4 2 , (1.32) which is the analog of (1.26) with a somewhat worse constant on the right, but the roles of ρ and σ interchanged there. Applying Lemma 1.6 once more, we obtain 1.8 Corollary. Let ρ and σ be two states on M. Let Eτ be the tracial conditional expectation onto a von Neumann subalgebra N , and let ρN = Eτ ρ and σN = Eτ σ. Then S(ρσ) − S(ρNσN ) ≥(cid:16) π 8(cid:17)4 k∆σ,ρk−2kρNk−2kσ−1 N k−2kRσ(ρN ) − ρk4 1 . (1.33) As above, bounding the norms of states by 1, we get a constant that depends only on the smallest eigenvalues of ρ and σN 8(cid:17)4 S(ρσ) − S(ρNσN ) ≥(cid:16) π kρ−1k−2kσ−1 N k−2kRσ(ρN ) − ρk4 1 . (1.34) We noted above that σ solves the Petz equation if and only if (1.30) is satisfied, and then since (1.30) is symmetric in ρ and σ, σ = RρσN if and only if ρ = RσρN , and hence S(ρσ) = S(ρNσN ) ⇐⇒ S(σρ) = S(σNρN ) . (1.35) The reasoning leading to Corollary 1.8 show that moreover, kρ = RσρNk1 and kσ = RρσNk1 are comparable in size. To state our results on the structure of the solution set of the Petz equation, we introduce the fixed point set (1.36) Standard results (see Section 3) show that C is a von Neumann subalgebra of N . Let ∆ρ denote the modular operator on M, C := {X ∈ M : Aρ(X) = X }. (1.37) Then C may also be characterized (Theorem 3.3) as the largest von Neumann subalgebra of N that is invariant under ∆ρ. ∆ρ(X) = ρXρ−1 . Let Z denote the center of C. Then by standard results (see Section 3), Z is generated by a finite family {P1, . . . , PJ} of mutually orthogonal projections. Define H(j), j = 1, . . . , J, to be the range of Pj. The restriction of C to each H(j) is a factor, and hence each H(j) factors as Hj,ℓ⊗Hj,r, and the general element A of C has the form A = JXj=1 1Hj,ℓ ⊗ Aj,r , 11 LJ where Aj,r ∈ B(Hj,r). All of this follows from the standard theory of the structure of finite dimensional von Neumann algebras, and we emphasize that C and the decomposition H = j=1 Hj,ℓ ⊗ Hj,r is canonically associated to ρ. Since C is invariant under ∆ρ, the orthogonal projection from M onto N in the GNS inner product induced by ρ is a conditional expectation (see Theorem 4.1) that we denote by EC,ρ, and which we call the conditional expectation given C under ρ. We shall prove: 1.9 Theorem. Let ρ be a faithful state on M, and let N be a von Neumann subalgebra of M. Let C be the fixed-point algebra of the Accardi-Cecchini coarse graining operator Aρ. Let H be the finite dimensional Hilbert space on which M acts, and let H = Hj,ℓ ⊗ Hj,r induced by the decomposition of C as a direct sum of factors. Then there are uniquely determined density matrices {γ1,r, . . . , γJ,r}, where γj,r acts on Hj,r and γj,ℓ ⊗ 1Hj,r ∈ M so that JMj=1 Moreover, Eτ (γj,ℓ ⊗ 1Hj,r ) has the formeγj,ℓ ⊗ 1Hj,r and ρ = JMj=1 γj,ℓ ⊗ TrHj,ℓ(PjρPj) . ρN = JMj=1eγj,ℓ ⊗ TrHj,ℓ(PjρPj) . (1.38) (1.39) A state σ on M solves the Petz equation RρσN = σ if and only if for all X in M, σ(X) = σ(EC,ρ); i.e., if and only if expectations with respect to σ are preserved under the the conditional expectation given C under ρ. Every such state σ has the form σ = same {γ1,r, . . . , γJ,r} 1.10 Corollary. Let ρ, σ be faithful states on M, and let N be a von Neumann subalgebra of M. Let C be the fixed-point algebra of the Accardi-Cecchini coarse graining operator Aρ for N . Let ECρ be the conditional expectation given C under ρ. Define ρC = ECρρ, σC = ECρσ. Then γj,ℓ ⊗ TrHj,ℓ(PjσPj) for the JMj=1 S(ρσ) = S(ρNσN ) ⇐⇒ S(ρσ) = S(ρCσC) . (1.40) In particular, if C is spanned by 1, S(ρσ) = S(ρNσN ) if and only if ρ = σ. We close the introduction with some further comments on recovery map stability bounds for the Data Processing Inequality. In physical applications, instead of the trace distance, one often consider an alternative measure of the closeness between two quantum states, the fidelity [39]. For two states ρ and σ on B(H), the fidelity between them is defined as F (ρ, σ) = k√ρ√σk2 1. (1.41) For all states ρ and σ, we have 0 ≤ F (ρ, σ) ≤ 1. The fidelity equal to one if and only if the states are equal, and it is equal to zero if and only is the support of ρ is orthogonal to the support of σ. So in other words, the fidelity is zero when states are perfectly distinguishable, and zero when they cannot be distinguished. Note that the fidelity itself satisfies the monotonicity relation under a completely positive trace preserving maps, but we will not discuss it here. Moreover, there is a relation between the trace distance kρ − σk1 and fidelity 1 −pF (ρ, σ) ≤ 1 2kρ − σk1 ≤p1 − F (ρ, σ) . From here and the Corollary 1.7 we obtain the quantitative version of the Petz's Theorem involving the fidelity between states S(ρσ) − S(ρNσN ) ≥(cid:16) π 4(cid:17)4 k∆σ,ρk−2(cid:18)1 −qF (σ, Rρ(σN ))(cid:19)4 . (1.43) Recent results [13, 17, 45, 46] provide sharpening of the monotonicity inequality, but the lower bounds provided there involve quantities that are hard to compute, e.g. rotated and twirled Petz recovery maps. for another fidelity type bound not explicitly involving the recover map, see [8, Theorem 2.2]. The appeal of the above bound is that it involves simple distance measure between the original state σ and Petz recovered state Rρ(σN ). The proof of the theorem 1.5 also implies that satisfaction of the Petz equation RρσN = σ is the necessary and sufficient condition for cases of equality in the monotonicity inequality for a large class of quasi-relative entropies, as we now briefly explain. Let f : (0, +∞) → R be an operator convex function, so that for all n ∈ N, and all positive 2 f (B). We say that f is strictly operator convex 2 f (A)) + 1 n× n matrices A and B, f ( 1 in case there is equality if and only if A = B. 2 B) ≤ 1 2A + 1 Petz [25], [26] has defined the f -relative quasi-entropy as Sf (ρσ) = Tr[f (∆σ,ρ)ρ] = hρ1/2, f (∆σ,ρ)ρ1/2iHS . (1.44) Since − log(∆σ,ρ)ρ1/2 = ρ1/2 log ρ− log σρ1/2, the choice f (x) = − log x yields the Umegaki relative entropy. Since for each t > 0, the function x 7→ (t + x)−1 is operator convex, this construction yields a one parameter family a quasi relative entropies, S(t), defined by 12 (1.42) (1.45) (1.46) From the integral representation of the logarithm it follows that S(t)(ρσ) = Tr(cid:2)(t + ∆σ,ρ)−1ρ(cid:3) . 1 + t(cid:19) dt , − log(A) =Z ∞ 1 + t(cid:19) dt , S(ρσ) =Z ∞ 0 (cid:18) 1 0 (cid:18)S(t)(ρσ) − t + A − 1 1 13 and we may use this representation to study monotonicity for the Umegaki relative entropy in terms of monotonicity for the one parameter family of quasi relative entropies S(t). The proof of Theorem 1.5 is ultimately derived from a stability bound for the variant of the Data Processing Inequality that is valid for the quasi relative entropies S(t). Since the R´enyi relative entropies can be expressed in terms of a similar integral representation, the Petz equation σ = RρσN again characterizes the condition for cases of equality in these variants of the Data Processing Inequality; this will be developed in detail in a companion paper. 2 Stability for the Data Processing Inequality We begin this section by recalling Petz's proof of the monotonicity of the quasi relative entropies Sf for operator convex f . Throughout this section N is a von Neumann subalgebra of the finite dimensional von Neumann algebra M, and ρ and σ are two density matrices in M. Eτ is the tracial conditional expectation onto N , and ρN = Eτ ρ and σN = Eτ σ. Finally H denotes (M,h·,·iHS), Define the operator U mapping H to H by U(X) = Eτ (X)ρ−1/2 N ρ1/2 . Note that for all X ∈ N , U(X) = Xρ−1/2 N ρ1/2. The adjoint operator on H is given by U ∗(Y ) = Eτ (Y ρ1/2)ρ−1/2 N (2.1) (2.2) for all Y ∈ H = M. For X ∈ M, U ∗U(X) = Eτ (ρ−1/2 N ρ) = Eτ (X). Hence U ∗U = Eτ , the orthogonal projection in H onto N . That is, U, restricted to N , is an isometric embedding of N into H = M, but it is not the trivial isometric embedding by inclusion. Also, we see that on N the map U is isometric. Eτ (X)ρ−1/2 N Now observe that for all X ∈ N , ∆1/2 σ,ρ (U(X)) = σ1/2Xρ−1/2 N , and hence for all X ∈ N , σ,ρ (U(X)), ∆1/2 h∆1/2 σ,ρ (U(X))i = Tr((ρN )−1/2X ∗σX(ρN )−1/2) = Tr((ρN )−1/2X ∗σN X(ρN )−1/2) = h∆1/2 σN ,ρN (X)i . σN ,ρN (X), ∆1/2 That is, on N , U ∗∆σ,ρU = ∆σN ,ρN . By the operator Jensen inequality, as operators on (N ,h·,·iHS), U ∗f (∆σ,ρ)U ≥ f (U ∗∆σ,ρU) . (2.3) (2.4) 14 Combining (2.3) and (2.4), and using the fact that U(ρN )1/2 = ρ1/2, Sf (ρNσN ) = h(ρN )1/2, f (∆σN ,ρN )(ρN )1/2i ≤ (cid:10)U(ρN )1/2, f (∆σ,ρ)U(ρN )1/2(cid:11) = (cid:10)ρ1/2, f (∆σ,ρ)ρ1/2(cid:11) = Sf (ρσ) . This proves, following Petz, his monotonicity theorem for the quasi relative entropy Sf for the operator convex function. Now consider the family of quasi relative entropies defined by functions ft(x) = (t + x)−1. Our immediate goal is to prove the inequality S(t)(ρσ) = hρ1/2, (t + ∆σ,ρ)−1ρ1/2i ≥ hρ1/2 N , (t + ∆σN ,ρN )−1ρ1/2 N i = S(t)(ρNσN ) . (2.5) 2.1 Lemma. Let U be a partial isometry embedding a Hilbert space K into a Hilbert space H. Let B be an invertible positive operator on K, A be an invertible positive operator on H, and suppose that U ∗AU = B. Then for all v ∈ K, where hv, U ∗A−1U vi = hv, B−1vi + hw, Awi , w := U B−1v − A−1U v . (2.6) (2.7) Proof. We compute, using U ∗U = 1K, hw, Awi = hU B−1v − A−1U v, AU B−1v − U v)i = hv, B−1U ∗AU B−1vi − 2hv, B−1vi + hv, U ∗A−1U vi = −hv, B−1vi + hv, U ∗A−1U vi Proof of Theorem 1.5. We apply Lemma 2.1 with A := (t + ∆σ,ρ), B = (t + ∆σN ,ρN ) and v := (ρN )1/2, and with U defined as above. The lemma's condition, U ∗AU = B, follows from (2.3) and the fact that U ∗U = 1K. Therefore, applying Lemma 2.1 with U(ρN )1/2 = ρ1/2, S(t)(ρσ) − S(t)(ρNσN ) = hρ1/2, (t + ∆σ,ρ)−1ρ1/2i − hρ1/2 N , (t + ∆σN ,ρN )−1ρ1/2 N i = hwt, (t + ∆σ,ρ)wti ≥ tkwtk2, where, recalling that U(ρN )1/2 = ρ1/2, wt := U(t + ∆σN ,ρN )−1(ρN )1/2 − (t + ∆σ,ρ)−1ρ1/2 . Using the integral representation of the square root function, (2.8) (2.9) X 1/2 = 1 πZ ∞ 0 t1/2(cid:18)1 t − 1 t + X(cid:19) dt, and U(N ρ)1/2 = ρ1/2 once more, we conclude that U(∆σN ,ρN )1/2(ρN )1/2 − (∆σ,ρ)1/2ρ1/2 = On the other hand, 1 πZ ∞ 0 15 t1/2wtdt . U(∆σN ,ρN )1/2(ρN )1/2 − (∆σ,ρ)1/2ρ1/2 = U(σN )1/2 − σ1/2 = (σN )1/2(ρN )−1/2ρ1/2 − σ1/2 . Therefore, combining the last two equalities and taking the Hilbert space norm associated with H, for any T > 0, k(σN )1/2(ρN )−1/2ρ1/2 − σ1/2k2 = ≤ 0 1 π(cid:13)(cid:13)(cid:13)(cid:13)Z ∞ πZ T 1 0 t1/2wtdt(cid:13)(cid:13)(cid:13)(cid:13)2 t1/2kwtk2dt + 1 π(cid:13)(cid:13)(cid:13)(cid:13)Z ∞ T t1/2wtdt(cid:13)(cid:13)(cid:13)(cid:13)2 . (2.10) We estimate these two terms separately. For the first term, by the Cauchy-Schwarz inequality, (cid:18)Z T 0 t1/2kwtk2dt(cid:19)2 ≤ TZ T ≤ TZ ∞ 0 0 tkwtk2 2dt (cid:0)S(t)(ρσ) − S(t)(ρNσN )(cid:1) dt = T (S(ρσ) − S(ρNσN )) . For the second term in (2.10), note that for any positive operator X (2.11) t − t1/2(cid:18) 1 t1/2(cid:18)1 t − and hence Z ∞ T 1 t − t + X(cid:19) ≤ t1/2(cid:18) 1 t + X(cid:19) dt ≤ kXk1/2(cid:18)Z ∞ 1 1 t + kXk(cid:19) 1 = kXk t1/2(kXk + t) 1, 1 t1/2(1 + t) dt(cid:19) 1 ≤ 2kXk T 1/2 1 . T /kXk The spectra of σN and ρN lie in the convex hulls of the spectra of σ and ρ respectively. It follows that k∆σN ,ρNk ≤ k∆σ,ρk. Therefore, recalling the definition of wt in (2.9), we obtain Combining (2.10), (2.11) and (2.12) we obtain (cid:13)(cid:13)(cid:13)(cid:13)Z ∞ T t1/2wtdt(cid:13)(cid:13)(cid:13)(cid:13)2 ≤ 4k∆σ,ρk T 1/2 . (2.12) k(σN )1/2(ρN )−1/2ρ1/2 − σ1/2k2 ≤ 1 π T 1/2(S(ρσ) − S(ρNσN ))1/2 + 4k∆σ,ρk πT 1/2 . 16 Optimizing in T , k(σN )1/2(ρN )−1/2ρ1/2 − σ1/2k2 ≤ 4 πk∆σ,ρk1/2(S(ρσ) − S(ρNσN ))1/4 Rearranging terms 4(cid:17)4 S(ρσ) − S(ρNσN ) ≥(cid:16) π k∆σ,ρk−2k(σN )1/2(ρN )−1/2ρ1/2 − σ1/2k4 2 . (2.13) We now prove the lemma leading from (1.26) to (1.27). 2.2 Lemma. For any operators X and Y with Tr[X ∗X] = Tr[Y ∗Y ] = 1. Then kX ∗X − Y ∗Y k1 ≤ 2kX − Y k2 . (2.14) Proof. Recall that for any operator A, kAk1 = sup{Tr[ZA] operator norm. For any contraction Z, using cyclicity of the trace we have : kZk ≤ 1} where k · k denotes the Tr[Z(X ∗X − Y ∗Y )] ≤ Tr[Z(X ∗ − Y ∗)X + ZY ∗(X − Y )] ≤ Tr[(X ∗ − Y ∗)XZ + Tr[ZY ∗(X − Y )] ≤ (Tr(X ∗ − Y ∗)(X − Y )])1/2(Tr[X ∗Z ∗ZX])1/2 + (Tr(X ∗ − Y ∗)(X − Y )])1/2(Tr[Y ∗Z ∗ZY ])1/2 ≤ 2kX − Y k2 . Applying this with X = (σN )1/2(ρN )−1/2ρ1/2 and Y = σ1/2, we get k(σN )1/2(ρN )−1/2ρ1/2 − σ1/2k2 ≥ 1 2kRρ(σN ) − σk1 . 3 Structure of the solution set of the Petz equation Define the CPTP map Φ : M → M by Φ := Rρ ◦ Eτ . (3.1) The Petz equation (1.29) can be written as Φ(σ) = σ. The adjoint of Φ, is the completely positive unital map Φ† = Ψ : N → N given by Ψ := ιN ,M ◦ Aρ (3.2) where ιN ,M is the inclusion of N in M. 17 The problem of determining all of the states fixed by Φ is closely related to the problem of determining all of the fixed points of Ψ in M. This problem has been investigated in a general context by Lindblad [21] in the proof of his General No-cloning Theorem, drawing on earlier work by Choi and Kadison. It was also investigated in this specific context by Accardi and Cecchini. For now, we need not assume that Ψ is given by (3.2). For now, all we require is that Ψ is a unital completely positive map from M to M, and that its Hilbert-Schmidt dual Φ has a faithful invariant state ρ. Then, by an often used argument, the map Ψ is a contraction on (M,h·,·iGN S,ρ): By the operator Schwarz inequality, for all X ∈ M, Ψ(X)∗Ψ(X) ≤ Ψ(X ∗X). Then kΨ(X)k2 GN S,ρ = ρ(Ψ(X)∗Ψ(X)) ≤ ρ(Ψ(X ∗X)) = Φ(ρ)(X ∗X) = kXk2 GN S,ρ . Define (3.3) which is evidently a subspace. Let EC be the orthogonal projection in (M,h·,·iGN S,ρ) onto C. Then, arguing as in [21], by the von Neumann Mean Ergodic Theorem, C = {X ∈ N : Ψ(X) = X } , EC(X) = lim N→∞ Ψj(X) , 1 N NXj=1 The following lemma may be found in [21]; we give the short proof for the reader's convenience. 3.1 Lemma. Let Φ be a CPTP map on M, and let Ψ = Φ†. A density matrix τ ∈ M satisfies Φ(τ ) = τ if and only if it satisfies E † C (τ ) = τ . Proof. Let τ be any density matrix τ in M such that Φ(τ ) = τ For all N and all X ∈ M, Tr(τ X) = 1 N NXj=1 (Tr(Φj(τ )X) = 1 N NXj=1 Tr(τ Ψj(X)) . In the limit, we obtain Tr(τ X) = Tr(τ EC(X)) = Tr(E † C (τ )X). Hence τ = E † C τ . Now suppose that τ is any density matrix in M satisfying τ = E † C τ . Since evidently, Ψ ◦ EC = EC ◦ Ψ = EC, for all X ∈ M, Tr(τ X) = Tr(E † C (τ )X) = Tr(τ, (EC(X))) = Tr(τ, (EC(ΨX))) = Tr(τ, Ψ(X))) = Tr(Φ(τ )X) . since X is arbitrary, Φ(τ ) = τ . Furthermore, by results of Choi [5] and Lindblad [21], C is a unital ∗-subalgebra of N . Let Z denote the center of C, which is commutative von Neumann algebra. Because Z is commutative, it has a particularly simple structure: If P and Q are two orthogonal projections in Z, then P Q = QP is also an orthogonal projection in Z. Since Z is the closed linear span of the projections contained in it, one easily deduces the existence of a family {P1, . . . , PJ} of mutually orthogonal projections summing to the identity such that Z is the span of these projections. 18 JMj=1 Define H(j), j = 1, . . . , J, to be the range of Pj. Then the Hilbert space H on which M acts can H(j). (The notation is chosen to avoid confusion with tensor product be decomposed as H = decompositions such as, e.g., H = H1 ⊗ H2 for bipartite systems.) Then each H(j) is invariant under C, and the center of C restricted to each H(j) is trivial -- it is spanned by Pj, the identity on H(j). Therefore, the restriction of C to each H(j) is a factor -- a ∗-subalgebra of B(H) with a trivial center. By the well-known structure theorem for finite dimensional factors, H(j) can be factored as H(j) = Hj,ℓ ⊗ Hj,r and Cj = 1Hj,ℓ ⊗ B(Hj,r) . Using this decomposition and structure theorem, Lindblad proves [21, Section 4] the following, stated here in terms of the notation set above: 3.2 Lemma. Let Ψ be a unital completely positive map on M, where M acts on a finite dimen- sional Hilbert space H, and where Ψ† leaves a faithful state ρ invariant. Let EC be the orthogonal projection onto C, the C ∗ algebra of fixed points of Ψ, with respect to the GNS inner product in- duced by ρ. Then there are uniquely determined density matrices {γ1,ℓ, . . . , γJ,ℓ}, where γj,r acts on Hj,r, such that for all Y ∈ M, EC(Y ) = JXj=1 1Hj,ℓ ⊗ TrHj,ℓ[(γj,ℓ ⊗ 1Hj,r )PjY Pj] , where TrHj,ℓ denotes the trace over Hj,ℓ. From this explicit description of EC, one readily deduces that E † C (τ ) = JXj=1 γj,ℓ ⊗ TrHj,ℓ(Pjτ Pj) (3.4) where γj,ℓ⊗ TrHj,ℓ(Pjτ Pj) is defined as operators on all of H by setting it to zero on the orthogonal complement of H(j). Hence τ = E † C (τ ) if and only if τ is given by the right hand side of (3.4). Now return to the case at hand, in which Φ and Ψ are given by (3.1) and (3.2) respectively. Proof of Theorem 1.9. Since Φρ = ρ, Lemma 3.1, Lemma 3.2 and (3.4) yield ρ = JXj=1 γj,ℓ ⊗ TrHj,ℓ(PjρPj) (3.5) with the set {γ1,ℓ, . . . , γJ,ℓ} determined by C, the fixed point algebra of Ψ = ιN ,M · Aρ. Next observe that C is a von Neumann subalgebra of N , and is in fact that fixed point algebra of eΨ := Aρ ◦ ιN ,M as well as of Ψ = ιN ,M ◦ Aρ. Since eΦ := eΨ† = Eτ ◦ Rρ, we have eΦρN = ρN and eΦσN = σN . Using Lemma 3.1, Lemma 3.2 and (3.4) once more, we see that for some {eγi,ℓ, . . . ,eγJ,ℓ}, where for each j,eγj,ℓ is a density matrix on Hj,ℓ (3.6) ρN = JXj=1eγj,ℓ ⊗ TrHj,ℓ(PjρN Pj) . Now observe that we may factor 19 ρ = JXj=1 (γj,ℓ ⊗ 1Hj,r )(1Hj,ℓ ⊗ (TrHj,ℓ(PjρPj))) For each j, 1Hj,ℓ ⊗ TrHj,ℓ(PjρPj) ∈ C ⊂ N . Therefore, ρN = Eτ ρ = JXj=1 Eτ (γj,ℓ ⊗ 1Hj,r )(1Hj,ℓ ⊗ (TrHj,ℓ(PjρPj))) We now claim that for each j, (3.7) (3.8) Eτ (γj,ℓ ⊗ 1Hj,r ) =eγj,ℓ ⊗ 1Hj,r . To see this note that since Z ⊂ C ⊂ N , N ′ ⊂ C′ ⊂ Z ′ = {P1, . . . , PJ}′, every unitary in N ′ commutes with each Pj and thus has the block form U =PJ j=1 PjU Pj. Moreover, again using the fact that N ′ ⊂ C′, and that the commutator of 1Hj,ℓ ⊗ B(Hj,r) is B(Hj,ℓ) ⊗ 1Hj,r , we see that U has the form (3.9) U = Pj(Uj,ℓ ⊗ 1Hj,r )Pj, JXj=1 where Uj,ℓ is unitary on Hj,ℓ, though in general, only a subset of the block unitaries of this form belong to N ′. In any case, representing Eτ as an average over appropriate unitaries of this form [12, 38], we obtain (3.9). Now combining (3.7), (3.8) and (3.9) yields Combing (3.7) and (3.10) it follows that ρ−1ρN = (γ−1 ρN = JXj=1 (eγj,ℓ ⊗ 1Hj,r )(1Hj,ℓ ⊗ (TrHj,ℓ(PjρPj))) . (3.10) JXj=1 j,ℓeγj,ℓ ⊗ 1Hj,r ). The general element A of C has the from A = JXj=1 1Hj,ℓ ⊗ Aj,r where each Aj,r ∈ B(Hj,r). For such A, ∆−1 ρ (∆ρN (A)) = ρ−1 N ρAρ−1ρN = A , (3.11) and this verifies that ∆ρ(A) = ∆ρN (A) for all A ∈ C, which we know must be valid by Theorem 4.1. j=1(γj,ℓ⊗ 1Hj,r )(1Hj,ℓ ⊗ (TrHj,ℓ(PjσPj))) The same analysis applies to σ and σN yielding σ =PJ and σN =PJ j=1(eγj,ℓ⊗1Hj,r )(1Hj,ℓ⊗(TrHj,ℓ(PjσPj))) , and then ∆σ(A) = ∆σN (A) for all A ∈ C. 20 3.3 Theorem. Let C be defined by (3.3) and Ψ := ιN ,M◦ Aρ, and let B be any other von Neumann subalgebra of N that is invariant under ∆ρ. Then B ⊂ C. Proof. Let Eτ,B denote the tracial conditional expectation onto B, and define ρB := Eτ,Bρ. Let Aρ,B,N , Aρ,N ,M and Aρ,B,M be the Accardi-Cecchini coarse graining operators from N to B, M to N and M to B. A simple computation shows that Aρ,B,M = Aρ,B,N ◦ Aρ,N ,M . (3.12) Let PB denote the orthogonal projection of M onto B with respect to the GNS inner product induced by ρ. Since ∆ρ leaves B invariant, by (2) of Theorem 4.1 Aρ,B,M = PB. We claim that Aρ,B,N is the restriction of Pρ to N . Indeed, by the defining relation (1.18), for all X ∈ M and all Y ∈ B, (3.13) Tautologically, this holds for all X ∈ N and all Y ∈ B, and so for all X ∈ N , Aρ,B,N (X) = Aρ,B,M(X). We therefore have that for all B ∈ B hX, Y iKM S,ρ = hAρ,B,M(X), Y iKM S,ρB . B = PρB = PρB ◦ Aρ,N ,M(B) , (3.14) and this implies that B = Aρ,N ,M(B), which, by definition, means that B ∈ C. 4 Conditional expectations Recall from the introduction that if ρ is a faithful state on M and N is a von Neumann subalgebra of M, then there exists a conditional expectation E from M to N such that for all X ∈ M, ρ(X) = ρ(E (X)) if and only if the orthogonal projection onto N in the GNS inner product induced by ρ is a conditional expectation. This raises the question: For which faithful states ρ is the orthogonal projection onto N in the GNS inner product induced by ρ is actually a conditional expectation? 4.1 Theorem. Let M be a finite dimensional von Neumann algebra, and let N be a von Neumann subalgebra of M. Let ρ be a faithful state on M, and let ∆ρ be the modular operator on M defined by ∆ρ(X) = ρXρ−1. Let Pρ be the orthogonal projection from M onto N in the GNS inner product induced by ρ. Then: (1) Pρ is real; i.e., it preserves self-adjointness, if and only if N is invariant under ∆ρ. (2) N is invariant under ∆ρ if and only if for all A ∈ N , ∆ρ(A) = ∆ρN (A) , (4.1) in which case ∆t only if Aρ(A) = A for all A ∈ N . ρ(A) = ∆t ρN (A) for all t ∈ R. Furthermore, (4.1) is valid for all A ∈ N if and 21 4.2 Remark. Part (2) of Theorem 4.1 is due to Accardi and Cecchini [2, Theorem 5.1]. In our finite dimensional context, we give a very simple proof; most of the proof below is devoted to (1). Proof of Theorem 4.1. Suppose that Pρ is real. Then for all X ∈ Null(Pρ), 0 = (Pρ(X))∗ = Pρ(X ∗), so that Null(Pρ) is a self adjoint subspace of M. Let m denote the dimension of Null(Pρ). Then, applying the Gram-Schmidt Algorithm, one can produce an orthonormal basis {H1, . . . , Hm} of Null(Pρ) consisting of self-adjoint elements of M. The map X 7→ Xρ1/2 is unitary from (M,h·,·iGN S,ρ) to (M,h·,·iHS,ρ). Therefore for all A ∈ N , and each j = 1, . . . , m, hAρ1/2, Hjρ1/2iHS = 0. Then since the map X 7→ X ∗ is an (antilinear) isometry on (M,h·,·iHS,ρ), 0 = h(Hjρ1/2)∗, (Aρ1/2)∗,iHS = Tr[HjρA∗] = Tr[Hj∆ρ(A∗)ρ] = hHj, ∆ρ(A∗),iGN S,ρ . For the converse, suppose that N is invariant under ∆ρ. Then N is invariant under ∆s Therefore, ∆ρ(A∗) is orthogonal to Null(Pρ) in (M,h·,·iGN S,ρ), and hence ∆ρ(A∗) ∈ N . Since A is arbitrary in N , it follows that N is invariant under ∆ρ. ρ for all s ∈ R, and in particular, N is invariant under ∆/12 . Then ρ1/2N = N ρ1/2 as subspaces of M; let K denote this subspace of M, which is evidently self-adjoint. Let H = H ∗ ∈ M. Then there are uniquely determined A, B ∈ N such that Hρ1/2 − Aρ1/2 and ρ1/2H − ρ1/2B are both orthogonal to K in the Hilbert-Schmidt inner product. Thus, and ρ ρ1/2H = (ρ1/2H − ρ1/2B) + ρ1/2B Hρ1/2 = (Hρ1/2 − Aρ1/2) + Aρ1/2 are the orthogonal decompositions of Hρ1/2 and ρ1/2H with respect to K. Again since X 7→ Xρ1/2 is unitary from (M,h·,·iGN S,ρ) to (M,h·,·iHS,ρ), Pρ(H) = A. We must show that A = A∗. Since X 7→ X ∗ is an isometry for the Hilbert-Schmidt inner product, and since K is self adjoint, ρ1/2H = (ρ1/2H − ρ1/2A∗) + ρ1/2A∗ is again an orthogonal decomposition of ρ1/2H with respect to K, and by uniqueness, B = A∗. Thus, ρ1/2H = (ρ1/2H − ρ1/2A∗) + ρ1/2A∗ . ρ Now apply ∆−1/2 to both sides to obtain Hρ1/2 = (Hρ1/2 − A∗ρ1/2) + A∗ρ1/2. We claim that Hρ1/2 − A∗ρ1/2 is orthogonal to K. Once this is shown, it will follow that Hρ1/2 = (Hρ1/2 − A∗ρ1/2) + A∗ρ1/2 is the orthogonal decomposition of Hρ1/2 with respect to K. Again by uniqueness of the orthogonal decomposition, it will follow that A = A∗. Hence it remains to show that Hρ1/2 − A∗ρ1/2 is orthogonal to K in the Hilbert-Schmidt inner product. The general element of K can be written as ρ1/2Z for Z inN . Then hρ1/2Z, Hρ1/2 − A∗ρ1/2iHS = Tr[Z ∗(ρ1/2H − ρ1/2A∗)ρ1/2 = hZρ1/2, (ρ1/2H − ρ1/2A∗)iHS . But we have seen above that ρ1/2H − ρ1/2A∗ is orthogonal to K, and Zρ1/2 ∈ K. This proves (1). To prove (2), note first of all that when (4.1)) is valid for all A ∈ N , then ∆ρ preserves N since the right side evidently belongs to N . Now suppose the ∆ρ preserves N . Let A, B ∈ N . Then A∗∆ρN (δρ(B))) ∈ N , and then by the definition of Eτ and cyclicity of the trace, 22 Tr[ρ(A∗ρN ρ−1Bρρ−1 N )] = Tr[A∗ρN ρ−1Bρ] = Tr[ρN (ρ−1BρA∗)] . In the same way, using the fact that (ρ−1BρA) ∈ N and cyclicity of the trace, Tr[ρN (ρ−1BρA∗)] = Tr[BρA∗] = Tr[ρA∗B] . Altogether, hA, ∆ρN (∆−1 in N , ∆ρN (∆−1 and then it follows that ∆t ρ (B)) = B, and hence ∆−1 ρ (B))iGN S,ρ = hA, BiGN S,ρ. Since ∆ρN (∆−1 ρN (B). Then ∆−n ρ (B) = ∆−1 ρ(B) = ∆t ρN (B) for all t ∈ R. ρ (B)) ∈ N , and A is arbitrary ρ (B) = ∆−n ρN (B) for all n ∈ N, Finally, we show that (4.1) is valid for all A ∈ N , then Aρ(A) = A for all A ∈ N : Eτ (ρ1/2Aρ1/2) = Eτ (∆1/2 ρ (A)ρ) = ∆1/2 ρN (A)Eτ (ρ) = ρcN A1/2Aρ1/2 N . By (2) of Theorem 4.1, for all A ∈ N , ∆1/2 ρN (A) = ∆1/2 Eτ,N (ρ1/2Aρ1/2) = Eτ (∆1/2 ρN (A)ρ) = ∆1/2 ρ (A), and therefore ρ (A)Eτ (ρ) = ρ1/2 N Aρ1/2 N . N N Eτ (ρ1/2Aρ1/2)ρ−1/2 That is, A = ρ−1/2 = Aρ(A). On the other hand, when A = Aρ(A) for all A ∈ N , Aρ is a norm one projection onto N , and by Tomiyama's Theorem [37], it is a conditional expectation, and it satisfies ρ(Aρ(X))X for all X ∈ M. Therefore, it must coincide with Pρ, the orthogonal projection form M onto N in the GNS inner product induced by ρ. Hence Pρ is a conditional expectation. By what we proved earlier, this means that N is invariant under ∆ρ, and then that (4.1) is valid for all A ∈ N . 4.3 Theorem. Let Pρ denote the orthogonal projection of M onto N in the GNS inner product induced by ρ. Then (1) Pρ is a conditional expectation if and only if N is invariant under ∆ρ. (2) Pρ is a conditional expectation if and only if Pρ is real. Proof. Theorem 4.1 says that when ∆ρ does not leave N invariant, Pρ is not even real, and hence is not a conditional expectation. On the other hand, when ∆ρ leaves N invariant, a theorem of Takesaki says that there exists a projection E with unit norm from M onto N that satisfies (1.15). By Tomiyama's Theorem and remarks we have made in the introduction, this means that E = Pρ, and that Pρ is a conditional expectation in the sense of Umegaki. This proves (1). It is evident that if Pρ is a conditional expectation, this Pρ is real. On the other hand, if Pρ is real, then by Theorem 4.1, N is invariant under ∆ρ, and now (2) follows from (1). 5 Strong Subadditivity We recall the proof of equivalence of the strong-subadditivity relation and the monotonicity of relative entropy under partial traces, according to [18], in which it is shown that strong sub- additivity relation can be written in the following form: for H = H1 ⊗ H2 ⊗ H3 and ρ123 ∈ B(H), (5.1) S(ρ12ρ1 ⊗ ρ2) ≤ S(ρ123ρ1 ⊗ ρ23), where ρ12 = TrH3ρ123 etc. (See [18]). With N := B(H1 ⊗ H2), S(ρ12ρ1 ⊗ ρ2) = S((ρ123)N(ρ1 ⊗ ρ23)N ) . The DPI inequality yields S(ρNσN ) ≤ S(ρσ), 23 (5.2) for ρ, σ ∈ B(H). 5.1 Lemma. (Lieb, Ruskai [18]) Let H = H1 ⊗ H2 ⊗ H3. The monotonicity of the relative entropy under partial traces holds for all states ρ, σ ∈ B(H) if and only if the strong sub-additivity inequality S(ρ12) + S(ρ23) − S(ρ123) − S(ρ2) ≥ 0. (5.3) holds for all states ρ123 ∈ H1 ⊗ H2 ⊗ H3. Proof. (MONO⇒ SSA) From here it is clear that taking the CPTP to be a partial trace over the third space and ρ = ρ123, σ = ρ1 ⊗ ρ23 in (5.2) leads to (5.1). (SSA⇒ MONO) Let us take the space H3 to be 2-dimensional and the state ρ123 in the following form ρ123 = λρ′ (5.4) where E3 and F3 are orthogonal one-dimensional projections on H3 and λ ∈ [0, 1]. Then the SSA relation (5.1) for this ρ123 is equivalent to the concavity property of the conditional entropy S(ρ12kρ1), i.e. for any λ ∈ [0, 1] and any ρ′ 12 ⊗ E3 + (1 − λ)ρ′′ 12 ⊗ F3, 12 and ρ′′ 2)) + (1 − λ)(S(ρ′′ 12 above we have 12) − S(ρ′′ 2)) ≤ S(ρ12) − S(ρ2). λ(S(ρ′ 12) − S(ρ′ (5.5) Recall that a function f : [0, 1] → R is called operator convex if for all matrices A, B with eigenvalues in [0, 1] and 0 < λ < 1 the following holds f (λA + (1 − λ)B) ≤ λf (A) + (1 − λ)f (B). (5.6) Note that if an operator concave function f is homogeneous (i.e. f (tA) = tf (A) for all t > 0), then for positive matrices A and B 1 t{f (A + tB) − f (A)} ≥ f (B) the above limit exits. To see this, use the homogeneity first, and then the concavity of f in the following way d f (A + tB) := lim t→0 dt(cid:12)(cid:12)(cid:12)(cid:12)t=0 f (A + tB) = (1 + t)f(cid:18) 1 ≥ (1 + t)(cid:18) 1 1 + t = f (A) + tf (B). A + 1 + t t 1 + t t B(cid:19) f (B)(cid:19) f (A) + 1 + t (5.7) (5.8) (5.9) (5.10) 24 Take a conditional entropy as this function f : Then the derivative is f (γ12) := S(γ12) − S(γ2). d dt f (γ12 + tω12) = −Trω12 ln(γ12 + tω12) + Trω2 ln(γ2 + tω2). Since the conditional entropy is concave and homogeneous, applying inequality (5.7) leads to the monotonicity of the relative entropy under partial traces (5.2). The stability bound proved here has obvious consequences for the SSA inequality, and can be used to give a quantitative version of the result [14] of Hayden, Josza, Petz and Winter. For another improvement to the SSA inequality, namely S(ρ12) + S(ρ23) − S(ρ123) − S(ρ2) ≥ 2 max{S(ρ1) − S(ρ13), S(ρ3) − S(ρ13)} , (5.11) see [7]. Acknowledgments. The authors are grateful to Mark Wilde and Lin Zhang for comments and questions that have led us to add some reformulations in this version. EAC was partially supported by NSF grant DMS 1501007. AV is grateful to EAC for hosting her visits to Rutgers University, during which this work was partially completed. AV is partially supported by NSF grant DMS 1812734. References [1] L. Accardi, Non commutative Markov chains, Proc. School of Math. Phys. Camerino (1974). [2] L. Accardi and C. Cecchini, Conditional Expectations in von Neumann algebras and A The- orem of Takesaki Jor. Func. Analysis 45, 245 - 273 (1982) [3] H. Araki, Relative entropy of state of von Neumann algebras, Publ. RIMS Kyoto Univ. 9, 809 - 833 (1976) [4] Bhatia, Matrix analysis, Springer-Verlag, New York, 1997 [5] M.-D. Choi, A Schwarz inequality for positive linear maps on C ∗-algebras, Illinois J. Math 18: 4, 565-574 (1974) [6] E. A. Carlen and E. H. Lieb, Optimal Hypercontractivity for Fermi Fields and Related Non-commutative Integration Inequalities, Comm. Math. Phys., 155, 1993 pp. 27-46. [7] E. A. Carlen and E. H. Lieb, Bounds for entanglement via an extension of strong subaddi- tivity of entropy. Lett. Math. Phys. 101 (2012), no. 1, 1 -- 11 25 [8] E. A. Carlen and E. H. Lieb, Remainder Terms for Some Quantum Entropy Inequalities, Jour. Math. Phys., 55, 042201 (2014) [9] E. A. Carlen and A. Vershynina, Recovery and the Data Processing Inequality for Quasi- Entropies, IEEE Trans. Info. Thy., 64, 6929 - 6938 (2018) [10] M.-D. Choi, Completely positive linear maps on complex matrices, Linear Algebra and its Applications, 10: 3, 285-290 (1975) [11] E. B. Davies, Markovian master equations, Commun. Math. Phys., 39, 91- 110 (1974) [12] C. Davis, Various averaging operations onto subalgebras, Illinois J. Math. 3 538-553 (1959) [13] O. Fawzi, R. Renner, Quantum conditional mutual information and approximate Markov chains, Commun. Math. Phys. 340(2), 2015 [14] P. Hayden, R. Jozsa, D. Petz, A. Winter, Structure of states which satisfy strong subadditivity of quantum entropy with equality, Communications in mathematical physics, 246:2, 359-374 (2004) [15] R. V. Kadison, A generalized Schwarz inequality and algebraic invariants for operator alge- bra, Ann. of Mafh. 56 494 - 503 (1952) [16] E. H. Lieb, Convex trace functions and the Wigner-Yanase-Dyson conjecture, Advances in Math., 11, 267-288 (1973) [17] M. Junge, R. Renner, D. Sutter, M. Wilde, A. Winter, Universal recovery from a decrease of quantum relative entropy arXiv:1509.07127, 2015 [18] E. H. Lieb, M. B. Ruskai, Proof of the strong subadditivity of quantum-mechanical entropy, Journal of Mathematical Physics 14:12, 1938-1941 (1973) [19] G. Lindblad, Expectations and Entropy Inequalities for Finite Quantum Systems, Commun. math. Phys. 39, 111-119 (1974) [20] G. Lindblad, Completely Positive Maps and Entropy Inequalities, Commun. Math. Phys. 40, 147-151 (1975) [21] G. Lindblad The general no-cloning theorem, Lett. Math. Phys., 47:2, 189-196 (1999) [22] M. Koashi, N. Imoto, Operations that do not disturb partially known quantum states, Phys. Rev. A, 66: 2, 022318 (2002) [23] M. Mosonyi and D. Petz, Structure of Sufficient Quantum Coarse-Grainings, Lett. in Math. Phys., 66, 19-30, (2004) 26 [24] M. Nielsen, D. Petz, A simple proof of the strong subadditivity inequality, Quantum Infor- mation & Computation, 6, 507 - 513 (2005) [25] D. Petz, Quasi-entropies for states of a von Neumann algebra, Publ. RIMS. Kyoto Univ. 21, 781-800 (1985) [26] D. Petz, Quasi-entropies for finite quantum systems, Rep. Math. Phys. 23, 57 - 65 (1986) [27] D. Petz, Sufficient subalgebras and the relative entropy of states of a von Neumann algebra, Comm. Math. Phys. 105:1, 123-131 (1986). [28] D. Petz, Sufficiency of channels over von Neumann algebras, Quart. J. Math. Oxford Ser. (2), 39:153, 97-108 (1988) [29] D. Petz, D. Virosztek. Some inequalities for quantum Tsallis entropy related to the strong subadditivity, arXiv:1403.7062 ( 2014) [30] D. W. Robinson, D. Ruelle, Mean Entropy of States in Classical Statistical Mechanis, Com- munications in Mathematical Physics 5, 288 (1967) [31] M. B. Ruskai, Inequalities for Quantum Entropy: A Review with Conditions for Equality, J. Math. Phys. 43, 4358-4375 (2002); erratum 46, 019901 (2005) [32] N. Sharma, Equality Conditions for Quantum Quasi-Entropies Under Monotonicity and Joint-Convexity, Nat. Conf. Commun., 2014 [33] W. F. Stinespring, Positive Functions on C∗-algebras, Proc. American Math. Soc., 6, 211-216 (1955) [34] D. Sutter, M. Berta and M. Tomamichel, Multivariate Trace Inequlaities, Comm. Math. Phys., 352, 37-58 (2017) [35] D. Sutter, Approximate quantum Markov chains, Springer Briefs in Mathematical Physics, Springer, Nerlin, 2018. [36] M. Takesaki, Conditional Expectations in von Neumann Algebras, J. Funct. Anal. 9, 306 - 321. (1972) [37] J. Tomiyama, On the projection of norm. one in W∗-algebras, Proc. Jupnn Acad. 33 608 - 612 (1957) [38] A. Uhlmann, Endlich Dimensionale Dichtematrizen, II, Wiss. Z. Karl-MarxUniversity Leipzig, Math-Naturwiss. 22:2, 139 (1973) [39] A. Uhlmann, The "transition probability" in the state space of a *-algebra, Reports on Math- ematical Physics, 9(2):273?279, (1976) 27 [40] A. Uhlmann. Relative entropy and the Wigner-Yanase-Dyson-Lieb concavity in an interpo- lation theory, Communications in Mathematical Physics 54: 1, 21-32 (1977) [41] H. Umegaki, Conditional expectation in an operator algebra, Tokohu Math. J. 6 177-181. (1954). [42] H. Umegaki, Conditional expectation in an operator algebra, II, Tokohu Math. J. 8, 86-100 (1956), [43] H. Umegaki, Conditional expectation in an operator algebra, III, Kodai Marh. Sem. Rep. 11 51-64 (1959) 39. H. UMEGAKI, Conditional [44] H. Umegaki, Conditional Expectation in an Operator Algebra. IV. Entropy and Information, Kodai Math. Sem. Rep. 14, 59-85 (1962) [45] M. Wilde, Recoverability in quantum information theory, Proc. R. Soc. A. 471(2182) The Royal Society, (2015) [46] L. Zhang, A Strengthened Monotonicity Inequality of Quantum Relative Entropy: A Unifying Approach Via R´enyi Relative Entropy, Letters in Mathematical Physics 106(4): 557-573, (2016)
1209.6227	1	1209	2012-09-27T13:41:32	W*-superrigidity of mixing Gaussian actions of rigid groups	[ "math.OA", "math.FA" ]	We generalize W-superrigidity results about Bernoulli actions of rigid groups to general mixing Gaussian actions. We thus obtain the following: If \Gamma\ is any ICC group which is w-rigid (i.e. it contains an infinite normal subgroup with the relative property (T)) then any mixing Gaussian action \sigma\ of \Gamma\ is W-superrigid. More precisely, if \rho\ is another free ergodic action of a group \Lambda\ such that the crossed-product von Neumann algebras associated with \rho\ and \sigma\ are isomorphic, then \Lambda\ and \Gamma\ are isomorphic, and the actions \rho\ and \sigma\ are conjugate. We prove a similar statement whenever \Gamma\ is a non-amenable ICC product of two infinite groups.	math.OA	math	W∗-SUPERRIGIDITY OF MIXING GAUSSIAN ACTIONS OF RIGID GROUPS R´EMI BOUTONNET Abstract. We generalize W∗-superrigidity results about Bernoulli actions of rigid groups to general mixing Gaussian actions. We thus obtain the following: If Γ is any ICC group which is w-rigid (i.e. it contains an infinite normal subgroup with the relative property (T)) then any mixing Gaussian action Γ y X is W∗-superrigid. More precisely, if Λ y Y is another free ergodic action such that the crossed-product von Neumann algebras are isomorphic L∞(X) ⋊ Γ ≃ L∞(Y ) ⋊ Λ, then the actions are conjugate. We prove a similar statement whenever Γ is a non-amenable ICC product of two infinite groups. 1. Introduction Most known examples of finite von Neumann algebras are constructed from discrete groups or equivalence relations. Thus the question of understanding which data of the initial group or equivalence relation is remembered in the construction of the associated von Neumann algebra is fundamental if one wants to classify finite von Neumann algebras. This problem is usually very hard, but a dramatic progress has been made possible in the last decade thanks to Sorin Popa's deformation/rigidity theory (see [Po07b, Ga10, Va10a] for surveys). The first W∗-rigidity result in the framework of group-measure space constructions is Popa's strong rigidity theorem [Po06a, Po06b]. Assume that Γ y X = X Γ 0 is a Bernoulli action and that Λ y Y is a probability measure preserving (pmp) free ergodic action of an ICC w-rigid group (i.e. which contains an infinite normal subgroup with the relative property (T), [Po06a]). Popa shows in [Po06b] that if the crossed- product von Neumann algebras of these actions are isomorphic, then the actions are conjugate. This is the first result that deduces conjugacy of two actions out of an isomorphism of their crossed product von Neumann algebra. Later on, Ioana managed to prove the following very general W∗-superrigidity result about Bernoulli shifts, which is a natural continuation to Popa's strong rigidity result. For more historical information and results on W ∗-superrigidity, see for instance [Pe09, PV09, HPV10, Io11, IPV11] and the introductions therein. Theorem (Ioana, [Io11]). Let Γ be an ICC w-rigid group. Then the Bernoulli action Γ yσ [0, 1]Γ is W∗-superrigid: Let ρ be any free ergodic measure-preserving action of any group Λ. Denote by M and N the crossed-product von Neumann algebras associated with σ and ρ respectively, and assume that M ≃ N. Then Λ is isomorphic to Γ, and the actions σ and ρ are conjugate. The rigidity of the von Neumann algebra in the theorem above comes from a tension between the property (T) of the group Γ and the deformability of Bernoulli actions. 1 2 R ´EMI BOUTONNET This tension is exploited via Popa's deformation/rigidity strategy. Using a similar strategy of proof, Ioana, Popa and Vaes later proved the W∗-superrigidity of Bernoulli actions for other groups, relying this time on the spectral gap type rigidity discovered by Popa in [Po08]. Theorem (Ioana-Popa-Vaes, [IPV11]). Let Γ be a non-amenable ICC group which is the product of two infinite groups Γ = Γ1 × Γ2. Then the Bernoulli action Γ yσ [0, 1]Γ of Γ is W∗-superrigid. Both of the proofs of Ioana's theorem and Ioana-Popa-Vaes' theorem seemed to deeply rely on the very particular structure of Bernoulli actions. We will show that this is not the case, and generalize these results to Gaussian actions. Let Γ be a countable group and π : Γ → O(H) an orthogonal representation of Γ on a real Hilbert space H. Recall that there exist (see [PS10] for instance) a standard probability space (X, µ) and a pmp action of Γ on X, such that H ⊂ L2(X), as repre- sentations of Γ. This action is called the Gaussian action induced by the orthogonal representation π. We generalize Ioana's result as follows. Theorem A. Let Γ be an ICC w-rigid group and π : Γ → O(H) any mixing orthogonal representation of Γ. Then the Gaussian action σπ associated with π is W∗-superrigid. However, in order to apply Popa's spectral gap argument, one has to make an extra assumption on the initial representation π. Ioana-Popa-Vaes' theorem then becomes a particular case of the following result. Theorem B. Let Γ be a non-amenable ICC group which is the product of two infinite groups, and consider a mixing orthogonal representation π : Γ → O(H) of Γ. Assume that some tensor power of π is weakly contained in the regular representation. Then the Gaussian action σπ associated with π is W∗-superrigid. To prove Theorem A and Theorem B, we will adapt the proof used by Ioana and Ioana-Popa-Vaes to the context of Gaussian action. Let us recall the general strategy of their proof. Steps of the proof in the Bernoulli case. Let Γ be a group as in theorem A or B and Γ y X = [0, 1]Γ the corresponding Bernoulli action. Assume that Λ y (Y, ν) is another pmp, free ergodic action such that Put A = L∞(X), B = L∞(Y ) and M = A ⋊ Γ. L∞(X) ⋊ Γ = L∞(Y ) ⋊ Λ. Thanks to Popa's orbit equivalence superrigidity theorems [IPV11, 5.2 and 5.6] and [Po08, Theorem 1.3], one only has to show that the two actions are orbit equivalent. More concretely it is enough to prove, by a result of Feldman and Moore [FM77], that B is unitaly conjugate to A inside M. The main idea of the proof, due to Ioana, is to exploit the information given by the isomorphism M = B ⋊ Λ via the dual co-action ∆ : M → M ⊗ M 7→ bvs ⊗ vs, bvs W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 3 b ∈ B, s ∈ Λ (vs, s ∈ Λ, denote the canonical unitaries corresponding to the action of Λ). This ∗-homomorphism ∆ allows us to play against each other two data of the single action Γ y X: the rigidity of ∆(LΓ), and the malleability of the algebra M ⊗ M = (A⊗ A) ⋊ (Γ × Γ). Assume that B is not unitarily conjugate to A, or equivalently that B ⊀M A by [Po06c, Theorem A.1]. We refer to Section 2.1 for the definition of Popa's intertwining symbol "≺". The rest of the proof can be divided into four steps, which lead to a contradiction. Step (1). One shows that there exists a unitary u ∈ M ⊗ M such that u∆(LΓ)u∗ ⊂ LΓ⊗ LΓ. Step (2). One further proves that the algebra C := ∆(A)′ ∩ (M ⊗ M) satisfies C ≺M ⊗ M A⊗ A. Step (3). The previous steps, and an enhanced version of Popa's conjugacy criterion [Po06b, Theorem 5.2] imply that roughly there exist a unitary v ∈ M ⊗ M, a group homomorphism δ : Γ → Γ × Γ, and a character ω : Γ → C such that v∆(C)v∗ = A⊗ A and v∆(ug)v∗ = ω(g)uδ(g), ∀g ∈ Γ. Step (4). Using Step (3), one can now show that if a sequence (xn) in M has Fourrier coefficient (with respect to the decomposition M = A ⋊ Γ) which tend to zero pointwise in norm k · k2, then this is also the case of the sequence ∆(xn), with respect to the decomposition M ⊗ M = (M ⊗ A) ⋊ Γ. This easily contradicts the fact that B ⊀M A. What has to be adapted. First note that Popa's orbit equivalence superrigidity results ([IPV11, 5.2 and 5.6] and [Po08, Theorem 1.3]) are still valid for Gaussian actions as in Theorem A or Theorem B. Thus we only have to prove Steps (1)-(4) for such Gaussian actions. Steps (3) and (4) are very general, and will work for any mixing action satisfying the conclusions of steps (1) and (2). Step (1) is the result of Popa's deformation/rigidity strategy so it should not be specific to Bernoulli shifts. In [IPV11], it was a direct consequence of [IPV11, Corollary 4.3]. Using the results in [Bo12], one can easily get the Gaussian counterpart of [IPV11, Corollary 4.3], namely Corollary 2.8. Finally, Step (2) relies on a beautiful localization theorem due to Ioana, [Io11, Theorem 6.1]. That theorem states that if D is an abelian subalgebra of M ⊗ M which is normalized by "enough" unitaries in L(Γ × Γ), then either D′ ∩ (M ⊗ M) ≺ A⊗ A, or D ≺ M ⊗ LΓ, or D ≺ LΓ⊗ M. This theorem is applied to D = ∆(A). Using mixing properties, and the fact that B ⊀ A, the last two cases cannot hold and Step (2) follows. So the point of the whole proof of theorems A and B is to generalize Ioana's localization theorem [Io11, Theorem 6.1]. It will be done in Section 3, Theorem 3.1. We explain below the main difficulties to obtain such a generalization. 4 R ´EMI BOUTONNET Main difficulties in the generalization. Unlike Bernoulli shifts, general mixing Gaussian actions do not satisfy the following properties, which were crucial in Ioana's argument. • Cylinder structure: If Γ y [0, 1]Γ is a Bernoulli action, we call finite cylinder subalgebra a subalgebra of A = L∞([0, 1]Γ) of the form AF = L∞([0, 1]F ), for some finite subset F ⊂ Γ. Then the union of all finite cylinder subalgebras is a strongly dense ∗-subalgebra A0 of A, which is stable under the action of Γ. In fact, A0 is a graded CΓ-module; • "strong compactness" property: If Γ y [0, 1]Γ is a Bernoulli action, there exist a strongly dense ∗-subalgebra A0 of L∞([0, 1]Γ) such that for any a, b ∈ A0⊖ C, hσg(a), bi = 0 if g ∈ Γ large enough. The use of the strong compactness property can be avoided using ε-orthogonality and a trick involving convex combinations. To avoid using the cylinder structure, the idea is to replace cylinders by general finite dimensional subsets of L∞(X), and use a multiple mixing property automatically enjoyed by mixing Gaussian actions. Definition. A trace-preserving action Γ yσ A of a countable group on an abelian von Neumann algebra is 2-mixing if for any a, b, c ∈ A, the quantity τ (aσg(b)σh(c)) tends to τ (a)τ (b)τ (c) as g, h, g−1h tend to infinity. In fact, each step of the proof still holds for general s-malleable actions (in the sense of Popa [Po08]) which are 2-mixing: Theorem C. Let Γ be an ICC group and Γ yσ (X, µ) be a free ergodic action of Γ. Assume that σ is 2-mixing and s-malleable in the sense of Popa [Po08] and that one of the following two conditions holds. • Γ is w-rigid or • Γ is non-amenable and is isomorphic to the product of two infinite groups, and some tensor power of the Koopman representation σL2(X)⊖C is weakly contained in the regular representation of Γ. Then σ is W∗-superrigid. Organization of the article. Apart from the introduction, this article contains three other sections. Section 2 is the preliminary section, in which we recall Popa's intertwining techniques, and several facts on Gaussian actions that we proved in [Bo12]. Section 3 is devoted to proving Theorem 3.1, which generalizes [Io11, Theorem 6.1]. Finally, an extra-section is devoted to constructing a large class of II1-factors which are not isomorphic to group von Neumann algebras. This will be a direct application of Theorem 3.1 and the work of Ioana-Popa-Vaes. Acknowledgement. We warmly thank Cyril Houdayer and Adrian Ioana for their valuable advice and comments about this work. 2. Preliminaries 2.1. Intertwining by bimodules. We recall here an essential tool introduced by Popa, the so-called intertwining by bimodules' theorem. W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 5 Theorem 2.1 (Popa, [Po06a, Po06d]). Let P, Q ⊂ M be finite von Neumann algebras (with possibly non-unital inclusions). Then the following are equivalent. • There exist projections p ∈ P , q ∈ Q, a normal ∗-homomorphism ψ : pP p → qQq, and a non-zero partial isometry v ∈ pMq such that xv = vψ(x), for all x ∈ pP p; • There exists a P -Q subbimodule H of L2(1P M1Q) which has finite index when regarded as a right Q-module; • There is no net of unitaries (ui) ∈ U(P ) such that for all x, y ∈ M, kEQ(1Qx∗uiy1Q)k2 → 0. Following [Po06a], if P, Q ⊂ M satisfy these conditions, we say that a corner of P embeds into Q inside M, and we write P ≺M Q. Assume that we are in the concrete situation where M is of the form M = B ⋊ Γ for some trace preserving action of Γ on a finite von Neumann algebra and Q = B. Denote by ug, g ∈ Γ the canonical unitaries in M implementing the action of Γ. Then it is easy to check that a subalgebra P ⊂ M satisfies P ⊀ B if and only if there exists a net of unitaries vi ∈ U(P ) such that kEB(viu∗ This result can be improved as follows. g)k2 → 0, ∀g ∈ Γ. Lemma 2.2 (Ioana, [Io11], Theorem 1.3.2). Let Γ y B be a trace preserving action on a finite von Neumann algebra (B, τ ). Put M = B ⋊ Γ, and let P ⊂ M be a von Neumann subalgebra. Then P ⊀ B if and only if there exists a net of unitaries vi ∈ U(P ) such that lim n (cid:18)sup g∈Γ kEB(viu∗ g)k2(cid:19) = 0. Another natural question one may wonder: What does it mean to embed into the group algebra LΓ inside a crossed-product algebra M = A ⋊ Γ? In some specific circumstances, this implies the unitary conjugacy into LΓ, as the following standard result shows. We denote by NM (Q) = {u ∈ U(M), uQu∗ = Q} the normalizer of a subalgebra Q of a von Neumann algebra M. The quasi-normalizer QN M (Q) of Q in M is the -subalgebra of M formed by Q-Q finite elements. We recall that an element x ∈ M is Q-Q finite if there exist x1,· · · , xk ∈ M such that Xi=1 Qxi and Qx ⊂ xQ ⊂ Xi=1 xiQ. k k Proposition 2.3. Let Γ y A be a free mixing action of an ICC group Γ on an abelian von Neumann algebra, and let N be a type II1 factor. Put M = (A ⋊ Γ)⊗ N, and assume that Q ⊂ pMp is a von Neumann subalgebra such that Q ⊀M 1 ⊗ N. Put P = QN pM p(Q)′′. (1) If Q ≺M LΓ⊗ N then there exists a non-zero partial isometry v ∈ pM such (2) If rQ ≺M LΓ⊗ N for all r ∈ Q′ ∩ pMp then there exists a unitary u ∈ U(M) that vv∗ ∈ Z(P ) and v∗P v ⊂ LΓ⊗ N. such that uP u∗ ⊂ LΓ⊗ N. 6 R ´EMI BOUTONNET Proof. (1) By assumption, there exist projections p0 ∈ Q, q ∈ LΓ⊗ N, a non-zero partial isometry v ∈ p0Mq and a -homomorphism ϕ : p0Qp0 → q(LΓ⊗ N)q such that for all x ∈ p0Qp0, one has xv = vϕ(x). By [Va08, Remark 3.8], one can assume that ϕ(p0Qp0) ⊀M 1 ⊗ N. Hence [Po06a, Theorem 3.1] implies that QN qM q(ϕ(p0Qp0))′′ ⊂ LΓ⊗ N. But we see that v∗P v ⊂ QN qM q(ϕ(p0Qp0))′′. Moreover vv∗ ∈ p0(Q′ ∩ M) ⊂ P . However vv∗ is not necessarily in Z(P ) but one can modify v as follows to obtain such a condition. i=1 viv∗ Take partial isometries v1,· · · , vk ∈ P such that v∗ is a central projection in P . Since LΓ⊗ N is a factor, there exist partial isometries w1,· · · , wk ∈ LΓ⊗ N such that wiw∗ j = 0, for all 1 ≤ i 6= j ≤ k. Define a non-zero partial isometry by w = Pi vivwi ∈ pM. We get i vi ≤ vv∗, i = 1,· · · , k andPk i viv and wiw∗ i = v∗v∗ i • ww∗ = Pi vivwiw∗ • w∗P w ⊂ Pi w∗ i v∗v∗ i = Pi viv∗ i v∗P vwi ⊂ LΓ⊗ N. i ∈ Z(P ); (2) Consider a maximal projection r0 ∈ Q′ ∩ pMp for which there exists a unitary u ∈ U(M) such that u(r0P r0)u∗ ⊂ LΓ⊗ N. One has to show that r0 = p. Otherwise we can cut by r = p−r0, and we obtain an algebra rQ ⊂ rMr such that rQ ≺M LΓ⊗ N and rQ ⊀M 1⊗N. Remark that rP r ⊂ QN rM r(rQ)′′. Applying (1), we get that there exists a non-zero partial isometry v ∈ rM, such that vv∗ ∈ (rP r)′∩rMr ⊂ (rQ)′∩rMr and v∗(rP r)v ⊂ LΓ⊗ N. Since LΓ⊗ N is a factor, modifying v if necessary, one can assume that v∗v ⊥ ur0u∗. Now the following "cutting and pasting" argument contradicts the maximality of r0. The partial isometry w0 = ur0 + v∗ satisfies w∗ 0w0 = r0 + vv∗ ∈ Q′ ∩ pMp and w0(r0 + vv∗)Qw∗ 0 ⊂ LΓ. Extending w0 into a unitary, we obtain a w ∈ U(M) satisfying w(r0 + vv∗)Qw∗ ⊂ LΓ. 2.2. Gaussian actions. To any orthogonal representation π : Γ → O(H) of a discrete countable group, one can associate a trace preserving action σπ : Γ y A on an abelian von Neumann algebra, called the Gaussian action associated with π. This Gaussian action can be constructed as follows. For more explicit constructions, see [BHV08, Appendix A.7] or [PS10]. Consider the unique abelian finite von Neumann algebra (A, τ ) generated by unitaries (w(ξ))ξ∈H such that: (cid:3) • w(0) = 1 and w(ξ + η) = w(ξ)w(η), w(ξ)∗ = w(−ξ), for all ξ, η ∈ H; • τ (w(ξ)) = exp(−kξk2), for all ξ ∈ H. It is easy to check that these conditions imply that the vectors (w(ξ))ξ∈HR are linearly independent and span a weakly dense ∗-subalgebra of A. Then the Gaussian action σπ is defined by (σπ)g(w(ξ)) = w(π(g)ξ), for all g ∈ Γ, ξ ∈ H. As explained in [Fu07] or [PS10], Gausssian actions are s-malleable in the sense of Popa [Po08]: the rotation operators θt, t ∈ R on H ⊕ H and the symmetry ρ defined by θt = (cid:18)cos(πt/2) − sin(πt/2) cos(πt/2) (cid:19) and ρ = (cid:18)1 sin(πt/2) 0 0 −1(cid:19) give rise to a one-parameter group of automorphisms αt and an automorphism β of A⊗ A, which commute with the diagonal action of Γ, and satisfy β ◦ αt = α−t ◦ β. Now consider the von Neumann algebras M = A ⋊ Γ and M = (A⊗ A) ⋊σ⊗σ Γ. View M as a subalgebra of M using the identification M ≃ (A⊗ 1)⋊Γ. The automorphisms W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 7 defined above then extend to automorphisms of M still denoted (αt) and β, in such a way that αt(ug) = β(ug) = ug, for all g ∈ Γ. (αt) is then easily seen to be an s-malleable deformation of the action, so it satisfies the so-called transversality property. Lemma 2.4 ([Po08], Lemma 2.1). For any x ∈ M and t ∈ R one has kx − α2t(x)k2 ≤ 2kαt(x) − EM ◦ αt(x)k2. With more conditions on the representation π, we also get the spectral gap property. Lemma 2.5 (Spectral gap, [Bo12]). Assume that the representation π is such that π⊗l is weakly contained in the regular representation for some l ≥ 1. Let ω ∈ βN \ N be a free ultrafilter on N. Then for every subalgebra Q ⊂ M with no amenable direct summand, one has Q′ ∩ M ω ⊂ M ω In fact this lemma admits a relative version. Recall from [OP07] that if (M, τ ) is a finite von Neumann algebra, p ∈ M a projection, and Q ⊂ M and P ⊂ pMp are subalgebras, one says that P is amenable relative to Q inside M if there exists a P -central state ϕ on phM, eQip such that ϕ(pxp) = τ (pxp)/τ (p) for any x ∈ M. Here hM, eQi denotes Jones' basic construction associated with the inclusion Q ⊂ M. Following [IPV11, Section 2.4], P is said to be strongly non-amenable relative to Q if for all non-zero projection p1 ∈ P ′ ∩ pMp, P p1 is not amenable relative to Q. Lemma 2.6. Assume that the representation π is such that π⊗l is weakly contained in the regular representation for some l ≥ 1. Let ω ∈ βN\ N be a free ultrafilter on N. Then for any subalgebra Q ⊂ M ⊗ N which is strongly non-amenable relative to 1⊗N, one has Q′ ∩ ( M ⊗ N)ω ⊂ Q′ ∩ (M ⊗ N)ω. 2.3. Deformation/rigidity results for Gaussian actions. We mention here dif- ferent versions of statements that we proved in [Bo12] using deformation/rigidity ar- guments. The following result is a variation of [Bo12, Theorem 3.4], with a formulation closer to [IPV11, Theorem 4.2]. Theorem 2.7. Assume that Γ y A is the Gaussian action associated with a mixing representation of an ICC group Γ. Let N be a II1 factor. Put M = A ⋊ Γ and define (αt) as in section 2.2. Let p ∈ M ⊗ N and Q ⊂ p(M ⊗ N)p be a subalgebra such that there exist t0 = 1/2n, z ∈ M ⊗ N and c > 0 satisfying τ ((αt0 ⊗ id)(u∗)zu) ≥ c, for all u ∈ U(Q). Put P = QN p(M ⊗ N )p(Q)′′. Then at least one of the following assertions holds. (1) Q ≺ 1 ⊗ N; (2) P ≺ A⊗ N; (3) There exists a non-zero partial isometry v ∈ pM such that vv∗ ∈ Z(P ) and v∗P v ⊂ LΓ⊗ N. 8 R ´EMI BOUTONNET Proof. Assume that (2) is not satisfied. Using the fact that π is mixing, the same proof that the one of [Bo12, Theorem 3.4] gives that Q ≺ LΓ⊗ N. Now if Q ⊀ 1⊗ N, Proposition 2.3(1) implies that (3) holds true. (cid:3) Now one can deduce the Gaussian version of [IPV11, Corollary 4.3], which implies Step (1) of the proof of Theorems A and B. Corollary 2.8. Assume that Γ is ICC and let Γ y A be a mixing Gaussian action. Put M = A ⋊ Γ. Let N be a II1 factor and Q ⊂ p(M ⊗ N)p a subalgebra, for some p ∈ M ⊗ N. Assume that we are in one of the following situations: • Q ⊂ p(M ⊗ N)p has the relative property (T); • Q′ ∩ p(M ⊗ N)p is strongly non-amenable relative to 1⊗ N (see end of Section 2.2), and some tensor power of π is weakly contained in the regular represen- tation of Γ. Denote by P = QN p(M ⊗ N )p(Q)′′. Then one of the following assertions is true. (1) Q ≺ 1 ⊗ N; (2) P ≺ A⊗ N; (3) There exists a unitary v ∈ M ⊗ N such that v∗P v ⊂ LΓ⊗ N. Proof. The assumptions imply that the deformation αt ⊗ id converges to the identity Indeed, if Q ⊂ p(M ⊗ N)p has uniformly on (Q)1 = {x ∈ Q,kxk ≤ 1} in k · k2. relative property (T), this is almost by definition. If Q′ ∩ p(M ⊗ N)p is strongly non- amenable relative to 1 ⊗ N, then this is a consequence of spectral gap lemma 2.6, and transversality property 2.4 (see the proof of [Po08, Lemma 5.2]). Hence, for all r ∈ Q′ ∩ pMp, the subalgebra rQ ⊂ rMr satisfies the assumpions of Theorem 2.7. Then if Q ⊀ 1 ⊗ N and P ⊀ A⊗ N, Theorem 2.7 applied to all such rQ's implies in particular that for all r ∈ Q′ ∩ pMp, rQ ≺ LΓ⊗ N. Now (3) follows from Lemma 2.3(2). (cid:3) In [Bo12], we also obtained a localization result (Theorem 3.8) for subalgebras of M that commute inside M ω with rigid subalgebras of M ω, for some free ultrafilter ω on N. In fact, the same proof leads to the following improvement. We include a sketch of the proof for convenience. Theorem 2.9. Let Γ y A be a mixing Gaussian action. Put M = A ⋊ Γ and consider a II1 factor N. Assume that (vn) is a bounded sequence of elements in M ⊗ N such that αt ⊗ id converges to the identity uniformly on the set {vn, n ∈ N}. Choose a free ultrafilter ω on N, and denote by D ⊂ M ⊗ N ⊂ (M ⊗ N)ω the subalgebra of elements that commute with the element (vn)n ∈ (M ⊗ N)ω. Put P = QN M ⊗ N (D)′′. Then one of the following is true. (1) (vn)n ∈ (A⊗ N)ω ⋊ Γ; (2) D ≺ LΓ⊗ N; (3) P ≺M A⊗ N. Sketch of proof. Assume that (vn)n /∈ (A⊗ N)ω ⋊ Γ. We will show that the D satisfies the assumptions of Theorem 2.7. Define x = (xn) = (vn) − E(A ⊗ N )ω ⋊Γ((vn)) 6= 0. Dividing x if necessary by kxk2, one can assume that kxk2 ≤ 1. For F ⊂ Γ finite, denote by PF : L2(M) → L2(M) the W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 9 projection onto the closed linear span of elements of the form xug, x ∈ A g ∈ F . One checks that: • αt ⊗ id converges to identity uniformly on {xn, n ∈ N}; • limn k[xn, u]k2 → 0, for any u ∈ U(D); • limn k(PF ⊗ id)(xn)k2 → 0, for any finite subset F ⊂ Γ. Using [Va10b, Lemma 3.8], one can show that this last condition implies that n hxnξx∗ lim n, ηi = 0,∀ξ, η ∈ (L2( M ) ⊖ L2(M)) ⊗ L2(N). Fix ε > 0. Then there exists a t = 1/2k such that k(αt ⊗ id)(xn) − xnk2 < ε, ∀n. Fix u ∈ U(D) and put δt(u) = (αt ⊗ id)(u) − EM ⊗N ((αt ⊗ id)(u)). Then δt(u) ∈ (L2( M) ⊖ L2(M)) ⊗ L2(N), and we get n kδt(u)xnk2 lim 2 ≈2ε lim ≈2ε lim ≈4ε lim n hδt(uxn), δt(u)xni n hδt(xnu), δt(u)xni n hxnδt(u)x∗ n, δt(u)i = 0. Thus limn kδt(u)xnk2 one shows that for all n ∈ N, 2 ≤ 4ε. But exactly as in the proof of Popa's transversality lemma, k(α2t ⊗ id)(u)xn − uxnk2 ≤ k(αt ⊗ id)(u)xn − (α−t ⊗ id)(u)xnk2 + 2ε Hence, if ε < 1, we get that limn k(α2t ⊗ id)(u)xn − uxnk2 < 6√ε. Put z = EM ⊗ N ((xnx∗ n)n). We have ≤ 2kδt(u)xnk2 + 2ε. 2 lim n kxnk2 2 − 2ℜ(τ ((α2t ⊗ id)(u∗)zu)) = lim n k(α2t ⊗ id)(u)xn − uxnk2 2 < 36ε. If ε was chosen to be small enough, this implies the result. (cid:3) Corollary 2.10. For i = 1, 2, consider mixing Gaussian actions Γi y Ai and put Mi = Ai ⋊ Γi, A = A1 ⊗ A2, Γ = Γ1 × Γ2 and M = M1 ⊗ M2 = A ⋊ Γ. For i = 1, 2, define Mi and (αi equipped with the deformation (αt) = (α1 t), as in section 2.2, and denote by M = M1 ⊗ M2 t ⊗ α2 t ). Assume that (vn) is a bounded sequence of elements in M such that αt converges uniformly to the identity on the set {vn, n ∈ N}. Choose a free ultrafilter ω on N, and denote D ⊂ M ⊂ M ω the subalgebra of elements that commute with the element (vn)n ∈ M ω. Put P = QN M (D)′′. Then one of the following is true. (1) (vn)n ∈ Aω ⋊ Γ; (2) D ≺M LΓ1 ⊗ M2 or D ≺M M1 ⊗ LΓ2; (3) P ≺M A1 ⊗ M2 or P ≺M M1 ⊗ A2. Proof. Exactly as in the proof of [Io11, Theorem 3.2] Claim 2, we get that if αt con- verges uniformly on {vn, n ∈ N}, then so do α1 t . Thus if (2) and (3) are not satisfied, Theorem 2.9 implies that (vn) ∈ ((A1 ⊗ M2)ω ⋊ Γ1)∩((M1 ⊗ A2)ω ⋊ Γ2) = Aω ⋊ Γ. t ⊗id and id⊗α2 (cid:3) 10 R ´EMI BOUTONNET 2.4. 2-mixing property. Definition 2.11. A trace-preserving action Γ yσ A of a countable group on an abelian von Neumann algebra is said to be 2-mixing if for any a, b, c ∈ A, the quantity τ (aσg(b)σh(c)) tends to τ (a)τ (b)τ (c) as g, h, g−1h tend to infinity. Proposition 2.12. An action Γ yσ A is 2-mixing if and only if for all a, b, c ∈ A, one has τ (aσg(b)σh(c)) − τ (a)τ (σg(b)σh(c)) → 0, when g → ∞, h → ∞. Proof. The if part is straightforward. For the converse, assume that σ is 2-mixing. It is sufficient to show that if a, b, c ∈ A, with τ (a) = 0, then τ (aσg(b)σh(c)) → 0, as g, h → ∞. Assume by contradiction that there exist sequences gn, hn ∈ Γ going to infinity, and δ > 0 such that τ (aσgn(b)σhn(c)) ≥ δ, for all n. Then two cases are possible: Case 1. The sequence g−1 if necessary, one can assume that g−1 n hn is contained in a finite set. Then taking a subsequence n hn = k is constant. Then for all n, we get But since σ is mixing this quantity tends to 0 as n tends to infinity. τ (aσgn(b)σhn(c)) = τ (aσgn(bσk(c)). Case 2. The sequence g−1 if necessary, one can assume that g−1 that τ (aσgn(b)σhn(c)) → 0. In both cases, we get a contradiction. n hn is not contained in a finite set. Then taking a subsequence n hn → ∞ when n → ∞. Then the 2-mixing implies (cid:3) Of course any 2-mixing action is mixing. The converse holds for Gaussian actions. Proposition 2.13. If Γ yσ A is the Gaussian action associated with a mixing repre- sentation π on H, then σ is 2-mixing. Proof. By a linearity/density argument, it is enough to prove that for all ξ, η, δ ∈ H, and all sequences gn, hn ∈ Γ tending to infinity, one has lim n [τ (ω(ξ)σgn(ω(η))σhn(ω(δ))) − τ (ω(ξ))τ (σgn(ω(η))σhn(ω(δ)))] = 0. But one checks that: • τ (ω(ξ)σgn(ω(δ))σhn(ω(δ))) = exp(−kξ + π(gn)η + π(hn)δk2); • τ (ω(ξ))τ (σgn(ω(η))σhn(ω(δ))) = exp(−kξk2 − kπ(gn)η + π(hn)δk2). The difference is easily seen to tend to 0 as n → ∞. (cid:3) 3. The key step We now state the key theorem from which Theorems A and B follow as explained in the introduction. W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 11 Theorem 3.1. For i = 1, 2, consider mixing Gaussian actions Γi y Ai of discrete countable groups Γi, and put Mi = Ai ⋊ Γi, A = A1 ⊗ A2, Γ = Γ1 × Γ2 and M = M1 ⊗ M2 = A ⋊ Γ. Let t > 0. Realize (LΓ)t ⊂ M t by fixing an integer n ≥ t and p ∈ LΓ ⊗ Mn(C) with trace t/n. Let D ⊂ M t be an abelian subalgebra, and denote by Λ = NM t(D)∩U((LΓ)t) and make the following assumptions: (i) Λ′′ ⊀M LΓ1 ⊗ 1 and Λ′′ ⊀M 1 ⊗ LΓ2; (ii) D ⊀ LΓ1 ⊗ M2 and D ⊀M M1 ⊗ LΓ2. Denote by C = D′ ∩ M t. Then for all projections q ∈ Z(C), Cq ≺M A. Proof. Exactly as in the proof of [IPV11, Theorem 5.1], we first show that it is sufficient to prove that C ≺M A. Indeed, assume that we have shown that the assumptions of the theorem imply that C ≺M A. Consider the set of projections P = {q1 ∈ Z(C) Cq ≺M A, for all non-zero subprojections q ∈ Z(C)q1}. Then P admits a unique maximal element p1 ∈ Z(C). By uniqueness, p1 commutes with the normalizer of C, and in particular with Λ′′. Using [Va10b, Lemma 3.8] and assumption (i), we get that p1 ∈ (LΓ)t. We want to show that p1 = p(= 1C). Otherwise, we can cut by p−p1 and we see that (p−p1)D ⊂ (p−p1)(M⊗Mn(C))(p−p1) satisfies the assumptions of the theorem. Thus (p − p1)C ≺M A. This contradicts the maximality of p1. So the rest of the proof is devoted to showing that C ≺M A. As in the proof of [IPV11, Theorem 5.1], we assume that t ≤ 1, so that n can be chosen to be equal to 1. This assumption largely simplifies notations, and does not hide any essential part of the proof. Note that the assumption (i) implies that there exists a sequence of unitary elements vn ∈ U(pLΓp) that normalize D and such that (3.1) We will proceed in two steps to prove that C ≺M A. In a first step we collect properties regarding the sequence (vn) or sequences of the form (vnav∗ n), a ∈ D. In the second step we show the result, reasoning by contradiction. Before moving on to these two steps, we introduce some notations: kELΓ1⊗1(avnb)k2 → 0 and kE1⊗LΓ2(avnb)k2 → 0, ∀a, b ∈ M. of Γ; its Fourier decomposition. • We denote by ug, g ∈ Γ the canonical unitaries in M implementing the action • For any element x ∈ M, we denote by x = Pg∈Γ xgug (xg ∈ A for all g ∈ Γ) • If S ⊂ Γ is any subset, denote by PS : L2(M) → L2(M) the projection onto the linear span of the vectors aug, a ∈ A, g ∈ S. • If K ⊂ A is a closed subspace, we denote by QK : L2(M) → L2(M) the projection onto the linear span of the vectors aug, a ∈ K, g ∈ Γ. Warning for the sequel: "g, h ∈ Γ" means that g and h are two elements (g1, g2) and (h1, h2) of the product group Γ = Γ1 × Γ2. It is different from "(g, h) ∈ Γ"! 12 R ´EMI BOUTONNET Step 1: Properties of the sequences (vnav∗ Lemma 3.2. For any free ultrafilter ω on N, and any a ∈ D, the element (vnav∗ M ω belongs to Aω ⋊ Γ. n), a ∈ D. n)n ∈ Proof. We will apply Corollary 2.10. Fix a ∈ D. Since the vn's are in LΓ, the deformation αt introduced in the statement of Corollary 2.10 converges uniformly on the set {vnav∗ n, n ∈ N}. Thus Corollary 2.10 implies that one of the following holds true: n)n ∈ Aω ⋊ Γ; • (vnav∗ • D ≺M LΓ1 ⊗ M2 or D ≺M M1 ⊗ LΓ2; • P ≺M A1 ⊗ M2 or P ≺M M1 ⊗ A2, where P = NpM p(D)′′. The second case is excluded by assumption, so we are left to showing that the third case is not possible. By symmetry, it is sufficient to show that P ⊀M A1 ⊗ M2. But we claim that for all x, y ∈ M, kEA1 ⊗ M2(xvny)k2 → 0. Since vn ∈ U(P ), this claim implies the result. By Kaplansky's density theorem, and by linearity it is sufficient to prove the claim for x and y of the form ug ⊗ 1, g ∈ Γ1. In particular xvny lies in LΓ. So using the fact that A1 ⊗ M2 ⊂ M ∪ 1 ⊗ LΓ2 ⊂ LΓ ∪ is a commuting square, 3.1 directly implies that kEA1 ⊗ M2(xvny)k2 → 0. For an element x ∈ M = LΓ, denote by h(x) the height of x: h(x) = supg∈Γ xg, where x = P xgug is the Fourier decomposition of x. Lemma 3.3. There exists δ > 0 such that h(vn) > δ for all n. (cid:3) Proof. Assume that the result is false. Taking a subsequence if necessary, we get that h(vn) → 0. Then we claim that for all finite subset S ⊂ Γ, and all a ∈ (M ⊖ (LΓ1 ⊗ M2)) ∩ (M ⊖ (M1 ⊗ LΓ2)), n kPS(vnav∗ lim n)k2 = 0. Note that (M ⊖ (LΓ1 ⊗ M2))∩ (M ⊖ (M1 ⊗ LΓ2)) is the subset of elements in M whose Fourier coefficients lie in the weak closure of (A1 ⊖ C1) ⊗ (A2 ⊖ C1). By a linearity/density argument, to prove this claim it is sufficient to show that for any sequence of unitaries wn ∈ U(pLΓp) and a ⊗ b ∈ (A1 ⊖ C1) ⊗ (A2 ⊖ C1), kEA(vn(a ⊗ b)wn)k2 → 0. Write vn = Pg∈Γ vn,gug and wn = Pg∈Γ wn,gug. We have EA(vn(a ⊗ b)wn) = Xg∈Γ vn,gwn,g−1σg(a ⊗ b), which leads to the formula: kEA(vn(a ⊗ b)wn)k2 (3.2) 2 = Xg,g′∈Γ vn,gwn,g−1vn,g′wn,g′−1τ (σg(a ⊗ b)σg′(a∗ ⊗ b∗)). W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 13 Fix ε > 0. Since the action Γi y Ai is mixing for i = 1, 2, there exist finite sets Fi ⊂ Γi such that τ ((a⊗ b)σ(s,t)(a∗⊗ b∗)) = τ (aσs(a∗))τ (bσt(b∗)) < ε, if (s, t) /∈ F = F1× F2. Now 3.2 and Cauchy-Schwarz inequality imply kEA(vn(a ⊗ b)wn)k2 2 ≤ Xg∈Γ Xg′∈gF ≤ kak2 2kbk2 2kbk2 ≤ kak2 vn,gwn,g−1vn,g′wn,g′−1τ (σg(a ⊗ b)σg′(a∗ ⊗ b∗)) + ε. 2h(vn)FXg∈Γ 2h(vn)F + ε. vn,gwn,g−1 + ε Hence, lim supn kEA ⊗ A(vn(a ⊗ b)wn)k2 2 ≤ ε. Since ε was arbitrary, we get the claim. Now take ε′ < kpk2/4. By assumption, D ⊀M LΓ1 ⊗ M2 and D ⊀M M1 ⊗ LΓ2, so there exists a ∈ U(D) such that kELΓ1 ⊗ M2(a)k2 < ε′ and kEM1 ⊗ LΓ2(a)k2 < ε′. By Lemma 3.2, the sequence (vnav∗ subset F ⊂ Γ such that kPF (vnav∗ ELΓ1 ⊗ M2(a) − EM1 ⊗ LΓ2(a − ELΓ1 ⊗ M2(a)), we get n)n belongs to Aω ⋊ Γ, so that there exists a finite n)k2 ≥ kpk2 − ε′. Thus if we Define a0 = a − kpk2 − ε′ ≤ kPF (vnav∗ n)k2 ≤ kPF (vn(a − ELΓ1 ⊗ M2(a))v∗ ≤ kPF (vna0v∗ n)k2 + 3ε′. n)k2 + ε′ But a0 is orthogonal to LΓ1 ⊗ M2 and M1 ⊗ LΓ2, because the conditional expectations ELΓ1 ⊗ M2 and EM1 ⊗ LΓ2 commute. Therefore, when n goes to infinity, the claim implies that kPF (vna0v∗ n)k2 → 0 which leads to the absurd statement that kpk2 ≤ 4ε′ < kpk2. We end this paragraph by a lemma that localizes the Fourier coefficients of elements vnav∗ n inside A, for a particular (fixed) a ∈ D. In fact, this lemma will be the starting point of our reasoning by contradiction in Step 2 below, being the initialization of an induction process. Lemma 3.4. For a well chosen a ∈ U(D), there exists a δ0 > 0, a finite dimensional subspace K ⊂ (A1 ⊖ C1) ⊗ (A2 ⊖ C1), and a sequence (gn, hn) ∈ Γ such that: (cid:3) • gn, hn → ∞, as n → ∞; • lim inf kQσ(gn ,hn)(K)(vnav∗ n)k2 > δ0. Proof. Put δ1 = lim inf h(vn) > 0 and consider a sequence (gn, hn) ∈ Γ such that vn,(gn,hn) = h(vn) for all n. Now 3.1 implies that the sequences (gn) and (hn) go to infinity with n. Moreover, we have lim sup n kvn − vn,(gn,hn)u(gn,hn)k2 = qkpk2 2 − δ2 1. 2 − δ2 Take ε > 0 such that pkpk2 1 + 4ε < kpk2. By assumption (ii), there exists a ∈ U(D) such that kELΓ1 ⊗ M2(a)k2 < ε and kEM1 ⊗ LΓ2(a)k2 < ε. Thus the element a1 = a − ELΓ1 ⊗ M2(a) − EM1 ⊗ LΓ2(a − ELΓ1 ⊗ M2(a)) satisfies ka − a1k2 < 3ε, and its Fourier coefficients are in (A1⊖ C1)⊗ (A2⊖ C1). We conclude that there exists a finite dimensional K ⊂ (A1 ⊖ C1) ⊗ (A2 ⊖ C1) such that, ka − QK(a)k2 < 4ε. 14 R ´EMI BOUTONNET Finally, we get that kvnav∗ n − vn,(gn,hn)u(gn,hn)QK(a)v∗ nk2 < qkpk2 2 − δ2 1 + 4ε. n belongs to the image of the projection Qσ(gn ,hn)(K), we (cid:3) Since vn,(gn,hn)u(gn,hn)QK(a)v∗ get the result with δ0 > 0 defined by kpk2 Step 2: We show that C ≺M A. Notation. For a finite subset G ⊂ Γ, finite dimensional subspaces K1, K2 ⊂ A and λ > 0, define 0 = (pkpk2 1 + 4ε)2. 2 − δ2 2 − δ2 [K1 × σG(K2)]λ = conv{λaσg(b) a ∈ K1, b ∈ K2, g ∈ G,kak2 ≤ 1,kbk2 ≤ 1}. We have that [K1 × σG(K2)]λ is a closed convex subset C of A (being the convex hull of a compact subset in a finite dimensional vector space). Then the set C = {Xg∈Γ ξg ⊗ δg ∈ L2(M) ∀g ∈ Γ, ξg ∈ C} is a closed convex subset of L2(M). Hence one can define the "orthogonal projection onto this set" QC : L2(M) → L2(M) as follows. For x ∈ L2(M), QC(x) is the unique point of C such that kx − QC(x)k = inf y∈ C kx − yk. Remark 3.5. This notation is consistent with the previous notation QK: If K ⊂ A is a finite dimensional subspace, then QK(a) = QC(a), where C = [C1 × σ{e}(K)]λ as soon as λ ≥ kak2. Before getting into the heart of the proof, we check some easy properties of these convex sets. Lemma 3.6. Fix λ > 0 and finite dimensional subspaces K1, K2 ⊂ A. Then there exists a constant κ > 0 such that for all finite G ⊂ Γ, and all x ∈ [K1 × σG(K2)]λ, kxk∞ ≤ κ. Proof. Since K1 and K2 are finite dimensional, there exists a constant c > 0 such that kak∞ ≤ ckak2 for all a ∈ K1 or a ∈ K2. One sees that κ = λc2 satisfies the conclusion of the lemma. Lemma 3.7. For finite subsets F, G ⊂ Γ, and finite dimensional subspaces K1, K2, K ′ A and λ, λ′ > 0, we have (cid:3) 1, K ′ 2 ⊂ [K1 × σF (K2)]λ + [K ′ Proof. This is straightforward. 1 × σG(K ′ 2)]λ′ ⊂ [(K1 + K ′ 1) × σG∪F (K2 + K ′ 2)]λ+λ′ . (cid:3) From now on, we assume by contradiction that C ⊀M A. The contradiction we are looking for is then a direct consequence of the following implication. Indeed, using Lemma 3.4, and iterating the implication enough times, we get the absurd statement that there exist unitaries an = vnav∗ n and elements bn of the form QCn(an) such that lim inf n kan − bnk2 2 is negative. W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 15 Implication. Fix a ∈ U(D) and put an = vnav∗ n for all n. Assume that there exists a sequence of finite subsets Fn×Gn ⊂ Γ = Γ1×Γ2, finite dimensional subspaces K1 ⊂ A, K2 ⊂ (A1 ⊖ C1) ⊗ (A2 ⊖ C1), λ > 0 and δ > 0 such that: • supn FnGn < ∞; • Fn → ∞, Gn → ∞; • lim supn kan − QCn(an)k2 2 < kpk2 2 − δ2, where Cn = [K1 × σFn×Gn(K2)]λ. Then there exists a sequence of finite subsets F ′ K ′ 1 ⊂ A, K ′ 2 ⊂ (A1 ⊖ C1) ⊗ (A2 ⊖ C1), and λ′ > 0 such that: n×G′ n ⊂ Γ, finite dimensional subspaces nG′ n → ∞, G′ • supn F ′ n < ∞; • F ′ n → ∞; • lim supn kan − QC ′ n(an)k2 2 < kpk2 2 − 3δ2/2, where C′ n = [K ′ 1 × σF ′ n(K ′ 2)]λ′. n×G′ The multiple mixing property will be used in the proof of this implication through the following lemma. Lemma 3.8. Let x, y, z ∈ A1 ⊗ A2. For any sequences gn = (g1 (h1 n) ∈ Γ and hn = n, g2 n, g2 n, h1 n, h2 n, h2 n) ∈ Γ such that g1 n → ∞, we have τ (xσgn(y)σhn(z)) − τ (x)τ (σgn(y)σhn(z)) → 0. Proof. Without loss of generality, one can assume that x = x1 ⊗ x2, y = y1 ⊗ y2, z = z1 ⊗ z2. We have • τ (xσgn(y)σhn(z)) = τ (x1σg1 n(y1)σh1 • τ (x)τ (σgn(y)σhn(z)) = τ (x1)τ (σg1 n(z1))τ (x2σg2 n(y1)σh1 n(y2)σh2 n(z2)); n(z1))τ (x2)τ (σg2 n(y2)σh2 n(z2)). So the result follows directly from the multiple mixing property of the Gaussian actions Γi y Ai, i = 1, 2. (cid:3) Proof of the implication. Let a, Fn, Gn, K1, K2, λ, δ and Cn be as in the implication. Fix ε > 0, with ε ≪ δ. By Lemma 3.2 one can find S ⊂ Γ finite such that kan − PS(an)k2 ≤ ε, for all n. Hence we get that lim supn kan−PS◦QCn(an)k2 < pkpk2 2 − δ2+ ε. Now following Ioana's idea, we will consider an element d ∈ U(C) with sufficiently spread out Fourier coefficients so that for n large enough, dPS ◦ QCn(an)d∗ is almost orthogonal to PS◦QCn(an), while it is still close to an. Then the sum dPS◦QCn(an)d∗ + PS ◦ QCn(an) should be even closer to an. Let α > 0 be a (finite) constant such that kxk∞ ≤ αkxk2, for all x ∈ K1. Since K2 ⊂ (A1 ⊖ C1) ⊗ (A2 ⊖ C1) is finite dimensional, the set L = {g ∈ Γ , ∃a, b ∈ K2,kak2 ≤ 1,kbk2 ≤ 1 : hσg(a), bi ≥ ε/S2λ2α2} is finite. Hence for all n, Ln = ∪g,h∈Fn×GngLh−1 is finite, with cardinal smaller or equal to Fn2Gn2L, which is itself majorized by some N, not depending on n. Since C ⊀ A, Ioana's intertwining criterion (Lemma 2.2) implies that there exists d ∈ U(C) such that kPF (d)k2 ≤ ε/κS, whenever F ≤ N, where κ is given by Lemma 3.6 applied to K1, K2 and λ. By Kaplansky's density theorem, one can find d0, d1 ∈ M, and T = T1 × T2 ⊂ Γ finite such that: 16 R ´EMI BOUTONNET • di = PT (di), i = 0, 1; • kd0 − dk2 ≤ min(ε, ε/κS), kd1 − d∗k2 ≤ ε; • kdik∞ ≤ 1, i = 0, 1. Since an ∈ D for all n and d ∈ C = D′ ∩ M, we have dand∗ = an. Thus for all n, kan − d0and1k2 ≤ 2ε, and so lim sup n kan − d0PS ◦ QCn(an)d1k2 ≤ qkpk2 2 − δ2 + 3ε. Now, for all n, put Tn = T \ Ln. By definition of d, kPT (d)− PTn(d)k2 ≤ ε/κS, hence kd0 − PTn(d0)k2 ≤ 3ε/κS. Notice that kPS ◦ QCn(an)k∞ ≤ κS, which implies that n lim sup kan − PTn(d0)PS ◦ QCn(an)d1k2 ≤ qkpk2 Denote by xn = PS ◦ QCn(an) and yn = PTn(d0)PS ◦ QCn(an)d1. We want to show that lim supn hxn, yni is small. Write d0 = Pg∈T d0,gug, an = Ph an,huh, and d1 = Pk∈T d1,kuk. We get τ (d0,gσgh(d1,k)σg(QCn(an,h))QCn(an,ghk)∗) 2 − δ2 + 6ε. hyn, xni = Xg∈Tn,h∈S,k∈T = Xg∈T,h∈S,k∈T ghk∈S ghk∈S 1{g∈Tn}τ (d0,gσgh(d1,k)σg(QCn(an,h))QCn(an,ghk)∗). Claim. For all fixed x, y ∈ A, and g ∈ T , there exists n0 such that for all n ≥ n0, and all a, b ∈ Cn, 1{g∈Tn}τ (xyσg(a)b∗)i ≤ 2εkxk2kyk2/S2. To prove this claim, first recall that for all n, Cn = [K1 × σFn×Gn(K2)]λ. Denote by K1 = span{xyσg(a)b∗, a, b ∈ K1}. Since K1 and K2 have finite dimension and since Fn, Gn → ∞, Lemma 3.8 implies that there exists n0 such that for n ≥ n0, and for all s, t ∈ Fn × Gn one has (3.3) τ (aσgs(b)σt(c∗)) − τ (a)τ (σgs(b)σt(c∗)) ≤ εkxk2kyk2/S2λ2. sup a∈ K1,kak2≤1 b,c∈K2,kbk2≤1,kck2≤1 Thus take n ≥ n0. By definition of Cn, it is sufficient to prove that for all a, b ∈ K1, c, d ∈ K2, with kak2,kbk2,kck2,kdk2 ≤ 1, and all s, t ∈ Fn × Gn, 1{g∈Tn}τ (xyσg(λaσs(c))λb∗σt(d∗)) ≤ 2εkxk2kyk2/S2. We can assume that g ∈ Tn. An easy calculation gives τ (xyσg(λaσs(c))λb∗σt(d∗)) ≤ εkxk2kyk2/S2 + λ2τ (xyσg(a)b∗)τ (σgs(c)σt(d∗)) ≤ εkxk2kyk2/S2 + λ2kxk2kyk2kak∞kbk∞ε/S2λ2α2 ≤ 2εkxk2kyk2/S2, where the first inequality is deduced from 3.3, while the second is because g /∈ Ln. So the claim is proven. W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 17 Now we can estimate hxn, yni, for n large enough. hxn, yni ≤ Xg∈T,h∈S,k′∈S ≤ Xg∈T,h∈S,k′∈S ≤ 2εkd0k2kd1k2 ≤ 2ε. Therefore, we obtain: 1{g∈Tn}τ (d0,gσgh(d1,h−1g−1k′)σg(QCn(an,h))QCn(an,k′)∗) 2εkd0,gk2kd1,h−1g−1k′k2/S2 • lim supn kan − xnk2 < pkpk2 • lim supn kan − ynk2 < pkpk2 • lim supn hxn, yni ≤ 2ε. 2 − δ2 + ε; 2 − δ2 + 6ε; Thus using the formula kx − (y + z)k2 2 = kx − yk2 one checks that lim supn kan − (xn + yn)k2 Now observe that 2 + kx − zk2 2 ≤ kpk2 2 + 2ℜhy, zi, 2 − kxk2 2 − 3δ2/2, if ε is small enough. yn = Xg∈Tn,h∈S,k∈T d0,gσgh(d1,k)σg(QCn(an,h))ughk. So let us check that yn has its Fourier coefficients in [K0 × σ(T1Fn)×(T2Gn)(K2)]λST , where K0 = span{d0,gσgh(d1,k)σg(c), c ∈ K1, g, k ∈ T, h ∈ S}. Fix n ∈ N, and s ∈ Γ. Denote by yn,s = EA(ynu∗ s). We have yn,s = Xg∈Tn,h∈S,k∈T ghk=s d0,gσgh(d1,k)σg(QCn(an,h)). Thus it is a convex combination of terms of the form d0,gσgh(d1,k)σg(λahσth(bh)) T = Xg∈T,h∈S,k∈T ghk=s 1 = ST Xg∈T,h∈S,k∈T ghk=s STd0,gσgh(d1,k)σg(λahσth(bh)), for elements ah ∈ K1, bh ∈ K2, with kahk2,kbhk2 ≤ 1 and th ∈ Fn × Gn, for all h ∈ S. But such terms T are themselves convex combinations of elements of the form λSTxσgt(y), with x ∈ K0, y ∈ K2, kxk2,kyk2 ≤ 1 and gt ∈ T (Fn × Gn) = (T1Fn) × (T2Gn). Therefore, as pointed out in Lemma 3.7, xn + yn has Fourier coefficients in C′ σF ′ G′ n = Gn ∪ T2Gn. We conclude that: 2 = K2, λ′ = λ + λST, and F ′ 1 × n = Fn ∪ T1Fn, 1 = K1 + K0, K ′ 2)]λ′, with K ′ n = [K ′ n(K ′ n×G′ which proves the implication. kan − QC ′ n(an)k2 2 ≤ kpk2 2 − 3δ2/2, The proof of Theorem 3.1 is complete. (cid:3) (cid:3) 18 R ´EMI BOUTONNET Taking Γ2 = {e} and A2 = C we obtain a similar statement for a single mixing action Γ y A. Corollary 3.9. Assume that Γ y A is a mixing Gaussian action. Denote by M = A ⋊ Γ. Consider an abelian subalgebra D ⊂ pMp, p ∈ LΓ, which is normalized by a sequence of unitaries (vn) ∈ U(pLΓp) with vn → 0 weakly. Put C = D′ ∩ pMp. Then one of the following is true: • D ≺M LΓ • For all q ∈ Z(C), qC ≺M A. In fact, S. Vaes asked during his series of lectures at the IHP in Paris (spring 2011) whether such a corollary could hold for any mixing action. A. Ioana showed that this is true for Bernoulli shifts [Io11, Theorem 6.2], and as we just showed, the proof can be adapted to Gaussian actions. In our proof, we only used the following properties of Gaussian actions: • The 2-mixing property; • The malleability property. Moreover, the malleability of Gaussian actions is only used to prove Lemma 3.2 (i.e. to show that the sequences (vnav∗ n), a ∈ D lie in Aω ⋊ Γ). We suspect that this lemma might be shown only using multiple mixing properties, but we were not able to reach this conclusion. We end this section by mentioning a generalization of Theorems A and B that considers some amplifications. The proof is the same, and still works because Popa's orbit equivalence superrigidity theorems ([IPV11, Theorem 5.2 and Theorem 5.6] and [Po08, Theorem 1.3]) handle such amplifications. Theorem 3.10. Let Γ be an ICC countable discrete group, and π : Γ → O(HR) an orthogonal representation of Γ. Make one of the following two assumptions: • Γ is w-rigid and ICC, and π is mixing; • Γ is an ICC non-amenable product of two infinite groups and π is mixing and admits a tensor power which is weakly contained in the regular representation. Let Γ y A be the Gaussian action associated with π and put M = A ⋊ Γ. Let Λ y B be another free ergodic action on an abelian von Neumann algebra, and put N = B ⋊Λ. If for some t ≥ 1, M ≃ N t, then t = 1, Γ ≃ Λ and the actions Γ y A and λ y B are conjugate. 4. An application to group von Neumann algebras As another application of Theorem 3.1, we construct a large class of II1 factors which are not stably isomorphic to group von Neumann algebras. These factors are the crossed-product von Neumann algebras of Gaussian actions associated with represen- tations π as in Theorem A or Theorem B, with the extra-assumption that π is not weakly contained in the regular representation. In [Bo12, Proposition 2.8], such Gaussian actions were shown not to be conjugate to generalized Bernoulli shifts. Using Theorems A and B, we get that the associated factors are not isomorphic to crossed-product factors of Bernoulli actions, and in par- ticular, to von Neumann algebras of certain wreath-product groups. However, showing W∗-SUPERRIGIDITY OF GAUSSIAN ACTIONS 19 that such factors are not isomorphic to algebras LΛ, with no assumptions on the group Λ is much harder, and will require the work of Ioana, Popa and Vaes [IPV11]. Theorem 4.1. Let Γ be an ICC group and π : Γ → O(H) a mixing orthogonal representation of Γ such that one of the following two conditions holds. • Γ is w-rigid or • Γ is non-amenable and is isomorphic to the product of two infinite groups, and some tensor power of π is weakly contained in the regular representation of Γ. Assume moreover that π itself is not weakly contained in the regular representation. Let Γ yσ A be the Gaussian action associated with π and put M = A ⋊ Γ. Then M is not stably isomorphic to a group von Neumann algebra. Proof. Let π be an orthogonal representation as in the statement of the theorem. Assume by contradiction that there exists a countable group Λ such that M ≃ (LΛ)t for some t > 0. Then adapting the proof of [IPV11, Theorem 8.2], we get that t = 1, and Λ ≃ Σ ⋊ Γ, for some infinite abelian group Σ and some action Γ y Σ by automorphisms. Moreover, the initial Gaussian action σ is conjugate to the action of Γ on LΣ. Now, since σ is mixing, the action Γ y Σ \ {e} has finite stabilizers. But then the Koopmann representation Γ → U(ℓ2(Σ \ {e})) is weakly contained in the left regular representation. Thus, Proposition 1.7 in [PS10] implies that π is weakly contained in the regular representation, which contradicts our assumptions on π. (cid:3) By [Bo12, Proposition 2.9], we know that for each n ≥ 3, PSL(n, Z) admits a repre- sentation as in Theorem C. Thus we obtain the existence of a II1 factor Mn, which is not stably isomorphic to a group von Neumann algebra. But using Theorem 3.10, we get that the Mn's are pairwise non-stably isomorphic : Mn ≇ (Mm)t, ∀t > 0, ∀n 6= m. References [BHV08] B. Bekka, P. de la Harpe & A. Valette, Kazhdan's property (T). New Mathematical Monographs, 11. Cambridge University Press, Cambridge, 2008. [Bo12] R. Boutonnet, On solid ergodicity for Gaussian actions. J. Funct. Anal., 263 (2012) 1040- 1063. [FM77] J. Feldmann & C.C. Moore, Ergodic equivalence relations, cohomology, and von Neumann algebras, II. Trans. Amer. Math. Soc. 234 (1977), 325-359. [Fu07] A. Furman, On Popa's Cocycle Superrigidity Theorem. Inter. Math. Res. Notices IMRN, 2007 (2007), 1 -- 46, Art. ID rnm073. [Ga10] D. Gaboriau, Orbit Equivalence and Measured Group Theory. Proceedings of the ICM (Hy- derabad, India, 2010), Vol. III, Hindustan Book Agency (2010), 1501 -- 1527. [HPV10] C. Houdayer, S. Popa & S. Vaes, A class of groups for which every action is W∗- superrigid, Groups Geom. Dyn., to appear. arXiv:1010.5077. [Io11] A. Ioana, W∗-superrigidity for Bernoulli actions of property (T) groups. J. Amer. Math. Soc., 24 (2011), 1175 -- 1226. [IPV11] A. Ioana, S. Popa & S. Vaes, A class of superrigid group von Neumann algebras, Ann. of Math., to appear. [OP07] N. Ozawa & S. Popa, On a class of II1 factors with at most one Cartan subalgebra. Ann. of Math. (2), 172 (2010), 713 -- 749. [Pe09] J. Peterson, Examples of groups which are virtually W∗-superrigid. preprint (2009). arXiv:1002.1745 [PS10] J. Peterson & T. Sinclair, On cocycle superrigidity for Gaussian actions. Erg. Th. & Dyn. Sys. 32 (2012), 249 -- 272. 20 R ´EMI BOUTONNET [Po06a] S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups I. Invent. Math. 165 (2006), 369 -- 408. [Po06b] S. Popa, Strong rigidity of II1 factors arising from malleable actions of w-rigid groups II. Invent. Math. 165 (2006), 409 -- 453. [Po06c] S. Popa, On a class of type II1 factors with Betti numbers invariants. Ann. of Math. 163 (2006), 809 -- 899. [Po06d] S. Popa, Some rigidity results for non-commutative Bernoulli Shifts. J. Funct. Anal. 230 (2006), 273 -- 328. [Po07a] S. Popa, Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups. Invent. Math. 170 (2007), 243 -- 295. [Po07b] S. Popa, Deformation and rigidity for group actions and von Neumann algebras. Pro- ceed- ings of the ICM (Madrid, 2006), Vol. I, European Mathematical Society Publishing House (2007), 445 -- 477. [Po08] S. Popa, On the superrigidity of malleable actions with spectral gap. J. Amer. Math. Soc. 21 (2008), 981 -- 1000. [PV09] S. Popa & S. Vaes, Group measure space decomposition of II1 factors and W∗-superrigidity Invent. Math. 182 (2010), 371 -- 417. [Va08] S. Vaes, Explicit computations of all finite index bimodules for a family of II1 factors. Ann. Sci. ´Ecole Norm. Sup. 41 (2008), 743 -- 788. [Va10a] S. Vaes, Rigidity for von Neumann algebras and their invariants. Proceedings of the ICM (Hyderabad, 2010), Vol. III, Hindustan Book Agency (2010), 1624 -- 1650. [Va10b] S. Vaes, One-cohomology and the uniqueness of the group measure space of a II1 factor. Math. Ann., to appear. arXiv:1012.5377. ENS Lyon, UMPA UMR 5669, 69364 Lyon cedex 7, France E-mail address: [email protected]
1712.03234	2	1712	2018-09-04T21:54:12	Primitive ideal space of Higher-rank graph $C^*$-algebras and decomposability	[ "math.OA", "math.FA" ]	In this paper, we describe primitive ideal space of the $C^$-algebra $C^(\Lambda)$ associated to any locally convex row-finite $k$-graph $\Lambda$. To do this, we will apply the Farthing's desourcifying method on a recent result of Carlsen, Kang, Shotwell, and Sims. We also characterize certain maximal ideals of $C^(\Lambda)$. Furthermore, we study the decomposability of $C^(\Lambda)$. We apply the description of primitive ideals to show that if $I$ is a direct summand of $C^(\Lambda)$, then it is gauge-invariant and isomorphic to a certain $k$-graph $C^$-algebra. So, we may characterize decomposable higher-rank $C^$-algebras by giving necessary and sufficient conditions for the underlying $k$-graphs. Moreover, we determine all such $C^$-algebras which can be decomposed into a direct sum of finitely many indecomposable $C^*$-algebras.	math.OA	math	PRIMITIVE IDEAL SPACE OF HIGHER-RANK GRAPH C ∗-ALGEBRAS AND DECOMPOSABILITY HOSSEIN LARKI Abstract. In this paper, we describe primitive ideal space of the C ∗- algebra C ∗(Λ) associated to any locally convex row-finite k-graph Λ. To do this, we will apply the Farthing's desourcifying method on a recent result of Carlsen, Kang, Shotwell, and Sims. We also characterize certain maximal ideals of C ∗(Λ). Furthermore, we study the decomposability of C ∗(Λ). We apply the description of primitive ideals to show that if I is a direct summand of C ∗(Λ), then it is gauge-invariant and isomorphic to a certain k-graph C ∗-algebra. So, we may characterize decomposable higher-rank C ∗- algebras by giving necessary and sufficient conditions for the underly- ing k-graphs. Moreover, we determine all such C ∗-algebras which can be decomposed into a direct sum of finitely many indecomposable C ∗- algebras. 1. Introductions Motivated from [15], the C ∗-algebras of higher-rank graphs (or k-graphs) were introduced by Kumjian and Pask in [8] as higher-rank analogous of the graph C ∗-algebras. They were first considered in [8] only for row-finite k-graphs with no sources, and then generalized for locally convex row-finite and finitely aligned setting [12, 13]. Since then, they have received a great deal of attention and provided a very interesting source of examples rather than ordinary graph C ∗-algebras (see [10, 11] among others). For any countable directed graph E, Hong and Szyma´nski described in [5] primitive ideal space of the C ∗-algebra C ∗(E) and its hull-kernel topology. After that, there have been many efforts to characterize primitive ideals of higher-rank graph C ∗-algebras (see [16, 6] for example). The substantial work of Carlsen, Kang, Shotwell and A. Sims in [2] is to catalogue all prim- itive (two-sided and closed) ideals of the C ∗-algebra C ∗(Λ) of a row-finite higher-rank graph Λ with no sources. Despite some similarities, the struc- ture of primitive ideals in higher-rank graph C ∗-algebras are much more complicated compared with that of ordinary graphs. Although the main Date: September 6, 2018. 2010 Mathematics Subject Classification. 46L05. Key words and phrases. k-graph, higher-rank graph C ∗-algebra, primitive ideal, decomposability. 1 2 HOSSEIN LARKI result of [2] is a generalization of the Hong-Szyma´nski's description, but its methods and computations are quite different from [5]. In this paper, we let Λ be a locally convex row-finite k-graph with possible sources. Our primary aim is to characterize all primitive ideals of C ∗(Λ). To this end, we apply the Farthing's desourcification [4] on the results of [2]. Recall that Farthing in [4] constructed a specific k-graph Λ without sources which contains Λ as a subgraph. She showed in [4, Theorem 3.30] that C ∗(Λ) is a full corner in C ∗(Λ), and therefore they are Morita-equivalent (see Section 2.4 below for details). So, we may characterize the structure of primitive ideal space of C ∗(Λ) using that of C ∗(Λ). Note that the Farthing's desourcification was modified by Webster in [18, Section 4]. However, [18, Proposition 4.12] shows that the desourcifications constructed in [4, Section 3] and [18, Section 4] are isomorphic when Λ is row-finite. Furthermore, there is a mistake in the proof of [4, Theorem 3.30] (the proof works only for locally convex row-finite k-graphs), and Webster resolved it in [18, Theorem 6.3] (see [18, Remark 6.2]). The rest of article will be devoted to some applications of the primitive ideal structure of C ∗(Λ). First, as any ideal is an intersection of primitive ideals, certain maximal ideals of C ∗(Λ) will be determined. Then, we will describe ideals which are direct summands of C ∗(Λ), and study the decom- posability of C ∗(Λ). This paper is organized as follows. In Section 2, we recall some elementary definitions and basic facts about k-graphs and their C ∗-algebras from [8, 12]. We also review the Farthing's desourcification of a higher-rank graph Λ. In Section 3, we define an equivalent relation on a row-finite k-graph Λ, which will be used in Section 4 to describe generators of primitive ideals. We then discuss on relationships between equivalent paths in Λ and in its desourcification Λ. In Section 4, for any locally convex row-finite k-graph Λ, we characterize primitive (two-sided and closed) ideals of C ∗(Λ) and define specific irreducible representations whose kernels generate such ideals. Then, in Section 5, some certain maximal primitive ideals of C ∗(Λ) are determined. As a consequence, we see that when Λ is a cofinal k-graph, then primitive ideals of C ∗(Λ) are all maximal. In Section 6, by applying the description of primitive ideals, we give some graph theoretic conditions for the decomposability of C ∗(Λ). In particular, we show that if C ∗(Λ) decomposes as A ⊕ B, then A and B are gauge- invariant ideals of C ∗(Λ), which are themselves isomorphic to certain k- graph C ∗-algebras. Finally, in Section 7, we consider the question "when is C ∗(Λ) a direct sum of finitely many indecomposable C ∗-algebras?". We describe all such higher-rank graph C ∗-algebras by giving necessary and sufficient conditions on the underlying k-graphs. PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 3 2. Higher-rank graphs and their C ∗-algebras In this section, we review some basic facts about higher-rank graphs and their C ∗-algebras, which will be needed in the paper. We refer the reader to [8] and [12] for more details. 2.1. Higher-rank graphs. Fix a positive integer k > 0. We regard Nk := {n = (n1, . . . , nk) : ni ≥ 0} as an additive semigroup with identity 0. We denote by e1, . . . , ek the standard generators of Nk. The relation m ≤ n ⇐⇒ mi ≤ ni for 1 ≤ i ≤ k, puts a partial order on Nk. We write m ∨ n and m ∧ n for the coordinatewise maximum and minimum of m, n ∈ Nk, respectively. Definition 2.1 ([8]). A k-graph (or graph of rank k) is a countable small category Λ = (Λ0, Λ1, r, s) equipped with a degree functor d : Λ → Nk for each λ ∈ Λ and m, n ∈ Nk with satisfying the factorization property: d(λ) = m + n, there exist unique µ, ν ∈ Λ such that d(µ) = m, d(ν) = n, and λ = µν. Note that for µ, ν ∈ Λ, the composition µν makes sense only if s(µ) = r(ν). Note that every directed graph may be considered as a 1-graph (and vise versa), in the usual manner. With this example in mind, we make some notations. For each n ∈ Nk, we think of Λn := d−1(n) as the paths of degree n; in particular, Λ0 = d−1(0) is the vertices in Λ. For v ∈ Λ0, H ⊆ Λ0, and E, F ⊆ Λ, we write vE := {µ ∈ E : r(µ) = v}, HE := {µ ∈ E, r(µ) ∈ H} and EF := {µν : µ ∈ E, ν ∈ F, and s(µ) = r(ν)}, and we define Ev and EH analogously. A vertex v ∈ Λ0 is called source if vΛei = ∅ for some 1 ≤ i ≤ k. Definition 2.2. Let Λ be a k-graph. We say that Λ is row-finite if the sets vΛn are all finite for v ∈ Λ0, n ∈ Nk. Also, Λ is called locally convex, if for every v ∈ Λ0, i 6= j ∈ {1, 2, . . . , k}, λ ∈ vΛei and µ ∈ vΛej , we have s(λ)Λej 6= ∅ and s(µ)Λei 6= ∅. Observe that if Λ has no sources, then Λ is locally convex. Throughout the paper, we work only with locally convex, row-finite k-graphs. 2.2. Boundary Paths. Let Λ be a locally convex row-finite k-graph. For n ∈ Nk, we write Λ≤n := {λ ∈ Λ : d(λ) ≤ n, and d(λ)i < ni =⇒ s(λ)Λei = ∅} . Note that if Λ has no sources, then Λ≤n = Λn. Example 2.3. For any m ∈ (N ∪ {∞})k, let Ωk,m := {(p, q) ∈ N × N : p ≤ q ≤ m}. If we define r(p, q) := (p, p), s(p, q) := (q, q), and d(p, q) := q − p, then Ωk,m is a k-graph. For simplicity, each (p, p) is denoted by p. Thus, we regard Ω0 k,m = {p : p ≤ m} as the object set of Ωk,m. For m = (∞, . . . , ∞), the k-graph Ωk,m is denoted by Ωk in [8]. 4 HOSSEIN LARKI A boundary path in Λ is a degree preserving functor x : Ωk,m → Λ such that p ≤ m and pi = mi imply x(p)Λei = ∅. Then, m is called the degree of x and we write d(x) = m. When m = (∞, . . . , ∞), x is an infinite path in the sense of [8]. We denote by Λ≤∞ the set of boundary paths in Λ. Note that if Λ has no sources, then every boundary path is an infinite path, so Λ≤∞ = Λ∞. Given any x ∈ Λ≤∞ and n ≤ d(x), we may define the n-shifted boundary path σn(x) : Ωk,d(x)−n → Λ by σn(x)(p, q) := x(p + n, q + n) for p ≤ q ≤ d(x)−n. Moreover, if λ ∈ Λx(0), there is a unique boundary path λx ∈ Λ≤∞ such that λx(0, d(λ)) = λ and (λx)(d(λ), d(λ) + p) = x(0, p) for all p ≤ d(x). So, we have σd(λ)(λx) = x. 2.3. The C ∗-algebra of a higher-rank graph. Let Λ be a locally convex row-finite k-graph. A Cuntz-Krieger Λ-family is a set of partial isometries {Sλ : λ ∈ Λ} satisfying the following relations: (1) SvSw = δv,wSv for all v, w ∈ Λ0; (2) SλSλ′ = Sλλ′ for all λ, λ′ ∈ Λ with s(λ) = r(λ′); (3) S∗ λSλ = Ss(λ) for all λ ∈ Λ; λ for all v ∈ Λ0 and n ∈ Nk. (4) Sv =Pλ∈vΛ≤n SλS∗ The associated C ∗-algebra C ∗(Λ) is the universal C ∗-algebra generated by a Cuntz-Krieger Λ-family {sλ : λ ∈ Λ}. The universality implies that there is a gauge action γ : Tk → AutC ∗(Λ) such that γt(sλ) = td(λ)sλ for t ∈ Tk, where td(λ) := td(λ)1 . . . td(λ)k . By ideal we mean a closed and two-sided one. An ideal I of C ∗(Λ) is called gauge-invariant if γz(I) ⊆ I for every z ∈ Tk. It is well-know by [12, Theorem 5.2] that gauge-invariant ideals of C ∗(Λ) are associated to hereditary and saturated subsets of Λ0. 1 k Definition 2.4. A subset H of Λ0 is called hereditary if r(λ) ∈ H implies s(λ) ∈ H for every λ ∈ Λ. Also, we say that H is saturated if v ∈ Λ0 and s(vΛ≤n) ⊆ H for some n ∈ Nk, then v ∈ H. If H is a subset of Λ0, the saturated closure of H is the smallest saturated subset Σ(H) of Λ0 containing H. Recall from [12, Lemma 5.1] that if H ⊆ Λ0 is hereditary, then so is Σ(H). We denote by HΛ the collection of saturated hereditary subsets of Λ0. Note that HΛ has a lattice structure by the operations H1 ∧ H2 := H1 ∩ H2 H1 ∨ H2 := Σ(cid:0)H1 ∪ H2(cid:1). For every H ∈ HΛ, we denote by IH the gauge-invariant ideal of C ∗(Λ) generated by {sv : v ∈ H}, which is IH = span{sλs∗ λ′ : λ, λ′ ∈ Λ, s(λ) = s(λ′) ∈ H}. Also, [12, Theorem 5.2(b)] shows that Λ \ ΛH is a locally convex k-subgraph of Λ and we have C ∗(Λ \ ΛH) ∼= C ∗(Λ)/IH . According to [12, Theorem 5.2(a)] the map H 7→ IH is a lattice isomorphism from HΛ onto the set of PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 5 gauge-invariant ideals of C ∗(Λ). Moreover, for each ideal I of C ∗(Λ), the set HI := {v ∈ Λ0 : sv ∈ I} is a saturated hereditary subset of Λ0. 2.4. Removing sources from a higher-rank graph. Here, we briefly review the desourcification constructions of Farthing [4] and Webster [18] with some minor modifications. We refer the reader to [4] and [18] for details and proofs. Note that, in case Λ is a row-finite k-graph, [18, Proposition 4.12] shows that the constructions of [18, Section 4] and [4, Section 3] produce isomorphic desourcifications. Let Λ be a locally convex row-finite k-graph. Let PΛ be the set PΛ := {(x; (m, n)) : x ∈ Λ≤∞ and m ≤ n ∈ Nk}. We define (x; (m, n)) ≈ (y; (p, q)) if and only if P1) x(m ∧ d(x), n ∧ d(x)) = y(p ∧ d(y), q ∧ d(y)), P2) m − m ∧ d(x) = p − p ∧ d(y), and P3) n − m = q − p. Then, ≈ is an equivalence relation on PΛ and we denote the equivalence class of (x; (m, n)) by [x; (m, n)]. According to [4, Theorem 3.24], Λ := PΛ/ ≈ equipped with r([x; (m, n)]) := [x; (m, m)], s([x; (m, n)]) := [x; (n, n)], and d([x; (m, n)]) := n − m. is a k-graph containing no sources. Also, the composition of paths in Λ is of the form [x; (m, n)] ◦ [y; (p, q)] = [x(0, n ∧ d(x))σp∧d(y); (m, n + q − p)], which makes sense only if [x; (n, n)] = [y; (p, p)]. For simplicity, we usually 0 = {[x; m] : x ∈ Λ≤∞, m ∈ Nk} is denote each (x; (m, m)) by (x; m), so Λ the vertex set of Λ. Note that, for (x; m) ≈ (y; n) it suffices to check only (P1) and (P2) because (P3) is trivial. The correspondence λ 7→ [λx; (0, d(λ))], where x ∈ s(λ)Λ≤∞, is an in- jective k-graph morphism from Λ into Λ [18, Poposition 4.13]. Hence, we may regard Λ as a subgraph of Λ. Also, the map π : Λ → Λ defined by π([x; (m, n)]) = [x; (m ∧ d(x), n ∧ d(x))], for x ∈ Λ≤∞ and m ≤ n ∈ Nk, is a well-defined surjective k-graph morphism such that π ◦ π = π. Recall from [4, Theorem 3.26] that if Λ is row-finite, then so is Λ. The k-graph Λ is called the desourcification of Λ. For every x ∈ Λ≤∞ and p ∈ Nk, we may define an infinite path [x; (p, ∞)] : Ωk → Λ by [x; (p, ∞)](m, n) := [x; (p + m, p + n)] for all m ≤ n ∈ Nk. The following Lemma says that every boundary path in Λ (which is an infinite path) is of the form [x; (p, ∞)]. 6 HOSSEIN LARKI Lemma 2.5 (See [14, Proposition 2.3]). Let Λ be a locally convex row-finite k-graph and let Λ be the desourcification of Λ. For each z ∈ Λ , there exist π(z) ∈ Λ≤∞ and pz ∈ Nk such that pz ∧ d(π(z)) = 0 and z = [π(z); (pz , ∞)]. Moreover, if z(0) ∈ Λ0, then pz = 0. ∞ Proof. The first statement is [14, Proposition 2.3]. For the second, if [π(z); pz] = z(0) ∈ Λ0, then π([π(z); pz ]) = [π(z); pz]. So, we have [π(z); pz] = π(cid:0)[π(z); pz ](cid:1) = [π(z); pz ∧ d(π(z))] = [π(z); 0]. Now condition (P2) for (π(z); pz) ≈ (π(z); 0) implies that pz = pz − pz ∧ d(π(z)) = 0. (cid:3) In other words, we may define the map π : Λ ∞ → Λ≤∞ by π([x; (p, ∞)])(m, n) := x(p ∧ d(x) + m, p ∧ d(x) + n) for m ≤ n ≤ d(x) − p ∧ d(x). Thus, we have π([x; (p, ∞)]) = σp∧d(x)(x). If {sλ : λ ∈ Λ} is a generating Cuntz-Krieger Λ-family for C ∗(Λ), then {sλ : λ ∈ Λ} is a universal Cuntz-Krieger Λ-family [4, Theorem 3.28]. Hence, C ∗({sλ : λ ∈ Λ}) = C ∗(Λ). Moreover, projection in the multiplier algebra M(C ∗(Λ)), then [4, Theorem 3.30] and [18, Theorem 6.3] say that P C ∗(Λ)P is a full corner in C ∗(Λ) such that P C ∗(Λ)P = C ∗(Λ). In particular, the C ∗-algebras C ∗(Λ) and C ∗(Λ) are Morita-equivalent. Therefore, the map J 7→ P JP is an isomorphism from the lattice of ideals in C ∗(Λ) onto that of C ∗(Λ). if we set P := Pv∈Λ0 sv as a Remark 2.6. There is an error in the proof [4, Theorem 3.30]; however, the proof may be considered for locally convex row-finite k-graphs. Webster resolved this result in [18, Theorem 6.3] (see [18, Remark 6.2]). 3. equivalent paths and periodicity For any 1-graph E, Hong and Szyma´nski in [5] described primitive ideals of C ∗(E) by specific collections of vertices T in E, called maximal tails, and periodicity in the quotient graphs ET . Recall that periodicity in 1-graphs can be determined by cycles with no entrances [6, Lemma 2.9]. Although we know that if a k-graph Λ is periodic (see Definition 3.2 below), then it contains a generalized cycle with no entrances [9, Lemma 4.4], but structure of periodic k-graphs is more complicated compared with 1-graphs and the arguments of [5] could not be generalized for k-graphs in general. To deal with the periodicity, Carlsen et al. in [2] used the following equivalent relation on Λ which is inspired from [3]. Definition 3.1 ([2]). Let Λ be a row-finite k-graph. We set an equivalent relation on Λ by µ ∼Λ ν ⇐⇒ s(µ) = s(ν) and µx = νx for all x ∈ s(µ)Λ≤∞. (The subscript in ∼Λ indicates the underlying k-graph.) We then define Per(Λ) := {d(µ) − d(ν) : µ, ν ∈ Λ and µ ∼Λ ν}, PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 7 which is a subset of Zk. There are several conditions for aperiodicity in the literature, which are [17, Proposition 2.11]). We equivalent for finitely aligned k-graphs (cf. consider the following from [12] which was called Condition (B) there. Definition 3.2. Let Λ be a locally convex row-finite k-graph. We say that Λ is aperiodic if for every v ∈ Λ0, there exists x ∈ vΛ≤∞ such that µ 6= ν ∈ Λv implies µx 6= νx. If Λ is not aperiodic, it is called periodic. Proposition 3.3. Let Λ be a locally convex row-finite k-graph. Then the following are equivalent. (1) Λ is aperiodic. (2) Per(Λ) = {0}. (3) There are no distinct µ, ν ∈ Λ such that µ ∼Λ ν. Proof. Implications (1) ⇔ (3) and (3) ⇒ (2) are immediate. To see (2) ⇒ (3), let µ ∼Λ ν. Then d(µ) − d(ν) ∈ Per(Λ) = {0}, and so d(µ) = d(ν). Fix x ∈ s(µ)Λ≤∞. Then for m := d(µ) = d(ν) we have µ = µx(0, m) = νx(0, m) = ν, giving (3). (cid:3) In the next proposition, we will relate equivalent paths in a k-graph Λ with those in its desourcification Λ. Before that, we state a simple lemma. Lemma 3.4. Let Λ be a row-finite k-graph and Λ the desourcification of Λ. Suppose µ, ν ∈ Λ such that µ ∼Λ ν. If µ = [x; (m, m + d(µ))] for some x ∈ Λ≤∞ and m ∈ Nk, then ν = [x; (m, m + d(ν))]. Proof. By µ ∼Λ ν, we can write [x; (m, ∞)] = [x; (m, m + d(µ))] ◦ [x; (m + d(µ), ∞)] = µ ◦ [x; (m + d(µ), ∞)] = ν ◦ [x; (m + d(µ), ∞)]. So, we must have ν = [x; (m, m + d(ν))] by the factorization property. (cid:3) Proposition 3.5. Let Λ be a locally convex, row-finite k-graph and let Λ be the desourcification of Λ. For any µ, ν ∈ Λ with r(µ) = r(ν), the following are equivalent: (1) µ ∼Λ ν; (2) π(µ) ∼Λ π(ν) and d(µ) − d(π(µ)) = d(ν) − d(π(ν)); (3) π(µ) ∼Λ π(ν) and d(µ) − d(π(µ)) = d(ν) − d(π(ν)). Proof. Write µ = [x; (m, m + d(µ))] and ν = [y; (n, n + d(ν))]. For q := m ∧ d(x), we have [x; (m, m + d(µ))] = [σq(x); (m − q, m − q + d(µ))] such that (m−q)∧d(σq(x)) =(cid:0)m−m∧d(x)(cid:1)∧(cid:0)d(x)−m∧d(x)(cid:1) = m∧d(x)−m∧d(x) = 0. 8 HOSSEIN LARKI So, without loss of generality, we may suppose µ = [x; (m, m + d(µ))] and ν = [y; (n, n + d(ν))] such that m ∧ d(x) = n ∧ d(y) = 0. Then π(µ) = [x; (m ∧ d(x), (m + d(µ)) ∧ d(x))] = [x; (0, d(µ) ∧ d(x))] and similarly π(ν) = [y; (0, d(ν) ∧ d(y))]. Moreover, we have [x; (0, m)] = [y; (0, m)]. Indeed, conditions (P1) and (P2) for [x; m] = r(µ) = r(ν) = [y; n] imply x(0) = y(0) and m = m − m ∧ d(x) = n − n ∧ d(y) = n, giving (x; (0, m)) ≈ (y; (0, m)). (1) =⇒ (2): We suppose µ ∼Λ ν. By Lemma 3.4, ν = [x; (m, m + d(ν))] and π(ν) = [x; (0, d(ν) ∧ d(x))]. So, (P2) for [x; m + d(µ)] = s(µ) = s(ν) = [x; m + d(ν)] yields that m + d(µ) −(cid:0)m + d(µ)(cid:1) ∧ d(x) = m + d(ν) −(cid:0)m + d(ν)(cid:1) ∧ d(x) d(µ) − d(π(µ)) = d(ν) − d(π(ν)). =⇒ Now we prove π(µ) ∼Λ π(ν). Using Lemma 2.5, every infinite path in Λ ∞ is of the form [z; (pz; ∞)] with pz ∧ d(z) = 0. Fix [z; (pz, ∞)] ∈ s(π(µ))Λ . In view of Lemma 2.5, we can assume pz = 0 because [z; pz] = s(π(µ)) ∈ Λ0. Set p := m + d(µ) − d(π(µ)) = m + d(ν) − d(π(ν)). Since d(σd(π(µ))(x)) ∧ p =(cid:0)d(x) − d(π(µ))(cid:1) ∧(cid:0)m + d(µ) − d(π(µ))(cid:1) =(cid:0)d(x) ∧ (m + d(µ))(cid:1) − d(π(µ)) = 0, [14, Lemma 2.2] implies [σd(π(µ))(x); (0, p)] = [z; (0, p)]. Thus, one may compute π(µ) ◦ [z; (0, ∞)] = [x; (0, d(π(µ)))] ◦ [z; (0, p)] ◦ [z; (p, ∞)] = [x; (0, d(π(µ)))] ◦ [σd(π(µ))(x); (0, p)] ◦ [z; (p, ∞)] = [x; (0, m + d(µ))] ◦ [z; (p, ∞)] = [x; (0, m)] ◦ [x; (m, m + d(µ))] ◦ [z; (p, ∞)] = [x; (0, m)] ◦ µ ◦ [z; (p, ∞)]. Analogously, we have π(ν) ◦ [z; (0, ∞)] = [x; (0, m)] ◦ ν ◦ [z; (p, ∞)], and hence π(µ) ◦ [z; (0, ∞)] = π(ν) ◦ [z; (0, ∞)] by applying µ ∼Λ ν. This establishes π(µ) ∼Λ π(ν). (2) =⇒ (1): Assume that π(µ) ∼Λ π(ν) and d(µ) − d(π(µ)) = d(ν) − d(π(ν)). We first claim [x; (d(π(µ)), m + d(π(µ)))] = [y; (d(π(ν)), m + d(π(ν)))]. To this end, we can check conditions (P1)-(P3) for(cid:0)x; (d(π(µ)), m+d(π(µ)))(cid:1) ≈ (cid:0)y; (d(π(ν)), m+d(π(ν)))(cid:1). Indeed, the fact [x; d(π(µ))] = s(π(µ)) = s(π(ν)) = PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 9 [y; d(π(ν))] gives (P1), whereas (P2) is trivial because d(π(µ)) ∧ d(x) = d(π(µ)) and d(π(ν)) ∧ d(y) = d(π(ν)). Also, the relation m + d(µ) − d(π(µ)) = m + d(ν) − d(π(ν)) yields (P3), therefore the claim holds. Furthermore, since [x; (d(π(µ)), ∞)] ∈ s(π(µ))Λ ∞ , we can write [x; (0, ∞)] = π(µ) ◦ [x; (d(π(µ)), ∞)] = π(ν) ◦ [x; (d(π(µ)), m + d(µ))] ◦ [x; (m + d(µ), ∞)] = [y; (0, d(π(ν)))] ◦ [y; (d(π(ν)), m + d(ν))] ◦ [x; (m + d(µ), ∞)] (by π(µ) ∼Λ π(ν)) (by the claim) = [y; (0, m + d(ν))] ◦ [x; (m + d(µ), ∞)]. So, we get [x; (0, m + d(ν))] = [y; (0, m + d(ν))] by the factorization property. In particular, ν = [y; (m, m + d(ν))] = [x; (m, m + d(ν))]. Now fix an arbitrary z ∈ s(µ)Λ Hence ∞ . Then [x; (d(π(µ)), d(µ))]◦z ∈ s(π(µ))Λ ∞ . [x; (0, m)] ◦ µ ◦ z = [x; (0, m)] ◦ [x; (m, m + d(µ))] ◦ z = [x; (0, d(π(µ)))] ◦ [x; (d(π(µ)), m + d(µ))] ◦ z = π(µ) ◦ [x; (d(π(µ)), m + d(µ))] ◦ z = π(ν) ◦ [x; (d(π(µ)), m + d(µ))] ◦ z (by π(µ) ∼Λ π(ν)) = [y; (0, d(π(ν)))] ◦ [y; (d(π(ν)), m + d(ν))] ◦ z = [y; (0, m)] ◦ [y; (m, m + d(µ))] ◦ z = [x; (0, m)] ◦ ν ◦ z. We therefore obtain µ ◦ z = ν ◦ z, following statement (1). (2) =⇒ (3): Fix y ∈ s(π(µ))Λ≤∞. Then [y; (0, ∞)] ∈ s(π(µ))Λ that π(cid:0)[y; (0, ∞)](cid:1) = y because for every p ≤ q ≤ d(y), π(cid:0)[y; (0, ∞)](cid:1)(p, q) = π(cid:0)[y; (p, q)](cid:1) = [y; (p, q)] = y(p, q). So, we have ∞ . Note π(µ) ◦ [y; (0, ∞)] = π(ν) ◦ [y; (0, ∞)] π(cid:0)π(µ) ◦ [y; (0, ∞)](cid:1) = π(cid:0)π(ν) ◦ [y; (0, ∞)](cid:1) π(µ) ◦ π(cid:0)[y; (0, ∞)](cid:1) = π(ν) ◦ π(cid:0)[y; (0, ∞)](cid:1) π(µ) ◦ y = π(ν) ◦ y. (by π(µ) ∼Λ π(ν)) (by π ◦ π = π) =⇒ =⇒ =⇒ This follows statement (3). (3) =⇒ (2): Since s(π(µ)) ∈ Λ0, by Lemma 2.5, every infinite path in is of the form [z; (0, ∞)] with z(0) = s(π(µ)). For every such s(π(µ))Λ ∞ 10 HOSSEIN LARKI [z; (0, ∞)] ∈ s(π(µ))Λ ∞ , we have π(µ) ◦ [z; (0, ∞)] = [π(µ)z; (0, ∞)] = [π(ν)z; (0, ∞)] (by π(µ) ∼Λ π(ν)) = π(ν) ◦ [z; (0, ∞)], giving π(µ) ∼Λ π(ν). (cid:3) The initial definition of maximal tails comes from [1] for the 1-graph C ∗- algebras (see [6, 2] also). Definition 3.6. Let Λ be a row-finite k-graph. A maximal tail nonempty subset T of Λ0 satisfying the following conditions: in Λ is a (1) for every λ ∈ Λ, s(λ) ∈ T implies r(λ) ∈ T , (2) for every v ∈ T and n ∈ Nk, there exists λ ∈ vΛ≤n such that s(λ) ∈ T , and (3) for every v, w ∈ T , there exist µ ∈ vΛ and ν ∈ wΛ such that s(µ) = s(ν). It is clear that if T is a maximal tail in Λ, then Λ0 \ T is hereditary and saturated. We write MT(Λ) for the collection of maximal tails in Λ. Corollary 3.7. Let Λ be a locally convex row-finite k-graph, and let Λ be the desourscification of Λ. Then Per(Λ) = Per(Λ). Moreover, if Λ0 is a maximal tail, then Per(Λ) is a subgroup of Zk. Proof. For every µ, ν ∈ Λ, Proposition 3.5 implies that µ ∼Λ ν if and only if µ ∼Λ ν. This follows Per(Λ) ⊆ Per(Λ). For the reverse containment, suppose µ, ν ∈ Λ, µ ∼Λ ν and d(µ)−d(ν) ∈ Per(Λ). Implication (1) ⇒ (3) of Proposition 3.5 yields that π(µ) ∼Λ π(ν) and d(µ)−d(π(µ)) = d(ν)−d(π(ν)). Thus d(µ) − d(ν) = d(π(µ)) − d(π(ν)) ∈ Per(Λ), and we have Per(Λ) ⊆ Per(Λ). 0 By [7, Lemma 4.3], if Λ0 is a maximal tail, then so is Λ statement follows immediately from the first and [2, Theorem 4.2(1)]. . Thus, the second (cid:3) Definition 3.8 ([2]). Given a row-finite k-graph Λ, we define the set HPer(Λ) :=(cid:8)v ∈ Λ0 : for all µ ∈ vΛ and m ∈ Nk with d(µ) − m ∈ Per(Λ), there exists ν ∈ vΛm such that µ ∼Λ ν(cid:9). We will show in Corollary 3.10 that HPer(Λ) is a hereditary subset of Λ0. Lemma 3.9. Let Λ be a locally convex row-finite k-graph such that Λ0 is 0 a maximal tail. Let Λ be the desourcification of Λ. Then for each v ∈ Λ , v ∈ HPer(Λ) if and only if π(v) ∈ HPer(Λ). PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 11 0 Λ 0 Proof. We first fix v ∈ Λ with π(v) ∈ HPer(Λ). Since Λ has no sources and is a maximal tail by [7, Lemma 4.3], Theorem 4.2 of [2] says that HPer(Λ) . So, we have v ∈ HPer(Λ) because π(v)Λv 6= ∅. Conversely, assume v ∈ HPer(Λ). Let µ ∈ π(v)Λ and m ∈ Nk such that is a hereditary subset of Λ 0 d(µ) − m ∈ Per(Λ). To show π(v) ∈ HPer(Λ), we must find ν ∈ π(v)Λ such that µ ∼Λ ν. Write v = [x; p] for some x ∈ Λ≤∞ and p ∈ Nk with p ∧ d(x) = 0. Let µ = [y; (0, d(µ))] for y ∈ π(v)Λ≤∞. Since x, y ∈ π(v)Λ≤∞ and p ∧ d(x) = 0, [14, Lemma 2.2] implies p ∧ d(y) = 0 and [x; (0, p)] = [y; (0, p)]. In particular, π(v) = [x; 0] = [y; 0] and v = [x; p] = [y; p]. If we and d(µ′) − m = d(µ) − m ∈ set µ′ := [y; (p, p + d(µ))], then µ′ ∈ vΛ such that µ′ ∼Λ ν′, which is of the form Per(Λ). So, there exists ν′ ∈ vΛ ν′ = [y; (p, p + m)] by Lemma 3.4. d(µ) m m We will show that ν := [y; (0, m)] is the desired path (i.e., µ ∼Λ ν). For, by Proposition 3.5, it suffices to prove π(µ) ∼Λ π(ν) and d(µ) − d(π(µ)) = d(ν) − d(π(ν)). Note that π(ν) = [y; (0, m ∧ d(y))] = π(ν′) and d(ν) − d(π(ν)) = m − m ∧ d(y) = d(ν′) − d(π(ν′)). Thus d(µ) − d(π(µ)) = d(µ) − d(µ) ∧ d(y) = d(µ′) − d(π(µ′)) = d(ν′) − d(π(ν′)) = d(ν) − d(π(ν)). (by µ′ ∼Λ ν′ and Proposition 3.5) (cid:3) Furthermore, since π(µ′) = [y; (0, d(µ) ∧ d(y))] = π(µ), implication (1) ⇒ (2) of Proposition 3.5 yields π(µ) = π(µ′) ∼Λ π(ν′) = π(ν). Consequently, µ ∼Λ ν and we have π(v) ∈ HPer(Λ). Corollary 3.10. Let Λ be a locally convex row-finite k-graph and Λ its desourcification. Then HPer(Λ) ∩ Λ0 = HPer(Λ). In particular, if Λ0 is a maximal tail, then HPer(Λ) is a nonempty hereditary subset of Λ0. Proof. To see HPer(Λ) ∩ Λ0 ⊆ HPer(Λ), take v ∈ HPer(Λ) ∩ Λ0. Suppose λ ∈ vΛ and m ∈ Nk such that d(λ)−m ∈ Per(Λ). Since d(λ)−m ∈ Per(Λ) = Per(Λ), such that λ ∼Λ µ. As s(µ) = s(λ) ∈ Λ0, we have µ = there exists µ ∈ vΛ π(µ) ∈ Λ, and hence λ ∼Λ µ by Proposition 3.5. This follows v ∈ HPer(Λ). To show HPer(Λ) ⊆ HPer(Λ) ∩ Λ0, we fix v ∈ HPer(Λ). Let λ ∈ vΛ and m ∈ Nk such that d(λ) − m ∈ Per(Λ). Then π(λ) ∈ vΛ, and for m′ := m −(cid:0)d(λ) − d(π(λ))(cid:1) we have d(π(λ)) − m′ = d(λ) − m ∈ Per(Λ) = Per(Λ). Since v ∈ HPer(Λ), there exists µ′ ∈ vΛm′ such that π(λ) ∼Λ µ′. Let λ = [x; (0, d(λ))] for some x ∈ vΛ≤∞. Then for µ := µ′ ◦ [x; (d(π(λ)), d(λ))] ∈ vΛ we have m m π(µ) = π(cid:0)µ′ ◦ [x; (d(π(λ)), d(λ))](cid:1) 12 HOSSEIN LARKI = π(µ′) ◦ π(cid:0)[x; (d(π(λ)), d(λ))](cid:1) = µ′ ◦ [x; d(π(λ))] = µ′, establishing π(λ) ∼Λ π(µ) and d(λ) − d(π(λ)) = d(µ) − d(π(µ)). Now Proposition 3.5 implies λ ∼Λ µ, and therefore v ∈ HPer(Λ). The second statement follows from [2, Theorem 4.2(2)]. Indeed, if Λ0 0 is a maximal tail, then so is Λ by [7, Lemma 4.3]. Hence, [2, Theorem 0 4.2] implies that HPer(Λ) is a nonempty hereditary subset of Λ . Therefore, HPer(Λ) = HPer(Λ) ∩ Λ0 is a hereditary subset of Λ0 as well. Furthermore, we may apply Lemma 3.9 to see that HPer(Λ) = HPer(Λ) ∩ Λ0 6= ∅. (cid:3) 4. primitive ideals in higher-rank graph C ∗-algebras In this section, we characterize the primitive ideal space of C ∗(Λ) for any locally convex, row-finite k-graph Λ. Our results generalize [2, Theorem 5.3] and [2, Corollary 5.4]. Definition 4.1. Let Λ be a locally convex row-finite k-graph. A boundary path x ∈ Λ≤∞ is called cofinal in case for every v ∈ Λ0, there exists n ∈ Nk such that n ≤ d(x) and vΛx(n) 6= ∅. If x ∈ Λ≤∞ is cofinal, we set F (x) := {λσn(x) : n ≤ d(x) and λ ∈ Λx(n)}. Notice that y ∈ F (x) if and only if σm(x) = σn(y) for some m, n ∈ Nk. Remark 4.2. In [2], the set F (x) was denoted by [x]. Since Λ is defined by equivalent classes, we use the notation F (x) instead of [x] here. The proof of [14, Proposition 3.5] derives the following. Lemma 4.3. Let x be a boundary path in Λ. Then x is cofinal in Λ if and only if the infinite path [x; (0, ∞)] in Λ is cofinal. Recall from Corollary 3.7 that if T is a maximal tail in Λ, then Per(ΛT ) is a subgroup of Zk. Let \Per(ΛT ) denote the character space of Per(ΛT ). So, for each character η of Per(ΛT ), there exists t ∈ Tk ∼= cZk such that η(m) = tm for all m ∈ Per(ΛT ). Lemma 4.4. Let Λ be a locally convex, row-finite k-graph and T a maximal tail in Λ. Suppose that x is a cofinal boundary path in the k-subgraph ΛT and consider the set F (x) in ΛT as in Definition 4.1. Let η ∈ \Per(ΛT ) and select some t ∈ Tk such that η(m) = tm for m ∈ Per(ΛT ). Then the representation πx,t : C ∗(Λ) → B(ℓ2(F (x))) defined by s(λ) = y(0) otherwise (λ ∈ Λ, y ∈ F (x)) (4.1) πx,t(sλ)ξy =(cid:26) td(λ)ξλy 0 is an irreducible representation on C ∗(Λ). PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 13 Proof. We will apply the desourcifying method on [2, Theorem 5.3(1)]. Let Λ be the desourcification of Λ. Recall that H := Λ0 \ T is a hereditary and saturated subset of Λ0. We first claim that H := (HΛ)0 is hereditary and saturated in Λ. Indeed, the hereditary property of H follows from that of H and the fact π(H) = (π(HΛ))0 = (Hπ(Λ))0 = (HΛ)0 = H. To see that 0 ) ⊆ H for some n ∈ Nk. Then, by H is saturated, fix v ∈ Λ n π(vΛ ) = π(v)Λ≤n, we have with s(vΛ n s(π(v)Λ≤n) = s(π(vΛ n )) = π(s(vΛ n )) ⊆ π(H) = H, and thus π(v) ∈ H ⊆ H as H is saturated. The fact π(v)Λv 6= ∅ combining with the hereditary property of H yield v ∈ H. Consequently, H is saturated and the claim holds. 0 Write T := Λ \ H. Since Λ T = Λ \ Λ H contains no sources, it is the desourcification of ΛT ; which means ΛT = Λ T . In particular, T is a maximal tail [7, Lemma 4.3]. Moreover, x := [x; (0, ∞)] is a cofinal infinite path in ΛT by Lemma 4.3. As η ∈ \Per(ΛT ) = \Per(ΛT ), [2, Theorem 5.3(1)] ensures that the representation πx,t : C ∗(Λ) → B(ℓ2(F (x)) satisfying πx,t(sλ)ξy =(cid:26) td(λ)ξλy 0 s(λ) = y(0) otherwise (λ ∈ Λ, y ∈ Λ ∞ ) is irreducible. Let φ : C ∗(Λ) → C ∗(Λ) be the natural embedding map and define πx,t := πx,t ◦ φ : C ∗(Λ) → B(ℓ2(F (x))). Observe that we may regard ℓ2(F (x)) as a subspace of ℓ2(F (x)) by identifying ξy with ξ[y;(0,∞)] for y ∈ F (x). By the fact πx,t(sλ)ξy = 0 unless s(λ) = y(0), ℓ2(F (x)) is an invariant subspace for operators π(sλ), λ ∈ Λ. Therefore, we can restrict operators πx,t(sλ) on ℓ2(F (x)) to get the desired representation πx,t : C ∗(Λ) → B(ℓ2(F (x))) satisfying formula (4.1). (cid:3) Now we have all requirements to characterize the primitive ideals of C ∗(Λ). For any ideal J of C ∗(Λ), we write TJ := Λ0\HJ = {v ∈ Λ0 : sv /∈ J} in the following. Theorem 4.5. Let Λ be a locally convex row-finite k-graph. union (1) If J is a primitive ideal of C ∗(Λ), then T := TJ = Λ0 \ HJ is a maximal tail. Also, there exists a unique η ∈ \Per(ΛT ) such that the ideal J ∩ (IHJ ∪HPer(ΛT )) in C ∗(Λ) is generated by (cid:8)sv : v ∈ Λ0 \ T(cid:9) (cid:8)sµ − η(cid:0)d(µ) − d(ν)(cid:1)sν : µ ∼ΛT ν and r(µ) = r(ν) ∈ HPer(ΛT )(cid:9). (2) The map (Tker πx,t, η) 7→ ker πx,t is a bijection betweenST ∈MT(Λ)(cid:0){T }× \Per(ΛT )(cid:1) and Prim(C ∗(Λ)), where t ∈ Tk satisfies η(m) = tm for Moreover, if x ∈ (ΛT )≤∞ is cofinal and t ∈ Tk with η(m) = tm for m ∈ Per(ΛT ), then we have J = ker πx,t. every m ∈ Per(ΛT ). 14 HOSSEIN LARKI 0 , T := Λ0\H, and T := Λ Proof. For statement (1), we fix a primitive ideal J of C ∗(Λ). Let Λ be the desourcification of Λ. Let J denote the ideal of C ∗(Λ) corresponding with J. (Recall that C ∗(Λ) is a full corner of C ∗(Λ) and J = hJi as an ideal of C ∗(Λ).) As the primitivity is preserved under Morita-equivalence, J is a primitive ideal of C ∗(Λ). For simplicity in notation, we set H := HJ ⊆ Λ0, 0 H := HJ ⊆ Λ \H. Since Λ is a row-finite k-graph 0 with no sources, [2, Theorem 5.3(2)] implies that T is a maximal tail in Λ and there exists unique η ∈ \Per(Λ T ) such that J = ker πx,t for any cofinal boundary path x in (HPer(Λ T )Λ T )∞ and t ∈ Tk satisfying η(m) = tm for m ∈ Per(ΛT ). By Corollary 3.10 we have HPer(Λ T ) ∩ (ΛT )0 = HPer(ΛT ) 6= ∅, so we may select such x with x(0) ∈ HPer(ΛT ). As seen in Lemma 2.5, x is of the form [x; (0, ∞)] for some x ∈ (ΛT )≤∞. Let I H∪HPer(ΛT ) denote the ideal of C ∗(Λ) generated by {sv : v ∈ H ∪ HPer(ΛT )}. The argument of [2, Theorem 5.3(2)] proves that the ideal ker πx,t ∩ I H∪HPer(ΛT ) in C ∗(Λ) is generated by the set B :=(cid:8)sv : v ∈ H(cid:9)∪(cid:8)sµ−η(cid:0)d(µ)−d(ν)(cid:1)sν : µ ∼Λ T ν, r(µ) = r(ν) ∈ HPer(Λ T )(cid:9). Pv∈Λ0 sv ∈ M(C ∗(Λ)). Observe that PΛ is a full projection, so ψ induces Let ψ : C ∗(Λ) → C ∗(Λ), a 7→ PΛaPΛ, be the restriction map, where PΛ := a one-to-one correspondence between ideal spaces of C ∗(Λ) and C ∗(Λ). In particular, we have ψ(J ) = J and ψ(J ∩ I H∪HPer(Λ T ) ) = J ∩ IH∪HPer(ΛT ). Therefore, the set ψ(B) =(cid:8)sv : v ∈ H(cid:9)∪(cid:8)sµ−η(cid:0)d(µ)−d(ν)(cid:1)sν : µ ∼ΛT ν, r(µ) = r(ν) ∈ HPer(ΛT )(cid:9) generates the ideal J ∩ IHPer(ΛT )∪H in C ∗(Λ). Furthermore, if φ : C ∗(Λ) → C ∗(Λ) is the embedding map, Lemma 4.4 shows that πx,t := πx,t ◦ φ : C ∗(Λ) → B(ℓ2(F (x))) is an irreducible represen- tation on C ∗(Λ). Since ker πx,t = (πx,t ◦ φ)−1(0) = φ−1(π−1 x,t (0)) = φ−1(J ) = J, this completes the proof of statement (1). Statement (2) follows immediately from (1) combining with [2, Corollary (cid:3) 5.4]. Notation 4.6. For every (T, η) ∈ ST ∈MT(Λ)(cid:0){T } × \Per(ΛT )(cid:1), we will de- note J(T,η) the primitive ideal of C ∗(Λ) associated to (T, η), as described in Theorem 4.5(2). We say that Λ is strongly aperiodic if for every saturated hereditary subset H ⊆ Λ0, the k-subgraph Λ\ΛH is aperiodic. According to [12, Theorem 5.3], PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 15 in case Λ is strongly aperiodic, then every ideals of C ∗(Λ) is gauge-invariant and of the form IH for some H ∈ HΛ. In the following, we write MTa(Λ) := {T ∈ MT(Λ) : ΛT is aperidic}. Corollary 4.7. Let Λ be a locally convex, row-finite k-graph. Then (1) The map T 7→ IΛ0\T is a bijection from MTa(Λ) onto the collection of primitive gauge-invariant ideals of C ∗(Λ). (2) If Λ is strongly aperiodic, then T 7→ IΛ0\T is a bijection between MT(Λ) and Prim(C ∗(Λ)). Proof. Statement (1): Let T be a maximal tail in Λ such that the sub- graph ΛT is aperiodic. We will show that the gauge-invariant ideal IΛ0\T is primitive. As Per(ΛT ) = {0} by Proposition 3.3, we have \Per(ΛT ) = {1} and hence HPer(ΛT ) = T . If ΛT is the desourcification of ΛT , [2, Lemma 5.2] says that the k-graph ΛT contains a cofinal infinite path [x; (0, ∞)]. Moreover, x is a cofinal boundary path in ΛT by Lemma 4.4. If we set J := ker πx,1, Theorem 4.5(1) yields {v ∈ Λ0 : sv ∈ J} = Λ0 \ T . Write H := Λ0 \ T for simplicity. Then ideal J + IH in the quotient C ∗-algebra ∼= C ∗(ΛT ) contains no vertex projections. Since ΛT is aperiodic, C ∗(Λ)/IH the Cuntz-Krieger uniqueness theorem implies that J = IH (cf. [14, Propo- sition 4.3]). Therefore IΛ0\T = ker πx,1, which is a primitive ideal of C ∗(Λ) by Lemma 4.4. Moreover, we may apply the fact HIH = H for every H ∈ HΛ and conclude injectivity of the map. For surjectivity, we fix a primitive gauge-invariant ideal IΛ0\T (Theorem 4.5(1) says that every primitive gauge-invariant deal of C ∗(Λ) is of the form IΛ0\T for some maximal tail T in Λ). Suppose on the contrary ΛT is pe- riodic. To derive a contradiction, we will show IΛ0\T 6= ker πx,t for every x ∈ (HPer(ΛT )ΛT )≤∞ and t ∈ Tk described in Theorem 4.5(1). So, let us fix some such x and t. Since ΛT is periodic, Proposition 3.3 implies that there are two distinct paths µ, ν ∈ ΛT such that µ ∼ΛT ν. By The- orem 4.5(1), a = sµ − td(µ)−d(ν)sν is a nonzero element in ker πx,t. But a /∈ IΛ0\T because s(µ), s(ν) ∈ T and image of a under the quotient map C ∗(Λ) → C ∗(Λ)/IΛ0\T = C ∗(ΛT ) is nonzero. Consequently, ker πx,t 6= IΛ0\T as desired. For statement (2), we notice that in case Λ is strongly aperiodic, each quotient k-graph ΛT = Λ\ΛH is aperiodic and every ideal of C ∗(Λ) is gauge- invariant [12, Theorem 5.3]. So, statement (2) is an immediate consequence of (1). (cid:3) 5. Maximal ideals We know that every ideal of C ∗(Λ) is the intersection of a family of In particular, maximal ideals of C ∗(Λ) are all primitive. primitive ones. Here, we want to use the characterization of Theorem 4.5 to determine certain maximal ideals of C ∗(Λ). 16 HOSSEIN LARKI Lemma 5.1. Let Λ be a locally convex row-finite k-graph. Suppose that T is a maximal tail in Λ and η1, η2 are two distinct characters of Per(ΛT ). If J(T,η1) and J(T,η2) are respectively the primitive ideals of C ∗(Λ) corresponding with (T, η1) and (T, η2) (see Notation 4.6), then neither J(T,η1) ⊆ J(T,η2) nor J(T,η2) ⊆ J(T,η1). Proof. We suppose J(T,η1) ⊆ J(T,η2) and derive a contradiction; the other case may be discussed analogously. Then, by Theorem 4.5(1), for every µ, ν ∈ ΛT with µ ∼ΛT ν, both elements sµ − η1(cid:0)d(µ) − d(ν)(cid:1)sν and sµ − η2(cid:0)d(µ) − d(ν)(cid:1)sν belong to J(T,η2). But, Theorem 4.5(1) says that such η1, η2 ∈ \Per(ΛT ) are unique for J(T,η2), hence η1 = η2. This contradicts our hypothesis. (cid:3) Proposition 5.2. Let Λ be a locally convex, row-finite k-graph. Let T be a maximal tail in Λ such that T is minimal in MT(Λ) under ⊆. Then for every η ∈ \Per(ΛT ), the primitive ideal J(T,η) associated to (T, η) is a maximal ideal of C ∗(Λ). Proof. Recall that HJ(T ,η) = Λ0 \ T for every η ∈ \Per(ΛT ). Take an ideal I of C ∗(Λ) such that J(T,η) ⊆ I ( C ∗(Λ). As I is an intersection of primitive ideals, without loss of generality, we can assume I is primitive. Then by Theorem 4.5(1), K := Λ0 \ HI is a maximal tail in Λ such that K ⊆ T . The minimality of T yields either K = ∅ or K = T . If K = ∅, then I = C ∗(Λ) which was not assumed. Thus we must have K = T . Since I is primitive, Theorem 4.5(1) implies that there exists η′ ∈ \Per(ΛT ) such that I = J(T,η′). Now apply Lemma 5.1 to have η = η′, and hence I = J(T,η). Consequently, J(T,η) is a maximal ideal of C ∗(Λ). (cid:3) Definition 5.3. Let Λ be a row-finite k-graph. We say Λ is cofinal in case all boundary paths x ∈ Λ≤∞ are cofinal (in the sense of Definition 4.1). By definition, one may easily verify that Λ is cofinal if and only if HΛ = {∅, Λ0}. Corollary 5.4. Let Λ a locally convex row-finite k-graph which is cofinal. Then all primitive ideals of C ∗(Λ), which are of the form JΛ0,η for η ∈ \Per(Λ), are maximal. Proof. We first claim that Λ0 is a maximal tail. To see this, it suffices to check condition (3) of Definition 3.6 only. So, fix some v, w ∈ Λ0. If x ∈ wΛ≤∞, then x is cofinal in Λ, so there exists n ≤ d(x) such that vΛx(n) 6= ∅. If we select some µ ∈ vΛx(n) and set ν := x(0, n), we then have µ ∈ vΛ and ν ∈ wΛ with s(µ) = s(ν). This follows the claim. Moreover, since Λ is cofinal, we have HΛ = {∅, Λ0}. Hence MT(Λ) = {Λ0}, and Theorem 4.5 says that primitive ideals of C ∗(Λ) are of the form JΛ0,η for η ∈ \Per(Λ). Now Propositions 5.2 follows immediately that such ideals are all maximal. (cid:3) PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 17 6. Decomposability of C ∗(Λ) In this section, we investigate the decomposability of a higher-rank graph C ∗-algebra C ∗(Λ). Our aim here is to find necessary and sufficient conditions on the underlying k-graph Λ such that C ∗(Λ) is decomposable. Furthermore, we show that direct summands in any decomposition of C ∗(Λ) are themselves isomorphic to higher-rank graph C ∗-algebras. Definition 6.1. We say that C ∗(Λ) is decomposable if there exist two non- zero C ∗-algebras A, B such that C ∗(Λ) = A ⊕ B. Otherwise, C ∗(Λ) is indecomposable. It is clear that in the case C ∗(Λ) = A ⊕ B, then A and B are two (closed) ideals of C ∗(Λ). The key step in our analysis is Corollary 6.4 below, which shows that any direct summand in a decomposition of C ∗(Λ) is a gauge-invariant ideal. To prove this, we use the structure of primitive ideals described in Section 4. Before that, we establish the following two lemmas. Lemma 6.2. Let T be a maximal tail in Λ. Then the collection {J(T,η) : η ∈ \Per(ΛT )} of primitive ideals of C ∗(Λ) is invariant under the gauge action γ. Proof. Fix η ∈ \Per(ΛT ) and take some t ∈ Tk such that η(m) = tm for all m ∈ Per(ΛT ). For every s ∈ Tk, we may define the character η′ : Per(ΛT ) → T, by m 7→ (ts−1)m. Note that we use the multi-index notation (ts−1)m := i )mi ∈ T here. Thus, for µ ∼ΛT ν with r(µ) = r(ν) ∈ HPer(ΛT ) we i=1(tis−1 Qk have γs(cid:0)sµ − η(cid:0)d(µ) − d(ν)(cid:1)sν(cid:1) = sd(µ)sµ − td(µ)−d(ν)(cid:0)sd(ν)sν(cid:1) = sd(µ)(cid:0)sµ − (ts−1)d(µ)−d(ν)sν(cid:1) = sd(µ)(cid:0)sµ − η′(cid:0)d(µ) − d(ν)(cid:1)sν(cid:1). In view of Theorem 4.5(1), this yields that generators of γs(cid:0)J(T,η)(cid:1) and J(T,η′) are the same, so γs(cid:0)J(T,η)(cid:1) = J(T,η′). We are done. Lemma 6.3. Let Λ be locally convex row-finite k-graph. If C ∗(Λ) decom- poses as A ⊕ B, then the collection D = {J ∈ Prim(C ∗(Λ)) : A ⊆ J} is invariant under the gauge action γ. (cid:3) Proof. First, note that the decomposability implies that Prim(C ∗(Λ)) = Prim(A)⊕Prim(B). In particular, Prim(A) and Prim(B) are clopen subsets of Prim(C ∗(Λ)). Moreover, we have J ∈ Prim(B) if and only if A ⊕ J ∈ Prim(C ∗(Λ)), hence Prim(B) is homeomorphic to D. Let us fix an arbitrary J0 ∈ Prim(C ∗(Λ)) with A ⊆ J0 (i.e. J0 ∈ D). By Theorem 4.5, T := {v ∈ Λ0 : sv /∈ J0} is a maximal tail and there exists η0 ∈ \Per(ΛT ) such that J0 = J(T,η0). Using Theorem 4.5(2), we may define the homeomorphic embedding Ψ : \Per(ΛT ) → Prim(C ∗(Λ)), by η 7→ J(T,η). 18 HOSSEIN LARKI Corollary 3.7 implies that \Per(ΛT ) is a subgroup of cZk ∼= Tk, and thus \Per(ΛT ) ∼= Tl for some 0 ≤ l ≤ k. In particular, Ψ( \Per(ΛT )) is a connected subset of Prim(C ∗(Λ)). Since Ψ(η0) = J0 ∈ D and D is clopen (because Prim(B) is), this follows that Ψ( \Per(ΛT )) must be entirely contained in D. Now, for every t ∈ Tk, Lemma 6.2 yields that γt(J0) = J(T,η) for some η ∈ \Per(ΛT ). Since such J(T,η) lies in Rang(Ψ) ⊆ D, this concludes the result. Corollary 6.4. Let Λ be a locally convex row-finite k-graph. If C ∗(Λ) de- composes into C ∗(Λ) = A⊕ B, then both A and B are gauge-invariant ideals of C ∗(Λ). Proof. We know that every ideal of C ∗(Λ) is the intersection of primitive ideals containing it. So, it suffices to show that the collections D = {J ∈ Prim(C ∗(Λ)) : A ⊆ J} and D′ = {J ∈ Prim(C ∗(Λ)) : B ⊆ J} are invariant under the gauge action. However, this follows from Lemma 6.3 immediately. (cid:3) (cid:3) Once we find that C ∗(Λ) can be decomposed only by gauge-invariant ideals, we may investigate its decomposability by properties of the underly- ing k-graph Λ. Definition 6.5. Let Λ be a row-finite k-graph. For v ∈ Λ0, denote T (v) := {s(µ) : µ ∈ vΛ} the smallest hereditary subset of Λ0 containing v. When H1, H2 ∈ HΛ with H1 ⊇ H2, we also set the following subsets of Λ0: ∆(H1, H2) := {v ∈ H1 : T (v) ∩ H2 = ∅}, and Ω(H1, H2) := {v ∈ H1 \ H2 : T (v) ∩ H2 6= ∅} = H1 \(cid:0)H2 ∪ ∆(H1, H2)(cid:1). Clearly, ∆(H1, H2) is always a hereditary subset of Λ0 such that ∆(H1, H2)∩ H2 = ∅. So, we have Σ(∆(H1, H2)) ∩ H2 = ∅ also. Now we define the relation ≻ on HΛ by: H1 ≻ H2 if and only if (1) H1 ⊇ H2, (2) ∆(H1, H2) 6= ∅ (i.e., there exists v ∈ H1 \ H2 with T (v) ∩ H2 = ∅), and (3) for every v ∈ Ω(H1, H2), there exists n ∈ Nk such that s(vΛ≤n) ⊆ H2 ∪ ∆(H1, H2). We now determine higher-rank graph C ∗-algebras C ∗(Λ) which are de- It is the generalization of [5, Theorem 4.1] for higher-rank composable. graph C ∗-algebras. Theorem 6.6. Let Λ be a locally convex row-finite k-graph. Then the fol- lowing are equivalent. (1) C ∗(Λ) is decomposable. (2) There exist nonempty, disjoint H1, H2 ∈ HΛ with this property that for every v ∈ Λ0 \ (cid:0)H1 ∪ H2(cid:1), there is some n ∈ Nk such that s(vΛ≤n) ⊆ H1 ∪ H2. PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 19 (3) There exists nonempty H ∈ HΛ such that Λ0 ≻ H. Moreover, in the case (2) we have C ∗(Λ) = IH1 ⊕ IH2 ∼= C ∗(Λ \ ΛH2) ⊕ C ∗(Λ \ ΛH1), and in the case (3), C ∗(Λ) ∼= C ∗(Λ \ ΛH) ⊕ IH ∼= C ∗(Λ \ ΛH) ⊕ C ∗(Λ \ ΛH ′) where H ′ = Σ(∆(Λ0, H)). Proof. We will first prove (1) ⇐⇒ (2), and then (2) ⇐⇒ (3). (1) =⇒ (2): Let C ∗(Λ) = A⊕B be a decomposition for C ∗(Λ). According to Corollary 6.4, A and B are gauge-invariant ideals of C ∗(Λ); so, there exist H1, H2 ∈ HΛ such that A = IH1 and B = IH2 [12, Theorem 5.2(a)]. = ∅, while H1 ∪ H2 = Λ0 because Moreover, we have H1 ∩H2 = HIH1 C ∗(Λ) = IH1 + IH2 = IH1∪H2 . This follows statement (2). ∩HIH2 (2) =⇒ (1): If H1, H2 ∈ HΛ satisfy conditions (2), then H1 ∪ H2 = Λ0. Thus, [12, Theorem 5.2(a)] implies IH1 ∩IH2 = IH1∩H2 = (0) and IH1 +IH2 = IH1∪H2 = C ∗(Λ). Consequently, C ∗(Λ) decomposes as IH1 ⊕ IH2. (2) =⇒ (3): If H1, H2 ∈ HΛ satisfy (2), it is routine to check that Λ0 ≻ H1, H2. (3) =⇒ (2): Let Λ0 ≻ H for some nonempty H ∈ HΛ. Set H ′ := Σ(∆(Λ0, H)) the saturated closure of ∆(Λ0, H). Since ∆(Λ0, H) is hered- itary, H ′ is a hereditary and saturated subset of Λ0 [12, Lemma 5.1]. For each v ∈ H ′, there is n ∈ Nk such that s(vΛ≤n) ⊆ ∆(Λ0, H), which con- cludes H ∩ H ′ = ∅ because ∆(Λ0, H) ∩ H = ∅. Now the property Λ0 ≻ H implies that H and H ′ satisfy the conditions of (2). For the last statement, it suffices to note that if C ∗(Λ) = IH1 ⊕ IH2, then and analogously IH2 ∼= C ∗(Λ)/IH2 IH1 ∼= C ∗(Λ \ ΛH1). Now the proof is complete. ∼= C ∗(Λ \ ΛH2) (cid:3) 7. decomposing C ∗(Λ) with indecomposable components In the final section, we want to determine higher-rank graph C ∗-algebras C ∗(Λ) which are direct sums of finitely many indecomposable C ∗-algebras. To this end, we use specific chains of hereditary and saturated subsets of Λ0. Definition 7.1. Let Λ be a row-finite k-graph. A sequence C := {Hi}n i=1, for n ∈ N ∪ {∞}, of saturated hereditary subsets of Λ0 is called a chain in HΛ whenever H1 = Λ0 and Hi ≻ Hi+1 for all 1 ≤ i < n (see Definition 6.5). Then, n = C is length of C. A refinement of C is a chain C′ in HΛ such that C ( C′. In case there are no refinements for C, we say C is a maximal chain. If C : Λ0 = Lemma 7.2. Let Λ be a locally convex row-finite k-graph. H1 ≻ · · · ≻ Hn is a finite chain in HΛ, then there exists a decomposition C ∗(Λ) = IK1 ⊕ · · · ⊕ IKn such that Ki's are pairwise disjoint and Hi = 20 HOSSEIN LARKI Σ(Sn IKj are indecomposable. j=i Kj) for each i. Moreover, C is a maximal chain if and only if all Proof. Since H1 ≻ H2, implication (3) ⇒ (1) of Theorem 6.6 implies that ∼= C ∗(Λ \ ΛK1), C ∗(Λ) = IK1 ⊕ IH2 where K1 = Σ(∆(H1, H2)). Then IH2 and the saturation of H2 in the subgraph Λ \ ΛK1 is all (Λ \ ΛK1)0. So ∼= (Λ \ ΛK1)0 ≻ H3 in HΛ\ΛK1 because H2 ≻ H3, and we have again IH2 C ∗(Λ \ ΛK1) = IK2 ⊕ IH3 for some K2 ∈ HΛ\ΛK1. Since IK2 is a direct summand of C ∗(Λ), it is a gauge-invariant ideal of C ∗(Λ) by Corollary 6.4. Hence, K2 ∈ HΛ. Continuing this process gives a decomposition C ∗(Λ) = IK1 ⊕ IK2 ⊕ · · · ⊕ IKn for C ∗(Λ), where Kn = Hn. The above process says that IHi = Ln for each 1 ≤ i ≤ n, which follows Hi = Σ(Sn j=i IKj j=i Kj) by Theorem 5.2(a) of [12]. For the last statement, if some IKj decomposes as A ⊕ B, then A and for some ∅ 6= B are two direct summands of C ∗(Λ); so IKj = IK ′ K ′ j ∈ HΛ by applying Corollary 6.4. Therefore, we have a refinement ⊕ IK ′′ j, K ′′ j j Λ0 = H1 ≻ · · · ≻ Hj ≻ Σ(cid:0)Hj+1 ∪ K ′ j(cid:1) ≻ Hj+1 ≻ · · · ≻ Hn of C. Conversely, if C′ : Λ0 = H1 ≻ · · · ≻ Hj ≻ H ≻ Hj+1 ≻ · · · ≻ Hn is a refinement of C, then IKj will be decomposable by (3) ⇒ (1) in Theorem 6.6. This completes the proof. (cid:3) Definition 7.3. Let Λ be a locally convex row-finite k-graph. We say that C ∗(Λ) is n-decomposable, for n ≥ 1, if there exists an n-term decomposition C ∗(Λ) = A1 ⊕ · · · ⊕ An for C ∗(Λ) such that each direct summand Ai is indecomposable (1-decomposable equals to indecomposable). Note that such n, if exists, is unique (see the last paragraph in Proof of Theorem 7.4 below). We now characterize n-decomposable higher-rank graph C ∗-algebras. Theorem 7.4. Let Λ be a locally convex row-finite k-graph. Then C ∗(Λ) is n-decomposable for some n ≥ 1 if and only if n = max{C : C is a chain in HΛ}. Moreover, if this is the case, then there is a unique decomposition (up to per- mutation) C ∗(Λ) ∼= C ∗(Λ1) ⊕ · · · ⊕ C ∗(Λn), where Λi are k-subgraphs of Λ and each C ∗(Λi) is indecomposable. Remark 7.5. In view of Lemma 6.2, every maximal chain of length n ≥ 1 in HΛ induces an n-term decomposition for C ∗(Λ) with indecomposable direct summands. Since such decompositions are unique up to permutation, Theorem 7.4 follows that maximal chains in HΛ are all of length n. In particular, in this case, any chain in HΛ has a refinement with length n. Proof of Theorem 7.4. (=⇒): Let C ∗(Λ) = A1 ⊕· · ·⊕An be a decomposition of C ∗(Λ) with indecomposable summands. By Corollary 6.4, each Aj is a PRIMITIVE IDEALS AND DECOMPOSABILITY OF k-GRAPH C ∗-ALGEBRAS 21 gauge-invariant ideal of C ∗(Λ), so Aj = IKj for some Kj ∈ HΛ. If we set j=i Kj) for 1 ≤ i ≤ n, Lemma 7.2 implies that Hi := Σ(Sn C : Λ0 = H1 ≻ · · · ≻ Hn is a chain in HΛ. We hence have n ≤ max{C : C is a chain in HΛ}. On the other hand, assume Λ0 = H1 ≻ · · · ≻ Hl is a chain in HΛ. By Lemma 7.2, there is a decomposition C ∗(Λ) = IK1 ⊕ · · · ⊕ IKl such that j=1 IKi ∩Aj, and thus either IKi ∩ Aj = (0) or Aj ⊆ IKi because Aj is indecomposable. This follows l ≤ n, and consequently n = max{C : C is a chain in HΛ}. j=i Kj) for 1 ≤ i ≤ l. So, for each i, we have IKi =Ln Hi := Σ(Sl The "⇐=" part follows from Lemma 7.2. : C is a chain in HΛ} and C is a maximal chain in HΛ with C = n, then C induces a decomposition C ∗(Λ) = A1 ⊕ · · · ⊕ An with indecomposable summands. Therefore, C ∗(Λ) is n-decomposable. Indeed, if n = max{C For the last statement, suppose that C ∗(Λ) is n-decomposable and C ∗(Λ) = IK1 ⊕ · · · ⊕ IKn. Then, for every 1 ≤ i ≤ n, we have C ∗(Λ) = IKi ⊕ IK ′ where K ′ i i := Σ(Sj6=i Kj); in particular, C ∗(Λ) IKi ∼= ∼= C ∗(Λ \ ΛK ′ i). IK ′ i Ln Ln := Λ \ ΛK ′ i=1 C ∗(Λi). ∼=Ln Therefore, Λi i are k-subgraphs of Λ and we have C ∗(Λ) = i=1 IKi To see that this decomposition is unique, assume C ∗(Λ) = A1 ⊕ · · · ⊕ Am with indecomposable components. For any 1 ≤ j ≤ m, Aj = Aj ∩ C ∗(Λ) = i=1 Aj ∩ IKi. Since Aj is indecomposable, there is 1 ≤ lj ≤ n such that and Aj ∩ IKi = (0) for i 6= lj. By a same argument we have ⊆ Aj also, and hence they are equal. This follows that m = n and j=1 is a permutation of {IKi}n (cid:3) Aj ⊆ IKlj IKlj {Aj}n i=1. The following is an immediate consequence of Theorem 7.4. Corollary 7.6. Let Λ be a locally convex row-finite k-graph. If Λ contains only finitely many saturated hereditary subsets, then there exist finitely many k-subgraphs Λ1, . . . , Λn of Λ such that C ∗(Λ) ∼= C ∗(Λ1) ⊕ · · · ⊕ C ∗(Λn) and each C ∗(Λi) is indecomposable. Acknowledgement. The author would like to kindly acknowledge the referee for his/her useful comments and pointing the related article [18]. References [1] T. Bates, D. Pask, I. Raeburn and W. Szyma´nski, The C ∗-algebras of row-finite graphs, New York J. Math. 6 (2000), 307-324. [2] T. Carlsen, S. Kang, J. Shotwell, and A. Sims, The primitive ideals of the Cuntz- Krieger algebra of a row-finite higher-rank graph with no sources, J. Funct. Anal. 266 (2014), 2570-2589. [3] K.R. Davidson and D. Yang, Periodicity in rank 2-graph algebras, Canad. J. Math. 61(6) (2009), 1239-1261. 22 HOSSEIN LARKI [4] C. Farthing, Removing sources from higher rank graphs, J. Operator Theory. 60 (2008), 165-198. [5] J.H. Hong and W. Szyma´nski, The primitive ideal space of the C ∗-algebras of infinite graphs, J. Math. Soc, Japan bf 56 (2004), 45-64. [6] S. Kang and D. Pask, Aperiodicity and primitive ideals of row-finite k-graphs, Inter- nat. J. Math. 25(3) (2014), 1450022 (25 pages). [7] M. Kashoul-Radjabzadeh, H. Larki, and A. Aminpour, Prime and primitive Kumjian- Pask algebras, J. Algebra Appl. 16(9) (2017), 1750169 [14 pages]. [8] A. Kumjian and D. Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1-20. [9] H. Larki, Purely infinite simple Kumjian-Pask algebras, Forum Math. 30(1) (2018), 253-268. [10] D. Pask, I. Raeburn, M. Rørdam, and A. Sims, Rank-two graphs whose C ∗-algebras are direct limits of circle algebras, J. Funct. Anal. 239 (2006), 137-178. [11] D. Pask, I. Raeburn, N.A. Weaver, A family of 2-graphs arising from two-dimensional subshifts, Ergodic Theory Dynam. Systems 29 (2009), 1613-1639. [12] I. Raeburn, A. Sims, and T. Yeend, Higher rank graphs and their C ∗-algebras, Proc. E dinb. Math. Soc. 46 (2003), 99-115. [13] I. Raeburn, A. Sims, and T. Yeend, The C ∗-algebras of finitely aligned higher-rank graphs, J. Funct. Anal. 213 (2004), 206-240. [14] D. Robertson and A. Sims, Simplicity of C ∗-algebras associated to row-finite locally convex higher-rank graphs, Israel J. Math. 172 (2009), 171-192. [15] G. Robertson and T. Steger, Affine buildings, tiling systems and higher rank Cuntz- Krieger algebras, J. Reine Angew. Math. 513 (1999), 115-144. [16] J. Shotwell, Ideals in k-graph algebras, PhD thesis, Arizona State University, 2009. [17] J. Shotwell, Simplicity of finitely aligned k-graph C ∗-algebras, J. Operator Theory 67 (2012), 335- 347. [18] S.B.G. Webster, The path space of a higher-rank graph, Studia Math. 204 (2011), 155-185. Department of Mathematics, Shahid Chamran University of Ahvaz, Ahvaz, Iran E-mail address: [email protected]
1103.5356	8	1103	2011-12-12T14:03:29	Examples of mixing subalgebras of von Neumann algebras and their normalizers	[ "math.OA" ]	We discuss different mixing properties for triples of finite von Neumann algebras $B\subset N\subset M$, and we introduce families of triples of groups $H<K<G$ whose associated von Neumann algebras $L(H)\subset L(K)\subset L(G)$ satisfy $\mathcal{N}_{L(G)}(L(H))"=L(K)$. It turns out that the latter equality is implied by two conditions: the equality $\mathcal{N}_G(H)=K$ and the above mentioned mixing properties. Our families of examples also allow us to exhibit examples of pairs $H<G$ such that $L(\mathcal{N}_G(H))\not=\mathcal{N}_{L(G)}(L(H))"$.	math.OA	math	EXAMPLES OF MIXING SUBALGEBRAS OF VON NEUMANN ALGEBRAS AND THEIR NORMALIZERS PAUL JOLISSAINT Abstract. We discuss different mixing properties for triples of finite von Neu- mann algebras B ⊂ N ⊂ M , and we introduce families of triples of groups H < K < G whose associated von Neumann algebras L(H) ⊂ L(K) ⊂ L(G) satisfy NL(G)(L(H))′′ = L(K). It turns out that the latter equality is implied by two conditions: the equality NG(H) = K and the above mentioned mixing properties. Our families of examples also allow us to exhibit examples of pairs H < G such that L(NG(H)) 6= NL(G)(L(H))′′. . A O h t a m [ 8 v 6 5 3 5 . 3 0 1 1 : v i X r a 1. Weakly and strongly mixing finite von Neumann algebras The main purpose of the present paper is to present families of triples of groups H ⊳ K < G whose associated von Neumann algebras have mixing properties which imply that the L(K) is the von Neumann algebra generated by the normalizer of L(H) in L(G). Thus this section is devoted to the discussion of mixing properties for arbitrary finite von Neumann algebras. Let 1 ∈ B ⊂ N ⊂ M be finite von Neumann algebras being endowed with a normal, finite, faithful, normalized trace τ . The normalizer of B in M is denoted by NM (B), and it is the group of all unitary elements u ∈ U (M ) such that uBu∗ = B. We assume for simplicity that M has separable predual. We denote by EB (resp. EN ) the trace-preserving conditional expectation from M onto B (resp. N ), and we set M ⊖ N = {x ∈ M : EN (x) = 0}. For v ∈ U (B), let σv ∈ Aut(M ) be defined by σv(x) = vxv∗. This gives an action of U (B) on M that preserves the subspace M ⊖ N . If G is a group, every element x of the associated von Neumann algebra L(G) admits a Fourier series decomposition x = Pg∈G x(g)λg where x(g) = τ (xλ−1 2. If H is a subgroup of G, then L(H) identifies to the von Neumann subalgebra of L(G) formed by all elements y such that y(g) = 0 for every g ∈ G \ H. The corresponding conditional expectation EL(H) satisfies EL(H)(x) = Ph∈H x(h)λh for every x ∈ L(G). g ) and Pg x(g)2 = kxk2 Our first definition is an extension of Definitions 2.1 and 3.4 of [6] to the case of triples as above. Definition 1.1. Consider a triple 1 ∈ B ⊂ N ⊂ M of finite von Neumann algebras as above and the action σ of U (B) on M by conjugation. Date: October 24, 2018. 2010 Mathematics Subject Classification. Primary 46L10; Secondary 22D25. Key words and phrases. Finite von Neumann algebras, relative weak mixing subalgebras, rel- ative weak asymptotic homomorphism property, discrete groups. To appear in the Bulletin of the Belgian Mathematical Society Simon Stevin. 1 2 PAUL JOLISSAINT (1) We say that B is weakly mixing in M relative to N if, for every finite set F ⊂ M ⊖ N and every ε > 0, one can find v ∈ U (B) such that kEB(xσv(y))k2 = kEB(xvy)k2 < ε ∀x, y ∈ F. If B = N , we say that B is weakly mixing in M . (2) If B is diffuse, we say that B is strongly mixing in M relative to N if lim n→∞ kEB(xuny)k2 = 0 for all x, y ∈ M ⊖ N and all sequences (un) ⊂ U (B) which converge to 0 in the weak operator topology. If B = N , we say that B is strongly mixing in M . The next definition introduces a relative version of the so-called weak asymptotic homomorphism property; the latter was used first by Robertson, Sinclair and Smith in [9] for MASA's in order to get an easily verifiable criterion for singularity. It was proved next by Sinclair, Smith, White and Wiggins in [10] that, conversely, any singular MASA has the weak asymptotic homomorphism property. The relative version of the above property was introduced by Chifan in [1] in order to prove that if A is a masa in a separable type II1 factor M then the triple A ⊂ NM (A)′′ ⊂ M has the relative weak asymptotic homomorphism property. He used it to prove that, if (Mi)i≥1 is a sequence of finite von Neumann algebras and if (Ai)i≥1 is a sequence such that Ai ⊂ Mi is a MASA for every i, then Oi≥1 NMi (Ai)′′ = (NNiMi (Oi Ai))′′. This relative property is related to one-sided quasi-normalizers, as it was proved by Fang, Gao and Smith in [3]. See Theorem 1.4 below. Definition 1.2. Let 1 ∈ B ⊂ N ⊂ M be a triple of finite von Neumann algebras as in Definition 1.1 above. (1) The triple of algebras 1 ∈ B ⊂ N ⊂ M has the relative weak asymptotic homomorphism property if there exists a net of unitaries (ui)i∈I in B such that lim i∈I kEB(xuiy) − EB(EN (x)uiEN (y))k2 = 0 for all x, y ∈ M . (2) The one-sided quasi-normalizer of B in M is the set of all elements x ∈ M for which there exists a finite set {x1, . . . , xn} ⊂ M such that Bx ⊂ n X i=1 xiB. Following [3], we denote the set of these elements by qN (1) M (B). Remark 1.3. (1) If B ⊂ N ⊂ M and if B is diffuse and strongly mixing in M relative to N , then it is obviously weakly mixing in M relative to N . (2) If B is strongly mixing in M relative to N , then every diffuse von Neumann algebra 1 ∈ D ⊂ B is also strongly mixing in M relative to N . EXAMPLES OF MIXING SUBALGEBRAS OF VON NEUMANN ALGEBRAS 3 (3) The following identity EB(xuy) − EB(EN (x)uEN (y)) = EB([x − EN (x)]u[y − EN (y)]), which holds for every u ∈ U (B) and all x, y ∈ M , implies that the relative weak mixing property and the relative weak asymptotic homomorphism property are equivalent. The following theorem is the main result of [3]. Theorem 1.4. (J. Fang, M. Gao and R. R. Smith) Let 1 ∈ B ⊂ N ⊂ M be a triple of finite von Neumann algebras with separable predual. Then the following conditions are equivalent: (1) The triple B ⊂ N ⊂ M has the relative weak asymptotic homomorphism property, or, equivalently, B is weakly mixing in M relative to N . (2) The one-sided quasi-normalizer qN (1) M (B) is contained in N . The next technical result is inspired by Proposition 4.1 of [8] and Lemma 2.1 of [9]; it reminds also heredity properties of relative weak mixing from [3]. Proposition 1.5. Let B ⊂ N ⊂ M be a triple as above. (1) Assume that B is weakly mixing in M relative to N . Then for every nonzero projection e ∈ B, the reduced algebra eBe is weakly mixing in eM e relative to eN e. Moreover, one has for every u ∈ U (M ): kEB − EuBu∗ k∞,2 ≥ ku − EN (u)k2. (2) If B is strongly mixing in M relative to N , then for every diffuse unital von Neumann subalgebra D of B, one has for every u ∈ U (M ): kED − EuDu∗ k∞,2 ≥ ku − EN (u)k2. (3) If Bi ⊂ Ni ⊂ Mi are triples of finite von Neumann algebras, i = 1, 2, and if Bi is weakly mixing in Mi relative to Ni, then B1⊗B2 is weakly mixing in M1⊗M2 relative to N1⊗N2. Proof. The first part of claim (1) is a straightforward consequence of Corollary 6.3 of [3] and claim (3) follows from Proposition 6.1 of the same article. We prove claim (2) because the proof of the last assertion in (1) is similar to that of statement (2). Thus fix a diffuse von Neumann subalgebra D of B, a unitary operator u ∈ U (M ), and let us consider x = u∗ − EN (u∗) and y = u − EN (u) ∈ M ⊖ N and let ε > 0. By the above remark, one has for every v ∈ U (D): ED(xvy) = ED(u∗vu) − ED(EN (u∗)vEN (u)) As D is diffuse, there exists a sequence of unitaries (vn) ⊂ U (D) which converges to 0 with respect to the weak operator topology. Since B is strongly mixing in M relative to N , there exists a positive integer n such that kEB(xvny)k2 ≤ ε. As ED = EDEB, we have kED(xvny)k2 ≤ ε as well. The above computations give kED(u∗vnu)k2 ≤ kED(EN (u∗)vnEN (u))k + ε ≤ kEN (u)k2 + ε. 4 PAUL JOLISSAINT We get then: kED − EuDu∗ k2 ∞,2 ≥ kvn − EuDu∗ (vn)k2 2 = ku∗vnu − ED(u∗vnu)k2 2 = 1 − kED(u∗vnu)k2 2 ≥ 1 − (kEN (u)k2 + ε)2 = 1 − kEN (u)k2 = ku − EN (u)k2 2 − 2εkEN (u)k2 − ε2 2 − 2εkEN (u)k2 − ε2. As ε is arbitrary, we get the conclusion. (cid:3) We end the present section with a first class of examples of relative strongly mixing algebras; its proof is inspired by that of Lemma 2.2 in [2]. Proposition 1.6. Let 1 ∈ B ⊂ N and Q be arbitrary finite von Neumann algebras with separable preduals. Assume moreover that B is diffuse. Then B is strongly mixing relative to N in the free product algebra M = N ∗ Q. Proof. Let us recall that the free product N ∗ Q is the von Neumann algebra gen- erated by the unital ∗-algebra P := C1 ⊕ M n≥1   M i16=···6=in M 0 i1 ⊗ · · · ⊗ M 0 in   1 = {x ∈ N : τ (x) = 0} and M 0 where M 0 (We have chosen and fixed finite, normal, faithful normalized traces on N and Q.) Thus every element of P is a finite linear combination of words w of the following form: either w ∈ N , or w is a finite product of letters of zero trace and at least one letter belongs to Q. 2 = {x ∈ Q : τ (x) = 0}. Then fix a sequence (un) ⊂ U (B) which converges weakly to zero and two words x, y ∈ P as above. If x, y ∈ N , then EB(xuny) − EB(EN (x)unEN (y)) = 0. If x, y ∈ Q0, then x, y ∈ M ⊖ N and xuny decomposes as a sum xuny = x(un − τ (un))y + τ (un)xy where the first term is a reduced word, hence EB(x(un − τ (un))y) = 0, and kEB(xuny)k2 ≤ τ (un)kxyk2 → 0 as n → ∞. If x = x1b1 and y = b2y2 where b1, b2 ∈ N and x1 ends with an element in Q0 and y2 starts with an element in Q0, then EB(xuny) − EB(EN (x)unEN (y)) = EB(x1b1unb2y2) −EB(EN (x1)b1unb2EN (y2)) = EB(x1{b1unb2 − τ (b1unb2)}y2) +τ (b1unb2)EB(x1y2) = τ (b1unb2)EB(x1y2). As the words x1{b1unb2 − τ (b2unb1)}y2, x1 and y2 are reduced, they belong to the kernels of the conditional expectations EB and EN . Hence we get kEB(xuny) − EB(EN (x)unEN (y))k2 ≤ τ (b1unb2)kx1y2k → 0 as n → ∞. The remaining case, when exactly one of x and y∗ ends with a letter from Q, is dealt with as in the previous case. (cid:3) EXAMPLES OF MIXING SUBALGEBRAS OF VON NEUMANN ALGEBRAS 5 2. The case of group algebras As will be seen below, the associated von Neumann algebras of suitable triples of groups H < K < G give rise to examples of von Neumann algebras which satisfy the relative weak or strong mixing properties. Our next definition generalizes the so-called conditions (SS) and (ST) of [6] to not necessarily abelian groups. We are indebted to the referee for having suggested the following more intuitive formulation. Definition 2.1. Let G be a countable group and let H < K be two infinite subgroups of G. (a) We say that the triple H < K < G satisfies condition (SS) if all orbits of the natural action H y (G \ K)/H are infinite. (b) The triple H < K < G satisfies condition (ST) if all stabilizers of the natural action H y (G \ K)/H are finite. When H < G we say that the pair H < G satisfies condition (SS) (resp. (ST)) if it is the case for H = K < G. We present below technical characterizations of both properties. In order to do that, recall on the one hand that all orbits of H y (G \ K)/H are infinite if and only if, for every finite subset Y ⊂ G \ K, one can find h ∈ H such that hY ∩ Y = ∅ (see for instance Lemma 2.2 in [7]). On the other hand, for g, h ∈ G \ K, set E(g, h) = {γ ∈ H : gγh ∈ H} = g−1Hh−1 ∩ H and E(g) = E(g−1, g) = gHg−1 ∩ H. The latter is a subgroup of G. Then it is easy to see that, for arbitrary γ0 ∈ E(g, h), one has E(g, h) ⊂ E(g−1)γ0. Making use of these observations, the proof of the following lemma is straight- forward. Lemma 2.2. Let H < K < G be three infinite, countable groups. (a) The triple H < K < G satisfies condition (SS) if and only if for every nonempty finite set F ⊂ G \ K, there exists h ∈ H such that F hF ∩ H = ∅. (b) The following conditions are equivalent: (1) H < K < G satisfies condition (ST); (2) for every finite set F ⊂ G \ K, there exists an exceptional finite set E ⊂ H such that F hF ∩ H = ∅ for every h ∈ H \ E. (3) E(g, h) is a finite set for all g, h ∈ G \ K; (4) E(g) is a finite group for every g ∈ G \ K. Using the same type of arguments as in [6], one proves the following generaliza- tion of Proposition 2.3 and of Theorem 3.5 in [6]: Proposition 2.3. Let H < K < G be a triple of countable groups. Then: (a) The triple H < K < G satisfies condition (SS) if and only if L(H) is weakly mixing in L(G) relative to L(K). (b) The triple H < K < G satisfies condition (ST) if and only if L(H) is strongly mixing in L(G) relative to L(K). Moreover, if H < K < G satisfies condition (SS) (resp. (ST)) and if σ : G → Aut(Q, τ ) is a trace-preserving action of G on some finite von Neumann algebra (Q, τ ), then the crossed product algebra Q ⋊σ H is weakly (resp. strongly) mixing in Q ⋊σ G relative to Q ⋊σ K. 6 PAUL JOLISSAINT Remark 2.4. (1) An infinite subgroup H of a group G is called malnormal if, for every g ∈ G \ H, one has H ∩ gHg−1 = {1}. Hence such a pair H < G satisfies condition (ST). More generally, H is said to be almost malnormal if, for every g ∈ G \ H, the subgroup H ∩ gHg−1 is finite. Thus, if H < K < G is a triple that satisfies condition (ST), one can say equivalently that H is almost malnormal in G relative to K. (2) Following [11], we say that a subgroup H of a group G is relatively malnormal if there exists an intermediate subgroup K < G of infinite index such that gHg−1 ∩H is finite for all g ∈ G \ K. This means precisely that the triple H < K < G satisfies condition (ST). If G is a group and H is a subgroup of G, there is a straightforward analogue of the one-sided quasi-normalizer of a subalgebra: we denote by qN (1) G (H) the set of elements g ∈ G for which there exists finitely many elements g1, . . . , gn ∈ G such that Hg ⊂ Sn i=1 giH. In view of Theorem 1.3, it is natural to ask whether the triple of algebras L(H) ⊂ L(K) ⊂ L(G) has the relative asymptotic homomorphism property if and only if qN (1) G (H) ⊂ K. In the special case H = K, the authors of [3] proved that it is indeed true, but their proof relied heavily on the main result of the article (i.e. Theorem 1.3 in the present notes). We succeeded in providing a self contained proof in the context of group algebras in [5], and we recall some characterizations in the next theorem (where λ denotes the left regular representation of G on ℓ2(G).) Theorem 2.5. ([5], Theorem 2.1) Let H < K < G be a triple of groups and let B = L(H) ⊂ N = L(K) ⊂ M = L(G) be their associated von Neumann algebras. Then the following conditions are equivalent: (1) There exists a net (hi)i∈I ⊂ H such that, for all x, y ∈ M ⊖ N , one has lim i∈I kEB(xλhi y)k2 = 0, the net of unitaries in the relative weak asymptotic homomorphism i.e. property may be chosen in the subgroup λ(H) of U (B). (2) The triple B ⊂ N ⊂ M has the relative weak asymptotic homomorphism property. (3) The subspace of H-fixed vectors ℓ2(G/H)H in the quasi-regular representa- tion mod H is contained in ℓ2(K/H). (4) The one-sided quasi-normalizer qN (1) (5) The triple H < K < G satisfies condition (SS), i.e. for every non empty G (H) is contained in K. finite set F ⊂ G \ K, there exists h ∈ H such that F hF ∩ H = ∅. Remark 2.6. (1) Let G be a group and let H be a subgroup of G. Then it is interesting to note the following description of the one-sided quasi-normalizer of H in G; it is certainly known to specialists: qN (1) G (H) = {g ∈ G : [H : H ∩ gHg−1] < ∞}. Indeed, let g ∈ G be arbitrary; let u1, . . . , un, . . . ∈ H be such that H = G j uj(H ∩ gHg−1). EXAMPLES OF MIXING SUBALGEBRAS OF VON NEUMANN ALGEBRAS 7 Then it is easy to see that HgHg−1 = G j ujgHg−1. It implies that HgH = Fj ujgH, hence that [H : H ∩ gHg−1] < ∞ if and only if g ∈ qN (1) (2) Let H < K < G be a triple of groups. Condition (SS) is equivalent to the following apparently weaker condition (wSS): G (H). For every nonempty finite set F ⊂ G \ K and for every g ∈ G \ K, there exists h ∈ H such that F hg ∩ H = ∅. Indeed, it is easy to see that condition (wSS) is a reformulation of condition (4) of Theorem 2.5, namely that qN (1) G (H) ⊂ K. We use Theorem 2.5 to give examples of triples of algebras B ⊂ N ⊂ M with N = NM (B)′′ in the case of group von Neumann algebras and crossed products. Theorem 2.7. Let H < K < G be a triple of groups such that K = NG(H) is the normalizer of H in G. We denote by W ∗(qN (1) L(G)(L(H))) the von Neumann algebra generated by the one-sided quasi-normalizer of L(H) in L(G). Then the following two conditions are equivalent: (1) L(K) = W ∗(qN (1) (2) H < K < G satisfies condition (SS). L(G)(L(H))); Thus, if the triple H < K < G satisfies condition (SS), then NL(G)(L(H))′′ = L(K), and, moreover, if G acts on some finite von Neumann algebra Q, then Q ⋊ K = NQ⋊G(Q ⋊ H)′′. Proof. (1) ⇒ (2). By hypothesis, the triple of algebras L(H) ⊂ L(K) ⊂ L(G) has the relative weak asymptotic homomorphism property, hence, by Theorem 2.5, H < K < G satisfies condition (SS). (2) ⇒ (1). By assumption, one has the following chain of inclusions and equalities: K = NG(H) ⊂ qN (1) G (H) ⊂ K where the first inclusion follows from the definition and the second one from con- dition (SS). At the level of von Neumann algebras, this gives: L(K) = L(NG(H)) ⊂ NL(G)(L(H))′′ ⊂ W ∗(qN (1) L(G)(L(H))) ⊂ L(K) where the last inclusion follows from the relative weak asymptotic homomorphism property. For the case of crossed products, the unitary groups U (Q) and K are contained in NQ⋊G(Q ⋊ H), hence Q ⋊ K ⊂ NQ⋊G(Q ⋊ H)′′. As the latter algebra is trivially contained in W ∗(qN (1) Q⋊G(Q⋊H)), the assertion follows from Proposition 2.3 above, and from Theorem 3.1 of [3]. (cid:3) Remark 2.8. (1) Let H < G be a pair of groups; one may ask whether it is possible to have NL(G)(L(H))′′ 6= W ∗(qN (1) L(G)(L(H))). It is indeed the case, as explained below. Following Section 5 of [3], we denote by H1 the quasi-normalizer of H in G, G (H)−1, and we let H2 denote the i.e., the maximal subgroup of qN (1) G (H) ∩ qN (1) 8 PAUL JOLISSAINT subgroup of G generated by qN (1) G (H). Finally, let qNL(G)(L(H))′′ be the quasi- normalizer algebra of L(H) in L(G), i.e., the two-sided version of the von Neumann subalgebra W ∗(qN (1) L(G)(L(H))). Then Corollary 5.2 of [3] shows that and that qNL(G)(L(H))′′ = L(H1) W ∗(qN (1) L(G)(L(H))) = L(H2). Example 5.3 of the same article shows that it may happen that H1 6= H2, hence that it is possible to have a triple H < K = H2 < G that satisfies condition (SS) (by Corollaries 4.1 and 5.2 of [3]) but with NL(G)(L(H))′′ $ W ∗(qN (1) L(G)(L(H))). (2) Let H < G be a pair of groups as above, H being abelian. Corollary 5.7 of [3] shows that if, furthermore, L(H) is a MASA in L(G), then NL(G)(L(H))′′ = L(NG(H)). As will be seen in the next section, the last equality can fail if we do not assume that L(H) is a MASA in L(G). In fact, the family of examples of triples of groups in the next section allows us to propose examples as well as counterexamples to the above equality, namely, we exhibit triples H < K < G such that K = NG(H) and L(K) = NG(L(H))′′ on the one hand, and, on the other hand, we will see that there are triples H < K < G with K = NG(H) but L(K) $ NL(G)(L(H))′′. We end the present section with a result that applies in the framework of con- dition (ST), but its relationship with strong mixing is still unclear. It is a gen- eralization of Corollary 2.6 of [8]. In order to state it, we recall the definition of commuting squares. Let M be a finite von Neumann algebra endowed with a normal, finite, faithful, normalized trace τ and let B0 and B1 be von Neumann subalgebras of M with the same unit. The diagram B0 ∪ ⊂ M ∪ B0 ∩ B1 ⊂ B1 is a commuting square if EB0∩B1(b0b1) = EB0∩B1 (b0)EB0∩B1 (b1) for all bj ∈ Bj, j = 0, 1, or, equivalently, if EB0EB1 = EB1EB0 = EB0∩B1. See Chapter 4 in [4] for other equivalent conditions and further details. As is well known, if G1 and G2 are subgroups of G, the system of inclusions L(G1) ∪ ⊂ L(G) ∪ L(G1 ∩ G2) ⊂ L(G2) is a commuting square. Thus, for every subgroup H of G and for every g ∈ G, the following diagram is a commuting square L(gHg−1) ⊂ L(G) ∪ ∪ L(gHg−1 ∩ H) ⊂ L(H). EXAMPLES OF MIXING SUBALGEBRAS OF VON NEUMANN ALGEBRAS 9 In particular, if H < K < G is a triple that satisfies condition (ST), if g ∈ G \ K, the commuting square L(gHg−1) ⊂ L(G) ∪ ∪ L(gHg−1 ∩ H) ⊂ L(H) satisfies the hypotheses of the following proposition since L(gHg−1 ∩ H) is finite- dimensional, hence atomic for every g ∈ G \ K. Thus, every such g is orthogonal to NL(G)(L(H))′′. Proposition 2.9. Let M be a finite von Neumann algebra endowed with some nor- mal, faithful, finite, normalized trace τ , let 1 ∈ B ⊂ M be a diffuse von Neumann subalgebra of M and let u ∈ U (M ) be such that uBu∗ ⊂ M ∪ uBu∗ ∩ B ⊂ B ∪ is a commuting square, and assume that uBu∗ ∩ B has the following properties: its center is atomic and its relative commutant (uBu∗ ∩ B)′ ∩ M is diffuse (the latter conditions are automatically satisfied if uBu∗ ∩ B itself is atomic). Then u is orthogonal to NM (B)′′. Proof. Let us first fix v ∈ NM (B) and let us prove that τ (vu) = 0. As vBv∗ = B, the diagram vuBu∗v∗ ⊂ M ∪ vuBu∗v∗ ∩ B ⊂ B ∪ is a commuting square as well, C := vuBu∗v∗ ∩ B = v(uBu∗ ∩ B)v∗ has atomic center and its relative commutant C ′ ∩ M is still diffuse. Moreover, recall that one has EC (xb) = EC (bx) for all b ∈ C ′ ∩ B, x ∈ M , and that EC (b) ∈ Z(C). If (zj)j≥1 is the set of minimal projections of the center Z(C), then each reduced algebra Czj is a finite subfactor of the reduced von Neumann algebra zjM zj, and its relative commutant is the diffuse algebra zj(C ′ ∩ B)zj = (Czj)′ ∩ zjBzj. If τj is the normalized trace on zjM zj defined by τj(zjxzj) = 1 τ (zj) τ (zj xzj) for all x ∈ M , then the associated conditional expectation ECzj satisfies the following identity: ECzj (zjxzj) = EC (zjxzj )zj for all x ∈ M . In particular, one has ECzj (y) = τj (y)zj for every y ∈ zj(C ′ ∩ B)zj. Let us fix ε > 0. We claim that there exists a partition of unity (ei)1≤i≤n in C ′∩B so that kEC(ei)k ≤ ε for every i. Indeed, choose a positive integer m such that 2−m ≤ ε and then, for every j, a partition of the unity (ej,i)1≤i≤2m ⊂ zj(C ′ ∩ B)zj such that τj(ej,i) = 2−m for all i. Finally, set n = 2m and ei = Pj ej,i. Then (ei)1≤i≤n is a partition of the unity in C ′ ∩ B and, by the above considerations, EC (ei) = X j τj(ej,i)zj = 2−m ≤ ε ∀i. Let D be the abelian von Neumann algebra generated by the projections (ei). We recall that ED′∩M (x) = X i eixei ∀x ∈ M, 10 PAUL JOLISSAINT and we will make use of the following identity EC (vueiu∗v∗ei) = EC (vueiu∗v∗)EC (ei) which is true since vueiu∗v∗ ∈ vuBu∗v∗, ei ∈ B and by the commuting square condition. One has: τ (vu)2 = τ (ED′ ∩M (vu))2 ≤ kED′∩M (vu)k2 2 = τ (ED′ ∩M (vu)2) = τ (EC (ED′∩M (vu)2)). But, EC (ED′ ∩M (vu)2) = EC   = EC (X (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) X i 2  eivuei(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) eivueieju∗v∗ej) i,j EC (eivueiu∗v∗ei) EC (vueiu∗v∗)EC (ei) EC (vueiu∗v∗)1/2EC (ei)EC (vueiu∗v∗)1/2 = X i = X = X i i ≤ εX i EC (vueiu∗v∗) = εEC (vuu∗v∗) = ε. Thus, one has τ (vu)2 ≤ ε. By weak density, we get τ (xu) = 0 for every x ∈ NM (B)′′. (cid:3) Remark 2.10. Let 1 ∈ C ⊂ M be a pair of finite von Neumann algebras. As we have seen in the proof above, the existence, for every ε > 0 of partitions of the unity (ei)1≤i≤n in C ′ ∩ M such that kEC (ei)k ≤ ε for all i, is implied by two conditions: (i) the center Z(C) is atomic, and (ii) the relative commutant C ′ ∩ M is diffuse. In fact, the existence of such partitions of the unity requires both conditions. Indeed, suppose for instance that M is a type II1 factor; firstly, if C is a MASA in M , then its center and its relative commutant are equal and diffuse, and EC (e) = e for every nonzero projection e ∈ C ′ ∩ M = C, thus kEC (e)k = 1 for all such projections. Secondly, if C is a subfactor of finite index in M , then its center is atomic, but its relative commutant is finite dimensional. Hence there exists a constant c > 0 such that for every nonzero projection e ∈ C ′ ∩ M . kEC (e)k = τ (e) ≥ c 3. Families of examples As indicated above, we are going to give examples of triples of groups which satisfy conditions (SS) or (ST). They will be based on semidirect products. Thus, let K be a group, let H < K be a subgroup of K and assume that K acts on some group A through an action α. Put G = A ⋊ K, and identify K with the subgroup {e} × K of G. For future use, we put A∗ = A \ {e}. A case that might be of interest comes from generalized wreath product groups as in the next example. EXAMPLES OF MIXING SUBALGEBRAS OF VON NEUMANN ALGEBRAS 11 Example 3.1. Let K be an arbitrary countable group. Assume that K acts on some countable set X, and take any nontrivial group Z. Let A = Z (X) be the group of all maps a : X → Z such that a(x) = e except for some finite subset of X. Then K acts by left translation on A, and the corresponding group G is the generalized wreath product group Z ≀X K. We present first a condition which implies that K is the normalizer of H in G. Proposition 3.2. Let H < K < G = A ⋊ K be a triple as above. Assume furthermore that H is a normal subgroup of K and that e ∈ A is the only element a ∈ A such that αh(a) = a for all h ∈ H. Then NG(H) = K. Proof. One has obviously K ⊂ NG(H). Conversely, if g = (a, k) ∈ NG(H), then one has for every h ∈ H: ghg−1 = (a, k)(e, h)(αk−1 (a−1), k−1) = (a, kh)(αk−1 (a−1), k−1) = (aαkhk−1 (a−1), khk−1) which belongs to H if and only if αkhk−1 (a) = a for every h ∈ H, if and only if αh(a) = a for every h ∈ H. Hence a = e. (cid:3) Let us see now under which conditions the triple H < K < G = A ⋊ K satis- fies condition (SS) or (ST). It is partly inspired by Theorem 2.2 of [9]. See also Proposition 4.6 of [6]. Theorem 3.3. Let H < K < G = A ⋊ K be a triple as above. (a) The triple H < K < G satisfies condition (ST) if and only if, for every a ∈ A∗, {h ∈ H : αh(a) = a} < ∞. (b) The triple H < K < G satisfies condition (SS) if and only if, for every finite set E ⊂ A∗, there exists h ∈ H such that E ∩ αh(E) = ∅. Equivalently, for every a ∈ A∗, the H-orbit H · a is infinite. Proof. (a) (⇒) Let a ∈ A∗. Then (a, e) ∈ G \ K, and H ∩ (a, e)H(a−1, e) is finite. But, if h ∈ H, one has (a, e)(e, h)(a−1, e) = (aαh(a−1), h)) ∈ H if and only if αh(a) = a. The set of such elements h ∈ H must be finite. (⇐) Let g ∈ G \ K; let us prove that H ∩ gHg−1 is finite. Put g = (a, k). If h ∈ H, one has ghg−1 = (a, k)(e, h)(αk−1 (a−1), k−1) = (a, kh)(αk−1 (a−1), k−1) = (aαkhk−1 (a−1), khk−1). Thus, ghg−1 ∈ H if and only if khk−1 ∈ H and αkhk−1 (a) = a. Hence H ∩ gHg−1 is finite. (b) (⇒) Let E ⊂ A∗ be finite. Replacing E by E ∪E−1 if necessary, we assume that E = E−1. Then F := {(a, e) : a ∈ E} ⊂ G \ K and there exists h ∈ H such that F (e, h)F ∩ H = ∅. In particular, (e, h)(a, e) = (αh(a), h) /∈ F H for every a ∈ E and E ∩ αh(E) = ∅. (⇐) Let F ⊂ G \ K be finite. There exist F1 ⊂ A∗ and F2 ⊂ K finite such that F ⊂ F1 × F2. Observe that, for (aj, kj ) ∈ F1 × F2, j = 1, 2, and for h ∈ H, one has (a1, k1)(e, h)(a2, k2) = (a1, k1h)(a2, k2) = (a1αk1h(a2), k1hk2) 12 PAUL JOLISSAINT hence, αh(a2) = αk−1 such that E ∩ αh(E) = ∅, then necessarily, F (e, h)F ∩ H = ∅. if it belongs to H, one must have a1αk1 (αh(a2)) = 1, or equivalently, 1 ). Taking E = F1 ∪ {αk−1 (a−1) : k ∈ F2, a ∈ F1}, if h ∈ H is (cid:3) (a−1 1 Remark 3.4. Let us consider the case where A = Z (X) as in Example 3.1: K acts on the countable set X. Then {h ∈ H : αh(a) = a} < ∞ for every a ∈ A∗ if and only if, for every x ∈ X, the stabilizer Hx = {h ∈ H : h · x = x} is finite. By Proposition 2.3 of [7], it is equivalent to the fact that the action of H by generalized Bernoulli shifts on (Y X , νX ) is mixing, where (Y, ν) is some standard probability space. In the same vein, for every nonempty, finite set E ⊂ A∗ one can find h ∈ H such that E ∩ αh(E) = ∅ if and only if the action of H on X has infinite orbits. By Proposition 2.1 of [7], it is equivalent to the fact that the corresponding Bernoulli shift action of H on (Y X , νX ) is weakly mixing. The following corollary is a straightforward consequence of Theorem 2.7, Propo- sition 3.2 and of Theorem 3.3. Corollary 3.5. Let H, K and A be as above, denote by H < K < G the associated triple of groups and assume that: (1) for every a ∈ A∗, there exists h ∈ H such that αh(a) 6= a; (2) for every finite set E ⊂ A∗, there exists h ∈ H such that E ∩ αh(E) = ∅. Then L(K) = NL(G)(L(H))′′. More generally, if G acts on some finite von Neu- mann algebra (Q, τ ), then Q ⋊ K = NQ⋊G(Q ⋊ H)′′. As promised in the preceeding section, we give examples of triples H < K < G such that NG(H) = K but L(K) $ NL(G)(L(H))′′. Example 3.6. Let H ⊳ K be infinite, countable groups, let A be an infinite, count- able group endowed with an action of K, and assume that: (i) for every a ∈ A∗, there exists h ∈ H such that αh(a) 6= a; (ii) there exists a0 ∈ A∗ whose orbit H · a0 is finite. Let H < K < G be the associated triple of groups. Then the first condition implies that NG(H) = K by Proposition 3.2, and the second one implies that the triple does not satisfy condition (SS) by Theorem 3.3. Let us prove that L(K) $ NL(G)(L(H))′′. Put F = {(αh(a0), e) : h ∈ H} which is a finite set by the second condition above. Define x = X g∈F ∪F −1 λg. Then x is a selfadjoint element of L(G) ⊖ L(K) since F ∩ K = ∅. Furthermore, it is easy to see that x ∈ L(H)′ ∩ L(G). As x /∈ L(K), there exists a spectral projection e of x which does not belong to L(K) either. Then u := 2e − 1 is a unitary element of L(H)′ ∩ L(G) hence it belongs to the normalizer NL(G)(L(H)), but u /∈ L(K). Thus L(K) $ NL(G)(L(H))′′. Observe that H can be an abelian group, and this shows that, in order to have the equality NL(G)(L(H))′′ = L(NG(H)) in Corollary 5.7 of [3], one must assume that L(H) is a MASA in L(G). 1. I. Chifan, On the normalizing algebra of a masa in a II1 factor, arXiv: math. OA/0606225, 2006. References EXAMPLES OF MIXING SUBALGEBRAS OF VON NEUMANN ALGEBRAS 13 2. K. J. Dykema, A. M. Sinclair, and R. R. Smith, Values of the Pukansky invariant in free group factors and the hyperfinite factor, J. Funct. Anal. 240 (2006), 373 -- 398. 3. J. Fang, M. Gao, and R.R. Smith, The relative weak asymptotic homomorphism property for inclusions of finite von Neumann algebras., Internat. J. Math. 22 (2011), 991 -- 1011. 4. F.M. Goodman, P. de la Harpe, and V.F.R Jones, Coxeter graphs and towers of algebras, Springer-Verlag, New York, 1989. 5. P. Jolissaint, On the relative weak homomorphism property for triples of group von Neumann algebras, To appear in Proc. Amer. Math. Soc., 2011. 6. P. Jolissaint and Y. Stalder, Strongly singular MASAs and mixing actions in finite von Neu- mann algebras, Ergod. Th. & Dynam. Sys. 28 (2008), 1861 -- 1878. 7. A.S. Kechris and T. Tsankov, Amenable actions and almost invariant sets, Proc. Amer. Math. Soc. 136 (2008), 687 -- 697. 8. S. Popa, Orthogonal pairs of ∗-subalgebras in finite von Neumann algebras, J. Operator The- ory 9 (1983), 253 -- 268. 9. G. Robertson, A. M. Sinclair, and R. R. Smith, Strong singularity for subalgebras of finite factors, Int. J. Math. 14 (2003), 235 -- 258. 10. A. M. Sinclair, R. R. Smith, S. A. White, and A. Wiggins, Strong singularity of singular masas in II1 factors, Illinois J. Math. 51 (2007), 1077 -- 1084. 11. S. Vaes, One-cohomology and the uniqueness of the group measure space decomposition of a II1 factor, ArXiv:math.OA/1012.5377 v1, 2010. Universit´e de Neuchatel, Institut de Math´emathiques, Emile-Argand 11, CH-2000 Neuchatel, Switzerland E-mail address: [email protected]
1301.0640	2	1301	2013-04-02T19:35:34	Further remarks on an order for quantum observables	[ "math.OA" ]	S. Gudder and, later, S. Pulmanova and E. Vincekova, have studied in two recent papers a certain ordering of bounded self-adjoint operators on a Hilbert space. We present some further results on this ordering and show that some structure theorems of the ordered set of operators can be obtained in a more abstract setting of posets having the upper bound property and equipped with a certain orthogonality relation.	math.OA	math	FURTHER REMARKS ON AN ORDER FOR QUANTUM OBSERVABLES J ¯ANIS C¯IRULIS Abstract. S. Gudder and, later, S. Pulmanov´a and E. Vincekov´a, have stud- ied in two recent papers a certain ordering of bounded self-adjoint operators on a Hilbert space. We present some further results on this ordering and show that some structure theorems of the ordered set of operators can be obtained in a more abstract setting of posets having the upper bound property and equipped with a certain orthogonality relation. 1. Introduction In [13], S. Gudder introduced a certain order for quantum observables, in fact, on the set S(H) of bounded self-adjoint operators on a complex Hilbert space H, and suggested to call it the logical order. He demonstrated, in particular, that S(H) is a generalized orthoalgebra; the logical order is by definition the natural order of this algebra. He also showed that this is not a lattice order, but nevertheless every pair of observables having a common upper bound has a join and a meet with respect to this order. Such a poset was called by him a near-lattice (but see the beginning of the next section). Actually, every initial segment of S(H) is even a σ-orthomodular lattice. Also the commutative case (in which observables are represented by random variables on a probability space) was considered in [13]; as noted by the author of that paper, this case actually served as motivation and a source of intuition for results and proofs of the general case. The properties of the new ordering were studied in more detail by S. Pulmannov´a and E. Vincekov´a in [19], where several results of [13] were essentially improved. These authors observed that the logical order is actually a restriction of Drazin's order (nowdays usually called star order or *-order) introduced by him in [10] for all bounded operators on H. They proved, for example, that S(H) is even a weak generalized orthomodular poset. Moreover, this poset is bounded complete, i.e., every subset bounded from above has a join and, correspondingly, every nonempty subset has a meet. More recently, existence conditions of joins and meets in S(H) (under the logical ordering), and representations of these operations have been discussed, e.g., in [12, 15, 16, 20, 21]. We present here some further results on the order structure of S(H), and also fill two small gaps in proofs in [13]. In particular, we obtain explicit descriptions for the Gudder join and meet operations in terms of operator composition and 1991 Mathematics Subject Classification. Primary 81Q10; Secondary 06A12, 46L10, 47L30, 81P10. Key words and phrases. bounded self-adjoint operator, generalized orthoalgebra, nearsemilat- tice, orthogonality, orthomodular lattice, quasi-orthomodular, quantum observable, skew meet. This work was supported by ESF project No. 2009/0216/1DP1.1.1.2.0/09/APIA/VIAA/044. 1 2 J ¯ANIS C¯IRULIS lattice operations on projectors, and a simple proof of bounded completeness of S(H). On the other hand, we discuss some of the properties of the poset S(H) in a more abstract setting, and show that the aforementioned properties of the logical order of observables are not quite independent. For instance, any poset having the least element and possessing the so called upper bound property (any pair of elements has a join if they have a common upper bound), if equipped with an appropriate orthogonality relation, carries a structure of a generalized orthoalgebra, in which every initial segment is an orthomodular lattice. Moreover, there is a non- commutative total binary operation (called skew meet ) on S(H) which, considered together with the partial join operation, turns the poset into a so called skew nearlattice. This allows, on the one hand, to establish a link between structures arising in quantum logic and some branches of the theory of information systems (where the notion of skew nearlattice has emerged; see [5]), and on the other hand, to apply to the algebra S(H) certain general decomposition and structure theorems from [6]. We do not address, however, these questions in the present paper. The structure of the paper is as follows. The subsequent section contains the necessary background on posets having the upper bound property, known also as nearsemilattices in algebra and as a (simple version of) domains in database theory. Three principal examples of such posets, including S(H), are considered in Section 3. Some order properties of S(H) are discussed also in Section 4. Section 5 deals with so called quasi-orthomodular nearsemilattices, which mimic, in a sense, the generalized orthomodular lattices of [14], and Section 6, with skew nearlattices. 2. Preliminaries: nearsemilattices A nearlattice is usually defined as a meet semilattice having the upper bound property. Therefore, every bounded complete poset is an example of a nearlattice. Equivalently, a nearlattice is a meet semilattice in which every initial segment is a lattice. Since early eighties, such structures have been intensively studied by W. Cornish and his collaborators; see, e.g., [8, 9, 18]. Some authors prefer order duals of such algebras [2]. Arbitrary posets having the upper bound property were named (upper) near- semilattices in [3]. Thus, a nearlattice may be viewed also as a nearsemilattice that happens to be a meet semilattice. Near-lattices mentioned in [13] (see Introduction) is a weaker concept. We shall always assume that a near(semi)lattice has the least element 0, and consider such structures as partial algebras of kind (A, ∨, 0), resp., (A, ∧, ∨, 0), where ∧ is, as usual, the meet operation and ∨ is the partial join operation. The following axiomatic description of nearsemilattices goes back to [3, Section 1] (for arbitrary near(semi)lattice terms s and t, we write s ◦ t to mean that s∨t is defined under that or other intended assignment of values to variables occurring in these terms). Proposition 1. An algebra (A, ∨, 0), where 0 is a nullary operation and ∨ is a partial binary operation, is a nearsemilattice if and only if it fulfils the conditions (∨1) : x ◦ x and x ∨ x = x, (∨2) : (∨3) : (∨4) : x ◦ 0 and x ∨ 0 = x. if x ◦ y, then y ◦ x and x ∨ y = y ∨ x, if x ◦ y and x ∨ y ◦ z, then y ◦ z, x ◦ y ∨ z and (x ∨ y) ∨ z = x ∨ (y ∨ z), FURTHER REMARKS ON AN ORDER FOR QUANTUM OBSERVABLES 3 The order relation on a nearsemilattice A is recovered from the partial join operation as follows: x ≤ y if and only if x ◦ y and x ∨ y = y, (1) and then x ∨ y (when defined) is the join of x and y w.r.t. this order, while 0 is the least element. Corollary 2. An algebra (A, ∧, ∨, 0) is a nearlattice if and only if (A, ∨, 0) is a nearsemilattice, (A, ∧) is a semilattice, and the following absorption laws are fulfilled: if x ◦ y, then x ∧ (x ∨ y) = x, (∨5) : (∨6) : x ∧ y ◦ y and (x ∧ y) ∨ y = y. Observe that these laws can also be rewritten in a ∨-free form, where ≤ is the ordering (1): if x ≤ y, then x ∧ y = x, and x ∧ y ≤ y. (2) A nearsemilattice A is said to be distributive [7], if every initial segment of it is a distributive semilattice: (∨7) : if y ◦ z and x ≤ y ∨ z, then x = y′ ∨ z′ for some y′ ≤ y and z′ ≤ z. In the case when A happens to be a nearlattice, this reduces to the standard notion of a distributive nearlattice (every initial segment is a distributive lattice). A De Morgan complementation, or just m-complementation, on a poset is a unary operation − such that x−− = x, and if x ≤ y, then y− ≤ x−. By a sectionally m-complemented poset we mean a poset with the least element, in which every initial segment [0, p] is equipped with an m-complementation − p [7, Section 2]. In such a poset, a partial subtraction operation ⊖ may be defined by x ⊖ y = z if and only if y ≤ x and z = y− x in [0, x]. Lemma 2.5 of [7], together with Proposition 2.2, Theorem 2.3 and Corollary 2.4 of that paper, leads us to the following characterization of certain nearlattices. Proposition 3. A sectionally m-complemented poset is a nearlattice if and only if the partial subtraction ⊖ on it can be extended to a total operation − satisfying conditions (−1) : (−2) : x − (x − y) ≤ y, (−3) : x − 0 = x. if x ≤ y, then z − y ≤ z − x, Namely, in an m-complemented nearlattice, x − y = x ⊖ (x ∧ y) (3) (see equation (7) in [7]). Explicit definitions of join and meet in terms of subtraction are not given in [7]. We omit the tedious calculations and note without proof that x ∧ y = x − (x − y), and x ∨ y = z − ((z − x) ∧ (z − y)) whenever x, y ≤ z. An ordered algebra (A, −, 0), where 0 is the least element and − is a binary opera- tion satisfying (−1) -- (−3), is a weak BCK-algebra in the sense of [7]; see Proposition 2.2 therein. 4 J ¯ANIS C¯IRULIS 3. Examples Let as now consider three examples of nearlattices. The first example may be considered as motivating the concept; the other two are borrowed from [13, 19]. Example 1 (partial functions). Let I and V be nonempty sets, and letPF (I, V ) be the set of all partial functions from I to V . It is partially ordered by set inclusion, and (PF (I, V ), ∩, ∪, λ), where λ is the nonwhere defined function, is a nearlattice with respect this order and usual operations ∩ and ∪. Actually, PF(I, V ) is even bounded complete, and total functions are just its maximal elements. Observe that the union of two functions in PF (I, V ) is defined if and only if they agree on the common part of their domains. It is easily seen that dom(ϕ ∪ ψ) = dom ϕ ∪ dom ψ and dom(ϕ ∩ ψ) ⊆ dom ϕ ∩ dom ψ. Any nearlattice (Φ, ∪, ∩, λ), where Φ is a subset of some PF (I, V ), is called a functional nearlattice. If the nearlattice contains all pairwise unions (of elements of Φ) existing in PF(I, V ), it is said to be closed. Every initial segment of a functional nearlattice Φ is even a Boolean lattice. Then the corresponding BCK-subtraction (3) coincides in Φ with set subtraction. Example 2 (random variables). Let (Ω, A, µ) be a probability space, and let M(A) be the set of random variables on this space (i.e., measurable functions Ω → R). Put supp f := {ω ∈ Ω : f (ω) 6= 0}. For f, g ∈ M(A), write f ⊥ g if f g = 0 (i.e., supp f ∩ supp g = ∅), and put f (cid:22) g if there is a function h ∈ M(A) such that f ⊥ h and f + h = g. The relation (cid:22) is an order on M(A). By [13, Theorem 3.1], f (cid:22) g iff f (ω) = g(ω) whenever ω ∈ supp f ; equivalently, iff f = gχsupp f (χK stands for the characteristic function of a subset K ⊆ Ω). Direct calculations show that it is a nearlattice ordering with the zero function 0 as its least element (Theorem 3.5 in [13]). The corresponding meet and join operations, f and g, may be defined as follows: f f g := gχ{ω∈(supp f ∩supp g) : f (ω)=g(ω)} = gχ{supp f ∩supp grsupp(f −g)} and, if f, g (cid:22) h, f g g := hχ(supp f ∪supp g). By the way, f f g = hχ(supp f ∩supp g) in this case. (We have changed the notation of [13], where these operations were denoted by ∧ and ∨, respectively.) Similarly, if F is a subset of M(A) with an upper bound h, then hχ(S{supp f : f ∈F }) is the least upper bound of F : M(A) is bounded complete. Observe that this example is essentially subsumed under the previous one: the nearlattice M(A) may be identified with the closed subalgebra MA of PF (Ω, R0) (where R0 := R r {0}) obtained by replacing every function in M(A) with its codomain restriction to R0; this transfer is an isomorphism of M(A) into PF(Ω, R0). In particular, every initial segment [0, g] in M(A) is a Boolean lattice, where the complementation of f is g − f . Then the BCK-subtraction in M(A) (which we denote by −. ) is given by g −. f = g − (f f g) = gχsupp(f −g). Example 3 (quantum observables). We return to the set S(H) of bounded self- adjoint operators on a Hilbert space. The ensuing conventions follow to [13]. Let P(H) stand for the set of all projections (idempotent operators). If A ∈ S(H), denote by ran A the range of A and by ran A, its closure. Denote the projection FURTHER REMARKS ON AN ORDER FOR QUANTUM OBSERVABLES 5 onto ran A by PA. For A, B ∈ S(H), put A ⊥ B if the composition AB is the zero operator O (equivalently, if ran A ⊥ ran B, or PAPB = O; see Lemma 4.1 in [13]), and put A (cid:22) B if B = A + C for some C ∈ S(H) with C ⊥ A. By [13, Lemma 4.3], A (cid:22) B if and only if Ax = Bx whenever x ∈ ran A (then ran A ⊆ ran B), or, equivalently, if A = BPA (then B and PA commute). The relation (cid:22) is the logical order on S(H) mentioned in Introduction. We shall say that the closed subspace ran A of H corresponds to A. Recall that the transfer PA 7→ ran A from projection operators to closed subspaces is one-to-one and onto; moreover, P(H) is ordered by P1 ≤ P2 if and only if P1P2 = P1 if and only if P1 = P2P1, and then PA ≤ PB if and only if ran A ⊆ ran B. In particular, P(H) is lattice ordered, and PA ≤ PB whenever A (cid:22) B. Corollary 4.4 in [13] implies that PA ≤ PB if and only if PA (cid:22) PB. We denote the join and meet operations in the lattice P(H) by ∨ and ∧, and these operations in S(H) (which may be partial), by g and f, respectively (in [13, 19], the same symbols ∨ and ∧ are used for both purposes). According to [13, Corollary 4.13], meets and joins exist in S(H) for every pair of elements bounded above (but see the beginning of the next section). Using this result as the base, it is shown in [19, Corollary 4.7] that S(H) is even bounded complete, in particular, a nearlattice (evidently, O is the least element in it). It is easily seen that, in every interval [O, B] of S(H), B − A is an m-complement of A. If the symbol −. stands for the BCK-subtraction (3) in the nearlattice S(H), then (for arbitrary A, B ∈ S(H)) B −. A = B − (A f B) = B − BPAfB = B(I − PAfB). This example is not fully subsumed under Example 1. The definition of (cid:22) shows that A = B if and only if A ran A = B ran B. Therefore, an operator in S(H) is completely determined by its restriction to the corresponding subspace of H. We thus can identify every operator A with the partial function A ran A from PF(H, H), and come in this way to a nearlattice of functions SH (ordered by set inclusion) isomorphic to S(H). However, as the join of two subspaces generally differs from their set-theoretical union, this nearlattice need not be functional. The set SH of partial operators admits also an immediate description. Theorem 4.12 of [13] introduces, for any B ∈ S(H), a subset LB := {P : P ≤ PB and BP = P B} of P(H). For example, if A (cid:22) B, then PA ∈ LB. It is shown in the proof of the theorem that the mapping φ : A 7→ PA is an order isomorphism of the initial segment [O, B] of S(H) onto LB. Therefore, LB = {P : P = PA and A (cid:22) B for some A ∈ S(H)} = {PA : A (cid:22) B}. It is now easily seen that SH = {B ran C : PC ∈ LB}. Indeed, B ran C belongs to SH , by definition, if and only if B ran C = A ran A for some A ∈ S(H), i.e. if and only if ran C = ran A and B ran A = A ran A, i.e., if and only if PC = PA and A (cid:22) B. It is also demonstrated in the proof of the mentioned theorem that if P ≤ PB, then P = PBP and, further, BP (cid:22) B (4) 6 J ¯ANIS C¯IRULIS whenever BP ∈ S(H) i.e., whenever BP = P B. It follows that the inverse of the isomorphism φ takes a projection P from LB into the operator BP ∈ [O, B]. Therefore, [O, B] = {BP : P ∈ LB}. 4. More on the ordering of S(H) In this section we discuss in more detail existence of joins and meets in the poset S(H). It is noticed in [13, Theorem 4.12] that every poset LB is a lattice with respect to ≤. According to this theorem, the initial segment [O, B] of S(H) is order isomorphic to LB; thus, the segment itself is a lattice. Corollary 4.13 to this theorem then asserts that, in S(H), every pair of elements bounded above has a meet and a join. Of course, the meet of two elements in an initial segment of a poset is also their meet in the poset itself (and conversely). However, this may be not the case with joins; this fact seems to be overlooked in [13] when drawing the corollary. The subsequent theorem confirms that the corollary is nevertheless correct, and provides explicit descriptions of the corresponding partial operations (Corollary 6). Recall that the commutant of an operator B (i.e., the set of all bounded operators commuting with B) is a von Neumann algebra. The lattice LB, being an initial segment in the complete orthomodular lattice of all projections of this algebra, is therefore complete. This observation allows us to obtain a simple direct proof of existence of meets and joins in S(H) for all bounded sets of operators. Theorem 4. Suppose that T ⊆ S(H), T 6= ∅, and that B ∈ S(H) is an upper bound of T. Then (a) B(V(PA : A ∈ T )) is the greatest lower bound of T, (b) B(W(PA : A ∈ T )) is the least upper bound of T. Proof. Assume that A (cid:22) B, i.e., A = BPA, for all A ∈ T. Then all projections PA belong to the complete lattice LB. This implies that the symbolic expressions in (a) and (b) present operators from S(H). (a) Let PT stand for the meet (in P(H)) of all projections PA with A ∈ T . For every A ∈ T, we have BPT = BPAPT = APT = APBPT (by (4), as PT ≤ PB), i.e., BPT (cid:22) A. Hence, BPT is a lower bound of T. Further, if D is one more lower bound, then D (cid:22) B, PD ≤ PT (as PD ≤ PA for all A ∈ T ), and BPT PD = BPD = D. Thus, D (cid:22) BPT . Therefore, BPT is the greatest lower bound of T. (b) Let P T stand for the join (in P(H)) of all projections PA with A ∈ T . For every A ∈ T, we have A = BPA = BP T PA, i.e., A (cid:22) BP T . Hence, BP T is an upper bound of T; moreover, BP T (cid:22) B, for P T belongs to the complete lattice LB containing all PA. Further, if D is one more upper bound, then likewise DPA = A for all A ∈ T and DP T (cid:22) D. Now, (B − D)PA = 0 and, consequently, PB−DPA = 0 for every A ∈ T . Then in P(H) also PA ≤ I − PB−D for every A. Hence, P T ≤ I − PB−D, PB−DP T = 0 and, finally, (B − D)P T = 0. Therefore, BP T = DP T (cid:22) D, and BP T is the least upper bound of T . (cid:3) Notice that sup T = O if T = ∅. Therefore, item (b) of the theorem immediately implies the result of [19] that the poset S(H) is actually bounded complete (see Example 3 above), every initial segment of S(H) is a complete sublattice of S(H), and the embedding P 7→ BP of LB into S(H) mentioned at the end of the previous section preserves all meets and joins. FURTHER REMARKS ON AN ORDER FOR QUANTUM OBSERVABLES 7 As noted in the previous section, the ordering (cid:22) agrees on P(H) with the stan- dard ordering ≤ of projections. In [13], it is implicitly assumed without proof that then the meet (join) of two projection operators in P(H) is also their meet (resp., join) in the more extensive poset S(H). It follows from the theorem that this is indeed the case. Corollary 5. P(H) is a complete sublattice of S(H). Proof. Consider that T ⊆ P(H) (then PA = A for every A ∈ T ) and B = I. (cid:3) We write out the following significant particular case of the above theorem. Corollary 6. If A, B, C ∈ S(H) and A, B (cid:22) C, then A f B and A g B exist, and A f B = C(PA ∧ PB), A g B = C(PA ∨ PB). By help of (4), we also conclude that PAfB = PA ∧ PB, PAgB = PA ∨ PB (5) whenever the pair A, B is bounded. Consequently, ran(A g B) = ran A ⊔ ran B and ran(A f B) = ran A ∩ ran B in this case. Examples 1 and 2 suggest that generally ran(A f B) ⊆ ran A ∩ ran B. The bounded completeness of S(H) implies that item (a) of Theorem 4 may be strengthened: every nonempty subset of S(H) has the greatest lower bound. In particular, the set of all lower bounds of operators A, B ∈ S(H) has the join, which is also its maximum element, i.e., A f B. Observe that {PC : C (cid:22) A and C (cid:22) B} = LAfB, so that PAfB belongs to this set and is the greatest element in it. Therefore, we come to the following description of meet of observables in a general case. Corollary 7. For every pair of observables A and B, their meet exists and A f B = B(max{PC : C (cid:22) A and C (cid:22) B}). This particular result admits also a simple direct proof (cf. Proposition 3.2 in [1]). Let us consider the following subset of P(H): LA,B := {P : P ≤ PA, P ≤ PB, AP = P A, BP = P B and P ⊥ PA−B}. If C (cid:22) A, B, then APC = C = BPC , (A − B)PC = O and PC ⊥ PA−B. Moreover, then also PC ∈ LA, LB. Therefore, PC ∈ LA,B. On the other hand, if P ∈ LA,B, then P ⊥ PA−B and P ⊥ (A − B), i.e., (A − B)P = O and AP = BP =: C. As also P ∈ LA and P ∈ LB, we conclude by (4) that C (cid:22) A, B. Therefore, LA,B = {P : P = PC and C (cid:22) A, B for some C ∈ S(H)} and, finally, LA,B = {PC : C (cid:22) A and C (cid:22) B} ⊆ LA ∩ LB. The definition of LA,B can be rewritten in the form LA,B = {P ∈ {A, B}′ : P ≤ PA ∧ PB ∧ (I − PA−B)}, where {A, B}′ is the commutant of {A, B} and I is the identity operator. Therefore, LA,B is a bounded subset of the complete lattice of all projections of the von Neumann algebra {A, B}′ and, hence, has the join P ∗. Evidently, P ∗ is even the maximum element of LA,B. It is easily seen that then the operator BP ∗ is the g.l.b. of A and B in S(H). Indeed, since P ∗ ∈ LA,B, it follows that P ∗ ∈ LA, LB and P ∗ = PC for some C with C (cid:22) A, B. So, C = AP ∗ = BP ∗. Suppose that D is one more lower bound of A and B. As PD ∈ LA,B, it follows that PD ≤ P ∗ = PC . On 8 J ¯ANIS C¯IRULIS the other hand, PD ∈ LA, LB. Recall that the transfer φ : E 7→ PE (Section 3) is an order isomorphism of [O, B] onto LB; therefore D (cid:22) C, and C = A f B. Of course, AP ∗ also is the g.l.b. of A and B. We already know that actually P ∗ = PAfB. 5. Quasi-orthomodular nearsemilattices In [14], M.F. Janowitz introduced the notion of a generalized orthomodular lat- tice, which was defined as a lattice equipped with an appropriate orthogonality relation. We take up the idea and consider in this section nearsemilattices with orthogonality. A binary relation ⊥ on a poset A with the least element 0 is said to be an orthogonality, if it satisfies the conditions if x ⊥ y, then y ⊥ x, if x ≤ y and y ⊥ z, then x ⊥ z, (⊥1) : (⊥2) : (⊥3) : x ⊥ 0. For example, if − is an m-complementation on A and a relation ⊥ is defined by x ⊥ y iff y ≤ x−, then ⊥ is an orthogonality on A. We say that it is induced by the m-complementation −. Evidently, the induced orthogonality is additive in the sense that (⊥4) : if x ⊥ y, x ⊥ z and y ◦ z, then x ⊥ y ∨ z. Definition 1. Suppose that (A, ∨, 0) is a nearsemilattice and that ⊥ is an or- thogonality on it. The algebraic system (A, ∨, ⊥, 0) is called a quasi-orthomodular nearsemilattice if the following additional conditions are fulfilled: (⊥5) : (⊥6) : (⊥7) : if x ⊥ y, then x ◦ y, if x ≤ y, then y = x ∨ z for some z with x ⊥ z, if x ⊥ y, x ⊥ z and y ≤ x ∨ z, then y ≤ z. In a quasi-orthomodular nearsemilattice, the following cancellation law holds: (⊥8) : if x ⊥ y, x ⊥ z and x ∨ y = x ∨ z, then y = z. In particular, (⊥9) : if x ⊥ x, then x = 0. Also, (⊥4) together with (⊥5) implies that (⊥10) : every finite set of mutually orthogonal elements has a join (recall that in a nearsemilattice x ◦ y iff x and y have a common upper bound). Example 1 (continuation). In a functional nearsemilattice, put ϕ ⊥ ψ if dom ϕ ∩ dom ψ = ∅. Then (PF(I, V ), ∪, ⊥, λ) is a quasi-orthomodular nearsemilattice with an additive orthogonality. The existential condition (⊥6) may fail in an arbitrary closed functional nearsemilattice. The other existential condition (⊥5) is ensured by closedness and may fail in an arbitrary functional nearsemilattice. Example 2 (continuation). The relation ⊥ on M(A) introduced in the Section 3 is an orthogonality, and (M(A), g, ⊥, 0) is a quasi-orthomodular nearsemilattice. The orthogonality satisfies also (⊥4) (Theorem 3.2(b) in [19]). Taking into account that (⊥6) follows directly from the definition of (cid:22), these facts can be derived from the previous example, because the isomorphism of M(A) onto the closed functional nearsemilattice M (A), which was described in Section 3, preserves orthogonality. FURTHER REMARKS ON AN ORDER FOR QUANTUM OBSERVABLES 9 Example 3 (continuation). Likewise, (S(H), g, ⊥, O) is a quasi-orthomodular nearsemilattice for ⊥ defined as in Section 3. Indeed, (i) if AB = O, then BA = O [13, Lemma 4.1]; (ii) if A (cid:22) B and BC = O, then A = BPA = PAB and AC = O; (iii) AO = O; thus, ⊥ is an orthogonality. Further, (v) if AB = O, then A (cid:22) A + B and B (cid:22) A + B by the definition of (cid:22), and A ◦ B (as S(H) is a nearsemilattice) [in addition, (v') A + B = A g B; see the proof of Theorem 4.3 in [19]]; (vi) if A (cid:22) B, then, by the definition and (v'), B = A + C = A g C for some C with A ⊥ C; (vii) if AB = O, AC = O and B (cid:22) A g C, then A + C = B + D for some D with DB = O, (D − A)B = O and, further, C = B + (D − A), i.e., B (cid:22) C. It was proved in Theorem 4.3 of [19] that a weakened version of (⊥4) holds in S(H). The orthogonality on S(H) is in fact additive. Indeed, suppose that D ⊥ A, D ⊥ B and A ◦ B. Then also PD ⊥ PA and PD ⊥ PB. As the lattice P(H) is orthomodular, it follows that PD ⊥ PA∨PB. On the other hand, PA∨PB = PAgB -- see (5). Therefore, PD ⊥ PAgB and D ⊥ A g B. The isomorphism of S(H) onto the nearsemilattice SH described in the Section 3 preserves orthogonality; so we may conclude that SH is quasi-orthomodular. Theorem 4.12 of [13] asserts that every initial segment of S(H) is isomorphic to an orthomodular lattice. We are now going to generalize this structure theorem to arbitrary quasi-orthomodular nearsemilattices. Recall that an orthocomplemen- tation, or an o-complementation, for short, on a bounded poset (P, ≤, 0, 1) is an m-complementation − such that 1 = x ∨ x− (equivalently, 0 = x ∧ x−) for all x ∈ P . Observe that then the induced orthogonality satisfies (⊥9). An o-complemented poset is orthomodular if this orthogonality satisfies also (⊥5) and (⊥6); conditions (⊥7) and (⊥4) are fulfilled in every such a poset. An orthomodular lattice is a lattice-ordered orthomodular poset. Theorem 8. Every initial segment of a quasi-orthomodular nearsemilattice A is an orthomodular lattice, where joins and meets agree with those existing in A. Proof. Due to (⊥6) and (⊥8), there is, for every x ≤ p, a unique element y ∈ [0, p] such that x ⊥ y and x ∨ y = p; we denote this element by x− p . The mapping x 7→ x− p )− p )− p ∨ x− p is an m-complementation on [0, p]. By the definition, p by p = x. If x ≤ y ≤ p, then x ⊥ y− (x− p ⊥ x− (⊥2) and y− As x ∨ x− p = p; hence (x− p and (x− p ≤ x− p by (−7) (observe that y− p = p, the m-complementation − p )− p ≤ x ∨ x− p is even an o-complementation in [0, p], and the interval [0, p] is an o-complemented poset. In virtue of (⊥5) and (⊥6), the poset is orthomodular. Since an initial segment [0, p] of a nearsemilattice is actually an upper semilattice by definition, the o-complemetantion − p turns it into a lattice. The final assertion is now trivial. (cid:3) p ). The following conclusion is immediate (recall that a distributive semilattice that happens to be a lattice is also distributive as a lattice). Corollary 9. In a distributive quasi-orthomodular nearsemilattice, every initial segment is a Boolean algebra. It is observed in [13] that M(A) is a generalized orthoalgebra, and [13, The- orem 4.2] asserts that so is also S(H). This result extends to arbitrary quasi- orthomodular nearsemilattices. 10 J ¯ANIS C¯IRULIS Definition 2. A generalized orthoalgebra is a system (A, ⊕, 0), where ⊕ is a partial binary operation and 0 is a nullary operation on A, satisfying the conditions (we write here, for arbitrary terms s and t, s ⊥ t to mean that s ⊕ t is defined): (⊕1) : (⊕2) : if x ⊥ y, then y ⊥ x and x ⊕ y = y ⊕ x, if x ⊥ y and x ⊕ y ⊥ z, then y ⊥ z, x ⊥ y ⊕ z and (x ⊕ y) ⊕ z = x ⊕ (y ⊕ z), (⊕3) : x ⊥ 0 and x ⊕ 0 = x, (⊕4) : (⊕5) : if x ⊥ y, x ⊥ z and x ⊕ y = x ⊕ z, then y = z, if x ⊥ x, then x = 0. The relation ≤ defined by x ≤ y if and only if y = x ⊕ z for some z with x ⊥ z, (6) is an order on a genralized orthoalgebra A and is called its natural ordering. Theorem 10. Suppose that (A, ∨, ⊥, 0) is a quasi-orthomodular nearsemilattice and that ⊕ is a binary operation on A defined by x ⊕ y = z if and only if x ⊥ y and z = x ∨ y. (7) Then (A, ⊕, 0) is a generalized orthoalgebra, and its natural ordering coincides with the nearsemilattice ordering of A. Proof. In virtue of (⊥5), the axioms (⊕1) -- (⊕5) are easy consequences of (∨2)+(⊥1), (∨3)+(⊥2), (∨4)+(⊥3), (⊥8) and (⊥9), respectively. We only note in connection with (⊕2) that the supposition x ⊕ y ⊥ z implies that (i) x, y ⊥ z (see (⊥2)) and (ii) x ∨ y ◦ z, i.e., x, y, z ≤ p for some p. Then (in the orthomodular lattice [0, p]) y, z, y ∨ z ≤ x− p and x ⊥ y ∨ z, i.e., x ⊥ y ⊕ z. The equivalence (6) follows from (⊥6). (cid:3) Example 2 (continuation). The operation ⊕ defined by f ⊕ g := f + g for f ⊥ g turns M(A) into a generalized orthoalgebra [13]. See the item (i) in the proof of [19, Theorem 3.2] for (7). Example 3 (continuation). If A ⊥ B, then put A⊕ B := A+ B [13]. Then S(H) is a generalized orthoalgebra [13, Theorem 4.2]. Evidently, A ⊕ B is an upper bound of A and B, and Corollary 4.5 in [13] says that it is actually a least upper bound. Thus, (7) also holds in S(H). The theorem implies that an orthomodular nearsemilattice satisfying (⊥4) is a generalized orthomodular poset (see [19] for an appropriate version of the latter notion). We say that a quasi-orthomodular nearsemilattice has the Riesz decomposition property if it satisfies the condition (⊕6) : if y ⊥ z and x ≤ y ⊕ z, then x = y′ ⊕ z′ for some y′ ≤ y and y′ ≤ z, where ⊕ is the operation (7) (cf. [11] or [19, Sect. 2]). By (⊥5) and (7), this property turns out to be equivalent to distributivity. Theorem 11. A quasi-orthomodular nearsemilattice is distributive if and only if it has the Riesz decomposition property. Proof. In a quasi-orthomodular nearsemilattice A, if y ⊥ z and x ≤ y ⊕ z, then x ≤ y ∨ z and, by distributivity, x = y′ ∨ z′ for some y′ ≤ y and z′ ≤ z. But y′ ⊥ z′ in virtue of (⊥2); therefore, x = y′ ⊕ z′. Now suppose that A has the Riesz FURTHER REMARKS ON AN ORDER FOR QUANTUM OBSERVABLES 11 decomposition property and that y ◦ z and x ≤ y ∨ z. By (⊥6), y ∨ z = y ∨ z0 for some z0 with z0 ⊥ y. As z ≤ y ∨ z0, there are y1 ≤ y and z1 ≤ z0 such that y1 ⊥ z1 and z = y1 ∨ z1. Then y ∨ z = y ∨ z1, z1 ≤ z and, in virtue of (⊥2), y ⊥ z1. It follows that x = y′ ∨ z′ for some y′ ≤ y and z′ ≤ z1 ≤ z. (cid:3) 6. Skew meets on quasi-orthomodular nearsemilattices We concentrate in this section on quasi-orthomodular nearsemilattices that ad- mit a non-commutative meet-like operation, and demonstrate that the nearsemi- lattices in Examples 1 -- 3 belong to this type of nearsemilattices. Let A be a quasi-orthomodular nearsemilattice, and assume that orthogonality in it is additive. For x, y ∈ A, we write x ⊏ y to mean that, for every z ∈ A, z ⊥ y only if z ⊥ x. We say that x is overridden by y, if x ⊏ y. The overriding relation ⊏ has the following properties: (⊏1) : ⊏ is reflexive and transitive, (⊏2) : (⊏3) : (⊏4) : if x ≤ y, then x ⊏ y. if x ⊏ y and x ◦ y, then x ≤ y, if x ⊏ z, y ⊏ z and x ◦ y, then x ∨ y ⊏ z. Only (⊏3) requires some comment. By (⊥6), x ∨ y = x0 ∨ y for some x0 with x0 ⊥ y. Since x ⊏ y, then x0 ⊥ x and, further, x0 ⊥ x ∨ y (see (⊥4)). Now, x0 = 0 by (⊥2) and (⊥9). Therefore, ⊏ is a preorder. We denote by k the equivalence relation on A induced by it: ⊏: x k y iff x ⊏ y and y ⊏ x Abstract overriding relations satisfying (⊏1) -- (⊏4) and one more condition (⊏5) : if x ⊏ y, then x k x′ ≤ y for some x′, were introduced in [4, Definition 2.2]. Observe that the element x′ in the right side of (⊏5) is uniquely defined; in fact, x′ = max{u : u ⊏ x and u ≤ y}. Indeed, x′ ⊏ x and x′ ≤ y by the definition of k; on the other hand, if u ⊏ x and u ≤ y, then u ⊏ x′ and u ◦ x′, whence u ≤ x′ by (⊏3). If the element ←− ∧ y := max{u : u ⊏ x and u ≤ y} x exists for some x and y, we call it a (right-handed) skew meet of x and y. Example 1 (continuation). In PF(I, V ), ϕ ⊏ ψ if and only if dom ϕ ⊆ dom ψ. This overriding relation on PF(I, V ) satisfies (⊏5): if ϕ ⊏ ψ and ϕ′ := ψ dom ϕ, then ϕ k ϕ′ ⊆ ψ. However, (⊏5) may fail in an arbitrary functional nearsemilattice ←− (even if the latter is closed). More generally, the skew intersection ∩ on PF(I, V ) ←− is totally defined and is given by ϕ ∩ ψ = ψ dom(ϕ ∩ ψ) (see [5]). Observe that ϕ ⊆ ψ iff ϕ ←− ∩ ψ = λ and ϕ ⊏ ψ iff ψ ←− ∩ ψ = ϕ, ϕ ⊥ ψ iff ϕ ←− ∩ ϕ = ϕ. Example 2 (continuation). In M(A), f ⊏ g if and only if supp f ⊆ supp g. In connection with (⊏5), observe that if f ⊏ g, then the function f ′ which agrees with g on supp f and vanishes outside supp f belongs to M(A). More generally, the skew meet ←−f on M(A) is a total operation and is given by f ←−f g = gχ(supp f ∩supp g). Example 3 (continuation). In S(H), A ⊏ B if and only if ran A ⊆ ran B. Indeed, suppose that C ⊥ A whenever C ⊥ B, and choose x ∈ ran A. Take for C the pro- jection onto the closed subspace [x] spanned by x. Then C 6⊥ A and, consequently, C 6⊥ B. Hence, [x] ⊆ ran B, and x ∈ ran B. Conversely, if ran A ⊆ ran B and C ⊥ B, then ran C ∩ ran A ⊆ ran C ∩ ran B and C ⊥ A. 12 J ¯ANIS C¯IRULIS Equivalently, A ⊏ B if and only if PA ≤ PB. The axiom (⊏5) means that if PA ≤ PB, then PA = PA′ and A′ (cid:22) B for some A′ ∈ S(H) (equivalently, PA ∈ LB, or B ran A ∈ SH ). A natural candidate for A′ is the operator BPA; we, however cannot prove that it is self-adjoint, i.e., that B and PA commute. ←−f in S(H) is total. The bounded subset However, the skew meet operation B := {P : P ≤ PA, PA ≤ PB and BP = P B} of the complete lattice LB always LA has the greatest element P ∗. On the other hand, LA B = {PC ∈ LB : PC ≤ PA} = {PC : C ⊏ A and C (cid:22) B}. Now, a reasoning similar to that at the end of Section 4 demonstrates that BP ∗ is the skew meet of A and B. We end with a theorem describing characteristic properties of the skew meet operation ←− ∧ . Theorem 12. Suppose that A is a quasi-orthomodular nearsemilattice with an ←− ∧ y in A exist. Then the algebra additive orthogonality and that all skew meets x (A, ←− ∧ ) is an idempotent semigroup, and the following conditions are fulfilled: x ←− ∧ y ≤ y, ←− ∧ y = x, x ≤ y iff x ←− ∧ y ⊏ x, x y ⊏ x iff x ←− ∧ y = y. (8) (9) Moreover, ←− ∧ is commutative in every initial segment of A. Proof. Evidently, the operation trivial. To prove that the operation is associative, observe that ←− ∧ is idempotent, and both inequalities (8) are ←− ∧ y) ←− ∧ z = max(v : v ⊏ max(u : u ⊏ x and u ≤ y) and v ≤ z) (x = max(v : v ⊏ u ⊏ x and u ≤ y for some u, and v ≤ z) (10) and ←− ∧ (y x ←− ∧ z) = max(v : v ⊏ x and v ≤ max(u : u ⊏ y and u ≤ z)) = max(v : v ⊏ x, and v ≤ u ≤ z and u ⊏ y for some u) = max(v : v ⊏ x, v ≤ z and v ⊏ y). (11) Now notice that if an element v satisfies, for some u, the conditions v ⊏ u ⊏ x, u ≤ y and v ≤ z from (10), then it satisfies also the conditions v ⊏ x, v ≤ z and v ⊏ y from (11). Conversely, if the latter triple of conditions is satisfied, then also the ←− ∧ y. We only note that if v ⊏ x, y, then there is former one is satisfied with u = x v′ such that v k v′ ≤ y (see (⊏5)), and, further, v′ ≤ u; it follows that v ⊏ u. Therefore, the maxima in (10) and (11) are equal. Next, if x ≤ y and u ≤ y, ←− ∧ y = x. The converse is evident ←− ∧ y, then x ◦ u and, if also u ⊏ x, then u ≤ x. Thus, x by (8), and this proves the first identity in (9). Further, if y ⊏ x, then y ≤ x and the converse again comes from (8); this proves the other identity. At last, it follows from (⊏3) by virtue of (⊥2) that x ←− ∧ y = x ∧ y = y ←− ∧ x when (cid:3) x ◦ y. The theorem shows that the algebra (A, ∨, ←− ∧ , 0) is a right normal skew nearlat- tice in the sense of [5, 6]. Thus, in particular, S(H) is such an algebra. FURTHER REMARKS ON AN ORDER FOR QUANTUM OBSERVABLES 13 References [1] ANTEZANA, J. -- CANO, C. e.a.: A note on the star order in Hilbert spaces, Linear Multi- linear Algebra 58, (2010), 1037 -- 1051. [2] CHAJDA, I. -- KOLA R´IK, M.: Nearlattices, Discrete Math. 308, (2008), 4906-4913. [3] C¯IRULIS, J.: Subtractive nearsemilattices, Proc. Latv. Acad. Sci., Sect. B, 52, (1998), 228- 233. [4] C¯IRULIS, J.: Knowledge representation in extended Pawlak's information systems: algebraic aspects, in: Eiter, T., Schewe, K.-D. (eds): FoIKS 2002, Lect. Notes Comput. Sci. 2284, Springer-Verl., Berlin e.a., 2002, 250 -- 267. [5] C¯IRULIS, J.: Knowledge representation systems and skew nearlattices, in: Chajda, I. (ed.) et al., Contributions to General Algebra 14, Verl. Johannes Heyn, Klagenfurt, 2004, 43 -- 51. [6] C¯IRULIS, J: Skew nearlattices: some structure and representation theorems, in: Chajda, I. (ed.) et al., Contributions to General Algebra 19, Verl. Johannes Heyn, Klagenfurt, 2010, 33 -- 44. [7] C¯IRULIS, J.: Subtraction-like operations in nearsemilattices, Dem. Math. 43 (2010), 725 -- 738. [8] CORNISH, W.H.: Conversion of nearlattices into implicative BCK-algebras, Math. Semin. Notes, Kobe Univ. 10 (1982), 1-8. [9] CORNISH, W.H. -- NOOR, A.S.A.: Standard elements in a nearlattice, Bull. Austral. Math. Soc. 26 (1982), 185-213. [10] DRAZIN, M.F.: Natural structures on semigroups with involution, Bull. Amer. Math. Soc. 84 (1978), 139 -- 141. [11] DVURE CENSKIJ, A. -- PULMANNOV ´A, S.: New Trends in Quantum Structures, Kluwer Acad. Publ./Ister Sci., Dordrecht/Bratislava, 2000. [12] DU, H. -- DOU, Y.: A spectral representation of the infimum of selfadjoint operators in the logic order (Chinese), Acta Math. Sin., Chin. Ser. 52 (2009), 1141-1146. [13] GUDDER, S: An order for quantum observables, Math. Slovaca 56 (2006), 573 -- 589. [14] JANOWITZ, M.F.: A note on generalized orthomodular lattices, J. Natur. Sci. Math. 8 (1968), 89 -- 94. [15] LIU, W. -- WU, J.: A representation theorem of infimum of bounded observables, J. Math. Phys. 49 (2008), paper No. 073521, 5 pp. [16] LIU, W. -- WU, J.: A supremumum of bounded quantum observables, J. Math. Phys. 50 (2009), paper No. 083513, 4 pp. [17] MAYET-IPPOLITO, A.: Generalized orthomodular posets, Demonstr. Math. 24 (1991), 263 -- 274. [18] NOOR, A.S.A, -- CORNISH, W.H.: Multipliers on a nearlattice, Commentat. Math. Univ. Carol. 27 (1986), 815 -- 827. [19] PULMANNOV ´A, S. -- VINCEKOV ´A, E.: Remarks on the order for quantum observables, Math. Slovaca 57 (2007), 589 -- 600. [20] SHEN, J, -- WU, J.: Spectral representation of infimum of bounded quantum observables J. Math. Phys. 50 (2009), paper No. 1135014, 4 pp. [21] XU, X. -- DU, H. -- FANG, X.: An explicit expression of supremum of bounded quantum ob- servables, J. Math. Phys. 50 (2009), paper No. 033502, 9 pp. Faculty of Computing University of Latvia Rain¸a b., 19 Riga LV-1586 LATVIA E-mail address: [email protected]
1611.03396	3	1611	2019-08-29T01:39:05	On a spectral theorem of Weyl	[ "math.OA", "math.CA", "math.SP" ]	We give a geometric proof of a theorem of Weyl on the continuous part of the spectrum of Sturm-Liouville operators on the half-line with asymptotically constant coefficients. Earlier proofs due to Weyl and Kodaira depend on special features of Green's functions for linear ordinary differential operators; ours might offer better prospects for generalization to higher dimensions, as required for example in noncommutative harmonic analysis.	math.OA	math	ON A SPECTRAL THEOREM OF WEYL NIGEL HIGSON AND QIJUN TAN ABSTRACT. We give a new proof of a theorem of Weyl on the continuous part of the spectrum of Sturm-Liouville operators on the half-line with asymptotically constant coefficients. Earlier arguments, due to Weyl and Kodaira, depended on particular features of Green's functions for linear ordinary differential operators. Ours uses a concept of asymptotic containment of C∗-algebra representations that has geometric origins. We apply the concept elsewhere to the Plancherel formula for spherical functions on reductive groups. 1. INTRODUCTION The purpose of this paper is to present a new approach to an old theorem of Hermann Weyl on the spectral theory of self-adjoint Sturm-Liouville operators on a half-line. Our aim is to introduce methods that are more geometric and more amenable to generalization than the originals. We show elsewhere [Tan19] that the same arguments apply to Harish-Chandra's Plancherel formula for spherical functions (in fact Harish-Chandra was very much inspired by Weyl's work; com- pare [Bor01, p. 38] and [Ban08]). Sturm-Liouville theory is of course concerned with the eigenvalues and eigen- functions of linear differential operators (1.1) D = − d dx · p(x) · d dx + q(x), initially on a closed interval [a, b]. Assume for simplicity that p(x) and q(x) are smooth, real-valued functions on [a, b], with p(x) positive everywhere. In exam- ining the nonzero solutions of the eigenvalue problem (1.2) Dfλ = λfλ, it is appropriate to impose suitable self-adjoint boundary conditions. For the sake of this introduction let us choose the simplest of these, namely (1.3) fλ(a) = 0 = fλ(b). The elements of Sturm-Liouville theory can then be summarized as follows: 1.4. Theorem. The eigenvalues λ for the above problem are real numbers, and each has multiplicity one. The set of all eigenvalues is a discrete subset of R, bounded below, and if h is any smooth function on [a, b], then for x ∈ (a, b). h(x) = λ (cid:104)fλ, h(cid:105) (cid:104)fλ, fλ(cid:105) fλ(x) (cid:88) 1 2 NIGEL HIGSON AND QIJUN TAN In an influential paper from early in his career, Weyl developed an analogous theory for Sturm-Liouville operators on a half-line rather than a bounded interval [Wey10]. Weyl's paper addresses many issues, but our concern here is his treat- ment of the continuous spectrum of certain Sturm-Liouville operators, and espe- cially his version, for the continuous spectrum, of the expansion theorem above. Assume that the coefficient functions p(x) and q(x) in (1.1) are defined on [0,∞) and converge sufficiently rapidly to the constants 1 and 0, respectively, as x tends to infinity. For the purpose of this introduction, let us assume more than Weyl, namely that (1.5) p(x) ≡ 1 and q(x) ≡ 0 if x (cid:29) 0 (this assumption is too strong to be interesting in applications, but it allows us to quickly introduce Weyl's result). For each λ ∈ C there is a one-dimensional space of eigenfunctions Fλ for D that satisfy the boundary condition (1.6) Fλ(0) = 0. If we focus on the case where λ > 0, and if we choose, as we may, Fλ to be nonzero and real-valued, then our assumptions on the coefficient functions p(x) and q(x) imply that √ √ λ x + c(λ)e−i Fλ(x) = c(λ)ei (1.7) for some nonzero c(λ) ∈ C and all x (cid:29) 0. Weyl's result for the continuous spec- trum, expressed in L2-terms, is as follows (there is also a pointwise result that is analogous to Theorem 1.4, which may be derived from the L2-result): 1.8. Theorem. If h and g are smooth and compactly supported functions on (0,∞), then λ x (cid:90)∞ (cid:88) λ<0 (cid:104)f, g(cid:105) = (cid:104)g, Fλ(cid:105)(cid:104)Fλ, h(cid:105) (cid:104)Fλ, Fλ(cid:105) + 1 4π (cid:104)g, Fλ(cid:105)(cid:104)Fλ, h(cid:105) 0 c(λ)2 dλ√ λ , The sum is over the square-integrable eigenfunctions associated to negative eigenvalues that satisfy the boundary condition (1.6), and there are finitely many of these. The integral is absolutely convergent. All the inner products in the formula have the standard L2-form. We shall approach Weyl's theorem by comparing the Sturm-Liouville operator D to the simpler operator But we shall regard D0 as an operator on the full line (−∞,∞), rather than the d2 dx2 D0 = − half-line, and the translation-invariance of D0 on the full line will be crucial. To explain why, it is helpful to make the following general definition: 1.9. Definition. Let A be a C∗-algebra and let . π : A −→ B(H) and π0 : A −→ B(H0) be nondegenerate Hilbert space representations of A. We shall say that π0 is asymp- totically contained in π if (i) There is a one-parameter group of unitary operators Ut : H0→H0 that com- mute with the operators in π0[A]. (ii) There is a bounded operator W : H0→H such that for every a ∈ A, and for ON A SPECTRAL THEOREM OF WEYL 3 every u, v ∈ H0, t→+∞(cid:104)(cid:10)u, π0(a)v(cid:11) lim H0 −(cid:10)WUtu, π(a)WUtv(cid:11) (cid:105) = 0. H (1.10) tional calculus representations of representations [Dix77, Definition 3.4.5]. Note that asymptotic containment of representations implies weak containment For Weyl's theorem, we take A=C0(R), and we define π and π0 to be the func- π : ϕ (cid:55)−→ ϕ(D) and π0 : ϕ (cid:55)−→ ϕ(D0), on H=L2(0,∞) and H0=L2(−∞,∞) respectively. We define {Ut} to be the stan- dard one-parameter unitary group of translations on L2(−∞,∞) and take W : L2(−∞,∞) −→ L2(0,∞) to be the obvious restriction operator. The asymptotic containment of π0 in π follows from the condition (1.5) on the coefficients of D; see Lemma 3.3. Returning to the general case, we shall assume that the C∗-algebra A is separa- ble and commutative, as it is in the examples of concern to us, and that the Hilbert spaces H and H0 are separable, too. Then the representations π and π0 may be decomposed into direct integrals (cid:90)⊕ (cid:90)⊕ H = (1.11) over the spectrum of A, so that the action of a∈A on each space in either integral is through the character a (cid:55)→ λ(a). We shall assume that the spaces Hλ and H0,λ Hλ dµ(λ) and H0 = H0,λ dµ0(λ) are finite-dimensional, as again is the case in the examples of concern to us. H H0 (cid:10)u, π0(a)v(cid:11) =(cid:10)Wu, π(a)Wv(cid:11) Now let us assume temporarily that the asymptotic containment relation (1.10) is replaced by the exact containment relation (1.12) for all u, v ∈ H0 and all a ∈ A, so that the operator W is necessarily an isometric inclusion of π0 as a subrepresentation of π. It follows that the operator W decom- poses into a field of operators (1.13) and that each W∗ multiple is Trace(W∗ λWλ is a multiple of the identity operator on H0,λ. Of course that Wλ : H0,λ −→ Hλ. λWλ)/dim(H0,λ). Using the family {Wλ}, the direct integral decomposition of π in (1.11) gives rise to an alternative direct integral decomposition of π0. Comparing the two, we find that the measure µ0 in (1.11) is absolutely continuous with respect to µ, and that indeed (1.14) dµ0 dµ (λ) = Trace(W∗ dim(H0,λ) λWλ) for µ-almost all λ in the support of the representation π0. The main observation of this paper, which, aside from some functional-analytic details, is very simple, is that even when π0 is only asymptotically contained in π, we can still derive a version of the formula (1.14) in essentially the same way, if we 4 assume the existence of operators Wλ : H0,λ→Hλ that asymptotically decompose W NIGEL HIGSON AND QIJUN TAN in the sense that (1.15) t→+∞ Wλ(Utv)0,λ − (WUtv)λ = 0. lim See Section 2 for a precise account of the assumptions we make. As for (1.15), it is easiest to understand its meaning by examining the adjoint operators λ : Hλ −→ H0,λ. Cλ = W∗ In the context of the Sturm-Liouville problem, the spaces Hλ and H0,λ may be understood as λ-eigenspaces for the operators D and D0, respectively, and the asymptotic formula (1.15) simply asserts that each eigenfunction of D is mapped by Cλ to an eigenfunction of D0 to which it is asymptotic in the sense of (1.7). This proves the existence of the operators Cλ in this context, and also computes the trace in (1.14) in terms of c(λ)2. Weyl's formula in Theorem 1.8 is an immediate consequence.1 Other interesting examples of asymptotic containment of representations come from representation theory. In brief, if G is a real reductive group with maximal compact subgroup K, and if P = MpApNp is a minimal parabolic subgroup, then the representation of C∗(G) on L2(G/MpNp) is asymptotically contained in the representation of C∗(G) on L2(G/K). The case of SL(2, R) is illustrated in Figure 1. The figure should make it clear that the asymptotic containment in this example has a very geometric origin. FIGURE 1. The homogeneous spaces G/K and G/MpNp for the reduc- tive group G = SL(2, R), realized as coadjoint orbits. To fit this example within the framework of this paper we take the C∗-algebra A in our asymptotic containment to be the commutative C∗-subalgebra of C∗(G) that is generated by the K-bi-invariant compactly supported smooth functions on G. It is naturally represented on the Hilbert spaces H and H0 of K-fixed vectors within L2(G/K) and L2(G/MpNp), respectively. The K-invariant functions on G/K and G/MpNp can be identified with functions on A+ p is a dominant chamber in Ap. A suitable operator W may be defined using restriction of functions from Ap to A+ p (followed by a translation away from the walls of A+ p to make W bounded). The minimal parabolic is defined by a one-parameter sub- group of A, and right-translation on G/MpNp by this one-parameter group gives p and Ap respectively, where A+ 1To be accurate, the Radon-Nikodym derivative formula only determines µ on the positive part of the spectrum because the necessary assumptions on Wλ only hold there. A separate argument is required for the nonpositive spectrum; see Section 4. ON A SPECTRAL THEOREM OF WEYL 5 the necessary one-parameter unitary group on H0. The counterpart of Weyl's the- orem in this example is Harish-Chandra's Plancherel theorem for spherical func- tions. See [Tan19]. In fact the present paper grew out of a project in noncommutative geometry involving the Plancherel formula [CCH16, CH16]. The reader is referred to [Ban08] for an interesting and thorough discussion of the relation between Weyl's theorem and harmonic analysis on symmetric spaces. After this paper was written the authors were lucky to enjoy a very stimulat- ing conversation with Joseph Bernstein, who pointed out that he had obtained very similar results in unpublished work from the 1980's (see the final remarks in [Ber88, Sec. 0.2] for hints of this). Some of the spectral-theoretic methods from [SV17], which studies harmonic analysis on p-adic spherical varieties, are also very closely related to the methods of this paper. See especially Section 8 of [SV17]. Here is a brief outline of the present paper. We shall discuss asymptotic con- tainment of representations in Section 2. The main result is 2.14. We shall apply our method to the positive, continuous spectrum of Sturm-Liouville operators in Section 3, and we shall briefly address the nonpositive spectrum in Section 4. We shall consider not only the operators discussed in this introduction, but also a non- trivial example related to the representation theory of SL(2, R). In an appendix we quickly review the Weyl-Kodaira approach to Theorem 1.8 for the sake of compar- ison. 2. ASYMPTOTIC CONTAINMENT OF REPRESENTATIONS In this section we shall describe our approach to Weyl's theorem. We shall for- mulate the method in fairly general terms, applicable to examples beyond Weyl's theorem, although we shall not strive for the upmost generality in the assumptions that we make. Let A be a separable, commutative C∗-algebra with Gelfand spectrum Λ, so that of course A ∼= C0(Λ) for some locally compact space Λ. We shall view elements of A as continuous functions on Λ without further comment. Let us suppose that we are given two non-degenerate representations of A on unitary group {Ut} on H0 and operator W : H0→ H as described in the definition. We shall assume that π0 is asymptotically contained in π, as in Definition 1.9, with Our analysis of the relation between π and π0 will be based on following formula, which is an immediate consequence of (1.10). 2.1. Lemma. If a ∈ A and if g, h ∈ H0, then T→+∞ 1 (cid:104)WUtg, π(a)WUth(cid:105)H dt. (cid:104)g, π0(a)h(cid:105)H0 = lim (cid:90) (cid:3) T 0 T We shall make the following assumptions concerning the representation π0; in the case of Sturm-Liouville operator they will be easy to verify using Fourier anal- ysis. 2.2. Assumption. We shall suppose that we are given: separable Hilbert spaces: π : A −→ B(H) and π0 : A −→ B(H0). 6 NIGEL HIGSON AND QIJUN TAN (i) An open subset Λ0 ⊆ Λ and a locally trivial continuous field of finite-dimen- sional Hilbert spaces {H0,λ}λ∈Λ0 over Λ0 (or in other words a Hermitian vec- tor bundle). (ii) A dense subspace H0 ⊆ H0 and a linear map h (cid:55)→ {h0,λ} from H0 into the continuous sections of {H0,λ} such that H0,λ = { h0,λ : h ∈ H0 } (iii) A Borel measure µ0 on Λ0 such that (cid:104)h0,λ, g0,λ(cid:105) is a µ0-integrable function of for every λ ∈ Λ0. λ, for every h, g ∈ H0, and such that (cid:90) (cid:10)h, π0(a)g(cid:11) (cid:10)h0,λ, g0,λ (cid:11) for every h, g ∈ H0 and every a ∈ A. = H0 Λ0 a(λ) dµ0(λ) H0,λ 2.3. Assumption. We shall assume that the action of the one-parameter unitary group {Ut} on the Hilbert space H0 maps the subspace H0 into itself, and that the continuous field {H0,λ}λ∈Λ0 carries a continuous, unitary action {Ut,λ} of R such that for every h ∈ H0 and every λ ∈ Λ0. (Uth)0,λ = Ut,λh0,λ Next, we shall make assumptions on the representation π that are similar to those in Assumption 2.2, except that we shall in addition assume that π has multi- plicity one: the fibers in the field of Hilbert spaces that decomposes π have dimen- sion one. This assumption isn't altogether necessary (finite-dimensionality would suffice), but it simplifies the statements of the results that follow, along with their proofs, and it is satisfied in the situations of interest to us. Here are the details. 2.4. Assumption. We shall suppose that there are given: over Λ (that is, a Hermitian line bundle over Λ). (i) A locally trivial continuous field of one-dimensional Hilbert spaces {Hλ}λ∈Λ (ii) A dense subspace H ⊆ H such that if h ∈ H0 then WUth ∈ H for all t (cid:29) 0. (iii) A family of linear maps ελ : h (cid:55)→ hλ from H into Hλ so that {hλ} is a contin- uous section, and a Borel measure on Λ such that (cid:104)hλ, gλ(cid:105) is a µ-integrable function of λ, for every h, g ∈ H, and such that (cid:10)h, π(a)g(cid:11) (cid:90) (cid:10)hλ, gλ (cid:11) = H Λ a(λ) dµ(λ) Hλ for every h, g ∈ H and every a ∈ A. Finally, we shall make the following assumption concerning the asymptotic re- lation between the fields {Hλ} and {H0,λ}. As we noted in the introduction, and as we shall see clearly in the next section, in the Sturm-Liouville context this means that the operator Cλ below maps each λ-eigenfunction for D to a λ-eigenfunction for D0 to which it is asymptotic. 2.5. Assumption. We shall assume that there is given a continuous family of injec- tive linear maps Cλ : Hλ −→ H0,λ (λ ∈ Λ0) ON A SPECTRAL THEOREM OF WEYL 7 with the property that if h belongs to H0, and if {vλ} is a continuous section of {Hλ}, and K is a compact subset of Λ0, then (cid:12)(cid:12)(cid:12)(cid:10)Cλvλ, (Uth)0,λ (cid:11) H0,λ −(cid:10)vλ, (WUth)λ (cid:11) (cid:12)(cid:12)(cid:12) = 0. Hλ t→+∞ sup lim λ∈K Using the four assumptions listed above we shall prove: 2.6. Theorem. The measure µ0 is absolutely continuous with respect to µ on the open set Λ0, with Radon-Nikodym derivative dµ0 dµ Trace(C∗ λCλ) dim(H0,λ) (λ) = . Here is the proof in outline. We are assuming that Λ0 (cid:90) (cid:10)h0,λ, h0,λ (cid:11) (cid:104)h, π0(ϕ)h(cid:105)H0 = (cid:90) (cid:10)h0,λ, Av(cid:2)CλC∗ (cid:3)h0,λ λ] : H0,λ→ H0,λ is defined by the averaging formula (cid:3) = lim T→∞ 1 Av(cid:2)CλC∗ U−t,λCλC∗ ϕ(λ) dµ0(λ). ϕ(λ)dµ(λ), λUt,λ dt (cid:11) (cid:90) H0,λ H0,λ Λ0 λ λ T 0 T We shall obtain from our remaining assumptions a new integral formula, namely (2.7) (cid:104)h, π0(ϕ)h(cid:105)H0 = (2.8) where the operator Av[CλC∗ (since we are dealing here with operators on the finite-dimensional space H0,λ, the limit certainly exists). At this point, we can appeal to the following uniqueness result for spectral decompositions: 2.9. Lemma. Let {Tλ} be a measurable field of positive operators on {H0,λ}λ∈Λ0. Suppose that (cid:90) ϕ(λ) dµ(λ) = H0,λ Λ0 ϕ(λ) dµ0(λ) H0,λ for every h ∈ H0 and every continuous and compactly supported function ϕ. Then Tλ is a scalar multiple of the identity for µ-almost all λ ∈ Λ0. In addition µ0 is absolutely continuous with respect to µ on Λ0 and (cid:90) Λ0 (cid:10)h0,λ, Tλh0,λ (cid:11) (cid:10)h0,λ, h0,λ (cid:11) Tλ = (λ) · IH0,λ dµ0 dµ µ-almost everywhere on Λ0. Proof. For each point λ0 ∈ Λ0 there exists h ∈ H0 for which the section h0,λ is nonzero at λ0. It follows immediately from the uniqueness part of the Riesz repre- sentation theorem that µ0 is absolutely continuous with respect to µ near λ0 with Radon-Nikodym derivative (cid:11) (cid:10)h0,λ, Tλh0,λ (cid:10)h0,λ, h0,λ (cid:11) H0,λ . H0,λ dµ0 dµ (λ) = Since the derivative is independent of {h0,λ} this implies that almost everywhere, as required. Tλ = (λ) · IH0,λ , dµ0 dµ (cid:3) 8 NIGEL HIGSON AND QIJUN TAN Av[CλC∗ tation Returning to the proof of Theorem 2.6, Lemma 2.9 tells us that the operator λ] is a scalar multiple of the identity for µ-almost-all λ ∈ Λ0. The compu- Trace(cid:0)Av(cid:2)CλC∗ λ (cid:3)(cid:1) = Trace(CλC∗ λ) = Trace(C∗ λCλ), determines the multiple, and the theorem follows. So it remains to establish the integral formula (2.8): 2.10. Lemma. If h ∈ H0 and if ϕ is a continuous and compactly supported function on Λ0, then (cid:104)h, π0(ϕ)h(cid:105)H0 = ϕ(λ)dµ(λ). H0,λ (cid:90) Λ0 (cid:10)h0,λ, Av(cid:2)CλC∗ λ (cid:3)h0,λ (cid:11) Hλ Hλ dt. H (2.11) (cid:11) , Hλ (cid:11) Hλ = H Λ0 T 0 T ϕ(λ)dµ(λ). According to Lemma 2.1, Since the space Hλ are one-dimensional, we can write (cid:10)Wth, π(ϕ)Wth(cid:11) Now use Assumption 2.4 to write the integrand in the right hand side of (2.11) as Proof. It suffices to prove this formula for all functions ϕ that are supported on compact sets K ⊆ Λ0 over which the field {Hλ} is trivializable, and so we shall assume that here. In addition we shall use the notation where {vλ} is a continuous section of {Hλ} with (cid:107)vλ(cid:107)Hλ = 1 for all λ ∈ K. So (2.12) (cid:104)h, π0(ϕ)h(cid:105)H0 Wt = WUt : H0 −→ H. (cid:90) T→+∞ 1 (cid:104)h, π0(ϕ)h(cid:105)H0 = lim (cid:90) (cid:10)Wth, π(ϕ)Wth(cid:11) (cid:10)(Wth)λ, (Wth)λ (cid:10)(Wth)λ, (Wth)λ (cid:11) =(cid:10)(Wth)λ, vλ (cid:90) (cid:90) (cid:10)(Wth)λ, vλ (cid:11) ·(cid:10)Cλvλ, (Uth)0,λ (cid:11) (cid:11) ·(cid:10)vλ, (Wth)λ (cid:11) ·(cid:10)vλ, (Wth)λ (cid:11) ·(cid:10)vλ, (Wth)λ (cid:11) −(cid:10)(Wth)λ, vλ (cid:105) (cid:11) −(cid:10)vλ, (Wth)λ (cid:105)(cid:10)vλ, (Wth)λ (cid:11) −(cid:10)(Wth)λ, vλ (cid:12)(cid:12) ≤ (cid:107)hλ(cid:107) · (cid:107)Cλvλ(cid:107), the expression (2.13) converges to zero as t→∞, uniformly in λ ∈ K. (cid:10)(Uth)0,λ, Cλvλ The terms in the square brackets converge to 0, as t→ +∞, uniformly in λ ∈ K. In T→+∞ 1 (cid:10)(Uth)0,λ, Cλvλ (cid:11) (cid:104)(cid:10)Cλvλ, (Uth)0,λ (cid:11) (cid:104)(cid:10)(Uth)0,λ, Cλvλ (cid:11) (cid:11) (cid:12)(cid:12)(cid:10)(Uth)0,λ, Cλvλ we see that (cid:104)(Uth)0,λ, Cλvλ(cid:105)H0,λ is uniformly bounded in t and λ ∈ K. It follows from this and Assumption 2.5 that (cid:104)(Wth)λ, vλ(cid:105)Hλ is uniformly bounded too. So (cid:12)(cid:12) =(cid:12)(cid:12)(cid:10)Ut,λh0,λ, Cλvλ Λ0 Consider now the difference 0 which we can write as addition, since ϕ(λ)dµ(λ)dt. (cid:11) (cid:11) = lim H0,λ + T T Hλ Hλ Hλ Hλ H0,λ H0,λ H0,λ H0,λ , Hλ . Hλ (cid:11) H0,λ (2.13) H0,λ Hλ Observe next that (cid:10)(Uth)0,λ, Cλvλ (cid:11) H0,λ ·(cid:10)Cλvλ, (Uth)0,λ (cid:11) ON A SPECTRAL THEOREM OF WEYL 9 ·(cid:10)Cλvλ, Ut,λh0,λ (cid:11) (cid:11) H0,λ H0,λ λUt,λh0,λ . H0,λ H0,λ =(cid:10)Ut,λh0,λ, Cλvλ (cid:11) =(cid:10)h0,λ, U−t,λCλC∗ (cid:11) (cid:90) (cid:10)h0,λ, U−t,λCλC∗ (cid:90) (cid:16)(cid:90) T = lim T T→+∞ 1 (cid:90) T→+∞ = lim 0 Λ0 (cid:90) T (cid:16) 1 It follows from our analysis of (2.13) that the inner integral in (2.12) is asymptotic to the integral (that is, the difference converges to zero as t→ +∞). As a result, λUt,λh0,λ ϕ(λ) dµ(λ) H0,λ Λ0 (cid:104)h, π0(ϕ)h(cid:105)H0 (cid:10)h0,λ, U−t,λCλC∗ λUt,λh0,λ (cid:11) H0,λ ϕ(λ) dµ(λ) (cid:17) dt, It now follows from Fubini's theorem, that (cid:104)h, π0(ϕ)h(cid:105)H0 (cid:10)h0,λ, U−t,λCλC∗ (cid:11) (cid:17) terchange the limit as T → +∞ and the integral over Λ0 to obtain (2.8), as re- The integral in the parentheses is uniformly bounded in T. Therefore we can in- (cid:3) H0,λ Λ0 T 0 λUt,λh0,λ dt ϕ(λ)dµ(λ). quired. Theorem 2.6 gives a formula for the measure µ0 in terms of the measure µ. But since our goal is to obtain information about the measure µ, we should invert this formula: 2.14. Theorem. The measue µ is absolutely continuous with respect to the measure µ0 on Λ0, and the Radon-Nikodym derivative of µ with respect to µ0 on Λ0 is dµ dµ0 (λ) = dim(H0,λ) Trace(C∗ λCλ) . 3. STURM-LIOUVILLE OPERATORS (cid:3) In this section we shall apply the approach of Section 2 to Sturm-Liouville op- erators on the half-line. So let · p(x) · d dx c (0,∞). ∞ where the coefficient functions p(x) and q(x) are smooth and real-valued on (0,∞), + q(x), D = − d dx and where p(x) is everywhere positive. We shall assume that D is a self-adjoint operator on some domain that includes C ably). We shall study the following examples (which may be generalized consider- 3.1. Example. If p and q are in fact smooth on [0,∞) and eventually constant, smooth, compactly supported functions on [0,∞) that vanish at 0. with p positive, as in Section 1, then D is essentially self-adjoint on the domain of 10 NIGEL HIGSON AND QIJUN TAN 3.2. Example. If G=SL(2, R) and K=SO(2), then the symmetric space G/K may be identified with the hyperbolic plane (with G acting as isometries on the plane). The Laplace-Beltrami operator ∆ is essentially self-adjoint on the space of smooth and compactly supported functions on G/K. On K-invariant functions it acts as d dr2 d dr , ∆ = − − coth(r) action of K. Now identify the K-fixed part of L2(G/K) with L2(0,∞) using the where r is the radial coordinate in the polar coordinate system associated to the radial coordinate and multiplying by sinh(r) 1 dArea = sinh(r)drdθ). We obtain an essentially-self adjoint operator 2 (the latter comes from the formula 4 − 1 4 = − mind). d dx2 ∆ = D + 1 4 csch2(x) + 1 Associated to the unbounded self-adjoint operator D on the Hilbert space H = on L2(0,∞) (we have subtracted the term 1/4 from ∆ with Lemma 3.3 below in L2(0,∞) is the functional calculus representation π : C0(R) −→ B(H) π : ϕ (cid:55)−→ ϕ(D). π0 : C0(R) −→ B(H0) π0 : ϕ (cid:55)−→ ϕ(D0), where D0 = −d2/dx2 and H0 = L2(−∞,∞). Here we view −d2/dx2 is an essen- Define Ut : H0→ H0 to be the translation operator tially self-adjoint operator on the domain of smooth, compactly supported func- tions, and we take D0 to be its self-adjoint extension. We shall compare π to the functional calculus representation (Uth)(x) = h(x−t). W : H0 −→ H Obviously each ϕ(D0) commutes with each Ut. Denote by the orthogonal projection (which restricts functions on (−∞,∞) to functions on (0,∞), of course). The following computation checks that π0 is asymptotically lim lim x→∞ p(x) = 1, contained in π, assuming that the coefficients of D converge to constant values. 3.3. Lemma. Assume that the coefficients of D satisfy and x→∞ p(cid:48)(x) = 0 If ϕ ∈ C0(R), and if g, h ∈ L2(−∞,∞), then t→+∞(cid:104)(cid:10)WUtg, ϕ(D)WUth(cid:11) x→∞ q(x) = 0. L2(−∞,∞) t→+∞(cid:13)(cid:13)ϕ(D)WUth − WUtϕ(D0)h(cid:13)(cid:13)L2(0,∞) = 0 L2(0,∞) −(cid:10)g, ϕ(D0)h(cid:11) Proof. We shall prove that (3.4) for every h ∈ L2(−∞,∞), which will suffice. The set of all ϕ ∈ C0(R) satisfying (3.4) is a norm-closed subalgebra of C0(R), and it therefore suffices to show that the = 0. lim lim lim (cid:105) ON A SPECTRAL THEOREM OF WEYL Let ϕ(x) = (x ± i)−1. We shall calculate the limit (3.4) when resolvent functions ϕ(λ) = (λ±i)−1 belong to it. Moreover it suffices to check (3.4) for each of these two functions ϕ and for a dense set of functions h in L2(−∞,∞). c (−∞,∞), and if t (cid:29) 0, then WUtf is a smooth and compactly supported ∞ function on (0,∞), and we compute that c (−∞,∞). ∞ and f ∈ C h = (D0 ± iI)f If f ∈ C 11 ϕ(D)WUth − WUtϕ(D0)h = (D ± iI)−1WUt(D0 ± iI)f − WUtf smooth and compactly supported functions on (0,∞). Our assumptions on the where, in the last line, we are regarding D0 as a differential operator acting on the = (D ± iI)−1(D − D0)WUtf, coefficients of D imply that t→+∞(cid:107)(D − D0)WUtf(cid:107) = 0, and so (3.4) is proved for ϕ(x) = (x ± i)−1, as required. lim (cid:3) (cid:98)h(ξ) = (cid:90)∞ −∞ h(x)e−iξx dx, Assumptions 2.2 and 2.3 about the representation π0 from the previous section are easily obtained from the Fourier transform as follows. To begin, let Λ0 = (0,∞), and for λ ∈ Λ0 define H0,λ to be the two- λx. dimensional vector space of functions on the line spanned by ei Equip H0,λ with the inner product that makes these two functions an orthonormal basis. The family {H0,λ}λ>0 obviously forms a continuous field of Hilbert spaces over Λ0 with constant and finite fiber dimension. Now let H0 be space of smooth and compactly supported functions in H0. The Fourier transform associates to each h ∈ H0 a continuous section {h0,λ} of the continuous field, namely λx and e−i √ √ h0,λ =(cid:98)h( √ λ)ei √ (cid:90) √ √ λx +(cid:98)h(− (cid:104)h0,λ, g0,λ(cid:105)H0,λ ϕ(λ) dµ0(λ), λ)e−i λx. Moreover it follows from Plancherel's formula that (cid:104)h, ϕ(D0)g(cid:105)L2(−∞,∞) = where (3.5) Λ0 dµ0(λ) = 1 4π dλ√ λ . So Assumption 2.2 is satisfied. The unitary actions √ on the fibers H0,λ decompose the translation action on L2(−∞,∞), as in Assump- Ut,λ : aei λx + be−i λx + ei λtbe−i λtaei √ √ √ √ √ λx λx (cid:55)−→ e−i Let us turn now to the representation π of C0(R). General theory guarantees that π has a measurable direct integral decomposition (cid:90)⊕ L2(0,∞) ∼= Hλ dµ(λ). R tion 2.3. (3.6) 12 NIGEL HIGSON AND QIJUN TAN This means that there exists: Chapter 1]. (i) A Borel-measurable field of Hilbert spaces, {Hλ}λ∈R, as in [Dix81, Part II, (ii) A Borel measure µ on R. (iii) A unitary isomorphism from L2(0,∞) to the Hilbert space of square-inte- grable sections of the measurable field, h (cid:55)→ {hλ}λ∈R, under which the repre- sentation π corresponds to the representation of C0(R) on square-integrable sections by pointwise multiplication. Thus if g ∈ H and if ϕ ∈ C0(R), then for µ-almost every λ ∈ R. (π(ϕ)g)λ = ϕ(λ)gλ See [Dix81, Part II, Chapter 6, Theorem 2]. We need to upgrade this measurable decomposition to a continuous decomposition, as required by Assumption 2.4. We don't know the full extent to which this is possible, but the Gelfand-Kostyuchenko method, which we shall now review, handles the examples of concern to us. (See [Ber88, Section 1] for a concise account of the Gelfand-Kostyuchenko method, as well as applications that are closely related to those in this paper.) The inclusion of the topological vector space C through a Hilbert-Schmidt operator. That is, there is a commuting diagram c (0,∞) into L2(0,∞) factors ∞ L2(0,∞) c (0,∞) ∞ C inclusion (3.7) continuous Hilbert-Schmidt K ελ : C c (0,∞) −→ Hλ ∞ where K is a Hilbert space. This has the following consequence: 3.8. Lemma (See for example [Mau67, Chapter VII, Section 1]). For all λ ∈ R there exist continuous linear operators (3.9) such that if h ∈ C {Hλ}λ∈R, then hλ = ελ(h) for µ-almost every λ ∈ R. c (0,∞), and if {hλ}λ∈R is the associated square-integrable section of ∞ c (0,∞) is dense in the Hilbert space L2(0,∞), the maps ελ have dense ∞ range for µ-almost every λ. The adjoint operators (3.10) are therefore injective for µ-almost every λ. That is, for almost every λ∈R the map λ is defined and embeds H∗ ε∗ Keeping in mind the Hilbert space isomorphism H∗ Lemma 3.8 and the definitions that if h ∈ C section of {Hλ}λ∈R, then c (0,∞)∗ ∞ c (0,∞), and if {vλ}λ∈R is a measurable ∞ λ into the space of distributions on R. ∼= Hλ, it follows from λ −→ C ε∗ λ : H∗ Since C (cid:3) λ (cid:90)∞ 0 (cid:104)vλ, hλ(cid:105)Hλ = (cid:104)vλ, ελ(h)(cid:105)Hλ = λ(vλ) · h, ε∗ for µ-almost every λ ∈ R, where the right-hand integral is the pairing between distributions and test functions. Using this and the propery (iii) above, we find that if Vλ = ε∗ λ(vλ), then (cid:90)∞ (cid:90)∞ DVλ · h = λVλ · h 0 0 / / $ $ ; ; ON A SPECTRAL THEOREM OF WEYL for µ-almost every λ ∈ R (the operator D is applied to Vλ in the sense of distribu- tions) and since C c (0,∞) is separable it follows that ∞ 13 for µ-almost every λ. Thus for almost every λ, the morphism ε∗ the space of λ-eigendistributions for D on (0,∞)). The latter is 2-dimensional and consists of smooth functions on (0,∞). λ embeds Hλ into DVλ = λVλ Let us study the implications of all this for the operators in Example 3.1. 3.11. Lemma. Let D be as in Example 3.1. For µ-almost every λ ∈ R the operator ε∗ λ maps Hλ isomorphically to the one-dimensional space of (smooth) solutions of the differential equation DGλ = λGλ that satisfy the boundary condition Gλ(0) = 0. Proof. We can repeat the Gelfand-Kostyuchenko method above using the space of functions in C this space, then for almost every λ we can write c [0,∞) that vanish at 0 in place of C ∞ c (0,∞). If f and g belong to ∞ (3.12) (cid:104)ελ(Dh), gλ(cid:105)Hλ − (cid:104)ελ(h), (Dg)λ(cid:105)Hλ = (cid:104)ελ(Dh), gλ(cid:105)Hλ − (cid:104)ελ(h), λgλ(cid:105)Hλ (Dh)(x)Gλ(x) dx − h(x)λGλ(x) dx (Dh)(x)Gλ(x) dx − h(x)(DGλ)(x) dx, = = 0 (cid:90)∞ (cid:90)∞ (cid:90)∞ 0 (cid:90)∞ (cid:90)∞ 0 0 where Gλ = ε∗ difference of integrals using the fundamental theorem of calculus we find that λ(gλ). Assume now that in addition h(cid:48)(0) = 1. Calculating the (cid:90)∞ (Dh)(x)Gλ(x) dx − h(x)(DGλ)(x) dx = p(0)Gλ(0). 0 0 The top expression in (3.12) is an integrable function of λ, and therefore so is Gλ(0). If ϕ is any continuous and compactly supported function on R, then by (iii) above the integral of the left-hand side of (3.12), times ϕ(λ), is equal to zero, and so (cid:90)∞ Gλ(0) ϕ(λ) dµ(λ) = 0, 0 It follows that Gλ(0) = 0 for almost every λ. The lemma follows from this because (cid:3) the elements gλ span Hλ, for almost all λ. Now form the family of one-dimensional eigenfunction spaces { Fλ : [0,∞)→C : DFλ = λFλ, Fλ(0) = 0 }. (3.13) These assemble to form the fibers of a smooth vector bundle using the usual topol- ogy of convergence of smooth functions. Equip each with the norm (cid:107)Fλ(cid:107) = F(cid:48) λ(0) to obtain a continuous field of one-dimensional Hilbert spaces over R for which λ (cid:55)→ Fλ is a continuous section if λ (cid:55)→ F(cid:48) Lemma 3.11 shows that for almost every λ the morphism ε∗ λ is a vector space isomorphism from the Hilbert space fiber Hλ in the direct integral decompostion (3.6) to the fiber (3.13) above. The morphism is not necessarily isometric, but we can remedy this possible shortcoming by changing the inner products on the Hλ, and the measure µ, using λ(0) is continuous. (cid:104) , (cid:105)Hλ := (cid:107)vλ(cid:107)−2 Hλ · (cid:104) , (cid:105)Hλ and dµ(λ) := (cid:107)vλ(cid:107)2 Hλ · dµ(λ) NIGEL HIGSON AND QIJUN TAN 14 where vλ is chosen so that if Fλ = ε∗ obtain a new direct integral decomposition of the form (3.6) (the map h (cid:55)→ {hλ} from L2(0,∞) to square-integrable sections is not changed), and now the mor- c (0,∞), and if h ∈ H, then according to the definitions, if ∞ λ are unitary isomorphisms, for almost every λ. λ(0) = 1. With these changes, we λ(vλ) then F(cid:48) Now take H = C phisms ε∗ F(cid:48) λ(0) = 1, then ε∗ λ(hλ) = Fλ(x)h(x) dx · Fλ. (cid:90)∞ 0 for almost every λ. The right hand side is a continuous section of the field {Hλ} since the function Fλ depends continuously (in fact analytically) on λ. This verifies Assumption 2.4 for the Sturm-Liouville operators from Section 1. c (0,∞), and obtain, for almost every λ, embeddings of Hλ ∞ As for the Laplace-Beltrami operator from Example 3.2, we can repeat the Gel- fand-Kostyuchenko analysis, as in Lemma 3.8 and the discussion following the lemma, using the space of smooth, compactly supported, K-invariant functions on G/K in place of C into the K-invariant λ-eigenfunctions of D. But the latter space is actually one- dimensional already (the λ-eigenfunctions are distinguished from one another by their values at eK ∈ G/K) and the family of all such eigenspaces spaces carries the structure of a continuous field of one-dimensional Hilbert spaces, since there are explicit formulas for the eigenfunctions that vary smoothly with λ. See [Hel72, Chap. 2, Thms 1.1 & 1.2]. The argument above then handles Assumption 2.4 in this case. Finally, we need to verify Assumption 2.5. For this purpose we shall assume a (cid:90)∞ bit more about the coefficients of D, namely that 1 − p(x)−1 dx <∞ and (3.14) x0 x0 (cid:90)∞ q(x) dx <∞. Certainly these conditions hold in our examples. 3.15. Proposition. Let λ > 0 and let Fλ be the λ-eigenfunction of D with F(cid:48)(0) = 1. If (3.14) holds, then there is a unique nonzero λ-eigenfunction F0,λ of D0 such that The convergence is uniform over compact sets of eigenvalues λ in (0,∞). lim x→∞(cid:12)(cid:12)Fλ(x) − F0,λ(x)(cid:12)(cid:12) = 0. (cid:3) This is standard in differential equations and we will omit the proof here, but see for example Weyl's paper [Wey10]). Of course Proposition 3.15 is obvious for the operators from Example 3.1. In any case, using Proposition 3.15 we define injective operators by Cλ : Fλ (cid:55)→ F0,λ where Fλ and F0,λ are as in Proposition 3.15. If h is a smooth, compactly supported function on R, and if vλ = Fλ, then Cλ : Hλ −→ H0,λ −(cid:10)vλ, (Wth)λ (cid:11) = (cid:10)Cλvλ, (Uth)0,λ (cid:11) (cid:90)∞ (3.16) axis). Proposition 3.15 implies that if λ is confined to a compact set in (0,∞), then (this formula holds as long as t is large that h(x−t) is supported on the positive x- H0,λ Hλ 0 the integral converges to zero, uniformly in λ, as required by Assumption 2.5. (F0,λ(x) − Fλ(x))h(x−t)dx ON A SPECTRAL THEOREM OF WEYL 15 (cid:90) β We arrive therefore the following result, which is Weyl's theorem for the posi- tive spectrum of D: 3.17. Theorem. Let D be one of the operators from Example 3.1. Let g and h be smooth and compactly supported functions on [0,∞). If 0<α<β, and if P[α,β] is the spectral projection for D associated to the interval [α, β], then (cid:104)g, P[α,β]h(cid:105) = dλ√ λ where Fλ is the nonzero λ-eigenfunction with Fλ(0) = 0 and F(cid:48) λ(0) = 1, and c(λ) is characterized by (cid:104)g, Fλ(cid:105)(cid:104)Fλ, h(cid:105) c(λ)2 1 4π 1 α x→+∞(cid:0)Fλ(x) − c(λ)ei lim √ λx(cid:1) = 0 √ λx − c(λ)e−i (the inner products are standard L2-inner products and the integral is absolutely conver- gent). Proof. We shall compute (cid:107)P[α,β]h(cid:107)2 (the formula in the statement of the theorem will follow by polarization). First, according to the definition of a direct integral decomposition, (cid:90) (cid:107)hλ(cid:107)2 Now let {vλ} be the section of {Hλ} for which ε∗ ment of the theorem. Then (cid:107)P[α,β]h(cid:107)2 = α β dµ(λ). Hλ (cid:90) (cid:90) (cid:107)hλ(cid:107)2 Hλ β α dµ(λ) = (cid:90) and applying Theorem 2.14 we get (cid:107)hλ(cid:107)2 Hλ β α dµ(λ) = (cid:90) β α (cid:12)(cid:12)2 (cid:12)(cid:12)(cid:104)vλ, hλ(cid:105)Hλ (cid:90) (cid:104)vλ, vλ(cid:105)Hλ (cid:90) β α β (cid:104)Fλ, hλ(cid:105)L2 2 (cid:104)vλ, vλ(cid:105)Hλ 2dµ0(λ) Trace(C∗ λCλ) (cid:104)Fλ, hλ(cid:105)L2 2 (cid:104)Cλvλ, Cλvλ(cid:105)H0,λ = 2 α dµ0(λ). λ(vλ) = Fλ, with Fλ as in the state- dµ(λ) = (cid:104)Fλ, hλ(cid:105)L2 2 (cid:104)vλ, vλ(cid:105)Hλ β α dµ(λ), It follows from our definition of Cλ that this is (cid:104)Fλ, hλ(cid:105)L2 2 β c(λ)2 α dµ0(λ), (cid:90) and the theorem follows from the explicit formula for µ0 in (3.5). There is a similar theorem for the operator in Example 3.2. The only change is that Fλ is taken to be the λ-eigenfunction of D on (0,∞) corresponding to the K- equivariant λ eigenfunction G/K→C of the shifted Laplace-Beltrami operator with (cid:3) value 1 at eK. 4. NON-POSITIVE SPECTRUM In this section we shall look at the non-positive part of the spectrum of a Sturm- Liouville operator D of the types considered in the previous section. The methods of this paper really have nothing to contribute here, and for that reason we shall be extremely brief. 16 NIGEL HIGSON AND QIJUN TAN The value λ=0 belongs to the spectrum of D of any of the operators from Sec- tion 1 because the spectrum is closed. But for the purposes of fully determining the measure µ we need to determine whether or not 0 is an eigenvalue of the self- adjoint operator D, or in other words whether or not µ({0}) > 0. The answer is that λ=0 is not an eigenvalue. For the Sturm-Liouville differential operators from Section 1, the λ=0 eigenfunctions have the form F0(x) = c1 + c2x (x (cid:29) 0), and the only possibility for a square-integrable eigenfunction is c1=c2=0, in which case F0 is identically zero. But any eigenfunction for the self-adjoint operator D would in particular be a square-integrable eigenfunction for the differential oper- ator D. For the Laplace-Beltrami operator from Example 3.2 one can employ a similar argument, using a version of Proposition 3.15 in place of the simple asymptotic formula for F0 given above. The negative part of the spectrum for the operators that we discussed in Sec- tion 1 needs to be handled differently, since square-integrable eigenfunctions are certainly possible in this case. But one can prove, using the same methods that go into the proof of Proposition 3.15, that: 4.1. Proposition. If we assume that eαx1 − p(x)−1 dx <∞ and eαxq(x) dx <∞. (cid:90)∞ (cid:90)∞ 1 1 for some α > 0, then the operator D has at most finitely many L2-eigenfunctions satisfying (cid:3) the boundary condition Fλ(0) = 0. One can say more using perturbation theory. The operators D from Section 1 are relatively compact perturbations of the positive operators −d/dx · p(x) · d/dx. So the negative parts of their spectra consist of at most countably many eigen- values, accumulating only at 0. Compare [Kat76, Chapter IV, Theorem 5.35]. But Proposition 4.1 rules out the possibility of accumulation at 0. Hence the negative spectra are finite in this case. that D ≥ 0, so there is no negative spectrum at all. As for the Laplace-Beltrami operator from Example 3.2, it is not difficult to show APPENDIX: REVIEW OF KODAIRA'S APPROACH In this appendix we shall review Weyl's approach to Theorem 1.8 , as improved by Kodaira [Kod49] (see also [Wey50]). Our aim in doing so is to indicate how different it is from the approach taken in the body of this paper. Let D be a self-adjoint Hilbert space operator. If α<β, and if both α and β are absent from the spectrum of D, then according to the Riesz functional calculus, the spectral projection for D associated to the interval (α, β) is (4.2) P(α,β) = (ν − D)−1 dν, where the contour Γ is indicated in Figure 2. The contributions to the integral in (cid:90) 1 2πi Γ ON A SPECTRAL THEOREM OF WEYL 17 FIGURE 2. The contour for the integral in (4.2) (4.2) from the vertical components of the contour Γ decrease to zero in norm as the height of the contour decreases to zero, and so β−iε (ν − D)−1 dν − (ν − D)−1 dν , (cid:90) β+iε α+iε (cid:33) (cid:32)(cid:90) α−iε (cid:90) (4.3) P(α,β) = lim ε→0+ 1 2πi or equivalently (4.4) P(α,β) = lim ε→0+ 1 2πi β α (D−λ−iε)−1 − (D−λ+iε)−1 dλ (these are norm limits). The integrand on the right-hand side of (4.4) is uniformly bounded in ε>0 and in λ ∈ R, and by approximating a general self-adjoint opera- tor D by operators that do not contain α or β in their spectrum, we find that: 4.5. Lemma (Kodaira). The formula (4.4) holds for any self-adjoint operator D and any interval [α, β], as long as as α and β do not belong to the point spectrum of D (the limit (cid:3) in (4.4) is now a strong limit of a uniformly bounded family of operators). The formula (4.4) is of particular value when D is a Sturm-Liouville operator because, as we shall see, the resolvent operators (D−λ±iε)−1 may be computed quite explicitly. Let us consider then D = − + q(x) d dx · p(x) · d dx ther conditions on p and q later on). Assume that D defines a self-adjoint operator where p and q are smooth, real-valued functions on (0,∞) (we shall impose fur- on L2(0,∞) on a given domain that (i) includes the smooth, compactly supported functions on (0,∞) and (ii) is invariant under multiplication by smooth functions on (0,∞) that are locally constant outside of a compact set. with an element of dom(D) near∞. Indeed, if h is any smooth and compactly supported function on (0,∞), then the function If ν /∈ R (or more generally if ν belongs to the resolvent set of D), then there exist nonzero ν-eigenfunctions Fν and Gν that vary smoothly with ν, the first of which agrees with an element in dom(D) near 0 and the second of which agrees f = (D − ν)−1h belongs to dom(D), and moreovoer Df = νf + h. 18 NIGEL HIGSON AND QIJUN TAN It follows that Df=νf near 0 and near∞. Because the set of all (D−ν)−1h is dense in L2(0,∞), we obtain, for at least some h, functions f that are nonzero near 0 and near∞. They agree there with functions in dom(D), and they extend to nonzero ν-eigenfunctions Fν and Gν on (0,∞), as required. Note that eigenfunctions Fν and Gν must be linearly independent, for otherwise they would belong to dom(D), which is impossible if ν /∈ Spec(D). Consider now the integral kernel defined by (4.6) kν(x, y) = smooth, compactly supported functions h on (0,∞), and moreover since The associated integral operator Kν can certainly be defined on the domain of Fν(y)Gν(x) x ≥ y Fν(x)Gν(y) x ≤ y. (cid:14) (cid:90) (cid:90)∞ (Kνh)(x) = Fν(x) Gν(y)h(y) dy + Gν(x) x x 0 Fν(y)h(y) dy. the range consists of smooth functions in dom(D). We compute directly that (4.7) where Wr(Fν, Gν) is the Wronskian (D − ν)Kνh = Wr(Fν, Gν)h, As is well known, this is a constant function of x∈(0,∞); moreover the constant ν(x)Gν(x) − Fν(x)G(cid:48) Wr(Fν, Gν)(x) = p(x)(cid:0)F(cid:48) ν(x)(cid:1). value determines a nondegenerate bilinear form on the 2-dimensional space of ν-eigenfunctions. Since Fν and Gν are linearly independent, we can therefore nor- malize them so that (4.8) in which case it follows from (4.7) that (4.9) for all smooth and compactly supported functions h on (0,∞). (D−ν)−1h = Kνh Wr(Fν, Gν) = 1, In order to apply (4.9) to the limit formula (4.4) we shall make the following additional assumptions concerning the eigenfunctions Gν: (G1) For all λ > 0 the limits λ = lim G+ ε(cid:38)0 Gλ+iε and G− λ = lim Gλ−iε their derivatives on compact sets of (0,∞); note that, using D, this implies exist in the usual C1-topology (uniform convergence of the functions and convergence in the C2-topology, and indeed in the C over the convergence is uniform over compact sets of positive λ. λ are linearly independent. (G2) For all λ > 0 the functions G+ λ and G− -topology). More- ∞ ε(cid:38)0 λ obey the relation The limit functions G± (4.10) for λ > 0. Indeed G+ and as a result, the Wronskian Wr(G+ λ and G− λ ) · Fλ = G+ λ , G− Wr(G+ λ are λ-eigenfunctions for the differential operator D, λ − G− λ λ , G− Hλ = Wr(G+ λ ) is a constant function, so if we write λ , G− λ ) · Fλ , ON A SPECTRAL THEOREM OF WEYL 19 then the three functions Hλ , G+ of λ-eigenfunctions for D. To verify that Hλ = G+ observe that λ and G− λ all belong to the two-dimensional space λ we therefore just need to Wr(Hλ , G± λ ) = Wr(G+ λ −G− λ −G− λ , G± λ ), which is a consequence of (4.8). 4.11. Theorem. If 0<α<β, then under the assumptions (G1) and (G2) above, the spectral projection P(α,β) for D is given by the formula for all smooth and compactly supported functions h on (0,∞), where p(α,β)(x, y)h(y) dy, (P(α,β)h)(x) = 0 p(α,β)(x, y) = Fλ(x)Fλ(y) Wr(G+ λ , G− λ ) dλ. Proof. It follows from Lemma 4.5,(4.9) that (P(α,β)h)(x) = lim ε(cid:38)0 1 2πi (cid:17) (cid:0)kλ+iε(x, y)−kλ−iε(x, y)(cid:1)h(y) dy dλ (cid:90)∞ β (cid:90) (cid:16)(cid:90)∞ α 1 2πi (cid:90) β 0 α (cid:0)kλ+iε(x, y)−kλ−iε(x, y)(cid:1) = Wr(G+ ε(cid:38)0 lim and from (4.6), together with the assumption (G1) and (4.10), that The convergence is uniform over compact subsets of y ∈ (0,∞) and compact sub- sets of λ ∈ (0,∞). Hence λ )Fλ(x)Fλ(y). λ , G− (P(α,β)h)(x) = 1 2πi as required. Wr(G+ λ , G− λ )Fλ(x)Fλ(y) dλ h(y) dy (cid:3) (cid:90)∞ (cid:16)(cid:90) β 0 α (cid:17) At this point we finally turn to Sturm-Liouville operators with eventually con- stant coefficient functions, as in Section 1. If we write √ νx) + c(−ν) exp(−i Fν(x) = c(ν) exp(i √ νx) (x (cid:29) 0) using the usual principal branch of the square root function, equal to the positive square root on the positive axis (we shall avoid the eigenvalues ν < 0), then using (4.8) we compute that for λ > 0 and ν = λ±iε, √ exp(±i (x (cid:29) 0). Gν(x) = (4.12) νx) √ i 2c(∓ν) ν The sign in the exponential is needed to ensure that Gν is an L2-function at infinity, which is of course necessary if Gν is to agree with a function in dom(D) at infinity. It follows easily from (4.12) that (G1) and (G2) are satisfied, that (x (cid:29) 0), √ exp(±i G± λ (x) = λx) √ ±i 2c(∓λ) λ and that Wr(G+ λ , G− λ ) = i 2c(λ)2 √ . λ 20 NIGEL HIGSON AND QIJUN TAN (cid:90) Therefore Theorem 4.11 gives p(α,β)(x, y) = 1 4π β α Fλ(x)Fλ(y) 1 c(λ)2 dλ√ λ . This is a reformulation of Weyl's theorem, as stated in Section 1. REFERENCES [Ban08] E. van den Ban. Weyl, eigenfunction expansions and harmonic analysis on non-compact sym- metric spaces. In Groups and analysis, volume 354 of London Math. Soc. Lecture Note Ser., pages 24 -- 62. Cambridge Univ. Press, Cambridge, 2008. J. N. Bernstein. On the support of Plancherel measure. J. Geom. Phys., 5(4):663 -- 710 (1989), 1988. [Ber88] [CH16] T. Crisp and N. Higson. A second adjoint [Bor01] A. Borel. Essays in the history of Lie groups and algebraic groups, volume 21 of History of Mathe- matics. American Mathematical Society, Providence, RI; London Mathematical Society, Cam- bridge, 2001. [CCH16] P. Clare, T. Crisp, and N. Higson. Parabolic induction and restriction via C∗-algebras and Hilbert C∗-modules. Compos. Math., 152(6):1286 -- 1318, 2016. theorem for SL(2, R). Preprint, 2016. arXiv:1603.08797. J. Dixmier. C∗-algebras. North-Holland Publishing Co., Amsterdam-New York-Oxford, 1977. Translated from the French by Francis Jellett, North-Holland Mathematical Library, Vol. 15. J. Dixmier. von Neumann algebras, volume 27 of North-Holland Mathematical Library. North- Holland Publishing Co., Amsterdam-New York, 1981. With a preface by E. C. Lance, Trans- lated from the second French edition by F. Jellett. S. Helgason. Analysis on Lie groups and homogeneous spaces. American Mathematical Society, Providence, R.I., 1972. Conference Board of the Mathematical Sciences Regional Conference Series in Mathematics, No. 14. [Hel72] [Dix77] [Dix81] [Kat76] T. Kato. Perturbation theory for linear operators. Springer-Verlag, Berlin-New York, second edi- tion, 1976. Grundlehren der Mathematischen Wissenschaften, Band 132. [Kod49] K. Kodaira. The eigenvalue problem for ordinary differential equations of the second order and Heisenberg's theory of S-matrices. Amer. J. Math., 71:921 -- 945, 1949. [Mau67] K. Maurin. Methods of Hilbert spaces. Translated from the Polish by Andrzej Alexiewicz and Waclaw Zawadowski. Monografie Matematyczne, Tom 45. Pa ´nstwowe Wydawnictwo Naukowe, Warsaw, 1967. Y. Sakellaridis and A. Venkatesh. Periods and harmonic analysis on spherical varieties. Ast´erisque, (396):viii+360, 2017. [SV17] [Tan19] Q. Tan. Asymptotically contained representations and the spherical Plancherel formula. PhD thesis, Penn State University, 2019. [Wey10] H. Weyl. Uber gewohnliche Differentialgleichungen mit Singularitaten und die zugehorigen Entwicklungen willk urlicher Funktionen. Math. Ann., 68(2):220 -- 269, 1910. [Wey50] H. Weyl. Ramifications, old and new, of the eigenvalue problem. Bull. Amer. Math. Soc., 56:115 -- 139, 1950. Department of Mathematics, Penn State University, University Park, PA 16802. Email: [email protected] and [email protected].
1605.03543	2	1605	2016-10-04T14:54:06	C*-envelopes of tensor algebras arising from stochastic matrices	[ "math.OA", "math.FA" ]	In this paper we study the C-envelope of the (non-self-adjoint) tensor algebra associated via subproduct systems to a finite irreducible stochastic matrix $P$. Firstly, we identify the boundary representations of the tensor algebra inside the Toeplitz algebra, also known as its non-commutative Choquet boundary. As an application, we provide examples of C-envelopes that are not -isomorphic to either the Toeplitz algebra or the Cuntz-Pimsner algebra. This characterization required a new proof for the fact that the Cuntz-Pimsner algebra associated to $P$ is isomorphic to $C(\mathbb{T}, M_d(\mathbb{C}))$, filling a gap in a previous paper. We then proceed to classify the C-envelopes of tensor algebras of stochastic matrices up to -isomorphism and stable isomorphism, in terms of the underlying matrices. This is accomplished by determining the K-theory of these C-algebras and by combining this information with results due to Paschke and Salinas in extension theory. This classification is applied to provide a clearer picture of the various C*-envelopes that can land between the Toeplitz and the Cuntz-Pimsner algebras.	math.OA	math	C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES ADAM DOR-ON AND DANIEL MARKIEWICZ Abstract. In this paper we study the C-envelope of the (non-self-adjoint) tensor algebra associated via subproduct systems to a finite irreducible stochastic matrix P . Firstly, we identify the boundary representations of the tensor algebra inside the Toeplitz algebra, also known as its non-commutative Choquet boundary. As an application, we provide examples of C-envelopes that are not -isomorphic to either the Toeplitz algebra or the Cuntz- Pimsner algebra. This characterization required a new proof for the fact that the Cuntz-Pimsner algebra associated to P is isomorphic to C(T, Md(C)), filling a gap in a previous paper. We then proceed to classify the C-envelopes of tensor algebras of stochastic matrices up to - isomorphism and stable isomorphism, in terms of the underlying matrices. This is accomplished by determining the K-theory of these C-algebras and by combining this information with results due to Paschke and Salinas in extension theory. This classification is applied to provide a clearer picture of the various C-envelopes that can land between the Toeplitz and the Cuntz-Pimsner algebras. 6 1 0 2 t c O 4 ] . A O h t a m [ 2 v 3 4 5 3 0 . 5 0 6 1 : v i X r a Introduction Given a C-correspondence E, the operator algebras associated to shift operators (also called creation operators) over the Fock correspondence F(E) have been the subject of considerable attention by too many researchers to appropriately list here. By an operator algebra in this paper we mean a (not necessarily self-adjoint) closed unital subalgebra A of a unital C∗-algebra B. The operator algebra generated by the shifts in L(F(E)) is called the tensor algebra T+(E), It is closely and it provides a very successful prototype for the study of operator algebras. related to the Toeplitz algebra T (E), which is the C-algebra generated by the shifts, and its celebrated quotient, the Cuntz-Pimsner algebra O(E). Analogously, given a a subproduct system X in the sense of Shalit and Solel [SS09] of C- correspondences over a C-algebra A parametrized by N, one obtains the operator algebras associated to shifts on F(X): the tensor algebra T+(X), the Toeplitz algebra T (X) and the Cuntz-Pimsner algebra O(X), where the latter was defined in [Vis12]. This new framework generalizes the previous one, in the sense that a C-correspondence E gives rise to a product system X whose Fock correspondence and associated operator algebras are precisely the ones discussed in the previous paragraph. There has been important work on the operator algebras arising from subproduct systems over C, or equivalently, the special case of subproduct systems whose C-correspondence fibers are actually Hilbert spaces, see for example [SS09, DRS11, KS15]. In our previous paper [DOM14], we turned to the simplest case for which the fibers of the subproduct system are not Hilbert spaces. Namely, we considered the case of subproduct systems of C-correspondences over ℓ∞(Ω) when Ω is countable with more than one point. Such a subproduct system and its associated operator algebras are conveniently parametrized by a stochastic matrix P over the state space Ω. In [DOM14], we considered isomorphism problems of the tensor algebras associated to stochastic matrices, via these subproduct systems. Date: September 4, 2018. 2000 Mathematics Subject Classification. Primary: 47L30, 46L55, 46L35. Secondary: 46L80, 60J10. Key words and phrases. C-envelope; boundary representations; classification; Cuntz-Pimsner algebra; sto- chastic matrix. This work was partially supported by the ISF with the ISF-UGC joint research program framework (grant No. 1775/14). The work of the first author was also partially supported by an Ontario Trillium Scholarship. 1 2 ADAM DOR-ON AND DANIEL MARKIEWICZ In this paper we focus on the C-envelope of the tensor algebra associated to finite irreducible stochastic matrices. Recall that given an operator algebra A, a C-cover is a pair (C, ι) where C is a C-algebra and ι : A → C is a unital completely isometric homomorphism, such that C∗(ι(A)) = C. A C-cover is called the C-envelope for A if for any other C-cover (C′, ι′), the map ι′(a) 7→ ι(a) extends uniquely to a surjective -homomorphism C′ → C. In this precise sense, the C-envelope is the smallest C-algebra which contains a completely isometric copy of A, and usually the algebra C is denoted C∗env(A) and the map ι is suppressed. The existence of the C-envelope of an operator algebra was first proven by Hamana [Ham79], by way of proving the existence of an injective envelope for operator systems. An alternative proof via dilation theory was found by Dritchel and McCullough in [DM05]. This new idea allowed Arveson [Arv08] to follow the original strategy he envisioned in [Arv69, Arv72] to prove the existence of the C-envelope via boundary representations in the separable case. Davidson and Kennedy finally realized Arveson's vision in full in [DK15] by providing a proof without the assumption of separability. We are motivated in this paper by the known results in the determination of the C-envelope of the tensor algebra of a subproduct system: (1) Given a C-correspondence E, we have that C∗env(T+(E)) = O(E). This was first proven by Muhly and Solel [MS98, Corollary 6.6] when the left action on E is faithful, essential and acts by compacts, and in the general case (without extra assumptions) by Katsoulis and Kribs [KK06]. (2) Let I be a homogeneous ideal in the ring of polynomials in finitely many commuting vari- ables. The universal commuting row contraction subject to the polynomial relations in I gives rise to a subproduct system of Hilbert spaces X I , and it was shown by Davdison, Ramsey and Shalit [DRS11] that C∗env(T+(X I )) = T (X I ). (3) Let I be a monomial ideal in the ring of polynomials in non-commuting variables. Simi- larly to the commutative case, a subproduct system X I can be defined. Kakariadis and Shalit [KS15] have shown that for many cases (depending on the monomial ideal) either C∗env(T+(X I )) = T (X I ) or C∗env(T+(X I )) = O(X I ). In summary, for all these cases, when the subproduct system X is irreducible in the appro- priate sense (i.e. no nontrivial reducing projections, see Definition 1.14) , C∗env(T+(X)) has been found to be isomorphic either to T (X) or O(X). In [Vis12, Section 6], this phenomenon was observed, and it was asked if one can find a general description for the behavior of C-envelopes of tensor algebras associated with subproduct systems. In this paper we shed some light on this question: we show that the evidence for the dichotomy witnessed above is misleading, and that the situation is more mysterious than previously thought, by providing an example of stochastic matrix with subproduct system X such that the C∗env(T+(X)) is not -isomorphic to either T (X) or O(X) (See Example 3.18). Our first main result is as follows. Let P be a finite irreducible stochastic matrix. In this case we show that the C-envelope lands between the Toeplitz and Cuntz-Pimsner algebras in the sense that it fits in the following sequence of quotient maps: (∗) and in fact T+(P ) is hyperrigid inside C∗env(T+(P )) (See [Arv11]). Moreover, in the case when P has the multiple arrival property (see Definition 3.13), we identify the boundary representations of T+(P ) inside T (P ), also known as the non-commutative Choquet boundary in the sense of Arveson [Arv08]. This enables us to describe the C-envelope C∗env(T+(P )) in terms of boundary representations. For details see Corollary 3.14 and Theorem 3.15. The fact that the Cuntz-Pimsner algebra O(X) as defined by Viselter [Vis12] is not always isomorphic to the C-envelope of the tensor algebra T+(X) in the subproduct system case, and even a dichotomy as above fails to hold, suggests that perhaps an alternative definition of Cuntz-Pimsner algebra for subproduct systems is desirable. T (P ) −→ C∗env(T+(P )) πP−→ O(P ) The concrete description of the C-envelope and lack of dichotomy lead to an unexpected richness of possibilities. Our second main result concerns the classification of C-envelope up to -isomorphism and stable isomorphism, so as to clarify the situation. For a finite irreducible C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 3 stochastic matrix P over ΩP , the ideal Ker(πP ) in the sequence (∗) is -isomorphic to a direct sum of nP ≤ ΩP copies of the algebra of compact operators. Given two irreducible stochastic matrices P and Q over finite state sets ΩP and ΩQ we have that details see Theorem 5.5. (1) T+(P ) and T+(Q) have stably isomorphic C-envelopes if and only if nP = nQ. For more (2) T+(P ) and T+(Q) have -isomorphic C-envelopes if and only if ΩP = ΩQ, nP = nQ and up to a reordering of ΩQ, the matrices P and Q have the same column nullity in every column. For more details, see Definition 5.3 and Theorem 5.6. Therefore, we see that instead of a dichotomy, we actually have a profusion of possibilities. These results are obtained by leveraging the surprisingly simple form of the Cuntz-Pimsner algebra as obtained in [DOM14, Corollary 5.16] to compute the K-theory of C∗env(T+(P )). How- ever, in our case K-theory does not completely resolve the issue by itself, and we then use extension theory (especially work by Paschke and Salinas [PS79]) to complete the task. We should note that Dilian Yang pointed out to us that there was a gap in the proof of [DOM14, Corollary 5.16], which we resolve in Section 2 of this paper. Finally, its natural to ask about the relationship between C∗env(T+(P )) and the graph algebra OGr(P ) associated to the unweighted directed graph obtained from a finite irreducible stochastic matrix P . We apply our classification results for the C-envelope and K-theory for graph algebras to show that these two objects are generally incomparable in the sense that we exhibit 3 × 3 irreducible stochastic matrices P , Q and R such that and C∗env(T+(P )) 6∼ C∗env(T+(Q)) ∼= C∗env(T+(R)) OGr(P ) ∼= OGr(Q) 6∼ OGr(R) where ∼= stands for -isomorphism and ∼ stands for stable isomorphism. This paper has six sections. In Section 1 we give some preliminary background required from [DOM14] and on extension theory. In Section 2 we fill the gap pointed out to us by Dilian Yang in the proof of [DOM14, Corollary 5.16] and compute the extension groups for the Cuntz-Pimsner algebra of a finite irreducible stochastic matrix. In Section 3 we determine the non-commutative Choquet boundary of T+(P ) inside T (P ), which then allows us to compute C-envelopes C∗env(T+(P )) associated to finite irreducible stochastic matrices. In Section 4 we compute the K-theory of C∗env(T+(P )) in terms of boundary representations. Finally, in Section 5 we prove our main classification results mentioned above, and compare C∗env(T+(P )) and OGr(P ) as invariants for the graph of P . 1. Preliminaries Boundary representations and Shilov ideal. Suppose that A is a unital operator algebra, and (B, ι) is a C-cover. A unital completely contractive (c.c.) map φ : ι(A) → B(H) has the unique extension property if there exists a unique unital completely positive (c.p.) extension eφ : B → B(H) which is also a -representation. We will say that a unital -representation ρ : B → B(H) is a boundary representation for A if ρ is irreducible and ρ ↾ι(A) has the unique extension property. For a unital c.c. map φ : A → B(H), we say that a unital c.c. map φ′ : A → B(H′) is a dilation of φ if there is an isometry V : H → H′ such that for any a ∈ A we have φ(a) = V ∗φ′(a)V . We will call a unital c.c. map φ : ι(A) → B(H) maximal if whenever φ′ is a dilation of φ, then φ′ = φ ⊕ ψ for some unital c.c. map ψ. It turns out that for a unital c.c. map φ : ι(A) → B(H) we have that φ is maximal if and only if φ has the unique extension property [Arv06, Proposition 2.2], and that maximality is invariant under change of C-cover [Arv06, Proposition 3.1]. Thus, for the definitions of the unique extension property and maximality, it makes no differ- ence which C-cover (B, ι) we work inside. Next, for a unital operator algebra A and (B, ι) a C-cover, an ideal I of B is called a boundary ideal for A if the canonical quotient map qI : B → B/I is completely isometric on ι(A). The Shilov ideal SA of A in B is the largest boundary ideal. The Shilov ideal is a tractable tool for finding the C-envelope, since B/SA must then be the C-envelope of A (See [Kak13, Proposition 1.9]). However, there is a way to compute the 4 ADAM DOR-ON AND DANIEL MARKIEWICZ C-envelope from boundary representations. By the theorem of Davidson and Kennedy from [DK15], the boundary representations of every unital operator systems completely norm it, so that by [Arv69, Theorem 2.2.3] we have that the Shilov ideal is the intersection of all kernels of boundary representations. In [Arv11] Arveson investigated a closely related notion for C-covers called hyperrigidity. For a unital operator algebra A and (B, ι) a C-cover for it, one of the equivalent formulations for hyperrigidity of (B, ι) is that for any ∗-representation π : B → B(H), the restriction πι(A) has the unique extension property. Suppose now that A is hyperrigid in (B, ι). Then any irreducible -representation of B must be a boundary representation with respect to ι(A), so by taking the direct sum ρ of all irreducible representations of B, by the above we have that the Shilov ideal of ι(A) in B is trivial. This means that the C-envelope of A is B. Hence, when we know that a C-cover is hyperrigid, this C-cover must then be the C-envelope for our operator algebra. By invariance of C-envelope, we see that up to -isomorphism, A can only be hyperrigid in at most one C-algebra. A as a subalgebra of some particular B(H). Hilbert modules and subproduct systems. We assume that the reader is familiar with the basic theory of Hilbert C-modules, which can be found in [Pas73, Lan95, MT05]. We only give a quick summary of basic notions and terminology as we go, so as to clarify our conventions. We will often suppress the notation for the C-cover that we use, and in many cases think of Let A be a C-algebra and E a Hilbert C-module over A. We denote by L(E) the collection If in addition E has a left A-module structure given by a - of adjointable operators on E. homomorphism φ : A → L(E), we call E a Hilbert C-correspondence over A. We often suppress notation and write a · ξ := φ(a)ξ. If E is a C-correspondence over A with left action φ, and F is a C-correspondence over A with left action ψ, then on the algebraic tensor product E ⊗alg F one defines an A-valued pre-inner product satisfying hx1 ⊗ y1, x2 ⊗ y2i = hy1, ψ(hx1, x2i)y2i on simple tensors. The usual completion process with respect to the norm induced by this inner product, yields the internal Hilbert C-module tensor product of E and F , denoted by E ⊗ F or E ⊗ψ F , which is a C-correspondence over A with left action φ ⊗ IdF . The following is the C-algebraic version of [SS09, Definition 1.1] for the semigroup N, which was also given in [Vis11, Definition 1.4]. Definition 1.1. Let A be a C-algebra, let {Xn}n∈N be a family of Hilbert C-correspondences over A and let U = {Un,m : Xn ⊗ Xm → Xn+m} be a family of bounded bimodule maps. We will say that (X, U ) is a subproduct system over A if the following conditions are met: (1) X0 = A (2) The maps U0,n and Un,0 are given by the left and right actions of A on Xn respectively (3) Un,m is an adjointable coisometric map for every n, m ∈ N (4) For every n, m ∈ N we have the associativity identity Un+m,p(Un,m ⊗ IdXp) = Un,m+p(IdXn ⊗ Um,p) In case the maps Un,m are unitaries, we say that X is a product system. Operator algebras associated to subproduct systems. We describe the construction of the tensor, Toeplitz and Cuntz-Pimsner algebras arising from subproduct systems (see [Vis11, Vis12]). Let (X, U ) be a subproduct system over a C-algebra A. There is a canonical product system containing (X, U ) as a subproduct subsystem as follows (See [SS09, Definition 5.1 & Proposition 5.2]). We define E := X1, so that Prod(X) := {E⊗n}n∈N constitutes a product systems where the unitaries from E⊗n ⊗ E⊗m to E⊗n+m are the usual associativity unitaries. One can then construct canonical adjointable coisometries Vn : En → Xn which, by asso- ciativity of U = {Un,m}, are uniquely determined inductively by the equations V1 = IdX1 and Vn+m = Un,m ◦ (Vn ⊗ Vm). C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 5 The X-Fock correspondence is the C - correspondence direct sum of the fibers of the sub- product system (1.1) FX :=Mn∈N Xn The X-shifts are the operators S(n) ξ ∈ L(FX ) uniquely determined between fibers by S(n) ξ (η) := Denote by Qn ∈ L(FX ) the projection of FX onto the n-th fiber Xn, and define Q[0,n] = Q0 + ... + Qn, and Q[n,∞) := IdFX − Q[0,n−1]. We then obtain an adjointable coisometric map V : FProd(X) → FX given by V = ⊕∞n=0Vn V ∗ Un,m(ξ ⊗ η) where n, m ∈ N and ξ ∈ Xn, η ∈ Xm. n (ξ)V ∗, so that S(n) ξ = V S(n) We note that S(n) n (ξ)V ∗, where S(n) = n (ξ) is a product system shift and is hence an adjointable operator in V S(n)∗V ∗ L(FProd(X)). Definition 1.2. The tensor and Toeplitz algebras are the non-self-adjoint and self-adjoint sub- algebras of L(FX ) generated by a copy of A and all X-shifts respectively, is adjointable with adjoint given by S(n)∗ V ∗ ξ ξ T+(X) := Alg(A ∪ {S(n) T (X) := C∗(A ∪ {S(n) ξ ξ ξ ∈ Xn, n ∈ N}) ξ ∈ Xn, n ∈ N}) Remark 1.3. When a subproduct system is comprised of W ∗-correspondences, since each S(n) is adjointable, the last part of [DOM14, Proposition 2.14] allows us to take the W ∗-correspondence (weak) direct sum of fibers as our Fock space in equation (1.1), and get that the operator algebras T (X) and T+(X) are the same as those considered in [DOM14, Definition 4.1 & Definition 6.1]. The algebra L(FX ) admits a natural action α of the unit circle T called the gauge action, defined by αλ(T ) = WλT W ∗λ for all λ ∈ T where Wλ : FX → FX is the unitary defined by ξ Wλ(⊕∞n=0ξn) = ⊕∞n=0λnξn Since αλ(S(n) ) = S(n) ξ subalgebras, and so the action restricts to them, and we shall still denote it by α. λnξ , we see that the algebras T+(X) and T (X) are α-invariant closed The circle action on T (X) then enables the definition of a faithful conditional expectation Φ given by Φ(S) =RT αλ(T )dλ where dλ is normalized Haar measure on T. One then defines T (X)k to be the closure of all homogeneous polynomials of degree k (see [DOM14, Definition 4.5]), which then coincides with the collection of operators T ∈ T (X) satisfying αλ(T ) = λkT as shown in [DOM14, Corollary 4.6]. This makes both T (X) and T+(X) into Z-graded and N-graded algebras respectively, and Φ on T (X) and T+(X) is then onto T (X)0 and A respectively. Another algebra associated to the subproduct system arises as a special quotient of T (X). The subset J ⊂ L(FX ) given by J = { T ∈ L(FX ) lim n→∞kT Qnk = 0 } is a closed α-invariant left ideal inside L(FX) according to [DOM14, Proposition 4.8]. It was proven by Viselter in [Vis12, Theorem 2.5] that J (T (X)) := J ∩ T (X) is a closed two sided ideal. Definition 1.4. Let (X, U ) be a subproduct system. Define the Cuntz-Pimsner ideal of T (X) to be J (T (X)) := J ∩T (X), and the Cuntz-Pimsner algebra of X is then O(X) := T (X)/J (T (X)). We note that the circle action on T (X) passes naturally to O(X) since J (T (X)) is gauge invariant, and the fixed point algebras are then T (X)0 and O(X)0 respectively. We shall later need the following formula for the norm of an element in the quotient Ms(O(X)), in terms of representatives in Ms(T (X)). Denote by q : T (X) → O(X) the canonical quotient map. When Qn ∈ T (X), it follows from item (1) of [Vis12, Theorem 3.1] that {Is · Q[0,m]} is an 6 ADAM DOR-ON AND DANIEL MARKIEWICZ approximate identity for Ms(J (T (X))), one may then invoke [Arv76, Exercise 1.8.C] to obtain the following. Proposition 1.5. Let (X, U ) be a subproduct system, and suppose that Qn ∈ T (X) for all n ∈ N. Then for any T = [Tij] ∈ Ms(T (X)) we have kq(s)(T )k = lim m→∞k[Tij Q[m,∞)]k Cuntz-Pimsner algebras and subproduct systems arising from stochastic matrices. In our previous paper [DOM14] we studied the tensor algebra T+(P ) associated to a certain subproduct system construction applied to a stochastic matrix P . This subproduct system construction can be applied to any unital normal completely positive map on a von-Neumann algebra, and is called the Arveson-Stinespring subproduct system construction. After characterizing isomorphism classes for Arveson-Stinespring subproduct systems in terms of the underlying stochastic matrices, we used this characterization to study the dependence of the isomorphism classes of the algebra T+(P ) on the matrix P (with respect to various concepts of isomorphism), which ended up coinciding with the respective isomorphism classes for the subproduct systems. We will now discuss some of the preliminaries and results in [DOM14] for such subproduct systems and their Cuntz-Pimsner algebras. For the basic theory of stochastic matrices and Markov chains, we recommend [Sen06] and [Dur10, Chapter 6]. Definition 1.6. Let Ω be a countable set. A stochastic matrix is a function P : Ω × Ω → R+ such that for all i ∈ Ω we havePj∈Ω Pij = 1. Elements of Ω are called states of P . To every stochastic matrix, one can associate a set of edges E(P ) := { (i, j) Pij > 0 } and a {0, 1} - matrix Gr(P ) representing the directed graph of P as an incidence matrix by way of Gr(P )ij =(cid:26) 1 : Pij > 0 0 : Pij = 0 Many dynamical properties of P can be put in terms of the directed graph (Ω, E(P )) of P . Definition 1.7. Let P be a stochastic matrix over Ω. A path of length ℓ in (Ω, E(P )), i.e. a path in the directed graph of P , is a function γ : {0, ..., ℓ} → Ω such that Pγ(k)γ(k+1) > 0 for every 0 ≤ k ≤ ℓ − 1. The path γ is said to be a cycle if γ(0) = γ(ℓ). We will say that a state i leads to a state j if there is a path γ (of some length ℓ) as above with γ(0) = i and γ(ℓ) = j. We next give the main definitions that we shall use in the context of stochastic matrices in this paper. Definition 1.8. Let P be a stochastic matrix over Ω, and let i ∈ Ω. (1) The period of i is r(i) = gcd{ n P (n) ii > 0 }. If no finite such r(i) exists, or if r(i) = 1 (2) P is said to be irreducible if for any pair i, j ∈ Ω, we have that i leads to j (and so j we say that i is aperiodic. also leads to i). r, so we define the periodicity of P to be r. If P is an irreducible stochastic matrix over Ω, then every state i ∈ Ω is of the same periodicity Let us recall the statement of the cyclic decomposition of irreducible stochastic matrices [Sen06, Theorem 1.3] which justifies the notion of periodicity of an irreducible stochastic matrix P . Theorem 1.9. (Cyclic decomposition for periodic irreducible matrices) Let P be an irreducible stochastic matrix over a state set Ω with period r, and let ω ∈ Ω. For each ℓ = 0, . . . r − 1, let Ωℓ = {j ∈ Ω P (n) ωj > 0 =⇒ n ≡ ℓ mod r}. Then, (1) The family (Ωℓ)r−1 (2) If j ∈ Ωℓ then there exists N (j) such that for all n ≥ N (j) we have P (nr+ℓ) ℓ=0 is a partition of Ω. ωj > 0.  0 P0 ... ... ··· Pr−1 ··· 0 0 ··· ... 0 Pr−2 0 ... 0  where the rows (columns) of Pℓ in this matrix decomposition are indexed by Ωℓ (Ωℓ+1 respectively) for all for ℓ ∈ Zr, where Zr is the cyclic group of order r. In this paper we shall restrict our attention to finite irreducible stochastic matrices. For this class of stochastic matrices, we have that the more generally stated [DOM14, Theorem 2.10], yields the following cleaner formulation, which is a combination of [Sen06, Theorem 4.1] and [Dur10, Theorem 6.7.2]. Theorem 1.10. (Convergence theorem for finite irreducible matrices) Let P be a finite irreducible stochastic matrix over Ω with period r ≥ 1, and Ω0, ..., Ωr−1 a cyclic decomposition for it as in item (3) of Theorem 1.9. There exists a unique probability vector ν = (νi)i∈Ω so that when we are given i ∈ Ωl1 and j ∈ Ωl2, for 0 ≤ ℓ < r such that ℓ ≡ (l2 − l1) mod r, we have that P (mr+ℓ) ij = νjr. lim m→∞ Let Ω be a finite set and ℓ∞(Ω) = C(Ω) = CΩ the C-algebra of finite sequences indexed by Ω. We denote by {pj}j∈Ω the collection of pairwise perpendicular projections on ℓ∞(Ω) given by pj(i) = δij. Notation 1.11. We denote by ∗ the Schur (entrywise) multiplication of matrices A = [aij] and B = [bkl] given by A ∗ B = [aijbij], and let Diag be the map on matrices given by Diag([aij ]) = (aii)i∈Ω ∈ ℓ∞(Ω). Next, for a non-negative matrix P = [Pij] indexed by Ω, we denote by √P and P ♭ the matrices with (i, j)-th entry given by C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 7 (3) Up to re-enumeration of Ω, there exist rectangular stochastic matrices P0, ...Pr−1 such that P has the following cyclic block decomposition: (√P )ij :=pPij, and (P ♭)ij :=((Pij )−1, 0, if Pij > 0 else In [DOM14, Theorem 3.4] the Arveson-Stinespring subproduct system associated to a sto- chastic matrix P on countable Ω was computed. When Ω is finite, we arrive at the following simpler version of the theorem. Theorem 1.12. Let P be a stochastic matrix over finite Ω. The following is a subproduct system Arv(P ) over C(Ω) ∼= ℓ∞(Ω) and is the one given in [DOM14, Theorem 3.4]. (1) The n-th fiber is a C∗-correspondence over C(Ω) given by Arv(P )n := { [aij] aij = 0 if (i, j) /∈ E(P n) } with left and right actions of C(Ω) on Arv(P )n as a bimodule are given by diagonal left and right matrix multiplication and the C(Ω)-valued inner product is given by hA, Bi = Diag(cid:2)A∗B(cid:3) for A, B ∈ Arv(P )n. (2) The subproduct maps are given by Un,m(A ⊗ B) = (√P n+m)♭ ∗(cid:2)(√P n ∗ A) · (√P m ∗ B)(cid:3) for n, m 6= 0 and A ∈ Arv(P )n and B ∈ Arv(P )m. Remark 1.13. Since the subproduct systems we shall consider in this work will be with fi- nite dimensional fibers and over finite dimensional C∗-algebras, they will automatically be W ∗- correspondences. Hence, by Remark 1.3, the theories of subproduct systems over C∗-algebras and their operator algebras discussed here and of subproduct systems over W ∗-algebras and their operator algebras discussed in [DOM14] will coincide. In this paper we choose to discuss our theories only in the C* (norm closed) context for the sake of brevity and a cleaner exposition. 8 ADAM DOR-ON AND DANIEL MARKIEWICZ Remark 1.14. Let (X, U ) be a subproduct system of a C-algebra A. A projection p ∈ A is said to be reducing for X if U∗n,m(pXn+mp) ⊂ pXnp ⊗ pXmp This is the C / norm-closed version of [DOM14, Definition 6.19]. Using [DOM14, Proposi- tion 7.4] we characterized the reducing projections of Arv(P ) for any stochastic matrix P over Ω. That is, there is a 1-1 bijection between reducing projections and subsets Cp ⊂ Ω such that whenever γ : {0, ..., ℓ} → Ω is a path in the directed graph of P that both begins and ends at Cp, then in fact for every 0 ≤ k ≤ ℓ one has that γ(k) ∈ Cp. Hence, for a finite irreducible stochastic matrix P over Ω, the only reducing non-zero pro- considering irreducible subproduct systems in the above dynamical sense. jection is 1 =Pi∈Ω pi ∈ ℓ∞(Ω). Hence, by considering irreducible stochastic matrices, we are For A ∈ Arv(P )n we defined in [DOM14] the shift operator S(n) A uniquely determined on A (B) = Un,m(A⊗ B) for B ∈ Arv(P )m. The Toeplitz and tensor algebras of Arv(P ) fibers by S(n) are given respectively by T (P ) = C∗(cid:16)ℓ∞(Ω) ∪ { S(n) T+(P ) = Alg(cid:16)ℓ∞(Ω) ∪ { S(n) A n ∈ N, A ∈ Arv(P )n }(cid:17) A n ∈ N, A ∈ Arv(P )n }(cid:17) We note in passing that T (P ) and T+(P ) are generated (as a C-algebra, and as a norm-closed algebra respectively) by {pi}i∈Ω and {SEij}(i,j)∈E(P ), where Eij = [δij(k, l)] is the zero matrix, except for the (i, j) entry at which it is 1. Indeed, since P is a finite matrix, each S(n) A can be written as a finite linear combination of S(n) with (i, j) ∈ E(P n). Then choose a path of length Eij n, say i = j0 → j1 → ... → jn = j, and we have that S(n) for some Ej0j1 · ... · S(1) = c · S(1) Ejn−1jn Eij c > 0. Next, for a finite stochastic matrix P over Ω, and for every n ∈ N and A ∈ Arv(P )n we defined operators in L(FArv(P )) mapping each Arv(P )m to Arv(P )n+m, one denoted by T (n) and given by T (n) A which is uniquely determined on fibers Arv(P )m by W (n) A (B) = A · B. For the purposes of computing the Cuntz-Pimsner algebra, we defined in [DOM14, Section 5] the auxiliary C∗-algebra , and the other denoted by W (n) A = S(n) (√P n)♭∗A A and we noted that due to finiteness of P we have that T ∞(P ) := C∗(cid:16)ℓ∞(Ω) ∪(cid:8) W (n) T (P ) = C∗(cid:16)ℓ∞(Ω) ∪(cid:8) T (n) A n ∈ N, A ∈ Arv(P )n (cid:9)(cid:17) A n ∈ N, A ∈ Arv(P )n (cid:9)(cid:17) [DOM14, Proposition 5.6] was then used to show that in fact O(P ) is -isomorphic to T ∞(P )/J (T ∞(P )), thereby reducing the computation of the Cuntz-Pimsner algebra to com- puting a quotient of an algebra generated by operators W (n) A which do not depend on weights of entries of P . Extension theory. We recall some facts from the theory of primitive ideal spectra and exten- sion theory for C-algebras, to be used in later sections. More details on primitive ideal spectra of C-algebras can be found in [Dix77, Chapter 3] and [Arv76, Section 1.5]. For an account on the Busby invariant and extension theory for C-algebras see [Arv77], [Bla98, Section 15], [BD96, Section 1], [ELP99, Section 2] and [PS79]. Let A be a C-algebra. We denote by A the collection of unitary equivalence classes of irreducible representations of A. On the other hand, we define P rim(A) to be the set of primitive ideals of A, where a primitive ideal is the kernel of an irreducible representation of A. The set P rim(A) comes equipped with a lattice structure determined by set inclusion. Next, since any two unitarily equivalent -representations have the same kernel, the map π 7→ Ker π factors through to yield a surjective map κ : A → P rim(A). C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 9 It turns out that for type I C-algebras, the above map κ is a bijection, as every primitive ideal J uniquely determines, up to unitary equivalence, an irreducible representation π such that J = Ker π (See [Dix77, Theorem 4.3.7]). When we have a -isomorphism ϕ : A → B between two C-algebras, we denote by ϕ∗ : Suppose we have the following exact sequence of C-algebras P rim(A) → P rim(B) the induced lattice isomorphism between the spectra. ι→ A π→ B → 0 0 → K (1.2) and denote by q : M (K) → M (K)/K =: Q(K) the Calkin map. Then there is a -homomorphism θ : A → M (K) into the multiplier algebra of K, uniquely determined by θ(a)c = ι−1(aι(c)) for c ∈ K and a ∈ A. Hence, a -homomorphism η : B → Q(K) will be induced, and we call this map η the Busby invariant of the exact sequence above. We say that the above exact sequence is essential if K is an essential ideal in A, that is, if the intersection of K with any non-trivial ideal in A is non-trivial. The above association turns out to be a bijection between exact sequences of C-algebras given as in (1.2) and -homomorphisms η : B → Q(K), where the inverse map sends a - homomorphism η : B → Q(K) to the exact sequence where the pre-image A := q−1(η(B)) under the Calkin quotient q, yield an exact sequence as in (1.2), with π replaced by the restriction of q to A. Under this bijection, an exact sequence as in (1.2) is essential if and only if its associated Busby invariant is an injective -homomorphism. Definition 1.15. Suppose Ki, Ai, Bi are C-algebras for i = 1, 2, and that π2→ B2 → 0 π1→ B1 → 0 and 0 → K2 0 → K1 ι1→ A1 ι2→ A2 (1.3) are two short exact sequences. We say that these two short exact sequences are isomorphic if there exists a -isomorphism α : A1 → A2 such that α(ι1(K1)) = ι2(K2). Suppose η1 and η2 are Busby maps for exact sequences as in (1.3). [ELP99, Theorem 2.2] then yields that these two short exact sequences are isomorphic if and only if there exist - isomorphisms κ : K1 → K2 and β : B1 → B2 such that eκη1 = η2β whereeκ : Q(K1) → Q(K2) is the induced -isomorphism between the Calkin algebras. In the context of extensions by a single copy of compact operators on separable infinite dimen- sional Hilbert space, that is when K = K(H), the Calkin quotient map q : M (K(H)) → Q(K(H)) discussed above is just the regular quotient map into the Calkin algebra, since M (K(H)) = B(H), so that M (K(H))/K(H) = Q(H). We denote by K = K(H) the compact operators on separable infinite dimensional Hilbert space H. Let B be a C-algebra. We write E(B) for the collection of all injective -homomorphisms of B into Q(H). We call elements in E(B) extensions, as they are in bijection, under (the inverse of) the Busby map, with essential exact sequences of C∗-algebras of the form 0 → K → A → B → 0 We then say that two extensions η1, η2 ∈ E(B) are (1) Strongly (unitarily) equivalent if there is a unitary U ∈ B(H) such that η1(b) = (2) Weakly (unitarily) equivalent if there is a unitary element u ∈ Q(H) such that η1(b) = q(U )η2(b)q(U∗) for all b ∈ B. uη2(b)u∗ for all b ∈ B. When B is unital we write Exts(B) and Extw(B) for the strong and weak equivalence classes of unital extensions in E(B), respectively. When B is non-unital, we write Exts(B) and Extw(B) for the strong and weak equivalence classes of all extensions in E(B), respectively, however in this case Exts(B) = Extw(B) by [Bla98, Proposition 15.6.4]. We denote by [η]s and [η]w the equivalence classes of an extension η in Exts(B) and Extw(B), respectively Given η1, η2 ∈ E(B), we may define η1 ⊕ η2 ∈ E(A) (via some fixed identification Q(H) ⊕ Q(H) ⊆ Q(H ⊗ C2) ∼= Q(H)) by specifying (η1 ⊕ η2)(b) = η1(b) ⊕ η2(b). This operation induces a well-defined addition + on Exts(B) and Extw(B) given for two extensions η1 and 10 ADAM DOR-ON AND DANIEL MARKIEWICZ η2 by [η1]s + [η2]s := [η1 ⊕ η2]s and [η1]w + [η2]w := [η1 ⊕ η2]w, and makes them into abelian semigroups. An extension τ is called trivial if it lifts to a -homomorphism τ : B → B(H) such that its composition with the Calkin quotient map yields q ◦ τ = τ . Such a trivial extension τ is called strongly unital if the map τ can be chosen to be unital (in particular this is relevant only when B itself is unital and τ is unital). Trivial extensions correspond to split essential exact sequences via (the inverse of) the Busby map. It is straightforward to construct injective -homomorphisms of a C-algebra B into B(H) which do not intersect K(H), hence trivial extensions always exist. Moreover, the same argument yields strongly unital trivial extensions. Voiculescu [Voi76] showed that when B is separable, the semigroup Exts(B) has a zero ele- ment. When B is non-unital, the zero element consists precisely of the trivial extensions. When B is unital, it consists of the strongly unital trivial extensions. For more details, see [Bla98, Section 15.12], especially [Bla98, Theorem 15.12.3]. Although Exts(B) and Extw(B) are not always groups, it follows from a theorem of Choi and Effros that when B is separable and nuclear, both semigroups are actually groups (see [Bla98, Corollary 15.8.4]). Suppose now that B is unital. There is an action ε of Z on Exts(B) given by ε(m)[η]s = [Adu◦η]s where u ∈ Q(H) is a unitary of Fredholm index −m, and Adu(a) = u∗au for a ∈ Q(H). By definition of addition, we have that ε(n + m)([η1]s + [η2]s) = [Adu⊕v(η1 ⊕ η2)] = ε(n)[η1]s + ε(m)[η2]s where u and v are unitaries in Q(H) of indices −n and −m respectively. In particular, if τ is a strongly unital trivial extension then ε(m)[η]s = ε(0 + m)([η]s + [τ ]s) = [η]s + ε(m)[τ ]s. Hence, when we denote by λB : Exts(B) → Extw(B) the canonical quotient map, we have that Ker λB = {ε(m)[τ ]s m ∈ Z}. Let γB : Extw(B) → Hom(K1(B), Z) be the so-called index invariant of B, given by γB([η]w) = ind◦η∗, where η∗ : K1(B) → K1(Q(H)) is the map induced between the K1 groups and ind : K1(Q(H)) → Z is the Fredholm index. Hence, for a unital C- algebra B, we always have the following sequence of maps (1.4) Exts(B) λB−→ Extw(B) γB−→ Hom(K1(B), Z) We next give the details of two particular examples, which will turn out to be useful to us later in the end of Section 2 and in Section 5. Example 1.16. Take B = C(T). In this case B is nuclear and separable, so both the weak and strong extension semigroups are groups. We note that Hom(K1(B), Z) ∼= Z as K1(B) ∼= Z, and every homomorphism is determined on the generator 1. We next show that in this case, Indeed, for every m ∈ Z there is a unitary u ∈ Q(H) with the map γB ◦ λB is surjective. σ(u) = T, and Fredholm index m, we may define a -homomorphism ηm : C(T) → Q(H) given by ηm(z 7→ z) = u which implements a -isomorphism C(T) ∼= C∗(u). Thus we obtain an extension with index invariant k 7→ k · m ∈ Hom(K1(B), Z). Next, we show that γB ◦ λB is injective. Indeed, if γB ◦ λB[η]s = 0, then ind(η(z 7→ z)) = 0 and hence there is a unitary U ∈ B(H) with σ(U ) = T s.t q(U ) = η(z 7→ z). Thus, η lifts to a unital ∗-homomorphism η : C(T) → B(H), so that η is a strongly unital trivial extension, and the map γB ◦ λB is injective. We conclude that Exts(C(T)) ∼= Extw(C(T)) ∼= Z, and that ε(n) acts trivially on Exts(C(T)) for each n. Example 1.17. Take B = Md. Again in this case B is nuclear and separable so that both weak and strong extension semigroups are groups. We already know that K1(Md) ∼= {0}, so that the right most group in (1.4) vanishes. Let η : Md → Q(H) be a unital extension. We reiterate the construction in [Bla98, Example 15.4.1 (b)] lifting η : Md → Q(H) to a -homomorphism η : Md → B(H), and measuring how far η is from being unital. That is, how far is η from being a strongly unital trivial extension. Let {eij} be a system of matrix units for η(Md). By standard essential spectrum arguments, one can find projections pii ∈ B(H) that lift each eii. Next, by appealing to [PS79, Lemma 1.1], for all 2 ≤ i ≤ d we may find partial isometries e1i lifting e1i such that e∗1ie1i ≤ pii and C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 11 e1ie∗1i ≤ p11. We set eij = e∗1ie1j so that {eij} is a lifted set of matrix units in pB(H)p, where p =P eii. We note that p is a projection of finite dimensional cokernel, say of dimension ℓ, so that by adding a homomorphism from Md to (1 − p)B(H)(1 − p) if necessary, we may arrange for 0 ≤ ℓ < d. The defect of η is then defined to be ℓ ∈ Zd, and up to strong equivalence it is independent It is then easy to see that two unital extensions of the choice made in the process above. η1, η2 ∈ E(Md) are strongly equivalent if and only if they have the same defect, and are always weakly equivalent. Hence, we conclude that Exts(Md) ∼= Zd and Extw(Md) ∼= {0}. 2. Cuntz-Pimsner algebra of a stochastic matrix We next close a gap kindly pointed out to us by Dilian Yang in the proof of the character- ization of the Cuntz-Pimsner algebra of a finite irreducible stochastic matrix, which is one of the theorems of [DOM14, Section 5]. The theorem at stake, which corresponds to [DOM14, Corollary 5.16] is as follows. Theorem 2.1. Let P be an irreducible stochastic matrix of size d. Then O(P ) ∼= Md(C)⊗C(T). The main issue is that the cyclic decomposition of periodic irreducible stochastic matrices need not be realized in square blocks as claimed in [DOM14, Remark 2.9]. Consider the following example kindly brought to our attention by Dilian Yang: let Ω = {1, 2, 3} and set P = 0 0 1 2 0 1 0 1 1 2 0 =(cid:20) 0 P0 0(cid:21) , P1 where P0 =(cid:20)1 1(cid:21) , P1 =(cid:2) 1 2 1 2(cid:3) The matrix P has period 2, Ω0 = {1, 2} and Ω1 = {3}, and both P0 and P1 are not square. Therefore, the results in [DOM14, Section 5] after [DOM14, Proposition 5.7], and in particular [DOM14, Proposition 5.15], only apply in the case of square blocks. Therefore a gap remains in the proof of Theorem 2.1 in its stated form, and we now provide a different proof for it, which works in all cases, and resolves any remaining gaps with the rest of [DOM14, Section 5]. The issue above does not affect the remainder of the paper, namely [DOM14, Sections 6 and 7]. Recall that the adjoint W (n)∗A determined on fibers by of W (n) A , which maps Arv(P )n+m to Arv(P )m, is uniquely W (n)∗A (B) = Gr(P m) ∗ (A∗B), B ∈ Arv(P )m+n where the reason for Schur-multipling A∗·B with Gr(P m) is to make sure that the product lands in Arv(P )m with its given entry constraints (See the discussion preceeding [DOM14, Proposition 5.7]). We note that ℓ∞(Ω) acts on Arv(P )m as left multiplication by diagonal matrices. Therefore, W (0) Ekk The following proposition, which works in all cases, replaces [DOM14, Remark 5.10] and the = pk as the adjointable operator on FArv(P ). discussion preceding it. Proposition 2.2. Let q ∈ N and suppose that A ∈ Arv(P )q. Then there exists m0 ∈ N such that for all m ≥ m0 we have that W (q)∗A (B) = A∗B, ∀B ∈ Arv(P )q+m That is, if m ≥ m0 and B ∈ Arv(P )q+m, then the matrix A∗B has support contained in the support of P m. Proof. Suppose this fails. Then there exists a sequence of matrices B(n) ∈ Arv(P )q+mn, with n 7→ mn increasing, such that the support of A∗B(n) is not contained in the support of P mn. By finiteness of P , perhaps by replacing B(n) by a subsequence, we may assume that there exist i, j, k ∈ Ω independent of n such that B(n) ∈ Arv(P )q+mn and both piA∗pk 6= 0 and pkB(n)pj 6= 0 while P (mn) = 0. Again by moving to a subsequence, we may assume that there exists 0 ≤ ℓ < r independent of n such that mn ≡ ℓ mod r. ij 12 ADAM DOR-ON AND DANIEL MARKIEWICZ of Theorem 1.9, we must have that P (m) are no paths from i to j whose length is of residue ℓ mod r. Let Ω0, . . . , Ωr−1 be the cyclic decomposition of P with respect to k. Note that by item (2) ij = 0 for all m such that m ≡ ℓ mod r. Therefore there Let σ(i), σ(j) be such that i ∈ Ωσ(i) and j ∈ Ωσ(j). Since piA∗pk 6= 0, we have that pkApi 6= 0 and hence P (q) ki > 0. It follows from the cyclic decomposition theorem that paths from k to i have length with residue σ(i) (mod r, which we will suppress). Since paths from k to k must have lengths with zero residue by periodicity, we must have that paths from i to k will have length with residue r−σ(i). Therefore paths from i to j have lengths with residue r−σ(i)+σ(j) ≡ σ(j)−σ(i) mod r. Next, since A ∈ Arv(P )q, and pkApi 6= 0, we have by the definition of the cyclic decomposition that σ(i) ≡ q mod r. Similarly, since B ∈ Arv(P )q+mn , and pkBpj 6= 0, we have that σ(j) ≡ q + mn ≡ q + ℓ mod r. Therefore, σ(j) − σ(i) ≡ q + ℓ − q ≡ ℓ mod r and we conclude that all paths from i to j must have residue ℓ mod r. But this is impossible since we have noted before that there are no paths from i to j whose length is of residue ℓ mod r. (cid:3) Definition 2.3. Let P be a finite irreducible r-periodic stochastic matrix over Ω of size d. We will say that a cyclic decomposition Ω0, ..., Ωr−1 for P is properly enumerated if Ω is enumerated in such a way that for every 0 ≤ m < k < r, i ∈ Ωm and j ∈ Ωk we have that i < j. For i ∈ Ω, denote by σ(i) the unique index 0 ≤ σ(i) < r such that i ∈ Ωσ(i). Given a properly enumerated cyclic decomposition Ω0, ..., Ωr−1 for P , we define operators U and (Sij)i,j∈Ω in L(FArv(P )) as follows. The operator U has degree r with respect to the grading, i.e. for every m ∈ N, U (Arv(P )m) ⊆ Arv(P )m+r, and it is uniquely determined by U (B) = Gr(P m+r) ∗ B, m ∈ N, B ∈ Arv(P )m. If i ≤ j, then σ(i) ≤ σ(j), and denote by ℓ = σ(j) − σ(i). Then Sij is an operator of degree ℓ, i.e. for all m ≥ 0, Sij(Arv(P )m) ⊆ Arv(P )m+ℓ and it is given by Sij(B) = Gr(P m+ℓ) ∗ (Eij · B), m ∈ N, B ∈ Arv(P )m. If i > j we define Sij = S∗ji. The family (U, (Sij)i,j∈Ω) is called the standard family associated to the properly enumerated cyclic decomposition Ω0, ..., Ωr−1. Recall the following auxiliary C-algebra considered in [DOM14, Section 5]: We recall that T ∞(P ) = C∗(ℓ∞(Ω) ∪ {W (n) A n ∈ N, A ∈ Arv(P )n}) J (T ∞(P )) := { T ∈ T ∞(P ) lim n→∞kT Qnk = 0 } is a two sided ideal in T ∞(P ) with O(P ) ∼= T ∞(P )/J (T ∞(P )) by [DOM14, Theorem 5.6]. We denote by T ∈ O(P ) the image of T ∈ T ∞(P ) under the associated canonical quotient map q : T ∞(P ) → O(P ) ∼= T ∞(P )/J (T ∞(P )). Lemma 2.4. Let P be an irreducible r-periodic stochastic matrix over Ω of size d with properly enumerated cyclic decomposition Ω0, ..., Ωr−1, and let (U, (Sij )i,j∈Ω) be its associated standard family. (1) Let i, j ∈ Ω be such that i ≤ j in the properly enumerated decomposition of Ω, so that ℓ := σ(j) − σ(i) ≥ 0. Then there exists n0 ∈ N such that for all n ≥ n0 we have W (nr+ℓ) Sij = W (nr)∗ and U = W (nr)∗ Hence, U ∈ T ∞(P ) and Sij ∈ T ∞(P ) for all i, j ∈ Ω. that Eij Id W (nr+r) Id Id (2) Let i, j ∈ Ω. There exists m0 ∈ N such that for all m ≥ m0, and B ∈ Arv(P )m we have Sij(B) = EijB, U (B) = B and U∗(B) = B (3) For all i, j, t, k ∈ Ω we have SijStk − δjtSik ∈ J (T ∞(P )) (4) U∗U − I, U U∗ − I ∈ J (T ∞(P )) (5) For all i, j ∈ Ω we have SijU − U Sij ∈ J (T ∞(P )) C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 13 (6) The family (U ,{Sij}i,j∈Ω) generates O(P ). Therefore, {Sij}i,j∈Ω is a system of d× d matrix units in O(P ) and U is a unitary in O(P ) that commutes with them and together they generate O(P ). Proof. (1) Let i, j ∈ Ω be such that i ≤ j in the properly enumerated decomposition of Ω, so that ℓ := σ(j) − σ(i) ≥ 0. By item (2) of the cyclic decomposition Theorem 1.9, there exists n0 ∈ N such that Eij ∈ Arv(P )nr+ℓ and Id ∈ Arv(P )nr for all n ≥ n0. Then for all n ≥ n0, m ∈ N and B ∈ Arv(P )m we have W (nr)∗ W (nr+ℓ) Eij Id W (nr)∗ Id W (nr+r) Id (B) = Gr(P m+ℓ) ∗ (EijB) = Sij(B) (B) = Gr(P m+r) ∗ B = U (B) W (nr+ℓ) and U = W (nr)∗ W (nr+r) Id Id Eij so that Sij = W (nr)∗ Id (2) Let i ≤ j ∈ Ω be given. By the previous item and by Proposition 2.2 there exists m0 ∈ N such that for all m ≥ m0 and B ∈ Arv(P )m we have Sij(B) = W (nr)∗ U (B) = W (nr)∗ Id W (nr+ℓ) Eij (B) = EijB W (nr+r) (B) = B Id Similarly, by taking adjoints, we have that Id Sji(B) = W (nr+ℓ)∗ U∗(B) = W (nr+r)∗ Eji W (nr) Id (B) = EjiB W (nr) Id (B) = B Id proving the statement in all cases. (3) Let i, j, t, k ∈ Ω be given. By item (2), there exists m0 ∈ N such that for all m ≥ m0 and B ∈ Arv(P )m we have that SijStk(B) = EijEtkB = δjtEikB = δjtSik(B) (4) By item (2), there exists m0 ∈ N such that for all m ≥ m0 and B ∈ Arv(P )m we have Thus we have that SijStk − δjtSik ∈ J (T ∞(P )). that U∗U (B) = U∗(B) = B ∈ Arv(P )m Thus we have that U∗U − I, U U∗ − I ∈ J (T ∞(P )) B ∈ Arv(P )m we have the following element in Arv(P )m+r+ℓ where ℓ = σ(j) − σ(i). (5) Let i, j ∈ Ω be given. By item (2), there exists m0 ∈ N such that for all m ≥ m0 and U U∗(B) = U∗(B) = B ∈ Arv(P )m and SijU (B) = Sij(B) = EijB = U (EijB) = U Sij(B) Thus we have that SijU − U Sij ∈ J (T ∞(P )). (6) We first observe that since we are dealing with stochastic matrices over a finite state space Ω, it is in fact the case that T ∞(P ) is generated by ℓ∞(Ω) and {W (1) Eij}(i,j)∈E(P ), where E(P ) = { (i, j) Pij > 0 }. Indeed, since every Arv(P )n is finite dimensional, A can be written as a linear combination of elements of the form W (n) every W (n) . Now, if Eik W (n) Eik given by i = j0 → j1 → ... → jn = k and we would have that W (n) so that every element W (n) Eik and so is in the algebra generated by ℓ∞(Ω) and {W (1) ik > 0 and so there is a path of length n from i to k is non-zero, this means that P (n) Ejn−1 jn Eij}(i,j)∈E(P ), Ej0j1 ·...·W (1) = W (1) Eik T ∞(P ) = C∗(cid:0)ℓ∞(Ω) ∪ {W (1) Eij}(i,j)∈E(P )(cid:1) Therefore O(P ) is generated as a C-algebra by the images of ℓ∞(Ω) and {W (1) under q : T ∞(P ) → O(P ). Eij}(i,j)∈E(P ) 14 ADAM DOR-ON AND DANIEL MARKIEWICZ Let us denote by A the C-subalgebra of O(P ) ∼= T ∞(P )/J (T ∞(P )) generated by U and Sij for i, j ∈ Ω. It follows from item (2) that Sii − pi ∈ J (T ∞(P )), therefore, we have that q(ℓ∞(Ω)) ⊆ A. In order to complete the proof that A = O(P ), it suffices to show that W (1) Eij ∈ A for all (i, j) ∈ E(P ). Let (i, j) ∈ E(P ), and suppose that r > 1. If i ≤ j, then we must have by the cyclic decomposition theorem that σ(j) − σ(i) = 1 and Sij is an operator of degree one and by item (2) we have that Sij − W (1) Eij ∈ J (T ∞(P )). On the other hand, if i > j, then also by the cyclic decomposition theorem we must have σ(i) − σ(j) = r − 1 and in that case Sij is an operator of degree −(r − 1). Therefore U Sij has degree 1, and by item (2) we have that U Sij − W (1) Eij ∈ A. Finally, if (i, j) ∈ E(P ), and r = 1, we have that Sij is an operator of degree zero and by item (2) we have that U Sij − W (1) Eij ∈ J (T ∞(P )). Therefore, we also obtain that W (1) Eij ∈ A. Eij ∈ J (T ∞(P )). Therefore, in both cases we obtain that W (1) (cid:3) Recall from the discussion preceding [DOM14, Proposition 5.5] that there is a natural gauge group action α on T ∞(P ) uniquely determined by αλ(W (n) A . Since J (T ∞(P )) is gauge invariant by [DOM14, Theorem 5.6], this gauge action passes to the quotient O(P ), and we denote by O(P )0 the fixed point algebra. A ) = λnW (n) Let A =Lr−1 Proof of Theorem 2.1. First note that Md(C) ⊗ C(T) is the universal C-algebra generated by a system of d × d matrix units eij and a unitary u that commutes with them. Hence, by Lemma 2.4 we obtain a surjective -homomorphism ψ : Md(C) ⊗ C(T) → O(P ) that sends eij to Sij and u to U . It remains to show that ψ is injective. ℓ=0 MΩℓ(C)⊗ 1 ⊆ Md(C)⊗ C(T). First we note that ψ restricted to A is injective, since ψ is already injective when restricted to the larger simple subalgebra Md(C) ⊗ 1. We now show that ψ(A) = O(P )0. First note that O(P )0 is generated by monomials of degree zero (according to the gauge action) in the matrix units (Sij) and the unitary U , which commutes with the latter. Let X ∈ O(P )0 be such a monomial. Products of matrix units are also matrix units, therefore there exists i, j ∈ Ω, n ∈ Z such that X = Sij U . Hence, the only way that X has degree zero is if n = 0 and σ(i) = σ(j). Moreover, A is precisely generated by all eij, i, j ∈ Ω such that σ(i) = σ(j). Hence ψ(A) = O(P )0. Next, we show ψ is injective on the entire algebra Md(C) ⊗ C(T). Given the identifications Md(C) ⊗ C(T) ∼= C(T; Md(C)) and Md(C) ⊗ 1 ∼= Md(C), let us consider the faithful conditional expectation Γ0 : Md(C) ⊗ C(T) → Md(C) ⊗ 1 given by n Γ0(T ) =ZT T (z) dz where dz represents normalized Haar measure on the circle. Note that in particular, for all i, j ∈ Ω and n ∈ Z, Γ0(eij un) = δ0,n eij We now take E0 to be the faithful conditional expectation from Md(C)⊗ 1 toLr−1 ℓ=0 MΩℓ(C)⊗ 1, and let Φ0 : O(P ) → O(P )0 denote the canonical conditional expectation into the fixed point algebra associated with the gauge action. We then have that Φ0ψ = ψE0Γ0. Indeed, since for all i, j ∈ Ω, n ∈ N, Φ0ψ(eijun) = Φ0(Sij U n ) = δ0,nδσ(i),σ(j) Sij = δ0,nδσ(i),σ(j) ψ(eij ) = δ0,nψ(E0(eij)) = ψ(E0(Γ0(eijun))), and since monomials are total in the algebra, we have Φ0 ◦ ψ = ψE0Γ0. itive non-zero T ∈ Md(C) ⊗ C(T) such that ψ(T ) = 0. Finally, suppose towards a contradiction that ψ is not injective. Then there exists a pos- In that case Φ0(ψ(T )) = 0. Hence C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 15 ψ(E0(Γ0(T ))) = Φ0(ψ(T )) = 0. By injectivity of ψ on the image of E0, which is the algebra A, we obtain E0(Γ0(T )) = 0. We reach a contradiction since E0 and Γ0 are faithful conditional expectations. (cid:3) Now that we have filled the gap in the computation of the Cuntz-Pimsner algebra of a finite irreducible stochastic matrix, we compute the extension groups for it, which will be useful to us later in Section 5. Based on the work of [PS79], one has a description of Exts(B ⊗ Md) for any unital C-algebra B, for which Exts(B) contains no elements of order d, as follows. For any unital extension η ∈ E(B ⊗ Md), we define a map [η]s 7→ ([ι∗η]s, [j∗η]s) into Exts(B)⊗ Zd by setting ι∗η = ηB⊗I and j∗η = ηI⊗Md. Then [PS79, Proposition 2.2] shows that this map induces an isomorphism of semigroups Exts(B ⊗ Md) ∼= { (d[η] + ε(ℓ)[τ ], ℓ) ∈ Exts(B) ⊗ Zd η ∈ E(B), ℓ ∈ Z } where τ is a trivial strongly unital extension. By Example 1.16 we have that ε(ℓ)[η]s = [η]s for all η ∈ Exts(C(T)), so that Exts(C(T) ⊗ Md) ∼= {(ds, ℓ) ∈ Z × Zd s ∈ Z, ℓ ∈ Z } so that Exts(C(T) ⊗ Md) ∼= dZ × Zd and Extw(C(T) ⊗ Md) ∼= Z as the projection (and division by d) onto the first coordinate of Exts(C(T)⊗ Md). Since Extw(C(T)⊗ Md) ∼= Z is the quotient of Exts(C(T) ⊗ Md) ∼= dZ × Zd by the subgroup { ε(n)[τ ]s n ∈ Z } ∼= Zd, we can identify the subgroup { ε(n)[τ ]s n ∈ Z } of Exts(C(T) ⊗ Md) with the image { [j∗η]s [η]s ∈ Exts(C(T) ⊗ Md) } ∼= Zd. Note that any automorphism β of C(T)⊗Md induces an automorphism βs of Exts(C(T)⊗Md) by composition [η]s 7→ [η ◦ β]s. Furthermore, every unitary element u ∈ U (C(T) ⊗ Md) defines an automorphism Adu of C(T) ⊗ Md by way of Adu(f )(z) = u∗(z)f (z)u(z) for f ∈ C(T; Md) and z ∈ T. Denote by AutC(T)(C(T) ⊗ Md) the collection of C(T)-bimodule -automorphisms of C(T) ⊗ Md. Proposition 2.5. Let η be a unital extension and let β ∈ Aut(C(T)⊗ Md) be an automorphism. Up to the identification Exts(C(T) ⊗ Md) ∼= dZ ⊗ Zd given above, we have that either βs[η] = [η] = ([ι∗η], [j∗η]) or βs[η] = (−[ι∗η], [j∗η]). Proof. Let β ∈ Aut(C(T) ⊗ Md) be some -automorphism. Then β induces an automorphism β∗ on the primitive ideal spectrum T, which then induces an automorphism (β∗)∗ back on ∗ (z)). It is easy to see that [j∗η] = [j∗η ◦ (β∗)∗] since C(T) ⊗ Md given by (β∗)∗(f )(z) = f (β−1 (β∗)∗(I ⊗ Md) = I ⊗ Md. Since the induced map ((β∗)∗)s on Exts(C(T) ⊗ Md) ∼= dZ × Zd is the identity on the second coordinate Zd, we must have that [ι∗(η ◦ (β∗)∗)] is either [ι∗η] or −[ι∗η]. Hence, by composing with the inverse of (β∗)∗ if necessary, we may assume that β∗ = IdT. By [RW98, Corollary 5.46] we have that β ∈ AutC(T)(C(T)⊗ Md), so that by [RW98, Lemma 4.28], there is a point-norm continuous map σ : T → Aut(Md) such that β(f )(z) = σz(f (z)). Since the second cohomology group of the torus H 2(T; Z) vanishes, by [RW98, Theorem 5.42], there is a unitary element u ∈ U (C(T) ⊗ Md) such that β = Adu. Then Adu induces a map on Exts(C(T) ⊗ Md), so that by the homomorphism property of the Fredholm index, we get that, [ι∗η ◦ Adu] = ind(η(Adu(z ⊗ I))) = ind(η(z ⊗ I)) = [ι∗η] Next, since the image { [j∗η]s [η]s ∈ Exts(C(T) ⊗ Md) } ∼= Zd can be identified with the subgroup { ε(n)[τ ]s n ∈ Z }, in order to show that [j∗η ◦ β] = [j∗η], it will suffice to show that βs(ε(n)[τ ]s) = ε(n)[τ ]s. However, since βs commutes with ε(n), it will suffice to show that βs([τ ]s) = [τ ]s. But βs is a group homomorphism, so it must send [τ ]s to itself. Hence, we obtain that [j∗η ◦ β] = [j∗η]. (cid:3) 3. Non-commutative Choquet boundary In this section we first find all the irreducible representations of T (P ) for a finite irreducible stochastic matrix P . We then determine the boundary representations with respect to T+(P ) among them. We show that any representation annihilating J (P ) := J (T (P )) has the unique 16 ADAM DOR-ON AND DANIEL MARKIEWICZ A ) = λnS(n) A ) = λnW (n) A and αλ(W (n) extension property when restricted to T+(P ), and find conditions that guarantee that an irre- ducible representation supported on J (P ) is boundary or not. As given in [DOM14, Theorem 5.6], the C-algebra T c(P ) is the one generated by both T ∞(P ) and T (P ), and it too has a gauge action which is the restriction of the gauge action of L(FArv(P )), which satisfies αλ(S(n) A , so that T c(P ) is gauge invariant, and J (T c(P )) is a closed gauge invariant two-sided ideal by [DOM14, Theorem 5.6]. As discussed in [DOM14, Section 4] for general subproduct systems, Fourier coeficients Φk on T c(P ) may be defined in such a way that every T ∈ T c(P ) can be written as P∞k=−∞ Φk(T ), k=−n(cid:0)1 − kn+1(cid:1)Φk(T ) converges in norm where this sum convergens Cesaro. That is, wherePn to T . From now on, we will denote Wij := WEij for i, j ∈ Ω. Proposition 3.1. Let P be an irreducible stochastic matrix on Ω of size d. Then J (T c(P )) is the two sided ideal generated by {Qn}n∈N inside T c(P ). Proof. By [DOM14, Proposition 5.2] we see that Qn ∈ T (P ) ⊂ T c(P ), and since kQnQmk → 0 as m goes to infinity, we see that Qn ∈ J (T c(P )). For the reverse inclusion, let T ∈ J (T c(P )), and write T = P∞k=−∞ Φk(T ) as a Cesaro convergent sum where Φk(T ) maps Arv(P )n to Arv(P )n+k if n + k ≥ 0 and {0} other- wise. Further notice that Φk(T ) ∈ J (T c(P )) for all k ∈ Z, since by [DOM14, Theorem 5.6] we have that J (T c(P )) is gauge invariant. In this case, we have that kΦk(T )Q[n+1,∞)k = supm≥n+1 kΦk(T )Qmk → 0. Hence, since Φk(T )Q[0,n] is in the ideal generated by {Qn}n∈N, we see that Φk(T ) is in the closed ideal generated by {Qn}n∈N and so must be T by Cesaro approximation. (cid:3) For a finite irreducible stochastic matrix P with state set Ω of size d, we have that ℓ∞(Ω) is faithfully represented in B(ℓ2(Ω)) by diagonal matrix multiplication on columns. Hence by [RW98, Corollary 2.74], this faithful -representation promotes to a faithful -representation π : L(FArv(P )) → B(FArv(P )⊗id ℓ2(Ω)) given by π(T )(ξ⊗h) = T ξ⊗h. Note that FArv(P )⊗id Cek is a reducing subspace for π(T c(P )) for each k ∈ Ω. Notation 3.2. For a state k ∈ Ω we will find it useful to denote Arv(P )n,k := Arv(P )n ⊗ Cek, and FP,k := ⊕∞n=0Arv(P )n,k = FArv(P ) ⊗id Cek, the reducing Hilbert space for π(T c(P )) mentioned above, so that FArv(P ) ⊗ ℓ2(Ω) = ⊕k∈ΩFP,k. For fixed n we also denote for i ∈ Ω with (i, k) ∈ Gr(P n) the elements e(n) ik := Eik ⊗ ek ∈ Arv(P )n,k which comprise a finite orthonormal basis for each Arv(P )n,k, so that for varying n ∈ N and i ∈ Ω with (i, k) ∈ E(P n) the collection {e(n) ik } is an orthonormal basis for FP,k. Proposition 3.3. Let P be an irreducible stochastic matrix over Ω of size d. Then for each πk : T c(P ) → B(FP,k) given by πk(T ) = π(T )FP,k we have that πk(T (P )) is an irreducible subalgebra of B(FP,k). Proof. By [DOM14, Proposition 5.2] we see that Qn ∈ T (P ) for every n ∈ N. Let 0 6= H′ ⊆ FP,k be some non-zero invariant subspace. Since {πk(Q[0,n])} converges SOT to the identity on FP,k, there is some minimal n0 ∈ N such that πk(Qn0)ξ 6= 0 for some ξ ∈ H′. In this case, 0 6= πk(Qn0)ξ = A ⊗ ek ∈ H′ ∩ Arv(P )n0,k for some A ∈ Arv(P )n0, so that there exists j ∈ Ω and some non-zero scalar c ∈ C with 0 6= e(n0) jk = c · πk(pjQn0)ξ ∈ H′ where (j, k) ∈ E(P n0). This means that e(0) )(e(n0) jk ) ∈ H′, where c1 > 0 is some scalar. kk = c1πk(S(n0)∗Ejk ik Thus, for m ≥ 0 if e(m) is some vector in Arv(P )m,k, we see that e(m) kk ) ∈ H′ where c2 > 0 is some scalar. This shows that the set of elements e(m) for all m ≥ 0 and ik (i, k) ∈ E(P m) is in H′, and this set of elements is an orthonormal basis for FP,k, and so H′ = FP,k. (cid:3) Hence, we see that π decomposes into d = Ω irreducible representations πk as above, so that π = ⊕k∈Ωπk : T c(P ) → ⊕k∈ΩB(FP,k). We next show that each πkT (P ) is in a distinct unitary equivalence class of irreducible representations for T (P ). ik = c2πk(S(m) Eik )(e(0) C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 17 Proposition 3.4. Let P be a finite irreducible stochastic matrix on Ω and k, k′ ∈ Ω be distinct indices. Then πkT (P ) and πk′T (P ) are not unitarily equivalent. Proof. Suppose that k, k′ ∈ Ω are such that πkT (P ) and πk′T (P ) are unitarily equivalent. Then there is a unitary U : FP,k → FP,k′ such that U πk(T ) = πk′(T )U for all T ∈ T (P ). For j ∈ Ω, we have that pjQ0 ∈ T (P ), so that U πk(pjQ0) = πk′(pjQ0)U . Apply this operator to e(0) kk ∈ Arv(P )0,k ⊂ FP,k and get πk′(pjQ0)U (e(0) kk ) = U πk(pjQ0)(e(0) kk ) = U (δjke(0) kk ) On the other hand πk′(Q0)U (e(0) for some non-zero c ∈ C. But after applying πk′(pj) we would obtain that kk ) must have image in Arv(P )0,k′ so that πk′(Q0)U (e(0) kk ) = c·e(0) k′k′ πk′(pjQ0)U (e(0) kk ) = c · πk′(pj)(e(0) k′k′) = c · δjk′e(0) k′k′ k′k′ = (cid:3) kk ) 6= 0 in contradiction. Hence, πkT (P ) and πk′T (P ) are not unitarily equivalent. Thus, we see that if k 6= k′ then by taking j = k we would obtain that 0 = c · δjk′e(0) U (δjke(0) Proposition 3.5. Let P be a finite irreducible stochastic matrix on Ω. Then J (T (P )) = J (T c(P )) and is -isomorphic to ⊕k∈ΩK(FP,k). Thus, we have that T ∞(P ) ⊆ T c(P ) = T (P ). Proof. By Proposition 3.1, we have that J (T c(P )) is the ideal generated by {Qn}n∈N inside T c(P ), and since π(Qn) is a finite rank operator, we see by Proposition 3.3 that πk(J (T c(P ))) and πk(J (T (P ))) are irreducible compact operator subalgebras of B(FP,k) and hence by [Arv76, Theorem 1.3.4] they must both be equal to K(FP,k). Write the identity representation Id : π(J (T c(P ))) → ⊕k∈ΩB(FP,k) as a direct sum of irreducible representations with multiplic- ity Id = L n(ζ) · ζ, where each ζ is a representative in the equivalence class of irreducible representation given by restriction to some FP,k for some k. Then by Proposition 3.4 we have that n(ζ) = 1 for all ζ and that Idπ(J (T (P ))) has the same decomposition into irre- ducible representations as the one above. Since π = L πk is injective on J (T c(P )), we have that π(J (T c(P ))) = ⊕k∈ΩK(FP,k) = π(J (T (P ))), and by taking the inverse of the faithful -representation π, we obtain J (T (P )) = J (T c(P )). Finally, by [DOM14, Proposition 5.5] we have that T (P ) = T (P ) + J (T (P )) = T (P ) + J (T c(P )) = T c(P ) so that T ∞(P ) ⊆ T c(P ) = T (P ). (cid:3) We next wish to parametrize all irreducible representations of T (P ). Under the identification O(P ) ∼= C(T, Md) and J (T (P )) ∼= ⊕k∈ΩK(FP,k) we have the following exact sequence 0 → ⊕k∈ΩK(FP,k) → T (P ) → C(T, Md) → 0 If ρ : T (P ) → B(H) is a unital representation, by the discussion preceding [Arv76, Theorem 1.3.4] it decomposes uniquely into a central direct sum of representations ρ = ρJ ⊕ ρO, where ρJ is the unique extension to T (P ) of the restriction of ρ to J (T (P )), and ρO annihilates J (T (P )). Hence, the spectrum of T (P ) decomposes into a disjoint union of the spectrum of J (T (P )) ∼= ⊕k∈ΩK(FP,k) and the spectrum of O(P ) ∼= C(T, Md). For λ ∈ T, we define evλ : C(T, Md) → Md given by evλ([fij]) = [fij(λ)]. Since evλ has range Md, we obtain that evλ ◦ q is an irreducible representation of T (P ) where q : T (P ) → O(P ) is the quotient map. Note that every evλ ◦ q is a d dimensional representation. Corollary 3.6. Let P be an irreducible stochastic matrix over Ω of size d. Then the spectrum of T (P ) is parameterized by d irreducible representations of infinite dimension, each unitarily equivalent to some πk, and a torus T of irreducible representations of dimension d that annihilate J (T (P )), each unitarily equivalent to evλ ◦ q for some λ ∈ T. Proof. If ρ is an irreducible representation of T (P ) that does not annihilate J (T (P )), we have by [Arv76, Theorem 1.3.4] that ρJ (T (P )) is also irreducible. We use π−1 to obtain an irreducible representation ρ◦ π−1 of π(J (T (P ))). Since π(J (T (P ))) is a C-algebra of compact operators, by [Arv76, Theorem 1.4.4] every irreducible representation of it is unitarily equivalent to some 18 ADAM DOR-ON AND DANIEL MARKIEWICZ restriction to some FP,k. Pushing this back via π we obtain that ρ is unitarily equivalent to some πk. For the other part, if ρ does annihilate J (T (P )), it induces an irreducible representation of O(P ) ∼= C(T, Md) by taking the quotient by J (T (P )). Since the irreducible representations of C(T) are just point evaluations, and since C(T) is strongly Morita equivalent to C(T, Md), we see that ρ must be unitarily equivalent to the composition evλ ◦ q of an evaluation evλ : C(T, Md) → Md given by evλ([fij]) = [fij(λ)] and the natural quotient map q : T (P ) → O(P ). Thus, the spectrum of T (P ) is parametrized by d irreducible representations of infinite di- mension, and a torus T of irreducible representations of dimension d. (cid:3) Lemma 3.7. Let P be an irreducible stochastic matrix over a finite set Ω, and let ε > 0. There exists m ≥ 1 and M > 0 such that for every (i, j) ∈ E(P ) we have T (1) Eij − M · Q[0,m] (1 + ε)pj ≥ T (1)∗Eij Proof. For Ekℓ ∈ Arv(P )m and m ≥ 1, by definition of T (1) Eij , we see that T (1) Eij So that (Ekℓ) = δj,kvuut P (m) P (m+1) iℓ kℓ Eiℓ and T (1)∗Eij (Ekℓ) = δi,kvuut P (m) P (m+1) kℓ jℓ Ejℓ T (1)∗Eij T (1) Eij (Ekℓ) = δj,kvuut P (m) P (m+1) iℓ kℓ T (1)∗Eij (Eiℓ) = P (m) jℓ P (m+1) iℓ pj(Ekℓ) By Theorem 1.10, there exists m such that such that (j, ℓ) ∈ E(P m). Hence, if we take M = kQ[0,m]T (1)∗Eij T (1)∗Eij T (1) Eij − M · Q[0,m] as required. jℓ P (m) P (m+1) iℓ ≤ 1 + ε for all (i, j) ∈ E(P ) and ℓ ∈ Ω T (1) Eijk it follows that (1 + ε)pj ≥ (cid:3) We next show that representations annihilating J (P ) have u.e.p. when restricted to T+(P ). Proposition 3.8. Let P be a finite irreducible stochastic matrix over Ω, and let ρ : T (P ) → B(H) be a -representation such that ρ(J (P )) = {0}. Then ρT+(P ) has u.e.p. Proof. Let eρ : T+(P ) → B(K) be a maximal dilation of ρT+(P ) such that H is a subspace of K, and let ψ : T (P ) → B(K) be its (unique) extension to a -representation. Denote First note that since pi is a self-adjoint projection, we get that Yi ) ="ρ(T (1) Zi(cid:21) and ψ(T (1) Zij# ψ(pi) =(cid:20)ρ(pi) Xi Zi(cid:21) = ψ(pi) = ψ(pi)ψ(pi)∗ =(cid:20)ρ(pi)ρ(pi)∗ + XiX∗i Eij Yij ) Xij ∗ Eij (cid:20)ρ(pi) Xi Yi ∗ ∗(cid:21) So that by taking the (1, 1) compression, we obtain that XiX∗i = 0, so that Xi = 0. Now, since ψ(pi) is self-adjoint, we see that we must also have that Yi = 0. Next, for (i, j) ∈ E(P m), Suppose ψ(S(m) Eij ) ="ρ(S(m) Y (m)ij Z(m)ij# ) X(m)ij Eij Observe that for all m ≥ 1, by the proof of [DOM14, Proposition 5.5] we have that 0 ≤ Q[0,m−1] = Id − X(i,j)∈E(P m) S(m) Eij S(m)∗Eij C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 19 Hence, by applying ψ to this equation, we obtain that 0 ≤ ψ(Q[0,m−1]) = Id − X(i,j)∈E(P m) ψ(S(m) Eij )ψ(S(m) Eij )∗ Then by compressing to the (1, 1) corner we get Eij X(m)ijX(m)∗ij 0 ≤ Id − X(i,j)∈E(P m)(cid:2)ρ(S(m) where the last equality follows due to the fact that ρ annihilates J . Hence we must have that X(m)ij = 0 for all (i, j) ∈ E(P m), so that the (1, 1) compression of ψ(Q[0,m]) is 0, and if we specify m = 1, and note that S(1) , the above also yields that Xij = 0 for all Eij (i, j) ∈ E(P ). we have that Next, let ε > 0. By Lemma 3.7 there exists m ≥ 1 and M > 0 such that for all (i, j) ∈ E(P ) )ρ(S(m) Eij )∗ + X(m)ij X(m)∗ij(cid:3) = − X(i,j)∈E(P m) = pPij · T (1) (1 + ε)pj ≥ T (1)∗Eij (1 + ε)ψ(pj ) ≥ ψ(T (1) T (1) Eij − M · Q[0,m] ) − M · ψ(Q[0,m]) )∗ψ(T (1) Eij Hence, Eij Eij By compressing to the (1, 1) corner, we obtain that (1 + ε)ρ(pj) ≥ ρ(T (1) W (1) Eij Eij )∗ρ(T (1) Eij ) + Y ∗ijYij )∗ρ(T (1) Eij ) = ρ(W (1)∗Eij ) = ρ(pj), so for every ε > 0 we have that ε · ρ(pj) ≥ Y ∗ijYij. but ρ(T (1) Eij Hence we have that Yij = 0 for all (i, j) ∈ E(P ). Since T (P ) is generated by {pi}i∈Ω and {T (1) summand, so that eρ is a trivial dilation of ρT+(P ), and hence ρT+(P ) is maximal, and must then Eij}(i,j)∈E(P ), we must have that ψ has ρ as a direct We next define a notion that will help us detect when an irreducible πk is not a boundary have the unique extension property. (cid:3) representation for T+(P ). Definition 3.9. Let P be a finite irreducible stochastic matrix over Ω. A state k ∈ Ω is called exclusive if whenever for i ∈ Ω and n ∈ N we have P (n) ik = 1. We denote by Ωe the set of all exclusive states in Ω. ik > 0, then P (n) One should think of exclusive states as those states k such that for any n for which i leads to k in n steps, it cannot lead anywhere else in n steps. Lemma 3.10. Let P be a finite irreducible r-periodic stochastic matrix over Ω, and Ω0, ..., Ωr−1 be a cyclic decomposition for P . Suppose that k ∈ Ω0 is a state. (1) Ω0 > 1 if and only if k is non-exclusive. In this case, any state in Ω0 is non-exclusive (2) Assume k is non-exclusive and k 6= s ∈ Ω is some different state. If there is k 6= k′ ∈ Ω0 ks > 0 for all m ∈ N, then there exists n ∈ N such that and there is an n0 such that for any n ≥ n0 and i, j ∈ Ω0 we have 0 < P (rn) such that P (m) 0 < P (rn) kk < 1 and for all m ∈ N with (k, s) ∈ E(P m) we have P (rn) k′s > 0 whenever P (m) ks < P (rn+m) kk P (m) ij < 1. ks . Proof. (1): Suppose Ω0 > 1 and let k ∈ Ω0. By item (2) of Theorem 1.9 there is n0 such that for all n ≥ n0 we would have that P (rn) ij > 0 for all i, j ∈ Ω0 and n ≥ n0. Thus, for some j ∈ Ω0 we have that P (nr) jk < 1, and we conclude that k is non-exclusive. jk > 0, and since the j-th row sums up to 1 we get that 0 < P (nr) , P (nr) jj For the converse, suppose k is non-exclusive. We show that Ω0 > 1. Let k′ ∈ Ω and n0 be so that 0 < P (n0) k′k < 1, and let m0 be large enough so that P (m0) kk′ > 0. Then P (m0) kj = 1 kj P (n0) P (m0) P (m0+n0) kk =Xj∈Ω jk <Xj∈Ω 20 ADAM DOR-ON AND DANIEL MARKIEWICZ =Xj∈Ω kj P (n0) P (m0) jk ≥ P (m0) kk′ P (n0) k′k > 0 < 1. Since the k-th row sums up to 1, there must be an i ∈ Ω > 0 and by definition of the cyclic decomposition we have On the other hand, P (m0+n0) kk kk So we see that 0 < P (m0+n0) differnt from k such that P (m0+n0) that i ∈ Ω0. This shows that Ω0 > 1. would have P (rn) that P (rn) ki Next, by item (2) of Theorem 1.9 we may find n0 large enough so that for any n ≥ n0 we ij > 0 for all i, j ∈ Ω0. As Ω0 > 1, and all rows sum up to 1, we must also have ij < 1 for all i, j ∈ Ω0. Hence, we see that all states in Ω0 are non-exclusive. (2): By item (1) we can find n0 so that 0 < P (rn) m ∈ N with P (m) ks > 0, so that by assumption P (m) P (rn+m) is ≥ P (rn) ks P (rn) ki P (m) kk′ P (m) =Xi∈Ω ij < 1 for all i, j ∈ Ω0 and n ≥ n0. Now fix k′s > 0. Then k′s + P (rn) ks > P (rn) kk P (m) kk P (m) ks (cid:3) Proposition 3.11. Let P be a finite r-periodic irreducible matrix over Ω and Ω0, ..., Ωr−1 a cyclic decomposition for P . Let k ∈ Ω. (1) If k ∈ Ω0 is non-exclusive and for any other non-exclusive s 6= k there is some k 6= k′ ∈ (2) If k is exclusive then πk is not a boundary representation. ks > 0, then πk is a boundary representation. k′s > 0 whenever P (m) Ω0 such that P (m) Proof. (1): Assume k is non-exclusive. We use [Arv11, Theorem 7.2] to show that πk is a strongly peaking representation according to [Arv11, Definition 7.1]. Since the irreducible representations of T (P ) are given by Corollary 3.6, it suffices to find an element T ∈ T+(P ) such that kπk(T )k > k(evλ ◦ q)(T )k for any λ ∈ T and such that kπk(T )k > kπs(T )k for any k 6= s. , and wait until prescribing n is necessary. Recall Notation 3.2, so that Choose T = T (n) Ekk kπk(T )k ≥ kπ(T (n) Ekk )(e(0) kk )k = k 1 P (n) kk e(n) kk k = 1 P (n) kk On the other hand, q(T (n) Ekk ) = (z 7→ zmEkk) for m ∈ N satisfying n = rm, so that k(evλ ◦ q)(T )k = kevλ(z 7→ zmEkk)k = λm = 1 So we see that kπk(T )k > supλ∈T k(evλ ◦ q)(T )k. Next, fix s ∈ Ω with k 6= s. Since T ∗T ∈ L(FArv(P )) sends Arv(P )m to Arv(P )m, it is a finite-block diagonal operator, so we must have that T ∗TFP,s = (T (n) )FP,s is also finite-block diagonal. Denote I(k, s) = { m ∈ N (k, s) ∈ E(P m), m ≥ 1 }, and note that since TFP,s(Arv(P )0,s) = 0, we have that )∗(T (n) Ekk Ekk kπs(T )k2 = kπs(T ∗T )k = kT ∗TFP,sk = sup m∈NkT ∗TArv(P )m,sk = Ekk sup )∗(T (n) Ekk m∈I(k,s)k(T (n) ke(m) ks k = sup m∈I(k,s) By Theorem 1.10 we see that as m ∈ I(k, s) goes to infinity, the fraction the constant νsr ks )k = sup m∈I(k,s) )(e(m) P (m) ks P (n+m) ks P (m) ks P (n+m) ks P (m) P (n+m) ks ks converges to ≤ 1, as k is non-exclusive, we have that P (n) kk < 1 for large enough νsr = 1. Hence, if supm∈I(k,s) ks P (m) P (n+m) ks n, and so that kπk(T )k > kπs(T )k. On the other hand, if supm∈I(k,s) P (m) P (n+m) ks ks and s must be non-exclusive. By item (2) of Lemma 3.10 there is n large enough (which we now > 1, then the supremum above is in fact a maximum, C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 21 prescribe) so that 0 < P (n) kk < 1 and P (n) kk P (m) ks < P (n+m) ks for all m ∈ I(k, s). Hence, we see that 1 P (n) kk > ks P (m) P (n+m) ks for all m ∈ I(k, s) so that still we obtain kπk(T )k > kπs(T )k. so that by [Arv11, Theorem 7.2] we have that πk is a boundary representation. To conclude, we have shown that kπk(T )k > max{sups6=k{kπs(T )k}, supλ∈T k(evλ ◦ q)(T )k} A ) = πk(W (n) (2): Suppose that k is exclusive. By the formula for T (n) A ). Indeed, this follows since any weights appearing in an application of T (n) A to a k-th column of a matrix B ∈ Arv(P )m arise only from entries of the k-th columns of P n, which are either 0 or 1 by assumption on k. We will use the above to show that πk is not strongly peaking anywhere by showing that it is not strongly peaking at anyPN n=−N [Tij] ∈ Ms(T+(P )∗ + T+(P )) where each Tij ∈ T+(P )∗ + A or T (n)∗A ). We also T+(P ) is of degree n ∈ [−N, N ] (which must then be either of the form T (n) denote by Wij the element (which is either of the form W (n) respectively) satisfying πk(Tij) = πk(Wij) above. A , we see that πk(T (n) A or W (n)∗A We note finally that there exists m0 such that for all m ≥ m0 we have that (U m)∗WijU m = Wij for all i, j. We then have that [Tij])k = k NXn=−N NXn=−N [(U m ⊗ Id)∗πk(Wij)(U m ⊗ Id)]k = k [πk(Tij)]k = k kπ(s) k ( NXn=−N ≤ k[U m∗( k Wij)U m]k ≤ k[Q[m,∞)( NXn=−N NXn=−N [πk(Wij)]k = NXn=−N NXn=−N [πk(U m∗WijU m)]k NXn=−N NXn=−N Wij)]k = kq(s)( Wij)Q[m,∞)]k So we see that by Proposition 1.5 m→∞k[( [Tij])k ≤ lim Wij)Q[m,∞)]k = k[q( NXn=−N kπ(s) k ( Since q(s)(PN n=−N [Tij]) ∈ C(T, Md) ⊗ Ms, there exists λ ∈ T such that kπ(s) k ( NXn=−N Since elements of the formPN n=−N [Tij] with Tij of degree n are dense inside Ms(T+(P )∗+T+(P )), we see that for any [Vij] ∈ Ms(T+(P )∗ +T+(P )) we have k[πk(Vij)]k ≤ supλ∈T k[(evλ ◦ q)(Vij)k so that πk cannot be strongly peaking. By [Arv11, Theorem 7.2] we see that πk is not a boundary representation. (cid:3) [Tij])k = k(evλ ◦ q)(s)( [Tij])k ≤ kq(s)( NXn=−N NXn=−N NXn=−N [Tij])k [Tij])k Remark 3.12. It is clear that for exclusive k ∈ Ω the representation πk is not boundary by item (2) of Propositon 3.11. Item (1) in Proposition 3.11 above provides a sufficient condition for πk to be boundary when k is non-exclusive. We believe that this condition is not necessary, however we do not have examples to that effect. We next introduce a class of stochastic matrices for which we can completely identify the non-commutative Choquet boundary of T+(P ) inside T (P ) in terms of the matrix P . Definition 3.13. Let P be a finite r-periodic irreducible stochastic matrix over Ω. We say that P has the multiple-arrival property if whenever k, s ∈ Ω are distinct non-exclusive states such that whenever k leads to s in n steps, then there exists k 6= k′ ∈ Ω such that k′ leads to s in n steps. 22 ADAM DOR-ON AND DANIEL MARKIEWICZ Corollary 3.14. Let P be a finite irreducible stochastic matrix over Ω, and k ∈ Ω. If P has the multiple-arrival property, then πk is a boundary representation if and only if k is non-exclusive. Hence, the non-commutative Choquet boundary of T+(P ) inside T (P ) is parameterized by a circle T of irreducible representations of dimension d, and irreducible representations πk of infinite dimension associated to non-exclusive states k ∈ Ω. Proof. This follows directly since if P has multiple-arrival, then the conditions of Proposition 3.11 item (1) are automatically satisfied for any non-exculsive k ∈ Ω, and item (2) of Proposition 3.11 then gives the reverse implication. (cid:3) There is an easy class of examples which automatically has the multiple arrival property. Sup- pose that P is an irreducible r-periodic stochastic matrix with cyclic decomposition Ω0, ..., Ωr−1. Then we may write 0 0  0 P0 ··· ... ... ... 0 Pr−2 ··· Pr−1 ··· 0 ... 0  for rectangle stochastic matrices P0, ..., Pr−1. If all entries of the matrices P0, ..., Pr−1 are non- zero, then P is called fully-supported, and has the multiple-arrival property. Suppose P is a finite irreducible stochastic matrix P over Ω of size d. We next discuss C∗env(T+(P )) and its spectrum. Denote by Ωb the set of states k for which πk is a boundary representation, which is a subset of Ω − Ωe. Since for all k ∈ Ω and λ ∈ T we have that Ker πk ⊂ J ⊂ ker evλ ◦ q, and since the intersection of kernels of all boundary representations is the Shilov ideal, we must have that π−1(⊕k∈Ω−ΩbK(FP,k)) is the Shilov ideal of T+(P ) inside T (P ), thought of as a subalgebra of π−1(⊕k∈ΩK(FP,k)) = J . We hence get the following short exact sequence 0 −→ ⊕k∈ΩbK(FP,k) −→ C∗env(T+(P )) −→ C(T, Md) −→ 0 (3.1) while we identify qe(J (P )) ⊂ C∗env(T+(P )) with ⊕k∈ΩbK(FP,k), where qe : T (P ) → C∗env(T+(P )) is the quotient map by the Shilov ideal. If ρ : C∗env(T+(P )) → B(H) is a unital -representation, it decomposes uniquely into a central direct sum of representations ρ = ρJ ⊕ρO, where ρJ is the unique extension to C∗env(T+(P )) of the restriction of ρ to qe(J (P )), and ρO annihilates qe(J (P )). Hence, the spectrum of C∗env(T+(P )) decomposes into a disjoint union of the spectrum of ⊕k∈ΩbK(FP,k) and the spectrum of C(T, Md). That is, the spectrum of C∗env(T+(P )) is comprised of Ωb irreducible representations of infinite dimension, and a torus T of irreducible representations of dimension d that annihilate qe(J (P )). Theorem 3.15. Suppose that P is a finite irreducible matrix over Ω. Then T+(P ) is hyperrigid in C∗env(T+(P )). Moreover, if P has multiple-arrival, the Shilov ideal for T+(P ) inside T (P ) is given by SP = \k∈Ω−Ωe { T ∈ J (P ) πk(T ) = 0 } and is -isomorphic via π to ⊕k∈ΩeK(FP,k) Proof. Let ρ : C∗env(T+(P )) → B(H) be a -representation. By the above discussion, we may decompose it into a central direct sum of representations ρ = ρJ ⊕ ρO, where ρJ is the unique extension to C∗env(T+(P )) of the restriction of ρ to qe(J (P )), and ρO annihilates qe(J (P )). By Proposition 3.8 we have that ρO ◦ qe has u.e.p. when restricted to T+(P ), so that ρO has u.e.p. when restricted to qe(T+(P )) by invariance of maximal UCP maps. Next, since ρJ ◦ qe = ⊕k∈Ωbnk · πk is a direct sum of -representations, with certain multiplicities nk, that have u.e.p. when restricted to T+(P ), by [Arv11, Theorem 4.4] ρJ ◦qe has u.e.p. when restricted to T+(P ). Hence, again by invariance of maximal UCP maps, ρJ has u.e.p. when restricted to qe(T+(P )). By another application of [Arv11, Theorem 4.4] we obtain that ρ = ρJ ⊕ ρO also has u.e.p. when restricted to qe(T+(P )), so that T+(P ), which is completely isometric to qe(T+(P )) via qe, is hyperrigid within C∗env(T+(P )). C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 23 For the second part, by Corollary 3.14 we know that Ωe = Ω−Ωb. Furthermore, by Proposition 3.8, we have that evλ ◦ q is a boundary representation for T+(P ) for any λ ∈ T, and since J (P ) =Tλ∈T Ker(evλ ◦ q), by the discussion preceding the theorem, the Shilov ideal must equal J (P ) ∩ \k∈Ωb Ker(πk) = \k∈Ω−Ωe { T ∈ J (P ) πk(T ) = 0 } We now give equivalent conditions that guarantee that the C-envelope of T+(P ) is either the Toeplitz algebra, or the Cuntz-Pimsner algebra. (cid:3) Corollary 3.16. Let P be a finite irreducible stochastic matrix of size d with multiple-arrival. In particular, if P aperiodic and of size d ≥ 2 with multiple-arrival, we have that C∗env(P ) ∼= Then we have that C∗env(T+(P )) ∼= T (P ) if and only if all k ∈ Ω are non-exclusive. T (P ). Proof. By Theorem 3.15 we see that if all k ∈ Ω are non-exclusive, then SP = {0} and so T (P ) is the C-envelope of T+(P ). Conversely, if T (P ) ∼= C∗env(T+(P )), then C∗env(T+(P )) has d irreducible representations of infinite dimension, which can only occur if Ωb = d. Since P has multiple-arrival, we see by Corollary 3.14 that Ωb = Ω − Ωe, and so all states k ∈ Ω are non-exclusive. For the second part, our assumptions guarantee that Ω = Ω0 is the cyclic decomposition for P , and Ω > 1, so by Lemma 3.10 all elements in Ω are non-exclusive. Since P has multiple- arrival, by Corollary 3.14 we have that πk is a boundary representation for any k ∈ Ω, and so SP = {0} and C∗env(T+(P )) ∼= T (P ). (cid:3) Corollary 3.17. Let P be a finite r-periodic irreducible stochastic matrix of size d, and let Ω0, ..., Ωr−1 be a cyclic decomposition for P . The following are equivalent: (1) All k ∈ Ω are exclusive. (2) Ωℓ = 1 for all ℓ ∈ Zr, or equivalently r = d. (3) P : ℓ∞(Ω) → ℓ∞(Ω) is a -homomorphism. (4) C∗env(T+(P )) ∼= O(P ) Proof. (1) ⇒ (2): By item (1) of Lemma 3.10 we see that Ωℓ = 1 for all ℓ ∈ Zr. in fact a permutation matrix of a single-cycle permutation, and is hence a homomorphism. (2) ⇒ (3): If all Ωℓ are of size 1, we see that the cyclic decomposition for P yields that P is (3) ⇒ (4): The Arveson-Stinespring construction of a subproduct system generally yields a product system when applied to a -homomorphism. Hence, Arv(P ) is a product system, and its tensor algebra is the tensor algebra of a single correspondence, and by [Vis12, Proposition 2.8] this is also true for the Cuntz-Pimsner algebra in our case. By [KK06, Theorem 3.7] we have C∗env(T+(P )) ∼= O(P ). this case, let n be so that 0 < P (n) Ekk ∈ Arv(P )0 we have (4) ⇒ (1): Assume towards contradiction that there is some k ∈ Ω that is non-exclusive. In )k = 1, while for kk < 1, and observe that kq(T (n) )k = kq(W (n) Ekk Ekk kT (n) Ekkk ≥ kT (n) Ekk (E(0) kk )k = 1 P (n) kk kE(n) kk k = 1 P (n) kk This means that q : T (P ) → O(P ) is not isometric on T+(P )∗ + T+(P ), and in particular, not completely isometric on T+(P )∗ + T+(P ). By [Arv11, Theorem 7.2], there is a boundary representation for T+(P ) coming from an extension to T (P ), of an element in the spectrum of J (P ), which then must be equivalent to one of the πk. This means that C∗env(T+(P )) has an irreducible representation of infinite dimension, which is impossible since C∗env(T+(P )) ∼= C(T, Md) only has irreducible representations of dimension d. (cid:3) 24 ADAM DOR-ON AND DANIEL MARKIEWICZ Example 3.18. We next give an example of 3 × 3 stochastic matrix for which C∗env(T+(P )), T (P ) and O(P ) are pairwise non -isomorphic. Let P = 0 0 1 2 0 1 0 1 1 2 0 The matrix P is fully-supported and we see that states 1 and 2 are non-exclusive, while 3 is exclusive. Hence, Ωb = Ω − Ωe ( Ω. Therefore, the Shilov ideal SP ∼= K(H1) ⊕ K(H2) 6∼= ⊕j∈ΩK(Hj) ∼= J (P ). This yields a quotient C∗env(T+(P )) for which C∗env(T+(P )), T (P ) and O(P ) are pairwise non -isomorphic. Without the irreducibility assumption on P , it is easy to construct intermediary C-envelopes from "extremal" C-envelopes. Indeed, if for finite stochastic matrices P and Q of sizes at least 2 we have that C∗env(T+(P )) = T (P ), and C∗env(T+(Q)) = O(Q), then R = P ⊕ Q is a finite stochastic matrix such that C∗env(T+(R)) = C∗env(T+(P )) ⊕ C∗env(T+(Q)) = T (P ) ⊕ O(Q), and one can similarly use representation theory to show that C∗env(T+(R)) is non -isomorphic to T (R) nor O(R). However, when P is irreducible, the subproduct system Arv(P ) associated to it, cannot have any non-trivial reducing projections in the sense of Remark 1.14, so we have irreducibility both in a dynamical sense and in a sense of its subproduct system. The above example then shows that even under this minimality / irreducibility assumptions on a dynamical object which is equivalent to this irreducibility assumptions on the subproduct system above, up to -isomorophism the C-envelope may be distinct from both the Cuntz-Pimsner algebra, and the Toeplitz algebra. 4. K-Theory In this section we compute the K-theory of C∗env(T+(P )). Recall Notation 3.2. Let Ωb be the collection of k ∈ Ω for which πk is a boundary representation. We note immediately that we identify T (P ) with its image under π : T (P ) → B(⊕k∈ΩFP,k), where FP,k are the invariant subspaces of π and πk : T (P ) → B(FP,k) the irreducible, pairwise non-unitarily equivalent representations given by restriction πk(T ) = TFP,k , for each k ∈ Ω. Recall the short exact sequence from (3.1). We know from [RLL00] that K0 and K1 are additive functors, and that for any k ∈ Ω we have K1(K(FP,k)) = {0}, and K0(C(T, Md)) ∼= K0(K(FP,k)) ∼= K1(C(T, Md)) ∼= Z. Hence, the six-term exact sequence of K-theory induced from the exact sequence (3.1) yields (4.1) 0 −→ K1(C∗env(T+(P ))) −→ Z ↑ ↓ δ1 Z ←− K0(C∗env(T+(P ))) ←− ZΩb Our goal in this section is to compute the index map δ1 : K1(C(T, Md)) → K0(⊕k∈ΩbK(FP,k)), which will then enable the computation of the K0 and K1 groups for C∗env(T+(P )). It will suffice to compute the value of δ1 on a generator of K1(C(T, Md)) ∼= Z, and in our computations we will work with the unitary element w := z 7→ diag(z, 1, ..., 1) ∈ C(T, Md), as [w]1 is a generator for K1(C(T, Md)) ∼= Z. Lemma 4.1. Let P be an r-periodic irreducible stochastic matrix over Ω of size d, with properly enumerated cyclic decomposition Ω0, ..., Ωr−1 such that 1 ∈ Ω0 is the first element, and let (U, (Sij)i,j∈Ω) be its associated standard family. that commute with U . (1) For all i ∈ Ω we have that (Sii = pi)i∈Ω is a family of pairwise orthogonal projections (2) The element w := z 7→ diag(z, 1, ..., 1) ∈ C(T, Md) lifts to a partial isometry V := U S11 + S22 + ...Sdd inside T (P ). Proof. (1): By definition, for any m ∈ N and Ejk ∈ Arv(P )m we have that Sii(Ejk) = Gr(P m)∗ (Eii · Ejk) = δi,jEjk = pi(Ejk) so that Sii = pi. Next, note that U Sii(Ejk) = Gr(P m+r) ∗ (δijEjk) = δijU (Ejk) = SiiU (Ejk) C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 25 So that U and Sii commute on the dense subset of FArv(P ), and hence commute. (2): It is clear that w lifts to V inside T (P ) since under the identification C(T) ⊗ Md ∼= C(T; Md), the element V in the quotient is identified with w. Hence, we need only verify that V is a partial isometry. Indeed, since by item (1), U commutes with S11, and since U is a partial isometry, we have that V V ∗V = S11U U∗U S11 + S22 + ...Sdd = V so that V is also a partial isometry. (cid:3) Let v be the image of V under the C-envelope quotient map qe : T (P ) → C∗env(T+(P )). By Lemma 4.1 we know that V is a partial isometry that lifts w, and hence v is a partial isometry that lifts w. By item (ii) of [RLL00, Proposition 9.2.5] we have that 1 − v∗v and 1 − vv∗ are projections in ⊕k∈ΩbK(Hk) ∼= qe(J (P )) with δ1([w]1) = [1 − v∗v]0 − [1 − vv∗]0 But due to the identification K0(⊕k∈ΩbK(FP,k)) ∼= ⊕k∈ΩbK0(K(FP,k)), we obtain that δ1([w]1) = [1 − v∗v]0 − [1 − vv∗]0 =(cid:16)[(1 − V ∗V )FP,k ] − [(1 − V V ∗)FP,k ](cid:17)k∈Ωb = (cid:16) dim Ker(V FP,k ) − dim Ker(V ∗FP,k )(cid:17)k∈Ωb =(cid:0) ind(V FP,k )(cid:1)k∈Ωb Where we are then left with computing the Fredholm indices of V FP,k for k ∈ Ωb. Proposition 4.2. Let P be an r-periodic irreducible stochastic matrix over Ω = {1, ..., d}. Suppose that Ω0, ..., Ωr−1 is a properly enumerated cyclic decomposition such that s ∈ Ω is its first element, and let (U, (Sij)i,j∈Ω) be its associated standard family. Let Vs := S11 + ... + Ss−1,s−1 + U Sss + Ss+1,s+1 + ... + Sdd. Then for every k ∈ Ω we have that ind(VsFP,k ) = −1. Proof. Up to conjugating P with a permutation matrix, we may assume that s = 1 is the first element. For each state k ∈ Ω, let ℓ = σ(k) = σ(k) − σ(1), and denote by bn =(1 : P (nr+ℓ) 0 : P (nr+ℓ) 1k 1k > 0 = 0 where P (0) 1k = 1 if k = 1, and is 0 otherwise. Recall Notation 3.2. Since FP,k = ⊕∞n=0Arv(P )n,k, and since V shifts only the first rows of the matrix A in an element A ⊗ ek ∈ Arv(P )n,k, we have for all n ∈ N that and for all n ≥ 1 that due to the support of elements in Arv(P )n,k. Note also that for n = 0, and we get dim Ker V Arv(P )n,k = bn − bn+1bn dim Ker V ∗Arv(P )n,k = bn+1 − bn+1bn dim Ker V ∗Arv(P )0,k =(0 : k 6= 1 1 : k = 1 Hence, if we sum up dimensions, we obtain that and Thus, dim Ker V FP,k = ∞Xn=0 bn − bn+1bn dim Ker V ∗FP,k =(P∞n=0 bn+1 − bn+1bn 1 +P∞n=0 bn+1 − bn+1bn ind(V FP,k ) = dim Ker V FP,k − dim Ker V ∗FP,k : k 6= 1 : k = 1 26 ADAM DOR-ON AND DANIEL MARKIEWICZ =(P∞n=0 bn − bn+1 −1 +P∞n=0 bn − bn+1 if k 6= 1 if k = 1 = (0 − 1 −1 + 1 − 1 if k 6= 1 if k = 1 Hence, we see that in any case, ind(V FP,k ) = −1, as required. Corollary 4.3. Let P be an irreducible stochastic matrix over finite Ω. Then the index map δ1 : Z → ZΩb is given by δ1(n) = −(n, ..., n) (cid:3) We then obtain the K-theory of C∗env(T+(P )) in terms of Ωb. Theorem 4.4. Let P be a finite irreducible stochastic matrix over Ω. Then (1) If P has a non-exclusive state then K0(C∗env(T+(P ))) ∼= ZΩb and K1(C∗env(T+(P ))) ∼= {0} (2) If all states of P are exclusive then K0(C∗env(T+(P ))) ∼= Z and K1(C∗env(T+(P ))) ∼= Z Proof. If all states of P are exclusive, then by Corollary 3.17 we have that C∗env(T+(P )) ∼= C(T, Md) so that the K0 and K1 groups of C∗env(T+(P )) must both be Z. Next, if P has a non-exclusive state, since δ1 is injective, by exactness at K1(C(T, Md)) in the six-term exact sequence (4.1), we see that K1(C∗env(T+(P ))) = {0}. the single exact sequence Since δ1(1) = (−1, ...,−1), we see that the six-term exact sequence (4.1) can be reduced to 0 ← Z ← K0(C∗env(T+(P ))) ← ZΩb/SpZ(cid:0)(−1, ...,−1)(cid:1) ← 0 Since ZΩb/SpZ(cid:0)(−1, ...,−1)(cid:1) ∼= ZΩb−1, we see that K0(C∗env(T+(P ))) ∼= ZΩb, and the proof is through (4) of Corollary 3.17 are all equivalent to K1(C∗env(T+(P ))) ∼= Z. Corollary 4.5. Let P be a finite irreducible stochastic matrix over Ω. Then conditions (1) complete. (cid:3) 5. Classification of C-envelope We are now in a position to apply the theory in the previous sections to obtain classifica- tion results up to -isomorphism and stable isomorphisms of C-envelopes arising from finite irreducible stochastic matrices. For every finite irreducible stochastic matrix P over ΩP , which has at least one non-exclusive b be the (non-empty) set of indices k ∈ ΩP such that πk : T (P ) → B(FP,k) state, let ΩP is a boundary representation for T+(P ). We note that C∗env(T+(P )) is Type I (equivalently GCR), being an extension of a CCR algebra by a CCR algebra. Thus, we may identify an irre- ducible representation with its kernel when discussing elements of the primitive ideal spectrum of C∗env(T+(P )). The analysis done around the exact sequence (3.1) shows that the spectrum of C∗env(T+(P )) as a set is comprised of ΩP b irreducible representations of infinite dimensions induced from πk, which we still denote by πk : C∗env(T+(P )) → B(FP,k) for k ∈ ΩP b , and a torus T of irreducible representations of dimension ΩP given by evλ ◦ q for every λ ∈ T, where q : C∗env(T+(P )) → C(T, MΩP) is the quotient map. Moreover, we have the exact sequence 0 −→ ⊕k∈ΩP b K(FP,k) ι−→ C∗env(T+(P )) q −→ C(T, MΩP ) −→ 0 Since Ker πk ⊆ Ker(evλ ◦ q) for every k ∈ ΩP b and every λ ∈ T, and Ker(evλ ◦ q) is not a subset of any Ker(evλ′ ◦ q) for λ′ 6= λ ∈ T, we see that for every λ ∈ T, each Ker(evλ ◦ q) is a maximal element in the lattice P rim(C∗env(T+(P ))). Notation 5.1. For a finite irreducible stochastic matrix P , we denote from now on KP := ⊕k∈ΩP b K(FP,k), BP := C(T, MΩP ) and AP := C∗env(T+(P )). C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 27 Let P and Q be irreducible stochastic matrices over finite sets ΩP and ΩQ respectively. Then we have the following exact sequences (5.1) 0 → KP → AP → BP → 0 and 0 → KQ → AQ → BQ → 0 b b η(∞) P k τ (k)η(∞) b → ΩQ with Busby invariants ηP and ηQ, and the stabilized exact sequences (5.2) 0 → KP ⊗ K → AP ⊗ K → BP ⊗ K → 0 and 0 → KQ ⊗ K → AQ ⊗ K → BQ ⊗ K → 0 with Busby invariants η(∞) and η(∞) Q given by η(∞) B(FP,ℓ) → B(FP,k) the restriction map, which then promotes P ([Tij]) = [ηP (Tij)] for [Tij] ∈ BP ⊗ K. For k ∈ Ωb denote by ρk : ⊕ℓ∈ΩP P b b Q(FP,ℓ) → Q(FP,k). Proposition 5.2. Let P and Q be finite irreducible stochastic matrices over ΩP and ΩQ respec- tively. to a restriction eρk : ⊕ℓ∈ΩP (1) C∗env(T+(P )) and C∗env(T+(Q)) are -isomorphic if and only if there exists a -isomorphism b such that for all k ∈ ΩP β : C(T, MΩP) → C(T, MΩQ) and a bijection τ : ΩP the extensions eρkηP and eρτ (k)ηQβ are strongly equivalent. (2) C∗env(T+(P )) and C∗env(T+(Q)) are stably isomorphic if and only if there exists a - isomorphism β : C(T, MΩP ) ⊗ K → C(T, MΩQ)⊗ K and a bijection τ : ΩP b such and eρ(∞) that for all k ∈ ΩP b → ΩQ Q β are weakly equivalent. the extensions eρ(∞) λ′ ◦ q) for some λ′ ∈ T in bijection. In particular, since KP = ∩λ∈T Ker(evP Proof. We first show (1). Suppose that α : AP → AQ is a -isomorphism. Let α∗ : P rim(AP ) → P rim(AQ) be the induced lattice isomorphisms between the spectra. Since α∗ must send max- imal elements to maximal elements, we see that for λ ∈ T we have that α∗ sends Ker(evP λ ◦ q) to Ker(evQ λ ◦ q) and KQ = ∩λ∈T Ker(evQ λ ◦ q), we see that α(KP ) = KQ. Hence C∗env(T+(P )) and C∗env(T+(Q)) are -isomorphic if and only if the exact sequences of (5.1) are isomorphic, which happens if and only if the restriction κ := αKP : KP → KQ and the induced map β satisfyeκηP = ηQβ, where So suppose eκηP = ηQβ for κ and β as above. Since κ : ⊕k∈ΩP eρτ (k)ηQβ =eρτ (k)eκηP = ]AdUkeρkηP β : BP → BQ is the induced -isomorphism from α between the quotients by KP and by KQ. b K(FQ,k), b and a unitaries Uk : FP,k → FQ,τ (k) such that κK(FP,k) = For the converse, if Uk are unitaries implementing the strong conjugacy between eρτ (k)ηQβ and ]AdUkeρkηP , by setting κ = ⊕k∈Ωb eκηP = ⊕k∈ΩP Next, we show (2). Since stabilizing an algebra does not change its primitive ideal spectrum, the same argument as used in (1) shows that C∗env(T+(P )) and C∗env(T+(Q)) are stably isomorphic if and only if the exact sequences in (5.2) are isomorphic, which happens if and only if there are Q β. Then a similar argument to the one used for item (1) shows that this happens if and only if there is a bijection τ : ΩP Q β are strongly equivalent. Since these are non-unital extensions, this happens if and only if they are weakly equivalent. (cid:3) -isomorphisms κ : KP ⊗ K → KQ ⊗ K and β : BP ⊗ K → BQ ⊗ K such that eκη(∞) and eρ(∞) b the extensions eρ(∞) there is a bijection τ : ΩP AdUk : K(FP,k) → K(FQ,τ (k)), so that b → ΩQ AdUkeρkηP = ⊕k∈ΩQ b eρτ (k)ηQβ = ηQβ b such that for all k ∈ ΩP b K(FQ,k), we have that b K(FP,k) → ⊕k∈ΩQ AdUk : ⊕k∈ΩP b K(FP,k) → ⊕k∈ΩQ P P = η(∞) η(∞) P k τ (k)η(∞) b b → ΩQ For an irreducible finite stochastic matrix P over ΩP with period rP , and k ∈ ΩP . Let Ω0, ..., ΩrP −1 be a cyclic decomposition for P . Then there exists m0, such that for all m ≥ m0 we have Ωσ(k)−m = { i ∈ Ωσ(k)−m P (m) ik > 0 } ∞Xm=1 28 ADAM DOR-ON AND DANIEL MARKIEWICZ Indeed, fix 0 ≤ ℓ ≤ rP − 1. By item (2) of Theorem 1.9 there is n(ℓ) we have that P (nrP +ℓ) such that for all n ≥ n(ℓ) > 0 for i, j ∈ ΩP with σ(i)− σ(j) = ℓ. Hence, if we fix j = k, we see that Ωσ(k)−(nrP +ℓ) = { i ∈ Ωσ(k)−(nrP +ℓ) P (nrP +ℓ) ij 0 0 > 0 } 0 rP + ℓ} to obtain the desired claim above. ik Then simply take m0 = maxℓ{n(ℓ) Definition 5.3. Let P be an r-periodic finite irreducible stochastic matrix over Ω of size d, and k ∈ Ω. Let Ω0, ..., Ωr−1 be a cyclic decomposition for P , so that σ(k) is the unique index such that k ∈ Ωσ(k). We define the k-th column nullity of P to be { i ∈ Ωσ(k)−m P (m) ik = 0 } NP (k) = where σ(k) − m is taken as an element in the cyclic group Zr of order r. We say that k ∈ Ω is a fully supported column if NP (k) = 0. columns of iterations of P , that lie in the support of a cyclic decomposition for P . Put in other words, the column nullity of a state k ∈ Ω is the number of zeros in all k-th The above infinite sum is in fact always finite by the discussion preceding Definition 5.3 and is hence convergent. For a finite irreducible stochastic matrix P , we find the element in Exts(C(T) ⊗ Md) repre- b , appearing in Proposition 5.2. Note that the exact sequence corresponding to the extension ηP,k is senting each extension ηP,k := eρkηP , for each k ∈ ΩP 0 → K(FP,k) → πk(T (P )) → C(T, MΩP ) → 0 Recall the computation of Exts(C(T) ⊗ Md) and Extw(C(T) ⊗ Md) preceding Proposition 2.5. Proposition 5.4. Let P be a finite irreducible stochastic matrix over ΩP with period rP , and let Ω0, ..., ΩrP −1 be a properly enumerated cyclic decomposition for P . Then for each k ∈ ΩP b , there exists n0 large enough so that for all n ≥ n0 we have that [j∗ηP,k]s is identified with 0 ≤ s < ΩP given by s ≡ nrP −1Xm=0 { i ∈ Ωσ(k)−m P (m) ik > 0} mod ΩP and [ι∗ηP,k]s is identified with −ΩP. In particular, [ηP,k]w = −1. Proof. To compute the class of [j∗ηP,k], we apply the algorithm in Example 1.17 to j∗ηP,k. Let {Sij} be the system of matrix units for C(T, MΩP) associated to a properly enumerated cyclic decomposition Ω0, ..., ΩrP −1, and let 1 ∈ Ω be the first element in this enumeration. There then exists m0 such that for all m ≥ m0 we have Ωσ(k)−m = { i ∈ Ωσ(k)−m P (m) ik > 0 }. We abuse notation for sake of brevity and write T instead of πk(T ) = TFP,k for T ∈ T (P ). Lift each Sii to pi·Q[nrP ,∞) ∈ πk(T (P )). Then we may lift each S1j to S1jQ[nrP ,∞) ∈ πk(T (P )). Hence, we get that eij := Q[nrP ,∞)S∗1jS1iQ[nrP ,∞), so that for all n ∈ N with nrP ≥ m0, p = Q[nrP ,∞)S∗1iS1iQ[nrP ,∞) ΩP Xi=1 Denote by b(m) ik the indicator, which is 1 if and only if P (m) ik > 0 and 0 otherwise. Then, the dimension of the cokernel of p is congruent mod ΩP to ΩP Xi=1 nrP −1Xm=0 nrP −1Xm=0 b(m) ik = { i ∈ Ωσ(k)−m P (m) ik > 0 } so we may take n0 = ⌈ m0 rP ⌉. As for ι∗ηP,k, a lift for z ⊗ I ∈ C(T) ⊗ I can be taken to be U P , where U P is the unitary associated to the properly enumerated cyclic decomposition Ω0, ..., ΩrP −1 (restricted to FP,k). C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 29 Then by Proposition 4.2 and the notation there, ind(U P ) = ind(Πi∈ΩVi) = −ΩP. Finally, recall that the image [ι∗ηP,k]s ∈ ΩP · Z is identified with [ηP,k]w ∈ Z up to dividing by ΩP, so that [ηP,k]w = −1. (cid:3) We now reach the two main results of this paper, which classify stable isomorphism and -isomorphism of C-envelopes in terms of the underlying stochastic matrices and boundary representations supported on different copies of compact operator subalgebras. b = ΩQ P,k (cid:3) and η(ΩP ) Q,k P,k and η(ΩP ) Q,k For the converse, suppose Ωb := ΩP b = ΩQ Theorem 5.5. Let P and Q be finite irreducible stochastic matrices over ΩP and ΩQ respectively. Then ΩP Proof. If C∗env(T+(P )) and C∗env(T+(Q)) are stably isomorphic, since K0 and K1 are stable func- tors, we must have that ΩP b if and only if C∗env(T+(P )) and C∗env(T+(Q)) are stably isomorphic. b by Theorem 4.4. b = ΩQ are weakly unitarily equivalent. Hence, η(∞) b . For k ∈ Ωb, denote by η(ΩQ) P,k and η(∞) the ampliations of these extensions to C(T) ⊗ MΩP ⊗ MΩQ. By Proposition 5.4 we then have that η(ΩQ) Q,k are also weakly equivalent, so that by item (2) of Proposition 5.2 (with β = Id) we have that C∗env(T+(P )) and C∗env(T+(Q)) are stably isomorphic. Theorem 5.6. Let P and Q be finite irreducible stochastic matrices over ΩP and ΩQ respectively. Then C∗env(T+(P )) and C∗env(T+(Q)) are -isomorphic if and only if d := ΩP = ΩQ and there is a bijection τ : ΩP Proof. Suppose C∗env(T+(P )) and C∗env(T+(Q)) are -isomorphic. By item (1) of Proposition 5.2 there is an -isomorphism β ∈ C(T, MΩP ) → C(T, MΩQ) (so that d := ΩP = ΩQ) and a bijection τ : ΩP b such that ηP,k and ηQ,τ (k)β are strongly equivalent. By Proposition 2.5 βs is the identity on the second coordinate of Exts(C(T) ⊗ Md) ∼= dZ × Zd. Hence, we see that [j∗ηP,k] = [j∗ηQ,τ (k)], so that k ∈ ΩP b via some bijection τ : ΩP b , and that ΩP = ΩQ. We see by Proposition 5.4 that j∗ηP,k and j∗ηQ,τ (k) are strongly equivalent. Again by Proposition 5.4 we have that [ι∗ηP,k] and [ι∗ηQ,τ (k)] are represented by the numbers −ΩP and −ΩQ which are equal by assumption. Hence, we have that ηP,k and ηQ,τ (k) are strongly equivalent. Thus, by item (1) of Proposition 5.2 (with β = Id) we have that C∗env(T+(P )) and C∗env(T+(Q)) are -isomorphic. (cid:3) For the converse, suppose NP (k) ≡ NQ(τ (k)) mod d for all k ∈ ΩP b and NP (k) ≡ NQ(τ (k)) mod d by Proposition 5.4. b we have NP (k) ≡ NQ(τ (k)) mod d. b → ΩQ b such that for all k ∈ ΩP b → ΩQ b → ΩQ It is interesting to try and compare these invariants with the one obtained from the graph C- algebra of the graph of the stochastic matrix P . Given an irreducible graph matrix A = (aij) over Ω, where aij ∈ {0, 1}, in their first paper [CK80], Cuntz and Krieger defined a C-algebra OA generated by partial isometries {Si}i∈Ω with pairwise orthogonal ranges, satisfying the relation S∗i Si =Xj∈Ω aij · SjS∗j For a stochastic matrix P , one has the {0, 1}-matrix Gr(P ) representing the directed graph of P . Since the C-correspondence Arv(P )1 is exactly the graph C-correspondence of Gr(P ), we get that the Cuntz-Pimsner algebra O(Arv(P )1) is -isomorphic to the Cuntz-Krieger algebra OGr(P ). In particular, by [Kat04, Corollary 7.4] we see that OGr(P ) is nuclear. In [Cu81], Cuntz computed the K-theory of these C-algebras. He showed that for finite {0, 1} matrix A over Ω where every column and row is non-zero, the K0 and K1 groups of OA are given as the cokernel and kernel of the map I − At : ZΩ → ZΩ. In the case where A is an irreducible finite matrix which is not a permutation matrix, Cuntz and Krieger establish in [CK80] that OA is simple and purely infinite. Hence, for a finite irreducible stochastic matrix P which is not a permutation matrix, the Cuntz-Krieger algebra OGr(P ) is separable, unital, nuclear, simple and purely infinite, or in other words a Kirchberg algebra. 30 ADAM DOR-ON AND DANIEL MARKIEWICZ A famous classification theorem of Kirchberg and Phillips then comes into play to show that for two finite irreducible stochastic matrices P and Q which are not permutation matrices, the Cuntz-Krieger algebras OGr(P ) and OGr(Q) are -isomorphic ( or stably isomorphic) if and only if (K0(OGr(P )), [1P ]0) ∼= (K0(OGr(Q)), [1Q]0) and K1(OGr(P )) ∼= K1(OGr(Q)) ( or K0(OGr(P )) ∼= K0(OGr(Q)) and K1(OGr(P )) ∼= K1(OGr(Q)) respectively). That is, the -isomorphism and stable isomorphism class are completely determined by K-theory. Example 5.7. In this example, we will use the above to show that for a finite irreducible sto- chastic matrix P , the Cuntz-Krieger algebra OGr(P ) and the C-envelope C∗env(T+(P )) generally yield incomparable invariants for P . If we restrict to matrices P with multiple-arrival, we have that Ωe = Ω − Ωb and the invariant C∗env(T+(P )) will only depend on the graph Gr(P ). Hence, we will only specify the {0, 1} graph incidence matrices of three stochastic matrices P, Q, R. Suppose the graph matrices for P, Q, R are given respectively by Gr(P ) = 0 0 1 0 0 1 1 1 0 , Gr(Q) = 1 1 0 1 1 1 1 1 1 , Gr(R) = 1 1 1 1 1 1 1 1 0 then P, Q and R have multiple-arrival, and it is clear that NP (j) = NQ(j) = NR(j) = 0 for j = 1, 2, and that NP (3) = 0. We also see that NQ(3) = NR(3) = 1, so that C∗env(T+(Q)) ∼= e = ∅, and hence C∗env(T+(Q)) is not stably C∗env(T+(R)). However, ΩP isomorphic to C∗env(T+(P )). For the Cuntz-Krieger C-algebras the situation is reversed. The maps I − Gr(P )t, I − Gr(Q)t and I − Gr(R)t on Z3 determining K0 and K1 for the Cuntz-Krieger algebras are given respectively by the matrices e = {3} whereas ΩQ e = ΩR  0 −1 1 −1 1 0 −1 −1 1  ,  0 −1 −1 −1 0 −1 0 −1 0  and  0 −1 −1 −1 0 −1 −1 −1 1  Hence, we see that the K1 groups for OGr(P ), OGr(Q) and OGr(R) are trivial, and that Ran(I − Gr(P )t) = Ran(I − Gr(Q)t) = Z3, so that K0(OGr(P )) = K0(OGr(Q)) are trivial. Hence, by the above mentioned result of Kirchberg and Phillips, we have that OGr(P ) is -isomorphic to OGr(Q). However, since Ran(I − Gr(R)t) ( Z3, we see that the cokernel K0(OGr(R)) is non-trivial, and hence OGr(R) is not stably isomorphic to OGr(P ). Altogether, we obtain that C∗env(T+(P )) 6∼ C∗env(T+(Q)) ∼= C∗env(T+(R)) and OGr(P ) ∼= OGr(Q) 6∼ OGr(R) where ∼= stands for -isomorphism and ∼ stands for stable isomorphism. Note that the C- envelope loses considerable information about the tensor algebra, for instance, the graphs of P and Q are not isomorphic so by [DOM14, Theorem 7.29] T+(Q) and T+(R) are not even algebraically isomorphic. Acknowledgments We would like to thank Orr Shalit for his many helpful remarks and suggestions on a draft version of this paper. References [Arv69] William B. Arveson, Subalgebras of C ∗-algebras, Acta Math. 123 (1969), 141 -- 224. MR 0253059 (40 #6274) [Arv72] William B. Arveson, Subalgebras of C ∗-algebras. II, Acta Math. 128 (1972), no. 3-4, 271 -- 308. MR 0394232 (52 #15035) [Arv76] William B. Arveson, An invitation to C ∗-algebras, Springer-Verlag, New York-Heidelberg, 1976, Grad- uate Texts in Mathematics, No. 39. MR 0512360 (58 #23621) [Arv77] William B. Arveson, Notes on extensions of C ∗ -algebras, Duke Math. J. 44 (1977), no. 2, 329 -- 355. MR 0438137 (55 #11056) [Arv06] William B. Arveson, Notes on the unique extension property, https://math.berkeley.edu/ arve- son/Dvi/unExt.pdf (2006). C-ENVELOPES OF TENSOR ALGEBRAS ARISING FROM STOCHASTIC MATRICES 31 [Arv08] William B. Arveson, The noncommutative Choquet boundary, J. Amer. Math. Soc. 21 (2008), no. 4, 1065 -- 1084. MR 2425180 (2009g:46108) [Arv11] William B. Arveson, The noncommutative Choquet boundary II: hyperrigidity, Israel J. Math. 184 [Bla98] (2011), 349 -- 385. MR 2823981 (2012h:46084) Bruce Blackadar, K-theory for operator algebras, second ed., Mathematical Sciences Research Institute Publications, vol. 5, Cambridge University Press, Cambridge, 1998. MR 1656031 (99g:46104) [BRS90] David P. Blecher, Zhong-Jin Ruan, and Allan M. Sinclair, A characterization of operator algebras, J. [BD96] [CK80] [Cu81] [DK15] Funct. Anal. 89 (1990), no. 1, 188 -- 201. MR 1040962 (91b:47098) Lawrence G. Brown and Marius Dadarlat, Extensions of C ∗-algebras and quasidiagonality, J. London Math. Soc. (2) 53 (1996), no. 3, 582 -- 600. MR 1396721 (97d:46086) Joachim Cuntz and Wolfgang Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. (1980) 56: 251. doi:10.1007/BF01390048 Joachim Cuntz, A class of C ∗-algebras and topological Markov chains. II. Reducible chains and the Ext-Functor for C ∗-algebras, Invent. Math. (1981), 63: 25. doi:10.1007/BF01389192 Kenneth R. Davidson and Matthew Kennedy, The Choquet boundary of an operator system, Duke Math. J. 164 (2015), no. 15, 2989 -- 3004. MR 3430455 [DRS11] Kenneth R. Davidson, Christopher Ramsey, and Orr Moshe Shalit, The isomorphism problem for some [Dix77] universal operator algebras, Adv. Math. 228 (2011), no. 1, 167 -- 218. MR 2822231 (2012j:46083) Jacques Dixmier, C ∗-algebras, North-Holland Publishing Co., Amsterdam-New York-Oxford, 1977, Translated from the French by Francis Jellett, North-Holland Mathematical Library, Vol. 15. MR 0458185 (56 #16388) [DOM14] Adam Dor-On and Daniel Markiewicz, Operator algebras and subproduct systems arising from stochastic [Dur10] matrices, J. Funct. Anal. 267 (2014), no. 4, 1057 -- 1120. MR 3217058 Rick Durrett, Probability: theory and examples, fourth ed., Cambridge Series in Statistical and Prob- abilistic Mathematics, Cambridge University Press, Cambridge, 2010. MR 2722836 (2011e:60001) [DM05] Michael A. Dritschel and Scott A. McCullough, Boundary representations for families of representa- [ELP99] tions of operator algebras and spaces, J. Operator Theory 53 (2005), no. 1, 159 -- 167. MR 2132691 Søren Eilers, Terry A. Loring, and Gert K. Pedersen, Morphisms of extensions of C ∗-algebras: pushing forward the Busby invariant, Adv. Math. 147 (1999), no. 1, 74 -- 109. MR 1725815 (2000j:46104) [ERR09] Søren Eilers, Gunnar Restorff, and Efren Ruiz, Classification of extensions of classifiable C ∗-algebras, Adv. Math. 222 (2009), no. 6, 2153 -- 2172. MR 2562779 [ERRS15] Søren Eilers, Gunner Restorff, Efren Ruiz, and Adam P.W. Sørensen, Geometric classification of unital graph C-algebras of real rank zero, http://arxiv.org/abs/1505.06773 [math.OA] (2015). [Ham79] Masamichi Hamana, Injective envelopes of operator systems, Publ. Res. Inst. Math. Sci. 15 (1979), [Kak13] [KS15] [KK06] [Kat04] no. 3, 773 -- 785. MR 566081 (81h:46071) Evgenios T. A. Kakariadis, The Silov boundary for operator spaces, Integral Equations Operator Theory 76 (2013), no. 1, 25 -- 38. MR 3041719 Evgenios T. A. Kakariadis and Orr Moshe Shalit, On operator algebras associated with monomial ideals in noncommuting variables, arXiv:1501.06495 [math.OA] (2015). Elias G. Katsoulis and David W. Kribs, Tensor algebras of C ∗-correspondences and their C ∗-envelopes, J. Funct. Anal. 234 (2006), no. 1, 226 -- 233. MR 2214146 (2007a:46061) Takeshi Katsura, On C ∗-algebras associated with C ∗-correspondences, J. Funct. Anal. 217 (2004), no. 2, 366 -- 401. MR 2102572 (2005e:46099) [KPR98] Alex Kumjian, David Pask, and Iain Raeburn, Cuntz-Krieger algebras of directed graphs, Pacific J. [Lan95] [MT05] [MS98] [MS00] Math. 184 (1998), no. 1, 161 -- 174. MR 1626528 E. C. Lance, Hilbert C ∗-modules, London Mathematical Society Lecture Note Series, vol. 210, Cam- bridge University Press, Cambridge, 1995, A toolkit for operator algebraists. MR 1325694 (96k:46100) V. M. Manuilov and E. V. Troitsky, Hilbert C ∗-modules, Translations of Mathematical Monographs, vol. 226, American Mathematical Society, Providence, RI, 2005, Translated from the 2001 Russian original by the authors. MR 2125398 (2005m:46099) Paul S. Muhly and Baruch Solel, Tensor algebras over C ∗-correspondences: representations, dilations, and C ∗-envelopes, J. Funct. Anal. 158 (1998), no. 2, 389 -- 457. Paul S. Muhly and Baruch Solel, On the Morita equivalence of tensor algebras, Proc. London Math. Soc. (3) 81 (2000), no. 1, 113 -- 168. MR 1757049 (2001g:46128) [Pas73] William L. Paschke, Inner product modules over B ∗-algebras, Trans. Amer. Math. Soc. 182 (1973), 443 -- 468. MR 0355613 (50 #8087) [PS79] William L. Paschke and Norberto Salinas, Matrix algebras over On, Michigan Math. J. 26 (1979), no. 1, 3 -- 12. MR 514956 (81a:46075) [Pim97] Michael V. Pimsner, A class of C ∗-algebras generalizing both Cuntz-Krieger algebras and crossed prod- ucts by Z, Free probability theory (Waterloo, ON, 1995), Fields Inst. Commun., vol. 12, Amer. Math. Soc., Providence, RI, 1997, pp. 189 -- 212. MR 1426840 (97k:46069) 32 [Rae05] [RS04] [RW98] ADAM DOR-ON AND DANIEL MARKIEWICZ Iain Raeburn, Graph algebras, CBMS Regional Conference Series in Mathematics, vol. 103, Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathe- matical Society, Providence, RI, 2005. MR 2135030 Iain Raeburn and Wojciech Szyma´nski, Cuntz-Krieger algebras of infinite graphs and matrices, Trans. Amer. Math. Soc. 356 (2004), no. 1, 39 -- 59 (electronic). MR 2020023 Iain Raeburn and Dana P. Williams, Morita equivalence and continuous-trace C ∗-algebras, Mathe- matical Surveys and Monographs, vol. 60, American Mathematical Society, Providence, RI, 1998. MR 1634408 (2000c:46108) [Sen06] [RLL00] M. Rørdam, F. Larsen, and N. Laustsen, An introduction to K-theory for C ∗-algebras, London Math- ematical Society Student Texts, vol. 49, Cambridge University Press, Cambridge, 2000. MR 1783408 (2001g:46001) E. Seneta, Non-negative matrices and Markov chains, Springer Series in Statistics, Springer, New York, 2006, Revised reprint of the second (1981) edition [Springer-Verlag, New York; MR0719544]. MR 2209438 Orr Moshe Shalit and Baruch Solel, Subproduct systems, Doc. Math. 14 (2009), 801 -- 868. MR 2608451 (2011k:46095) Ami Viselter, Covariant representations of subproduct systems, Proc. Lond. Math. Soc. (3) 102 (2011), no. 4, 767 -- 800. MR 2793449 (2012d:46170) Ami Viselter, Cuntz-Pimsner algebras for subproduct systems, Internat. J. Math. 23 (2012), no. 8, 1250081 (32 pp.). MR 2949219 Dan Voiculescu, A non-commutative Weyl-von Neumann theorem, Rev. Roumaine Math. Pures Appl. 21 (1976), no. 1, 97 -- 113. MR 0415338 (54 #3427) [Voi76] [SS09] [Vis11] [Vis12] Adam Dor-On, Pure Mathematics Department, University of Waterloo, Waterloo, ON, Canada N2L 3G1 E-mail address: [email protected] Daniel Markiewicz, Department of Mathematics, Ben-Gurion University of the Negev, P.O.B. 653, Be'er Sheva 8410501, Israel. E-mail address: [email protected]
1306.6107	1	1306	2013-06-25T23:47:58	Simplicity of the C*-algebras of skew product k-graphs	[ "math.OA", "math.FA" ]	We consider conditions on a $k$-graph $\Lambda$, a semigroup $S$ and a functor $\eta : \Lambda \to S$ which ensure that the $C^$-algebra of the skew-product graph $\Lambda \times_\eta S$ is simple. Our results allow give some necessary and sufficient conditions for the AF-core of a $k$-graph $C^{}$-algebra to be simple.	math.OA	math	SIMPLICITY OF THE C ∗-ALGEBRAS OF SKEW PRODUCT k-GRAPHS BEN MALONEY AND DAVID PASK Abstract. We consider conditions on a k-graph Λ, a semigroup S and a functor η : Λ → S which ensure that the C ∗-algebra of the skew-product graph Λ×η S is simple. Our results allow give some necessary and sufficient conditions for the AF-core of a k-graph C ∗-algebra to be simple. 1. Introduction In [24] Robertson and Steger investigated C ∗-algebras which they considered to be higher-rank versions of the Cuntz-Krieger algebras. Subsequently, in [9] Kumjian and Pask introduced higher-rank graphs, or k-graphs, as a graphical means to provide combinatorial models for the Cuntz-Krieger algebras of Robertson and Steger. They showed how to construct a C ∗-algebra that is associated to a k-graph. Since then k-graphs and their C ∗-algebras have attracted a lot of attention from many authors (see [1, 3–5, 9, 12–14, 17– 19, 21, 23]). Roughly speaking, a k-graph is a category Λ together with a functor d : Λ → Nk satisfying a certain factorisation property. A 1-graph is then the path category of a directed graph. Given a functor η : Λ → S, where S is a semigroup with identity, we may form the skew product k-graph Λ ×η S. Skew product graphs play an important part in the development of k-graph C ∗-algebras. For example [9, Corollary 5.3] shows that C ∗(Λ ×d Zk) is isomorphic to C ∗(Λ) ×γ Tk where γ : Tk → Aut C ∗(Λ) is the canonical gauge action. Skew product graphs feature in nonabelian duality: In [13] it is shown that if a right-reversible semigroup (Ore semigroup) S acts freely on a k-graph Λ then the crossed product C ∗(Λ) × S is stably isomorphic to C ∗(Λ/S). On the other hand if S is a group G then C ∗(Λ ×η G) is isomorphic to the crossed product C ∗(Λ) ×δη G where δη is the coaction of G on C ∗(Λ) induced by η. The main purpose of this paper is to investigate necessary and sufficient conditions for the C ∗-algebra of a skew product k-graph to be simple. We will be particularly interested in the specific case when S = Nk and η = d. It can be shown that simplicity of C ∗(Λ×d Nk) is equivalent to simplicity of the fixed point algebra (AF core) C ∗(Λ)γ. This is important as many results in the literature apply particularly when AF core is simple; see [1]. We begin by introducing some basic facts we will need during this paper. 2. Background 2.1. Basic facts about k-graphs. All semigroups in this paper will be countable, can- cellative and have an identity, hence any semigroup may be considered as a category with a single object. The semigroup Nk is freely generated by {e1, . . . , ek} and comes with the usual order structure: if n = Pk i=1 miei then m > n (resp. i=1 niei and m = Pk Date: January 16, 2018. 1991 Mathematics Subject Classification. Primary 46L05. Key words and phrases. C ∗-algebra; Graph algebra; k-graph. 1 2 BEN MALONEY AND DAVID PASK m ≥ n) if mi > ni (resp. mi ≥ ni) for all i. For m, n ∈ Nk we define m ∨ n ∈ Nk by (m ∨ n)i = max{mi, ni} for i = 1, . . . , k. A directed graph E is a quadruple (E0, E1, r, s) where E0, E1 are countable sets of vertices and edges. The direction of an edge e ∈ E1 is given by the maps r, s : E1 → E0. A path λ of length n ≥ 1 is a sequence λ = λ1 · · · λn of edges such that s(λi) = r(λi+1) for i = 1, . . . , n − 1. The set of paths of length n ≥ 1 is denoted En. We may extend r, s to En for n ≥ 1 by r(λ) = r(λ1) and s(λ) = s(λn) and to E0 by r(v) = v = s(v). A higher-rank graph or k-graph is a combinatorial structure, and is a k-dimensional analogue of a directed graph. A k-graph consists of a countable category Λ together with a functor d : Λ → Nk, known as the degree map, with the following factorisation property: for every morphism λ ∈ Λ and every decomposition d(λ) = m + n, there exist unique morphisms µ, ν ∈ Λ such that d(µ) = m, d(ν) = n, and λ = µν. For n ∈ Nk we define Λn := d−1(n) to be those morphisms in Λ of degree n. Then by the factorisation property Λ0 may be identified with the objects of Λ, and are called vertices. For u, v ∈ Λ0, X ⊆ Λ and n ∈ Nk we set uX = {λ ∈ X : r(λ) = u} Xv = {λ ∈ X : s(λ) = v} uXv = uX ∩ Xv. A k-graph Λ is visualised by a k-coloured directed graph EΛ with vertices Λ0 and edges ⊔k i=1Λei together with range and source maps inherited from Λ called its 1-skeleton. The 1- skeleton is provided with square relations CΛ between the edges in EΛ, called factorisation rules, which come from factorisations of morphisms in Λ of degree ei + ej where i 6= j. By convention the edges of degree e1 are drawn blue (solid) and the edges of degree e2 are drawn red (dashed). For more details about the 1-skeleton of a k-graph see [21]. On the other hand, if G is a k-coloured directed graph with a complete, associative collection of square relations C completely determines a k-graph Λ such that EΛ = G and CΛ = C (see [6]). A k-graph Λ is row-finite if for every v ∈ Λ0 and every n ∈ Nk, vΛn is finite. A k-graph has no sources if vΛn 6= ∅ for all v ∈ Λ0 and nonzero n ∈ Nk. A k-graph has no sinks is Λnv 6= ∅ for all v ∈ Λ0 and nonzero n ∈ Nk. For λ ∈ Λ and m ≤ n ≤ d(λ), we define λ(m, n) to be the unique path in Λn−m obtained from the k-graph factorisation property such that λ = λ′(λ(m, n))λ′′ for some λ′ ∈ Λm and λ′′ ∈ Λd(λ)−n. Examples 2.1. (a) In [9, Example 1.3] it is shown that the path category E∗ = ∪i≥0Ei of a directed graph E is a 1-graph, and vice versa. For this reason we shall move seamlessly between 1-graphs and directed graphs. (b) For k ≥ 1 let Tk be the category with a single object v and generated by k commuting morphisms {f1, . . . , fk}. Define d : Tk → Nk by d(f n1 k ) = (n1, . . . , nk) then it is 1 straightforward to check that Tk is a k-graph. We frequently identify Tk with Nk via the map f n1 1 7→ (n1, . . . , nk). . . . f nk · · · f nk k (c) For k ≥ 1 define a category ∆k as follows: Let Mor ∆k = {(m, n) ∈ Zk × Zk : m ≤ n} and Obj ∆k = Zk; structure maps r(m, n) = m, s(m, n) = n, and composition (m, n)(n, p) = (m, p). Define d : ∆k → Nk by d(m, n) = n − m, then one checks that (∆k, d) is a row-finite k-graph. We identify Obj ∆k with {(m, m) : m ∈ Zk} ⊂ Mor ∆k. (d) For n ≥ 1 let n = {1, . . . , n}. For m, n ≥ 1 let θ : m × n → m × n a bijection. Let F2 θ be the 2-graph which has 1-skeleton which consists of with single vertex v and edges f1, . . . , fm, g1, . . . , gn, such that fi have the same colour (blue) for i ∈ m and gj have the same colour (red) for j ∈ n together with complete associative square relations figj = gj ′fi′ where θ(i, j) = (i′, j′) for (i, j) ∈ m × n (for more details see [3, 4, 19]). SIMPLICITY OF THE C ∗-ALGEBRAS OF SKEW PRODUCT k-GRAPHS 3 2.2. Skew product k-graphs. Let Λ be a k-graph and η : Λ → S a functor into a semigroup S. We can make the cartesian product Λ × S into a k-graph Λ ×η S by taking (Λ ×η S)0 = Λ0 × S, defining r, s : Λ ×η S → (Λ ×η S)0 by (1) r(λ, t) = (r(λ), t) and s(λ, t) = (s(λ), tη(λ)), defining the composition by (λ, t)(µ, u) = (λµ, t) when s(λ, t) = r(µ, u) (so that u = tη(λ) ), and defining d : Λ ×η S → Nk by d(λ, t) = d(λ). As in [13] it is straightforward to show that this defines a k-graph. Remark 2.2. If Λ is row-finite with no sources and η : Λ → S a functor then Λ ×η S is row-finite with no sources. A k-graph morphism is a degree preserving functor between two k-graphs. If a k-graph morphism is bijective, then it is called an isomorphism. Examples 2.3. (i) Let Λ be a k-graph and η : Λ → S a functor, where S is a semigroup and Λ ×η S the associated skew product graph. Then the map π : Λ ×η S → Λ given by π(λ, s) = λ is a surjective k-graph morphism. (ii) For k ≥ 1 the map (k, m) 7→ (m, m + k) gives an isomorphism from Tk ×d Zk to ∆k. Definition 2.4. Let Λ, Γ be row-finite k-graphs. A surjective k-graph morphism p : Λ → Γ has r-path lifting if for all v ∈ Λ0 and λ ∈ p(v)Γ there is λ′ ∈ vΛ such that p(λ′) = λ. If λ′ is the unique element with this property then p has unique r-path lifting. Example 2.5. Let Λ be a row-finite k-graph and η : Λ → S a functor where S is a semigroup, and Λ ×η S the associated skew product graph. The map π : Λ ×η S → Λ described in Examples 2.3(i) has unique r-path lifting. 2.3. Connectivity. A k-graph Λ is connected if the equivalence relation on Λ0 generated by the relation {(u, v) : uΛv 6= ∅} is Λ0 × Λ0. The k-graph Λ is strongly connected if for all u, v ∈ Λ0 there is N > 0 such that uΛN v 6= ∅. If Λ is strongly connected, then it is connected and has no sinks or sources. The k-graph Λ is primitive is there is N > 0 such that uΛN v 6= ∅ for all u, v ∈ Λ0. If Λ is primitive then it is strongly connected. Examples 2.6. The graphs Tk and F2 have one vertex. θ from Examples 2.1 are primitive since they only The connectivity of a k-graph may also be described in terms of its component ma- trices as defined in [9, §6]: Given a k-graph Λ, for 1 ≤ i ≤ k and u, v ∈ Λ0, we de- fine k non-negative Λ0 × Λ0 matrices Mi with entries Mi(u, v) = uΛeiv. Using the k-graph factorisation property, we have that uΛei+ej v = uΛej+eiv for all u, v ∈ Λ0, and so MiMj = MjMi. For m = (m1, . . . , mk) ∈ Nk and u, v ∈ Λ0, we have uΛmv = (M m1 )(u, v) = M m(u, v), using multiindex notation. The following lemma follows directly from the definitions given above. · · · M mk 1 k Lemma 2.7. Let Λ be a row-finite k-graph with no sources. (a) Then Λ is strongly connected if and only if for all pairs u, v ∈ Λ0 there is there is N ∈ Nk such that M N (u, v) > 0. (b) Then Λ is primitive if and only if there is N > 0 such that M N (u, v) > 0 for all pairs u, v ∈ Λ0. Remarks 2.8. Following [18, §4], a primitive 1-graph Λ is strongly connected with period 1; that is, the greatest common divisor of all n such that vΛnv for some v ∈ Λ0 is 1. 4 BEN MALONEY AND DAVID PASK Lemma 2.9. Let Λ be a k-graph with no sinks, and Λ0 finite. Then for all v ∈ Λ0, there exists w ∈ Λ0 and α ∈ wΛw such that d(α) > 0 and wΛv 6= ∅. Proof. Let p = (1, . . . , 1) ∈ Nk. Since v is not a sink, there exists β1 ∈ Λpv. Since r(β1) is not a sink, there exists β2 ∈ Λpr(β1). Inductively, there exist infinitely many βi such that d(βi) = p and r(βi) = s(βi+1). Since Λ0 is finite, there exists w ∈ Λ0 such that r(βi) = w for infinitely many i. Suppose r(βn) = w = r(βm) with m > n. Then α = βm . . . βn+1 has the requisite properties, and wΛv 6= ∅, since βn . . . β1 ∈ wΛv. (cid:3) 2.4. The graph C ∗-algebra. Let Λ be a row-finite k-graph with no sources, then fol- lowing [9], a Cuntz-Krieger Λ-family in a C ∗-algebra B consists of partial isometries {Sλ : λ ∈ Λ} in B satisfying the Cuntz-Krieger relations: (CK1) {Sv : v ∈ Λ0} are mutually orthogonal projections; (CK2) SλSµ = Sλµ whenever s(λ) = r(µ); (CK3) S∗ (CK4) Sv = P{λ∈vΛn} SλS∗ λ for every v ∈ Λ0 and n ∈ Nk. λSλ = Ss(λ) for every λ ∈ Λ; The k-graph C ∗-algebra C ∗(Λ) is generated by a universal Cuntz-Krieger Λ-family {sλ}. By [9, Proposition 2.11] there exists a Cuntz-Krieger Λ-family such that each vertex projection Sv (and hence by (CK3) each Sλ) is nonzero and so there exists a nonzero universal k-graph C ∗-algebra for a Cuntz-Krieger Λ-family. Moreover, C ∗(Λ) = span{sλs∗ µ : λ, µ ∈ Λ, s(λ) = s(µ)} (see [9, Lemma 3.1]). We will use [23, Theorem 3.1] by Robertson and Sims when considering the simplicity of graph C ∗-algebras: Theorem 2.10 (Robertson-Sims). Suppose Λ is a row-finite k-graph with no sources. Then C ∗(Λ) is simple if and only if Λ is cofinal and aperiodic. We now focus on the two key properties involved in the simplicity criterion of Theorem 2.10, namely aperiodicity and cofinality. Our attention will be directed towards applying these conditions on skew product graphs. 3. Aperiodicity Our definition of aperiodicity is taken from Robertson-Sims, [23, Theorem 3.2]. Definitions 3.1. A row-finite k-graph Λ with no sources has no local periodicity at v ∈ Λ0 if for all m 6= n ∈ Nk there exists a path λ ∈ vΛ such that d(λ) ≥ m ∨ n and λ(m, m + d(λ) − (m ∨ n)) 6= λ(n, n + d(λ) − (m ∨ n)). Λ is called aperiodic if every v ∈ Λ0 has no local periodicity. Examples 3.2. (a) The k-graph ∆k is aperiodic for all k ≥ 1. First observe that there is no local periodicity at v = (0, 0). Given m 6= n ∈ Nk, let N ≥ m ∨ n; then λ = (0, N) is the only element of v∆k. Then λ(m, m) = (m, m) 6= (n, n) = λ(n, n). A similar argument applies for any other vertex w = (n, n) in ∆k and so there is no local periodicity at w for all w ∈ ∆0 k. (b) The k-graph Tk is not aperiodic for all k ≥ 1. For all n ∈ Nk one checks that f n1 1 · · · f nk k k . Hence given m 6= n ∈ Nk it follows that for all λ ∈ vΛN is the only element of vT n with N ≥ m ∨ n we have λ(m, m + (m ∨ n)) = λ(n, n + (m ∨ n)). SIMPLICITY OF THE C ∗-ALGEBRAS OF SKEW PRODUCT k-GRAPHS 5 Since the map π : Λ ×η S → Λ has unique r-path lifting, we wish to know if we can deduce the aperiodicity of Λ ×η S from that of Λ. A corollary of our main result Theorem 3.3, shows that this is true. Theorem 3.3. Let Λ, Γ be row-finite k-graphs and p : Λ → Γ have r-path lifting. If Γ is aperiodic, then Λ is aperiodic. Proof. Suppose that Γ is aperiodic. Let v ∈ Λ0 and m 6= n ∈ Nk. Since Γ is aperiodic, there exists λ ∈ p(v)Γ with d(λ) ≥ m ∨ n such that λ(m, m + d(λ) − (m ∨ n)) 6= λ(n, n + d(λ) − (m ∨ n)). By r-path lifting there is λ′ ∈ vΛ with p(λ′) = λ such that d(λ′) ≥ m ∨ n and λ′(m, m + d(λ) − (m ∨ n)) 6= λ′(n, n + d(λ) − (m ∨ n)), and so Λ is aperiodic. The converse of Theorem 3.3 is false: (cid:3) Example 3.4. The surjective k-graph morphism p : ∆k → Tk given by p(m, m + ei) = fi for all m ∈ Zk and i = 1, . . . , k has r-path lifting. However by Examples 3.2 we see that ∆k ∼= Tk ×d Zk is aperiodic but Tk is not. Corollary 3.5. Let Λ be a row-finite k-graph with no sources, η : Λ → S a functor where S is a semigroup and Λ ×η S the associated skew product graph. If Λ is aperiodic then Λ ×η S is aperiodic. Proof. Follows from Theorem 4.7 and Example 2.5. (cid:3) In some cases the aperiodicity of a skew product graph Λ ×η S can be deduced directly from properties of η. Proposition 3.6. Suppose S is a semigroup, Λ is a row-finite k-graph, η : Λ → S is a functor, and there exists a map φ : S → Zk such that d = φ ◦ η. Then Λ ×η S is aperiodic. Proof. Fix (v, s) ∈ (Λ ×η S)0 and m 6= n ∈ Nk. Let λ ∈ (v, s)(Λ ×η S) be such that d(λ) ≥ m∨n. Observe that λ(m, m) = s(λ(0, m)), λ(m, m) is of the form (w, sη(λ(0, m))) for some w ∈ Λ0. Similarly, λ(n, n) is of the form (w′, sη(λ(0, n))) for some w′ ∈ Λ0. We claim λ(m, m) 6= λ(n, n): Suppose, by hypothesis, η(λ(0, n)) = η(λ(0, m)). Then n = φ◦η(λ(0, n)) = φ◦η(λ(0, m)) = m, which provides a contradiction, and m 6= n. Then η(λ(0, m)) 6= η(λ(0, n)), and so λ(m, m) 6= λ(n, n), and hence λ(m, m + d(λ) − (m ∨ n)) 6= λ(n, n + d(λ) − (m ∨ n)). (cid:3) Corollary 3.7. Suppose Λ is a row-finite k-graph. Then Λ×dNk and Λ×dZk are aperiodic. Proof. Apply Proposition 3.6 with η = d and S = N, Z respectively. (cid:3) 4. Cofinality We will use the Lewin-Sims definition of cofinality, [12, Remark A.3]: Definition 4.1. A row-finite, k-graph Λ with no sources is cofinal if for all pairs v, w ∈ Λ0 there exists N ∈ Nk such that vΛs(α) 6= ∅ for every α ∈ wΛN . Lemma 4.2. Let Λ be a row-finite k-graph with no sources. (a) If Λ is cofinal then Λ is connected. (b) Suppose that for all pairs v, w ∈ Λ0 there exists N ∈ Nk such that vΛs(α) 6= ∅ for every α ∈ wΛN . Then for n ≥ N we have vΛs(α) 6= ∅ for every α ∈ wΛn. 6 BEN MALONEY AND DAVID PASK Proof. Fix v, w ∈ Λ0. If Λ is cofinal it follows that there is α ∈ wΛ such that wΛs(α) and vΛs(α) are non-empty. It then follows that (v, w) belongs to the equivalence relation described in Section 2.3. Since v, w were arbitrary it follows that the equivalence relation is Λ0 × Λ0 and so Λ is connected. Fix v, w ∈ Λ0, then there is N ∈ Nk such that vΛs(α) 6= ∅ for every α ∈ wΛN . Let n ≥ N and consider β ∈ wΛn then β′ = β(0, N) ∈ wΛN and so by hypothesis there is λ ∈ vΛs(β′). Then λβ(N, n) ∈ vΛs(β) and the result follows. (cid:3) Lemma 4.3. Let Λ be a row-finite k-graph with skeleton EΛ. If EΛ is cofinal then Λ is cofinal. Furthermore, Λ is strongly connected if and only if EΛ strongly connected Proof. Fix v, w ∈ Λ0 = E0 all α ∈ wEn α′ ∈ En Λ. Since EΛ is cofinal there is n ∈ N such that vEΛs(α) 6= ∅ for i=1 Ni = n. Then for all α′ ∈ wΛN we have Λ. Let N ∈ Nk be such that Pk Λ and so vΛN s(α′) 6= ∅. Suppose that Λ is strongly connected and v, w ∈ E0 Λ = Λ0. Since Λ is strongly connected there is α ∈ vΛw with d(α) > 0. Let n = Pn i=1 d(α)i then n > 0 and vEΛw 6= ∅, so EΛ is strongly connected. Suppose that EΛ is strongly connected, and v, w ∈ Λ0 = E0 Λ. Since Λ has no sources, there is a path α ∈ vEk Λ which uses an edge of each of the k-colours. Let u = s(α). Since EΛ is strongly connected there is β ∈ uEn Λw. Let λ be the element of Λ which may be represented by αβ ∈ EΛ. Then λ ∈ vΛw and d(λ) > 0 and so Λ is strongly connected. (cid:3) Remark 4.4. The converse to the first part of Lemma 4.3 is not true: Let Λ be the 2-graph which is completely determined by its 1-skeleton as shown: ... • ... • ... • ... • ... • ... • ... • w • • • v • • • • • • • • • • ... • ... • ... • Then Λ is cofinal: For example for v, w as shown, N = (1, 0) will suffice. However EΛ is not cofinal: For example for v, w as shown, for any n ≥ 0 the vertex which is the source of the vertical path of length n with range w does not connect to v. The following proposition establishes a link between cofinality and strongly connectivity for a row-finite k-graph. Proposition 4.5. Suppose Λ is a row-finite k-graph with no sources. (1) If Λ is strongly connected then Λ is cofinal. (2) If Λ is cofinal, has no sinks and Λ0 finite then Λ is strongly connected. Proof. Suppose Λ is strongly connected. Fix v, w ∈ Λ0 then for N = e1 we have vΛs(α) 6= ∅ for all α ∈ wΛN since Λ is strongly connected, and so Λ is cofinal. Suppose Λ is cofinal. Fix u, v ∈ Λ0. Then by Lemma 2.9, there exists w ∈ Λ0 and α ∈ wΛw such that d(α) > 0 and wΛv 6= ∅. Let α′ ∈ wΛv. Given u, w ∈ Λ0, since Λ is cofinal and has no sources, by Lemma 4.2(ii) there exists N ∈ Nk such that for all n ≥ N and all α′′ ∈ wΛn, there exists β ∈ uΛs(α′′). Since d(α) > 0 we may choose t ∈ N such that td(α) > N. Then αt ∈ wΛn where n > N, and so by cofinality of Λ exists β ∈ uΛs(αt) = uΛw. Hence βαα′ ∈ uΛv with d(βαα′) > d(α) > 0 and so Λ is strongly connected. (cid:3) SIMPLICITY OF THE C ∗-ALGEBRAS OF SKEW PRODUCT k-GRAPHS 7 Example 4.6. The condition that Λ0 is finite in Proposition 4.5(2) is essential: For instance ∆k is cofinal by Lemma 4.3 since its skeleton is cofinal; however it is not strongly connected by Lemma 4.3 since its skeleton is not strongly connected. Since the map π : Λ ×η S → Λ has unique r-path lifting, we wish to know if we can deduce the cofinality of Λ ×η S from that of Λ. By Theorem 4.7 the image of a cofinal k-graph under a map with r-path lifting is cofinal, however Example 4.9 shows that the converse is not true. For a cofinal k-graph Λ, we must then seek additional conditions on the functor η which guarantees that Λ ×η S is cofinal. In Definition 4.10 we introduce the notion of (Λ, S, η) cofinality to address this problem. Theorem 4.7. Suppose Λ, Γ be row-finite k-graphs and p : Λ → Γ have r-path lifting. If Λ is cofinal then Γ is cofinal. Proof. Suppose that Λ is cofinal. Fix v, w ∈ Γ0. Let v′, w′ ∈ Λ0 be such that p(v′) = v and p(w′) = w. Since Λ is cofinal there is an N such that for all α′ ∈ w′ΛN there is β′ ∈ v′Λs(α′). Then for α ∈ vΓN there is α′ ∈ v′ΛN with p(α′) = α. By hypothesis there is β′ ∈ v′Λs(α′), and so β = p(β′) is such that s(β) = s(α) and r(β) = v which implies that vΛs(α) 6= ∅ as required. (cid:3) Corollary 4.8. Let Λ be a row-finite k-graph with no sources, η : Λ → S a functor where S is a semigroup and Λ ×η S the associated skew product graph. If Λ ×η S is cofinal then Λ is cofinal. The converse of Theorem 4.7 is false: Example 4.9. Consider the following 2-graph Λ with 1-skeleton c . u e f a . v b t1 t2 g h . w d and factorisation rules: ec = t1e and ha = t2e for paths from u to v; cf = f t1 and bg = f t2 for paths from v to u. Also hd = t1h and eb = t2h for paths from w to v; dg = gt1 and af = gt2 for paths from v to w. By Lemma 4.3 Λ is strongly connected as its skeleton is strongly connected. Note there are no paths of degree e1 + e2 from a vertex to itself. Since M1 = (cid:16) 0 1 0 0 1 0(cid:17) and M2 = (cid:16) 1 0 1 1 0 1 0 2 0 M (2j1+1,j2) = 2j1+j2+1M1. Hence M (2j−1,2j−1) = (cid:16) 0 4j 0 particular by Lemma 2.7 (b) Λ is not primitive, even though it is strongly connected. 0 4j 0 (cid:17) and M (2j,2j) = (cid:16) 4j 0 4j 1 0 1(cid:17), we calculate that M (2j1,j2) = 2j1+j2−1M2 and 4j 0 4j (cid:17). In 4j 0 4j 0 8j 0 We claim that the skew product graph Λ ×d Z2 is not cofinal. Consider v1 = (v, (m, n)) and v2 = (v, (m + 1, n)) in (Λ ×d Z2)0. We claim that for all N ∈ N2, for all α ∈ v1(Λ ×d Z2)N , we have v2(Λ ×d Z2)s(α) 6= ∅. Let N = (N1, N2). Suppose N1 is even. Then for all α ∈ v1(Λ ×d Z2)N , s(α) = (v, (m + N1, n + N2)). In order for this vertex to connect to (v, (m + 1, n)), we have M (N1−1,N2)(v, v) 6= 0. But N1 − 1 is odd, and this matrix entry is zero. If N1 is odd, then s(α) = (u, (m + N1, n + N2)) or s(α) = (w, (m + N1, n + N2)). In order for either of these vertices to connect to (v, (m + 1, n)), we must have M (N1−1,N2)(u, v) 6= 0, or M (N1−1,N2)(w, v) 6= 0. But N1 − 1 is even, and so both of these matrix entries are zero. Hence Λ ×d Z2 is not cofinal, even though Λ is cofinal. 8 BEN MALONEY AND DAVID PASK To establish a sufficient condition for Λ ×η S to be cofinal, we need Λ to be cofinal and an additional condition on η. Definition 4.10. Let Λ be a k-graph with no sources and η : Λ → S a functor, where S is a semigroup. The system (Λ, S, η) is cofinal if for all v, w ∈ Λ0, a, b ∈ S, there exists N ∈ Nk such that for all α ∈ wΛN , there exists β ∈ vΛs(α) such that aη(β) = bη(α). Proposition 4.11. Let Λ be a k-graph with no sources and η : Λ → S a functor, where S is a semigroup and Λ ×η S the associated skew product graph. Then the system (Λ, S, η) is cofinal if and only if Λ ×η S is cofinal. Proof. Suppose Λ ×η S is cofinal. Fix a, b ∈ S and v, w ∈ Λ0. By hypothesis there is N ∈ Nk such that (v, a)(Λ ×η S)s(α, b) is non-empty for every (α, b) ∈ (w, b)(Λ ×η S)N . In particular for all α ∈ wΛN there exists β ∈ wΛN such that aη(β) = bη(α), and so (Λ, S, η) is cofinal. Now suppose (Λ, S, η) is cofinal. Fix (v, a), (w, b) ∈ (Λ ×η S)0. By hypothesis there exists N ∈ Nk such that for all α ∈ wΛN , there exists β ∈ vΛs(α) with aη(β) = bη(α). In particular for all (α, b) ∈ (w, b)(Λ ×η S)N there is (β, a) ∈ (v, a)Λs(α, b), and so Λ ×η S is cofinal. (cid:3) Theorem 4.12. Let Λ be an aperiodic row-finite k-graph with no sources, η : Λ → S a functor, where S is a semigroup and Λ ×η S the associated skew product graph. Then C ∗(Λ ×η S) is simple if and only if the system (Λ, S, η) is cofinal. Proof. Suppose that the system (Λ, S, η) is cofinal. Then by Proposition 4.11, Λ ×η S is cofinal. By Corollary 3.5, Λ ×η S is aperiodic and so by [23, Theorem 3.1], C ∗(Λ ×η S) is simple. Now suppose that C ∗(Λ ×η S) is simple. Then by [23, Theorem 3.1], Λ ×η S is cofinal. (cid:3) By Proposition 4.11 this implies that (Λ, S, η) is cofinal. The condition of (Λ, S, η) cofinality is difficult to check in practice. For 1-graphs it was shown in [18, Proposition 5.13] that Λ ×d Zk is cofinal if Λ is primitive1. We seek an equivalent condition for k-graphs which guarantees (Λ, S, η) cofinality. 5. Primitivity and left-reversible semigroups A semigroup S is said to be left-reversible if for all s, t ∈ S we have sS ∩ tS 6= ∅. It is more common to work with right-reversible semigroups, which are then called Ore semigroups (see [13]). In analogy with the results of Dubriel it can be shown that a left-reversible semigroup has an enveloping group Γ such that Γ = SS−1. In equation (1) we see that functor η : Λ → S multiplies on the right in the semigroup coordinate in the definition of the source map in a skew product graph Λ ×η S. This forces us to consider left-reversible semigroups here. In order to avoid confusion we have decided not to call them Ore. Examples 5.1. (i) Any abelian semigroup is automatically right- and left-reversible. Moreover, any group is a both a right- and left-reversible semigroup. (ii) Let N denote the semigroup of natural numbers under addition and N× denote the semigroup of nonzero natural numbers under multiplication. Let S = N × N× be gifted with the associative binary operation ⋆ given by (m1, n1) ⋆ (m2, n2) = (m1n2 + m2, n1n2), then one checks that S is a nonabelian left-reversible semigroup. reversible; for example, S(m, n) ∩ S(p, q) = ∅ when n = q = 0 and m 6= p. It is not right- 1Actually strongly connected with period 1 which is equivalent to primitive SIMPLICITY OF THE C ∗-ALGEBRAS OF SKEW PRODUCT k-GRAPHS 9 (iii) The free semigroup F+ n ∩ tF+ n on n ≥ 2 generators is not an left-reversible semigroup since for all s, t ∈ F+ n = ∅ as there is no cancellation, and so not only the left-reversibility but also the right-reversibility conditions cannot be satisfied. n with s 6= t we have sF+ A preorder is a reflexive, transitive relation ≤ on a set X. A preordered set (X, ≤) is directed if the following condition holds: for every x, y ∈ X, there exists z ∈ X such that x ≤ z and y ≤ z. A subset Y of X is cofinal if for each x ∈ X there exists y ∈ Y such that x ≤ y. We say that sets X ≤ Y if x ≤ y for all x ∈ X and for all y ∈ Y . We say that t ∈ S is strictly positive if {tn : n ≥ 0} is a cofinal set in S. The following result appears as [15, Lemma 2.2] for right-reversible semigroups. Lemma 5.2. Let S be a left-reversible semigroup with enveloping group Γ, and define ≥l on Γ by h ≥l g if and only if g−1h ∈ S. Then ≥l is a left-invariant preorder that directs Γ, and for any t ∈ S, tS is cofinal in S. Our first attempt at a condition on η which guarantees cofinality of (Λ, S, η) is one which ensures that η takes arbitrarily large values on paths which terminate a given vertex. Definition 5.3. Let Λ be a k-graph with no sources and η : Λ → S be a functor where S is a left-reversible semigroup. We will say that η is upper dense if for all w ∈ Λ0 and a, b ∈ S there exists N ∈ Nk such that bη(wΛN ) ≥l a. Lemma 5.4. Let (Λ, d) be a row-finite k-graph with no sources then d is upper dense for Λ. Proof. Since Λ has no sources it is immediate that wΛN 6= ∅ for all w ∈ Λ0 and N ∈ Nk. For any b, a ∈ Nk we have b + d(wΛN ) = b + N ≥ a provided N ≥ a. (cid:3) Examples 5.5. (i) Let B2 be the 1-graph which is the path category of the directed graph with a single vertex v and two edges e, f . Define a functor η : B2 → N by η(e) = 1 and η(f ) = 0. We may form the skew product B2 ×η N with 1-skeleton: . (v, 0) . (v, 1) . (v, 2) . (v, 3) . . . 2 ) for all n ∈ N it follows that if we choose N = a, 2 ) ≥ a and so η is upper dense. However (B2, N, η) is not cofinal: Fix a, b ∈ N, then since n ∈ η(vBn then b + η(vBN Choose a = 1, b = 0, then for all N ≥ 0 there is f N ∈ vBN b + η(f N ) = 0 6= 1 + η(β) for all β ∈ B2v. 2 is such that (ii) Define a functor η from T2 to N2 such that η(f1) = (2, 0), and η(f2) = (0, 1). We may form the skew product T2 ×η N2 with the following 1-skeleton: . (0, 3) . (0, 2) . (0, 1) . (0, 0) . (1, 3) . (1, 2) . (1, 1) . (0, 1) . (2, 3) . (2, 2) . (1, 2) . (0, 2) . (3, 3) . (3, 2) . (1, 3) . (0, 3) . (4, 3) . (4, 2) . (1, 4) . (0, 4) 10 BEN MALONEY AND DAVID PASK We claim that the functor η is not upper dense: Fix b = (b1, b2) and a = (a1, a2) in N2. Let N1 be such that b1 + 2N1 ≥ a1 and N2 be such that b2 + N2 ≥ a2 then 2 ) ≥l u where N = (N1, N2). Moreover (T2, N2, η) is not cofinal: Let b = (0, 0) bη(vT N and a = (1, 0) then since η(f N1 2 ) = (2N1, N2) it follows that there cannot be N = (N1, N2) ∈ N2 such that for α ∈ vT N 2 there is β ∈ vT2v with bη(α) = aη(β). 1 f N2 (iii) Taking T2 again, we define a functor η : T2 → N2 by η(f1) = (1, 0) and η(f2) = (1, 1). The skew product graph has 1-skeleton: . . . . . . . . . . . . . . . We claim that η is upper dense: Fix b = (b2, b2) and a = (a1, a2) in N2 then there is N1 such that b1 + N1 ≥ a1 and N2 such that b2 + N1 + N2 ≥ a2. Then with 2 we have bη(α) ≥l a. In this case (T2, N2, η) is cofinal: N = (N1, N2) for all α ∈ vT N Fix b = (b1, b2) and a = (a1, a2) in N2. Then there is N1 such that b1 + N1 = a1 + m1 for some m1 ∈ N and N2 such that b2 + N1 + N2 = a2 + m2 for some m2 ∈ N. Hence for all α ∈ vT N ) ∈ vT2v such that bη(α) = aη(β). 2 where N = (N1, N2) there is β = (f m1 , f m2 1 2 It is clear from these last two examples that η being upper dense is not sufficient to guarantee cofinality of (Λ, S, η). The following definition allows for the interaction of the values of η at different vertices of Λ and the following result gives us the required extra condition. Definition 5.6. Let Λ be a k-graph and η : Λ → S be a functor where S is a left-reversible semigroup. We say that η is S-primitive for Λ if there is a strictly positive t ∈ S such that for all v, w ∈ Λ0 we have vη−1(s)w 6= ∅ for all s ∈ S such that s ≥l t. Remarks 5.7. (i) The condition that t is strictly positive in the above definition guar- antees that η(vΛw) is cofinal in S for all v, w ∈ Λ0. (ii) If η : Λ → S is S-primitive for Λ where S is a left-reversible semigroup, then if we extend η to Γ = SS−1 then η is Γ-primitive for Λ. Examples 5.8. (i) Let Λ be a k-graph. Then the degree functor d : Λ → Nk is Nk– primitive for Λ if and only if Λ is primitive as defined in Section 2.3. For this reason we will say that Λ is primitive if d is Nk primitive for Λ. (ii) As in Examples 5.5 (i) let η : B2 → N be defined by η(e) = 1, η(f ) = 0. Then the functor η is N-primitive since η−1(n) is nonempty for all n ∈ N. Hence N-primitivity does not, by itself, guarantee cofinality. (iii) As in Examples 5.5 (ii) let η be the functor from T2 to N2 such that η(f1) = (2, 0), and η(f2) = (0, 1). Then the functor η is not N2-primitive for T2: Take t = (2m, n) ≥ 0 then if s = (2m+1, n) we have vη−1(s)v = ∅ and s ≥l t. Similarly if t = (2m+1) ≥ 0 then if s = (2m + 2, n) we have vη−1(s)v = ∅ and s ≥l t. (iv) As in Examples 5.5 (iii) let η : T2 → N2 be defined by η(f1) = (1, 0) and η(f2) = (0, 1). Then η is not N2-primitive for T2 as vη−1(m, n)v = ∅ whenever n > m. The last two examples above illustrate that upper density and primitivity are unrelated conditions on a k-graph. Together they provide a necessary condition for cofinality. SIMPLICITY OF THE C ∗-ALGEBRAS OF SKEW PRODUCT k-GRAPHS 11 Proposition 5.9. Let Λ be a k-graph with no sources and η : Λ → S be a functor where S is a left-reversible semigroup. If η is S-primitive for Λ and upper dense then (Λ, S, η) is cofinal. If (Λ, S, η) is cofinal then η is upper dense. Proof. Suppose that (Λ, S, η) is cofinal. Fix w ∈ Λ0 and a, b ∈ S and let v be any vertex of Λ. By cofinality of (Λ, S, η) there exists N ∈ Nk such that for all α ∈ wΛN there is β ∈ vΛs(α) such that aη(β) = bη(α). Then any element of bη(wΛN ) is of the form bη(α) = aη(β) ≥l a. Suppose η is S-primitive and upper dense for Λ. Since η is S–primitive for Λ there exists t ∈ S such that for all v, w ∈ Λ0 we have vη−1(s)w 6= ∅ for all s ≥l t. Fix v, w ∈ Λ0 and a, b ∈ S. Since η is upper dense there exists N ∈ Nk such that bη(α) ≥l at for all α ∈ wΛN . Since S is left-reversible, it is directed, and so by definition bη(α) = atu for some u ∈ S. But tu ≥l t and so since η is S–primitive there exists β ∈ vΛs(α) such that η(β) = tu and hence bη(α) = aη(β). (cid:3) Corollary 5.10. Let Λ be a row-finite k-graph such that d is Nk primitive for Λ then (Λ, Nk, d) is cofinal. Proof. Since d is Nk primitive for Λ it follows that Λ has no sources. The result then follows from Lemma 5.4 and Proposition 5.9. (cid:3) Example 5.11. Let η : T2 → S be any functor, then η(S) is a subsemigroup of S since T2 has a single vertex; moreover η is η(S)–primitive for T2. Hence if η is upper dense for T2, it follows that (T2, η(S), η) is cofinal. In particular, in Example5.5 (ii) one checks that (T2, η(N2, η) is cofinal. Theorem 5.12. Let Λ be an aperiodic k-graph, η : Λ → S be a functor into a left- reversible semigroup, and η be S–primitive for Λ. Then C ∗(Λ ×η S) is simple if and only if η is upper dense. Proof. If η is upper dense then the result follows from Proposition 5.9. On the other hand if C ∗(Λ ×η S) is simple then the result follows from Theorem 4.12 and Corollary 3.5. (cid:3) 6. Skew products by a group Let Λ be a row-finite k-graph. A functor η : Λ → G defines a coaction δη on C ∗(Λ) determined by δη(sλ) = sλ ⊗ η(λ). It is shown in [14, Theorem 7.1] that C ∗(Λ ×η G) is isomorphic to C ∗(Λ) ×δη G. Hence we may relate the simplicity of the C ∗-algebra of a skew product graph to the simplicity of the associated crossed product. This can be done by using the results of [20]. Following [14, Lemma 7.9], for g ∈ G the spectral subspace C ∗(Λ)g of the coaction δη is given by C ∗(Λ)g = span{sλs∗ µ : η(λ)η(µ)−1 = g}. We define sp(δη) = {g ∈ G : C ∗(Λ)g 6= ∅}, to be the collection of non-empty spectral subspaces. The fixed point algebra, C ∗(Λ)δη of the coaction is defined to be C ∗(Λ)1G. For more details on the coactions of discrete groups on k-graph algebras, see [14, §7] and [20]. We give necessary and sufficient conditions for the skew product graph C ∗-algebra to be simple in terms of the fixed-point algebra as our main result in Theorem 6.3. We are particularly interested in the case when η is the degree functor. Definition 6.1. Let Λ be a row-finite k-graph, G be a discrete group and η : Λ → G a functor, then we define Γ(η) = {g ∈ G : g = η(λ)η(µ)−1 for some λ, µ ∈ Λ with s(λ) = s(µ)}. 12 BEN MALONEY AND DAVID PASK Lemma 6.2. Let Λ be a row-finite graph with no sources and η : Λ → G a functor, where G is a discrete group. (a) If (Λ, G, η) is cofinal then Γ(η) = G. (b) sp(δη) = G if and only if Γ(η) = G. Proof. Fix g ∈ G and write g = b−1a for some a, b ∈ G. Now fix v, w ∈ Λ0; since (Λ, G, η) is cofinal there exist λ, µ ∈ Λ with s(λ) = s(µ) such that aη(µ) = bη(λ). Hence b−1a = η(λ)η(µ)−1 and so g ∈ Γ(η). Since g was arbitrary the result follows. The second statement follows by definition. (cid:3) Theorem 6.3. Let Λ be an aperiodic row-finite k-graph with no sources, η : Λ → G a functor and δη the associated coaction of G on C ∗(Λ). Then C ∗(Λ ×η G) is simple if and only if C ∗(Λ)δη is simple and Γ(η) = G. Proof. By [14, Theorem 7.1] it follows that C ∗(Λ ×η G) is isomorphic to C ∗(Λ) ×δη G. Then by [20, Theorem 2.10] C ∗(Λ) ×δη G is simple if and only if C ∗(Λ)δη is simple and sp(δη) = G. The result now follows from Lemma 6.2. (cid:3) Example 6.4. Let Λ be a row-finite k-graph with no sources and d : Λ → Nk be the degree functor. We claim that Γ(d) = Zk. Fix p ∈ Zk, and write p = m − n where m, n ∈ Nk. Since Λ has no sources, for every v ∈ Λ0 there is λ ∈ Λmv and µ ∈ Λnv. Then d(λ) − d(µ) = m − n = p ∈ Γ(d), and so Γ(d) = Zk. Since Γ(d) = Zk, and (Λ, Zk, d) is aperiodic, we have that C ∗(Λ)δd is simple if and only (Λ, Nk, d) is cofinal. We seek conditions on Λ that will guarantee (Λ, Nk, d) is cofinal. 7. The gauge coaction The coaction δd of Zk on C ∗(Λ) defined in Section 6 is such that the fixed point algebra C ∗(Λ)δd is precisely the fixed point algebra C ∗(Λ)γ for the canonical gauge action of Tk on C ∗(Λ) by the Fourier transform (cf. [2, Corollary 4.9]. By [9, Lemma 3.3] the fixed point algebra C ∗(Λ)γ is AF, and is usually referred to as the AF core. In Theorem 7.2 we use the results of the last two sections to give necessary and sufficient conditions for the AF core C ∗(Λ)γ to be simple when Λ0 is finite. When there are infinitely many vertices we show, in Theorem 7.8 that in many cases the AF core is not simple. The AF core of a k-graph algebra plays a significant role in the development of crossed products by endomorphisms. Results of Takehana and Katayama [8] show that when Λ is a finite 1-graph such that the core C ∗(Λ) is simple, then every nontrivial automorphism of C ∗(Λ) is outer (see [17, Proposition 3.4]). We saw in Example 4.9 that a k-graph being strongly connected is not enough to guarantee that Λ ×d Zk is cofinal, and hence by [23, Theorem 3.1] C ∗(Λ ×d Zk) is not simple and then by Theorem 6.3 the AF core is not simple. Another condition is required to guarantee that Λ ×d Zk is cofinal, which is suggested by [18] and was introduced in Section 5: Theorem 7.1. Let Λ be a row-finite k-graph with no sinks and sources and Λ0 finite. If (Λ, d, Zk) is cofinal then Λ is primitive. Proof. We claim that for v ∈ Λ0 there is N(v) ∈ Nk such that for all n ≥ N(v) we have vΛnv 6= ∅. Fix (v, 0) ∈ (Λ ×d Zk)0 then for each w ∈ Λ0, when we apply the cofinality condition to (w, 0) ∈ (Λ ×d Zk)0 we obtain Nw ∈ Nk such that (v, 0)(Λ ×d Zk)s(α, 0) 6= ∅ SIMPLICITY OF THE C ∗-ALGEBRAS OF SKEW PRODUCT k-GRAPHS 13 for all (α, 0) ∈ (w, 0)(Λ ×d Zk)Nw. Define N = maxw∈Λ0{Nw}, which is finite since Λ0 is finite. By Proposition 4.5 it follows that Λ is strongly connected, hence there exists α ∈ vΛv with d(α) = r > 0. Hence, there exists t ≥ 1 such that tr ≥ N. Let N(v) = tr. Let m = n − tr ≥ 0. Since Λ has no sources, vΛm 6= ∅; hence there exists γ ∈ vΛm. Let w = s(γ). For (v, 0), (w, 0) ∈ (Λ ×d Zk)0, we have (αt, 0) ∈ (v, 0)(Λ ×d Zk)tr where tr ≥ N ≥ Nw. By cofinality and Lemma 4.2 (b), there exists (β, 0) ∈ (w, 0)(Λ×d Zk)(v, tr) as s(αt, 0) = (v, tr). As β ∈ wΛtrv it follows that γβ ∈ vΛnv, which proves the claim. (cid:3) The following result generalises results from [18]: Theorem 7.2. Let (Λ, d) be a row-finite k-graph with no sinks or sources, and Λ0 finite. Then C ∗(Λ)δd is simple if and only if Λ is primitive. Proof. Suppose that Λ is primitive. Then (Λ, Zk, d) is strongly connected and cofinal by Remarks 2.8. Hence C ∗(Λ ×d Zk) is simple and so C ∗(Λ)δd is simple by Theorem 6.3. Suppose that C ∗(Λ)δd is simple. Recall from Example 6.4 that since Λ has no sources then Γ(d) = Zk. Then by Theorem 6.3, C ∗(Λ ×d Zk) is simple, and hence (Λ, d, Zk) is cofinal by [23, Theorem 3.1] and Proposition 4.11. By Theorem 7.1 this implies that Λ is primitive. (cid:3) Example 7.3. Since it has a single vertex it is easy to see that the 2-graph F2 Examples 2.1 (d) is primitive. Hence by Theorem 7.2 we see that C ∗(F2 all θ. Indeed in [4, §2.1] it is shown that C ∗(F2 We now turn our attention to the case when Λ0 is infinite. We adapt the technique used in [18] to show that, in many cases the AF core is not simple. Definition 7.4. Let Λ be a row-finite k-graph with no sources. For v ∈ Λ0, n ∈ Nk let θ defined in θ)γ is simple for θ)γ ∼= UHF(mn)∞. V (n, v) = {s(λ) : λ ∈ vΛm, m ≤ n} F V (n, v) = V (n, v)\ ∪k i=1 V (n − ei, v). Remarks 7.5. For v ∈ Λ0, m ≤ n ∈ Nk we have, by definition, that V (m, v) ⊆ V (n, v). For v ∈ Λ0, n ∈ Nk the set F V (n, v) denotes those vertices which connect to v with a path of degree n and there is no path from that vertex to v with degree less than n. Lemma 7.6. Let Λ be a row-finite k-graph with no sources. For v ∈ Λ0, n ∈ Nk then V (n, v) is finite and if V (n) = V (n − ei) for some 1 ≤ i ≤ k then V (n + rei) = V (n − ei) for all r ≥ 0. Proof. Fix, v ∈ Λ0, n ∈ Nk, since Λ row-finite it follows that ∪m≤nvΛm is finite and hence so is V (n, v). Suppose, without loss of generality that V (n) = V (n − e1). Let w ∈ V (n + e1), then there is λ ∈ vΛn+e1w. Now λ(0, n) ∈ vΛn and so s(λ(0, n)) ∈ V (n) = V (n − e1). Hence there is µ ∈ vΛms(λ(0, n)) for some m ≤ n − e1 and so µλ(n, n + e1) ∈ vΛm+e1. Since s(µλ(n, e + e1)) = s(λ) = w and m + e1 ≤ n it follows that w ∈ V (n). As w was an arbitrary element of V (n + e1) it follows that V (n + e1) ⊆ V (n) = V (n − e1). By Remarks 7.5 we have V (n − e1) ⊆ V (n + e1) and so V (n + e1) = V (n − e1). It then follows that V (n + re1) = V (n − e1) for r ≥ 0 by an elementary induction argument. (cid:3) We adopt the following notation, used in [11]: Let Λ be a k-graph for 1 ≤ i ≤ k we set ΛNei = ∪r≥0Λrei. Proposition 7.7. Let Λ be a row-finite k-graph with no sources such that for all w ∈ Λ0 and for 1 ≤ i ≤ k, the set s−1 (cid:0)wΛNei(cid:1) is infinite. Then for all n ∈ Nk, v ∈ Λ0 we have F V (n, v) 6= ∅. 14 BEN MALONEY AND DAVID PASK Proof. Suppose, for contradiction, that F V (n, v) = ∅ for some n ∈ Nk and v ∈ Λ0. Then, without loss of generality we may assume that V (n) = V (e − e1). Let λ ∈ vΛn, then s(λ) ∈ V (n) = V (n − e1). Fix r ≥ 0, then since Λ has no sources there is µ ∈ s(λ)Λre1. Then λµ ∈ vΛn+re1 and so s(λµ) = s(µ) ∈ V (n + re1, v). By Lemma 7.6 it follows that V (n + re1) = V (n − e1) and so for any µ ∈ s(λ)ΛNe1 we have s(µ) ∈ V (n − e1). By Remarks 7.5 V (n − e1) is finite and so we have contradicted the hypothesis that s−1 (cid:0)wΛNe1(cid:1) is infinite. (cid:3) Note that k-graphs satisfying the hypothesis of Proposition 7.7 must have infinitely many vertices. The following result generalises results from [18]: Theorem 7.8. Let Λ be a row-finite k-graph with no sources such that for all w ∈ Λ0 and for 1 ≤ i ≤ k, the set s−1 (cid:0)wΛNei(cid:1) is infinite. Then Λ ×d Zk is not cofinal. Proof. Suppose, for contradiction, that Λ ×d Zk is cofinal. Fix v ∈ Λ0 then since Λ is row-finite and has no sources W = s−1 (vΛe1) is finite and nonempty. Without loss of generality let W = {w1, . . . , wn}. Since Λ×dZk is cofinal, for 1 ≤ i ≤ n if we consider (wi, 0) and (v, 0) ∈ Λ0×Zk then there is Ni ∈ Nk such that for all (α, 0) ∈ (wi, 0)(cid:0)Λ ×d Zk(cid:1)Ni we have (v, 0)(cid:0)Λ ×d Zk(cid:1) (s(α), Ni) 6= ∅. Let N = max{N1, . . . , Nn}. By Proposition 7.7 F V (N + e1, v) 6= ∅, hence there is λ ∈ vΛN +e1 such that there is no path of degree less than N + e1 from s(λ) to v. Without loss of generality s(λ(0, e1)) = w1, and so (λ(e1, N + e1), 0) ∈ (w1, 0)(cid:0)Λ ×d Zk(cid:1)N . Since N ≥ N1 and Λ has no sources, by Lemma 4.2(ii) there is (α, 0) ∈ (v, 0)(cid:0)Λ ×d Zk(cid:1) (s(λ), N) which implies that α ∈ vΛN s(λ), contradicting the defining property of λ ∈ vΛN +e1. (cid:3) Examples 7.9. (1) Let Λ be a strongly connected k-graph with Λ0 infinite, then Λ has no sources and for all w ∈ Λ0 we have s−1 (cid:0)wΛNei(cid:1) is infinite for 1 ≤ i ≤ k. Hence by Theorem 7.8 it follows that Λ ×d Zk is not cofinal. (2) Let Λ be a k-graph with Λ0 infinite, no sources and no paths with the same source and range. Then for all w ∈ Λ0 we have s−1 (cid:0)wΛNei(cid:1) is infinite for 1 ≤ i ≤ k. Hence by Theorem 7.8 it follows that Λ ×d Zk is not cofinal. References [1] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on C ∗-algebras associated to higher rank graphs, J. Math. Anal. Appl., 405 (2013), 388–399. [2] T. Crisp. Corners of graph algebras, J. Operator Theory, 60 (2008), 253–271. [3] K. R. Davidson, S. C. Power and D. Yang, Atomic representations of rank 2 graph algebras, J. Funct. Anal., 255 (2008), 819–853. [4] K. R. Davidson and D. Yang, Periodicity in rank 2 graph algebras, Canad. J. Math., 61 (2009), 1239–1261. [5] D. G. Evans, On the K-theory of higher rank graph C*-algebras, New York J. Math., 14 (2008), 1–31. [6] R. Hazlewood, I. Raeburn, A. Sims and S. Webster, Remarks on some fundamental results about higher-rank graphs and their C ∗-algebras, Proc. Edinburgh Math. Soc., 56 (2013), 575–597. [7] S. Kang and D. Pask. Aperiodicity and the primitive ideal space of a row-finite k-graph C ∗-algebra, arXiv:math/1105.1208 [math.OA]. [8] Y. Katayama and H. Takehana, On automorphisms of generalized Cuntz algebras, Internat. J. Math., 9 (1998), 493–512. [9] A. Kumjian and D. Pask, Higher Rank Graph C ∗-algebras, New York J. Math., 6 (2000), 1–20. [10] A. Kumjian and D. Pask, Actions of Zk associated to higher rank graphs, Ergod. Th. & Dynam. Sys., 23 (2003), 1153–1172. [11] A. Kumjian, D. Pask and A. Sims, C ∗-algebras associated to covering of k-graphs, Doc. Math., 13 (2008), 161–205. SIMPLICITY OF THE C ∗-ALGEBRAS OF SKEW PRODUCT k-GRAPHS 15 [12] P. Lewin and A. Sims, Aperiodicity and Cofinality for Finitely Aligned Higher-Rank Graphs, Math. Proc. Cambridge Philosophical Soc., 149 (2010), 333–350. [13] B. Maloney, D. Pask and I. Raeburn, Skew products of higher-rank graphs and crossed products by semigroups, Semigroup Forum (to appear). [14] D. Pask, J. Quigg and I. Raeburn, Coverings of k-graphs, J. Alg., 289 (2005), 161–191. [15] D. Pask, I. Raeburn and T. Yeend, Actions of semigroups on directed graphs and their C ∗-algebras, J. Pure Appl. Algebra, 159 (2001), 297–313. [16] D. Pask and I. Raeburn, On the K-Theory of Cuntz-Krieger algebras, Publ. RIMS Kyoto, 32 (1996), 415–443. [17] D. Pask, I. Raeburn and N. Weaver, Periodic 2-graph arising from subshifts, Bull. Aust. Math. Soc., 82 (2010), 120–138. [18] D. Pask and S-J. Rho. Some intrinsic properties of simple graph C ∗-algebras, Operator Algebras and Mathematical Physics, Constanza, Romania, 2001, Theta Foundation, 2003, 325–340. [19] S. Power, Classifying higher rank analytic Toeplitz algebras, New York J. Math., 13 (2007), 271–298. [20] J. Quigg Discrete C ∗–coactions and C ∗–algebraic bundles. J. Austral. Math. Soc., 60 (1996), 204– 221. [21] I. Raeburn, A. Sims and T. Yeend, Higher-rank graphs and their C ∗-algebras, Proc. Edinburgh Math. Soc., 46 (2003), 99–115. [22] I. Raeburn, A. Sims and T. Yeend, The C ∗-algebras of finitely aligned higher-rank graphs J. Funct. Anal., 213 (2004), 206–240. [23] D. I. Robertson and A. Sims, Simplicity of C ∗-algebras associated to higher rank graphs, Bull. London Math. Soc., 39 (2007), 337–344. [24] G. Robertson, T. Steger, Affine buildings, tiling systems and higher rank Cuntz-Krieger algebras, J. reine angew. Math., 513 (1999), 115–144. School of Mathematics & Applied Statistics, University of Wollongong, Northfields Avenue, NSW 2522, AUSTRALIA E-mail address: [email protected], [email protected]
1505.00646	2	1505	2015-09-20T13:57:34	Half-liberated manifolds, and their quantum isometries	[ "math.OA", "math.QA" ]	We discuss the half-liberation operation $X\to X^$, for the algebraic submanifolds of the unit sphere, $X\subset S^{N-1}_\mathbb C$. There are several ways of constructing this correspondence, and we take them into account. Our main results concern the computation of the affine quantum isometry group $G^+(X^)$, for the sphere itself.	math.OA	math	HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES TEODOR BANICA Abstract. We discuss the half-liberation operation X → X ∗, for the algebraic sub- manifolds of the unit sphere, X ⊂ SN −1 . There are several ways of constructing this correspondence, and we take them into account. Our main results concern the compu- tation of the affine quantum isometry group G+(X ∗), for the sphere itself. C 5 1 0 2 p e S 0 2 ] . A O h t a m [ 2 v 6 4 6 0 0 . 5 0 5 1 : v i X r a Introduction The notion of noncommutative space goes back to an old theorem of Gelfand, which states that any commutative C∗-algebra must be of the form C(X), for a certain compact space X. One can therefore define the category of "noncommutative compact spaces" to be the category of C∗-algebras, with the arrows reversed. The category of usual compact spaces embeds then covariantly into this category, via X → C(X). We will be interested here in noncommutative analogues of the compact algebraic man- ifolds X ⊂ CN . These are by definition the duals of the universal C∗-algebras defined with generators z1, . . . , zN , subject to (noncommutative) polynomial relations: C(X) = C∗(cid:16)z1, . . . , zN(cid:12)(cid:12)(cid:12)Pi(z1, . . . , zN ) = 0(cid:17) The Gelfand theorem tells us that this construction covers all the compact algebraic manifolds X ⊂ CN . In general, the axiomatization of the algebras on the right is quite a tricky problem. Instead of getting into details here, let us just say that the family of noncommutative polynomials {Pi} must be by definition such that the biggest C∗-norm on the universal ∗-algebra < z1, . . . , zNPi(z1 . . . , zN ) = 0 > is bounded. The compact quantum Lie groups, axiomatized by Woronowicz in [23], [24], and their homogeneous spaces, provide some key examples of such manifolds. Technically speaking, one problem with such quantum groups is that they lack an analogue of a Lie algebra. As explained in [23], [24], one solution to this issue comes from the intensive use of representation theory, in order to overcome the lack of geometric techniques. The aim of this paper is to use some quantum group ideas, coming from representation theory, in the complex manifold setting. Let X be as above, and consider its classical 2000 Mathematics Subject Classification. 46L65 (46L54). Key words and phrases. Half-liberation, Quantum isometry. 1 2 TEODOR BANICA version Xclass ⊂ CN , obtained by dividing the algebra C(X) by its commutator ideal: C(Xclass) = C∗comm(cid:16)z1, . . . , zN(cid:12)(cid:12)(cid:12)Pi(z1, . . . , zN ) = 0(cid:17) We can think then of X as being a "liberation" of Xclass, and the problem is that of understanding how the correspondence Xclass → X can appear. This latter question was solved in the quantum group case in [5], by using some in- spiration from Wang's papers [20], [21], from the Weingarten formula [3], [12], [22], and from free probability theory [7], [18], [19]. Among the findings there, and from the related papers [6], [8], [10], is the fact that, for liberation purposes, the usual commutation re- lations ab = ba can be succesfully replaced by the half-commutation relations abc = cba. This is actually a quite non-trivial phenomenon, which comes from the fact that the half-commutation relations abc = cba have a deep categorical meaning. See [5]. As explained in [8], [10], there are several possible ways of half-liberating a manifold , and we will take this into account. We will show here that, under suitable C X ⊂ SN−1 assumptions on X ⊂ SN−1 C , we have a half-liberation diagram for it, as follows: X / X∗∗ / X∗ X− X◦ / X # Our main results will concern the sphere X = SN−1 itself. More specifically, we will be interested in computing the quantum isometry groups of its various half-liberations. Our approach here will be based on the affine quantum isometry group formalism [11], [13], [15], [16], with various technical ingredients from [1], [6], [10], [17]. We will prove that the affine quantum isometry groups of the 6 half-liberated spheres are as follows: C SN−1 C / SN−1 C,∗∗ / SN−1 C,∗ UN / U∗∗N / U∗N TSN−1 R SN−1 C,◦ / SN−1 C,# TON U◦N / U # N −→ In other words, our result will state that, for the sphere X = SN−1 itself, the quantum isometry groups of the half-liberations are the half-liberations of the usual isometry group. This could be thought of as being related to the various rigidity results in [9], [14]. C The paper is organized as follows: in 1-2 we discuss the half-liberation operation for the complex sphere itself, in 3-4 we study the associated quantum isometry groups, and in 5-6 we discuss the case of more general algebraic manifolds X ⊂ SN−1 C . / / / / O O O O / O O / / / / O O O O / O O / / / / O O O O / O O HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 3 Acknowledgements. I would like to thank Julien Bichon for several useful discussions, and the anonymous referee for a careful reading of the manuscript. This work was partly supported by the NCN grant 2012/06/M/ST1/00169. 1. Noncommutative spheres According to [2], [4], which were based on the previous work of Wang in [20], the free analogue of the complex unit sphere SN−1 C is constructed as follows: Definition 1.1. Associated to any N ∈ N is the universal C∗-algebra C(SN−1 C,+ ) = C∗ z1, . . . , zN(cid:12)(cid:12)(cid:12)Xi ziz∗i =Xi z∗i zi = 1! whose abstract spectrum SN−1 C,+ is called free analogue of SN−1 C . Observe that the classical version of SN−1 standard coordinates zi and their adjoints z∗i commute, is the usual sphere SN−1 follows indeed from the Stone-Weierstrass and Gelfand theorems. See [2], [4]. C C,+ , obtained by assuming in addition that the . This We will be interested in what follows in various half-liberated analogues of SN−1 C . We have the following constructions here, which go back to the work in [2]: Definition 1.2. We have the following subspheres of SN−1 C,+ : (1) SN−1 C,∗ (2) SN−1 C,∗∗ (3) SN−1 C,# : obtained via the relations ab∗ = ba∗, a∗b = b∗a, with a, b ∈ {zi}. (4) SN−1 C,◦ : obtained via the relations ab∗c = cb∗a, with a, b, c ∈ {zi}. : obtained via the relations abc = cba, with a, b, c ∈ {zi, z∗i }. : obtained as an intersection, SN−1 C,◦ C,# ∩ SN−1 C,∗∗ = SN−1 . Once again, we use here the general C∗-algebra philosophy, which allows us to define C,+ ) by noncommutative compact subspaces SN−1 various algebraic relations, and then by taking the abstract spectrum. See [2]. In addition to the above 4 noncommutative spheres, and to the sphere SN−1 C,+ , by dividing the algebra C(SN−1 × ⊂ SN−1 itself, we C have as well the following "subsphere" of SN−1 C , which is of interest for us: TSN−1 R =n(ux1, . . . , uxN ) ∈ SN−1 C (cid:12)(cid:12)(cid:12)u ∈ T, (x1, . . . , xN ) ∈ SN−1 R o Here, and in what follows, T is the unit circle in the complex plane. When adding the above new "sphere" to the 5 examples that we have so far, we obtain a set of objects which is stable by intersections, as follows: 4 TEODOR BANICA Proposition 1.3. We have the following diagram, with all maps being inclusions: SN−1 C / SN−1 C,∗∗ / SN−1 C,∗ TSN−1 R SN−1 C,◦ / SN−1 C,# In addition, this is an intersection diagram, in the sense that any intersection X ∩ Y appears on the diagram, as the biggest object contained in both X, Y . Proof. The upper horizontal inclusions are all clear. The lower horizontal inclusions are clear as well, with TSN−1 coming from the fact that the standard coordinates zi = uxi on TSN−1 satisfy the relations ab∗ = ba∗ = a∗b = b∗a. R ⊂ SN−1 C,◦ R The two vertical inclusions on the left are clear. The remaining vertical inclusion, on the right, comes from the fact that, by using ab∗ = ba∗, a∗b = b∗a, we obtain: ab∗c = ba∗c = bc∗a = cb∗a . Regarding the C The intersection claim on the right is clear from the definition of SN−1 C,◦ equals TSN−1 intersection claim on the left, this states that SN−1 C,− ∩ SN−1 C,◦ = SN−1 We have SN−1 C,− ⊂ SN−1 , so consider a point z ∈ SN−1 C,− C,# , the coordinates of z must satisfy the relations ab∗ = ba∗, a∗b = b∗a, so we have zi ¯zj = zj ¯zi. In the case zi, zj 6= 0 we obtain zi/¯zi = zj/¯zj, and we deduce that the numbers zi/¯zi are all equal, independently of the index i satisfying zi 6= 0. Now by multiplying by a suitable scalar u ∈ T, we can assume that we have zi/¯zi = 1, for any i such that zi 6= 0. Thus, up to the multiplication by a scalar u ∈ T, we have z ∈ SN−1 . Since we have SN−1 , as desired. R . C,− ⊂ SN−1 C R As already mentioned, the above 6 spheres were introduced in [2]. In order to explain where these spheres come from, let us recall from [2] that we have: Proposition 1.4. We have the following intersection diagram, SN−1 R,+ SN−1 R,∗ / SN−1 C,+ SN−1 C / SN−1 C,∗∗ / SN−1 C,∗ SN−1 R TSN−1 R SN−1 C,◦ / SN−1 C,# with the spheres on the left being the real versions (zi = z∗i ) of the spheres on the right. (cid:3) / / / / O O O O / O O / & & O O / / O O / / O O / / O O O O / O O HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 5 Proof. Observe first that SN−1 the coordinates are self-adjoint, the relations ab∗ = ba∗, a∗b = b∗a read ab = ba. is the real version of SN−1 C,# , because when assuming that R Also, we have an inclusion SN−1 of the sphere SN−1 C,∗ R,∗ ⊂ SN−1 C,∗∗ , because when taking the real version SN−1 R,∗ , the defining relations ab∗c = cb∗a read abc = cba. With these observations in hand, the fact that we have the diagram in the statement, (cid:3) and that this is an intersection diagram, are clear from Proposition 1.3. The point now is that the above 10 spheres have a number of common features: Proposition 1.5. The above 10 spheres appear from SN−1 C,+ via relations of type i1 . . . zek ze1 ik = zd1 iσ(1) . . . zdk iσ(k) ,∀i1, . . . , ik where σ ∈ Sk is a permutation, and where ei, di ∈ {1,∗} are exponents. Proof. The 10 spheres appear indeed from SN−1 C,+ via the following relations: a = a∗, ab = ba, ab∗ = b∗a, ab∗ = ba∗, a∗b = b∗a abc = cba, abc∗ = c∗ba, ab∗c = cb∗a Now since all these relations are as in the statement, this proves the result. (cid:3) As explained in [2], the formalism in Proposition 1.5 is in fact too wide. The solution R,+ , which are conjecturally the proposed in [2] is that of starting with SN−1 R,∗ ⊂ SN−1 only real examples, and then by performing 3 operations: C ⊂ SN−1 R ⊂ SN−1 (1) Mirroring: this produces the spheres SN−1 C,∗∗ ⊂ SN−1 C,+ . (2) Free complexification: this produces the extra spheres SN−1 C,# ⊂ SN−1 C,∗ (3) Taking intersections: this produces the remaining spheres TSN−1 Summarizing, the above 10 spheres are expected to be the "only ones", under some R ⊂ SN−1 C,◦ . . strong axioms, which are however not available yet. See [2]. Let us try now to better understand the half-liberated spheres. Given SN−1 associated projective space is the quotient SN−1 C(SN−1 × ) is the subalgebra generated by the variables pij = ziz∗j . We have then: × → P N × × ⊂ SN−1 given by the fact that C(P N × C,+ , the ) ⊂ Theorem 1.6. The projective spaces for the 6 half-liberated spheres are P N C P N C P N C P N R P N R P N R where P N R , P N C are the usual real and complex projective spaces. O O O O O O 6 TEODOR BANICA Proof. We use the following presentation results, coming from the Gelfand theorem: C(P N C(P N R ) = C∗comm(cid:16)(pij)i,j=1,...,N(cid:12)(cid:12)(cid:12)p = pt = p∗ = p2, T r(p) = 1(cid:17) C ) = C∗comm(cid:16)(pij)i,j=1,...,N(cid:12)(cid:12)(cid:12)p = p∗ = p2, T r(p) = 1(cid:17) By functoriality, the projective spaces for our 6 spheres are as follows: P N C / P N C,∗∗ / P N C,∗ P N R P N C,◦ / P N C,# In order to finish, it is enough to prove that we have P N P N C,∗ ⊂ P N C . From ab∗c = cb∗a we obtain ab∗cd∗ = cb∗ad∗ = cd∗ab∗, so the variables pij = ziz∗j commute. In addition we have p = p∗ = p2, T r(p) = 1, and we are done. C,∗ ⊂ P N C , P N C,# ⊂ P N R . C follows to be a subspace of P N R . From ab∗ = ba∗ we deduce that the matrix pij = ziz∗j is symmetric, and so = P N (cid:3) P N C,# ⊂ P N C,# ⊂ P N P N C,∗ We should mention that the above result has an extension to the 10-sphere framework of Proposition 1.4, with the 3 rows of spheres corresponding to the 3 types of projective spaces (real, complex, free). Indeed, we have P N R,+ ⊂ P N R,∗ is known to be an isomorphism at the level of reduced versions. See [2], [4]. C , and the inclusion P N R , as desired. = P N C,+ 2. Matrix models We further advance now on the understanding of the 6 half-liberated spheres. Given a subspace X ⊂ SN−1 generated by the elements wi = uzi, where u ∈ C(T) is the standard generator. Since we C,+ , called free complexification of X. See [1], [17]. With this notion in hand, we have: havePi wiw∗i =Pi w∗i wi = 1, we obtain in this way a closed subspace eX ⊂ SN−1 C,+ , we can consider the subalgebra C(eX) ⊂ C(T) ∗ C(X) Proposition 2.1. We have inclusions and equalities as follows, C eSN−1 eSN−1 R C,∗ / eSN−1 eSN−1 C,# SN−1 C,∗ SN−1 C,# making correspond standard coordinates to standard coordinates. / / / / O O O O / O O / / / O O O O O O HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 7 Proof. Consider the diagram in Proposition 1.3, with TSN−1 toriality, we have inclusions as follows: R replaced by SN−1 R . By func- C,◦ C,∗ C,∗ C,# C R C,∗∗ eSN−1 eSN−1 / eSN−1 / eSN−1 Thus we have the square on the left in the statement. / eSN−1 eSN−1 isomorphisms on the right, consider the space eSN−1 Thus we have eSN−1 In order to establish now the lower right isomorphism, consider the space eSN−1 coordinates wi = uzi. We have then eSN−1 following composition, with ε ∗ id on the right, where ε : C(T) → C is the counit: C,∗ ⊂ SN−1 C,∗ C(eSN−1 wiw∗j wk = uziz∗j zk = uzkz∗j zi = wkw∗j wi ) ⊂ C(T) ∗ C(SN−1 C,∗ ) → C(SN−1 C,∗ C,# ⊂ SN−1 C,# , because: wiw∗j = uxi · x∗j u∗ = uxj · x∗i u∗ = wjw∗i w∗i wj = x∗i u∗ · uxj = x∗j u∗ · uxi = w∗j wi C,∗ ) C,# , with In order to prove now the , with coordinates wi = uzi. We have: . As for the converse inclusion, this follows by using the , SN−1 C,◦ As for the converse inclusion, this follows by using the counit, as before. Regarding now SN−1 C,∗∗ , we can use here some 2 × 2 matrix tricks, inspired from C,+ , with coordinates denoted zi, we can consider [10]. Given a closed subspace X ⊂ SN−1 the subalgebra C(X) ⊂ M2(C(X)) generated by the following elements: z′i =(cid:18) 0 0(cid:19) Since these elements are self-adjoint, and their squares sum up to 1, we have X ⊂ SN−1 R,+ . We call this space X doubling of X. We have then the following result: Proposition 2.2. We have inclusions and equalities as follows, z∗i zi (cid:3) SN−1 R,∗ SN−1 R,∗ / SN−1 R,+ / SN−1 C,+ mapping the standard coordinates to the standard coordinates. SN−1 C,∗ SN−1 C / / / / O O O O / O O / / / O O O O / O O 8 TEODOR BANICA Proof. The inclusion on the right appears as the particular case X = SN−1 X ⊂ SN−1 R,+ constructed above. Regarding now the middle inclusion, we have: C,+ of the inclusion z′iz′jz′k =(cid:18) 0 z∗i zi 0(cid:19)(cid:18) 0 z∗j zj 0(cid:19)(cid:18) 0 z∗k zk 0(cid:19) =(cid:18) 0 z∗i zjz∗k ziz∗j zk 0 (cid:19) Now by assuming that the elements zi are the standard coordinates of SN−1 C,∗ , we conclude that we have z′iz′jz′k = z′kz′jz′i, and this gives the middle inclusion. Finally, the inclusion on the left follows by restricting the inclusion in the middle. (cid:3) In order to extend the above notions to the complex case, we begin with a technical result, regarding the relation between the real and the complex spheres. We denote by xi the coordinates on the real spheres. In the odd-dimensional case, we can split half-half the coordinates, and denote them xi, yi. We have then: Proposition 2.3. We have the following diagram, given by zi = xi + iyi, S2N−1 R S2N−1 R SN−1 C / S2N−1 R,∗ S2N−1 R,∗ / SN−1 C,∗∗ / S2N−1 R,+ S2N−1 R,+ / SN−1 C,+ where each S2N−1 R,× ⊂ S2N−1 R,× is obtained via the relationsPi[xi, yi] = 0. Proof. The composition on the left corresponds to the isomorphism SN−1 by zi = xi + iyi. Observe that we have indeed S2N−1 C = S2N−1 , by commutativity. = S2N−1 R R R given We construct now the maps on the right. With z = x + iy we have: Thus, with zi = xi + iyi, we have the following formulae: (x2 i + y2 zz∗ = (x + iy)(x − iy) = x2 + y2 − i[x, y] z∗z = (x − iy)(x + iy) = x2 + y2 + i[x, y] Xi Xi ziz∗i =Xi ziz∗i = Xi z∗i zi = Xi z∗i zi = 1 ⇐⇒ Xi i ) − iXi i ) + iXi i = 1,Xi i + y2 [xi, yi] [xi, yi] (x2 Xi We conclude that we have the following equivalence: x2 i + y2 [xi, yi] = 0 / / / / / / O O O O / / HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 9 But this gives a quotient map C(S2N−1 and this map factorizes as C(S2N−1 R,+ ) → C(SN−1 R,+ ) → C( S2N−1 R,+ ) = C(SN−1 C,+ ), as desired. C,+ ), given by xi = Re(zi), yi = Im(zi), Regarding now the middle maps, we must show that, with zi = xi + iyi, we have: nxi, yi half − commuteo ⇐⇒ nzi, z∗i half − commuteo The " =⇒ " assertion being clear, let us discuss now the "⇐=" assertion. Here the half-commutation relations abc = cba with a, b, c ∈ {zi, z∗i } can be written as follows, in terms of a = x + iy, b = z + it, c = u + iv, with x, y, z, t, u, v self-adjoint: (x + αy)(z + βt)(u + γv) = (u + γv)(z + βt)(x + αy) ∀α, β, γ ∈ {i,−i} Now by looking at the real and imaginary parts, we obtain the following system of equations, once again valid for any choice of α, β, γ ∈ {i,−i}: ((xzu − uzx) + αβ(ytu − uty) + βγ(xtv − vtx) + αγ(yzv − vzy) = 0 α(yzu − uzy) + β(xtu − utx) + γ(xzv − vzx) + αβγ(ytv − vty) = 0 From the 8 possible choices of α, β, γ ∈ {i,−i}, we select now the 4 ones having at most one −i among α, β, γ. The corresponding 4 × 4 determinants being both nonzero, we conclude that the global system, formed by the above 2×8 = 16 equations, is equivalent to the vanishing of all 8 quantities of type xzu − uzx, and we are done. (cid:3) Let us go back now to the question of finding a complex analogue of Proposition 2.2. C,+ , with coordinates denoted xi, yi, we can consider the Given a closed subspace X ⊂ S2N−1 subalgebra C([X]) ⊂ M2(C(X)) generated by the following elements: zi =(cid:18) 0 xi 0(cid:19) + i(cid:18) 0 x∗i y∗i yi 0(cid:19) We call this space [X] complex doubling of X. Observe that we do not have in general [X] ⊂ SN−1 C,+ , because the formulaePi ziz∗i =Pi z∗i zi = 1 are not satisfied. , let us introduce the following manifolds: In relation now with SN−1 C,∗∗ , SN−1 C,◦ consisting of the points of the form S2N−1 C S2N−1 C = ((x, y) ∈ S2N−1 = n(x, y) ∈ S2N−1 C C xi ¯yi ∈ R) (cid:12)(cid:12)(cid:12)Xi (cid:12)(cid:12)(cid:12)xi ¯xj + yi ¯yj ∈ R, xi ¯yj − yi ¯xj ∈ iRo ⊂ S2N−1 C . We have then: Consider as well the manifold T2SN−1 R u(λp, µp), with u ∈ T, (λ, µ) ∈ S1 R ≃ T, and p ∈ SN−1 R 10 TEODOR BANICA Theorem 2.4. We have inclusions of noncommutative spaces as follows, [TS2N−1 R ] / [ S2N−1 C ] SN−1 C / SN−1 C,∗∗ −→ [T2SN−1 R ] / [ S2N−1 C ] TSN−1 R / SN−1 C,◦ mapping the standard coordinates to the standard coordinates. Proof. We have to prove that the 2 × 2 matrix model construction zi = x′i + iy′i, with w′ = (0 ¯w w 0 ), induces morphisms of algebras as follows: C(SN−1 C,∗∗ ) C(SN−1 C,◦ ) C(SN−1 C ) M2(C( S2N−1 C )) M2(C(TS2N−1 R )) −→ / C(TSN−1 R ) M2(C( S2N−1 R )) / M2(C(T2SN−1 R )) We will first construct the morphism C(SN−1 C,∗∗ C obtain the remaining 3 morphisms by factorizing this morphism. ) → M2(C( S2N−1 )), and then we will 1. We first construct the morphism at top left. We recall from Proposition 2.3 above and its proof that with zi = xi + iyi, we have the following equivalence: Xi ziz∗i =Xi i = 1,Xi In our situation now, with zi = x′i + iy′i, and (x, y) ∈ S2N−1 0 z∗i zi = 1 ⇐⇒ Xi x2 i + y2 0 C x′2 i + y′2 i = Xi (cid:18)xi2 Xi [x′i, y′i] = Xi (cid:18)xi ¯yi Xi 0 0 xi2(cid:19) +(cid:18)yi2 ¯xiyi(cid:19) −(cid:18)yi¯xi C,+ ) → M2(C( S2N−1 0 0 0 C [xi, yi] = 0 , we have: yi2(cid:19) =(cid:18)1 0 0 1(cid:19) ¯yixi(cid:19) =(cid:18)0 0 0 0(cid:19) 0 Thus, we have a morphism C(SN−1 )). Now since the matrices x′i, y′i half-commute, the variables zi = x′i + iy′i and their adjoints z∗i = x′i − iy′i half-commute as well, and we therefore obtain a factorization C(SN−1 C,∗∗ 2. We prove now that, when restricting attention to S2N−1 C ) → M2(C( S2N−1 ⊂ S2N−1 C,◦ ⊂ SN−1 C,∗∗ )). , we obtain a model C appears via the relations C . For this purpose, we recall that SN−1 for SN−1 C,◦ ⊂ SN−1 C,∗∗ / O O / O O / O O / O O / / / / / / HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 11 ab∗ = ba∗, a∗b = b∗a. With a = x + iy, b = z + it, these relations are: ((x + iy)(z − it) = (z + it)(x − iy) (x − iy)(z + it) = (z − it)(x + iy) These relations read [x, z] + [y, t] = ±i(xt + tx − yz − zy), so they are equivalent to: Now in terms of our variables zi = x′i + iy′i, we must have: xt + tx = yz + zy x′iy′j + y′jx′i = y′ix′j + x′jy′i ([x, z] + [y, t] = 0 ([x′i, x′j] + [y′i, y′j] = 0 x′y′ =(cid:18)0 x ¯x 0(cid:19)(cid:18)0 y ¯y 0(cid:19) =(cid:18)x¯y (xi ¯xj − xj ¯xi + yi ¯yj − yj ¯yi = 0 (xi ¯xj + yi ¯yj = xj ¯xi + yj ¯yi xi ¯yj − yi ¯xj = xj ¯yi − yj ¯xi ⊂ S2N−1 xi ¯yj + yj ¯xi = yi ¯xj + xj ¯yi 0 C C 0 ¯xy(cid:19) In order to apply these equations to our 2 × 2 matrices, we use the following formula: We are therefore led to the following equations, for the parameter space for SN−1 C,◦ : These latter equations can be written more conveniently, as follows: But these are exactly the equations for S2N−1 3. We prove now that, when restricting attention to TS2N−1 a model for SN−1 commute, and so the variables x′i, y′i must commute. Thus we must have: ⊂ SN−1 C,∗∗ , we obtain . In order to obtain such a model, the variables zi, z∗i must , and we are done. ⊂ S2N−1 C C R xi ¯xj ∈ R, yi ¯yj ∈ R, xi ¯yj ∈ R With λ = x, µ = y the first two conditions read x ∈ λTSN−1 , so let us write x = λup, y = µvq with u, v ∈ T and p, q ∈ SN−1 . The third condition tells us then that we must have u¯v ∈ R, and so v = ±u, and by changing if necessary q → −q, we can assume that we have u = v. We conclude that we have (x, y) = u(λp, µq), and since the point (λp, µq) must belong to the real sphere S2N−1 , y ∈ µTSN−1 , we are done. R R R R 4. We prove now that T2SN−1 R latter model space appears as an intersection, TS2N−1 is the model space for TSN−1 ∩ S2N−1 R C R . By functoriality, this . So, let us pick a point 12 TEODOR BANICA (x, y) ∈ TS2N−1 R , and apply to it the equations for S2N−1 C (xi ¯xj + yi ¯yj ∈ R xi ¯yj − yi¯xj ∈ iR . These equations are: The first equations are automatic, and since the variables in the second equations are real as well, these equations tell us that we must have xi ¯yj = yi¯xj, for any i, j. Now with (x, y) = u(p, q) these latter equations read piqj = qipj, for any i, j. We deduce that we must have (p, q) = (λr, µr) with (λ, µ) ∈ S1 (cid:3) R and r ∈ SN−1 , and we are done. R As an application of the above methods, we have the following result: Proposition 2.5. The inclusions between the 6 half-liberated spheres SN−1 C / SN−1 C,∗∗ / SN−1 C,∗ TSN−1 R SN−1 C,◦ / SN−1 C,# are all proper, at any N ≥ 2. Proof. By using Theorem 1.6, the vertical maps are all proper. For the horizontal maps, we can use Proposition 2.1, Proposition 2.2 and Theorem 2.4: SN−1 C ⊂ SN−1 C,∗∗ is not classical. Thus, SN−1 C,∗∗ found there shows that SN−1 R,∗ . This follows from Proposition 2.2, because the inclusion SN−1 is not classical either. C,# . Here we can use the inclusion eSN−1 C,# from Proposition 2.1. In- don't 6⊂ SN−1 , and so SN−1 , as subspaces C,∗∗ C,# is indeed proper. SN−1 C,◦ ⊂ SN−1 deed, since the standard coordinates wi = uxi on the free complexification eSN−1 satisfy the relations abc = cba, we have eSN−1 of SN−1 C,# ) ⊂ SN−1 C,+ . It follows that (SN−1 SN−1 C,∗∗ ⊂ SN−1 C,∗ . Assuming that this inclusion is an equality, by intersecting with SN−1 C,# R ⊂ SN−1 C,# 6⊂ SN−1 C,∗∗ C,∗∗ ∩ SN−1 ⊂ SN−1 R,∗ C R R x1 ¯y1 + x2 ¯y2 ∈ R x1 ¯x2 + y1 ¯y2 ∈ R x1 ¯y2 − y1¯x2 ∈ iR  C,# is an equality too, contradiction. we would obtain that SN−1 C,◦ ⊂ SN−1 TSN−1 R ⊂ SN−1 C,◦ . Here we must show that SN−1 C,◦ we have embeddings between spheres S1 it is enough to solve the problem at N = 2. So, consider the manifold S3 Theorem 2.4. The equations defining it, over (x1, x2, y1, y2) ∈ S3 C,× ⊂ SN−1 C,× C ⊂ S3 C, are as follows: = SN−1 C,∗∗ ∩ SN−1 C,# is not classical. Since given by x3 = x4 = . . . = xN = 0, C used in / / / / O O O O / O O HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 13 Observe now that these equations are satisfied for the following point: (x1, x2, y1, y2) = 1 √2 (i, 0, 0, 1) The corresponding matrices z1, z2 for this special point are then: z1 = 1 √2(cid:18) 0 −i 0(cid:19) i z2 = 1 √2(cid:18)0 i i 0(cid:19) Now since these two matrices do not commute, this finishes the proof. (cid:3) 3. Quantum groups In this section and in the next one we further advance on the understanding of the 6 half-liberated spheres, by studying the associated quantum isometry groups. Our starting point is the following definition, due to Wang [20]: Definition 3.1. The free analogue of C(UN ) is the universal C∗-algebra C(U + N ) = C∗(cid:16)(uij)i,j=1,...,N(cid:12)(cid:12)(cid:12)u, ut = unitaries(cid:17) with Hopf algebra maps ∆(uij) =Pk uik ⊗ ukj, ε(uij) = δij, S(uij) = u∗ji. As explained in [20], the above formulae define indeed a comultiplication, counit and antipode, and we have a Hopf C∗-algebra in the sense of Woronowicz [23], [24]. Observe that the square of the antipode is the identity, S2 = id. The underlying noncommutative space U + N is a compact quantum group, called free analogue of UN . Observe the analogy with Definition 1.1. We can build on this analogy, by introducing "quantum group analogues" of the spheres in Definition 1.2, simply by imposing the relations there to the standard coordinates of U + N . We obtain in this way: Proposition 3.2. We have an intersection diagram of compact quantum groups UN / U∗∗N / U∗N TON U◦N / U # N N , U◦N being defined inside U + with U∗N , U∗∗N , U # Proof. The quantum groups U∗N , U∗∗N were introduced and studied in [8], [10]. Regarding U # N , our first claim is that its defining relations can be reformulated as follows: N via the relations in Definition 1.2. (cid:16)ab∗ = ba∗, a∗b = b∗a(cid:17) ⇐⇒ (cid:16)ab∗c depends only on {a, b, c}(cid:17) / / / / O O O O / O O 14 TEODOR BANICA Indeed, the implication " =⇒ " can be checked by alternatively using the relations ab∗ = ba∗, a∗b = b∗a, on left and on the right, as follows: ab∗c = ba∗c = bc∗a = cb∗a = ca∗b = ac∗b As for the converse implication, "⇐=", the first formula follows from the following computation, and the proof of the second formula is similar: ab∗c = ba∗c =⇒ Xc ab∗cc∗ =Xc ba∗cc∗ =⇒ ab∗ = ba∗ With the above claim in hand, the construction of ε, S is clear. Concerning now the comultiplication ∆, observe that with Uij =Pk uik ⊗ ukj, we have: uiau∗jbukc ⊗ uaxu∗byucz UixU∗jyUkz =Xabc to permute a, b, c, in a similar way, and this gives the existence of ∆. Now let us permute (ix), (jy), (kz). We can use the same permutation σ ∈ S3 in order Finally, if we set U◦N = U∗∗N ∩ U # Thus, we have the 6 quantum groups in the statement. The inclusions are clear, and (cid:3) the intersection claim UN ∩ U◦N = TON follows as in the proof of Proposition 1.3. N , we obtain as well a compact quantum group. We have as well analogues of the other basic results regarding spheres. First, we have the following analogue of Theorem 1.6 above, basically known since [8]: Proposition 3.3. The projective versions of the 6 quantum groups are: P UN P UN P UN P ON P ON P ON In addition, we have P ON = ON /Z2 and P UN = UN /T. Proof. By functoriality, it is enough to prove that we have inclusions P U∗N ⊂ P UN and P U # N ⊂ P ON . As explained in [8], the first inclusion can be deduced as follows: P U∗N ⊂ (P U∗N )class ⊂ (P U + N )class = P UN Indeed, the first inclusion follows from the fact that the projective version coordinates N , and wia,jb = uiju∗ab commute, the second inclusion follows by functoriality from U∗N ⊂ U + the third inclusion follows from Tannakian duality, as explained in [6]. Regarding now the second inclusion, this follows from P U # the fact that the variables wia,jb = uiju∗ab are self-adjoint over P U # N . N ⊂ P U∗N ⊂ P UN , and from (cid:3) Regarding now the free complexifications, we have the following result, which is much more precise than the one for the spheres, from Proposition 2.1 above: O O O O O O HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 15 Proposition 3.4. The free complexifications of the 6 quantum groups are U∗N U∗N U∗N U # N U # N U # N with all the isomorphisms mapping standard coordinates to standard coordinates. Proof. The arguments in the proof of Proposition 2.1 extend to the quantum group case, and provide us with the following diagram: U∗N U # N N eUN eON Indeed, since the projective version P U∗N is Proposition 3.3, and technology from [1]. classical, we obtain that U∗N , as well as all its subgroups, are amenable. Thus, we can indeed use the results in [1], established there at the level of reduced versions. / eU∗N eU # We must prove now that we have eON = U # N , eUN = U∗N . For this purpose, we can use With notations and terminology from [1], the quantum groups eUN , U∗N , U # N are all easy (called "free" there), of infinite level, and appear as free complexifications. Thus the main result in [1] applies, and shows that these 3 quantum groups must appear as free complexifications of certain intermediate easy quantum groups ON ⊂ O×N ⊂ O+ N . tum group ON ⊂ G ⊂ O+ groups O×N constructed above must satisfy O×N ∈ {ON , O∗N , O+ N}. In order to finish we use the fact, once again from [6], that the projective versions of the quantum groups ON ⊂ O∗N ⊂ O+ N are the quantum groups P ON ⊂ P UN ⊂ P O+ N . In particular, the projective version determines the quantum group. Now since we have PeUN = P UN , P U∗N = P UN , P U # eUN = eO∗N , Thus we have indeed eON = U # On the other hand, we know from [6] that the only non-trivial intermediate easy quan- N is the half-liberation G = O∗N . Thus, each of the 3 quantum Let us discuss now the analogues of the matrix model constructions from section 2 N = P ON , we conclude that we have: above. Following [10], we consider the following compact group: U # U∗N = eO∗N , N = eON N , eUN = U∗N , and we are done. −B A(cid:19) ∈ U2N(cid:12)(cid:12)(cid:12)A, B ∈ MN (C)(cid:27) U2,N =(cid:26)(cid:18) A B (cid:3) O O O O O O / / / O O O O O O 16 TEODOR BANICA We have then the following result, basically from [10]: Proposition 3.5. We have a morphism C(U∗∗N ) → M2(C(U2,N )), given by uij →(cid:18) 0 aij 0(cid:19) + i(cid:18) 0 bij 0(cid:19) ¯bij where aij, bij denote the standard coordinates on U2,N . Proof. The group elements U ∈ U2,N , written U = ( A −B UU∗ = U∗U = U t ¯U = ¯U U t = 1, and we deduce that the matrices A, B satisfy: ¯aij B A) as above, satisfy the relations AA∗ + BB∗ = A∗A + B∗B = At ¯A + Bt ¯B = ¯AAt + ¯BBt = 1 AB∗ = BA∗, A∗B = B∗A, At ¯B = Bt ¯A, ¯ABt = ¯BAt Consider now the target elements wij = a′ij + ib′ij appearing in the statement. The matrix w = (wij) that they form, and its adjoint, are then given by: A + iB 0 (cid:19) 0 A∗ + iB∗ ¯A + i ¯B w =(cid:18) 0 wt =(cid:18) 0 A∗ − iB∗ w∗ =(cid:18) ¯w =(cid:18) 0 (cid:19) 0 At − iBt (cid:19) 0 (cid:19) ¯A − i ¯B Also, the transpose of this matrix, and its complex conjugate, are given by: A − iB By using now the above formulae relating A, B, we obtain: At + iBt 0 ww∗ = w∗w = wt ¯w = ¯wwt = 1 Thus, we have obtained a morphism of algebras C(U + Now since the 2 × 2 matrices a′ij, b′ij half-commute, so do the elements wij, w∗ij, and so N ) → M2(C(U2,N )). our morphism factorizes through the algebra C(U∗∗N ), as claimed. (cid:3) With the above result in hand, we can suitably modify the "complex doubling" opera- tion X → [X] constructed in section 2 above, as follows: Definition 3.6. Given X ⊂ U2,N , we define [[X]] ⊂ U∗∗N by stating that C([[X]]) is the image of the representation C(U∗∗N ) → M2(C(X)), given by uij → a′ij + ib′ij. In other words, our construction is defined by the following diagram: C(U∗∗N ) / M2(C(U2,N )) C([[X]]) / M2(C(X)) As an example here, the results in [10] show that we have [[U2,N ]] = U∗∗N . We can now formulate an analogue of Theorem 2.4 above, as follows: / / HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 17 Theorem 3.7. We have inclusions of noncommutative spaces [[TO2,N ]] / [[U2,N ]] UN / U∗∗N −→ [[T2ON ]] / [[U′N ]] TON / U◦N with T2ON and U′N being certain closed subgroups of U2,N . Proof. We follow the method in the proof of Theorem 2.4. The computations there apply to the present situation, with a 2N rescaling factor for the spheres, and we obtain that the "parameter spaces" for the quantum groups G = UN , U◦N , TON , i.e. the biggest closed subspaces X ⊂ U2,N producing embeddings [[X]] ⊂ G, are as follows: UN → U2,N ∩ 2N · TS4N 2−1 U◦N → U2,N ∩ 2N · S4N 2−1 TON → U2,N ∩ 2N · T2S2N 2−1 R C R We will compute these three spaces, and then show that they are indeed groups. 1. We first compute the parameter space for UN . We know that a matrix U ∈ U2,N belongs to this space precisely when there exists z ∈ T such that V = zU is real. Thus V must belong to the group O2,N = U2,N ∩ O2N , and the parameter space is: 2. Regarding now the parameter space for U◦N , this appears from U2,N via the defining relations for S4N 2−1 C , from section 2 above, which are as follows: TO2,N =(cid:26)z(cid:18) A B −B A(cid:19) ∈ U2N(cid:12)(cid:12)(cid:12)z ∈ T, A, B ∈ MN (R)(cid:27) (aij¯akl + bij¯bkl ∈ R aij¯bkl − bij¯akl ∈ iR 3. Finally, the parameter space for TON is best obtained by intersecting the parameter B A) as above. spaces for UN , U◦N . Indeed, let us pick a matrix U ∈ TO2,N , written U = z( A −B satisfy the above two equations. As in the sphere case, the variable z ∈ T cancels, and the first equation is automatic, and the second equation reads aijbkl = bijakl. We therefore conclude, as in the sphere case, that the parameter space for TON is: Then U belongs to the parameter space for TON when its entries eaij = zaij,ebij = zbij T2ON =(cid:26)z(cid:18) cA sA −sA cA(cid:19)(cid:12)(cid:12)(cid:12)z ∈ T,(cid:18) c s −s c(cid:19) ∈ SO2 ≃ T, A ∈ ON(cid:27) / O O / O O / O O / O O 18 TEODOR BANICA 4. We are left with checking that the parameter spaces are indeed groups. Since this is clear for TO2,N , T2ON , it remains to verify that the following space is a group: U′N =(cid:26)(cid:18) A B −B A(cid:19) ∈ U2N(cid:12)(cid:12)(cid:12) aij¯akl + bij aij¯bkl − bij¯akl ∈ iR(cid:27) ¯bkl ∈ R We have 1 ∈ U′N , and U ∈ U′N =⇒ U∗ ∈ U′N is clear as well, because at the level of coordinates, the passage U → U∗ is given by (aij, bij) → (¯aji,−¯bji), and this transformation preserves the solutions of the defining equations for U′N . Regarding now the multiplication axiom, we use the following formula: (cid:18) A B −B A(cid:19)(cid:18) C D −D C(cid:19) =(cid:18) AC − BD AD + BC −AD − BC AC − BD(cid:19) Assuming now that the two matrices on the left belong to U′N , we have: (AC − BD)ij(AC − BD)kl + (AD + BC)ij(AD + BC)kl (aipcpj − bipdpj)(¯akq¯cql − ¯bkq ¯dql) + (aipdpj + bipcpj)(¯akq ¯dql − ¯bkq¯cql) ¯dql) ¯bkq)(cpj¯cql + dpj ¯dql) + (aip (aip¯akq + bip ¯bkq − bip¯akq)(dpj¯cql − cpj Now since the above 4 quantities are respectively in R, R, iR, iR, the summand is real, and hence the whole sum is real as well. Thus, we have checked the first equations. For the second equations, the proof is similar. We have indeed: = Xpq = Xpq = Xpq = Xpq (AC − BD)ij(AD + BC)kl − (AD + BC)ij(AC − BD)kl ¯dql + ¯bkq¯cql) − (aipdpj + bipcpj)(¯akq¯cql − ¯bkq (aipcpj − bipdpj)(¯akq (aip¯akq + bip¯bkq)(cpj ¯dql − dpj¯cql) + (aip¯bkq − bip¯akq)(cpj¯cql − dpj ¯dql) ¯dql) Now the quantities which appear are respectively in R, iR, iR, R, so the summand is (cid:3) imaginary, and hence the whole sum is imaginary as well, and we are done. 4. Affine isometries In this section we show that the 6 quantum groups introduced above appear as affine quantum isometry groups of the 6 spheres, and we deduce some consequences. We use the following formalism, inspired from [13]: Definition 4.1. We say that G ⊂ U + N acts affinely on X ⊂ SN−1 zi →Xa uia ⊗ za C,+ when defines a morphism of algebras Φ : C(X) → C(G) ⊗ C(X). HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 19 Observe that such a morphism Φ is automatically coassociative and counital, in the sense that we have (id ⊗ Φ)Φ = (∆ ⊗ id)Φ and (ε ⊗ id)Φ = id. Thus, we have a coaction, in the usual sense. The basic example is UN y SN−1 We agree to denote the 6 half-liberated quantum groups by U×N , and the corresponding , via Φ(f )(U, x) = f (Ux). C 6 half-liberated spheres by SN−1 × . First, we have the following result: Proposition 4.2. We have an affine action U×N y SN−1 × . Proof. We must prove that the formula in Definition 4.1 defines a morphism of algebras C(SN−1 × ). For this purpose, we just have to show that the elements × ) → C(U×N )⊗ C(SN−1 Zi =Pa uia ⊗ za satisfy the defining relations for SN−1 As a first observation, the quadratic relationsPi ZiZ∗i =Pi Z∗i Zi = 1 follow from the biunitarity of u. For the remaining relations, we perform a case-by case analysis. we have indeed the following computation: × . SN−1 C,∗∗ , SN−1 C,∗ . For SN−1 C,∗ ZiZ∗j Zk =Xabc uiau∗jbukc ⊗ zaz∗b zc =Xabc ukcu∗jbuia ⊗ zcz∗b za = ZkZ∗j Zi For SN−1 C,∗∗ SN−1 , SN−1 C,◦ the proof is similar, by removing all the ∗ exponents. C,# . It is enough to do the verification for SN−1 C,# , and here we have: ZiZ∗j =Xab uiau∗jb ⊗ zaz∗b =Xab ujbu∗ia ⊗ zbz∗a = ZjZ∗i The proof of Z∗i Zj = Z∗j Zi is similar, by moving the ∗ exponents on the left. TSN−1 . It is enough to do the verification for SN−1 , SN−1 C R C because UN is known to act on SN−1 C , with coaction map as in the statement. . But the result here is clear, (cid:3) We will prove now that the actions in Proposition 4.2 are universal. For this purpose, we use an old 3-step method from [9], where the result was established for SN−1 . The idea is to: (1) establish linear independence results for the products of coordinates, (2) deduce from this the precise conditions on G ⊂ U + N which allow an action, and (3) solve the quantum group question left, by using an antipode/relabel trick. C In our case, the linear independence lemma that we will need is: Lemma 4.3. The following variables are linearly independent: (1) {zaz∗b1 ≤ a ≤ b ≤ N}, over SN−1 C,◦ (2) {zazbzc1 ≤ a ≤ c ≤ N, 1 ≤ b ≤ N}, over SN−1 . C,◦ (3) {zaz∗b zc1 ≤ a ≤ c ≤ N, 1 ≤ b ≤ N}, over SN−1 . C,∗∗ . (1) Here we can use the isomorphism P N−1 Proof. This follows by using various 2 × 2 matrix models for the spheres: the variables {paba ≤ b} are linearly independent over P N C,◦ ≃ P N R , this gives the result. R given by pab = zaz∗b . Indeed, since , found in Theorem 2.4 20 TEODOR BANICA , R C R C , 2 (SN−1 p − q Indeed, since for p, q ∈ SN−1 pi+qi (SN−1 2 and finally the defining relations for S2N−1 )2 ⊂ S2N−1 C above. Our first claim is that we have an inclusion, as follows: (2) We use here the model z = x′ + iy′, with (x, y) ∈ S2N−1 2i (cid:19) (p, q) →(cid:18)p + q )2 ⊂ S2N−1 +(cid:0) pi−qi 2 (cid:1)2 2 (cid:1)2 R we havePi(cid:0) pi+qi . Moreover, sincePi · pi−qi are both trivially satisfied. When restricting the parameter space to (SN−1 2(cid:18) 0 2(cid:18) 0 0 (cid:19) + zizjzk =(cid:18) 0 pi 0(cid:19)(cid:18) 0 0(cid:19)(cid:18) 0 qi − pi Observe now that we have the following formula: 0 (cid:19) =(cid:18) 0 pi 0(cid:19) 0 (cid:19) 0(cid:19) =(cid:18) 0 )2, the model becomes: pi − qi pi + qi pi + qi piqjpk qipjqk zi = pk qk pj qj qi qi 1 1 C R 2 = 0, we have in fact (SN−1 = 1, we obtain an embedding )2 ⊂ S2N−1 , C R Now since the variables {wi ¯wjwki ≤ k} on the right are linearly independent over , so are the 2 × 2 matrices {ziz∗j zki ≤ k}, and this gives the result. (cid:3) SN−1 C We will need as well, several times, the following lemma: Lemma 4.4. If the standard coordinates uij on a compact quantum group G ⊂ U∗N satisfy the relations abc = cba, then we have G ⊂ U∗∗N . Proof. We must prove that abc = cba for any a, b, c ∈ {uij, u∗ij}, and by using the involu- tion, it is enough to check that the following relations hold, for any a, b, c ∈ {uij}: abc = cba, ab∗c = cb∗a, abc∗ = c∗ba The first two relations hold by assumption, and we must therefore deduce the third relations from them. For this purpose, we can use the diagrammatic formalism in [5], or (SN−1 R )2, so are the 2 × 2 matrices {zizjzki ≤ k}, and this gives the result. Now since the variables {piqjpki ≤ k} on the right are linearly independent over (3) Here we can use the model z = x′ + iy′, with (x, y) ∈ S2N−1 , from Theorem 2.4, C . Indeed, if we denote by with the parameter space restricted to SN−1 wi = xi + iyi the coordinates on SN−1 C , the matrix model formula becomes: R R xi + iyi ⊂ S2N−1 C ≃ S2N−1 0 (cid:19) =(cid:18) 0 wi 0(cid:19) 0(cid:19) =(cid:18) 0 0(cid:19)(cid:18) 0 wk wk wi ¯wj wi ¯wjwk xi + iyi zi =(cid:18) 0 0(cid:19)(cid:18) 0 ¯wj wi ¯wjwk 0 (cid:19) Now observe that we have the following formula: ziz∗j zk =(cid:18) 0 wi wi HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 21 rather its unitary extension, which applies to the easy quantum group G ⊂ U + from the first two relations. Indeed, in the Tannakian category of G, we have: N coming • • ◦ ◦ ●●●●●●●●●●● ✇✇✇✇✇✇✇✇✇✇✇ ◦ ◦ ◦ ◦ • • = ◦ • ■■■■■■■■■■■■ ✉✉✉✉✉✉✉✉✉✉✉✉ ◦ ◦ • ◦ Thus the relations abc = cba imply the relations abc∗ = c∗ba, and we are done. (cid:3) Now back to the quantum isometries, and to the 3-step method from [9], Lemma 4.3 and Lemma 4.4 provide us with the first step. We will perform the second and third step altogether, first for SN−1 , and then for SN−1 C,# . First, we have: C,∗∗ ⊂ SN−1 C,∗ Proposition 4.5. The affine actions of U∗∗N , U∗N on SN−1 C,∗∗ are universal. C,◦ ⊂ SN−1 , SN−1 C,∗ Proof. This is a routine computation, based on the antipode/relabel trick in [9]. Consider indeed a compact quantum group G ⊂ U + in mind, let us fix as well a symbol × ∈ {∅,∗}. We have then: ZiZ×j Zk =Xabc uiau×jbukc ⊗ zaz×b zc N , and let Zi =Pa uia ⊗ za. With Lemma 4.4 Assuming now that the variables z1, . . . , zN are subject to the relations zaz×b zc = zcz×b za, some of the terms on the right coincide. By taking into account the various cases, and by merging these terms, we can write the above formula as follows: ZiZ×j Zk = Xa<c,b6=a,c + Xa<c + Xa6=b + Xa (uiau×jbukc + uicu×jbuka) ⊗ zaz×b zc (uiau×jaukc + uicu×jauka) ⊗ zaz×a zc uiau×jbuka ⊗ zaz×b za uiau×jauka ⊗ zaz×a za By interchanging i ↔ k, we have as well a similar formula for ZkZ×j Zi. Now by using the linear independence of the variables on the right, coming from Lemma 4.3 (2) and (3) above, we conclude that the relations ZiZ×j Zk = ZkZ×j Zi are equivalent to the following system of equations, where [x, y, z] = xy×z − zy×x: (1) [uia, ujb, ukc] = [uka, ujb, uic], for a, b, c distinct. (2) [uia, uja, ukc] = [uka, uja, uic]. (3) [uia, ujb, uka] = 0. 22 TEODOR BANICA Here we have merged the relation coming by comparing the fourth sums, namely [uia, uja, uka] = 0 for any a, with the relations coming from the second and third sums, in order to drop the assumptions a 6= c, a 6= b appearing there. equivalent to [uia, ujb, ukc] = 0, regardless of the indices i, j, k and a, b, c. Our claim, which will prove the result, is that the above equations (1,2,3) are in fact Let us first process the relations (1). By applying the antipode and then the involution we obtain [uck, ubj, uai] = [uci, ubj, uak], for any a, b, c distinct, and then by relabelling we obtain [ukc, ujb, uia] = [uka, ujb, uic], for any i, j, k distinct. Now by comparing with the original relations (1), we have several cases, and we are led to the following relations: (1a) [uia, ujb, ukc] = 0, for a, b, c distinct, and i, j, k distinct. (1b) [uia, ujb, ukc] = [uka, ujb, uic], for a, b, c distinct, and i, j, k not distinct. We further process now the relations (1b). Since the relations at i = k are trivial, and those at i = j, j = k are equivalent, we can assume that we have i = j, and we get: (1b') [uia, uib, ukc] = [uka, uib, uic], for a, b, c distinct. Let us process now the above relations (2). By applying the antipode and the involution we obtain [uck, uaj, uai] = [uci, uaj, uak], and by relabelling, we obtain: (2') [ukc, uib, uia] = [uka, uib, uic]. The point now is that the relations (1b'), (2') can be merged. Indeed, in view of (2'), the relations (1b') simplify to: (1b") [uia, uib, ukc] = 0, for a, b, c distinct. Now, with these relations (1b") in hand, the relations (2') are automatic for a, b, c distinct. Thus, what is left from the relations (2') is: (2") [ukc, uib, uia] = [uka, uib, uic], for a, b, c not distinct. As a partial conclusion, the relevant relations are (1a), (1b"), (2"), (3). Now let us further process the relations (2"). Since these relations are automatic at a = c, and are equivalent at a = b, b = c, we can assume a = b, and we obtain: (2∗) [uka, uia, uic] = [ukc, uia, uia]. Now by applying the antipode and then the involution we obtain [uci, uai, uak] = [uai, uai, uck], and by relabelling we obtain [uka, uia, uic] = [uia, uia, ukc]. By comparing now with the original relations (2∗) we are led to the following two relations: (2∗a) [uka, uia, uic] = 0. (2∗b) [ukc, uia, uia] = 0. Summarizing, the relevant relations are (1a), (1b"), (2∗a), (2∗b), (3). Now observe that all these relations are of the form [uia, ujb, ukc] = 0, the precise assumptions being: (1a) i, j, k distinct, and a, b, c distinct. (1b") i = j, and a, b, c distinct. (2∗a) i = j, and b = c. (2∗b) i = j, and a = b. (3) a = c. HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 23 Our claim is that, from these relations, we can deduce that we have [uia, ujb, ukc] = 0, regardless of the indices. Indeed, let us look first at (1b"), (2∗a), (2∗b). These relations are of the same nature, involving the assumption i = j, and since by (3) the relation [uia, ujb, ukc] = 0 holds as well for i = j, a = c, we can merge them. We conclude that the relations [uia, ujb, ukc] = 0 hold, under the following assumptions: (1a) i, j, k distinct, and a, b, c distinct. (2+) i = j. (3) a = c. Now by using the antipode, the relations (2+), (3) tell us precisely that we have [uia, ujb, ukc] = 0, whenever i, j, k are not distinct, or when a, b, c are not distinct. But this is exactly the complementary of the case covered by (1a), and we are done. (cid:3) Let us discuss now the remaining spheres, SN−1 C,# . We have here: Proposition 4.6. The affine actions of U◦N , U # , SN−1 C,# are universal. C,◦ ⊂ SN−1 N on SN−1 C,◦ Proof. We use the same method as in the proof of Proposition 4.5. We first discuss the case of the sphere SN−1 C,# . With Zi =Pu uia ⊗ za, we have: By using now the relations zaz∗b = z∗b za, this formula can be written as: uiau∗jb ⊗ zaz∗b ZiZ∗j =Xab (uiau∗jb + uibu∗ja) ⊗ zaz∗b +Xa (ujau∗ib + ujbu∗ia) ⊗ zaz∗b +Xa ZiZ∗j =Xa<b ZjZ∗i =Xa<b uiau∗ja ⊗ zaz∗a ujau∗ia ⊗ zaz∗a By interchanging i ↔ j, we have as well the following formula: Now since by Lemma 4.3 (1) the variables on the right are independent, we conclude that the relations ZiZ∗j = ZjZ∗i are equivalent to the following conditions: (1) ujau∗ib − uibu∗ja = ujbu∗ia − uiau∗jb. (2) ujau∗ia = uiau∗ja. Here we have dropped the assumption a < b for the first relations, because by symmetry we have them for a > b too, and these relations are automatic at a = b. By applying now the antipode to these relations, and then by relabelling, we succesively obtain: ubiu∗aj − uaju∗bi = uaiu∗bj − ubju∗ai ujau∗ib − uibu∗ja = uiau∗jb − ujbu∗ia Now by comparing with the original relations (1), we conclude that: ujau∗ib − uibu∗ja = ujbu∗ia − uiau∗jb = 0 24 TEODOR BANICA = SN−1 C,# ∩ SN−1 C,∗∗ C,# problem. In other words, the standard coordinates on a quantum group G y SN−1 C,# must satisfy ab∗ = ba∗. Similarly, these coordinates must satisfy as well a∗b = b∗a. We conclude that we must have G ⊂ U # is needed. Consider indeed a quantum group G y SN−1 C,◦ over SN−1 C,◦ , the above computations apply, and we obtain G ⊂ U # N . , our claim is that no new computation . Since Lemma 4.3 (1) was valid N , and we are therefore done with the SN−1 Regarding now the sphere SN−1 C,◦ On the other hand, since Lemma 4.3 (2) was valid as well over SN−1 C,◦ , the computations in the proof of Proposition 4.5 apply as well, with the choice × = ∅, and show that the standard coordinates on G must satisfy the relations abc = cba. In order to conclude, we use Lemma 4.4. We already know that the standard coordinates on G satisfy the relations abc = cba, and from G ⊂ U # N ⊂ U∗N we obtain that the relations ab∗c = cb∗a are satisfied as well. Thus Lemma 4.4 applies, and gives G ⊂ U∗∗N . We therefore conclude that we have G ⊂ U # N ∩ U∗∗N = U◦N , and we are done. (cid:3) We can now formulate our main result in this section, as follows: Theorem 4.7. We have the following correspondence SN−1 C / SN−1 C,∗∗ / SN−1 C,∗ UN / U∗∗N / U∗N TSN−1 R SN−1 C,◦ / SN−1 C,# TON U◦N / U # N between the 6 spheres, and their affine quantum isometry groups. −→ Proof. The result for SN−1 argument in [9] showing that we have indeed G+(TSN−1 remaining results follow from Proposition 4.5 and Proposition 4.6 above. is known since [9], the result for TSN−1 ) = G(TSN−1 is similar, with the ) = TON , and the (cid:3) C R R R Summarizing, we have now some basic understanding of the 6 half-liberated spheres. There are, however, many questions left. A first series of questions concerns the ergod- icity, uniqueness, and possible faithfulness of the U×N -invariant integration on SN−1 × . A second series of questions concerns the construction of the Laplacian, and notably of its eigenvalues, and the possible Riemannian structure of SN−1 × . Finally, a third series of questions concerns the possible twisting of the above results. See [2], [4]. 5. Half-liberated manifolds We discuss now the extension of some of the results in sections 1-4, with the complex . There is in fact a sphere SN−1 lot of work to be done here, and we have so far only very partial results. replaced by more general algebraic manifolds X ⊂ SN−1 C C / / / / O O O O / O O / / / / O O O O / O O HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 25 Generally speaking, the problem is that of constructing, under suitable assumptions on X ⊂ SN−1 C , a half-liberation diagram for it, as follows: X / X∗∗ / X∗ X− X◦ / X # The starting point is Theorem 4.7 above. Forgetting that on the right we have quantum itself, we have as well the isometry groups, we can see that, besides the sphere X = SN−1 rescaled unitary group X = 1√N Proposition 5.1. We have embeddings as follows, given by zij = 1√N / SN 2−1 C,∗∗ UN as example. Indeed, we have: SN 2−1 / U∗∗N / U∗N UN C C uij, / SN 2−1 C,∗ TON U◦N / U # N TSN 2−1 R SN 2−1 C,◦ / SN 2−1 C,# −→ . × U∗N ∩ SN 2−1 U×N = 1√N whose images are given by 1√N Proof. Since the fundamental corepresentation u = (uij) of the quantum group U + N is N ⊂ SN 2−1 given by zij = 1√N uij. Now since the quantum groups U×N in the statement appear by imposing to the standard coordinates uij the same relations as those for the coordinates zij on the corresponding spheres SN 2−1 (cid:3) biunitary, we havePij uiju∗ij =Pij u∗ijuij = N. Thus we have an embedding U + U∗N ∩ SN 2−1 UN , suggest an approach via "lifting projective versions". More precisely, given X ⊂ SN−1 , consider its projective version P X ⊂ P N R . The general idea is then to define X∗∗, X∗/X◦, X # as being the "biggest" submanifolds of the corresponding spheres, having P X/P X− as projective versions. C . Also, let X− = X ∩ TSN−1 The examples that we have so far, X = SN−1 , so that P X− = P X ∩ P N U×N = 1√N and X = 1√N , we obtain 1√N C,+ × × C R C . In order for this idea to work, X, X− themselves must be the lifts to SN−1 C , TSN−1 R of their projective versions P X, P X−. So, let us first recall that we have: Proposition 5.2. For a subspace X ⊂ SN−1 (1) X is the lift to SN−1 (2) X is invariant under the action of T, given by u · z = (uzi)i. of its projective version P X ⊂ P N C . is the lift to TSN−1 In addition, in this case, X− = X ∩ TSN−1 C R R C , the following are equivalent: of P X− = P X ∩ P N R . / / / / O O O O / O O / / / / O O O O / O O / / / / O O O O / O O 26 TEODOR BANICA Proof. Since the quotient map π : SN−1 lifting condition X = {x ∈ SN−1 π(x) ∈ P X} is equivalent to the T-invariance of X. R → P N so the lifting condition X− = {x ∈ TSN−1 of X−. But if X is T-invariant, then so is X−, and this gives the last assertion. C satisfies π(z) = π(z′) ⇐⇒ z′ ∈ Tz, the R satisfies as well σ(z) = σ(z′) ⇐⇒ z′ ∈ Tz, σ(x) ∈ P X−} is equivalent to the T-invariance Also, the quotient map σ : TSN−1 C → P N (cid:3) C R The other problem is that the general noncommutative manifolds Z ⊂ SN−1 C,+ have in fact two projective versions, one given by pij = ziz∗j , and the other one given by qij = z∗j zi. In order to deal with this issue, best is to assume that all our manifolds Z are "conjugation- stable", in the sense that C(Z) has an anti-automorphism given by zi → z∗i . itself, the stability under conjugation, which comes from an action of Z2, can be combined with the stability under the action of T, coming from Proposition 5.2 above. In this case, we say that X is O2-invariant. Observe that for the manifold X ⊂ SN−1 C We can now formulate our half-liberation construction, as follows: Definition 5.3. If X ⊂ SN−1 C is O2-invariant, we set X− = X ∩ TSN−1 R , and we define X / X∗∗ / X∗ X− X◦ / X # by the fact that X∗∗, X∗/X◦, X # are the conjugation-stable lifts of P X/P X−. As a basic example, for the sphere X = SN−1 C we have P X = P N C , P X− = P N R , the lifting problem is trivial, and we obtain the 6 half-liberated spheres themselves. Observe also that, due to our T-invariance assumption on X, all the 6 spaces appearing in the above diagram are the lifts of their projective versions P X, P X−. In general, the fact that the above lifts exist indeed follows by dividing the corresponding algebras by suitable ideals. Let us record a more precise result here: Proposition 5.4. The spaces X× appear via C(X×) = C(SN−1 × )/ < I, J >, where I/J =(ker[C(P N ker[C(P N C ) → C(P X)] R ) → C(P X−)] at × = ∗∗,∗ at × = ◦, # regarded as linear subspaces of C(SN−1 × ), via the embeddings pij = ziz∗j /qij = z∗j zi. / / / / O O O O / O O HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 27 Proof. At the algebra level, the lifts at × = ∗∗,∗ and at × = ◦, # in Definition 5.3 above are by definition the universal solutions to the following problems: C(P N C ) C(SN−1 × ) C(P N R ) C(SN−1 × ) C(P X) / C(X×) C(P X−) / C(X×) But the solutions to these problems are given by formula in the statement. (cid:3) As an illustration, consider the space X = 1√N TN formed by the points z ∈ SN−1 C satisfying zi = 1√N Proposition 5.5. We have the following (rescaled) half-liberation diagram, for any i. Here we have X− = 1√N 2 , and the result is: TZN TN [Z⋄⋄N TZN 2 dZ◦N / dZ⋄N / dZ#N where ⋄,⋄⋄, #,◦ are the group-theoretic analogues of the operations ∗,∗∗, #,◦. Proof. Observe first that given a discrete group Γ =< g1, . . . , gN >, we have an embedding C,+ , given by zi = 1√N In our case, we deduce from pii = qii = 1 gi, and that over PbΓ we have pii = qii = 1 bΓ ⊂ SN−1 various lifts. Thus the rescaled lifts are group duals, given by dZ×N = cFN ∩ √N SN 2−1 N , over the . But this gives the result, with the ⋄, # constructions obtained respectively by imposing the conditions ab−1c = cb−1a, ab−1 = ba−1 to the standard generators of FN , and with the ⋄⋄,◦ constructions being obtained by further imposing the relations abc = cba. (cid:3) N that we have ziz∗i = z∗i zi = 1 N . × Let us check the fact that the example X = 1√N UN is covered as well: Proposition 5.6. For the rescaled unitary group X = 1√N construction produces the (rescaled) 6 half-liberated quantum groups U×N . UN , the abstract half-liberation / / / / / / / / / / / O O O O / O O 28 TEODOR BANICA Proof. We first discuss the lifting problem for P UN ⊂ P N 2 rescaling of the lift inside SN−1 C,× , we have the following series of implications: C . If we denote by U××N the uu∗ = u∗u = ut¯u = ¯uut = 1, over UN =⇒ Xk =⇒ Xk =⇒ Xk uik ¯ujk =Xk pik,jk =Xk uiku∗jk =Xk uki¯ukj =Xk pki,kj =Xk ukiu∗kj =Xk ¯ukiukj =Xk qki,kj =Xk u∗kiukj =Xk Thus we have an inclusion U××N ⊂ U + N ∩ √NSN 2−1 For the lifting problem for P ON ⊂ P N 2 middle relations, over N · P UN , hold over N · P ON as well. =⇒ uu∗ = u∗u = ut¯u = ¯uut = 1, over U××N U + C,× ¯uikujk = δij, over UN qik,jk = δij, over N · P UN u∗ikujk = δij, over U××N , we conclude that we have U××N = U×N , as desired. N . But since U×N is by definition given by U×N = R we can use the same proof, because the above (cid:3) Let us work out as well a "discrete" analogue of Proposition 5.6. Consider the group KN ⊂ UN of matrices which are monomial, in the sense that each row and each column has exactly one nonzero entry. Its free version K + N is then defined via the relations ab∗ = a∗b = 0, for any a 6= b on the same row or column of u. See [1]. N ⊂ U + With the notations, we have the following result: Proposition 5.7. With X = 1√N KN we obtain the following diagram, KN / K∗∗N / K∗N THN K◦N / K # N , where on the right we have the quantum groups K # rescaled by 1√N Proof. The space K−N = √N X− is given by K−N = KN ∩ √N TSN 2−1 obtain K−N = KN ∩ TON = THN , where HN ⊂ ON is the hyperoctahedral group. that these lifts satisfy K××N ⊂ U×N . Also, for j 6= k we have: Let us first compute the various lifts of P KN . We already know from Proposition 5.6 , and we therefore N ∩ U # N . N = K + R uij ¯uik = ¯uijuik = uji¯uki = ¯ukiuki = 0, over KN =⇒ pij,ik = qij,ik = pji,ki = qji,ki = 0, over P KN =⇒ uiju∗ik = u∗ijuik = ujiu∗ki = u∗jiuki = 0, over K××N / / / / O O O O / O O HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 29 We conclude that the lifts appear inside U×N via the relations ab∗ = a∗b = 0, for any a 6= b on the same row or column of u. Thus we have K××N ⊂ K + The lifting problem for P THN = P HN is similar, by using the same computation. (cid:3) N , and we are done. Summarizing, the half-liberation operation that we constructed leads to quite natural objects, in all the cases investigated so far. In particular, we can now formulate: Theorem 5.8. For the half-liberations of the sphere X = SN−1 C we have G+(X×) = G(X)× with the quantum isometry groups being taken in an affine sense. Proof. This is just a reformulation of the results that we proved before, in Theorem 4.7, by using the abstract half-liberation formalism developed above. (cid:3) Observe that the formula established above could be thought of as being related to the various rigidity results of type G+(X) = G(X), from [9], [14]. In general, it is quite unclear what exact assumptions on X ⊂ SN−1 results. This is an interesting question, that we would like to raise here. C could lead to such 6. Real versions In this section we discuss a number of more specialized results, concerning the real ver- sions of our half-liberations, obtained by imposing the conditions zi = z∗i to the standard coordinates. First, we have the following elementary result: Proposition 6.1. The real versions of the half-liberations X× are XR / X∗R / X∗R XR XR , and X∗R = X∗∗ ∩ SN−1 R,+ . / XR R where XR = X ∩ SN−1 Proof. This follows indeed from the last assertion in Proposition 1.4 above, which tells us that taking the real versions amounts in intersecting with SN−1 (cid:3) R,∗ /SN−1 R . We can axiomatize the construction X → X∗R, as follows: Proposition 6.2. Given an O2-invariant closed subset X ⊂ SN−1 X∗R ⊂ SN−1 R,∗ Proof. This follows from Proposition 5.4 above, because the variables pij = ziz∗j and qji = z∗i zj being now equal, the conjugation-stable lift becomes a plain lift. (cid:3) appears by lifting the projective version P X ⊂ P N C . , the closed subset C / / / / O O O O / O O 30 TEODOR BANICA At the level of examples now, we have the following result: Proposition 6.3. We have the following plain/rescaled real half-liberations, (SN−1 R )∗ = SN−1 R,∗ / (ZN (ON )∗ = O∗N , (HN )∗ = H∗N coming respectively from the complex manifolds SN−1 C /TN , UN , KN . 2 , 2 )∗ = dZ⋄N Proof. The first assertion is clear from the comments made after Definition 5.3. Regarding the second assertion, we can use here Proposition 5.5, which tells us that the rescaled real-half liberation in question is: Finally, the last two assertions are clear from Proposition 5.6 and Proposition 5.7. (cid:3) We have as well the following matrix model result, obtained by using the doubling [Z⋄⋄N ∩ SN−1 R,+ = [Z⋄⋄N 2 2 = dZ⋄N C operation X → X, constructed in section 2 above: Proposition 6.4. If X ⊂ SN−1 Proof. We recall from Proposition 2.2 that we have SN−1 holds for X = SN−1 0(cid:19) =(cid:18)zi ¯zj z′i(z′j)∗ =(cid:18) 0 zi 0(cid:19)(cid:18) 0 is O2-invariant, then X ⊂ X∗R. itself. In general now, observe that we have: ⊂ SN−1 R,∗ ¯zizj(cid:19) ¯zj zj ¯zi 0 0 C C , and so the result C Now since X ⊂ SN−1 is O2-invariant, z → ¯z induces an automorphism of C(X), and so an automorphism of C(P X). We can therefore "cut" the lower part of the above matrix, and we obtain PX = P X. Thus X lifts P X, and so X ⊂ X∗R, as desired. (cid:3) Regarding now the quantum isometry groups, the fact that we have G+(SN−1 R,∗ was already known from [2]. We can improve now this result. We use: Definition 6.5. A closed subspace X ⊂ Y is called k-saturated when the dimension of span(ze1 ik ) does not decrease via C(Y ) → C(X), for any e1, . . . , ek ∈ {1,∗}. i1 . . . zek ) = O∗N Observe that the 1-saturation of X ⊂ SN−1 , which is equivalent to the fact that the coordinates z1, . . . , zN ∈ C(X) are linearly independent, is needed in order to define the affine quantum isometry group G+(X), as a closed subgroup of U + The 2-saturation condition is a familiar one as well, because for a subset X ⊂ SN−1 , We have the following result, regarding the 3-saturated sets: this condition implies that we have G+(X) = G(X), as shown in [9]. N . See [13]. C C Theorem 6.6. We have the "half-classical rigidity" formula provided that X ⊂ SN−1 C G+(X∗R) ⊂ O∗N is O2-invariant and 3-saturated. HALF-LIBERATED MANIFOLDS, AND THEIR QUANTUM ISOMETRIES 31 Proof. Assuming that X ⊂ SN−1 as well, and we conclude that the half-liberation X∗R ⊂ SN−1 R,∗ method in the proof of Proposition 4.5 applies, and gives the result. Thus, the variables {zaz×b zca ≤ c} with × = 1,∗ are linearly independent, and so the is 3-saturated too. (cid:3) C is 3-saturated, the doubling X ⊂ SN−1 R,∗ is 3-saturated Observe the similarity between the above result and Theorem 5.8. As a conclusion, our various results suggest that a certain analogue of the rigidity result in [14] should hold in real and complex half-liberated affine geometry. Finding such a general result, however, looks like a quite difficult question. References [1] T. Banica, A note on free quantum groups, Ann. Math. Blaise Pascal 15 (2008), 135 -- 146. [2] T. Banica, Liberations and twists of real and complex spheres, J. Geom. Phys. 96 (2015), 1 -- 25. [3] T. Banica and B. Collins, Integration over compact quantum groups, Publ. Res. Inst. Math. Sci. 43 (2007), 277 -- 302. [4] T. Banica and D. Goswami, Quantum isometries and noncommutative spheres, Comm. Math. Phys. 298 (2010), 343 -- 356. [5] T. Banica and R. Speicher, Liberation of orthogonal Lie groups, Adv. Math. 222 (2009), 1461 -- 1501. [6] T. Banica and R. Vergnioux, Invariants of the half-liberated orthogonal group, Ann. Inst. Fourier 60 (2010), 2137 -- 2164. [7] H. Bercovici and V. Pata, Stable laws and domains of attraction in free probability theory, Ann. of Math. 149 (1999), 1023 -- 1060. [8] J. Bhowmick, F. D'Andrea and L. Dabrowski, Quantum isometries of the finite noncommutative geometry of the standard model, Comm. Math. Phys. 307 (2011), 101 -- 131. [9] J. Bhowmick and D. Goswami, Quantum isometry groups: examples and computations, Comm. Math. Phys. 285 (2009), 421 -- 444. [10] J. Bichon and M. Dubois-Violette, Half-commutative orthogonal Hopf algebras, Pacific J. Math. 263 (2013), 13 -- 28. [11] A. Chirvasitu, On quantum symmetries of compact metric spaces, J. Geom. Phys. 94 (2015), 141 -- 157. [12] B. Collins and P. ´Sniady, Integration with respect to the Haar measure on the unitary, orthogonal and symplectic group, Comm. Math. Phys. 264 (2006), 773 -- 795. [13] D. Goswami, Existence and examples of quantum isometry groups for a class of compact metric spaces, Adv. Math. 280 (2015), 340 -- 359. [14] D. Goswami and S. Joardar, Rigidity of action of compact quantum groups on compact, connected manifolds, preprint 2013. [15] H. Huang, Faithful compact quantum group actions on connected compact metrizable spaces, J. Geom. Phys. 70 (2013), 232 -- 236. [16] J. Quaegebeur and M. Sabbe, Isometric coactions of compact quantum groups on compact quantum metric spaces, Proc. Indian Acad. Sci. Math. Sci. 122 (2012), 351 -- 373. [17] S. Raum, Isomorphisms and fusion rules of orthogonal free quantum groups and their complexifica- tions, Proc. Amer. Math. Soc. 140 (2012), 3207 -- 3218. [18] R. Speicher, Multiplicative functions on the lattice of noncrossing partitions and free convolution, Math. Ann. 298 (1994), 611 -- 628. [19] D.V. Voiculescu, K.J. Dykema and A. Nica, Free random variables, AMS (1992). 32 TEODOR BANICA [20] S. Wang, Free products of compact quantum groups, Comm. Math. Phys. 167 (1995), 671 -- 692. [21] S. Wang, Quantum symmetry groups of finite spaces, Comm. Math. Phys. 195 (1998), 195 -- 211. [22] D. Weingarten, Asymptotic behavior of group integrals in the limit of infinite rank, J. Math. Phys. 19 (1978), 999 -- 1001. [23] S.L. Woronowicz, Compact matrix pseudogroups, Comm. Math. Phys. 111 (1987), 613 -- 665. [24] S.L. Woronowicz, Tannaka-Krein duality for compact matrix pseudogroups. Twisted SU(N) groups, Invent. Math. 93 (1988), 35 -- 76. T.B.: Department of Mathematics, Cergy-Pontoise University, 95000 Cergy-Pontoise, France. [email protected]
1301.2507	3	1301	2015-07-30T06:24:16	Extremal unital completely positive normal maps and its symmetries	[ "math.OA", "math-ph", "math-ph" ]	We consider the convex set of ( unital ) positive ( completely ) maps from a $C^*$ algebra $\cla$ to a von-Neumann sub-algebra $\clm$ of $\clb(\clh)$, the algebra of bounded linear operators on a Hilbert space $\clh$ and study its extreme points via its canonical lifting to the convex set of ( unital ) positive ( complete ) normal maps from $\hat{\cla}$ to $\clm$, where $\hat{\cla}$ is the universal enveloping von-Neumann algebra over $\cla$. If $\cla=\clm$ and a ( complete ) positive operator $\tau$ is a unique sum of a normal and a singular ( complete ) positive maps. Furthermore, a unital complete positive map is a unique convex combination of unital normal and singular complete positive maps. We used a duality argument to find a criteria for extremal elements in the convex set of unital completely positive maps having a given faithful normal invariant state. In our investigation, gauge symmetry in Stinespring representation and Kadison theorem on order isomorphism played an important role.	math.OA	math	EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES ANILESH MOHARI 5 1 0 2 l u J 0 3 ] Abstract We consider the convex set of ( unital ) positive ( completely ) maps from a C ∗ algebra A to a von-Neumann sub-algebra M of B(H), the algebra of bounded linear operators on a Hilbert space H and study its extreme points via its canonical lifting to the convex set of ( unital ) positive ( complete ) normal maps from A to M, where A is the universal enveloping von-Neumann algebra over A. If A = M and a ( complete ) positive operator τ is a unique sum of a normal and a singular ( complete ) positive maps. Furthermore, a unital complete positive map is a unique convex combination of unital normal and singular complete positive maps. We used a duality argument to find a criteria for extremal elements in the convex set of unital completely positive maps having a given faithful normal invariant state. In our investigation, gauge symmetry in Stinespring representation and Kadison theorem on order isomorphism played an important role. . A O h t a m [ 3 v 7 0 5 2 . 1 0 3 1 : v i X r a 1. Introduction: An irreversible quantum dynamics in discrete time is governed by a complete posi- tive map [St] on a non commutative algebra of observables. A given physics problem often fixes the algebra of observables made of elements from a suitable unital C∗- algebra A over the field of complex numbers. Once a unital C∗-algebra A is fixed, a mathematically challenging problem is to classify or parametrize all complete positive maps by more elementary mathematical objects. E. Størmer [Stø] and W. Arveson [Ar] gave mathematical foundation to this classification problem within the frame of Krein-Milman-Choquet theory [Ph]. A simplified presentation of Størmer- Arveson's approach towards the same problem is also given by M. D. Choi [Ch] for A = Md(C), the algebra of d-dimensional matrices over the field of complex number C. In other similar context with possible extensions of these results, extreme points are studied in details in many follow up works, here we cite some of them [Pas], [HMP], [Ra], [Ts] for a historical account. On the other hand, a state of a bipartite system, either in quantum information theory [MW,Pa,PSa,Oh] or two sided infinite quantum spin chain models [Mo2,Mo3] on a lattice, give rises unital completely positive map as follows: Let AL and AR be two C∗-algebras of a unital C∗ algebra A such that A = AL ⊗ AR, where C∗ norm of A is taken to be the maximal one. For two faithful states ωL and ωR on AL and AR respectively, we consider the non empty convex set CωL,ωR = {ω : A → C, a state with ωAL = ωL, ωAR = ωR} Then each element ω ∈ CωL,ωR determines unique unital completely positive maps τω : AR → πωL (AL)′′ 1991 Mathematics Subject Classification. 46L. Key words and phrases. Operator system, Arveson-Hahn-Banach extension theorem, complete order isomorphism . ... 1 2 ANILESH MOHARI such that and τω : AL → πωR(AR)′′ ωLτω = ωR ωR τω = ωL Furthermore, the maps ω → τω, ω → τω are affine continuous maps in Bounded weak topology [Ar,Pa]. The map τω → τω is an affine one to one map satisfying the duality relation for all x ∈ AL and y ∈ AR (1) ω(x ⊗ y) =< JLτω(y)JLζωL , πωL (x)ζωL >=< JRπωR(x)JRζωR , τ (y)ζωR > where ωL and ωR are cyclic and separating vectors for πω(AL)′′ and πωR(AR)′′ in HωL and HωR respectively. Furthermore, JL and JR are conjugate operators of Tomita [BR1] associated with ωL and ωR respectively. Thus the convex set CP1(A, M) of unital completely positive maps from a C∗ algebra A to a von- Neumann algebra M plays an important role in quantum communication channels [Par,PSa,MW,Oh]. Our main aim in this paper is to find finer structures of its extreme points and investigate how two extreme points are related for a possible classification of its extreme points upto cocycle conjugation. In this paper, we start with a little different approach then [Stø,Ar] and generalize classical work of J. Tomiyama [To1,To2,To3] on Lebesgue decomposition of a norm one projection on a von-Neumann algebra M. We follow closely S. Sakai's Lebesgue decomposition [Ta] of a positive functional on a von Neumann algebra to prove: an extremal unital completely positive map on a von Neumann factor M is either a normal map or a singular map. Furthermore, any other unital completely positive map τ on a von Neumann factor M is a unique convex combination of a normal τσ and a singular τs unital completely positive map on M. In this analysis, Stinespring representation played no role. More generally, let CP1(A, M) be the convex set of unital completely positive map from a C∗-algebra A to a von Neumann algebra M, which is acting on separa- ble Hilbert space H over the field of complex numbers C and B(H) be the algebra of bounded operators on H. For a given unital completely positive map τ : A → M, we have unique upto isomorphism minimal Stinespring representation (K, π, V ) (2) τ (x) = V π(x)V ∗ for all x ∈ A, where π : A → B(K) is a unital ∗-homomorphism and K is the cyclic Hilbert space generated by π and H i.e. {π(x)f : x ∈ A, f ∈ H} is total in K and V ∗ : H → K is an isometry. Let π : A∗∗ → A be the universal lifting map so that πi = π on A, where A∗∗ is the universal von-Neumann algebra over A isomorphic to the dual Banach space of the dual space A∗ of A. Let zπ be the element in the centre of A∗∗, the support projection of the representation π, then we define elements τσ, τs ∈ CP (A, M) by τσ(x) = V π(i(x)zπ)V ∗ τs(x) = V π(i(x)zπ)V ∗ for all x ∈ A. Furthermore, we prove that the collection CP σ 1 (A, M) = {τσ : τ ∈ CP1(A, M)} EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 3 is also a sequentially closed convex set in Bounded Weak topology and any ele- ment τ ∈ CP1(A, M) is a convex combination of elements from CP σ 1 (A, M) and CP s 1 (A, M), where CP s 1 (S, M) = {τ ∈ CP1(A, M) : τσ = 0} In particular, using duality relation (1) and symmetries in Stinespring represen- tation (2), we prove τ is an extremal element in CPφ = {τ ∈ CP1(A, M) : φτ = φ} where M = πφ(A)′′ for a faithful normal state of M if, and only if there exists no non-trivial λ = (λk j ) with entries in M′ satisfying the relation Xα,β vαλα β = 0, Xα,β β v∗ vαλα β v∗ β = 0, where λ → λ is an unital order-isomorphism map on Mnτ (M′)), (v∗ α : 1 ≤ α ≤ nτ ) and (v∗ α : 1 ≤ α ≤ nτ ) are Krause elements in minimal Stinespring representation of τ and τ respectively. Furthermore, if τ admits inner representation, then same condition holds with entries tα β in the centre of M and order isomorphism is given by tα β = tβ α The proof relies on Størmer-Arveson version of Radon-Nikodym theorem for com- pletely positive maps and Kadison theorem [Ka1] on unital order isomorphism on C∗-algebras. In particular, this result generalize a well known criteria [LS] valid for M = Mn(C) and φ = tr, the normalize trace. In the last section, we deal with simplest situation Cφ = CωL,ωR where we have taken AL = Mn(C) and AR = Mn(C) with ωL = ωR = φ, where φ is a faithful state on Mn(C). We prove that an element ψ ∈ Cφ is a factor state if τψ is an extremal element in CPφ. Furthermore, ψ is a pure state in Cφ if, and only if τψ ∈ CPφ is an automorphism on M. The last statement is a generalization of a theorem proved in [Oh] with φ = tr. 2. A decomposition of completely positive maps: Let A be a unital C∗-algebra. A linear map τ : A → A is called positive if τ (x) ≥ 0 for all x ≥ 0. Such a map is automatically bounded with norm τ = τ (I). A linear map τ : A → A is called n-positive [St] (CP ) if τ ⊗ In : A ⊗ Mn → A ⊗ Mn is positive, where τ ⊗ In is defined by (xi j ) are in A. If τ is n-positive for each n ≥ 1, τ is called completely positive. j )) with elements (xi j ) → (τ (xi For each unital representation π : A → B(Hπ) of A, we denote by M(π) the von- Neumann algebra π(A)′′ generated by π(A) in B(Hπ). A representation (π, Hπ) is called universal if for any representation (ρ, Kρ) there exists a σ-weak continuous ∗-homomorphism ρ of M(π) onto M(ρ) such that ρ(x) = ρ(π(x)) If (π, Hπ) is a universal representation of A, M(π) is called universal enveloping von-Neumann algebra of A. The universal enveloping von-Neumann algebra is uniquely determined upto isomorphism. We recall now two standard constructions of universal enveloping von-Neumann algebras, each one has its own merit. Let A∗ be the Banach space dual of A and A∗∗ be the double dual of A i.e. u be the inclusion map of A dual Banach space of A∗. Let i : A → A∗∗ ≡ MA 4 ANILESH MOHARI into A∗∗. For a representation π : A → B(Hπ), we set von-Neumann algebra π → A∗. We set Mπ = π(A)′′ and the Banach space adjoint linear map by πt : M∗ now linear map (πt)∗ : (Mπ)∗ → A∗ by restricting πt to (Mπ)∗. Finally, we set notation π : A∗∗ → Mπ for the Banach space adjoint map of (πt)∗. Thus by our construction π is a normal map, being the dual of a bounded linear map on their pre-dual Banach spaces. The enveloping von-Neumann algebra A∗∗ of a C∗-algebra A is defined to be the double dual A∗∗ of A. That A∗∗, being a dual of a Banach space, is a von- Neumann algebra [Sa]. The following proposition says A∗∗ is indeed the universal von-Neumann algebra of A. Proposition 2.1. Let (π, Hπ) be a representation of A. We have the following properties: (a) π : A∗∗ → Mπ is a linear map which is continuous with respect to weak∗ topologies on A∗∗ and Mπ. The map π takes the norm closed unit ball of A∗∗ onto the norm closed unit ball of Mπ; (b) π ◦ i = π on A; (c) π is a unital completely positive map from A∗∗ onto Mπ; (d) For any central element z ∈ A∗∗, π(z) is an element in the center of Mπ. Proof. For statement (a) and (b), we refer to Lemma 2.2 in Chapter 3 of [Ta]. Statement (c) and (d) are as well known but could not find a suitable ready reference. Here we indicate a proof. Since π is a positive map and so is its transpose πt. Thus the restriction (πt)∗ is also positive. That π is positive follows as (πt)∗ is positive and onto. For n−positive property of π, we note that the universal enveloping algebra over Mn(A) is Mn(A∗∗) and the canonical map i ⊗ In is the inclusion map of Mn(A) into Mn(A∗∗). Furthermore, for a representation π : A → B(Hπ), we also have π ⊗ In ◦ i ⊗ In = π ⊗ In. This shows in particular that π is a completely positive map and its restriction on i(A) is a representation. By Kadison-Schwarz inequality [Ka2] for unital completely positive map we have π(x∗y) = π(x∗)π(y) for x ∈ A∗∗ and y ∈ i(A) since π(y∗y) = π(y∗)π(y) for y ∈ i(A). Taking adjoint in the above relation we also get π(y∗x) = π(y∗)π(x). Operator algebras involved are being ∗-closed and also π being onto, we get π(x) ∈ Mπ T M′ π if x ∈ A∗∗T A∗∗′ . Let S(A) be the convex set of states on A and (Hφ, πφ, ζφ) be the GNS repre- sentation associated with a state φ on A. The universal von-Neumann algebra of A is given by MA u = {πu(x) : x ∈ A}′′, where πu(x) = ⊕φ∈S(A)πφ(x) is the direct sum of representations on Hu = ⊕φ∈S(A)Hφ Since any representation (π, Hπ) is a direct sum of cyclic representation, MA universal von-Neumann algebra of A. u is the We defined left and right action of a C∗-algebra A on its dual A∗ by (3) < y, xω >=< yx, ω >, < y, ωx >=< xy, ω > where < ., . > denotes evaluation of a functional on a given element in the Banach space. The crucial property that A is an algebra is used here to define these actions on A. A subspace of A∗ is called A invariant if the subspace is invariant by both left and right action. Given a A-invariant subspace V of A∗ EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 5 A projection in the center of A∗∗ determines uniquely a representation of A upto quasi-equivalence i.e. a representation π is quasi-equivalent to the sub-representation x → πu(x)zπ for some central projection zπ in πu(A)′′. The projection zπ is called support of the representation in A∗∗, determined unique by the representation π upto isomorphism. Given a central projection zπ, there exists a uniquely deter- mined A-invariant norm closed subspace V (π) of A∗ given by πt(M∗(π)) i.e. The range of the map πt : M∗ → A∗ restricted to M∗. Conversely, given a norm closed A-invariant subspace of A∗, there exists a central projection z in A∗∗ such that the representation ρ : x → πu(x)z is non-degenerate and V (ρ) = V . For details, we once more refer to section 2 of Chapter 3 in [Ta]. Let M be a von-Neumann algebra acting on a Hilbert space (H, < ., . >) over the field of complex numbers, where the inner product is linear in second variable. Let M∗ be the pre-dual Banach space of M. For a von-Neumann algebra M, M∗ is a proper subset unless M is finite dimensional. However the unit ball of M∗ is a dense subset in the unit ball of M∗ in the weak∗ topology on M∗. Nevertheless M∗ is sequentially closed [Ta] and given any bounded linear functional ω on M there is a unique element ωσ ∈ M∗ so that ω = ωσ + ωs determined by ω − ωσ = minω′∈M∗ω − ω′. The element ωs is called purely singular on M unless it is the zero element. In the following we describe such a decomposition via universal enveloping algebra of M. Now we take A = M, a von-Neumann algebra in Proposition 2.1 and π : M → M ⊆ B(H) be the identity representation. Let zπ be the support projection of the representation π : M → M in M∗∗. Thus π(z) is an element in the center of M, where π is the lift of π to the universal enveloping algebra defined in Proposition 2.1. Proposition 2.2. Let M be a von-Neumann algebra with its pre-dual space M∗ and dual M∗. Then there exist a unique central projection z ∈ M∗∗ so that M∗ = M∗z, where M∗∗ acts natural on the Banach space M∗ given by < y, xω >=< yx, ω >, < y, ωx >=< xy, ω > for x, y ∈ M∗∗ and ω ∈ M∗. Proof. The von-Neumann algebra M is dense in weak∗ topology of M∗∗ ( since unit ball of M∗ is dense in the unit ball of M∗ ) and M∗ is a M∗∗-invariant norm closed subspace of M∗. Thus by Theorem 2.7 in Chapter -III [Ta], we conclude that M∗ = M∗z. Proposition 2.3. An element ω ∈ M∗ determines uniquely an element ωσ ∈ M∗ defined by ωσ(x) =< ω, zi(x) > for all x ∈ M and the map ω → ωσ ∈ M∗ is positive linear contractive map on the Banach space M∗. The element ωs(x) =< ω, (1 − z)i(x) > in M∗ is singular provided z 6= 1. The decomposition ω = ωσ + ωs is also uniquely determined and ω = ωσ + ωs. Proof. We refer chapter 3 in [Ta,Chapter 3.2] for details. 6 ANILESH MOHARI Let A be a C∗ algebra and M be a von-Neumann algebra acting on a Hilbert space H. For a positive map τ : A → M ⊆ B(H), we set a positive map τ : A∗∗ → M, the adjoint map of (τ t)∗ : M∗ → A∗, where τ t : M∗ → A∗ is the adjoint map of τ : A → M. Thus we have the lifting property where we recall i : A → A∗∗ is the canonical inclusion map. τ ◦ i = τ A positive linear map τ : M → M is called normal if l.u.b.τ (xα) = τ (l.u.b.xα) for any bounded increasing net xα in M where l.u.b. denotes least upper bound. We will use notations P σ and CP σ(M) for the convex set of positive and completely positive normal maps on M. When there is no room for confusion,, we will simply denote them by P σ and CP σ respectively. In contrast, a positive non-zero map τ : M → M is called singular if φτ is either zero or a purely singular positive functional for any positive normal functional φ of M. We will use notations P s(M) and CP s(M) for the convex set of positive and completely positive singular maps on M respectively. When there is no room for confusion, we will simply denote them by P s and CP s respectively. The following proposition says along the same vein now for the class of positive or completely positive maps on M. The result is well known for quite some time [To2] for completely positive map. Proposition 2.4. Given a positive map τ on M, there is a unique normal positive map τσ on M and a singular positive map τs such that τ = τσ + τs, where τσ, τs are determined uniquely by the decomposition ωτσ = (ωτ )σ, ωτs = (ωτ )s of a normal state ω. Furthermore, the set of normal positive maps on M is se- quentially closed in the set of positive maps on M. Furthermore, if τ is n-positive, then τσ and τs are also n-positive. Proof. Let M∗∗ be the universal von-Neumann algebra of M and i : M → M∗∗ be the canonical inclusion map of M into M∗∗. We write for an element ω ∈ M∗ (ωτ )σ(x) =< ωτ, zi(x) > where z is the central projection in M∗∗. The map ω → (ωτ )σ determines a normal positive linear map τ σ on M by ωτσ(x) =< ωτ, zi(x) > for all x ∈ M and ω ∈ M∗. It also shows clearly that τσ is n-positive as τσ ⊗ In = (τ ⊗ In)σ for n-positive map τ . Similarly we also have ωτs(x) =< ωτ, (1 − z)i(x) > for all x ∈ M, ω ∈ M∗ and the induced map τs is also n-positive if τ is so. Let τn be a sequence of such normal maps on M and its bounded weak limit is τ i.e. ω(τn(x)) → ω(τ (x)) for all ω ∈ M∗ and x ∈ M. Let τs be the singular part of τ . Then singular part of ωτn also converges to singular part of ωτ in weak∗ topology. Since each τn is normal and so is ω, we get ωτ n s = 0. Since this holds for all ω ∈ M∗, we arrive at our desired claim that τs = 0. A non-zero positive map is called purely singular in short singular if its normal part of the unique decomposition contributes the zero map. The convex set of EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 7 unital completely positive maps on M is compact in Bounded-Weak topology of William Arveson [Ar, Pa]. One natural question that arises at this point: when a unital completely positive map on M is a convex combination of unital normal and singular completely positive maps? Proposition 2.5. Let M be von-Neumann algebra and z be central projection in the universal enveloping algebra M∗∗ such that M∗ = zM∗. A positive map τ : M → M is a convex combination of two unital positive maps τσ ∈ CPσ and τs ∈ CPs with λ ∈ [0, 1] i.e. (4) τ = λτσ + (1 − λ)τs if and only if τ (z) is a scaler where τ is the unique completely positive unital normal map from M∗∗ to M such that following relation holds: (5) for all x ∈ M∗∗ and ω ∈ M∗. Furthermore, the map τ → τ is affine and one to one. < ωτ, x >=< ω, τ (x) > Proof. By Proposition 2.4, we have (ωτ )σ(x) =< ωτ, zi(x) >=< ω, τ (zi(x)) > for all x ∈ M and ω ∈ M∗. Uniqueness of the decomposition into normal and singular completely positive maps ensures that for x ∈ M. So we also have x ∈ M such that τ = τσ + τs. τσ(x) = τ (zi(x)) τs(x) = τ ((I − z)i(x)) If τ (z) ∈ M is a scaler say λ = τ (z) ∈ [0, 1], we can rewrite τ = τσ + τs as in (4) by obvious modifications. On the converse, if τ is a convex combination of two unital elements given as in (4), we conclude by uniqueness of the decomposition, τ (z) = λ is a scaler. A unital completely positive map τ : A → B(H) on a unital C∗ algebra A admits a minimal Stinespring representation [St] (6) where π : A → B(Hτ ) is a ∗-representation of A into B(Hτ ), which is the Hilbert space completion of the kernel given on the set A ⊗ H by k(x ⊗ ζ, y ⊗ η) =< ζ, τ (x∗y)η > τ (x) = V π(x)V ∗, and V ∗ : H → Hs is an isometry from H into Hs defined by V ∗ : ζ → I ⊗ ζ such that {π(x)V ∗ζ : x ∈ A, ζ ∈ H} spans Hτ . Such a minimal Stinespring triplet (Hτ , π, V ∗) is uniquely determined modulo unitary equivalence i.e. if (H′, π′, V ′∗) be another triplet associated with τ with cyclic property, then we get W ∗ : π(x)V ∗ζ → π′(x)V ′∗ζ, x ∈ A, ζ ∈ H is an unitary operator so that (7) W ∗V ∗ = V ′∗ and W ∗π(x)W = π′(x), x ∈ A An alternative constructions of τσ and τs for unital CP map τ : M → M are given as follows. 8 ANILESH MOHARI Let τ (x) = V π(x)V ∗ be the Stinespring minimal representation of τ : M → B(H) i.e. π : M → B(K) is the unique unital representation in a Hilbert space K and V ∗ : K → H is the unique isometry ( modulo unitary equivalence ). Let zπ be the support projection of the representation π : M → B(K) i.e. is quasi-equivalent to π′ : M → M∗∗ defined by π′(x) = πu(x)zπ Note that support projection does not depend on the choice of the minimal Stine- spring dilation π we choose as zπ = zπ′ by Proposition 2.12 in Chapter 3 [Ta] for any choice π′. Furthermore, by the duality relation (5), we verify the universal lifting property of τ → τ as follows: for ω ∈ M∗ and x ∈ M < ω, τ (i(x)) > =< ωτ, i(x) > =< ωτ, x > i.e. τ ◦ i = τ In particular, we have by Proposition 2.5, for all x ∈ M and thus and πσ(x) = π(i(x)zπ) πs(x) = π(i(x)(I − zπ)) τσ(x) = V π(zπi(x))V ∗ τs(x) = V π((I − zπ)i(x))V ∗ where zπ is the support of π in M∗∗ and π : M∗∗ → B(K) is the lifting map of the representation π : M → B(K) defined as in Proposition 2.1. By Proposition 2.1 (d), projection π(zπ) is an element in the center of Mπ = {π(x) : x ∈ M}′′. So in particular, we have τ (zπ) = V π(zπ)V ∗ ∈ M. We recall by Proposition 2.1 (d), π(zπ) is an element in the centre of M. Thus τσ(I) is scaler for a factor M and we arrive at our main result of this section. Theorem 2.6. Let τ be an unital completely positive map on a von-Neumann factor M. Then there exists a decomposition of τ given by for some λ ∈ [0, 1] and unital normal τσ and singular τs completely positive maps respectively. Furthermore, if λ ∈ (0, 1) then such a decomposition is unique. τ = λτσ + (1 − λ)τs 3. Radon-Nikodym theorem for completely positive unital maps: In this section, we review a characterization of unital complete positive maps τ : A → B(H) [Stø, Ar] to be extremal elements in CP1(A, M) with some finer additional results for our main results. Proposition 3.1. Let τ : A → B(H) be a completely positive map on a C∗ algebra A and η be another completely positive map on A so that for some positive constant c, η(x) ≤ cτ (x) for all x ∈ A+. Then there exists a non-negative element T ′ ∈ π(A)′ such that (8) η(x) = V π(x)T ′V ∗, EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 9 where (K, π, V ∗) is the minimal Stinespring triplet associated with τ . Furthermore, for each x ∈ M, π(x) and T ′ commutes with the representation ρ : τ (A)′ → B(K) defined by (9) ρ(y) : x ⊗ f = x ⊗ yf Proof. Without loss of generality we assume that A is a closed ∗-subalgebra of B(H), where H is a Hilbert space. We recall Stinespring minimal representation τ (x) = V π(x)V ∗ where π : A → B(K) is the representation induced by the map y ⊗ f → xy ⊗ f where H is von-Neumann's completion of algebraic tensor product A ⊗ H with the kernel k(x ⊗ f, y ⊗ g) =< f, τ (x∗y)g > Now we set another sesqui-linear form on A ⊗ H defined by s(x ⊗ f, y ⊗ g) =< f, η(x∗y)g > Being a kernel, we have Cauchy-Schwarz inequality to check the following steps: s(x ⊗ f, y ⊗ g)2 ≤ s(x ⊗ f, x ⊗ f )s(y ⊗ g, y ⊗ g) ≤ c2x ⊗ f y ⊗ g So there exists a bounder operator T ′ on K such that s(x ⊗ f, y ⊗ g) =< x ⊗ f, T ′y ⊗ g >. We claim that T ′ ∈ π(A)′. Proof follows once we check the following steps: < f, η(x∗zy)g >= s(x ⊗ f, zy ⊗ g >=< x ⊗ f, T ′π(z)y ⊗ g > and < f, η(x∗zy)g >= s(z∗x ⊗ f, y ⊗ g >=< π(z∗)x ⊗ f, T ′y ⊗ g > This completes the proof. We briefly recall Bounded-Weak (BW) topology [Chapter 7 in Pa] on the convex set P1(A, M) ( CP1(A, M) ) of unital positive maps ( completely map ) from a C∗- algebra A to a von-Neumann algebra M ⊆ B(H). A net of unital positive maps τα : A → M ⊆ B(H) converges in BW topology if and only if τα(x) converges to τ (x) in weak∗ topology of von-Neumann algebra M. Furthermore, these topology makes the convex set P1(A, M) (CP1(A, M)) compact and furthermore, BW topology is metrizable if A and as well as M∗ are separable as Banach spaces. Given an element τ ∈ P (A, M), there exists a unique lifting τ ∈ P (A∗∗, M) such that τ ◦ i = τ on A. By our construction τ is a normal map. Conversely, given a normal positive map τ : A∗∗ → M, τ is the unique lift of τ : A → M given by τ = τ ◦ i. Let τ : A → M be a completely positive map and τ (x) = V π(x)V ∗ be its Stinespring representation with π : A → B(K) be a ∗-representation and V ∗ : H → K be a bounded operator. Let π : A∗∗ → B(K) be the universal lifting map in Proposition 2.1 and zπ be the the support projection of the representation π in A∗∗. We define two completely positive maps from A → B(K) by πσ(x) = π(i(x)zπ) πs(x) = π(i(x)(I − zπ)) for all x ∈ A. We set also completely positive maps from A to B(H) by τσ(x) = V πσ(x)V ∗ τs(x) = V πs(x)V ∗ 10 ANILESH MOHARI for all x ∈ A. Thus by our construction we have (10) τ = τσ + τs We say τ is normal if τs = 0. Similarly, τ is called singular if τσ = 0. The uniqueness part of the lifting theorem in Proposition 2.1 says that τ : A → M has a unique decomposition τ = τσ + τs as in Proposition 2.4, where τσ and τs are normal and singular CP map from A to B(H) defined above. Thus this notions indeed generalize the notion of normal and singular complete positive map to C∗ algebras. We use notions CP σ(A, M) and CP s(A, M) respectively for normal and singular CP maps. A routine now also says that CP σ(A, M) is sequentially closed in CP (A, M). For a proof, we take a sequence τ n ∈ CP σ(A, M) and τ ∈ CP (A, M) such that φτ n(x) → φφτ (x) for all normal state φ on M. Since each φτ n is a normal state on M, we get (φτ )s = φτs = 0. This holds for all normal states, we get τs = 0. Proposition 3.2. CP σ 1 (A, M) is a convex face of CP1(A, M). Proof. Let τ0, τ1 ∈ CP σ 1 (A, M) and τ = λτ1 + (1 − λ)τ0. We recall elements π : A → B(K) and V ∗ : H → K in Proposition 3.1 and find τ0(x) = V (T0π(x))V ∗ and τ1(x) = V (T1π(x))V ∗ for unique non-negative elements T0, T1 ∈ π(A)′ such that λT1 + (1 − λ)T0 = I. Thus (τk)s(x) = V (Tk π(i(x)zπ)V ∗ = 0 for k = 0, 1. By adding them up, we get τs = 0. This shows CP σ 1 (A, M) is a convex set. By Proposition 3.1, if τ ∈ CP σ(A, M) and η ∈ CP (A, M) such that η ≤ τ 1 (A, M) and τ0, τ1 ∈ CP1(A, M) so that 1−λ τ + . Thus both τ1, τ1 ∈ CP σ(A, M). This completes the then η ∈ CP σ(A, M). Let τ ∈ CP σ τ = λτ1 + (1 − λ)τ0 for some λ ∈ (0, 1). Then we have τ1 ≤ 1 on non-negative elements A∗∗ proof. λ τ and also τ0 ≤ 1 As an application of Theorem 2.6 and Choquet theorem [Ph], we have the fol- lowing theorem. Theorem 3.3. Let τ ∈ CP1(A, M) be a unital CP map from a unital C∗ algebra A to a von-Neumann algebra M acting on a Hilbert space H. Then the following hold: (a) If M is a factor, then we have τ = λτσ + (1 − λ)τs for some unital normal τσ and singular τs CP maps from A into M. Furthermore, if λ ∈ (0, 1), then τσ and τs are determined uniquely. (b) If A and M∗ are separable as Banach space, then there exists a regular Borel probability measure µ on the set of extreme points CP e 1 (A, M) satisfying τ = Zτe∈CP e 1 (A,M) τedµ(e) Furthermore, if τ is normal (respectively singular), the support of the regular mea- sure µ is also on the set of normal (respectively singular) extremal elements. In particular, we also have the decomposition given in (a). EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 11 Proof. (a) is simple application of Theorem 2.6, where we use the uniqueness of the lifting τ → τ and affine property of the map. First part of (b) is a simple application of Choquet theorem [Ph]. The set of unital normal completely positive maps CP σ 1 (A, M) need not be a closed subset in BW topology and thus need not be compact in general in BW topology. However, for the last part, we only need to how that the set of normal extremal elements is Borel measurable. In fact, it is enough if we show that the set of unital normal maps are Borel measurable i.e. we need to show the set {φ(τ (x)) : τ ∈ CP σ 1 (A, M)} for each x ∈ A and normal state φ of M is a Borel measurable set in C. Since the BW topology is now metrizable and the map τ → φ(τ (x)) being sequentially continuous on CP σ 1 (A, M), we get the map τ → φ(τ (x)) is in fact continuous in BW topology and thus the set of unital normal CP maps is Borel measurable. The last part of the statement follows once we split the integral into sum of two integrals of disjoint measurable sets consists of normal and singular extreme points of CP1(A, M). Proposition 3.4. Let A be a C∗-algebra and τ : A → B(H) be a unital com- pletely positive map and τ (x) = V π(x)V ∗ be the unique upto isomorphism minimal Stinespring representation. Then following are equivalent (a) τ is extremal in CP1; (b) V ΛV ∗ = 0 for Λ ∈ π(A)′ if and only if Λ = 0; Proof. Let τ = λτ1 + (1 − λ)τ0 for some τ0, τ1 ∈ CP1 and λ ∈ (0, 1). By Proposition 2.7 we have τ0(x) = V π(x)T ′V ∗ for some non-negative T ′ ∈ π(A)′. τ0 being unital we also have V T ′V ∗ = I i.e. V (I − T ′)V ∗ = 0. In case (b) is true then T ′ = I and thus τ0 and so τ1 = τ0 = τ . This proves (b) implies (a). For the converse Λ ∈ π(A)′ such that V ΛV ∗ = 0. Same holds if we replace symmetric or anti-symmetric part of Λ. Thus it enough if prove (b) for self-adjoint Λ. Since it is a bounded operator, we assume without loss of generality that −I ≤ Λ ≤ I. So both 0 ≤ I − Λ ≤ 2I and 0 ≤ I + Λ ≤ 2I are operators in π(A)′. We write I − Λ = W ∗W and I − Λ = W ′∗W ′ for some W, W ′ in π(A)′ and check that ′∗V ∗ Since τ τ = 1 is extremal, we conclude that τW = τW ′ = τ . Thus (K, π, V ∗) and (K, π, W ∗V ∗) are two Stinespring minimum triplet. Thus W ∗V ∗ = U ∗V ∗ and U π(x)U ∗ = π(x) for all x ∈ A where U is an unitary operator on K. So we get (U W ∗ − I)V ∗ = 0. Since both U, W commutes with π(A), we get (U W ∗ − I)π(x)V ∗f = 0 for all f ∈ H and x ∈ A. By cyclic property {π(x)V ∗f : x ∈ A, f ∈ H} is total in K, thus U W ∗ − I = 0 i.e. W is unitary. So Λ = I − W ∗W = 0. 2 (τW + τW ′ ) where τW = V W π(x)W ∗V ∗ and τW ′ = V W ′π(x)W 4. Operator systems of unital normal complete positive maps: Proposition 4.1. Let τ : M → B(H) be a unital normal completely positive map on a von-Neumann algebra M acting on a Hilbert space H. Then there exists a family of linearly independent operators {vα : α ∈ Iτ } over the coefficients in M′, α for all x ∈ M. commutant of M such that Pα∈Iτ If {v′ β : β ∈ Jτ } is another such a family of operators representing τ then there exists a family of elements {wα β = Pα vαwα β and (wα β ) is a unitary operator on H ⊗ K0. Furthermore, τ is an extremal element α = I and τ (x) = Pα vαxv∗ β ∈ M′ : α ∈ Iτ , β ∈ Jτ } such that v′ vαv∗ 12 ANILESH MOHARI in CPσ if and only if there exists no non-trivial solution with elements λα satisfying Pα,β∈Iτ β vβ = 0. vαλα β ∈ M′ Furthermore, if τ : M → M and admits an inner representation i.e. with vα ∈ M then τ is extremal in the set of completely positive unital map on M if and only if there is no non-trivial solution (λα β ) in the center of M. Proof. τ being normal and unital, π is a normal unital representation of M into B(K) and thus we may write K ≡ H ⊗ K0 for some Hilbert space K0 and π(x) ≡ x ⊗ IK0 . Thus we conclude the first part by fixing an orthonormal basis eα αg >=< f ⊗ eα, V ∗g >. To show linear independence of for K0 and defining < f, v∗ α : α ∈ Iτ }, let Pα cαv∗ α = 0 for some P cα2 < ∞. Then we have {v∗ < f ⊗ Xα =< x∗f ⊗ Xα = Xα =< x∗f,Xα cαeα, (x ⊗ IK0)V ∗g > cαeα, V ∗g > cα < x∗f, v∗ cαv∗ αg > αg > = 0 i.e. f ⊗ Pα eα is orthogonal to the total set of vectors {(x ⊗ IK0 )V ∗g : g ∈ H, x ∈ M} in K and so f ⊗ Pα cαeα = 0. Thus Pα cαeα = 0 and (eα) being linearly independent we get cα = 0. A little modification of the proof works to show for the minimal representation, α = 0 αcα ∈ M′ if and only if cα = 0. It is enough if we prove for cα = 0 for all α ∈ Iτ except finitely many. To show linear independence of α = 0 for some cα ∈ M′ with finitely many non {v∗ zero elements. Then we have the family vα is linearly independent over coefficients in M′ i.e. Pα cαv∗ with cα ∈ M′ and Pα c∗ α : α ∈ Iτ } over M′, let Pα cαv∗ < Xα = Xα αf ⊗ eα, (x ⊗ IK0)V ∗g > c∗ < x∗c∗ αf ⊗ eα, V ∗g > αx∗f, v∗ αg > < c∗ = Xα =< x∗f,Xα cαv∗ αg > = 0 i.e. P c∗ M} in K and so Pα c∗ αf ⊗ eα is orthogonal to the total set of vectors {(x ⊗ IK0 )V ∗g : g ∈ H, x ∈ αf ⊗ eα, g ⊗ fβ >= 0 for all β ∈ Iτ and g ∈ H. The family (eα) being an orthonormal basis for K0, we get < c∗ αf ⊗ eα = 0. Thus < Pα c∗ αf, g >= 0 for all f, g ∈ H. Thus we get cα = 0. Choosing another such a family of elements (v′ β : β ∈ Jτ ) means that we are choosing another orthonormal basis (e′β) for K0 and thus there exists an unitary operator W on H ⊗ K0 by (7) such that W ∗V ∗ = V ′∗. Since by our construction EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 13 W x ⊗ IK0 W ∗ = x ⊗ IK0 for all x ∈ M, we conclude that W ∈ M′ ⊗ B(H0). We set matrix elements W α β ∈ M′ defined by < f, W α β g >=< f ⊗ eα, W e′ β ⊗ g > for all f, g ∈ H. We give now the proof of the last two statements. First statement is a simple corollary of Theorem 2.8 and normality of τ . For the last statement, we compute with any unitary element u ∈ M′ by uniqueness of the Radon-Nykodym repre- sentation of η for which η ≤ cτ on M+ with a scaler c > 0, η(x) = uη(x)u∗ = j u∗ = Pk vkxutk j v∗ uPk vkxtk j ) with entries j u∗ = tk j ) are elements in MT M′. This completes the proof. j and thus matrix entries in t = (tk j and thus by uniqueness of t = (tk in M′, we get utk j u∗v∗ Given a unital C∗-algebra A, a subspace S of A is called operator system if S is closed under involution ∗ and identity of A denoted by I ∈ S. Let S1, S2 be two operator systems in C∗-algebras A1 and A2 respectively. A linear one to one and onto map I : S1 → S2 is called order isomorphism if I and I −1 are both non-negative i.e. taking non-negative elements to non-negative elements. It is called a complete order isomorphism if I ⊗ In : Mn(S1) → Mn(S2) is also an order-isomorphism for each n ≥ 1. For a unital completely positive normal map τ : M → B(H) with minimal k for all x ∈ M, we set notations Mτ and Sτ for representation τ (x) = Pk vkxv∗ the operator space and system Mτ = {X λiv∗ Sτ = {X viλi j v∗ i : λi ∈ M′} j : λi j ∈ M′} The operator space Mτ over the algebra M′ is independent of the choice that we make for minimal representation of τ and thus the dimension dτ of Mτ is an invariance for τ . This integer dτ is called now onwards Arveson index of τ . The τ over the algebra M′, operator system Sτ is a two sided module of dimension d2 where left and right natural actions of M′ on Sτ are given by y ◦ X vixv∗ X vixv∗ j = X viyxv∗ j ◦ y = X vixyv∗ j j One simple question that arises here: does operator space Sτ depends on the choice that we make for (vk) to represent τ ? The following proposition says that Sτ is independent of the representation we choose and a little more. τ (x) = Pk vkxv∗ map. Then Sτ is determined uniquely modulo an unitary operator u ∈ A′ i.e. Proposition 4.2. Let τ : M → B(H) be a unital completely positive normal if k)∗ be two minimal representations of τ j )) ∈ Mn(M′) so that Λ∗V ∗ = V ′∗ then there exists an unitary operator Λ = ((λi and operator system Sτ is uniquely determined by τ by its minimal representation j ) with entries in M′ of representation i.e. there exists an unitary matrix λ = (λi order equal to numerical index so that k and τ (x) = Pk v′ kx(v′ (11) k = Xj v∗ j (v′ λk j)∗ 14 ANILESH MOHARI In particular, group of unitary elements u ∈ M′ acts on (v∗ i ) (12) u∗v∗ ku = Xj j (u)v∗ λk j , where u → Λ(u) = ((λj (13) k(u))) ∈ Udτ (M′) satisfies the following cocycle relation Λ(uv) = (v∗ ⊗ Idτ )Λ(u)(v ⊗ Idτ )Λ(v) Conversely, let τ (x) = Pk vkxv∗ element τ in CPσ and η = Pk wkxw∗ element η ∈ CPσ so that (w∗ then η = τ . k be a minimal representation of an extremal k be another minimal representation of an k) over M′ i.e. Mτ = Mη k) is in the linear span of (v∗ Proof. By the uniqueness part of Stinespring intertwining relation find an uni- tary operator Λ on H ⊗ K0 so that relation we have Λ(x ⊗ I)Λ∗ = x ⊗ I and ΛV ′∗ = V ∗. Thus we get in the basis v∗ k) unitary elements in Mdτ (M′) determined by < f ⊗ ei, Λg ⊗ ej >=< f, λi jg > where dτ is the cardinality of an orthonormal basis for K0. The relation (13) follows from (12) and linear independence of the family (v∗ k = Pj λk j where Λ = (λj j v′∗ k) over M′. Now we consider the relation ΛV ′∗V ′Λ∗ = V ∗V where Λ ∈ Udτ (M′). Thus the change of the basis for K will not change the operator systems Sτ and Sτ ′. For the last part we fix λi j ∈ M′ so that W ∗ = ΛV ∗ and claim that Λ is an isometry if τ is an extremal element. To that end we check I = W W ∗ = V Λ∗ΛV ∗ i.e. V (I − Λ∗Λ)V ∗ = 0. τ being extremal we get Λ∗Λ − I = 0 by Corollary 2.9. Since λi j ∈ M′, we get by a simple computation that for all x ∈ M, η(x) = X wkxw∗ k = W (x ⊗ I)W ∗ = V Λ(x ⊗ I)Λ∗V ∗ = V (x ⊗ I)V ∗ = Xk vkxv∗ k = τ (x) i.e. η = τ . Now we formulate a more deeper problem. Two unital UCP maps τ1 : M1 → B(H1) and τ2 : M2 → B(H2) are called cocycle conjugate if there exist automor- phisms α : M1 → M2 and β : M2 → M1 such that (14) τ2 ◦ α = β ◦ τ1 on M1. Two cocycle conjugate elements are called conjugate if (14) is valid with α = β−1. For two conjugate elements τ1 and τ2, the operator systems Sτ1 and Sτ2 are complete order isomorphic. For a proof we consider the lift of cocycle conjugate relation (14) to universal enveloping von-Neumann algebra M∗∗ 2 as M1 and M2 are isomorphic. In other words, we can assume with out loss of generality that M1 = M2 = M and M is in its standard form acting on H and thus for some unitary operators u, v on H, we have α(x) = uxu∗ and β(x) = vxv∗ for all x ∈ M. Thus operator spaces are cocycle conjugated by u, v i.e. uMτ1 v∗ = Mτ2. Since Sτ = {x∗y : x, y ∈ Mτ }, the operator systems Sτ1 and Sτ2 are complete order j u∗), where τ1(x) = j → Pi,j(uviv∗)vλi j v∗ isomorphic via the map Pi,j viλi Pj vj xv∗ 1 = M∗∗ j v∗(vv∗ j . However, the converse is false even in the simplest situation when M = M2(C) and a detail results are given in a recent paper [Mo4]. However, one additional EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 15 natural condition on involved operator system gives a positive result. This inverse problem is further addressed in a forth coming paper [Mo5]. 5. Tomita coupling and extremal marginal states: We now aim to describe extremal marginal states in a more general mathematical frame work then originally proposed in [Par] and subsequently followed in papers [PSa,Ru,Oh]. Let φ be a faithful normal state on a von-Neumann algebra M acting on a complex separable Hilbert space H with inner product < ., . > taken linear in the second variable and without loss of generality we assume M be in its standard form (M, P, J , ∆, Ω) [BR] associated with Ω where Ω is a cyclic and separating vector for M and ∆, J are modular and conjugate operators of Tomita associated 2 , which is the closer of of the densely defined with polar decomposition S = J ∆ anti-linear closable operator S0xΩ → x∗Ω and P = {xJ xJ Ω : x ∈ M} is the self- dual pointed positive cone in H. Here we recall that analytic elements Ma of the modular group σt(x) = ∆itx∆−it on M is dense in weak∗ topology in M and we have the following modular relation for any two x, y ∈ Ma given by 1 (15) φ(x∗y) = φ(σ i 2 (y)σ− i 2 (x∗)) We consider the algebraic tensor product M⊗M and complete it with C∗-cross norm of the von-Neumann algebra M⊗M. The set of states ψ on C∗ tensor product of von-Neumann algebras M◦M acting on H⊗H such that its restrictions on M◦I and I ◦ M are equal to φ. In short, we will call such a element in words a coupling state with marginal φ and denote the convex set by Cφ. Simplest example is the product state and simplest non-product state is given by a Tomita state ψ(x ◦ y) =< J xJ Ω, yΩ > by extending linearly and then to its norm closures M ◦ M. Important differ- ence that we note now that ψ may not have a normal extension to von-Neumann to M ⊗ M = (M ◦ M)′′. As an example we take completion of M ◦ M i.e. M = L∞[0, 1] and check that indicator function of the diagonal set in [0, 1] × [0, 1] can be expressed as limit of decreasing projections En ∈ M ◦ M with ψ(En) = 1 and Tn≥1 En = {[x, x] : 0 ≤ x ≤ 1}. In spite of this odd feature, we have a simple but beautiful observation. Lemma 5.1. A state ψ ∈ Cφ if and only if there exists a unital normal map τψ : M → M preserving φ so that (16) ψ(x ◦ y) =< J xJ Ω, τψ(y)Ω > for all x, y ∈ M. Further the map ψ → τψ is an one to one and onto affine map between two convex set Cφ and CPφ = {τ : M → M, completely positive unital map and φ ◦ τ = φ} Proof. We fix y ≥ 0 so that φ(y) = 1 and consider that state φy : x → ψ(x ⊗ y) and note that φy(x) ≤ yφ(x) for x ≥ 0 as 0 ≤ y ≤ yI. Since by our assumption φ is normal, φy is also normal. Thus by Dixmier lemma we conclude that φy(x) =< y′∗Ω, xΩ > for some non-negative element y′∗ ∈ M′. Thus we get ψ(x ◦ y) =< J xJ Ω, J y′∗J Ω > 16 ANILESH MOHARI We set τ (y) = J y′∗J ∈ M to arrive at (16). Since τ (y) is determined uniquely by the separating property of Ω for M, we may extend the map for an arbitrary element y using linearity by writing it as a linear combination of four non negative elements in M. That y → τ (y) is a normal map follows by invariance of the faithful normal state φ for τ and positivity of the map as follows: for an increasing net yα with least upper bound y, τ (yα) is also an increasing net with τ (y) as an upper bound and thus has a least upper bound say z. Then φ(z − τ (y)) = l.u.b.αφ(τ (yα) − τ (y)) by normality of φ and thus by invariance equal to l.u.b.αφ(yα − y) which is 0 by normality of the state. For complete positive property of the map τ , we first check that < Ω,Xi,j (z′ i)∗τ (y∗ j yi)z′ jΩ > < Ω, (z′ i)∗z′ jτ (yj y∗ i )Ω > < (z′ j)∗z′ iΩ, τ (yjy∗ i )Ω > < J z∗ j ziJ Ω, τ (yj y∗ i )Ω > = Xi,j = Xi,j = Xi,j j zi ◦ yjy∗ i ) ψ(z∗ = Xi,j = ψ(X ∗X) ≥ 0 i for all 1 ≤ i ≤ n, yi ∈ M, z′ where X = Pi zi ◦ y∗ iJ . Since Ω is cyclic for M′, we conclude that the operator ((τ (yj y∗ i ))) is a non-negative element in Mn(M). Thus τ is n−positive for each n ≥ 1. Rest of the statements are now obvious. i ∈ M′ and zi = J z′ Theorem 5.2. The affine map ψ → τψ defined in Lemma 3.1 takes extremal elements of the convex set Cφ to extremal elements of CPφ. Further (a) Cφ is a closed convex subset in the weak topology of M ◦ M; (b) CPφ is also compact once equipped with the point-wise topology inherited from σ-weak operator topology of M (i.e. we say a net τα ∈ CP converges to τ ∈ CP if the net τα(x) converges to τ (x) in σ−weak operator topology for each x ∈ M, where CP denotes the set of unital completely positive maps on M ). Proof. First part is obvious as the map is one to one and onto. Crucial obser- vation that we make here for a net τα in CPφ such that τα(x) → τ (x) in σ−weak operator topology for some τ ∈ CP . Then τ is also unital and φ preserving and τ is normal as φ is faithful. Thus τ ∈ Cφ. We consider a net of states ψα ∈ Cφ so that ψα → ψ in weak topology on M ◦M and since each ψα preserves marginals so is their limit. Thus ψ ∈ Cφ. Now we also check that (17) ψα(x ◦ y) = φ(J xJ Ω, τα(y)Ω >→ ψ(x ◦ y) for all x, y ∈ M, where the state ψ on M ◦ M defined by ψ(x ◦ y) =< J xJ Ω, τ (y)Ω > EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 17 determines a unique element τ ∈ CPφ by Lemma 3.1. These shows that τα(x) → τ (x) in weak operator topology. Since the family (τα) is uniformly bounded and Ω is cyclic and separating for M, the limit (17) says that Xk < fk, τα(x)gk >→ Xk < fk, τ (x)gk > as the net α → ∞ for Pk fk2 < ∞ and Pk gk2 < ∞ by dominated conver- gence theorem. Thus τα(x) → τ (x) in σ−weak operator topology. In other words, convergence holds in Bounded Weak topology of Arveson. Conversely for a given net τα in CPφ which converges to τ in point-wise σ−weak operator topology, τ ∈ CPφ by our remark at the beginning of the proof. We check that the associated elements ψα also converges to ψ in the weak topology of M ◦M first on a norm dense subspace of algebraic tensor product of M⊗M and then for any arbitrary elements by a standard argument as the net is uniformly bounded. The convex set of states on M ◦ M is compact in weak∗ topology and thus Cφ being a closed subset of the compact set in the weak∗ topology, Cφ is also compact. That CPφ is compact also follows from compactness of Cφ via the above correspondence which respect the topologies. One can as well use BW (Bounded Weak) topology [Ar,Pa, Chapter 7] of Arveson directly to give a direct proof that CPφ is a closed subset of the unit ball of bounded linear maps on M. Any limit points of a convergent net of φ-invariant unital completely positive normal maps τα will be also completely positive and φ-invariant. φ being faithful and normal, any positive φ invariant map is also normal. However such an argument will not be valid with BW topology for a more general situation, where φ is just a normal σ−finite weight [BR,Ta]. It is not clear what would be a possible truncation method along the classic work [Ke,Ko]. By our last theorem we conclude that extremal elements in CPφ exists and any element in CPφ admits Krein-Milman property. In the following text we aim to find a criteria for an extremal element τ ∈ CPφ. To that end, we recall KMS-adjoint completely positive map [OP]. Proposition 5.3. Let φ be a faithful normal state on a von-Neumann algebra M. Then τ is a positive normal map on M such that φτ ≤ φ if and only if exists a unique normal positive map τ on M satisfying the following duality relation (18) φ(τ (x)σ− i 2 (y)) = φ(σ i 2 (x)τ (y)) where x, y ∈ Ma, the ∗-algebra of analytic elements for modular group σ = (σt : t ∈ R) associated with φ and τ (I) ≤ I. Moreover, φτ ≤ φ if and only if τ (I) ≤ I. Furthermore, τ is completely positive if, and only if τ is completely positive. In such a case, the numerical indices of τ and τ are equal. Proof. For the first part of the statement we refer to chapter 8 of monograph [OP]. For the second part we re-investigate the proof of Stinespring representation with our special situation. With out loss of generality we assume that M is in the standard form associated with φ. i.e. φ(x) = (Ω, xΩ), where Ω is a cyclic and separating vector for M. We set a kernel on Ma ⊗ Ma defined by k(x ⊗ z, y ⊗ w) =< Ω, x∗τ (z∗w)yΩ > 18 ANILESH MOHARI That Hilbert space completion of the kernel is same as that of Stinespring follows by cyclic property of Ω. Similarly we set kernel k associated with τ by k(x ⊗ z, y ⊗ w) =< Ω, z∗τ (x∗y)wΩ > in reverse direction and check by KMS relation that < Ω, x∗τ (z∗w)yΩ >=< Ω, w∗ τ (y∗ x)zΩ > (x∗). Such a relation where for any analytic element x ∈ M we write x = σ− i is used already in [Mo1,Mo2,Mo3]. This clearly shows that Stinespring Hilbert spaces K and K associated with kernels k and k respectively are conjugated by an anti-unitary operator defined by U : x ⊗ y → y ⊗ x. Since anti-unitary operator U inter-twins the Stinespring representations (K, π, V ) and ( K, π, V ) it also inter- twins K0 and K0 and thus we get dimension of K0 and K0 are same. Thus index of τ and τ are equal. 2 In the following, we give a criteria for an element τ in CPφ to be extremal. If M is a matrix algebra and φ is the normalize trace, the criteria coincides with Landau- Streater criteria [LS] known in the literature for quite some time. Our proof follows quite a different method inspired by Proposition 4.1. Theorem 5.4. Let τ and τ be the elements in CPφ of equal numerical index admitting the following minimal representations and τ (x) = Xα vαxv∗ α τ (x) = Xα vαxv∗ α for all x ∈ M, where τ and τ are related by the duality relation (16) in Proposition 5.3. Then τ is an extremal element in CPφ if and only if there exists no non-trivial λ = (λk j ) with entries in M′ satisfying the relation (19) Xα,β vαλα β = 0, Xα,β β v∗ vαλα β v∗ β = 0, where λ → λ is an unital order-isomorphism map on Mnτ (M′)). Proof. Let η be an element in CP such that η ≤ kτ for some k ≥ 0. Then η0 = kτ − η is also a positive map and η0 ≤ kτ . The duality being an affine map, we get η + η0 = kτ , where η0 is a positive map by Proposition 5.3. Thus we also have η ≤ kτ . Then η is a normal map and there exists a unique non-negative element t = (tα β ) with entries in M′ such that (20) η(x) = Xα,β vαxtα β v∗ β for all x ∈ M. Conversely. a non-negative bounded operator T = (tα β ) with entries in M determines a normal map η by (20) for some k ≥ 0. Thus there is a one to one relation between completely positive map satisfying η ≤ kτ and bounded non-negative elements T = (tα β ) with entries in M′ determined by (20). EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 19 Similarly, for a give η ≤ kτ , there exists a unique T = (tα β ) with entries in M′ such that (21) for all x ∈ M. η(x) = Xα,β vαxtα β v∗ β This shows that T → T is a well defined affine map on the non negative elements of Mdτ (M′) determined by (20) and (21) for fixed minimal representations of τ and τ . +, S2 + − T 2 + ≥ 0. Then T 1 We extends the map T → T to self-adjoint elements T = T 1 +−T 2 +. To show that the map is well defined, let T = S1 + ≥ 0 to T = T 1 + be another +. Since T → T is an affine expression with S1 map onMnτ (M′)+, we get T 1 +. Rearranging the terms, we conclude the map T → T is well defined on self-adjoint elements of Mnτ (M′). That it is an injective map, follows readily as T = 0 implies T+ = T−. The map being injective on non-negative elements, we have T+ = T−. Thus T +2 = T 2 − = T+T− = 0 i.e. T = 0. The map is clearly onto on self adjoint elements. We extend now by linearity to all element from Mnτ (M′) to itself. + with T 1 + − S2 + + S2 + + T 2 + + S2 + = S1 + = S1 + + T 2 +, T 2 Since T → T is also an injective map on non-negative matrices, there exists an injective and onto extension of this map from self adjoint elements to self adjoint elements. Thus we conclude that T → T extends to an order isomorphic map uniquely on Mnτ (M′). Thus λα β = tα β − δα β I is a solution to (19) and it is non-trivial if and only if τ is not an extremal element in CPφ. This completes the proof. By a theorem of Kadison [Ka1], the map T → T is a disjoint sum of a morphism and anti-morphism determined by a projection e in the centre of M i.e. there exists a projection e in the centre of M such that T → T e ⊗ Inτ is a morphism and T → T (I − e) ⊗ Inτ is an anti-morphism. In case M is a factor, then T → T is either an isomorphism or an anti-isomorphism. If τ admits an inner minimal representation (22) τ (x) = Xα∈C vαxv∗ α, x ∈ M i.e. with elements vα ∈ M then τ also admits an inner minimal representation given by (23) τ (y) = Xα∈C vαyv∗ α, y ∈ M where vα ∈ M is defined as the bounded operator extending the densely defined 2 f for all f ∈ MΩ. For the last part, we refer section 7 and operator f → ∆ appendix given in [BJKW]. α∆− 1 2 v∗ 1 In particular, if η ≤ kτ for some k ≥ 0 then the representations of η and η β ) are with entries in the given in (20) and (22) are inner with matrices (tα β ) and (tα 20 ANILESH MOHARI centre of M. Thus the map T → T is an order isomorphism on Mdτ (Z), where Z = MT M′. A simple computation using modular relation (15) now also leads to the minimal representation η(x) = X vαxtα β v∗ β since modular group acts trivially on the center Z of M. By the uniqueness part of representation (21), we conclude that tα β = tβ α We sum up now our results in the following corollary. Corollary 5.5. Let τ be an element in CPφ with inner representation given by (22). Then its dual element τ also admits a inner representation given by (23). Furthermore, τ is an extremal element in CPφ if, and only if there exists no non- trivial λ = (λk j ) with entries in the centre of M satisfying the relation (24) Xα,β vαλα β = 0, Xα,β β v∗ vαλβ αv∗ β = 0 Krein-Milman theorem says that τ = R ηdµ(η) for some probability measure µ on the closer of extreme points CP e φ of CPφ in BW topology [Pa, Chapter 7], where µ may not be uniquely determined. A valid question that rises here when can we guarantee µ to have support on CP e φ only? The topology on the states of M ◦ M is not metrizable unless M is separable as a C∗ algebra. Such a restriction makes our choice for M rather limited since M is a von-Neumann algebra. When M = l∞(N), some clever truncated methods are used [Ke, Ko] to show the support of µ is confined to CP e φ0 (l∞(N)), where φ0 is the counting measure on N. 6. Pure marginal states: In the following, we answer a question raised in recent papers [Par,PSa,Oh]. Theorem 6.1. Let M = Mn(C) and φ be a faithful normal state on M. An extremal element ψ in Cφ is a pure state if and only if τψ is an automorphism on M. Proof. We may follow proof of Theorem 1.1 in [PSa] with obvious modification to prove: ψ ∈ Cφ is a pure state of M ⊗ M if and only if there are orthonormal bases (fi : 1 ≤ i ≤ n) and (gi : 1 ≤ i ≤ n) of Cn such that ψ(X) =< ζψ, Xζψ > i fi ⊗ gi, φ(x) = tr(ρx), x ∈ M and for all X ∈ M ⊗ M, where ζψ = Pi λ ρ = P1≤i≤n λifi >< fi = P1≤i≤n λigi >< gi. For two such pure states ψ and ψ′, we find unitary operators u, v on Cn which 1 2 takes bases (fi) → (f ′ i ) and (gi) → (g′ i) respectively. We claim that For a proof we use (14) and u ⊗ vζψ = ζψ′ to compute τψ′ (x) = uτψ(vxv∗)u∗ < J xJ Ω, τψ′(y)Ω > = ψ′(x ⊗ y) = ψ(u ⊗ v(x ⊗ y)u∗ ⊗ v∗) EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 21 = ψ(uxu∗ ⊗ vyv∗) =< J uxu∗J Ω, τψ(vyv∗)Ω > =< J xJ u∗Ω, u∗τψ(vyv∗)Ω > (since φ(uxu∗) = φ(x), x ∈ M, u commutes with Tomita's modular operator and conjugate operator and uΩ = Ω ). =< J xJ Ω, u∗τψ(vyv∗)uΩ > Thus we have τψ′ (y) = u∗τψ(vyv∗)u for all y ∈ M. Since ψ(x⊗y) =< J xJ Ω, yΩ > is a pure state on M ⊗ M, we get the required result. Theorem 6.2. Let φ be the normalized trace on M = Mn(C) and ψ be a state on M ⊗ M with marginal states φ i.e. ψ ∈ Cφ. If ψ is an extremal element in Cφ then ψ is a factor state of M ⊗ M. Proof. A state ψ, ψ(X) = tr(Xhψ), X ∈ M ⊗ M for some non-negative d Id d Id, where E1, E2 are conditional expectation with respect to the density operator hψ ∈ M ⊗ M, is an element in Cφ if and only if E1(hψ) = 1 and E2(hψ) = 1 normalized trace from M ⊗ M onto M ⊗ In and In ⊗ M respectively. Let ψ be an extremal element in Cφ and (Hψ, πψ, ζψ) be its GNS representation. We claim that ψ is a factor state i.e. the centre of πψ(M ⊗ M)′′ is trivial. Suppose it is not. Let E be a non trivial projection in the centre of πψ(M ⊗ M)′′ and ψ = λφ1 + (1 − λ)φ0, where ψ1, ψ0 are states on M ⊗ M defined by λψ1(X) = ψE(X) =< ζψ, XEζψ >= tr(hψXE) and (1 − λ)ψ0(X) = ψI−E(X) =< ζψ, X(I − E)ζψ >= tr(hψX(I − E)) for λ = ψ(E) ∈ (0, 1) and h0, h1 ∈ M ⊗ M since E is an element in M ⊗ M as well. Thus we have h = λh1 + (1 − λ)h0, where ψk(X) = tr(hkX) for k = 0, 1. So 1 d Id = E1(hψ) = λE1(h1) + (1 − λ)E1(h0), where we have used h0, h1 ∈ M ⊗ M. The operator 1 d Id being an extremal element in the set of non-negative definite matrices of trace 1, we get E1(h1) = E1(h0) = n Id. Similarly, we also have E2(h1) = E2(h0) = 1 1 d Id. This shows in particular, ψ0, ψ1 ∈ Cφ and ψ = λψ1 +(1−λ)ψ0 for some λ ∈ (0, 1). This brings a contradiction to extremal property of ψ in our hypothesis at the beginning. The converse statement of Theorem 6.2 is obviously false. For a counter example, we consider the product state ψ(x ⊗ y) = φ(x)φ(y) on M ⊗ M, where φ is a faithful state on M. In such a case, τψ = φ and φ is a factor state but need not be an extremal element unless M is itself C i.e. n = 1 in Theorem 6.2. However, for a state ψ in Cφ and we may consider the extremal decomposition of τψ in CPφ given by τψ = ZCP e φ τedµτ (e) 22 ANILESH MOHARI for some regular Borel probability measure µτ on the set of extremal elements. By what we have proved above, each τe gives a unique factor state ψe of M ⊗ M and so that (25) ψ = ZCP e φ ψedµτ (e) If ψ is a factor state but not an extremal element, then support of µτ is not atomic and the decomposition (25) is not central i.e. associated GNS representations is not a central decomposition [BRI]. Theorem 6.1 and Theorem 6.2 give raises the following interesting question. Let ψ and ψ′ be two states of M ⊗ M in Cφ with density matrices hψ and hψ′ respectively. Then τψ and τψ′ are cocycle conjugate if, and only if hψ and hψ′ are unitary equivalent by an unitary matrix of the form u ⊗ v. So a classification of extremal elements is equivalent to classify all density matrices h in M ⊗ M for which the state ψh(x) = tr(xh) is atleast a factor state and E1(h) = 1 d Id and E2(h) = 1 d Id. One necessary condition is equality of ranks of hψ and hψ′ . Theorem 6.1 suggests rank of hψ is possibly a complete invariance for an extremal element τψ in CPφ. We defer a possible answer of this finer question to [Mo5] which takes some hint from recent results proved in [Mo4]. Any faithful state φ on M ⊆ B(H) admits a representation φ(x) = Xk λk < fk, xfk > for all x ∈ M with some λk > 0 for all k ≥ 1, where H is assumed to be a separable Hilbert space. If M = B(H), an obvious modification of the argument used in the proof of Theorem 6.1, also proves that any extremal element ψ in Cφ is pure if, and only if τψ is an automorphism on M. Same questions for an arbitrary von- Neumann algebra M is rather delicate since a state ψ ∈ Cφ may not have a normal extension to a state of M ⊗ M. REFERENCES • [Ar] Arveson, W.: Sub-algebras of C∗-algebras, Acta Math. 123, 141-224, 1969. • [BR] Bratteli, Ola., Robinson, D.W. : Operator algebras and quantum statistical mechanics I,II, Springer 1981. • [BJKW] Bratteli, Ola,; Jorgensen, Palle E.T.; Kishimoto, Akitaka and Werner Reinhard F.: Pure states on Od, J.Operator Theory 43 (2000), no-1, 97-143. • [Ch] Choi, M.D.: Completely positive linear maps on complex matrices, Linear Algebra and App (10) 1975 285-290. • [HMP] Hopenwasser, Alan; Moore, Robert L.; Paulsen, V. I. : C∗-extreme points, Trans. Amer. Math. Soc. 266 (1981), no. 1, 291-307. • [Ka1] Kadison, Richard V.: Isometries of operator algebras, Ann. Math. 54(2)(1951) 325-338. • [Ka2] Kadison, Richard V.: A generalized Schwarz inequality and algebraic invariants for operator algebras, Ann. of Math. (2) 56, 494-503 (1952). • [Ke] Kendall, D.G. : On infinite doubly stochastic matrices and Birkhoff's problem, III. London Math. Soc. J., 35 (1960):81-84. • [Ko] Konig, D. : The theory of finite and infinite graphs, Taubner 1936, Birkhaser, Boston, 1990, p-327. EXTREMAL UNITAL COMPLETELY POSITIVE MAPS AND ITS SYMMETRIES 23 • [LS] Landau, L.J., Streater, R.F.: On Birkhoff theorem for doubly stochas- tic completely positive maps of matrix algebras, Linear Algebra and its Applications, Vol 193, 1993, 107-127 • [MW] Mendl, Christian B., Wolf, Michael M.: Unital quantum channel's convex structure and revivals of Birkhoff's theorem. Comm. Math. Phys. 289 (2009), no. 3, 1057 -- 1086. • [Mo1] Mohari, A.: Pure inductive limit state and Kolmogorov property. II Journal of Operator Theory. vol 72, issue 2, 387-404. • [Mo2] Mohari, A.: Translation invariant pure state on ⊗ZMd(C) and Haag duality, Complex Anal. Oper. Theory 8 (2014), no. 3, 745-789. • [Mo3] Mohari, A.: Translation invariant pure state on B = ⊗ZMd(C) and its split property, J. Math. Phys. 56, 061701 (2015). • [Mo4] Mohari, A.: Hann-Banach-Arveson extension theorem and Kadison isomorphism, arXiv:1304.6849 (2015). • [Mo5] Mohari, A.: G. Birkhoff problem in irreversible quantum dynamics, in preparation (2015). • [Oh] Ohno, H.: Maximal rank of extremal marginal tracial states. J. Math. Phys. 51 (2010), no. 9, 092101, 9 pp. • [OP] Ohya, M., Petz, D.: Quantum entropy and its use, Text and mono- graph in physics, Springer-Verlag 1995. • [Pa] Paulsen, V.: Completely bounded maps and operator algebras, Cam- bridge Studies in Advance Mathematics 78, Cambridge University Press. 2002 • [Pas] Paschke, W. L. : Inner Product Modules Over B∗-Algebras, Transac- tions of the American Mathematical Society, Vol. 182 (1973), pp 443-468 • [Ph] Phelps, Robert R.: Lectures on Choquet's Theorem, Lecture notes in Mathematics 1757, Springer 2001. • [Par] Parthasarathy, K.R.: Extremal quantum states in coupled states in coupled systems, Ann.Inst. H. Poincar´e, 41, 257-268 (2005). • [PSa] Price, G.L., Sakai, S.: Extremal marginal tracial states in couple systems, Operators and matrices, 1, 153-163 (2007). • [Ra] Raginsky, M. Radon-Nikodym derivatives of quantum operations. J. Math. Phys. 44 (2003), no. 11, 50035020. • [Ru] Rudolph, O.: On extremal quantum states of composite systems with fixed marginals, J. Math. Phys. 45, 4035-4041 (2004). • [St] Stinespring, W. F.: Positive functions on C∗ algebras, Proc. Amer. Math. Soc. 6 (1955) 211-216. • [Stø] Størmer, E.: Positive linear maps of operator algebras. Acta Math. 110, 233-278 (1963). • [Ta] Takesaki, M. : Theory of Operator algebras I, Springer, 2001. • [To1] Tomiyama, J.: On the projection of norm one in W∗-algebras. Proc. Japan Acad. 33 1957 608-612. • [To2] Tomiyama, J.: On the projection of norm one in W∗-algebras. II. Thoku . Math J. (2) 10 1958 204-209. • [To3] Tomiyama, J.: On the projection of norm one in W∗-algebras. III. Thoku Math. J. (2) 11 1959 125-129. ( • [Ts] Tsui, S. K.: Completely positive module maps and completely positive extreme maps. Proc. Amer. Math. Soc. 124 (1996), no. 2, 437-445. The Institute of Mathematical Sciences, CIT Campus, Taramani, Chennai-600113 E-mail : [email protected]
1801.07802	1	1801	2018-01-23T22:56:03	Phase transitions on C*-algebras arising from number fields and the generalized Furstenberg conjecture	[ "math.OA" ]	In recent work, Cuntz, Deninger and Laca have studied the Toeplitz type C-algebra associated to the affine monoid of algebraic integers in a number field, under a time evolution determined by the absolute norm. The KMS equilibrium states of their system are parametrized by traces on the C-algebras of the semidirect products $J \rtimes O^$ resulting from the multiplicative action of the units $O^$ on integral ideals $J$ representing each ideal class. At each fixed inverse temperature $\beta > 2$, the extremal equilibrium states correspond to extremal traces of $C^(J\rtimes O^)$. Here we undertake the study of these traces using the transposed action of $O^$ on the duals $\hat J$ of the ideals and the recent characterization of traces on transformation group C-algebras due to Neshveyev. We show that the extremal traces of $C^(J\rtimes O^)$ are parametrized by pairs consisting of an ergodic invariant measure for the action of $O^$ on $\hat{J}$ together with a character of the isotropy subgroup associated to the support of this measure. For every ideal the dual group $\hat {J}$ is a d-torus on which $O^$ acts by linear toral automorphisms. Hence, the problem of classifying all extremal traces is a generalized version of Furstenberg's celebrated $\times 2$ $\times 3$ conjecture. We classify the results for various number fields in terms of ideal class group, degree, and unit rank, and we point along the way the trivial, the intractable, and the conjecturally classifiable cases. At the topological level, it is possible to characterize the number fields for which infinite $O^$-invariant sets are dense in $\hat{J} $, thanks to a theorem of Berend; as an application we give a description of the primitive ideal space of $C^(J\rtimes O^*)$ for those number fields.	math.OA	math	PHASE TRANSITIONS ON C-ALGEBRAS ARISING FROM NUMBER FIELDS AND THE GENERALIZED FURSTENBERG CONJECTURE MARCELO LACA AND JACQUELINE M. WARREN ABSTRACT. In recent work, Cuntz, Deninger and Laca have studied the Toeplitz type C- algebra associated to the affine monoid of algebraic integers in a number field, under a time evolution determined by the absolute norm. The KMS equilibrium states of their system are parametrized by traces on the C-algebras of the semidirect products Jγ ¸ O K resulting from the multiplicative action of the units O K on integral ideals Jγ representing each ideal class γ P CℓK. At each fixed inverse temperature β ą 2, the extremal equilibrium states corre- spond to extremal traces of C(Jγ ¸ O K ). Here we undertake the study of these traces using K on the duals ^Jγ of the ideals and the recent characterization of the transposed action of O traces on transformation group C-algebras due to Neshveyev. We show that the extremal traces of C(Jγ ¸ O K ) are parametrized by pairs consisting of an ergodic invariant measure K on ^Jγ together with a character of the isotropy subgroup associated to for the action of O the support of this measure. For every class γ, the dual group ^Jγ is a d-torus on which O K acts by linear toral automorphisms. Hence, the problem of classifying all extremal traces is a generalized version of Furstenberg's celebrated 2 3 conjecture. We classify the re- sults for various number fields in terms of ideal class group, degree, and unit rank, and we point along the way the trivial, the intractable, and the conjecturally classifiable cases. At the topological level, it is possible to characterize the number fields for which infinite O K - invariant sets are dense in ^Jγ, thanks to a theorem of Berend; as an application we give a description of the primitive ideal space of C(Jγ ¸ O K ) for those number fields. 1. INTRODUCTION Let K be an algebraic number field and let OK denote its ring of integers. The associated K := OK \ {0} of nonzero integers acts by injective endomorphisms K , the affine multiplicative monoid O on the additive group of OK and gives rise to the semi-direct product OK ¸ O monoid (or 'b + ax monoid') of algebraic integers in K. Let {ξ(x,w) : (x, w) P OK ¸ O K ). The left regular representation L of OK ¸ O K } be the standard orthonormal basis of the Hilbert space K by isometries on ℓ2(OK ¸ O ℓ2(OK ¸ O K ) is determined by L(b,a)ξ(x,w) = ξ(b+ax,aw). In [2], Cuntz, Deninger and Laca studied the Toeplitz-like C-algebra T[OK] := C(L(b,a) : (b, a) P OK ¸ O K ) generated by this repre- sentation and analyzed the equilibrium states of the natural time evolution σ on T[OK] determined by the absolute norm Na := O K {(a) via σt(L(b,a)) = Nit a L(b,a) a P O K , t P R. One of the main results of [2] is a characterization of the simplex of KMS equilibrium states of this dynamical system at each inverse temperature β P (0,∞]. Here we will be interested in the low-temperature range of that classification. To describe the result briefly, K be the group of units, that is, the elements of O let O K whose inverses are also integers, and recall that by a celebrated theorem of Dirichlet, O K -- WK Zr+s−1, where WK (the Date: January 23, 2018. 1 2 MARCELO LACA AND JACQUELINE M. WARREN C(Jγ ¸ O ÀγPCℓK group of roots of unity in O K ) is finite, r is the number of real embeddings of K, and s is equal to half the number of complex embeddings of K. Let CℓK be the ideal class group of K, which, by definition, is the quotient of the group of all fractional ideals in K modulo the principal ones, and is a finite abelian group. For each ideal class γ P CℓK let Jγ P γ be an integral ideal representing γ. By [2, Theorem 7.3], for each β ą 2 the KMSβ states of C(OK ¸ O K ) are parametrized by the tracial states of the direct sum of group C-algebras K ), where the units act by multiplication on each ideal viewed as an additive group. It is intriguing that exactly the same direct sum of group C-algebras also plays a role in the computation of the K-groups of the semigroup C-algebras of algebraic integers in the work of Cuntz, Echterhoff and Li, see e.g. [3, Theorem 8.2.1]. Considering as well that the group of units and the ideals representing different ideal classes are a measure of the failure of unique factorization into primes in OK, we feel it is of interest to investigate the tracial states of the C-algebras C(Jγ ¸ O K ) that arise as a natural parametrization of KMS equilibrium states of C(OK ¸ O K ). This work is organized as follows. In Section 2 we review the phase transition from [2] and apply a theorem of Neshveyev's to show in Theorem 2.2 that the extremal KMS states arise from ergodic invariant probability measures and characters of their isotropy subgroups for the actions O K ý ^Jγ of units on the duals of integral ideals. We begin Section 3 by showing that for imaginary quadratic fields, the orbit space of the action of units is a compact Hausdorff space that parametrizes the ergodic invariant probability measures. All other number fields have infinite groups of units leading to 'bad quotients' for which noncommutative geometry provides convenient tools of anal- ysis. Units act by toral automorphisms and so the classification of equilibrium states is intrinsically related to the higher-dimensional, higher-rank version of the question, first asked by H. Furstenberg, of whether Lebesgue measure is the only nonatomic ergodic invariant measure for the pair of transformations 2 and 3 on R{Z. Once in this frame- work, it is evident from work of Sigmund [22] and of Marcus [15] on partially hyperbolic toral automorphisms and from the properties of the Poulsen simplex [13], that for fields whose unit rank is 1, which include real quadratic fields, there is an abundance of ergodic measures, Proposition 3.5, and hence of extremal equilibrium states, see also [11]. We also show in this section that there is solidarity among integral ideals with respect to the ergod- icity properties of the actions of units, Proposition 3.6. In Section 4, we look at the topological version of the problem and we identify the num- ber fields for which [1, Theorem 2.1] can be used to give a complete description of the invariant closed sets. In Theorem 4.11 we summarize the consequences, for extremal equi- librium at low temperature, of the current knowledge on the generalized Furstenberg con- jecture. For fields of unit rank at least 2 that are not complex multiplication fields, i.e. that have no proper subfields of the same unit rank, we show that if there is an extremal KMS state that does not arise from a finite orbit or from Lebesgue measure, then it must arise from a zero-entropy, nonatomic ergodic invariant measure; it is not known whether such a measure exists. For complex multiplication fields of unit rank at least 2, on the other hand, it is known that there are other measures, arising from invariant subtori. As a byproduct, we also provide in Proposition 4.9 a proof of an interesting fact stated in [25], namely the units acting on algebraic integers are generic among toral automorphism groups that have Berend's ID property. We conclude our analysis in Section 5 by computing the topology of the quasi-orbit space of the action O K ý pOK for number fields satisfying Berend's conditions. As an PHASE TRANSITIONS OF C-ALGEBRAS AND FURSTENBERG CONJECTURE 3 application we also obtain an explicit description of the primitive ideal space of the C- algebra C(OK ¸ O K ), Theorem 5.2. For the most part, sections 3 and 4 do not depend on operator algebra considerations other than for the motivation and the application, which are discussed in sections 2 and 5. Acknowledgments: This research started as an Undergraduate Student Research Award project and the authors acknowledge the support from the Natural Sciences and Engineer- ing Research Council of Canada. We would like to thank Martha Ł acka for pointing us to [13], and we also especially thank Anthony Quas for bringing Z. Wang's work [25] to our attention, and for many helpful comments, especially those leading to Lemmas 3.7 and 3.8. 2. FROM KMS STATES TO INVARIANT MEASURES AND ISOTROPY Our approach to describing the tracial states of the C-algebras ÀγPCℓK K ) is shaped by the following three observations. First, the tracial states of a group C-algebra form a Choquet simplex [23], so it suffices to focus our attention on the extremal traces. Second, there is a canonical isomorphism C(J¸O K , which we may combine with the Gelfand transform for C(J), thus obtaining an isomorphism of C(J ¸ O K ) to the transformation group C-algebra C(^J) ¸ O K on the continuous complex-valued functions on the compact dual group ^J. Specifically, the action of O K , associated to the transposed action of O K ) -- C(J)¸O K on ^J is determined by C(Jγ ¸ O u P O K , χ P ^J, j P J, (u χ)(j) := χ(uj), (2.1) or by xj, u χy = xuj, χy, if we use x , y to denote the duality pairing of J and ^J. Third, this puts the problem of describing the tracial states squarely in the context of Neshveyev's characterization of traces on crossed products, so our task is to identify and describe the relevant ingredients of this characterization. In brief terms, when [16, Corollary 2.4] is interpreted in the present situation, it says that that for each integral ideal J, the extremal traces on C(^J) ¸ O K , χ K -invariant measure on ^J such that the set of points is a character of H, and µ is an ergodic O in ^J whose isotropy subgroups for the action of O K are parametrized by triples (H, χ, µ) in which H is a subgroup of O K are equal to H has full µ measure. Recall that, by definition, an O K if µ(A) P {0, 1} for every O K -invariant probability measure µ onpJ is ergodic invariant K -invariant Borel set A Ă ^J. Our first K on ^J automatically has µ-almost everywhere constant for the action of O simplification is that the action of O isotropy with respect to each ergodic invariant probability measure µ. Lemma 2.1. Let K be an algebraic number field with ring of integers OK and group of units O let J be a nonzero ideal in OK. Suppose µ is an ergodic O there exists a unique subgroup Hµ of O is equal to Hµ for µ-a.a. characters χ P ^J. Proof. For each subgroup H ď O K )χ = H} be the set of characters of J with isotropy equal to H. Since the isotropy is constant on orbits, each MH is O K -invariant, and clearly the MH are mutually disjoint. By Dirichlet's unit theorem O K -- WK Zr+s−1 with WK finite, and r and 2s the number of real and complex embeddings of K, respectively. Thus every subgroup of O K is generated by at most WK + (r + s − 1) generators, and hence K } is a countable partition of ^J O K has only countably many subgroups. Thus {MH : H ď O into subsets of constant isotropy. K and K -invariant probability measure on ^J. Then K : uχ = χ} K such that the isotropy group (O K , let MH := {χ P ^J (O K )χ := {u P O We claim that each MH is a Borel measurable set in ^J. To see this, observe: 4 MARCELO LACA AND JACQUELINE M. WARREN MH ={χ P ^J : u χ = χ for all u P H and u χ ‰ χ for all u P O K \ H} =(cid:16) čuPH {χ P ^J χ−1(u χ) = 1}(cid:17)č(cid:16) čuPO K \H {χ P ^J χ−1(u χ) ‰ 1}(cid:17) because u χ = χ iff χ−1(u χ) = 1. Since the map χ Þ→ χ−1(u χ) is continuous on ^J, the sets in the first intersection are closed and those in the second one are open. By above, the intersection is countable, so MH is Borel-measurable, as desired. For every Borel measure µ on ^J, we have ÿHďO K µ(MH) = µ(cid:16) ďHďO K MH(cid:17) = 1, so at least one MH has positive measure. If µ is ergodic O unique Hµ ď O µ-a.a points χ P ^J. K -invariant, then there exists a K such that µ(MHµ ) = 1 and thus Hµ is the (constant) isotropy group of (cid:3) Since each ergodic invariant measure determines an isotropy subgroup, the characteri- zation of extremal traces from [16, Corollary 2.4] simplifies as follows. Theorem 2.2. Let K be an algebraic number field with ring of integers OK and group of units O K and let J be a nonzero ideal in OK. Denote the standard generating unitaries of C(J ¸ O K ) by δj for j P J and νu for u P O K ) there exists a unique probability measure µτ on ^J such that K . Then for each extremal trace τ on C(J ¸ O (2.2) xj, xydµτ(x) = τ(δj) for j P J. Z^J The probability measure µτ is ergodic O K -invariant, and if we denote by Hµτ its associated isotropy subgroup from Lemma 2.1, then the function χτ defined by χτ(h) := τ(νh) for h P Hµτ is a character on Hµτ. onto the set of pairs (µ, χ) consisting of an ergodic O Furthermore, the map τ Þ→ (µτ, χτ) is a bijection of the set of extremal traces of C(J ¸ O character χ P pHµ. The inverse map (µ, χ) Þ→ τ(µ,χ) is determined by K ) K -invariant probability measure µ on ^J and a xj, xydµ(x) if u P Hµ (2.3) 0 otherwise, χ(u)Z^J τ(µ,χ)(δjνu) =  for j P J and u P O K . Proof. Recall that equation (2.1) gives the continuous action of O the compact abelian group ^J obtained on transposing the multiplicative action of O J. There is a corresponding action α of O continuous functions on ^J; it is given by αu(f)(χ) = f(u−1 χ). K by automorphisms of K on K by automorphisms of the C-algebra C(^J) of K as follows. For a given extremal tracial state τ of C(J ¸ O The characterization of traces [16, Corollary 2.4] then applies to the crossed product C(^J) ¸α O K ) there is a probabil- ity measure µτ on ^J that arises, via the Riesz representation theorem, from the restriction of τ to C(J) -- C(^J) and is characterized by its Fourier coefficients in equation (2.2). By Lemma 2.1, there is a subset of ^J of full µτ measure on which the isotropy subgroup is PHASE TRANSITIONS OF C-ALGEBRAS AND FURSTENBERG CONJECTURE 5 K ) inside C(^J) ¸α O automatically constant, and is denoted by Hµτ. The unitary elements νu generate a copy of C(O K and the restriction of τ to these generators determines a charac- ter χτ of Hµτ given by χτ(u) := τ(νu). See the proof of [16, Corollary 2.4] for more details. By Lemma 2.1, the condition of almost constant isotropy is automatically satisfied for ev- ery ergodic invariant measure on ^J, hence every ergodic invariant measure arises as µτ for some extremal trace τ. The parameter space for extremal tracial states is thus the set of all K -invariant probability measure µ on ^J and a char- pairs (µ, χ) consisting of an ergodic O acter χ of the isotropy subgroup Hµ of µ. Formula (2.3) is a particular case of the formula in [16, Corollary 2.4] with f equal to the character function f() = xj, y on ^J associated to j P J. Since for a fixed u P O K the right hand side of (2.3) is a continuous linear functional of the integrand and the character functions span a dense subalgebra, this particular case is enough to imply χ(u)Z^J τ(µ,χ)(fνu) =  0 f(x)dµ(x) if u P Hµ otherwise, (cid:3) (2.4) for every f P C(^J). 3. THE ACTION OF UNITS ON INTEGRAL IDEALS Combining [2, Theorem 7.3] with Theorem 2.2 above, we see that for β ą 2, the extremal KMSβ equilibrium states of the system (T[OK], σ) are indexed by pairs (µ, κ) consisting of an ergodic invariant probability measure µ and a character κ of its isotropy subgroup relative to the action of the unit group O K on a representative of each ideal class. If the field K is imaginary quadratic, that is, if r = 0 and s = 1, then the group of units is finite, consisting exclusively of roots of unity. In this case, things are easy enough to describe because the space of O K -orbits in ^J is a compact Hausdorff topological space. Proposition 3.1. Suppose K is an imaginary quadratic number field, let J Ă OK be an integral ideal and write WK for the group of units. Then the orbit space WK\^J is a compact Hausdorff space and the closed invariant sets in ^J are indexed by the closed sets in WK\^J. Moreover, the ergodic invariant probability measures on ^J are the equiprobability measures on the orbits and correspond to unit point masses on WK\^J. Proof. Since WK is finite, distinct orbits are separated by disjoint invariant open sets, so the quotient space WK\^J is a compact Hausdorff space. Since ^J is compact, the quotient map q : ^J → WK\^J given by q(χ) := WK χ is a closed map by the closed map lemma, and so invariant closed sets in ^J correspond to closed sets in the quotient. For each probability measure µ on ^J, there is a probability measure µ on WK\^J defined by µ(E) := µ(q−1(E)) for each measurable E Ď WK\^J. This maps the set of WK-invariant probability measures on ^J onto the set of all probability measures on WK\^J. Ergodic invariant measures correspond to unit point masses on WK\^J, and their WK-invariant lifts are equiprobability measures on single orbits in ^J. (cid:3) As a result we obtain the following characterization of extremal KMS equilibrium states. 6 MARCELO LACA AND JACQUELINE M. WARREN Corollary 3.2. Suppose K is an imaginary quadratic algebraic number field and let Jγ be an integral ideal representing the ideal class γ P CℓK. For β ą 2, the extremal KMSβ states of the system (T[OK], σN) are parametrized by the triples (γ, W χ, κ), where γ P CℓK, χ is a point in ^Jγ, with orbit W χ and κ is a character of the isotropy subgroup of χ. Before we discuss invariant measures and isotropy for fields with infinite group of units, we need to revisit a few general facts about the multiplicative action of units on the alge- braic integers and, more generally, on the integral ideals. The concise discussion in [25] is particularly convenient for our purposes. As is customary, we let d = [K : Q] be the degree of K over Q. The number r of real embeddings and the number 2s of complex embeddings satisfy r + 2s = d. We also let n = r + s − 1 be the unit rank of K, namely, the free abelian rank of O K according to Dirichlet's unit theorem. We shall denote the real embeddings of K by σj : K → R for j = 1, 2, r and the conjugate pairs of complex embeddings of K by σr+j, σr+s+j : K→ C for j = 1, , s. Thus, there is an isomorphism σ : K bQ R→ Rr Cs such that σ(k b x) = (σ1(k)x, σ2(k)x, , σr(k)x; σr+1(k)x, , σr+s(k)x). The ring of integers OK is a free Z-module of rank d, and thus OK bZ R -- Rd -- Rr ' Cs. We is given by χt(x) = exp 2πi(θ(x) t) for x P OK. Thus, the action of a unit u P O at the level of Zd, the action of each u P O K is implemented as left multiplication by a matrix Au P GLd(Z). Moreover, once this basis has been fixed, the usual duality pairing j=1 njtj, gives temporarily fix an integral basis for OK, which fixes an isomorphism θ : OK → Zd. Then, xZd, Rd{Zdy given by xn, ty = exp 2πi(n t), with n P Zd, t P Rd and n t =řd an isomorphism of Rd{Zd to pOK, in which the character χt P pOK corresponding to t P Rd{Zd ý pOK is implemented, at the level of Rd{Zd, by the repre- K → GLd(Z). For ease of reference we state the following fact K → GLd(Z) defined by ρ(u) = AT (u χt)(x) = χt(u x) = exp 2πi(Auθ(x) t) = exp 2πi(θ(x) AT This implies that the action O K sentation ρ : O Similar considerations apply to the action of O giving a representation ρJ : O about this matrix realization ρJ of the action of O Proposition 3.3. The collection of matrices {ρ(u) : u P O (over C) , and for each u P O σk(u) : k = 1, 2, r + 2s. K on ^J. K } is simultaneously diagonalizable K the eigenvalue list of ρ(u) is the list of its archimedean embeddings u, cf. [24, Theorem 0.15]. K on ^J for each integral ideal J Ă OK, K is ut). See e.g. the discussion in [14, Section 2.1], and [25, Section 2.1] for the details. Multi- plication of complex numbers in each complex embedding is regarded as the action of 2x2 matrices on R + iR -- R2, and the 2x2 blocks corresponding to complex roots simultane- ously diagonalize over Cd. The self duality of Rr ' Cs can be chosen to be compatible with the isomorphism mentioned right after (2.1) in [14] and with multiplication by units. See also [21, Ch7]. When the number field K is not imaginary quadratic, then O K is infinite and so the analysis of orbits and invariant measures is much more subtle; for instance, most orbits are infinite, some are dense, and the orbit space does not have a Hausdorff topology. We summarize for convenience of reference the known basic general properties in the next proposition. PHASE TRANSITIONS OF C-ALGEBRAS AND FURSTENBERG CONJECTURE 7 Proposition 3.4. Let K be a number field with rank(O normalized Haar measure on ^J is ergodic O K -invariant, and for each χ P ^J, K ) ě 1, and let J be an ideal in OK. Then (1) the orbit O K χ is finite if and only if χ corresponds to a point with rational coordinates in the identification ^J -- Rd{Zd; in this case the corresponding isotropy subgroup is a full-rank subgroup of O K ; K χ is infinite if and only if χ corresponds to a point with at least one irrational (2) the orbit O coordinate in Rd{Zd; (3) the characters χ corresponding to points (w1, w2, . . . , wd) P Rd such that the numbers 1, w1, w2, . . . wd are rationally independent have trivial isotropy. Proof. By Proposition 3.3, for each u P O K , the eigenvalues of the matrix ρ(u) encoding the action of u at the level of Rd{Zd are precisely the various embeddings of u in the archimedean completions of K. Since rank(O K ) ě 1, there exists a non-torsion element u P O K , whose eigenvalues are not roots of unity. Hence normalized Haar measure is ergodic for the action of {ρ(u) : u P O K } by [24, Corollary 1.10.1] and the first assertion now follows from [24, Theorem 5.11]. The isotropy is a full rank subgroup of O K because O K {(O Let w = (w1, w2, , wd) be a point in Rd{Zd such that 1, w1, . . . , wd are rationally independent. Suppose w is a fixed point for the matrix ρ(u) P GLd(Z) acting on Rd{Zd. Then ρ(u)w = w (mod Zd) and hence (ρ(u) − I)w P Zd, i.e. K x ă∞. K )x = O [(ρ(u) − I)w]i = dÿj=1 (ρ(u) − I)ijwj P Z for all 1 ď i ď d. Since (ρ(u) − I)ij P Z for all i, j, the rational independence of 1, w1, . . . , wd implies that ρ(u) = I, so u = 1, as desired. (cid:3) We see next that for the number fields with unit rank 1 there are many more ergodic invariant probability measures on pOK than just Haar measure and measures supported on finite orbits. In fact, a smooth parametrization of these measures and of the corresponding KMS equilibrium states of (T[OK], σ) seems unattainable. Proposition 3.5. Suppose the number field K has unit-rank equal to 1, namely, K is real quadratic, is isomorphic to the Poulsen simplex [13]. mixed cubic, or complex quartic. Then the simplex of ergodic invariant probability measures on pOK Proof. The fundamental unit gives a partially hyperbolic toral automorphism of pOK, for which Haar measure is ergodic invariant. By [15, 22], the invariant probability measures of such an automorphism that are supported on finite orbits are dense in the space of all invariant probability measures. This remains true when we include the torsion elements of O K . Since these equiprobabilities supported on finite orbits are obviously ergodic invariant and hence extremal among invariant measures, it follows from [13, Theorem 2.3] that the simplex of invariant probability measures on pOK is isomorphic to the Poulsen simplex. (cid:3) For fields with unit rank at least 2, whether normalized Haar measure and equiproba- bilities supported on finite orbits are the only ergodic O K -invariant probability measures is a higher-dimensional version of the celebrated Furstenberg conjecture, according to which Lebesgue measure is the only non-atomic probability measure on T = R{Z that is jointly 8 MARCELO LACA AND JACQUELINE M. WARREN ergodic invariant for the transformations 2 and 3 on R modulo Z. As stated, this re- mains open, however, Rudolph and Johnson have proved that if p and q are multiplica- tively independent positive integers, then the only probability measure on R{Z that is ergodic invariant for p and q and has non-zero entropy is indeed Lebesgue measure [18, 7]. Number fields always give rise to automorphisms of tori of dimension at least 2, so, strictly speaking the problem in which we are interested does not contain Furstenberg's original formulation as a particular case. Nevertheless, the higher-dimensional problem is also interesting and open as stated in general, and there is significant recent activity on it and on closely related problems [9, 8, 10]. In particular, see [4] for a summary of the history and also a positive entropy result for higher dimensional tori along the lines of the Rudolph -- Johnson theorem. We show next that the toral automorphism groups arising from different integral ideals have a solidarity property with respect to the generalized Furstenberg conjecture. Proposition 3.6. If for some integral ideal J in OK the only ergodic O K -invariant probability mea- sure on ^J having infinite support is normalized Haar measure, then the same is true for every integral ideal in OK. The proof depends on the following lemmas. Lemma 3.7. Let J Ď I be two integral ideals in OK and let r : ^I→ ^J be the restriction map. Denote by λ^I normalized Haar measure on ^I. For each γ P ^J, there exists a neighborhood N of γ in ^J and homeomorphisms hj of N into ^I for j = 1, 2, . . . , I{J, with mutually disjoint images and such that (1) λ^I(hj(E)) = λ^I(hk(E)) for every measurable E Ď N and 1 ď j, k ď I{J; (2) r hj = idN; of r on N. {Aκ : κ P JK} of mutually disjoint open subsets of ^I such that κ P Aκ for each κ P JK. Define (3) r−1(E) =Ůj hj(E) for all E Ď N, that is, the hj's form a complete system of local inverses Proof. Let JK := {κ P ^I : κ(j) = 1, @j P J} be the kernel of the restriction map r : ^I → ^J. Since JK is a subgroup of order I{J ă ∞, and since ^I is Hausdorff, we may choose a collection B1 := ŞκPJK κ−1Aκ and for each κ P JK let Bκ := κB1. Then {Bκ : κ P JK} is a collection of that the restrictions r : Bκ → ^J are homeomorphisms onto their image. Since the Bκ are translates of B1 and since r is continuous and open, it suffices to verify that r is injective on B1. This is easy to see because if r(ξ1) = r(ξ2) for two distinct elements ξ1, ξ2 of B1, then ξ2 = κξ1 for some ρ P JK \ {1}, and this would contradict B1 X κB1 = H. This proves the claim. We may then take N := γ r(B1) and define hρ := (rBρ)−1, for which properties (1)-(3) are now easily verified. (cid:3) mutually disjoint open sets such that κ P Bκ and r(Bκ) = r(B1) for every κ P JK. We claim Lemma 3.8. Let X be a measurable space and let T : X → X be measurable. Suppose that λ is an ergodic T-invariant probability measure on X. If µ is a T-invariant probability measure on X such that µ ! λ, then µ = λ. Proof. Fix f P L∞(λ) and define (Anf)(x) = 1 n f(T kx). Let S = {x P X : (Anf)(x) → RX fdλ}. By the Birkhoff ergodic theorem, we have that λ(Sc) = 0, and so µ(Sc) = 0 as well, that is, (Anf)(x) → RX fdλ µ-a.e.. Since f P L∞(λ) and µ ! λ, we have that f P ∞. Observe that (Anf)(x) ď }f}λ L∞(µ) as well, with }f}µ ∞ for µ-a.e. x, and so by n−1řk=0 ∞ ď }f}λ PHASE TRANSITIONS OF C-ALGEBRAS AND FURSTENBERG CONJECTURE 9 equality because µ(X) = 1. the dominated convergence theorem, RX Anfdµ → RX(cid:0)RX fdλ(cid:1) dµ = RX fdλ, with the last Because µ is T -invariant, we have thatRX Anfdµ =RX fdµ for all n. Combining this with the above implies that RX fdλ = RX fdµ for all f P L∞(λ). In particular, this holds for the Lemma 3.9. Let J Ď I be two integral ideals in pOK and let r : ^I → ^J be the restriction map. If K -invariant probability measure on ^I, then µ := µ r−1 is an ergodic invariant µ is an ergodic O probability measure on ^J. Moreover, the support of µ is finite if and only if the support of µ is finite. indicator function of each measurable set, and so µ = λ. (cid:3) K -equivariant, r−1(E) is also O Proof. Assume µ is ergodic invariant on ^I and let E Ď ^J be an O K -invariant measurable set. Since r is O K -invariant so µ(E) := µ(r−1(E)) P {0, 1} because µ is ergodic invariant. Thus, µ is also ergodic invariant. The statement about the support follows immediately because r has finite fibers. (cid:3) Lemma 3.10. Suppose J Ď I are integral ideals in OK, and let λ^J, λ^I be normalized Haar measures K -invariant probability measure on ^J with infinite support on ^J, ^I, respectively. If the only ergodic O is λ^J, then the only ergodic O Proof. Let µ be an ergodic O By Lemma 3.9, µ r−1 is an ergodic O support, and so by assumption must equal λ^J. In particular, λ^I r−1 = λ^J. K -invariant probability measure on ^I with infinite support. K -invariant probability measure on ^J with infinite K -invariant probability measure on ^I with infinite support is λ^I. Since ^J is compact, the open cover {Nγ : γ P ^J} given by the sets constructed in Lemma 3.7 has a finite subcover, that is, there exist γ1, . . . , γn P ^J so that ^J = Nγk , where Nγk is a nŤk=1 I{JŮj=1 neighborhood of γk P ^J satisfying the conditions stated in Lemma 3.7, with corresponding maps h (j) γk, for 1 ď j ď I{J and 1 ď k ď n. We will first show that if B Ď ^I is such that rB is a homeomorphism with r(B) Ď Nγk for some k, and if λ^I(B) = 0, then µ(B) = 0. Suppose B is such a set and λ^I(B) = 0. By part (3) of Lemma 3.7, r−1(r(B)) = h (j) γk(r(B)), so r−1(r(B)) is a disjoint union of I{J (j) γk (r(B)) = B, because the h sets, all having the same measure under λ^I. Moreover, there exists some 1 ď j ď I{J such (j) γk's form a complete set of local inverses for r, and r is that h injective on B. Putting these together yields λ^I(r−1(r(B))) = I{Jλ^I(h (j) γk (r(B))) = I{Jλ^I(B) = 0. Since µ r−1 = λ^J = λ^I r−1, this implies that µ(r−1(r(B))) = 0 as well, and since B Ď r−1(r(B)), we have that µ(B) = 0. Now, since r : ^I → ^J is a covering map, for each χ P ^I, there exists an open neighbour- hood Uχ of χ such that rUχ is a homeomorphism. Let 1 ď k ď n be such that r(χ) P Nγk, and let Wχ := Uχ X r−1(Nγk ). This forms another open cover of ^I, and so by compactness of ^I, there exists a finite subcover W1, . . . , Wm. Finally, let A Ď ^I be such that λ^I(A) = 0. Then A X Wi is a set on which r acts as a homeomorphism, and there exists 1 ď k ď n such that r(A X Wi) Ď Nγk. Thus, by the above, we conclude µ(A X Wi) = 0 for all 1 ď i ď m. Since these sets cover A, we have that µ(A) = 0, and hence µ ! λ^I, as desired. By Lemma 3.8 it follows that µ = λ^I. (cid:3) 10 MARCELO LACA AND JACQUELINE M. WARREN Proof of Proposition 3.6: Suppose J is an integral ideal such that the only O K -invariant prob- ability measure on ^J having an infinite orbit is normalized Haar measure. By Lemma 3.10 applied to the inclusion J Ă OK, the only ergodic O with infinite support is normalized Haar measure. K -invariant probability measure on pOK Suppose now I Ď OK is an arbitrary integral ideal. Since the ideal class group is finite, K such that qOK Ď I. The action K -invariant a power of I is principal and thus we may choose q P O of O probability measure on yqOK with infinite support is normalized Haar measure. Thus, by Lemma 3.10 again with yqOK Ď ^I, we conclude that the only ergodic O K -invariant probabil- (cid:3) ity measure on ^I is λ^I. K on pOK is conjugate to the action of O K on yqOK, and so the only ergodic O In order to understand the situation for number fields with unit rank higher than 1, we review in the next section the topological version of the problem of ergodic invariant measures, namely, the classification of closed invariant sets. 4. BEREND'S THEOREM AND NUMBER FIELDS An elegant generalization to higher-dimensional tori of Furstenberg's characterization [5, Theorem IV.1] of closed invariant sets for semigroups of transformations of the cir- cle was obtained by Berend [1, Theorem 2.1]. The fundamental question investigated by Berend is whether an infinite invariant set is necessarily dense, and his original formula- tion is for semigroups of endomorphisms of a torus. Here we are interested in the specific situation arising from an algebraic number field K in which the units O K act by automor- phisms on ^J for integral ideals J Ď OK representing each ideal class, so we paraphrase Berend's Property ID for the special case of a group action on a compact space. Definition 4.1. (cf. [1, Definition 2.1].) Let G be a group acting on a compact space X by homeomorphisms. We say that the action G ý X satisfies the ID property, or that it has the infinite invariant dense property, if the only closed infinite G-invariant subset of X is X itself. The first observation is a topological version of the measure-theoretic solidarity proved ý ^J has the ID in Proposition 3.6; namely, if K is a given number field, then the action O K property either for all integral ideals J, or for none. Proposition 4.2. Suppose K is an algebraic number field, and let J be an ideal in OK. Then the action of O K on ^J is ID if and only if the action of O K on ^OK is ID. Proof. Suppose first that J1 Ď J2 are ideals in OK and assume that the action of O is ID. The restriction map r : ^J2 → ^J1 is O finite fibers. Thus, if E were a closed, proper, infinite O would be a closed, proper, infinite O that the action of O is also ID. K on ^J2 K -equivariant, continuous, surjective, and has K -invariant subset of ^J1, then r−1(E) K -invariant subset of ^J2, contradicting the assumption K on ^J1 K on ^J2 is ID. So no such set E exists, proving that the action of O In particular, if the action of O K on ^OK is ID, then the action on ^J is also ID for every integral ideal J Ă OK. For the converse, recall that, as in the proof of Proposition 3.6, there exists an integer q P O K such that qOK Ď J, so we may apply the preceding paragraph to this inclusion. Since the action of O proof. K on yqOK is conjugate to that on pOK, this completes the (cid:3) PHASE TRANSITIONS OF C-ALGEBRAS AND FURSTENBERG CONJECTURE 11 In order to decide for which number fields the action of units on the integral ideals is ID, we need to recast Berend's necessary and sufficient conditions in terms of properties of the number field. Recall that, by definition, a number field is called a complex multiplication (or CM) field if it is a totally imaginary quadratic extension of a totally real subfield. These fields were studied by Remak [17], who observed that they are exactly the fields that have a unit defect, in the sense that they contain a proper subfield L with the same unit rank. Theorem 4.3. Let K be an algebraic number field and let J be an ideal in OK. The action of O ^J is ID if and only if K is not a CM field and rank O K on K ě 2. For the proof we shall need a few number theoretic facts. We believe these are known but we include the relatively straightforward proofs below for the convenience of the reader. Lemma 4.4. Suppose F is a finite family of subgroups of Zd such that rank(F) ă d for every F P F. Then there exists m P Zd such that m + F is nontorsion in Zd{F for every F P F. Proof. Recall that for each subgroup F there exists a basis {nF a1, a2, . . . , arank(F) such that j }j=1,2,...,d of Zd and integers F =(cid:8)řrank(F) i=1 kinF i : ki P aiZ, 1 ď i ď rank(F)(cid:9). 1, . . . , nF The associated vector subspaces SF := spanR{nF rank(F)} of Rd are proper and closed so Rd \ YFSF is a nonempty open set, see e.g. [19, Theorem 1.2]. Let r be a point in Rd \ YFSF with rational coordinates. If k denotes the l.c.m. of all the denominators of the coordinates of r, then m := kr P Zd and its image m + F P Zd{F is of infinite order for every F because m R SF. (cid:3) Proposition 4.5. Let K be an algebraic number field. Then there exists a unit u P O K = Q(uk) for every k P N if and only if K is not a CM field. Proof. Assume first K is not a CM field. Then rank O K for every proper subfield F of K. Since there are only finitely many proper subfields F of K, Lemma 4.4 gives a unit u P O F for every F. Thus uk R F for every proper subfield F of K and every k P N. K with nontorsion image in O F ă rank O K such that K {O Assume now K is a CM field, and let F be a totally real subfield with the same unit rank F is finite and there exists a fixed integer m such that (cid:3) as K [17]. Then the quotient O um P F for every u P O K . Lemma 4.6. Let k be an algebraic number field with rank O K ě 1. Then for every embedding K {O Proof. Assume for contradiction that σ is an embedding of k in C such that σ(O K ) Ď {z P C : z = 1}. Let K = σ(k) and let UK = σ(O K ). Then K X R is a real subfield of K with UKXR = {1}, so K X R = Q. Also K X R is the maximal real subfield of K, and since we are assuming rank O K ě 1, K cannot be a CM field. To see this, suppose that k were CM. Let ℓ Ď k be a totally real subfield such that [k : ℓ] = 2. Since ℓ is totally real, σ(ℓ) Ď R, and since K X R = Q, it must be that σ(ℓ) = Q. Then ℓ = Q, so k is quadratic imaginary, contradicting rank O K ě 1. By Proposition 4.5, there exists u P UK such that K = Q(u). Since u = 1, we have that K = Q(u) = Q(u−1) = Q(u) = K, so K is closed under complex conjugation. Write σ : k→ C, there exists u P O K such that σ(u) ą 1. 12 MARCELO LACA AND JACQUELINE M. WARREN u = a + ib. Then u + u = 2a P K X R = Q, so a P Q. Thus, K = Q(u) = Q(ib). Since u = 1, a2 + b2 = 1, and so we have that Q(ib) -- Q(a−b2) -- Q(aa2 − 1) -- Q c m2 − n2 n2 ! -- Q(am2 − n2), where a = m{n P Q. Thus, K is a quadratic field. But it cannot be quadratic imaginary because rank UK ě 1, and it cannot be quadratic real because all the units lie on the unit circle. This proves there can be no such embedding. (cid:3) Proof of Theorem 4.3. By Proposition 4.2, it suffices to prove the case J = OK. Let d = [K : Q] conditions for ID [1, Theorem 2.1], when interpreted for the automorphic action of O and recall that pOK -- Td. All we need to do is verify that Berend's necessary and sufficient pOK, characterize non-CM fields of unit rank 2 or higher. Since the action of O K on K by linear K is faithful by [10, p. 729], Berend's conditions are: (1) (totally irreducible) there exists a unit u such that the characteristic polynomial of toral automorphisms ρ(u) with u P O ρ(un) is irreducible for all n P N; (2) (quasi-hyperbolic) for every common eigenvector of {ρ(u) : u P O K }, there is a unit K such that the corresponding eigenvalue of ρ(u) is outside the unit disc; and K such that if m, n P N satisfy ρ(um) = (3) (not virtually cyclic) there exist units u, v P O u P O ρ(vn), then m = n = 0. Suppose first that the action of O (3) above hold, i.e. the action of O By Proposition 4.5, K is not a CM field and since ρ : O restatement of rank O K ě 2. K on OK is ID. By [1, Theorem 2.1] conditions (1) and K on OK is totally irreducible and not virtually cyclic. K → GLd(Z) is faithful, (3) is a Suppose now that K is not CM and has unit-rank at least 2. By Proposition 4.5, there exists u P O K such that Q(un) = K for every n P N. Hence the minimal polynomial of ρ(un) has degree d, and so it coincides with the characteristic polynomial. This proves the action of ρ(u) is totally irreducible. We have already that condition (1) holds, i.e. observed that condition (3) holds iff the unit rank of K is at least 2, so it remains to see that the hyperbolicity condition (2) holds too. In the simultaneous diagonalization of the matrix group ρ(O K ), the diagonal entries of ρ(u) are the embeddings of u into R or C, see e.g. [10, p.729]. Then condition (2) follows from Lemma 4.6. (cid:3) Remark 4.7. Notice that for units acting on algebraic integers, Berend's hyperbolicity con- dition (2) is automatically implied by the rank condition (3). Remark 4.8. Since the matrices representing the actions of O K on ^J and on ^OK are conjugate over Q, Proposition 4.2 can be derived from the implication (1) =⇒ (3) in [10, Proposition 2.1]. We may also see that the matrices implementing the action on ^J and on ^OK have the same sets of characteristic polynomials, so the questions of expansive eigenvalues (condi- tion (2)) and of total irreducibility are equivalent for the two actions. The third condition is independent of whether we look at ^J or ^OK, so this yields yet another proof of Proposition 4.2. By Theorem 4.3, for each non-CM algebraic number field K with unit rank at least 2, the K } acting on Rd{Zd, is an example of an abelian action O K on pOK, transposed as {ρ(u) : u P O PHASE TRANSITIONS OF C-ALGEBRAS AND FURSTENBERG CONJECTURE 13 toral automorphism group for which one may hope to prove that normalized Haar mea- sure is the only ergodic invariant probability measure with infinite support. So it is natural to ask which groups of toral automorphisms arise this way. A striking observation of Z. Wang [25, Theorem 2.12], see also [14, Proposition 2.2], states that every finitely gener- ated abelian group of automorphisms of Td that contains a totally irreducible element and whose rank is maximal and greater than or equal to 2 arises, up to conjugacy, from a finite index subgroup of units acting on the integers of a non-CM field of degree d and unit rank at least 2, cf. [25, Condition 1.5]. We wish next to give a proof of the converse, which was also stated in [25]. Proposition 4.9. Suppose G is an abelian subgroup of SLd(Z) satisfying [25, Condition 2.8]. Specifically, suppose there exist ‚ a non-CM number field K of degree d and unit rank at least 2; K of G into a finite index subgroup of O K ; ‚ a co-compact lattice Γ in K Ă K bQ R -- Rd invariant under multiplication by φ(G); and ‚ an embedding φ : G→ O ‚ a linear isomorphism ψ : Rd → K bQ R -- Rd mapping Zd onto Γ that intertwines the actions G ý Rd and φ(G) ý (K bQ R){Γ. Then G satisfies [25, Condition 1.5], namely (1) rank(G) ě 2; (2) the action g ý Rd{Zd -- Td is totally irreducible for some g P G; (3) rank G1 = rank G for each abelian subgroup G1 Ă SLd(Z) containing G. Proof. Suppose K is a non-CM algebraic number field of degree d with unit rank at least 2, and assume G is a subgroup of SLd(Z) that satisfies the assumptions with respect to K. Part (1) of [25, Condition 1.5] is immediate, because φ(G) is of full rank in O K . By Proposition 4.5, there exists a unit u P O K such that the characteristic polynomial of ρ(um) is irreducible over Q for all m P N. This is equivalent to the action of um on pOK being irreducible for all m P N, see, e.g. [10, Proposition 3.1]. Since φ(G) is of finite index in O K , there exists N P N such that uN P φ(G). We claim that g := φ−1(uN) is a totally irreducible element in G ý Rd{Zd. To see this, it suffices to show that the characteristic polynomial of gk is irreducible over Q for every positive integer k. Since the linear isomorphism ψ inter- twines the actions gk ý Td and ρ(φ(g))k ý (K bQ R){Γ , the characteristic polynomial of gk equals the characteristic polynomial of ρ(φ(g))k = ρ(ukN), which is irreducible because it coincides with the characteristic polynomial of ukN as an element of the ring OK. This proves part (2) of Condition 1.5. Suppose now that G1 is an abelian subgroup of SLd(Z) containing G and apply the construction from [25, Proposition 2.13] (see also [21, 4]) to the irreducible element g P G Ă G1 ý Td. Up to an automorphism, the resulting number field arising from this construction is K = Q(uN), and the embedding φ1 : G1 → O O K . Since φ(G) Ă φ1(G1) Ă O K and φ(G) is of finite index in O K , K is an extension of φ : G → rank(G1) = rank φ1(G1) = rank O K = rank φ(G) = rank G, and this proves proves part (3) of Condition 1.5. (cid:3) As a consequence, we see that the action of units on the algebraic integers of number fields are generic for group actions with Berend's ID property in the following sense, cf. [25, 14]. 14 MARCELO LACA AND JACQUELINE M. WARREN Corollary 4.10. If G is a finitely generated abelian subgroup of SLd(Z) of torsion-free rank at least 2 that contains a totally irreducible element and is maximal among abelian subgroups of SLd(Z) containing G, then G is conjugate to a finite-index toral automorphism subgroup of the action of O K ý pOK for a non-CM algebraic number field K of degree d and unit rank at least 2. Finally, we summarize what we can say at this point for equilibrium states of C-algebras associated to number fields with unit rank strictly higher than one. If the generalized Furstenberg conjecture is verified, the following result would complete the classification started in Proposition 3.1 and Proposition 3.5. Let K be a number field and for each γ P CℓK define Fγ to be the set of all pairs (µ, χ) with K in ^Jγ, and χ P ^Hµ, where µ an equiprobability measure on a finite orbit of the action of O the µ-a.e. isotropy group Hµ is a finite index subgroup of O K . Also let (λJ, 1) denote the pair consisting of normalized Haar measure on ^J and the trivial character of its trivial a.e. state of C(Jγ ¸ O C(Jγ ¸O isotropy group. Then the map (µ, χ) Þ→ τµ,χ from Theorem 2.2 gives an extremal tracial Recall that the map τ Þ→ ϕτ from [2, Theorem 7.3] is an affine bijection of all tracial states ofÀγPCℓK K ) onto Kβ, the simplex of KMSβ equilibrium states of the system K ) for each pair (µ, χ) P Fγ \ {(λJγ , 1)}. (T[OK], σ) studied in [2]. Theorem 4.11. Suppose K is an algebraic number field with unit rank at least 2 and define Φ : (µ, χ) Þ→ ϕτµ,χ to be the composition of the maps from Theorem 2.2 and from [2, Theorem 7.3], assigning a state ϕτµ,χ P Extr(Kβ) to each pair (µ, χ) consisting of an ergodic invariant probability measure µ in one of the ^Jγ and an associated character of the µ-almost constant isotropy Hµ. Let FK := ğγPCℓK(cid:0)Fγ \ {(λJγ , 1)}(cid:1) be the set of pairs whose measure µ has finite support or is Haar measure. Then (1) if K is a CM field, then the inclusion Φ(FK) Ă Extr(Kβ) is proper; and (2) if K is not a CM field, and if there exists φ P Extr(Kβ) \ Φ(FK) then the measure µ on ^Jγ arising from φ has zero-entropy and infinite support. Proof. To prove assertion (1), recall that when K is a CM field Berend's theorem implies that there are invariant subtori, which have ergodic invariant probability measures on the fibers, cf. [8, 9]. These measures give rise to tracial states and to KMS states not accounted for in Φ(FK). Assertion (2) follows from [4, Theorem 1.1]. (cid:3) 5. PRIMITIVE IDEAL SPACE The computation of the primitive ideal spaces of the C-algebras C(J ¸ O K ) associated to the action of units on integral ideals lies within the scope of Williams' characterization in [27]. We briefly review the general setting next. Let G be a countable, discrete, abelian group acting continuously on a second countable compact Hausdorff space X. We define an equivalence relation on X by saying that x and y are equivalent if x and y have the same orbit closure, i.e. if G x = G y. The equivalence class of x, denoted by [x], is called the quasi-orbit of x, and the quotient space, which in general is not Hausdorff, is denoted by Q(G ý X) and is called the quasi-orbit space. It is important to distinguish the quasi-orbit of a point from the closure of its orbit, as the latter may contain other points with strictly smaller orbit closure. PHASE TRANSITIONS OF C-ALGEBRAS AND FURSTENBERG CONJECTURE 15 Let ǫx denote evaluation at x P X, viewed as a one-dimensional representation of C(X). For each character κ P ^Gx, the pair (ǫx, κ) is clearly covariant for the transformation group (C(X), Gx), and the corresponding representation ǫx κ of C(X) ¸ Gx gives rise to an induced representation IndG (ǫx κ) of C(X) ¸ G, which is irreducible because ǫx κ is. Gx Since G is abelian and the action is continuous, whenever x and y are in the same quasi- orbit, [x] = [y], the corresponding isotropy subgroups coincide: Gx = Gy. Thus, we may consider an equivalence relation on the product Q(G ý X) ^G defined by ([x], κ) ∼ ([y], λ) ⇐⇒ [x] = [y] and κaeGx= λaeGx . By [27, Theorem 5.3], the map (x, κ) Þ→ ker IndG (ǫx κ) induces a homeomorphism of (Q(G ý X) ^G){∼ onto the primitive ideal space of the crossed product C(X) ¸ G, see e.g. [12, Theorem 1.1] for more details on this approach. We wish to apply the above result to actions O K ý ^J for integral ideals J of non-CM number fields with unit rank at least 2, as in Theorem 4.3. Notice that by Proposition 3.4 if the orbit O K χ is finite, then it is equal to the quasi-orbit [χ]. The first step is to describe the quasi-orbit space for the action of units. We focus on the case J = OK; ideals representing nontrivial classes behave similarly because of the solidarity established in Proposition 4.2. Gx Proposition 5.1. Suppose K is a non-CM algebraic number field with unit rank at least 2. Then the quasi-orbit space of the action O K Q(O K ý pOK) = {[x] : O K x ă∞} Y {ω∞}. The point ω∞ is the unique infinite quasi-orbit [α] of any α P pOK -- Rd{Zd having at least one irrational coordinate. The closed proper subsets are the finite subsets all of whose points are finite (quasi-)orbits. Infinite subsets and subsets that contain the infinite quasi-orbit ω∞ are dense in the whole space. ý pOK is Proof. By Theorem 4.3, the closure of each infinite orbit is the whole space. Thus, the points with infinite orbits collapse into a single quasi-orbit K x =∞} = {x P pOK : O That this is the set of points with at least one irrational coordinate is immediate from [24, Theorem 5.11]. When the orbit of x is finite, it is itself a quasi-orbit, which we view as a point in Q(O K K x = pOK}. ω∞ := {x P pOK : O ý pOK). In this case x P pOK has all rational coordinates. tive, continuous and open by the Lemma in page 221 of [6], see also the proof of Proposi- tion 2.4 in [12]. and as such can be separated by disjoint open sets V and W, so that [x] Ă V and [y] Ă W. Passing to the quotient space, we have [x] R q(W) and [y] R q(V), so [x] and [y] are T1- separated, which implies that finite sets of finite quasi-orbits are closed in Q(O K ý pOK) is surjec- To describe the topology, recall that the quotient map q : pOK → Q(O Any two different finite quasi-orbits [x] and [y] are finite, mutually disjoint subsets of pOK ý pOK). ý pOK) because every infinite orbit in pOK is dense ý pOK) consisting of finite quasi-orbits, thenŤ[x]PA[x] is an infinite invariant set in pOK, hence is dense by Theorem 4.3. This implies that ω∞ is in the closure of A, and hence A is dense in Q(O K by Theorem 4.3. If A is an infinite subset of Q(O K The singleton {ω∞} is dense in Q(O K (cid:3) K ý pOK). 16 MARCELO LACA AND JACQUELINE M. WARREN (5.1) Theorem 5.2. Let K be a non-CM algebraic number field with unit rank at least 2, and let G = O K . ğ[x] (cid:0){[x]} ^Gx(cid:1) The primitive ideal space of C(pOK) ¸ G is homeomorphic to the space in which a net ([xι], γι) converges to ([x], γ) iff [xι]→ [x] in Q(G ý pOK) and γιGx → γGx in ^Gx. [x] = ω∞, then the condition γιGω∞ → γGω∞ is trivially true because Gω∞ = {1}. Proof. Consider the diagram below, where f is the quotient map and the vertical map g is defined by g([([x], γ)]) = ([x], γGx ), where [([x], γ)] denotes the equivalence class of ([x], γ) with respect to ∼. Notice that if [x] is a finite quasi-orbit, then the net {[xι]} is eventually constant equal to [x], and if Q(G ý pOK) ^G f gf g Q(G ý pOK) ^G{ ∼ Ů[x](cid:0){[x]} ^Gx(cid:1) By the fundamental property of the quotient map, see e.g. [26, Theorem 9.4], g f is con- tinuous if and only if g is continuous. It is clear that g is a bijection. We show next that g f is continuous. Suppose that converging to ([x], γ). converges to g f([x], γ) = ([x], γGx ), as desired. It remains to show that g−1 is continuous, or equivalently, that g is a closed map. Sup- ([xι], γι) is a net in Q(G ý pOK) ^G converging to ([x], γ). Then [xι] → [x] in Q(G ý pOK), and γι → γ in ^G. Then clearly also γιGx → γGx in ^Gx as well. Hence the net g f([xι], γι) pose that W Ď Q(G ý pOK) ^G{ ∼ is closed, and suppose that ([xι], γι) is a net in g(W) convergent subnet γιη with limit γ. Then γιη Gx → γGx as well, so γιη → γGx . Since ^Gx is The net ([xιη], γιη ) converges to ([x], γ) in Q(G ý pOK) ^G, and since f is continuous, f([xιη ], γιη ) → f([x], γ). Moreover, f([xιη], γιη ) = [([xιη ], γιη )] P W because g is injective and g([([xιη ], γιη )]) = ([xιη ], γιη ) P g(W) by assumption. Since W is closed, [([x], γ)] P W, and so its image ([x], γ) P g(W), as desired. (cid:3) Consider any net ( γι) in ^G such that γιGx = γι. By the compactness of ^G, there exists a Hausdorff, limits are unique, and hence γGx = γ. Remark 5.3. Recall that G -- W Zd with W the roots of unity in G, and that the isotropy subgroup Gx is constant on the quasi-orbit [x] of x. If [x] is finite, then Gx is of full rank in G, and thus Gx -- Vx Zd, with Vx Ă W the torsion part of Gx. Hence, for every finite quasi-orbit [x], we have ^Gx -- ^V[x] Td. Notice that ^V[x] -- V[x] (noncanonically) because Vx is finite. REFERENCES [1] D. Berend, Multi-invariant sets on tori, Trans. Amer. Math. Soc. 280 (1983), 509 -- 532. [2] J. Cuntz, C. Deninger and M. Laca, C-algebras of Toeplitz type associated with algebraic number fields, Math. Ann. 355 (2013), 1383 -- 1423. [3] J. Cuntz, S. Echterhoff, and X. Li, On the K-theory of the C-algebra generated by the left regular representation of an Ore semigroup, J. Eur. Math. Soc. 17 (2015), 645 -- 687. PHASE TRANSITIONS OF C*-ALGEBRAS AND FURSTENBERG CONJECTURE 17 [4] M. Einsiedler and E. Lindenstrauss, Rigidity properties of Zd actions on tori and solenoids, Electr. Res. An- nounc. Amer. Math. Soc. 9 (2003), 99 -- 110. [5] H. Furstenberg, Disjointness in ergodic theory, minimal sets, and a problem in diophantine approximation, Math. Systems Theory 1 (1967), 1 -- 49. [6] P. Green, The local structure of twisted covariance algebras, Acta Math. 140 (1978), 191 -- 250. [7] A. Johnson, Measures on the circle invariant under multiplication by a nonlacunary subsemigroup of the integers, Israel J. Math. 77 (1992), 211 -- 240. [8] A. Katok, B. Kalinin, Invariant measures for actions of higher rank abelian groups, Smooth Ergodic Theory and its Applications (Seattle, WA, 1999), Proc. Sympos. Pure Math., 69 (2001), 593 -- 637. [9] A. Katok, R. Spatzier, Invariant measures for higher-rank hyperbolic abelian actions, Ergod. Th. and Dynam. Syst. 16 no. 4 (1996), 751 -- 778. [10] A. Katok, S. Katok, and K. Schmidt, Rigidity of measurable structure for Zd actions by automorphisms of a torus, Comment. Math. Helv. 77 (2002), 718 -- 745. [11] Y. Katznelson, Ergodic automorphisms of T n are Bernoulli shifts, Israel J. Math. 10 (1971), 186 -- 195. [12] M. Laca and I. Raeburn, The ideal structure of the Hecke C-algebra of Bost and Connes, Math. Ann. 318 (2000), 433-451. [13] J. Lindenstrauss, G. Olsen, and Y. Sternfeld, The Poulsen simplex, Ann. Inst. Fourier 28 (1978), 91 -- 114. [14] E. Lindenstrauss and Z. Wang, Topological self-joinings of Cartan actions by toral automorphisms, Duke Math. J. 161 No. 7 (2012), 1305 -- 1350. [15] B. Marcus, A note on periodic points for ergodic toral automorphisms, Mh. Math. 89 (1980), 121 -- 129. [16] S. Neshveyev, KMS states on the C-algebras of non-principal groupoids, J. Operator Theory 70 (2013), 513 -- 530. [17] R. Remak, Über algebraische Zahlkörper mit schwachem Einheitsdefekt, Compositio Math. 12 (1954), 35 -- 80. [18] D.J. Rudolph, 2 and 3 invariant measures and entropy, Ergod. Th. and Dynam. Syst. 10 no. 2 (1990), 395 -- 406. [19] S. Roman, Advanced Linear Algebra, 3rd Ed. Graduate Texts in Mathematics, 2008 Springer. [20] I.N. Stewart and D.O. Tall, Algebraic Number Theory, 1987, Chapman and Hall. [21] K. Schmidt, Dynamical Systems of Algebraic Origin, 3rd Ed. Graduate Texts in Mathematics, 2008, Springer. [22] K. Sigmund, Generic properties of invariant measures for Axiom A diffeomorphisms, Invent. math. 11 (1970), 99-109. [23] E. Thoma, Über unitäre Darstellungen abzählbarer, diskreter Gruppen, Math. Ann. 153 (1964), 111 -- 138. [24] P. Walters, An Introduction to Ergodic Theory, Graduate Texts in Mathematics no. 79, 1982, Springer. [25] Z. Wang, Quantitative Density under Higher Rank Abelian Algebraic Toral Actions, Int. Math. Res. Not. No. 16 (2011), 3744 -- 3821. [26] S. Willard, General Topology, 1970, Addison-Wesley, . [27] D.P. Williams, The topology on the primitive ideal space of transformation group C-algebras and C.C.R. trans- formation group C-algebras, Trans. Amer. Math. Soc. 226 (1981), 335 -- 359. (M. Laca) DEPARTMENT OF MATHEMATICS AND STATISTICS, UNIVERSITY OF VICTORIA, CANADA (J. M. Warren) DEPARTMENT OF MATHEMATICS, UNIVERSITY OF CALIFORNIA, SAN DIEGO
1709.08996	1	1709	2017-09-26T13:12:58	On a theorem of Kucerovsky for half-closed chains	[ "math.OA", "math.FA", "math.KT" ]	Kucerovsky's theorem provides a method for recognizing the interior Kasparov product of selfadjoint unbounded cycles. In this paper we extend Kucerovsky's theorem to the non-selfadjoint setting by replacing unbounded Kasparov modules with Hilsum's half-closed chains. On our way we show that any half-closed chain gives rise to a multitude of twisted selfadjoint unbounded cycles via a localization procedure. These unbounded modular cycles allow us to provide verifiable criteria avoiding any reference to domains of adjoints of symmetric unbounded operators.	math.OA	math	ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS JENS KAAD AND WALTER D. VAN SUIJLEKOM Abstract. Kucerovsky's theorem provides a method for recog- nizing the interior Kasparov product of selfadjoint unbounded cy- cles. In this paper we extend Kucerovsky's theorem to the non- selfadjoint setting by replacing unbounded Kasparov modules with Hilsum's half-closed chains. On our way we show that any half- closed chain gives rise to a multitude of twisted selfadjoint un- bounded cycles via a localization procedure. These unbounded modular cycles allow us to provide verifiable criteria avoiding any reference to domains of adjoints of symmetric unbounded opera- tors. 1. Introduction In recent years a lot of attention has been given to the non-unital framework for noncommutative geometry, where the absence of a unit is interpreted as a non-compactness condition on the underlying non- commutative space, [Con94, Lat13, CGRS14, MeRe16]. For a more detailed analysis of the non-compact setting it is important to distin- guish between the complete and the non-complete case, [MeRe16]. Whereas the complete case is still modelled by a (non-unital) spec- tral triple or more generally an unbounded Kasparov module, the lack of completeness leads to the non-selfadjointness of symmetric differen- tial operators. A noncommutative geometric framework that captures the non-complete setting is provided by Hilsum's notion of a half- closed chain, where the selfadjointness condition on the unbounded operator is replaced by a more flexible symmetry condition, [Hil10]. This framework is supported by results of Baum, Douglas, Taylor and Hilsum showing that any first-order symmetric elliptic differential op- erator on any Riemannian manifold gives rise to a half-closed chain, [BDT89, Hil10]. Unbounded Kasparov modules give rises to classes in Kasparov's KK-theory via the Baaj-Julg bounded transform and this result has been extended by Hilsum to cover half-closed chains, [BaJu83, Hil10]. Date: November 13, 2018. 2010 Mathematics Subject Classification. 19K35; 58B34. Key words and phrases. Unbounded Kasparov modules, Half-closed chains, Un- bounded modular cycles, KK-theory, Unbounded KK-theory, Kasparov product, Unbounded Kasparov product. 1 2 JENS KAAD AND WALTER D. VAN SUIJLEKOM This transform contains information about the algebraic topology of the original geometric situation described by a half-closed chain. The main structural property of Kasparov's KK-theory is the interior Kasparov product, [Kas80]: (cid:98)⊗B : KK(A, B) × KK(B, C) → KK(A, C) . The interior Kasparov product is however not explicitly constructed and it is therefore important to develop tools for computing the interior Kasparov product of two given Kasparov modules. Given three classes in KK-theory, Connes and Skandalis developed suitable conditions for verifying whether one of these three classes factorizes as an interior Kasparov product of the remaining two classes, [CoSk84]. The conditions of Connes and Skandalis were translated to the un- bounded setting by Kucerovsky, [Kuc97]. Thus, given three unbounded Kasparov modules, Kucerovsky's theorem provides criteria for verify- ing that one of these unbounded Kasparov modules factorizes as an unbounded Kasparov product of the remaining two unbounded Kas- parov modules. In many cases, the conditions are easier to verify di- rectly at the unbounded level, using Kucerovsky's theorem, instead of first applying the bounded transform and then relying on the results of Connes and Skandalis. Indeed, in the unbounded setting we are usually working with first-order differential operators whereas their bounded transforms are zeroth-order pseudo-differential operators involving a square root of the resolvent. In this paper we extend Kucerovsky's theorem to cover the non- complete setting, where the unbounded Kasparov modules are replaced by half-closed chains. The main challenge in carrying out such a task is that the domain of the adjoint of a symmetric unbounded operator can be difficult to describe. The original proof of Kucerovsky does therefore not translate to the non-selfadjoint setting as the correct conditions have to be formulated without any reference to maximal domains of symmetric unbounded operators. The main technique that we apply is a localization procedure relat- ing to the work of the first author in [Kaa15, Kaa17]. This procedure allows us to pass from a symmetric regular unbounded operator D to an essentially selfadjoint regular unbounded operator of the form xDx∗ for an appropriate bounded adjointable operator x. In the case where D is a Dirac operator, the localization corresponds to a combination of two operations: restricting all data to an open subset and passing from the non-complete Riemannian metric on this open subset to a confor- mally equivalent but complete Riemannian metric. The size of the open neighborhood and the relevant conformal factor are both determined by the positive function xx∗. In particular, our technique allows us to construct a multitude of un- bounded modular cycles out of a given half-closed chain. We interpret ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 3 this localization procedure in terms of the unbounded Kasparov prod- uct by the module generated by the localizing element x. In this way, we may work with selfadjoint unbounded operators and hence elimi- nate the difficulties relating to the description of maximal domains. On the other hand, the "conformal factor" (xx∗)−2 produces a twist of the commutator condition and this twist is described by the modular au- tomorphism σ(·) = (xx∗)(·)(xx∗)−1. We refer to Connes and Moscovici for further discussion of this issue in the case where x is positive and invertible, see [CoMo08]. The present paper is motivated by the geometric setting of a proper Riemannian submersion of spinc-manifolds, and the criteria that we develop here have already been applied in [KavS17] to obtain fac- torization results involving the corresponding fundamental classes in KK-theory. Our results may also be of importance for the further development of the unbounded Kasparov product as initiated by Connes in [Con96] and developed further by Mesland and others in [Mes14, KaLe13, BMvS16, Kaa15, MeRe16, Kaa16]. The structure of this paper is as follows: In Section 2 and Section 3 we review the concept of a half-closed chain and of an unbounded modular cycle. In Section 4, Section 5 and Section 6 we prove our results on the localization procedure and investigate how it relates to the Kasparov product. In Section 7 we prove Kucerovsky's theorem for half-closed chains. Acknowledgements. We would like to thank Georges Skandalis for a highly stimulating remark concerning the "locality" of Kucerovsky's theorem. This work also benefited from various conversations with Magnus Goffeng and Bram Mesland. The authors gratefully acknowledge the Syddansk Universitet Odense and the Radboud University Nijmegen for their financial support in fa- cilitating this collaboration. During the initial stages of this research project the first author was supported by the Radboud excellence fellowship. The first author was partially supported by the DFF-Research Project 2 "Automorphisms and Invariants of Operator Algebras", no. 7014- 00145B and by the Villum Foundation (grant 7423). The second author was partially supported by NWO under VIDI- grant 016.133.326. 2. Half-closed chains Let us fix two σ-unital C∗-algebras A and B. Let E be a countably generated Hilbert C∗-module over B. We recall that a closed (densely defined) unbounded operator D : Dom(D) → 4 JENS KAAD AND WALTER D. VAN SUIJLEKOM E is said to be regular when it has a densely defined adjoint D∗ : Dom(D∗) → E and when 1 + D∗D : Dom(D∗D) → E has dense range. It follows from this definition that 1 + D∗D : Dom(D∗D) → E is in fact densely defined and surjective, [Lan95, Lemma 9.1]. In particular we have a bounded adjointable inverse (1 + D∗D)−1 : E → E. For two countably generated Hilbert C∗-modules E and F over B, we let L(E, F ) and K(E, F ) denote the bounded adjointable operators from E to F and the compact operators from E to F , respectively. When E = F we put L(E) := L(E, F ) and K(E) := K(E, F ). We let (cid:107) · (cid:107)∞ : L(E, F ) → [0,∞) denote the operator norm. The following definition is due to Hilsum, [Hil10, Section 3]: Definition 1. A half-closed chain from A to B is a triple (A , E, D), where A ⊆ A is a norm-dense ∗-subalgebra, E is a countably generated C∗-correspondence from A to B and D : Dom(D) → E is a closed, symmetric and regular unbounded operator such that (2) a(cid:0)Dom(D∗)(cid:1) ⊆ Dom(D) for all a ∈ A ; (1) a · (1 + D∗D)−1 is a compact operator on E for all a ∈ A; (3) [D, a] : Dom(D) → E extends to a bounded operator d(a) : E → E for all a ∈ A . A half-closed chain (A , E, D) from A to B is said to be even when E comes equipped with a Z/2Z-grading operator γ : E → E (γ = γ∗, γ2 = 1), such that [a, γ] = 0 for all a ∈ A and Dγ = −γD. A half-closed chain which is not even is said to be odd. Let (A , E, D) be a half-closed chain from A to B. A few observations are in place: −d(a∗). (1) d(a) : E → E, a ∈ A , is automatically adjointable with d(a)∗ = (2) The difference Da − aD∗ : Dom(D∗) → E a ∈ A extends to the bounded adjointable operator d(a) : E → E. (3) a · (1 + DD∗)−1 ∈ K(E) for all a ∈ A. (Remark that D∗ is automatically regular by [Lan95, Proposition 9.5]). We recall that a Kasparov module from A to B is a pair (E, F ) where E is a countably generated C∗-correspondence from A to B and F : E → E is a bounded adjointable operator such that a · (F − F ∗) , a · (F 2 − 1) , [F, a] ∈ K(E) , for all a ∈ A. A Kasparov module (E, F ) from A to B is even when it comes equipped with a Z/2Z-grading operator γ : E → E such that [a, γ] = 0 for all a ∈ A and F γ + γF = 0. Otherwise we say that (E, F ) is odd. ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 5 For an unbounded regular operator D : Dom(D) → E we let FD := D(1 + D∗D)−1/2 ∈ L(E) denote the bounded transform of D. We have that F ∗ D = FD∗ = D∗(1 + DD∗)−1/2. The next result creates the main link between half-closed chains and Kasparov modules. This result is due to Hilsum, [Hil10], and it generalizes the corresponding result of Baaj and Julg for unbounded Kasparov modules, [BaJu83]. Remark however that the condition [FD, a] ∈ K(E), a ∈ A, is for some reason left unproved in [Hil10, Theorem 3.2]. We therefore give a full proof of this commutator con- dition here: Theorem 2. Suppose that (A , E, D) is a half-closed chain from A to B. Then (E, FD) is a Kasparov module from A to B of the same parity as (A , E, D) and with the same Z/2Z-grading operator γ : E → E in the even case. Proof. We have to show that [FD, a] ∈ K(E) for all a ∈ A. Since the ∗-algebra A ⊆ A is dense in C∗-norm and since the C∗-algebra K(E) ⊆ L(E) is closed in operator norm it suffices to show that [FD, a]·b ∈ K(E) for all a, b ∈ A . We recall that (1 + D∗D)−1/2 = 1 π λ−1/2(1 + λ + D∗D)−1 dλ , (cid:90) ∞ 0 where the integral converges absolutely in operator norm and where the integrand is continuous in operator norm. Remark here that (cid:107)(1 + λ + D∗D)−1(cid:107)∞ ≤ (1 + λ)−1 for all λ ≥ 0. For a ∈ A and λ ≥ 0 we then compute that (cid:2)D(1 + λ + D∗D)−1, a(cid:3) = −DD∗(1 + λ + DD∗)−1d(a)(1 + λ + D∗D)−1 − D(1 + λ + D∗D)−1d(a)D(1 + λ + D∗D)−1 + d(a)(1 + λ + D∗D)−1 . In particular, it holds for each a, b ∈ A that the map M : (0,∞) → L(E) M (λ) := λ−1/2[D(1 + λ + D∗D)−1, a]b is continuous in operator norm and that M (λ) ∈ K(E) for all λ ∈ (0,∞). Moreover, we have the estimate (cid:107)M (λ)(cid:107)∞ ≤ λ−1/2 · (cid:107)d(a)(cid:107) · 3 · (1 + λ)−1 , (cid:90) ∞ 0 for all λ > 0. We may thus conclude that [FD, a]b = 1 π M (λ) dλ ∈ K(E) , for all a, b ∈ A . This proves the theorem. (cid:3) 6 JENS KAAD AND WALTER D. VAN SUIJLEKOM 3. Unbounded modular cycles Let us fix σ-unital C∗-algebras A and B together with a dense ∗- subalgebra A ⊆ A. The following definition is from [Kaa15, Section 3]: Definition 3. An unbounded modular cycle from A to B is a triple (E, D, ∆) where E is a countably generated C∗-correspondence from A to B, D : Dom(D) → E is an unbounded selfadjoint and regular operator, and ∆ : E → E is a bounded positive and selfadjoint operator with norm-dense image such that (1) a(i + D)−1 : E → E is a compact operator for all a ∈ A; (2) (a + λ)∆ has Dom(D) ⊆ E as an invariant submodule and D(a + λ)∆ − ∆(a + λ)D : Dom(D) → E extends to a bounded adjointable operator d∆(a, λ) : E → E for all a ∈ A , λ ∈ C. (3) The supremum (cid:107)(∆1/2 + ε)−1d∆(a, λ)(∆1/2 + ε)−1(cid:107)∞ sup ε>0 (4) The sequence {∆(∆ + 1/n)−1a} converges in operator norm to is finite for all a ∈ A , λ ∈ C. a for all a ∈ A. An unbounded modular cycle is even when E comes equipped with a Z/2Z-grading operator γ : E → E (γ = γ∗, γ2 = 1), such that [a, γ] = 0 for all a ∈ A and Dγ = −γD. An unbounded modular cycle is odd when it is not even. Remark 4. Note that if ∆ has a bounded inverse then (3) and (4) are automatic. If, in addition, A is unital, ∆, ∆−1 ∈ A and B = C then the modular cycle (E, D, ∆) defines a twisted spectral triple in the sense of [CoMo08], with the twisting automorphism σ : A → A given by σ(a) = ∆a∆−1 for all a ∈ A . Remark 5. In [Kaa15] it is assumed that A is equipped with a fixed operator space norm (cid:107) · (cid:107)1 : Mn(A ) → [0,∞), n ∈ N, such that the inclusion A → A is completely bounded. In the above definition it is then required that the supremum in (3) is completely bounded in the sense that (cid:107)(∆1/2 + ε)−1d∆(a, 0)(∆1/2 + ε)−1(cid:107)∞ ≤ C · (cid:107)a(cid:107)1 sup ε>0 for all a ∈ Mn(A ), n ∈ N (thus, the constant C is independent of the size of the matrices). This structure is relevant for the construction of the unbounded Kasparov product, but will not play a role in the present text. ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 7 As in the case of half-closed chains, each unbounded modular cycle represents an explicit class in KK-theory. This result can be found as [Kaa15, Theorem 9.1]. We state it here for the convenience of the reader. We recall that FD := D(1 + D2)−1/2 denotes the bounded transform of D : Dom(D) → E (but now D is selfadjoint and regular). Theorem 6. Suppose that (E, D, ∆) is an unbounded modular cycle from A to B. Then (E, FD) is a Kasparov module from A to B of the same parity as (E, D, ∆) and the same Z/2Z-grading operator γ : E → E in the even case. 4. Localization of regular unbounded operators Let E be a countably generated Hilbert C∗-module over a σ-unital C∗-algebra B and let D : Dom(D) → E be a closed, symmetric and regular unbounded operator. Assumption 1. It will be assumed that ∆ : E → E is a bounded selfadjoint operator such that (1) ∆(cid:0)Dom(D∗)(cid:1) ⊆ Dom(D) ; (2) D∆−∆D : Dom(D) → E extends to a bounded operator d(∆) : E → E. Remark that it follows by the above assumption and the inclusion D ⊆ D∗ that D∆ = D∗∆ D∆ − ∆D∗ : Dom(D∗) → E also has d(∆) : E → E as a bounded extension. Moreover, d(∆) : E → E is automatically adjointable with d(∆)∗ = −d(∆). Before proving our first result, we notice that D∆ : Dom(D∆) → E is a closed unbounded operator on the domain Dom(D∆) :=(cid:8)ξ ∈ E ∆(ξ) ∈ Dom(D)(cid:9) . A similar remark holds for D∗∆ : Dom(D∗∆) → E. Proposition 7. Suppose that the conditions in Assumption 1 hold. Then and D∆ : Dom(D∆) → E is a regular unbounded operator with core Dom(D) and with (D∆)∗ = D∆ − d(∆) . In particular, we have that Dom((D∆)∗) = Dom(D∆) . Proof. We first claim that the unbounded operators D∆ : Dom(D∆) → E and D∗∆ : Dom(D∗∆) → E are regular with cores Dom(D∗) and Dom(D), respectively, and with adjoints (D∆)∗ = D∆ − d(∆) and (D∗∆)∗ = D∗∆ − d(∆) . 8 JENS KAAD AND WALTER D. VAN SUIJLEKOM To prove this claim, we recall that D : Dom(D) → E is regular by assumption, and we thus have that is selfadjoint and regular. Moreover, we have that D 0 (cid:19) : Dom(D) ⊕ Dom(D∗) → E ⊕ E (cid:18) 0 D∗ (cid:19)(cid:0)Dom(D) ⊕ Dom(D∗)(cid:1) ⊆ Dom(D) ⊕ Dom(D∗) (cid:19) (cid:18) 0 ∆ (cid:19) 0 (cid:19)(cid:3) = D∆ − ∆D∗ , ∆ 0 (cid:18) D∆ − ∆D (cid:18) d(∆) 0 0 (cid:18) 0 ∆ (cid:2)(cid:18) 0 D∗ ∆ 0 D 0 and the identities = 0 (cid:19) (cid:18) 0 D∗ d(∆) (cid:19) (cid:18) 0 ∆ (cid:19) hold on Dom(D)⊕Dom(D∗). This means that ∆ 0 satisfy the conditions of [Kaa17, Section 6] and we may conclude that D 0 and (cid:18) 0 D∗ (cid:19)(cid:18) 0 ∆ (cid:19) D 0 ∆ 0 = (cid:18) D∗∆ 0 (cid:19) 0 D∆ : Dom(D∗∆) ⊕ Dom(D∆) → E ⊕ E is a regular unbounded operator with (cid:18) D∗∆ 0 (cid:19)∗ 0 D∆ = (cid:18) D∗∆ 0 (cid:18) D∗∆ 0 0 D∆ (cid:19) (cid:19) (cid:18) d(∆) − 0 0 d(∆) (cid:19) : Dom(D∗∆) ⊕ Dom(D∆) → E ⊕ E Moreover, we know that E ⊕ E has Dom(D) ⊕ Dom(D∗) as a core. This proves the claim. D∆ = D∗∆. To this end, we notice that To end the proof of the proposition, it now suffices to prove that 0 D∆ : Dom(D∗∆)⊕ Dom(D∆) → for all ξ ∈ Dom(D∗) (D∗∆)(ξ) = (D∆)(ξ) (4.1) Since Dom(D) ⊆ Dom(D∗) is a core for D∗∆ we obtain from Equation (4.1) that D∗∆ ⊆ D∆. Moreover, since Dom(D∗) is a core for D∆ we also obtain from Equation (4.1) that D∆ ⊆ D∗∆. We conclude that (cid:3) D∆ = D∗∆. Assumption 2. It will be assumed that x : E → E is a bounded adjointable operator such that (1) x(cid:0)Dom(D∗)(cid:1) ⊆ Dom(D) and x∗(cid:0)Dom(D∗)(cid:1) ⊆ Dom(D); (2) Dx − xD and Dx∗ − x∗D : Dom(D) → E extend to bounded operators d(x) and d(x∗) : E → E, respectively. ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 9 As above, d(x) and d(x∗) : E → E are automatically adjointable with d(x)∗ = −d(x∗). Moreover, d(x) and d(x∗) are bounded extensions of Dx − xD∗ and Dx∗ − x∗D∗ : Dom(D∗) → E, respectively. We define the localization of E (with respect to x : E → E) as the Hilbert C∗-submodule Ex ⊆ E given by the norm-closure of the image of x: We define ∆ := xx∗ : E → E. Ex := cl(Im(x)) . Lemma 8. Suppose that the conditions of Assumption 2 are satisfied. Then the unbounded operator D∆ − d(x)x∗ : Dom(D∆) → E is selfadjoint and regular and it has Dom(D) ⊆ Dom(D∆) as a core. Moreover, we have that (D∆ − d(x)x∗)(ξ) = (xDx∗)(ξ) , for all ξ ∈ Dom(Dx∗) ⊆ Dom(D∆). Proof. Clearly, ∆ = xx∗ : E → E satisfied the conditions of As- sumption 1 and it therefore follows from Proposition 7 that D∆ : Dom(D∆) → E is regular with core Dom(D) and that (D∆)∗ = D∆ − d(∆) = D∆ − d(x)x∗ − xd(x∗) . Since d(x)x∗ : E → E is a bounded adjointable operator, it follows by [Wor91, Section 2, Example 1] that D∆− d(x)x∗ : Dom(D∆) → E is regular. It is moreover clear that Dom(D) is also a core for D∆−d(x)x∗ and that (D∆ − d(x)x∗)∗ = (D∆)∗ − (d(x)x∗)∗ = D∆ − d(x)x∗ − xd(x∗) + xd(x∗) = D∆ − d(x)x∗ , proving that our unbounded operator is selfadjoint as well. The final (cid:3) statement of the lemma is obvious. Definition 9. Suppose that the conditions of Assumption 2 are satis- fied. We define the localization of D : Dom(D) → E (with respect to x : E → E) as the closure of the unbounded symmetric operator xDx∗ : Dom(D) ∩ Ex → Ex . The localization of D is denoted by Remark that x(cid:0)Dom(D)(cid:1) ⊆ Dom(D) ∩ Ex, implying that the local- Dx : Dom(Dx) → Ex . ization Dx is densely defined. 10 JENS KAAD AND WALTER D. VAN SUIJLEKOM Lemma 10. Suppose that the conditions of Assumption 2 are satis- fied and let r ∈ R with r > (cid:107)d(x∗)x(cid:107)∞ be given. Then ir + Dx∗x : Dom(Dx∗x) → E is a bijection and the resolvent is a bounded ad- jointable operator (ir + Dx∗x)−1 : E → E satisfying the relation (4.2) Proof. By replacing x with x∗ in Assumption 2 we see from Lemma 8 that the unbounded operator (ir + D∆ − d(x)x∗)−1x = x(ir + Dx∗x)−1 . Dx∗x − d(x∗)x : Dom(Dx∗x) → E is selfadjoint and regular. In particular, we know that the resolvent (ir + Dx∗x − d(x∗)x)−1 : E → E is a well-defined bounded adjointable operator. Since (cid:107)d(x∗)x(ir + Dx∗x − d(x∗)x)−1(cid:107)∞ ≤ (cid:107)d(x∗)x(cid:107)∞ · r−1 < 1 we may conclude that ir + Dx∗x : Dom(Dx∗x) → E is a bijection and that the resolvent is a bounded adjointable operator. In fact, we have that (ir + Dx∗x)−1 = (ir + Dx∗x − d(x∗)x)−1 ·(cid:0)1 + d(x∗)x(ir + Dx∗x − d(x∗)x)−1(cid:1)−1 . The relation in Equation Equation (4.2) now follows since (ir + D∆ − d(x)x∗)x = (ir + xDx∗)x = x(ir + Dx∗x) on Dom(Dx∗x). Proposition 11. Suppose that the conditions of Assumption 2 are sat- isfied. Then the localization of D : Dom(D) → E with respect to x : E → E is a selfadjoint and regular unbounded operator (cid:3) Dx : Dom(Dx) → Ex , (iµ + Dx)−1(ξ) = (iµ + D∆ − d(x)x∗)−1(ξ) , with core x(Dom(D)) ⊆ Dom(Dx). Moreover, we have the identity (4.3) for all ξ ∈ Ex and all µ ∈ R\{0}. In particular, Ex ⊆ E is an invariant submodule for (iµ + D∆ − d(x)x∗)−1 : E → E for all µ ∈ R \ {0}. Proof. To show that Dx : Dom(Dx) → Ex is selfadjoint and regular, it suffices to verify that ir + xDx∗ : x(cid:0)Dom(D)(cid:1) → Ex has dense image whenever r ∈ R satisfies r > (cid:107)d(x∗)x(cid:107)∞, see [Lan95, Lemma 9.7 and Lemma 9.8]. Let such an r ∈ R be given. Clearly, x∗x : E → E satisfies the condition of Assumption 1 and it therefore follows from Proposition 7 that Dx∗x : Dom(Dx∗x) → E is regular with core Dom(D) ⊆ E. Combining this with Lemma 10 we may find a norm-dense submodule E ⊆ E such that (ir + Dx∗x)−1(E ) = Dom(D) . ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 11 Moreover, we have that for all ξ ∈ E . (ir + xDx∗)x(ir + Dx∗x)−1(ξ) = x(ξ) Since x(E ) ⊆ Ex is norm-dense and x(ir + Dx∗x)−1(E ) = x(cid:0)Dom(D)(cid:1), the image of iµ+xDx∗ : x(cid:0)Dom(D)(cid:1) → Ex, but here it follows immedi- ately since (xDx∗)(ξ) = (D∆ − d(x)x∗)(ξ) for all ξ ∈ x(cid:0)Dom(D)(cid:1). (cid:3) this proves the desired density result and hence that the localization Dx : Dom(Dx) → Ex is selfadjoint and regular. Let µ ∈ R\{0}. The identity in Equation (4.3) can now be verified on Remark 12. The result of Proposition 11 can be generalized by re- placing the bounded adjointable operator x : E → E by a sequence of bounded adjointable operators xn : E → E, n ∈ N, each of them satisfying the conditions of Assumption 2. Suppose moreover that the sums xnx∗ n and d(xn)d(xn)∗ ∞(cid:88) n=1 ∞(cid:88) n=1 ∞(cid:88) are norm-convergent in L(E) (this can of course always be obtained by rescaling the operators xn : E → E, n ∈ N). sequence x = {xn} as the closed submodule In this context, we define the localization of E with respect to the Ex := cl(cid:0)spanC{xn(ξ) n ∈ N, ξ ∈ E}(cid:1) ⊆ E . The localization Dx of D : Dom(D) → E is defined as the closure of the symmetric unbounded operator xnDx∗ n : Dom(D) ∩ Ex → Ex . n=1 As in Proposition 11, we then obtain that Dx : Dom(Dx) → Ex is a selfadjoint and regular unbounded operator. 5. Localization of half-closed chains Let A and B be σ-unital C∗-algebras. Throughout this section (A , E, D) will be a half-closed chain from A to B. We denote by φ : A → L(E) the ∗-homomorphism that provides the left action of A on E. Moreover, x ∈ A will be a fixed element. Notice that φ(x) : E → E satisfies the condition of Assumption 2 with respect to the symmetric and regular unbounded operator D : Dom(D) → E. Recall then that the localization of E is the norm- Dx of D : Dom(D) → E is the closure of the symmetric unbounded operator closed submodule Ex := cl(cid:0)Im(φ(x))(cid:1) ⊆ E and that the localization φ(x)Dφ(x∗) : Dom(D) ∩ Ex → Ex . 12 JENS KAAD AND WALTER D. VAN SUIJLEKOM By Proposition 11, the localization Dx : Dom(Dx) → Ex is selfadjoint and regular. We put By definition, the localization of A with respect to x ∈ A is the hereditary C∗-subalgebra of A defined by ∆ := xx∗ ∈ A . Ax := cl(xAx∗) ⊆ A . The ∗-homomorphism φ : A → L(E) restricts to a ∗-homomorphism φx : Ax → L(Ex) and in this way Ex becomes a C∗-correspondence from Ax to B. We remark that ∆ ∈ Ax and that φx(∆) : Ex → Ex is a bounded positive and selfadjoint operator with norm-dense image. We define the ∗-subalgebra Ax ⊆ Ax as the intersection Ax := A ∩ Ax . Remark that Ax ⊆ Ax is automatically norm-dense. When the half-closed chain (A , E, D) is even with Z/2Z-grading operator γ : E → E, then Ex can be equipped with the Z/2Z-grading operator γEx : Ex → Ex obtained by restriction of γ : E → E. We are going to prove the following: Theorem 13. Suppose that (A , E, D) is a half-closed chain and that x is an element in A . Then the triple (Ex, Dx, φx(∆)) is an unbounded modular cycle from Ax to B of the same parity as (A , E, D) and with grading operator γEx : Ex → Ex in the even case. Proof. Clearly the C∗-correspondence Ex is countably generated (since E is countably generated by assumption). Moreover, we have already established that the unbounded operator Dx : Dom(Dx) → Ex is self- adjoint and regular in Proposition 11 and that φx(∆) : Ex → Ex is bounded positive and selfadjoint with norm-dense image. So it only remains to check conditions (1), (2), (3) and (4) of Definition 3. The last condition (4) follows immediately since ∆(∆ + 1/n)−1a → a in C∗-norm for all a ∈ Ax. The remaining three conditions are proved in (cid:3) Proposition 15, Proposition 16 and Proposition 17 below. We will refer to the unbounded modular cycle (Ex, Dx, φx(∆)) as the localization of the half-closed chain (E, φ, D) with respect to x ∈ A . We start by proving the compactness condition (1) of Definition 3. We put (cid:102)Dx := Dφ(∆) − d(x)φ(x∗) : Dom(Dφ(∆)) → E and recall that (cid:102)Dx is a selfadjoint and regular unbounded operator by Lemma 8. We remark that (cid:102)Dx agrees with Dx if and only if the image of φ(x) : E → E is norm-dense. In fact, when the image of φ(x) is not norm-dense then these two unbounded operators do not even act on the same Hilbert C∗-module. Lemma 14. We have the resolvent identity Proof. It suffices to notice that the identities (cid:18) 0 φ(∆) 0 = ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 13 0 0 − φ(∆) D i (cid:33) (i +(cid:102)Dx)−1 (cid:19)(cid:32) (i +(cid:102)Dx)−1 (cid:18) i D∗ (cid:19)−1(cid:18) d(x)φ(x∗) − i (cid:19) (cid:19)(cid:18) 0 (cid:18) i D∗ i +(cid:102)Dx (cid:32) Dφ(∆) − i −(cid:102)Dx Dφ(∆) − i −(cid:102)Dx (cid:18) d(x)φ(x∗) − i (cid:19) (cid:32) iφ(∆) iφ(∆) iφ(∆) iφ(∆) D i φ(∆) φ(∆) − 0 0 d(x)φ(x∗) − i iφ(∆) iφ(∆) d(x)φ(x∗) − i = = (cid:18) i D∗ (cid:19) D i (cid:19)−1 (i +(cid:102)Dx)−1 . (cid:33) 0 i +(cid:102)Dx (cid:33) hold on Dom(Dφ(∆)) ⊕ Dom(Dφ(∆)). Recall in this respect that (cid:3) Dφ(∆) = D∗φ(∆) by Proposition 7. Proposition 15. The bounded adjointable operator φx(a)(i + Dx)−1 : Ex → Ex is compact for all a ∈ Ax. Proof. Notice that ∆ ∈ Ax and that the left ideal Ax · ∆ ⊆ Ax is norm-dense. It thus suffices to show that φx(∆) · (i + Dx)−1 ∈ K(Ex). We apply the notation K(E, Ex) ⊆ K(E) for the closed right ideal generated by all compact operators on E of the form ξ(cid:105)(cid:104)η with ξ ∈ Ex and η ∈ E. Similarly, we let K(Ex, E) ⊆ K(E) denote the closed left ideal generated by all compact operators of the form η(cid:105)(cid:104)ξ for ξ ∈ Ex and η ∈ E. We remark that K(Ex, E) = K(E, Ex)∗. Since (E, φ, D) is a half-closed chain we know that (cid:19)(cid:18) i D∗ (cid:19)−1 ∈ K(E ⊕ E) 0 (cid:18) φ(∆) and it therefore follows from Lemma 14 that 0 φ(∆) D i φ(∆)2(i +(cid:102)Dx)−1 ∈ K(E, Ex) . Since (φ(∆) + 1/n)−1φ(∆)2 → φ(∆) as n → ∞ this implies that also φ(∆)(i+(cid:102)Dx)−1 ∈ K(E, Ex) and thus that (−i+(cid:102)Dx)−1φ(∆) ∈ K(Ex, E). We may thus conclude that φ(∆)(1+(cid:102)Dx )−1φ(∆) ∈ K(E, Ex)·K(Ex, E) restricts to a compact operator on the Hilbert C∗-module Ex ⊆ E. But this proves the present proposition since we have from Proposition 11 that 2 x)−1φx(∆) =(cid:0)φ(∆)(1 +(cid:102)Dx )−1φ(∆)(cid:1)Ex . 2 φx(∆)(1 + D2 (cid:3) 14 JENS KAAD AND WALTER D. VAN SUIJLEKOM We continue by proving the twisted commutator condition (2) of Definition 3. Proposition 16. Let a ∈ Ax, λ ∈ C. Then (φx(a) + λ)φx(∆) : Ex → Ex has Dom(Dx) ⊆ Ex as an invariant submodule and Dx(φx(a) + λ)φx(∆) − φx(∆)(φx(a) + λ)Dx : Dom(Dx) → Ex extends to a bounded adjointable operator d∆(a, λ) : Ex → Ex. In fact we have that d∆(a, λ) =(cid:0)φ(x)d(x∗(a + λ)x)φ(x∗)(cid:1)Ex . Proof. Let ξ ∈ Dom(D) ∩ Ex. We then have that (φx(a) + λ)φx(∆)(ξ) ∈ Dom(D) ∩ Ex and that Dx(φx(a) + λ)φx(∆)(ξ) − φx(∆)(φx(a) + λ)Dx(ξ) = φ(x)Dφ(x∗)(φ(a) + λ)φ(xx∗)(ξ) − φ(xx∗)(φ(a) + λ)φ(x)Dφ(x∗)(ξ) = φ(x)d(x∗(a + λ)x)φ(x∗)(ξ) . Since Dom(D) ∩ Ex is a core for the localization Dx : Dom(Dx) → Ex, (cid:3) this proves the proposition. We finally prove the supremum condition (3) of Definition 3. Proposition 17. Let a ∈ Ax, λ ∈ C. Then we have that (cid:107)(φx(∆)1/2 + ε)−1d∆(a, λ)(φx(∆)1/2 + ε)−1(cid:107)∞ < ∞ . sup ε>0 Proof. This follows immediately from Proposition 16. Indeed, the op- erator norm of (φx(∆)1/2 + ε)−1φ(x) : E → Ex (cid:3) is bounded by 1 for all ε > 0. Remark 18. One may equip Ax with the operator space norm (cid:107) · (cid:107)1 : Mn(Ax) → [0,∞), n ∈ N, defined by (cid:107)a(cid:107)1 := sup{(cid:107)a(cid:107),(cid:107)d(a)(cid:107)∞} for all a ∈ Mn(Ax) , where the norms inside the supremum are the C∗-norm on Mn(A) and the operator-norm on L(E⊕n), respectively. Clearly, the inclusion Ax → Ax is then completely bounded. It is moreover possible to find a constant C > 0 such that (cid:107)(φx(∆)1/2 + ε)−1d∆(a, 0)(φx(∆)1/2 + ε)−1(cid:107)∞ ≤ C · (cid:107)a(cid:107)1 , sup ε>0 for all a ∈ Mn(Ax). Cf. Remark 5. ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 15 6. Localization as an unbounded Kasparov product In this section we continue under the conditions spelled out in the beginning of Section 5. We thus have a half-closed chain (A , E, D) and an element x ∈ A . The element x ∈ A provides us with a closed right ideal Ix ⊆ A defined as the norm-closure: Ix := cl(xA) . In particular, we may consider Ix as a countably generated Hilbert C∗- module over A. The hereditary C∗-subalgebra Ax = cl(xAx∗) ⊆ A can be identified with the compact operators on Ix via the ∗-homomorphism ψ : Ax → L(Ix) induced by the multiplication in A. We thus obtain an even Kasparov module (Ix, 0) from Ax to A with corresponding class [Ix, 0] ∈ KK0(Ax, A) in KK-theory. Moreover, by Theorem 2, our half-closed chain (A , E, D) (of par- ity p ∈ {0, 1}) yields a Kasparov module (E, FD) from A to B with corresponding class [E, FD] ∈ KKp(A, B). Finally, the unbounded modular cycle (A ∩Ax, Ex, φx(∆)) constructed in Section 5 yields a Kasparov module (Ex, FDx) from Ax to B with corresponding class [Ex, FDx] ∈ KKp(Ax, B), see Theorem 6. In this section we will prove the following theorem: Theorem 19. Suppose that (A , E, D) is a half-closed chain, that x ∈ A and that Ax is separable. Then we have the identity [Ex, FDx] = [Ix, 0](cid:98)⊗A[E, FD] in KKp(Ax, B), where (cid:98)⊗A : KK0(Ax, A)× KKp(A, B) → KKp(Ax, B) to the interior tensor product of C∗-correspondences Ix(cid:98)⊗φE (via the unitary isomorphism xa(cid:98)⊗ξ (cid:55)→ φ(xa)(ξ)). For each a ∈ A, we define denotes the Kasparov product. Proof. The C∗-correspondence Ex from Ax to A is unitarily isomorphic the bounded adjointable operator Txa : E → Ex by ξ (cid:55)→ φ(xa)(ξ). By [CoSk84, Theorem A.3] it suffices to prove the connection condition, thus that FDxTxa − TxaFD , FDxTxa − TxaFD∗ ∈ K(E, Ex) (6.1) for all a ∈ A. Indeed, the positivity condition of [CoSk84, Theorem A.3] is obviously satisfied since the bounded adjointable operator in the Kasparov module (Ix, 0) from Ax to A is trivial. See also Section 7 for more details. However, since Txa = Txφ(a) : E → Ex and φ(a)(FD − FD∗) ∈ K(E) it suffices to prove the first of these inclusions. This proof will occupy (cid:3) the remainder of this section, see Proposition 26. 16 JENS KAAD AND WALTER D. VAN SUIJLEKOM Remark 20. In the case where xA ⊆ A is norm-dense and A is sep- arable, we have that (Ix, 0) = (A, 0) and it therefore follows from the above theorem that the two Kasparov modules (Ex, FDx) and (E, FD) represents the same class in KKp(A, B). 6.1. The modular transform. We continue working under the gen- eral assumptions stated in the beginning of Section 5. We recall that ∆ := xx∗. We will in the following suppress the ∗-homomorphism φx : Ax → L(Ex). For each λ ≥ 0, we introduce the notation Rx(λ∆2) := (1 + λ∆2 + D2 Rx(λ) := (1 + λ + D2 x)−1 ∈ L(Ex) . x)−1 ∈ L(Ex) In general, we are not able to estimate the norm of Rx(λ∆2) from above by (1 + λ)−1 since ∆ : Ex → Ex may have zero in the spectrum. Instead, we recall the following basic estimate from [Kaa15, Section 11]: (6.2) (cid:107)∆Rx(λ∆2)∆(cid:107)∞ ≤ 2 (1 + λ) ∀λ ≥ 0 . The next definition is from [Kaa15, Section 8]: Definition 21. The modular transform of the unbounded modular cy- cle (Ex, Dx, ∆) is the unbounded operator G(Dx,∆) : ∆(Dom(Dx)) → Ex defined by (6.3) G(Dx,∆) : η (cid:55)→ 1 π (cid:90) ∞ 0 λ−1/2∆(1 + λ∆2 + D2 x)−1Dx(η) dλ . We remark that G(Dx,∆) : ∆(Dom(Dx)) → Ex is well-defined. In- deed, for η = ∆(ξ) with ξ ∈ Dom(Dx) we have from Proposition 16 that ∆Rx(λ∆2)Dx(η) = ∆Rx(λ∆2)∆Dx(ξ) + ∆Rx(λ∆2)xd(x∗x)x∗(ξ) . Using the estimate from Equation (6.2), we may thus find a constant C > 0 such that (cid:107)∆(1 + λ∆2 + D2 x)−1Dx(η)(cid:107) ≤ C · (1 + λ)−3/4 ∀λ ≥ 0 , implying that the integral in Equation (6.3) converges absolutely in the norm on Ex. The following result is a consequence of [Kaa15, Theorem 8.1]: Theorem 22. The difference FDx∆6 − G(Dx,∆)∆6 : Dom(Dx) → Ex extends to a compact operator on Ex. ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 17 Notice that the above result implies that the unbounded operator G(Dx,∆)∆6 : Dom(Dx) → Ex extends to a bounded adjointable operator on Ex. 6.2. The connection condition. We will continue working under the assumptions of Section 5. We recall from Lemma 8 that is a selfadjoint and regular unbounded operator and we put (cid:102)Dx := Dφ(∆) − d(x)φ(x∗) : Dom(Dφ(∆)) → E (cid:102)Rx(λφ(∆2)) := (1 + λφ(∆2) + ((cid:102)Dx)2)−1 ∈ L(E) R(λ) := (1 + λ + D∗D)−1 ∈ L(E) , for all λ ≥ 0. Lemma 23. For each λ ≥ 0, we have the identity R(λ) −(cid:102)Rx(λφ(∆2))φ(∆2) =(cid:102)Rx(λφ(∆2))(cid:0)1 − φ(∆2) + φ(x)d(x∗xx∗)D(cid:1)R(λ) +(cid:0)(cid:102)Dx(cid:102)Rx(λφ(∆2))(cid:1)∗ 1 −(cid:102)Rx(λφ(∆2))φ(∆2)(1 + λ + D∗D) of bounded adjointable operators on E. Proof. We have the identities = 1 −(cid:102)Rx(λφ(∆2))(cid:0)1 + λφ(∆2) + φ(x)Dφ(x∗xx∗)D(cid:1) +(cid:102)Rx(λφ(∆2))(1 − φ(∆2) + φ(x)d(x∗xx∗)D) =(cid:0)(cid:102)Dx(cid:102)Rx(λφ(∆2))(cid:1)∗ +(cid:102)Rx(λφ(∆2))(1 − φ(∆2) + φ(x)d(x∗xx∗)D) φ(x)d(x∗) φ(x)d(x∗)R(λ) on Dom(D∗D). But this proves the lemma after multiplying with (cid:3) R(λ) = (1 + λ + D∗D)−1 from the right. For each y ∈ Ix = cl(xA), we recall that Ty : E → Ex denotes the bounded adjointable operator Ty : ξ (cid:55)→ φ(y)(ξ). Notice then that it follows from Proposition 11 that T∆(cid:102)Rx(λφ(∆2))φ(∆) = ∆Rx(λ∆2)T∆ : E → Ex . Lemma 24. The difference T∆R(λ)Dφ(∆) − ∆Rx(λ∆2)DxT∆2 : Dom(Dφ(∆)) → Ex extends to a compact operator Mλ : E → Ex for all λ ≥ 0. Moreover, there exists a constant C > 0 such that (cid:107)Mλ(cid:107)∞ ≤ C · (1 + λ)−3/4 ∀λ ≥ 0 . 18 JENS KAAD AND WALTER D. VAN SUIJLEKOM Proof. Since (A , E, D) is a half-closed chain and (Ex, Dx, ∆) is an un- bounded modular cycle we obtain that the difference T∆R(λ)Dφ(∆) − ∆Rx(λ∆2)DxT∆2 : Dom(Dφ(∆)) → Ex extends to a compact operator Mλ : E → Ex for all λ ≥ 0. Indeed, this is already true for each of the terms viewed separately. So we only need to prove the norm-estimate. To this end, we let ξ ∈ Dom(Dφ(∆)) and compute that (T∆R(λ)Dφ(∆) − ∆Rx(λ∆2)DxT∆2)(ξ) = T∆R(λ)Dφ(∆)(ξ) − T∆(cid:102)Rx(λφ(∆2))φ(x)Dφ(x∗∆2)(ξ) = T∆R(λ)Dφ(∆)(ξ) − T∆(cid:102)Rx(λφ(∆2))φ(∆2)Dφ(∆)(ξ) − T∆(cid:102)Rx(λφ(∆2))φ(x)d(x∗xx∗)φ(∆)(ξ) . Since (cid:107)T∆(cid:102)Rx(λφ(∆2))φ(x)(cid:107)∞ ≤ 23/4 · (1 + λ)−3/4 by the estimate in Equation (6.2) we may focus on the difference T∆R(λ)Dφ(∆)(ξ) − T∆(cid:102)Rx(λφ(∆2))φ(∆2)Dφ(∆)(ξ) . = T∆(cid:102)Rx(λφ(∆2))(cid:0)1 − φ(∆2) + φ(x)d(x∗xx∗)D(cid:1)R(λ)Dφ(∆)(ξ) T∆R(λ)Dφ(∆)(ξ) − T∆(cid:102)Rx(λφ(∆2))φ(∆2)Dφ(∆)(ξ) However, using Lemma 23 we get that (cid:0)(cid:102)Dx(cid:102)Rx(λφ(∆2))(cid:1)∗ + T∆ φ(x)d(x∗)R(λ)Dφ(∆)(ξ) . The result of the lemma then follows from the basic estimate (cid:107)DR(λ)(cid:107)∞ ≤ (1 + λ)−1/2 and the estimate in Equation (6.2) a few times. (cid:3) Proposition 25. The difference T∆2FD − G(Dx,∆)T∆2 : Dom(D) → Ex extends to a compact operator from E to Ex. Proof. Since φ(∆)FD − FD∗φ(∆) : E → E is compact, we only need to show that T∆FD∗φ(∆) − G(Dx,∆)T∆2 : Dom(D) → Ex extends to a compact operator from E to Ex. Now, recall that T∆FD∗φ(∆)(ξ) = 1 π λ−1/2T∆(1 + λ + D∗D)−1Dφ(∆)(ξ) dλ (cid:90) ∞ 0 ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 19 for all ξ ∈ Dom(D). The result of the proposition now follows by Lemma 24 since T∆D∗(1 + DD∗)−1/2φ(∆)(ξ) − G(Dx,∆)T∆2(ξ) λ−1/2(cid:0)T∆(1 + λ + D∗D)−1Dφ(∆) − ∆Rx(λ∆2)DxT∆2 (cid:1)(ξ) dλ (cid:90) ∞ (cid:90) ∞ 0 = = 1 π 1 π 0 λ−1/2Mλ(ξ) dλ . (cid:3) Remark that it follows from the above proposition that the un- bounded operator G(Dx,∆)T∆2 : Dom(D) → Ex extends to a bounded adjointable operator on Ex. Proposition 26. The difference FDxTxa − TxaFD : E → Ex is a compact operator for all a ∈ A. Proof. Since [φ(b), FD] ∈ K(E) for all b ∈ A and since ∆7(1/n + ∆7)−1x → x in the norm on A, it suffices to show that FDxT∆7 − T∆7FD : E → Ex is a compact operator. But now Proposition 25 and Theorem 22 imply that the following identities hold modulo K(E, Ex): FDxT∆7 − T∆7FD ∼ FDxT∆7 − T∆2FDφ(∆5) ∼ FDxT∆7 − cl(G(Dx,∆)T∆2)φ(∆5) = FDx∆6T∆ − cl(G(Dx,∆)∆6)T∆ ∼ 0 . (cid:3) 7. Kucerovsky's theorem Let us fix three C∗-algebras A, B and C with A separable and B and C both σ-unital. Throughout this section we will assume that (A , E1, D1), (B, E2, D2) and (A , E, D) are even half-closed chains from A to B, from B to C and from A to C, respectively. We denote the associated ∗-homomorphisms by φ1 : A → L(E1), φ2 : B → L(E2) and φ : A → L(E) and the Z/2Z-grading operators by γ1 : E1 → E1, γ2 : E2 → E2 and γ : E → E, respectively. We will moreover as- the C∗-correspondences E1 and E2. sume that E := E1(cid:98)⊗φ2E2 agrees with the interior tensor product of φ(a) = φ1(a)(cid:98)⊗1 for all a ∈ A and that γ = γ1(cid:98)⊗γ2. We will denote the bounded transforms of our half-closed chains by (E1, FD1), (E2, FD2) and (E, FD) and the corresponding classes in KK-theory by [E1, FD1] ∈ KK0(A, B), [E2, FD2] ∈ KK0(B, C) and In particular, we assume that 20 JENS KAAD AND WALTER D. VAN SUIJLEKOM [E, FD] ∈ KK0(A, C). We may then form the interior Kasparov prod- uct [E1, FD1](cid:98)⊗B[E2, FD2] ∈ KK0(A, C) and it becomes a highly relevant question to find an explicit formula for this class in KK0(A, C). In this section we shall find conditions on the half-closed chains (A , E1, D1), (B, E2, D2) and (A , E, D) entailing that the identity [E, FD] = [E1, FD1](cid:98)⊗B[E2, FD2] holds in KK0(A, C). This kind of theorem was first proved by Kucerovsky in [Kuc97] under the stronger assumption that the half-closed chains (A , E1, D1), (B, E2, D2) and (A , E, D) were in fact unbounded Kas- parov modules. Thus under the strong assumption that all the in- volved symmetric and regular unbounded operators were in fact self- adjoint. As in the case of Kucerovsky's theorem we rely on the work of Connes and Skandalis for computing the interior Kasparov product, see [CoSk84]. is positive in the Calkin algebra L(E)/K(E) for all a ∈ A. The condition in Equation (7.1) is often referred to as the connection condition and the condition in Equation (7.2) is referred to as the positivity condition. Before we state our conditions on half-closed chains we recall that the odd symmetric and regular unbounded operator D1 : Dom(D1) → E1 can be promoted to an odd symmetric and regular unbounded operator D1(cid:98)⊗1 : Dom(D1(cid:98)⊗1) → E1(cid:98)⊗φ2E2 with resolvent (1 + D∗ L(E1(cid:98)⊗φ2E2). 1D1)−1(cid:98)⊗1 ∈ We now introduce the analogues for the above connection and posi- tivity condition for half-closed chains. They will be shown in Theorem 34 below to indeed correspond to the above two conditions for Kasparov modules. Definition 27. Given three even half-closed chains (A , E1, D1), (B, E2, D2) and (A , E1(cid:98)⊗φ2E2, D) as above, the connection condition We recall from [CoSk84, Theorem A.3] that an even Kasparov mod- ule (E, F ) from A to C is the Kasparov product of the even Kasparov modules (E1, F1) and (E2, F2) from A to B and from B to C, respec- tively, when the following holds: • For every homogeneous ξ ∈ E1 we have that F Tξ − (−1)∂ξTξF2 , F ∗Tξ − (−1)∂ξTξF ∗ • E = E1(cid:98)⊗φ2E2, φ = φ1(cid:98)⊗1. where Tξ : E2 → E is defined by Tξ(y) := ξ(cid:98)⊗η for all η ∈ E2 (cid:0)(F1(cid:98)⊗1)∗ · F ∗ + F · (F1(cid:98)⊗1)(cid:1) · φ(a∗a) + ν · φ(a∗a) and where ∂ξ ∈ {0, 1} denotes the degree of ξ ∈ E1. • There exists a ν < 2 such that (7.1) (7.2) 2 ∈ K(E2, E) , ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 21 demands that there exist a dense B-submodule E1 ⊆ E1 and cores E2 and E for D2 : Dom(D2) → E2 and D : Dom(D) → E, respectively, such that (a) For each ξ ∈ E1: Tξ(E2) ⊆ Dom(D) , (b) For each homogeneous ξ ∈ E1, the graded commutator ξ (E ) ⊆ Dom(D2) , T ∗ γ1(ξ) ∈ E1 . DTξ − (−1)∂ξTξD2 : E2 → E extends to a bounded operator Lξ : E2 → E. (B, E2, D2) and (A , E1(cid:98)⊗φ2E2, D) as above, a localizing subset is a Definition 28. Given three even half-closed chains (A , E1, D1), countable subset Λ ⊆ A with Λ = Λ∗ such that (a) The subspace Λ · A := spanC{x · a x ∈ Λ , a ∈ A} ⊆ A is norm-dense. (b) The commutator is trivial for all x ∈ Λ. (c) We have the domain inclusion [D1(cid:98)⊗1, φ(x)] : Dom(D1(cid:98)⊗1) → E Dom(D) ∩ Im(φ(x∗x)) ⊆ Dom(D1(cid:98)⊗1) , for all x ∈ Λ. (B, E2, D2) and (A , E1(cid:98)⊗φ2E2, D) and a localizing subset Λ ⊆ A , the Definition 29. Given three even half-closed chains (A , E1, D1), local positivity condition requires that for each x ∈ Λ, there exists a constant κx > 0 such that (cid:10)(D1(cid:98)⊗1)φ(x∗)ξ, Dφ(x∗)ξ(cid:11) + (cid:104)Dφ(x∗)ξ, (D1(cid:98)⊗1)φ(x∗)ξ(cid:105) ≥ −κx · (cid:104)ξ, ξ(cid:105) , for all ξ ∈ Im(φ(x)) ∩ Dom(Dφ(x∗)). Note that the local positivity condition makes sense because of (d) Indeed, for each ξ ∈ Im(φ(x)) ∩ Dom(Dφ(x∗)) we in Definition 28. have that φ(x∗)ξ ∈ Im(φ(x∗x)) ∩ Dom(D) ⊆ Dom(D1(cid:98)⊗1) . Remark 30. Suppose that A ⊆ A is unital and that φ1(A) · E1 ⊆ E1 is norm-dense. Then the half-closed chains (A , E, D) and (A , E1, D1) are in fact unbounded Kasparov modules (thus D = D∗ and D1 = D∗ 1). The choice Λ := {1} ⊆ A automatically satisfies the conditions (a) and (b) for a localizing subset in Definition 28 and the last condition (c) amounts to the requirement Dom(D) ⊆ Dom(D1(cid:98)⊗1) . 22 JENS KAAD AND WALTER D. VAN SUIJLEKOM Moreover, in this case, the local positivity condition in Definition 29 means that there exists a constant κ > 0 such that (cid:10)(D1(cid:98)⊗1)ξ, Dξ(cid:11) +(cid:10)Dξ, (D1(cid:98)⊗1)ξ(cid:11) ≥ −κ · (cid:104)ξ, ξ(cid:105) , for all ξ ∈ Dom(D). Finally, the connection condition in Definition 27 can be seen to be equivalent to the connection condition applied by Kucerovsky in [Kuc97]. In this setting, we therefore recover the as- sumptions applied by Kucerovsky in [Kuc97, Theorem 13] (except that the domain condition in [Kuc97, Theorem 13] is marginally more flex- ible). The corresponding special case of Theorem 34 here below, is therefore in itself an improvement to [Kuc97, Theorem 13] because of the extra flexibility in the choice of localizing subset Λ ⊆ A (if one is willing to disregard the minor domain issue mentioned earlier in this remark). We record the following convenient lemma, which can be proved by standard techniques: Lemma 31. Suppose that the connection condition of Definition 27 holds. Then the connection condition holds for E2 := Dom(D2) and E := Dom(D). Moreover, Lξ : E2 → E is adjointable with (Lξ)∗(η) = (T ∗ ξ D − (−1)∂ξD2T ∗ whenever ξ ∈ E1 is homogeneous. ξ )(η) ∀η ∈ Dom(D) The next lemma provides a convenient sufficient condition for veri- fying the inequality in Definition 29: Lemma 32. Let x ∈ A and suppose that Im(φ(x∗x)) ∩ Dom(D) ⊆ Dom(D1(cid:98)⊗1) and that there exists a constant κx > 0 such that (cid:104)(D1(cid:98)⊗1)η, Dη(cid:105) + (cid:104)Dη, (D1(cid:98)⊗1)η(cid:105) ≥ −κx(cid:104)η, η(cid:105) , (cid:10)(D1(cid:98)⊗1)φ(x∗)ξ, Dφ(x∗)ξ(cid:11) +(cid:10)Dφ(x∗)ξ, (D1(cid:98)⊗1)φ(x∗)ξ(cid:11) for all η ∈ Im(φ(x∗x)) ∩ Dom(D). Then we have that ≥ −(cid:107)φ(x)(cid:107)2κx(cid:104)ξ, ξ(cid:105) , for all ξ ∈ Im(φ(x)) ∩ Dom(Dφ(x∗)). Proof. This follows immediately since −κx(cid:104)φ(x∗)ξ, φ(x∗)ξ(cid:105) ≥ −(cid:107)φ(x)(cid:107)2κx(cid:104)ξ, ξ(cid:105) ∀ξ ∈ E . (cid:3) The next lemma is straightforward to prove by rescaling the elements in Λ by elements in (0,∞). It will nonetheless play a very important role: Lemma 33. Suppose that the local positivity condition of Definition 29 holds with localizing subset Λ ⊆ A . Then we may rescale the elements ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 23 in Λ and obtain a localizing subset Λ(cid:48) ⊆ A such that the local positivity condition of Definition 29 holds with the additional requirement that κx = 1/4 and (cid:107)d(x∗)φ(x)(cid:107)∞ < 1 ∀x ∈ Λ(cid:48) . Theorem 34. Suppose that the three even half-closed chains (A , E1, D1), (B, E2, D2) and (A , E1(cid:98)⊗φ2E2, D) satisfy the connection condition and the local positivity condition. Then (E, FD) is the Kasparov product of (E1, FD1) and (E2, FD2). In particular we have the identity [E, FD] = [E1, FD1](cid:98)⊗B[E2, FD2] in the KK-group KK0(A, C). Proof. Without loss of generality we may assume that κx = 1/4 and that (cid:107)d(x∗)φ(x)(cid:107)∞ < 1 for all x ∈ Λ. We need to prove the connection condition in Equation (7.1) and the positivity condition in Equation (7.2) for the even Kasparov modules (E, FD), (E1, FD1) and (E2, FD2). But these two conditions are proved in Proposition 35 and Proposi- tion 43 below, respectively. The positivity condition will be satisfied with ν = 1 = 4 · κx. (cid:3) 7.1. The connection condition. We continue working in the setting explained in the beginning of Section 7. Before proving our first proposition on the connection condition in Equation (7.1), it will be convenient to introduce some extra notation. For λ ∈ [0,∞), define the bounded adjointable operators R(λ) := (1 + λ + D∗D)−1 , R(λ) := (1 + λ + DD∗)−1 : E → E 2)−1 : E2 → E2 . R2(λ) := (1 + λ + D∗ Proposition 35. Suppose that the connection condition of Definition 27 holds. Then we have that 2D2)−1 , R2(λ) := (1 + λ + D2D∗ FDTξ − (−1)∂ξTξFD2 , F ∗ DTξ − (−1)∂ξTξF ∗ D2 ∈ K(E2, E) , for all homogeneous ξ ∈ E1. Proof. Without loss of generality we may assume that ξ = η · b1b2 with η ∈ E1 homogeneous and b1, b2 ∈ B. Using Lemma 31 we compute as follows, for each λ ∈ [0,∞): R(λ)Tη·b1 − Tη·b1R2(λ) = R(λ)Tη·b1D∗ 2D2R2(λ) − D∗DR(λ)Tη·b1R2(λ) = −R(λ)Tη · d2(b1) · D2R2(λ) − (−1)∂ηR(λ)Lη · φ2(b1) · D2R2(λ) + (−1)∂ηR(λ)DTη·b1 · D2R2(λ) − D∗DR(λ)Tη·b1R2(λ) = −R(λ)(cid:0)Tη · d2(b1) + (−1)∂ηLη · φ2(b1)(cid:1) · D2R2(λ) − D∗R(λ)Lη·b1R2(λ) , 24 JENS KAAD AND WALTER D. VAN SUIJLEKOM where d2(b1) : E2 → E2 is the bounded extension of the commutator 2) → E. In particular, we may find a D2φ2(b1) − φ2(b1)D∗ constant C > 0 such that (cid:13)(cid:13)DR(λ)Tη·b1 − DTη·b1R2(λ)(cid:13)(cid:13)∞ ≤ C · (1 + λ)−1 , 2 : Dom(D∗ (7.3) for all λ ≥ 0. We now use the integral formulae (cid:90) ∞ (cid:90) ∞ 0 0 FD = FD2 = 1 π 1 π D · D2 · λ−1/2R(λ) dλ λ−1/2R2(λ) dλ = 1 π for the bounded transforms. Indeed, using Lemma 31 one more time, these formulae allow us to compute that (cid:90) ∞ (7.4) FDTξ = FDTη·b1 · φ2(b2) (cid:90) ∞ D · Tη·b1 · λ−1/2(cid:0)R(λ)Tη·b1 − Tη·b1R2(λ)(cid:1) · φ2(b2) dλ D · (cid:90) ∞ λ−1/2D ·(cid:0)R(λ)Tη·b1 − Tη·b1R2(λ)(cid:1) · φ2(b2) dλ . λ−1/2Lη·b1 · R2(λ) · φ2(b2) dλ = (−1)∂ξTη·b1FD2 · φ2(b2) + λ−1/2R2(λ) · φ2(b2) dλ (cid:90) ∞ 1 π + 0 0 + 0 1 π 1 π 0 The fact that D2R2(λ)φ2(b2) and R2(λ)φ2(b2) ∈ K(E2), for all λ ∈ [0,∞), combined with the estimate in Equation (7.3) now imply that both of the integrals on the right hand side of Equation (7.4) converge absolutely to elements in K(E2, E) (remark that the integrands also depend continuously on λ ∈ (0,∞) with respect to the operator norm). We thus conclude that FDTξ − (−1)∂ξTη·b1FD2 · φ2(b2) ∈ K(E2, E) . Since [FD2, φ2(b2)] ∈ K(E2) we have proved that FDTξ−(−1)∂ξTξFD2 ∈ K(E2, E). D2 ∈ K(E2, E) as (cid:3) A similar argument shows that F ∗ DTξ − (−1)∂ξTξF ∗ well. 7.2. Localization. Throughout this subsection the conditions stated in the beginning of Section 7 are in effect. We are now going to apply the localization results obtained in Section 4, 5 and 6. Recall from Definition 9 and Proposition 11 that whenever x ∈ A , then the localization Dx : Dom(Dx) → Ex is the selfadjoint and regular unbounded operator defined as the closure of φ(x)Dφ(x∗) : Dom(D) ∩ Ex → Ex , ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS where Ex := cl(cid:0)Im(φ(x))(cid:1) ⊆ E. The core idea is to replace the bounded transform of D : Dom(D) → E by the bounded transforms of suffi- ciently many localizations Dx : Dom(Dx) → Ex, when verifying the positivity condition in Equation (7.2). The precise result is given here: 25 x T ∗ Proposition 36. Suppose that conditions (a) and (b) of Definition 28 hold for some localizing subset Λ ⊆ A and that ν ∈ R is given. Suppose moreover that (cid:0)(F ∗ D1(cid:98)⊗1)Ex · FDx + FDx · (FD1(cid:98)⊗1)Ex φ(a∗)(cid:0)(F ∗ (cid:1)Tx + ν · φ(x∗x) D + FD(FD1(cid:98)⊗1)(cid:1)φ(a) + ν · φ(a∗a) D1(cid:98)⊗1)F ∗ is positive in L(E)/K(E) for all x ∈ Λ. Then we have that is positive in L(E)/K(E) for all a ∈ A. Proof. For x ∈ Λ we have that [FD1(cid:98)⊗1, φ(x)] = 0 and the closed submodule Ex ⊆ E is thus invariant under FD1(cid:98)⊗1. The restriction (FD1(cid:98)⊗1)Ex : Ex → Ex is therefore a well-defined bounded adjointable operator. The same observation holds for the adjoint F ∗ {x1, x2, x3, . . .}. For each n ∈ N, we choose a constant Since Λ is countable we may write the elements in Λ as a sequence D1(cid:98)⊗1. Cn > 2 + (cid:107)xn(cid:107)2 + (cid:107)FDT ∗ xn − T ∗ xnFDx(cid:107)∞ · (cid:107)xn(cid:107) and define the element Γ := ∞(cid:88) n=1 1 n2Cn nxn ∈ A , x∗ where the series is absolutely convergent. Since Λ·A ⊆ A is norm-dense and Λ = Λ∗ we have that We now compute modulo K(E), using Proposition 26, that Γ com- Γ · A ⊆ A is norm-dense as well. It therefore suffices to show that is positive in the Calkin algebra L(E)/K(E). D + FD(FD1(cid:98)⊗1)(cid:1) · Γ + ν · φ(Γ2) mutes with FD1(cid:98)⊗1 and that (FD, E) is a Kasparov module: Γ ·(cid:0)(F ∗ Γ ·(cid:0)(F ∗ D1(cid:98)⊗1)F ∗ D + FD(FD1(cid:98)⊗1)(cid:1) · Γ D1(cid:98)⊗1)F ∗ D1(cid:98)⊗1)FD + FD(FD1(cid:98)⊗1)(cid:1) · Γ3/2 ∼ Γ1/2(cid:0)(F ∗ ∞(cid:88) ∞(cid:88) D1(cid:98)⊗1)FD + FD(FD1(cid:98)⊗1)(cid:1)T ∗ (cid:0)(F ∗ (cid:0)(F ∗ D1(cid:98)⊗1)ExFDx + FDx(FD1(cid:98)⊗1)Ex ∼ Γ1/2 = Γ1/2 xnTxn n2Cn T ∗ n=1 xn 1 1 n2Cn n=1 (cid:1)TxnΓ1/2 . 26 JENS KAAD AND WALTER D. VAN SUIJLEKOM But this proves the present proposition since ∞(cid:88) n=1 Γ1/2 1 n2Cn = Γ1/2 T ∗ xn (cid:0)(F ∗ D1(cid:98)⊗1)ExFDx + FDx(FD1(cid:98)⊗1)Ex (cid:16) (cid:1)TxnΓ1/2 (cid:0)(F ∗ D1(cid:98)⊗1)ExFDx + FDx(FD1(cid:98)⊗1)Ex (cid:17) · Γ1/2 + νT ∗ + νφ(Γ2) T ∗ n2Cn xn 1 ∞(cid:88) n=1 xnTxn is positive in L(E)/K(E) by assumption. (cid:1)Txn (cid:3) 7.3. The positivity condition. We remain in the setup described in the beginning of Section 7. Before continuing our treatment of the positivity condition in Equa- tion (7.2) we introduce some further notation: Definition 37. For each x ∈ A satisfying condition (c) in Definition 28 we put Dom(Qx) := Dom(Dφ(x∗)) ∩ Im(φ(x)) and define the map Qx : Dom(Qx) → C by Qx(ξ) := 2 · Re(cid:104)Dφ(x∗)ξ, (D1(cid:98)⊗1)φ(x∗)ξ(cid:105) , where Re : C → C takes the real part of an element in the C∗-algebra C. For each λ ≥ 0 and x ∈ A satisfying condition (b) of Definition 28 we define the bounded adjointable operators on Ex: R1(λ)Ex :=(cid:0)1 + λ + (D∗ S1(λ)Ex := (D1(cid:98)⊗1)(cid:0)1 + λ + (D∗ 1(cid:98)⊗1)(D1(cid:98)⊗1)(cid:1)−1Ex 1(cid:98)⊗1)(D1(cid:98)⊗1)(cid:1)−1Ex Rx(λ) := (1 + λ + D2 Sx(λ) := Dx(1 + λ + D2 x)−1 x)−1 . The next lemma follows by standard functional calculus arguments: Lemma 38. Suppose that x ∈ A satisfies condition (b) of Definition 28. Then the maps [0,∞)2 → L(Ex) defined by M2(λ, µ, x) := Sx(λ)R1(µ)Ex ·(cid:112)1 + µ M1(λ, µ, x) := Sx(λ)S1(µ)Ex M3(λ, µ, x) := Rx(λ)S1(µ)Ex · √ M4(λ, µ, x) := Rx(λ)R1(µ)Ex ·(cid:112)(1 + λ)(1 + µ) 1 + λ are all continuous in operator norm and satisfy the estimate (cid:107)Mj(λ, µ, x)(cid:107)∞ ≤ (1 + λ)−1/2 · (1 + µ)−1/2 j ∈ {1, 2, 3, 4} , ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 27 for all λ, µ ∈ [0,∞). In particular, it holds that the integral (λµ)−1/2 · (M∗ j Mj)(λ, µ, x) dλdµ converges absolutely to a bounded adjointable operator Kj(x) ∈ L(Ex) with 0 ≤ Kj(x) ≤ 1 for all j ∈ {1, 2, 3, 4}. In order to ensure that later computations are well-defined we prove the following: Lemma 39. Suppose that x ∈ A satisfies condition (c) of Definition 28 and that (cid:107)d(x∗)φ(x)(cid:107)∞ < 1. Then (7.5) (cid:1) ⊆ Dom(Qx) , Im(cid:0)Rx(λ)Tx (cid:90) ∞ (cid:90) ∞ 0 0 1 π2 Im(cid:0)Sx(λ)Tx (cid:1) ⊆ Dom(Qx) Im(cid:0)Mj(λ, µ, x)Tx) ⊆ Dom(Qx) , and for all λ ≥ 0. In particular, if x ∈ A moreover satisfies condition (b) of Definition 28, then for all j ∈ {1, 2, 3, 4} and all λ, µ ∈ [0,∞). Proof. Recall from Lemma 10 and Proposition 11 that (ir + Dx)−1Tx = Tx(ir + Dφ(x∗x))−1 , for all r ∈ R with r ≥ 1 > (cid:107)d(x∗)φ(x)(cid:107)∞. We thus see that (cid:1) ⊆ Im(φ(x)) ∩ Dom(Dφ(x∗)) = Dom(Qx) . Im(cid:0)(ir + Dx)−1Tx The inclusions in Equation (7.5) now follow since Rx(λ)Tx = (−i √ √ 1 + λ + Dx)−1(i 1 + λ + Dx)−1Tx and since √ Sx(λ)Tx = DxRx(λ)Tx = (i for all λ ≥ 0. √ 1 + λ + Dx)−1Tx + i 1 + λ · Rx(λ)Tx , (cid:3) We now start a more detailed computation of the application Qx : Dom(Qx) → C from Definition 37. Lemma 40. Suppose that x ∈ A satisfies condition (b) and (c) of Definition 28 and that (cid:107)d(x∗)φ(x)(cid:107)∞ < 1. Then (cid:0)Sx(λ)Tx(ξ)(cid:1) = 2 · Re(cid:10)(D1(cid:98)⊗1)φ(x)ξ, Sx(λ)Txξ(cid:11) (cid:0)Rx(λ)Tx(ξ)(cid:1) , for all λ ∈ [0,∞) and ξ ∈ Dom(cid:0)(D1(cid:98)⊗1)φ(x)(cid:1). Proof. Let λ ∈ [0,∞) and let ξ ∈ Dom(cid:0)(D1(cid:98)⊗1)φ(x)(cid:1) be given. We − (1 + λ)Qx Qx first claim that x Sx(λ)Txξ ∈ Dom((D1(cid:98)⊗1)φ(x∗x)) DT ∗ 28 JENS KAAD AND WALTER D. VAN SUIJLEKOM and that (D1(cid:98)⊗1)φ(x∗x)DT ∗ = (D1(cid:98)⊗1)φ(x∗x)ξ − (1 + λ)(D1(cid:98)⊗1)T ∗ x Sx(λ)Txξ x Rx(λ)Txξ . But this follows since φ(x∗x)DT ∗ x Sx(λ)Txξ = T ∗ = φ(x∗x)ξ − (1 + λ)T ∗ x DxSx(λ)Txξ x Rx(λ)Txξ ∈ Dom(D1(cid:98)⊗1) , where we remark that φ(x∗x)ξ ∈ Dom(D1(cid:98)⊗1) since x∗ ∈ A and that x Rx(λ)Txξ ∈ Dom(D)∩ Dom(D1(cid:98)⊗1) by condition (c) and Lemma 39. T ∗ Notice now that condition (b) and Proposition 7 implies that (D1(cid:98)⊗1)φ(x∗x) : Dom(cid:0)(D1(cid:98)⊗1)φ(x∗x)(cid:1) → E the following computation: is selfadjoint and regular. Putting η := T ∗ Dom(D1(cid:98)⊗1) and using the above claim, the lemma is then proved by x Rx(λ)Tx(ξ) ∈ Dom(D) ∩ (cid:0)Sx(λ)Tx(ξ)(cid:1) = Re(cid:10)DT ∗ x Sx(λ)Tx(ξ)(cid:11) = Re(cid:10)(D1(cid:98)⊗1)φ(x∗x)DT ∗ = Re(cid:10)(D1(cid:98)⊗1)φ(x∗x)ξ, Dη(cid:11) − (1 + λ)Re(cid:10)(D1(cid:98)⊗1)η, Dη(cid:11) = Re(cid:10)(D1(cid:98)⊗1)φ(x)ξ, Sx(λ)Txξ(cid:11) − (1 + λ)Re(cid:10)(D1(cid:98)⊗1)η, Dη(cid:11) . x Sx(λ)Tx(ξ), (D1(cid:98)⊗1)T ∗ x Sx(λ)Txξ, Dη(cid:11) · Qx 1 2 (cid:3) Definition 41. For each x ∈ A satisfying condition (b) and (c) of Definition 28 and that (cid:107)d(x∗)φ(x)(cid:107)∞ < 1, we define the assignment Qj(λ, µ, x) : Im(Tx) → C for all λ, µ ∈ [0,∞), j ∈ {1, 2, 3, 4}. (cid:0)Mj(λ, µ, x)Txξ(cid:1) , Qj(λ, µ, x)(Txξ) := Qx The main algebraic result of this section can now be stated and proved: Lemma 42. Suppose that x ∈ A satisfies condition (b) and (c) of Definition 28 and that (cid:107)d(x∗)φ(x)(cid:107)∞ < 1. Then we have the identity 4(cid:88) Qj(λ, µ, x)(Txξ) = 2 · Re(cid:10)Txξ, Sx(λ)S1(µ)ExTxξ(cid:11) , j=1 for all λ, µ ∈ [0,∞) and all ξ ∈ E. ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS Proof. Let λ, µ ∈ [0,∞) and ξ ∈ E be given. Remark that S1(µ)ξ , R1(µ)ξ ∈ Dom((D1(cid:98)⊗1)φ(x)). We may thus use Lemma 40 to compute as follows: 4(cid:88) 29 Qj(λ, µ, x)(Txξ) j=1 = Qx(Sx(λ)TxS1(µ)ξ) + Qx(Sx(λ)TxR1(µ)ξ)(1 + µ) + Qx(Rx(λ)TxS1(µ)ξ)(1 + λ) + Qx(Rx(λ)TxR1(µ)ξ)(1 + λ)(1 + µ) = 2 · Re(cid:10)(D1(cid:98)⊗1)φ(x)S1(µ)ξ, Sx(λ)TxS1(µ)ξ(cid:11) + 2 · Re(cid:10)(D1(cid:98)⊗1)φ(x)R1(µ)ξ, Sx(λ)TxR1(µ)ξ(cid:11) · (1 + µ) 1(cid:98)⊗1)S1(µ)ξ, Sx(λ)TxS1(µ)ξ(cid:11) = 2 · Re(cid:10)Tx(D∗ + 2 · Re(cid:10)TxS1(µ)ξ, Sx(λ)TxR1(µ)ξ(cid:11) · (1 + µ) = 2 · Re(cid:10)Txξ, Sx(λ)S1(µ)ExTxξ(cid:11) This proves the present lemma. (cid:3) D1(cid:98)⊗1)F ∗ Proof. By Proposition 36, it suffices to show that holds in the quotient C∗-algebra L(E)/K(E) for all a ∈ A. We are now ready to treat the positivity condition in Equation (7.2): Proposition 43. Suppose that Λ ⊆ A is a localizing subset satisfying the local positivity condition, that (cid:107)d(x∗)φ(x)(cid:107)∞ < 1 for all x ∈ Λ and that there exists a κ > 0 such that κx ≤ κ for all x ∈ Λ. Then the inequality is positive in L(E)/K(E) for all x ∈ Λ. Let thus x ∈ Λ be fixed. We will prove the inequality φ(a)∗(cid:0)(F ∗ D + FD(FD1(cid:98)⊗1)(cid:1)φ(a) ≥ −4κ · φ(a∗a) (cid:0)(F ∗ (cid:1)Tx + 4κφ(x∗x) D1(cid:98)⊗1)ExFDx + FDx(FD1(cid:98)⊗1)Ex 2 · Re(cid:10)FDx(FD1(cid:98)⊗1)ExTxξ, Txξ(cid:11) ≥ −4κ(cid:104)Txξ, Txξ(cid:105) in the C∗-algebra C, for all ξ ∈ Dom(D) ∩ Dom(D1(cid:98)⊗1). Remark that this is enough since Dom(D) ∩ Dom(D1(cid:98)⊗1) ⊆ E is norm-dense. Let thus ξ ∈ Dom(D) ∩ Dom(D1(cid:98)⊗1) be given. We have that 2 · Re(cid:10)FDx(FD1(cid:98)⊗1)Ex (cid:1)Txξ, Txξ(cid:11) (cid:90) ∞ (λµ)−1/2 · Re(cid:10)Sx(λ)S1(µ)ExTxξ, Txξ(cid:11) dλdµ , (cid:90) ∞ T ∗ x = 2 π2 0 0 where the integral converges absolutely in the norm on C and the integrand is norm-continuous from [0,∞)2 to C. Now, by Lemma 42 30 JENS KAAD AND WALTER D. VAN SUIJLEKOM and the local positivity condition we have that 2 · Re(cid:10)Sx(λ)S1(µ)ExTxξ, Txξ(cid:11) = 4(cid:88) j=1 It therefore follows by Lemma 38 that Qj(λ, µ, x)(Txξ) (cid:104)Mj(λ, µ, x)Txξ, Mj(λ, µ, x)Txξ(cid:105) . 4(cid:88) (λµ)−1/2 · Re(cid:10)Sx(λ)S1(µ)ExTxξ, Txξ(cid:11) (cid:90) ∞ 4(cid:88) 4(cid:88) (cid:104)Txξ, Kj(x)Txξ(cid:105) ≥ −4κ(cid:104)Txξ, Txξ(cid:105) . (cid:90) ∞ (λµ)−1/2 0 0 j=1 (cid:104)Mj(λ, µ, x)Txξ, Mj(λ, µ, x)Txξ(cid:105) dλdµ ≥ −κ · j=1 (cid:90) ∞ (cid:90) ∞ 2 π2 0 0 ≥ −κ · 1 π2 · = −κ · j=1 But this proves the proposition. (cid:3) References [BaJu83] S. Baaj and P. Julg, Th´eorie bivariante de Kasparov et op´erateurs non born´es dans les C∗-modules hilbertiens, C. R. Acad. Sci. Paris S´er. I Math. 296 (1983), no. 21, 875 -- 878. MR 715325 (84m:46091) [BDT89] P. Baum, R. G. Douglas, and M. E. Taylor, Cycles and relative cycles in analytic K-homology, J. Differential Geom. 30 (1989), no. 3, 761 -- 804. MR 1021372 [BMvS16] S. Brain, B. Mesland, and W. D. van Suijlekom, Gauge theory for spectral triples and the unbounded Kasparov product, J. Noncommut. Geom. 10 (2016), no. 1, 135 -- 206. MR 3500818 [CGRS14] A. L. Carey, V. Gayral, A. Rennie, and F. A. Sukochev, Index theory for locally compact noncommutative geometries, Mem. Amer. Math. Soc. 231 (2014), no. 1085, vi+130. MR 3221983 [CoMo08] A. Connes and H. Moscovici, Type III and spectral triples, Traces in number theory, geometry and quantum fields, Aspects Math., E38, Friedr. Vieweg, Wiesbaden, 2008, pp. 57 -- 71. [Con94] A. Connes, Noncommutative geometry, Academic Press, Inc., San [Con96] Diego, CA, 1994. MR 1303779 (95j:46063) , Gravity coupled with matter and the foundation of non- commutative geometry, Comm. Math. Phys. 182 (1996), no. 1, 155 -- 176. MR 1441908 [CoSk84] A. Connes and G. Skandalis, The longitudinal index theorem for foliations, Publ. Res. Inst. Math. Sci. 20 (1984), no. 6, 1139 -- 1183. MR 775126 (87h:58209) [Hil10] M. Hilsum, Bordism invariance in KK-theory, Math. Scand. 107 (2010), no. 1, 73 -- 89. MR 2679393 ON A THEOREM OF KUCEROVSKY FOR HALF-CLOSED CHAINS 31 [Kaa15] [Kaa16] [Kaa17] J. Kaad, The unbounded Kasparov product by a differentiable module, arXiv:1509.09063 [math.KT]. arXiv:1612.08405. , Morita invariance of unbounded bivariant KK-theory, , Differentiable absorption of Hilbert C∗-modules, connections, and lifts of unbounded operators, To appear in J. Noncommut. Geom. 11 (2017), 1 -- 32. arXiv:1407.1389 [math.OA] product, Adv. Math. 248 (2013), 495 -- 530. MR 3107519 [KaLe13] J. Kaad and M. Lesch, Spectral flow and the unbounded Kasparov [Kas80] G. G. Kasparov, The operator K-functor and extensions of C∗- algebras, Izv. Akad. Nauk SSSR Ser. Mat. 44 (1980), no. 3, 571 -- 636, 719. MR 582160 (81m:58075) [KavS17] J. Kaad and W. D. van Suijlekom, Factorization of Dirac operators on almost-regular fibrations of spinc manifolds, Preprint, 2017. [Lan95] [Kuc97] D. Kucerovsky, The KK-product of unbounded modules, K-Theory 11 (1997), no. 1, 17 -- 34. MR 1435704 (98k:19007) E. C. Lance, Hilbert C∗-modules, London Mathematical Society Lec- ture Note Series, vol. 210, Cambridge University Press, Cambridge, 1995, A toolkit for operator algebraists. MR 1325694 (96k:46100) F. Latr´emoli`ere, Quantum locally compact metric spaces, J. Funct. Anal. 264 (2013), no. 1, 362 -- 402. MR 2995712 [Lat13] [Mes14] [MeRe16] B. Mesland and A. Rennie, Nonunital spectral triples and metric completeness in unbounded KK-theory, J. Funct. Anal. 271 (2016), no. 9, 2460 -- 2538. MR 3545223 B. Mesland, Unbounded bivariant K-theory and correspondences in noncommutative geometry, J. Reine Angew. Math. 691 (2014), 101 -- 172. MR 3213549 S. L. Woronowicz, Unbounded elements affiliated with C∗-algebras and noncompact quantum groups, Comm. Math. Phys. 136 (1991), no. 2, 399 -- 432. MR 1096123 (92b:46117) [Wor91] Department of Mathematics and Computer Science, Syddansk Uni- versitet, Campusvej 55, 5230, Odense M, Denmark E-mail address: [email protected] Institute for Mathematics, Astrophysics and Particle Physics, Rad- boud University Nijmegen, Heyendaalseweg 135, 6525 AJ Nijmegen, The Netherlands E-mail address: [email protected]
1109.2379	2	1109	2012-03-29T07:17:37	Decomposable approximations of nuclear C*-algebras	[ "math.OA" ]	We show that nuclear C-algebras have a refined version of the completely positive approximation property, in which the maps that approximately factorize through finite dimensional algebras are convex combinations of order zero maps. We use this to show that a separable nuclear C-algebra A which is closely contained in a C*-algebra B embeds into B.	math.OA	math	DECOMPOSABLE APPROXIMATIONS OF NUCLEAR C ∗-ALGEBRAS ILAN HIRSHBERG, EBERHARD KIRCHBERG, AND STUART WHITE Abstract. We show that nuclear C ∗-algebras have a refined version of the com- pletely positive approximation property, in which the maps that approximately factorize through finite dimensional algebras are convex combinations of order zero maps. We use this to show that a separable nuclear C ∗-algebra A which is closely contained in a C ∗-algebra B embeds into B. The decomposition rank and nuclear dimension of a C ∗-algebra are noncommu- tative notions of covering dimension which play a prominent role in the structure and classification theory of C ∗-algebras (see, e.g. [ET08, Win10, Win12]). These dimensions are defined in terms of uniformly decomposable completely positive ap- proximations of nuclear C ∗-algebras. The main theorem of this paper places these definitions in a broader context by providing a sharpening of the completely positive approximation property: nuclear C ∗-algebras always have decomposable completely positive approximations. We apply our approximation theorem to resolve a problem in perturbation theory of nuclear C ∗-algebras. Given two C ∗-algebras A and B concretely represented on the same Hilbert space, say that A ⊆γ B if operators in the unit ball of A can be approximated in B up to γ. In [Chr80], Christensen showed that a sufficiently close near containment M ⊆γ N of von Neumann algebras with M injective implies that there is an embedding M ֒→ N . In the C ∗-context, [CSS+12, Theorem 6.10] (see also [CSS+10]) produces embeddings from sufficiently close near containments A ⊆γ B when A is separable and has finite nuclear dimension, however the estimates in this result depend on the nuclear dimension of A. The approximation theorem enables us to remove the hypothesis of finite nuclear dimension and establish a C ∗- version of Christensen's embedding theorem with a universal constant valid for all separable nuclear C ∗-algebras. This is the subject of Section 2. Recall that a C ∗-algebra A is nuclear if and only if it has the completely positive approximation property ([CE78, Kir77]), that is, if for any finite subset F ⊆ A and for any ε > 0 there is a finite dimensional C ∗-algebra A0 and completely positive contractions 2 1 0 2 r a M 9 2 ] . A O h t a m [ 2 v 9 7 3 2 . 9 0 1 1 : v i X r a A ψ / A0 ϕ / A such that kϕ◦ ψ(a)− ak < ε for all a ∈ F . We call the triple (A0, ψ, ϕ) a CP approx- imation for (F, ε). The decomposition rank was introduced in [KW04] using order zero maps and decomposable maps. A completely positive contraction ϕ : A → B is said to be an order zero map if ϕ preserves orthogonality, i.e. whenever self-adjoint elements x, y ∈ A satisfy xy = 0 then we also have ϕ(x)ϕ(y) = 0. A completely positive map ϕ : A → B is said to be n-decomposable if A decomposes into a direct This research was supported by Israel Science Foundation grant 1471/07 and UK Engineering and Physical Sciences Research Council grant EP/I019227/1. 1 / / 2 ILAN HIRSHBERG, EBERHARD KIRCHBERG, AND STUART WHITE sum A = A0 ⊕A1⊕···⊕An such that ϕAk is an order zero map for all k. If ϕ is n- decomposable for some n, we will say that ϕ is decomposable. A nuclear C ∗-algebra A is said to have decomposition rank n if n is the smallest number such that for every finite subset F ⊆ A and ε > 0, there exist CP approximation (A0, ψ, ϕ) where ϕ is n-decomposable. The definition of nuclear dimension in [WZ10] uses a slightly weaker condition on approximating triples, asking that we can approximate (F, ε) by a triple (A0, ψ, ϕ) with A0 finite dimensional, ψ a completely positive contrac- tion and ϕ an n-decomposable map (but not necessarily contractive). This notion reaches purely infinite C ∗-algebras, whereas C ∗-algebras with finite decomposition rank are quasidiagonal and therefore stably finite ([KW04]). In Theorem 1.4, we refine the completely positive approximation theorem and show that if A is nuclear, then one can always chose CP approximations such that the maps ϕ going back into A are decomposable and contractive, only without an upper bound on the number of summands involved in the decomposition. In fact, the theorem is stronger - the map ϕ can be chosen to be a convex combination of order zero maps. Thus, one can always choose decomposable CP approximations, and the definitions of decomposition rank and nuclear dimension require placing a uniform bound on the number of summands present. In Section 3, we present an approximation theorem for weakly nuclear completely positive contractions Φ : B → M from separable exact C ∗-algebras into properly infinite von Neumann algebras. Such maps can be approximated in the point-weak∗ topology by a convex combination of ∗-homomorphisms B → M. In fact we can find a fixed ∗-homomorphism ϕ : B → M so that Φ can be approximated by convex combinations of two unitary conjugates of ϕ. As a consequence we obtain another proof of a special case of Theorem 1.4, under the additional assumption that A∗∗ is properly infinite (e.g. if A is stable or simple and purely infinite). Part of the work on this paper was done while the authors visited the CRM in Barcelona and we thank the CRM for its hospitality. We thank Erik Christensen and the referee for their remarks on an earlier version of the paper. I.H. also wishes to thank Nate Brown for some helpful conversations regarding this paper. 1. Approximation via convex combinations of order zero maps In this section we establish our strengthening of the completely positive approxi- mation property for nuclear C ∗-algebras. The first step is to establish a Kaplansky density lemma for order zero maps. Recall from [KW04, Remark 2.4] that order zero maps are projective in the sense that whenever J ⊳D is an ideal and ϕ : F → D/J is an order zero map from a finite dimensional C ∗-algebra, there exists an order zero map ϕ : F → D lifting ϕ, i.e. q ◦ ϕ = ϕ, where q : D → D/J is the quotient map. As noted in [KW04], the projectivity of order zero maps arises from the correspon- dence between order zero maps F → A and ∗-homomorphisms from the cone on F into A and Loring's projectivity of the cones on finite dimensional C ∗-algebras (see [Lor97]). Lemma 1.1. Let A ⊂ B(H) be a C ∗-algebra, and let M be the strong∗-closure of A. Let F be a finite dimensional C ∗-algebra and ϕ : F → M be an order zero map. Then there exists a net (ϕλ) of order zero maps F → A such that ϕλ(x) → ϕ(x) in strong∗-topology for all x ∈ F . In particular ϕλ(x) → ϕ(x) in weak∗-topology for all x ∈ F . DECOMPOSABLE APPROXIMATIONS OF NUCLEAR C ∗-ALGEBRAS 3 ∞ for the strong∗-limit of (x(n) Proof. Let Λ denote the directed set of all strong∗-neighborhoods of 0 ∈ H ordered for V, W ∈ Λ, write V ≥ W when V ⊆ W . Let D consist by containment, i.e. of all bounded nets (xV )V ∈Λ in A indexed by Λ which converge in the strong∗- topology to some x ∈ M (that is, such that for any strong∗-neighborhood W of 0 there exists W0 ≥ W such that for any V ≥ W0 we have that xV − x ∈ W ). Since multiplication is jointly strong∗-continuous on bounded sets, D forms a ∗- subalgebra of ℓ∞(Λ,A). Given a sequence (x(n))∞ n=1 in D converging (in norm) to x ∈ ℓ∞(Λ,A), write x(n) V )V ∈Λ. It is easy to check that (x(n) ∞ )∞ n=1 is norm-Cauchy, and therefore converges to x∞ in M, say. By writing xV − x∞ = (xV − x(n) ∞ − x∞), we see that xV converges in strong∗-topology to x∞ and so D is a C ∗-subalgebra of ℓ∞(Λ,A). Now let J be the ideal in D consisting of strong∗-null nets and note that the argument above shows that J is norm closed in D. The quotient map D → D/J is given by evaluating the strong∗-limit of a net in D. By Kaplansky's density theorem every element of M is the strong∗-limit of a net indexed by Λ and so D/J is canonically ∗-isomorphic to M. Given a finite-dimensional C ∗-algebra F and an order zero map ϕ : F → M, projectivity enables us to lift ϕ to an order zero map ϕ : F → D. Writing ϕ as a net (ϕU )U ∈Λ of contractions F → A, each of the maps ϕU is order zero and the lifting ensures that ϕU (x) → ϕ(x) in strong∗-topology for all x ∈ F . ∞ ) + (x(n) V ) + (x(n) V − x(n) (cid:3) Our next lemma uses the fact that AF algebras are locally reflexive to show that hyperfinite von-Neumann algebras M (i.e. those containing a weak∗-dense AF C ∗- subalgebra) have a stronger version of semi-discreteness: the maps coming back into M can be taken to be ∗-homomorphisms. If one defined a 'semi-discreteness dimension' analogously to the nuclear dimension, then, as expected, all injective von Neumann algebras would be zero dimensional. Recall that a C ∗-algebra B is said to be locally reflexive if for any finite dimensional operator system E ⊂ B∗∗ there is a net of CP contractions ψν : E → B such that ψν (x) → x weak∗ for any x ∈ E (see [BO08, Chapter 9]). We recall that any nuclear C ∗-algebra is locally reflexive, and in particular any AF algebra is locally reflexive. Lemma 1.2. Let M be a hyperfinite von-Neumann algebra. Then there exists a net of finite dimensional subalgebras Aν and CP contractions Φν : M → Aν such that Φν(a) → a in the weak∗ topology for any a ∈ M. Proof. Given x1,··· , xn ∈ M and a weak∗-neighborhood V of 0, let E be the finite dimensional operator system in M spanned by the xi. It suffices to produce a finite dimensional subalgebra A ⊂ M and a completely positive contraction Φ : M → A with Φ(xi) − xi ∈ V for i = 1,··· , n, as then we can produce a net of such maps indexed by finite subsets of M and a weak∗-neighborhood basis of 0. Let B be a weak∗-dense AF subalgebra of M and regard M as a von Neumann subalgebra of B∗∗. By local reflexivity there exists a completely positive contraction ϕ : E → B with ϕ(xi) − xi ∈ V for all i. By choosing a finite dimensional subalgebra A ⊂ B which ε-contains the image under ϕ of the unit ball of E for a sufficiently small ε and composing ϕ by a conditional expectation from B onto A, we may additionally assume that ϕ takes values in A. Arveson's extension theorem then enables us to extend this map to the required map Φ. (cid:3) 4 ILAN HIRSHBERG, EBERHARD KIRCHBERG, AND STUART WHITE The previous lemmas together with Connes' theorem from [Con76] that injectivity implies hyperfiniteness will now be used to derive an approximation property for nuclear C ∗-algebras in the weak topology in which the maps going into the algebra are order zero. Lemma 1.3. If A is a separable nuclear C ∗-algebra, then there is a net of CP contractions A ψν / Aν ϕν / A with Aν finite dimensional such that for all a ∈ A, ϕν ◦ ψν(a) → a weakly and ϕν are order zero maps. Proof. Since A is nuclear, A∗∗ is injective. If π is a GNS representation of A, then π(A)′′ is a cutdown of A∗∗ by a central projection, and therefore is injective as well. Since A is separable, π(A)′′ has separable predual, and hence is hyperfinite by Connes' theorem. Since A∗∗ is a direct sum of (possibly uncountably many) hyperfinite W ∗-algebras, it is hyperfinite itself (see [Tak03, Chapter XVI]). For a1,··· , an ∈ A and a weak neighborhood V of 0 in A, take a weak neighbor- hood W of 0 with W + W ⊆ V . By Lemma 1.2, find a finite dimensional subalgebra A0 ⊂ A∗∗ and a CP contraction ψ : A∗∗ → A0 such that ψ(ai) − ai ∈ W for i = 1,··· , n. Applying Lemma 1.1 to the inclusion map A0 ֒→ A∗∗, we can find an order zero map ϕ : A0 → A with ϕ(ψ(ai)) − ψ(ai) ∈ W for i = 1,··· , n. Thus ϕ(ψ(ai)) − ai ∈ V for i = 1,··· , n. Using nets indexed by finite subsets of A and weak-neighborhoods of 0 gives the result. (cid:3) We are now in a position to deduce the main result of the paper from the preceding lemmas. This works in the same way as the completely positive approximation property of a C ∗-algebra is deduced from semidiscreteness of its bidual. Theorem 1.4. If A is a nuclear C ∗-algebra, then for any finite set F ⊆ A and any ε > 0 there is a CP-approximation (A0, ψ, ϕ), ψ : A → A0, ϕ : A0 → A for F to within ε such that ϕ, ψ are CP contractions, A0 is finite dimensional, and ϕ is a convex combination of finitely many contractive order zero maps. Proof. Since any finite subset of a nuclear C ∗-algebra is contained in a separable nuclear subalgebra (see [BO08, Exercise 2.3.9]), we may assume without loss of generality that A is separable. It follows from the Hahn-Banach theorem that any convex subset of B(A) has the same point-norm and point-weak closures. Thus, let K0 ⊆ B(A) be the set of all contractive CP maps T : A → A which admit a factorization T = ϕ ◦ ψ, where −→ A, with A0 finite dimensional, ψ a CP contraction and ϕ an order A zero map, and let K = conv(K0). By Lemma 1.3, the identity map A → A lies in the point-weak closure of K0 and so there is a net of elements Tλ ∈ K such that Tλ(a) → a in norm for all a ∈ A. −→ A0 We finally note that any T ∈ K can be decomposed as ψ ϕ T A ψ B ϕ / A / / / / ' ' / DECOMPOSABLE APPROXIMATIONS OF NUCLEAR C ∗-ALGEBRAS 5 with B finite dimensional and ϕ a finite convex combination of order zero maps, as i=1 tiTi, a convex combination of elements Ti ∈ K0. Decom- follows. Write T = Pn pose each Ti as Ti A ψi with ϕi order zero. Setting B = Ln ϕi(a1 ⊕ a2 ⊕ ... ⊕ an) = ϕi(ai), we can take ϕ = Pn decomposition of T . i=1 Bi, ψ(a) = Ln Bi ϕi / A i=1 ψi(a) and, denoting i=1 ti ϕi to obtain the required (cid:3) 2. Near Inclusions In this section we apply Theorem 1.4 to near containments. First we recall the definition of a near inclusion. Definition 2.1. Let A and B be C ∗-subalgebras of B(H). For γ > 0, write A ⊆γ B if for each a ∈ A, there exists b ∈ B with ka − bk ≤ γkak. Write A ⊂γ B if there exists γ′ < γ with A ⊆γ B. Combining Theorem 1.4 with a perturbation theorem for order zero maps from [CSS+12], we can linearize a near inclusion A ⊂γ B when A is nuclear. Lemma 2.2. Let A ⊂γ B be a near inclusion of C ∗-algebras and suppose that A is nuclear. Let η = 2(2γ + γ2)(2 + 2γ + γ2). Then there exists η0 < η with the property that for each finite subset Z of the unit ball of A, there exists a contractive CP map Φ : A → B such that kΦ(z) − zk ≤ η0 for z ∈ Z. Proof. Fix γ0 < γ so that A ⊆γ0 B and choose ε > 0 such that 0 ) + ε < η. 0 )(2 + 2γ0 + γ2 η0 = 2(2γ0 + γ2 Given a finite subset Z of the unit ball of A, use Theorem 1.4 to produce a CP approximation (A0, ψ, ϕ), ψ : A → A0, ϕ : A0 → A for Z to within ε such that ϕ and ψ are CP contractions, A0 is finite dimensional, and ϕ = Pn i=1 λiϕi is a convex combination of order zero maps. For each i, use [CSS+12, Theorem 6.4] to find a CP map θi : A0 → B with k θi − ϕikcb ≤ (2γ0 + γ2 0 ). By rescaling each θi if necessary we can find contractive CP maps θi : A0 → B with kϕi − θikcb ≤ 2kϕi − θikcb. Define θ = Pn i=1 λiθi. This is a contractive CP map with kθ − ϕkcb ≤ 2(2γ0 + γ2 0). Thus Φ = θ ◦ ψ is a contractive CP map from A into B with 0)(2 + 2γ0 + γ2 0)(2 + 2γ0 + γ2 kΦ(z) − zk ≤ kθ(ψ(z)) − ϕ(ψ(z))k + kϕ(ψ(z)) − zk ≤ η0, z ∈ Z. (cid:3) The embedding theorem below now follows immediately from [CSS+12, Lemma 4.1] (taking the µ of that lemma to be small enough that 8√6(η0)1/2 + η0 + µ < 8√6η1/2 + η). Theorem 2.3. Let A ⊂γ B be a near inclusion of C ∗-algebras for some γ satisfying η = 2(2γ + γ2)(2 + 2γ + γ2) < 1/210000. Suppose that A is separable and nuclear. Then for each finite subset X of the unit ball of A, there exists an embedding θ : A ֒→ B with kθ(x) − xk ≤ 8√6η1/2 + η, for x ∈ X. / / ' ' / 6 ILAN HIRSHBERG, EBERHARD KIRCHBERG, AND STUART WHITE Note that the examples of Johnson [Joh82] show that in the situation of Theorem 2.3 it is not generally possible to construct embeddings θ : A → B which are uniformly close to the inclusion of A into the underlying B(H). 3. Completely positive maps from exact C ∗-algebras In this last section, we examine weakly nuclear completely positive maps from exact algebras into properly infinite von Neumann algebras. Via the O2-embedding theorem, such maps are automatically nuclear (see Remark 3.4 below). More surpris- ingly we can approximate all such maps by convex combinations of ∗-homomorphisms. In fact more is true. An average of two unitary conjugates of a given ∗-homomorphism can be used to make this approximation. This can be used to give a different proof of Theorem 1.4 under the stronger assumption that A∗∗ is properly infinite (as happens when A is stable). This alternative approach avoids using Connes' theorem on the equivalence of injectivity and hyperfiniteness. We begin by recalling the standard fact that properly infinite von Neumann algebras absorb B(ℓ2) tensorially. Lemma 3.1. Let M be a properly infinite von-Neumann algebra. Then M ∼= M⊗B(ℓ2). Proof. Since 1 is properly infinite, M has a countably infinite family of mutually equivalent orthogonal projections which sum up to 1. Denoting one of those pro- jections by e, it follows from [Tak02, Proposition V.1.22], that M ∼= eMe⊗ B(ℓ2). Since B(ℓ2) ∼= B(ℓ2)⊗ B(ℓ2), we see that M ∼= eMe⊗ B(ℓ2)⊗B(ℓ2) ∼= M⊗B(ℓ2). (cid:3) We next prove a dilation lemma for unital completely positive maps into properly infinite von-Neumann algebras. In this lemma and the following theorem, we write ∼ for the Murray-von Neumann equivalence relation on projections. Lemma 3.2. Let A be a separable unital C ∗-algebra, M be a properly infinite von Neumann algebra and ϕ : A → M a unital completely positive map. Given a projection p ∈ M with 1 ∼ p ∼ 1 − p, there exists a ∗-homomorphism θ : A → M with pθ(x)p = pϕ(x)p for all x ∈ A. Proof. First note that the multiplier algebra M (K ⊗ M) of K ⊗ M is contained in B(ℓ2)⊗M. Indeed, representing M non-degenerately on a Hilbert space H so that K⊗M is non-degenerately represented on ℓ2⊗ H, take an increasing sequence (Pn)n of finite rank projections in K converging strongly to 1ℓ2(N). Given any idealizer x ∈ B(ℓ2 ⊗ H) (i.e. x satisfying x(K ⊗ M), (K ⊗ M)x ⊆ K ⊗ M), it follows that (Pn ⊗ 1)x(Pn ⊗ 1) ∈ K ⊗ M. By taking strong operator limits we see that x ∈ B(ℓ2)⊗M. Let e be a rank one projection in K and, by identifying M with the subalgebra (e⊗1)(K⊗M)(e⊗1) of K⊗M, view ϕ as a completely positive map into M (K⊗M). Write p0 = e ⊗ p, and apply Kasparov's Stinespring theorem [Kas80, Theorem 3] to the map x 7→ p0ϕ(x)p0 to obtain a partial isometry v ∈ M (K ⊗ M) and a ∗- homomorphism π : A → M (K ⊗ M) with p0ϕ(x)p0 = vπ(x)v∗ for x ∈ A. By the previous paragraph, we can view π as taking values in B(ℓ2)⊗M and v as a partial isometry in this algebra. We may assume that v = vπ(1) so that vv∗ = p0. Set q0 = v∗v. Write N0 = B(ℓ2)⊗M. Under the identification of M with B(ℓ2)⊗M of Lemma 3.1, we see that e ⊗ 1M is identified with e ⊗ 1B(ℓ2) ⊗ 1M ∼ 1B(ℓ2) ⊗ 1B(ℓ2) ⊗ 1M. DECOMPOSABLE APPROXIMATIONS OF NUCLEAR C ∗-ALGEBRAS 7 Undoing this identification, we have e ⊗ 1M ∼ 1B(ℓ2) ⊗ 1M in N0. Thus q0 ∼ p0 = e ⊗ p ∼ e ⊗ 1M ∼ 1B(ℓ2) ⊗ 1M. Now consider N1 = B(ℓ2)⊗N0 and let p1 = e ⊗ p0, q1 = e ⊗ q0. Arguing as above, we see that p1 ∼ q1 ∼ 1N1. Furthermore 1N1 − q1 = (1B(ℓ2) − e) ⊗ q0 + (1B(ℓ2) ⊗ (1N0 − q0)) ∼ 1B(ℓ2) ⊗ q0 + (1B(ℓ2) ⊗ (1N0 − q0)) = 1N1. In conclusion, we obtain the equivalence e⊗e⊗(1M−p) ∼ e⊗e⊗1M ∼ 1N1 ∼ 1N1−q1 and so there is a partial isometry w ∈ N1 with ww∗ = e ⊗ e ⊗ (1M − p) and w∗w = 1N1 − q1. Since e ⊗ v and w have orthogonal domain projections and orthogonal range projections, u = e⊗ v + w is an isometry in N1 with uu∗ = e⊗ e⊗ 1M. Define a ∗-homomorphism θ : A → (e ⊗ e ⊗ 1M)N1(e ⊗ e ⊗ 1M) by θ(x) = u(1 ⊗ π(x))u∗. As (e ⊗ e ⊗ p)u = e ⊗ v, we have (e ⊗ e ⊗ p)θ(x)(e ⊗ e ⊗ p) = (e ⊗ v)(1 ⊗ π(x))(e ⊗ v)∗ = e ⊗ p0ϕ(x)p0, x ∈ A. Identifying M with (e⊗e⊗1M)N1(e⊗e⊗1M), we can view θ as a ∗-homomorphism from A into M with pθ(x)p = pϕ(x)p for x ∈ A, as required. (cid:3) We can now establish our approximation result for weakly nuclear completely positive contractions from separable exact algebras into properly infinite von Neu- mann algebras. Given a state λ on a von Neumann algebra M, we write kxkλ for λ(x∗x)1/2. Theorem 3.3. Let M be a properly infinite von-Neumann algebra and let B be a separable exact C ∗-algebra. Let Φ : B → M be a weakly nuclear completely positive contraction. (1) There exists an injective ∗-homomorphism ϕ : B → M such that Φ is in the point-weak∗ closure of the convex hull of the family of maps {u∗ϕ(·)u u ∈ U (M)}. The ∗-homomorphism ϕ can be taken to be unital when Φ is unital. (2) Furthermore, if A ⊆ M is a weak∗-dense C ∗-algebra such that Φ(B) ⊆ A and O2 is unitally contained in the multiplier algebra M (A) ⊆ M (e.g. if A is stable) then one can find an injective ∗-homomorphism ϕ : B → M (A) such that Φ is in the point-σ(M (A),M∗)-closure of the convex hull of the family of maps {e−ihϕ(·)eih h ∈ As.a.,khk < π}. Again ϕ can be taken to be unital when Φ is unital. Proof. We first reduce to the case in which both B and Φ are unital. If B is not unital, then we can add a unit to B and extend Φ uniquely to a unital CP map from the unitization. If B was already unital but Φ is not, then we can adjoin a new unit to B and extend Φ to a unital map as well (see [BO08, Section 2.2], for example). We thus assume that B is unital and Φ is unital. Next, we reduce to the case where B is the Cuntz algebra O2. By [KP00, Theorem 2.8], we can find a unital embedding of B in O2. Since Φ is weakly nuclear, we can take a net (Φτ ) of CP-approximations to Φ in the point-weak∗ topology which factorize through finite dimensional algebras as Φτ Mnτ B Sτ Tτ / M where Sτ , Tτ are CP and unital. Use Arveson's extension theorem to extend Sτ to a unital CP map Sτ : O2 → Mnτ . Then any point-weak∗ cluster point of (Tτ ◦ Sτ ) / / ( ( / 8 ILAN HIRSHBERG, EBERHARD KIRCHBERG, AND STUART WHITE is an extension of Φ to O2. Thus we may assume without loss of generality that B = O2. Since M is properly infinite, O2 embeds unitally into M. Fix such a unital em- bedding ϕ : O2 → M. We will show that convex combinations of unitary conjugates of ϕ can be used to approximate Φ. To this end, fix a finite dimensional operator system X ⊆ M, normal states ρ1, ...ρn ∈ M∗ and ε > 0. We will find unitaries u1, u2 ∈ U (M) such that ρi(cid:18) 1 2 (cid:12)(cid:12)(cid:12)(cid:12) (u∗ 1ϕ(x)u1 + u∗ 2ϕ(x)u2) − Φ(x)(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) < εkxk, x ∈ X. Since X, ε and the states ρ1,··· , ρn are arbitrary, this will establish the result. Note that it actually proves more; Φ can be approximated by a convex combination of two unitary conjugates of ϕ. By the Radon-Nikodym theorem for states (see [BO08, Proposition 3.8.3]), there is a normal state λ ∈ M∗ such that in the corresponding GNS representation πλ with cyclic vector ξλ we can find positive operators y1, ..., yn ∈ πλ(M)′ satisfying ρi(x) = hπλ(x)ξλ, yiξλi for i = 1,··· , n. Fix ε′ > 0 such that 3kyik√ε′ +ε′ < ε for all i. As M is properly infinite, there exists a projection p1 ∈ M with p1 ∼ 1 − p1 ∼ 1 such that k(1 − p1)Φ(x)k2 (3.1) By Lemma 3.2, there is a ∗-homomorphism κ : O2 → M such that (3.2) λ = λ(Φ(x)∗(1 − p1)Φ(x)) ≤ ε′kxk2, p1κ(a)p1 = p1Φ(a)p1, x ∈ X. a ∈ O2. Let s = 1 − 2p1, then s is a self-adjoint unitary and we have (s∗κ(a)s + κ(a)) = p1κ(a)p1 + (1 − p1)κ(a)(1 − p1), 1 2 a ∈ O2. For any y ∈ M, the estimate (3.1) gives ky(1 − p1)k2 λ = λ((1 − p1)y∗y(1 − p1)) ≤ kyk2λ(1 − p1) ≤ kyk2ε′, as 1 ∈ X. Using (3.2), we have k(1 − p1)κ(x)(1 − p1) + p1κ(x)p1 − Φ(x)kλ √ε′, ≤k(1 − p1)κ(x)(1 − p1)kλ + k(1 − p1)Φ(x)kλ + kp1Φ(x)(1 − p1)kλ ≤ 3kxk for any x ∈ X, using (3.1) to estimate the middle term and the previous estimate for the first (with y = (1 − p1)κ(x)) and third (with y = p1Φ(x)). Therefore ρi(p1κ(x)p1 + (1 − p1)κ(x)(1 − p1) − Φ(x)) ≤ 3kxkkyik i = 1,··· n, x ∈ X. Since any two unital embeddings O2 → M are approximately unitary equivalent (as any properly infinite von Neumann algebra M satisfies the hypotheses of [Rør93, Theorem 3.6]), we can choose a unitary u ∈ U (M) such that x ∈ X. ku∗ϕ(x)u − κ(x)k < ε′kxk, √ε′, Taking u1 = us and u2 = u, we have (cid:12)(cid:12)(cid:12)(cid:12) ρi(cid:18) 1 2 (u∗ 1ϕ(x)u1 + u∗ 2ϕ(a)u2) − Φ(x)(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) 2 ≤ (cid:12)(cid:12)(cid:12)(cid:12) ρi(cid:18) 1 ≤ 3kxkkyik ≤kxkε, (s∗κ(x)s + κ(x)) − Φ(x)(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) √ε′ + kxkε′ i = 1,··· , n, x ∈ X, + ε′kxk DECOMPOSABLE APPROXIMATIONS OF NUCLEAR C ∗-ALGEBRAS 9 as claimed. For part 2, note that we had the freedom to choose any embedding ϕ : O2 → M and so we can ensure that the range of ϕ lies in M (A). As for the unitaries, as M is a von-Neumann algebra, we can write the unitaries in the form eib for b ∈ Msa with kbk ≤ π. By Kaplansky's density theorem, we can find a net of self-adjoint elements hτ ∈ A with norm smaller than π which converge to b in the SOT. Since the exponential function is SOT-continuous (see [Dav96, Lemma I.7.2], for example), it follows that eihτ → eib in SOT, and therefore one could replace the unitaries by exponentials as in the statement. (cid:3) We now show how the previous result gives a second proof of Theorem 1.4 under the stronger assumption that A∗∗ is properly infinite (e.g. if A is stable or simple and purely infinite). To see this, note that if F is a finite dimensional C ∗-algebra then any completely positive contraction Φ : F → A ⊆ A∗∗ is in the point-weak∗ closure of the convex hull of the ∗-homomorphisms from F to A∗∗ by Theorem 3.3. Using Lemma 1.1, it follows that Φ is in the point-weak closure of the convex hull of the order zero maps from F to A. As before, it follows from the Hahn-Banach theorem that in fact Φ is in the point-norm closure of convex combinations of order zero maps from F to A. Since A is nuclear, we can find a net of CP approximations Ψλ : A → Fλ, Φλ : Fλ → A to the identity. By the argument above, we may assume without loss of generality that Φλ is in fact a convex combination of order zero maps, giving us the required result. Remark 3.4. The fact that separable exact algebras embed in O2 can be used to give another characterization of exactness as follows. A C ∗-algebra B is exact if and only every weakly nuclear CP contraction Φ from B into a von Neumann algebra M is in fact nuclear. Proof. If B satisfies the hypothesis of the remark, then take Φ to be any faithful embedding of B into B(H) to see that B is nuclearly embeddable and therefore exact (see [BO08]). Conversely, let B be exact and take a weakly nuclear CP contraction Φ : B → M. Fix a finite dimensional operator system E ⊂ B and consider the separable C ∗-algebra B0 generated by E which is also exact. Thus we can embed B0 in O2. As in the first part of the proof of Theorem 3.3, ΦE can be extended to O2 and so is nuclear as O2 is nuclear. Since E was arbitrary, we can approximate Φ in the point-norm topology. (cid:3) Theorem 3.3, requires M to be properly infinite. It is natural to ask what happens when M is a finite von Neumann algebra. Question 3.5. Let M be type II1 factor, let F be a finite dimensional C ∗-algebra and let Φ : F → M be a completely positive contraction. Can Φ be approximated in the point-weak topology by convex combinations of order zero maps? One could ask the question about finite von-Neumann algebras which are not factors and do not have a type I part. One should note, though, that if M is finite dimensional then one would not necessarily be able to find such approximations for completely positive maps, due to dimension restrictions. For example, if n > m there are many CP maps from Mn to Mm but no non-zero order zero maps. This question is also interesting when we restrict to trace preserving maps. Note that different factorization properties of trace preserving contractive CP maps between 10 ILAN HIRSHBERG, EBERHARD KIRCHBERG, AND STUART WHITE finite von Neumann algebras have been considered in the literature and can be somewhat delicate, see [HM11], for example. References [BO08] Nathanial P. Brown and Narutaka Ozawa. C ∗-algebras and finite-dimensional approxima- tions, volume 88 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2008. [CE78] Man Duen Choi and Edward G. Effros. Nuclear C ∗-algebras and the approximation property. Amer. J. Math., 100(1):61 -- 79, 1978. [Chr80] Erik Christensen. Near inclusions of C ∗-algebras. Acta Math., 144(3-4):249 -- 265, 1980. [CSS+10] Erik Christensen, Allan M. Sinclair, Roger R. Smith, Stuart A. White, and Wilhelm Winter. The spatial isomorphism problem for close separable nuclear C∗-algebras. Proc. Natl. Acad. Sci. USA, 107(2):587 -- 591, 2010. [CSS+12] Eric Christensen, Allan M. Sinclair, Roger R. Smith, Stuart A. White, and Wilhelm Winter. Perturbations of nuclear C∗-algebras. Acta. Math., 208:93 -- 150, 2012. [Con76] Alain Connes. Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1. Ann. of Math. (2), 104(1):73 -- 115, 1976. [Dav96] Kenneth R. Davidson. C ∗-algebras by example, volume 6 of Fields Institute Monographs. [ET08] American Mathematical Society, Providence, RI, 1996. George A. Elliott and Andrew S. Toms. Regularity properties in the classification program for separable amenable C ∗-algebras. Bull. Amer. Math. Soc. (N.S.), 45(2):229 -- 245, 2008. [HM11] Uffe Haagerup and Magdalena Musat. Factorization and dilation problems for completely [Joh82] positive maps on von neumann algebras. Comm. Math. Phys., 303(2):555 -- 594, 2011. Barry E. Johnson. A counterexample in the perturbation theory of C ∗-algebras. Canad. Math. Bull., 25(3):311 -- 316, 1982. [KK72] Richard V. Kadison and Daniel Kastler. Perturbations of von Neumann algebras. I. Stability of type. Amer. J. Math., 94:38 -- 54, 1972. [Kas80] Gennadi Kasparov. Hilbert C ∗-modules: theorems of Stinespring and Voiculescu. J. [Kir77] [KP00] Operator Theory, 4(1):133 -- 150, 1980. Eberhard Kirchberg. C ∗-nuclearity implies CPAP. Math. Nachr., 76:203 -- 212, 1977. Eberhard Kirchberg and N. Christopher Phillips. Embedding of exact C ∗-algebras in the Cuntz algebra O2. J. Reine Angew. Math., 525:17 -- 53, 2000. [KW04] Eberhard Kirchberg and Wilhelm Winter. Covering dimension and quasidiagonality. In- [Lor97] ternat. J. Math., 15(1):63 -- 85, 2004. Terry A. Loring. Lifting solutions to perturbing problems in C ∗-algebras, volume 8 of Fields Institute Monographs. American Mathematical Society, Providence, RI, 1997. [Rør93] Mikael Rørdam. Classification of inductive limits of Cuntz algebras. J. Reine Angew. Math., 440:175 -- 200, 1993. [Tak02] Masamichi Takesaki. Theory of operator algebras. I, volume 124 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, Berlin, 2002. Reprint of the first (1979) edition, Operator Algebras and Non-commutative Geometry, 5. [Tak03] Masamichi Takesaki. Theory of operator algebras. III, volume 127 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, Berlin, 2003. Operator Algebras and Non- commutative Geometry, 8. [Win10] Wilhelm Winter. Decomposition rank and Z-stability. Invent. Math., 179(2):229 -- 301, 2010. [Win12] Wilhelm Winter. Nuclear dimension and Z-stability of pure C ∗-algebras. Invent. Math., 187(2):259 -- 342, 2012. [WZ10] Wilhelm Winter and Joachim Zacharias. The nuclear dimension of C ∗-algebras. Adv. Math., 224(2):461 -- 498, 2010. DECOMPOSABLE APPROXIMATIONS OF NUCLEAR C ∗-ALGEBRAS 11 Department of Mathematics, Ben Gurion University of the Negev, P.O.B. 653, Be'er Sheva 84105, Israel E-mail address: [email protected] Institut fur Mathematik, Humboldt-Universitat zu Berlin, Unter den Linden 6, D-10099 Berlin, Germany E-mail address: [email protected] School of Mathematics and Statistics, University of Glasgow, Glasgow, G12 8QW, Scotland E-mail address: [email protected]
1509.05701	1	1509	2015-09-18T16:58:51	2-Local derivations on algebras of matrix-valued functions on a compact	[ "math.OA" ]	In the present paper 2-local derivations on various algebras of infinite dimensional matrix-valued functions on a compact are considered. It is proved that every 2-local derivation on such algebra is a derivation. Also we explain that the method developed in the given paper can be applied to associative, Jordan and Lie algebras of infinite dimensional matrix-valued functions on a compact.	math.OA	math	2-Local derivations on algebras of matrix-valued functions on a compact Ayupov Shavkat Abdullayevich Institute of Mathematics, National University of Uzbekistan, Tashkent, Uzbekistan sh−[email protected] Arzikulov Farhodjon Nematjonovich (a corresponding author) Faculty of Mathematics, Andizhan State University, Andizhan, Uzbekistan [email protected] Abstract In the present paper 2-local derivations on various algebras of infinite dimensional matrix-valued functions on a compact are considered. It is proved that every 2- local derivation on such algebra is a derivation. Also we explain that the method developed in the given paper can be applied to associative, Jordan and Lie algebras of infinite dimensional matrix-valued functions on a compact. Keywords: derivation, 2-local derivation, associative algebra, C∗-algebra, a von Neumann algebra AMS 2000: 46L57, 46L40 Introduction The present paper is devoted to 2-local derivations on algebras. Recall that a 2-local derivation is defined as follows: given an algebra A, a map △ : A → A (not linear in general) is called a 2-local derivation if for every x, y ∈ A, there exists a derivation Dx,y : A → A such that △(x) = Dx,y(x) and △(y) = Dx,y(y). In 1997, P. Semrl introduced the notion of 2-local derivations and described 2-local derivations on the algebra B(H) of all bounded linear operators on the infinite-dimensional separable Hilbert space H. A similar description for the finite- dimensional case appeared later in (Kim and Kim 2004). In the paper (Lin and Wong 2006) 2-local derivations have been described on matrix algebras over finite- dimensional division rings. In (Ayupov and Kudaybergenov 2012) the authors suggested a new technique and have generalized the above mentioned results of (Semrl 1997) and (Kim and Kim 2004) for arbitrary Hilbert spaces. Namely they considered 2-local derivations on the algebra B(H) of all linear bounded operators on an arbitrary (no separability is assumed) Hilbert space H and proved that every 2-local derivation on B(H) is a derivation. After it is also published a number of paper devoted to 2-local derivations on associative algebras. In the present paper we also suggest another technique and apply to various associative algebras of infinite dimensional matrix-valued functions on a compact. As a result we will have that every 2-local derivation on such an algebra is a deriva- tion. As the main result of the paper it is established that every 2-local derivation on a ∗-algebra C(Q, Mn(F )) or C(Q, Nn(F )), where Q is a compact, Mn(F ) is the ∗-algebra of infinite dimensional matrices over complex numbers (real numbers or quaternoins) (see section 1), Nn(F ) is the ∗-algebra of infinite dimensional matri- ces over complex numbers (real numbers or quaternoins) defined in section 2, is a derivation. Also we explain that the method developed in the given paper can be 1 2 applied to Jordan and Lie algebras of infinite dimensional matrix-valued functions on a compact. We conclude that there are a number of various associative algebras of infinite di- mensional matrix-valued functions on a compact every 2-local derivation of which is a derivation. The main results of this paper are new and never proven. The method of proving of these results represented in this paper is sufficiently universal and can be applied to associative, Lie and Jordan algebras. Its respective modification al- lows to prove similar problem for Jordan and Lie algebras of infinite dimensional matrix-valued functions on a compact. 1. Preliminaries Let M be an associative algebra. Definition. A linear map D : M → M is called a derivation, if D(xy) = D(x)y + xD(y) for every two elements x, y ∈ M . A map ∆ : M → M is called a 2-local derivation, if for every two elements x, y ∈ M there exists a derivation Dx,y : M → M such that ∆(x) = Dx,y(x), ∆(y) = Dx,y(y). It is known that each derivation D on a von Neumann algebra M is an inner derivation, that is there exists an element a ∈ M such that D(x) = ax − xa, x ∈ M. Therefore for a von Neumann algebra M the above definition is equivalent to the following one: A map ∆ : M → M is called a 2-local derivation, if for every two elements x, y ∈ M there exists an element a ∈ M such that ∆(x) = ax − xa, ∆(y) = ay − ya. Let throughout the paper n be an arbitrary infinite cardinal number, Ξ be a set of indexes of the cardinality n. Let {eij} be a set of matrix units such that eij is a n × n-dimensional matrix, i.e. eij = (aαβ)αβ∈Ξ, the (i, j)-th component of which is 1, i.e. aij = 1, and the rest components are zeros. Let {mξ}ξ∈Ξ be a set of n × n-dimensional matrixes. By Pξ∈Ξ mξ we denote the matrix whose components are sums of the corresponding components of matrixes of the set {mξ}ξ∈Ξ. Let throughout the paper F = C (complexes), R(reals) or H(quaternions) and Mn(F ) = {{λijeij} : f or all indices i, j λij ∈ F, and there exists such number K ∈ R, that f or all n ∈ N and {ekl}n kl=1 ⊆ {eij}k n X kl=1 λkleklk ≤ K}, where k k is a norm of a matrix. It is easy to see that Mn(F ) is a vector space. In the vector space Mn(F ) = {{λijeij} : f or all indices i, j λij ∈ F } of all n × n-dimensional matrices (indexed sets) over F we introduce an associative if x = {λijeij }, y = {µijeij} are elements of Mn(F ) multiplication as follows: then xy = {Pξ∈Ξ λiξµξjeij }. With respect to this operation Mn(F ) becomes an associative algebra and Mn(F ) ∼= B(l2(Ξ)), where l2(Ξ) is a Hilbert space over F with elements {xi}i∈Ξ, xi ∈ F for all i ∈ Ξ, B(l2(Ξ)) is the associative algebra of all bounded linear operators on the Hilbert space l2(Ξ). Then Mn(F ) is a von Neumann algebra of infinite n × n-dimensional matrices over F . 3 Similarly, if we take the algebra B(H) of all bounded linear operators on an arbitrary Hilbert space H and if {qi} is an arbitrary maximal orthogonal set of minimal projections of B(H), then B(H) = P⊕ ij qiB(H)qj (see (Arzikulov 2008)). Let throughout the paper X be a hyperstonean compact, C(X) denote the al- gebra of all F -valued continuous functions on X and M = {{λij(x)eij }ij∈Ξ : (∀ij λij (x) ∈ C(X)) (∃K ∈ R)(∀m ∈ N )(∀{ekl}m kl=1 ⊆ {eij})k X λkl(x)eklk ≤ K}, kl=1...m where k Pkl=1...m λkl(x)eklk ≤ K means (∀xo ∈ X)k Pkl=1...m λkl(xo)eklk ≤ K. The set M is a vector space with point-wise algebraic operations. The map k k : M → R+ defined as kak = sup kl=1⊆{eij } {ekl}n n k X kl=1 λkl(x)eklk, is a norm on the vector space M, where a ∈ M and a = {λij(x)eij }. In the vector space V = {{λijeij}ij : {λij} ⊂ C(X)} of all infinite n × n- dimensional matrices (indexed sets) over C(X) we introduce an associative mul- if x = {λij (x)eij }, y = {µij(x)eij } are elements of V then tiplication as follows: xy = {Pξ λiξ(x)µξj (x)eij }. With respect to this multiplication M becomes an associative algebra and M ∼= C(X) ⊗ Mn(F ). Thus M is a real or complex von Neumann algebra of type In. Let M be a C∗-algebra, △ : M → M be a 2-local derivation. Now let us show that △ is homogeneous. Indeed, for each x ∈ M , and for λ ∈ C there exists a derivation Dx,λx such that △(x) = Dx,λx(x) and △(λx) = Dx,λx(λx). Then △(λx) = Dx,λx(λx) = λDx,λx(x) = λ △ (x). Hence, △ is homogenous. At the same time, for each x ∈ M , there exists a derivation Dx,x2 such that △(x) = Dx,x2(x) and △(x2) = Dx,x2(x2). Then △(x2) = Dx,x2(x2) = Dx,x2(x)x + xDx,x2(x) = △(x)x + x △ (x). In (Bresar 1988) it is proved that every Jordan derivation on a semi-prime algebra is a derivation. Since M is semi-prime (i.e. aM a = {0} implies that a = {0}), the map △ is a derivation if it is additive. Therefore, to prove that the 2-local derivation △ : M → M is a derivation it is sufficient to prove that △ : M → M is additive in the proof of theorem 1. 2. 2-local derivations on some associative algebras of matrix-valued functions Let Q be a compact. Then there exists a hyperstonean compact X such that for the algebra C(Q) of all continuous complex number-valued functions on Q we have C(Q)∗∗ ∼= C(X). If we take the ∗-algebra C(Q, Mn(C)) of all continuous maps of Q to Mn(C), then we may assume that C(Q, Mn(C)) ⊆ M. In this case the set {eij} of constant functions belongs to C(Q, Mn(C)) and the weak closure of C(Q, Mn(C)) in M coincides with M. Hence by separately weakly continuity of multiplication every derivation of C(Q, Mn(C)) has a unique extension to a derivation on M. Therefore, if △ is a 2-local derivation on C(Q, Mn(C)), then for every two elements x, y ∈ C(Q, Mn(C)) there exists a derivation Dx,y : M → M 4 such that △(x) = Dx,y(x), △(y) = Dx,y(y), i.e. Dx,y is a derivation of M (not only of C(Q, Mn(C))). The following theorem is the key result of this section. Theorem 1. Let △ be a 2-local derivation on C(Q, Mn(C)). Then △ is a derivation. First let us prove lemmata which are necessary for the proof of theorem 1. By the above arguments for every 2-local derivation △ on C(Q, Mn(F )) and for each x ∈ C(Q, Mn(F )) there exist a ∈ M such that Put △(x) = ax − xa. eij := {λξηeξη}, where for all ξ, η, if ξ = i, η = j then λξη = 1, else λξη = 0, 1 is unit of the algebra C(Q). Let {a(ij)} ⊂ M be a subset such that △(eij) = a(ij)eij − eija(ij). for all i, j, let aij eij, aij ∈ C(Q), be the (i,j)-th component of the element eiia(ji)ejj of M for all pairs of different indexes i, j and let {aξηeξη}ξ6=η be the matrix in V with all such components, the diagonal components of which are zeros. Lemma 2. For each pair i, j of different indices the following equality is valid △(eij) = {aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η + a(ij)iieij − eija(ij)jj , (1) where a(ij)ii, a(ij)jj are functions in C(Q) which are the coefficients of the Peirce components eiia(ij)eii, ejja(ij)ejj . Proof. Let k be an arbitrary index different from i, j and let a(ij, ik) ∈ M be an element such that △(eik) = a(ij, ik)eik − eika(ij, ik) and △ (eij) = a(ij, ik)eij − eija(ij, ik). Then ekk △ (eij)ejj = ekk(a(ij, ik)eij − eija(ij, ik))ejj = ekka(ij, ik)eij − 0 = ekka(ik)eij − ekkeij{aξηeξη}ξ6=ηejj = ekkakieij − ekkeij{aξηeξη}ξ6=ηejj = ekk{aξηeξη}ξ6=ηeij − ekkeij{aξηeξη}ξ6=ηejj = ekk({aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η)ejj. Similarly, ekk △ (eij )eii = ekk({aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η)eii. Let a(ij, kj) ∈ M be an element such that △(ekj) = a(ij, kj)ekj − ekja(ij, kj) and △ (eij) = a(ij, kj)eij − eija(ij, kj). Then eii △ (eij)ekk = eii(a(ij, kj)eij − eija(ij, kj))ekk = 0 − eija(ij, kj)ekk = 0 − eija(kj)ekk = 0 − eij ajkekk = eii{aξηeξη}ξ6=ηeijekk − eij{aξηeξη}ξ6=ηekk = eii({aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η)ekk. Also similarly we have ejj △ (eij)ekk = ejj ({aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η)ekk, eii △ (eij )eii = eii({aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η)eii, 5 ejj △ (eij)ejj = ejj ({aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η)ejj . Hence the equality (1) is valid. ⊲ We take elements of the sets {{eiξ}ξ}i and {{eξj}ξ}j in pairs ({eαξ}ξ, {eξβ}ξ) such that α 6= β. Then using the set {({eαξ}ξ, {eξβ}ξ)} of such pairs we get the set {eαβ}. Let xo = {eαβ} be a set {vijeij}ij such that for all i, j if (α, β) 6= (i, j) then vij = 0 ∈ C(Q) else vij = 1 ∈ C(Q). Then xo ∈ C(Q, Mn(C)). Fix different indices io, jo. Let c ∈ M be an element such that △(eiojo ) = ceiojo − eiojo c and △ (xo) = cxo − xoc. Put c = {cijeij} ∈ M and ¯a = {aijeij}i6=j ∪ {aiieii}, where {aiieii} = {ciieii}. Lemma 3. Let ξ, η be arbitrary different indices, and let b = {bijeij} ∈ M be an element such that △(eξη) = beξη − eξηb and △ (xo) = bxo − xob. Then cξξ − cηη = bξξ − bηη. Proof. We have that there exist ¯α, ¯β such that eξ ¯α, e ¯βη ∈ {eαβ} (or e ¯αη, eξ ¯β ∈ {eαβ}, or e ¯α, ¯β ∈ {eαβ}), and there exists a chain of pairs of indexes (α, β) in Ω, where Ω = {(α, β) : e α, β ∈ {eαβ}}, connecting pairs (ξ, ¯α), ( ¯β, η) i.e., (ξ, ¯α), (¯α, ξ1), (ξ1, η1), . . . , (η2, ¯β), ( ¯β, η). Then cξξ − c ¯α ¯α = bξξ − b ¯α ¯α, c ¯α ¯α − cξ1ξ1 = b ¯α ¯α − bξ1ξ1 , cξ1ξ1 − cη1η1 = bξ1ξ1 − bη1η1 , . . . , cη2η2 − c Hence ¯β ¯β = bη2η2 − b ¯β ¯β, c ¯β ¯β − cηη = b ¯β ¯β − bηη. cξξ − bξξ = c ¯α ¯α − b ¯α ¯α, c ¯α ¯α − b ¯α ¯α = cξ1ξ1 − bξ1ξ1 , ¯β ¯β − b ¯β ¯β − b cξ1ξ1 − bξ1ξ1 = cη1η1 − bη1η1 , . . . , cη2η2 − bη2η2 = c and cξξ − bξξ = cηη − bηη, cξξ − cηη = bξξ − bηη. ¯β ¯β, c ¯β ¯β = cηη − bηη. Therefore cξξ − cηη = bξξ − bηη. ⊲ Lemma 4. Let x be an element of the algebra C(Q, Mn(C)). Then where ¯a is defined as above. Proof. Let d(ij) ∈ M be an element such that △(x) = ¯ax − x¯a, △(eij) = d(ij)eij − eijd(ij) and △ (x) = d(ij)x − xd(ij) and i 6= j. Then △(eij) = d(ij)eij − eijd(ij) = eiid(ij)eij − eijd(ij)ejj + (1 − eii)d(ij)eij − eijd(ij)(1 − ejj ) = a(ij)iieij − eija(ij)jj + {aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η for all i, j by lemma 2. Since we have eiid(ij)eij − eijd(ij)ejj = a(ij)iieij − eija(ij)jj (1 − eii)d(ij)eii = {aξηeξη}ξ6=ηeii, ejj d(ij)(1 − ejj ) = ejj{aξηeξη}ξ6=η 6 for all different i and j. Let b = {bijeij} ∈ M be an element such that △(eij ) = beij − eijb and △ (xo) = bxo − xob. Then bii − bjj = cii − cjj by lemma 3. We have bii − bjj = d(ij)ii − d(ij)jj since beij − eijb = d(ij)eij − eijd(ij). Hence cii − cjj = d(ij)ii − d(ij)jj , cjj − cii = d(ij)jj − d(ij)ii. Therefore we have ejj △ (x)eii = ejj (d(ij)x − xd(ij))eii = ejj d(ij)(1 − ejj )xeii + ejjd(ij)ejj xeii − ejjx(1 − eii)d(ij)eii − ejj xeiid(ij)eii = ejj{aξηeξη}ξ6=ηxeii − ejj x{aξηeξη}ξ6=ηeii + ejjd(ij)ejj xeii − ejjxeiid(ij)eii = ejj{aξηeξη}ξ6=ηxeii − ejj x{aξηeξη}ξ6=ηeii + cjj ejjxeii − ejjxeiiciieii = ejj{aξηeξη}ξ6=ηxeii − ejj x{aξηeξη}ξ6=ηeii+ ejj(X aξξeξξ)eii = aξξeξξ)xeii − ejjx(X ξ ξ ejj {aξηeξη}xeii − ejj x{aξηeξη}eii = ejj(¯ax − x¯a)eii. Let d(ii), v, w ∈ M be elements such that △(eii) = d(ii)eii − eiid(ii) and △ (x) = d(ii)x − xd(ii), △(eii) = veii − eiiv, △(eij) = veij − eijv, and Then and By lemma 2 and Similarly Hence △(eii) = weii − eiiw, △(eji) = weji − ejiw. (1 − eii)a(ij)eii = (1 − eii)veii = (1 − eii)d(ii)eii, eiia(ji)(1 − eii) = eiiw(1 − eii) = eiid(ii)(1 − eii). △(eij) = a(ij)eij − eija(ij) = {aξηeξη}ξ6=ηeij − eij{aξηeξη}ξ6=η + a(ij)iieij − eij a(ij)jj (1 − eii)a(ij)eii = {aξηeξη}ξ6=ηeii. eiia(ji)(1 − eii) = eii{aξηeξη}ξ6=η. eii △ (x)eii = eii(d(ii)x − xd(ii))eii = eiid(ii)(1 − eii)xeii + eiid(ii)eiixeii − eiix(1 − eii)d(ii)eii − eiixeiid(ii)eii = eiia(ji)(1 − eii)xeii + eiid(ii)eiixeii − eiix(1 − eii)a(ij)eii − eiixeiid(ii)eii = eii{aξηeξη}ξ6=ηxeii − eiix{aξηeξη}ξ6=ηeii + eiid(ii)eiixeii − eiixeiid(ii)eii = eii{aξηeξη}ξ6=ηxeii − eiix{aξηeξη}ξ6=ηeii + ciieiixeii − eiixciieii = eii{aξηeξη}ξ6=ηxeii − eiix{aξηeξη}ξ6=ηeii+ eii(X aξξeξξ)eii = aξξeξξ)xeii − eiix(X ξ ξ 7 eii{aξηeξη}xeii − eiix{aξηeξη}eii = eii(¯ax − x¯a)eii. Hence for all x ∈ C(Q, Mn(C)). ⊲ △(x) = ¯ax − x¯a Proof of theorem 1. By lemma 4 △(eii) = ¯aeii − eii¯a ∈ M. Hence X aξieξi − X aiξeiξ ∈ M. ξ ξ Then and eii(X aξieξi − X aiξeiξ) = aiieii − X aiξeiξ ∈ M ξ ξ ξ (X aξieξi − X aiξeiξ)eii = X aξieξi − aiieii ∈ M. ξ ξ ξ Therefore Pξ aξieξi, Pξ aiξeiξ ∈ M i.e., ¯aeii, eii¯a ∈ M. Hence eii¯ax, x¯aeii ∈ M for each i and for each element x = {xij eij} ∈ C(Q, Mn(C)), i.e., ¯ax, x¯a ∈ V X aiξxξjeij , X xiξaξj eij ∈ C(Q)eij ξ ξ for all i, j. Therefore for all x, y ∈ C(Q, Mn(C)) we have that the elements ¯ax, x¯a, ¯ay, y¯a, ¯a(x + y), (x + y)¯a belong to V . Hence △(x + y) = △(x) + △(y) by lemma 4. Similarly for all x, y ∈ C(Q, Mn(C)) we have (¯ax + x¯a)y = ¯axy − x¯ay ∈ M, ¯axy = ¯a(xy) ∈ V. Then x¯ay = ¯axy − (¯ax − x¯a)y and x¯ay ∈ V . Therefore ¯a(xy) − (xy)¯a = ¯axy − x¯ay + x¯ay − xy¯a = (¯ax − x¯a)y + x(¯ay − y¯a). Hence △(xy) = △(x)y + x △ (y) by lemma 4. By section 1 △ is homogeneous. Hence, △ is a linear operator and a derivation. The proof is complete. ⊲ If we take the ∗-algebra C(Q, Mn(F )), F = R or H, then we can similarly prove the following theorem. Theorem 2. Let △ be a 2-local derivation on C(Q, Mn(F )). Then △ is a derivation. Note, for theorem 2 to be proved the proof of theorem 1 will be repeated with very minor modification. Let o X ij F eij = {{λijeij} : f or all indices i, j λij ∈ F, and ∀ε > 0∃no ∈ N such that ∀n ≥ m ≥ no 8 n k X (λkieki + λikeik) + λiieii]k < ε}. [ X k=1,...,i−1 i=m where k k is a norm of a matrix. Then Po ij F eij is a C∗-algebra with respect to com- ponentwise algebraic operations, the bilinear operation and the norm (Arzikulov 2012). Since F eij is a simple C∗-algebra for all i, then by the proof of theorem 8 in (Arzikulov 2012) the C∗-algebra Po ij F eij. Then C(Q, Nn(F )) is a real or complex C∗-algebra, where (F = C, R or H) and C(Q, Nn(F )) ⊆ M. Hence similar to theorems 1, 2 we can prove the following theorem ij F eij is simple. Let Nn(F ) = Po Theorem 3. Let △ be a 2-local derivation on C(Q, Nn(F )). Then △ is a derivation. It is known that the set Msa of all self-adjoint elements (i.e. a∗ = a) of M forms 2 (ab + ba). a Jordan algebra with respect to the operation of multiplication a · b = 1 The following problem can be similarly solved. Problem 1. Develop a Jordan analog of the method applied in the proof of theorem 1 and prove that every 2-local derivation △ on the Jordan algebra Msa or C(Q, Mn(F )sa) or C(Q, Nn(F )sa) is a derivation. It is known that the set Mk = {a ∈ M : a∗ = −a} forms a Lie algebra with respect to the operation of multiplication [a, b] = ab−ba. So it is natural to consider the following problem. Problem 2. Develop a Lie analog of the method applied in the proof of theorem 1 and prove that every 2-local derivation △ on the Lie algebra Mk or C(Q, Mn(F )k) or C(Q, Nn(F )k) is a derivation. Acknowledgements The authors want to thank K.K.Kudaybergenov for many stimulating conversa- tions on the subject. References Semrl P (1997) Local automorphisms and derivations on B(H). Proc Amer Math Soc 125:2677-2680 Kim S O, Kim J S (2004) Local automorphisms and derivations on Mn. Proc Amer Math Soc 132:1389-1392 Lin Y, Wong T (2006) A note on 2-local maps. Proc Edinb Math Soc 49:701-708. Ayupov Sh A, Kudaybergenov K K (2012) 2-local derivations and automor- phisms on B(H) 395:15-18 Arzikulov F N (2008) Infinite order and norm decompositions of C-algebras. Int Journal of Math Analysis 2(5):255-262 Arzikulov F N (2012) Infinite norm decompositions of C-algebras. In: Operator Theory: Advances and Applications, vol 220. Springer Basel AG pp. 11-21 Bresar M (1988) Jordan derivations on semiprime rings. Proc Amer Math Soc 104:1003-1006
1511.08014	1	1511	2015-11-25T10:58:18	A characterization of reflexive spaces of operators	[ "math.OA" ]	We show that for a linear space of operators ${\mathcal M}\subseteq {\mathcal B}(H_1,H_2)$ the following assertions are equivalent. (i) ${\mathcal M} $ is reflexive in the sense of Loginov--Shulman. (ii) There exists an order-preserving map $\Psi=(\psi_1,\psi_2)$ on a bilattice $Bil({\mathcal M})$ of subspaces determined by ${\mathcal M}$, with $P\leq \psi_1(P,Q)$ and $Q\leq \psi_2(P,Q)$, for any pair $(P,Q)\in Bil({\mathcal M})$, and such that an operator $T\in {\mathcal B}(H_1,H_2)$ lies in ${\mathcal M}$ if and only if $\psi_2(P,Q)T\psi_1(P,Q)=0$ for all $(P,Q)\in Bil( {\mathcal M})$. This extends to reflexive spaces the Erdos--Power type characterization of weakly closed bimodules over a nest algebra.	math.OA	math	A CHARACTERIZATION OF REFLEXIVE SPACES OF OPERATORS JANKO BRA CI C AND LINA OLIVEIRA Abstract. We show that for a linear space of operators M ⊆ B(H1, H2) the following assertions are equivalent. (i) M is reflexive in the sense of Loginov -- Shulman. (ii) There exists an order-preserving map Ψ = (ψ1, ψ2) on a bilattice Bil(M) of subspaces determined by M, with P ≤ ψ1(P, Q) and Q ≤ ψ2(P, Q), for any pair (P, Q) ∈ Bil(M), and such that an operator T ∈ B(H1, H2) lies in M if and only if ψ2(P, Q)T ψ1(P, Q) = 0 for all (P, Q) ∈ Bil(M). This extends to reflexive spaces the Erdos -- Power type characterization of weakly closed bimodules over a nest algebra. 1. Introduction and preliminaries In [2], Erdos and Power characterized the weakly closed bimodules of a nest algebra in terms of order homomorphisms on the lattice of invariant subspaces of the algebra. Deguang showed in [1] that, given any reflexive subalgebra σ- weakly generated by its rank one operators, the σ-weakly closed bimodules over the algebra could analogously be characterized in terms of order homomorphisms on the lattice of invariant subspaces of the algebra. Li and Li [6, Proposition 2.6] have extended the mentioned results to the realm of Banach spaces. It is worth noticing that the bimodules considered in the Erdos -- Power theorems are implicitly reflexive subspaces in the sense of Loginov -- Shulman (cf. [7]). The aim of the present paper is to extend this type of characterization to all such reflexive subspaces. The main result Theorem 3.5 shows that, for every reflexive space M of operators between two complex Hilbert spaces, there exists an order homomorphism on a bilattice of subspaces determined by M which describes this subspace in the sense of Erdos -- Power [2, Theorem 1.5]. The proof of Theorem 3.5 requires some auxiliary results appearing in Section 2. In the rest of the present section, apart from the notation, we shall also establish the facts about bilattices needed in the sequel. Let H be a complex Hilbert space, let B(H) be the Banach algebra of all bounded linear operators on H, and let P(H) be the set of all orthogonal projections on H. It is well known that P(H) is a lattice when endowed with the partial order relation defined, for all P1, P2 ∈ H, by P1 ≤ P2 ⇐⇒ P1H ⊆ P2H. The join P1 ∨ P2 is the orthogonal projection onto P1H + P2H and the meet P1 ∧ P2 is the orthogonal projection onto P1H ∩ P2H. In fact, P(H) is a complete lattice whose top and bottom elements are, respectively, the identity operator I and the zero operator 0. 2010 Mathematics Subject Classification. Primary 47A15. Key words and phrases. Reflexive space of operators, order-preserving map. The first author was supported by the Slovenian Research Agency through the research program P1-0222. The second author is partially funded by FCT/Portugal through the projects PEst- OE/EEI/LA0009/2013 and EXCL/MAT-GEO/0222/2012. 1 2 JANKO BRA CI C AND LINA OLIVEIRA Recall that the lattice Lat(U) of invariant subspaces of a subset U of B(H) is given by Lat(U) = {P ∈ P(H); P ⊥T P = 0, for all T ∈ U}, where P ⊥ = I − P . It is clear that Lat(U) is a sublattice of P(H) which is strongly closed and therefore complete, i.e., for every subset F ⊆ Lat(U), the supremum ∨F and the infimum ∧F lie in Lat(U) (see [5]). If U ⊆ B(H) is a non-empty subset, then let U ∗ = {T ∗; T ∈ U}. We say that U is selfadjoint if U ∗ = U. It is obvious that P ∈ Lat(U) if and only if P ⊥ ∈ Lat(U ∗), i.e., Lat(U ∗) = Lat(U)⊥. Let H1, H2 be complex Hilbert spaces. We endow the Cartesian product P(H1) ×P(H2) with the partial order (cid:22) which is defined, for all (P1, Q1), (P2, Q2) ∈ P(H1) × P(H2), by (1.1) (P1, Q1) (cid:22) (P2, Q2) if and only if P1 ≤ P2 and Q1 ≥ Q2. Hence the operations of join and meet are given, respectively, by (1.2) (P1, Q1) ∨ (P2, Q2) = (P1 ∨ P2, Q1 ∧ Q2) and (P1, Q1) ∧ (P2, Q2) = (P1 ∧ P2, Q1 ∨ Q2). It follows that P(H1) × P(H2) together with (cid:22) is a lattice as it contains all the binary joins and meets. From now on we write P(H1) ×(cid:22) P(H2) whenever we consider the Cartesian product to be endowed with the partial order (1.1), i.e., with the lattice structure (1.2). The corresponding notation will also be used for Cartesian products of subsets of P(H1) × P(H2). Unless otherwise stated, it is assumed that the partial order under consideration will always be (cid:22). Following [8], we call a subset L of P(H1) ×(cid:22) P(H2) a bilattice if it is closed under the lattice operations (1.2) and contains the pairs (0, 0), (0, I), and (I, 0). Examples of bilattices are P(H1) ×(cid:22) P(H2), of course, and (1.3) BIL(U) = {(P, Q) ∈ P(H1) ×(cid:22) P(H2); QT P = 0, for any T ∈ U}, where U ⊆ B(H1, H2) is an arbitrary non-empty set. Recall that, for a non-empty family F ⊆ P(H), Alg(F ) = {T ∈ B(H); P ⊥T P = 0, for all P ∈ F } is a weakly closed subalgebra of B(H) that contains the identity operator. A sub- algebra A of B(H) is said to be reflexive if AlgLat(A) = A. The notion of reflexive algebras has been generalized in several different directions; see [3] for a general view of reflexivity and [4] for a recently introduced generalization. The concept of reflexivity is naturally extended to spaces of operators as follows. For a non-empty family F ⊆ P(H1) ×(cid:22) P(H2), let Op(F ) = {T ∈ B(H1, H2); QT P = 0, for all (P, Q) ∈ F }. It is easily seen that Op(F ) is a weakly closed linear subspace of B(H1, H2). A subspace M ⊆ B(H1, H2) is said to be reflexive if OpBIL(M) = M. This definition is equivalent to that of Loginov and Shulman in [7], where a subspace M is said to be reflexive if M coincides with its reflexive cover Ref (M) = {S ∈ B(H1, H2); Sx ∈ Mx, for all x ∈ H1}. In fact, OpBIL(M) = Ref (M) (cf. [8, p. 298]). A CHARACTERIZATION OF REFLEXIVE SPACES OF OPERATORS 3 Remark 1.1. Notice that, if A ⊆ B(H) is an algebra containing the identity operator, then Ref (A) = AlgLat(A). 2. Subspaces and modules For a linear subspace M ⊆ B(H1, H2), let (2.1) and (2.2) AM = {A ∈ B(H1); T A ∈ M, for all T ∈ M} BM = {B ∈ B(H2); BT ∈ M, for all T ∈ M}. It is easily seen that AM and BM are algebras containing the identity operator and that M is BM-AM-bimodule. It is clear that these are the largest subalgebras of B(H1), respectively B(H2), such that M is a bimodule over them. If M is closed (respectively, weakly closed), then BM and AM are closed (respectively, weakly closed). Next we show that AM and BM are reflexive whenever M is a reflexive space. Proposition 2.1. If M ⊆ B(H1, H2) is a reflexive space, then AM and BM are reflexive algebras. Proof. It will only be shown that AM is reflexive since the reflexivity of BM can be similarly proved. In view of Remark 1.1, it suffices to show that Ref (AM) = AM. In other words, fixing S ∈ Ref (AM), we need to show that, for all T ∈ M, the operator T S lies in M. Since this is trivially satisfied by T = 0, henceforth we shall assume that T 6= 0. Let x ∈ H1 and ε > 0 be arbitrary. Since S ∈ Ref (AM), there exists Ax,ε ∈ AM such that kSx − Ax,εxk < ε/kT k. Hence kT Sx − T Ax,εxk ≤ kT kkSx − Ax,εxk < ε. The operator T Ax,ε lies in M and, therefore, we can conclude that T S ∈ Ref (M) = M, as required. (cid:3) Corollary 2.2. Let M be a linear subspace of B(H1, H2). Then Ref(cid:0)AM(cid:1) ⊆ ARef (M) and Ref(cid:0)BM(cid:1) ⊆ BRef (M). Proof. Let A ∈ AM. If T ∈ Ref (M), then, for any x ∈ H1 and any ε > 0, there exists Sx,ε ∈ M such that kT Ax − Sx,εAxk < ε. Since Sx,εA ∈ M, we conclude that T A ∈ Ref (M). By Proposition 2.1, the algebra ARef (M) is reflexive, from which follows that Ref(cid:0)AM(cid:1) ⊆ Ref(cid:0)ARef (M)(cid:1) = ARef (M). The proof of the second inclusion is similar. (cid:3) Let tr(·) be the trace functional and let C1(H) ⊆ B(H) be the ideal of trace-class operators. The dual of C1(H) can be identified with B(H) by means of the pairing hC, Ai = tr(CA∗), with C ∈ C1(H), A ∈ B(H). The preannihilator of a subset U ⊆ B(H) is U⊥ = {C ∈ C1(H); tr(CA∗) = 0, for all A ∈ U} and the annihilator of V ⊆ C1(H) is V ⊥ = {A ∈ B(H); tr(CA∗) = 0, for all C ∈ V}. It is obvious that U⊥ and V ⊥ are linear spaces and that a linear subspace M ⊆ B(H) is σ-weakly closed if and only if M = (M⊥)⊥. If U, V are two non-empty sets of operators, then we denote by UV the set of all products T S, where T ∈ U and S ∈ V. Proposition 2.3. Let M be a linear subspace of B(H). Then the following asser- tions hold. (i) (AM)∗ = BM∗ . 4 JANKO BRA CI C AND LINA OLIVEIRA (ii) If M is σ-weakly closed, then AM = (M∗M⊥)⊥ and BM = (MM⊥)⊥. (iii) If M is selfadjoint and σ-weakly closed, then AM = BM is a C ∗-algebra. (iv) If M is selfadjoint, σ-weakly closed, and reflexive, then AM = BM is a von Neumann algebra. Proof. (i) An operator A ∈ B(H) lies in (AM)∗ if and only if T A∗ ∈ M for every T ∈ M. However this is equivalent to AT ∗ ∈ M∗ for any T ∗ ∈ M∗, which by the definition means that A ∈ BM∗ . (ii) Let A ∈ AM. For arbitrary T ∈ M and C ∈ M⊥, we have tr(T ∗CA∗) = tr(C(T A)∗) = 0, since T A ∈ M. This proves that A ∈ (M∗M⊥)⊥. On the other hand, if A ∈ (M∗M⊥)⊥ and T ∈ M, then tr(C(T A)∗) = tr(T ∗CA∗) = 0, for any C ∈ M⊥. Hence T A ∈ (M⊥)⊥ = M. The second equality is similarly proved. (iii) It follows from (ii) that AM = BM. By (i), the algebra AM is selfadjoint and closed since M itself is closed. Hence AM is a C ∗-algebra. (iv) By (iii), AM = BM is a C ∗-algebra. However, since M is reflexive, it is (cid:3) weakly closed and, therefore, AM is also weakly closed. 3. A characterization of reflexivity Let M ⊆ B(H1, H2) be a linear subspace and let AM and BM be the algebras defined in (2.1) -- (2.2). The associated bilattice BIL(M) (see (1.3)) is very large. For our purposes it suffices to consider a smaller bilattice to be defined below. Firstly, we state the following lemma which is just a formalization of a remark in [8, p. 298]. We include a short proof. Lemma 3.1. Let M be a linear subspace of B(H1, H2). For any pair (P, Q) ∈ BIL(M), there exists a pair (P ′, Q′) ∈ BIL(M) such that P ′ ∈ Lat(AM), Q′ ∈ Lat(BM)⊥, P ≤ P ′, and Q ≤ Q′. Proof. Let P ′ be the orthogonal projection onto AMP H1 and let Q′ be the or- thogonal projection onto B∗ MQH2. It is obvious that AMP H1 is invariant for any M. Hence P ′ ∈ Lat(AM) and A ∈ AM and that B∗ Q′ ∈ Lat(B∗ MQH2, since both algebras contain the identity operator. Consequently, P ≤ P ′ and Q ≤ Q′. M) = Lat(BM)⊥. Observe that P H1 ⊆ AMP H1 and QH2 ⊆ B∗ MQH2 is invariant for any B∗ ∈ B∗ To prove that (P ′, Q′) lies in BIL(M), we have to see that, for any T ∈ M, the equality Q′T P ′ = 0 holds, i.e., T P ′H1 ⊥ Q′H2. Let x ∈ H1 be arbitrary. For any ε > 0, there exist Aε ∈ AM and xε ∈ H1 such that kP ′x − AεP xεk < ε, and therefore kT P ′x − T AεP xεk < εkT k. For arbitrary B∗ ∈ B∗ M and y ∈ H2, we have hT AεP xε, B∗Qyi = hQBT AεP xε, yi = 0, since BT Aε ∈ M. Hence hT P ′x, B∗Qyi = hT P ′x − T AεP xε, B∗Qyi ≤ kT P ′x − T AεP xεkkB∗Qyk < εkT kkB∗Qyk, yielding T P ′x ⊥ B∗QH2, from which it follows that T P ′H1 ⊥ Q′H2. (cid:3) Let Bil(M) = BIL(M) ∩ (cid:0)Lat(AM) ×(cid:22) Lat(BM)⊥(cid:1). It is clear that Bil(M) is a bilattice. Proposition 3.2. Let M be a linear subspace of B(H1, H2). Then OpBIL(M) = OpBil(M). A CHARACTERIZATION OF REFLEXIVE SPACES OF OPERATORS 5 Proof. Since Bil(M) is a subset of BIL(M), it follows that OpBIL(M) ⊆ OpBil(M). To show that the reverse inclusion also holds, we begin by fixing an operator T ∈ OpBil(M) and a pair of projections (P, Q) ∈ BIL(M). By Lemma 3.1, there exists a pair (P ′, Q′) ∈ Bil(M) such that P ≤ P ′ and Q ≤ Q′. Hence P ′P = P and QQ′ = Q. It follows that QT P = QQ′T P ′P = 0 and, therefore, T lies in OpBIL(M), as required. (cid:3) Let M ⊆ B(H1, H2) be a linear space. Define φ : Lat(AM) → Lat(BM)⊥ by (3.1) φ(P ) = ∨{Q ∈ Lat(BM)⊥; (P, Q) ∈ Bil(M)}, and similarly define θ : Lat(BM)⊥ → Lat(AM) by (3.2) θ(Q) = ∨{P ∈ Lat(AM); (P, Q) ∈ Bil(M)}. Observe that none of the sets appearing in (3.1) -- (3.2) is empty as (P, 0), (0, Q) ∈ Bil(M), for any P ∈ Lat(AM), Q ∈ Lat(BM)⊥. The next proposition lists some properties of the maps φ and θ. Proposition 3.3. Let M be a linear subspace of B(H1, H2) and let φ, θ be the maps defined in (3.1) -- (3.2). Then the following assertions hold. (i) φ and θ are order-reversing maps. (ii) (P, φ(P )), (θ(Q), Q) ∈ Bil(M), for any P ∈ Lat(AM) and Q ∈ Lat(BM)⊥. (iii) If C ⊆ Lat(AM) and D ⊆ Lat(BM)⊥ are non-empty sets, then φ(∨C) = ∧φ(C) and θ(∨D) = ∧θ(D). (iv) P ≤ θφ(P ) and Q ≤ φθ(Q), for all P ∈ Lat(AM) and Q ∈ Lat(BM)⊥. (v) φθφ = φ and θφθ = θ. Proof. Assertions (i) -- (iv) will only be proved for the map φ, since the corresponding assertions concerning the map θ can be similarly proved. For the same reason, only the first equality in (v) will be proved. (i) If P1, P2 ∈ Lat(AM) are such that P1 ≤ P2, then P1P2 = P1 = P2P1. Hence, if Q is a projection in P(H2) with (P2, Q) ∈ Bil(M), then, for every T ∈ M, we have QT P1 = QT P2P1 = 0, yielding (P1, Q) ∈ Bil(M). It follows that φ(P2) = ∨{Q ∈ Lat(BM)⊥; (P2, Q) ∈ Bil(M)} ≤ ∨{Q ∈ Lat(BM)⊥; (P1, Q) ∈ Bil(M)} = φ(P1). (ii) Let P ∈ Lat(AM). We have to show that φ(P )T P = 0, for every T ∈ M. Let T ∈ M, x ∈ H1, y ∈ H2 be arbitrary, and let Q ∈ P(H2) be a projection such that (P, Q) ∈ Bil(M). Then hT P x, Qyi = hQT P x, yi = 0, that is to say that T P H1 ⊥ QH2. Since φ(P ) is the orthogonal projection onto the closed linear span of all the spaces QH2, where Q is an orthogonal projection in P(H2) such that (P, Q) ∈ Bil(M), we conclude that T P H1 ⊥ φ(P )H2, i.e., φ(P )T P = 0. (iii) Let C ⊆ Lat(AM) be a non-empty set. Then, for all P ∈ C, P ≤ ∨C and, since Lat(AM) is complete, ∨C ∈ Lat(AM). It follows that, for all P ∈ C, φ(∨C) ≤ φ(P ), as φ is an order-reversing map. Therefore φ(∨C) ≤ ∧φ(C). To show that this inequality can be reversed, we shall prove firstly that (∨C, ∧φ(C)) ∈ Bil(M). Let T ∈ M be arbitrary. Then, for every P ∈ C, we have ∧φ(C) ≤ φ(P ), from which it follows that (cid:0)∧φ(C)(cid:1)φ(P ) = ∧φ(C). Hence, for all P ∈ C, (cid:0)∧φ(C)(cid:1)T P = (cid:0)∧φ(C)(cid:1)φ(P )T P = 0 6 JANKO BRA CI C AND LINA OLIVEIRA and, consequently, (cid:0)∧φ(C)(cid:1)T (cid:0)∨C(cid:1) = 0, i.e., (cid:0)∨C, ∧φ(C)(cid:1) ∈ Bil(M). It follows, by the definition of φ(∨C), that ∧φ(C) ≤ φ(∨C). (iv) Let P ∈ Lat(AM). By assertion (ii), we have (P, φ(P )), (θ(φ(P )), φ(P )) ∈ Bil(M). Since, by the definition (3.1), the projection θ(φ(P )) is the largest P ′ ∈ Lat(AM) such that (P ′, φ(P )) ∈ Bil(M), we conclude that P ≤ θ(cid:0)φ(P )(cid:1). (v) Let P ∈ Lat(AM) be arbitrary. By assertion (iv), we know that φ(P ) ≤ φθφ(P ). Moreover, since by (ii) of this proposition, (P, φ(P )) and (θφ(P ), φθφ(P )) lie in the bilattice Bil(M), we have (cid:0)P ∧ θφ(P ), φ(P ) ∨ φθφ(P )(cid:1) ∈ Bil(M). Notice however that (iv) implies P ∧ θφ(P ) = P and φ(P ) ∨ φθφ(P ) = φθφ(P ). Thus, (P, φθφ(P )) ∈ Bil(M). By the definition of φ, the projection φ(P ) is the largest Q ∈ Lat(BM)⊥ having the property (P, Q) ∈ Bil(M). Hence, φθφ(P ) ≤ φ(P ). Consequently, for all P ∈ Lat(AM), we have φθφ(P ) = φ(P ). (cid:3) Let Ψ1, Ψ2 : Bil(M) → Bil(M) be defined by (3.3) Ψ1(P, Q) = (θφ(P ), φ(P )) Ψ2(P, Q) = (θ(Q), φθ(Q)) and (P, Q) ∈ Bil(M). Observe that Proposition 3.3 (ii) guarantees that the maps Ψ1 and Ψ2 are well defined. Corollary 3.4. Let M be a linear subspace of B(H1, H2) and let Ψ1, Ψ2 : Bil(M) → Bil(M) be the maps defined in (3.3). Then Ψ1, Ψ2 are order-preserving maps and Ψ1(Bil(M)) = Ψ2(Bil(M)). Proof. It easily follows from Proposition 3.3 (i) that Ψ1 and Ψ2 are order-preser- ving maps. That the images of Ψ1 and Ψ2 coincide is an immediate consequence of Proposition 3.3 (v). (cid:3) We are now able to prove our main result. Theorem 3.5. Let M be a linear subspace of B(H1, H2) and let AM, BM be the algebras defined in (2.1) -- (2.2). The following assertions are equivalent. (i) M is a reflexive space. (ii) There exists a map Ψ = (ψ1, ψ2) : Bil(M) → Bil(M) such that P ≤ ψ1(P, Q) and Q ≤ ψ2(P, Q), for any pair (P, Q) ∈ Bil(M), and M = {T ∈ B(H1, H2); ψ2(P, Q)T ψ1(P, Q) = 0, for all (P, Q) ∈ Bil(M)}. (iii) There exists a map ψ1 : Lat(BM)⊥ → Lat(AM) such that P ≤ ψ1(Q), for any pair (P, Q) ∈ Bil(M), and M = {T ∈ B(H1, H2); QT ψ1(Q) = 0, for all Q ∈ Lat(BM)⊥}. (iv) There exists a map ψ2 : Lat(AM) → Lat(BM)⊥ such that Q ≤ ψ2(P ), for any pair (P, Q) ∈ Bil(M), and M = {T ∈ B(H1, H2); ψ2(P )T P = 0, for all P ∈ Lat(AM)}. Proof. Firstly we show that (i) ⇐⇒ (ii). Assume that M is a reflexive space. Let Ψ be the map Ψ1 defined in (3.3), and let F = Ψ(Bil(M)). Clearly, F ⊆ Bil(M) and, therefore, Op(F ) ⊇ OpBil(M) = M. To reverse the inclusion, fix T ∈ Op(F ). Observe that, by Proposition 3.3 (iv), for any pair (P, Q) ∈ Bil(M), P ≤ θφ(P ) = ψ1(P, Q) and, by the definition of A CHARACTERIZATION OF REFLEXIVE SPACES OF OPERATORS 7 the map φ, Q ≤ φ(P ) = ψ2(P, Q). Hence, for all (P, Q) ∈ Bil(M), P = θφ(P )P , Q = Qφ(P ) and, consequently, QT P = Qφ(P )T θφ(P )P = 0. It follows that T ∈ OpBil(M) = M, as required. Conversely, suppose that there exists a map Ψ = (ψ1, ψ2) as stated in (ii). It has to be shown that M = OpBil(M). Since it is clear that M ⊆ OpBil(M), it remains to show that M ⊇ OpBil(M). Let S ∈ OpBil(M) be arbitrary. Hence, for any pair (P ′, Q′) ∈ Bil(M), we have Q′SP ′ = 0. In particular, since for (P, Q) ∈ Bil(M), the image (ψ1(P, Q), ψ2(P, Q)) lies also in BilM, it follows that ψ2(P, Q)T ψ1(P, Q) = 0. Finally, this yields that S lies in the set {T ∈ B(H1, H2); ψ1(P, Q)T ψ2(P, Q) = 0 ∀ (P, Q) ∈ Bil(M)}, which coincides with M, by the assumption. The remaining equivalences are similarly proved. Notice that to prove the impli- cation (i)⇒(iii) (respectively, (i)⇒(iv)), we set ψ1 = θ (respectively, ψ2 = φ). (cid:3) Observe that the maps appearing in the equalities characterizing a reflexive space M in Theorem 3.5 need not be unique (see [2, Remark, p. 223]). In particular, the map Ψ in Theorem 3.5 (ii) can be chosen to be order-preserving. References [1] H. Deguang, On A-submodules for reflexive operator algebras, Proc. Amer. Math. Soc. 104 (4) (1988), 1067 -- 1070. [2] J. A. Erdos and S. C. Power, Weakly closed ideals of nest algebras, J. Operator Theory 7 (1982), 219 -- 235. [3] D. Hadwin, A general view of reflexivity, Trans. Amer. Math. Soc., 344 (1994), 325 -- 360. [4] D. Hadwin, I. Iona¸scu, and H. Yousefi, R-orbit reflexive operators, Oper. Matrices 7 (3) (2013), 619 -- 631. [5] P. Halmos, Reflexive lattices of subspaces, J. London Math. Soc. (2) 4 (1971), 257 -- 263. [6] P. Li and F. Li, Jordan modules and Jordan ideals of reflexive algebras, Integr. Equ. Oper. Theory 74 (2012), 123 -- 136. [7] A. I. Loginov and V. S. Shulman, Hereditary and intermediate reflexivity of W ∗-algebras, Izv. Akad. Nauk SSSR Ser. Mat. 39 (1975) 1260-1273 (in Russian); Trans. Math. USSR-Izv. 9 (1975) 1189 -- 1201. [8] V. S. Shulman and L. Turowska, Operator synthesis. I. Synthetic sets, bilattices and tensor algebras, J. Funct. Anal. 209 (2004), 293 -- 331. Janko Bracic, University of Ljubljana, NTF, Askerceva c. 12, SI-1000 Ljubljana, Slovenia E-mail address: [email protected] Lina Oliveira, Center for Mathematical Analysis, Geometry and Dynamical Systems and, Department of Mathematics, Instituto Superior T´ecnico, Universidade de Lisboa, Av. Rovisco Pais,, 1049-001 Lisboa, Portugal E-mail address: [email protected]
1007.2616	3	1007	2011-02-11T21:50:58	Group actions on topological graphs	[ "math.OA" ]	We define the action of a locally compact group $G$ on a topological graph $E$. This action induces a natural action of $G$ on the $C^$-correspondence ${\mathcal H}(E)$ and on the graph $C^$-algebra $C^(E)$. If the action is free and proper, we prove that $C^(E)\rtimes_r G$ is strongly Morita equivalent to $C^(E/G)$. We define the skew product of a locally compact group $G$ by a topological graph $E$ via a cocycle $c:E^1\to G$. The group acts freely and properly on this new topological graph $E\times_cG$. If $G$ is abelian, there is a dual action on $C^(E)$ such that $C^(E)\rtimes \hat{G}\cong C^(E\times_cG)$. We also define the fundamental group and the universal covering of a topological graph.	math.OA	math	GROUP ACTIONS ON TOPOLOGICAL GRAPHS VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG Abstract. We define the action of a locally compact group G on a topological graph E. This action induces a natural action of G on the C∗-correspondence H(E) and on the graph C∗-algebra C∗(E). If the action is free and proper, we prove that C∗(E) (cid:111)r G is strongly Morita equivalent to C∗(E/G). We define the skew product of a locally compact group G by a topological graph E via a cocycle c : E1 → G. The group acts freely and properly on this new topological graph E ×c G. If G is abelian, there is a dual action on C∗(E) such that C∗(E) (cid:111) G ∼= C∗(E ×c G). We also define the fundamental group and the universal covering of a topological graph. 1. Introduction Topological graphs generalize discrete directed graphs. They have a richer struc- ture, since the vertex and edge spaces may be arbitrary locally compact spaces, and it is natural to study their symmetries. The corresponding C∗-algebras provide a wealth of intriguing examples. Katsura proved in [Kat08] that all Kirchberg alge- bras can be obtained from topological graphs. In [KP99] the authors consider free actions of discrete groups G on directed graphs E and on their associated C∗-algebras C∗(E). They prove that the crossed product C∗(E) (cid:111) G is strongly Morita equivalent to C∗(E/G), where E/G is the quotient graph. They also consider the notion of skew product graph associated to a cocycle c : E1 → G and the universal covering tree of a graph. In this paper, we prove an analogous result for the reduced crossed product for G a locally compact group acting freely and properly on a topological graph E. Unlike [KP99], where groupoids were used extensively, the C∗-algebra C∗(E) is defined here using a C∗- correspondence H(E) and the Cuntz-Pimsner construction (see [Kat04]). The main ingredients of the proof are a structure theorem of Palais about principal G-bundles, the generalized fixed point algebras of Rieffel, and the use of multiplier bimodules introduced by Echterhoff, Raeburn, and others. We need multiplier bimodules because we must construct a homomorphism of C∗(E/G) into the multiplier algebra M (C∗(E)), and as an intermediate step we map the correspondence H(E/G) into the multipliers of the correspondence H(E). We also define the notion of fundamental group of a topological graph and the universal covering, using a space R(E), called the geometric realization, which is a kind of double mapping torus. In the discrete case, the geometric realization is the usual 1-dimensional CW-complex that we associate with a graph, and the fundamental group is free. For a topological graph E, the space R(E) could be a Date: October 22, 2018. 1991 Mathematics Subject Classification. Primary 46L05; Secondary 46L55. Key words and phrases. C∗-algebra, group action, topological graph. 1 2 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG higher dimensional CW-complex, and π1(E) is not necessarily free, as we will see in our examples. Our paper is organized as follows. In section 2 we recall the notion of topological graph and the construction of its C∗-algebra as a Cuntz-Pimsner algebra, giving several examples. In section 3 we define group actions on topological graphs, and prove a structure theorem about topological graphs on which a locally compact group acts freely and properly. The fundamental group of a topological graph and the corresponding universal covering are defined in section 4; we also consider a number of examples. In section 5 we discuss group actions on C∗-correspondences, recall the notion of generalized fixed point algebra, and prove the main theorem stating that C∗(E) (cid:111)r G and C∗(E/G) are strongly Morita equivalent for a free and proper action. The appendix contains basic results concerning multipliers of correspondences which are needed in section 5. The authors would like to thank the referee for a number of helpful comments. 2. Topological graphs and their C∗-algebras Let E = (E0, E1, s, r) be a topological graph. Recall that E0 and E1 are locally compact (Hausdorff) spaces, and that s, r : E1 → E0 are continuous maps such that s is a local homeomorphism. We think of points in E0 as vertices, and of points e ∈ E1 as edges from s(e) to r(e). A path of length n in E = (E0, E1, s, r) is a sequence e1e2 ··· en of edges such that s(ei) = r(ei+1) for all i. We denote by En the set of paths of length n. The space of infinite paths is denoted E∞. The maps s and r extend naturally to En, and (E0, En, s, r) becomes a topological graph. A vertex v ∈ E0 is viewed as a path of length 0. We define two open subsets E0 sce, E0 fin of E0 by sce = {v ∈ E0 : v has a neighborhood V such that r−1(V ) = ∅}, E0 fin = {v ∈ E0 : v has a neighborhood V such that r−1(V ) is compact}, E0 fin \ E0 rg := E0 and the set of regular vertices by E0 sce. It is proved in Proposition 2.8 of [Kat04] that v ∈ E0 rg if and only if v has a neighborhood such that r−1(V ) is compact and r(r−1(V )) = V . In particular, if r is proper and surjective, we have rg = E0. E0 The C∗-algebra C∗(E) of a topological graph is defined as the Cuntz-Pimsner algebra OH of the C∗-correspondence H = H(E) over the C∗-algebra A = C0(E0), which is obtained as a completion of Cc(E1) using the inner product (cid:88) (cid:104)ξ, η(cid:105)(v) = ξ(e)η(e), ξ, η ∈ Cc(E1) and the multiplications s(e)=v (ξ · f )(e) = ξ(e)f (s(e)), (f · ξ)(e) = f (r(e))ξ(e). The Hilbert C0(E0)-module H can be identified with {ξ ∈ C0(E1) (cid:104)ξ, ξ(cid:105) ∈ C0(E0)}. There exists an injective ∗-homomorphism πE : Cb(E1) → L(H) given by (πE(f )ξ)(e) = f (e)ξ(e) such that πE(f ) ∈ K(H) if and only if f ∈ C0(E1). For more details, see [Kat04]. Recall that a Toeplitz representation of a C∗-correspondence H over A in a C∗-algebra C is a pair (τ, π) with τ : H → C a linear map and π : A → C a GROUP ACTIONS ON TOPOLOGICAL GRAPHS 3 ∗-homomorphism, such that τ (ξa) = τ (ξ)π(a), τ (ξ)∗τ (η) = π((cid:104)ξ, η(cid:105)), τ (ϕ(a)ξ) = π(a)τ (ξ), where ϕ : A → L(H) is the left multiplication. The corresponding universal C∗- algebra is called the Toeplitz algebra TH. A representation (τ, π) is covariant if π(a) = τ (1)(ϕ(a)) for all a in the ideal ϕ−1(K(H))∩(ker ϕ)⊥, where τ (1) : K(H) → C is such that τ (1)(θξ,η) = τ (ξ)τ (η)∗. The Cuntz-Pimsner algebra OH is universal with respect to covariant representations, and it is a quotient of TH. Let (kH, kA) denote the canonical Toeplitz representation of the A-correspondence H in OH. A covariant representation (τ, π) of H in a C∗-algebra C gives rise to a ∗-homomorphism τ × π : OH → C such that τ = (τ × π) ◦ kH and π = (τ × π) ◦ kA. Remark 2.1. In the case of a C∗-correspondence H(E) associated with a topologi- cal graph as above, it is proved in Proposition 1.24 of [Kat04] that ker ϕ = C0(E0 sce) fin), therefore the ideal ϕ−1(K(H))∩(ker ϕ)⊥ coincides and that ϕ−1(K(H)) = C0(E0 with C0(E0 rg). Example 2.2. If the vertex space E0 is discrete, then the edge space E1 is also discrete, and E = (E0, E1, s, r) is an usual (discrete) graph. For a discrete graph E, its C∗-algebra C∗(E) was initially defined as the universal C∗-algebra generated by mutually orthogonal projections pv for v ∈ E0 and partial isometries te for e ∈ E1 with orthogonal ranges such that t∗ e ≤ pr(e) and (cid:88) ete = ps(e), tet∗ tet∗ e if 0 < r−1(v) < ∞. (1) pv = If H = H(E) is the associated C∗-correspondence described above, then r(e)=v ϕ−1(K(H)) = C0({v ∈ E0 : r−1(v) < ∞}), ker(ϕ) = C0({v ∈ E0 : r−1(v) = 0}), fin = {v ∈ E0 : r−1(v) is finite} and E0 sce = {v ∈ E0 : r−1(v) = ∅} is the set (cid:88) e∈E1 since E0 of sources. The maps (cid:88) v∈E0 π(f ) = f (v)pv, f ∈ C0(E0), τ (ξ) = ξ(e)te, ξ ∈ Cc(E1) give a Toeplitz representation of H into C∗(E) if and only if t∗ ete = ps(e) and e ≤ pr(e), which is covariant if and only if (1) is satisfied. Hence, the C∗-algebra tet∗ of a discrete graph E, defined using generators and relations, is isomorphic to the Cuntz-Pimsner algebra OH (see [FLR00, Proposition 12]) and [Kat04, Example 1]). Example 2.3. Let E0 = E1 = T, s(z) = z, and r(z) = e2πiθz for θ ∈ [0, 1] irrational. Then C∗(E) ∼= Aθ, the irrational rotation algebra. More generally, let E0 = E1 = X, where X is a locally compact metric space, let s = id and let r = h for h : X → X a homeomorphism. Then C∗(E) ∼= C0(X) (cid:111) Z, since C∗(E) is the universal C∗-algebra generated by a copy of C0(X) and a unitary u satisfying h(f ) = u∗f u for f ∈ C0(X), where h(f ) = f ◦ h. Example 2.4. Let n ∈ N \ {0} and let m ∈ Z \ {0}. Take E0 = E1 = T, s(z) = zn, r(z) = zm. We get a topological graph with both s and r local homeomorphisms. When m /∈ nZ, C∗(E) is simple and purely infinite, see [Kat08, Example A.6]. 4 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG Example 2.5. (The Cayley graph of a finitely generated locally compact group). A locally compact group G is called finitely generated if there is a finite subset S of G such that the closure of the subgroup generated by S is all G. Given a set of generators S = {h1, h2, ..., hn}, define the associated Cayley graph E = E(G, S) with E0 = G, E1 = S × G, s(h, g) = g, and r(h, g) = gh. Note that E(G, S) becomes a topological graph and that both s and r are local homeomorphisms. For G discrete and finitely generated, we recover the usual notion of Cayley graph. The Cayley graph may change if we change the set of generators. For G = (R, +) and S = {1, θ}, where θ is irrational, the corresponding Cayley graph E has E0 = R, E1 = {1, θ} × R, s(1, x) = x, r(1, x) = x + 1, s(θ, x) = x, and r(θ, x) = x+θ. Its C∗-algebra is simple and, if θ < 0, it is purely infinite; moreover, C∗(E) ∼= O2 (cid:111)α R, where αt(S1) = eitS1, αt(S2) = eitθS2 for t ∈ R and for the standard generators S1, S2 of the Cuntz algebra O2 (see Theorem 2, section 3 in [KK97] and [Kat02, Proposition 4.3]). The above example should be regarded as a special case of the following. Example 2.6. (Skew products of topological graphs). Let E = (E0, E1, s, r) be a topological graph, let G be a locally compact group, and let c : E1 → G be a continuous function. We call c a cocycle, and define the skew product graph E ×c G = (E0 × G, E1 × G, s, r) (cf. [KP99, Rae05]), where s(e, g) = (s(e), g), r(e, g) = (r(e), gc(e)). Notice that E ×c G becomes a topological graph using the product topology, since s is a local homeomorphism and r is continuous. The situation of a graph with one vertex and n loops {e1, ..., en} and a set of generators S = {h1, ..., hn} of a group G such that c(ei) = hi, i = 1, ..., n gives the Cayley graph E(G, S) as a skew product. In particular, if E is the graph with one vertex and two edges {e, f} and G = (R, +), the map c : {e, f} → R, c(e) = 1, c(f ) = θ for θ irrational determines the topological graph described in the above example. If E is an arbitrary discrete graph and G = Z with c(e) = 1 for all e ∈ E1, then E×c G is the product graph E×Z, where Z 0 = Z 1 = Z, with s(k) = k, r(k) = k+1. There is an isomorphism C∗(E ×c G) ∼= C∗(E) (cid:111) T (see [KP99, Proposition 2.6]). 3. Group actions on topological graphs Definition 3.1. Let E, F be two topological graphs. A graph morphism ϕ : E → F is a pair of continuous maps ϕ = (ϕ0, ϕ1) where ϕi : Ei → F i, i = 0, 1 such that ϕ0 ◦ r = r ◦ ϕ1 and ϕ0 ◦ s = s ◦ ϕ1, i.e. the diagram s←−−−− E1 r−−−−→ E0 E0 (cid:121) ϕ0 (cid:121) (cid:121) ϕ1 ϕ0 s←−−−− F 1 r−−−−→ F 0 F 0 is commutative. A graph morphism has the unique path lifting property for s if ϕ0, ϕ1 are sur- jective and for every v ∈ E0 and every f ∈ F 1 with s(f ) = ϕ0(v) there is a unique e ∈ E1 such that ϕ1(e) = f and s(e) = v. A graph morphism ϕ is a graph cover- ing if both ϕ0, ϕ1 are covering maps. In that case, ϕ has the unique path lifting property with respect to both s and r. GROUP ACTIONS ON TOPOLOGICAL GRAPHS 5 An isomorphism of topological graphs is a graph morphism ϕ = (ϕ0, ϕ1) such that ϕi is a homeomorphism for i = 0, 1. It follows that ϕ−1 = ((ϕ0)−1, (ϕ1)−1) is also a graph morphism. We denote by Aut(E) the group of automorphisms of a topological graph E. A locally compact group G acts on E if there are continuous maps λi : G× Ei → g(x) for i = 0, 1, such that g (cid:55)→ λg is a group Ei, and we write λi(g, x) = λi homomorphism from G into Aut(E). The action λ is called free if λ0 g(v) = v for some v ∈ E0 implies g = 1G. Note that in this case the action of G is also free on the edges of E. The action is called proper if the maps G× E0 → E0 × E0, (g, v) (cid:55)→ (λ0 g(v), v) and G×E1 → E1×E1, (g, e) (cid:55)→ (λ1 g(e), e) are proper. In fact, by Proposition 2.1.14 in [ADR00], it is sufficient to require properness of the first map. We will frequently write g · e for λ1 g(e), and similarly for the action on vertices. Definition 3.2. Let P, X be locally compact Hausdorff spaces, and let G be a q−→ X is called a principal G-bundle if there is a locally compact group. A map P free and proper action of G on P and q induces an identification of P/G with X. Note that we make no assumption regarding local triviality. Principal bundles are thoroughly discussed in [Pal61] and [Hus94] using various definitions. By [RW98, Remark 4.64] the notion of principal G-bundle in [Hus94] is equivalent to our notion determined by a free and proper action. q−→ X and a continuous function f : Y → X, we may view the pull-back f∗(P ) = {(y, p) ∈ Y × P : f (y) = q(p)} as a principal G-bundle f∗(P ) π1−→ Y where the action of G is given by g · (y, p) = (y, g · p). Given a principal G-bundle P The following result may be found in [Hus94, Theorem 4.4.2] (see also the dis- cussion preceding [Pal61, Proposition 1.3.4]). qi−→ Xi for i = 1, 2 and Lemma 3.3. Suppose we are given principal G-bundles Pi a pair of continuous maps f : X1 → X2, f : P1 → P2 such that f is equivariant and f = f q1. Then θf : P1 → f∗(P2) given by θf (p) = (q1(p), f (p)) is an isomorphism q2 of G-bundles. The following is a structure theorem for topological graphs equipped with a free and proper action of a locally compact group: the quotient is a topological graph, and moreover the original graph can be reconstructed from the quotient graph and certain additional data. Theorem 3.4. Let (F 0, F 1, s, r) be a topological graph and G be a locally compact q−→ F 0 and an isomorphism of pull-backs group. Given a principal G-bundle P θ : s∗(P ) ∼= r∗(P ), we may construct a topological graph (E0, E1, s, r) with a free and proper action of G by setting E0 := P and E1 := s∗(P ) (with structure maps defined in the obvious way). Conversely, every topological graph (E0, E1, s, r) on which G acts freely and properly is isomorphic to one arising in this way. Proof. First suppose we are given P , q, and θ. Define s : s∗(P ) → P by s = π2, i.e., s(e, v) = v, and r : s∗(P ) → P by r = π2◦θ. To see that s is a local homeomorphism, fix a point in e = (e, v) ∈ s∗(P ). Since s is a local homeomorphism, there is an open set U ⊂ F 1 containing e such that the restriction sU is injective. It follows that the restriction sπ 1 (U ) is also injective, and hence that s is a local homeomorphism. −1 6 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG The following commutative diagram illustrates the above construction: r∗(P ) zttttttttt sfffffffffff π2 ffffffffffffff r θ ∼= 333333333333333 π1 π1 E1 = s∗(P ) &NNNNNNNNNN s=π2 E0 = P q F 0 E0 = P q F 0 r F 1 s Since π1 ◦ θ = π1, θ has the following form: for θ(e, v) = (e, r(e, v)) (e, v) ∈ s∗(P ). The converse essentially follows from Lemma 3.3 above; the only non-obvious bit is showing that the quotient is a topological graph. So, suppose G acts freely and properly on a topological graph E = (E0, E1, s, r). Let F i = Ei/G be the quotient spaces, with quotient maps qi : Ei → F i, for i = 0, 1. Since the G-action on E is compatible with the range and source maps, we have the following diagram of principal G-bundles: E0 r E1 s E0 q0 q1 F 0 F 1 r s q0 F 0, where r, s : F 1 → F 0 are continuous, and s is open. We must show that F = (F 0, F 1, s, r) is a topological graph, and that there is an isomorphism of pull-backs θ : s∗(E0) → r∗(E0) such that E is isomorphic to the topological graph with the same vertex space E0, but with edge space s∗(E0), source map π2, and range map π2 ◦ θ. By Lemma 3.3 there are G-bundle isomorphisms θr, θs making the following diagram commute (without the top arrow θ): r∗(E0) o_ _ _ _ _ _ _ θ s∗(E0) ∼= cFFFFFFFF θr∼= ;xxxxxxxx θs ∼= #GGGGGGGGG π2 s π2 {xxxxxxxx r E1 q1 F 1 π1 r π1 s E0 q0 F 0. E0 q0 F 0 Of course we define the G-bundle isomorphism θ so that the top triangle (and hence the entire diagram) commutes. It only remains to show that s : F 1 → F 0 is locally injective. Let e ∈ F 1, and put v = s(e). Fix u ∈ E0 with q0(u) = v, so that (e, u) ∈ s∗(E0). Since E is a topological graph, we can find open neighborhoods U of e in F 1 and V of u in E0 such that π2 : s∗(E0) → E0 is injective on (U × V ) ∩ s∗(E0). Since q0 : E0 → F 0 is open, we can shrink U if necessary so that s(U ) ⊂ q0(V ). z & o o s o o / / o o / / o o / / { $ $ # o z z o o / / c ; o o / / GROUP ACTIONS ON TOPOLOGICAL GRAPHS 7 Let e1, e2 ∈ U , and assume that s(e1) = s(e2) = v(cid:48). Choose u(cid:48) ∈ V such that q0(u(cid:48)) = v(cid:48). Then (e1, u(cid:48)), (e2, u(cid:48)) ∈ (U × V ) ∩ s∗(E0), so (cid:3) we must have e1 = e2. Example 3.5. If we are given a finitely generated locally compact group G with generating set S, then G acts freely and properly on its Cayley graph E = E(G, S) g(h, g(cid:48)) = (h, gg(cid:48)). The quotient graph E/G has S loops and by λ0 one vertex. g(g(cid:48)) = gg(cid:48) and λ1 Example 3.6. If E = (E0, E1, s, r) is a topological graph, G is a locally compact group, and c : E1 → G is a continuous function, then ϕ : E ×c G → E, ϕ(g, x) = Indeed, G acts freely and properly on E ×c G x is a principal G-bundle map. by λ0 g(e, h) = (e, gh). The source and range maps are equivariant, and the quotient graph is isomorphic to E. g(v, h) = (v, gh) and λ1 The following result characterizes skew products: Corollary 3.7. The topological graph constructed in Theorem 3.4 from an isomor- phism θ : s∗(P ) ∼= r∗(P ) is G-equivariantly isomorphic to a skew product F ×c G if and only if the G-bundle P over F 0 is trivial. Proof. The forward direction is trivial, so assume that the G-bundle q : P → F 0 is trivial, and let E be the topological graph constructed from an isomorphism θ : s∗(P ) → r∗(P ). The pull-back of a trivial bundle is trivial, so most of what we have to do is quite straightforward; the only slightly non-obvious bit is that the isomorphism s∗(P ) ∼= F 1 × G of G-bundles can be promoted to a topological-graph isomorphism E ∼= F ×c G for a suitable cocycle c. Let ϕ0 : P → F 0 × G be a G-bundle isomorphism. We must show that there are: • a G-bundle isomorphism ϕ1 : E1 → F 1 × G, and • a cocycle c : F 1 → G such that ϕ = (ϕ0, ϕ1) : E → F ×c G is an isomorphism of topological graphs. The diagram (2) P q ϕ0 ∼= #GGGGGGGGG {wwwwwwwww π1 F 0 × G F 0 where F 1 × G ϕ1 ∼= {wwwwwwwww #GGGGGGGGG π1 r r(cid:48) r ϕ1 ∼= #GGGGGGGGG {wwwwwwwww π1 π1 F 1 × G E1 F 1 s=π2 s×id / s P q F 0 × G ϕ0 ∼= zvvvvvvvvv #GGGGGGGGG π1 F 0, r(cid:48) : (F ×c G)1 = F 1 × G → F 0 × G = (F ×c G)0 is the range map for the skew product, given by r(cid:48)(e, g) = (r(e), gc(e)), illustrates the hypothesis and conclusion. Keep in mind that E0 = P and E1 = s∗(P ). We are given the map ϕ0, and since it is a G-bundle isomorphism the extreme left and right triangles commute. # { # o o / / z { # o o { / # o o / / 8 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG It is an elementary fact about pull-backs that there is a G-bundle isomorphism ϕ1 : s∗(P ) → F 1 × G making the right half of the diagram (2) commute. We can now define r(cid:48) : F 1 × G → F 0 × G so that the left half of (2) commutes. It remains to show that there is a continuous map c : F 1 → G such that r(cid:48) has the form r(cid:48)(e, g) = (r(e), gc(e)). Since π1 ◦ r(cid:48) = r ◦ π1 : F 1 × G → F 0, we have π1(r(cid:48)(e, 1)) = r(e). Define c : F 1 → G by c(e) = π2(r(cid:48)(e, 1)). Then c is continuous, and we have r(cid:48)(e, g) = r(cid:48)(g · (e, 1)) = g · r(cid:48)(e, 1) = g · (r(e), c(e)) = (r(e), gc(e)). (cid:3) qi−→ Xi for i = 1, 2 and a Lemma 3.8. If we are given principal G-bundles Pi pair of continuous maps f : X1 → X2, f : P1 → P2 such that f is equivariant and f = f q1, and compact sets K ⊂ X1 and L ⊂ P2, then (q1)−1(K) ∩ f−1(L) is a q2 compact subset of P1. Proof. Observe that f∗(P2) is a closed subset of X1×P2 and hence (K×L)∩f∗(P2) is compact. Therefore, 1 (K) ∩ f−1(L) = (θf )−1((K × L) ∩ f∗(P2)) q−1 (cid:3) is also compact. Corollary 3.9. Let a locally compact group G act freely and properly on a topo- logical graph E = (E0, E1, s, r), let q1 : E1 → E1/G be the quotient map, and let K ⊂ E1/G be compact. Then: (1) For every compact subset L ⊂ E0, both intersections 1 (K) ∩ s−1(L) q−1 1 (K) ∩ r−1(L) and q−1 are compact. (2) There exists d ≥ 0 such that (cid:12)(cid:12)q−1(K) ∩ s−1(v)(cid:12)(cid:12) ≤ d for all v ∈ E0. Proof. (1) follows immediately from Lemma 3.8. For (2), note first of all that the conclusion is unaffected by replacing E with an isomorphic topological graph, so by Theorem 3.4 we may assume that E is constructed from a topological graph F , a principal G-bundle q : P → F 0, and an isomorphism θ : s∗(P ) → r∗(P ). Thus E1 = s∗(P ), with quotient maps π1 : s∗(P ) → F 1 on edges and q : P → F 0 on vertices, and source map π2 : s∗(P ) → P . Note that for v ∈ P , if we put u = q(v) ∈ F 0 then we have 1 (K) ∩ s−1(v) = {(e, v) ∈ s∗(P ) : e ∈ K} =(cid:0)K ∩ s−1(u)(cid:1) × {v}. π−1 1 (K) ∩ s−1(v)(cid:12)(cid:12) =(cid:12)(cid:12)K ∩ s−1(u)(cid:12)(cid:12). (cid:12)(cid:12)π−1 Thus GROUP ACTIONS ON TOPOLOGICAL GRAPHS 9 Since s : F 1 → F 0 is a local homeomorphism and K is compact, the cardinalities of the intersections K ∩ s−1(u) for u ∈ F 0 are bounded above by some fixed real (cid:3) number. Remark 3.10. If G is discrete and G acts freely and properly on E, then the morphism q : E → E/G is a graph covering, since G acts properly discontinuously (see 4.69 in [RW98]). Remark 3.11. If E and G are discrete and G acts freely on E, then by Theo- rem 2.2.2 in [GT01], there is a cocycle c : (E/G)1 → G such that (E/G) ×c G ∼= E in an equivariant way. This result can be obtained from Corollary 3.7, since any principal G-bundle over E0 is trivial. In general, not every free and proper action on a topological graph is associated to a skew product. Recall that by [Kat06, Proposition 8.9] a topological graph E is minimal iff every orbit space Orb(v, e) (see [Kat06, Definition 4.9]) is dense in E0, where v ∈ E0 and e is a negative orbit of v. If E0 = E0 rg, a negative orbit of v is an infinite path e = e1e2 ··· ∈ E∞ ending at v and Orb(v, e) = {r(e(cid:48)) e(cid:48) ∈ Em, s(e(cid:48)) = s(en) for some n}. Theorem 3.12. Let E be a minimal topological graph and let q : E → F be a graph covering. Suppose that E0 is not discrete and E0 = E0 rg. Then F is also minimal, rg. In particular, both C∗(E) and C∗(F ) are simple. F 0 is not discrete and F 0 = F 0 Proof. Indeed, the unique path lifting property ensures that every orbit space in F (cid:3) is dense. The second part follows from Corollary 8.13 in [Kat06]. Corollary 3.13. Let E be a minimal topological graph and suppose that E0 is not discrete and E0 = E0 rg. If the discrete group G acts freely and properly on E, then C∗(E/G) is simple. Example 3.14. Now let E = E(G, S) be as in Example 2.5 where G = R and S = {1, θ}, with θ is irrational. Then E0 = R, E1 = {1, θ} × R. Translation by Z gives a free and proper action on both E0 and E1 which intertwines the structure maps r, s. Hence, translation induces a free and proper action of Z on E. Since E is a minimal topological graph, E0 is not discrete and E0 = E0 rg, we have by the above corollary that C∗(E/Z) is simple. 4. The fundamental group of a topological graph In this section, we define the fundamental group of a topological graph E = (E0, E1, s, r) and the notion of a universal covering of E, using a single topological space R(E) called the geometric realization. We need to assume that the geometric realization has a universal covering. In this section we need not assume that the source map is a local homeomorphism, even though in all of our examples this is the case. For each e ∈ E1 we formally denote the reversed edge by e, where s(e) = r(e) and r(e) = s(e). The set of reversed edges is denoted E . A walk in E is a sequence such that s(ei) = r(ei+1) for i = 1, ..., n − 1. We w = e1 ··· en where ei ∈ E1 ∪ E define s(w) := s(en) and r(w) := r(e1). A vertex will be considered as a trivial walk. A walk w is reduced if it does not contain the subword ee for any e ∈ E1∪E 1 . 1 1 10 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG Denote by Erw the space of reduced walks, with product topology. Note that E0 is viewed as a subset of Erw. For w = e1 ··· en ∈ Erw, the reverse walk w is en ··· e1. The following definition is modeled on [BH99, Definition III.G.3.1]. Definition 4.1. An E-path c = (w0, c1, w1, ..., ck, wk) over a partition 0 = t0 ≤ ... ≤ tk = 1 of the interval [0, 1] consists of continuous maps ci : [ti−1, ti] → E0 and reduced walks wi ∈ Erw such that r(wi) = ci+1(ti) for i = 0, 1, ..., k − 1 and s(wi) = ci(ti) for i = 1, 2, ..., k. The initial point of c is x = s(w0); its terminal point is y = r(wk). We say that c joins x to y. If w0 and wk are trivial walks (vertices), they can be dropped in the notation for c. In particular, a map c : [0, 1] → E0 can be considered as an E-path. Definition 4.2. A topological graph E is said to be connected if given any two vertices x, y ∈ E0, there is an E-path c joining x and y. Definition 4.3. The geometric realization of a topological graph E is the topologi- cal space R(E) obtained from the disjoint union E1× [0, 1](cid:116) E0 by identifying (e, 0) with s(e) and (e, 1) with r(e) (a kind of double mapping torus). We will identify E0 with a subspace of R(E) in the obvious way. Also, we embed E1 in R(E) by e ∈ E1 (cid:55)→ (e, 1 2 ). Remark 4.4. Notice that if E is connected, then R(E) is path connected. Let ϕ : E → F be a graph morphism. Then ϕ yields a natural map E1 × [0, 1] (cid:116) E0 → F 1 × [0, 1] (cid:116) F 0 which then induces a map R(ϕ) : R(E) → R(F ). Moreover, for i = 0, 1 the following diagram commutes: Ei −−−−→ R(E) (cid:121) ϕi (cid:121) R(ϕ) F i −−−−→ R(F ) where the horizontal maps are the canonical embeddings given in Definition 4.3. Observe that R(ϕ) is continuous; moreover, R(ϕ) is a covering if ϕ is. g(e), t) for e ∈ E1, t ∈ [0, 1], and by g · v = λ0 Remark 4.5. If the group G acts on the topological graph E, then G acts on R(E) g(v) for v ∈ E0. Since by g · (e, t) = (λ1 λi g commute with s and r for i = 0, 1, the action is well defined. Definition 4.6. Let E = (E0, E1, s, r) be a a connected topological graph. We define the fundamental group of E by π1(E) := π1(R(E)). We say that E is simply connected if π1(E) is trivial (i.e. R(E) is simply connected). We say that a graph morphism p : E → E is a universal covering if p is a covering and E is connected and simply connected (or briefly that E is a universal cover). Remark 4.7. The fundamental group π1(E) acts freely on E, and the orbit space is isomorphic to E. Any subgroup H of π1(E) will determine an intermediate covering of E, by taking the graph E/H. Recall that the fundamental group of a finite graph E is the free group with E1−E0 + 1 generators, see [KP99, Lemma 4.10]. Proposition 4.8. Let E be a connected topological graph and suppose that R = R(E) is locally path-connected and semi-locally simply connected. Then R has a GROUP ACTIONS ON TOPOLOGICAL GRAPHS 11 universal covering space R. Moreover, R ∼= R( E) where E is a universal cover of E. Proof. By [Mas91, Theorem 10.2] R has a universal covering space R. Let π : R → R be the canonical map, let E0 = π−1(E0), and let E1 = π−1(E1×{1/2}). In order to define the range and the source maps, we use the unique path lifting property of the map π. Let e ∈ E1, then π(e) = (e, 1/2) for some e ∈ E1. Join (e, 1/2) with the image of (e, 0) in R by the path t (cid:55)→ (e, 1 2 (1 − t)), t ∈ [0, 1]. Lift this path to a path in R, and define s(e) to be the endpoint of the lifted path, which belongs to E0. The range map is obtained similarly by joining (e, 1/2) with the image of (cid:3) (e, 1). Corollary 4.9. Consider a connected topological graph E. If both r, s : E1 → E0 are finite-to-one covering maps and E0 is locally contractible (each point has a local base of contractible neighborhoods), then R(E) is also locally contractible. In particular, E has a universal cover. Proof. It suffices to show that R(E) is locally contractible since the hypotheses of the above proposition will be satisfied. A point (e, t) in the image of E1 × (0, 1) has a local base of contractible neighborhoods of the form V × W , where V is a contractible neighborhood of e in E1 and W is a contractible neighborhood of t in (0, 1). For the image of v ∈ E0 in R(E), take V ⊂ E0 a contractible neighborhood of v which is evenly covered by the maps s and r. Then we claim that the image N of s−1(V ) × [0, ε) ∪ r−1(V ) × (1 − ε, 1] in R(E) is a contractible neighborhood of the image of v in R(E) for any 0 < ε < 1/2. Indeed, s−1(V ) is a disjoint union V1 ∪ ... ∪ Vp , and r−1(V ) is a disjoint union U1 ∪ ... ∪ Uq of homeomorphic copies of V . Let S be the star with p + q branches obtained by gluing p + q copies of [0, ε) at 0. Then N is homeomorphic to S × V , which is contractible. (cid:3) It would be nice to have a weaker condition on E that would ensure the existence of the universal cover. Remark 4.10. In order to obtain other coverings of a graph E, we may consider a subgroup H of the fundamental group π1(E) which will act on R, and take the corresponding topological graph of the quotient space R/H, constructed as in the proof of Proposition 4.8. Remark 4.11. The fundamental group of (E0, E1, s, r) is isomorphic to the fun- damental group of the opposite graph (E0, E1, r, s), obtained by interchanging the maps s and r. The natural embeddings described in 4.3 induce maps π1(E0, v) → π1(E) and π1(E1, e) → π1(E) for fixed v ∈ E0 and e ∈ E1. Example 4.12. Consider the topological graph E with E0 = E1 = T and source and range maps s(z) = z, r(z) = e2πiθz for θ irrational. The geometric realization is homeomorphic to the 2-torus T2, hence the fundamental group π1(E) is isomorphic to Z2. The universal covering graph is E = ( E0, E1, s, r), where E0 = E1 = R× Z, with s(y, k) = (y, k), r(y, k) = (y + θ, k + 1). The action of Z2 on E is given by (j, m) · (y, k) = (j + y + mθ, k + m), and E/Z2 ∼= E. Any other connected covering of E is of the form E/H, where H is a subgroup of Z2. More generally, let X be a compact space which admits a universal covering space X, and let h : X → X be a homeomorphism. The geometric realization of the corresponding topological graph E with E0 = E1 = X, s = id and r = h is 12 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG homeomorphic to the mapping torus of h, obtained from X × [0, 1] by identifying (x, 1) with (h(x), 0). The universal covering graph E has E0 = E1 = X × Z. The source and range maps are given by s(y, k) = (y, k), r(y, k) = (h(y), k + 1), where h : X → X is a lifting of h, y ∈ X and k ∈ Z. The map h induces an automorphism h∗ of π1(X), and the fundamental group of E is isomorphic to the semi-direct product π1(X) (cid:111) Z defined using h∗. The action of π1(X) (cid:111) Z on X × Z is given by (g, m) · (y, k) = (g · hm(y), k + m). Example 4.13. Let again E0 = E1 = T with source and range maps s(z) = zn, r(z) = zm for n, m positive integers. The geometric realization R(E) is obtained from the cylinder E1 × [0, 1], where the two boundary circles are identified using the source and range maps. Alternatively, we may start with a rectangle, such that the left and right edges are labeled by a. The top edge E1 × {0} is divided into n segments called b, and the bottom edge E1 × {1} is divided into m segments also called b. By making the identifications, we get the geometric realization R(E). Note that here the arrows are for identification purposes. Figure 1. The case n = 2, m = 3. The segments a, b become generators in the fundamental group π1(R(E)), which is isomorphic to the Baumslag-Solitar group B(n, m) = (cid:104)a, b abna−1 = bm(cid:105). For n = 1 or m = 1, this group is a semi-direct product and it is amenable. For n (cid:54)= 1, m (cid:54)= 1 and n, m relatively prime, it is not amenable (see Example E.11 in [BO08]). The universal covering space of R(E) is the Cayley complex R of B(n, m), ob- tained from the Cayley graph by filling out the rectangles (see, for example page 122 in [LS77]). It is the cartesian product T × R, where T is the Bass-Serre tree of B(n, m), viewed as an HNN-extension. Recall that, given a group G, a subgroup H ⊂ G, and a monomorphism τ : H → G, then the HNN extension G ∗H τ is generated by G and a letter a such that aha−1 = τ (h) for h ∈ H (see for example [Bau93]). In our case, G = (cid:104)b(cid:105) ∼= Z, H = (cid:104)bn(cid:105) ∼= nZ, and τ (bn) = bm. The free action of the group B(n, m) on R is the extension to the 2-cells of the usual left action on the Cayley graph. A piece of the Cayley complex of B(2, 3) together with the action of the generators is illustrated in Figure 2. The 1-skeleton is the directed Cayley graph of B(2, 3), where the generators a, b In the multiply on the right. The group action is given by left multiplication. GROUP ACTIONS ON TOPOLOGICAL GRAPHS 13 Figure 2. Cayley complex for B(2, 3). corresponding tree T , each vertex has 5 edges. The vertex set T 0 is identified with the left cosets g(cid:104)b(cid:105) ∈ B(2, 3)/(cid:104)b(cid:105), and the edge set T 1 with the left cosets g(cid:104)b2(cid:105) ∈ B(2, 3)/(cid:104)b2(cid:105). The source and range maps are given by s(g(cid:104)b2(cid:105)) = g(cid:104)b(cid:105), r(g(cid:104)b2(cid:105)) = ga−1(cid:104)b(cid:105) for g ∈ B(2, 3). Using Proposition 4.8, we can describe the universal covering graph E of E. We have E0 ∼= T 0 × R, E1 ∼= T 1 × R with s(t, y) = (s(t), ny), r(t, y) = (r(t), my) for t ∈ T 1 and y ∈ R. The group B(n, m) acts freely and properly on E, and the quotient graph E/B(n, m) is isomorphic to E. In particular, note that the topological graph E is not a skew product. In the next section, we show that C∗( E) (cid:111)r B(n, m) is strongly Morita equivalent to C∗(E). 5. Group actions on C∗-correspondences Definition 5.1. Let G be a locally compact group, and let H be a C∗-correspondence over A. We say that G acts on H if there is a map where ϕ : A → L(H) defines the left multiplication. Remark 5.2. There is an action of G on L(H) given by (g · T )(ξ) = g · (T (g−1 · ξ)), G × H → H, (g, ξ) (cid:55)→ g · ξ such that g (cid:55)→ g · ξ is continuous and ξ (cid:55)→ g · ξ is linear, and a continuous action of G on A by ∗-automorphisms such that (cid:104)g · ξ, g · η(cid:105) = g · (cid:104)ξ, η(cid:105), g · (ξa) = (g · ξ)(g · a), g · (ϕ(a)ξ) = ϕ(g · a)(g · ξ), b5ba-1b2ba-1b3b2a-1bb2a-1b2b2a-1b3b4ebb2b3aabab2a2a2ba2b2aabab2ab3ebb2b3bb2b3b4aabab2bababbab2abbb2b3aabab2=b3abababbab2=b4ab2ab2abb2ab2=b5aba-1eba-1bb2a-114 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG such that g (cid:55)→ g · T is continuous with respect to the strict topology on L(H). We have g · (θξ,η) = θg·ξ,g·η, which gives an action on K(H). Indeed, g · (θξ,η)(ζ) = g · (ξ(cid:104)η, g−1 · ζ(cid:105)) = (g · ξ)(cid:104)g · η, ζ(cid:105) = (θg·ξ,g·η)(ζ). An action of G on the C∗-correspondence H defines an action on the Toeplitz algebra TH by g·Tξ = Tg·ξ, and on the Cuntz-Pimsner algebra OH, since all defining relations are equivariant. The full crossed product H (cid:111) G can be defined by H (cid:111) G = H⊗ϕ (A (cid:111) G), where ϕ : A → L(A (cid:111) G) is the embedding of A in the multiplier algebra of the crossed product A (cid:111) G, regarded as a Hilbert module over itself. g ), where ξ ∈ H, f ∈ A(cid:111)G, Note that G acts on H(cid:111)G by g·(ξ⊗f ) = g·ξ⊗(ugf u−1 and g·a = ugau−1 g . Here u : G → U M (A(cid:111)G) is the canonical map into the unitary multipliers. The crossed product H (cid:111) G becomes a C∗-correspondence over A (cid:111) G in a natural way. over A (cid:111)r G; moreover, there is a natural action of G on H (cid:111)r G defined as above. Remark 5.3. This crossed product H (cid:111) G is isomorphic to the one defined using a completion of Cc(G,H), as in [Kas88] and [HN08]. The isomorphism is induced by the function Similarly, the reduced crossed product H(cid:111)rG can be defined as a C∗-correspondence which sends ξ ⊗ f to the map g (cid:55)→ ξf (g). It is proved in [HN08, Theorem 2.10] that if G is amenable, then H ⊗ Cc(G, A) → Cc(G,H) OH(cid:111)G ∼= OH (cid:111) G. Note that we also consider actions and crossed products by non-amenable groups. Proposition 5.4. If G acts on the topological graph E = (E0, E1, s, r), then G acts on the C∗-correspondence H = H(E) and hence on C∗(E). Proof. Define g · ξ(e) = ξ(g−1e) for ξ ∈ Cc(E1), and g · f (v) = f (g−1v) for f ∈ C0(E0). Then this action is compatible with the bimodule structure since s and r (cid:3) are equivariant. Definition 5.5. Recall from [Rie90] that the action α of a locally compact group G on a C∗-algebra A is proper if there is a dense α-invariant ∗-subalgebra A0 of A such that for every a, b ∈ A0 the functions x (cid:55)→ aαx(b) and x (cid:55)→ ∆(x)−1/2aαx(b) are integrable and there exists a (right) inner product (cid:104)·,·(cid:105)r with values in the subalgebra of M (A), which Rieffel denotes by M (A0), comprising those elements that multiply A0 into itself, such that c(cid:104)a, b(cid:105)r = cαx(a∗b)dx for all c ∈ A0. For such an action, G Aα := span{(cid:104)a, b(cid:105)r : a, b ∈ A0} ⊂ M (A) is called the generalized fixed-point algebra. Define a (left) inner product on A0 with values in A (cid:111)α,r G by (cid:96)(cid:104)a, b(cid:105)(x) = ∆(x)−1/2aαx(b∗). (cid:90) GROUP ACTIONS ON TOPOLOGICAL GRAPHS 15 The set I := span{(cid:96)(cid:104)a, b(cid:105) : a, b ∈ A0} is an ideal in A (cid:111)α,r G, and the closure Z of A0 in the norm (cid:107)a(cid:107)2 := (cid:107)(cid:104)a, a(cid:105)r(cid:107) is an I − Aα imprimitivity bimodule. The action is called saturated if I = A (cid:111)α,r G. We review now the averaging process of [KQR08, Section 2], which was built upon work in [OP78, OP80, Qui92, QR95]. Let a locally compact group G act freely and properly on a locally compact Hausdorff space T , and let γ be the associated action of G on C0(T ):(cid:0)γg(f )(cid:1)(t) = f (g−1 · t) for g ∈ G, f ∈ C0(T ), t ∈ T. Also let α be an action of G on a C∗-algebra A, and suppose that we have an equivariant nondegenerate1 homomorphism ϕ : C0(T ) → M (A). For g ∈ G let αg denote the canonical extension of αg to M (A), and write M (A)α = {a ∈ M (A) : αg(a) = a for all g ∈ G} (not to be confused with Rieffel's generalized fixed-point algebras discussed above). It was shown in [KQR08] that A0 := span ϕ(Cc(T ))Aϕ(Cc(T )) is a dense ∗-subalgebra of A, and that there is a linear map Φ : A0 → M (A)α such that ω(Φ(a)) = ω(αg(a)) dg for all a ∈ A0, ω ∈ A∗. (cid:90) G We write Φα when confusion is possible. [KQR08] also shows that Φ(A0) is a ∗-subalgebra of M (A), and so its norm closure, denoted by Fix(A, α, ϕ), is a C∗- subalgebra. [Rie04, Theorem 5.7] implies that the action α is proper and saturated with respect to A0, and so by [Rie90] the reduced crossed product A ×α,r G is Morita equivalent to a generalized fixed-point algebra Aα. It was shown in [KQR08, Propo- sition 3.1] that Fix(A, α, ϕ) coincides with the algebra Aα of [Rie90]. We record a few properties of this averaging process, which can be found in [KQR08, Section 2]: • If a ∈ A0 and b ∈ M (A)α then ab ∈ A0 and Φ(ab) = Φ(a)b. • If f ∈ Cc(T ) and a ∈ A0 then the function g (cid:55)→ ϕ(f )αg(a) belongs to Cc(G, A) and ϕ(f )αg(a) dg = ϕ(f )Φ(a). • If (B, β) is another action, and ψ : C0(T ) → M (B) and σ : A → M (B) are nondegenerate and G-equivariant, and if the canonical extension σ : M (A) → M (B) satisfies σ ◦ φ = ψ, then σ restricts to a nondegenerate homomorphism σ : Fix(A, α, ϕ) → M (Fix(B, β, ψ)). (cid:90) G Moreover, • For fixed f, h ∈ Cc(T ) the map Φβ ◦ σ = σ ◦ Φα. a (cid:55)→ Φ(ϕ(f )aϕ(h)) : A0 → M (A) is norm continuous. 1recall that this means that ϕ(C0(T ))A = A 16 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG For the proof of Theorem 5.6 below, we will need to make extensive use of the multiplier bimodules of [EKQR06]; the relevant facts are collected in Appendix A. Theorem 5.6. If a locally compact group G acts freely and properly on a topological graph E, then C∗(E) (cid:111)r G and C∗(E/G) are strongly Morita equivalent. Proof. Since G acts freely and properly on E0 and there is a nondegenerate equivari- ant homomorphism C0(E0) → M (C∗(E)), it follows from [Rie04] and Lemma 4.1 in [HRW05] that the corresponding action α of G on C∗(E) is proper and saturated with respect to the ∗-subalgebra A0 = Cc(E0)C∗(E)Cc(E0) of C∗(E). Therefore the reduced crossed product C∗(E)(cid:111)rG is strongly Morita equivalent to the general- ized fixed point algebra C∗(E)α. Thus it suffices to show C∗(E)α ∼= C∗(E/G), and for this we will construct an injective homomorphism from C∗(E/G) to M (C∗(E)) whose image is C∗(E)α. As before, denote the quotient topological graph by F = E/G, so that F 0 = E0/G and F 1 = E1/G. It will also be convenient to denote both quotient maps E0 → F 0 and E1 → F 1 by q (and no confusion will occur since the particular q that is intended will be clear from the context), and to use the following abbreviations: • A = C0(E0); • B = C0(F 0); • X = H(E); • Y = H(F ). Of course, our homomorphism C∗(F ) → M (C∗(E)) will be of the form τ × π for a Cuntz-Pimsner covariant Toeplitz representation (τ, π) of Y in M (C∗(E)). We will construct (τ, π) : Y → M (C∗(E)) by, roughly speaking, first mapping the B-correspondence Y into the multiplier bimodule M (X), then composing with the canonical Toeplitz representation (kX , kA) of the A-correspondence X in C∗(E). Actually, this needs a little tweaking; our strategy is more accurately indicated by the diagram (3) Y (µ,ν) /____ (τ,π) M (C∗(E)) 8qqqqqqqqqq (kX ,kA) MA(X). Note the appearance in the above diagram of the "A-multiplier" bimodule MA(X) (see Definition A.8 and Remark A.10); this is necessary because the canonical Toeplitz representation (kX , kA) can be degenerate -- see Example A.1. Our proof is rather long, and to improve readability we break it into the following steps: (1) construct a correspondence homomorphism (µ, ν) as in Diagram 3; (2) define (τ, π) to make Diagram 3 commute, and show that it is a Toeplitz representation; (3) show that the Toeplitz representation (τ, π) is Cuntz-Pimsner covariant; (4) show that the associated homomorphism τ × π : C∗(F ) → M (C∗(E)) is (5) show that the image of τ × π is C∗(E)α. injective; We take each step in order: / 8 GROUP ACTIONS ON TOPOLOGICAL GRAPHS 17 Step 1. Clearly, we should define (µ, ν) by composing with the quotient map q. On B this obviously does the right thing: ν : B → M (A), ν(f ) := f ◦ q. On Y , things are a little more delicate: for ξ ∈ Cc(F 1) and f ∈ Cc(E0) define µ(ξ) · f : E1 → C by (cid:0)µ(ξ) · f(cid:1)(e) = ξ(q(e))f (s(e)). Then µ(ξ) · f ∈ Cc(E1) by Lemma 3.9. (µ, ν) : Y → MA(X), and we will accomplish this by: We must show that we can extend µ to get a correspondence homomorphism • extending to a correspondence homomorphism into M (X), and then • showing that it actually takes values in MA(X). We first need to know that for ξ ∈ Cc(F 1) the map f (cid:55)→ µ(ξ) · f : Cc(E0) → Cc(E1) extends to an adjointable map A → X, i.e., an element of M (X). For technical purposes, we need the following Claim: If ξ, η ∈ Cc(F 1) and v ∈ E0, then (cid:88) (cid:104)ξ, η(cid:105)(q(v)) = ξ(q(e))η(q(e)). s(e)=v To see this, just observe that, for every v(cid:48) ∈ F 0, no matter which element v ∈ q−1(v(cid:48)) we choose, the set of values (cid:8)ξ(q(e))η(q(e)) : s(e) = v(cid:9) coincides with the set of values(cid:8)ξ(e(cid:48))η(e(cid:48)) : s(e(cid:48)) = v(cid:48)(cid:9), because G acts freely and the source map s : E1 → E0 is G-equivariant, and we have proved the claim. We use this to show that the linear map f (cid:55)→ µ(ξ) · f : Cc(E0) → X is bounded, and hence extends uniquely to a bounded linear map For v ∈ E0 we have (cid:104)µ(ξ) · f, µ(ξ) · f(cid:105)(v) = µ(ξ) : A → X. (cid:88) µ(ξ)(q(e))2f (v)2 s(e)=v = (cid:104)ξ, ξ(cid:105)(q(e))f (v)2 ≤ (cid:107)ξ(cid:107)2(cid:107)f(cid:107)2, so (4) verifying the boundedness. (cid:107)µ(ξ) · f(cid:107) ≤ (cid:107)ξ(cid:107)(cid:107)f(cid:107), 18 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG We need to show that the linear map µ(ξ) : A → X is adjointable. For η ∈ Cc(E1) define µ(ξ)∗ · η : E0 → C by (cid:0)µ(ξ)∗ · η(cid:1)(v) = (cid:12)(cid:12)(cid:0)µ(ξ)∗ · η(cid:1)(v)(cid:12)(cid:12)2 ≤ (cid:88) s(e)=v ξ(q(e))η(e), (cid:88) ξ(q(e))2 (cid:88) η(e)2 where the sum is finite by Corollary 3.9. An argument similar to the proof of [Kat04, Lemma 1.5] shows that µ(ξ)∗ · η ∈ Cc(E0). We have s(e)=v s(e)=v = (cid:104)ξ, ξ(cid:105)(q(v))(cid:104)η, η(cid:105)(v) ≤ (cid:107)ξ(cid:107)2(cid:107)η(cid:107)2, so Thus the map (cid:107)µ(ξ)∗ · η(cid:107) ≤ (cid:107)ξ(cid:107)(cid:107)η(cid:107). η (cid:55)→ µ(ξ)∗ · η : Cc(E1) → Cc(E0) extends uniquely to a bounded linear map µ(ξ)∗ : X → A. For f ∈ Cc(E0), η ∈ Cc(E1), and v ∈ E0 we have (cid:104)µ(ξ) · f, η(cid:105)(v) = (cid:88) = f (v)(cid:0)µ(ξ)∗ · η(cid:1)(v), s(e)=v ξ(q(e))f (v)η(e). so (cid:104)µ(ξ) · f, η(cid:105) = (cid:104)f, µ(ξ)∗ · η(cid:105), and it follows by continuity that µ(ξ)∗ : X → A is an adjoint of µ(ξ). Therefore µ(ξ) ∈ M (X). We now have a linear map µ : Cc(F 1) → M (X), and then the estimate (4) shows µ is bounded, hence extends uniquely to a bounded linear map µ : Y → M (X). We can now show that the pair (µ, ν) is a Toeplitz representation. First we show that for ξ, η ∈ Y we have We next show that for f ∈ B and ξ ∈ Y we have µ(f · ξ) = ν(f ) · µ(ξ), (cid:104)µ(ξ), µ(η)(cid:105) = ν((cid:104)ξ, η(cid:105)), and by continuity the following computation is sufficient: f ∈ Cc(E0) and v ∈ E0 we have (cid:0)(cid:104)µ(ξ), µ(η)(cid:105)f(cid:1)(v) =(cid:0)µ(ξ)∗ · µ(η) · f(cid:1)(v) for ξ, η ∈ Cc(F 1) and ξ(q(e))(cid:0)µ(η) · f(cid:1)(e) = = s(e)=v (cid:88) (cid:88) =(cid:0)ν((cid:104)ξ, η(cid:105))f(cid:1)(v). = (cid:104)ξ, η(cid:105)(q(v))f (v) s(e)=v ξ(q(e))η(q(e))f (s(e)) GROUP ACTIONS ON TOPOLOGICAL GRAPHS 19 and again by continuity the following computation suffices: for f ∈ Cc(F 0), ξ ∈ Cc(F 1), h ∈ Cc(E0), and e ∈ E1 we have = f (r(q(e)))ξ(q(e))h(s(e)) (cid:0)µ(f · ξ) · h(cid:1)(e) = (f · ξ)(q(e))h(s(e)) = f (q(r(e)))(cid:0)µ(ξ) · h(cid:1)(e) = ν(f )(r(e))(cid:0)µ(ξ) · h(cid:1)(e) ν(f ) ·(cid:0)µ(ξ) · h(cid:1)(cid:17) (cid:16) (cid:16)(cid:0)ν(f ) · µ(ξ)(cid:1) · h (cid:17) (e). (e) = = We thus have a homomorphism (µ, ν) from the B-correspondence Y to the M (A)- correspondence M (X), and it remains to show that for ξ ∈ Y we actually have µ(ξ) ∈ MA(X), i.e., for f ∈ A we have f · µ(ξ) ∈ X (because µ(ξ)·f ∈ X automatically since µ(ξ) is a module multiplier). By continuity it suffices to show that if ξ ∈ Cc(F 1) and f ∈ Cc(E0) then For h ∈ Cc(E0) and e ∈ E1 we have (cid:17) (cid:16)(cid:0)f · µ(ξ)(cid:1) · h f · µ(ξ) ∈ Cc(E1). (cid:16) f ·(cid:0)µ(ξ) · h(cid:1)(cid:17) = f (r(e))(cid:0)µ(ξ) · h(cid:1)(e) (e) (e) = so the multiplier f · µ(ξ) of X coincides with the element of Cc(E1) given by = f (r(e))ξ(q(e))h(s(e)), e (cid:55)→ f (r(e))ξ(q(e)) (which has compact support by Corollary 3.9). Step 2. By Corollary A.14 the Toeplitz representation (kX , kA) : X → C∗(E) extends to a Toeplitz representation (kX , kA) : MA(X) → MA(C∗(E)), because kA is nondegenerate. Now define (τ, π) : Y → MA(C∗(E)) to make diagram (3) commute, i.e., τ := kX ◦ µ and π := kA ◦ ν. Then (τ, π) is a Toeplitz representation, being the composition of the correspon- dence homomorphism (µ, ν) and the Toeplitz representation (kX , kA). Step 3. Cuntz-Pimsner covariance of (τ, π) will follow from an analogous property of the correspondence homomorphism (µ, ν). Our strategy is illustrated by the 20 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG diagram (5) Brg ϕB,rg MA(C∗(E)) ϕA,rg K(Y ) MA(K(X)), πrg &LLLLLLLLLLL 8rrrrrrrrrr τ (1) ν(cid:48) rg µ(1) MA(Arg) kA,rg wooooooooooo gOOOOOOOOOOO k(1) X which requires explanation. We define Brg = C0(F 0 rg), and we let ϕB,rg denote the restriction to the ideal Brg of the natural homomorphism ϕB : B → L(Y ) implementing the left module multiplication, and similarly for Arg and ϕA,rg. Also similarly for the restrictions πrg of π : B → MA(C∗(E)) and kA,rg of kA : A → C∗(E). The bars in the right-hand triangle of (5) denote the unique A-strictly continuous extensions guaranteed by Lemma A.5 and Corollary A.7. For the bottom arrow, note that the correspondence homomorphism µ : Y → MA(X) naturally induces a homomorphism µ(1) : K(Y ) → K(MA(X)), and by Remark A.10 we have K(MA(X)) ⊂ MA(K(X)). morphism ν : B → M (A), and we denote the restriction to Brg by The part of (5) requiring the most explanation is ν(cid:48) rg. We have defined a homo- νrg : Brg → M (A). For the purposes of the rest of (5), what we need is to have a homomorphism into MA(Arg). will show that νrg actually maps into the C∗-subalgebra The problem is that MA(Arg) doesn't embed naturally in M (A). However, we MA,Arg (A) := {f ∈ M (A) : f A ⊂ Arg}, of M (A). Since every f ∈ MA,Arg(A) vanishes on E0 \ E0 f (cid:55)→ fE0 so that the diagram rg, the restriction map gives an embedding of this subalgebra into MA(Arg), and we define ν(cid:48) rg rg Brg $IIIIIIIII ν(cid:48) rg νrg MA,Arg (A) xqqqqqqqqqqq MA(Arg). commutes. The desired Cuntz-Pimsner covariance of (τ, π) follows from commutativity of the left triangle of (5) (see Remark 2.1), which we will deduce from that of the outer square and the other three triangles. / / & w / / 8 g / / $ K k x GROUP ACTIONS ON TOPOLOGICAL GRAPHS 21 The top triangle. The top triangle of (5) coincides with the lower left triangle of the diagram (6) Brg MA,Arg(A) rg ν(cid:48) #GGGGGGGGGGGGGGGGGGG 5555555555555555555555555555555 πrg νrg yssssssssssssssssssss kA kA,rg MA(Arg) MA(C∗(E)), in which kA denotes the restriction of kA : M (A) → MA(C∗(E)) to the subalgebra MA,Arg(A). Once we have verified that νrg maps Brg into MA,Arg (A), it will suffice to observe that all the other triangles in (6) commute: the top triangle of (6) commutes by definition of ν(cid:48) rg, the outer triangle commutes by definition of πrg, and the lower-right triangle commutes by a routine check of the definitions. We must show that if f ∈ C0(F 0 rg) and g ∈ C0(E0) then ν(f )g = (f ◦ q)g ∈ C0(E0 rg). We have ν(f )g ∈ C0(E0) since ν(f ) is bounded and g ∈ C0(E0). We must show rg, equivalently if ν(f )g is nonzero at an element v ∈ E0, that ν(f )g = 0 on E0 \ E0 then we must have v ∈ E0 rg. So, we must show that v is in the image of the range map r : E1 → E0, and that there is a neighborhood U of v such that r−1(U ) is compact, i.e., v ∈ E0 fin. By Lemma 3.3, we may assume that E1 = r∗(E0), i.e., that the diagram coincides with a pull-back E0 q F 0 π2 r r E1 q F 1 E1 = r∗(E0) π1 F 1 r E0 q F 0 Since ν(f )(v) (cid:54)= 0 and f ∈ C0(F 0 rg. Thus we can choose e ∈ F 1 such that r(e) = q(v). Then (e, v) ∈ E1 = r∗(E0) and v = π2(e, v), showing that v is in the image of the range map on E1. rg), we have q(v) ∈ F 0 fin, start with a compact neighborhood U of v in E0, and put V = q(U ), which is a compact neighborhood of q(v) in F 0. Shrink U if necessary To see that v ∈ E0 / / # K k y o o o o o o o o 22 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG so that W := r−1(V ) is compact in F 1. We will show that π−1 Suppose π2(e, u) ∈ U , i.e., u ∈ U and r(e) = q(u). Then e ∈ W . Thus 2 (U ) is compact. (e, u) ∈ (W × U ) ∩ r∗(E0), which is compact. We have verified the inclusion νrg(Brg) ⊂ MA,Arg (A). The right triangle. Cuntz-Pimsner covariance of the canonical homomorphism (kX , kA) is expressed by the commutative diagram C∗(E) Arg ϕA,rg kA,rg vnnnnnnnnnnnnnn hPPPPPPPPPPPP k(1) X K(X), and it follows from A-strict continuity that the right triangle of (5) commutes. The bottom triangle. For ξ, η ∈ Y we have (cid:0)µ(1)(θξ,η)(cid:1) = k(1) X (cid:0)µ(ξ)µ(η)∗(cid:1) k(1) X = kX (µ(ξ))kX (µ(η))∗ = τ (ξ)τ (η)∗ = τ (1)(θξ,η). (by Corollary A.14) The outer square. We need to show that if f ∈ C0(F 0 rg(f ) = µ(1) ◦ ϕB,rg(f ). ϕA,rg ◦ ν(cid:48) rg), then As in the commutative diagram (6) we can show φA,rg ◦ ν(cid:48) rg(f ) = φA ◦ νrg(f ) = φA(ν(f )). Also, ϕB,rg(f ) = ϕB(f ). Thus we must show ϕA(ν(f )) = µ(1)(ϕB(f )). Since f ◦ r ∈ C0(F 1), we can approximate f ◦ r uniformly by a function h ∈ Cc(F 1), and by [Kat04, Lemmas 1.15 and 1.16] there are ξi, ηi ∈ Cc(F 1) for 1 = 1 . . . n such that • h = n(cid:88) 1 ξiηi; n(cid:88) • πF (h) = • for each i we have ξi(e)ηi(e(cid:48)) = 0 for all e (cid:54)= e(cid:48) in F 1 with s(e) = s(e(cid:48)). θξi,ηi; 1 v h GROUP ACTIONS ON TOPOLOGICAL GRAPHS 23 For ζ ∈ Cc(E1) and e ∈ E1 we have = ξi(q(e))(cid:0)µ(ηi)∗ζ(cid:1)(s(e)) (cid:0)µ(1)(θξi,ηi)ζ(cid:1)(e) =(cid:0)µ(ξi)µ(ηi)∗ζ(cid:1)(e) (cid:88) (cid:88) = ξi(q(e)) s(e(cid:48))=s(e) = ξi(q(e)) s(q(e(cid:48)))=s(q(e)) = ξi(q(e))ηi(q(e))ζ(e). ηi(q(e(cid:48)))ζ(e(cid:48)) ηi(q(e(cid:48)))ζ(e(cid:48)) Thus (cid:0)µ(1)(πF (h)ζ(cid:1)(e) = n(cid:88) 1 ξi(q(e))ηi(q(e))ζ(e) By our choice of h we have(cid:0)ν(f ) · ζ(cid:1)(e) = ν(f )(r(e))ζ(e) = h(q(e))ζ(e). = f ◦ q(r(e))ζ(e) = f ◦ r(q(e))ζ(e) ≈ h(q(e))ζ(e). Let K := supp ζ. The above approximation is uniform over ζ in the subspace CK(E1) := {ξ ∈ Cc(E1) : supp ξ ⊂ K}, so by the elementary Lemma 5.7 below we have ν(f ) · ζ ≈ µ(1)(πF (h))ζ in X. Now, by definition we have ν(f ) · ζ = ϕA(ν(f ))ζ, so we get Since f ◦ r ≈ h, we have ϕB(f ) ≈ πF (h), so ϕA(ν(f ))ζ ≈ µ(1)(πF (h))ζ. µ(1)(ϕB(f ))ζ ≈ µ(1)(πF (h))ζ ≈ ϕA(ν(f ))ζ. Thus (cid:107)µ(1)(ϕB(f ))ζ − ϕA(ν(f ))ζ(cid:107) is arbitrarily small, so we must have µ(1)(ϕB(f ))ζ = ϕA(ν(f ))ζ. Since ζ was an arbitrary element of Cc(E1), which is dense in X, we have shown that the outer square commutes. In the above argument we used the following lemma: Lemma 5.7. For any compact subset K ⊂ E1, on the subspace CK(E1) of X the Hilbert-module norm from X is equivalent to the uniform norm. 24 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG Proof. The uniform norm on CK(E1) is of course smaller than the Hilbert-module norm from X. On the other hand, by Lemma 3.9 (ii) we can let d be an upper bound for the cardinalities of the intersections K ∩ s−1(v) for v ∈ E0. Then for any v ∈ E0 we have (cid:88) ξ(e)2 ≤ d(cid:107)ξ(cid:107)2 u, where (cid:107) · (cid:107)u denotes the uniform norm. s(e)=v (cid:3) It now follows that the left triangle of (5) commutes, and hence that the Toeplitz representation (τ, π) : Y → M (C∗(E)) is Cuntz-Pimsner covariant. Step 4. To see that τ × π : C∗(F ) → M (C∗(E)) is injective, we apply the Gauge- Invariant Uniqueness Theorem [Kat04, Theorem 4.5]. For this we need to show that: σz(τ (ξ)) = zτ (ξ) for ξ ∈ Cc(F 1) and σz(π(f )) = π(f ) for f ∈ Cc(F 0). (1) π : B → M (C∗(E)) is injective, and (2) for all z ∈ T there is an automorphism σz of (τ × π)(C∗(F )) such that For (1), just note that both ν : B → M (A) and kA : M (A) → M (C∗(E)) are injective (the latter because kA : A → C∗(E) is), and hence so is kA ◦ ν. For (2), we extend the gauge automorphism γz of C∗(E) to an automorphism γz of the multiplier algebra M (C∗(E)), and note that by strict and A-strict continuity we have • γz ◦ kX = zkX on MA(X), and • γz ◦ kA = kA on M (A). Step 5. Finally, we need to show that the image (τ × π)(C∗(F )) coincides with the generalized fixed-point algebra of C∗(E) under the action α of G. We have an action β of G on C0(E0) corresponding to the free and proper action on E0, and a G-equivariant nondegenerate homomorphism kA : A = C0(E0) → C∗(E) ⊂ M (C∗(E)), so we can form the generalized fixed-point algebra and we will show that it coincides with (τ ×π)(C∗(F )). We begin with the inclusion C∗(E)α = Fix(C∗(E), α, kA), Fix(C∗(E), α, kA) ⊂ (τ × π)(C∗(F )), which will occupy us for some time. For n ∈ N = {0, 1, 2, . . .} let En denote the space of paths in E of length n (with the relative product topology). Then by [Kat04], En is naturally a topological graph with vertex space E0, and the associated A-correspondence H(En) is isomorphic to the n-fold balanced tensor product X ⊗A ··· ⊗A X, which we denote simply by X n. We let kn corresponding to the n-fold tensor power of kX . Then C∗(E) = span{kn X (Cc(En))km X : X n → C∗(E) be the representation X (Cc(Em))∗ : n, m ∈ N}. Our hypotheses on E, namely that G acts freely and properly, carry over to En. We use the same notational conventions as for the quotient graph F = E/G, and we write τ n = kn X ◦ µ. GROUP ACTIONS ON TOPOLOGICAL GRAPHS 25 Now let f, h ∈ Cc(E0) and a ∈ C∗(E). We must show that Φ(cid:0)kA(f )akA(h)(cid:1) ∈ (τ × π)(C∗(F )). We may assume that a = kn X (ξ)km Y (η)∗, where ξ ∈ Cc(En) and η ∈ Cc(Em). By the elementary Lemma 5.8 below, we can find finitely many functions ξi ∈ Cc(F n), ηj ∈ Cc(F m), fi, gj ∈ Cc(E0) such that f · ξ ≈(cid:88) µ(ξi) · fi and η ≈(cid:88) i j gj · µ(ηj). Then for each i, j, again by Lemma 5.8 there exist finitely many fijl ∈ Cc(E0), ηijl ∈ Cc(F m) such that µ(ηj) · fi ≈(cid:88) fijl · µ(ηijl). We have l kA(f )a = kA(f )kn X (η)∗ X (µ(ηj))∗kA(gj) X = = kn X = kn X (ξ)km X (η)∗ kn X (µ(ξi))kA(fi)km (cid:0)µ(ξi) · fi (cid:1)km (cid:0)gj · µ(ηj)(cid:1)∗ X (µ(ηj))kA(fi)(cid:1)∗ X (µ(ξi))(cid:0)km (cid:1)(cid:17)∗ (cid:16) (cid:0)µ(ηj) · fi (cid:0)fijl · µ(ηijl)(cid:1)∗ ≈(cid:88) X (f · ξ)km (cid:88) (cid:88) (cid:88) ≈(cid:88) (cid:88) X (µ(ξi))τ m(ηijl)∗kA(fijlgj). kn X (µ(ξi))km kn kn X (µ(ξi)) km X ij ij ij ij kn = = ijl X kA(gj) kA(gj) kA(gj) Now, we can choose h(cid:48) ∈ Cc(E0) such that ijl Then we have h(cid:48)f = f and h(cid:48) · µ(ξi) = µ(ξi) Φ(cid:0)kA(f )akA(h)(cid:1) = Φ(cid:0)kA(h(cid:48))kA(f )akA(h)(cid:1) for all i. X (µ(ξi))τ m(ηijl)∗kA(fijlgj)(cid:1) Φ(cid:0)kA(h(cid:48))kn Φ(cid:0)kn X (µ(ξi))τ m(ηijl)∗kA(fijlgj)(cid:1) τ n(ξi)τ m(ηijl)∗Φ(cid:0)kA(fijlgj)(cid:1). ≈(cid:88) (cid:88) (cid:88) = ijl ijl = ijl 26 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG Since τ n(ξi)τ m(ηijl)∗ ∈ (τ × π)(C∗(F )) and Φ ◦ kA(Cc(E0)) = kA ◦ Φβ(Cc(E0)) ⊂ kA ◦ ν(Cc(F 0)) = π(Cc(F 0)) ⊂ (τ × π)(C∗(F )), it follows that Φ(kA(f )akA(h)) ∈ (τ × π)(C∗(F )). The above argument used the following elementary lemma: Lemma 5.8. In the norm of X, Cc(E1) ⊂ span(cid:8)µ(Cc(F 1)) · Cc(E0)(cid:9) ∩ span(cid:8)Cc(E0) · µ(Cc(F 1))(cid:9). Proof. By Theorem 3.4 we may assume that E1 = s∗(E0). Let θ : s∗(E0) → r∗(E0) be the isomorphism of Theorem 3.4. Note that µ(η) · f = (η ⊗ f )s∗(E0) and f · µ(η) = (η ⊗ f ) ◦ θ, where (η ⊗ f )(e, v) = η(e)f (v) for (e, v) ∈ F 1 × E0. Let ξ ∈ Cc(E1). Then there is ξ(cid:48) ∈ Cc(F 1 × E0) such that ξ = ξ(cid:48)s∗(E0), because s∗(E0) is a closed subset of F 1 × E0. Then there are finitely many ηi ∈ Cc(F 1), fi ∈ Cc(E0) such that ξ(cid:48) ≈(cid:88) (ηi ⊗ fi). Thus ξ ≈(cid:88) i (ηi ⊗ fi)s∗(E0) = (cid:88) i µ(ηi) · fi. This shows the inclusion in the first set. For the other one, since ξ ◦ θ−1 ∈ Cc(r∗(E0)), there is ζ ∈ Cc(F 1 × E0) such that i and then as above we can approximate ξ ◦ θ−1 = ζr∗(E0), i ζ ≈(cid:88) ξ ◦ θ−1 ≈(cid:88) ξ ≈(cid:88) i (ηi ⊗ fi), (ηi ⊗ fi)r∗(E0), (cid:88) so that and hence (ηi ⊗ fi) ◦ θ = fi · µ(ηi). (cid:3) i i We have shown that Fix(C∗(E), α, kA) ⊂ (τ × π)(C∗(F )), and we turn to the opposite inclusion. We need the following easy lemma. Lemma 5.9. For all f ∈ Cc(F 0) there exists h ∈ Cc(E0) such that ν(f ) = Φβ(h). GROUP ACTIONS ON TOPOLOGICAL GRAPHS 27 Proof. Put M = supp f , a compact subset of F 0. Since the quotient map q : E0 → F 0 is a continuous open surjection, it is a standard fact that there exists a compact set K ⊂ E0 such that q(K) = M . Choose h1 ∈ Cc(E0) such that h1(v) (cid:54)= 0 for all v ∈ K. Then Φβ(h1) (cid:54)= 0 on q−1(M ), and there is a unique h2 ∈ Cc(F 0) such that Φβ(h1) = ν(h2). Then h2 (cid:54)= 0 on M , so there exists h3 ∈ Cc(F 0) such that h3 = f h2 on M. Then f = h2h3, so ν(f ) = ν(h2)ν(h3) = Φβ(h1)ν(h3) = Φβ(cid:0)h1ν(h3)(cid:1), so we can take h = h1ν(h3). (cid:3) To finish, it suffices to show that τ (ξ) and π(f ) are in Fix(C∗(E), α, kA) for all ξ ∈ Cc(F 1) and f ∈ Cc(F 0). For ξ, choose h ∈ Cc(F 0) such that ξ = ξ · h. By Lemma 5.9 we can choose h1 ∈ Cc(E0) such that ν(h) = Φβ(h1). We have τ (ξ) = τ (ξ · h) = τ (ξ)π(h) = τ (ξ)kA(ν(h)) = τ (ξ)kA(Φβ(h1)) = τ (ξ)Φ(kA(h1)) = Φ(cid:0)τ (ξ)kA(h1)(cid:1) = Φ(cid:0)kX (µ(ξ))kA(h1)(cid:1) = Φ(cid:0)kX (µ(ξ) · h1)(cid:1). Now, by Corollary 3.9 we have µ(ξ) · h1 ∈ Cc(E1), so we can choose h2 ∈ Cc(E0) such that and then we can factor h2 = h3h4 with h3, h4 ∈ Cc(E0). Then µ(ξ) · h1 = h2 · µ(ξ) · h1, (cid:16) (cid:16) (cid:0)h3h4 · µ(ξ) · h1 (cid:1)(cid:17) (cid:17) (cid:0)h4 · µ(ξ)(cid:1)kA(h1) . τ (ξ) = Φ kX = Φ kA(h3)kX Since h4 · µ(ξ) ∈ Cc(E1) (again by Corollary 3.9), we have (cid:0)h4 · µ(ξ)(cid:1) ∈ C∗(E), kX and therefore τ (ξ) ∈ Fix(C∗(E), α, kA). Our strategy for f ∈ Cc(F 0) is similar, but things are somewhat easier this factor f = f h1 with h1 ∈ Cc(F 0), then choose h2 ∈ Cc(E0) such that time: 28 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG ν(h1) = Φβ(h2), then factor h2 = h3h4h5 with hi ∈ Cc(E0). We have π(f ) = π(f h1) = π(f )π(h1) = π(f )kA(ν(h1)) = π(f )kA(Φβ(h2)) = π(f )Φ(kA(h3h4h5)) = Φ(cid:0)π(f )kA(h3h4h5)(cid:1) = Φ(cid:0)kA(ν(f ))kA(h3h4h5)(cid:1) = Φ(cid:0)kA(h3)kA(f h3)kA(h5)(cid:1), which is in Fix(C∗(E), α, kA) because kA(f h3) ∈ C∗(E). (cid:3) Let E be a topological graph and let c : E1 → G be a continuous map. Recall that G acts freely and properly on G ×c E. Since (G ×c E)/G ∼= E, we have by Theorem 5.6: Corollary 5.10. The C∗-algebras C∗(E×c G)(cid:111)r G and C∗(E) are strongly Morita equivalent. In particular, for a finitely generated locally compact group G with gen- erators S = {h1, h2, ..., hn} and Cayley graph E(G, S), we get that C∗(E(G, S))(cid:111)rG is strongly Morita equivalent to the Cuntz algebra On. Remark 5.11. If G is abelian, there is an action αc of G on C∗(E) such that (αc for ξ ∈ Cc(E1) and χ ∈ G. Then χξ)(e) = (cid:104)χ, c(e)(cid:105)ξ(e) C∗(E) (cid:111)αc G ∼= C∗(E ×c G). For G non-abelian, we need to use the notion of coaction. This will be investigated in [KQR10]. Appendix A. multipliers of correspondences In Section 5 we make extensive use of the multiplier bimodules of [EKQR06, Chapter 1]. The bimodules in [EKQR06] were called right-Hilbert A− B bimodules, which are also commonly referred to as A − B correspondences, where A and B are C∗-algebras. In most of this paper we are interested in correspondences over a single C∗-algebra, but occasionally we need the added generality of allowing A and B to be possibly different. Warning: there is a nondegeneracy issue; in the literature, A− B correspondences are often (usually?) allowed to be degenerate, in the sense that the closed span of A · X might be smaller than X, but in [EKQR06] the right-Hilbert bimodules are required to be nondegenerate. This will cause no problem here, though, because the correspondences associated to topological graphs are all nondegenerate, so that the results of [EKQR06] may be applied freely. To emphasize, throughout this appendix we will assume that all our correspon- dences are nondegenerate. A and B are the coefficient algebras of the correspondence X, and we will some- times denote the correspondence by AXB to indicate what the coefficient algebras are. Note that a Hilbert B-module is also a C − B correspondence. GROUP ACTIONS ON TOPOLOGICAL GRAPHS 29 We recall briefly the theory of multiplier bimodules; see [EKQR06, Chapter 1] for more details. A multiplier of AXB is by definition an adjointable map from B to X, and M (X) denotes the set of all multipliers of X. Thus by definition we have M (X) = L(B, X). Note that the definition of M (X) does not involve the left A-module structure at all; it is, in essence, a Hilbert B-module concept. However, M (X) is an L(X) − M (B) correspondence, hence an M (A) − M (B) correspondence because of our assumption that X is nondegenerate as a left A- module. The restriction of the right M (B)-module action to B is just the evaluation of adjointable maps; i.e., if m ∈ M (X) and b ∈ B then we write m · b to mean the value of the adjointable map m : B → X at the element b of its domain. There is an embedding of X in M (X) such that the adjointable map associated to ξ ∈ X is M (X) has both an L(X)-valued inner product b (cid:55)→ ξ · b for b ∈ B. L(X)(cid:104)m, n(cid:105) = mn∗ for m, n ∈ M (X) and an M (B)-valued inner product (cid:104)m, n(cid:105)M (B) = (cid:104)m, n(cid:105) = m∗n. If either m or n lies in X, then L(X)(cid:104)m, n(cid:105) ∈ K(X) and (cid:104)m, n(cid:105) ∈ B, because, more generally, composing an adjointable operator on either side with a compact operator gives a compact operator. The restriction of the left inner product L(X)(cid:104)·,·(cid:105) to X gives the rank-one operators: L(X)(cid:104)ξ, η(cid:105) = K(X)(cid:104)ξ, η(cid:105) = θξ,η, and the restriction of the right inner product to X is the given one. The left and right module actions of multipliers of the coefficient algebras A and B, respectively, on an element m ∈ M (X) are as follows: if n ∈ M (A) then n · m is the element of M (X) defined by (n · m) · b = n · (m · b) b ∈ B, for and if n ∈ M (B) then m · n is defined by Of course, if n ∈ B then m · n ∈ X. (m · n) · b = m · (nb) b ∈ B. for The strict topology on M (X) is generated by the seminorms m (cid:55)→ (cid:107)T · m(cid:107) and m (cid:55)→ (cid:107)m · b(cid:107) for T ∈ K(X), b ∈ B. The operations on the M (A)− M (B) correspondence M (X) are separately strictly continuous, and M (X) is the strict completion of X. Given another correspondence CYD and homomorphisms π : A → M (C) and θ : B → M (D), a linear map ψ : X → M (Y ) is a π − θ compatible correspondence homomorphism if the two conditions (cid:104)ψ(ξ), ψ(η)(cid:105)M (D) = θ((cid:104)ξ, η(cid:105)B); ψ(a · ξ) = π(a) · ψ(ξ) hold. Note that ψ will automatically satisfy ψ(ξ · b) = ψ(ξ)θ(b). 30 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG π and θ are the coefficient homomorphisms of ψ, and sometimes we write πψθ for the correspondence homomorphism to indicate what the coefficient homomorphisms are. A correspondence homomorphism πψθ is nondegenerate if π and θ are nonde- generate and span{ψ(X) · D} = Y, in which case πψθ extends uniquely to a strictly continuous homomorphism πψθ : M (A)M (X)M (B) → M (C)M (Y )M (D). (In fact, as for C∗-algebras, the extension is unique as just a correspondence ho- momorphism, and is automatically strictly continuous, although we will not need this fact.) If A = B, C = D, and π = θ we will write (ψ, π) = πψθ, and if in addition Y = C is a correspondence over itself in the canonical way, and if both ψ and π map into C (and not just M (C)), then (ψ, π) reduces to the familiar concept of a Toeplitz representation of the A-correspondence X into C. If πψθ : AXB → M (CYD) is a nondegenerate homomorphism, then there is a unique nondegenerate homo- morphism ψ(1) : K(X) → M (K(Y )) = L(Y ) such that ψ(1)(θξ,η) = ψ(ξ)ψ(η)∗ ξ, η ∈ X, for and then, as usual, by nondegeneracy ψ(1) extends uniquely to a homomorphism ψ(1) : L(X) → L(Y ). Note that even if (ψ, π) is degenerate, there is still a homomorphism ψ(1) : K(X) → K(M (Y )) ⊂ M (K(Y )), but then there may not be an extension ψ(1) to L(X) (although we will define a partial extension, under additional hypotheses, in Corol- lary A.14). C-multipliers. Unfortunately, we must deal with possibly degenerate homomor- phisms of correspondences, and for this reason we have been lead to develop a generalization of the "C-multipliers" of [EKQR06, Section 1.4]. Example A.1. We can see why degeneracy is an issue for us already in the case of (discrete) directed graphs: let E be the directed graph with a single non-loop edge, i.e., u• e •v We have: • C∗(E) = M2; • X = H(E) = C · χ{e}; • kX (χ{e}) = ( 0 1 0 0 ); • kX (X)C∗(E) = ( ∗ ∗ 0 0 ), so the canonical Toeplitz representation (kX , kA) is degenerate (where A = C2). o o GROUP ACTIONS ON TOPOLOGICAL GRAPHS 31 Remark A.2. Of course, there are situations in which degeneracy is not an issue; for example, if the correspondence X is actually an imprimitivity bimodule (so that ϕA : A → K(X) is an isomorphism), then every Cuntz-Pimsner covariant Toeplitz representation (ψ, π) of X to a C∗-algebra B for which π is nondegenerate is automatically nondegenerate as a correspondence homomorphism, i.e., ψ(X)B = B, by [KQR97, Lemma 5.1]. Note also that whenever E is a topological graph with s and r both surjective and r proper, then kX : X → C∗(E) will be nondegenerate, hence will extend uniquely to the multiplier bimodule M (X). The following is similar to that in [EKQR06]. To prepare for the general frame- work introduced below, we begin with some prefatory comments. First, to establish context, we recall that the "C-multipliers" in [EKQR06] involved tensor products: if A and C are C∗-algebras, then [EKQR06] defined MC(A⊗C) to be all multipliers of A⊗C that multiply 1⊗C into A⊗C. Here we need a generalization to situations where there are no tensor products. Definition A.3. Let κ : C → M (A) be a nondegenerate homomorphism. (1) A C-multiplier of A is a multiplier m ∈ M (A) such that κ(C)m ∪ mκ(C) ⊂ A, and MC(A) denotes the set of all C-multipliers of A. (2) The C-strict topology on MC(A) is generated by the seminorms m (cid:55)→ (cid:107)κ(c)m(cid:107) and (cid:55)→ (cid:107)mκ(c)(cid:107) for c ∈ C. Often C will be a nondegenerate C∗-subalgebra of M (A) and κ will be the inclusion map. The following generalizes [EKQR06, Proposition A.5]: Lemma A.4. (1) The C-strict topology on MC(A) is stronger than the relative strict topology (2) MC(A) is a C∗-subalgebra of M (A), and multiplication and involution are from M (A). separately C-strictly continuous on MC(A). (3) MC(A) is the C-strict completion of A; (4) MC(A) is an M (C)-subbimodule of M (A). Proof. (1) Let mi → 0 C-strictly in MC(A). We must show that if a ∈ A then (cid:107)ami(cid:107) and (cid:107)mia(cid:107) both tend to 0. By the Hewitt-Cohen factorization theorem (see, for example, [RW98, Proposition 2.33]) we can factor a = a(cid:48)κ(c) for some a(cid:48) ∈ A and c ∈ C, and then (cid:107)ami(cid:107) = (cid:107)a(cid:48)κ(c)mi(cid:107) ≤ (cid:107)a(cid:48)(cid:107)(cid:107)κ(c)mi(cid:107) → 0, and similarly for mia. (2) It follows straight from the definitions that MC(A) is a C∗-subalgebra. Let mi → 0 C-strictly in MC(A), and let n ∈ M (A). We must show that min → 0 C-strictly in MC(A), and it will follow similarly that nmi → 0 C-strictly. For c ∈ C we have (cid:107)κ(c)min(cid:107) ≤ (cid:107)κ(c)mi(cid:107)(cid:107)n(cid:107) → 0 because mi → 0 C-strictly, and (cid:107)minκ(c)(cid:107) → 0 32 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG because nκ(c) ∈ A and mi → 0 strictly by (1). It is even easier to verify that m∗ (3) We first show that A is C-strictly dense in MC(A). Let m ∈ MC(A), and let {ei} be an approximate identity for C. We will show that κ(ei)m → m C-strictly. For c ∈ C we have i → 0 C-strictly. (cid:107)κ(c)(κ(ei)m − m)(cid:107) = (cid:107)κ(cei − c) · m(cid:107) ≤ (cid:107)cei − c(cid:107)(cid:107)m(cid:107) → 0, while (cid:107)(κ(ei)m − m)κ(c)(cid:107) = (cid:107)κ(ei)mκ(c) − mκ(c)(cid:107) → 0 because κ(ei) → 0 strictly in M (A). Now we verify the C-strict completeness. Let {mi} be a C-strictly Cauchy net in MC(A). Then it is also strictly Cauchy, so converges strictly to some m ∈ M (A) by strict completeness. We will show that m ∈ MC(A) and mi → m C-strictly. Let > 0 and c ∈ C. Choose i0 such that (cid:107)κ(c)(mi − mj)(cid:107) ≤ i, j ≥ i0. for all Then for all a ∈ A with (cid:107)a(cid:107) ≤ 1 we have (cid:107)(mj − m)a(cid:107) → 0, so for all i ≥ i0 we have (cid:107)κ(c)(mi − m)a(cid:107) = lim j (cid:107)κ(c)(mi − mj)a(cid:107) ≤ lim sup j (cid:107)κ(c)(mi − mj)(cid:107) ≤ , and hence (cid:107)κ(c)(mi− m)(cid:107) ≤ . It follows that (cid:107)κ(c)(mi− m)(cid:107) → 0. Since κ(c)mi ∈ A for all i, we have κ(c)m ∈ A. Similarly, (cid:107)(mi − m)κ(c)(cid:107) → 0 and mκ(c) ∈ A. Thus m ∈ MC(A), and the above arguments then show that mi → m C-strictly. (4) We have and κ(M (C))MC(A)κ(C) ⊂ κ(M (C))A ⊂ A κ(C)κ(M (C))MC(A) ⊂ κ(C)MC(A) ⊂ A, so κ(M (C))MC(A) ⊂ MC(A), and similarly for MC(A)κ(M (C)). (cid:3) The whole point of C-multipliers is to extend degenerate homomorphisms. The following generalizes [EKQR06, Proposition A.6]: Lemma A.5. Let κ : C → M (A), σ : D → M (B), π : A → MD(B), and λ : C → M (σ(D)) be homomorphisms, with κ, σ, and λ nondegenerate, such that π(κ(c)a) = λ(c)π(a) for c ∈ C, a ∈ A. Then π extends uniquely to a C-strict to D-strictly continuous homomorphism π : MC(A) → MD(B). Moreover, for n ∈ M (C) and m ∈ MC(A) we have π(κ(n)m) = λ(n)π(m) and π(mκ(n)) = π(m)λ(n). Note that, in the above lemma, M (σ(D)) is identified with the idealizer of σ(D) in M (B) (and this is valid since σ is nondegenerate), and λ and κ denote the canonical extensions to the multiplier algebras (which exist by nondegeneracy). GROUP ACTIONS ON TOPOLOGICAL GRAPHS 33 Proof. We first show that π is C-strict to D-strictly continuous. Let mi → 0 C- strictly in A, and let d ∈ D. By the Hewitt-Cohen factorization theorem we can factor σ(d) = σ(d(cid:48))λ(c) for some d(cid:48) ∈ D and c ∈ C, and then (cid:107)σ(d)π(mi)(cid:107) = (cid:107)σ(d(cid:48))λ(c)π(mi)(cid:107) = (cid:107)σ(d(cid:48))π(κ(c)mi)(cid:107) ≤ (cid:107)d(cid:48)(cid:107)(cid:107)κ(c)mi(cid:107) → 0, and similarly for π(mi)σ(d). Thus, by completeness there is a unique C-strict to D-strictly continuous linear extension π : MC(A) → MD(B). Since the algebraic operations on MC(A) and MD(B) are separately continuous for the C-strict and D-strict topologies, respec- tively, π is a ∗-homomorphism. (cid:3) The other part follows from continuity and density. Remark A.6. (1) In fact, the above extension π is unique as a homomorphism, but we will not need this. (2) The above lemma could be the starting point for an investigation of what one might call "decorated C∗-algebras" (where A is "decorated" by C and B is decorated by D). However, we will not pursue this any further here. One of our applications of the above result is described by the following special case, in which we will actually need π to take values in B itself: Corollary A.7. Let I be an ideal of A, and let π : A → B be a nondegenerate homomorphism. Then the restriction πI of π to I extends uniquely to an A-strictly continuous homomorphism πI : MA(I) → MA(B). Proof. Apply Lemma A.5 with A, I playing the role of C, A, with the canonical homomorphism ρ : A → M (I) playing the role of κ, with D = A and σ = λ = π, and with πI playing the role of π. Since πI (I) ⊂ B ⊂ MA(B), the hypotheses of the lemma are satisfied. (cid:3) C-multiplier bimodules. The preceding subsection was really only a prelude for the current one, where we generalize the C-multiplier bimodules of [EKQR06, Section 1.4] for tensor products. In addition to the preceding nondegenerate homomorphism κ : C → M (A), suppose that we also have a nondegenerate A-correspondence X, and let ϕA : A → L(X) be the associated homomorphism. Then we can compose to get a nondegenerate homomorphism ϕC := ϕA ◦ κ : C → L(X), where ϕA denotes the canonical extension of ϕA : A → L(X) to M (A), so X becomes a nondegenerate C − A correspondence. Similarly on the right, so X becomes a nondegenerate C-bimodule (but not a correspondence, because the A- valued inner product cannot be turned into a C-valued one). Since M (X) is an M (A)-correspondence, by composing with κ we get an M (C)-bimodule structure on M (X) (but again, not a correspondence). 34 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG Definition A.8. Let X be a nondegenerate A-correspondence, and let κ : C → M (A) be a nondegenerate homomorphism. (1) A C-multiplier of X is a multiplier m ∈ M (X) such that κ(C) · m ∪ m · κ(C) ⊂ X, and MC(X) denotes the set of all C-multipliers of X. (2) The C-strict topology on MC(X) is generated by the seminorms m (cid:55)→ (cid:107)κ(c) · m(cid:107) and m (cid:55)→ (cid:107)m · κ(c)(cid:107) c ∈ C. for The C-multiplier algebra MC(A) from the preceding subsection is the special case of MC(X) where X = A regarded as an A-correspondence in the usual way. The following generalizes [EKQR06, Lemma 1.40]: Lemma A.9. (1) The C-strict topology is stronger than the relative strict topology on MC(X). (2) MC(X) is an MC(A)-correspondence with respect to the restrictions of the operations of the M (A)-correspondence M (X), and the operations are sep- arately C-strictly continuous. (3) K(MC(X)) ⊂ MC(K(X)). (4) MC(X) is the C-strict completion of X. (5) MC(X) is an M (C)-subbimodule of M (X). Proof. Many of the arguments are quite similar to those in the proof of Lemma A.4. (1) If mi → 0 C-strictly in MC(X), we must show that if T ∈ K(X) and a ∈ A then both (cid:107)T mi(cid:107) and (cid:107)mi·a(cid:107) tend to 0, and the argument is similar to Lemma A.4. (2) If n ∈ MC(A) and m ∈ MC(X) then n · m is in M (X) because M (X) is an M (A)-correspondence; we must show that it is in MC(X). For c ∈ C, note that κ(c)n ∈ A, so because κ : C → M (A) is nondegenerate we can factor κ(c)n = aκ(c(cid:48)) with a ∈ A and c(cid:48) ∈ C, and then we have κ(c) · (n · m) = (κ(c)n) · m = (aκ(c(cid:48))) · m = a · (κ(c(cid:48)) · m) ∈ X because n · κ(c) ∈ X and (cid:104)M (X), X(cid:105) ⊂ A, and similarly κ(c)(cid:104)m, n(cid:105) ∈ A. strictly in MC(X), n ∈ MC(X), and c ∈ C, then The separate continuity is similar to Lemma A.4. For example, if mi → 0 C- (cid:107)(cid:104)mi, n(cid:105)κ(c)(cid:107) = (cid:107)(cid:104)mi, n · κ(c)(cid:105)(cid:107) → 0 because n · κ(c) ∈ X, mi → 0 strictly by (1), and the operations in M (X) are separately strictly continuous, while (cid:107)κ(c)(cid:104)mi, n(cid:105)(cid:107) = (cid:107)(cid:104)mi · κ(c∗), n(cid:105)(cid:107) → 0 because (cid:107)mi · κ(c∗)(cid:107) → 0. because κ(c(cid:48)) · m ∈ X. On the other hand, we have (n · m) · κ(c) = n · (m · κ(c)) ∈ X because m · κ(c) ∈ X and n ∈ M (A). The M (A)-valued inner product on M (X) restricts to one on MC(X); we must show that for m, n ∈ MC(X) the inner product (cid:104)m, n(cid:105) is in MC(A). For c ∈ C we have (cid:104)m, n(cid:105)κ(c) = (cid:104)m, n · κ(c)(cid:105) ∈ A, GROUP ACTIONS ON TOPOLOGICAL GRAPHS 35 (3) Here the only issue is keeping straight where everything is: we have MC(X) ⊂ M (X) ⊂ L(A, X), and for m, n ∈ MC(X) the rank-one operator θm,n = mn∗ is therefore an element of L(X) = M (K(X)). We need to know that mn∗ is a C-multiplier of K(X). So, take c ∈ C. Since m ∈ MC(X) we have so φC(c)m ∈ X = K(A, X), φC(c)mn∗ ∈ K(X). mn∗ΦC(c) ∈ K(X). Similarly, φC(c∗)n ∈ K(X, A), so n∗ΦC(c) ∈ K(A, X), and hence Thus we have shown that mn∗ ∈ MC(K(X)). (4) and (5) are similar to Lemma A.4. For example, (cid:0)κ(M (C)) · MC(X)(cid:1) · κ(C) ⊂ κ(M (C)) · X ⊂ X, κ(C) ·(cid:0)κ(M (C)) · MC(X)(cid:1) ⊂ κ(C) · MC(X) ⊂ X. because C acts nondegenerately on X, while (cid:3) Remark A.10. It will be useful to explicitly note the following special case of (3) above: taking C = A and κ = idA we get K(MA(X)) ⊂ MA(K(X)). Note that MA(X) = {m ∈ M (X) : φA(a)m ∈ X for all a ∈ A}, so that, unlike M (X), the set MA(X) depends upon the left module map φA. The following generalizes [EKQR06, Proposition 1.42]: Proposition A.11. Let X and Y be nondegenerate correspondences over A and B, respectively, let κ : C → M (A) and σ : D → M (B) be nondegenerate homo- morphisms, let (ψ, π) : X → MD(Y ) be a correspondence homomorphism, and let λ : C → M (σ(D)) be a nondegenerate homomorphism, such that π(κ(c)a) = λ(c)π(a) for c ∈ C, a ∈ A. Then (ψ, π) extends uniquely to a C-strict to D-strictly continuous correspondence homomorphism (ψ, π) from the MC(A)-correspondence MC(X) to the MD(B)- correspondence MD(Y ). Proof. We take π : MC(A) → MD(B) as in Lemma A.5. The unique existence of ψ is proved similarly to Lemma A.5: one shows first that ψ is C-strict to D- strictly continuous, so that by completeness there is a unique continuous linear extension, which must be a correspondence homomorphism by separate continuity (cid:3) of the operations. Remark A.12. Again, the above proposition could be the start of an investigation of what one might call "decorated correspondences", where C decorates the A- correspondence X, etc. Here is the Toeplitz version of Proposition A.11: 36 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG Corollary A.13. Let X be a nondegenerate A-correspondence, let κ : C → M (A) and σ : D → M (B) be nondegenerate homomorphisms, let (ψ, π) : X → B be a Toeplitz representation, and let λ : C → M (σ(D)) be a nondegenerate homomor- phism, such that π(κ(c)a) = λ(c)π(a) for c ∈ C, a ∈ A. Then (ψ, π) extends uniquely to a C-strict to D-strictly continuous Toeplitz repre- sentation (ψ, π) from the MC(A)-correspondence MC(X) to MD(B). Proof. Apply Proposition A.11 with Y = B, regarded as a B-correspondence in the (cid:3) usual way. Corollary A.14. Let X be a nondegenerate A-correspondence, and let (ψ, π) : X → B be a Toeplitz representation with π nondegenerate. Then: (1) (ψ, π) extends uniquely to an A-strictly continuous Toeplitz representation (2) ψ(1) : K(X) → B extends uniquely to an A-strictly continuous homomor- (ψ, π) from the M (A)-correspondence MA(X) to MA(B). phism ψ(1) : MA(K(X)) → MA(B). Moreover, we have: (a) ψ(1)(n · m) = π(n)ψ(1)(m) and ψ(1)(m · n) = ψ(1)(m)π(n) for all n ∈ M (A) and m ∈ MA(K(X)); (b) ψ(1)(mn∗) = ψ(m)ψ(n)∗ for all m, n ∈ MA(X). Proof. (1) Apply Corollary A.13 with C = D = A, κ = λ = idA, and σ = π. (2) First we need to justify the meaning of (b): for m, n ∈ MA(X) we need to know that mn∗ ∈ MA(K(X)). But this follows easily from work we have already done: since X is a K(X) − B imprimitivity bimodule, arguments similar to those in the proof of Lemma A.9 (2) show that the left-hand inner product on MA(X) takes values in MA(K(X)). To prove (2), we apply Corollary A.5 with C, A, B, D replaced by A,K(X), B, A, respectively; recall that for a ∈ A and T ∈ K(X) we have ψ(1)(ϕ(a)T ) = π(a)ψ(1)(T ). Now everything in (2) except for (b) follows from Corollary A.5. Then (b) follows from separate continuity of the operations, density of A, X, and K(X) in M (A), MA(X), and MA(K(X)), respectively, and the corresponding property of ψ(1). (cid:3) Example A.15. We should justify our need for (2) in above corollary: if (kX , kA) is the canonical Toeplitz representation of the correspondence X = H(E) associated X : K(X) → C∗(E) can be degenerate. For to a topological graph E, then k(1) example, let E be the directed graph in Example A.1. Then X (K(X)) = ( ∗ 0 k(1) 0 0 ) , so X (K(X))C∗(E) = ( ∗ 0 k(1) 0 0 ) M2 = ( ∗ ∗ 0 0 ) , and so k(1) X is degenerate. Remark A.16. Although we will have no use for it in the main body of this paper, we point out that all of the above can be done with an "asymmetric" version of the C-multiplier bimodules; namely allowing correspondences with different left and GROUP ACTIONS ON TOPOLOGICAL GRAPHS 37 right coefficient algebras. It takes no extra effort to establish the more general concepts and results. Since we feel that they may be useful elsewhere, we record here the asymmetric versions: Definition A.17. Let X be a nondegenerate A − B correspondence, and let κC : C → M (A) and κD : D → M (B) be nondegenerate homomorphisms. (1) A C, D-multiplier of X is a multiplier m ∈ M (X) such that κC(C) · m ∪ m · κD(D) ⊂ X, and MC,D(X) denotes the set of all C, D-multipliers of X. (2) The C, D-strict topology on MC,D(X) is generated by the seminorms m (cid:55)→ (cid:107)κC(c) · m(cid:107) and m (cid:55)→ (cid:107)m · κD(d)(cid:107) c ∈ C, d ∈ D. for Lemma A.18. (1) The C, D-strict topology is stronger than the relative strict topology on MC,D(X). (2) MC,D(X) is an MC(A) − MD(B) correspondence with respect to the re- strictions of the operations of the M (A) − M (B) correspondence M (X), and the operations are separately continuous for the C, D-strict, C-strict, and D-strict topologies. (3) MC,D(X) is the C, D-strict completion of X. (4) MC,D(X) is an M (C) − M (D) subbimodule of M (X). Proposition A.19. Let X and Y be nondegenerate A − B and P − Q corre- spondences, respectively, let κC : C → M (A), κD : D → M (B), κR : R → M (P ), and κS : S → M (Q) be nondegenerate homomorphisms, let πψθ : AXB → MR(P )MR,S(Y )MS (Q) be a correspondence homomorphism, and let λC : C → M (κR(R)) and λD : D → M (κS(S)) be nondegenerate homomorphisms, such that for c ∈ C, a ∈ A, b ∈ B, and d ∈ D we have π(κC(c)a) = λC(c)π(a) and θ(κD(d)b) = λD(d)θ(b). Then πψθ extends uniquely to a C, D-strict to R, S-strictly continuous correspon- dence homomorphism πψθ from the MC(A) − MD(B) correspondence MC,D(X) to the MR(P ) − MS(Q) correspondence MR,S(Y ). References [BH99] [Bau93] [ADR00] C. Anantharaman-Delaroche and J. Renault, Amenable groupoids, Monographies de L'Enseignement Math´ematique [Monographs of L'Enseignement Math´ematique], vol. 36, L'Enseignement Math´ematique, Geneva, 2000, With a foreword by Georges Skandalis and Appendix B by E. Germain. G. Baumslag, Topics in combinatorial group theory, Lectures in Mathematics ETH Zurich, Birkhauser Verlag, Basel, 1993. M. R. Bridson, A. Haefliger, Metric spaces of non-positive curvature, Grundlehren der Mathematischen Wissenschaften, 319. Springer-Verlag, Berlin, 1999. N. P. Brown, N. Ozawa, C∗-algebras and finite-dimensional approximations, Graduate Studies in Mathematics, 88. American Mathematical Society, Providence, RI, 2008. [EKQR06] S. Echterhoff, S. Kaliszewski, J. Quigg, and I. Raeburn, A Categorical Approach to Imprimitivity Theorems for C∗-Dynamical Systems, vol. 180, Mem. Amer. Math. Soc., no. 850, American Mathematical Society, Providence, RI, 2006. [FLR00] N. J. Fowler, M. Laca, I. Raeburn, The C∗-algebras of infinite graphs, Proc. Amer. [BO08] Math. Soc. 128 (2000), 2319 -- 2327. 38 VALENTIN DEACONU, ALEX KUMJIAN, AND JOHN QUIGG [GT01] [HN08] J. L. Gross and T. W. Tucker, Topological graph theory, Dover Publications Inc., Mineola, NY, 2001, Reprint of the 1987 original [Wiley, New York] with a new preface and supplementary bibliography. G. Hao and C.-K. Ng, Crossed products of C∗-correspondences by amenable group actions, J. Math. Anal. Appl. 345 (2008), no. 2, 702 -- 707. [HRW05] A. an Huef, I. Raeburn, and D. P. Williams, A symmetric imprimitivity theorem for [Hus94] [KQR97] commuting proper actions, Canad. J. Math. 57 (2005), 983 -- 1011. D. Husemoller, Fibre bundles, third ed., Graduate Texts in Mathematics, vol. 20, Springer-Verlag, New York, 1994. S. Kaliszewski, J. Quigg, and I. Raeburn, Duality of restriction and induction for C∗-coactions, Trans. Amer. Math. Soc. 349 (1997), 2085 -- 2113. [KQR08] , Proper actions, fixed-point algebras and naturality in nonabelian duality, J. [KQR10] [Kas88] [Kat02] [Kat04] [Kat06] [Kat08] [KK97] [KP99] [LS77] Funct. Anal. 254 (2008), 2949 -- 2968. , Skew products and coactions for topological graphs, preprint, 2010. G. G. Kasparov, Equivariant KK-theory and the Novikov conjecture, Invent. Math. 91 (1988), no. 1, 147 -- 201. T. Katsura, Continuous graphs and crossed products of Cuntz algebras, S¯urikaisekikenky¯usho K¯oky¯uroku (2002), no. 1291, 73 -- 83, Recent aspects of C∗- algebras (Japanese) (Kyoto, 2002). , A class of C∗-algebras generalizing both graph algebras and homeomorphism C∗-algebras. I. Fundamental results, Trans. Amer. Math. Soc. 356 (2004), no. 11, 4287 -- 4322. , A class of C∗-algebras generalizing both graph algebras and homeomorphism C∗-algebras. III. Ideal structures, Ergodic Theory Dynam. Systems 26 (2006), no. 6, 1805 -- 1854. , A class of C∗-algebras generalizing both graph algebras and homeomorphism C∗-algebras. IV. Pure infiniteness, J. Funct. Anal. 254 (2008), no. 5, 1161 -- 1187. A. Kishimoto and A. Kumjian, Crossed products of Cuntz algebras by quasi-free au- tomorphisms, Operator algebras and their applications (Waterloo, ON, 1994/1995), Fields Inst. Commun., vol. 13, Amer. Math. Soc., Providence, RI, 1997, pp. 173 -- 192. A. Kumjian and D. Pask, C∗-algebras of directed graphs and group actions, Ergod. Th. and Dynam. Sys. 19 (1999), 1503 -- 1519. R. C. Lyndon, P. E. Schupp, Combinatorial group theory, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 89. Springer-Verlag, Berlin-New York, 1977. [Mas91] W. S. Massey, A basic course in algebraic topology, Graduate Texts in Mathematics, vol. 127, Springer-Verlag, New York, 1991. D. Olesen and G. K. Pedersen, Applications of the Connes spectrum to C∗-dynamical systems, J. Funct. Anal. 30 (1978), no. 2, 179 -- 197. , Applications of the Connes spectrum to C∗-dynamical systems. II, J. Funct. [OP78] [OP80] [Pal61] [QR95] [Qui92] [Rae05] [Rie90] [Rie04] [RW98] Anal. 36 (1980), no. 1, 18 -- 32. R. S. Palais, On the existence of slices for actions of non-compact Lie groups, Ann. of Math. (2) 73 (1961), 295 -- 323. J. C. Quigg and I. Raeburn, Induced C∗-algebras and Landstad duality for twisted coactions, Trans. Amer. Math. Soc. 347 (1995), 2885 -- 2915. J. C. Quigg, Landstad duality for C∗-coactions, Math. Scand. 71 (1992), 277 -- 294. I. Raeburn, Graph algebras, CBMS Regional Conference Series in Mathematics, vol. 103, Published for the Conference Board of the Mathematical Sciences, Washington, DC, 2005. M. A. Rieffel, Proper actions of groups on C∗-algebras, Mappings of operator algebras (Philadelphia, PA, 1988) (Boston, MA), Birkhauser Boston, 1990. , Integrable and proper actions on C∗-algebras, and square-integrable represen- tations of groups, Expo. Math. 22 (2004), 1 -- 53. I. Raeburn and D. P. Williams, Morita equivalence and continuous-trace C∗-algebras, Math. Surveys and Monographs, vol. 60, American Mathematical Society, Providence, RI, 1998. GROUP ACTIONS ON TOPOLOGICAL GRAPHS 39 Department of Mathematics, University of Nevada, Reno NV 89557-0084, USA E-mail address, Valentin Deaconu: [email protected] Department of Mathematics, University of Nevada, Reno NV 89557-0084, USA E-mail address, Alex Kumjian: [email protected] School of Mathematical and Statistical Sciences, Arizona State University, Tempe AZ 85287-1804, USA E-mail address, John Quigg: [email protected]
1511.06723	1	1511	2015-11-20T18:42:07	Rank Constrained Homotopies	[ "math.OA", "math.AT" ]	For any $n\geq k\geq l\in\mathbb{N},$ let $S(n,k,l)$ be the set of all those non-negative definite matrices $a\in M_{n}(\mathbb{C})$ with $l\leq\text{rank }a\leq k$. Motivated by applications to $C^{}$-algebra theory, we investigate the homotopy properties of continuous maps from a compact Hausdorff space $X$ into sets of the form $S(n,k,l).$ It is known that for any $n,$ if $k-l$ is approximately 4 times the covering dimension of $X$ then there is only one homotopy class of maps from $X$ into $S(n,k,l)$, i.e. $C(X,S(n,k,l))$ is path connected. In our main Theorem we improve this bound by a factor of 8. By combining classical homotopy theory methods with $C^{}$-algebraic techniques we also show that if $\pi_{r}(S(n,k,l))$ vanishes for all $r\leq d$ then $C(X,S(n,k,l))$ is path connected for any compact Hausdorff $X$ with covering dimension not greater than $d$.	math.OA	math	RANK CONSTRAINED HOMOTOPIES KAUSHIKA DE SILVA Abstract. For any n ≥ k ≥ l ∈ N, let S(n, k, l) be the set of all those non-negative definite matrices a ∈ Mn(C) with l ≤ rank a ≤ k. Motivated by applications to C ∗-algebra theory, we investigate the homotopy properties of continuous maps from a compact Hausdorff space X into sets of the form S(n, k, l). It is known that for any n, if k − l is approximately 4 times the covering dimension of X then there is only one homotopy class of maps from X into S(n, k, l), i.e. C(X, S(n, k, l)) is path connected. In our main Theorem we improve this bound by a factor of 8. By combining classical homotopy theory methods with C ∗-algebraic techniques we also show that if πr(S(n, k, l)) vanishes for all r ≤ d then C(X, S(n, k, l)) is path connected for any compact Hausdorff X with covering dimension not greater than d. 1. Introduction. For n ∈ N, let (Mn)+ stand for the n × n non-negative definite matrices over C and for fixed k, l ∈ N with n ≥ k ≥ l define the space S(n, k, l) by, S(n, k, l) = {b ∈ (Mn)+ l ≤ rank(b) ≤ k} . Let S(n, k, l) be given the subspace topology induced by the norm topology of Mn. We focus on homotopy properties of S(n, k, l). Our main theorem is the following; (cid:5) ≤ k − l. Then Theorem 3.5. Let X be a compact Hausdorff space X with (cid:4) dim X there is only one homotopy class of functions f : X → S(n, k, l), i.e. the function space C(X, S(n, k, l)) is path connected. 2 Here, by dim X we mean the covering dimension of the space X and ⌊m⌋ stands for the largest integer which is less than or equal to m ∈ R. An immediate consequence of Theorem 3.5 is that πr(S(n, k, l)) = 0, ∀r ≤ 2(k − l) + 1. The spaces S(n, k, l), or rather their homotopy properties have relevance in C∗- algebra theory and our motivation for this study is mainly due to this natural connection. For a unital C∗-algebra A, let T (A) denote the tracial state space. In other words T (A) is the set of all normalized positive linear functionals τ on A which satisfy the trace property τ (ab) = τ (ba), ∀a, b ∈ A. An important object of study for a C∗-algebraist is the space LAf fb(T (A))++ of all bounded, strictly positive, lower semi-continuous affine maps on T (A). For a positive element a in A, τ 7→ ιa(τ ) = limn τ (a1/n) defines a positive lower semi-continuous affine map on T (A). This definition extends to positive elements in Mn(A), n ≥ 2, in which case τ in the right hand side stands for the non-normalized canonical extension of τ to Mn(A). If A is simple ιa is strictly positive for all a ∈ Mn(A) for any n ∈ N. For sufficiently regular simple C∗-algebras (as in [2],[13]), {ιa : a ∈ M∞(A)+} is dense in LAf fb(T (A))++. 1 RANK CONSTRAINED HOMOTOPIES 2 On the other hand, density of {ιa : a ∈ M∞(A)+} ⊆ LAf f (T (A))++ provides one with useful tools to work with to derive classification results and to establish regularity properties such as Z-stability [3, 13]. In [13, Theorem 3.4], Toms shows that for unital simple ASH algebras with slow dimension growth {ιa : a ∈ M∞(A)} is dense in LAf fb(T (A))++. He then uses this fact to show that such algebras are Z-stable and hence derives classification results. The proof of [13, Theorem 3.4] heavily depends on homotopy properties of the spaces S(n, k, l) [13, Section 2]. Our main result here is an improvement of [13, Proposition 2.5], which is a key ingredient in proving that {ιa : a ∈ M∞(A)+} ⊆ LAf f (T (A))++ is dense for the class of algebras mentioned above. However, one should note that this improvement does not have a direct impact on the main result of [13]. Spaces S(n, k, l) as a generalization of complex Grammarians. Let us consider the special case k = l. Recall that Gk(Cn) stands for the complex Grassmann variety of k-dimensional subspaces of Cn. Given a k -dimensional sub- space V of Cn, one may identify it uniquely with the orthogonal projection of Cn on V . This identification leads to a natural homeomorphism from Gk(Cn) to Pk(Cn), where Pk(Cn) is the set of all rank k projections in Mn(C) equipped with the norm topology. It is not hard to see that the inclusion Pk(Cn) ⊂ S(n, k, k) induces a homotopy equivalence. Hence, Gk(Cn) is homotopy equivalent to S(n, k, k). Thus in a sense, the spaces S(n, k, l) can be viewed as a generalization of the Grassmann varieties, at least for homotopy interests. By setting k = l in Theorem 3.5 and using the homotopy equivalence of S(n, k, k) with Gk(Cn) one can derive that Gk(Cn) is simply connected (i.e. π1(Gk(Cn)) = 0) for any pair of k, n, which is a well known classical result. However, it should be noted that our proof of Theorem 3.5 does not in provide a alternate proof of this fact, rather we use a stronger classical result in our proof. Rank varying bundles associated with the spaces S(n, k, l). Given a continuous map f : X → Gk(Cn), one has the associated locally trivial vector bundle f ∗(γ) over X, which is the pullback of the canonical k-dimensional vector bundle γ over Gk(Cn) to X. In a similar vein, for each map a ∈ C(X, (Mn)+) there is a naturally associated bundle (in the sense of [7, Chapter 2]) ǫa over X with the total space, Ea = {(x, v) ∈ X × Cnv ∈ a(x)(Cn)} and π1 : Ea → X being the natural coordinate projection. Admittedly, a typical bundle obtained in this nature is not necessarily a vector bundle in the classical sense ([1, 7]). In fact such a bundle may not even have a constant fiber. Still, for a bundle ǫa each fiber admits a natural vector space structure, which to an extent preserves local triviality. Moreover these bundles occur naturally in C∗- algebra theory (see [14],[15]) and understanding this association would make the techniques used in section 3 more intuitive. Thus, before moving on to the next section we outline the structure of these bundles. RANK CONSTRAINED HOMOTOPIES 3 The structure of bundles associated with the spaces S(n, k, l). Observe that if a(x) ∈ S(n, k, k), ∀x ∈ X, then the bundle ǫa is indeed a lo- cally trivial vector bundle over X. However, if a is not of constant rank the associated bundle ǫa is not locally trivial in the usual sense. Still, if we set Ei = {x ∈ X : rank(a(x)) = ni}, then continuity of a implies that the support projection of a is continuous on Ei. Here, by support projection of a we mean the function which assigns each x to the orthogonal projection on the subspace a(x)(Cn). Hence, the restriction of ǫa to Ei - denoted by ǫa ↾Ei, is a locally trivial vector bundle over Ei. In this manner we can partition X into a finite collection of subsets, such that the restriction of ǫa to each subset is a locally trivial bun- dle of constant rank. One may now consider the possibility of applying classical vector bundle theory [1, 7] to establish the structure of ǫa in some local sense, i.e the structure of ǫa ↾Ei for each i. But the issue with this is that the subsets Ei formed here are highly non regular where as to apply classical theory of vector bun- dles the base spaces need to satisfy regularity properties such as compactness. We can overcome this if qi- the support projection of a on Ei, extends to a projection pi ∈ Mn(C(Ei)) for each i, where Ei is the closure Ei. If this is the case, then we can apply classical vector bundle theory (see [1, 7]) to understand the structure of bundles ǫpi. Moreover, if its also the case that for any two distinct i, j with Ei ∩ Ej 6= ∅ the extended projections pi and pj are comparable on Ei ∩ Ej, then we can expect to use structural properties of the bundles ǫpi (i.e local structure of ǫa) to establish the global structure of ǫa. The above considerations would not hold true for arbitrary a ∈ Mn(C(X))+. How- ever, [[15], c.f. [10]] introduces a special class of elements in Mn(C(X))+ called well supported positive elements, which are well behaved in the above sense. For an element a in this class (Definition 2.2), each qi extends to pi ∈ Mn(C(Ei)) in such a way that for i < j with Ei ∩ Ej 6= ∅, pi is a sub-projection of pj of on Ei ∩ Ej. From results in [14] it follows that for any n, k, l ∈ N the set of all well supported positive elements contained in C(X, S(n, k, l)) is homotopy equivalent to C(X, S(n, k, l)) [Lemma 3.1]. This fact plays a crucial role in the proof of our main theorem. In section 2, we briefly recall some well known definitions and results from vector bundle theory [1, 7] as well as some other tools and notations that are necessary. In section 3, following the ideas from [10] and [13] we prove the main result. Sec- tion 4 (Theorem 4.6) shows that for any n, k, l ∈ N the path connectedness of C(X, S(n, k, l)) depends solely on homotopy of the space S(n, k, l). In section 4 we do not assume that k −l ≥ ⌊ dim X ⌋ and the proof of Theorem 4.6 is achieved mainly through applications of classical homotopy theory results (see [16]), to the space S(n, k, l). 2 Acknowledgments. I thank my adviser Professor Andrew Toms for his encourage- ment, for several helpful conversations and suggestions that were invaluable. This work wouldn't have been possible without his efforts. I am also thankful to the reviewer of an earlier version of this article, for providing thoughtful suggestions, comments and references. RANK CONSTRAINED HOMOTOPIES 4 2. Notations and Preliminaries. Notations and Conventions. Unless stated otherwise we assume X to be a compact Hausdorff space. C(X, Y ), Mn×m(C(X)), Mn(C(X)) all have the usual meanings and C(X) = C(X, C). We will often identify Mn(C(X)) with C(X, Mn), where Mn denotes Mn(C). (Mn)+ denotes all non negative definite matrices in Mn and its customary to call these as positive elements of Mn. Given x ∈ Mn, x∗ denotes the conjugate transpose of x. By a projection p in Mn we mean a self adjoint idempotent, i.e p = p2 = p∗. Note that the notions of the conjugate of an element, positive elements and projections carry over to Mn(C(X)) via point wise defined operations. We will use Mn(C(X))+ to denote positive elements in Mn(C(X)). The base field for all vector spaces and vector bundles that we will consider is the field of complex numbers. Murray-von Neumann semi group of C(X), semi group of isomorphism classes of vector bundles over X and Serre-Swan Theorem. m∈N P (Mm(C(X))), where P (Mm(C(X))) denote the set of all Let P∞(C(X)) = S projections in Mm(C(X)). A pair p ∈ P (Mm(C(X))) and q ∈ P (Mn(C(X))) are said to be Murray-von Neumann equivalent, denoted p ∼ q, if there is some v ∈ Mn,m(C(X)) with p = v∗v and q = vv∗. The Murray-von Neumann equivalence class of p ∈ P∞((C(X))) is denoted by [p]0 and the Murray-von Neumann semigroup of C(X) is the semigroup D(C(X)) = (P∞(C(X))/ ∼, +) , where [p]0 + [q]0 = [p ⊕ q]0. Let Bun(X) denote the set of locally trivial complex vector bundles over X, and let Bunk(X) denote the set of all ǫ ∈Bun(X) of constant fiber dimension k. Write Vect(X) to denote the semigroup of all isomorphism classes of bundles in Bun(X) with addition being induced via the direct sum of bundles. The well known natural identification of D(C(X)) with Vect(X), induced through the Serre-Swan Theorem [12, Theorem 2] (which states that Bun(X) as a category with morphisms being the morphisms of vector bundles is equivalent to the category of finitely generated projective C(X)-modules), is central to our study. For the sake of completeness and to introduce some of the notations and terms that we will use, let us describe this identification. For p ∈ P (Mm(C(X))) ⊂ Mn(C(X))+, let ǫp = (Ep, π1, X) be the bundle over X defined as in the introduction. Since p is a continuous projection, ǫp is a locally trivial vector bundle over X. Moreover, p ∼ q for some q ∈ P∞(C(X)), if and only if ǫp ∼= ǫq. This gives a well defined injection ψ : D(C(X)) → Vect(X), which is a semigroup morphism and preserves dimensions. On the other hand, suppose ǫ is a locally trivial vector bundle over X, with total space E and the fiber at x ∈ X being Ex. From [12, Corollary 5], ǫ is a direct summand of a trivial bundle. That is, there is some n ∈ N and a bundle ǫ⊥ over X such that ǫ ⊕ ǫ⊥ ∼= θn, where θn is the n dimensional product bundle. Bundle ǫ⊥ is called a complementary bundle for ǫ. Let F denote the total space of ǫ⊥and let Fx denote the fiber at x ∈ X. Then, we may assume that Ex ⊕ Fx = Cn, ∀x ∈ X. Let pǫ(x) denote the orthogonal RANK CONSTRAINED HOMOTOPIES 5 projection of Cn on the fiber Ex ⊂ Cn. Then, the local triviality of ǫ ensures that the map x 7→ pǫ(x) is continuous. Indeed, ψ(pǫ) = ǫ and thus ψ is a surjection. Recall that a vector bundle η over X is called trivial if η is isomorphic to a product bundle over X. We say that a projection p in Mn(C(X)) is a trivial projection if ǫp is trivial, where ǫp is defined as before. From the preceding paragraph it is clear that if p, q ∈ Mn(C(X)) with p ∼ q, then p is trivial iff q is trivial. Bounding the dimension of ǫ⊥. From the preceding for each ǫ ∈ Bun(C(X)) there is ǫ⊥ so that ǫ ⊕ ǫ⊥ ∼= θn for some n ∈ N. We would like to have a bound on the dimension on ǫ⊥. Having such a bound will be useful for our work in section 3 as there we focus on S(n, k, l) for a fixed triple (n, k, l). Let us now address this concern and provide a bound for the dimension of ǫ⊥ which depends only on the dimension of the base space X. Recall the following Theorem on locally trivial vector bundles over a compact Haus- dorff space. Here, ⌈x⌉ stands for the smallest integer n with x ≤ n. Theorem 2.1. Let X be a finite dimensional compact Hausdorff space. Let ǫ, γ ∈ Bun(X) and θn denote the product bundle of dimension n. Let k be the dimension of ǫ and write m = ⌊ dim X 1. ǫ ∼= η ⊕ θk−m for some η ∈ Bunm(X). 2. If k ≥ (cid:6) dim X (cid:7) and ǫ ⊕ δ ∼= γ ⊕ δ, where δ is a trivial bundle over X, then ǫ ∼= γ. ⌋. The following hold, 2 2 A proof of Theorem 2.1 in the case of X being a CW -complex is given by Husemoller [7, Chap. 9, Theorems 1.2 and 1.5]. Any compact Hausdorff space X of dimension d can be realized as the inverse limit of compact metric spaces Xα, with dim Xα ≤ d for each α [9, Chap 27, Theorem 8] and any compact metric space Y is homeomor- phic to an inverse limit of finite simplicial complexes of dimension not greater than that of Y [9, Chap. 27, Theorem 12]). In [6, Theorem 2.5], Goodearl uses these identifications and Husmoller's results [7, Chap. 9, Theorems 1.2 and 1.5] to prove the Theorem for arbitrary compact Hausdorff spaces. 2 2 2 2 2 2 If t > (cid:4) dim X (cid:5), by part 1 of the Theorem, δ = γ ⊕ θt−⌊ dim X ⌋(X). If dimension of ǫ ⊕ γ is k1, then k1 = k + (cid:4) dim X One immediate consequence of Theorem 2.1 is that, given ǫ ∈ Bunk(X) we may choose a complementary bundle ǫ⊥ so that the dimension of ǫ⊥ is at most (cid:4) dim X (cid:5). To observe this, first choose some δ ∈ Bun(X), say of dimension t, so that ǫ ⊕ δ ∼= ⌋ for some θk+t. (cid:5) ≥ (cid:6) dim X (cid:7) γ ∈ Bun⌊ dim X and (ǫ ⊕ γ) ⊕ θt−⌊ dim X ⌋ ∼= θk+t. Hence by part 2 of Theorem 2.1, it follows that ǫ ⊕ γ ∼= θk+⌊ dim X ⌋. Thus for each ǫ ∈ Bunk(X) we may choose ǫ⊥ to be of dimension k + (cid:4) dim X (cid:5) . (cid:5), by following Upshot of all this is that for fixed n, k ∈ N with n ≥ k + (cid:4) dim X the methods discussed in the preceding subsection we can now construct a bijective correspondence between the isomorphism classes of Bunk(X) and Murray-von Neu- mann equivalence classes of projections in Pk(Mn(X)). For convenience let us call this the Serre-Swan correspondence. In section 3, we will combine this correspon- dence with part 1 of Theorem 2.1 to construct continuous maps a : X → (Mn)+, which satisfy specified rank constrains. 2 2 2 2 RANK CONSTRAINED HOMOTOPIES 6 Well supported positive elements. Given a ∈ Mn(C(X))+, the rank function of a - denoted ra, is the (lower semi- continuous) function defined on X by x 7→ rank(a(x)). For a, b ∈ Mn(C(X)) if a − b ∈ Mn(C(X))+, we write a ≤ b. We now give the definition of a well supported positive element in Mn(C(X)), given in [15]. Definition 2.2. Let X be a compact Hausdorff space and let a ∈ Mn(C(X))+. Suppose that n1 < n2 < · · · < nL denote all the values that ra takes and set Ei = {x ∈ X : ra(x) = ni}. We say that a is well supported if, for each 1 ≤ i ≤ L there is a projection pi ∈ Mn(C(Ei)) such that limr→∞ a(x)1/r = pi(x), ∀x ∈ Ei, and pi(x) ≤ pj(x) whenever x ∈ Ei ∩ Ej, and i ≤ j. Following Theorem can be used to replace arbitrary positive elements by well sup- ported ones up to homotopy (see Lemma 3.1). Theorem 2.3. [14, Theorem 3.9]. Let X be a compact Hausdorff space and let a ∈ Mn(C(X))+. Then, for every δ > 0, there exists a well supported element b ∈ Mn(C(X))+ such that b ≤ a and a − b < δ, with the range of rb equal to the range of ra. Remark 2.4. The Theorem stated above is only a part of Theorem 3.9 of [14]. There, in the hypothesis X is assumed to be a finite simplicial complex. However, the simplicial structure of X is required only for the second part of the Theorem, which guarantees that each Ei corresponding to b (as defined in 2.2) can assumed to be a sub-complex of X. For the Theorem as stated here, X being compact and Hausdorff suffices. Some useful extension results. Given a function f on X and a Z ⊂ X, we use f ↾Z to denote the restriction of f to Z. Let a ∈ Mn(C(X))+ be well supported and fix a closed subset Y ⊂ X. Let q be a trivial projection on Y , majorized by the support projection of a ↾Y . In section 3 we would require to extend q to a trivial projection defined on X, such that extension is also majorized by the support projection of a. To accomplish this, we follow ideas developed by Toms [15] based on the work of Phillips [10]. Theorem 2.5. [10, Proposition 4.2 (1)]. Let X be a compact Hausdorff space of dimension d < ∞, and let Y ⊂ X be closed. Let p, q ∈ Mn(C(X)) be projections with the property that rank(q(x)) +(cid:4) d 2(cid:5) ≤ rank(p(x)), ∀x ∈ X. Let s0 ∈ Mn(C(Y )) be such that s∗ 0 ≤ p ↾Y . It follows that there is s ∈ Mn(C(X)) such that s∗s = q, ss∗ ≤ p, and s0 = s ↾Y . 0s0 = q ↾Y and s0s∗ Corollary 2.6. [15, corollary 2.7. (ii)]. Let X be a compact Hausdorff space of dimension d < ∞, and let E1, ..., Ek be a cover of X by closed sets. Let q ∈ P (Mn(C(X))) and for each i ∈ 1, ..., k let pi ∈ Mn(C(Ei)) be a projection of constant rank ni. Assume that n1 < n2 < · · · < nk and pi(x) ≤ pj(x) whenever 2(cid:5) for every i. Then x ∈ Ei ∩ Ej and i ≤ j. Finally, suppose that ni − rank (q) ≥ (cid:4) d the following hold: RANK CONSTRAINED HOMOTOPIES 7 If Y ⊂ X is closed, q ↾Y is trivial and ∀y ∈ Y, q(y) ≤ ^ pi(y), {iy∈Ei} then q ↾Y can be extended to a projection eq on X which is also trivial and satisfies, eq(x)≤ ^ {ix∈Ei} pi(x), ∀x ∈ X. The following probably is a well known result, but we could not find a direct reference and thus we include a short proof by applying Theorem 2.5. Corollary 2.7. Let X be a compact Hausdorff space of dimension d < ∞, and let Y ⊂ X be closed. Then any trivial projection r ∈ Mn(C(Y )) extends to a trivial projection on X, provided that rank (r) ≤ n − (cid:4) d 2(cid:5) . Proof. Suppose that the rank of r is k and take q to be any projection in Mn of rank k. Then, q represents a trivial projection in Mn(C(X)) and since r is trivial on Y, from Serre-Swan correspondence there exists s0 ∈ Mn(C(Y )) such that 0. Let p be the unit of Mn(C(X)). Then, as s0s* 0s0 = q ↾Y and r = s0s∗ s∗ 0 ≤ p ↾Y 2(cid:5), by Theorem 2.5 s0 extends to s ∈ Mn(C(X)) such and rank(q) ≤ rank(p) − (cid:4) d that, s∗s = q. Thus, r = ss∗ is a trivial projection on X with r ↾Y = s0s∗ 0 = r. (cid:3) Remark. In the hypothesis of 2.6 its assumed that q is defined on X. However, 2.7 tell us that the conclusion of 2.6 is valid even when q is defined only on Y . In section 3, we will use this observation without further mention. The final result of this section will not be required in proving Theorem 3.5 but we will need this to prove Lemma 4.2, which is essential to our proof of the main Theorem of section 4. For a ∈ Mn(C(X)), let spec(a) = {λ ∈ C : a − λ1n is not invertible}. For a ∈ Mn(C(X))+, let Γa : X → Rn be the map given by Γa(x) = (λ1(x), λ2(x)........, λn(x)), where λ1(x) ≤ λ2(x) ≤ ..... ≤ λn(x) are the eigenvalues of a(x). Proposition 2.8 is a straightforward consequence of the continuity of Γa on X. Proposition 2.8. Let a ∈ Mn(C(X))+ and η ≥ 0. The map x 7→ rank[χ(η,a](a(x))] where χ(η,a] denotes the characteristic map on (η, a], is lower semi-continuous. Proof. It is clear that Γa is continuous on X for any a ∈ Mn(C(X))+. Let x ∈ X and suppose rank[χ(η,a](a(x))] = m. Then, there are exactly m (with possible repetitions) eigenvalues of a(x), which are grater than η. Moreover, as λi(x) are in increasing order, λn−m+1(x), λn−m+2(x), ....λn(x) are exactly the eigenvalues of a(x) which are greater than η. Set ǫ = λn−m+1(x) − η > 0 and by continuity of Γa choose a neighborhood Ux of x such that Γa(x) − Γa(y) < ǫ, ∀y ∈ Ux. Then for all 1 ≤ i ≤ n, and therefore by the choice of ǫ, λi(x) − λi(y) < ǫ ∀y ∈ Ux, ∀n − m + 1 ≤ i ≤ n, λi(y) > η. RANK CONSTRAINED HOMOTOPIES Thus, for each y ∈ Ux, rank[χ(η,a](a(y))] ≥ m = rank[χ(η,a](a(x))]. 8 (cid:3) 3. Proof of the main result Our first aim is to prove Lemma 3.2, which is the main technical result. Next we prove Proposition 3.4 which extends [7, Chap. 8, Theorem 7.2] to compact Hausdorff spaces. Theorem 3.5 follows immediately by combining the two Lemmas. The following is a direct consequence of Theorem 2.3. Lemma 3.1. Let X be a compact Hausdorff space and a ∈ Mn(C(X))+. Then, there exists a continuous path t 7→ at in Mn(C(X))+ connecting a0 = a to a1, where a1 is well supported in the sense of Definition 2.2 and has the same rank values as that of a. The path is such that rank(at(x)) = rank(a(x)), ∀t ∈ (0, 1), ∀x ∈ X. Proof. Applying Theorem 2.3, choose a well supported positive element b ≤ a such that b has the same rank values as that of a, let at = (1 − t)a + tb. We only have to verify that rank (at(x)) = rank (a(x)) for every t ∈ (0, 1) and x ∈ X. But this is immediate. Since b ≤ a, if 0 < t < 1, (1 − t)a ≤ at ≤ (1 − t)a + ta = a. Therefore, rank(at(x)) = rank(a(x)), ∀t ∈ (0, 1), ∀x ∈ X. (cid:3) Lemma 3.2. Let X be a compact Hausdorff space with dim X < ∞. Suppose (cid:5) and let a ∈ C(X, S(n, k, l)). n, k, l ∈ N are such that k ≤ n and k − l ≥ (cid:4) dim X 2 Then, there is a continuous path h : [0, 1] → C(X, S(n, k, l)) such that h(0) = a and h(1) is a trivial projection of rank l. Proof. Let X, n, k, l and a be as given in the hypothesis. By Lemma 3.1 we can clearly assume that a is well supported. Let the rank values of a be n1 < n2 < ..... < nL and let E1, E2 . . . , EL and p1, p2, . . . , pL be as in Definition 2.2. For convenience we will write Fi = Ei and d = dim X. We first consider the case nL ≤ ⌊ d projection of rank l and let 2 ⌋. Then, choose p ∈ Mn(C(X)) to be any trivial Now for each t ∈ [0, 1], x ∈ X, h(t) = (1 − t)a + tp. rank [h(t)(x)] ≤ rank a(t)(x) + rank p ≤ nL + l ⌋ + l ≤ ⌊ d 2 ≤ k, and clearly rank [h(t)(x)] ≥ l. Thus, we get the required path. RANK CONSTRAINED HOMOTOPIES 9 Now let us assume nL > ⌊ d Fix r such that nr > ⌊ d set n0 = 0, F0 = ∅. 2 ⌋. 2 ⌋and nr−1 ≤ ⌊ d 2 ⌋, where we allow the possibility r = 1 and In what proceeds, we will construct a trivial projection R ∈ Mn(C(X)) of rank l such that, rank (R + a)(x) ≤ k, ∀x ∈ X. Once we have such R, we define h : [0, 1] → Mn(C(X)) by, h(t) = (1 − t)a + tR, ∀t ∈ [0, 1]. Then its immediate that this path satisfies the said rank constrains (i.e. remains in side C(S(n, k, l)). We focus on constructing R. To this end, we first define a trivial projection qL ∈ Mn(C( S Fj)) such that r≤j≤L and rank qL = nL − ⌊ d 2 ⌋ rank (a + qL)(x) ≤ nL, ∀x ∈ [ Fj. r≤j≤L We follow an inductive argument to define qL. Since Fr is compact Hausdorff with dim Fr ≤ d and pr ∈ Mn(C(Fr)) is a projection of rank nr > ⌊ d 2 ⌋, using Theorem 2 .1 (1) and Serre-Swan correspondence we find a trivial projection qr ∈ Mn(C(Fr)) such that, rank qr = nr − ⌊ d 2 ⌋ and qr ≤ pr. By the requirements for well supportedness of a, each pi ∈ Mn(C(Fi)) is of constant rank ni and whenever r ≤ i ≤ j with Fi ∩ Fj 6= ∅, (3.1) pi(x) ≤ pj(x), ∀x ∈ Fi ∩ Fj . Also for all j ≥ r, rank pj − rank qr ≥ nr − rank q1 ≥ ⌊ d 2 ⌋. Hence, we apply 2.6 with X = S 2.7, q in 2.6 need not be defined on X a priori) to extend qr to a trivial projection in Mn(C( S Fj))) - which is again called qr, such that whenever r ≤ j, Fj , Y = F1, q = qr (by the remark following r≤j≤L r≤j≤L Then, for each r ≤ j ≤ L, qr(x) ≤ pj(x), ∀x ∈ Fj. rank (qr + a)(x) ≤ nj, ∀x ∈ Fj. RANK CONSTRAINED HOMOTOPIES 10 If r = L then we are done (defining qL). Thus, let us assume r < L. Suppose that for some r ≤ t < L we have defined a trivial projection qt ∈ Mn(C( S Fj)) such that the following hold, r≤j≤L (3.2) (3.3) (3.4) rank qt = nt − ⌊ d 2 ⌋, qt(x) ≤ pj(x), ∀x ∈ Fj, ∀t ≤ j ≤ L, rank (qt + pj) ≤ nt, ∀r ≤ j ≤ t. Now whenever t + 1 ≤ j ≤ L, (pj − qt) ↾Ft+1 ∈ Mn(C(Fi+1)) is a projection constant rank (nj − nt) + ⌊ d 2 ⌋. Thus, since dim Ft+1 ≤ d, by applying 2.1 we choose a trivial projection qt,t+1 ∈ Mn(C(Ft+1)) such that, rank qt,t+1 = nt+1 − nt and qt,t+1 ≤ pt+1 − qt. Moreover, by applying 2.6 with X = S qt,t+1 to a trivial projection in Mn(C( S t+1≤j≤L Fj, Y = Ft+1 and q = qt,t+1 we extend Fj)) (which we again name qt,t+1) t+1≤j≤L such that whenever j ≥ t + 1, (3.5) qt,t+1(x) ≤ pj(x) − qt(x), ∀x ∈ Fj . Set qt+1 = qt + qt,t+1. Then, since qt, qt,t+1 are orthogonal trivial projections, qt+1 is a trivial projection in Mn(C( S Fj )). t+1≤j≤L Moreover, by 3.2, rank qt+1 = (nt − ⌊ d 2 ⌋) + (nt+1 − nt) = nt+1 − ⌊ d 2 ⌋ and whenever j ≥ t + 1, ∀x ∈ Fj (by 3.3 and 3.5 ), qt+1(x) = [qt(x) + qt,t+1(x)] ≤ pj(x) and finally for each r ≤ j ≤ t + 1 (by 3.4), rank (qt+1 + pj) ≤ rank qt,t+1 + rank (qt + pj) ≤ (nt+1 − nt) + nt = nt+1. RANK CONSTRAINED HOMOTOPIES By proceeding in this manner we construct a trivial projection qL ∈ Mn(C( S of rank nL − ⌊ d 2 ⌋ such that, r≤j≤L 11 Fj )) rank (qL + pj)(x) ≤ nL, ∀x ∈ Fj, ∀r ≤ j ≤ L and qL(x) ≤ pL(x), ∀x ∈ FL. Choose R1 ∈ Mn(C(X)) to be any trivial projection (of rank nL − ⌊ d extends qL. Note that such R1 exists by Corollary 2.7. 2 ⌋) which By the choice of r, whenever j < r, ∀x ∈ Fj, rank (R1 + a)(x) ≤ (nL − ⌊ d 2 ⌋) + ⌊ d 2 ⌋ ≤ nL. Thus, since R1 ↾Fj = qL ↾Fj whenever j ≥ r we conclude that , rank (R1 + a)(x) ≤ nL, ∀x ∈ X. If nL = k, then rank R1 = k − ⌊ d 2 ⌋ ≥ l and we choose R to be any trivial sub-projection of rank l . Hence we are left with the case k > nL. Then, rank (1n − R1) = n − (nL − ⌊ ⌋) d 2 d 2 ≥ (k − nL) + ⌊ ⌋ and we apply Theorem 2.1 (1) and Serre-Swan for one last time to choose a trivial projection R2 ∈ Mn(C(X)) of rank k − nL with R2 ≤ (1n − R1). Now R1 + R2 is a trivial projection of rank k − ⌊ d 2 ⌋ and rank (R1 + R2 + a)(x) ≤ k, ∀x ∈ X. To complete the proof we choose R to be a trivial sub-projection of R1 + R2 of rank l. (cid:3) Recall the following theorem for locally trivial vector bundles bundles over CW -complexes. Theorem 3.3. [7, Chp. 8, Theoreom 7.2] Let X be a CW-complex and n, l be non-negative integers. Then, the function that assigns to each homotopy class [f ] : X → Gl(Cn) the isomorphism class of the k-dimensional vector bundle f ∗(γn k ) over X is a bijection, if n ≥ l + (cid:6) dim X (cid:7). 2 RANK CONSTRAINED HOMOTOPIES 12 Combining 3.3 with Lemma 3.2 proves Theorem 3.5 for all CW -complex. The fact that Theorem 3.3 extends to the case of X being Compact Hausdorff is probably well known. However, we could not find a clear reference for this in the literature. Note that since Gl(C∞) (the l-dimensional Grassmannian over C∞) is the classifying space for l-dimensional vector bundles over paracompact spaces, if one replaces Gl(Cn) by Gl(C∞) the conclusion of 3.3 holds even for paracompact spaces. But the application we have in mind require the target space to be Gl(Cn). In Proposition 3.4, based on the proof of [6, Theorem 2.5] we apply dimension theory results [5, 9] and few C∗-algebraic techniques to extend the conclusion of 3.3 to compact Hausdorff spaces. If A, B are C∗-algebras and φ : A → B is a ∗-homomorphism, then for any n ∈ N we have an induced ∗-homomorphism from Mn(A) to Mn(B) given by [aij ] 7→ [φ(aij )]. We will use φ to denote this ∗-homomorphism as well. For two projections p, q ∈ Mn(C(X)), we write p ∼h q if there is a projection valued continuous path in Mn(C(X)) which connects p and q. Proposition 3.4. Let X be a compact Hausdorff space with dim X < ∞ and sup- (cid:7). Then the isomorphism classes of k-dimensional pose n, k ∈ N with n−k ≥ (cid:6) dim X locally trivial complex vector bundles over X are in bijective correspondence with the homotopy classes of maps p : X → Pk(Cn), where Pk(Cn) stands for rank k projections in Mn(C). 2 Proof. Let [X, Pk(Cn)] stand for the homotopy classes of maps in Pk(Mn(C(X))) and let V ectk(X) denote the set of all isomorphic classes of locally trivial k- dimensional vector bundles over X. Define ψ : [X, Pk(Cn)] → V ectk(X) by ψ([p]) = [ǫp], ∀p ∈ Pk(Mn(C(X))), where ǫp is defined as in section 2. Since the homotopy equivalence of projections implies Murray-von Neumann equivalence (see [8, Prop. 2.2.7]), map ψ is well defined by the discussion in section 2. Moreover, the discussion following Theorem 2.1 shows that ψ is surjective. To complete the proof of 3.4, we only have to show that if the vector bundles associated with two projections in Pk(Mn(C(X))) are isomorphic then the two projections are homotopic in Pk(Mn(C(X))). Let us first assume that X is a compact metric space. Then by [9, Chap. 27, Theorem 8.] or [5, Chap. X, Sec. 10], X is homeomorphic to an inverse limit of finite simplicial complexes Xα, with dim Xα ≤ dim X for each α. Let ψα : X → Xα be the canonical induced maps. We have the induced homomor- phisms, ψT α : C(Xα) → C(X) given by ψT α (f ) = f ◦ ψα. Moreover, by the inverse limit structure (3.6) [ α ψT α (C(Xα)) = C(X), RANK CONSTRAINED HOMOTOPIES 13 i.e. S α ψT α (C(Xα)) is a sup dense ∗-subalgebra of C(X). Note that if α < β then, (3.7) ψT α (C(Xα)) ⊂ ψT β (C(Xβ)) α ψT α (C(Xα)) is indeed a ∗-subalgebra. and hence S Suppose p, q ∈ Mn(C(X)) are projections of constant rank k such that the corre- sponding vector bundles are isomorphic. This means, p = v∗v, q = vv∗ for some v ∈ Mn(C(X)). Fix 0 < ǫ < 1/3. By (3.6), there is some α and vα ∈ Mn(C(Xα)))1 such that, ψT α (vα) − v < ǫ/2. Write Yα = ψα(X) and vα = vα ↾Yα, aα = v∗ Now aα ∈ Mn(C(Yα)) and moreover by the choice of vα it follows that, αvα, bα = vαv∗ α. Similarly, spec(aα) ⊂ [0, ǫ) ∪ (1 − ǫ, 1]. spec(b) ⊂ [0, ǫ) ∪ (1 − ǫ, 1]. Let f : [0, 1] → [0, 1] be the continuous function which vanishes on [0, ǫ], is equal to 1 on [1 − ǫ, 1] and is linear on (ǫ, 1 − ǫ). Then f (aα), f (bα) are projections in Mn(C(Yα)) with aα − f (aα) < ǫ, bα − f (bα) < ǫ. Furthermore, as a, b are Murray-von Neumann equivalent, by Lemma 3.3 of [4] there is some s0 ∈ Mn(C(Yα)) such that, f (aα) = s∗ 0s0, f (bα) = s0s∗ 0. Since Yα is a closed in Xα, we may choose some open neighborhood U of Yα so that f (aα), f (bα) extends to projections in Mn(C(U )). Let pα, qα denote these exten- sions respectively. Moreover, since f (aα) ∼ f (bα), we may choose the extensions in such a way that pα ∼ qα. As Xα is a finite simplicial complex, after a finite simplicial refinement of Xα via barycentric subdivisions, choose a sub complex Z of Xα with Y ⊂ Z ⊂ U . For convenience let us denote the restrictions of pα, qα to Z by pα, qα. Note that pα, qα generate isomorphic vector bundles over Z, each of rank k. Hence, form Theorem 3.3 and the identification of Gk(Cn) with Pk(Cn), there is a contin- uous path, such that hα(0) = pα, h(1) = qα. t 7→ hα(t) ∈ Pk(Mn(C(Z))), This gives a path t 7→ h(t) ∈ Pk(Mn(C(X))), given by h(t)(x) = hα(t)(ψα(x)). Note that for all x ∈ X, p(x) − h(0)(x) = p(x) − f (aα)(ψα(x)) ≤ v∗v(x) − (v∗ αvα)(ψα(x)) + (v∗ ≤ 2ǫ + aα(ψα(x)) − f (aα)(ψα(x)) ≤ 2ǫ + ǫ αvα)(ψα(x)) − f (aα)(ψα(x)) RANK CONSTRAINED HOMOTOPIES 14 Thus, p − h(0) < 1 and similarly q − h(1) < 1. Therefore, from [8, Proposition 2.2.4] This completes the proof for compact metric spaces. p ∼h h(0) ∼h h(1) ∼h q. The proof of the Proposition in the case of X being an arbitrary compact Hausdorff space follows almost identically. In this case X is the inverse limit of compact metric spaces Xλ, with dim Xλ ≤ dim X for each λ. By the preceding the result holds for each Xλ, and now we can argue as in the preceding case. (cid:3) Theorem 3.5. Let X be a compact Hausdorff space X with (cid:4) dim X (cid:5) ≤ k − l. There is only one homotopy class of functions f : X → S(n, k, l), i.e. the function space C(X, S(n, k, l)) is path connected. 2 Proof. Observe that the case n = k is straightforward. For any a ∈ C(X, S(n, k, l)) we have the linear path t 7→ (1 − t)a + 1n connecting a to 1n. So we assume n > k. a to p and b to q. Since n > k and k−l ≥ (cid:4) dim X Let a, b ∈ C(X, S(n, k, l)). As dim X ≤ k − l, by applying Lemma 3.2 choose trivial projections p, q of rank l such that there are paths inside C(X, S(n, k, l)) connecting ⌉. Therefore, as p, q are both trivial, by Proposition 3.4 there is a path inside C(X, S(n, k, l)) connecting p and q. Thus, there is a path between a and b in C(X, S(n, k, l)). (cid:3) (cid:5), we get n−l ≥ ⌈ dim X 2 2 Corollary 3.6. For every r ≤ 2(k − l) + 1, πr(S(n, k, l)) = 0. Proof. Follows directly from Theorem 3.5 as dim Sr = r. (cid:3) Even though we derived the above result as a Corollary to Theorem 3.5, from a topological view point it might seem more natural (and technically easier) to first prove the conclusion of 3.6 independently and then apply techniques from homotopy theory and dimension theory to prove Theorem 3.5. This is indeed possible and applies in a slightly wider scope. In section 4, we show that for compact Hausdorff spaces of finite covering dimension the path connectedness of C(X, S(n, k, l)) depends solely on homotopy groups πr(S(n, k, l)), r ≤ dim X. We chose not to follow this method for the proof of 3.5, due to two reasons. Firstly, with the techniques we used, the proof of Lemma 3.2 would not be any simpler if we assumed X to be a sphere instead of a general compact Hausdorff spaces. Secondly, if we used such an argument (i.e proving 3.6 independently and then showing 3.5) it would have avoided the use of Proposition 3.4, but we thought 3.4 could be of independent interest for some readers. 4. Homotopy groups of S(n, k, l) and Path Connectedness of C(X, S(n, k, l)). In Theorem 4.6, combining well known homotopy theory techniques and classical C∗-algebraic results we prove that for a fixed integer d if πr(S(n, k, l)) = 0, ∀r ≤ d, then C(X, S(n, k, l)) is path connected for every compact Hausdorff space X with dim X ≤ d. Note that in this section we do not assume k − l ≥ (cid:4) dim X (cid:5) . 2 RANK CONSTRAINED HOMOTOPIES 15 The proof of Lemma 4.5 is of the same flavor as that of Proposition 3.4. We need few more technical results. The following is well known and was used in the proof of 3.4 as well. We state it here for convenience. Proposition 4.1. Let X be a finite simplicial complex. Let Y ⊂ X be closed and U be a open neighborhood of Y in X. Then, after a finite simplicial refinement of X, there is a sub-complex Z of X such that Y ⊂ Z ⊂ U . Lemma 4.2 is a simpler version of [13, Lemma 2.1]. For the sake of completeness, we provide a proof. Lemma 4.2. [13, Lemma 2.1] Let X be a compact Hausdorff space and suppose that a ∈ Mn(C(X))+. Let l ∈ N be such that rank(a(x)) ≥ l, ∀x ∈ X. Then, there is some η > 0 such that for each x ∈ X, the spectral projection χ(η,∞)(a(x)) has rank at least l. Proof. For each x ∈ X, let ηx = 1 ηx exists for each x ∈ X. Then, 2 min{λ ∈ spec a(x):λ > 0}. Note that since l > 0, rank(a(x)) = rank[χ(ηx,∞)(a(x))], ∀x ∈ X. By Proposition 2.8, for each x ∈ X, the map y 7→ rank[χ(ηx,∞)(a(y))] is lower semi-continuous. So, for each x ∈ X, there exists an open neighborhood Ux of x such that, rank[χ(ηx,∞)(a(y))] ≥ rank[χ(ηx,∞)(a(x))], ∀y ∈ Ux. By compactness of X, choose some finite set of points {x1, x2, ...xL} such that X = S Uxi. 1≤i≤L By setting η = min1≤i≤L(ηxi) > 0, for every y ∈ Uxi we get, rank[χ(η,∞)(a(y))] ≥ rank[χ(ηxi ,∞)(a(y))] ≥ rank[χ(ηxi ,∞)(a(x))] = rank(a(x)) ≥ l. Since X = S 1≤i≤L Uxi, this completes the proof. (cid:3) α (C(Xα, S(n, k, l))). Here, ψT Let X = lim←− Xα, where (Xα, ψαβ) is a inverse system of compact Hausdorff spaces and ψα : X → Xα be the natural maps. In the proof of Lemma 4.5, given a ∈ C(X, S(n, k, l)) ⊂ Mn(C(X))+ we need to approximate a in norm, by elements in ψT α is defined as in the proof of Proposition 3.4. To achieve this, we will use the previous Lemma with two other results from C∗- algebra theory. First we prove Lemma 4.3, which follows as a special case of a well known fact in C∗-algebras. For the sake of completeness we include the proof for the special case. RANK CONSTRAINED HOMOTOPIES 16 Lemma 4.3. Let X = lim←− Xα, for a inverse system of compact Hausdorff spaces (Xα, ψαβ). Let ψα : X → Xα be the natural maps and ǫ > 0. Given a ∈ Mn(C(X))+, there is some index α and some b ∈ Mn(C(Xα))+ such that, where ψT α (b) = b ◦ ψα. a − ψT α (b) < ǫ Proof. We may assume a = 1 and ǫ < 1. Since a is positive, by the functional calculus of a there is some c ∈ Mn(C(X))+ such that c2 = a. As in the proof of 3.4, pick some α and d ∈ Mn(C(Xα)) such that, ǫ 3 α (d) < c − ψT (4.1) . Since c is positive, c∗ = c where c∗ is the conjugate transpose of c. Then, (4.2) c − ψT α (d∗)) = (c∗ − ψT = (c − ψT = c − ψT α (d∗)) α (d))∗ α (d) < ǫ 3 Put b = d∗d then b ∈ Mn(C(Xα))+ and moreover, a − ψT α (d∗d) α (d∗)ψT α (b)) = c2 − ψT = c2 − ψT ≤ c2 − c · ψT ≤ c · c − ψT < ǫ α (d) α (d) + c · ψT α (d) + c − ψT α (d) − ψT α (d∗)ψT α (d) α (d∗) · ψT α (d) The last inequality follows from (4.1), (4.2) and the fact that ψT α (d) ≤ 2. (cid:3) For ǫ ≥ 0, let fǫ : [0, 1] → [0, 1] be defined by, fǫ(t) = max{ǫ, t}, ∀t ∈ [0, 1]. For any a ∈ Mn(C(X))+, let (a−ǫ)+ denote the element fǫ(a) ∈ Mn(C(X))+ given by the functional calculus of a. Recall that by support projection of d ∈ Mn(C(X)) we mean the function (not necessarily continuous) which maps x to the orthogonal projection of Cn onto d(x)(Cn). Note that the support projection of fǫ(a) is the spectral projection χ(ǫ,1](a). To prove 4.5, we need the following proposition. Proposition 4.4. [11, Proposition 2.2] Let X be compact Hausdorff and a, b ∈ Mn(C(X))+. Let ǫ > 0 and suppose that b − a < ǫ. Then there exists c ∈ Mn(C(X)) such that, (a − ǫ)+ = c∗bc. We now prove Lemma 4.5. Lemma 4.5. Suppose for each finite simplicial complex Z with dim Z ≤ d, the function space C(Z, S(n, k, l)) is path connected. Then, for every compact Hausdorff space X of covering dimension d, space C(X, S(n, k, l)) is path connected. RANK CONSTRAINED HOMOTOPIES 17 Proof. Like in the proof of 3.4, we first prove the result for the case of X being a compact metric space. In this case X = lim←− Xα, where (Xα, ψαβ) is a inverse system finite simplicial complexes with dim Xα ≤ d. Let ψα : X → Xα be the natural maps. Fixed a ∈ C(X, S(n, k, l)), our first goal is to show that there is some index α and c ∈ C(Xα, (Mn)+) such that ψT α (c) ∈ C(X, S(n, k, l)) and there is a path in C(X, S(n, k, l)) connecting a to ψT α (c). To do this, first note that we may assume a = 1 and use Lemma 4.2 to pick η > 0 such that, (4.3) rank [χ(2η,1](a(x))] ≥ l, ∀x ∈ X. By Lemma 4.3, pick α and b ∈ Mn(C(Xα))+ such that, (4.4) a − ψT α (b) < η. Let us write aα = ψT α (b). Note that from Proposition 4.4 and (4.4), (aα − η)+ = d∗ad, for some d ∈ Mn(C(X)). Therefore, for every x ∈ X, (4.5) ≤ k rank [(aα − η)+(x)] ≤ rank (a(x)) From the functional calculus for aα and 4.4, (aα − η)+ − a ≤ (aα − η)+ − aα + aα − a < η + η = 2η. Therefore, by Proposition 4.4 it follows that, rank [(a − 2η)+(x)] ≤ rank [(aα − η)+(x)], ∀x ∈ X. Now, from (4.3) and the discussion preceding Proposition 4.4, (4.6) rank [(aα − η)+(x)] ≥ l, ∀x ∈ X. Put c = ψT α ((b − η)+). Then, c = (b − η)+ ◦ ψα = ((b ◦ ψα) − η)+ = (aα − η)+. Thus by (4.5) and (4.6), c ∈ C(X, S(n, k, l)). To complete our first goal consider h : [0, 1] → C(X, S(n, k, l)) given by, We have h(0) = (a − η)+ and h(1) = c. h(t) = [((1 − t)a + taα) − η]+. RANK CONSTRAINED HOMOTOPIES 18 Note that, a − ((1 − t)a + taα) = ta − aα < η, ∀t ∈ [0, 1]. Thus, by following the same type of argument we used to show c ∈ C(X, S(n, k, l)), it follows that h(t) = [((1 − t)a + taα) − η]+ ∈ C(X, S(n, k, l)), ∀t ∈ [0, 1]. The continuity of h can be established using functional calculus arguments, see [8, Lemma 1.25]. First goal is now achieved once when we notice that a is homotopic to (a − η)+ as maps in C(X, S(n, k, l)). But this is immediate. Indeed, one can use the linear path t 7→ (1 − t)a + t(a − η)+. Now suppose a, b ∈ C(X, S(n, k, l)). We need to construct a path in C(X, S(n, k, l)) joining a to b. From the first part w.l.o.g. we may assume that a = ψT α (c), b = ψT β (d), for some α = β and c, d ∈ Mn(C(Xα))+. Put Y = ψα(X) ⊂ Xα. Then as X is compact so is Y . Moreover, Y is closed as each Xα is Hausdorff. For each y = ψα(x), Hence, Similarly, rank (c(y)) = rank (c(ψα(x))) = rank (a(x)) l ≤ rank (c(y)) ≤ k, ∀y ∈ Y. l ≤ rank (d(y)) ≤ k, ∀y ∈ Y. Therefore as Y is closed, by Lemma 2.7 of [13] there is some open neighborhood U of Y in Xα and c, d ∈ C(U, S(n, k, l)) such that c, d are extensions of c, d respectively. By Proposition 4.1, after a refinement of the simplicial structure of Xα, we have a finite sub complex Z of Xα such that, Y ⊂ Z ⊂ U . Hence, we may view c, d ∈ C(Z, S(n, k, l)). Now, as Z is a finite simplicial complex with dim Z ≤ d, by the hypothesis there is a path g : [0, 1] → C(Z, S(n, k, l)) such that g(0) = c and g(1) = d. Define a path g : [0, 1] → C(X, Mn) by, g(t)(x) = g(t)(ψα(x)). By definition of g it is clear that g(0) = ψT g(t) ∈ C(X, S(n, k, l)). α (c) = a , g(1) = ψT α (d) = b and moreover, This proves the result in the case of X being a compact metric space such that dim X ≤ d. To get the result for X being compact Hausdorff, write X = lim←− Xα where now Xα is compact metric with dim Xα ≤ d. Since C(Xα, S(n, k, l)) is path connected for each α by the first part of the proof, following a similar argument as before give the result. (Note here that the argument is simpler than in the first step. Since ψα(X) ⊂ Xα is compact metric for any α, we do not require Lemma 2.7 of [13]. (cid:3) RANK CONSTRAINED HOMOTOPIES 19 Theorem 4.6. Let X be compact Hausdorff with dim X ≤ d. If πr(S(n, k, l)) = 0 for each r ≤ d then, C(X, S(n, k, l)) is path connected. Proof. From Lemma 4.5, it suffices to show that C(K, S(n, k, l)) is path connected for every finite simplicial complex K with dim K ≤ d. We will use induction on the number of simplexes in the complex K to show this. If K consists of a single simplex then result is true since K is contractible. Suppose now that result is true for every simplicial complex which contains r num- ber of simplexes. To complete the inductive step, let K = L ∪ {s} where L is a sub complex of K containing r number of simplexes and s is a n-simplex for some n ≤ d. Observe that {K, L} is a NDR pair in the sense of [16]. Since S(n, k, l) is locally compact, S(n, k, l) is compactly generated. By [16, Diagram 6.3], following sequence is exact in the category of sets with base points, (4.7) [C(L, S(n, k, l))] ←− [ C(K, S(n, k, l))] ←− [ C((K/L), S(n, k, l))], i∗ p∗ where K/L stands for the quotient space. For a space Z, by [C(Z, S(n, k, l))] we mean the set of all homotopy classes of maps in C(Z, S(n, k, l)) and as the base point we may choose any constant map z 7→ a, for a fixed a ∈ S(n, k, l). The maps i∗ and p∗ are the maps induced by the inclusion i : L → K and the quotient map p : K → K/L. By the induction hypothesis [C(L, S(n, k, l))] consists of a single point. Since (K/L) ∼= Sn and n ≤ d, by assumption [C((K/L), S(n, k, l))] is also a single- ton. Thus, by exactness of (4.7), [C(K, S(n, k, l))] contains only one point, i.e. C(K, S(n, k, l)) is path connected. (cid:3) References 1. M. F. Atiyah (notes by D. W. Anderson) K-Theory, W.A. Benjamin, Inc. New york, New York 1967. 2. N.P. Brown, F. Perera, A.S. Toms, The Cuntz semigroup, the Elliott conjecture, and dimen- sion functions on C ∗ -algebras, J. Reine Angew. Math 621, (2008), 191-211 3. N.P. Brown, A.S. Toms, There applications of the Cuntz semigroup, Int. Math Res. Not., (2007), doi: 10.1093/imrnirnm068 4. M. Dadarlat, A.S. Toms, Ranks of operators in simple C ∗-algebras, J.Funct.Anal.259, (2010), 1209-1229. 5. S. Eilenberg and N. Steenrod, Foundations of Algebraic Topology, Princeton University Press, Princeton, New Jersey, 1952. 6. K. R. Goodearl, Riesz decomposition in inductive limit C ∗-algebras, Rocky Mountain J. Math. 24, (1994), 1405 -- 1430. 7. D. Husemoller, Fiber Bundles (Third Ed.), Springer-Verlag, New York, Inc., 1994. 8. F. Larsen, N. J. Lausten and M. Rørdam, An introduction to K-Theory for -C ∗-algebras, Cambridge University press, Cambridge, 2000. 9. K. Nagami, Dimension Theory, Academic Press, New York, 1970. 10. N. C. Phillips, Recursive subhomogeneous algebras, Trans. Am. Math. Soc. 359 (2007), no. 10, 4595-4623. 11. M. Rørdam, On the structure of simple C ∗-algebras tensored with a UHF-algebra, II, J. Funct. Anal.107, (1992), 255-269 12. R. G. Swan, Vector bundles and projective modules, Trans. Amer. Math. Soc. 105 (1962), 264-277. RANK CONSTRAINED HOMOTOPIES 20 13. A. Toms, K-theoretic rigidity and slow dimension growth, Invent. Math. 183 (2011), no. 2, 225 -- 244. 14. A. S. Toms, Stability in the Cuntz semigroup of a commutative C ∗-algebra, Proc. Lond. Math. Soc. (3) 96 (2008), no. 1, 1 -- 25. 15. A. S. Toms, Comparison theory and smooth minimal C ∗-dynamics, Comm. Math. Phys 96 (2009), no. 2, 401 -- 433. 16. G. W. Whitehead, Elements of Homotopy Theory, Springer-Verlag, New York, 1978. Department of Mathematics, Purdue University, 150 N. University St., West Lafayette IN, USA, 47907 E-mail address: [email protected]
1505.04784	1	1505	2015-05-17T13:38:00	$C^*$-index of observable algebra in the field algebra determined by a normal group	[ "math.OA" ]	Let $G$ be a finite group and $H$ a normal subgroup. $D(H;G)$ is the crossed product of $C(H)$ and ${\Bbb C}G$ which is only a subalgebra of $D(G)$, the quantum double of $G$. One can construct a $C^$-subalgebra ${\mathcal{F}}_{_H}$ of the field algebra $\mathcal{F}$ of $G$-spin models, such that ${\mathcal{F}}_{_H}$ is a $D(H;G)$-module algebra. The concrete construction of $D(H;G)$-invariant subalgebra ${\mathcal{A}}_{_{(H,G)}}$ of ${\mathcal{F}}_{_H}$ is given. By constructing the quasi-basis of conditional expectation $z_{_H}$ of ${\mathcal{F}}_{_H}$ onto ${\mathcal{A}}_{_{(H,G)}}$, the $C^$-index of $z_{_H}$ is given.	math.OA	math	C ∗-index of observable algebra in the field algebra determined by a normal group ∗ Xin Qiaoling, Jiang Lining † School of Mathematics and Statistics, Beijing Institute of Technology, Beijing 100081, P. R. China Abstract: Let G be a finite group and H a normal subgroup. D(H; G) is the crossed product of C(H) and CG which is only a subalgebra of D(G), the quantum double of G. One can construct a C ∗-subalgebra FH of the field algebra F of G-spin models, such that FH is a D(H; G)-module algebra. The concrete construction of D(H; G)-invariant subalgebra A(H,G) of FH is given. By constructing the quasi-basis of conditional expectation zH of FH onto A(H,G) , the C ∗-index of zH is given. Keywords: quantum double, C ∗-index, quasi-basis, conditional expectation Mathematics Subject Classification (2010): 46L05, 16S35 5 1 0 2 y a M 7 1 ] . A O h t a m [ 1 v 4 8 7 4 0 . 5 0 5 1 : v i X r a 1 Introduction Assume that G is a finite group with a unit e. The G-valued spin configuration on the two-dimensional square lattices is the map σ : Z2 → G with Euclidean action functional: S(σ) = X(x,y) f (σ−1 x σy), in which the summation runs over the nearest neighbour pairs in Z2 and f : G → R is a function of the positive type. This kind of classical statistical systems or the corresponding quantum field theories are called G-spin models, see [1, 2, 3]. Such models provide the simplest examples of lattice field theories exhibiting quantum symmetry. In general, G-spin models with an Abelian group G are known to have a symmetry group G × bG, where bG is the group of characters of G, namely the Pontryagin dual of G. If G is non-Abelian, the Pontryagin dual loses its meaning, and the models have a symmetry of a quantum double D(G) ([4, 5]). Here D(G) is defined as the crossed product of C(G), the algebra of complex functions on G, and group algebra CG with respect to the adjoint action of the latter on the former. Then D(G) is a Hopf -algebra of finite dimension ([6, 7, 8]). Also as in ∗This work is supported by National Science Foundation of China (10971011,11371222) †E-mail address: [email protected] 1 the traditional quantum field theory, one can define a field algebra F associated with this model ([9]). There is a natural action γ of D(G) on F such that F is a D(G)-module algebra with respect to the map γ. Namely, there is a bilinear map γ : D(G) × F → F satisfying: ∀ a, b ∈ D(G), F1, F2, F ∈ F, (ab)(F ) = a(b(F )), a(F1F2) = P(a) a(F ∗) = (S(a∗)(F ))∗. a(1)(F1)a(2)(F2), Here and from now on, by a(F ) we always denote γ(a × F ) in F. Under the action of γ on F, the observable algebra A(G,G) as the D(G)-invariant subalgebra of F is obtained. And there exists a duality between AG and D(G), i.e., there is a unique C ∗-representation of D(G) such that D(G) and A(G,G) are commutants with each other. In [10], we consider a more general situation. Let H be a normal subgroup of G, then D(H; G) is defined as the crossed product of C(H) and CG with respect to the adjoint action of the latter on the former. One can construct a C ∗-subalgebra FH of the field algebra F of G-spin models, called the field algebra of G-spin models determined by H, such that FH is a D(H; G)-module algebra even though D(H; G) is not a Hopf subalgebra of D(G). Then the observable algebra A(H,G), which is the set of fixed points of FH under the action of D(H; G) is given. Also there exists a duality between D(H; G) and A(H,G). In this paper, we continue to discuss this model. In Section 2, we find algebraic generators for A(H,G) by means of discussing the local net structure to A(H,G). In Section 3, we construct a quasi- basis for the conditional expectation zH : FH → A(H,G), and then obtain the corresponding C ∗-index Index zH = GHI, where G and H denote the order of the group G and H, respectively. Throughout this paper, all algebras are complex unital associative algebras. For more details on Hopf algebras one can refer to the books of Sweedler [11] and Abe [12]. We shall adopt its notation, such as S, △, ε for the antipode, the comultiplication and the counit, respectively. Also we shall use the summation convention, which is standard in Hopf algebra theory: a(1) ⊗ a(2), △(a) = P(a) △(2)(a) = △ ◦ (id ⊗ △)(a) =P(a) △(n)(a) = △(n−1) ◦ (id ⊗ △)(a) =P(a) a(1) ⊗ a(2) ⊗ a(3), a(1) ⊗ a(2) ⊗ · · · ⊗ a(n+1), where the second one holds since △ ◦ (id ⊗ △) = △ ◦ (△ ⊗ id), and so on. 2 The structure of the observable algebra in FH Suppose that H is a normal subgroup of G. In the previous paper [10], we defined a Hopf -algebra D(H; G) and then constructed a C ∗-subalgebra FH in the field algebra F of G-spin models. Under the 2 action γ of D(H; G) on it, FH becomes a D(H; G)-module algebra and the observable algebra A(H;G) as the D(H; G)-invariant subalgebra of FH is given. This section will discuss a local net structure to A(H;G), which can be achieved by finding algebraic generators for A(H;G) with local commutation relations. Let us begin with the following definition. Definition 2.1. [10] D(H; G) is the crossed product of C(H) and group algebra CG, where C(H) denotes the set of complex functions on H, with respect to the adjoint action of the latter on the former. Using the linear basis elements (h, g) of D(H; G), the multiplication can be written as: (h1, g1)(h2, g2) = δh1g1,g1h2(h1, g1g2). (h, e) is the unit of D(H; G). Also, the structure maps are defined as Clearly, Ph∈H (h, g)∗ = (g−1hg, g−1), (t, g) ⊗ (t−1h, g), △(h, g) = Pt∈H ε(h, g) = δh,e, S(h, g) = (g−1h−1g, g−1), (-operation) (coproduct) (counit) (antipode) where h ∈ H, g ∈ G and δg,h =( 1, 0, if g = h if g 6= h unique element . One can prove D(H; G) is a Hopf -algebra with a called a cointegral, satisfying zH = 1 G Pg∈G (e, g), azH = zH a = ε(a)zH , ∀a ∈ D(H; G), and ε(zH ) = 1. As a result, D(H; G) is a semisimple -algebra of finite dimension. Consequently it can be a C ∗-algebra in a natural way ([10]). Remark 2.1. (1) If H is a subgroup of G, not a normal subgroup. One can prove there is not the adjoint action of CH on C(G), and then D(H; G) can not be defined. (2) Different from the case of D(G; H), which is the crossed product of C(G) and CH with respect to the adjoint action of the latter on the former ([9, 13]), D(H; G) is not a Hopf subalgebra of D(G), even though it is a subalgebra of D(G). (3) Also, the relation S2 = id holds in D(H; G), which implies that ∀a ∈ D(H; G), P(a) S(a(2))a(1) =P(a) a(2)S(a(1)) = ε(a)1D(H;G). 3 As in the traditional case, one can define the local quantum field algebra associated with the model. Definition 2.2. [10] FH,loc is an associative algebra with a unit I generated by {δg(x), ρh(l) : g ∈ G, h ∈ H; x ∈ Z, l ∈ Z + 1 2 } subject to δg(x) = I = ρe(l), Pg∈G δg1(x)δg2 (x) = δg1,g2δg1(x), ρh1(l)ρh2(l) = ρh1h2(l), δg1(x)δg2(x′) = δg2(x′)δg1 (x), ρh(l)δg(x) = ( δhg(x)ρh(l), δg(x)ρh(l), l < x, l > x, ρh1(l)ρh2(l′) = ρh2(l′)ρh2 −1h1h2(l), l > l′. for x, x′ ∈ Z; l, l′ ∈ Z + 1 2 and h1, h2 ∈ H, g1, g2 ∈ G. In particular, if H = G, by Floc we denote FH,loc. The -operation is defined on the generators as δ∗ g (x) = δg(x), ρ∗ h(l) = ρh−1(l) and is extended antilinearly and antimultiplicatively to FH,loc. In this way, FH,loc becomes a unital *-algebra. For any finite subset Λ ⊆ 1 2 Z, let FH (Λ) be the subalgebra of FH,loc generated by {δg(x), ρh(l) : g ∈ G, h ∈ H, x, l ∈ Λ}. In particular, we consider an increasing sequence of intervals Λn, n ∈ N, where Λ2n = {s ∈ 1 Λ2n+1 = {s ∈ 1 2 Z : −n + 1 2 Z : −n − 1 2 ≤ s ≤ n} 2 ≤ s ≤ n} In [3], the authors have shown that F(Λn), n ∈ N are full matrix algebras, they can be identified with MGn. Moreover, under the induced norm, F(Λn) are finite dimensional C ∗-algebras. Hence FH (Λn), n ∈ N are subalgebras of full matrix algebras, and then they are finite dimensional C ∗- algebras. The natural embeddings ιn : FH (Λn) → FH (Λn+1), that identify the δ and ρ generators, are norm preserving. Using the C ∗-inductive limit ([14]), a C ∗-algebra FH can be given by called the field algebra of G-spin models determined by a normal subgroup H. FH =Sn FH (Λn), There is an action γ of D(H; G) on FH in the following. For x ∈ Z; l ∈ Z + 1 2 and h ∈ H, g ∈ G, set (h, g)δf (x) = δh,eδgf (x), ∀f ∈ G, (h, g)ρt(l) = δh,gtg−1 ρh(l), ∀t ∈ H. 4 Then the γ can be extended continuously to an action of D(H; G) on FH , such that FH is a D(H; G)- module algebra with respect to the γ. Namely, the γ satisfies the following relations: (ab)(F ) = a(b(F )), a(F1F2) = P(a) a(1)(F1)a(2)(F2), a(F ∗) = (S(a∗)(F ))∗ for a, b ∈ D(H; G), F1, F2, F ∈ FH . Set A(H,G) = {F ∈ FH : zH (F ) = F }. Then A(H,G) is a subalgebra of FH . Lemma 2.1. [10] zH : FH → A(H,G) satisfies the following conditions: (1) zH (I) = I where I is the unit of FH ; (2) (bimodular property) ∀ F1, F2 ∈ A(H,G), F ∈ FH , zH (F1F F2) = F1zH (F )F2; (3) zH is positive. In the following a linear map Γ from a unital C ∗-algebra B onto its unital C ∗-subalgebra A with properties (1)-(3) in Lemma 2.1 is called a conditional expectation. If Γ is a conditional expectation from B onto A, then Γ is a projection of norm one ([14]). The conditional expectation zH : FH → A(H,G) will be addressed in the next section. From the above lemma, one can prove A(H,G) is the C ∗-subalgebra of FH , called an observable algebra related to H in the field algebra F of G-spin models. Moreover, if H1 ( H2 with Hi ⊳ G for i = 1, 2, then A(H1,G) ( A(H2,G), since △(n) (zH2 ) as projections on FH1 (zH1) ≤ △(n) H2 for n ∈ N. ⊗n+1 H1 In this section, we will give the concrete construction of A(H,G). In order to do this, for g ∈ G, x ∈ Z, and l ∈ Z + 1 2 , set vg(x) = Ph∈G wg(l) = Ph∈G hg−1h−1(x − 1 2 )δh(x)hgh−1(x + 1 2 ), δh(l − 1 2 )δhg(l + 1 2 ). Lemma 2.2. Let Λn− 1 of FH (Λn− 1 2 ,m ⊆ 1 2 ,m) is generated by 2 Z be a finite interval for n, m ∈ Z. The D(H; G)-invariant subalgebra That is 2o . nωg(x), vh(l) : g ∈ G, h ∈ H, x, l ∈ Λn,m− 1 2o . 2 ,m)(cid:17) =nωg(x), vh(l) : g ∈ G, h ∈ H, x, l ∈ Λn,m− 1 zH(cid:16)FH (Λn− 1 5 Proof. We know that FH (Λ 1 2 ,m) is a C ∗-subalgebra of FH,loc, generated by i = 1, 2, · · · m with h1h2 · · · hm = e, nδg(x), h(l) : g ∈ G, h ∈ H, x, l ∈ Λ 1 2 )(cid:1) 2 ) · · · hm(m − 1 2 )h2( 3 2 )h2( 3 2 ,mo . 2 ) · · · hm(m − 1 2 , f )δg2(2) · · · (tm−1t−1 m , f )δgm(m) Notice that for hi ∈ H, = 1 = 1 = 1 = 1 (tmt−1 (f, e)(cid:0)δg1(1)δg2 (2) · · · δgm(m)h1( 1 zH(cid:0)δg1(1)δg2 (2) · · · δgm(m)h1( 1 G Pf ∈G G Pf ∈G Pti∈H(cid:16)(t1, f )δg1(1)(t1t−1 G Pf ∈G(cid:16)δf g1(1)δf g2 (2) · · · δf gm(m)f h1f −1( 1 G Ps∈G(cid:16)δs(1)δsg−1 (2) · · · δsg−1 1 g2 1 g1s−1(m − 1 m+1, f )h1( 1 m+2, f )h2( 3 2 )(tm+1t−1 (m)sg−1 m−2···h−1 m−1h−1 sg−1 1 h−1 1 gm 2 )(cid:17), which together with the following equation 2 )(cid:1) 2 )(cid:17) 2 ) · · · (t−1 2m−1, f )hm(m − 1 2 )f h2f −1( 3 2 ) · · · f h−1 m−1h−1 m−2···h−1 1 f −1(m − 1 2 )(cid:17) 1 h1g1s−1( 1 2 )sg−1 1 h2g1s−1( 3 2 ) · · · wx1( 3 2 )wx2( 5 2 ) · · · wxm−1(m − 1 2 )vy1(1)vy2 (2) · · · vym−1 (m − 1) 2 2 t−1 ρt2y−1 = Psi∈G Pti∈G = Psi∈G Pti∈G = Ps∈G δtm−1y−1 ρtm−2y−1 δs1(1)δs1x1(2)δs2 (2)δs2x2(3) · · · δsm−1(m − 1)δsm−1xm−1(m)ρt1y−1 ( 3 2 )δt2 (2)ρt2y2t−1 δs1(1)δt1 y−1 2 )δtm−1 (m − 1)ρtm−1ym−1t−1 (1) · · · δsm−2xm−2(m − 1)δsm−1 (m − 1) ( 5 2 ) · · · ρtm−1y−1 (m − 3 m−1t−1 1 t−1 m−1 2 1 m−1 1 m−1 (m − 1)δsm−1xm−1(m)ρt1y−1 m−2y−1 (m − 3 m−1t−1 m−1 m−2t−1 2 1 t−1 (1)δs1x1(2)δs2 (2)δt2 y−1 ( 1 2 )ρt1y1t−1 (m − 1 2 )ρtm−1ym−1t−1 2 ) 1 s−1( 1 2 )ρsx1y2x2···xm−3xm−2ym−1x−1 ( 3 2 )ρt2y2t−1 2 )ρsy1x1y−1 1 s−1( 3 1 t2y−1 2 t3y−1 2 x−1 3 t−1 2 t−1 m−1 1 2 3 δs(1)δsx1 (2)δsx1x2(3) · · · δsx1x2···xm−1(m)ρsy−1 1 s−1(m − 3 · · · ρsx1y2x2···xm−3ym−2xm−2y−1 m−2···x−1 m−1x−1 ( 5 2 ) · · · 2 )ρsx1y2x2y−1 2 x−1 1 s−1( 5 2 ) 3 x−1 1 s−1(m − 1 2 ) m−2···x−1 ( 1 2 )δt1 (1)ρt1y1t−1 1 ( 3 2 ) (m − 1 2 ) yields that (1)vg−1 2 h−1 2 h−1 1 g2 (2) = 1 2 )(cid:1) 1 g3 3 h−1 2 h−1 3 h−1 2 )wg2 2 )vg−1 −1g2( 3 2 ) · · · wg−1 G wg1 vg−1 −1g3( 5 (3) · · · vg−1 2 )h2( 3 m−1gm m−1h−1 2 ) · · · hm(m − 1 (m − 1 1 h−1 1 g1 (m − 1). zH(cid:0)δg1(1)δg2(2) · · · δgm(m)h1( 1 ,m)(cid:17) is a C ∗-subalgebra of A(H,G), generated by 2o . nωg(x), vh(l) : g ∈ G, h ∈ H, x, l ∈ Λ1,m− 1 m−1···h−1 m−1h−1 1 gm−1 2 Hence, zH(cid:16)FH (Λ 1 By induction, one can show zH(cid:16)FH (Λn− 1 2 ,m)(cid:17) is generated by 2o . nωg(x), vh(l) : g ∈ G, h ∈ H, x, l ∈ Λn,m− 1 6 Remark 2.2. For Λ ⊆ 1 2 Z, let AH (Λ) = hvh(x), wg(l) : h ∈ H, g ∈ G, x, l ∈ Λi. Lemma 2.2 together with Lemma 2.1 implies that zH : FH (Λn− 1 expectation. 2 ,m) → AH (Λn,m− 1 2 ) is a conditional Theorem 2.1. The observable algebra A(H,G) is the C ∗-algebra given by the C ∗-inductive limit A(H,G) =SΛ AH (Λ). Proof. If A ∈ A(H,G) and ε > 0, then from Lemma 2.2 and the continuity of the projection zH , we know A = zH (A) and there is B ∈ FH (Λn− 1 ,m) with kA − Bk < ε, which implies that 2 kA − zH (B)k = kzH (A − B)k ≤ kA − Bk < ε, and zH (B) ∈ AH (Λn,m− 1 2 ). 3 C ∗-index This section will give the C ∗-index of conditional expectation zH : FH → A(H,G), where H is a normal subgroup of G with [G : H] = k and t1 = e, t2, · · · , tk is a left coset representation of H in G, namely G = tiH and i 6= j induces that tiH ∩ tj H = ∅, where e is the unit of G. Definition 3.1. Let Γ be a conditional expectation from a unital C ∗-algebra B onto its unital C ∗- subalgebra A. A finite family {(u1, u∗ if for all a ∈ B, n)} ⊆ B × B is called a quasi-basis for Γ 1), (u2, u∗ 2), · · · , (un, u∗ kSi=1 Furthermore, if there exists a quasi-basis for Γ, we call Γ of index-finite type. In this case we define the index of Γ by Index Γ = uiΓ(u∗ i a) = a = Γ(aui)u∗ i . nPi=1 nPi=1 uiu∗ i . nPi=1 Remark 3.1. (1) If Γ is a conditional expectation of index-finite type, then Index Γ is in the center of A and does not depend on the choice of quasi-basis ([15]). (2) Let N ⊆ M be factors of type II1 and Γ : M → N the canonical conditional expectation determined by the unique normalized trace on M , then Index Γ is exactly Jones index [M, N ] based on the coupling constant ([16]). More generally, let M be a (σ-finite) factor with a subfactor N and Γ a normal conditional expectation from M onto N , then Γ is of index-finite if and only if Index Γ is finite in the sence of [17], and the values of Index Γ are equal. 7 Theorem 3.1. For fixed k ∈ Z, x ∈ G, y ∈ H, set Then {(ux,y, u∗ x,y) : x ∈ G, y ∈ H} is a quasi-basis of zH : FH → A(H,G). ux,y =pGδx(k)ρy(k + 1 2 ). Proof. Without loss of generality, one can consider the case k = 1. Firstly, one can show that {(ux,y, u∗ ), for any m ∈ Z and m > 1. AH (Λ1,m− 1 2 Note that x,y) : x ∈ G, y ∈ H} is a quasi-basis of zH : FH (Λ 1 2 ,m) → 2 ) · · · hm(m − 1 2 )h1( 1 2 )h2( 3 2 ) 2 )i 2 )h−1 1 y−1h1 ( 3 2 ) 2 ) · · · hm(m − 1 δx(1)y( 3 Px∈G Py∈H = G Px∈G Py∈H = G Px∈G Py∈H = G Px∈G Py∈H = G Px∈G Py∈H h2( 3 · · · hm(m − 1 2 )i 2 )h2( 3 2 ) · · · hm(m − 1 x,yδg1(1)δg2 (2) · · · δgm(m)h1( 1 2 )δg1(1)δg2 (2) · · · δgm(m)h1( 1 ux,yzHhu∗ ux,yzHhδx(1)y−1 ( 3 ux,yzHhδx(1)δg1 (1)δy−1g2(2)δy−1g3(3) · · · δy−1gm(m)y−1 ( 3 2 )i ux,yzHhδx(1)δg1 (1)δy−1g2(2)δy−1g3(3) · · · δy−1gm(m)h1( 1 2 )h2( 3 2 )i 2 )zHhδx(1)δg1 (1)δy−1g2(2)δy−1g3(3) · · · δy−1gm(m)h1( 1 2 )i 2 )zHhδg1(1)δh−1 2 )h Ps∈G m h−1 2 ) · · · hm(m − 1 ( 3 2 )h3( 5 δs(1)δsg−1 2 )i (2)δsg−1 m−1···h−1 m−1···h−1 m−1···h−1 (2)δh−1 1 h−1 4 h−1 m h−1 1 g3 1 g2 1 g3 3 (3) · · · δh−1 m−1···h−1 h3( 5 2 ) · · · hm(m − 1 = Gδg1 (1)h1h2···hm( 3 h1( 1 2 )h−1 m h−1 = δg1(1)h1h2···hm( 3 1 h−1 sg−1 1 h1g1s−1( 1 2 )sg−1 m h−1 m−1···h−1 1 h−1 2 )δh−1 1 h−1 m h−1 = δg1(1)h1h2···hm( 3 m−1···h−1 h1( 1 2 )h−1 1 g2 ( 3 2 )h3( 5 = δg1(1)δg2 (2) · · · δgm(m)h1( 1 2 )h−1 = δg1(1)δg2 (2) · · · δgm(m)h1( 1 2 )h2( 3 m−1···h−1 1 h−1 4 h−1 m h−1 3 m h−1 1 h−1 2 )sg−1 m−1···h−1 1 g2 3 g1s−1( 3 m h−1 m h−1 1 h3g1s−1( 5 4 h−1 (3) · · · δh−1 (2)δh−1 m−1···h−1 1 g3 2 ) · · · hm(m − 1 2 ) ( 3 2 )h−1 1 h−1 m−1···h−1 m h−1 2 ) · · · hm(m − 1 2 ) 2 )h3( 5 1 h1h2···hmh1 2 )h−1 1 y−1h1h2 m h−1 m−1···h−1 1 gm ( 3 2 ) (m) m h−1 m−1···h−1 1 gm (m) (3) · · · δsg−1 1 h−1 1 hmg1s−1(m − 1 1 gm (m) 2 )i 2 ) · · · sg−1 m h−1 m−1···h−1 4 h−1 3 ( 3 2 )h3( 5 2 ) · · · hm(m − 1 2 ) which yields that for any a ∈ FH (Λ 1 2 ,m), ux,yzH (u∗ x,ya) = a. Px∈G Py∈H Similarly, one can verify Px∈G Py∈H zH (aux,y)u∗ x,y = a, ∀a ∈ FH (Λ 1 2 ,m). 8 By induction, we can show that {(ux,y, u∗ ), for any n, m ∈ Z and n < m. AH (Λn,m− 1 2 x,y) : x ∈ G, y ∈ H} is a quasi-basis of zH : FH (Λn− 1 2 ,m) → 2 AH (Λn,m− 1 ) by continuity, and then {(ux,y, u∗ Since zH is a projection of norm one, zH can therefore be extended to the map of Sn<m onto Sn<m zH : Sn<m Finally, the uniqueness of the C ∗-inductive limit ([14]) implies that FH = Sn<m and A(H,G) = Sn<m 2 ,m) → Sn<m ). As a result, {(ux,y, u∗ AH (Λn,m− 1 AH (Λn,m− 1 FH (Λn− 1 ). 2 2 zH : FH → A(H,G). FH (Λn− 1 2 ,m) FH (Λn− 1 2 ,m) x,y) : x ∈ G, y ∈ H} is a quasi-basis of x,y) : x ∈ G, y ∈ H} is a quasi-basis of From Theorem 3.1, we know zH is a conditional expectation of index-finite type, which can guarantee that zH is non-degenerate. Remark 3.2. (1) For k, l ∈ Z with k ≤ l, x ∈ G, y ∈ H, set One can verify that {(wx,y, w∗ x,y) : x ∈ G, y ∈ H} is a quasi-basis of zH : FH → A(H,G). wx,y =pGδx(k)ρy(l + 1 2 ). (2) Let k, l ∈ Z with k > l, x ∈ G, y ∈ H, put vx,y =pGδx(k)ρy(l + 1 2 ), x,y) : x ∈ G, y ∈ H} is not a quasi-basis of zH : FH → A(H,G). In fact, one can x,y) : x ∈ G, y ∈ H} is not a quasi-basis of zH : FH (Λ 1 ,2) → AH (Λ1, 3 2 ), where 2 2 ). then {(vx,y, v∗ show that {(νx,y, ν∗ Notice that νx,y =pGδx(1)ρy( 1 Px∈G Py∈H = G Px∈G Py∈H = G Px∈G Py∈H = G Px∈G Py∈H 2 )h2( 3 x,yδg1(1)δg2 (2)h1( 1 νx,yzHhν∗ 2 )i νx,yzHhδx(1)y−1 ( 1 νx,yzHhδx(1)δy−1g1(1)δy−1g2(2)y−1 ( 1 νx,yzHhδx(1)δy−1g1(1)δy−1g2(2)y−1h1( 1 2 )δg1(1)δg2(2)h1 ( 1 2 )h2( 3 2 )h1( 1 2 )h2( 3 2 )h2( 3 2 )i 2 )i 2 )i 2 h−1 = Gδh−1 = δh−1 2 h−1 1 g1 1 g1 (1)h1h2( 1 (1)h1h2( 1 2 )zHhδh−1 2 )h Ps∈G 2 h−1 1 g1 δs(1)δsg−1 1 g2 (2)h−1 2 ( 1 2 )h2( 3 (1)δh−1 2 h−1 (2)sg−1 1 g2 1 h1h−1 2 h−1 1 g1s−1( 1 1 g1s−1( 1 2 )sg−1 1 h1h2h−1 2 )sg−1 1 h1h2h−1 2 )i 1 g1s−1( 3 1 g1s−1( 3 2 ) 2 )i = Ps∈G = δh−1 δh−1 2 h−1 1 g1 (1)δh−1 (1)δh1h2s(1)δh1h2sg−1 1 g2 ( 1 2 )h−1 (2)h−1 2 h−1 1 g2 2 h1h2 2 h−1 1 g1 (2)h1h2sg−1 2 h−1 1 h2h1h2 1 h1h−1 2 h−1 ( 3 2 ), 9 which can implies that for some a ∈ FH (Λ 1 2 ,2), we have νx,yzH (ν∗ x,ya) 6= a. Px∈G Py∈H Theorem 3.2. The C ∗-index of zH : FH → A(H,G) is GHI. Proof. Since Index zH does not depend on the choice of quasi-basis, then ux,yu∗ x,y Index zH = Px∈G Py∈H = G Px∈G Py∈H = G Px∈G Py∈H = G Px∈G Py∈H = GHI. δx(1)ρy( 3 2 )δx(1)ρ−1 2 )ρ−1 y ( 3 2 ) y ( 3 2 ) δx(1)δx(1)ρy( 3 δx(1)ρe( 3 2 ) Remark 3.3. In particular, if H = G, then zG = 1 (e, g) : F → A(G,G) is a conditional expectation of index-finite type, and Index zG = G2I. References G Pg∈G [1] Doplicher, S., Roberts, J. Fields, statistics and non-abelian gauge group. Comm. Math. Phys., 28: 331-348 (1972) [2] Jones, V.F.R. Subfactors and Knots, CBMS, No. 80, American Mathematical Society Providence, Rhode Island, 1991. [3] Szlach´anyi, K., Vecsernyes, P. Quantum symmetry and braided group statistics in G-spin models. Comm. Math. Phys., 156: 127-168 (1993) [4] Dancer, K.A., Isaac, P.S., Links, J. Representations of the quantum doubles of finite group algebras and spectral parameter dependent solutions of the Yang-Baxter equations. J. Math. Phys., 47: 103511 (2006) [5] Mason, G. The quantum double of a finite group and its role in conformal field theory, London Mathematical Society Lecture Notes, 212, 405-417, Cambridge Univ. Press, Cambridge, 1995. [6] B´antay, P. Orbifolds and Hopf algebras. Phys. Lett. B., 245: 477-479 (1990) [7] Kassel, C. Quantum groups, Springer, New York, 1995, GTM 155. 10 [8] Nill, F., Szlach´anyi, K. Quantum chains of Hopf algebras with quantum double cosymmetry. Comm. Math. Phys., 187: 159-200 (1997) [9] Jiang, L.N. C ∗-index of observable algebras in G-spin models. Science in China. Ser.A Mathe- matics, 48: 57-66 (2005) [10] Xin, Q.L., Jiang, L.N. Symmetric structure of field algebra of G-spin models determined by a normal subgroup. J. Math. Phys., 55: 091703 (2014) [11] Sweedler, M.E. Hopf algebras, W.A. Benjamin, New York, 1969. [12] Abe, E. Hopf Algebras, Cambridge Tracts in Mathematics, No. 74, Cambridge Univ. Press, New York, 1980 [13] Jiang, L.N. Towards a quantum Galois theory for quantum double algebras of finite groups. Proc. Amer. Math. Soc., 138: 2793-2801 (2010) [14] Li, B.R. Operator Algebras, Scientific Press, Beijing, 1998 (in Chinese). [15] Watatani, Y. Index for C ∗-subalgebras, Mem. Am. Math. Soc. No. 424, 1990. [16] Pimsner, M., Popa, S. Entropy and index for subfactors, Ann. Sci. Ecole. Norm. Sup., 19: 57-106 (1986) [17] Kosaki, H. Extension of Jones' theory on index to arbitrary factors. J. Func. Anal., 66: 123-140 (1986) 11
1510.03967	3	1510	2017-02-03T18:59:20	Duality for symmetric Hardy spaces of noncommutative martingales	[ "math.OA", "math.FA" ]	We show the dual spaces of conditional Hardy space and symmetric Hardy space of noncommutative martingales. We derive relationship between the symmetric Hardy space of noncommutative martingales and its conditioned version.	math.OA	math	DUALITY FOR SYMMETRIC HARDY SPACES OF NONCOMMUTATIVE MARTINGALES TURDEBEK N. BEKJAN Abstract. We show the dual spaces of conditional Hardy space and symmetric Hardy space of noncommutative martingales. We derive relationship between the symmetric Hardy space of noncommutative martingales and its conditioned version. 0. Introduction The theory of noncommutative martingale inequalities has been rapidly developed. Many of the classical martingale inequalities have been transferred to the noncommutative setting. Here we mention only three of them directly related with the objective of this paper. The first one is the noncommutative Burkholder-Gundy inequality. In the fundamental work [28], Pisier/Xu established the noncommutative Burkholder-Gundy inequalities of noncommutative martingales in noncommutative Lp-spaces. In [13], the authors proved Burkholder-Gundy inequalities for martingale difference sequences in certain noncommutative symmetric spaces. Some versions can be found in [3, 4, 10]. The second one is the noncommutative Burkholder-Rosenthal inequality. Junge/Xu [20, 21]) obtained the noncommutative analogue of the classical Burkholder/Rosenthal inequalities on the conditioned (or little) square function in noncommutative Lp-spaces. In [13], the authors gave a generalization of Rosenthal inequality to noncommutative symmetric spaces. Dirksen [11] proved that if symmetric space's Boyd index satisfy 2 < pE ≤ qE < ∞, then the Burkholder-Rosenthal inequalities hold in corresponding noncommutative symmetric space. Randrianantoanina and Wu [29] was proved the Burkholder-Rosenthal inequalities hold if symmetric space has the Fatou prop- erty and its Boyd index satisfy 1 < pE ≤ qE < 2. The third result is that independently Junge/Mei[19] and Perrin[27] obtained the following relationship between Hardy space of noncommutative martingales and its conditioned version: 7 1 0 2 b e F 3 ] . A O h t a m [ 3 v 7 6 9 3 0 . 0 1 5 1 : v i X r a (0.1) for all 1 ≤ p ≤ 2. H c p(M) p(M) = hc p(M) + hd 2010 Mathematics Subject Classification. 46L53, 46L51. Key words and phrases. Noncommutative martingale, Burkholder-Rosenthal inequalities, noncommutative sym- metric spaces. This work is supported by project 3606/GF4 of Science Committee of Ministry of Education and Science of Republic of Kazakhstan. 1 2 T. N. BEKJAN The main novelty of our paper is the following dualities for Hardy spaces and its conditional version: If E is a separable symmetric space with 1 < pE ≤ qE < 2 and (E×( 1 2 ))× is separable, then and (H c E(M))∗ = H c E× (M), (H r E(M))∗ = H r E×(M) (hc E(M))∗ = hc E×(M), (hr E(M))∗ = hr E×(M). As applications of this result, we extended (0.1) to the noncommutative symmetric case. We use this result and Burkholder-Gundy inequalities to obtain that the Burkholder-Rosenthal inequalities: (0.2) hold. LE(M) = hc E(M) + hd E(M) + hr E(M) The organization of the paper is as follows. In Section 1, we give some preliminaries and notations on noncommutative martingales and the noncommutative Hardy spaces. We prove the main results in Section 2. 1. Preliminaries Now let E be a quasi-Banach lattice. Let 0 < α < ∞. E is said to be α-convex (resp. α-concave) if there exists a constant C > 0 such that for all finite sequence (xn) in E 1 k(P xnα) (resp. k(P xnα) α kE ≤ C(P kxnkα α kE ≥ C−1(P kxnkα E) 1 1 α , 1 α ). E) The least such constant C is called the α-convexity (resp. α-concavity) constant of E and is denoted by M (α)(E) (resp. M(α)(E)). For 0 < p < ∞, E(p) will denote the quasi-Banach lattice defined by equipped with the quasi-norm E(p) = {x : xp ∈ E}, kxkE(p) = kxpk 1 p E. It is easy to check that if E is α-convex and β-concave then E(p) is pα-convex and pβ-concave with M (pα)(E(p)) ≤ M (α)(E) 1 p and M(pβ)(E(p)) ≤ M(β)(E) 1 p . Therefore, if E is α-convex then E( 1 α ) is 1-convex, so it can be renormed as a Banach lattice (see [17, 25, 31]). Let ([01], Σ, m) be the Lesbegue measure space and L0[0, 1] be the space of all classes of Lebesgue measurable real-valued functions defined on [0, 1]. Let x ∈ L0[0, 1]. Recall that the distribution function of x is defined by λs(x) = m({ω ∈ [0, 1] : x(ω) > s}), s > 0 and its decreasing rearrangement by µt(x) = inf {s > 0 : λs(x) ≤ t}, t > 0. DUALITY FOR SYMMETRIC HARDY SPACES OF NONCOMMUTATIVE MARTINGALES 3 For x, y ∈ L0[0, 1] we say x is submajorized by y, and write x (cid:22) y, if Z t 0 µs(x)ds ≤ Z t 0 µs(y)ds, for all t > 0. By a symmetric quasi Banach space on [0, 1] we mean a Banach lattice E of measurable functions on [0, 1] satisfying the following properties: (i) E contains all simple functions; (ii) if x ∈ E and y is a measurable function on [0, 1] such that µ(y) ≤ µ(x), then y ∈ E and kykE ≤ kxkE. A symmetric quasi Banach space E on [0, 1] is said to have the Fatou property if for every net (xi)i∈I in E satisfying 0 ≤ xi ↑ and supi∈I kxikE < ∞ the supremum x = supi∈I xi exists in E and kxikE ↑ kxkE. We say that E has order continuous norm if for every net (xi)i∈I in E such that xi ↓ 0 we have kxikE ↓ 0. A symmetric Banach space E on [0, 1] is called fully symmetric if, in addition, for x ∈ L0([0, 1]) and y ∈ E with x (cid:22) y it follows that x ∈ E and kxkE ≤ kykE. The Kothe dual of a symmetric Banach space E on [0, 1] is the symmetric Banach space E× given by E× = (cid:26)x ∈ L0([0, 1]) : sup{Z 1 0 x(t)y(t)dt : kxkE ≤ 1} < ∞(cid:27) ; kykE× = sup{Z 1 0 x(t)y(t)dt : kxkE ≤ 1}, y ∈ E×. The space E× is fully symmetric and has the Fatou property. It is isometrically isomorphic to a closed subspace of E∗ via the map y → Ly, Ly(x) = Z 1 0 x(t)y(t)dt (x ∈ E). A symmetric Banach space E on [0, 1] has the Fatou property if and only if E = E×× isometrically. It has order continuous norm if and only if it is separable, which is also equivalent to the statement E∗ = E×. Moreover, a symmetric Banach space which is separable or has the Fatou property is automatically fully symmetric (cf. [5, 6, 22, 25]). If E is a symmetric (quasi-)Banach space on [0, 1] which is β-concave for some β < ∞ then E has order continuous (quasi-)norm (cf. [1, 13]). Let Ei be a symmetric quasi Banach space on [0, 1], i = 1, 2. We define the pointwise product space E1 ⊙ E2 as (1.1) E1 ⊙ E2 = {x : x = x1x2, xi ∈ Ei, i = 1, 2} with a functional kxkE1⊙E2 defined by kxkE1⊙E2 = inf{kx1kE1kx2kE2 : x = x1x2, , xi ∈ Ei, i = 1, 2}. By Theorem 2 in [23], we know that if Ei is a quasi symmetric Banach space on [0, 1], i = 1, 2. Then E1 ⊙ E2 is a quasi symmetric Banach space on [0, 1]. 4 T. N. BEKJAN Using Holder type inequality to Banach lattice (see Proposition 1.d.2 in [25] or (3.1) in [31]) we obtain the following: If E is a symmetric Banach space on [0, 1], then (1.2) E( 1 2 ) = E ⊙ E. We use the following result( see (iii) of Theorem 1 and Corollary 2 in [23]). Theorem 1.1. Let E, F be symmetric Banach spaces on [0, 1]. (i) If 0 < p < ∞, then (E ⊙ F )(p) = E(p) ⊙ F (p). (ii) If 1 < p < ∞, then (E(p))× = (E×)(p) ⊙ Lp′ [0, 1], where 1 p + 1 p′ = 1. We also use the following Lozanovskii factorization theorem (Theorem 6 in [26]). Theorem 1.2. Let E be a symmetric Banach space on [0, 1], then L1[0, 1] = E ⊙ E×. Proposition 1.3. Let E be a symmetric Banach space on [0, 1]. If E is α-convex and β-concave for some α > 1 and 1 < β < 2 then E = F ⊙ E×, where F = (E×( 1 2 ))× is separable. Proof. Since E∗ = E×, by Proposition 1.d.4 in [25], it follows that E× is β′-convex and α′- 2 α′-concave. Since 2 ) becomes a symmetric Banach 2 ) can be renormed, with this equivalent norm, E×( 1 β ′ = 1. Hence E×( 1 2 β′-convex and 1 α + 1 β + 1 2 ) is 1 1 α′ = 1 and 1 concave, where 1 2 β′ > 1, E×( 1 space. On the other hand, from 1 2 ))×. Since E×( 1 γ + 1 2 β′ (i.e. 1 By (ii) of Theorem 1.1, we have γ is conjugate of 1 Let F = (E×( 1 2 α′ < ∞ follows that E×( 1 2 ) is separable, F = (E×( 1 2 ) is separable. 2 ))∗. Therefore, F is γ-concave, where 2 β ′ = 1). It follows that γ < ∞, so F is separable. 1 E = (E×)× = (cid:0)[E×( 1 2 )](2)(cid:1)× = (cid:0)[E×( 1 2 )]×(cid:1)(2) ⊙ L2[0, 1] = F (2) ⊙ L2[0, 1]. Using Theorem 1.2 and (i) of Theorem 1.1 we obtain that E = F (2) ⊙ L2[0, 1] = F (2) ⊙ [F ⊙ E×( 1 2 )](2) = F (2) ⊙ [F (2) ⊙ E×] = F ⊙ E×. (cid:3) 1.1. Noncommutative symmetric Banach spaces. We use standard notation and notions from theory of noncommutative Lp-spaces. Our main references are [28] and [18] (see also [28] for more historical references). Throughout this paper, we denote by M a finite von Neumann algebra on the Hilbert space H with a faithful normal normalized finite trace τ . The closed densely defined linear operator x in H with domain D(x) is said to be affiliated with M if and only if u∗xu = x for all unitary operators u which belong to the commutant M′ of M. If x is affiliated with M, then x is said to be τ -measurable if for every ε > 0 there exists a projection e ∈ M such that e(H) ⊑ D(x) and τ (e⊥) < ε (where for any projection e we let e⊥ = 1 − e ). The set of all DUALITY FOR SYMMETRIC HARDY SPACES OF NONCOMMUTATIVE MARTINGALES 5 τ -measurable operators will be denoted by L0(M). The set L0(M) is a ∗-algebra with sum and product being the respective closure of the algebraic sum and product. Let x be a positive measurable operator and x = R ∞ 0 s des(x) its spectral decomposition. Com- posed with the spectral measure (es(x)), the trace τ induces a positive measure dτ (es(x)) on R+. Thus we are reduced to the commutative case by regarding x as a function. In view of the preceding discussion on functions, this justifies the following definition. Recall that e⊥ s (x) = 1(s,∞)(x) is the spectral projection of x corresponding to the interval (s, ∞). Let x ∈ L0(M). Define λs(x) = τ(cid:0)e⊥ s (x)), s > 0 and µt(x) = inf{s > 0 : λs(x) ≤ t}, t > 0. The function s 7→ λs(x) is called the distribution function of x and the µt(x) the generalized singular numbers of x. For more details on generalized singular value function of measurable operators we refer to [18]. Let E be a symmetric Banach space on [0, 1]. We define LE(M) = {x ∈ L0(M) : µ.(x) ∈ E}; kxkLE(M) = kµ.(x)kE, x ∈ LE(M). Then (LE(M), k.kLE(M)) is a Banach space (cf. [14, 30, 31]). We need the following result (Theorem 5.6 in [16], p. 745). Theorem 1.4. Let E be a separable symmetric Banach space on [0, 1], then LE(M)∗ = LE×(M) isometrically, with associated duality bracket given by hx, yi = τ (xy) (x ∈ LE(M), y ∈ LE×(M)). In what follows, unless otherwise specified, we always denote by E a symmetric Banach space on [0, 1]. Let a = (an)n≥0 be a finite sequence in LE(M), define kakLE(M,ℓ2 c) = k(Xn≥0 an2)1/2kLE(M), kakLE(M,ℓ2 r) = k(Xn≥0 a∗ n2)1/2kLE (M). This gives two noms on the family of all finite sequences in LE(M). To see this, denoting by B(ℓ2) the algebra of all bounded operators on ℓ2 with its usual trace tr, let us consider the von Neumann algebra tensor product M ⊗ B(ℓ2) with the product trace τ ⊗ tr. τ ⊗ tr is a semi-finite normal faithful trace, the associated noncommutative LE space is denoted by LE(M ⊗ B(ℓ2)). Now, any finite sequence a = (an)n≥1 in LE(M) can be regarded as an element in LE(M ⊗ B(ℓ2)) via the 6 T. N. BEKJAN following map a 7−→ T (a) = a0 a1 ...   0 0 ... . . . . . . . . .   , that is, the matrix of T (a) has all vanishing entries except those in the first column which are the an's. Such a matrix is called a column matrix, and the closure in LE(M ⊗ B(ℓ2)) of all column matrices is called the column subspace of LE(M ⊗ B(ℓ2)). Then kakLE(M,ℓ2 c) = kT (a)kLE(M⊗B(ℓ2)) = kT (a)kLE(M⊗B(ℓ2)). Therefore k.kLE(M,ℓ2 sponding completion is a Banach space, denoted by LE(M, ℓ2 property, then a sequence a = (an)n≥1 in LE(M) belongs to LE(M, ℓ2 c) defines a norm on the family of all finite sequences of LE(M). The corre- c). It is clear that if E has the Fatou c) iff If this is the case, ( ∞ Pk=1 n sup n≥1 k( Xk=1 ak2)1/2kLE (M) < ∞. ak2)1/2 can be appropriately defined as an element of LE(M). Similarly, we may show that k.kLE(M,ℓ2 defines a Banach space LE(M, ℓ2 r) is a norm on the family of all finite sequence in LE(M). As above, it r), which now is isometric to the row subspace of LE(M ⊗ B(ℓ2)) consisting of matrices whose nonzero entries lie only in the first row. Observe that the column and row subspaces of LE(M ⊗ B(ℓ2)) are 1-complemented subspace (by the definition of E and Theorem 3.4 in [15]). We also need Ld E(M), the space of all sequences a = (an)n≥1 in LE(M) such that kakLd E(M) = kdiag(an)kLE (M⊗B(ℓ2)) < ∞, where diag(an) is the matrix with the an on the diagonal and zeroes elsewhere. 1.2. Noncommutative martingales. Let M be a finite von Neumann algebra with a normalized normal faithful trace τ. Let (Mn)n≥1 be an increasing sequence of von Neumann subalgebras of M such that ∪n≥1Mn generates M (in the w∗-topology). (Mn)n≥1 is called a filtration of M. The restriction of τ to Mn is still denoted by τ. Let En = E(·Mn) be the conditional expectation of M with respect to Mn. Then En is a norm 1 projection of LE(M) onto LE(Mn) (see Theorem 3.4 in [15]) and En(x) ≥ 0 whenever x ≥ 0. A noncommutative LE-martingale with respect to (Mn)n≥1 is a sequence x = (xn)n≥1 such that xn ∈ LE(Mn) and En(xn+1) = xn for any n ≥ 1. Let kxkLE(M) = supn≥1 kxnkLE(M). If kxkLE(M) < ∞, then x is said to be a bounded LE-martingale. DUALITY FOR SYMMETRIC HARDY SPACES OF NONCOMMUTATIVE MARTINGALES 7 Let x be a noncommutative martingale. The martingale difference sequence of x, denoted by dx = (dxn)n≥1, is defined as dx1 = x1, dxn = xn − xn−1, n ≥ 2. Set Sc n(x) = (cid:16) n Xk=1 1 2 dxk2(cid:17) and Sr n(x) = (cid:16) n Xk=1 1 2 . dx∗ k2(cid:17) By the preceding discussion, if E has the Fatou property, then dx belongs to LE(M, ℓ2 LE(M, ℓ2 n(x))n≥1) is a bounded sequence in LE(M); in this r)) if and only if (Sc n(x))n≥1 (resp. (Sr c) (resp. case, Sc(x) = (cid:16) ∞ Xk=1 1 2 dxk2(cid:17) and Sr(x) = (cid:16) ∞ Xk=1 1 2 dx∗ k2(cid:17) are elements in LE(M). These are noncommutative analogues of the usual square functions in the commutative martingale theory. It should be pointed out that the two sequences Sc n(x) and Sr n(x) may not be bounded in LE(M) at the same time. We define H c LE(M, ℓ2 E(M) (resp. H r c) (resp. dx ∈ LE(M, ℓ2 E(M)) to be the space of all LE-martingales such that dx ∈ r) ), equipped with the norm kxkHc E (M) = kdxkLE(M,ℓ2 c) (cid:0)resp. kxkHr r)(cid:1). E (M) = kdxkLE (M,ℓ2 H c E(M) and H r E(M) are Banach spaces. Note that if x ∈ H c E(M), kxkHc E (M) = sup n≥0 kSc n(x)kLE (M) = kSc(x)kLE (M). Similar equalities hold for H r E(M). We now consider the conditioned versions of square functions and Hardy spaces developed in [20]. For a finite noncommutative LE-martingale x = (xn)n≥1 define (with E0 = E1) and kxkhc E(M) = k(cid:16)Xk≥1 Ek−1(dxk2)(cid:17) 1 2 kLE(M) kxkhr E (M) = k(cid:16)Xk≥1 Ek−1(dx∗ k2)(cid:17) 1 2 kLE(M). Let hc E(M) and hr E(M) are Banach spaces. We define the column and row conditioned square functions as follows. For any finite E(M) be the corresponding completions. Then hc E(M) and hr martingale x = (xn)n≥1 in LE(M ), we set sc(x) = (cid:16)Xk≥1 Ek−1(dxk2)(cid:17) 1 2 and sr(x) = (cid:16)Xk≥1 Ek−1(dx∗ k2)(cid:17) 1 2 . Then kxkhc E (M) = ksc(x)kLE (M) and kxkhr E(M) = ksr(x)kLE (M). 8 T. N. BEKJAN Let x = (xn)n≥0 be a finite LE-martingale, we set sd(x) = diag(dxn) We note that ksd(x)kLE (M⊗B(ℓ2)) = kdxnkLd E(M) Let hd E(M) be the subspace of Ld E(M) consisting of all martingale difference sequences. 1.3. The space LE(M; ℓ∞). Recall that LE(M; ℓ∞) is defined as the space of all sequences (xn)n≥1 in LE(M) for which there exist a, b ∈ LE(2) (M) and a bounded sequence (yn)n≥1 in M such that xn = aynb for all n ≥ 1. For such a sequence, set (1.3) k(xn)n≥1kLE (M,ℓ∞) := inf(cid:8)kakL E(2) (M) sup n kynk∞kbkL E(2) (M)(cid:9), where the infimum runs over all possible factorizations of (xn)n≥1 as above. This is a norm and LE(M; ℓ∞) is a Banach space (see [11]). As in [21], we usually write We warn the reader that this suggestive notation should be treated with care. It is used for possibly (cid:13)(cid:13)sup n +xn(cid:13)(cid:13)E = k(xn)n≥1kLE(M,ℓ∞). nonpositive operators and (cid:13)(cid:13)sup n +xn(cid:13)(cid:13)E 6= (cid:13)(cid:13)sup n +xn(cid:13)(cid:13)E in general. However it has an intuitive description in the positive case: A positive sequence (xn)n≥1 of LE(M) belongs to LE(M; ℓ∞) if and only if there exists a positive a ∈ LE(M) such that xn ≤ a for any n ≥ 1 and in this case, (1.4) (cid:13)(cid:13)sup n +xn(cid:13)(cid:13)E ≤ inf kakLE(M) ≤ 2(cid:13)(cid:13)sup n +xn(cid:13)(cid:13)E , where the infimum runs over all possible positives a ∈ LE(M) as above. Indeed, let (xn)n≥1 ∈ LE(M; ℓ∞). Then for ε > 0, there exist a, b ∈ LE(2) (M)+ and a bounded sequence (yn)n≥1 E(2) (M), supn kynk∞ = 1 and in M such that xn = aynb for all n ≥ 1, kakL +xnkE + ε. Define the operator c = (a2 + b2) ∈ LE(M)+. Then E(2) (M) = kbkL kakL E(2) (M)kbkL E(2) (M) < ksupn there are contractions u, v ∈ M such that a = c 1 2 u, b = vc 1 1 1 2 Re(uynv)c c So, the second inequality of (1.4) holds. Conversely, if xn ≤ a for some a ∈ L+ 2 , for all n ≥ 1. It follows that xn ≤ c for all n ≥ 1 and kckLE(M) < 2ksupn 1 2 2 (see Remark 2.3 in [9]). Hence, xn = +xnkE+2ε. E(M ), then x n = una +xnkE ≤ kakLE(M). Hence, the 1 2 for a contraction un ∈ M, and so xn = a first inequality of (1.4) holds. 1 2 u∗ nuna 1 2 . Thus ksupn We define LE(M; ℓ1) to be the space of all sequences x = (xn) in LE(M) which can be decom- posed as xn = ∞ Xk=1 uknvnk (n ≥ 1) DUALITY FOR SYMMETRIC HARDY SPACES OF NONCOMMUTATIVE MARTINGALES 9 for two families (ukn)k,n≥1 and (vnk)n,k≥1 in LE(2) (M) satisfying ∞ Xk,n=1 uknu∗ kn ∈ LE(M) and ∞ Xn,k=1 v∗ nkvnk ∈ LE(M), where the series converge in norm. LE(M; ℓ1) is a Banach space when equipped with the norm kxkLE(M;ℓ1) = inf{k ∞ Xk,n=1 uknu∗ knk1/2 LE(M)k ∞ Xn,k=1 nkvnkk1/2 v∗ LE(M)}, where the infimum runs over all decompositions of (xn) as above. We will use the following fact (see [11]). Let x = (xk)k≥1 ∈ LE(M; ℓ1) for which xk ≥ 0 for all k. Then (1.5) kxkLE (M;ℓ1) = kXk≥1 xkkLE (M). We need the following result (Theorem 5.3 and in Remark 5.4 in [11]). Theorem 1.5. Let E be a separable symmetric Banach space on [0, 1]. (i) If positive sequence (xn)n≥1 of LE×(M) belongs to LE×(M; ℓ∞), then (1.6) (cid:13)(cid:13)sup n +xn(cid:13)(cid:13)E× = sup(cid:8)Xk≥1 τ (xkyk) : yk ∈ LE(M)+, kXk≥1 ykkLE(M) ≤ 1(cid:9). (ii) LE(M; ℓ1)∗ = LE×(M; ℓ∞) isometrically, with respect to the duality bracket hx, yi = Xk≥1 where x ∈ LE(M; ℓ1) and y ∈ LE×(M; ℓ∞). τ (xkyk), 2. Main results Lemma 2.1. Let E, E1, E2 be symmetric Banach spaces on [0, 1] such that E = E1 ⊙ E2. x ∈ LE(M)+, then for ε > 0, there exist a ∈ L+ E1 (M) and b ∈ L+ E2 (M) such that x = If ab, kakLE1 (M)kbkLE2 (M) < kxkLE (M) + ε and a is invertible with bounded inverse. Proof. Let N be the commutative von Neumann subalgebra of M generated by the spectral projection of x. Then N is isometric isomorphic to L∞(Ω, Σ, µ) where (Ω, Σ, µ) a finite measure space. Hence, x ∈ LE(N ) = LE(Ω, µ). Since E = E1 ⊙ E2, for every ε > 0, there are x1 ∈ LE1(Ω, µ)+ = LE1(N )+ and x2 ∈ LE2(Ω, µ)+ = LE2(N )+ such that x = x1x2 and kxkLE (N ) + ε 2 > kx1kLE1 (N )kx2kLE2 (N ). Let δ > 0. Set a = x1 + δ and b = x1(x1 + δ)−1x2. Then a ∈ LE1(N ) ⊂ LE1(M), b ∈ LE2(N ) ⊂ LE2(M), x = ab and a is invertible with bounded inverse. For enough small δ we have that kxkLE(M) + ε > kakLE1 (M)kbkLE2 (M). Hence we obtain the desired result. (cid:3) Theorem 2.2. Let E be a separable symmetric Banach space on [0, 1] with 1 < pE ≤ qE < 2. If F = (E×( 1 2 ))× is separable, then we have (i) (hc (ii) (H c E(M))∗ = hc E(M))∗ = H c E×(M) with equivalent norms. E×(M) with equivalent norms. Similarly, (hr E(M))∗ = hr E×(M) and (H r E(M))∗ = H r E× (M) with equivalent norms. 10 T. N. BEKJAN Proof. (i) 1◦ From 1 < pE ≤ qE < 2, we obtain that L2(M) ⊂ LE(M) with continuous inclu- sions, i.e. there exists a constant C > 0 such that kxkE ≤ CkxkL2(M) for all x ∈ L2(M). We identify an element x ∈ L2(M ) with the martingale (En(x))n≥1. By the trace-preserving property of conditional expectations and the orthogonality in L2(M) of martingale difference sequences, we get kxkhc E(M) = k(Pk≥1 Ek−1(dxk2)) ≤ Ck(Pk≥1 Ek−1(dxk2)) = CkxkL2(M), 1 2 kLE(M) 1 2 kL2(M) i.e. this martingale is in hc E(M). Let y ∈ hc E× (M). Since 2 < pE× ≤ qE× < ∞, it follows that E× ⊂ L2([0, 1]) with continuous inclusions. Hence, kykL2(M) ≤ C1kykhc E(M) < ∞, i.e. y is an L2- martingale. Define φy by φy(x) = τ (y∗x), ∀x ∈ L2(M). We must show that φy induces a continuous linear functional on hc E(M). for ε > 0, there exist a ∈ L+ From the proof of Proposition 1.3, it follows that E = F ⊙E×. Using Lemma 2.1, we obtain that E×(M) such that sc(x) = ab, kakLF (M)kbkLE× (M) < ksc(x)kLE (M)+ε and a is invertible with bounded inverse. Then, by the Cauchy-Schwarz inequality F (M) and b ∈ L+ and the tracial property of τ , we have φy(x) = Pn≥1 τ (dy∗ 1 1 ≤ hPn≥1 τ (En−1(a) = hPn≥1 τ (aEn−1(dyn2))i ndxn) = Pn≥1 τ (En−1(a) 2 )i 2 dyn2En−1(a) 2 dy∗ ndxnEn−1(a)− 1 2hPn≥1 τ (En−1(a)− 1 2hPn≥1 τ (En−1(a)−1dxn2)i 1 2 1 1 1 2 ) = I · II. 2 dxn2En−1(a)− 1 2 ] 1 2 By Theorem 1.4, we have I2 = Pn≥1 τ (aEn−1(dyn2)) = τ (aPn≥1 En−1(dyn2)) ≤ kakLF (M)kPn≥1 En−1(dyn2)kL = kakLF (M)k(Pn≥1 En−1(dyn2)) = kakLF (M)kyk2 E× (M). hc 1 ×( 1 2 E ) (M) 2 k2 LE× (M) To estimate II, we set and sc,0(x) = dx1. Then sc,n(x) = (cid:0) n Xk=1 Ek−1(dxk2)(cid:1) 1 2 , ∀n ≥ 1 Xn≥1 En−1(dxn2) = Xn≥0 (sc,n+1(x)2 − sc,n(x)2)). DUALITY FOR SYMMETRIC HARDY SPACES OF NONCOMMUTATIVE MARTINGALES 11 Applying Corollary 2.3 of [8] we obtain that En−1(a−1) ≥ En−1(a)−1. Hence, by Theorem 1.4, we have II2 = Pn≥1 τ (En−1(a)−1dxn2) ≤ Pn≥1 τ (En−1(a−1)dxn2) = Pn≥1 τ (a−1En−1(dxn2)) = Pn≥0 τ (a−1(sc,n+1(x)2 − sc,n(x)2)) = τ (a−1sc(x)2) = τ (bsc(x)) ≤ kbkLE× (M)ksc(x)kLE (M) = kbkLE× (M)kxkhc E (M). Combining the precedent estimations, for any finite L2-martingale x, we deduce that φy(x) ≤ kykhc E× (M)kxkhc E(M) Thus φ extends to an element of hc E(M)∗ with norm at most kykhc E× (M). 2◦ Let φ ∈ hc E(M)∗ of norm one. As L2(M) ⊂ hc E(M), it follows that φ induces a continuous functional φ on L2(M ). Consequently, φ is given by an element y of L2(M), φ(y) = τ (y∗x), ∀x ∈ L2(M). As finite L2- martingales are dense in hc hc E(M). We have E(M) and in L2(M ), we deduce that L2(M ) is dense in (2.1) kφkhc E(M)∗ = sup τ (y∗x) ≤ 1. x∈L2(M),kxkhc E (M)≤1 We want to show that y ∈ hc E×(M) and kykhc E× (M) ≤ C. Set zn = En−1(dyn2), ∀n ≥ 1. Using Theorem 1.4 we obtain that kyk2 hc E× (M) = kPn≥1 En−1(dyn2)kL ×( 1 2 E ) (M) = kPn≥1 znkL F (M) and kakLF (M) ≤ 1(cid:9) ×( 1 2 E ) (M) = sup(cid:8)Pn≥1 τ (zna) : a ∈ L+ = sup(cid:8)Pn≥1 τ (znEn−1(a)) : a ∈ L+ F (M) kakLF (M) ≤ 1(cid:9). Let a ∈ L+ F (M) and kakF ≤ 1. Let b be the martingale defined as follows: dbn = dynEn−1(a), ∀n ≥ 1. τ (y∗b) ≤ kbkhc E(M). Using (2.1) we obtain that We have that τ (y∗b) = Xn≥1 τ (dyn2En−1(a)) = Xn≥1 τ (En−1(dyn2)En−1(a)) = Xn≥1 τ (znEn−1(a)). 12 T. N. BEKJAN On the other hand, by the definition of b, we find sc(b)2 = Pn≥1 En−1(En−1(a)dyn2En−1(a)) = Pn≥1 En−1(a)En−1(dyn2)En−1(a) = Pn≥1 En−1(a)znEn−1(a). Let c ∈ L+ F (M) such that En−1(a) ≤ c for all n ≥ 1. Then for each n, there exists a contraction un ∈ M such that En−1(a) 1 2 = unc 1 2 , and so sc(b)2 = Xn≥1 c Whence 1 2 u∗ nEn−1(a) 1 2 znEn−1(a) 1 2 unc 1 2 = c 1 2(cid:0)Xn≥1 u∗ nEn−1(a) 1 2 znEn−1(a) 1 2 . 1 2 un(cid:1)c µt(sc(b)2) ≤ µ t 3 (c 1 2 )µ t 3 u∗ nEn−1(a) 1 2 znEn−1(a) (Xn≥1 1 2 un)µ t 3 (c 1 2 ), ∀t ∈ [0, t]. On the other hand, from the proof of Proposition 1.3, we know that E = F (2) ⊙ L2[0, 1]. By (1.2), we get E( 1 2 ) = E ⊙ E = F (2) ⊙ L2[0, 1] ⊙ L2[0, 1] ⊙ F (2) = F (2) ⊙ L1[0, 1] ⊙ F (2). Hence, kbk2 hc E(M) = ksc(b)k2 LE(M) = ksc(b)2kL ( 1 2 E ≤ kµ t 3 ≤ kµ t 3 ≤ Kkck nEn−1(a) (c (c 1 2 1 1 2 )µ t 3 2 )kF (2) kµ t (Pn≥1 u∗ F(cid:0)Pn≥1 τ (u∗ 3 (Pn≥1 u∗ nEn−1(a) ≤ KkckF(cid:0)Pn≥1 τ (znEn−1(a))(cid:1), ) (M) = kµt(sc(b)2)k 2 znEn−1(a) (c 2 un)µ t 1 1 E( 1 2 ) 1 2 )k ) 3 1 E( 1 2 un)k1kµ t 2 3 nEn−1(a) 1 2 znEn−1(a) (c 1 2 )kF (2) 1 2 znEn−1(a) 1 2 un)(cid:1)kck 1 2 F where K is a constant depending only on norms of the delation operator on the spaces F (2), L1[0, 1]. Using (1.4) and Corollary 5.4 of [12] we deduce that kbkhc E(M) ≤ Ckak 1 2 F(cid:0)Xn≥1 τ (znEn−1(a))(cid:1) 1 2 , where C is a constant. Combining the preceding inequalities, we obtain that τ (znEn−1(a)) ≤ C2kakF ≤ C2. Xn≥1 It follows that kykhc (ii) 1◦ Let y ∈ H c E(M) ≤ C. Thus we have finished the proof of (i). E×(M) and define φy by φy(x) = τ (y∗x), ∀x ∈ L2(M). We must show that E(M). Let x be a finite L2-martingale such that φy induces a continuous linear functional on H c kxkHc E(M) < ∞. φy(x) = Pn≥1 τ (dy∗ ndxn) = τ ⊗ tr(cid:0)   dy1 dy2 ... · · · · · · . . . ∗     dx1 dx2 ... 0 · · · 0 · · · ... . . . dx1 dx2 ...   0 0 ...   0 0 ... · · · · · · . . .   (cid:17) kLE(M⊗B(ℓ2)) kLE× (M⊗B(ℓ2))k ≤ k   = kykHc dy1 0 · · · 0 ... · · · . . . dy2 ... E× (M)kxkHc ∗   E (M). DUALITY FOR SYMMETRIC HARDY SPACES OF NONCOMMUTATIVE MARTINGALES 13 Thus φy extends to an element of H c E(M)∗ with norm at most kykHc E× (M). 2◦ Let φ ∈ H c E(M)∗ be such that kφkHc E (M)∗ ≤ 1. There exists y ∈ L2(M) such that φ(y) = τ (y∗x), ∀x ∈ L2(M). By the density of L2(M) in H c E(M), we have (2.2) kφkHc E (M)∗ = sup τ (y∗x) ≤ 1. x∈L2(M),kxkHc E (M)≤1 We use the same method in 2◦ of the proof of (i) to show that y ∈ H c E×(M) and kykHc E× (M) ≤ C. By Theorem 1.4, we find that kyk2 Hc E× (M) = kPn≥1 dyn2kL ×( 1 2 E ) (M) = sup(cid:8)Pn≥1 τ (dyn2a) : a ∈ L+ = sup(cid:8)Pn≥1 τ (dyn2En(a)) : a ∈ L+ F (M) and kakLF (M) ≤ 1(cid:9) F (M) kakLF (M) ≤ 1(cid:9). Let a ∈ L+ F (M) and kakF ≤ 1. Set dbn = dynEn(a) − En−1(dynEn(a)), ∀n ≥ 1. Then b is a martingale. By (2.2), it follows that τ (y∗b) ≤ kbkHc E(M). Since (dyn)n≥1 is a martingale difference sequence, we obtain that τ (y∗b) = Pn≥1 τ (dyn2En(a)) −Pn≥1 τ (dy∗ = Pn≥1 τ (dyn2En(a)) −Pn≥1 τ (En−1(dy∗ = Pn≥1 τ (dyn2En(a)). nEn−1(dynEn(a))) n)dynEn(a)) Using the triangular inequality in LE(M, ℓ2 c), we get kbkHc E(M) = k(dbn)n≥1kLE(M,ℓ2 c) ≤ k(dynEn(a))n≥1kLE(M,ℓ2 c) + k(En−1(dynEn(a)))n≥1kLE(M,ℓ2 c). On the other hand, by Lemma 2.2 of [2], there is a constant rE such that k(En−1(dynEn(a)))n≥1kLE(M,ℓ2 c) ≤ rEk(dynEn(a))n≥1kLE(M,ℓ2 c). Hence, kbkHc E(M) ≤ (1 + rE)k(dynEn(a))n≥1kLE(M,ℓ2 c). Let c ∈ L+ F (M) such that En(a) ≤ c for all n ≥ 1. Then for each n, there exists a contraction un ∈ M such that En(a) 1 2 = unc 1 2 . As before, by E( 1 2 ) = F (2) ⊙ L1[0, 1] ⊙ F (2), we have that k(dynEn(a))n≥1k2 LE(M,ℓ2 c) = kc 1 2(cid:0)Pn≥1 u∗ 1 2 nEn(a) F(cid:0)Pn≥1 τ (u∗ ≤ Kkck 1 2 dyn2En(a) 1 2 un(cid:1)c 2 dyn2En(a) 1 nEn(a) ≤ KkckF(cid:0)Pn≥1 τ (dyn2En(a))(cid:1). 1 2 kL ( 1 2 E ) (M) 1 2 un)(cid:1)kck 1 2 F 14 T. N. BEKJAN By (1.4) and Corollary 5.4 of [12] we deduce that kbkHc E(M) ≤ Ckak 1 2 F(cid:0)Xn≥1 τ (dyn2En(a))(cid:1) 1 2 , where C is a constant. Combining the preceding inequalities, we obtain that dyn2a) = Xn≥1 τ (Xn≥1 E(M) ≤ C. Thus we have finished the proof of (ii). τ (dyn2En(a)) ≤ C2kakF ≤ C2. It follows that kykHc Passing to adjoint, we obtain the identities (hr E(M))∗ = hr E×(M) and (H r E(M))∗ = H r E×(M). (cid:3) Lemma 2.3. Let E be a separable symmetric Banach space on [0, 1]. Then we have (hd E(M))∗ = hd E×(M) with equivalent norms. Proof. Recall that hd 2-complemented in Ld E(M) consists of martingale difference sequences in Ld E(M)) via the projection E(M). So hd E(M) is Ld E(M)) −→ (an)n≥1 P :   hd E(M) 7−→ (En(an) − En−1(an))n≥1 By Theorem 1.4, we obtain the desired result. (cid:3) Proposition 2.4. Let E be a separable symmetric Banach space on [0, 1] with 2 < pE ≤ qE < ∞. Then we have (i) H c (ii) H r E(M)) = hc E(M) = H r E(M) ∩ hd E(M) ∩ hd E(M) with equivalent norms. E(M) with equivalent norms. Proof. (i) Let y ∈ H c E(M)). From the proof of Theorem 6.2 in [11] we have that kykHc E(M) ≤ kykhc E(M)∩hd E (M). On the other hand, from the proof of Theorem 6.2 in [11], it follows that kdynkLd E(M) ≤ C1kXn≥1 rn ⊗ dxnkLE(L∞⊗M). Hence, by Theorem 4.7 in [13], we get kdynkLd Corollary 4.13 in [11], we have kykhc E(M) ≤ C3kykHc E(M) ≤ C2kykHc E (M). Thus E(M). Since p ) = 1 2 pE > 1, by E( 1 2 kykhc E(M)∩hd E (M) ≤ C′kykHc E(M). (ii) Passing to adjoint, we obtain the desired result. (cid:3) Proposition 2.5. Let E be a separable symmetric Banach space on [0, 1] with 1 < pE ≤ qE < 2. If F = (E×( 1 2 ))× is separable, then we have (i) H c (ii) H r E(M)) = hc E(M) = hr E(M) + hd E(M) + hd E(M) with equivalent norms. E(M) with equivalent norms. DUALITY FOR SYMMETRIC HARDY SPACES OF NONCOMMUTATIVE MARTINGALES 15 Proof. (i) By Proposition 2.4, it follows that there exist constants C > 0 such that (2.3) max{kxkhc E× (M), kdynkLd E× (M)} ≤ CkxkHc E× (M), ∀x ∈ H c E×(M). Let y ∈ hc E(M). By (2.3) and Theorem 2.2, we deduce that kykHc E(M) = sup E× (M)≤1 kxkHc τ (x∗y) ≤ sup E× (M)≤C kxkhc τ (x∗y) = Ckykhc E(M). Similarly, Hence kzkHc E(M) ≤ Ckzkhd E(M), ∀z ∈ hd E(M). kykHc E(M) ≤ C inf{kwkhd E(M) + kzkhc E(M)}, where the infimum runs over all decomposition y = w + z with w in hc So, hc E(M) + hd E(M))∗ = hc hd E∗ = E× is separable. On the other hand, F = (E×( 1 E(M). E(M). Using Theorem 2.2 and Lemma 2.3 we obtain that (hc E(M) + E×(M). Sincer E is a separable symmetric Banach space on [0, 1], 2 ))× is separable, it follows that E× is E(M) ⊂ H c E×(M) ∩ hd E(M) and z in hd separable. Applying Proposition 2.4 we deduce that H c E(M)) = hc E(M) + hd E(M) with equivalent norms. (ii) Passing to adjoint, we obtain H r E(M) = hr E(M) + hd E(M) with equivalent norms. (cid:3) By Proposition 2.4, 2.5 and Proposition 4.18 in [13], we obtain the following: Theorem 2.6. Let E be a separable symmetric Banach space on [0, 1] with 1 < pE ≤ qE < 2. If F = (E×( 1 2 ))× is separable, then we have LE(M) = hc E(M) + hd E(M) + hr E(M). Remark 2.7. Using Theorem 2.2 and Theorem 6.2 in [11] we can obtain the result of Theorem 2.6. Remark 2.8. Should Theorem 2.6 be true whenever M is semi-finite and E has the Fatou property. This was proved in [29] by Randrianantoanina and Wu via a very different method. References [1] C.D. Aliprantis and O. Burkinshaw, Locally solid Riesz spaces, Pure and Applied Mathematics, Vol. 76, Academic Press, New York, 1978. [2] T.N. Bekjan, Φ-inequalities of noncommutative martingales, Rocky Mountain J. Math. 36(2) (2006), 401-412. [3] T. N. Bekjan, Z. Chen, P. Liu and Y. Jiao, Noncommutative weak Orlicz spaces and martingale inequalities, Studia Math. 204(3) (2011), 195-212. [4] T. N. Bekjan and Z. Chen, Interpolation and Φ-moment inequalities of noncommutative martingales, Probab. Theory Relat. Fields 152 (2012), 179-206. [5] C. Bennett and R. Sharpley, Interpolation of operators, Academic Press Inc., Boston, MA, 1988. [6] D. Boyd, Indices of function spaces and their relationship to interpolation, Canad. J. Math. 21 (1969), 1245-1254. 16 T. N. BEKJAN [7] A.P. Calder´on, Spaces between L1 and L∞ and the theorem of Marcinkiewicz, Studia Math. 26 (1966), 273-299. [8] M.D. Choi, A Schwarz inequality for positive linear maps on C ∗-algebras. Illinois J. Math. 18 (1974), 565-574. [9] A. Defant, M. Junge, Maximal theorems of Menchoff-Rademacher type in nonommutative Lp-spaces, J. Funct. Anal. 206 (2004), 322-355. [10] S. Dirksen and E. Ricard, Some remarks on noncommutative Khintchine inequalities, Bull. Lond. Math. Soc. 45(3) (2013), 618-624. [11] S. Dirksen, Noncommutative Boyd interpolation theorems. Trans. Amer. Math. Soc. 367 (2015) 4079-4110. [12] S. Dirksen, Weak-type interpolation for noncommutative maximal operators, J. Operator Theory 73(2) (2015), 515-532. [13] S. Dirksen, B. de Pagter, D. Potapov, and F. Sukochev, Rosenthal inequalities in noncommutative symmetric spaces, J. Funct. Anal. 261(2011), 2890-2925. [14] P. G. Dodds, T. K. Dodds and B. de Pager, Noncommutative Banach function spaces, Math. Z. 201 (1989), 583-587. [15] P. G. Dodds, T. K. Dodds and B. de Pager, Fully symmetric operator spaces, Integral Equations Operator Theory 15 (1992), 942-972. [16] P. G. Dodds, T. K. Dodds and B. de Pager, Noncommutative Kothe duality, Trans. Amer. Math. Soc. 339 (1993), 717-750. [17] P. G. Dodds, T. K. Dodds and F. A. Sukochev, On p-convexity and q- concavity in non-commutative sym- metric spaces, Integ. Equ. Oper. Theory 78(1) (2014), 91-114. [18] T. Fack and H. Kosaki, Generalized s-numbers of τ -measurable operators, Pac. J. Math. 123 (1986), 269-300. [19] M. Junge and T. Mei, Noncommutative Riesz transforms -- a probabilistic approach, Amer. J. Math. 132 (2010), 611-680. [20] M. Junge and Q. Xu, Noncommutative Burkholder/Rosenthal inequalities, Ann. Probab. 31 (2003), 948-995. [21] M. Junge and Q. Xu, Noncommutative maximal ergodic inequalities, J. Amer. Math. Soc. 20(2) (2007), 385-439. [22] S.G. Krein, J.I. Petunin and E.M. Semenov,Interpolation of linear operators, Translations of Mathematical Monographs, vol.54, AMS, 1982. [23] Pawel Kolwicz, Karol Le´snik and Lech Maligranda, Pointwise products of some Banach function spaces and factorization, J. Funct. Anal. 266 (2014), 616-659. [24] J. Lindentrauss and L. Tzafriri, Classical Banach space I, Springer-Verlag, Berlin, 1977. [25] J. Lindentrauss and L. Tzafriri, Classical Banach space II, Springer-Verlag, Berlin, 1979. [26] G.Ya. Lozanovskii, On some Banach lattices, Siberian Math. J. 10 (1969), 419-431. [27] M. Perrin, A noncommutative Davis' decomposition for martingales, J. Lond. Math. Soc. 80 (2009), 627-648. [28] G. Pisier and Q. Xu, Noncommutative Lp − spaces, In Handbook of the geometry of Banach spaces, Vol. 2 (2003), 1459-1517. [29] N. Randrianantoanina and L. Wu, Martingale inequalities in noncommutative symmetric spaces, J. Funct. Anal. 269 (2015), 2222-2253. [30] F. Sukochev, Completeness of quasi-normed symmetric operator spaces, Indag. Math. (N.S.) 25(2) (2014), 376-388. [31] Q. Xu, Analytic functions with values in lattices and symmetric spaces of measurable operators, Math. Proc. Camb. Phil. Soc. 109 (1991), 541-563. L. N. Gumilyov Eurasian national University, Astana 010000, Kazakhstan E-mail address: [email protected]
1201.4620	2	1201	2012-03-30T10:46:14	Cuntz-Li relations, Inverse semigroups and Groupoids	[ "math.OA" ]	In this paper we show that the universal C-algebra satisfying the Cuntz-Li relations is generated by an inverse semigroup of partial isometries. We apply Exel's theory of tight representations to this inverse semigroup. We identify the universal C-algebra as the C*-algebra of the tight groupoid associated to the inverse semigroup.	math.OA	math	CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS S. SUNDAR Abstract. In this paper we show that the universal C ∗-algebra satisfying the Cuntz- Li relations is generated by an inverse semigroup of partial isometries. We apply Exel's theory of tight representations to this inverse semigroup. We identify the universal C ∗-algebra as the C ∗-algebra of the tight groupoid associated to the inverse semigroup. 1. Introduction Let R be an integral domain with only finite quotients. Assume that R is not a field and let K be its field of fractions. We denote the set of non-zero elements in R (resp. K) by R× (resp. K ×). In [CL10], Cuntz and Li studied the C ∗-algebra, denoted Ar[R], on ℓ2(R) generated by the isometries induced by the multiplication and addition operations of the ring R. They showed that it is simple and purely infinite. It was also shown that this C ∗-algebra is the universal C ∗-algebra generated by isometries satisfying the relations reflecting the semigroup multiplication in R ⋊ R× and one more important relation satisfied by the range projections. Also it was shown that Ar[R] is Morita- equivalent to a crossed product of the form C0(R) ⋊ (K ⋊ K ×) where R is a locally compact Hausdorff space. For R = Z, R = Af is the space of finite adeles. Alternate approaches to the algebra Ar[R] were considered in [KLQ11], [BE10], and [Sun11]. In [KLQ11], the situation in [CL10] was abstracted. Consider a semidirect product N ⋊ H and a normal subgroup M of N. Let P := {a ∈ H : aMa−1 ⊂ M}. Then P is a semigroup. In [KLQ11],under certain hypotheses regarding the pair (G = N ⋊H, M), the crossed product algebra C0(N)⋊G was considered. Here N is the profinite completion of N with respect to the group topology induced by the neighbourhood base {aMa−1}a∈H at the identity. Let M be the closure of M in N. In [KLQ11], it was shown that the crossed product algebra C0(N ) ⋊ G is Morita-equivalent to the C ∗-algebra of the groupoid N ⋊ GM . In [KLQ11], It was shown that when H is abelian, C ∗(N ⋊ GM ) 2010 Mathematics Subject Classification. 46L05, 20M18. Key words and phrases. Inverse semigroups, Groupoids, Cuntz-Li relations, Tight representations. 1 2 S. SUNDAR is the universal C ∗-algebra generated by isometries satisfying the relations reflecting the semigroup multiplication in M ⋊ P and one more important relation among the range projections. They also obtained sufficient conditions which will ensure that the reduced C ∗-algebra C ∗ red(N ⋊ GM ) is simple and purely infinite. Our objective in this paper is to weaken the hypothesis that H is abelian. Instead we assume H = P P −1 = P −1P . This allows us to consider pairs like (Qn ⋊ GLn(Q), Zn). Also we start with the universal C ∗-algebra, denoted A[N ⋊ H, M], generated by isome- tries satisfying the Cuntz-Li relations (See Defn. 2.11.) We show that A[N ⋊ H, M] is generated by an inverse semigroup of partial isometries denoted by T . We show that A[N ⋊H, M] is isomorphic to the C ∗-algebra of the groupoid Gtight, considered in [Exe08], of the inverse semigroup T . We also identify the groupoid Gtight explicitly and show that Gtight is isomorphic to N ⋊ GM . The author had done a similar analysis for the Cuntz-Li algebra associated to the ring Z in [Sun11]. At the end of this paper, we prove a duality result analogous to the duality result obtained in [CL11]. 2. Semidirect products and the Cuntz-Li relations Let G = N ⋊ H be a semidirect product and let M be a normal subgroup of N. Let P := {a ∈ H : aMa−1 ⊂ M}. Then P is a semigroup containing the identity e. Assume that the following holds. (C1) The group H = P P −1 = P −1P . (C2) For every a ∈ P , the subgroup aMa−1 is of finite index in M. aMa−1 = {e} where e denotes the identity element of G. (C3) The intersection \a∈P Let U = {aMa−1 : a ∈ H}. In [KLQ11], the following conditions were required to be satisfied. (Cf. Section 2 in [KLQ11].) (E1) Given U, V ∈ U, there exists W ∈ U such that W ⊂ U ∩ V . (E2) If U, V ∈ U and U ⊂ V then U is of finite index in V . (E3) The intersection \U ∈U U = {e}. We claim that (E1) is equivalent to the condition H = P P −1. Assume (E1). Let a ∈ H be given. Then there exists c ∈ H such that a−1Ma ∩ M ⊇ cMc−1. Then c ∈ P and ac ∈ P . Note that a = (ac)c−1 ∈ P P −1. Thus we have H = P P −1. Now suppose H = P P −1. First note that for every a, b ∈ P , aP ∩ bP is non-empty. 2 with ai, bi ∈ P . Choose Now let c, d ∈ H be given. Write c = a1a−1 and d = b1b−1 2 CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 3 α, β ∈ P such that a1α = b1β. Let a := a1α. Then c−1a = a2α ∈ P . Similarly d−1a ∈ P . Hence aMa−1 ⊂ cMc−1 ∩ dMd−1. Thus (E1) holds. Given (E1), note that (E3) is equivalent to (C3). For if a ∈ H, there exists b ∈ P such that aMa−1 ∩ M ⊇ bMb−1. Thus for every a ∈ H, aMa−1 ⊇Tb∈P bMb−1. Hence TU ∈U U = Ta∈P aMa−1. Thus given (E1), (E3) is equivalent to (C3). Clearly (E2) is equivalent to (C2). Remark 2.1. In [KLQ11], the Cuntz-Li algebra associated to the pair ( Cf. defn 2.11) (N ⋊ H, M) was considered when H is abelian. (Cf. Hypothesis 9.2 and Theorem 9.11 in [KLQ11].) Here, we consider a slightly more general situation. We assume H = P −1P = P P −1. Remark 2.2. The condition H = P −1P = P P −1 is equivalent to saying that P generates H and P is right and left reversible i.e. given a, b ∈ P , the intersections P a ∩ P b and aP ∩ bP are non-empty. Cancellative semigroups which are right (or left) reversible are called Ore semigroups. For more details on Ore semigroups, we refer to [CP61]. A semigroup P is called right reversible (left reversible) if P a ∩ P b (if aP ∩ bP ) is non-empty for every a, b ∈ P . Throughout this article, whenever we write G = N ⋊ H and M is a normal subgroup of N, we assume that conditions (C1), (C2) and (C3) hold. For a ∈ P , let Ma = aMa−1. We will use this notation throughout. Lemma 2.3. Let G = N ⋊ H and M be a normal subgroup of N. Let N0 := [a∈P Then N0 is a subgroup of N and is invariant under conjugation by H. a−1Ma. Proof. First observe that N0 is closed under inversion. Let a, b ∈ P be given. Choose an element c in the intersection P a ∩ P b. Then a−1Ma ⊂ c−1Mc and b−1Mb ⊂ c−1Mc. Now it follows that N0 is closed under multiplication. Thus N0 is a subgroup of N. Obviously N0 is invariant under conjugation by P −1. Let a, b ∈ P be given. Since P is right reversible, there exists c, d ∈ P such that ab−1 = c−1d. Now observe that a(b−1Mb)a−1 = c−1(dMd−1)c ⊂ c−1Mc Thus it follows that N0 is closed under conjuga- tion by P . This completes the proof. ✷ Remark 2.4. As a consequence of Lemma 2.3, we may very well assume as in [KLQ11] that N = [a∈P a−1Ma. 4 S. SUNDAR Let us consider a few examples which fits the setup that we are considering. Example 2.5 ([CL10]). Let R be an integral domain such that for every non-zero m ∈ R, the ideal generated by m is of finite index in R. Assume that R is not a field. We denote the field of fractions of R by Q and the set of non-zero elements in Q by Q×. The multiplicative group Q× acts on Q by multiplication. Now let N := Q, H := Q× and M := R. Then P = R× where R× denotes the set of non-zero elements in R. Then conditions (C1)-(C3) hold for the pair (N ⋊ H, M). Example 2.6 ([KLQ11]). Let F be a finite group and consider the direct sum N := ⊕ZF . Then H := Z acts on N by shifting. Let M := ⊕NF be the normal subgroup of N. Then it is easily verifiable that the pair (N ⋊ H, M) satisfies the hypothesis (C1)-(C3). In the following two examples, we think of elements of Qn as column vectors. Example 2.7. Let A be a n × n integer dilation matrix. In other words, A is an n × n matrix with integer entries such that every complex eigen value of A has absolute value greater than 1. Note that A is invertible over Q and det(A) > 1. The matrix A acts on Qn by matrix multiplication and thus induces an action of Z on Qn. We let the generator 1 of Z act on Qn by 1.v = Av for v ∈ Qn. Let N := Qn, H := Z and M := Zn. Then P = N. Let us verify the hypothesis (C1)-(C3). (C1) Note that H is abelian and H = P P −1 = P −1P . (C2) For r ≥ 0, the index of ArZn is of finite index in Zn and in fact its index is det(A)r. (C3) Lemma 4.1 of [EaHR10] implies that the operator norm A−m converges to 0 as ArZn, then for every m ≥ 0, A−mv ∈ Zn. m tends to infinity. Thus if 0 6= v ∈ Thus we have 1 ≤ A−mv ≤ A−mv which is a contradiction. Thus (C3) holds. ∞\r=0 The case n = 1 and A = p where p is a prime number was discussed in [SL10a]. In the previous example, we can consider integer matrices other than dilation matrices. It is possible that (C3) is satisfied for an integer matrix A such that det(A) > 1 and Tr>0 ArZn = {0} without A being a dilation matrix. In fact we have the following nice characterisation of condition (C3) when n = 2. Lemma 2.8. Let A be a 2 × 2 matrix with integer entries. Assume that det(A) > 1. Then the following are equivalent. CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 5 ArZ2 is trivial. (2) Neither 1 nor −1 is an eigen value of A. If 1 is an eigen value of A then there exists a non-zero v ∈ Q2 such that Av = v. By clearing denominators, we can assume that ArZ2. Thus we have shown that 1 is not an eigen value of (1) The intersection \r≥0 Proof. Suppose Tr≥0 ArZ2 = {0}. v ∈ Z2. Then clearly v ∈\r≥0 Γ := Tr≥0 Γr. Since Γ ⊂ Γr ⊂ Z2, we have [Z2 : Γ] ≥ [Z2 : Γr] = det(A)r. Hence Γ Now assume that neither 1 nor −1 is an eigen value of A. Let Γr := ArZ2 and A. Similarly we can show −1 is not an eigen value of A. cannot be of finite index in Z2. This implies that Γ is of rank atmost 1. If Γ is rank 1 then there exists a non-zero v ∈ Z2 such that Γ = Zv. But A : Γ → Γ is a bijection. Thus it must either be multiplication by 1 or by −1. In other words, v is an eigen vector for A with eigen value 1 or −1. This is a contradiction. Thus Γ cannot be of rank 1 ✷. which in turn implies Γ = {0}. This completes the proof. 1 −2# has eigen values √3 − 1 and −√3 − 1. But A is not a The matrix A := "0 2 dilation matrix but still (C3) holds for A. Remark 2.9. It is not clear to the author whether (C3) can be characterised in terms of eigen values of the matrix in the higher dimensional case. Let us now consider an example where H is non-abelian. Example 2.10. Let N = Qn and H be a subgroup of GLn(Q) containing the non-zero scalars. Just as in Example 2.7, H acts on N by matrix multiplication. Let M = Zn. Then P consists of elements of H whose entries are integers. (C1) Let A ∈ H be given. Then there exists a non-zero integer m such that mA = Am ∈ P . Hence H = P P −1 = P −1P . (C2) For A ∈ P , the subgroup AZn is of finite index and its index is det(A). (C3) Since \m∈Z× mZn = {0}, it follows that \A∈P AZn = {0}. Definition 2.11. Let G := N ⋊H be a semidirect product and M be a normal subgroup of N such that (C1)-(C3) holds. We let A[N ⋊ H, M] be the universal C ∗-algebra generated by a set of isometries {sa : a ∈ P} and a set of unitaries {u(m) : m ∈ M} satisfying the 6 S. SUNDAR following relations. sasb = sab u(m)u(n) = u(mn) sau(m) = u(ama−1)sa u(k)eau(k)−1 = 1 Xk∈M/Ma where ea denotes the final projection of sa. Note that u(k)eau(k)−1 depends only on the coset k(Ma). Moreover if k1 and k2 lie in different cosets of Ma then u(k1)eau(k1)−1 and u(k2)eau(k2)−1 are orthogonal. For a ∈ P and m ∈ M, consider the operators Sa and U(m) on ℓ2(M)⊗ ℓ2(H) defined as follows Sa(δn ⊗ δb) : = δana−1 ⊗ δab U(m)(δn ⊗ δb) : = δmn ⊗ δb. Then sa → Sa and u(m) → U(m) gives a representation of A[N ⋊ H, M] on the Hilbert space ℓ2(M) ⊗ ℓ2(H). Let us call this representation the regular representation and denote its image by Ar[N ⋊ H, M]. Remark 2.12. It should be noted that the regular representation for integral domains considered in [CL10] is different from ours. 3. An Inverse semigroup for the Cuntz-Li relations The main aim of this section is to show that the C ∗-algebra A[N ⋊ H, M] is generated by an inverse semigroup of partial isometries. We begin with a lemma similar to Lemma 1 of Section 3.1 in [CL10]. Lemma 3.1. For every a, b ∈ P , one has ea = Xk∈M/Mb u(aka−1)eabu(aka−1)−1 CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 7 Proof. One has u(k)ebu(k)−1(cid:1)s∗ a u(aka−1)saebs∗ au(aka−1)−1 ea = sas∗ a = sa(cid:0) Xk∈M/Mb = Xk∈M/Mb = Xk∈M/Mb u(aka−1)eabu(aka−1)−1 This completes the proof. ✷ Let X be the linear span of {u(k)ebu(k)−1 : b ∈ P, k ∈ M}. Denote the set of projections in X by F . By Lemma 3.1 and the left reversibility of P , it follows that f ∈ F if and only if there exists b ∈ P such that f is in the linear span of {u(k)ebu(k)−1}. The following lemma is an immediate corollary of Lemma 3.1 and the fact that P is left reversible. Lemma 3.2. The set F is a commutative semigroup of projections. Moreover F is invariant under the maps x → sbxs∗ b for every b ∈ P and x → u(m)xu(m)−1 for every m ∈ M. Now we show that F is also invariant under conjugation by s∗ a for every a ∈ P . Lemma 3.3. Let a ∈ P be given. If f ∈ F , then s∗ is in the linear span of {u(k)ea−1cu(k)−1} where c is any element in aP ∩ bP . af sa ∈ F . Moreover, s∗ au(m)ebu(m)−1sa Proof. Let a ∈ P and f ∈ F be given. First observe that s∗ af sa is selfadjoint. Also (s∗ af sa)2 = s∗ = s∗ af sas∗ af sa af eaf sa = s∗ aeaf sa ( Since F is commutative ) = s∗ af sa af sa is a projection. Now to show that s∗ Thus s∗ af sa ∈ F , it is enough to consider the case when f = u(m)ebu(m)−1. Now let c ∈ aP ∩ bP and write c = aα = bβ with α, β ∈ P . 8 S. SUNDAR Let r1, r2,· · · , rn be distinct representatives of M/Mβ. Then by Lemma 3.1, it follows that s∗ au(m)ebu(m)−1sa = = nXi=1 nXi=1 s∗ au(mbrib−1)ebβu(mbrib−1)−1sa s∗ au(mbrib−1)eaαu(mbrib−1)−1sa au(mbrib−1)eaαu(mbrib−1)−1sa survives if and only if eaαu(mbrib−1)sa 6= 0 The term s∗ and that is if and only if eaαu(mbrib−1)eau(mbrib−1)−1 6= 0. But by Lemma 3.1 this happens precisely when there exists ti ∈ M/Mα such that mbrib−1 ≡ atia−1modMaα. Let A := {i : There exists ti such that mbrib−1 ≡ atia−1 mod Maα}. For every i ∈ A, choose ti such that mbrib−1 ≡ atia−1 mod Maα. Now we have s∗ au(m)ebu(m)−1sa = nXi=1 =Xi∈A =Xi∈A =Xi∈A =Xi∈A s∗ au(mbrib−1)eaαu(mbrib−1)−1sa s∗ au(mbrib−1)eaαu(mbrib−1)−1sa s∗ au(atia−1)eaαu(atia−1)−1sa u(ti)s∗ aeaαsau(ti)−1 u(ti)eαu(ti)−1 This completes the proof. ✷ Let us isolate the computation in the previous lemma in a remark. This will be used later. Remark 3.4. Let a, b ∈ P be given. Let c ∈ aP ∩ bP . Choose α and β in P such that c = aα = bβ. Conjugation by a sends Mα to Mc. Thus we get a map denoted πa α : M/Mα → M/Mc. Similarly conjugation by b gives a map πb β : M/Mβ → M/Mc. Note that both πa β are injective. Denote the quotient map M → M/Mc by qc. For m ∈ M, define α and πb Am := {r ∈ M/Mβ : qc(m)πb β(r) ∈ πa α(M/Mα)}. = s∗ s∗ a1u(m1)eu(m2)sa2s∗ = s∗ a1u(m1m2)u(m−1 a1u(m1m2)u(m−1 a1u(m1m2)s∗ a1s∗ = s∗ = s∗ b1u(n1)f u(n2)sb2 2 )eu(m2)s∗ 2 )eu(m2)s∗ αsαsa2s∗ αsαa2s∗ αsαu(m−1 2 )eu(m2)s∗ αscs∗ 1 )u(n1n2)sb2 βsβu(n1)f u(n−1 b1s∗ βb1sβu(n1)f u(n−1 csβu(n1)f u(n−1 1 )u(n1n2)sb2 1 )s∗ βsβu(n1n2)sb2 CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 9 Then the computation in Lemma 3.3 can be restated as follows s∗ au(m)ebu(m)−1sa = Xr∈Am u(cid:16)(cid:0)πa α(cid:1)−1(qc(m)πb β(r))(cid:17)eαu(cid:16)(cid:0)πa α(cid:1)−1(qc(m)πb β(r))(cid:17)−1 . Now we show that A[N ⋊ H, M] is generated by an inverse semigroup of partial isome- tries. : m, m′ ∈ M, a, a′ ∈ P, and f ∈ F}. Proposition 3.5. Let T := {s∗ Then T is an inverse semigroup of partial isometries containing 0. Moreover the set of projections in T coincides exactly with F . Also the linear span of T is a dense ∗- subalgebra of A[N ⋊ H, M]. au(m)f u(m′)sa′ Proof. The fact that T is closed under multiplication follows from the following cal- culation. Let a1, a2, b1, b2 ∈ P , m1, m2, n1, n2 ∈ M and e, f ∈ F be given. Choose c ∈ P b1 ∩ P a2 and write c as c = βb1 = αa2. Observe that αu(αm1m2α−1)(cid:0)sαu(m−1 αa1u(αm1m2α−1)(cid:0)sαu(m−1 αa1u(αm1m2α−1)(sαes∗ = s∗ = s∗ 2 )eu(m2)s∗ 2 )eu(m2)s∗ α(cid:1)ec(cid:0)sβu(n1)f u(n−1 α(cid:1)ec(cid:0)sβu(n1)f u(n−1 1 )s∗ 1 )s∗ β(cid:1)u(βn1n2β−1)sβsb2 β(cid:1)u(βn1n2β−1)sβb2 α)ec(sβ f s∗ β)u(βn1n2β−1)sβb2 2 )eu(m2) and f = u(n1)f u(n1)−1. The above calculation together with where e = u(m−1 Lemma 3.2 implies that T is closed under multiplication. Obviously T is closed under the involution ∗. Now let us show that every element of T is a partial isometry. Let v := s∗ au(m)f u(m′)sa′ be an element of T . Then vv∗ = s∗ ′ ′ )ea′ u(m a(cid:16)u(m)(cid:0)f u(m )−1f(cid:1)u(m)−1(cid:17)sa Now Lemma 3.2 and Lemma 3.3 implies that vv∗ ∈ F . Thus we have shown that every element of T is a partial isometry and the set of projections in T coincides with F . In other words T is an inverse semigroup. Since T is closed under multiplication and involution, it follows that the linear span of T is a ∗-algebra. Moreover T contains {sa : a ∈ P} and {u(m) : m ∈ M}. Thus the linear span of T is dense in A[N ⋊ H, M]. This completes the proof. ✷ 10 S. SUNDAR The following equality will be used later. Let a1, a2, b1, b2 ∈ P and m1, m2 ∈ M be given. Choose c ∈ P b1 ∩ P a2 and write c as c = βb1 = αa2. Now the computation in Proposition 3.5 gives the following equality (3.1) a1u(m1)sb1s∗ s∗ a2u(m2)sb2 = s∗ βa1u(βm1β−1)ecu(αm2α−1)sαb2 Remark 3.6. We also need the following fact. If v ∈ T , let us denote its image in the regular representation by V . Observe that v 6= 0 if and only if V 6= 0. This is clear for projections in T . Now let v ∈ T be a non-zero element. Then vv∗ ∈ F is non-zero. Thus V V ∗ 6= 0 which implies V 6= 0. In the remainder of this article, we reserve the letter T to denote the inverse semigroup in Proposition 3.5 and F to denote the set of projections in T . 4. Tight representations of inverse semigroups In this section, we show that the identity representation of T in A[N ⋊H, M] is tight in the sense of Exel and the C ∗-algebra of the tight groupoid associated to T is isomorphic to A[N ⋊H, M]. First let us recall the notion of tight characters and tight representations from [Exe08]. Definition 4.1. Let S be an inverse semigroup with 0. Denote the set of projections in S by E. A character for E is a map x : E → {0, 1} such that (1) the map x is a semigroup homomorphism, and (2) x(0) = 0. We denote the set of characters of E by cE0. We consider cE0 as a locally compact Hausdorff topological space where the topology on cE0 is the subspace topology induced from the product topology on {0, 1}E. For a character x of E, let Ax := {e ∈ E : x(e) = 1}. Then Ax is a nonempty set satisfying the following properties. (1) The element 0 /∈ Ax. (2) If e ∈ Ax and f ≥ e then f ∈ Ax. (3) If e, f ∈ Ax then ef ∈ Ax. Any nonempty subset A of E for which (1), (2) and (3) are satisfied is called a filter. Moreover if A is a filter then the indicator function 1A is a character. Thus there is a bijective correspondence between the set of characters and filters. A filter is called an ultrafilter if it is maximal. We also call a character x maximal or an ultrafilter if its CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 11 support Ax is maximal. The set of maximal characters is denoted bydE∞ and its closure in cE0 is denoted by [Etight. We refer to [Sun11] (Corollary 3.3) for the proof of the following lemma. Lemma 4.2. Let A be a unital C ∗-algebra and E ⊂ A be an inverse semigroup of projections containing {0, 1}. Suppose that E contains a finite set {e1, e2,· · · , en} of i=1 ei = 1. Then for every maximal character mutually orthogonal projections such thatPn x of E, there exists a unique ei for which x(ei) = 1. Let us recall the notion of tight representations of semilattices from [Exe08] and from [Exe09]. The only semilattice we consider is that of an inverse semigroup of projections or in other words the idempotent semilattice of an inverse semigroup. Also our semilattice contains a maximal element 1. First let us recall the notion of a cover from [Exe08]. Definition 4.3. Let E be an inverse semigroup of projections containing {0, 1} and Z be a subset of E. A subset F of Z is called a cover for Z if given a non-zero element z ∈ Z there exists an f ∈ F such that f z 6= 0. A cover F of Z is called a finite cover if F is finite. The following definition is actually Proposition 11.8 in [Exe08] Definition 4.4. Let E be an inverse semigroup of projections containing {0, 1}. A representation σ : E → B of the semilattice E in a Boolean algebra B is said to be tight if σ(0) = 0 and given e 6= 0 in E and for every finite cover F of the interval [0, e] := {x ∈ E : x ≤ e}, one has supf ∈F σ(f ) = σ(e). Let A be a unital C ∗ algebra and S be an inverse semigroup containing {0, 1}. Denote the set of projections in S by E. Let σ : S → A be a unital representation of S as partial isometries in A. Let σ(C ∗(E)) be the C ∗−subalgebra in A generated by σ(E). Then σ(C ∗(E)) is a unital, commutative C ∗−algebra and hence the set of projections in it is a Boolean algebra which we denote by Bσ(C ∗(E)). We say the representation σ is tight if the representation σ : E → Bσ(C ∗(E)) is tight. The proof of the following lemma can be found in [Sun11] (Lemma 3.6, page 7). Lemma 4.5. Let X be a compact metric space and E ⊂ C(X) be an inverse semi- group of projections containing {0, 1}. Suppose that for every finite set of projections 12 S. SUNDAR {f1, f2,· · · , fm} in E, there exists a finite set of mutually orthogonal non-zero projec- tions {e1, e2,· · · , en} in E and a matrix (aij) such that nXi=1 ei = 1 fi =Xj aijej. Then the identity representation of E in C(X) is tight. As in [Sun11], we prove that the identity representation of T in A[N ⋊ H, M] is tight. Proposition 4.6. The identity representation of T in A[N ⋊ H, M] is tight. Proof. We apply Lemma 4.5. Let {f1, f2,· · · , fn} be a finite set of projections in T . By definition, given i there exists ai ∈ P such that fi is in the linear span of {u(k)eaiu(k)−1}. aiP . By Lemma 3.1, it follows that for every i, fi is in the linear span of Let c ∈ {u(k)ecu(k)−1 : k ∈ M/cMc−1}. Appealing to Lemma 4.5, we can conclude that the identity representation of T in A[N ⋊ H, M] is tight. This completes the proof. n\i=1 ✷ Now we show that A[N ⋊ H, M] is isomorphic to the C ∗-algebra of the groupoid Gtight associated to T . For the convenience of the reader, we recall the construction of the groupoid Gtight, considered in [Exe08], associated to an inverse semigroup with 0. Let S be an inverse semigroup with 0 and let E denote its set of projections. Note that S acts on cE0 partially. For x ∈cE0 and s ∈ S, define (x.s)(e) = x(ses∗). Then • The map x.s is a semigroup homomorphism, and • (x.s)(0) = 0. But x.s is nonzero if and only if x(ss∗) = 1. For s ∈ S, define the domain and range of s as Ds : = {x ∈cE0 : x(ss∗) = 1} Rs : = {x ∈cE0 : x(s∗s) = 1} Note that both Ds and Rs are compact and open. Moreover s defines a homeomorphism from Ds to Rs with s∗ as its inverse. Also observe that [Etight is invariant under the action of S. CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 13 Consider the transformation groupoid Σ := {(x, s) : x ∈ Ds} with the composition and the inversion being given by: (x, s)(y, t) : = (x, st) if y = x.s (x, s)−1 : = (x.s, s∗) Define an equivalence relation ∼ on Σ as (x, s) ∼ (y, t) if x = y and if there exists an e ∈ E such that x ∈ De for which es = et. Let G = Σ/ ∼. Then G is a groupoid as the product and the inversion respects the equivalence relation ∼. Now we describe a topology on G which makes G into a topological groupoid. For s ∈ S and U an open subset of Ds, let θ(s, U) := {[x, s] : x ∈ U}. We refer to [Exe08] for the proof of the following proposition. We denote θ(s, Ds) by θs. Proposition 4.7. The collection {θ(s, U) : s ∈ S, U open in Ds} forms a basis for a topology on G. The groupoid G with this topology is a topological groupoid whose unit space can be identified with cE0. Also one has the following. (1) For s, t ∈ S, θsθt = θst, (2) For s ∈ S, θ−1 s = θs∗, (3) For s ∈ S, θs is compact, open and Hausdorff, and (4) The set {1θs : s ∈ T} generates the C ∗-algebra C ∗(G). We define the groupoid Gtight to be the reduction of the groupoid G to [Etight. In [Exe08], it is shown that the representation s → 1θs ∈ C ∗(Gtight) is tight and any tight representation of S factors through this universal one. Proposition 4.8. Let T be the inverse semigroup considered in Proposition 3.5. Denote the tight groupoid associated to T by Gtight. Then A[N ⋊H, M] is isomorphic to C ∗(Gtight). Proof. Let ta and v(m) be the images of sa and u(m) in C ∗(Gtight). By Proposition 4.6 and by the universal property of Gtight, it follows that there exists a homomorphism ρ : C ∗(Gtight) → A[N ⋊ H, M] such that ρ(ta) = sa and ρ(v(m)) = u(m). Given a ∈ P , the projections {u(k)eau(k)−1 : k ∈ M/Ma} cover the projections in T . Since the representation of T in C ∗(Gtight) is tight, it follows that Xk∈M/Ma v(k)(tat∗ a)v(k)−1 = 1 Now the universal property of A[N ⋊ H, M] implies that there exists a homomorphism σ : A[N ⋊ H, M] → C ∗(Gtight) such that σ(sa) = ta and σ(u(m)) = v(m). It is then clear that σ and ρ are inverses of each other. This completes the proof. ✷ 14 S. SUNDAR We identify the groupoid Gtight explicitly in the rest of the article. 5. Tight characters of the inverse semigroup T In this section, we determine the tight characters of the inverse semigroup T defined in Proposition 3.5. Let M :=(cid:8)(ra) ∈Ya∈P M/Ma : rab ≡ ra mod Ma(cid:9) We give M the subspace topology induced from the product topology onQa∈P M/Ma. Here the finite group M/Ma is given the discrete topology. Then M is a compact, Hausdorff topological space. Moreover M is a topological group. Note that M embeds naturally into M via the imbedding r → (ra := r). The map r → (ra := r) is an imbedding since we have assumed thatTa∈P Ma is trivial. For b ∈ P and k ∈ M, the set Ub,k := {(ra) ∈ M : rb ≡ k mod Mb} is an open set. Moreover the collection {Ub,k : b ∈ P, k ∈ M} forms a basis for M . If k ∈ M then clearly k ∈ Ub,k for any b ∈ P . As a consequence, M is dense in M . For r ∈ M , let Ar := {f ∈ F : f ≥ u(ra)eau(ra)−1 for some a ∈ P}. In the next lemma, we show that for every r ∈ M, Ar is an ultrafilter and all ultrafilters are of this form. Lemma 5.1. For r ∈ M, Ar is an ultrafilter. Moreover any ultrafilter is of the form Ar for some r ∈ M . Proof: Let r ∈ M be given. First let us show that Ar is a filter. Clearly 0 /∈ Ar. Also if f1 ≥ f2 and f2 ∈ Ar then f1 ∈ Ar. Now suppose that f1, f2 ∈ Ar. Then there exists a1, a2 ∈ P such that fi ≥ u(rai)eaiu(rai)−1 for i = 1, 2. Choose c ∈ a1P ∩ a2P . Then by Lemma 3.1, it follows that ec ≤ eai for i = 1, 2. Since r ∈ M , it follows that rc ≡ rai mod Mai for i = 1, 2. Now observe that f1f2 ≥ u(ra1)ea1u(ra1)−1u(ra2)ea2u(ra2)−1 = u(rc)ea1u(rc)−1u(rc)ea2u(rc)−1 = u(rc)ea1ea2u(rc)−1 ≥ u(rc)ecu(rc)−1 Thus f1f2 ∈ Ar. Thus we have shown that Ar is a filter. CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 15 Now we show Ar is maximal. Let A be a filter which contains Ar. Consider an element f ∈ A. By definition there exists a ∈ P and scalars αk ∈ {0, 1} such that f = Xk∈M/Ma αku(k)eau(k)−1. But both f and u(ra)eau(ra)−1 belong to A and hence their product belongs to A. Thus the product f u(ra)eau(ra)−1 is non-zero. This implies that αra = 1. Thus we have f ≥ u(ra)eau(ra)−1 or in other words f ∈ Ar. Hence A = Ar. This proves that Ar is maximal. Let A be an ultrafilter. By Lemma 4.2, it follows that for every a ∈ P , there exists a unique ra ∈ M/Ma such that u(ra)eau(ra)−1 ∈ A. Let r := (ra). We claim that r ∈ M . Let a, b ∈ P be given. By Lemma 3.1, we have (5.2) u(raaka−1)eabu(rakak−1)−1 u(ra)eau(ra)−1 = Xk∈M/Mb Since A is a filter containing u(ra)eau(ra)−1 and u(rab)eabu(rab)−1, it follows that their product is non-zero. This fact together with Equation 5.2 implies that there exists k ∈ M, such that rab ≡ ra(aka−1) mod Mab. Thus rab ≡ ra mod Ma for every a, b ∈ P . As a result, we have r ∈ M . Since A is a filter it follows that Ar ⊂ A. We have already proved that Ar is maximal. Thus A = Ar. This completes the proof. ✷ The following proposition identifies the tight characters of T . Proposition 5.2. The map M : r → Ar ∈ [Ftight is a homeomorphism. Proof. It is clear from the definition that r → Ar is one-one. Let us denote this map by φ. We show φ is continuous. Consider a net rα in M converging to r. We denote the indicator function of a set A by 1A. Let f ∈ F be given. Then there exists a ∈ P and scalars αk such that a = ra eventually, it follows that 1Arα (f ) converges to 1Ar (f ). This shows that Now Lemma 5.1 implies that φ has range cF∞. Since M is compact, it follows that bF∞ is compact and hence closed. Thus cF∞ = [Ftight. Thus φ : M → bF∞ is one-one, onto αku(k)eau(k)−1 f =Xk 1Arα (f ) =Xk αkδrα a ,k Then we have Since rα r → Ar is continuous. 16 S. SUNDAR and continuous. Since M is compact, it follows that φ is in fact a homeomorphism. This completes the proof. ✷ From now on we will simply denote Ar by r and 1Ar(f ) by r(f ). 6. The groupoid Gtight of the inverse semigroup T In this section, we will identify the tight groupoid Gtight associated to the inverse semigroup. Throughout this section, we assume N =Sa∈P a−1Ma. By Remark 2.4, we can very well assume this. There is another natural groupoid which arises out of the following construction. For every a ∈ P , the co-isometry s∗ a will give rise to an injection on M and the unitary u(m) for m ∈ M will act as a bijection on M. Thus we get an action of the semigroup M ⋊ P , as injections, on M . Now the space M can be enlarged to a space N and the action of M ⋊ P can be dilated to get an action of G = N ⋊ H on N. We can then consider the transformation groupoid N ⋊ G. But the unit space of Gtight is M . Thus we restrict the transformation groupoid N ⋊ G to M and prove that it is isomorphic to Gtight. This dilation procedure has appeared in several works [See [Lac00], [SL10b] ]. The basic principle goes back to [Ore31]. First let us explain the action of M ⋊ P on M . The action of M on M is by left multiplication as M is a subgroup of M. Let a ∈ P and r ∈ M be given. For b ∈ P , choose c ∈ aP ∩ bP and write c as c = aα = bβ. We will use the notation as in Remark 3.4. Note that Mc ⊂ Mb and we denote the induced quotient map M/Mc → M/Mb by qb,c. Define mb = qb,c(πa α(rα)). First let us show that mb depends only on a and b and not on the choices made. Suppose c1 = aα1 = bβ1 and c2 = aα2 = bβ2. Choose γ1, γ2 ∈ P such that α1γ1 = α2γ2. Note that this implies c1γ1 = c2γ2. Now we have qb,ciπa αi(cid:0)qαi,αiγi(rαiγi)(cid:1)(cid:17) αi(rαi) = qb,ci(cid:16)πa αiγi(rαiγi)(cid:1)(cid:17) = qb,ci(cid:16)qci,ciγi(cid:0)πa αiγi(rαiγi)(cid:1). = qb,ciγi(cid:0)πa Note that the right hand side is constant for i = 1, 2. Thus we have qb,c1(πa α1(rα1)) = qb,c2(πa α2(rα2)). This shows that mb is well defined. We leave it to the reader to check that m = (mb) ∈ M . CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 17 On M, the action of P is the usual conjugation. From now on, we denote the element m by ara−1. This way P acts on M injectively and continuously. This action of P together with the left multiplication action of M defines an action of M ⋊ P on M (as injective,continuous transformations). We leave the details to the reader. Lemma 6.1. For a ∈ P , the kernel of the projection map M ∋ (yb) → ya ∈ M/Ma is aM a−1. Proof. By definition, it follows that aM a−1 is in the kernel of the ath projection. Now let y = (yb) be such that ya = 1. Since M is dense in M, there exists a sequence yn ∈ M such that yn → y in M. As M/Ma is finite, we can without loss of generality assume that yn ∈ Ma for every n. Thus there exists xn ∈ M such that yn = axna−1. But M is compact. Thus, by passing to a subsequence if necessary, we can assume that xn converges to an element say x ∈ M. Since conjugation by a is continuous, it follows that yn = axna−1 converges to axa−1. But yn converges to y. Thus axa−1 = y. This completes the proof. ✷ Now let us explain the dilation procedure that we promised at the beginning of this section. Consider the set M × P and define a relation on M × P by (x, a) ∼ (y, b) if there exists α, β ∈ P such that αa = βb and αxα−1 = βyβ−1. We leave the following routine checking to the reader. (1) The relation ∼ is an equivalence relation. We denote the equivalence class con- taining (x, a) by [(x, a)]. (2) Let N := M × P/ ∼. Then N is a group. The multiplication on N is defined as follows. For a, b ∈ P , choose α and β such that αa = βb. Then [(x, a)][(y, b)] = [(αxα−1βyβ−1, αa)] The identity element of N is [(e, e)] where (e, e) is the identity element of M × P and the inverse of [(x, a)] is [(x−1, a)]. (3) The group N is a locally compact Hausdorff topological group when N is given the quotient topology. Here P is given the discrete topology. (4) The map M ∋ x → [(x, e)] ∈ N is a topological embedding. Thus M can be viewed as a subset of N. Moreover M is a compact open subgroup of N. (5) The map N ∋ a−1ma → [(m, a)] ∈ N is an embedding. When N is viewed as a subset of N via this embedding, N is dense in N . Also N ∩ M = M. (6) Let a ∈ P be given. Define a map φa : N → N as follows. Given [(x, b)] ∈ N , choose α, β ∈ P such that αa = βb. Define φa([(x, b)]) = [βxβ−1, α)]. One 18 S. SUNDAR given by φ−1 checks that φa is well defined. Moreover for a ∈ P , φa is a homeomorphism with φ−1 a [(x, b)] = [(x, ba)]. Note that φa restricted to N is the usual a conjugation. Also φaφb = φab for a, b ∈ P . For m ∈ M , define ψm : N → N as ψm([(x, a)]) = [(ama−1x, a)]. That is ψm is just left multiplication by m. One also has the following commutation relation. For a ∈ P and m ∈ M, φaψm = ψama−1 φa. (7) Since we have assumed that N = Sa∈P a−1Ma, it follows that any element of g ∈ G = N ⋊ H can be written as g = a−1mb with a, b ∈ P and m ∈ M. The map a−1mb → φ−1 If h = a−1b ∈ H and x ∈ N, we denote φ−1 a φb(x) as hxh−1. If n = a−1ma and x ∈ N, we denote φ−1 a ψmφb is well defined and defines an action of G on N . a ψmφa(x) as nx. on which H acts by group homomorphism. Suppose that K is a compact open (9) Universal Property: Let L be a locally compact Hausdorff topological group (8) Note that N =Sa∈P a−1M a. subgroup of L which is invariant under P and L =Sa∈P a−1K. If φ : M → K is a P -equivariant continuous bijection then the map N ∋ a−1xa → a−1.φ(x) ∈ L is a topological isomorphism and is H -equivariant. Remark 6.2. It is not difficult to show by using (9) that N is the pro-finite completion of N when N is given the topology induced by the neighbourhood base {aMa−1 : a ∈ H} at the identity. In [KLQ11], the pro-finite completion model of N is used. When considering transformation groupoids, we consider only right actions of groups and thus we change the above left action of G on N to a right action simply by defining x.g = g−1x for x ∈ N and g ∈ G. Now consider the transformation groupoid N ⋊ G and restrict it to M . We show that the groupoid Gtight of the inverse semigroup T is isomorphic to the groupoid N ⋊GM i.e. to the transformation groupoid N ⋊G restricted to the unit space M . We will start with two lemmas which will be extremely useful to prove this. Lemma 6.3. If a−1 1 m1b1 = a−1 2 m2b2 then s∗ a1u(m1)sb1 = s∗ a2u(m2)sb2. Proof. Suppose a−1 1 m1b1 = a−1 2 m2b2. Then a−1 1 m1a1 = a−1 2 m2a2 and a−1 1 b1 = a−1 2 b2. 2 . Hence Choose β1, β2 ∈ P such that β1b1 = β2b2. Then a1a−1 2 = β−1 1 β2 = b1b−1 CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 19 β1m1β−1 1 = β2m2β−1 2 . Now observe that a1u(m1)sb1 = s∗ s∗ = s∗ a1u(m1)s∗ β1sβ1sb1 β1u(β1m1β−1 a1s∗ β1a1u(β1m1β−1 β2a2u(β2m2β−1 a2s∗ β2u(β2m2β−1 a2u(m2)s∗ β2sβ2sb2 a2u(m2)sb2 = s∗ = s∗ = s∗ = s∗ = s∗ 1 )sβ1b1 1 )sβ1b1 2 )sβ2b2 2 )sβ2b2 This completes the proof. ✷ Lemma 6.4. In Gtight, [(r, s∗ au(m)f u(n)sb)] = [(r, s∗ au(mn)sb)]. a)][r.s∗ a, u(m)f u(n)sb] = [(r, s∗ Proof. First observe that [(r, s∗ au(m)f u(n)sb)]. Thus it is enough to consider the case when a is the identity element of P . Now let s = u(m)f u(n)sb, t = u(mn)sb and e = u(m)f u(m)−1. Observe that s = et. Thus ss∗ = ett∗e. Hence r(ss∗) = 1 implies r(e) = 1 and r(tt∗) = 1. Moreover es = s = et. Thus [(r, s)] = [(r, t)]. This completes the proof. ✷ Now we can state our main theorem. Theorem 6.5. Let φ : N ⋊ GM → Gtight be the map defined by Then φ is a topological groupoid isomorphism. φ(cid:0)(x, a−1mb)(cid:1) = [(x, s∗ au(m)sb)]. Proof. First let us show that φ is well defined. Let (x, a−1mb) ∈ N ⋊ GM . Then by definition, there exists y ∈ M such that m−1axa−1 = byb−1. Choose α and β in P such that c := aα = bβ. By definition, this means that πa β(yβ). Now Remark 3.4 implies that α(xα) ≡ qc(m)πb Hence x(s∗ au(m)ebu(m)−1sa) = 1. Thus we have shown that φ is well-defined. s∗ au(m)ebu(m)−1sa ≥ u(xα)eαu(xα)−1. Before we show φ is a surjection, let us show that if [(x, s∗ au(m)sb)] ∈ Gtight then (x, a−1mb) ∈ N ⋊ GM . To that effect, assume that x(s∗ au(m)ebu(m)−1sa) = 1. Choose c ∈ aP ∩ bP and write c = aα = bβ. By Remark 3.4, it follows that there exists y ∈ M/Mβ such that qc(m−1)πa β(y). This implies that the bth co-ordinate of α(xα) = πb 20 S. SUNDAR m−1axa−1 is 1 i.e. the identity element of M/Mb. Now Lemma 6.1 implies that there exists z ∈ M such that m−1axa−1 = bzb−1. Hence (x, a−1mb) ∈ N ⋊ GM . Surjectivity is then an immediate consequence of Lemma 6.4. Now we show φ is injective. Suppose [(x, s∗ a2u(m2)sb2)]. Then there exists a projection e ∈ F such that 0 6= e(s∗ a2u(m2)sb2). We can without loss of generality assume that e = u(rc)ecu(rc)−1. By Remark 3.6 and by reading the above equality in the regular representation, we immediately obtain a−1 2 b2 and a−1 a1u(m1)sb1)] = [(x, s∗ a1u(m1)sb1) = e(s∗ 2 m2b2. This implies that φ is injective. 1 m1b1 = a−1 1 b1 = a−1 Now let us show that φ is a groupoid morphism. First we show that φ preserves the range and source. By definition, φ preserves the range. Observe that φ is continuous and this is a direct consequence of Proposition 5.2. Let γ = (x, a−1mb) ∈ N ⋊ GM . Since M is dense in M there exists a sequence xn ∈ M such that xn converges to x. Moreover the action of G on N is continuous and M is compact and open. Thus we can assume that (xn, a−1mb) ∈ N ⋊ GM for every n. By definition, there exists y ∈ M such that axa−1 = mbyb−1. Also let yn be such that axna−1 = mbynb−1. To keep things clear, if z ∈ M, we denote the character determined by z as ξz. Let v := s∗ au(m)sb. Now if can show that ξxn.v = ξyn then it will follow from continuity of φ that ξx.v = ξy. Thus we only need to show that s(φ(γ)) = φ(s(γ)) for γ = (x, a−1mb) with x ∈ M. Now let (x, a−1mb) ∈ N ⋊ GM with x ∈ M. Then there exists y ∈ M such that axa−1 = mbyb−1. Let v = s∗ au(m)sb. To show ξx.v = ξy, as ξy is maximal, it is enough to show that the support of ξy is contained in ξx.v. Again it is enough to show that u(y)ecu(y)−1 is in the support of ξx.v. Choose α, β such that aα = bcβ. Note that vu(y)ecu(y)−1v∗ = s∗ au(m)sbu(y)ecu(y)−1s∗ au(mbyb−1)sbecs∗ au(axa−1)ebcu(axa−1)−1sa bu(m)−1sa bu(mbyb−1)−1sa = s∗ = s∗ aebcsau(x)−1 aebcβsau(x)−1 aeaαsau(x)−1 = u(x)s∗ ≥ u(x)s∗ = u(x)s∗ = u(x)eαu(x)−1 ∈ supp(ξx) Hence u(y)ecu(y)−1 is in the support of ξx.v. Thus we have shown that ξx.v = ξy. This proves that φ preserves the source. CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 21 Now we show φ preserves multiplication. Let γ1 = (x1, a−1 2 m2b2). Since φ preserves the range and source, it follows that γ1 and γ2 are composable if and only if φ(γ1) and φ(γ2) are composable. Choose α, β ∈ P such that βb1 = αa2. . Now 1 m1b1) and γ2 = (x2, a−1 φ(γ1)φ(γ2) = [(x1, s∗ a1u(m1)sb1s∗ a2u(m2)sb2)] = [(x1, s∗ = [(x1, s∗ βa1u(βm1β−1)eαa2u(αm2α−1)sαb2)](cid:0) by Eq. 3.1(cid:17) βa1u(βm1β−1αm2α−1)sαb2)](cid:0) by Remark 6.4(cid:17) = φ(γ1γ2). It is easily verifiable that φ preserves inversion. For an open subset U of M and g = a−1mb, consider the open set θ(U, g) := {x ∈ M : x.g ∈ M} The collection {θ(U, g)} forms a basis for N⋊GM . Moreover φ(θ(U, g)) = θ(U, s∗ au(m)sb). Thus φ is an open map. Thus we have shown that φ is a homeomorphism. This completes the proof. Corollary 6.6. The algebra A[N ⋊ H, M] is isomorphic to C ∗(N ⋊ GM ). Proof. This follows from Theorem 6.5 and Proposition 4.8. ✷ ✷ 7. Simplicity of Ar[N ⋊ H, M] Let us recall a few definitions from [AD97]. Let G be an r-discrete groupoid and we denote its unit space by G0. The relation ∼ defined by x ∼ y if and only if there exists γ ∈ G such that s(γ) = x and r(γ) = y is an equivalence relation on G0. A subset E ⊂ G0 is said to be invariant if given x ∈ E and y ∼ x then y ∈ E. For x ∈ G, let G(x) := {γ ∈ G : s(γ) = r(γ) = x} be the isotropy group of x. A subset S ⊂ G is said to be a bi-section if the range and source maps restricted to S are one-one. If S is a bisection, let αS : r(S) → s(S) be defined by αS := s ◦ r−1. The groupoid G is said to be • minimal if the only non-empty, open invariant subset of G0 is G0. • topologically principal if the set of x ∈ G0 for which G(x) = {x} is dense in G0. • locally contractive if for every non-empty open subset U of G0, there exists an open subset V ⊂ U and an open bisection S with V ⊂ s(S) and αS−1(V ) not contained in V . 22 S. SUNDAR Conjugation by P on M gives rise to a semigroup homomorphism from P to the semigroup of injective maps on M. In [KLQ11], the action of P on M is called an effective action if the above semigroup homomorphism is injective i.e. given h ∈ H with h 6= 1, then there exists s ∈ M such that hsh−1 6= s. In [KLQ11], the following facts were proved about the transformation groupoid N ⋊ G. (1) The groupoid N ⋊ G is minimal and locally contractive. (2) The groupoid N ⋊ G is topologically principal if and only if P acts effectively on M. (3) Thus the reduced C ∗-algebra C ∗ red(N ⋊ G) is simple and purely infinite if P acts effectively on M. [Refer to [AD97]]. Analogous statements hold for the groupoid Gtight associated to the inverse semigroup T . Remark 7.1. In [KLQ11], only the if part (in (2)) was proved. But then the other direction i.e. if N ⋊ G is topologically principal then P acts effectively on M is easy to verify. Also note that M is a closed subset of N which meets each G orbit of N . Moreover M is open as well. Hence by appealing to Example 2.7 in [MRW87] , we conclude that C ∗(N ⋊ G) and C ∗(N ⋊ GM ) are Morita-equivalent. We end this section by showing that Ar[N ⋊ H, M] is isomorphic to the reduced C ∗- algebra C ∗ red(Gtight). Proposition 7.2. Let G := N ⋊ GM . Then the reduced C ∗-algebra of the groupoid G is isomorphic to Ar[N ⋊ H, M]. Proof. Let e be the identity element of M . Define Ge := {γ ∈ G : r(γ) = e}. Then Ge := {(e, hm) : m ∈ M, h ∈ H}. Thus L2(Ge) can be identified with ℓ2(M) ⊗ ℓ2(H). Consider the representation πe of C ∗ red(G) on L2(Ge) defined as follows. For f ∈ Cc(G), define πe(f ) by the following formula. (πe(f )(ξ))(γ) := Xγ1∈Ge f (γ−1γ1)ξ(γ1) Since M is dense in M , it follows that the largest open invariant set not containing e is the empty set. Hence πe is faithful. For a ∈ P and m ∈ M, we let Sa and U(m) be the images of sa and u(m) in C ∗ red(G). Let {δm ⊗ δb : m ∈ M, b ∈ H} be the canonical basis of ℓ2(M) ⊗ ℓ2(H). Consider the CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 23 unitary operator V on ℓ2(M) ⊗ ℓ2(H) defined by V (δm ⊗ δb) := δm−1 ⊗ δb−1 For a ∈ P and k ∈ M, we leave it to the reader to check the following equality. V πe(Sa)V ∗(δm ⊗ δb) = δama−1 ⊗ δab V πe(U(k))V ∗(δm ⊗ δb) = δkm ⊗ δb Since {Sa : a ∈ P} and {U(k) : k ∈ M} generate C ∗ isomorphic to Ar[N ⋊ H, M]. This completes the proof. red(G), it follows that C ∗ red(G) is ✷ We now show that Corollary 6.6 and Proposition 7.2 can also be expressed in terms of crossed products as in [KLQ11]. We need to digress a bit before we do this. Let G be an r-discrete, locally compact and Hausdorff groupoid. Let Y ⊂ G0 be a compact open subset of the unit space. Assume that Y meets each orbit of G0. Let GY : = {γ ∈ G : s(γ) ∈ Y } GY Y : = {γ ∈ G : s(γ), r(γ) ∈ Y } Since Y is clopen, it follows that GY and GY Y are clopen. Thus if f ∈ Cc(GY ), then f can be extended to an element in Cc(G) by declaring its value to be zero outside GY . Thus we have the inclusion Cc(GY ) ⊂ Cc(G). Similarly, we have the inclusion Cc(GY Y ) ⊂ Cc(GY ). The algebra Cc(GY Y ) is a ∗-subalgebra of Cc(G). The space Cc(GY ) is a pre-Hilbert Cc(GY Y ) ⊂ C ∗(GY and the right multiplication given by Y ) module with the inner product < f1, f2 > (γ) = Xγ1γ2=γ (f.g)(γ) = Xγ1γ2=γ f1(γ−1 1 )f2(γ2) for γ ∈ GY Y , f1, f2 ∈ Cc(GY ) f (γ1)g(γ2) for γ ∈ GY , f ∈ Cc(GY ), g ∈ Cc(GY Y ) Moreover there is left action of Cc(G) on Cc(GY ) and it is given by (f.φ)(γ) = (f ∗ φ)(γ) = Xγ1γ2=γ f (γ1)φ(γ2) for γ ∈ GY , f ∈ Cc(G) and φ ∈ Cc(GY ). Now Theorem 2.8 and Example 2.7 of [MRW87] implies the following. The "com- Y ) imprimitivity bimodule pletion" of Cc(G)-Cc(GY implementing a strong Morita equivalence between C ∗(G) and C ∗(GY Y ). Y ) bimodule Cc(GY ) is a C ∗(G)-C ∗(GY red(GY Proof. Let f ∈ Cc(GY 24 S. SUNDAR Let us denote the completion of Cc(GY ) by E. For x, y ∈ E, let θx,y be the compact operator on E defined by θx,y(z) = x < y, z >. For x ∈ E, the operator norm of θx,x is x2. The following proposition has also appeared in [Li12]. (See Lemma 5.18 in [Li12].) The proof is exactly as in [Li12]. We include the proof for the sake of completeness. Proposition 7.3. The inclusion Cc(GY from C ∗(GY embedding from C ∗ Y ) to C ∗(G). Also the inclusion Cc(GY Y ) ⊂ Cc(G) extends to an isometric embedding Y ) ⊂ Cc(G) extends to an isometric Y ) to C ∗ red(G). Y ) be given. Consider f as an element of Cc(GY ) ⊂ E. Then θf,f restricted to Cc(GY ) is just multiplication by f ∗ f ∗. Since E is a C ∗(G)-C ∗(GY Y ) imprimitivity bimodule, it follows that f2 C ∗(G) = f ∗ f ∗C ∗(G) E = θf,f = f2 = f ∗ ∗ fC ∗(GY = f2 C ∗(GY Y ) Y ) For x ∈ G0, let G(x) := r−1(x). Consider ℓ2(G(x)) and let {δγ : γ ∈ G(x)} be the standard orthonormal basis. Consider the representation πx of Cc(G) on ℓ2(G(x)) defined by (7.3) πx(f )(δγ) = Xα∈G(x) f (α−1γ)δα. The reduced C ∗-algebra C ∗ fred = supx∈G0πx(f ). (We refer the reader to [Ren09].) red(G) is the completion of Cc(G) under the norm . given by := {γ ∈ G(x) : s(γ) ∈ Y }. If x ∈ Y , let πY Y Y ) defined by the same formula as in Eq. 7.3. Now observe the following. x be the representation of Cc(GY Y ) Let G(x) on ℓ2(G(x) (1) Let γ0 ∈ G be such that s(γ0) = x and r(γ0) = y. Then U : ℓ2(G(x)) → ℓ2(G(y)) defined by U(δγ) = δγ0γ is a unitary. Moreover Uπx(.)U ∗ = πy(.). (2) Since Y meets each orbit of G0, it follows from (1) that for f ∈ Cc(G), fred = sup x∈Y πx(f ). (3) If x ∈ Y , then write ℓ2(G(x)) as ℓ2(G(x)) = ℓ2(G(x) Y ), we have πx(f ) = πY decomposition, for f ∈ Cc(GY Y ) ⊕ (ℓ2(G(x) x (f ) ⊕ 0. Y ))⊥. With this CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 25 Now the above three observations imply that for f ∈ Cc(GY Y ), fC ∗ red(GY This completes the proof. Y ) = fC ∗ red(G). ✷ Remark 7.4. The representations used to define the regular representation in [Ren09] is different from what we have used. But the inversion map of the groupoid intertwines our representations with those used in [Ren09]. The C ∗-algebra of the groupoid N ⋊ G is naturally isomorphic to C0(N ) ⋊ G . Let Φ : Cc(N) ⋊ G → Cc(N ⋊ G) be the map defined by (7.4) Φ(f Ug)(x, h) := ( f (x) 0 if g = h, otherwise. for f ∈ Cc(N ) and g ∈ G. Here {Ug : g ∈ G} denotes the canonical unitaries (corre- sponding to the group elements) in the multiplier algebra of C0(N ) ⋊ G. Then Φ extends to an isomorphism from C0(N ) ⋊ G onto C ∗(N ⋊ G) (Cf. Corollary 2.3.19, Page 34, [Ren09]). Let p := 1M ∈ Cc(N) ⊂ C0(N ) ⋊ G where 1M is the characteristic function associated to the compact open subset M . Note that Φ(1M ) = 1M×{e}. Proposition 7.5. The full corner p(C0(N) ⋊ G)p is isomorphic to A[N ⋊ H, M]. Here the projection p is given by p = 1M . Proof. Let i : Cc(N ⋊ GM ) → Cc(N ⋊ G) be the natural inclusion. It is easy to verify that the image of i is 1M ×{e}Cc(N ⋊ G)1M ×{e}. Now from Proposition 7.3, it follows that C ∗(N ⋊ GM ) is isomorphic to 1M ×{e}C ∗(N ⋊ G)1M ×{e}. But we have the isomorphism Φ : C0(N) ⋊ G → C ∗(N ⋊ G) with Φ(1M ) = 1M ×{e}. Hence A[N ⋊ H, M] is isomorphic to the corner 1M (C0(N ) ⋊ G)1M . Let A = C0(N) ⋊ G. Then ApA is an ideal in A containing p = 1M . Note that for every g ∈ G, xg := Ug1M 1M ∈ ApA. Hence 1gM = Ug1M U ∗ g ∈ ApA. Hence for every g ∈ G, 1g.M ∈ ApA. Thus 1a−1M a ∈ ApA for every a ∈ P . Thus we have Cc(N) ⊂ ApA (See Remark 7.6) and hence C0(N) ⊂ ApA. As a consequence we have ApA = C0(N) ⋊ G. Thus the projection p is full. This completes the proof. g = xgx∗ ✷ Remark 7.6. If K ⊂ N is compact then there exists b ∈ P such that K ⊂ b−1M b. For {a−1M a : a ∈ P} is an open cover of N . Thus there exists a1, a2,· · · , an ∈ P such that i M ai ⊂ b−1Mb. (Reason: i=1 P ai. Then for every i, a−1 i=1 a−1 K ⊂Sn i Mai. Choose b ∈Tn M is dense in M and ba−1 i ∈ P ). Hence K ⊂ b−1M b. 26 S. SUNDAR Remark 7.7. Using the second half of Proposition 7.3, it can be shown that the C ∗- algebra Ared[N ⋊ H, M] is isomorphic to the full corner 1M (C0(N ) ⋊red G)1M . We leave the details to the reader. 8. Cuntz-Li Duality theorem The purpose of this section is to establish a duality result for the C ∗-algebra associated to Examples 2.7 and 2.10. This is analogous to the duality result obtained in [CL11] for the ring C ∗-algebra associated to the ring of integers in a number field. The proof is really a step by step adaptation of the arguments in [CL11] to our situation. Let Γ ⊂ GLn(Q) be a subgroup and let Γ+ := {γ ∈ Γ : γ ∈ Mn(Z)}. Assume that the following holds. (1) The group Γ = Γ+Γ−1 (2) The intersections \γ∈Γ+ γZn = \γ∈Γ+ + = Γ−1 + Γ+. γtZn = {0}. Let Γop := {γt : γ ∈ Γ}. Then Γop is a subgroup of GLn(Q). Also Γ satisfies (1) and (2) if and only if Γop satisfies (1) and (2). If Γ contains the non-zero scalars then (1) and (2) are satisfied. (2). The group Γ acts on Qn by left multiplication. Let NΓ :=Sγ∈Γ+ For the rest of this section, we let Γ be a subgroup of GLn(Q) which satisfies (1) and γ−1Zn. Then by Lemma 2.3, it follows that NΓ is a subgroup of Qn and Γ leaves NΓ invariant. Consider the semidirect product NΓ ⋊ Γ. Then the pair (NΓ ⋊ Γ, Zn) satisfies the hypotheses (C1), (C2) and (C3). Let us denote the C ∗-algebra A[NΓ ⋊ Γ, Zn] by AΓ. Note that NΓ ⋊ Γ acts on Rn on the right as follows. For ξ ∈ Rn and (v, γ) ∈ NΓ ⋊ Γ, let ξ.(v, γ) = γ−1(ξ − v). This right action of NΓ ⋊ Γ on Rn gives rise to a left action of NΓ ⋊ Γ on C0(Rn) as follows. For g ∈ NΓ ⋊ Γ and f ∈ C0(Rn), let (g.f )(x) = f (x.g). The main theorem of this section is the following. Theorem 8.1. The C ∗-algebras AΓop and C0(Rn) ⋊ (NΓ ⋊ Γ) are Morita-equivalent. To prove this we need a bit of preparation. If γ ∈ Γ+, then γ leaves Zn invariant and induces a map on the quotient NΓ Zn which we still denote by γ. Let NΓ := {(zγ)γ∈Γ+ ∈ Yγ∈Γ+ NΓ Zn : δzγδ = zγ for every γ, δ ∈ Γ+} CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 27 We give inherited from the product topology on Yγ∈Γ+ NΓ Zn the discrete topology. The abelian group NΓ is given the subspace topology NΓ Zn . The topological group NΓ is Hausdorff. Now we describe the action of Γ+ on NΓ. Let γ ∈ Γ+ and z ∈ NΓ be given. For δ ∈ Γ+, choose α, β ∈ Γ+ such that γα = δβ. Let (γ.z)δ = βzα. It is easily verifiable that γ is a homeomorphism. The inverse of γ is given by (γ−1z)δ = zγδ. This way Γ+ acts on NΓ and induces an action of Γ on NΓ. Proposition 8.2. We have the following. (1) The map NΓ ∋ v → (γ−1v)γ∈Γ+ ∈ NΓ is injective and is Γ-equivariant. Moreover, when NΓ is viewed as a subset of NΓ via this embedding, NΓ is dense in N Γ. (2) Let MΓ := {z ∈ NΓ : ze = 0} is a compact open subgroup of N Γ. Also the intersection MΓ ∩ NΓ = Zn. Hence Zn is dense in MΓ. γ−1MΓ. As a consequence, NΓ is locally compact. (3) Also NΓ =Sγ∈Γ+ Proof. The fact that v → (γ−1v)γ is injective follows from the assumption that Tγ∈Γ+ γZn = {0}. Let γ ∈ Γ+ and v ∈ NΓ be given. Let us denote the image of v in NΓ by v. We need to show that for δ ∈ Γ+, the δth co-ordinate of γ.v is δ−1γv. Choose α and β in Γ+ such that γα = δβ. Then by definition (γ.v)δ = βα−1v = δ−1γv. Thus we have shown that the embedding NΓ ∋ v → (γ−1v)γ∈Γ+ ∈ NΓ is Γ+ -equivariant and consequently is Γ -equivariant. For γ ∈ Γ+ and v ∈ NΓ, let Uγ,v := {z ∈ NΓ : zγ ≡ v mod Zn}. Clearly the collection {Uγ,v : γ ∈ Γ+, v ∈ NΓ} forms a basis for NΓ. Note that γ.v ∈ Uγ,v. Thus NΓ is dense in NΓ. Nγ Zn is finite. Now observe that Zn . Thus MΓ is compact. Since the projection onto the eth co-ordinate is a continuous homomorphism, it follows that MΓ is an open subgroup. The equality MΓ ∩ NΓ = Zn is obvious. For γ ∈ Γ+, let Nγ := γ−1Zn. Note that for γ ∈ Γ+, Nγ M Γ = NΓ ∩Qγ Let z ∈ NΓ be given. Since NΓ = Sγ∈Γ+ such that γze = 0. Then γ.z ∈ MΓ. Thus NΓ = Sγ∈Γ+ γ−1Zn, it follows that there exists γ ∈ Γ+ γ−1MΓ. As NΓ is a union of compact open subsets, it follows that NΓ is locally compact. This completes the proof. ✷ 28 S. SUNDAR Let N ′ and M ′ be the groups considered in Section 6 applied to the pair (NΓ ⋊ Γ, Zn). Let us now convince ourselves that the pair (N ′, M ′) is Γ-equivariantly isomorphic to the pair (NΓ, MΓ). Let γ, δ ∈ Γ+ be given. Zn γZn which we denote by qγ,δ. Multiplication by γ−1 maps Zn injectively onto γ−1Zn and Zn γZn by qγ. Then qγ descends to a map Denote the quotient map Zn → Zn γδZn → takes γZn onto Zn. We denote the resulting isomorphism from again by Zn γZn → γ−1Zn Zn γ−1. Then we have the following commutative diagram where the vertical arrows are isomorphisms. (8.5) Recall that qγ,δ Zn γδZn Zn γZn (γδ)−1 γ−1 (γδ)−1Zn Zn δ γ−1Zn Zn M ′ = {(zγ)γ∈Γ+ ∈ Yγ∈Γ+ MΓ = {(zγ)γ∈Γ+ ∈ Yγ∈Γ+ Zn γZn : qγ,δ(zγδ) = zγ} γ−1Zn Zn : δzγδ = zγ} Let i : Zn → M ′ be the embedding given by i(v) = (v)γ∈Γ+ and j : Zn → MΓ be the embedding described in Proposition 8.2. Then j(v) = (γ−1v)γ∈Γ+ for v ∈ Zn. Now the commutative diagram 8.5 implies that the map ϕ : M ′ → MΓ given by ϕ((zγ)) = (γ−1zγ) is an isomorphism and ϕ(i(v)) = j(v) for v ∈ Zn. It is also clear that ϕ is a homeomorphism. Claim: ϕ is Γ+-equivariant. First the embeddings i and j are Γ+-equivariant. Since ϕ ◦ i = j, it follows that ϕ(γ.i(v)) = γ.ϕ(i(v)) if γ ∈ Γ+ and v ∈ Zn. Since i(Zn) is dense in M ′ ( and the maps involved are continuous ), it follows that ϕ(γ.x) = γ.ϕ(x) for x ∈ M ′ and γ ∈ Γ+. γ−1M ′, it follows from the univer- sal property, as explained in Section 6 (item 9), that the map γ−1x → γ−1ϕ(x) (with γ−1MΓ and N ′ = Sγ∈Γ+ Now since NΓ = Sγ∈Γ+ CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS x ∈ M ′) extends to a Γ-equivariant isomorphism from N ′ → NΓ. 29 ✷ Now we describe the Pontryagin dual of the discrete group NΓ. For x, ξ ∈ Rn, let map ξ → χξ. If ξ ∈ Rn, restricting χξ to NΓ gives a character of NΓ. Moreover the map < x, ξ >:= xtξ . If x, ξ ∈ Rn, we let χξ(x) = e2πi<x,ξ>. We identity Rn with cRn via the Rn ∋ ξ → χξ ∈ cNΓ is continuous. Let z ∈ N Γop be given. Let χz : NΓ → T be defined as follows. For x ∈ γ−1Zn for some γ ∈ Γ+, , let χz(x) = e2πi<γx,zγ> = e2πi<x,γtzγ >. It is easy to verify that χz is well defined and χz is a character of NΓ. Clearly NΓop ∋ z → χz ∈ cNΓ is continuous. Note Proposition 8.3. The map Ψ : Rn × N Γop → cNΓ defined by that if z ∈ NΓop and x ∈ NΓ then χz(x) = e2πi<x,z>. Ψ(ξ, z) = χξχ−z is a surjective homomorphism with kernel ∆ = {(x, x) : x ∈ NΓop}. The induced map Rn × N Γop ∆ eΨ : → cNΓ is a topological isomorphism. Proof. Clearly Ψ is a continuous group homomorphism and Ψ(∆) = {1}. Now let us show that the kernel of Ψ is ∆. Let (ξ, z) be such that Ψ(ξ, z) = 1. Then for every γ ∈ Γ+ and x ∈ Zn, we have 1 = χξ(γ−1x)χ−z(γ−1x) = e2πi<x,(γt)−1ξ>e−2πi<x,zγ> = e2πi<x,(γt)−1ξ−zγ> Thus for every γ ∈ Γ+, we have zγ − (γt)−1ξ ∈ Zn. In other words, we have ξ ∈ NΓop and z = ξ in N Γop. Hence (ξ, z) ∈ ∆. Thus we have shown that the kernel of Ψ is ∆ Next we claim Rn×N Γop which implies that eΨ is one-one. map. We also write λ(ξ, z) as [(ξ, z)]. We claim that λ([0, 1]n × M Γop) = Rn×N Γop will prove that Rn×N Γop is compact. Let λ : Rn × N Γop → Rn×N Γop is compact. be the quotient ∆ . This ∆ ∆ Let [(ξ, z)] be an element in the quotient Rn×N Γop . Choose v ∈ Zn and γ ∈ Γ+ such that ze ≡ (γt)−1v. Then [(ξ, z)] = [(ξ − (γt)−1v, z − (γt)−1v)]. Choose w ∈ Zn such that ξ − (γt)−1v − w ∈ [0, 1]n. Let ξ ′ = ξ − (γt)−1v − w and z ′ = z − (γt)−1v − w. Then ξ ′ ∈ [0, 1]n and z ′ ∈ M Γop. Moreover λ(ξ, z) = λ(ξ ′, z ′). Thus the image of [0, 1]n × M Γop under λ is Rn×N Γop ∆ . ∆ ∆ 30 S. SUNDAR Consider the semidirect product Rn ⋊ Γop where Γop acts on Rn by left multiplication. on the right as follows. For The image of eΨ is a compact subgroup of cNΓ and it separates points of NΓ ( The image of Rn × {0} under Ψ separates points of NΓ). Hence eΨ is onto. Since Rn×N Γop compact, it follows that eΨ is a topological isomorphism. This completes the proof. ✷ The semidirect product Rn ⋊ Γop acts on cNΓ = Rn×NΓop [(ξ, z)] ∈ cNΓ and (v, γ) ∈ Rn ⋊ Γop, let [(ξ, z)].(v, γ) = [(γ−1(ξ + v), γ−1z)]. This right action of Rn ⋊ Γop on cNΓ induces a left action of Rn ⋊ Γop on C ∗(NΓ) ∼= C(cNΓ). The crossed product C ∗(NΓ)⋊(Rn ⋊Γop) is isomorphic to the iterated crossed product (C ∗(NΓ) ⋊ Rn) ⋊ Γop. (Cf. Proposition 3.11, Page 87, [Wil07].) But then the map Γ ∋ γ → (γt)−1 ∈ Γop is an isomorphism. Thus the crossed product (C ∗(NΓ)⋊Rn)⋊Γop ∼= (C ∗(NΓ) ⋊ Rn) ⋊ Γ. is ∆ ∆ Let us fix notations. Let τ be the action of Rn on C ∗(NΓ). Let β be the action of Γ on C ∗(NΓ) ∼= C(cNΓ), induced by the action of Γop and the identification Γ ∼= Γop. For v ∈ NΓ, ξ ∈ Rn and γ ∈ Γ, it is easy to verify the following. τξ(δv) = e−2πi<ξ,v>δv, βγ(δv) = δγv where {δv : v ∈ NΓ} denotes the canonical unitaries of C ∗(NΓ). The action of Γop on C ∗(NΓ) ⋊ Rn, induces an action of Γ ( via the identification Γ ∋ γ → (γt)−1) and let us denote it by eβ. For γ ∈ Γ, and f ∈ Cc(Rn, C ∗(NΓ)), we have eβγ(f )(x) = det(γ)βγ(f (γtx)). Now consider the crossed product C0(Rn) ⋊ (NΓ ⋊ Γ) ∼= C ∗(Rn) ⋊ (NΓ ⋊ Γ). Let us denote the action of NΓ and Γ on C ∗(Rn) by σ and α. For v ∈ NΓ, γ ∈ Γ and f ∈ Cc(Rn), we have (σvf )(ξ) = e2πi<ξ,v>f (ξ), (αγf )(ξ) = det(γ)f (γtξ). Denote the action of Γ on C ∗(Rn) ⋊ NΓ by eα. For γ ∈ Γ, v ∈ NΓ and f ∈ C ∗(Rn), one has Let us recall the following lemma which is Lemma 4.3 in [CL11]. fαγ(f δv) = αγ(f )δγv. Lemma 8.4 ([CL11]). Let G be a locally compact abelian group and H be a subgroup of the Pontryagin dual bG. Endow H with the discrete topology. Let σ be the action of CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 31 H on C ∗(G) and τ be the action of G on C ∗(H) given by σh(f ) = [g → h(g)f (g)] and τg( f ) = [h → h(−g) f (h)]. Then the map φ : Cc(H, Cc(G)) → Cc(G, Cc(H)) defined by φ(f )(g)(h) = h(−g)f (h)(g) extends to an isomorphism between C ∗(G) ⋊σ H and C ∗(H) ⋊τ G. We are now ready to prove the following proposition. Proposition 8.5. The crossed products C0(Rn) ⋊ (NΓ ⋊ Γ) and C(cNΓ) ⋊ (Rn ⋊ Γop) are isomorphic. Proof. It is enough to show that the crossed products (C ∗(Rn) ⋊σ NΓ) ⋊eα Γ and (C ∗(NΓ) ⋊τ Rn) ⋊ eβ Γ are isomorphic. We show that C ∗(Rn) ⋊σ NΓ and C ∗(NΓ) ⋊τ Rn are Γ-equivariantly isomorphic. Then the isomorphism between the crossed products will follow. Note that the action σ of NΓ on C ∗(Rn) and τ of Rn on C ∗(NΓ) are exactly as in Lemma 8.4. Identify Rn with cRn via the map ξ → χξ. (Recall that χξ is the character given by χξ(x) = e2πi<x,ξ>. ) Consider NΓ as a subgroup ofcRn via the natural inclusion NΓ ⊂ Rn. Thus Lemma 8.4 implies that C ∗(Rn)⋊σ NΓ ∼= C ∗(NΓ)⋊τ Rn. Let φ : C ∗(Rn)⋊σ NΓ → C ∗(NΓ)⋊τ Rn be the isomorphism prescribed by Lemma 8.4. We claim φ is Γ-equivariant. First note that φ(f δv)(ξ) = e−2πi<ξ,v>f (ξ)δv for f ∈ Cc(Rn) and v ∈ NΓ. Let γ ∈ Γ be given. Now observe that eβγ(φ(f δv))(ξ) = det(γ)βγ(φ(f δv)(γtξ)) = det(γ)e−2πi<γtξ,v>f (γtξ)δγv = det(γ)e−2πi<ξ,γv>f (γtξ)δγv. On the other hand, observe that φ(fαγ(f δv))(ξ) = φ(αγ(f )δγv)(ξ) = e−2πi<ξ,γv>αγ(f )(ξ)δγv = e−2πi<ξ,γv> det(γ)f (γtξ)δγv Γ-equivariant. This completes the proof. Hence for every γ ∈ Γ, eβγφ(f δv) = φfαγ(f δv). Since {f δv : f ∈ Cc(Rn), v ∈ NΓ} is total in C ∗(Rn) ⋊σ NΓ, it follows that for every γ, eβγφ = φfαγ. algebra of the groupoid eG := N Γop ⋊(NΓop ⋊Γop)M Γop . By Proposition 8.5, it follows that Proof of Theorem 8.1. By Corollary 6.6, it follows that AΓop is isomorphic to the C ∗- In other words, φ is ✷ 32 S. SUNDAR C0(Rn) ⋊ (NΓ ⋊ Γ) is isomorphic to the C ∗-algebra of the groupoid G := cNΓ ⋊ (Rn ⋊ Γop). We will show that G and eG are equivalent in the sense of [MRW87]. By Proposition 8.3, bNΓ = Rn×N Γop quotient map Rn × N Γop → Rn×N Γop subset of G0 and it is easy to verify that X meets each orbit of G0. Let where ∆ := {(x, x) : x ∈ NΓop}. Denote the by λ. Let X := λ({0} × M Γop). Then X is a closed ∆ ∆ GX := {α ∈ G : s(α) ∈ X} = s−1(X) We claim that the (restricted) source map s : GX → X and the range map r : GX → G0 are open. Let U ⊂ G be an open subset. Then s(U ∩ GX) = s(U) ∩ X. Since s : G → G0 is open, it follows that s : GX → X is open. Now we prove that r : GX → G0 is open. It is enough to show that r((U×V ×{γ})∩GX ) is open whenever U ⊂ Rn×N Γop ∆ and V ⊂ Rn are open and γ ∈ Γop. We claim that r((U × V × {γ}) ∩ GX ) = U ∩ λ(−V × γM Γop) Let [(ξ, z)] ∈ r((U × V × {γ}) ∩ GX ). Then there exists ([(η, y)], v, γ) ∈ U × V × {γ} such that [(η, y)].(v, γ) ∈ X and [(ξ, z)] = [(η, y)]. Thus there exists u ∈ NΓop such that γ−1(ξ + v) = u and γ−1z − u = x for some x ∈ M Γop. Hence [(ξ, z)] = [(−v, γx)]. Clearly [(ξ, z)] ∈ U. Hence [(ξ, z)] ∈ U ∩ λ(−V × γM Γop). Thus we have shown that r((U × V × {γ}) ∩ GX) ⊂ U ∩ λ(−V × γM Γop). Now let [(ξ, z)] ∈ U ∩ λ(−V × γM Γop). Then there exists (v, x) ∈ V × M Γop such that [(ξ, z)] = [(−v, γx)]. This is equivalent to saying that [(ξ, z)].(v, γ) ∈ X. Thus ([(ξ, z)], v, γ) ∈ (U × V × {γ}) ∩ GX and r([(ξ, z)], v, γ) = (ξ, z)]. This proves that U ∩ λ(−V × γM Γop) ⊂ r((U × V × {γ}) ∩ GX ). This proves the claim that r((U × V × {γ}) ∩ GX ) = U ∩ λ(−V × γM Γop). Now since λ is open and M Γop is open, it follows that r((U × V ×{γ})∩GX) is open. Thus we have shown that r : GX → G0 is open. Now by Example 2.7 of [MRW87], it follows that G and GX X := {α ∈ GX : r(α) ∈ X} Now we prove that Φ is a topological isomorphism. X be defined by Φ(x, v, γ) = ([(0, x)], v, γ). It is easy to check that Φ is a groupoid isomorphism and it is continuous. are equivalent. Recall that eG = N Γop ⋊ (NΓop ⋊ Γop)M Γop The right action of NΓop ⋊ Γop on N Γop is given by x.(v, γ) = γ−1(x − v). Let Φ : eG → GX Let (xn, vn, γ) be a sequence in eG such that Φ(xn, vn, γ) converges to ([(0, x)], v, γ)). First note that x → [(0, x)] is a topological embedding of M Γop into cNΓ. Thus, it follows that xn converges to x in M Γop. Now Φ(xn, vn, γ) converges to [(0, x)], v, γ) implies that CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 33 isomorphism. vn tends to v in Rn and γ−1(x− vn) tends to γ−1(x− v) in M Γop. Hence vn converges to v in N Γop. Thus (vn, vn) → (v, v) in Rn × N Γop. But ∆ is a discrete subgroup of Rn × N Γop. Hence vn = v eventually. Therefore, (xn, vn, γ) → (x, v, γ) in eG. So, Φ is a topological Since G and eG are equivalent in the sense of [MRW87], it follows from Theorem 2.8 in [MRW87] that C ∗(G) and C ∗(eG) are Morita-equivalent. This completes the proof. 8.1. Examples. We end this article by considering two examples. ✷ Example 1: First we show that the duality result for the ring C ∗-algebra associated to number fields obtained in [CL11] can be derived from Theorem 8.1. Consider a number field K of degree n. Denote the ring of integers in K by OK. Let {w1, w2,· · · , wn} be a Z-basis for OK. Then {w1, w2,· · · , wn} is a Q-basis for K. Identify i=1 xiwi ∈ K. By definition, K with Qn via the map β : Qn ∋ (x1, x2,· · · , xn)t → Pn β(Zn) = OK. If a ∈ K, then a acts on K by left multiplication and is Q-linear. Thus a gives rise to a matrix with respect to the basis {w1, w2,· · · , wn} which we denote by α(a). Explicitly, for 1 ≤ j ≤ n, let (8.6) awj := nXi=1 αij(a)wi. Let α(a) := (αij(a)). Then α : K → Mn(Q) is an injective ring homomorphism. We also have the following equivariance. For a ∈ K and x ∈ Qn, β(α(a)x) = aβ(x). Let Γ := α(K ×). Then Γ is a subgroup of GLn(Q). Now the pair (K ⋊ K ×, OK) is isomorphic to (Qn ⋊ Γ, Zn). Thus the ring C ∗-algebra associated to OK is nothing but A[Qn ⋊ Γ, Zn]. Hence Theorem 8.1 applies. The only thing that one needs to verify α(a)Zn = {0}. We produce a matrix X with rational entries whose determinant is non-zero and α(a)tZn = {0}. (See is Ta∈OK Xα(a)X −1 = α(a)t for every a ∈ OK. Then it will follow that \a∈OK aOK = {0}, it follows that Ta∈OK α(a)tZn is trivial. Since Ta∈OK also Lemma 8.8.) Let T r : Mn(Q) → Q be the usual trace and let tr := T r ◦ α. Denote the n × n matrix whose (i, j)th entry is tr(wiwj) by X. Then X has determinant non-zero and its determinant is called the discriminant of the number field K. Lemma 8.6. For every a ∈ K, Xα(a)X −1 = α(a)t. 34 S. SUNDAR Proof. Fix a ∈ K. Let Y = (tr(awiwj)). Multiplying Equation 8.6 by wk and taking trace, we get Yjk = αij(a)Xik nXi=1 In other words, we have Y = α(a)tX. But Y and X are symmetric. Thus taking transpose, we get Y = Xα(a). Hence Xα(a) = α(a)tX. This completes the proof. ✷ Let A∞ denote the ring of infinite adeles associated to K. Theorem 8.7 ([CL11]). For a number field K, the ring C ∗-algebra A[K ⋊ K ×, OK] is Morita-equivalent to C0(A∞) ⋊ (K ⋊ K ×). Proof. Note that for Γ = α(K ×), NΓ = Qn and NΓop = Qn ( since Γ contains the diagonal matrices with rational entries). Thus Lemma 8.6 implies that the matrix X = (tr(wiwj)) implements an isomorphism between the dynamical systems (Rn, Qn⋊Γ) and (Rn, Qn ⋊ Γop). The map (Rn, Qn ⋊ Γ) ∋ (ξ, (v, γ)) → (Xξ, (Xv, γt)) ∈ (Rn, Qn ⋊ Γop) is the required isomorphism. (Note that Γ is commutative.) number theoretic arguments, ( for example, using Theorem 13.5 (page 70) and Theorem Consider the map δ : Rn ∋ (x1, x2,· · · , xn) →Pi=1 xiwi ∈ A∞. Then from standard 4.4 (page 110) in [Jan96]), it follows that δ (together with identifications α and β) implements an isomorphism between (A∞, K × K ⋊) and (Rn, Qn ⋊ Γ). Now Theorem 8.1 yields the required result. This completes the proof. ✷ We claim that the duality result is applicable to this example. The only thing that Example 2: Let A be an n × n matrix with integer entries such that det(A) 6= 0 and T∞ r=0 ArZn = {0}. Let Γ := {Ar : r ∈ Z} ∼= Z. Denote the subgroup NΓ by NA and the Cuntz-Li algebra A[NΓ ⋊ Γ, Zn] by AA. Denote the transpose At by B. Then Γop = {Br : r ∈ Z} ∼= Z. needs verification isT∞ T∞ r=0 ArZn = {0} if and only ifT∞ Lemma 8.8. Let A be a n × n matrix with integer entries and denote At by B. Then r=0 BrZn = {0}. This follows from the following lemma. Proof. Since A and B are similar over Q, it follows that there exists Y ∈ GLn(Q) such that Y AY −1 = B. Choose a non-zero integer m such that X = mY ∈ Mn(Z). One has XA = BX. By induction, it follows that XAr = BrX for every r ≥ 0. First note that r=0 BrZn = {0}. it is enough to show thatT∞ r=0 ArZn 6= {0} impliesT∞ r=0 BrZn 6= {0}. CUNTZ-LI RELATIONS, INVERSE SEMIGROUPS AND GROUPOIDS 35 r=0 ArZn. Then Suppose v is a non-zero element inT∞ ∞\r=0 ∞\r=0 Xv ∈ = XArZn BrXZn ⊂ BrZn. ∞\r=0 Since X is invertible over Q, it follows that Xv is a non-zero element in T∞ Thus ifT∞ r=0 BrZn 6= {0}. This completes the proof. Now Theorem 8.1 and Proposition 8.5 implies the following proposition. r=0 ArZn 6= {0} thenT∞ r=0 BrZn. ✷ Proposition 8.9. The C ∗-algebra AAt is Morita-equivalent to C0(Rn) ⋊ (NA ⋊ Z). Also AAt is Morita-equivalent to (C ∗(NA) ⋊ Rn) ⋊ Z. Proposition 8.9 for the case when n = 1 and A = (2) was proved in [SL10a]. In this case, the C ∗-algebra AAt = AA is the C ∗-algebra Q2 considered in [SL10a]. The 2] in [SL10a]. The Morita equivalence between Q2 and 2] ⋊ (2)) is called the 2-adic duality theorem in [SL10a]. (Cf. Corollary 5.5 subgroupSr=0 2−rZ is denoted Z[ 1 C0(R) ⋊ (Z[ 1 and Theorem 7.5 in [SL10a].) References [AD97] Claire Anantharaman-Delaroche, Purely infinite C ∗-algebras arising from dynamical systems, Bull. Soc. Math. France 125 (1997), no. 2, 199 -- 225. MR 1478030 (99i:46051) [BE10] Giuliano Boavo and Ruy Exel, Partial crossed product description of the C ∗-algebras associ- ated to integral domains, arxiv:1010.0967v2/math.OA, 2010. [CL10] Joachim Cuntz and Xin Li, The regular C ∗-algebra of an integral domain, Quanta of maths, Clay Math. Proc., vol. 11, Amer. Math. Soc., Providence, RI, 2010, pp. 149 -- 170. MR 2732050 [CL11] , C ∗-algebras associated with integral domains and crossed products by actions on adele spaces, J. Noncommut. Geom. 5 (2011), no. 1, 1 -- 37. MR 2746649 [CP61] A. H. Clifford and G. B. Preston, The algebraic theory of semigroups. Vol. I, Mathematical Surveys, No. 7, American Mathematical Society, Providence, R.I., 1961. MR 0132791 (24 #A2627) [EaHR10] Ruy Exel, Astrid an Huef, and Iain Raeburn, Purely infinite simple C ∗-algebras associated to integer dilation matrices, arxiv.1003.2097/math.OA, 2010. [Exe08] Ruy Exel, Inverse semigroups and combinatorial C ∗ algebras, Bull.Braz.Math.Soc.(NS) 39 (2008), no. 2, 191 -- 313. [Exe09] , Tight representations of semilattices and inverse semigroups, Semigroup Forum 79 (2009), 159 -- 182. 36 S. SUNDAR [Jan96] Gerald J. Janusz, Algebraic number fields, second ed., Graduate Studies in Mathematics, vol. 7, American Mathematical Society, Providence, RI, 1996. MR 1362545 (96j:11137) [KLQ11] S. Kaliszewski, M. Landstad, and J. Quigg, A crossed-product approach to the Cuntz-Li algebras, arxiv:1012:5285v3, 2011. [Lac00] Marcelo Laca, From endomorphisms to automorphisms and back: dilations and full corners, J. London Math. Soc. (2) 61 (2000), no. 3, 893 -- 904. MR 1766113 (2002a:46094) [Li12] Xin Li, Nuclearity of semigroup C ∗-algebras and the connection to amenability, [MR82] arxiv/math.OA:1203.0021, 2012. Paul S. Muhly and Jean N. Renault, C ∗−algebras of multivariable Wiener-Hopf operators, Trans.Amer.Math.Soc. 274 (1982), no. 1, 1 -- 44. [MRW87] Paul S. Muhly, Jean N. Renault, and Dana P. Williams, Equivalence and isomorphism for groupoid C ∗-algebras, J. Operator Theory 17 (1987), no. 1, 3 -- 22. MR 873460 (88h:46123) [Ore31] Oystein Ore, Linear equations in non-commutative fields, Ann. of Math. (2) 32 (1931), no. 3, 463 -- 477. [Ren09] Jean Renault, C ⋆-algebras and dynamical systems, Publica¸coes Matem´aticas do IMPA. [IMPA Mathematical Publications], Instituto Nacional de Matem´atica Pura e Aplicada (IMPA), Rio de Janeiro, 2009, 27o Col´oquio Brasileiro de Matem´atica. [27th Brazilian Mathematics Colloquium]. [SL10a] Nadia S.Larsen and Xin Li, The 2−-adic ring C ∗-algebra of the integers and its representa- tions, arxiv:1011.5622v1/math.OA, 2010. [SL10b] Nadia S.Larsen and Xin Li, Dilations of semigroup crossed products as crossed products of dilations, arxiv:1009.5842/math.OA, 2010. [Sun11] .S Sundar, Inverse semigroups and the Cuntz-Li algebras, arxiv:1104.1085v2/math.OA, 2011. [Wil07] Dana P. Williams, Crossed products of C ∗-algebras, Mathematical Surveys and Monographs, vol. 134, American Mathematical Society, Providence, RI, 2007. MR 2288954 (2007m:46003) E-mail address: [email protected] Indian Statistical Institute, Delhi.
1110.6500	1	1110	2011-10-29T05:34:25	The carpenter and Schur--Horn problems for masas in finite factors	[ "math.OA" ]	Two classical theorems in matrix theory, due to Schur and Horn, relate the eigenvalues of a self-adjoint matrix to the diagonal entries. These have recently been given a formulation in the setting of operator algebras as the Schur-Horn problem, where matrix algebras and diagonals are replaced respectively by finite factors and maximal abelian self-adjoint subalgebras (masas). There is a special case of the problem, called the carpenter problem, which can be stated as follows: for a masa A in a finite factor M with conditional expectation E_A, can each x in A with 0 <= x <= 1 be expressed as E_A(p) for a projection p in M? In this paper, we investigate these problems for various masas. We give positive solutions for the generator and radial masas in free group factors, and we also solve affirmatively a weaker form of the Schur-Horm problem for the Cartan masa in the hyperfinite factor.	math.OA	math	THE CARPENTER AND SCHUR -- HORN PROBLEMS FOR MASAS IN FINITE FACTORS 1 1 0 2 t c O 9 2 ] Kenneth J. Dykema ∗ Junsheng Fang † Roger R. Smith ‡ September 26, 2018 Abstract Donald W. Hadwin . A O h t a m [ 1 v 0 0 5 6 . 0 1 1 1 : v i X r a Two classical theorems in matrix theory, due to Schur and Horn, relate the eigen- values of a self-adjoint matrix to the diagonal entries. These have recently been given a formulation in the setting of operator algebras as the Schur-Horn problem, where matrix algebras and diagonals are replaced respectively by finite factors and maximal abelian self-adjoint subalgebras (masas). There is a special case of the problem, called the carpenter problem, which can be stated as follows: for a masa A in a finite factor M with conditional expectation EA, can each x ∈ A with 0 ≤ x ≤ 1 be expressed as EA(p) for a projection p ∈ M ? In this paper, we investigate these problems for various masas. We give positive solutions for the generator and radial masas in free group factors, and we also solve affirmatively a weaker form of the Schur-Horm problem for the Cartan masa in the hyperfinite factor. 1 Introduction Two classical theorems due to Schur [17] and Horn [9], which relate the diagonal entries of an n × n self-adjoint matrix to its eigenvalues, have recently been reformulated in the setting of type II1 factors M with normalized trace τ [3]. A special case of the problem, termed the carpenter problem in [10, 11], asks whether each element x in a masa A ⊆ M satisfying 0 ≤ x ≤ 1 can be expressed as EA(p) for some projection p ∈ M. This entails τ (x) = τ (p), so the analogous problem in complex matrix algebras places a constraint on the value of τ (x). Subject to this, Horn's theorem gives a positive solution for matrices. The II1 -- factor analogue of the diagonal subalgebra in the n × n matrices is a maximal abelian (self -- adjoint) subalgebra, called a masa, A ⊆ M. We let EA denote the trace -- preserving conditional expectation of M onto A. The carpenter problem in a II1 -- factor is, ∗Partially supported by NSF grant DMS-0901220 †Partially supported by the Fundamental Research Funds for the Central Universities of China and NSFC(11071027) ‡Partially supported by NSF grant DMS-1101403 1 given x ∈ A with 0 ≤ x ≤ 1, to find a projection p ∈ M so that EA(p) = x; this problem remains open. The Schur -- Horn problem for a masa A ⊆ M may be stated as follows: for a suitable notion of spectral majorization of x ∈ A by z ∈ M (described in Section 5), does there exist an element y ∈ M having the same spectral distribution as z so that x = EA(y)? In this paper we address these two questions for specific choices of masas. We give positive solutions to both the carpenter problem and the Schur -- Horn problem when A is either a generator masa or the radial masa in a free group factor. We also investigate the Cartan masa in the hyperfinite factor, and obtain a version of the Schur -- Horn theorem which is slightly weaker than the one above. The paper is organized as follows. In Section 2 we present a technical result giving a sufficient condition for positive solutions of the carpenter problem (Lemma 2.1), and all of our subsequent results are based on this. The main results on masas in free group factors are contained in Section 3, while Section 4 is concerned with the carpenter problem for the Cartan masa in the hyperfinite factor. Here our results are less definitive, although we do present classes of elements in A for which a positive solution can be given. In a different direction, we also solve the carpenter problem for all elements of the Cartan masa A, but modulo an automorphism of A. In the final section, we consider the Schur -- Horn problem. We first consider a minor reformulation of Arveson and Kadison's version of the problem and show that it is equivalent to theirs. Then we give a positive solution for the generator masa and the radial masa in free group factors. We also investigate the Cartan masa, proving a weaker version of the Schur -- Horn problem as mentioned above. There has been considerable recent interest in these problems, and we have drawn heavily on the ideas and results presented in [1, 2, 3, 10, 11]. 2 An existence method In the first lemma below we will describe a sufficient condition for solving the carpenter problem positively, and in subsequent sections we will apply it in various situations. We fix a finite von Neumann algebra M with a normal normalized trace τ and a masa A ⊆ M. We denote the unique trace preserving conditional expectation of M onto A by EA. For each x ∈ A satisfying 0 ≤ x ≤ 1, we introduce the w∗-compact convex subset Γx ⊆ M, defined by Γx = {y ∈ M : 0 ≤ y ≤ 1, EA(y) = x}. (2.1) This set is nonempty since it contains x, and any projection p ∈ Γx is a solution of the carpenter problem for the element x ∈ A. Any such projection is automatically an extreme point of Γx, and so it suffices to consider the extreme points of Γx. These are abundant, by the Krein -- Milman theorem. For each nonzero projection e ∈ M, define a bounded map Φe : eMe → A by Φe(exe) = EA(exe), x ∈ M. (2.2) 2 Lemma 2.1. Let A be a masa in a finite von Neumann algebra M, and suppose that Φe is not injective for each nonzero projection e ∈ M. Given x ∈ A satisfying 0 ≤ x ≤ 1, there exists a projection p ∈ M such that EA(p) = x. Proof. Fix an arbitrary x ∈ A satisfying 0 ≤ x ≤ 1. Under the stated hypotheses, we will show that every extreme point of Γx is a projection and the result then follows. To obtain a contradiction, let y be an extreme point of Γx which is not a projection. For a sufficiently small choice of ε > 0, the spectral projection e of y for the interval (ε, 1 − ε) is nonzero. Since Φe is not injective we may choose a nonzero element z ∈ eMe so that Φe(z) = 0. By considering real and imaginary parts we may take z to be self-adjoint, and by scaling we may assume that kzk ≤ ε. Note that EA(y ± z) = EA(y) = x. Since εe ≤ ye ≤ (1 − ε)e, it follows that 0 ≤ y ± z ≤ 1, and so y ± z ∈ Γx with y = ((y + z) + (y − z))/2. This contradicts the assumption that y is an extreme point, showing that every extreme point is a projection. To illustrate the use of Lemma 2.1, we now show that the carpenter problem has a positive solution for any masa in a free group factor with an uncountable number of generators. Theorem 2.2. Let S be an uncountable set and let FS be the free group on a set of generators indexed by S. If A is a masa in L(FS) and x ∈ A satisfies 0 ≤ x ≤ 1, then there exists a projection p ∈ L(FS) such that EA(p) = x. Proof. From [15], any masa A in L(FS) is separable as a von Neumann algebra. Cardinality considerations then show that Φe must have a nontrivial kernel for each nonzero projection e ∈ L(FS), and the result follows from Lemma 2.1. Remark 2.3. (i) The maps Φe introduced above are normal and so have preduals. It is an easy calculation to see that (Φe)∗ : L1(A) → L1(eMe) is given by (Φe)∗(a) = eae, a ∈ A, and extended by continuity to L1(A). It then follows that noninjectivity of Φe is equivalent to the failure of eAe to be k · k1-dense in eMe, a potentially useful reformulation. (ii) In the case of type II1 factors, we have no example of a nonzero projection e for which Φe is injective. However, this can occur for type I factors. Take A to be the diagonal masa in B(H) and let e ∈ A be a rank one projection. Then eB(H)e = eAe and Φe is injective in this case. If e ∈ M is a projection such that e{e, A}′′e 6= eMe then the map Φe is not injective. (iii) To see this, let N = {e, A}′′ and observe that the condition eNe 6= eMe gives a nonzero element exe ∈ eMe so that EeN e(exe) = 0. Then EA(exe) = EA(EN (exe)) = EA(EeN e(exe)) = 0 and Φe is not injective. (2.3) (cid:3) The third part of this remark leads to a connection with another open problem, the question of whether separable von Neumann algebras must be singly generated. Lemma 2.4. Let M be a type II1 factor and let A be a separable masa. If there exists a nonzero projection e ∈ M such that Φe is injective, then M is singly generated. 3 Proof. Let N = {A, e}′′ and let z be the central support of e in N. Then z is the identity element for the w∗-closed ideal NeN in N. By Remark 2.3 (iii), the injectivity of Φe implies that eNe = eMe. For any m ∈ M, w∗ zmz ∈ NeN m NeN w∗ ⊆ NeMeN w∗ = NeNeN w∗ = Nz, (2.4) showing that zMz ⊆ zNz. The reverse containment is obvious and so zMz = Nz. Since z ∈ A, this gives zMz = {Az, e}′′, so the separability of A implies that zMz is generated by two self-adjoint elements x1 and x2. By adding a multiple of z and scaling, we may assume that 0 ≤ x1 ≤ z. Since M is a finite factor, we can find projections z2, . . . , zn ∈ M which are all equivalent n to subprojections of z and such that z+ zi = 1. Then choose partial isometries v2, . . . , vn ∈ M so that v∗ i vi = zi and viv∗ i ≤ z for 2 ≤ i ≤ n, and define Pi=2 y1 = x1 + 2z2 + · · · + nzn, y2 = e + v2 + v∗ 2 + · · · + vn + v∗ n. (2.5) By construction, z2, . . . , zn are spectral projections of y1 and so lie in {y1, y2}′′, showing that this algebra also contains x1. Since y2zi = vi, 2 ≤ i ≤ n, we see that {y1, y2}′′ also contains v2, . . . , vn and e, so in particular zMz ⊆ {y1, y2}′′. Now v∗ i zvi = zi, and so ziMzj ⊆ {y1, y2}′′, showing that M = {y1, y2}′′. Thus M is singly generated by y1 + iy2. It is currently unknown whether separable type II1 factors exist that are not singly generated. Lemmas 2.1 and 2.4 show that any such example would have a positive solution to the carpenter problem for any masa A. We conclude this section by presenting a class of masas for which the carpenter problem has a positive solution. We will need a preliminary lemma which gives a norm density result. Lemma 2.5. Let M be a separable type II1 factor and let A be a masa in M. (i) If r ∈ Q ∩ [0, 1] then there exists a projection p ∈ M so that EA(p) = r1. (ii) Given ε > 0 and x ∈ A satisfying 0 ≤ x ≤ 1, there exists a projection p ∈ M such that kx − EA(p)k < ε. Proof. (i) The cases r = 0 and r = 1 are trivial so we may assume that r = k/n where 1 ≤ k ≤ n − 1 for integers k, n. In A, choose n orthogonal projections e11, . . . , enn of trace 1/n and choose a matrix algebra Mn ⊆ M with diagonal Dn so that the eii's are the minimal diagonal projections. Since the eii's lie in Dn, the two conditional expectations EA and EDn agree on Mn. From [9], there is a projection p ∈ Mn ⊆ M so that EDn(p) = (k/n)In, and so EA(p) = (k/n)1 ∈ A. (ii) Now consider a fixed but arbitrary x ∈ A satisfying 0 ≤ x ≤ 1 and let ε > 0 be given. Since A is separable we may identify A with L∞[0, 1] and then we may choose projections ek, 1 ≤ k ≤ n, summing to 1, corresponding to disjoint measurable subsets of [0,1], and constants λk ∈ [0, 1] so that x − < ε. (2.6) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) n Xk=1 λkek(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)∞ 4 A further approximation allows us to assume that each λk is rational in [0,1]. Applying (i) to the containment Aek ⊆ ekMek, we find projections pk ≤ ek, 1 ≤ k ≤ n, so that n n EAek(pk) = λkek. If we define a projection by p = kx − EA(p)k < ε as required. pk, then EA(p) = Pk=1 λkek and Pk=1 Theorem 2.6. Let A be a masa in a type II1 factor M and let ω be a free ultrafilter on N. Then the carpenter problem has a positive solution for the masa Aω ⊆ M ω. Proof. Let x ∈ Aω satisfy 0 ≤ x ≤ 1 and choose a representative (x1, x2, . . .) for x where xn ∈ A and 0 ≤ xn ≤ 1 − 1/n. By Lemma 2.5, there exist elements yn ∈ A, 0 ≤ yn ≤ 1, and projections pn ∈ M such that kxn − ynk < 1 n and EA(pn) = yn. Then (y1, y2, . . .) is also a representative for x, p = (p1, p2, . . .) is a projection in M ω, and it follows that EAω (p) = (EA(p1), EA(p2), . . .) = x. 3 Free group factors In this section we consider the carpenter problem in free group factors. Let Fn denote the free group on n generators {g1, . . . , gn}, 2 ≤ n < ∞. There are types of masas in the free group factor L(Fn) that have been much studied. Each gi generates a masa Ai, called a generator masa. The second type is the radial or Laplacian masa, whose generator is the self-adjoint element (gi + g−1 i ). We consider first the generator masa. n Pi=1 Theorem 3.1. Let g1, . . . , gn be the generators for Fn, 2 ≤ n ≤ ∞, and let Ai be the ith generator masa, where i is fixed. Given x ∈ Ai, 0 ≤ x ≤ 1, there exists a projection p ∈ L(Fn) such that EAi(p) = x. Proof. We first consider the case n = 2, and without loss of generality we take i = 1. Let S0 be an uncountable set and let S = {1, 2} ∪ S0. Then the free group factor L(FS) with generators g1, g2, and gs for s ∈ S0 contains A1 as a masa. By Theorem 2.2, there is a projection q ∈ L(FS) such that EA1(q) = x. The underlying Hilbert space L2(L(FS)) has an orthonormal basis of group elements and the Fourier series for q can only have countably many nonzero terms. Thus there is a countable subset T ⊆ S, whose elements we list as t1, t2, . . . with t1 = 1, so that q ∈ L(FT ). Define an embedding φ : L(FT ) → L(F2) on generators by φ(gti) = gi−1 , i ≥ 1. Then φ is the identity on A1, and EA1(φ(q)) = x. The desired projection is then p = φ(q). 2 g1g1−i For the general case, choose an integer j 6= i. Then Ai ⊆ L({gi, gj}) ∼= L(F2) ⊆ L(Fn), 2 and the result follows from above since the desired projection can be chosen from L(F2). For the notion of freeness that is used below, see [19] or [20]. Corollary 3.2. In a type II1 factor M with tracial state τ , if A ⊆ M is a masa and if s ∈ M is a symmetry with τ (s) = 0 and such that A and {s} are free with respect to τ , then for every x ∈ A satisfying 0 ≤ x ≤ 1, there exists a projection p ∈ M so that EA(p) = x. Proof. Since A and sAs are free and together generate a copy of L(F2), this follows from Theorem 3.1. 5 Remark 3.3. Now it is clear that if a masa A ⊆ L(Fn) is supported on at most n − 1 generators, then the carpenter problem for A has a positive solution. (cid:3) We now consider the radial masa B in L(Fn) for 2 ≤ n < ∞. Theorem 3.4. Let B be the radial masa in L(Fn) for a fixed n in the range 2 ≤ n < ∞. Given x ∈ B, 0 ≤ x ≤ 1, there exists a projection p ∈ L(Fn) so that EB(p) = x. Proof. Let g1, . . . , gn be the generators of Fn and let Ai be the ith generator masa. For each i, let hi be gi + g−1 and let Li ⊆ Ai be the abelian von Neumann algebra generated by hi. We denote by L the von Neumann algebra generated by {Li : 1 ≤ i ≤ n} which can be regarded as the free product L1 ∗ L2 ∗ · · · ∗ Ln. i If we identify Ai with L∞[−1, 1], then Li is the subalgebra of even functions. let vi ∈ Ai be the self-adjoint unitary corresponding to the odd function 1 − 2χ[0,1]. For each f ∈ Li, f vi is an odd function and so has trace 0. We now wish to show that the algebras v1Lv1, v2Lv2, . . . , vnLvn are free. Recall that the centered elements of a type II1 factor N are ◦ N = {y ∈ N : τ (y) = 0}. In order to show freeness, of the algebras viLvi, it suffices to show that the trace vanishes on finite products of the form vi1y1vi1vi2y2vi2 . . . vik ykvik (3.1) where each yi ∈ ◦ L and ij 6= ij+1 for 1 ≤ j ≤ k − 1. Products of the form zr1zr2 . . . zrs with ◦ Lri, ri 6= ri=1, span a weakly dense subspace of L so we may assume that each yi has zri ∈ this form. Consider vi1y1vi1 = vi1zr1 . . . zrsvi1. (3.2) A cancellation is only possible if zr1 ∈ Li1 ⊆ Ai1 or zrs ∈ Li1 ⊆ Ai1. In the first case, vi1zr1 ◦ Ai1, is an element of Ai1 and so corresponds to an odd function on [−1, 1]. Thus vi1zr1 ∈ ◦ Ai1. Analyzing in the same way the behavior when and similarly, if zrs ∈ Li1, then zrsvi1 ∈ each vij is adjacent to a yj leads to the conclusion that the element of (3.1) has trace 0. Thus the algebras v1Lv1, . . . , vnLvn are free, implying that v1Bv1, . . . , vnBvn are free. Thus B and v1v2Bv2v1 are free subalgebras of L(Fn) and Corollary 3.2 finishes the proof. 4 Crossed products and tensor products One of the most important masas is the Cartan masa A in the hyperfinite II1 factor R. From [5], it is unique up to isomorphisms of R. While the carpenter problem is open in this case, significant progress has been made in [1, 2]. In this section we display classes of elements in A for which a positive solution can be given. There are many ways of constructing the hyperfinite factor R. One that we will employ below is to let Z act on L∞(T) by irrational rotation, whereupon the crossed product L∞(T)⋊ Z is isomorphic to R. In keeping with our earlier techniques, we will enlarge the crossed product and exploit the nonseparability of the resulting algebra. As a vector space over the field of rationals Q, the real field R has an uncountable Hamel basis {θα : α ∈ S}, where S is an uncountable index set. For integers n1, . . . , nk+1, the 6 k Pi=1 equation niθαi = nk+1 can only be satisfied by taking all the ni's to be 0. Then the group G, defined to be the set of all finite sums {n1θα1 + · · · + nkθαk : ni ∈ Z, αi ∈ S} under Gα, where Gα = {nθα : n ∈ Z} ∼= Z. The group G acts on addition, can be expressed as Pα∈S L∞(T) by irrational rotation, and the crossed product L∞(T) ⋊ G is a type II1 factor and so has a faithful trace. Theorem 4.1. Let R be the separable hyperfinite II1 factor and let A be the Cartan masa in R. Given x ∈ A, 0 ≤ x ≤ 1, there exists a trace preserving automorphism φ of A and a projection p ∈ R such that EA(p) = φ(x). Proof. Fix x ∈ A, 0 ≤ x ≤ 1, and let φ1 : A → L∞(T) be an isomorphism that takes the trace on A to integration by Lebesgue measure on T. The algebra L∞(T) is a separable masa in the nonseparable factor L∞(T) ⋊ G, and so by Lemma 2.1 there is a projection agg q ∈ L∞(T) ⋊ G so that Eφ1(A)(q) = φ1(x). Elements of L∞(T) ⋊ G have Fourier series Pg∈G for ag ∈ L∞(T), only a countable number of whose terms are nonzero. Thus there is a countable subgroup H of G so that q ∈ L∞(T) ⋊ H ⊆ L∞(T) ⋊ G. Now (after enlarging H if needed to contain an irrational element) L∞(T) ⋊ H is a copy of R, and so the uniqueness of Cartan subalgebras in R gives an isomorphism φ2 : L∞(T) ⋊H → R so that φ2(L∞(T)) = A, and note that φ2 is trace preserving. Then EA(φ2(q)) = φ2(EL∞(T)(q)) = φ2φ1(x). (4.1) Set p = φ2(q) and φ = φ2φ1 to conclude that EA(p) = φ(x). Remark 4.2. (i) p ∈ R such that EA(p) = φ(λ1) = λ1. If x is taken to be λ1 for any λ ∈ [0, 1], then there exists a projection In Theorem 4.1, an identical proof gives a more general result: given {xi}∞ i=1 ∈ A, 0 ≤ (ii) xi ≤ 1, there exists an automorphism φ of A and projections pi ∈ R so that EA(pi) = φ(xi), i ≥ 1. (cid:3) If A is the Cartan masa in R then, by uniqueness, the inclusions A ⊆ R and A⊗A ⊆ R⊗R are equivalent. In the latter formulation we now obtain a large class of elements for which we can solve the carpenter problem. Theorem 4.3. Let A be the Cartan masa in the hyperfinite factor R and let x ∈ A, 0 ≤ x ≤ 1. Then there exists a projection p ∈ R⊗R so that EA⊗A(p) = x ⊗ 1. Proof. Let S be an uncountable set, and for each α ∈ S let Aα be a copy of A inside Rα, a Rα(cid:1), and denote the abelian subalgebra A⊗(cid:0)Nα∈S Aα(cid:1) by N. Let e ∈ M be a nonzero projection. Then there exists a countable subset T ⊆ S such We identify R with R ⊗ 1 ⊆ M. copy of R. Form M = R⊗(cid:0)Nα∈S that e lies in R⊗(cid:0)Nα∈T Rα(cid:1). Then eMe = e R⊗ (cid:0)Oα∈T Rα!! e ⊗ 7 Rα  (4.2)  Oα∈S\T e{N, e}′′e = e (A⊗ Oα∈T Aα! , e)′′! e ⊗ Aα  .  Oα∈S\T (4.3) and Thus e{N, e}′′e 6= eMe, and we can find a nonzero element z ∈ eMe such that Ee{N,e}′′e(z) = 0. It follows that EN (z) = 0, so Lemma 2.1 applies to give a projection q ∈ M such that EN (q) = x ⊗ 1. Now q is supported by R⊗(cid:0)Nα∈S2 there is an isomorphism φ : R → Nα∈S2 R ⊗(cid:0)Nα∈S2 Rα(cid:1) for a countable subset S2 of S. Then Rα(cid:1) is also an isomorphism. If we define a projection p ∈ R⊗R by p = θ−1(q), Rα which maps A to Nα∈S2 then it follows that EA⊗A(p) = x ⊗ 1 as required. Aα, and θ = 1 ⊗ φ : R⊗R → 5 The Schur -- Horn theorem Let A be a self-adjoint n×n matrix, let α1 ≥ α2 ≥ · · · ≥ αn be a decreasing rearrangement of the diagonal entries and let λ1 ≥ λ2 ≥ · · · ≥ λn be a decreasing ordering of the eigenvalues. A classical theorem of Schur [17] states that αi ≤ k Xi=1 k Xi=1 λi, 1 ≤ k ≤ n, (5.1) with equality when k = n. These inequalities can be used to define a partial ordering on general n-tuples of real numbers by α (cid:22) λ if (5.1) holds for the decreasing rearrangements of the entries, with equality when k = n. A converse to Schur's theorem was proved by Horn in [9]: if two n-tuples α and λ satisfy α (cid:22) λ then there is a self-adjoint matrix A so that the diagonal is α and the eigenvalues are the entries of λ. Collectively, these two results are known as the Schur -- Horn theorem. If we denote by EDn the conditional expectation of Mn onto the diagonal Dn, then there is an equivalent reformulation of the Schur -- Horn theorem as follows (see [10, 11]). If α and λ are n-tuples of real numbers and Dα is the diagonal matrix with entries from α, then α (cid:22) λ if and only if there exists a unitary matrix U ∈ Mn so that EDn(UDλU ∗) = Dα. (5.2) When λ has entries that are 0 or 1, then Dλ is a projection, and (5.2) reduces to a solution of the carpenter problem for the masa Dn. An appropriate formulation of the Schur -- Horn theorem for type II1 factors M with a normalized trace τ was given by Arveson and Kadison in [3] (see also the work of Hiai [7, 8]), as we now describe. For each self-adjoint a ∈ M, the distribution of a is the unique Borel probability measure ma on R so that tn dma(t) = τ (an), n = 0, 1, 2, . . . . (5.3) ZR To each Borel subset B of R, there corresponds a spectral projection eB of a, and it follows from (5.3) that ma(B) = τ (eB). Moreover, ma is supported on the spectrum σ(a) of a, and 8 is called the spectral distribution of a. Following [3, Definition 6.2], we say that a compactly supported probability measure n on R dominates a similar probability measure m on R if ZR t dm(t) =ZR m((s, ∞)) ds ≤Z ∞ t t dn(t) and n((s, ∞)) ds, t ∈ R. (5.4) Z ∞ t We deviate slightly from [3] which uses closed intervals [s, ∞), but this makes no difference since m((s, ∞)) and m([s, ∞)) can only be unequal on a countable set of s-values. For a, b ∈ Ms.a., we can then define a (cid:22) b to mean ma (cid:22) mb. This relation a (cid:22) b can be rewritten to resemble more closely the classical condition (5.1) for matrices. Instead of the eigenvalue sequence of a matrix, for a = a∗ ∈ Ms.a., we have the eigenvalue function, which is the real -- valued, monotone nonincreasing, right -- continuous function µt(a) = inf{s ∈ R : ma((s, ∞)) ≤ t} of t ∈ [0, 1). This is the unique real -- valued, nonincreasing, right -- continuous function so 0 µt(a) dE(t) for some projection -- valued measure E on [0, 1) such that τ (E([0, t))) = t (the actual measure E is obtained by reparameterizing ea). This eigenvalue function µt(a) is analogous to the decreasing eigenvalue sequence, and we have, for example, that we have a = R 1 τ (ak) =Z 1 0 (µt(a))k dt, (k ≥ 1). It is, after a change of variable, the function defined by Murray and von Neumann [13, Lemma 15.2.1] and used in various forms by several authors (e.g. [12], [14], [4]). In terms of eigenvalue functions, the relation a (cid:22) b is characterized by the inequalities Z t 0 µs(a) ds ≤Z t 0 µs(b) ds for all 0 ≤ t ≤ 1 with equality at t = 1. This follows by combining Theorem 2.1 of [1] (which comes from [7]) with Proposition 6.1 of [3]. The analog of Schur's theorem was established in [3, Theorem 7.2]: Theorem 5.1. If A is a masa in a type II1 factor M, then EA(x) (cid:22) x for all self-adjoint elements x ∈ M. Let O(x) denote the norm closure of the unitary orbit of a self-adjoint x ∈ M. Then y ∈ O(x) if and only if x and y have the same spectral data, i.e., µt(x) = µt(y) for all t ∈ [0, 1) or, equivalently mx = my. This was shown by Kamei [12], and also in [3]. The analog of Horn's theorem is then the following problem. If A is a masa in a type II1 factor M and x ∈ Ms.a. and y ∈ As.a. satisfy y (cid:22) x, does y lie in EA(O(x))? In attempting to answer this question, it is unchanged by adding multiples of the identity to x and y, and so it suffices to assume that x, y ≥ 0. For x ∈ M +, the eigenvalue function's values µt(x) 9 are actually the generalized s -- numbers of [13] (see the account in [6]), which are defined for x ∈ M and t ≥ 0 by µt(x) = inf{kxek : e ∈ P(M), τ (e) ≥ 1 − t}, (5.5) where P(M) denotes the set of projections in M. This was established in [6, Proposition 2.2]. In particular, we have µ0(x) = kxk. For x ∈ M + we will need the distribution function λt(x), defined in [6, Definition 1.3] to be so that we have λt(x) = mx(t, ∞) = τ (e(t,∞)(x)), µt(x) = inf{s ≥ 0 : λs(x) ≤ t}. Finally we will need the Ky Fan norms kxk(t) =Z t 0 µs(x) ds, 0 ≤ t ≤ 1. (5.6) (5.7) (5.8) It is not obvious that these are norms for t > 0, but this fact is established in [6, Theorem 4.4 (ii)]. Since x ∈ O(y) means that the spectral data of x and y agree, we have: Lemma 5.2. Let M be a type II1 factor. Then the following are equivalent for elements x, y ∈ M +. (i) x ∈ O(y). (ii) kxk(t) = kyk(t), 0 ≤ t ≤ 1. We now use this to investigate the Schur -- Horn theorem. Let A be a masa in a type II1 factor M, let x ∈ A+, 0 ≤ x ≤ 1, and z ∈ M + be elements such that x (cid:22) z. Then define ∆x,z = {y ∈ M : 0 ≤ y ≤ 1, y (cid:22) z, EA(y) = x}, (5.9) which is nonempty since it contains x. From the Ky Fan norm characterization above, ∆x,z is convex and it is w∗-compact from [1, Corollary 3.5]. By the Krein -- Milman theorem, ∆x,z has extreme points. Recall the definition of Φe from (2.2). Theorem 5.3. Let A be a masa in a type II1 factor M and suppose that Φe is noninjective for each nonzero projection e ∈ M. Let x ∈ A, 0 ≤ x ≤ 1 and suppose that x (cid:22) z for some element z ∈ M +. Then every extreme point of ∆x,z lies in O(z). Proof. Fix an extreme point b of ∆x,z. Then kbk(t) ≤ kzk(t) for 0 ≤ t ≤ 1 since b (cid:22) z. To derive a contradiction, suppose that there exists t so that kbk(t) < kzk(t), for otherwise Lemma 5.2 gives the result. Since kbk(0) = kzk(0) = 0, and kbk(1) = kzk(1) = τ (b), this value of t lies in (0,1). The function kzk(t) − kbk(t) is continuous on [0,1] and so attains its maximum value on a closed nonempty subset Λ ⊆ [0, 1]. Let t0 be the least value in Λ, and let t1 ∈ Λ be the largest value for which [t0, t1] ⊆ Λ. Since 0, 1 /∈ Λ, we have 0 < t0 ≤ t1 < 1, 10 and it is possible to have t0 = t1. By continuity, there exist δ0 > 0 and ε < min{t0, 1 − t1} so that kzk(t) − kbk(t) > δ0 (t ∈ [t0 − ε, t1 + ε]). (5.10) On (t0 − ε, t0), the inequality µt(z) ≤ µt(b) cannot hold everywhere because we would then have kzk(t0−ε) − kbk(t0−ε) ≥ kzk(t0) − kbk(b0), implying t0 − ε ∈ Λ and contradicting the minimal choice of t0. Thus there exist δ1 > 0 and ε0 ∈ (0, ε) so that µt0−ε0(z) ≥ µt0−ε0(b) + δ1. (5.11) Similarly, if we had µt(z) ≥ µt(b) for all t ∈ (t1, t1+ε) then it would follow that [t0, t1+ε] ⊆ Λ, contradicting the maximal choice of t1. Thus there exists ε1 ∈ (0, ε) so that µt1+ε1(z) ≤ µt1+ε1(b) − δ2 for some δ2 > 0. Clearly, the values of δ0, δ1 and δ2 can be replaced by their minimum value which we denote by δ > 0. Thus we have the inequalities kzk(t) − kbk(t) > δ, t ∈ [t0 − ε, t1 + ε] (5.12) and µt0−ε0(z) ≥ µt0−ε0(b) + δ, µt1+ε1(z) ≤ µt1+ε1(b) − δ. (5.13) Now consider the interval (t0 − ε0, t1 + ε1) to which we will associate a nonzero spectral projection e of b. There are two cases to consider. Suppose first that µs(b) takes at least three distinct values on this interval. Then there are points α1 < α2 < α3 ∈ (t0 − ε0, t1 + ε1) so that µα1(b) > µα2(b) > µα3(b). Then the open interval I = (µt1+ε1(b), µt0−ε0(b)) contains a value µα2(b) in the spectrum σ(b) of b. Secondly, if µs(b) takes at most two distinct values on (t0 − ε0, t1 + ε1) then there exists an interval on which µs(b) is constant, taking a value in the interval I. Now from (5.6) and (5.7), we see that this value in I lies in the point spectrum of b. In both cases the spectral projection e of b for the interval [µt1+ε1(b), µt0−ε0(b)] is nonzero. By hypothesis there exists a nonzero self-adjoint element w ∈ eMe so that EA(w) = 0, and by scaling we may assume that kwk < δ/2. Note that EA(b ± w) = x. We now establish that b ± w ∈ ∆x,z which will contradict the assumption that b is an extreme point. By symmetry we need only consider b + w. There are several cases. Since µs(b) is nonincreasing and right continuous, there exists r ∈ [0, t0 − ε0] so that µs(b) ≤ µt0−ε0(b) + δ/2, s ∈ [r, t0 − ε0], (5.14) while µs(b) > µt0−ε0(b) + δ/2 for s ∈ [0, r). If t ∈ [r, t0 − ε0] then kb + wk(t) = kb + wk(r) +Z t r µs(b + w) ds ≤ kbk(r) +Z t µt0−ε0(b) + δ ds ≤ kzk(r) +Z t r r µs(b) + δ/2 ds µt0−ε0(z) ds ≤ kzk(r) +Z t ≤ kzk(r) +Z t r r µs(z) ds = kzk(t), (5.15) where we have used, respectively, kwk ≤ δ/2, (5.14), (5.13), and the fact that µs(z) is nonincreasing. 11 If r = 0, then (5.15) has already handled the interval [0, t0 − ε0], so we assume that r > 0 and we now examine the interval [0, r). Fix a value s in this interval, and let f be the spectral projection of b for the interval [0, µs(b)]. By [6, Prop. 2.2], τ (f ) ≥ 1 − s. Thus µs(b + δe/2) ≤ k(b + δe/2)f k. Since e is supported on σ(b) ∩ [0, µs(b) − δ/2], we have k(b + δe/2)f k = kbf k = µs(b), which, together with [6, Lemma 2.5(iii)] implies that µs(b+w) ≤ µs(b+δe/2) ≤ µs(b). It follows by integrating these inequalities that kb+wk(t) ≤ kbk(t) ≤ kzk(t) for t ∈ [0, r). On [t0 − ε0, t1 + ε1], using kwk ≤ δ/2 and (5.12), we have kb + wk(t) ≤ kbk(t) + δ/2 ≤ kzk(t), (5.16) so it remains to consider [t1 + ε1, 1], which is handled in a similar manner to [0, t0 − ε0]. Let r′ ∈ (t1 + ε1, 1] be the maximum value so that µs(b) > µt1+ε1(b) − δ/2 for s ∈ [t1 + ε1, r′). This value exists since µs(b) is monotone nonincreasing and right continuous. First consider the case r′ < 1. Then the inequality µs(b) ≤ µt1+ε1(b) − δ/2 holds for s ∈ [r′, 1]. We take s ∈ [r′, 1]. Then the spectral projection g of b for the interval [0, µs(b)] is orthogonal to e and has trace at least 1 − s, by [6, Prop. 2.2]. Thus µs(b + w) = k(b + δe/2)gk = kbgk = µs(b), (5.17) for s ∈ [r′, 1]. Thus if t ∈ [r′, 1], then (see [6, Prop. 2.7]) Z 1 t µs(b + w) ds =Z 1 t µs(b) ds = τ (b) − kbk(t) ≥ τ (z) − kzk(t) =Z 1 t µs(z) ds. (5.18) Since τ (b + w) = τ (b) = τ (z), we have kb + wk(t) ≤ kzk(t) on this interval. Let s ∈ [t1 + ε1, r′). Then µs(b + w) ≥ µs(b − δe/2) ≥ µs(b) − δ/2 ≥ µt1+ε1(b) − δ ≥ µt1+ε1(z) ≥ µs(z). (5.19) If t ∈ [t1 + ε1, r′), then Z 1 t µs(b + w) ds =Z r′ ≥Z r′ t t µs(b + w) ds +Z 1 µs(z) ds +Z 1 r′ r′ µs(b + w) ds µs(z) ds =Z 1 t µs(z) ds, (5.20) where we have used (5.18) with t = r′. Since τ (b + w) = τ (b) = τ (z), we have the inequality kb + wk(t) ≤ kzk(t) on this interval also. This shows that b + w ∈ ∆x,z, completing the proof in the case r′ < 1. If r′ = 1, then the proof of (5.20) for t ∈ [t1 + ε1, 1) is exactly as before, and this suffices to prove b + w ∈ ∆x,z. 12 As a consequence of Theorem 5.3, we can immediately deduce two corollaries whose proofs are so similar to those of Theorems 3.1, 3.4 and 4.1 that we omit the details. The only minor change is that when passing from the augmented algebras back to the original ones, instead of observing that a certain element is countably supported we need this for a countable set of elements, which is of course true. Corollary 5.4. Let A be either a generator masa or the radial masa in L(Fn), 2 ≤ n < ∞. If x ∈ A+, z ∈ L(Fn)+ and x (cid:22) z then x ∈ EA(O(z)). Corollary 5.5. If A is the Cartan masa in the hyperfinite factor R and x ∈ A+, z ∈ R+ satisfy x (cid:22) z, then there is a trace preserving automorphism θ of A so that θ(x) ∈ EA(O(z)). In [1], {x : x (cid:22) z} was shown to be the σ-SOT closure of EA(O(z)) for general masas A. In the case of the Cartan masa, we can improve this to norm closure. We will need a simple preliminary lemma. Lemma 5.6. Let A be the Cartan masa in the hyperfinite factor R. Given two sets of orthogonal projections {p1, . . . , pn} and {q1, . . . , qn} in A satisfying τ (pi) = τ (qi), 1 ≤ i ≤ n, there exists a unitary normalizer u of A so that upiu∗ = qi, 1 ≤ i ≤ n. Proof. We proceed by induction on the number n of projections. The case n = 1 is proved in [16] (see also [18, Lemma 6.2.6]), so suppose that the result is true for n − 1 projec- tions. Choose a unitary normalizer u1 of A so that u1p1u∗ 1 = q1, and consider the sets of projections {p1, . . . , pn} and {u∗ 1qnu1}. Since A(1 − p1) is a Cartan masa in (1−p1)R(1−p1), we may apply the induction hypothesis to the sets of projections {p2, . . . , pn} and {u∗ 1qnu1} in A(1 − p1) to obtain a unitary normalizer w ∈ (1 − p1)R(1 − p1) of A(1 − p1) so that wpiw∗ = u∗ 1qiu1 for 2 ≤ i ≤ n. This extends to a unitary normalizer v = w + p1 of A in R. The proof is completed by defining u to be u1v, so that upiu∗ = qi for 1 ≤ i ≤ n. 1q2u1, . . . , u∗ 1q1u1, . . . , u∗ Corollary 5.7. Let A be the Cartan masa in the hyperfinite factor R. If x ∈ A+, z ∈ R+ and x (cid:22) z, then x ∈ EA(O(z)) (norm closure). Proof. By Corollary 5.5 there exists a trace preserving automorphism θ of A so that θ(x) ∈ EA(O(z)). Given ε > 0, there exist projections pi ∈ A, 1 ≤ i ≤ n and positive constants λi, 1 ≤ i ≤ n, so that x − < ε (5.21) pi = 1. Since τ (pi) = τ (θ(pi)), Lemma 5.6 gives a unitary normalizer u of (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) n Xi=1 λipi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) and such that A satisfying n Pi=1 u∗piu = θ(pi), 1 ≤ i ≤ n. Choose unitaries vn ∈ R such that lim n→∞ kθ(x) − EA(vnzv∗ n)k = 0. 13 (5.22) (5.23) From (5.21), (5.22) and (5.23) it follows that lim sup n→∞ kx − uEA(vnzv∗ n)u∗k ≤ ε. (5.24) But uniqueness of the conditional expectation gives uEA(vnzv∗ result follows from (5.24). n)u∗ = EA(uvnzv∗ nu∗), and the References [1] M. Argerami and P. Massey, A Schur-Horn theorem in II1 factors. Indiana Univ. Math. J. 56 (2007), 2051 -- 2059. [2] M. Argerami and P. Massey, Towards the carpenter's theorem. Proc. Amer. Math. Soc. 137 (2009), 3679 -- 3687. [3] W. Arveson and R. V. Kadison, Diagonals of self-adjoint operators. Operator theory, operator algebras, and applications, 247 -- 263, Contemp. Math., 414, Amer. Math. Soc., Providence, RI, 2006. [4] H. Bercovici and W.S. Li, Eigenvalue inequalities in an embeddable factor, Proc. Amer. Math. Soc. 134 (2006), 75 -- 80. [5] A. Connes, J. Feldman and B. Weiss, An amenable equivalence relation is generated by a single transformation. Ergodic Theory Dynam. Systems 1 (1981), 431 -- 450. [6] T. Fack and H. Kosaki, Generalized s-numbers of τ -measurable operators, Pacific J.Math. 123 (1986), 269 -- 300. [7] F. Hiai, Majorization and stochastic maps in von Neumann algebras, J. Math. Anal. Appl. 127 (1987), 18 -- 48. [8] F. Hiai, Spectral majorization between normal operators in von Neumann algebras, Operator algebras and operator theory (Craiova, 1989), Pitman Res. Notes Math. Ser., 271 78 -- 115, Longman Sci. Tech., Harlow, 1992. [9] A. Horn, Doubly stochastic matrices and the diagonal of a rotation matrix. Amer. J. Math. 76 (1954). 620 -- 630. [10] R. V. Kadison, The Pythagorean theorem. I. The finite case, Proc. Natl. Acad. Sci. USA 99 (2002), 4178 -- 4184. [11] R. V. Kadison, The Pythagorean theorem. II. The infinite discrete case, Proc. Natl. Acad. Sci. USA 99 (2002), 5217 -- 5222. [12] E. Kamei, Majorization in finite factors, Math. Japon. 28 (1983), 495 -- 499. [13] F. J. Murray and J. von Neumann, On rings of operators. Ann. of Math. (2) 37 (1936), 116 -- 229 14 [14] D. Petz, Spectral scale of selfadjoint operators and trace inequalities, J. Math. Anal. Appl. 109 (1985), 74 -- 82. [15] S. Popa, Orthogonal pairs of *-subalgebras in finite von Neumann algebras, J. Oper. Theory, 9 (1983), 253-268. [16] S. Popa, Notes on Cartan subalgebras in type II1 factors. Math. Scand. 57 (1985), 171 -- 188. [17] I. Schur, Uber eine Klasse von Mittlebildungen mit Anwendungen auf der Determinan- tentheorie, Sitzungsber. Berliner Mat. Ges. 22 (1923), 9 -- 29. [18] A. M. Sinclair and R. R. Smith, Finite von Neumann algebras and masas. London Math- ematical Society Lecture Note Series, 351. Cambridge University Press, Cambridge, 2008. [19] D. Voiculescu, Symmetries of some reduced free product C∗ -- algebras, Operator Alge- bras, Unitary Representations, Enveloping Algebras, and Invariant Theory, 556 -- 588, Lecture Notes in Mathematics, vol. 1132, Springer -- Verlag, 1985. [20] D. Voiculescu, K. Dykema, A. Nica, Free random variables, CRM Monograph Series, vol. 1, American Mathematical Society, Providence, RI, 1992. Update: After circulating a version of this preprint, we learned of independent work of C. Akemann, D. Sherman and N. Weaver that may overlap with some of our results. Contact Information Kenneth J. Dykema Department of Mathematics, Texas A&M University, College Station, TX 77843 E-mail address: [Dykema] [email protected] Junsheng Fang School of Mathematical Sciences, Dalian University of Technology, Dalian, China E-mail address: [Fang] [email protected] Donald W. Hadwin Department of Mathematics, University of New Hampshire, Durham NH 03824 E-mail address: [Hadwin] [email protected] Roger R. Smith Department of Mathematics, Texas A&M University, College Station, TX 77843 E-mail address: [Smith] [email protected] 15
0911.2999	2	0911	2011-07-11T14:55:34	The Baum-Connes conjecture for free orthogonal quantum groups	[ "math.OA", "math.KT" ]	We prove an analogue of the Baum-Connes conjecture for free orthogonal quantum groups. More precisely, we show that these quantum groups have a $ \gamma $-element and that $ \gamma = 1 $. It follows that free orthogonal quantum groups are $ K $-amenable. We compute explicitly their $ K $-theory and deduce in the unimodular case that the corresponding reduced $ C^* $-algebras do not contain nontrivial idempotents. Our approach is based on the reformulation of the Baum-Connes conjecture by Meyer and Nest using the language of triangulated categories. An important ingredient is the theory of monoidal equivalence of compact quantum groups developed by Bichon, De Rijdt and Vaes. This allows us to study the problem in terms of the quantum group $ SU_q(2) $. The crucial part of the argument is a detailed analysis of the equivariant Kasparov theory of the standard Podle\'s sphere.	math.OA	math	THE BAUM-CONNES CONJECTURE FOR FREE ORTHOGONAL QUANTUM GROUPS CHRISTIAN VOIGT Abstract. We prove an analogue of the Baum-Connes conjecture for free orthogonal quantum groups. More precisely, we show that these quantum groups have a γ-element and that γ = 1. It follows that free orthogonal quantum groups are K-amenable. We compute explicitly their K-theory and deduce in the unimodular case that the corresponding reduced C ∗-algebras do not contain nontrivial idempotents. Our approach is based on the reformulation of the Baum-Connes conjecture by Meyer and Nest using the language of triangulated categories. An important ingredient is the theory of monoidal equivalence of compact quantum groups developed by Bichon, De Rijdt and Vaes. This allows us to study the problem in terms of the quantum group SUq(2). The crucial part of the argument is a detailed analysis of the equivariant Kasparov theory of the standard Podle´s sphere. 1. Introduction Let G be a second countable locally compact group and let A be a separable G-C ∗-algebra. The Baum-Connes conjecture with coefficients in A asserts that the assembly map µA : K top ∗ (G; A) → K∗(G ⋉r A) is an isomorphism [6], [7]. Here K∗(G ⋉r A) is the K-theory of the reduced crossed product of A by G. The validity of this conjecture has applications in topology, geometry and representation theory. In particular, if G is discrete then the Baum- Connes conjecture with trivial coefficients C implies the Novikov conjecture on higher signatures and the Kadison-Kaplansky idempotent conjecture. Meyer and Nest have reformulated the Baum-Connes conjecture using the lan- guage of triangulated categories and derived functors [20]. In this approach the left hand side of the assembly map is identified with the localisation LF of the functor F (A) = K∗(G ⋉r A) on the equivariant Kasparov category KK G. Among other things, this description allows to establish permanence properties of the conjecture in an efficient way. In addition, the approach in [20] is a natural starting point to study an analogue of the Baum-Connes conjecture for locally compact quantum groups. The usual definition of the left hand side of the conjecture is based on the universal space for proper actions, a concept which does not translate to the quantum setting in an obvious way. Following [20], one has to specify instead an appropriate subcategory of the equivariant Kasparov category corresponding to compactly induced actions in the group case. This approach has been implemented in [21] where a strong form of the Baum-Connes conjecture for duals of compact groups is established. Duals of compact groups are, in a sense, the most basic examples of discrete quantum groups. In this paper we develop these ideas further and prove an analogue of the Baum- Connes conjecture for free orthogonal quantum groups. These quantum groups, 2000 Mathematics Subject Classification. 20G42, 46L80, 19K35. 1 2 CHRISTIAN VOIGT introduced by Wang and van Daele [34], [31], can be considered as quantum analogs of orthogonal matrix Lie groups. If Q ∈ GLn(C) is a matrix satisfying QQ = ±1 then the free orthogonal quantum group Ao(Q) is the universal C ∗-algebra gener- ated by the entries of a unitary n × n-matrix u satisfying the relations u = QuQ−1. In this paper we will use the notation Ao(Q) = C ∗ f (FO(Q)) in order to emphasize that we view this C ∗-algebra as the full group C ∗-algebra of a discrete quantum group. Accordingly, we will refer to FO(Q) as the free orthogonal quantum group associated to Q. In the case that Q = 1n ∈ GLn(C) is the identity matrix we simply write FO(n) instead of FO(1n). In fact, this special case illustrates the link to classical orthogonal groups since the algebra C(O(n)) of functions on O(n) is the abelianization of C ∗ f (FO(n)). It is known [4] that the quantum group FO(Q) is not amenable if Q ∈ GLn(C) for n > 2. The main result of this paper is that FO(Q) has a γ-element and that γ = 1 for Q ∈ GLn(C) and n > 2. The precise meaning of this statement, also referred to as the strong Baum-Connes conjecture, will be explained in section 5 using the language of triangulated categories. However, we point out that triangulated cate- gories are not needed to describe the applications that motivated our study. Firstly, it follows that free orthogonal quantum groups are K-amenable. This answers in an affirmative way a question arising from the work of Vergnioux on quantum Cayley trees [33]. Secondly, by studying the left hand side of the assembly map we obtain an explicit calculation of the K-theory of FO(Q). In the same way as in the case of free groups, the result of this calculation implies that the reduced group C ∗-algebra of FO(n) does not contain nontrivial idempotents. This may be regarded as an analogue of the Kadison-Kaplansky conjecture. Our results support the point of view that free quantum groups behave like free groups in many respects. By work of Vaes and Vergnioux [30] it is known, for instance, that the reduced C ∗-algebra C ∗ r (FO(n)) of FO(n) is exact and simple for n > 2. Moreover, the associated von Neumann algebra L(FO(n)) is a full and prime factor. Let us note that, in contrast to the case of free groups, even the K-theory of the maximal C ∗-algebras of orthogonal quantum groups seems difficult to compute directly. As already mentioned above, our approach is motivated from the work of Meyer and Nest [20]. In fact, the definition of the assembly map for torsion-free quantum groups proposed by Meyer in [19] is the starting point of this paper. We proceed by observing that the strong Baum-Connes conjecture for torsion-free quantum groups is invariant under monoidal equivalence. The theory of monoidal equivalence for compact quantum groups was developed by Bichon, De Rijdt, and Vaes [8]. We use it to translate the Baum-Connes problem for free orthogonal quantum groups into a specific problem concerning SUq(2). This step builds on the results in [8] and the foundational work of Banica [3]. The crucial part of our argument is a precise study of the equivariant KK-theory of the standard Podle´s sphere. By definition, the Podle´s sphere SUq(2)/T is the homogeneous space of SUq(2) with respect to the classical maximal torus T ⊂ SUq(2). Our constructions in connection with the Podle´s sphere rely on the considerations in [25]. Finally, the K-theory computation for FO(Q) involves some facts from homological algebra for triangulated categories worked out in [19]. Let us describe how the paper is organized. In section 2 we discuss some pre- liminaries on compact quantum groups. In particular, we review the construction of spectral subspaces for actions of compact quantum groups on C ∗-algebras and Hilbert modules. Section 3 contains basic definitions related to SUq(2) and the standard Podle´s sphere SUq(2)/T . Moreover, it is shown that the dual of SUq(2) can be viewed as a torsion-free discrete quantum group in the sense of [19]. The BAUM-CONNES CONJECTURE 3 most technical part of the paper is section 4 which contains our results on the Podle´s sphere. In section 5 we explain the formulation of the Baum-Connes conjecture for torsion-free quantum groups proposed in [19]. Using the considerations from section 4 we prove that the dual of SUq(2) satisfies the strong Baum-Connes conjecture in section 6. Section 7 contains the definition of free orthogonal quantum groups and a brief review of the theory of monoidal equivalence for compact quantum groups [8]. In section 8 we show that monoidally equivalent compact quantum groups have equivalent equivariant KK-categories. This implies that the strong Baum-Connes property is invariant under monoidal equivalence. Due to the work in [8] and our results in section 6 it follows that free orthogonal quantum groups satisfy the strong Baum-Connes conjecture. Finally, in section 9 we discuss applications and conse- quences. Let us make some remarks on notation. We write L(E) for the space of adjointable operators on a Hilbert A-module E. Moreover K(E) denotes the space of compact operators. The closed linear span of a subset X of a Banach space is denoted by [X]. Depending on the context, the symbol ⊗ denotes either the tensor product of Hilbert spaces, the minimal tensor product of C ∗-algebras, or the tensor product of von Neumann algebras. We write ⊙ for algebraic tensor products. For operators on multiple tensor products we use the leg numbering notation. It is a pleasure to thank U. Krahmer, R. Meyer, R. Nest and N. Vander Vennet for helpful discussions on the subject of this paper. 2. Compact quantum groups and spectral decomposition Concerning the general theory of quantum groups, we assume that the reader is familiar with the definitions and constructions that are reviewed in the first section of [25]. For more information we refer to the literature [2], [17], [18], [29], [37]. Unless explicitly stated otherwise, our notation and conventions will follow [25] throughout the paper. The purpose of this section is to explain some specific preliminaries on compact quantum groups. In particular, we discuss the construction of spectral subspaces for actions of compact quantum groups on C ∗-algebras and Hilbert modules. Let G be a compact quantum group. Since G is compact the corresponding reduced C ∗-algebra of functions C r(G) is unital. A (unitary) representation π of G on a Hilbert space Hπ is an invertible (unitary) element uπ ∈ L(C r(G) ⊗ Hπ) satisfying the relation (∆ ⊗ id)(uπ) = uπ 13uπ 23. That is, a unitary representation of G is the same thing as a unitary corepresenta- tion of the Hopf-C ∗-algebra C r(G). Let π, η be representations of G on the Hilbert spaces Hπ, Hη, given by the invertible elements uπ ∈ L(C r(G) ⊗ Hπ) and uη ∈ L(C r(G) ⊗ Hη), respectively. An operator T in L(Hπ, Hη) is called an intertwiner between π and η if (id ⊗T )uπ = uη(id ⊗T ). We will denote the space of intertwiners between Hπ and Hη by Mor(Hπ, Hη). The representations π and η are equivalent iff Mor(Hπ, Hη) contains an invertible operator. Every unitary representation of a compact quantum group decomposes into a direct sum of irreducibles, and all irreducible representations are finite di- mensional. Every finite dimensional representation is equivalent to a unitary rep- resentation, and according to Schur's lemma a representation π is irreducible iff dim(Mor(Hπ, Hπ)) = 1. By slight abuse of notation, we will sometimes write G for the set of isomorphism classes of irreducible unitary representations of G. The trivial representation of G on the one-dimensional Hilbert space is denoted by ǫ. The tensor product of the representations π and η is the representation π ⊗ η on 4 CHRISTIAN VOIGT 13uπ Hπ ⊗ Hη given by uπ⊗η = uη 12 ∈ L(C r(G) ⊗ Hπ ⊗ Hη). The class of all fi- nite dimensional representations of G together with the intertwining operators as morphisms and the direct sum and tensor product operations yields a C ∗-tensor category Rep(G). This category is called the representation category of G. By construction, it comes equipped with a canonical C ∗-tensor functor to the category of finite dimensional Hilbert spaces. Let π be a finite dimensional representation given by uπ ∈ L(C r(G) ⊗ Hπ), and let dim(π) = dim(Hπ) = n be the dimension of the underlying Hilbert space. If eπ 1 , . . . , eπ n is an orthonormal basis for Hπ we obtain associated matrix elements uπ ij ∈ C r(G) given by j )i. The corepresentation identity for uπ corresponds to uπ ij = heπ i , uπ(eπ n ∆(uπ ij) = uπ ik ⊗ uπ kj Xk=1 for 1 ≤ i, j ≤ n. Conversely, a (unitary) invertible matrix u = (uij) ∈ Mn(C r(G)) = L(C r(G) ⊗ Cn) satisfying these relations yields a (unitary) representation of G. The linear span of the matrix elements of π ∈ G forms a finite dimensional coalgebra C[G]π ⊂ C r(G). Moreover C[G] = Mπ∈ G C[G]π is a dense Hopf ∗-algebra C[G] ⊂ C r(G) by the Peter-Weyl theorem. In subsequent sections we will use the fact that similar spectral decompositions exist for arbitrary G-C ∗-algebras and G-Hilbert modules. In order to discuss this we review some further definitions and results. Let π be an irreducible unitary representation of G, and let uπ ij be the matrix elements in some fixed basis. The contragredient representation πc is given by the matrix (uπc In general πc is not unitary, but as any finite dimensional representation of G it is unitarizable. The representations π and πcc are equivalent, and there exists a unique positive invertible intertwiner Fπ ∈ Mor(Hπ, Hπcc) satisfying tr(Fπ) = tr(F −1 π ). The trace of Fπ is called the quantum dimension of π and denoted by dimq(π). With this notation, the Schur orthogonality relations are ji) where S is the antipode of C[G]. )ij = S(uπ φ(uπ ij(uη kl)∗) = δπηδik 1 dimq(π) (Fπ)lj where π, η ∈ G and φ : C r(G) → C is the Haar state of G. In the sequel we shall fix bases such that Fπ is a diagonal operator for all π ∈ G. Let π be a unitary representation of G with matrix elements uπ ij. The element dim(π) χπ = uπ jj Xj=1 in C r(G) is called the character of π. The subring in C r(G) generated by the charac- ters of unitary representations is isomorphic to the (opposite of the) representation ring of G. Let us now fix our terminology concerning coactions. By an algebraic coaction of C[G] on a vector space M we mean an injective linear map γ : M → C[G] ⊙ M such that (id ⊙γ)γ = (∆ ⊙ id)γ. For an algebraic coaction one always has (ǫ ⊙ id)γ = id and (C[G] ⊙ 1)γ(M ) = C[G] ⊙ M . Hence a vector space together with an algebraic coaction of C[G] is the same thing as a (left) C[G]-comodule. Accordingly one de- fines right coactions on vector spaces. BAUM-CONNES CONJECTURE 5 By an algebraic coaction of C[G] on a ∗-algebra A we shall mean an injective ∗- homomorphism α : A → C[G] ⊙ A such that (id ⊙α)α = (∆ ⊙ id)α. Accordingly one defines right coactions on ∗-algebras. A ∗-algebra equipped with an algebraic coaction of C[G] will also be called a G-algebra. Let G be a compact quantum group and let A be a G-C ∗-algebra with coac- tion α : A → M (C r(G) ⊗ A). Since G is compact, such a coaction is an injec- tive ∗-homomorphism α : A → C r(G) ⊗ A satisfying the coassociativity identity (∆ ⊗ id)α = (id ⊗α)α and the density condition [(C r(G) ⊗ 1)α(A)] = C r(G) ⊗ A. For π ∈ G we let Aπ = {a ∈ Aα(a) ∈ C[G]π ⊙ A} be the π-spectral subspace of A. The subspace Aπ is a closed in A, and there is a projection operator pπ : A → Aπ defined by where pπ(a) = (θπ ⊗ id)α(a) dim(π) θπ(x) = dimq(π) Xj=1 (Fπ)−1 jj φ(x(uπ jj )∗). The spectral subalgebra S(A) ⊂ A is the ∗-subalgebra defined by S(A) = SG(A) = Mπ∈ G Aπ, and we note that S(A) is a G-algebra in a canonical way. From the Schur orthog- onality relations and [(C r(G) ⊗ 1)α(A)] = C r(G) ⊗ A it is easy to check that S(A) is dense in A, compare [27]. In a similar way one defines the spectral decomposition of G-Hilbert modules. Let EA be a G-Hilbert A-module over the G-C ∗-algebra A. Since G is compact, the cor- responding coaction γ : E → M (C r(G)⊗E) is an injective linear map E → C r(G)⊗E satisfying the coaction identity (∆ ⊗ id)γ = (id ⊗γ)γ and the density condition [(C r(G) ⊗ 1)γ(E)] = C r(G) ⊗ E. For π ∈ G we let Eπ = {ξ ∈ Eγ(ξ) ∈ C[G]π ⊙ E} be the corresponding spectral subspace. As in the algebra case, the spectral sub- space Eπ is closed in E, and there is a projection map pπ : E → Eπ. By definition, the spectral submodule of E is the dense subspace S(E) = Mπ∈ G Eπ of E. The spectral submodule S(E) is in fact a right S(A)-module, and the scalar product of E restricts to an S(A)-valued inner product on S(E). In this way S(E) becomes a pre-Hilbert S(A)-module. For a G-algebra A the spectral subspace Aπ for π ∈ G is defined in the same way. This yields a corresponding spectral decomposition of A, the difference to the C ∗- setting is that we always have S(A) = A in this case. The same remark applies to coactions of C[G] on arbitrary vector spaces. If H is another compact quantum group and M is a C[G]-C[H]-bicomodule, we will also write πM for the π-spectral subspace corresponding to the left coaction. Finally, we recall the definition of cotensor products. Let M be a right C[G]- comodule with coaction ρ : M → M ⊙ C[G] and N be a left C[G]-comodule with coaction λ : N → C[G] ⊙ N . The cotensor product of M and N is the equalizer M (cid:3)C[G]N / M ⊙ N / M ⊙ C[G] ⊙ N of the maps id ⊙λ and ρ ⊙ id. / / / / 6 CHRISTIAN VOIGT 3. The quantum group SUq(2) In this section we review some definitions and constructions related to SUq(2) [35]. Moreover we show that the dual discrete quantum group of SUq(2) is torsion- free in a suitable sense. Throughout we consider q ∈ [−1, 1] \ {0}, at some points we will exclude the cases q = ±1. By definition, C(SUq(2)) is the universal C ∗-algebra generated by elements α and γ satisfying the relations αγ = qγα, αγ ∗ = qγ ∗α, γγ ∗ = γ ∗γ, α∗α + γ ∗γ = 1, αα∗ + q2γγ ∗ = 1. These relations are equivalent to saying that the fundamental matrix u =(cid:18)α −qγ ∗ α∗ (cid:19) γ is unitary. The comultiplication ∆ : C(SUq(2)) → C(SUq(2)) ⊗ C(SUq(2)) is given on the generators by ∆(α) = α ⊗ α − qγ ∗ ⊗ γ, ∆(γ) = γ ⊗ α + α∗ ⊗ γ, and in this way the compact quantum group SUq(2) is defined. We remark that there is no need to distinguish between full and reduced C ∗-algebras here since SUq(2) is coamenable, see [5]. The Hopf ∗-algebra C[SUq(2)] is the dense ∗-subalgebra of C(SUq(2)) generated by α and γ, with counit ǫ : C[SUq(2)] → C and antipode S : C[SUq(2)] → C[SUq(2)] determined by ǫ(α) = 1, ǫ(γ) = 0 and S(α) = α∗, S(α∗) = α, S(γ) = −qγ, S(γ ∗) = −q−1γ ∗, respectively. We use the Sweedler notation ∆(x) = x(1) ⊗ x(2) for the comultiplica- tion and write f ⇀ x = x(1)f (x(2)), x ↼ f = f (x(1))x(2) for elements x ∈ C[SUq(2)] and linear functionals f : C[SUq(2)] → C. The antipode is an algebra antihomomorphism satisfying S(S(x∗)∗) = x for all x ∈ C[SUq(2)]. In particular the map S is invertible, and the inverse of S can be written as S−1(x) = f1 ⇀ S(x) ↼ f−1 where f1 : C[SUq(2)] → C is the modular character given by f1(α) = q−1, f1(α∗) = q, f1(γ) = 0, f1(γ ∗) = 0 and f−1 is defined by f−1(x) = f (S(x)). These maps are actually members of a canonical family (fz)z∈C of characters. The character f1 describes the modular properties of the Haar state φ of C(SUq(2)) in the sense that φ(xy) = φ(y(f1 ⇀ x ↼ f1)) for all x, y ∈ C[SUq(2)]. For q ∈ (−1, 1) \ {0} we denote by Uq(sl(2)) the quantized universal enveloping algebra of sl(2). This is the algebra generated by the elements E, F, K such that K is invertible and the relations KEK −1 = q2E, KF K −1 = q−2F, [E, F ] = K − K −1 q − q−1 BAUM-CONNES CONJECTURE 7 are satisfied. We consider Uq(sl(2)) with its Hopf algebra structure determined by ∆(K) = K ⊗ K, ∆(E) = E ⊗ K + 1 ⊗ E, ∆(F ) = F ⊗ 1 + K −1 ⊗ F ǫ(K) = 1, ǫ(E) = 0, ǫ(F ) = 0 S(K) = K −1, S(E) = − EK −1, S(F ) = −KF and the ∗-structure defining the compact real form, explicitly K ∗ = K, E∗ = sgn(q)F K, F ∗ = sgn(q)K −1E where sgn(q) denotes the sign of q. Let us remark that in the literature sometimes a wrong ∗-structure is used in the case q < 0. There is a nondegenerate skew-pairing between the Hopf-∗-algebras Uq(sl(2)) and C[SUq(2)], compare [16]. In particular, every finite dimensional unitary corepre- sentation of C(SUq(2)) corresponds to a finite dimensional unital ∗-representation of Uq(sl(2)). Recall that a discrete group is called torsion-free if it does not contain nontrivial elements of finite order. For discrete quantum groups the following definition was proposed by Meyer [19]. Definition 3.1. A discrete quantum group G is called torsion-free iff every finite dimensional G-C ∗-algebra for the dual compact quantum group G is isomorphic to a direct sum of G-C ∗-algebras that are equivariantly Morita equivalent to C. In other words, according to definition 3.1, a discrete quantum group G is torsion- free if for every finite dimensional G-C ∗-algebra A there are finite dimensional Hilbert spaces H1, . . . , Hl and unitary corepresentations uj ∈ L(C r( G) ⊗ Hj) such that A is isomorphic to K(H1) ⊕ · · · ⊕ K(Hl) as a G-C ∗-algebra. Here each matrix block K(Hj) is equipped with the adjoint action corresponding to uj. If G is a discrete group this is equivalent to the usual notion of torsion-freeness. Definition 3.1 is motivated from the study of torsion phenomena that occur for coactions of compact groups [21]. Hence it is not surprising that it also provides the correct picture for duals of q-deformations. We shall discuss explicitly the case of SUq(2). Proposition 3.2. Let q ∈ (−1, 1) \ {0}. Then the discrete quantum group dual to SUq(2) is torsion-free. Proof. The following argument was suggested by U. Krahmer. For simplicity we restrict ourselves to the case q > 0, the case of negative q is treated in a similar way. Let us assume that A is a finite dimensional SUq(2)-C ∗-algebra with coaction α : A → C(SUq(2)) ⊗ A. Since A is finite dimensional we may write A = Mn1(C) ⊕ · · · ⊕ Mnl(C). The task is to describe the coaction in terms of this decomposition. First consider the restriction of α to C(T ). Since the torus T is a connected group, the corresponding action of T preserves the decomposition of A into matrix blocks. Moreover, on each matrix block the action arises from a representation of T . Hence, if we view A as a Uq(sl(2))-module algebra, the action of K is implemented by conjugating with an invertible self-adjoint matrix k = k1 ⊕ · · · ⊕ kl. Moreover we may suppose that k has only positive eigenvalues. With these requirements each of the matrices kj is uniquely determined up to a positive scalar factor. Next consider the skew-primitive elements E and F . From the definition of the comultiplication in Uq(sl(2)) we obtain E · (ab) = (E · a)(K · b) + a(E · b), F · (ab) = F (a)b + (K −1 · a)(F · b) for all a, b ∈ A. Hence E and F can be viewed as Hochschild-1-cocycles on A with values in appropriate A-bimodules. Since A is a semisimple algebra the correspond- ing Hochschild cohomology groups vanish. Consequently there are e, f ∈ A such 8 that CHRISTIAN VOIGT E · a = ek−1(K · a) − aek−1 F · a = f a − (K −1 · a)f for all a ∈ A. It follows that E and F preserve the decomposition of A into matrix blocks. In particular, we may restrict attention to the case that A is a simple matrix algebra. Let us assume A = Mn(C) in the sequel. Then e and f are uniquely determined up to addition of a scalar multiple of 1 and k−1, respectively. The relation KEK −1 = q2E implies k(ek−1a − k−1akek−1)k−1 = kek−1ak−1 − akek−2 = q2(eak−1 − aek−1) for all a ∈ A and yields kek−1 − q2e = λ for some λ ∈ C. If we replace e by e − λ(1 − q2)−1 we obtain kek−1 = q2e. Similarly we may achieve kf k−1 = q−2f , and we fix e and f such that these identities hold. The commutation relation for E and F implies (cid:18)e(f a − k−1akf )k−1 − (f a − k−1akf )ek−1(cid:19) = (ef − f e)ak−1 − k−1a(ef − f e) −(cid:18)f (eak−1 − aek−1) − k−1(eak−1 − aek−1)kf(cid:19) q − q−1(cid:18)kak−1 − k−1ak(cid:19). 1 = As a consequence we obtain ef − f e − k q − q−1 = − µ q − q−1 k−1 for some constant µ. In fact, since k has only positive eigenvalues the scalar µ is strictly positive. Replacing k by λk and e by λe with λ = µ−1/2 yields [e, f ] = k − k−1 q − q−1 . It follows that there is a representation of Uq(sl(2)) on Cn which induces the given action on A by conjugation. We have (E · a)∗ = −F · a∗ for all a ∈ A and hence a∗(e∗ − kf ) = (e∗ − kf )a∗ which implies e∗ − kf = ν for some ν ∈ C. Conjugating with k yields ν = k−1(e∗ − kf )k = (kek−1)∗ − q2kf = q2(e∗ − kf ) = q2ν and thus ν = 0. It follows that the representation of Uq(sl(2)) given by e, f and k is a ∗-representation. We conclude that there exists a unitary corepresentation of C(SUq(2)) on Cn which implements the coaction on A as desired. (cid:3) Let us next discuss the regular representation of SUq(2) for q ∈ (−1, 1) \ {0}. We write L2(SUq(2)) for the Hilbert space obtained using the inner product hx, yi = φ(x∗y) on C(SUq(2)). By definition, the regular representation on L2(SUq(2)) is given by the multiplicative unitary W ∈ L(C(SUq(2)) ⊗ L2(SUq(2)). The Peter-Weyl theory describes the decomposition of this representation into ir- reducibles. As in the classical case, the irreducible representations of SUq(2) are labelled by half-integers l, and the corresponding Hilbert spaces have dimension BAUM-CONNES CONJECTURE 9 2l + 1. The matrix elements u(l) onal set in L2(SUq(2)). Moreover, if we write ij with respect to weight bases determine an orthog- [a] = qa − q−a q − q−1 for the q-number associated to a ∈ C, then the vectors e(l) i,j = qi[2l + 1] 1 2 u(l) i,j form an orthonormal basis of L2(SUq(2)), compare [16]. The regular representation of C(SUq(2)) on L2(SUq(2)) is given by α e(l) i,j = a+(l, i, j) e γ e(l) i,j = c+(l, i, j) e (l+ 1 i− 1 (l+ 1 i+ 1 2 ) 2 ,j− 1 2 ) 2 ,j− 1 2 2 + a−(l, i, j) e + c−(l, i, j) e (l− 1 i− 1 (l− 1 i+ 1 2 ) 2 ,j− 1 2 ) 2 ,j− 1 2 2 where the explicit form of a± and c± is a+(l, i, j) = q2l+i+j+1 (1 − q2l−2j+2) (1 − q4l+2) 1 2 1 2 (1 − q2l−2i+2) 1 2 (1 − q4l+4) 1 2 1 2 a−(l, i, j) = and (1 − q2l+2j ) 1 2 (1 − q4l+2) 1 2 (1 − q2l+2i) 1 2 (1 − q4l) c+(l, i, j) = −ql+j (1 − q2l−2j+2) (1 − q4l+2) 1 2 (1 − q2l+2i+2) 1 2 (1 − q4l+4) 1 2 1 2 c−(l, i, j) = ql+i (1 − q2l+2j) (1 − q4l) 1 2 (1 − q2l−2i) 1 2 1 2 (1 − q4l+2) 1 2 . Note here that a−(l, i, j) vanishes if i = −l or j = −l, and similarly, c−(l, i, j) vanishes for i = l or j = −l. We obtain α∗ e(l) i,j = a+(l − 1 2 ,i + 1 2 , j + 1 2 ) e (l− 1 i+ 1 2 ) 2 ,j+ 1 2 + a−(l + 1 2 , i + 1 2 , j + 1 2 ) e (l+ 1 i+ 1 2 ) 2 ,j+ 1 2 and γ ∗ e(l) i,j = c+(l − 1 (l+ 1 i− 1 for the action of α∗ and γ ∗, respectively. Let us also record the formulas + c−(l + 1 2 , j + 1 2 , j + 1 2 ) 2 ,j+ 1 2 , i − 1 2 ,i − 1 (l− 1 i− 1 2 ) e 2 ) e 2 2 ) 2 ,j+ 1 2 and i,j = (−1)2l+i+j qj−iu(l) u(l)∗ −i,−j i,j = (−1)2l+i+jqi+j e(l) e(l)∗ −i,−j for the adjoint. The classical torus T = S1 is a closed quantum subgroup of SUq(2). Explicitly, the inclusion T ⊂ SUq(2) is determined by the ∗-homomorphism π : C[SUq(2)] → C[T ] = C[z, z−1] given by π(α) = z, π(γ) = 0. By definition, the standard Podle´s sphere SUq(2)/T is the corresponding homoge- neous space [26]. In the algebraic setting, it is described by the dense ∗-subalgebra C[SUq(2)/T ] ⊂ C(SUq(2)/T ) of coinvariants in C[SUq(2)] with respect to the right coaction (id ⊗π)∆ of C[T ]. Equivalently, it is the unital ∗-subalgebra of C[SUq(2)] 10 CHRISTIAN VOIGT generated by the elements A = γ ∗γ and B = α∗γ. These elements satisfy the relations A = A∗, AB = q2BA, BB∗ = q−2A(1 − A), B∗B = A(1 − q2A), and we record the following explicit formulas for the action of A and B on L2(SUq(2)) for q ∈ (−1, 1) \ {0}. Firstly, γ ∗γ e(l) i,j = γ ∗(cid:18)c+(l, i, j) e 2 ) 2 ,j− 1 = c+(l − 1, i, j)c−(l, i, j) e(l−1) (l+ 1 i+ 1 2 ij + c−(l, i, j) e (l− 1 i+ 1 2 ) 2 ,j− 1 2(cid:19) + (c+(l, i, j)2 + c−(l, i, j)2) e(l) ij + c−(l + 1, i, j)c+(l, i, j) e(l+1) ij 1 2 (1 − q2l+2i) (1 − q4l−2) = −q2l+i+j−1 (1 − q2l−2j ) +(cid:18)q2l+2j (1 − q2l−2j+2)(1 − q2l+2i+2) − q2l+i+j+1 (1 − q2l+2j+2) (1 − q4l+2)(1 − q4l+4) 1 2 (1 − q2l−2i+2) determines the action of A. Similarly, we get (1 − q4l+4)(1 − q4l+6) 1 2 (1 − q2l+2j ) 1 2 (1 − q2l−2i) 1 2 1 2 (1 − q4l)(1 − q4l+2) 1 2 e(l−1) ij + q2l+2i (1 − q2l+2j )(1 − q2l−2i) (1 − q4l)(1 − q4l+2) (cid:19)e(l) ij 1 2 (1 − q2l−2j+2) 1 2 (1 − q2l+2i+2) 1 2 e(l+1) ij 1 2 (1 − q4l+2) 1 2 α∗γ e(l) i,j = α∗(cid:18)c+(l, i, j)e (l+ 1 i+ 1 2 ) 2 ,j− 1 2 + c−(l, i, j)e (l− 1 i+ 1 2 ) 2 ,j− 1 2(cid:19) = a+(l − 1, i + 1, j)c−(l, i, j)e(l−1) i+1,j +(cid:18)a+(l, i + 1, j)c+(l, i, j) + a−(l, i + 1, j)c−(l, i, j)(cid:19)e(l) i+1,j + a−(l + 1, i + 1, j)c+(l, i, j)e(l+1) i+1,j = q3l+2i+j (1 − q2l−2j) 1 2 (1 − q2l−2i−2) 1 2 (1 − q2l+2j) 1 2 (1 − q2l−2i) 1 2 (1 − q4l−2) 1 2 (1 − q4l)(1 − q4l+2) 1 2 e(l−1) i+1,j +(cid:18)ql+i (1 − q2l+2j )(1 − q2l+2i+2) (1 − q4l)(1 − q4l+2) 1 2 (1 − q2l−2i) 1 2 − q3l+i+2j+2 (1 − q2l−2j+2)(1 − q2l−2i) (1 − q4l+2)(1 − q4l+4) 1 2 (1 − q2l+2i+2) 1 2 − ql+j (1 − q2l+2j+2) 1 2 (1 − q2l+2i+4) 1 2 (1 − q2l−2j+2) 1 2 (1 − q2l+2i+2) (1 − q4l+4)(1 − q4l+6) 1 2 (1 − q4l+2) 1 2 (cid:19)e(l) i+1,j 1 2 e(l+1) i+1,j for the action of B. In the sequel we abbreviate SUq(2) = Gq. If Ck denotes the irreducible represen- tation of T of weight k ∈ Z, then the cotensor product Γ(Ek) = Γ(Gq ×T Ck) = C[Gq](cid:3)C[T ]Ck ⊂ C[Gq] is a noncommutative analogue of the space of sections of the homogeneous vector bundle G ×T Ck over G/T . The space Γ(Ek) is a C[Gq/T ]-bimodule in a natural way which is finitely generated and projective both as a left and right C[Gq/T ]- module. The latter follows from the fact that C[Gq/T ] ⊂ C[Gq] is a faithfully flat Hopf-Galois extension, compare [28]. We denote by C(Ek) the closure of Γ(Ek) inside C(Gq). The space C(Ek) is a Gq- equivariant Hilbert C(Gq/T )-module with coaction induced by comultiplication. BAUM-CONNES CONJECTURE 11 We write L2(Ek) for the Gq-Hilbert space obtained by taking the closure of Γ(Ek) inside L2(Gq). Let us recall the definition of the Drinfeld double D(Gq) of Gq. It is the locally compact quantum group determined by C0(D(Gq)) = C(Gq) ⊗ C ∗(Gq) with comul- tiplication ∆D(Gq) = (id ⊗σ ⊗ id)(id ⊗ad(W ) ⊗ id)(∆ ⊗ ∆). Here ad(W ) denotes the adjoint action of the left regular multiplicative unitary W ∈ M (C(Gq) ⊗ C ∗(Gq)) and σ is the flip map. Observe that both C(Gq) and C ∗(Gq) are quotient Hopf-C ∗-algebras of C0(D(Gq)). It is shown in [25] that a D(Gq)-C ∗-algebra A is uniquely determined by coactions α : A → M (C(Gq) ⊗ A) and λ : A → M (C ∗(Gq) ⊗ A) satisfying the Yetter-Drinfeld compatibility condition (σ ⊗ id)(id ⊗α)λ = (ad(W ) ⊗ id)(id ⊗λ)α. In a similar fashion on can study D(Gq)-equivariant Hilbert modules. For instance, the space C(Ek) defined above carries a coaction λ : C(Ek) → M (C ∗(Gq) ⊗ C(Ek)) given by λ(f ) = W ∗(1 ⊗ f ) W where W = ΣW ∗Σ. Together with the canonical coaction of C(Gq) this turns C(Ek) into a D(Gq)-equivariant Hilbert C(Gq/T )- module. 4. Equivariant KK-theory for the Podle´s sphere In this section we study the equivariant KK-theory of the standard Podle´s sphere SUq(2)/T . Background information on equivariant KK-theory for quantum group actions can be found in [1], [25]. Most of the ingredients in this section have already been introduced in [25] in the case q ∈ (0, 1). However, for the purposes of this paper we have to allow for negative values of q. In the sequel we consider q ∈ (−1, 1)\ {0}, and as in the previous section we abbreviate Gq = SUq(2). Let us first recall the definition of the Fredholm module corresponding to the Dirac operator on the standard Podle´s sphere Gq/T , compare [11], [25]. The underlying graded Gq-Hilbert space is H = H1 ⊕ H−1 = L2(E1) ⊕ L2(E−1) with its natural coaction of C(Gq). The covariant representation φ of C(Gq/T ) is given by left multiplication. We note that H1 and H−1 are isomorphic representations of Gq due to Frobenius reciprocity. Hence we obtain a self-adjoint unitary operator F on H by F =(cid:18)0 i,1/2 and e(l) 1 1 0(cid:19) if we identify the basis vectors e(l) i,−1/2 in H1 and H−1, respectively. Us- ing the explicit formulas for the generators of the Podle´s sphere from section 3 one checks that D = (H, φ, F ) is a Gq-equivariant Fredholm module defining an element in KK Gq(C(Gq/T ), C). Our first aim is to lift this construction to D(Gq)-equivariant KK-theory where D(Gq) denotes the Drinfeld double of Gq as in section 3. We recall that the C ∗- algebra C(Gq/T ) = indGq T (C) is naturally a D(Gq)-C ∗-algebra [25]. The corre- sponding coaction C(Gq/T ) → M (C ∗(Gq)⊗C(Gq/T )) is determined by the adjoint action h · g = h(1)gS(h(2)) of C[Gq] on C[Gq/T ]. The following lemma shows that the Hilbert spaces L2(Ek) become D(Gq)-Hilbert spaces in a similar way. 12 CHRISTIAN VOIGT Lemma 4.1. For every k ∈ Z the formula ω(x)(ξ) = x(1)ξ f1 ⇀ S(x(2)) determines a ∗-homomorphism ω : C(Gq) → L(L2(Ek)) which turns L2(Ek) into a D(Gq)-Hilbert space such that the representation of C(Gq/T ) by left multiplication is covariant. Proof. Let us write H = L2(Ek) and define a map ω : C[Gq] → L(H) by the above formula, where we recall that f1 : C[Gq] → C denotes the modular character. For ξ ∈ Γ(Ek) one obtains (id ⊗π)∆(ω(x)(ξ)) = (id ⊗π)∆(x(1)ξ f1 ⇀ S(x(2))) = x(1)ξ(1)S(x(4)) ⊗ π(x(2)ξ(2)f1 ⇀ S(x(3))) = x(1)ξ(1)S(x(4)) ⊗ π(x(2))π(ξ(2))π(f1 ⇀ S(x(3))) = x(1)ξS(x(4)) ⊗ π(x(2)f1 ⇀ S(x(3)))zk = x(1)ξS(x(4)) ⊗ π(S−1(x(3)) ↼ f1)π(x(2))zk = x(1)ξS(x(3)) ⊗ π(f−1(x(2))1)zk = x(1)ξf1 ⇀ S(x(2)) ⊗ zk = ω(x)(ξ) ⊗ zk using that C[T ] is commutative and S−1(x) = f1 ⇀ S(x) ↼ f−1 for all x ∈ C[Gq]. It follows that the map ω is well-defined. Using f1 ⇀ x∗ = (f−1 ⇀ x)∗ and the modular properties of the Haar state φ we obtain hω(x∗)(ξ), ηi = φ((ω(x∗)(ξ))∗η) (2)))∗η) (1)ξf1 ⇀ S(x∗ = φ((x∗ (1)ξf1 ⇀ S−1(x(2))∗)∗η) = φ((x∗ = φ(f−1 ⇀ S−1(x(2))ξ∗x(1)η) = φ(S(x(2)) ↼ f−1ξ∗x(1)η) = φ(ξ∗x(1)η f1 ⇀ S(x(2))) = φ(ξ∗ω(x)(η)) = hξ, ω(x)(η)i, and this shows that ω defines a ∗-homomorphism from C(Gq) to L(H). We deduce that ω corresponds to a coaction λ : H → M (C ∗(Gq) ⊗ H), and λ combines with the standard coaction of C(Gq) such that H becomes a D(Gq)-Hilbert space. Recall that the action of C(Gq/T ) on H by left multiplication yields a Gq-equivariant ∗-homomorphism φ : C(Gq/T ) → L(H). We have ω(x)(φ(g)(ξ)) = x(1)gξf1 ⇀ S(x(2)) = x(1)gS(x(2))x(3)ξf1 ⇀ S(x(4)) = φ(x(1) · g)(ω(x(2))(ξ)) for all x ∈ C[Gq], g ∈ C[Gq/T ] and ξ ∈ H, and this implies that φ is covariant with respect to λ. (cid:3) We shall now show that the Fredholm module constructed in the beginning of this section determines a D(Gq)-equivariant KK-element. Proposition 4.2. Let q ∈ (−1, 1) \ {0}. The Fredholm module D defined above induces an element [D] in KK D(Gq)(C(Gq/T ), C) in a natural way. BAUM-CONNES CONJECTURE 13 Proof. We have to verify that F commutes with the action of D(Gq) up to compact operators. Since F is Gq-equivariant this amounts to showing (C ∗(Gq) ⊗ 1)(1 ⊗ F − adλ(F )) ⊂ C ∗(Gq) ⊗ K(H) where λ : H → M (C ∗(Gq) ⊗ H) is the coaction on H = H1 ⊕ H−1 defined in lemma 4.1. It suffices to check that F commutes with the corresponding action ω : C(Gq) → L(H) up to compact operators. This is an explicit calculation using the formulas for the regular representation from section 3. In fact, we will obtain the assertion as a special case of our computations below. It turns out that F is actually D(Gq)-equivariant. (cid:3) In the sequel we need variants of the representations defined in lemma 4.1. Let t ∈ [0, 1] and consider the representation πt of C[Gq] on Γ(Ek) given by πt(x)(ξ) = x(1)ξ ft ⇀ S(x(2)) where ft : C[Gq] → C is the modular character given by ft(α) = qtα and ft(γ) = 0. This action has the correct algebraic properties to turn L2(Ek) into a D(Gq)-Hilbert space except that it is not compatible with the ∗-structures for t < 1. In order to proceed we need explicit formulas for the action of the generators. More precisely, a straightforward computation based on the formulas for the regular representation in section 3 yields πt(α)(e(l) i,j ) = qtαe(l) i,jα∗ + q−tq2γ ∗e(l) i,jγ = qtq−1a+(l, −i, −j)(cid:18)a+(l + 1 + qtq−1a−(l, −i, −j)(cid:18)a+(l − 1 − q−tq2c+(l − 1 2 , −i − 1 2 , −j + 1 2 , i + 1 2 , j + 1 2 )e(l+1) i,j + a−(l + 1 2 , i + 1 2 , j + 1 2 , i + 1 2 , j + 1 2 )e(l) i,j + a−(l − 1 2 , i + 1 2 , j + 1 i,j(cid:19) 2 )e(l) i,j (cid:19) 2 )e(l−1) − q−tq2c−(l + 1 2 , −i − 1 2 , −j + 1 i,j + c−(l + 1, i, j) e(l+1) i,j + c−(l, i, j)e(l) 2 )(cid:18)c+(l − 1, i, j)e(l−1) 2 )(cid:18)c+(l, i, j)e(l) i,j(cid:19) i,j (cid:19) and πt(α∗)(e(l) i,j) = q−tα∗e(l) ij α + qtγe(l) ij γ ∗ = q−tqa+(l − 1 2 , −i + 1 2 , −j + 1 + q−tqa−(l + 1 2 , −i + 1 2 , −j + 1 − qtc+(l, −i, −j)(cid:18)c+(l + 1 − qtc−(l, −i, −j)(cid:18)c+(l − 1 2 )(cid:18)a+(l − 1, i, j) e(l−1) 2 )(cid:18)a+(l, i, j) e(l) 2 ) e(l+1) 2 , j + 1 2 , i − 1 i,j + a−(l, i, j) e(l) i,j(cid:19) i,j (cid:19) i,j + a−(l + 1, i, j) e(l+1) i,j(cid:19) 2 ) e(l) i,j (cid:19). 2 ) e(l−1) 2 , j + 1 2 , j + 1 2 , i − 1 2 , i − 1 i,j + c−(l + 1 2 , i − 1 2 , j + 1 2 ) e(l) i,j + c−(l − 1 Similarly we obtain πt(γ)(e(l) ij ) = qtγe(l) ij α∗ − q−tqα∗e(l) ij γ = qtq−1a+(l, −i, −j)(cid:18)c+(l + 1 + qtq−1a−(l, −i, −j)(cid:18)c+(l − 1 2 , i + 1 2 , j + 1 2 ) e(l+1) i+1,j + c−(l + 1 2 , i + 1 2 , j + 1 2 , i + 1 2 , j + 1 2 ) e(l) i+1,j + c−(l − 1 2 , i + 1 2 , j + 1 i+1,j(cid:19) 2 ) e(l) i+1,j(cid:19) 2 ) e(l−1) 14 CHRISTIAN VOIGT + q−tqc+(l − 1 2 , −i − 1 2 , −j + 1 + q−tqc−(l + 1 2 , −i − 1 2 , −j + 1 i+1,j + a−(l, i + 1, j) e(l) 2 )(cid:18)a+(l − 1, i + 1, j) e(l−1) 2 )(cid:18)a+(l, i + 1, j) e(l) i+1,j(cid:19) i+1,j(cid:19) i+1,j + a−(l + 1, i + 1, j) e(l+1) and πt(γ ∗)(e(l) ij ) = q−tγ ∗e(l) ij α − qtq−1αe(l) ij γ ∗ = q−tqa+(l − 1 2 , −i + 1 2 , −j + 1 i−1,j + c−(l, i − 1, j) e(l) + q−tqa−(l + 1 2 , −i + 1 2 , −j + 1 + qtq−1c+(l, −i, −j)(cid:18)a+(l + 1 + qtq−1c−(l, −i, −j)(cid:18)a+(l − 1 2 )(cid:18)c+(l − 1, i − 1, j) e(l−1) 2 )(cid:18)c+(l, i − 1, j) e(l) 2 ) e(l+1) 2 , j + 1 2 , i − 1 i−1,j + c−(l + 1, i − 1, j) e(l+1) i−1,j + a−(l + 1 2 , i − 1 2 , j + 1 2 , i − 1 2 , j + 1 2 ) e(l) i−1,j + a−(l − 1 2 , i − 1 2 , j + 1 i−1,j(cid:19) i−1,j(cid:19) i−1,j(cid:19) 2 ) e(l) i−1,j(cid:19). 2 ) e(l−1) This may be written in the form πt(α)(e(l) πt(α∗)(e(l) πt(γ)(e(l) πt(γ ∗)(e(l) i,j ) = a1(t, l, i, j)e(l+1) i,j) = b1(t, l, i, j)e(l+1) i,j) = c1(t, l, i, j)e(l+1) i,j) = d1(t, l, i, j)e(l+1) i,j + a0(t, l, i, j)e(l) i,j + b0(t, l, i, j)e(l) i+1,j + c0(t, l, i, j)e(l) i−1,j + d0(t, l, i, j)e(l) i,j + a−1(t, l, i, j)e(l−1) i,j + b−1(t, l, i, j)e(l−1) i+1,j + c−1(t, l, i, j)e(l−1) i−1,j + d−1(t, l, i, j)e(l−1) i−1,j i+1,j i,j i,j where a1(t, l, i, j) = qtq4l+3 (1 − q2l+2j+2) − q−tq2l+3 (1 − q2l−2j+2) 1 2 (1 − q2l+2i+2) 1 2 (1 − q2l−2j+2) 1 2 (1 − q2l−2i+2) 1 2 (1 − q4l+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 (1 − q2l−2i+2) 1 2 (1 − q2l+2j+2) 1 2 1 2 1 2 (1 − q2l+2i+2) (1 − q4l+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 a0(t, l, i, j) = qtq2l−i−j (1 − q2l+2j+2)(1 − q2l+2i+2) (1 − q4l+2)(1 − q4l+4) + qtq2l+i+j (1 − q2l−2j )(1 − q2l−2i) (1 − q4l)(1 − q4l+2) + q−tq2l+i−j+2 (1 − q2l+2j )(1 − q2l−2i) (1 − q4l)(1 − q4l+2) + q−tq2l−i+j+2 (1 − q2l−2j+2)(1 − q2l+2i+2) (1 − q4l+2)(1 − q4l+4) a−1(t, l, i, j) = qtq−1 (1 − q2l−2j) 1 2 (1 − q2l−2i) 1 2 (1 − q2l+2j) 1 2 (1 − q2l+2i) 1 2 (1 − q4l)(1 − q4l+2) 1 2 (1 − q2l−2j) 1 2 (1 − q4l−2) 1 2 (1 − q2l+2i) 1 2 (1 − q2l−2i) 1 2 1 2 (1 − q4l)(1 − q4l+2) 1 2 (1 − q4l−2) 1 2 − q−tq2l+1 (1 − q2l+2j) BAUM-CONNES CONJECTURE 15 and b1(t, l, i, j) = q−tq (1 − q2l−2j+2) 1 2 (1 − q2l−2i+2) 1 2 (1 − q2l+2j+2) 1 2 (1 − q2l+2i+2) 1 2 (1 − q4l+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 − qtq2l+1 (1 − q2l+2j+2) 1 2 (1 − q2l−2i+2) 1 2 (1 − q2l−2j+2) 1 2 (1 − q2l+2i+2) 1 2 (1 − q4l+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 b0(t, l, i, j) = q−tq2l−i−j+2 (1 − q2l+2j)(1 − q2l+2i) (1 − q4l)(1 − q4l+2) + q−tq2l+i+j+2 (1 − q2l−2j+2)(1 − q2l−2i+2) (1 − q4l+2)(1 − q4l+4) + qtq2l+i−j (1 − q2l+2j+2)(1 − q2l−2i+2) (1 − q4l+2)(1 − q4l+4) + qtq2l−i+j (1 − q2l−2j)(1 − q2l+2i) (1 − q4l)(1 − q4l+2) 1 2 (1 − q2l−2j) 1 2 (1 − q2l+2i) b−1(t, l, i, j) = q−tq4l+1 (1 − q2l+2j) − qtq2l−1 (1 − q2l−2j ) 1 2 (1 − q2l−2i) 1 2 1 2 (1 − q4l−2) 1 2 (1 − q4l)(1 − q4l+2) 1 2 (1 − q2l+2i) 1 2 (1 − q2l+2j ) 1 2 (1 − q2l−2i) 1 2 , (1 − q4l)(1 − q4l+2) 1 2 (1 − q4l−2) 1 2 similarly c1(t, l, i, j) = −qtq3l−i+1 (1 − q2l+2j+2) + q−tql−i+1 (1 − q2l−2j+2) 1 2 (1 − q2l+2i+2) 1 2 (1 − q2l−2j+2) 1 2 (1 − q2l+2i+4) 1 2 (1 − q4l+2) 1 2 (1 − q2l+2i+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 1 2 (1 − q2l+2j+2) 1 2 (1 − q2l+2i+4) 1 2 (1 − q4l+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 c0(t, l, i, j) = qtq3l−j+1 (1 − q2l+2j+2)(1 − q2l+2i+2) (1 − q4l+2)(1 − q4l+4) 1 2 (1 − q2l−2i) 1 2 − qtql+j−1 (1 − q2l−2j)(1 − q2l−2i) (1 − q4l)(1 − q4l+2) 1 2 (1 − q2l+2i+2) 1 2 − q−tql−j+1 (1 − q2l+2j )(1 − q2l−2i) (1 − q4l)(1 − q4l+2) 1 2 (1 − q2l−2i) + q−tq3l+j+3 (1 − q2l−2j+2)(1 − q2l+2i+2) 1 2 1 2 (1 − q2l+2i+2) 1 2 (1 − q4l+2)(1 − q4l+4) 1 2 (1 − q2l−2i−2) 1 2 (1 − q2l+2j ) 1 2 (1 − q2l−2i) 1 2 (1 − q4l)(1 − q4l+2) − q−tq3l+i+1 (1 − q2l+2j ) 1 2 (1 − q2l−2i) 1 2 (1 − q4l−2) 1 2 (1 − q2l−2i−2) 1 2 (1 − q2l−2j ) 1 2 1 2 , (1 − q4l)(1 − q4l+2) 1 2 (1 − q4l−2) 1 2 c−1(t, l, i, j) = qtql+i−1 (1 − q2l−2j ) 16 and CHRISTIAN VOIGT d1(t, l, i, j) = q−tql+i+1 (1 − q2l−2j+2) 1 2 (1 − q2l−2i+2) 1 2 (1 − q2l+2j+2) 1 2 (1 − q2l−2i+4) 1 2 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 (1 − q4l+2) 1 2 (1 − q2l−2i+2) 1 2 (1 − q2l−2j+2) 1 2 (1 − q2l−2i+4) 1 2 − qtq3l+i+1 (1 − q2l+2j+2) (1 − q4l+2) d0(t, l, i, j) = q−tq3l−j+1 (1 − q2l+2j)(1 − q2l+2i) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 (1 − q2l−2i+2) 1 2 1 2 (1 − q4l)(1 − q4l+2) − q−tql+j+1 (1 − q2l−2j+2)(1 − q2l−2i+2) (1 − q4l+2)(1 − q4l+4) 1 2 (1 − q2l+2i) 1 2 − qtql−j−1 (1 − q2l+2j+2)(1 − q2l−2i+2) (1 − q4l+2)(1 − q4l+4) 1 2 (1 − q2l+2i) 1 2 + qtq3l+j−1 (1 − q2l−2j)(1 − q2l+2i) 1 2 (1 − q2l−2i+2) 1 2 d−1(t, l, i, j) = −q−tq3l−i+1 (1 − q2l+2j ) (1 − q4l)(1 − q4l+2) 1 2 (1 − q2l+2i) 1 2 (1 − q2l−2j ) 1 2 (1 − q2l+2i−2) 1 2 (1 − q4l)(1 − q4l+2) 1 2 (1 − q4l−2) 1 2 + qtql−i−1 (1 − q2l−2j ) 1 2 (1 − q2l+2i) 1 2 (1 − q2l+2j ) 1 2 (1 − q2l+2i−2) 1 2 (1 − q4l)(1 − q4l+2) 1 2 (1 − q4l−2) 1 2 . Let us now set m(t, l) = q−tq2 − qtq2l qt − q−tq2l+2 = q2 − q2tq2l q2t − q2l+2 for t ∈ [0, 1] and l ∈ N. Note that m(1, l) = 1 for all l if we interpret m(1, 0) = 1. We define A1(t, l, i) = m(t, l + 1)− 1 A0(t, l, i) = a0(t, l, i, 0) 2 a1(t, l, i, 0) A−1(t, l, i) = m(t, l) 1 2 a−1(t, l, i, 0), and by rescaling bk(t, l, i, 0), ck(t, l, i, 0) and dk(t, l, i, 0) for k = −1, 0, 1 in the same way we obtain constants Bk(t, l, i), Ck(t, l, i) and Dk(t, l, i). Inspection of the formu- las above shows that the expressions X1(t, 0, 0) for X = A, B, C, D are well-defined and depend continuously on t ∈ [0, 1] although m(0, 1) = 0. Lemma 4.3. Let q ∈ (−1, 1) \ {0}. For t ∈ [0, 1] the formulas ωt(α)(e(l) ωt(α∗)(e(l) ωt(γ)(e(l) ωt(γ ∗)(e(l) i,0) = A1(t, l, i)e(l+1) i,0) = B1(t, l, i)e(l+1) i,0) = C1(t, l, i)e(l+1) i,0) = D1(t, l, i)e(l+1) i,0 + A0(t, l, i)e(l) i,0 + B0(t, l, i)e(l) i+1,0 + C0(t, l, i)e(l) i−1,0 + D0(t, l, i)e(l) i,0 + A−1(t, l, i)e(l−1) i,0 + B−1(t, l, i)e(l−1) i+1,0 + C−1(t, l, i)e(l−1) i−1,0 + D−1(t, l, i)e(l−1) i+1,0 i−1,0 i,0 i,0 define a ∗-homomorphism ωt : C(Gq) → L(L2(E0)). BAUM-CONNES CONJECTURE 17 Proof. The main point is to show that ωt is compatible with the ∗-structures. In order to prove ωt(α)∗ = ωt(α∗) we have to verify A1(t, l, i) = B−1(t, l + 1, i) A0(t, l, i) = B0(t, l, i) A−1(t, l, i) = B1(t, l − 1, i). We obtain m(t, l + 1) 1 2 A1(t, l, i) = (qtq4l+3 − q−tq2l+3)× (1 − q2l+2)(1 − q2l+2i+2) 1 2 (1 − q2l−2i+2) 1 2 (1 − q4l+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 and m(t, l + 1) 1 2 B−1(t, l + 1, i, 0) = qtq4l+3 − q−tq2l+3 q−tq4l+5 − qtq2l+1 × (q−tq4l+5 − qtq2l+1) (1 − q2l+2)(1 − q2l+2i+2) (1 − q4l+4)(1 − q4l+6) 1 2 (1 − q2l−2i+2) 1 2 (1 − q4l+2) 1 2 1 2 . Similarly we find A0(t, l, i) = (qtq2l−i + q−tq2l−i+2) + (qtq2l+i + q−tq2l+i+2) = (qtq2l−i + q−tq2l−i+2) (1 − q2l+2)(1 − q2l+2i+2) (1 − q4l+2)(1 − q4l+4) (1 − q2l)(1 − q2l−2i) (1 − q4l)(1 − q4l+2) (1 − q2l+2i+2) (1 + q2l+2)(1 − q4l+2) + (qtq2l+i + q−tq2l+i+2) (1 − q2l−2i) (1 + q2l)(1 − q4l+2) and hence (1 + q2l)(1 + q2l+2)(1 − q4l+2)A0(t, l, i) = (qtq2l−i + q−tq2l−i+2)(1 − q2l+2i+2)(1 + q2l) + (qtq2l+i + q−tq2l+i+2)(1 − q2l−2i)(1 + q2l+2) = (qtq2l−i + q−tq2l−i+2)(1 − q2l+2i+2 + q2l − q4l+2i+2) + (qtq2l+i + q−tq2l+i+2)(1 − q2l−2i + q2l+2 − q4l−2i+2) = qt(q2l−i − q4l+i+2 + q4l−i − q6l+i+2 + q2l+i − q4l−i + q4l+i+2 − q6l−i+2) + q−t(q2l−i+2 − q4l+i+4 + q4l−i+2 − q6l+i+4 + q2l+i+2 − q4l−i+2 + q4l+i+4 − q6l−i+4) = qt(q2l−i − q6l+i+2 + q2l+i − q6l−i+2) + q−t(q2l−i+2 − q6l+i+4 + q2l+i+2 − q6l−i+4). 18 CHRISTIAN VOIGT Conversely, B0(t, l, i) = (q−tq2l+i+2 + qtq2l+i) (1 − q2l+2)(1 − q2l−2i+2) (1 − q4l+2)(1 − q4l+4) + (q−tq2l−i+2 + qtq2l−i) = (q−tq2l+i+2 + qtq2l+i) (1 − q2l)(1 − q2l+2i) (1 − q4l)(1 − q4l+2) (1 − q2l−2i+2) (1 − q4l+2)(1 + q2l+2) + (q−tq2l−i+2 + qtq2l−i) (1 − q2l+2i) (1 + q2l)(1 − q4l+2) and hence (1 + q2l)(1 + q2l+2)(1 − q4l+2)B0(t, l, i) = (q−tq2l+i+2 + qtq2l+i)(1 − q2l−2i+2)(1 + q2l) + (q−tq2l−i+2 + qtq2l−i)(1 − q2l+2i)(1 + q2l+2) = (q−tq2l+i+2 + qtq2l+i)(1 − q2l−2i+2 + q2l − q4l−2i+2) + (q−tq2l−i+2 + qtq2l−i)(1 − q2l+2i + q2l+2 − q4l+2i+2) = qt(q2l+i − q4l−i+2 + q4l+i − q6l−i+2 + q2l−i − q4l+i + q4l−i+2 − q6l+i+2) + q−t(q2l+i+2 − q4l−i+4 + q4l+i+2 − q6l−i+4 + q2l−i+2 − q4l+i+2 + q4l−i+4 − q6l+i+4) = qt(q2l+i − q6l−i+2 + q2l−i − q6l+i+2) + q−t(q2l+i+2 − q6l−i+4 + q2l−i+2 − q6l+i+4). Finally, and m(t, l) 1 2 A−1(t, l, i, 0) = qtq2l−1 − q−tq q−tq2l+1 − qtq−1 × (qtq−1 − q−tq2l+1) (1 − q2l)(1 − q2l−2i) (1 − q4l)(1 − q4l+2) 1 2 1 2 (1 − q2l+2i) 1 1 2 (1 − q4l−2) 2 m(t, l) 1 2 B1(t, l − 1, i, 0) = (q−tq − qtq2l−1) (1 − q2l)(1 − q2l−2i) 1 2 (1 − q2l+2i) 1 2 1 2 (1 − q4l)(1 − q4l+2) (1 − q4l−2) 1 2 . One verifies ωt(γ)∗ = ωt(γ ∗) in a similar fashion. It is then easy to check that the operators ωt(α) and ωt(γ) satisfy the defining relations of C(Gq). (cid:3) Lemma 4.4. Let q ∈ (−1, 1) \ {0}. For l → ∞ the expressions a1(1, l, i, ±1) − A1(t, l, i), a0(1, l, i, ±1), A0(t, l, i), a−1(1, l, i, ±1) − A−1(t, l, i) and c1(1, l, i, ±1) − C1(t, l, i), c0(1, l, i, ±1), C0(t, l, i), c−1(1, l, i, ±1) − C−1(t, l, i) tend to zero uniformly for t ∈ [0, 1] and independently of i. Proof. Since the constants ak(1, l, i, j) and ck(1, l, i, j) for k = −1, 0, 1 are symmet- ric in the variable j we may restrict attention to the case j = 1. BAUM-CONNES CONJECTURE 19 The estimates involving A1, A0 and a0 are easy. For A−1 it suffices to consider m(t, l) (cid:12)(cid:12)(cid:12)(cid:12) 1 1 2 (1 − q2l+2i) 2 qtq−1 (1 − q2l)(1 − q2l−2i) 1 1 (1 − q4l)(1 − q4l+2) 2 (1 − q4l−2) 2 1 2 (1 − q2l+2i) 1 2 (1 − q2l−2i) 1 2 (1 − q2l+2) 1 2 1 2 1 2 2 1 2 (1 − q2l+2) 1 qtq2l − q−tq2 1 2 (1 − q4l−2) 1 2 (1 − q4l)(1 − q4l+2) −qq−1 (1 − q2l−2) (qtq2l − q−tq2) (q−tq2l+2 − qt) 1 qt(1 − q2l) − q(1 − q2l−2) (cid:12)(cid:12)(cid:12)(cid:12) 2(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) q−tq2l+2 − qt q2t(1 − q2l)2 − q2(1 − q2l−2)(1 − q2l+2)(cid:12)(cid:12)(cid:12)(cid:12) (1 − q2l)2 − q2(1 − q2l−2)(1 − q2l+2)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12) (1 − q2l)2(cid:12)(cid:12)(cid:12)(cid:12) 1 − q−2tq2l+2 (1 − q2l)2 − q2(1 − q2l−2)(1 − q2l+2)(cid:12)(cid:12)(cid:12)(cid:12) q−2tq2l+2 − 1 q2 +(cid:12)(cid:12)(cid:12)(cid:12) q2tq2l − q2 q−2tq2l+2 − 1 q2tq2l . , . We may estimate this expression by which reduces to It is enough to estimate (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) and both terms converge to zero for l → ∞. The remaining assertions are verified in a similar fashion. (cid:3) In the sequel we write sgn(q) for the sign of q, that is, sgn(q) = 1 if q > 0 and sgn(q) = −1 if q < 0. Lemma 4.5. Let q ∈ (−1, 1) \ {0}. We have A1(0, l, i) = a1(1, l, i, ±1) A0(0, l, i) = sgn(q) a0(1, l, i, ±1) A−1(0, l, i) = a−1(1, l, i, ±1) C1(0, l, i) = c1(1, l, i, ±1) C0(0, l, i) = sgn(q) c0(1, l, i, ±1) C−1(0, l, i) = c−1(1, l, i, ±1) and similarly for l > 0 and all i. Proof. Since the coefficients ak(1, l, i, j) and ck(1, l, i, j) for k = −1, 0, 1 are sym- metric in the variable j it suffices again to consider the case j = 1. We have A1(0, l, i) = m(0, l + 1)− 1 2 (q4l+3 − q2l+3)× (1 − q2l+2)(1 − q2l+2i+2) 1 2 (1 − q2l−2i+2) 1 2 (1 − q4l+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 and a1(1, l, i, 1) = (qq4l+3 − q−1q2l+3)× (1 − q2l) 1 2 (1 − q2l+2i+2) 1 2 (1 − q2l+4) 1 2 (1 − q2l−2i+2) 1 2 (1 − q4l+2) 1 2 (1 − q4l+4)(1 − q4l+6) 1 2 . 20 Then CHRISTIAN VOIGT m(0, l + 1)− 1 2 (q2l+1 − q)(1 − q2l+2) = 1 2 (1 − q2l+4) (q2 − q2l+2) 1 2 (q2l+1 − q)(1 − q2l+2) 1 2 q−1(−q)(1 − q2l)(1 − q2l+2) = (1 − q2l+4) (1 − q2l) = (1 − q2l+4) 1 2 1 2 q−1(1 − q2l) 1 2 (q2l+3 − q) 1 2 (1 − q2l+4) 1 2 = (qq2l+1 − q−1q)(1 − q2l) yields the first claim. Next we note A0(0, l, i) = (q2l−i + q2l−i+2) (1 − q2l+2)(1 − q2l+2i+2) (1 − q4l+2)(1 − q4l+4) (1 − q2l)(1 − q2l−2i) (1 − q4l)(1 − q4l+2) + (q2l+i + q2l+i+2) and a0(1, l, i, 1) = (qq2l−i−1(1 − q2l+4) + q−1q2l−i+3(1 − q2l)) 1 − q2l+2i+2 (1 − q4l+2)(1 − q4l+4) + (qq2l+i+1(1 − q2l−2) + q−1q2l+i+1(1 − q2l+2)) 1 − q2l−2i (1 − q4l)(1 − q4l+2) . Since sgn(q)(q2l−i + q2l−i+2)(1 − q2l+2) = sgn(q)(q2l−i + q2l−i+2 − q4l−i+2 − q4l−i+4) = qq−1(q2l−i − q4l−i+4) + q−1q(q2l−i+2 − q4l−i+2) = qq2l−i−1(1 − q2l+4) + q−1q2l−i+3(1 − q2l) and sgn(q)(q2l+i + q2l+i+2)(1 − q2l) = sgn(q)(q2l+i + q2l+i+2 − q4l+i − q4l+i+2) = qq−1(q2l+i+2 − q4l+i) + q−1q(q2l+i − q4l+i+2) = qq2l+i+1(1 − q2l−2) + q−1q2l+i+1(1 − q2l+2) we obtain the second assertion. Moreover, A−1(0, l, i) = m(0, l) 1 2 (q−1 − q2l+1)(1 − q2l) (1 − q2l−2i) 1 2 (1 − q2l+2i) 1 2 (1 − q4l)(1 − q4l+2) 1 2 (1 − q4l−2) 1 2 and a−1(1, l, i, 1) = (qq−1 − q−1q2l+1)× (1 − q2l−2) 1 2 (1 − q2l−2i) 1 2 (1 − q2l+2) 1 2 (1 − q2l+2i) 1 2 (1 − q4l)(1 − q4l+2) 1 2 (1 − q4l−2) 1 2 . BAUM-CONNES CONJECTURE 21 Hence m(0, l) 1 2 (q−1 − q2l+1)(1 − q2l) = = (q2 − q2l) 1 2 (1 − q2l+2) (1 − q2l−2) 1 2 1 2 1 2 (q−1 − q2l+1)(1 − q2l) qq−1(1 − q2l+2)(1 − q2l) (1 − q2l+2) 1 2 qq−1(1 − q2l+2) = (1 − q2l−2) 1 2 (1 − q2l) = (qq−1 − q−1q2l+1)(1 − q2l−2) 1 2 (1 − q2l+2) 1 2 yields the third claim. The remaining assertions are verified in the same way. (cid:3) We need some further constructions. Recall that C(Ek) for k ∈ Z is a D(Gq)- equivariant Hilbert C(Gq/T )-module in a natural way. Left multiplication yields a D(Gq)-equivariant ∗-homomorphism ψ : C(Gq/T ) → K(C(Ek)), and (C(Ek), ψ, 0) defines a class [[Ek]] in KK D(Gq)(C(Gq/T ), C(Gq/T )). Moreover [[Em]] ◦ [[En]] = [[Em+n]] for all m, n ∈ Z. For k ∈ Z we define [Dk] ∈ KK D(Gq)(C(Gq/T ), C) by [Dk] = [[Ek]] ◦ [D] where [D] ∈ KK D(Gq)(C(Gq/T ), C) is the element obtained in proposition 4.2. Re- mark that [D0] = [D] since [[E0]] = 1. Evidently, the unit homomorphism u : C → C(Gq/T ) induces a class [u] in KK D(Gq)(C, C(Gq/T )). We define [Ek] in KK D(Gq)(C, C(Gq/T )) by restricting [[Ek]] along u, or equivalently, by taking the product [Ek] = [u] ◦ [[Ek]]. In the sequel we will interested in the elements αq ∈ KK D(Gq)(C(Gq/T ), C ⊕ C) and βq ∈ KK D(Gq)(C ⊕ C, C(Gq/T )) given by αq = [D0] ⊕ [D−1], βq = (−[E1]) ⊕ [E0], respectively. Theorem 4.6. Let q ∈ (−1, 1) \ {0}. Then C is a retract of C(Gq/T ) in KK D(Gq). More precisely, we have βq ◦ αq = id in KK D(Gq)(C ⊕ C, C ⊕ C). Proof. In order to prove the assertion we have to compute the Kasparov products [E0] ◦ [D] and [E±1] ◦ [D]. Let us first consider [E0] ◦ [D]. This class is obtained from the D(Gq)-equivariant Fredholm module D by forgetting the left action of C(Gq/T ). As already mentioned in the proof of proposition 4.2, the operator F intertwines the representations of C(Gq) on H1 and H−1 induced from the D(Gq)-Hilbert space structure. This can be read off from the fact that the coefficients xk(1, l, i, j) for x = a, b, c, d and k = −1, 0, 1 are symmetric in the variable j. It follows that the resulting D(Gq)-equivariant Kasparov C-C-module is degenerate, and hence [E0] ◦ [D] = 0 in KK D(Gq)(C, C). Let us now study [E−1] ◦ [D]. The underlying graded D(Gq)-Hilbert space of this Kasparov module is H = H0 ⊕ H−2 = L2(E0) ⊕ L2(E−2), and the corresponding action of C(Gq) is given by ω as defined in lemma 4.1. We choose the standard orthonormal basis vectors e(l) i,−1 for H−2. The i,0 for H0 and e(l) 22 operator is determined by CHRISTIAN VOIGT F =(cid:18) 0 F+ F− 0 (cid:19) F+(e(l) i,0) =(e(l) i,−1 0 l > 0 l = i = 0, F−(e(l) i,−1) = e(l) i,0. By construction, this operator is Gq-equivariant, but F does not commute with the action of the discrete part of D(Gq). Let us construct a D(Gq)-equivariant Kasparov C-C[0, 1]-module as follows. As underlying graded Gq-equivariant Hilbert C[0, 1]-module we take the constant field H ⊗ C[0, 1] = (H0 ⊗ C[0, 1]) ⊕ (H−2 ⊗ C[0, 1]) of Gq-Hilbert spaces. It follows from lemma 4.3 that Ω(x)(ξ)(t) = ωt(x)ξ(t) defines a ∗-homomorphism Ω : C(Gq) → L(H0 ⊗ C[0, 1]), and the corresponding coaction of C ∗(Gq) turns the even part H0⊗C[0, 1] into a D(Gq)-equivariant Hilbert C[0, 1]-module. In odd degree we consider the constant D(Gq)-Hilbert module structure induced from H−2. The left action of C on H⊗C[0, 1] is given by multiples of the identity operator, and as a final ingredient we take the constant operator F ⊗1 on H ⊗ C[0, 1]. By construction, this operator is Gq-equivariant, and it follows from lemma 4.4 that F ⊗ 1 commutes with the action of C(Gq) up to compact operators. Hence we have indeed defined a Kasparov module. Evaluation of this Kasparov module at t = 1 yields [E−1] ◦ [D]. The evaluation at t = 0 agrees with the cycle defining [E−1] ◦ [D] except that the action of C(Gq) on H0 is given by ω0 instead of ω1. We decompose H0 = C ⊕ C⊥ as a direct sum of the one-dimensional subspace C spanned by e(0) 0,0 and its orthogo- nal complement C⊥. Inspection of the explicit formulas shows that ω0 preserves this decomposition and implements the trivial representation ǫ : C(Gq) → C = L(C) on the first component. It follows that the Kasparov module decomposes as a direct sum of the trivial module C representing the identity and its orthogonal comple- ment which we denote by R. That is, the underlying graded D(Gq)-Hilbert space of R is C⊥ ⊕ H−2, the representation of C is given by multiples of the identity, and the operator F defines a Gq-equivariant isomorphism between C⊥ and H−2. For q > 0 we see from lemma 4.5 that F intertwines the representation ω0 on C⊥ with the representation ω on H−2. It follows that R is degenerate, and hence [E−1] ◦ [D] = id in KK D(Gq)(C, C) in this case. For q < 0 the module R fails to be degenerate since the representations of C(Gq) do not match. Let F be the operator on H−2 ⊕ H−2 given by F =(cid:18)0 1 1 0(cid:19) . We define a homotopy T by considering the constant field of graded Gq-Hilbert spaces H ⊗ C[0, 1] = (H−2 ⊕ H−2) ⊗ C[0, 1] ⊕ (H−2 ⊕ H−2) ⊗ C[0, 1], the representation φ : C → L(H ⊗ C[0, 1]) given by φ =(cid:18)φ+ 0 0 φ−(cid:19) BAUM-CONNES CONJECTURE 23 where φ±(λ)(cid:18)ξ η(cid:19) =(cid:18)λ ξ 0 (cid:19) and the constant operator (F ⊕ F ) ⊗ 1. Let us use the canonical identification of C⊥ with H−2 in order to view ω0 as a representation on H−2. We define a D(Gq)-Hilbert module structure on the even part (H−2 ⊕ H−2) ⊗ C[0, 1] of T by the action Θ+(x)(cid:18)ξ η(cid:19) (t) = U (t)(cid:18)ω0(x) 0 0 ω(x)(cid:19) U −1(t)(cid:18)ξ(t) η(t)(cid:19) of C(Gq) where U (t) is the rotation matrix U (t) =(cid:18) cos(πt/2) − sin(πt/2) sin(πt/2) cos(πt/2)(cid:19) . In odd degree we consider the constant D(Gq)-Hilbert module structure Θ−(x)(cid:18)ξ η(cid:19) =(cid:18)ω(x) 0 0 ω(x)(cid:19)(cid:18)ξ η(cid:19) =(cid:18)ω(x)(ξ) ω(x)(η)(cid:19) . Writing Θ = Θ+ ⊕ Θ− we see from lemma 4.4 and lemma 4.5 that the commutators [(F ⊕ F ) ⊗ 1, Θ(x)] are compact for all x ∈ C(Gq). It follows that T defines a D(Gq)-equivariant Kasparov C-C[0, 1]-module. The evaluation of T at t = 0 identifies with the sum of R and a degenerate module, and evaluation at t = 1 yields a degenerate module. We conclude that the Kasparov module R is homotopic to zero, and hence [E−1] ◦ [D] = id in KK D(Gq)(C, C). In a similar way one proves the relation (−[E1]) ◦ [D] = id in KK D(Gq)(C, C) for all q ∈ (−1, 1) \ {0}. The calculations are analogous and will be omitted. (cid:3) As a corollary of theorem 4.6 we obtain the following result. Theorem 4.7. Let q ∈ (0, 1). The standard Podle´s sphere C(Gq/T ) is isomorphic to C ⊕ C in KK D(Gq). Proof. According to theorem 6.7 in [25] the elements αq and βq considered above satisfy αq ◦ βq = id in KK D(Gq)(C(Gq/T ), C(Gq/T )). Hence due to theorem 4.6 these elements induce inverse isomorphisms in KK D(Gq) as desired. (cid:3) We have not checked wether the assertion of theorem 4.7 holds for q < 0 as well. For our purposes the following result is sufficient, compare [25]. Proposition 4.8. Let q ∈ (−1, 1) \ {0}. The standard Podle´s sphere C(Gq/T ) is isomorphic to C ⊕ C in KK Gq . Proof. For q > 0 this is an immediate consequence of theorem 4.7. Hence it re- mains to treat the case of negative q, and clearly it suffices to show αq ◦ βq = id in KK Gq (C(Gq/T ), C(Gq/T )) for all q ∈ (−1, 1) \ {0}. Note first that the definition of the C ∗-algebras C(Gq) and C(Gq/T ) makes sense also for q = 0. Moreover, the algebras C(Gq/T ) assemble into a T -equivariant con- tinuous field over [q, 1] for all q ∈ (−1, 1], see [9], [23]. Let us write C(G/T ) for the corresponding algebra of sections. The elements αt yield a class α ∈ KK T (C(G/T ), C[q, 1]), in particular, we have a well-defined element α0 for t = 0. Similarly, the elements βt determine a class β ∈ KK T (C[q, 1], C(G/T )). Using a comparison argument as in [25], the claim follows from the induction isomorphism KK Gq (C(Gq/T ), C(Gq/T )) ∼= KK T (C(Gq/T ), C) and the fact that α1 ◦ β1 = id ∈ KK G1(C(G1/T ), C(G1/T )). (cid:3) 24 CHRISTIAN VOIGT 5. The Baum-Connes conjecture for torsion-free discrete quantum groups In this section we recall the formulation of the Baum-Connes conjecture for torsion-free discrete quantum groups proposed by Meyer [19]. This involves some general concepts from homological algebra in triangulated categories. For more detailed information we refer to [24], [22], [19], [20]. Let G be a torsion-free discrete quantum group in the sense of definition 3.1. The equivariant Kasparov category KK G has as objects all separable G-C ∗-algebras, and KK G(A, B) as the set of morphisms between two objects A and B. Compo- sition of morphisms is given by the Kasparov product. The category KK G is tri- angulated with translation automorphism Σ : KK G → KK G given by the suspen- sion ΣA = C0(R, A) of a G-C ∗-algebra A. Every G-equivariant ∗-homomorphism f : A → B induces a diagram of the form ΣB / Cf / A f / B where Cf denotes the mapping cone of f . Such diagrams are called mapping cone triangles. By definition, an exact triangle is a diagram in KK G of the form ΣQ → K → E → Q which is isomorphic to a mapping cone triangle. Associated with the inclusion of the trivial quantum subgroup E in G we have the E : KK G → KK E = KK and an induction functor obvious restriction functor resG indG E(A) = C0(G) ⊗ A for A ∈ KK with action of G given by translation on the copy of C0(G). We consider the following full subcategories of KK G, E : KK → KK G. Explicitly, indG CCG = {A ∈ KK G resG CI G = {indG E(A)A ∈ KK}, E(A) = 0 ∈ KK} and refer to their objects as compactly contractible and compactly induced G-C ∗- algebras, respectively. Since G is torsion-free, it suffices to consider the trivial quantum subgroup in the definition of these categories. If there is no risk of confu- sion we will write CC and CI instead of CCG and CI G. The subcategory CC is localising, and we denote by hCIi the localising subcategory generated by CI. It follows from theorem 3.21 in [19] that the pair of localising subcategories (hCIi, CC) in KK G is complementary. That is, KK G(P, N ) = 0 for all P ∈ hCIi and N ∈ CC, and every object A ∈ KK G fits into an exact triangle ΣN / A / A / N with A ∈ hCIi and N ∈ CC. Such a triangle is uniquely determined up to isomor- phism and depends functorially on A. We will call the morphism A → A a Dirac element for A. The localisation LF of a homological functor F on KK G at CC is given by LF (A) = F ( A) where A → A is a Dirac element for A. By construction, there is an obvious map LF (A) → F (A) for every A ∈ KK G. In the sequel we write G ⋉f A and G ⋉r A for the full and reduced crossed products of A by G. Let us remark that in [25] these algebras are denoted by C ∗ f (G)cop ⋉f A and C ∗ r (G)cop ⋉r A, respectively. Definition 5.1. Let G be a torsion-free discrete quantum group and consider the functor F (A) = K∗(G ⋉r A) on KK G. The Baum-Connes assembly map for G with coefficients in A is the map µA : LF (A) → F (A). / / / / / / BAUM-CONNES CONJECTURE 25 We say that G satisfies the Baum-Connes conjecture with coefficients in A if µA is an isomorphism. We say that G satisfies the strong Baum-Connes conjecture if hCIi = KK G. Observe that the strong Baum-Connes conjecture implies the Baum-Connes con- jecture with arbitrary coefficients. Indeed, for A ∈ hCIi the assembly map µA is clearly an isomorphism. By the work of Meyer and Nest [20], the above terminology is consistent with the classical definitions in the case that G is a torsion-free discrete group. The strong Baum-Connes conjecture amounts to the assertion that G has a γ-element and γ = 1 in this case. In section 9 we will need further considerations from [22], [19] relying on the notion of a homological ideal in a triangulated category. Let us briefly discuss the relevant material adapted to our specific situation. We denote by J the homological ideal in KK G consisting of all f ∈ KK G(A, B) such that resG E(f ) = 0 ∈ KK(A, B). By definition, J is the kernel of the exact functor resG E : KK G → KK. The ideal J is compatible with countable direct sums and has enough projective objects. The J-projective objects in KK G are precisely the retracts of compactly induced G-C ∗-algebras. A chain complex · · · / Cn+1 dn+1 / Cn dn / / Cn−1 / · · · in KK G is J-exact if · · · / KK(A, Cn+1) (dn+1)∗/ / KK(A, Cn) (dn)∗/ / KK(A, Cn−1) / · · · is exact for every A ∈ KK. A J-projective resolution of A ∈ KK G is a chain complex · · · / Pn+1 dn+1 / / Pn dn / / Pn−1 / · · · d2 / P1 d1 / P0 of J-projective objects in KK G, augmented by a map P0 → A such that the augmented chain complex is J-exact. For our purposes it is important that a J-projective resolution of A ∈ KK G can be used to construct a Dirac element A → A. In general, this construction leads to a spectral sequence computing the derived functor LF (A). In the specific case of free orthogonal quantum groups that we are interested in, the spectral sequence reduces to a short exact sequence. This short exact sequence will be discussed in section 9 in connection with our K-theory computations. 6. The Baum-Connes conjecture for the dual of SUq(2) In this section we show that the dual of SUq(2) satisfies the strong Baum-Connes conjecture. We work within the general setup explained in the previous section, taking into account proposition 3.2 which asserts that the dual of SUq(2) is torsion- free. Let us remark that the strong Baum-Connes conjecture for the dual of the classical group SU (2) is a special case of the results in [21]. Theorem 6.1. Let q ∈ (−1, 1) \ {0}. The dual discrete quantum group of SUq(2) satisfies the strong Baum-Connes conjecture. Proof. In the sequel we write G = SUq(2). Due to Baaj-Skandalis duality it suffices to prove that every G-C ∗-algebra is contained in the localising subcategory T of KK G generated by all trivial G-C ∗-algebras. Let A be a G-C ∗-algebra. Theorem 4.6 implies that A is a retract of C(G/T ) ⊠G A in KK G, and according to theorem 3.6 in [25] there is a natural G-equivariant / / / / / / / / / 26 CHRISTIAN VOIGT T (A). T resG isomorphism C(G/T ) ⊠G A ∼= indG Since T = Z is a torsion-free discrete abelian group the strong Baum-Connes con- jecture holds for T . More precisely, the trivial T -C ∗-algebra C is contained in the localising subcategory hC0( T )i of KK T generated by C0( T ). Next observe that there is a T -equivariant ∗-isomorphism C0( T ) ⊗ B ∼= C0( T ) ⊗ Bτ for every T -C ∗- algebra B where Bτ denotes B with the trivial T -action. It follows that T ⋉ resG T (A) ∼= C ⊗ T ⋉ resG T (A) ∈ hC0( T ) ⊗ T ⋉ resG T (A)i in KK T . According to Baaj-Skandalis duality, this implies = hC0( T ) ⊗ (T ⋉ resG T (A))τ i resG T (A) ∼= T ⋉ T ⋉ resG T (A) ∈ h T ⋉(C0( T ) ⊗ (T ⋉ resG T (A))τ i = h(T ⋉ resG T (A))τ )i in KK T . Using proposition 4.8 we thus obtain indG T resG T (A) ∈ hC(G/T ) ⊗ (T ⋉ resG T (A))τ i ⊂ h(T ⋉ resG T (A))τ i ⊂ T in KK G since the induction functor indG Combining the above considerations shows A ∈ T and finishes the proof. (cid:3) Starting from theorem 6.1 it is easy to calculate the K-groups of C(SUq(2)) and C(SOq(3)). We shall not present these computations here. T is triangulated. 7. Free orthogonal quantum groups and monoidal equivalence In this section we review the definition of free orthogonal quantum groups and discuss the concept of monoidal equivalence for compact quantum groups. We begin with the definition of free orthogonal quantum groups. These quantum groups were introduced by Wang and van Daele [34], [31]. As usual, for a matrix u = (uij) of elements in a ∗-algebra we shall write u and ut for its conjugate and ij and (ut)ij = uji for transposed matrices, respectively. That is, we have (u)ij = u∗ the corresponding matrix entries. Definition 7.1. Let Q ∈ GLn(C) such that QQ = ±1. The group C ∗-algebra f (FO(Q)) of the free orthogonal quantum group FO(Q) is the universal C ∗-algebra C ∗ with generators uij, 1 ≤ i, j ≤ n such that the resulting matrix u is unitary and the relation u = QuQ−1 holds. In definition 7.1 we basically adopt the conventions in [4]. However, we write Q instead of F for the parameter matrix, and we deviate from the standard notation Ao(Q) = C ∗ f (FO(Q)). The latter is motivated from the fact that we shall view this C ∗-algebra as the group C ∗-algebra of a discrete quantum group. It is well-known that C(SUq(2)) can be written as the group C ∗-algebra of a free orthogonal quantum group for appropriate Q ∈ GL2(C). Moreover, the free quan- tum groups FO(Q) for Q ∈ GL2(C) exhaust up to isomorphism precisely the duals of SUq(2) for q ∈ [−1, 1] \ {0}. The quantum groups FO(Q) for higher dimensional matrices Q are still closely re- lated to quantum SU (2). In order to explain this, we shall discuss the notion of monoidal equivalence for compact quantum groups introduced by Bichon, de Rijdt and Vaes [8]. For the algebraic aspects of monoidal equivalences and Hopf-Galois theory we refer to [28]. As in section 2 we write Rep(G) for the C ∗-tensor category of finite dimensional representations of a compact quantum group G. Recall that the objects in Rep(G) are the finite dimensional representations of G, and the morphism sets consist of all intertwining operators. BAUM-CONNES CONJECTURE 27 Definition 7.2. Two compact quantum groups G and H are called monoidally equivalent if the representation categories Rep(G) and Rep(H) are equivalent as C ∗-tensor categories. We point out that in definition 7.2 the representation categories are only required to be equivalent as abstract C ∗-tensor categories. In fact, by the Tannaka-Krein reconstruction theorem [36], a compact quantum group G is determined up to isomorphism by the C ∗-tensor category Rep(G) together with its canonical fiber functor into the category of Hilbert spaces. Let H be a compact quantum group. An algebraic coaction λ : P → C[H] ⊙ P on the unital ∗-algebra P is called ergodic if the invariant subalgebra C(cid:3)C[H]P ⊂ P is equal to C. We say that P is a left Galois object if λ is ergodic and the Galois map γP : P ⊙ P → C[H] ⊙ P given by γP (x ⊙ y) = λ(x)(1 ⊙ y) is a linear isomorphism. Similarly one defines ergodicity for right coactions and the notion of a right Galois object. Like in Morita theory, it is important that monoidal equivalences can be imple- mented concretely. Definition 7.3. Let G and H be compact quantum groups. A bi-Galois object for G and H is a unital ∗-algebra P which is both a left C[H]-Galois object and a right C[G]-Galois object, such that the corresponding coactions turn P into a C[H]-C[G]- bicomodule. A linear functional ω on a unital ∗-algebra P is called a state if ω(x∗x) ≥ 0 for all x ∈ P and ω(1) = 1. A state ω is said to be faithful if ω(x∗x) = 0 implies x = 0. If P is in addition equipped with a coaction λ : P → C[H] ⊙ P then ω is called invariant if (id ⊙ω)λ(x) = ω(x)1 for all x ∈ P. The following result is proved in [8]. Theorem 7.4. Let G and H be monoidally equivalent compact quantum groups. Then there exists a bi-Galois object P for G and H such that F (H) = P (cid:3)C[G]H defines a monoidal equivalence F : Rep(G) → Rep(H), and there exists a canonical faithful state ω on P which is left and right invariant with respect to the coactions of C[G] and C[H], respectively. As a first application of the concept of monoidal equivalence let us record the following fact. Proposition 7.5. Let G and H be discrete quantum groups with monoidally equiv- alent duals. Then G is torsion-free iff H is torsion-free. Proof. As explained in [13], actions of monoidally equivalent compact quantum groups on finite dimensional C ∗-algebras are in a bijective correspondence. We will discuss this more generally, for arbitrary C ∗-algebras and on the level of equi- variant KK-theory, in section 8. Under this correspondence, actions associated to representations of G correspond to actions associated to representations of H. This immediately yields the claim. (cid:3) In the sequel we will make use of the following crucial result from [8], which in turn relies on the fundamental work of Banica [3], [4]. Theorem 7.6. Let Qj ∈ GLnj (C) such that QjQj = ±1 for j = 1, 2. Then the dual of FO(Q1) is monoidally equivalent to the dual of FO(Q2) iff Q1Q1 and Q2Q2 have the same sign and tr(Q∗ 1Q1) = tr(Q∗ 2Q2). 28 CHRISTIAN VOIGT In particular, for any Q ∈ GLn(C) such that QQ = ±1, the dual of FO(Q) is monoidally equivalent to SUq(2) for a unique q ∈ [−1, 1] \ {0}. Theorem 7.6 implies in particular that the dual of FO(Q) for Q ∈ GLn(C) and n > 2 is not monoidally equivalent to SU±1(2). With this in mind we obtain the following consequence of proposition 3.2 and proposition 7.5. Corollary 7.7. Let Q ∈ GLn(C) for n > 2 such that QQ = ±1. Then the free orthogonal quantum group FO(Q) is torsion-free. 8. Monoidal equivalence and equivariant KK-theory Extending considerations in [13], we discuss in this section the correspondence for actions of monoidally equivalent compact quantum groups. We show in par- ticular that the strong Baum-Connes property for torsion-free quantum groups is invariant under monoidal equivalence. Let G and H be monoidally equivalent compact quantum groups and let P be the bi-Galois object for G and H as in theorem 7.4. Moreover let A be a G-C ∗-algebra, and recall from section 2 that we write S(A) for the dense spectral ∗-subalgebra of A. The algebraic cotensor product F (A) = P (cid:3)C[G]S(A) ⊂ P ⊙ S(A) is again a ∗-algebra and carries an algebraic coaction λ : F (A) → C[H] ⊙ F (A) inherited from P. Consider the C ∗-algebra P ⊗ A where P denotes the minimal completion of P, that is, the C ∗-algebra generated by P in the GNS-representation of the invariant state ω. The left coaction on P turns P ⊗ A into an H-C ∗-algebra. Let F (A) = P (cid:3)GA be the closure of F (A) = P (cid:3)C[G]S(A) inside P ⊗ A. By construction, the coaction of P ⊗ A maps F (A) into C r(H) ⊗ F (A). In this way we obtain a coaction on F (A) which turns F (A) into an H-C ∗-algebra. If f : A → B is a G-equivariant ∗-homomorphism then id ⊗f : P ⊗ A → P ⊗ B induces an H-equivariant ∗-homomorphism id (cid:3)Gf : P (cid:3)GA → P (cid:3)GB. Conse- quently, we obtain a functor F : G-Alg → H-Alg by setting F (A) = P (cid:3)GA on objects and F (f ) = id (cid:3)Gf on morphisms. Here G-Alg and H-Alg denote the cat- egories of separable G-C ∗-algebras and H-C ∗-algebras, respectively. Note that a trivial G-C ∗-algebra A is mapped to the trivial H-C ∗-algebra F (A) ∼= A under the functor F . Moreover F (A ⊕ B) ∼= F (A) ⊕ F (B) for all G-C ∗-algebras A and B. By symmetry, we have the dual Galois object Q for H and G and a correspond- ing functor H-Alg → G-Alg. This functor sends an H-C ∗-algebra B to the G- C ∗-algebra Q(cid:3)HB. Here Q denotes the C ∗-algebra generated by Q in the GNS- representation associated to its natural invariant state η. Proposition 8.1. For every G-C ∗-algebra A there is a natural isomorphism of G-C ∗-algebras. Q(cid:3)H P (cid:3)GA ∼= A Proof. Consider first the case A = C r(G). In this case we have a canonical iso- morphism P (cid:3)GA ∼= P . By construction, Q(cid:3)HP ⊂ Q ⊗ P is the closure of Q(cid:3)C[H]P ∼= C[G]. Since η and ω are faithful states on Q and P , respectively, the state η ⊗ ω is faithful on Q ⊗ P , and hence also on Q(cid:3)HP . Moreover η ⊗ ω is both left and right invariant with respect to the natural coactions of C r(G). From this we conclude that the above inclusion C[G] → Q ⊗ P induces an equivariant ∗-isomorphism C r(G) ∼= Q(cid:3)HP . Now let A be an arbitrary G-C ∗-algebra. Using the previous discussion we obtain a well-defined injective ∗-homomorphism α : A → Q⊗P ⊗A by applying the coaction BAUM-CONNES CONJECTURE 29 followed with the isomorphism C r(G) ∼= Q(cid:3)HP ⊂ Q ⊗ P . Due to associativity of the cotensor product the coaction S(A) → C[G] ⊙ S(A) induces an isomorphism S(A) ∼= C[G](cid:3)C[G]S(A) ∼= Q(cid:3)C[H]P (cid:3)C[G]S(A). Now let π ∈ H. The spectral subspace πP is a finite dimensional right C[G]- comodule, and we observe that π(P ⊗ A) = (πP ) ⊙ A. In fact, we have π(P ⊗ A) = (pπ ⊗ id)(P ⊗ A) ⊂ (πP ) ⊙ A where pπ : P → πP is the projection operator, and the reverse inclusion is obvious. This implies π(P (cid:3)GA) = (πP )(cid:3)C[G]S(A) and hence S(P (cid:3)GA) = P (cid:3)C[G]S(A) for the spectral subalgebras. Using a symmetric argument for Q we conclude S(Q(cid:3)H P (cid:3)GA) = Q(cid:3)C[H]S(P (cid:3)GA) = Q(cid:3)C[H]P (cid:3)C[G]S(A). It follows that the image of S(A) in Q(cid:3)HP (cid:3)GA under the map α is dense. Hence α induces an equivariant ∗-isomorphism A ∼= Q(cid:3)HP (cid:3)GA as desired. (cid:3) We have thus proved the following theorem. Theorem 8.2. Let H and G be monoidally equivalent compact quantum groups. Then the categories G-Alg and H-Alg are equivalent. Our next aim is to extend the equivalence of theorem 8.2 to the level of equi- variant Kasparov theory. For this we need an appropriate version of the cotensor product for Hilbert mod- ules. Let E be a G-Hilbert A-module, and recall from section 2 that S(E) denotes the spectral submodule of E. We define P (cid:3)GE as the closure of P (cid:3)C[G]S(E) inside the Hilbert P ⊗ A-module P ⊗ E. It is straightforward to check that P (cid:3)GE is an H-Hilbert P (cid:3)GA-module. Moreover, the following assertion is proved in the same way as proposition 8.1. Proposition 8.3. For every G-Hilbert A-module E there is a natural isomorphism of G-Hilbert A-modules. Q(cid:3)HP (cid:3)GE ∼= E If E is a G-Hilbert A-module then K(E) is a G-C ∗-algebra in a natural way. The cotensor product constructions for C ∗-algebras and Hilbert modules are compatible in the following sense. Proposition 8.4. Let E be a G-Hilbert A-module. Then K(P (cid:3)GE) ∼= P (cid:3)GK(E) as H-C ∗-algebras. Proof. Note that there are canonical inclusions P (cid:3)GK(E) ⊂ P ⊗ K(E) ∼= K(P ⊗ E) and K(P (cid:3)GE) ⊂ K(P ⊗ E). These inclusions determine a homomorphism ιE : K(P (cid:3)GE) → P (cid:3)GK(E) of H-C ∗-algebras. Using proposition 8.3 and proposition 8.1 we obtain a map P (cid:3)GK(E) ∼= P (cid:3)GK(Q(cid:3)HP (cid:3)GE) → P (cid:3)GQ(cid:3)H K(P (cid:3)GE) ∼= K(P (cid:3)GE) where the middle arrow is given by id (cid:3)ιP (cid:3)G E. It is readily checked that this map is inverse to ιE . (cid:3) Let E and F be G-Hilbert-A-modules and let T ∈ L(E, F ) be G-equivariant. Then id ⊗T : P ⊗ E → P ⊗ F induces an adjointable operator id (cid:3)GT : P (cid:3)GE → P (cid:3)GF . If φ : A → L(E) is a G-equivariant ∗-homomorphism then id ⊗φ : P ⊗A → L(P ⊗E) induces an H-equivariant ∗-homomorphism id (cid:3)Gφ : P (cid:3)GA → L(P (cid:3)GE). Now let (E, φ, F ) be a G-equivariant Kasparov A-B-module with G-invariant op- erator F . Since G is compact we may restrict to such Kasparov modules in the definition of KK G. Using our previous observations it follows easily that 30 CHRISTIAN VOIGT (P (cid:3)GE, id (cid:3)Gφ, id (cid:3)GF ) is an H-equivariant Kasparov P (cid:3)GA-P (cid:3)GB-module. It is not difficult to check that this assignment is compatible with homotopies and Kasparov products. As a consequence we obtain the desired functor KK G → KK H extending the func- tor F : G-Alg → H-Alg defined above. This functor, again denoted by F , preserves exact triangles and suspensions. We may summarize our considerations as follows. Theorem 8.5. If H and G are monoidally equivalent compact quantum groups then the triangulated categories KK H and KK G are equivalent. Note that a trivial G-C ∗-algebra is mapped to the corresponding trivial H-C ∗- algebra under this equivalence. As a consequence we immediately obtain the fol- lowing assertion. Theorem 8.6. Let G and H be torsion-free discrete quantum groups whose dual compact quantum groups are monoidally equivalent. Then G satisfies the strong Baum-Connes conjecture iff H satisfies the strong Baum-Connes conjecture. Proof. Using Baaj-Skandalis duality and theorem 8.5 we see that KK G and KK H are equivalent triangulated categories. In addition, the compactly induced G-C ∗- algebra C0(G) ⊗ A corresponds to the compactly induced H-C ∗-algebra C0(H) ⊗ A under this equivalence. Hence hCI Gi = KK G holds iff hCI H i = KK H holds. (cid:3) Combining theorem 8.6 with theorem 6.1 yields the main result of this paper. Theorem 8.7. Let n > 2 and Q ∈ GLn(C) such that QQ = ±1. Then the free orthogonal quantum group FO(Q) satisfies the strong Baum-Connes conjecture. 9. Applications In this section we discuss consequences and applications of theorem 8.7. In particular, we show that free orthogonal quantum groups are K-amenable and compute their K-theory. The concept of K-amenability, introduced by Cuntz for discrete groups in [10], extends to the setting of quantum groups in a natural way [32]. More precisely, a discrete quantum group G is called K-amenable if the unit element in KK G(C, C) can be represented by a Kasparov module (E, π, F ) such that the representation of G on the Hilbert space E is weakly contained in the regular representation. As in the case of discrete groups, this is equivalent to saying that the canonical map G ⋉f A → G ⋉r A is an isomorphism in KK for every G-C ∗-algebra A. Of course, every amenable discrete quantum group is K-amenable. It is known [4] that FO(Q) is not amenable for Q ∈ GLn(C) with n > 2. The main application of theorem 8.7 is the following result. Theorem 9.1. Let n > 2 and Q ∈ GLn(C) such that QQ = ±1. Then the free orthogonal quantum group FO(Q) is K-amenable. In particular, the map K∗(C ∗ f (FO(Q))) → K∗(C ∗ r (FO(Q))) is an isomorphism. The K-theory of FO(Q) is K0(C ∗ f (FO(Q))) = Z, K1(C ∗ f (FO(Q))) = Z. These groups are generated by the class of 1 in the even case and the class of the fundamental matrix u in the odd case. BAUM-CONNES CONJECTURE 31 Proof. Let us write G = FO(Q). The reduced and full crossed product functors KK G → KK agree on hCIi because they agree for all compactly induced G-C ∗- algebras. Indeed, for B ∈ KK we have G ⋉f indG E(B) = G ⋉f (C0(G) ⊗ B) ∼= K ⊗ B ∼= G ⋉r (C0(G) ⊗ B) = G ⋉r indG f (G)) ∼= K∗(C ∗ E(B) by strong regularity. According to theorem 8.7 it follows that the canonical map G ⋉f A → G ⋉r A is an isomorphism in KK for every G-C ∗-algebra A, and this means precisely that G is K-amenable. As in section 5 we denote by J the homological ideal in KK G given by the ker- nel of the restriction functor resG E : KK G → KK. To compute the K-groups K∗(C ∗ r (G)) we shall construct a concrete J-projective resolution of the trivial G-C ∗-algebra C. Let G be the dual compact quantum group of G, and let us identify the set of irreducible representations of G with N. We write πk for the representation corre- sponding to k ∈ N and denote by Hk the underlying Hilbert space. In particular, π0 = ǫ is the trivial representation. Moreover, π1 identifies with the fundamental representation given by u, and we have dim(H1) = n. The representation ring R( G) of G is isomorphic to the polynomial ring Z[t] such that t corresponds to H1. Due to the Green-Julg theorem and the Takesaki-Takai duality theorem, we have a natural isomorphism KK G(C, G ⋉r B) ∼= K(KG ⊗ B) ∼= K(B) for B ∈ KK G. Consequently, taking into account KK G(C, C) ∼= R( G), the Kas- parov product KK G(C, C) × KK G(C, G ⋉r B) → KK G(C, G ⋉r B) induces an R( G)-module structure on K(B), and every element f ∈ KK G(B, C) defines an R( G)-module homomorphism f∗ : K(B) → K(C). For B = C0(G) this construction leads to the action of R( G) on itself by multipli- cation, and for B = C the corresponding module structure on Z is induced by the augmentation homomorphism ǫ : Z[t] → Z given by ǫ(t) = n. Let us now consider the Koszul complex 0 / C0(G) n−T / C0(G) λ / C in KK G defined as follows. The map λ : C0(G) → KG ∼= C is given by the regular representation. Moreover, n : C0(G) → C0(G) denotes the sum of n copies of the identity element. The morphism T : C0(G) → C0(G) ⊗ Mn(C) ∼= C0(G) is the ∗-homomorphism induced by the comultipliation ∆ : C0(G) → M (C0(G) ⊗ C0(G)) followed by projection onto the matrix block corresponding to π1 in the second factor. Let us determine the map T∗ : R( G) → R( G) induced by T on the level of K-theory. Consider the Hopf ∗-algebra C[ G] of matrix elements for G, and denote by h , i the natural bilinear pairing between C[ G] and C0(G). Under this pairing, the character χk ∈ C[ G] of the representation πk corresponds to the trace on K(Hk). Moreover we observe h∆(f ), χl ⊗ χ1i = hf, χlχ1i = hf, χl+1 + χl−1i = hf, χl+1i + hf, χl−1i for f ∈ K(Hk) ⊂ C0(G). According to the definition of T this implies T∗(Hk) = Hk+1 + Hk−1 for all k ∈ N, where we interpret H−1 = 0. multiplication by t under the identification Z[t] = R( G). It follows that T∗ corresponds to / / / 32 CHRISTIAN VOIGT Applying K-theory to the Koszul complex yields the exact sequence of Z[t]-modules 0 / Z[t] n−t / Z[t] ǫ / Z / 0 where ǫ is again the augmentation homomorphism given by ǫ(t) = n. Taking into account that C0(G) is J-projective, it follows easily that the Koszul complex yields a J-projective resolution of C. Since the Koszul resolution has length 1, we obtain a Dirac morphism C → C as in the proof of theorem 4.4 in [22]. The only piece of information that we need about this construction is that the G-C ∗-algebra C fits into an exact triangle C0(G) / C0(G) / C / ΣC0(G) in KK G. Here the first arrow C0(G) → C0(G) is given by n − T , but we will not make use of this fact. By applying the crossed product functor we obtain an exact triangle K / K / G ⋉f C / ΣK in KK. Hence the associated long exact sequence in K-theory takes the form Z K0(G ⋉f C) Z o K1(G ⋉f C) o / 0 0 f (G)) defines an element [u] ∈ K1(C ∗ f (G) has a counit the group K0(G ⋉f C) ∼= K0(C ∗ Since C ∗ f (G)) contains a direct summand Z generated by the unit element 1 ∈ C ∗ f (G). It follows that the upper left horizontal map in the diagram is an isomorphism. Hence the vertical arrow is zero, and the lower left horizontal map is an isomorphism as well. f (G)) ∼= K1(G ⋉f C). Clearly, the It remains to identify the generator of K1(C ∗ fundamental unitary u ∈ Mn(C ∗ f (G)). The discussion at the end of section 5 in [8] shows that we find a quotient homomor- phism π : C ∗ f (FO(M )) for some matrix M ∈ GL2(C). On the level of K-theory, the class [u] maps under π to the class of the fundamental matrix of C ∗ f (FO(M ))). Since M ∈ GL2(C), the quantum group FO(M ) is isomorphic to the dual of SUq(2) for some q ∈ [−1, 1] \ {0}. Let uq ∈ M2(C(SUq(2))) be the fundamental matrix of SUq(2). For positive q, the index pairing of uq with the Fredholm module corresponding to the Dirac operator on SUq(2) is known to be equal to 1, see [12]. The argument extends to the case of negative q in a straightforward way. As a consequence of these observations we conclude that [u] is a generator of K1(C ∗ (cid:3) In his work on quantum Cayley trees [33], Vergnioux has constructed an analogue of the Julg-Valette element for FO(Q). Our considerations imply that this element is homotopic to the identity, although we do not get an explicit homotopy. f (G) → C ∗ f (FO(M )) in K1(C ∗ f (G)) = Z, and this finishes the proof. Corollary 9.2. Let n > 2 and Q ∈ GLn(C) such that QQ = ±1. The Julg-Valette element for FO(Q) is equal to 1 in KK FO(Q)(C, C). Proof. Let us write G = FO(Q). An analogous argument with K-homology instead f (G), C) ∼= Z is gen- of K-theory as in the proof of theorem 9.1 shows that KK(C ∗ erated by the class of the counit ǫ : C ∗ f (G) → C. It follows that the forgetful map KK G(C, C) → KK(C, C) ∼= Z is an isomorphism. Since the Julg-Valette element in [33] has index 1 this yields the claim. (cid:3) / / / / / / / / / / / / O O / o o BAUM-CONNES CONJECTURE 33 In the special case Q = 1 ∈ GLn(C) the quantum group FO(Q) = FO(n) is uni- modular. Hence the Haar functional φ : C ∗ r (FO(n)) → C is a trace and determines an additive map φ0 : K0(C ∗ r (FO(n))) → Z. Theorem 9.3. Let n > 2. The free quantum group FO(n) satisfies the analogue of the Kadison-Kaplansky conjecture. That is, C ∗ r (FO(n)) does not contain nontrivial idempotents. r (FO(n)). Assume that p ∈ C ∗ Proof. The classical argument for free groups carries over. Since every idempo- tent is similar to a projection it suffices to show that C ∗ r (FO(n)) does not contain nontrivial projections. We know that the Haar functional is a faithful tracial state on C ∗ r (FO(n)) is a projection. Then from the pos- itivity of φ we obtain φ(p) ∈ [0, 1], and from the above considerations we know φ(p) = φ0([p]) ∈ Z. This implies p = 0 or 1 − p = 0. (cid:3) Finally, we note that the dual of SUq(2) does not satisfy the analogue of the Kadison-Kaplansky conjecture. In fact, there are lots of nontrivial idempotents in C(SUq(2)) for q ∈ (−1, 1) \ {0}. References [1] Baaj, S., Skandalis, G., C ∗-alg`ebres de Hopf et th´eorie de Kasparov ´equivariante, K-theory 2 (1989), 683 - 721 [2] Baaj, S., Skandalis, G., Unitaires multiplicatifs et dualit´e pour les produits crois´es des C ∗- alg`ebres, Ann. Sci. ´Ecole Norm. Sup. 26 (1993), 425 - 488 [3] Banica, T., Th´eorie des repr´esentations du groupe quantique compact libre O(n), C. R. Acad. Sci. Paris S´er. I Math. 322 (1996), 241 - 244 [4] Banica, T., Le group quantique compact libre U (n), Comm. Math. Phys. 190 (1997), 143 - 172 [5] Banica, T., Fusion rules for representations of compact quantum groups, Exposition. Math. 17 (1999), 313 - 337 [6] Baum, P., Connes, A., Geometric K-theory for Lie groups and foliations, Preprint IHES (1982) [7] Baum, P., Connes, A., Higson, N., Classifying space for proper actions and K-theory of group C ∗-algebras, in C ∗-algebras: 1943 - 1993 (San Antonio, TX, 1993), 241 - 291, Contemp. Math. 167, 1994 [8] Bichon, J., De Rijdt, A., Vaes, S., Ergodic coactions with large multiplicity and monoidal equivalence of quantum groups, Comm. Math. Phys. 262 (2006), 703 - 728 [9] Blanchard, E., D´eformations de C ∗-alg`ebres de Hopf, Bull. Soc. Math. France 124 (1996), 141 - 215 [10] Cuntz, J., K-theoretic amenability for discrete groups, J. Reine Angew. Math. 344 (1983), 180 - 195 [11] D ֒abrowski, L., Sitarz, A., Dirac operator on the standard Podle´s quantum sphere, in: Non- commutative geometry and quantum groups (Warsaw, 2001), 49 - 58, Banach Center Publ. 61, Polish Acad. Sci., Warsaw, 2003 [12] D ֒abrowski, L., Landi, G., Sitarz, A., van Suijlekom, W., V´arilly, J., The local index formula for SUq(2), K-theory 35 (2005), 375 - 394 [13] De Rijdt, A., Vander Vennet, N., Actions of monoidally equivalent compact quantum groups and applications to probabilistic boundaries, arXiv:math/0611175 (2006) [14] Kasparov, G. G., The operator K-functor and extensions of C ∗-algebras, Izv. Akad. Nauk SSSR Ser. Mat. 44 (1980), 571 - 636 [15] Kasparov, G. G., Equivariant KK-theory and the Novikov conjecture, Invent. Math. 91 (1988), 147 - 201 [16] Klimyk, A., Schmuedgen, K., Quantum groups and their representations, Texts and Mono- graphs in Physics, Springer, Berlin, 1997 [17] Kustermans, J., Locally compact quantum groups in the universal setting, Internat. J. Math. 12 (2001), 289 - 338 [18] Kustermans, J., Vaes, S., Locally compact quantum groups, Ann. Sci. ´Ecole Norm. Sup. 33 (2000), 837 - 934 [19] Meyer, R., Homological algebra in bivariant K-theory and other triangulated categories. II, arXiv:math.KT/0801.1344 (2008) 34 CHRISTIAN VOIGT [20] Meyer, R., Nest, R., The Baum-Connes conjecture via localisation of categories, Topology 45 (2006), 209 - 259 [21] Meyer, R., Nest, R., An analogue of the Baum-Connes isomorphism for coactions of compact groups, Math. Scand. 100 (2007), 301 - 316 [22] Meyer, R., Nest, R., Homological algebra in bivariant K-theory and other triangulated cate- gories, arXiv:math/0702146v2 (2007) [23] Nagy, G., Deformation quantization and K-theory, Perspectives on quantization (South Hadley, MA, 1996), 111 - 134, Contemp. Math. 214, 1998 [24] Neeman, A., Triangulated categories, Annals of Mathematics Studies 148, Princeton Univer- sity Press, 2001 [25] Nest, R., Voigt, C., Equivariant Poincar´e duality for quantum group actions, arXiv:0902.3987, to appear in J. Funct. Anal. [26] Podle´s, P., Quantum spheres, Lett. Math. Phys. 14 (1987), 193 - 202 [27] Podle´s, P., Symmetries of quantum spaces. Subgroups and quotient spaces of quantum SU (2) and SO(3) groups, Comm. Math. Phys. 170 (1995), 1 - 20 [28] Schauenburg, P., Hopf-Galois and bi-Galois extensions, in Galois theory, Hopf algebras, and semiabelian categories, 469 - 515, Fields Inst. Commun., 43, Amer. Math. Soc., Providence, RI, 2004 [29] Vaes, S., A new approach to induction and imprimitivity results, J. Funct. Anal. 229 (2005), 317 - 374 [30] Vaes, S., Vergnioux, R., The boundary of universal discrete quantum groups, exactness, and factoriality. Duke Math. J. 140 (2007), 35 - 84 [31] Van Daele, A., Wang, S., Universal quantum groups, Internat. J. Math. 7 (1996), 255 - 263 [32] Vergnioux, R., K-amenability for amalgamated free products of amenable discrete quantum groups, J. Funct. Anal. 212 (2004), 206 - 221 [33] Vergnioux, R., Orientation of quantum Cayley trees and applications, J. Reine Angew. Math. 580 (2005), 101 - 138 [34] Wang, S., Free products of compact quantum groups, Comm. Math. Phys. 167 (1995), 671 - 692 [35] Woronowicz, S. L., Twisted SU (2) group. An example of a noncommutative differential cal- culus, Publ. RIMS Kyoto 23 (1987), 117 - 181 [36] Woronowicz, S. L., Tannaka-Krein duality for compact matrix pseudogroups. Twisted SU(N ) groups, Invent. Math. 93 (1988), 35 - 76 [37] Woronowicz, S. L., Compact quantum groups, Sym´etries quantiques (Les Houches, 1995), 845 - 884, North-Holland, Amsterdam, 1998 Christian Voigt, Mathematisches Institut, Westfalische Wilhelms-Universitat Mun- ster, Einsteinstrasse 62, 48149 Munster, Germany E-mail address: [email protected]
1511.07049	2	1511	2017-01-10T11:45:42	Positive extensions of Schur multipliers	[ "math.OA" ]	We introduce partially defined Schur multipliers and obtain necessary and sufficient conditions for the existence of extensions to fully defined positive Schur multipliers, in terms of operator systems canonically associated with their domains. We use these results to study the problem of extending a positive definite function defined on a symmetric subset of a locally compact group to a positive definite function defined on the whole group.	math.OA	math	POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS RUPERT H. LEVENE, YING-FEN LIN, AND IVAN G. TODOROV Abstract. We introduce partially defined Schur multipliers and obtain necessary and sufficient conditions for the existence of extensions to fully defined positive Schur multipliers, in terms of operator systems canon- ically associated with their domains. We use these results to study the problem of extending a positive definite function defined on a symmetric subset of a locally compact group to a positive definite function defined on the whole group. 1. Introduction The problem of completing a partially defined matrix to a fully defined positive matrix has attracted considerable attention in the literature (see e.g. [2, 8, 12, 21]). Given an n by n matrix, only a subset of whose entries are specified, this problem asks whether the remaining entries can be determined so as to yield a positive matrix. The set κ of pairs (i, j) for which the (i, j)-entry is specified is called the pattern of the initial matrix; to avoid trivialities, κ is assumed to be symmetric and to contain the diagonal. One may then consider the operator system S(κ) of all (fully specified) matrices supported by κ. The operator systems arising in this way are precisely the operator subsystems of the space Mn of all n by n complex matrices, which are also bimodules over the algebra of all diagonal matrices. In [21], using operator space methods, Paulsen, Power and Smith formulated necessary and sufficient conditions for the existence of positive completions for a given pattern κ, in terms of S(κ), and related such completions to positivity and extendability of associated Schur multipliers with domain κ. In this paper, we study the corresponding extension problem in infinite dimensions. More precisely, we replace the Hilbert space Cn with the Hilbert space L2(X, µ) for some measure space (X, µ), and the algebra of diagonal matrices with the maximal abelian selfadjoint algebra D ≡ L∞(X, µ). Given a suitable measurable subset κ ⊆ X × X, we define a weak* closed D- bimodule S(κ), canonically associated with κ, and introduce the notion of a Schur multiplier with domain κ. We study the problem of extending such a (partially defined) Schur multiplier to a positive Schur multiplier defined on all of X × X, and relate it to the positivity structure of S(κ). Our motivation is two-fold. Firstly, we will see in Section 5 that the problem we consider is closely related to the problem of extending partially defined positive definite functions on locally compact groups. The latter problem has been studied in a variety of contexts and there is a rich bib- liography on its modern aspects as well as its connections with classical problems and applications (see [2], [3], [4], [7], [11], [19] and the references Date: 23 December 2016. 1 POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 2 therein). Since the main interest here lies in infinite groups, the passage to infinite dimensions becomes necessary. Secondly, bimodules over continuous maximal abelian selfadjoint algebras (masas, for short) have been instrumental in a number of contexts in Opera- tor Algebras. Introduced by Arveson in [1], they have proved useful in topics as diverse as spectral synthesis and uniqueness sets in Harmonic Analysis [27], closability of multipliers on Fourier algebras [26], finite rank approxi- mations [9] and structure of idempotents [15], among others. They are also closely related to Schur multipliers (see [20] and [23], as well as [29], where questions related to positivity of Schur multipliers were studied). However, masa-bimodules that are simultaneously operator systems have not received attention to date, despite their importance in modern Analysis [20]. The paper is organised as follows. In Section 2, we consider the discrete case and formulate a straightforward generalisation of several results in [21], which use a graph theoretic property of the pattern κ called chordality. We note that extension results for partially defined functions that are not necessarily Schur multipliers, in terms of chordality, were obtained in [30]. In Section 3, we study measurable versions of the patterns κ and their operator systems. Although these subsets κ can be thought of as measurable graphs, the passage from a discrete to a general measure space leads to substantial differences (see e.g. Corollary 5.3). In Section 4, we formulate necessary and sufficient conditions for the existence of an extension of a partially defined partially positive Schur multiplier to a fully defined positive Schur multiplier, in terms of approximation of positive operators by sums of rank one positive operators in the operator system S(κ). In Section 5, we study the problem of extending a positive definite func- tion defined on a symmetric subset E of a locally compact group to a pos- itive definite function defined on the whole group. The special case where E is a closed subgroup has attracted considerable attention previously (see e.g. [14]). Closely related problems about extension of Herz-Schur multipli- ers were recently considered in [6]. We use the results from Section 4 to give, in the case of discrete amenable groups, a different approach to the result of Bakonyi and Timotin [4] concerning positive definite extensions of partially defined functions. Our main result here (Theorem 5.6) concerns (classes of non-discrete) locally compact groups, where we formulate a sufficient con- dition for the existence of positive definite extensions in terms of operator approximations. For a Hilbert space H, we denote by B(H) (resp. K(H)) the space of all bounded linear (resp. compact) operators on H. We will often use basic concepts from Operator Space Theory, such as complete positivity; we refer the reader to [20] for the necessary background. As customary, the closure of a subset T in a topology τ will be denoted by T , and if T ⊆ B(H), then we will write T + for the set of all positive elements of T . τ 2. The discrete case Let X be a set and let H = ℓ2(X). With every element T of B(H), we associate the matrix (tx,y)x,y∈X given by tx,y = (T ey, ex), where (ex)x∈X is POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 3 the canonical orthonormal basis of H. For x, y ∈ X, denote by Ex,y the corresponding matrix unit in B(H) (so that Ex,yey = ex). For κ ⊆ X × X, define (1) S(κ) = span{Ex,y : (x, y) ∈ κ} w∗ . It is easy to see that an operator T ∈ B(H) is in S(κ) if and only if the corresponding matrix (tx,y) has tx,y = 0 whenever (x, y) is in the complement of κ. Throughout this section, we fix an additional (non-trivial) Hilbert space K and let H ⊗ K be the Hilbert space tensor product of H and K. The elements T ∈ B(H ⊗ K) can in a similar fashion be regarded as matrices (Tx,y)x,y∈X , where Tx,y ∈ B(K) is determined by the identity (Tx,yξ, η) = (T (ey ⊗ ξ), ex ⊗ η), x, y ∈ X, ξ, η ∈ K. Definition 2.1. A function ψ : κ → B(K) will be called an (operator valued) Schur multiplier if the matrix (tx,yψ(x, y))x,y∈X defines an element of B(H ⊗ K) for every (tx,y)x,y∈X ∈ S(κ) (here, we have set tx,yψ(x, y) = 0 provided (x, y) 6∈ κ). Let S0(κ) = span{Ex,y : (x, y) ∈ κ} k·k . We have that S0(κ) ⊆ K(H); let ι : S0(κ) → K(H) be the inclusion map. Lemma 2.2. (i) The second dual ι∗∗ of ι is a completely isometric weak* homeomorphism of S0(κ)∗∗ onto S(κ). (ii) Every bounded linear map Ψ : S0(κ) → B(H ⊗ K) has a bounded weak* continuous extension to a map Ψ : S(κ) → B(H ⊗ K). (iii) Every bounded linear map Ψ : K(H) → B(H) has a bounded weak* continuous extension to a map Ψ : B(H) → B(H). Moreover, if Ψ is com- pletely positive (resp. completely bounded) then Ψ is completely positive (resp. completely bounded). Proof. (i) The map ι∗∗ is a surjective completely isometric weak* homeo- morphism onto its range in B(H); since S0(κ) = S(κ), we conclude that the range of ι∗∗ is S(κ). w∗ (ii) Let E : B(H ⊗ K)∗∗ → B(H ⊗ K) be the canonical (weak* continuous) projection; thus, E(T ) = T whenever T ∈ B(H ⊗ K). Set Ψ = E ◦ Ψ∗∗ ◦ (ι∗∗)−1; thus, Ψ : S(κ) → B(H ⊗ K) is weak* continuous and its restriction to S0(κ) coincides with Ψ. (iii) The statement follows from (ii) after letting K = C and κ = X × X. The remaining statements follow from the facts that E is unital and com- pletely positive (and hence completely bounded) and that Ψ∗∗ is completely positive (resp. bounded) provided Ψ is so. (cid:3) Proposition 2.3. Let κ ⊆ X × X. If ψ : κ → B(K) is a Schur multiplier, then the map Sψ : S(κ) → B(H ⊗ K), is bounded and weak* continuous. (tx,y) 7→(cid:0)tx,yψ(x, y)(cid:1) POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 4 Proof. To show that Sψ is bounded, we use the Closed Graph Theorem. Suppose that Tn = (tn x,y)x,y∈X are operators in S(κ) such that Tn → 0 and Sψ(Tn) → S in norm as n → ∞, for some S ∈ B(H ⊗ K). Letting S = (Sx,y)x,y∈X , we have tn x,yψ(x, y) → Sx,y in norm for each (x, y) ∈ κ. It follows that Sx,y = 0 for each x, y ∈ X, and hence S = 0. Thus, Sψ is bounded. x,y → 0 and tn Let S0 ψ be the restriction of Sψ to the subspace S0(κ) defined before Lemma 2.2. Let Ψ be the weak* continuous extension of S0 ψ guaranteed by Lemma 2.2 (ii). Fix T = (tx,y)x,y∈X ∈ S(κ). For a finite set F ⊆ X, let PF be the projection on H whose range has basis {ex : x ∈ F } and set TF = PF T PF . The net (TF )F lies in S0(κ) and TF →F T in the weak* topology. For all x, y ∈ X and all ξ, η ∈ K, we have (Ψ(T )(ey ⊗ ξ), ex ⊗ η) = w∗- lim F (Ψ(TF )(ey ⊗ ξ), ex ⊗ η) ψ(TF )(ey ⊗ ξ), ex ⊗ η) = w∗- lim F (S0 = (tx,yψ(x, y)ξ, η) = (Sψ(T )(ey ⊗ ξ), ex ⊗ η). We conclude that Ψ = Sψ, and hence Sψ is weak* continuous. (cid:3) Note that if ψ : κ → B(K) is a Schur multiplier then the range of the map Sψ, defined in Proposition 2.3, is contained in the weak* closed spatial tensor product S(κ) ¯⊗B(K), of S(κ) and B(K). A Schur multiplier ϕ : X × X → B(K) will be called positive if the map Sϕ is positive, that is, if for every positive operator T ∈ B(H), the operator Sϕ(T ) ∈ B(H ⊗ K) is also positive. An application of [21, Proposition 1.2] shows that if ϕ is a positive Schur multiplier then the map Sϕ is in fact completely positive. Let κ ⊆ X × X. It is straightforward to verify that the subspace S(κ) is an operator system (i.e. a selfadjoint unital subspace of B(H)) if and only if (i) κ is symmetric (that is, (x, y) ∈ κ implies that (y, x) ∈ κ), and (ii) κ contains the diagonal of X × X (that is, (x, x) ∈ κ for every x ∈ X). Note that such subsets κ can be identified with (undirected) graphs with vertex set X and no loops: for distinct elements x, y ∈ X, the subset {x, y} is an edge if, by definition, (x, y) ∈ κ. Thus, subsets satisfying properties (i) and (ii) above will hereafter be referred to as graphs. Definition 2.4. Let κ ⊆ X ×X be a graph and let ψ : κ → B(K) be a Schur multiplier. We say that ψ is partially positive if for every subset α ⊆ X with α × α ⊆ κ, the Schur multiplier ψα×α is positive. Let κ ⊆ X × X be a graph. A Schur multiplier ϕ : X × X → B(K) will be called an extension of the Schur multiplier ψ : κ → B(K) if the restriction ϕκ of ϕ to κ coincides with ψ. We will be interested in the question of when a Schur multiplier ψ : κ → B(K) possesses a positive extension. Clearly, a necessary condition for this to happen is that ψ be partially positive. In Theorem 2.5, we will identify conditions which ensure that the converse implication holds true. We say that the vertices x1, . . . , xn form a cycle of κ (of length n) if (xi, xi+1) ∈ κ for all i (where addition is performed mod n). A chord in POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 5 such a cycle is an edge of the form (xi, xk), where 2 ≤ i − k ≤ n − 2. We say that κ is chordal (see e.g. [21]) if every cycle of length at least four has a chord. Theorem 2.5. Let κ ⊆ X × X be a graph. The following conditions are equivalent: (i) every partially positive Schur multiplier ψ : κ → B(K) has a positive extension; (ii) the graph κ is chordal; (iii) every positive operator in S(κ) is the weak* limit of sums of rank one positive operators in S(κ). Proof. We denote by F an arbitrary finite subset of X, and let κF = κ ∩ (F × F ). (i)⇒(ii) We claim that condition (i) is satisfied for the graph κF . Indeed, given a partially positive Schur multiplier ψF : κF → B(K), let ψ : κ → B(K) be the extension of ψF with ψ(x, y) = 0 if (x, y) ∈ κ \ κF . Since ψ If α ⊆ X is a set with has finite support, ψ is a Schur multiplier on κ. α × α ⊆ κ, then Sψα×α = SψF (α∩F )×(α∩F ) ⊕ 0. Since ψF is partially positive, SψF (α∩F )×(α∩F ) is positive, so Sψα×α is positive. Thus, ψ is partially positive. By assumption, ψ has a positive extension, whose restriction to F × F is a positive extension of ψF . It now follows from [21] that κF is chordal, and since this holds for every finite set F , we conclude that κ is chordal. (ii)⇒(iii) Let PF be the projection from H onto ℓ2(F ) (when the latter is viewed as a subspace of H in the natural way). If T ∈ S(κ) is a positive operator then T = limF PF T PF in the weak* topology. On the other hand, κF is chordal and hence, by [21], PF T PF is the sum of rank one positive operators in S(κ). (iii)⇒(i) Suppose that ψ : κ → B(K) is a partially positive Schur mul- tiplier. It is clear that ψF := ψκF is a partially positive Schur multiplier on κF . Since S(κF ) = PF S(κ)PF ⊆ S(κ), every positive operator in S(κF ) is the weak* limit of sums of rank one positive operators in S(κF ); since S(κF ) is finite dimensional, we have that, in fact, every positive operator in S(κF ) is the norm limit of sums of rank one positive operators in S(κF ). Now [21] implies that there exists a positive Schur multiplier ϕF : F × F → B(K) whose restriction to κF coincides with ψF . Let ϕF : X × X → B(K) be defined by ϕF (x, y) =(ϕF (x, y) 0 if x, y ∈ F ; otherwise. Then the map S ϕF : B(H) → B(H ⊗K) is completely positive. On the other hand, since ψ is a Schur multiplier on κ and I ∈ S(κ), by [20, Proposition 3.6], we have that kS ϕF kcb = kS ϕF (I)k = max x∈F kϕF (x, x)k = max x∈F kψF (x, x)k ≤ sup x∈X kψ(x, x)k = kSψ(I)k < ∞. So {S ϕF }F is a uniformly bounded family of completely positive maps from B(H) into B(H ⊗ K). By [20, Theorem 7.4], there exist a subnet (ΦF ′)F ′ and a completely positive map Φ : B(H) → B(H ⊗ K) such that ΦF ′(T ) → POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 6 Φ(T ) along F ′ in the weak* topology, for every T ∈ B(H). We have that Φ(AT B) = (A ⊗ I)Φ(T )(B ⊗ I) for all diagonal operators A, B and all T ∈ B(H), and this easily implies that Φ(Ex,y) = Ex,y ⊗ ϕ(x, y) for some ϕ(x, y) ∈ B(K). Since ΦF ′(Ex,y) = Ex,y ⊗ ψ(x, y) for (x, y) ∈ κF ′, the map ϕ : X × X → B(K) extends ψ. Now, for T = (tx,y)x,y∈X ∈ B(H), ξ, η ∈ K and x, y ∈ X, we have (Φ(T )(ey ⊗ ξ), ex ⊗ η) = ((Ex,x ⊗ I)Φ(T )(Ey,y ⊗ I)(ey ⊗ ξ), ex ⊗ η) = (Φ(tx,yEx,y)(ey ⊗ ξ), ex ⊗ η) = (tx,yϕ(x, y)ξ, η), so Φ(T ) = (tx,yϕ(x, y))x,y. It follows that ϕ is a Schur multiplier and Φ = Sϕ; since Φ is completely positive, ϕ is positive. (cid:3) 3. Positivity domains In this section, we study the domains of Schur multipliers over arbitrary standard σ-finite measure spaces. To set the stage, we recall some measure theoretic terminology [9]. Let (X, µ) be a standard σ-finite measure space. The set X × X will be equipped with the product σ-algebra and the product measure µ × µ. A subset E ⊆ X × X is called marginally null if E ⊆ (M ×X)∪(X ×M ), where M ⊆ X is null. We call two subsets E, F ⊆ X ×X marginally equivalent (resp. equivalent), and write E ∼= F (resp. E ∼ F ), if their symmetric difference is marginally null (resp. null with respect to product measure). We say that E is marginally contained in F (and write E ⊆ω F ) if the set difference E \ F is marginally null. A subset κ ⊆ X × X will be called • a rectangle if κ = α × β where α, β are measurable subsets of X; • a square if κ = α × α where α is a measurable subset of X; • ω-open if it is marginally equivalent to a countable union of rectan- gles, and • ω-closed if its complement κc is ω-open. Recall that, by [25], if E is any collection of ω-open sets, then there exists a smallest, up to marginal equivalence, ω-open set ∪ωE, called the ω-union of E, such that every set in E is marginally contained in ∪ωE. Given a measurable set κ, one defines its ω-interior to be intω(κ) = ∪ω{R : R is a rectangle with R ⊆ κ}. The ω-closure clω(κ) of κ is defined as the complement of intω(κc). For a set κ ⊆ X × X, we write κ = {(x, y) ∈ X × X : (y, x) ∈ κ}. Unless otherwise stated, we use the symbol H to denote the Hilbert space L2(X, µ). For each a ∈ L∞(X, µ), let Ma : H → H be the multiplication operator given by Maf = af and let D = {Ma : a ∈ L∞(X, µ)} be the multiplication algebra of L∞(X, µ); we have that D is a masa in B(H). For a measurable subset α ⊆ X, we let χα denote the characteristic function of α, and set P (α) = Mχα, a projection in D. A D-bimodule (or simply a masa-bimodule) is a subspace S ⊆ B(H) such that DSD ⊆ S. Let κ be a measurable subset of X × X. An operator T ∈ B(H) is said to be supported by κ if P (β)T P (α) = 0 whenever (α × β) ∩ κ ∼= ∅. Given any masa-bimodule U , there exists a unique (up to marginal equivalence) POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 7 smallest ω-closed set κ ⊆ X × X such that every operator in U is supported by κ [9]. The set κ is denoted by supp U and is called the support of U . For k ∈ L2(X × X), the Hilbert-Schmidt operator Tk : H → H with integral kernel k is defined by Tkf (y) =ZX k(y, x)f (x)dµ(x), f ∈ H, y ∈ X. For any measurable, ω-closed subset κ ⊆ X × X, let S(κ) = {Tk : k ∈ L2(κ)} w∗ , where L2(κ) is the space of functions in L2(X × X) which are supported It is easy to see that every operator T ∈ S(κ) is supported by κ, on κ. so supp S(κ) ⊆ω κ. Note that the latter inclusion may be strict. Indeed, if κ has product measure zero then S(κ) = {0} and hence supp S(κ) ∼= ∅; however, κ does not need to be marginally null. Suppose that X is equipped with the counting measure. Then D is the algebra of all diagonal operators on ℓ2(X) and S(κ) is the weak* closure of the linear span of the matrix units Ex,y, with (x, y) ∈ κ; it thus coincides with the space defined in Section 2 (see (1)). In particular, S(κ) is generated, as a weak* closed subspace, by the rank one operators it contains. In view of Proposition 3.2 below, in this general context it is therefore natural to restrict attention to sets κ which contain "plenty of rectangles". We will now make this intuitive idea precise. Definition 3.1. A measurable subset κ ⊆ X × X is said to be: (i) generated by rectangles if κ ∼= clω(intω(κ)); (ii) symmetric if κ ∼= κ. For ξ, η ∈ H, we denote by ξ ⊗ η∗ the rank one operator on H given by (ξ ⊗ η∗)(ζ) = (ζ, η)ξ. We denote by S1(κ) the set of all rank one operators in S(κ). Proposition 3.2. Let κ ⊆ X × X be an ω-closed set. (i) If ξ, η ∈ H, then the rank one operator ξ ⊗ η∗ belongs to S(κ) if and only if (supp ξ) × (supp η) is marginally contained in κ. (ii) The set S1(κ) generates S(κ) as a weak* closed subspace of B(H) if and only if κ is equivalent to clω(intω(κ)). Proof. (i) Suppose that ξ ⊗η∗ ∈ S(κ). Then the operator ξ ⊗η∗ is supported by κ. Writing α = supp ξ, β = supp η and κc ∼= ∪∞ i=1αi×βi, where αi, βi ⊆ X are measurable for i ∈ N, we have, in particular, that (βi × αi) ∩ κ ∼= ∅, so (χαiξ) ⊗ (χβiη)∗ = P (αi)(ξ ⊗ η∗)P (βi) = 0, hence either αi∩α or βi∩β is null, for each i ∈ N. It follows that (α×β)∩κc ∼= ∅, so α × β ⊆ω κ. Conversely, if α × β ⊆ω κ then the integral kernel of ξ ⊗ η∗ is clearly in L2(κ), and thus ξ ⊗ η∗ ∈ S(κ). (ii) Set λ = clω(intω(κ)). Let V = span(S1(λ)) and U = Ref(V) be the reflexive hull of V in the sense of [18]. By [9, Theorem 5.2], supp U ∼= λ; since supp V ∼= supp U [9], we conclude that supp V ∼= λ. Thus, w∗ V ⊆ S(λ) ⊆ U . POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 8 Since Hilbert-Schmidt operators are pseudo-integral in the sense of [1], we have that both V and S(λ) are masa-bimodules, generated as weak* closed subspaces by pseudo-integral operators supported by λ. It follows from [1] that V = S(λ). If κ ∼ λ then S(κ) = S(λ) = V = span(S1(κ)) w∗ . Conversely, suppose that span(S1(κ)) the previous paragraph implies that w∗ = S(κ). By (i), S1(λ) = S1(κ), and span(S1(κ)) w∗ = span(S1(λ)) w∗ = S(λ). Thus, S(κ) = S(λ) and it now easily follows that κ ∼ λ. (cid:3) Let ∆ = {(x, x) : x ∈ X} denote the diagonal of X. Proposition 3.3. If κ ⊆ X×X is generated by rectangles, then the following are equivalent: (i) S(κ) is an operator system; (ii) κ is symmetric and ∆ ⊆ω κ. w∗ Proof. (i)⇒(ii) By Proposition 3.2 and its proof, S(κ) = span(S1(κ)) and supp S(κ) ∼= κ. Similarly, letting S(κ)∗ = {T ∗ : T ∈ S(κ)}, we have that supp S(κ)∗ ∼= κ; thus κ ∼= κ. On the other hand, by assumption, I ∈ S(κ); since S(κ) is a masa-bimodule, we have that D ⊆ S(κ). Thus, ∆ ∼= supp D ⊆ω κ. (ii)⇒(i) If k ∈ L2(κ) then the function k∗ given by k∗(x, y) = k(y, x) is k = Tk∗ ∈ S(κ). in L2(κ). Since κ ∼= κ, we have that k∗ ∈ L2(κ) and thus T ∗ Hence, S(κ)∗ = S(κ). Since D consists of pseudo-integral operators in terms of [1] and supp D ∼= (cid:3) ∆, by Proposition 3.2 and its proof we have that D ⊆ S(κ). Definition 3.4. A measurable set κ ⊆ X × X will be called a positivity domain if κ is generated by rectangles, κ is symmetric and ∆ ⊆ω κ. By Proposition 3.2 and Proposition 3.3, an ω-closed set κ ⊆ X × X is equivalent to a positivity domain if and only if S(κ) is an operator system, which is generated (as a weak* closed linear space) by the rank one operators it contains. Note that every operator system is generated, as a linear space, by the positive operators it contains. However, as we will show in Corollary 5.3, S(κ) does not need to contain a positive rank one operator. It is worth noting that this phenomenon does not occur in the case of a discrete measure space X; indeed, in this case, any diagonal matrix unit belongs to S(κ). Remark 3.5. Recall that a Schur idempotent is a weak* continuous D- bimodule map Φ ∈ B(B(H)) such that Φ ◦ Φ = Φ. Suppose that Φ is a (completely) positive Schur idempotent; in this case, the range Ran Φ of Φ is selfadjoint. Trivially, Ran Φ is an operator system if and only if Φ is unital, that is, if and only if Φ(I) = I. By [20, Proposition 3.6], kΦk = kΦ(I)k, and it follows that if Ran Φ is an operator system then Φ is contractive. i=1 αi × αi i=1 of pairwise disjoint measurable subsets of X with Conversely, if kΦk = 1 then, by [15], Ran Φ = S(κ), where κ ∼=S∞ i=1αi = X, and hence S(κ) is in this case an operator system. for some sequence (αi)∞ ∪∞ POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 9 We next introduce a way to quantify the amount of positive rank one operators contained in S(κ), where κ is a positivity domain. To this end, we define the square interior sqintω(κ) of κ by sqintω(κ) def = [ ω{Q : Q is a square with Q ⊆ κ}. By Proposition 3.2 (i), S(κ) contains a positive rank one operator if and only if sqintω(κ) 6∼= ∅. We will say that κ is generated by squares if κ ∼= clω(sqintω(κ)). Theorem 3.6 below characterises the positivity domains with this property. For a positivity domain κ ⊆ X × X, let S + 1 (κ) denote the set of positive rank one operators in S(κ), and set [S + 1 (κ)] =( k Xi=1 Ri : k ∈ N and Ri ∈ S + 1 (κ) for 1 ≤ i ≤ k) . Theorem 3.6. Let κ ⊆ X × X be a positivity domain. The following are equivalent: (i) there is a unital contractive Schur idempotent Φ with Ran Φ ⊆ S(κ); i=1 αi of X into measurable sub- (ii) there is a countable partition X = S∞ sets αi so that αi × αi ⊆ω κ for each i ∈ N; w∗ 1 (κ)] (iii) I ∈ [S + (iv) ∆ ⊆ω sqintω(κ); (v) κ is generated by squares. ; Proof. (i)⇔(ii) follows from [15] (see also Remark 3.5), while the implication (ii)⇒(iv) follows from the inclusions ∆ ⊆ ∞ [i=1 αi × αi ⊆ω sqintω(κ). (ii)⇒(iii) Clearly, P (αj) ∈ [S + 1 (αj × αj)] hence w∗ ; thus P (αj) ∈ [S + 1 (κ)] w∗ and I = w∗- lim N→∞ P (αj) ∈ [S + 1 (κ)] w∗ . N Xj=1 element in Σ, say {Φi}i∈N. Define Φ by Φ(T ) = P∞ (iii)⇒(i) Let Σ be the collection of all countable sets {Ψi}i∈N where each Ψi is a contractive Schur idempotent with Ran Ψi ⊆ S(κ) and ΨiΨj = 0 for all i 6= j. We order Σ by inclusion. The set Σ is non-empty, since it contains the singleton consisting of the zero map. Suppose that Λ ⊆ Σ is a non-empty chain. Since H is separable, Λ is countable, and hence its union is its upper bound. By Zorn's Lemma, there exists a maximal i=1 Φi(T ), T ∈ B(H) (where the sum converges in the strong operator topology), so that Φ is a contractive Schur idempotent with range contained in S(κ). It remains to show that Φ is unital, or equivalently, that I ∈ Ran(Φ). Let (αi)i∈N be a (countable) family of mutually disjoint measurable subsets of X such that i=1 P (αi)T P (αi), T ∈ B(H) (see [15] or Remark 3.5). Set P = i=1 P (αi) (where the sum converges in the strong operator topology). We claim that P = I. To see this, suppose that P ⊥ = P (α) for some measurable subset α of X and note that P ⊥S(κ)P ⊥ = S((α × α) ∩ κ) is an operator Φ(T ) = P∞ P∞ POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 10 system on the Hilbert space P ⊥H. Since S(κ) satisfies the assumption in (iii), so does P ⊥S(κ)P ⊥; so if P ⊥ 6= 0, then S + 1 ((α × α) ∩ κ) 6= ∅. Hence by Proposition 3.2, there exists a non-trivial square β ×β, marginally contained in (α × α) ∩ κ. Letting Φ0 be the Schur idempotent given by Φ0(T ) = P (β)T P (β), we see that the family {Φi}∞ i=1, a contradiction. It follows that P = I and hence I ∈ Ran Φ. i=0 strictly contains {Φi}∞ (iv)⇒(ii) Suppose that ∆ ⊆ω sqintω(κ). By [25, Lemma 2.1] there exists a family (βi)i∈N of measurable (not necessarily mutually disjoint) subsets of X such that ∞ ∆ ⊆ω βi × βi ⊆ω κ. [i=1 This implies that, up to a null set, S∞ βn \ (Sn−1 j=1 αj). Then αn × αn ⊆ βn × βn ⊆ω κ and S∞ null set which we may adjoin to (say) α1. i=1 βi = X. Let α1 = β1 and αn = n=1 αn = X, up to a (ii)⇒(v) Let (αi)i∈N be the partition from (ii). Given a rectangle β ×γ ⊆ω κ, observe that γ × β ⊆ω κ ∼= κ since κ is symmetric, and for i, j ∈ N let βi = β ∩ αi and γj = γ ∩ αj. Then βi × γj ⊆ (βi ∪ γj) × (βi ∪ γj) = (βi × βi) ∪ (βi × γj) ∪ (γj × βi) ∪ (γj × γj) ⊆ (αi × αi) ∪ (β × γ) ∪ (γ × β) ∪ (αj × αj) ⊆ω κ. It follows that β × γ = ∞ [i,j=1 βi × γj ⊆ω κ. Taking the ω-union over all such rectangles, we conclude that intω(κ) ∼= sqintω(κ). Since κ is generated by rectangles, we have κ ∼= clω(intω(κ)) ∼= clω(sqintω(κ)), and hence κ is generated by squares. (v)⇒(iv) By [25, Lemma 2.1], we can write sqintω(κ) ∼= ∪∞ i=1(αi × αi), for some sequence (αi)i∈N of measurable subsets of X. Set Y = ∪i∈Nαi. Then sqintω(κ) ⊆ Y × Y and, since Y × Y is ω-closed, κ ∼= clω(sqintω(κ)) ⊆ Y × Y. Since ∆ ⊆ω κ, we have Y ∼ X. It follows that ∆ ⊆ω sqintω(κ). (cid:3) It is easy to see that if κ ⊆ X × X is a positivity domain and Remark. λ = clω(sqintω(κ)) then the operator system S(λ) coincides with the weak* closure of the linear span of [S + 1 (κ)]. Thus, the square interior of κ can be viewed as a measure of the quantity of the positive rank one operators in S(κ). POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 11 4. Partially defined Schur multipliers Let (X, µ) be a standard measure space. Recall that a function ϕ ∈ L∞(X × X) is called a Schur multiplier if the map Sϕ defined on the space of all Hilbert-Schmidt operators on H = L2(X, µ) by k ∈ L2(X × X), Sϕ(Tk) = Tϕk, (2) is bounded in the operator norm. If ϕ is a Schur multiplier, we denote again by Sϕ the bounded weak* continuous extension of this map to B(H); note that Sϕ is a masa-bimodule map in the sense that Sϕ(BT A) = BSϕ(T )A, for all T ∈ B(H), A, B ∈ D. By [28, Theorem 2.1], Sϕ is automatically completely bounded. We denote the set of all Schur multipliers on X × X by S(X). A Schur multiplier ϕ is called positive if the map Sϕ : B(H) → B(H) is positive. In this case, Sϕ is automatically completely positive (see [28, Theorem 2.1] and [29, Lemma 4.3]). We write S(X)+ for the set of positive Schur multi- pliers on X × X. We record the following well-known fact, which follows from results of Haagerup [13] and Smith [28]. Theorem 4.1. Let Φ : B(H) → B(H) be a linear map. The following are equivalent: (i) Φ is a bounded weak* continuous D-bimodule map; (ii) Φ is a completely bounded weak* continuous D-bimodule map; (iii) Φ = Sϕ for some ϕ ∈ S(X); (iv) there exist families (ai)i∈N, (bi)i∈N ⊆ L∞(X) such that esssup x∈X ∞ Xi=1 ϕ(x, y) = and Xi=1 ai(x)2 < ∞, ∞ esssup y∈X ∞ Xi=1 bi(y)2 < ∞ ai(x)bi(y), a.e. (x, y) ∈ X × X. We extend the definition of a Schur multiplier given above to functions defined on proper subsets of X × X. Let κ ⊆ X × X be a measurable set, equipped with the induced σ-algebra and, as before, identify L2(κ) with the space of all functions in L2(X × X) supported on κ. Definition 4.2. Let κ ⊆ X × X be a measurable subset generated by rect- angles. A measurable function ϕ : κ → C will be called a Schur multiplier if there exists C > 0 such that kTϕkk ≤ CkTkk, for every k ∈ L2(κ). Let S0(κ) = {Tk : k ∈ L2(κ)} k·k ; thus, S0(κ) ⊆ S(κ) ∩ K(H) and S0(κ) = S(κ). If κ is a positivity domain, we equip S0(κ) with the structure of a matrix ordered space (see e.g. [31]), arising from its inclusion into the operator system S(κ); thus, it makes sense to talk about positive or completely positive maps on S0(κ). w∗ For functions ϕ, ψ : κ → C, we write ϕ ∼ ψ if {(x, y) : ϕ(x, y) 6= ψ(x, y)} is a null set. POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 12 Proposition 4.3. Let κ ⊆ X × X be generated by rectangles and let ϕ : κ → C be a measurable function. The following are equivalent: (i) ϕ is a Schur multiplier; (ii) there exists a Schur multiplier ψ : X × X → C such that ψκ ∼ ϕ; (iii) there exists a unique completely bounded map Φ0 : S0(κ) → S0(κ) such that Φ0(Tk) = Tϕk, for each k ∈ L2(κ); (iv) there exists a unique completely bounded weak* continuous map Φ : S(κ) → S(κ) such that Φ(Tk) = Tϕk, for each k ∈ L2(κ). Proof. (i)⇒(ii) Let S2(κ) = {Tk : k ∈ L2(κ)}. Since ϕ is a Schur multiplier, the map Φ2 : S2(κ) → S2(κ), given by Φ2(Tk) = Tϕk, extends to a bounded linear map Φ0 : S0(κ) → S0(κ). Moreover, since Φ2 is a D-bimodule map, Φ0 is such as well. By Smith's theorem [28, Theorem 2.1], Φ0 is completely bounded. By [20, Exercise 8.6 (ii)], there exists a completely bounded D-bimodule map Ψ0 : K(H) → B(H) such that Ψ0S0(κ) = Φ0. Let Ψ be the weak* continuous extension of Ψ0 whose existence is guaranteed by Lemma 2.2 (iii); we have that Ψ is a completely bounded weak* continuous D-bimodule map. By Theorem 4.1, there exists ψ ∈ S(X) such that Ψ = Sψ. For every k ∈ L2(κ) we have It follows that and hence ψκ ∼ ϕ. Tψk = Sψ(Tk) = Sϕ(Tk) = Tϕk. ψk ∼ ϕk, for each k ∈ L2(κ), (ii)⇒(iv) Take Φ = SψS(κ). The uniqueness of Φ follows from the fact that the Hilbert-Schmidt operators with integral kernels in L2(κ) are dense in S(κ). (iv)⇒(iii)⇒(i) are trivial. (cid:3) If ϕ : κ → C is a Schur multiplier, then we write Sϕ : S(κ) → S(κ) for the ϕ be the restriction of Sϕ to S0(κ). map Φ appearing in (iv) above, and let S0 Definition 4.4. Let κ ⊆ X × X be a positivity domain. A Schur multiplier ϕ : κ → C will be called partially positive if ϕα×α is a positive Schur multiplier whenever α ⊆ X is a measurable set with α × α ⊆ κ. We can characterise partial positivity in this context using rank one op- erators, extending [21, Lemma 4.2], as follows. 1 (κ)) ⊆ B(H)+. Proposition 4.5. Let κ be a positivity domain. A Schur multiplier ϕ : κ → C is partially positive if and only if Sϕ(S + Proof. Suppose that Sϕ(S + 1 (κ)) ⊆ B(H)+ and that α × α ⊆ κ for a measur- able set α ⊆ X. If T is a positive rank one operator supported by α × α then, by assumption, Sϕ(T ) ≥ 0. Since Sϕ is weak* continuous and the weak* closed span of the positive rank one operators supported by α × α equals B(P (α)H)+, we conclude that ϕα×α is a positive Schur multiplier. Conversely, suppose that ϕ is partially positive and that T ∈ S(κ) is a positive rank one operator, say T = η ⊗ η∗ for some η ∈ H. If supp η = α then, by Proposition 3.2, α × α ⊆ω κ. By deleting a null set from α, we POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 13 may in fact suppose that α × α ⊆ κ. The assumption now implies that Sϕ(T ) ≥ 0. (cid:3) We note that the main interest in Proposition 4.5 is when κ is generated by squares; however, we formulate it in its present generality in view of Theorem 4.8. Definition 4.6. Let κ ⊆ X × X be a positivity domain and ϕ : κ → C be a Schur multiplier. We say that a measurable function ψ : X × X → C is a positive extension of ϕ if ψ is a positive Schur multiplier and ψκ ∼ ϕ. Remarks (i) If a Schur multiplier ϕ : κ → C has a positive extension then ϕ is necessarily partially positive. (ii) Recall [9] that a function ϕ : X × X → C is called ω-continuous if ϕ−1(U ) is an ω-open set for every open subset U ⊆ C. We note that if ϕ : κ → C has a positive extension, then ϕ has an ω-continuous positive extension. This follows from the fact that if ψ : X × X → C is a positive Schur multiplier then there exists an ω-continuous positive Schur multiplier ψ′ : X × X → C such that ψ and ψ′ differ on a set of product measure zero (see [29, Corollary 4.5]). Theorem 4.7. Let κ be a positivity domain. The following are equivalent, for a partially positive Schur multiplier ϕ : κ → C: (i) ϕ has a positive extension; (ii) the map Sϕ : S(κ) → S(κ) is positive; (iii) the map Sϕ : S(κ) → S(κ) is completely positive; (iv) the map S0 (v) the map S0 ϕ : S0(κ) → S0(κ) is positive; ϕ : S0(κ) → S0(κ) is completely positive. Proof. (i)⇒(ii) If ψ is a positive extension of ϕ then Sψ is positive and hence so is its restriction to S(κ); it is easily seen that this restriction coincides with Sϕ. (ii)⇒(iii) and (iv)⇒(v) follow from the operator system version of R. R. Smith's theorem [28] on automatic complete boundedness (see [29, Lemma 4.3]). (iii)⇒(iv) is trivial. (v)⇒(i) Let Φ = S0 ϕ; thus, Φ is a completely positive linear map on S0(κ). By [24, Theorem 3.16 and Lemma 3.12], Φ can be extended to a completely positive map Φ1 on the operator system S0(κ)+CI. By Arveson's Extension Theorem, there exists a completely positive map Ψ1 : K(H) + CI → B(H) extending Φ1. The restriction Ψ of Ψ1 to K(H) is then a completely positive extension of Φ. Let Ψ : B(H) → B(H) be the completely positive weak* continuous extension of Ψ whose existence is guaranteed by Lemma 2.2 (iii). Let Φ be the restriction of Ψ to S(κ); thus, Φ is a weak* continuous extension of Φ, and since Φ is a D-bimodule map, the same holds true for Φ. We now have that Ψ is a completely positive extension of Φ; by [20, Exercise 7.4], Ψ is a D-bimodule map. By Theorem 4.1, there exists ψ ∈ S(X) such that Ψ = Sψ; the function ψ is the desired positive extension of ϕ. (cid:3) The next theorem is a measurable version of one of the main results of [21] concerning positive completions of partially positive matrices. Recall POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 14 that the projective tensor product T (X) = L2(X) ⊗L2(X) can be canonically identified with the predual of B(H). Indeed, each element 2 < ∞ i=1 kfik2 h ∈ T (X) can be written as a series h =P∞ and P∞ i=1 kgik2 ∞ i=1 fi ⊗gi, whereP∞ 2 < ∞, and the duality pairing is then given by hT, hi = (T fi, gi), T ∈ B(H). Xi=1 defined up to a marginally null set, and given by h(x, y) =P∞ positive functionals on B(H), that is, functions h of the form h =P∞ fi, whereP∞ It follows [1] that h can be identified with a complex function on X × X, i=1 fi(x)gi(y). The positive cone T (X)+ consists of all functions h ∈ T (X) that define i=1 fi ⊗ 2 < ∞. It is well-known that a function ϕ ∈ L∞(X × X) is a Schur multiplier if and only if ϕh is equivalent (with respect to the product measure) to a function contained in T (X) for every h ∈ T (X) (see [22]). i=1 kfik2 Theorem 4.8. Let κ be a positivity domain. The following are equivalent: (i) every partially positive Schur multiplier ϕ : κ → C has a positive extension; (ii) S(κ)+ = [S + (iii) S0(κ)+ = [S + 1 (κ)] 1 (κ)] w∗ ; k·k . Proof. Assume (i) holds. We establish (ii) and (iii) simultaneously. It is clear that [S + ⊆ S0(κ)+. To show that the two cones are equal, it suffices by the Hahn-Banach Theorem to prove that if h ∈ T (X) is such that hA, hi ≥ 0 for all A ∈ [S + 1 (κ)], then hA, hi ≥ 0 for all A ∈ S(κ)+. 1 (κ)] k·k Thus, suppose that Then hA, hi ≥ 0, A ∈ [S + 1 (κ)]. hξ ⊗ ξ∗, hi ≥ 0, whenever ξ ∈ L2(X) and ξ ⊗ ξ∗ ∈ S1(κ), that is (see Proposition 3.2), hξ ⊗ ξ∗, hi ≥ 0, whenever (supp ξ) × (supp ξ) ⊆ω κ. Suppose first that the measure µ is finite and h ∈ S(X); recall that Sh denotes the corresponding bounded weak* continuous map on B(H), defined in (2). For every η ∈ L2(X), we have (Sh(ξ ⊗ξ∗)η, η) =ZX×X h(x, y)ξ(y)η(y)ξ(x)η(x)dxdy = h(ξη)⊗(ξη)∗, hi ≥ 0 whenever (supp ξ) × (supp ξ) ⊆ω κ. It follows that Sh(ξ ⊗ ξ∗) ≥ 0 for every rank one operator ξ ⊗ ξ∗ that belongs to S(κ)+; thus, hκ is a partially positive Schur multiplier. By assumption, hκ has an extension to a positive Schur multiplier h on X × X. We thus have (3) (Sh(A)χX , χX) ≥ 0, for all A ∈ B(H)+ POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 15 (note that χX ∈ L2(X) since µ is finite). We claim that if A ∈ S(κ) then (4) (Sh(A)χX , χX) = hA, hi. Indeed, suppose that k ∈ L2(κ). Then (Sh(Tk)χX, χX ) = ZX×X = Zκ h(x, y)k(y, x)dµ(x)dµ(y) h(x, y)k(y, x)dµ(x)dµ(y) = hTk, hi, so (4) holds if A is a Hilbert-Schmidt operator. Since Sh is weak* continuous and the Hilbert-Schmidt operators are dense in S(κ), (4) is established. It now follows that if A ∈ S(κ)+, then hA, hi ≥ 0. Next, relax the assumption that the measure µ be finite, but continue to suppose that h ∈ S(X). Let (Xn)n∈N be an increasing sequence of meaurable sets of finite measure such that X = ∪n∈NXn. By (4), (Sh(A)χXn, χXn) = hP (Xn)AP (Xn), hi ≥ 0 whenever A ∈ S(κ)+. Passing to the limit as n → ∞, we conclude that hA, hi ≥ 0 whenever A ∈ S(κ)+. Finally, relax the assumption that h be a Schur multiplier. Write h = 2 < ∞. For each N ∈ N, i=1 kfik2 i=1 kgik2 i=1 fi ⊗ gi where P∞ P∞ let 2 < ∞ and P∞ ∞ and XN =(x ∈ X : YN =(y ∈ X : fi(x)2 ≤ N) Xi=1 gi(y)2 ≤ N) . Xi=1 ∞ Then X \ (∪N ∈NXN ) and X \ (∪N ∈NYN ) have measure zero. Let ZN = XN ∩ YN , N ∈ N. Then X \ (∪N ∈NZN ) has measure zero and, by Theorem 4.1, the restriction of h to ZN × ZN is a Schur multiplier. By the previous paragraphs, hP (ZN )AP (ZN ), hi ≥ 0 whenever A ∈ S(κ)+, and passing to the limit as N → ∞ shows that hA, hi ≥ 0 for all A ∈ S(κ)+. (iii)⇒(i) and (ii)⇒(i) follow from Proposition 4.5 and Theorem 4.7. (cid:3) We note that, by Theorem 4.8, if a positivity domain admits positive extensions of partially positive Schur multipliers, then it necessarily satisfies the equivalent conditions of Theorem 3.6. 5. Extending positive definite functions In this section, we apply the results obtained previously to some questions arising in Abstract Harmonic Analysis. We assume, throughout the section, that G is a locally compact second countable amenable group, unless G is discrete, in which case no countability restriction is required. We denote by m the left Haar measure on G. We formulate sufficient conditions for unital symmetric subsets E ⊆ G which guarantee that every positive definite function defined on E can be extended to a positive definite function defined on the whole group G. POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 16 Let A(G) (resp. B(G)) be the Fourier (resp. the Fourier-Stieltjes) algebra of G (see [10] for the definition and basic properties of these algebras). Recall that T (G) = L2(G) ⊗L2(G) and let P : T (G) → A(G) be the contractive surjection given by P (ξ ⊗ η) = η ∗ ξ, ξ, η ∈ L2(G) (here ∗ denotes the usual convolution of functions and ξ(s) = ξ(s−1), s ∈ G). Let A(G)+ (resp. B(G)+) be the canonical positive cone of A(G) (resp. B(G)). Recall that a function f : G → C is called positive definite if for all finite subsets {s1, . . . , sn} ⊆ G, the matrix (f (sis−1 j ))i,j is positive; it is well-known that B(G)+ coincides with the cone of all continuous positive definite functions on G. More generally, if E ⊆ G is a symmetric set (that is, E−1 = E) containing the neutral element e of G, we say that a function f : E → C is positive definite if the matrix (f (sis−1 j ))i,j is positive for any s1, . . . , sn ∈ G with sis−1 j ∈ E for 1 ≤ i, j ≤ n. We say that f : G → C is positive definite on E if f E is positive definite. For a subset E ⊆ G, let E∗ = {(s, t) ∈ G × G : ts−1 ∈ E}, and for a function f : E → C, let N (f ) : E∗ → C be given by N (f )(s, t) = f (ts−1), (s, t) ∈ E∗. It is well-known that N maps B(G) isometrically onto the set Sinv(G) of all invariant Schur multipliers, that is, the Schur multipliers ϕ : G × G → C with the property that, for each r ∈ G, ϕ(sr, tr) = ϕ(s, t) for marginally almost every (s, t) ∈ G × G (see [5]), and that the image of B(G)+ under N consists of the positive elements of Sinv(G). Since the sets of the form E∗ play an important role in positive definite- ness, we first provide a description of when they are positivity domains (Theorem 5.2). We need an extra technical condition on the group G, known as Lebesgue density. If G is a locally compact group equipped with a metric, we say that t ∈ G is a point of density for a Borel set α ⊆ G m(α∩B(t,ǫ)) m(B(t,ǫ)) = 1, where B(t, ǫ) is the open ball with centre t and if limǫ→0 radius ǫ. We say that G satisfies the Lebesgue Density Theorem if there exists a left invariant metric on G which induces its topology, such that for every Borel set α, the subset α0 = {s ∈ α : s a point of density of α} has full measure in α. The classical Lebesgue Density Theorem states that the real line R satisfies this condition; it is also fulfilled for the groups Tn, Rm and their products. Lemma 5.1. Let G be a locally compact group that satisfies the Lebesgue Density Theorem, and let E ⊆ G be a Borel set. Suppose that α and β are non-null Borel subsets of G such that βα−1 ⊆ E. Let α0 (resp. β0) be the subset of α (resp. β) consisting of its points of density. Then β0α−1 0 ⊆ int(E). POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 17 Proof. For t ∈ G, we have (5) Let (6) P (χα ⊗ χβ)(t) =ZG χβ(s) χα(s−1t)ds = m(β ∩ tα). U := {t ∈ G : m(β ∩ tα) > 0}. Since P (χα ⊗ χβ) is a continuous function, (5) implies that U is an open set. Note that t ∈ βα−1 if and only if β ∩ tα 6= ∅. We show that (7) To this end, suppose that t ∈ β0α−1 0 , and write t = yx−1, where x ∈ α0 and y ∈ β0. Suppose that m(β ∩ tα) = 0; then m(y−1β ∩ x−1α) = 0. Thus, for every ǫ > 0, 0 ⊆ U. β0α−1 m((y−1β) ∩ B(e, ǫ)) + m((x−1α) ∩ B(e, ǫ)) ≤ m(B(e, ǫ)) and so 2 = lim ǫ→0 m((y−1β) ∩ B(e, ǫ)) m(B(e, ǫ)) + lim ǫ→0 m((x−1α) ∩ B(e, ǫ)) m(B(e, ǫ)) ≤ 1, a contradiction. Thus, (7) holds true. On the other hand, clearly U ⊆ E; since U is open, we have that β0α−1 (cid:3) 0 ⊆ int(E). Theorem 5.2. If G is a locally compact group satisfying the Lebesgue Den- sity Theorem and E ⊆ G is a Borel subset of positive measure, then (8) intω(E∗) ∼= int(E)∗ and clω(E∗) ∼= cl(E)∗. In particular, (i) E∗ is a positivity domain if and only if E is a symmetric set, e ∈ E, and E is the closure of its interior; (ii) E∗ is generated by squares if and only if E∗ contains a non-null square, if and only if E contains a symmetric open neighbourhood of e. Proof. If U ⊆ E is open then U ∗ is an open, and hence an ω-open, subset of E∗, and so int(E)∗ ⊆ω intω(E∗). Conversely, suppose that α × β is a rectangle, marginally contained in E∗. After deleting sets of measure zero from α and β, we may assume that α × β ⊆ E∗. Let α0 (resp. β0) be the set of all points of density in α (resp. β). By assumption, α \ α0 and β \ β0 have measure zero. By Lemma 5.1, (β0α−1 0 )∗ ⊆ int(E)∗, and since α0 × β0 ⊆ (β0α−1 0 )∗, we have that α × β ⊆ω int(E)∗. It follows that intω(E∗) ∼= int(E)∗. This easily implies that clω(E∗) ∼= cl(E)∗. Statement (i) is immediate from identities (8) and the definition of a pos- itivity domain. To prove (ii), assume first that E∗ is generated by squares. Since E∗ is non-null, E∗ contains a non-null square, say α × α. Let U be the open set defined in (6), where we have taken β = α. By the left invariance of the Haar measure, U is symmetric, and it clearly contains the neutral element of G. Conversely, suppose that U ⊆ E is symmetric, open and con- tains e. Since G is second countable, it admits a left invariant metric [16], say ρ. Let δ > 0 be such that B(e, δ) ⊆ U . Letting V = B(e, δ/2), we have that V × V ⊆ U ∗ (indeed, if s, t ∈ V then ρ(e, ts−1) ≤ ρ(e, t) + ρ(t, ts−1) = ρ(e, t) + ρ(e, s−1) < δ). POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 18 It follows that (V r) × (V r) ⊆ U ∗ for all r ∈ G and, letting R ⊆ G be a countable dense set, we have that ∆ ⊆ ∪r∈R(V r) × (V r). Thus, ∆ ⊆ω sqintω(E∗), and Theorem 3.6 implies that E∗ is generated by squares. (cid:3) We are now in a position to show that the operator systems of the form S(κ), where κ is a positivity domain, do not always contain positive rank one operators (see the paragraph after Definition 3.4). Corollary 5.3. There exists a positivity domain κ such that S(κ) does not contain a positive rank one operator. Proof. Realise the group of the circle T as the interval (−1, 1], equipped with addition modulo 2 and (normalised) Lebesgue measure; set H = L2(−1, 1). Let (tn)∞ n=1 ⊆ (0, 1) be a strictly decreasing sequence converging to zero and set ∞ E = {0} ∪ ([t2n, t2n−1] ∪ [−t2n−1, −t2n]). [n=1 It is clear that E does not contain a (symmetric) neighbourhood of 0. By Theorem 5.2, E∗ is a positivity domain and it does not contain a non-null square. By Proposition 3.2 (i), S(E∗) does not contain a positive rank one operator. (cid:3) We will need the following result by Lau [17]. It will be used in the proof of Proposition 5.5, and is the reason we require the assumption that G be amenable in most of the subsequent results. Lemma 5.4. Let G be a locally compact amenable group. Then there exists a net (ξi) ⊆ L2(G) of (positive) functions of norm one, such that the net (P (ξi ⊗ ξi)) is an approximate identity of A(G). If E ⊆ G is a set that is the closure of its interior, we set B(E) = {uE : u ∈ B(G)}. If E is in addition symmetric and contains the neutral element of G, we set B(E)+ = {uE : u ∈ B(G)+}. We next prove an invariant version of Proposition 4.3 and Theorem 4.7. Proposition 5.5. Let G be a locally compact amenable group satisfying the Lebesgue Density Theorem, E ⊆ G be a set that is the closure of its interior and u : E → C be a measurable function. The following hold: (i) N (u) is a Schur multiplier if and only if u is equivalent to a function that lies in B(E); (ii) If E is moreover symmetric and e ∈ E, then N (u) is a positive Schur multiplier if and only if u is equivalent to a function that lies in B(E)+. Proof. (i) Suppose that u is equivalent to a function u′ in B(E). We have that u′ is the restriction to E of a function v ∈ B(G); thus, N (u′) is the restriction to E∗ of the Schur multiplier N (v) on G × G. By (8) in Theorem 5.2, E∗ is generated by rectangles, and by Proposition 4.3, N (u′) is a Schur multiplier. Since N (u) ∼ N (u′), we conclude that N (u) is a Schur multiplier. Conversely, suppose that N (u) is a Schur multiplier on E∗. By Proposi- tion 4.3, there is a Schur multiplier ϕ : G × G → C whose restriction to E∗ POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 19 is equivalent to N (u). Let (ξi) ⊆ L2(G) be the net from Lemma 5.4, and ψi = ξi ⊗ ξi; thus, ψi ∈ T (G)+. We have that ϕψi ∈ T (G) for all i. Set vi = P (ϕψi); thus, (vi)i ⊆ A(G). Note that kvikA(G) = kP (ϕψi)kA(G) ≤ kϕψikT (G) ≤ kϕkS(G)kψikT (G) = kϕkS(G), for every i. Without loss of generality, we may assume that the net (vi) converges to a function v ∈ B(G) in the weak* topology of B(G) (here we view B(G) as the dual of the C-algebra C ∗(G) of G, see [10]). By [26, Lemma 2.3], for almost every t ∈ E, we have vi(t) =ZG ϕ(t−1s, s)ψi(t−1s, s)ds = u(t)ZG ψi(t−1s, s)ds → u(t), since (P (ξi ⊗ ξi)) converges to the constant function 1 uniformly on compact sets. Let U ⊆ E be open. Denoting by λ the left regular representation of L1(G) on L2(G), for every f ∈ Cc(G) with support contained in U , we have ZG vif dm = hλ(f ), vii −→ hλ(f ), vi =ZG vf dm. On the other hand, by the Lebesgue Dominated Convergence Theorem, ZG vif dm −→ZG uf dm and hence RG vf dm = RG uf dm. Since this holds for every f ∈ Cc(G) supported on U , we conclude that v = u almost everywhere on U , and since U was an arbitrary open subset of E, we have that v = u almost everywhere on E. It follows that u is equivalent to a function from B(E). (ii) By Theorem 5.2 (i), E∗ is a positivity domain. If u ∈ B(E)+ then u is the restriction to E of a function v ∈ B(G)+. Thus, N (v) is a positive Schur multiplier on G × G and N (u) is its restriction to E∗. It follows that N (u) is a positive Schur multiplier. Conversely, suppose that N (u) is a positive Schur multiplier. By Theorem 4.7, there exists ϕ ∈ S(G)+ whose restriction to E∗ is equivalent to N (u). But then (letting as above vi = P (ϕψi)), we have that vi ∈ A(G)+ for all i and hence the weak cluster point v of the net (vi) is a positive definite function. As in (i), we conclude that u is equivalent to the restriction of v to E. (cid:3) The next result gives a sufficient condition for the existence of positive definite extensions in operator-theoretic terms and is one of the main results of this section. Theorem 5.6. Let G be a locally compact amenable group satisfying the Lebesgue Density Theorem. Let E ⊆ G be a symmetric set which is the closure of its interior and contains the neutral element of G. Suppose that every positive operator in S(E∗) is a weak* limit of sums of rank one positive operators in S(E∗). If u ∈ B(G) is positive definite on E then there exists a positive definite function v ∈ B(G) such that vE = uE. Proof. Let u0 : E → C be the restriction of u and ϕ = N (u0). By Propo- sition 5.5, ϕ is a Schur multiplier. We claim that ϕ is partially positive. POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 20 Indeed, suppose first that α ⊆ G is a compact subset such that α × α ⊆ E∗. Since ϕα×α = N (u0)α×α, we have that ϕα×α is a continuous positive def- inite function. It follows that ϕα×α is a positive Schur multiplier (see the discussion preceding [29, Theorem 4.8] and the proof of that theorem). If α ⊆ G is a measurable set of finite measure with α × α ⊆ E∗ then, by the regularity of the Haar measure, there exists an increasing sequence (αk)k∈N of compact subsets of α such that ∪k∈Nαk has full measure in α. We have that Sϕ(T ) = sot-limk→∞ Sϕ(P (αk)T P (αk)) for every T ∈ B(P (α)H); it thus follows that ϕα×α is a positive Schur multiplier. Finally, if α ⊆ G is an arbitrary measurable set such that α × α ⊆ E∗ then by the σ-finiteness of the Haar measure, α is the union of an increasing sequence of sets of finite measure and it follows as before that ϕα×α is a positive Schur multiplier. We have thus shown that ϕ is a partially positive Schur multiplier. By Theorem 5.2, E∗ is a positivity domain and, by Theorem 4.8, there exists a positive Schur multiplier ψ : G × G → C extending ϕ. Now Proposition 5.5 implies the existence of a positive definite function v ∈ B(G) such that vE = u0. The proof is complete. (cid:3) We note that the function u in the statement of Theorem 5.6 does not need to be defined on the whole of G but just on the subset E, and it suffices that it possess an extension to a function from B(G). We now turn our attention to discrete groups; the assumption on the second countability will be dropped for the rest of the section. Definition 5.7. Let G be a discrete group. A subset E ⊆ G will be called a chordal subset of G if E is a symmetric set containing the neutral el- if n ≥ 4 and s1, . . . , sn ∈ E with ement e with the following property: sn · · · s2s1 = e, then there exist i, k ∈ {1, . . . , n} with 2 ≤ k − i ≤ n − 2 so that sk−1sk−2 · · · si ∈ E. Lemma 5.8. Let G be a discrete group and E ⊆ G be a symmetric subset containing the neutral element. The following are equivalent: (i) E is a chordal subset of G; (ii) the set E∗ is chordal. Proof. (i)⇒(ii) Suppose that n ≥ 4 and consider a cycle (x1, x2), · · · , (xn−1, xn), (xn, x1) ∈ E∗. Setting s1 = x2x−1 1 , · · · , sn−1 = xnx−1 n−1 and sn = x1x−1 n , we see that s1, . . . , sn ∈ E and sn · · · s2s1 = e. By the chordality of E, there exists i, k ∈ {1, . . . , n} with 2 ≤ k − i ≤ n − 2 so that xkx−1 i = sk−1 · · · si+1si ∈ E, so (xi, xk) ∈ E∗ is a chord in the given cycle. (ii)⇒(i) Suppose that n ≥ 4 and s1, . . . , sn ∈ E with sn · · · s2s1 = e. The edges (e, s1), (s1, s2s1), (s2s1, s3s2s1), · · · , (sn−1 · · · s1, e) form a cycle of E∗, which we label as (x1, x2), · · · (xn−1, xn), (xn, x1). Since E∗ is chordal, there exist 1 ≤ i, k ≤ n with 2 ≤ k − i ≤ n − 2 so that (xi, xk) ∈ E∗, so xkx−1 i = sk−1 · · · si+1si ∈ E. Hence E is a chordal subset of G. (cid:3) POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 21 As an application of our results on positive extensions of Schur multipliers, we now provide a different approach to the main result of [4]. Theorem 5.9. If E is a chordal subset of a discrete amenable group G, then every positive definite function u : E → C has a positive definite extension to a function v : G → C. Proof. Let u : E → C be positive definite and ϕ(s, t) = N (u)(s, t) = u(ts−1), (s, t) ∈ E∗. We note that ϕ is a Schur multiplier. Indeed, for every finite set F ⊆ G, let ϕF be the restriction of ϕ to E∗ ∩ (F × F ). Then the map SϕF acts on S(E∗ ∩ (F × F )). Since E is chordal, ϕF has a positive extension to a Schur multiplier ψF on F × F , and hence kSϕF k ≤ kSψF k = kSψF (I)k = max s∈F ϕF (s, s) = u(e). It follows that the Schur product by ϕ is a well-defined map from S(E∗) into B(H), and hence ϕ is a Schur multiplier. Note that, if α ⊆ G then α × α ⊆ E∗ if and only if αα−1 ⊆ E. The positive definiteness of u now implies that ϕ is a partially positive Schur multiplier. By Lemma 5.8 and Theorem 2.5, ϕ has an extension to a positive Schur multiplier ψ : G×G → C. By Proposition 5.5, there exists a positive definite extension v of u. (cid:3) As an illustration of Theorem 5.9, note that every symmetric arithmetic progression E ⊆ Z containing 0 is a chordal set; thus, Theorem 5.9 applies and gives the well-known fact that every positive definite function u : E → C has a positive definite extension to a function v : Z → C (see [11, Theorem 4.8]). Acknowledgements We wish to thank Yemon Choi for bringing the ref- erence [4] to our attention. We also thank the anonymous referee for the careful reading of the manuscript and his/her useful suggestions. References [1] W. B. Arveson, Operator algebras and invariant subspaces, Ann. Math. (2) 100 (1974), 433-532. [2] M. Bakonyi and G. Naevdal, On the matrix completion method for multidimen- sional moment problems, Acta Sci. Math. (Szeged) 64 (1998), 547-558. [3] M. Bakonyi and G. Naevdal, The finite subsets of Z2 having the extension prop- erty, J. London Math. Soc. (2) 62 (2000), 904-916. [4] M. Bakonyi and D. Timotin, Extensions of positive definite functions on amenable groups, Canad. Math. Bull. 54 (2011), no. 1, 3-11. [5] M. Bozejko and G. Fendler, Herz-Schur multipliers and completely bounded mul- tipliers of the Fourier algebra of a locally compact group, Boll. Un. Mat. Ital. A (6) 2 (1984), no. 2, 297-302. [6] M. Brannan and B. Forrest, Extending multipliers of the Fourier algebra from a subgroup, Proc. Amer. Math. Soc. 142 (2014), no. 4, 1181-1191. [7] A. Devinatz, On the extensions of positive definite functions, Acta Math. 102 (1959), 109-134. [8] H. Dym and I. Gohberg, Extensions of band matrices with band inverses, Linear Algebra Appl. 36 (1981), 1-24. [9] J. A. Erdos, A. Katavolos and V. S. Shulman, Rank one subspaces of bimodules over maximal abelian selfadjoint algebras, J. Funct. Anal. 157 (1998) no. 2, 554-587. POSITIVE EXTENSIONS OF SCHUR MULTIPLIERS 22 [10] P. Eymard, L'alg`ebre de Fourier d'un groupe localement compact, Bull. Soc. Math. France 92 (1964), 181-236. [11] J.-P. Gabardo, Trigonometric moment problems for arbitrary finite subsets of Zn, Trans. Amer. Math. Soc. 350 (1998), no. 11, 4473-4498. [12] R. Grone, C. R. Johnson, E. M. Sa, H. Wolkowicz, Positive definite completions of partial Hermitian matrices, Linear Algebra Appl. 58 (1984), 109-112. [13] U. Haagerup, Decomposition of completely bounded maps on operator algebras, un- published manuscript. [14] E. Kaniuth and A. T.-M. Lau, Extension and separation properties of positive definite functions on locally compact groups, Trans. Amer. Math. Soc. 359 (2007), no. 1, 447-463. [15] A. Katavolos and V. I. Paulsen, On the ranges of bimodule projections, Canad. Math. Bull. 48 (2005), no. 1, 97-111. [16] S. Kakutani, Uber die Metrisation der topologischen gruppen, Proc. Iup. Acad. 12 (1936), no. 4, 82-84. [17] A. T.-M. Lau, Uniformly continuous functionals on the Fourier algebra of any locally compact group, Trans. Amer. Math. Soc. 251 (1979), 39-59. [18] A. I. Loginov and V. S. Shulman, Hereditary and intermediate reflexivity of W*- algebras, Izv. Akad. Nauk SSSR 39 (1975), 1260-1273; Math. USSR-Izv. 9 (1975), 1189-1201. [19] J. R. McMullen, Extensions of positive-definite functions, Mem. Amer. Math. Soc. 117 (1972). [20] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge Univer- sity Press, 2002. [21] V. I. Paulsen, S. Power and R. R. Smith, Schur products and matrix completions, J. Funct. Anal. 85 (1989), 151-178. [22] V. V. Peller, Hankel operators in the theory of perturbations of unitary and selfad- joint operators, Funktsional. Anal. i Prilozhen. 19 (1985), no. 2, 37-51, 96. [23] G. Pisier, Similarity problems and completely bounded maps, Springer-Verlag, 2001. [24] T. Russell, Characterizations of ordered self-adjoint operaror spaces, preprint, arXiv: 1508.06272. [25] V. S. Shulman, I. G. Todorov and L. Turowska, Closable multipliers, Integral Equations Operator Theory 69 (2011), no. 1, 29-62. [26] V. S. Shulman, I. G. Todorov and L. Turowska, Sets of multiplicity and closable multipliers on group algebras, J. Funct. Anal. 268 (2015), no. 6, 1454-1508. [27] V. S. Shulman and L. Turowska, Operator synthesis I. Synthetic sets, bilattices and tensor algebras, J. Funct. Anal. 209 (2004), 293-331. [28] R. R. Smith, Completely bounded module maps and the Haagerup tensor product, J. Funct. Anal. 102 (1991), 156-175. [29] N. Steen, I. G. Todorov and L. Turowska, Local operator multipliers and posi- tivity, J. Funct. Anal. 267 (2014), 80-111. [30] D. Timotin, Completions of matrices and the commutant lifting theorem, J. Funct. Anal. 104 (1992), 291-298. [31] W. Werner, Subspaces of L(H) that are ∗-invariant, J. Funct. Anal. 193 (2002), 207-223. School of Mathematics and Statistics, University College Dublin, Belfield, Dublin 4, Ireland E-mail address: [email protected] Pure Mathematics Research Centre, Queen's University Belfast, Belfast BT7 1NN, United Kingdom E-mail address: [email protected] Pure Mathematics Research Centre, Queen's University Belfast, Belfast BT7 1NN, United Kingdom E-mail address: [email protected]
1705.06555	1	1705	2017-05-18T12:47:39	MF traces and the Cuntz semigroup	[ "math.OA" ]	A trace $\tau$ on a separable C-algebra $A$ is called matricial field (MF) if there is a trace-preserving morphism from $A$ to $Q_\omega$, where $Q_\omega$ denotes the norm ultrapower of the universal UHF-algebra $Q$. In general, the trace $\tau$ induces a state on the Cuntz semigroup $Cu(A)$. We show there is always a state-preserving morphism from $Cu(A)$ to $Cu(Q_\omega)$. As an application, if $A$ is an AI-algebra and $F$ is a free group acting on $A$, then every trace on the reduced crossed product $A \rtimes F$ is MF. This further implies the same result when $A$ is an AH-algebra with the ideal property such that $K_1(A)$ is a torsion group. We also use this to characterize when $A \rtimes F$ is MF (i.e. admits an isometric morphism into $Q_\omega$) for many simple, nuclear C-algebras $A$.	math.OA	math	MF TRACES AND THE CUNTZ SEMIGROUP CHRISTOPHER SCHAFHAUSER Abstract. A trace τ on a separable C∗-algebra A is called matricial field (MF) if there is a trace-preserving morphism A → Qω, where Qω denotes the norm ultrapower of the universal UHF-algebra Q. In general, the trace τ induces a state on the Cuntz semigroup Cu(A). We show there is always a state-preserving morphism Cu(A) → Cu(Qω). As an application, if A is an AI-algebra and F is a free group acting on A, then every trace on A ⋊λ F is MF. This further implies the same result when A is an AH- algebra with the ideal property such that K1(A) is a torsion group. We also use this to characterize when A ⋊λ F is MF (i.e. admits an isometric morphism into Qω) for many simple, nuclear C∗-algebras A. 1. Introduction A C∗-algebra A is called matricial field (MF) if there is a net of linear, -preserving maps ϕn : A → Mk(n) where kϕn(ab) − ϕn(a)ϕn(b)k → 0 and kϕn(a)k → kak for all a, b ∈ A. The MF-algebras were introduced by Blackadar and Kirchberg in [5] where it was shown such algebras admit a certain "generalized inductive limit" decomposition in terms of finite dimensional algebras. When A is further assumed to be simple and nuclear, a more refined version of this decomposition result is possible which has been crucial in Elliott's Classification Program. More precisely, a separable, simple, nuclear, MF C∗-algebras is the closed union of an increasing sequence of nuclear, residually finite dimensional C∗-subalgebras (see [6]). The MF property is still poorly understood and, despite having several examples of such algebras, there are very few techniques for verifying a given C∗-algebra if MF. The only obvious obstruction, and possible the only obstruction, is that if A is MF, then A ⊗ K does not contain an infinite projection, where A is the unitization of A. Recently, Tikuisis, White, and Winter have shown in [29] that every separable, nuclear C∗-algebra which satisfies the Universal Coefficient Theorem (UCT) and admits a faithful trace is MF (an alternate proof is given in [27]). In the non-nuclear setting, even less is known. A tracial state (hereafter referred to as a trace) on a C∗-algebra A is called matricial field if one can find linear, -preserving maps ϕn : A → Mk(n) which approximately preserve the multiplication and the trace in the same sense as above; note, however, the ϕn are not required to be approximately isometric. There are no traces which are known not to be MF, and, as with the isometric version, there are very few tools for verifying a Date: June 22, 2021. Key words and phrases. MF traces, Crossed Products, K-theoretic dynamics, Cuntz semigroup, AI- algebras. 1 2 CHRISTOPHER SCHAFHAUSER trace is MF aside from the Tikuisis-White-Winter Theorem mentioned above. Note that if every trace is MF, then Connes's Embedding Problem, which asks for the same type of approximations in the 2-norm, has a positive solution. In a sense, the MF-trace problem is a uniform version of Connes's Embedding Problem. For crossed products of C∗-algebras, slightly more is known, and, moreover, the existence of MF-approximations can often be detected in the dynamics. For example, a classical result of Pimsner in [23] characterizes when a crossed product C(X)⋊λ Z is MF in terms of the induced action of Z on X. Later, N. Brown has shown in [8] that if A is an AF-algebra, the MF-property for A ⋊λ Z is characterized by the induced action of Z on K0(A). Results of this nature for were obtained by Matui for actions of Z on simple, unital AT-algebras of real rank zero and by H. Lin in [21] for actions of Z on AH-algebras with a faithful invariant trace. After the integers, the next easiest class of groups has turned out to be non-abelian free groups. A deep result of Haagerup and Thorbjørnsen [18] shows that if F is a free group, then the group algebra C∗ λ(F ) is MF. Using this result, Kerr and Nowak have shown in [20] that if A is a commutative AF-algebra and F is a free group acting on A (including F = Z), the MF-property of A ⋊λ F is characterized in terms of the induced action of F on the Gelfand spectrum of A. Rainone has extended this result to possibly non-commutative AF-algebras A in [24] with the characterization given in terms of the induced action of F on K0(A). In Rainone's work in [24], a new technique was used. The MF approximations were first constructed at the level of K-theory and then lifted to the algebra using the classification of AF-algebras. This idea was later refined and formalized by Rainone and the author in [25]. The idea is essentially to show that if A ⋊λ F is stably finite, then there is a faithful approximate morphism K0(A ⋊λ F ) → K0(Mk(n)). From here, one uses the classification of AF-algebras to lift the compositions K0(A) → K0(A ⋊λ F ) → K0(Mk(n)) to approximate covariant representations (ϕn, un) : (A, F ) → Mk(n). These, in turn, induce approximate morphisms ϕn ⋊λ un : A ⋊λ F → Mk(n) ⊗ C∗ λ(F ) given by aλs 7→ ϕn(a)un(s) ⊗ λs. Composing these maps with MF-approximations for C∗ λ(F ) produces the MF-approximations of A ⋊λ F . The same technique works for traces on crossed products and with the help of deeper classification results, extends far beyond the class of AF-algebras. The following is the main result of [25]. Theorem 1.1 (Rainone, Schafhauser). Suppose A is an AH-algebra with real rank zero and F is a free group (possibly Z) acting on A. The following are equivalent: (1) A ⋊λ F is MF; (2) A ⋊λ F is stably finite; (3) spanZ{x − s · x : x ∈ K0(A), s ∈ F } ∩ K+ 0 (A) = {0}. Moreover, every trace on A ⋊λ F is MF. The real rank zero assumption in Theorem 1.1 is essential if one expect to relate the MF-property of the crossed product to the K0-group. For instance, if A = C[0, 1] and F = Z, then K0(A) = Z and hence there are no non-trivial actions of F on K0(A). On MF TRACES AND THE CUNTZ SEMIGROUP 3 the other hand, there are many actions of F on A for which the crossed product is not MF as can be seen from Pimsner's characterization of MF crossed products in [23]. The issue here is that A ⊗ K has no non-trivial projections and hence the K0-group contains very little information about the dynamics. To work around the real rank zero assumption, we apply the above argument with the Cuntz semigroup Cu(A) in place of K0(A). As the Cuntz semigroup is constructed using positive elements in A ⊗ K, the Cuntz semigroup contains rich information about the dynamics even in the absence of real rank zero. The following is the key technical result. Let Q = Nn Mn denote the universal UHF-algebra and let Qω denote the norm ultrapower of Q with respect to a free ultrafilter ω on the natural numbers. Theorem 1.2. Suppose A is a separable, unital C-algebra and τ is a trace on A. There is a unital Cu-morphism σ : Cu(A) → Cu(Qω) such that Cu(trQω ) ◦ σ = Cu(τ ). Recall an AI-algebra is a sequential direct limit of algebra C[0, 1] ⊗ An for finite dimen- sional C∗-algebras An. Following the same strategy as outlined before Theorem 1.1 above, the Ciuperca-Elliott classification of AI-algebras together with Theorem 1.2 produces the following result. Theorem 1.3. If A is an AI-algebra and F is a free group acting on A, then every trace on A ⋊λ F is MF. For actions of Z, the above theorem can be deduced from the Tikuisis-White-Winter Theorem. In fact, the result holds for crossed products of locally Type I algebras by free abelian groups (see Theorem 4.2). For arbitrary free groups F , the above theorem also holds assuming only that A ⊗ Q is an AI-algebra and hence holds whenever A is an AH-algebra with the ideal property and torsion K1-group (Corollary 4.5 below). Using a technical refinement of Theorem 1.3 (Theorem 4.3 below) together with recent progress in classification, it is in fact possible to characterize the MF property for crossed products of certain simple, nuclear C∗-algebras by free groups (Corollary 4.6 below). At this point, we are not able to characterize when the crossed products themselves are MF. The same techniques would work provided one could show that for every separable, stably finite C∗-algebra A, there is a faithful Cu-morphism Cu(A) → Cu(Qω). This is probably within reach, however we have been unable to resolve this question. In the case of minimal actions and, in particular, in the case when A is simple, more can be said since every invariant trace on A is faithful (see Corollary 4.4). In Section 2, we recall some terminology and results from [25] on MF approximations of dynamical systems and crossed products and also recall some definitions related to Cuntz semigroups. Theorem 1.2 is proven in Section 3 along with the analogous statement for abstract Cuntz semigroups. The final section, Section 4, contains our main result (Theorem 4.3) of which Theorem 1.3 is a special case. Some consequences of this result of a related nature are also discussed. 2.1. MF Dynamics. In this section, we recall some facts about MF algebras, MF traces, and the dynamical versions of these concepts. More details can be found in [25]. 2. Preliminaries 4 CHRISTOPHER SCHAFHAUSER Let A be a C∗-algebra. We say A is matricial field (MF) if for every finite set F ⊆ A and for every ε > 0, there is an integer k ≥ 1 and a linear, -preserving map ϕ : A → Mk such that kϕ(ab) − ϕ(a)ϕ(b)k < ε and kϕ(a)k ≥ kak − ε for all a ∈ F. Similarly, we say a trace τ on A is MF if for every finite set F ⊆ A and for every ε > 0, there is an integer k ≥ 1 and a linear, -preserving map ϕ : A → Mk such that kϕ(ab) − ϕ(a)ϕ(b)k < ε and trk(ϕ(a)) − τ (a) < ε for all a ∈ F. If A is a unital MF-algebra, then A admits an MF trace. Conversely, if A admits a faithful MF trace, then A is MF. n=1 Mn denote the universal UHF-algebra. For a free ultrafilter ω on the Let Q = N∞ natural numbers, define Qω = {(qn)∞ n=1 ⊆ Q : sup n kqnk < ∞}/{(qn)∞ n=1 ⊆ Q : lim n→ω kqnk = 0}. We will often identify an element q ∈ Qω with a bounded sequence (qn) ⊆ Q representing it. The trace on each Mn induces a trace trQ on Q which in turn induces a trace trω on Qω via trω(q) = limn→ω trQ(qn). A separable C∗-algebra A is MF if, and only if, there is a faithful -homomorphism ϕ : A → Qω. Similarly, a trace τ on a separable C∗-algebra A is MF if, and only if, there is a -homomorphism ϕ : A → Qω such that trω ◦ ϕ = τ . A dynamical system is a triple (A, G, α), where A is a C∗-algebra, G is a discrete group, and α : G → Aut(G) is a group homomorphism. A covariant representation of a dynamical system (ϕ, u) : (A, G, α) → B consists of a C∗-algebra B, a -homomorphism ϕ : A → B, and a group homomorphism u : G → U (B), where U (B) denotes the unitary group of B, such that usϕ(a)u∗ s = ϕ(αs(a)) for all a ∈ A. A dynamical system (A, G, α) is called MF if there is a covariant representation (ϕ, u) : (A, G, α) → Qω such that ϕ is faithful. A trace τ is called α-MF if there is a covariant representation (ϕ, u) : (A, G, α) → Qω such that trω ◦ ϕ = τ . Note that an α-MF trace is necessarily α-invariant in the sense that τ ◦ αs = τ for all s ∈ G. For a group G, we let trG denote the trace on the reduced group algebra C∗ λ(G) determined by trG(λs) = 0 for s 6= 1. For a C∗-dynamical system (A, G, α), we let E : A ⋊λ G → A denote the expectation on the reduced crossed product A ⋊λ G deter- mined by E(λs) = 0 for s 6= 1. Note that if τ is an invariant trace on A, then τ ◦ E is a trace on A ⋊λ G. The following result is Theorem 3.9 in [25]. Theorem 2.1 (Rainone, Schafhauser). Let (A, G, α) be a dynamical system such that C∗ λ(G) is exact. (1) If (A, G, α) is MF and C∗ (2) If τ is an α-MF trace on A and trG is MF, then E ◦ τ is MF. λ(G) is MF, then A ⋊λ G is MF. For our purposes, we will only be concerned with free groups. The following result is of fundamental importance in this paper. Theorem 2.2. If F is a free group, then C∗ λ(F ) is exact and MF and trF is MF. MF TRACES AND THE CUNTZ SEMIGROUP 5 Proof. If F = Z, the result follows since C∗ By Proposition 5.1.8 in [9], C∗ shows C∗ λ(F ) is MF. Hence C∗ Corollary VII.7.6 in [13]), we have that trF is MF. λ(Z) is commutative. Assume F is non-abelian. λ(Fr) is exact. A result of Haagerup and Thorbjørnsen [18] λ(F ) admits an MF trace. As C∗ λ(F ) has a unique trace (see (cid:3) 2.2. The Cuntz Semigroup. We recall the definition of and a few results about Cuntz semigroups which will be needed in Section 3. The modern definition of Cuntz semigroups (Definition 2.5 below) was introduced in [11]. The theory has been extensively developed in recent years. See, for example, [2] and [3]. Definition 2.3. A positively ordered monoid is a monoid M together with a partial order ≤ such that (1) if x ∈ M , then 0 ≤ x, and (2) if x, x′, y, y′ ∈ M , x ≤ x′, and y ≤ y′, then x + y ≤ x′ + y′. Definition 2.4. Let (M, ≤) be a partially ordered set. Given an x, y ∈ M , we say x is compactly contained in y, written x ≪ y, if whenever (yn) ⊆ M is an increasing sequence in M such that the supremum exists and y ≤ supn yn, there is an n ∈ Z+ such that x ≪ yn. A sequence (yn)n ⊆ M is called rapidly increasing if yn ≪ yn+1 for all n ∈ Z+. Definition 2.5. A Cuntz semigroup is a positively ordered monoid S such that (1) every increasing sequence in S has a supremum, (2) every element of S is the supremum of a rapidly increasing sequence, (3) if (xn) and (yn) are increasing sequences in S, then sup n (xn + yn) = (sup n xn) + (sup n yn), and (4) if x, x′, y, y′ ∈ S, x ≪ x′, and y ≪ y′, then x + y ≪ x′ + y′. A Cu-morphism σ : S → T between Cuntz semigroups S and T is a function preserving 0, +, ≤, ≪, and sup. We will need a few extra properties of Cuntz semigroups. First we need a few more definitions. Definition 2.6. Let S be a Cuntz semigroup. A basis for S is a set B ⊆ M such that for every x ∈ S, there is an increasing sequence in B with supremum x. We call S countably based if it admits a countable basis. Definition 2.7. Let S be a Cuntz semigroup. An element in x ∈ S is called compact if x ≪ x. We say S is algebraic if the submonoid Sc ⊆ S consisting of compact elements forms a basis for S. Note that if (xn) is an increasing sequence with compact supremum, then (xn) is even- tually constant. This shows that if B is a basis for S, then Sc ⊆ B and, in particular, if S is countably based, then Sc is countable. The following result follows from Proposition 5.5.4 in [2]. 6 CHRISTOPHER SCHAFHAUSER Theorem 2.8 (Antoine, Perera, Thiel). Suppose S and T are Cuntz semigroups and S is algebraic. Any ordered monoid morphism Sc → T extends uniquely to a Cu-morphism S → T . Definition 2.9. An order unit for a Cuntz semigroup S is a compact element e ∈ S such that for all x ∈ S, we have x ≤ sup{ne : n ≥ 1}. A pair (S, e) where S is a Cuntz semigroup and e is an order unit will be called a unital Cuntz semigroup. We often suppress the order unit from the notation and refer to S as a unital Cuntz semigroup. A unital Cu-morphism ϕ : (S, e) → (S′, e′) between unital Cuntz semigroups is a Cu- morphism ϕ : S → S′ such that ϕ(e) = e′. The definition of functional below is taken from [14] where the theory of functionals on Cuntz semigroups is extensively developed. We will not need anything more than the definition here. Definition 2.10. A functional of a Cuntz semigroup S is a function µ : S → [0, ∞] which preserves 0, +, ≤, and sup. If (S, e) is a unital Cuntz semigroup, a state of S is a functional µ on S such that µ(e) = 1. The most important example of Cuntz semigroups are defined as quotients of the posi- tive cone of a stable C∗-algebra. We recall the construction below. Example 2.11. Given a C∗-algebra A and positive elements a, b ∈ A, define a - b if there is a sequence (xn) ⊆ A such that x∗ nbxn → a and define a ∼ b if a - b and b - a. The Cuntz semigroup of A is defined as the set Cu(A) = (A ⊗ K)+/ ∼. Write hai for the class in Cu(A) represented by an element a ∈ (A ⊗ K)+. The partial order on Cu(A) is induced by the relation - and the addition is defined by orthogonal condition; that is, hai + hbi = ha′ + b′i where a′ and b′ are such that hai = ha′i, hbi = hb′i, and a′b′ = 0. The main result of [11] shows that Cu(A) is a Cuntz semigroup for all C∗-algebras A and that A 7→ Cu(A) defines a functor from C∗-algebras to Cuntz semigroups. For a - homomorphism ϕ : A → B, the map Cu(ϕ) : Cu(A) → Cu(B) is defined by Cu(ϕ)(hai) = h(ϕ ⊗ idK)(a)i for all positive a ∈ A ⊗ K. If A is separable, then Cu(A) is countably based. If A has real rank zero, then Cu(A) is algebraic. For a stably finite C∗-algebra A, there is a ordered monoid isomorphism V (A) → Sc, [p] 7→ hpi, where V (A) is the semigroup consisting of projections in A ⊗ K up to Murray-von Neumann equivalence. When A is unital, the element h1Ai is an order unit for Cu(A). We will always view Cu(A) as a unital Cu-semigroup with this order unit. Every trace τ on a unital C∗-algebra A induces a state Cu(τ ) on Cu(A) via the formula Cu(τ )(hai) = lim n→∞ (τ ⊗ Tr)(a1/n) for all positive a ∈ A ⊗ K, where Tr is the semifinite tracial weight on K normalized so that Tr(p) = 1 for any rank one projection p ∈ K. (cid:3) The connection between the Cuntz semigroup and Theorem 1.3 is given by the Ciuperca- Elliott Theorem in [10] which shows Cu classifies morphism from AI-algebras into stable rank one algebras. In fact, using this result, Robert has shown in [26] that the same result holds for sequential direct limits of Elliott-Thomsen algebras with trivial K1-group. The precise statement is given in 2.13 below. First we recall the definition of Elliott-Thomsen algebras. MF TRACES AND THE CUNTZ SEMIGROUP 7 Definition 2.12. Suppose A and B be C∗-algebras and ϕ0, ϕ1 : A → B are -homomor- phisms. Define M (A, B, ϕ0, ϕ1) = {(a, b) ∈ A ⊕ C([0, 1], B) : ϕ0(a) = b(0), ϕ1(a) = b(1)}. An Elliott-Thomsen algebra is a C∗-algebra of the form M (A, B, ϕ0, ϕ1) where A and B are finite dimensional. We let C denote the class of Elliott-Thomsen algebras and let C0 ⊆ C denote the subclass consisting of Elliott-Thomsen algebras with trivial K1-group. A C∗-algebra is an AC- algebra (resp. an AC0-algebra) if it is a sequential direct limit of algebras in C (resp. C0). Theorem 2.13 (Ciuperca, Elliott, Robert). Let A be a unital AC0-algebra and let B be a unital C-algebra with stable rank one. (1) If σ : Cu(A) → Cu(B) is a unital Cu-morphism, there is a unital -homomorphism ϕ : A → B such that Cu(ϕ) = σ. (2) If ϕ, ψ : A → B are unital -homomorphisms with Cu(ϕ) = Cu(ψ), then there is a sequence of unitaries (un) ⊆ B such that for all a ∈ A. kunϕ(a)u∗ n − ψ(a)k → 0 In our setting, we will have an action α of a group G on an AC0-algebra A. To pro- duce the dynamical MF-approximations needed in Theorem 2.1, we first produce a Cu- morphism σ : Cu(A) → Cu(Qω) such that σ ◦ Cu(αs) = σ for all s ∈ G. The existence statement above is used to lift σ to a -homomorphism ϕ : A → Qω and the uniqueness statement is used to compare the -homomorphisms ϕ and ϕ ◦ αs for s ∈ G. 3. MF Approximations on Cuntz Semigroups In this section, we will prove Theorem 1.2 in the introduction. We also prove the analogous statement for abstract Cuntz semigroups in Theorem 3.4 since it does not take much more work and may be of independent interest. For the applications to crossed products in Section 4, Theorem 1.2 will suffice. Both version of the theorem require handling the algebraic case first and then reducing the general case to the algebraic one. Lemma 3.1. If µ is a state on a countably based, algebraic, unital Cuntz semigroup S, then there is a unital Cu-morphism σ : S → Cu(Qω) such that Cu(trω) ◦ σ = µ. Proof. Since S is countably based, Sc is countable, and hence µ(Sc) ⊆ [0, ∞] is countable. Suppose x ∈ Sc. Since x ≤ sup{n · e : n ≥ 1} and x is compact, there is an n ≥ 1 such that x ≤ n · e and hence µ(x) ≤ n. Therefore, µ(Sc) ⊆ [0, ∞) and, in particular, µ(Sc) generates a countable subgroup of G ⊆ R. By Theorem 4.8 in [25], there is a positive, unital group morphism σ0 : G → K0(Qω) such that K0(trω)(σ0(g)) = g for all g ∈ G. Hence there is an ordered monoid morphism Sc → K+ 0 (Qω) which preserves the unit and the state. As Qω has stable rank one, the canonical map V (Qω) → K+ 0 (Qω) is an order isomor- phism and hence there is an order embedding K+ 0 (Qω) → Cu(A). Composing this map with σ0 yields a positive, unital monoid morphism σ : Sc → Qω. Since S is algebraic, 8 CHRISTOPHER SCHAFHAUSER there is a unique extension of σ to a Cu-morphism M → Qω by Theorem 2.8. Since σ0 is state-preserving and unital, so is σ. (cid:3) The previous lemma is enough to complete the proof of Theorem 1.2. Proof of Theorem 1.2. If πτ denotes the GNS-representation of τ , then πτ (A)′′ is a von Neumann algebra and hence has real rank zero. Since πτ (A) is separable, there is a separable C∗-algebra B with real rank zero such that πτ (A) ⊆ B ⊆ πτ (B)′′ (see Section II.8.5 of [4]). Then Cu(B) is algebraic and countably based so that there is a state- preserving, unital Cu-morphism Cu(B) → Cu(Qω). The composition Cu(A) Cu(πτ (A)) Cu(B) Cu(Qω). is the desired Cu-morphism. (cid:3) We now work towards proving a version of Theorem 1.2 for abstract Cuntz semigroups. The next two lemmas essentially provide an analogue of the argument above for abstract Cuntz semigroups. The role of the von Neumann algebra in the abstract setting is played by the Cuntz semigroup M1 introduced by Antoine, Perera, and Thiel in Example 4.14 of [3]. By Proposition 4.16 in [3], M1 = Cu(M ) for any II1-factor M . As II1-factors have real rank zero, M1 is algebraic. The trace on M induces a state on M1 denoted here by µ1. Although M1 is not countably based, we will see in Lemma 3.3, in the same spirit as the C∗-algebraic argument above, that any countably based Cuntz subsemigroup of M1 is contained in a countably based, algebraic Cuntz subsemigroup of M1. Lemma 3.2. If µ is a state on a countably based algebraic, unital Cuntz semigroup S, then there is a unit preserving Cu-morphism σ : S → M1 such that µ1 ◦ σ = µ. Proof. As in Example 4.14 of [3], define a relation ≺1 on [0, ∞] by x ≺1 y if, and only if, x ≤ y and x < ∞. Then ([0, ∞], ≺1) is a Q-semigroup in the sense of Definition 4.1 in [3]. The pairs (S, ≪) and (M1, ≪) are also Q-semigroups and every Q-morphism (S, ≪) → (M1, ≪) is a Cu-morphism. By the construction of M1 in Example 4.14 of [3], there is a Q-morphism σ0 : ([0, ∞], ≺1) → (M1, ≪) such that µ1(σ0(x)) = x for all x ∈ [0, ∞]. We claim µ : S → [0, ∞] is a Q-morphism. By definition, µ preserves zero, addition, order, and suprema of increasing sequences. It only remains to show µ preserves the auxiliary relation; that is, if x ≪ y in S, then µ(x) ≺1 µ(y) in [0, ∞]. To this end, suppose x ≪ y in S. Then x ≤ y and hence µ(x) ≤ µ(y). Also, since y ≤ sup{n · u : n ≥ 1} in S, there is an n ≥ 1 such that x ≤ n · u. Therefore, µ(x) ≤ n < ∞. This shows µ(x) ≺1 µ(y) and that µ is a Q-morphism. To complete the proof, define σ = σ0 ◦ µ : S → M0. (cid:3) Lemma 3.3. If S ⊆ T are Cuntz semigroups such that S is countably based and T is algebraic, then there is a countably based, algebraic Cuntz semigroup S′ with S ⊆ S′ ⊆ T . Proof. Let B denote a countable basis for S. For each x ∈ B, there is a sequence (yx,n)n of compact elements in T such that x = supn yx,n. Let S′ 0 ⊆ N denote the monoid generated by {yx,n : x ∈ B, n ≥ 1}. Let S′ denote the set of suprema of increasing sequences of 0 and endow S′ with the order structure inherited from T . Note that S′ elements in S′ 0 MF TRACES AND THE CUNTZ SEMIGROUP 9 is countable and each element of S′ countably based, algebraic Cuntz semigroup containing S. 0 is a compact element of T . We will show S′ is a If x, y ∈ S′, there are increasing sequences (xn) and (yn) in S′ 0 such that x = sup xn 0 and x + y = sup(xn + yn). Hence S′ is a submonoid and y = sup yn. Then xn + yn ∈ S′ of T . Now suppose (xn) is an increasing sequence in S′ and let x = sup xn, where the supremum is taken in T . We claim that x ∈ S′. For each n ≥ 1, there is an increasing sequence (yn,k)k in S′ 0 such that xn = supk yn,k. Since T is a Cuntz semigroup, there is a rapidly increasing sequence (x′ 2 ≪ x = supn xn, there is an integer n(1) ≥ 1 such that x′ 1 ≪ xn(1). Continuing in this fashion, we produce an increasing sequence (n(j))j of integers such that x′ 1 ≪ xn(1) = supk yn(1),k, there is an integer k(1) such that 1 ≤ yn(1),k(1). Also, x′ x′ 2 ≪ xn(2) = supk yn(2),k and yn(1),k(1) ≪ xn(1) ≤ xn(2) = supk yn(2),k (since yn(1),k(1) is compact). Hence there is an integer k(2) ≥ 1 such that x′ 2 ≤ yn(2),k(2) and yn(1),k(1) ≤ yn(2),k(2). Continue inductively and define zj = yn(j),k(j). Then zj is an increasing sequence in S′ j = sup xn(j) = x, it follows that supj zj = x and hence x ∈ S′. j) in T with supremum x. Since x′ 2 ≤ xn(1). Then x′ 1 ≪ x′ 2 ≤ xn(1) and, in particular, x′ 0 and x′ j ≤ zj ≤ xn(j) for every j ≥ 1. As sup x′ j ≪ xn(j). As x′ To show S′ is a Cuntz semigroup, it now suffices to show the relation ≪ computed with respect to S′ agrees with the relation ≪ computed with respect to T for all pairs of elements in S′. For the moment, let us denote compact containment relative to the order structure on S′ by ≪′. Then, with this notation, we mush show that for x, y ∈ S′, we have x ≪′ y if and only if x ≪ y. For x, y ∈ S′, it is clear that x ≪ y implies x ≪′ y. Suppose conversely, x, y ∈ S′ and x ≪′ y. Suppose zn is an increasing sequence in T such that y ≤ supn zn. Fix an increasing sequence yk in S′ 0 such that y = sup yk. As x ≪′ y, there is an integer k ≥ 1 such that x ≤ yk. As yk is compact, and yk ≤ y ≤ supn zn, there is an integer n ≥ 1 such that yk ≤ zn. Hence x ≤ zn and it follows that x ≪ y. This shows the relations ≪ and ≪′ agree on S′ as claimed. We have S′ is a Cuntz semigroup with the addition and order inherited from T and the 0 is countable and consists of compact elements 0, we have that S′ 0 contains a basis for S, it follows that S ⊆ S′, (cid:3) inclusion S′ ֒→ T is a Cu-embedding. As S′ and as every element in S′ is a supremum of an increasing sequence in S′ is countably based and algebraic. Since S′ and this completes the proof. Combining the previous three lemmas yields the following result which provides an abstract version of Theorem 1.2. Theorem 3.4. If S is a countably based, unital Cuntz semigroup and µ is a state on S, there is a unital Cu-morphism σ : S → Cu(Qω) such that Cu(trω) ◦ σ = µ. Proof. There is a state-preserving Cu-morphism S → M1 by Lemma 3.2 and hence after replacing S with the Cuntz semigroup generated by the image of S under this morphism, we may assume S ⊆ M1. As M1 is the Cuntz semigroup of a II1-factor and II1-factors have real rank zero, M1 is algebraic. By Lemma 3.3, there is a countably based, algebraic Cu-semigroup S′ ⊆ M1 containing S. Now Lemma 3.1 produces an state-preserving Cu- morphism S′ to Cu(Qω). Composing this morphism with the inclusion S ֒→ S′ yields the result. (cid:3) 10 CHRISTOPHER SCHAFHAUSER 4. Applications to Crossed Products We now work to prove Theorem 1.3 and some consequences of this result. The proofs in this section are nearly identical to those used in [25] in the real rank zero setting. The key difference is using the Ciuperca-Elliott-Robert classification (Theorem 2.13) in place of the classification of real rank zero AT-algebras and using the Cu-morphisms provided by Theorem 1.2 in place of the analogous result for K0-groups. Theorem 2.1 provides a method for showing traces on crossed products which factor through the expectation are MF, but to handle arbitrary traces, we use the following two results. The first is due to de la Harpe and Skandalis in [19] and shows for non-abelian free groups, all traces factor through the expectation. In the case of actions of the integers, the structure of traces is much more complicated, but in our setting, we can work around this using the Tikuisis-White-Winter Theorem (see [29] and [27]). Theorem 4.1 (de la Harpe, Skandalis). If A is a C-algebra and F is a non-abelian free group acting on A, then every trace on A ⋊λ F has the form τ ◦ E where τ is a F -invariant trace on A and E : A ⋊λ F → A is the canonical expectation. Recall that an algebra is locally Type I if the following approximation property holds: for every finite set F ⊆ A and for every ε > 0, there is a Type I subalgebra B ⊆ A such that for every a ∈ F, there is a b ∈ B such that ka − bk < ε. Note that as quotients of Type I algebras are Type I, we also have the quotients of locally Type I algebras are locally Type I. Theorem 4.2. Suppose A is a locally Type I algebra and G is a free abelian group. Then every trace on A ⋊λ G is quasidiagonal and, in particular, is MF. Proof. We may assume A is separable and that G is finitely generated; say G ∼= Zk. Let τ be a trace on A ⋊λ G and let J = {a ∈ A : τ (a∗a) = 0}. A standard application of the Cauchy-Schwarz inequality shows J is a two-sided ideal in A. Moreover, J is easily seen to be G-invariant. Now, J ⋊λ G is an ideal in A ⋊λ G and τ vanishes on J ⋊λ G. Hence τ induces a trace ¯τ on (A ⋊λ G)/(J ⋊λ G) ∼= (A/J) ⋊λ G and ¯τ A/J is faithful. To show τ is quasidiagonal, it suffices to show ¯τ is quasidiagonal. As A is locally Type I, so is A/J. By Theorem 1.1 in [12], A/J satisfies the UCT. Since the UCT is preserved by taking crossed products by Z, the algebra (A/J) ⋊λ G satisfies the UCT. The trace ¯τ A/J ◦ E is a faithful trace on (A/J) ⋊λ G and hence the Tikuisis-White-Winter Theorem implies every trace on (A/J) ⋊λ G is quasidiagonal (see Corollary 6.1 in [29]). (cid:3) The following contains Theorem 1.3 as a special case. Theorem 4.3. Suppose A is a C∗-algebra such that A ⊗ Q is a AC0-algebra. If F is a free group acting on A, then every trace on A ⋊λ F is MF. Proof. The case when F = Z follows from Theorem 4.2. When F is a non-abelian free group, every trace on A ⋊λ F has the form τ ◦ E for an invariant trace τ on A by Theorem 4.1. Hence it suffices to show every invariant trace on A is α-MF where α is the action of F on A. We may assume A is unital and F is freely generated by finitely many elements s1, . . . , sn. After replacing A with A ⊗ Q, we may assume A is an AC0-algebra. Let αj = αsj and MF TRACES AND THE CUNTZ SEMIGROUP 11 λj = λsj ∈ A ⋊λ F for j = 1, . . . , n. homomorphism ϕ : A → Qω and unitaries u1, . . . , un ∈ Qω such that ujϕ(a)u∗ It suffices to show there is a trace-preserving - j = ϕ(αj(a)). The trace τ induces a trace τ = τ ◦ E on A ⋊λ F where E : A ⋊λ F → A is the canonical conditional expectation. By Theorem 1.2, there is a state-preserving Cu-morphism σ : Cu(A ⋊λ F ) → Cu(Qω). Let ι : A → A⋊λ F denote the canonical inclusion and note that since ι is trace-preserving, the composition σ ◦ Cu(ι) is state-preserving. By the Ciuperca-Elliott-Robert Theorem (Theorem 2.13 above), there is a unital - homomorphism ϕ : A → Qω such that Cu(ϕ) = σ ◦ Cu(ι). For j = 1, . . . , n, we have Cu(ϕ ◦ αj) = σ ◦ Cu(ι ◦ αj) = σ ◦ Cu(ad(λj) ◦ ι) = σ ◦ Cu(ι) = Cu(ϕ). Applying the Ciuperca-Elliott-Robert Theorem again shows ϕ and ϕ◦αj are approximately unitarily equivalent and hence are unitarily equivalent by applying a reindexing argument in Qω. That is there are unitaries uj ∈ Qω such that ϕ ◦ αj = ad(uj) ◦ ϕ as desired. (cid:3) Corollary 4.4. Suppose A a unital C-algebra such that A ⊗ Q is an AC0-algebra and F is a free group acting minimally on A. The following are equivalent: (1) A ⋊λ F is MF; (2) A ⋊λ F is stably finite; (3) A admits an invariant trace. Proof. The implication (1) implies (2) is well known and follows immediately from the stable finiteness of Qω. Assuming (2) holds, A ⋊λ F admits a trace (Corollary 5.12 in [17]) which restricts to an invariant trace on A. If (3) holds, let τ be an invariant trace on A. By minimality, τ is faithful and hence induces a faithful trace on τ ◦ E on A ⋊λ F . The trace τ ◦ E is MF and hence A ⋊λ F is MF. (cid:3) Recall that a C∗-algebra A has the ideal property if every ideal of A is generated as an ideal by the projections it contains. Corollary 4.5. Suppose A is an AH-algebra with the ideal property such that K1(A) is a torsion group. If F is a free group acting on A, then every trace on A ⋊λ F is MF. If, moreover, A is unital and the action is minimal, then the following are equivalent: (1) A ⋊λ F is MF; (2) A ⋊λ F is stably finite; (3) A admits an invariant trace. Proof. As K∗(A ⊗ Q) ∼= K∗(A) ⊗ Q is torsion free, A ⊗ Q is an AT-algebra by the main result of [15]. As K1(A) is a torsion group, we have K1(A ⊗ Q) = 0. Theorem 2.5 in [28] implies A ⊗ Q is an AI-algebra and, in particular, is an AC0-algebra. The result follows from Theorem 4.3 and Corollary 4.4. (cid:3) Corollary 4.6. Suppose A is separable, simple, unital C-algebra satisfying the UCT such that A⊗Q has finite nuclear dimension and suppose F is a free group acting on A. Assume either that the projection in A separate traces on A or that K1(A) is a torsion group. The following are equivalent. (1) A ⋊λ F is MF; 12 CHRISTOPHER SCHAFHAUSER (2) A ⋊λ F is stably finite; (3) A admits an invariant trace. Moreover, every trace on A ⋊λ F is MF. Proof. If the projections in A separate traces on A, this Theorem 1.2 in [25]. If K1(A) is a torsion group, then A ⊗ Q is an AC0-algebra by Corollary 13.47 in [16] and Corollary D in [29]. (cid:3) Remark 4.7. The assumption that A satisfies the UCT and that A ⊗ Q has finite nuclear dimension in the above corollary may both be automatic when A is nuclear. For the finite nuclear dimension assumption, this is known to be the case whenever the set of extremal traces on A is weak-compact (see [7]). There is a good chance the conditions on the Elliott invariant in Corollary 4.6 are not necessary. We end this paper with a few remarks on potential further applications of these tech- niques. If one could produce a classification theorem for AT-algebras or AC-algebras (and morphisms between them), one could probably reduce to the setting of AI-algebras or AC0-algebras, respectively. An argument of this form was used in [25] as a way of reducing the MF-crossed product problem from real rank zero AT-algebras to AF-algebras. There is a good chance that if A is an AC0-algebra and F is a free group acting on A such that A⋊λ F is stably finite, then A⋊λ F is MF. The arguments would carry through in this setting if one could produce a faithful Cu-morphism Cu(A ⋊λ F ) → Cu(Qω) whenever the crossed product is stably finite. In particular, it would be enough to know the following. Conjecture 4.8. If A is a separable, stably finite C-algebra, then there is a faithful Cu-morphism Cu(A) → Cu(Qω). Note that if A is MF, then a faithful -homomorphism A → Qω induces a faithful Cu-morphism Cu(A) → Cu(Qω). Hence a negative answer to the above conjecture would provide a counter example to the Blackadar-Kirchberg problem. An interesting special case is A = C0(0, 1] ⊗ O2. The Cuntz semigroup Cu(A) con- sists of open subsets of (0, 1] with addition given by union and order given by inclusion. Voiculescu's homotopy invariance of quasidiagonality shows that A is MF and hence A satisfies the conjecture. It seems difficult to prove the existence of a faithful Cu-morphism using only Cuntz semigroup techniques, however. References [1] R. Antoine, J. Bosa, F. Perera, Completions of monoids with applications to the Cuntz semigroup, Inter. J. Math., 22 (2011), 837 -- 861. [2] R. Antoine, F. Perera, H. Thiel, Tensor products and regularity properties of Cuntz semigroups, preprint (2014). (arXiv:14110.0483 [math.OA]) [3] R. Antoine, F. Perera, H. Thiel R. Antoine, F. Perera, H. Thiel, An abstract bivariant Cuntz semigroup, preprint (2017). (arXiv:1702.01588 [math.OA]) [4] B. Blackadar, Operator Algbras: Theory of C∗-algebras and von Neumann algebras, Encyclopedia of Mathematical Sciences, Vol 122, Springer-Verlag, Berlin Heidelberg New York, 2006. [5] B. Blackadar, E. Kirchberg, Generalized inductive limits of finite dimensional C∗-algebras, Math. Ann., 307 (1997), 343 -- 380. [6] B. Blackadar, E. Kirchberg, Inner quasidiagonality and strong NF algebras, Pacific J. Math., 198 (2001), 307-329 MF TRACES AND THE CUNTZ SEMIGROUP 13 [7] J. Bosa, N. Brown, Y. Sato, A. Tikuisis, S. White, W. Winter, Covering dimension of C∗- algebras and 2-coloured classification, preprint (2015). (arXiv:1506.03974 [math.OA]) [8] N. Brown, AF embeddability of crossed products of AF algebras by the integers, J. Funct. Anal. 160 (1998), 150 -- 175. [9] N. Brown, N. Ozawa, C-Algebras and Finite-Dimensional Approximations, Graduate Studies in Mathematics, 88, Amer. Math. Soc., Providence, RI, 2008. [10] A. Ciuperca, G. Elliott, A remark on invariants for C∗-algebras of stable rank one, Int. Math. Res. Not. IMRN 2008, no. 5, Art. ID rnm 158, 33 pp. [11] K. Coward, G. Elliott, C. Ivanescu, The Cuntz semigroup as an invariant for C∗-algebras, J. Reine Angew. Math., 623 (2008), 161 -- 193. [12] M. Dadarlat, Some remarks on the Universal Coefficient Theorem in KK-theory, Operator algebras and mathematical physics (Constant¸a, 2001), 65-74, Theta, Bucharest, 2003. [13] K. Davidson, C*-Algebras by Example, Fields Institute Monographs, 6, Amer. Math. Soc., Provi- dence, RI, 1996. [14] G. Elliott, L. Robert, L. Santiago, The cone of lower semicontinuous traces on a C∗-algebra, American Journal of Mathematics, 133 (2011), 969 -- 1005. [15] G. Gong, C. Jiang, L. Li, C. Paniscu, AT-structure of AH-algebras with the ideal property and torsion free K-theory, J. Funct. Anal., 258 (2010), 2119 -- 2143. [16] G. Gong, H. Lin, Z. Niu, Classification of simple amenable Z-stable C∗-algebras, preprint (2015). (arXiv:1501.00135 [math.OA]). [17] U. Haagerup, Quasitraces on exact C∗-algebras are traces, C. R. Math. Acad. Sci. Soc. R. Can., 36 (2014), 67 -- 92. [18] U. Haagerup, S. Thorbjørnsen, A new application of random matrices: Ext(C ∗ red(F2)) is not a group, Ann. of Math., 162 (2005), 711 -- 775. [19] P. de la Harpe, G. Skandalis, Powers's property and simple C∗-algebras, Math. Ann., 273 (1985/86), 241 -- 250. [20] D. Kerr, P. Nowak, Residually finite actions and crossed products, Ergodic Theory Dynam. Systems, 32 (2012), 1585 -- 1614. [21] H. Lin Asymptotically unitary equivalence and asymptotically inner automorphisms, Amer. J. Math, 131 (2009), 1589 -- 1677. [22] Matui, H., AF embeddability of crossed products of AT algebras by the integers and its applications, J. Funct. Anal. 192 (2002), 562 -- 580. [23] M. Pimsner, Embedding some transformation group C∗-algebras into AF-algebras, Ergodic Theory Dynam. Systems, 3 (1983), 613 -- 626. [24] T. Rainone, MF actions and K-theoretic dynamics, J. Funct. Anal., 267 (2014), 542 -- 578. [25] T. Rainone, C. Schafhauser, Crossed products of nuclear C∗-algebras by free groups and their traces, preprint (2016). (arXiv:1601.06090v2 [math.OA]). [26] L. Robert, Classification of inductive limits of 1-dimensional NCCW complexes, Adv. Math., 231 (2012), 2802 -- 2836. [27] C. Schafhauser, A new proof of the Tikuisis-White-Winter Theorem, preprint (2017). (arxiv:1701.02180 [math.OA]) [28] K. Thomsen, Diagonalization in inductive limits of circle algebras, J. Opertor Theory, 27 (1992), 325 -- 340. [29] A. Tikuisis, S. White, W. Winter, Quasidiagonality of nuclear C∗-algebras, Ann. of Math., 185 (2017), 229 -- 284. Department of Pure Mathematics, University of Waterloo, 200 University Avenue West, Waterloo, ON, Canada, N2L 3G1 E-mail address: [email protected]
1212.3199	1	1212	2012-12-13T15:41:39	On K-theoretic invariants of semigroup C*-algebras attached to number fields	[ "math.OA", "math.NT" ]	We show that semigroup C-algebras attached to ax+b-semigroups over rings of integers determine number fields up to arithmetic equivalence, under the assumption that the number fields have the same number of roots of unity. For finite Galois extensions, this means that the semigroup C-algebras are isomorphic if and only if the number fields are isomorphic.	math.OA	math	ON K-THEORETIC INVARIANTS OF SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS XIN LI Abstract. We show that semigroup C-algebras attached to ax + b-semigroups over rings of integers determine number fields up to arithmetic equivalence, under the assumption that the number fields have the same number of roots of unity. For finite Galois extensions, this means that the semigroup C-algebras are isomorphic if and only if the number fields are isomorphic. 2 1 0 2 c e D 3 1 ] . A O h t a m [ 1 v 9 9 1 3 . 2 1 2 1 : v i X r a 1. Introduction Recently, there has been some progress in understanding the structure of semigroup C-algebras (see [Li1], [Li2], [Nor], [C-E-L1], [C-E-L2]). The construction of these semigroup C-algebras is easy to describe: Whenever P is a left cancellative semi- group, we can consider the action of P on itself by left multiplication, pass to the induced action of P via isometries on the Hilbert space ℓ2(P ) and finally form the C-algebra of bounded operators on ℓ2(P ) generated by all these isometries. This is the reduced semigroup C-algebra of P , denoted by C ∗ r (P ). A particular case of this construction, which served as one of the guiding examples for most of the recent work on semigroup C-algebras, has been the C-algebra of certain ax + b-semigroups: Given a number field K with ring of integers R, we take r (R⋊R×). the ax+b-semigroup R⋊R× over R and form the semigroup C-algebra C ∗ This C-algebra has been introduced and studied in [C-D-L] (the similar case of the particular semigroup N ⋊ N× was treated in [La-Rae] and [La-Nesh]), and it has been further investigated in [Ech-La] and [C-E-L1]. r (R ⋊ R×) as described Given the natural (and functorial) assignment K 7−→ C ∗ above, the following question immediately comes to mind: What information about r (R ⋊ R×)? The goal the number field K is encoded in the semigroup C-algebra C ∗ of the present paper is to address this question. Here are our main results: Let K and L be two number fields, i.e. finite field extensions of Q. Let R and S be the rings of integers of K and L, respectively. Theorem 1.1. Assume that K and L have the same number of roots of unity. If r (R ⋊ R×) ∼= C ∗ C ∗ r (S ⋊ S×), then K and L are arithmetically equivalent. 2010 Mathematics Subject Classification. Primary 46L05, 46L80; Secondary 11Rxx. Research supported by the ERC through AdG 267079. 1 2 XIN LI Arithmetic equivalence for number fields is defined in [Per, § 1]. If we restrict to Galois extensions, we obtain Theorem 1.2. Assume that K and L are finite Galois extensions of Q which have r (S ⋊ S×) if and the same number of roots of unity. Then we have C ∗ only if K ∼= L. r (R ⋊ R×) ∼= C ∗ The question whether it is possible to read off the number of roots of unity from the semigroup C-algebra is left open. However, we have the following partial answer: Theorem 1.3. Let K and L be finite Galois extensions of Q with rings of integers R and S, respectively. Assume that either both K and L have at least one real embedding, or that both K and L are purely imaginary. Then C ∗ r (S ⋊ S×) if and only if K ∼= L. r (R⋊ R×) ∼= C ∗ Alternatively, we could also formulate the following stronger version of Theorem 1.2: Theorem 1.4. Let K and L be finite Galois extensions of Q with rings of integers R and S, respectively. Assume that either both K and L have only the roots of unity +1 and −1, or that both K and L have more than two roots of unity. Then r (R ⋊ R×) ∼= C ∗ C ∗ r (S ⋊ S×) if and only if K ∼= L. Before we explain the strategy of the proofs, let us first put these results into context. First of all, we should remark that the analogous question for the Bost-Connes system from [Bo-Co], [Ha-Pa] and [L-L-N] has been studied in [Cor-Mar], where it is shown that the Bost-Connes system, together with the position of the so-called "dagger algebra", completely determines the number field (up to isomorphism). This is stronger than our result for the semigroup C-algebra. However, we should point out that the Bost-Connes system is a C-dynamical system, not only a C-algebra, and its construction involves not only the number field or its ring of integers as such, but class field theory. At the other extreme, it follows from [Cu-Li2], [Cu-Li3] and [Li-Lu] that if we take the ring C-algebra of the ring of integers (see [Cu-Li1], [Cu-Li2]) as such without additional structures, then we obtain very little information about the number field. Since the ring C-algebra of a ring of integers R is nothing else but the quotient r (R ⋊ R×) by its maximal primitive ideal, it is certainly not surprising that of C ∗ semigroup C-algebras contain more information about number fields than ring C- algebras. The striking observation is that the difference is huge. Our results could also be interesting from the point of view of classification. First of all, the semigroup C-algebras we examine here are strongly purely infinite, i.e. O∞-absorbing. This follows from [C-E-L1, Theorem 8.2.4] and [Pas-Rør, Propo- sition 2.14]. Therefore, by E. Kirchberg's result in [Kir], these C-algebras are classified by KK-theoretic invariants which keep track of the primitive ideal spaces. However, in general it is very complicated to go from these KK-theoretic results to K-theory, i.e, to prove a version of the UCT keeping track of primitive ideal spaces SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 3 (see [Me-Ne]). From this point of view, it is interesting that for our main results, all we need is ideal related K-theory (with the positions of the units). For the proof of our main results, the strategy is to extract information about the number field from the K-theory of certain primitive ideals and their quotients of the semigroup C-algebra. First, it turns out that given a number field K with ring of integers R, the minimal non-zero primitive ideals of the semigroup C-algebra r (R ⋊ R×) correspond one-to-one to the (non-zero) prime ideals of R. This is an C ∗ easy consequence of the results from [Ech-La]. Secondly, we consider the primitive ideal I{p} corresponding to a prime ideal p of R and form the quotient C ∗ r (R ⋊ R×)/I{p}. The K0-class of the unit of this quotient (in terms of modules this is the class of the trivial module) turns out to be a torsion element. Building on previous K-theoretic computations in [C-E-L1] and [C-E-L2], we then show how to read off the prime number p over which p lies (i.e. which satisfies pZ = p ∩ Z) from the torsion order of the K0-class of the unit. Once we know this, we can use the result from [dS-Per] that the notions of split and arithmetic equivalence coincide to deduce Theorem 1.1 and Theorem 1.2. To deduce the stronger versions Theorem 1.3 and Theorem 1.4, we study the maximal primitive ideal of our semigroup C-algebra C ∗ r (R ⋊ R×). The corresponding quotient is canonically isomorphic to the ring C- algebra A[R] from [Cu-Li1]. The canonical projection π : C ∗ r (R ⋊ R×) ։ A[R] induces a homomorphism in K0, say π∗. It turns out that for purely imaginary number fields, we can read off the number of roots of unity from the rank of the image of π∗. The paper is structured as follows: In a first preliminary section, we gather a few general facts about semigroup C-algebras and apply these to the particular case of r (R ⋊ R×), R being a ring of integers. In the next section, building on [Ech-La], C ∗ r (R ⋊ R×). Finally, we give a very concrete description of the primitive ideals of C ∗ we turn to K-theoretic invariants attached to the minimal non-zero primitive ideals. In Section 4, we prove Theorem 1.1 and Theorem 1.2, and in Section 5, we prove Theorem 1.3 and Theorem 1.4. The appendix consists of well-known results about restriction homomorphisms in K-theory which are useful for us. Acknowledgement: I am grateful to Gunther Cornelissen for helpful comments about arithmetic equivalence for number fields. 2. Preliminaries 2.1. Semigroup C-algebras. Let P be a left cancellative semigroup with identity element. We first of all recall the construction of the left regular C-algebra of P . Consider the Hilbert space ℓ2(P ) with its canonical orthonormal basis {εx: x ∈ P }, and define for every p ∈ P an isometry Vp by setting Vpεx = εpx. As in the group case, we simply take the C-algebra generated by the left regular representation of our semigroup: Definition 2.1. C ∗ r (P ) := C ∗ ({Vp: p ∈ P }) ⊆ L(ℓ2(P )). 4 XIN LI Moreover, it turns out that in the analysis of semigroup C-algebras as in [Li1], [Li2], [C-E-L1] and [C-E-L2], the following family of right ideals of P plays an important role. Definition 2.2. Let J :=(cid:8)q−1 1 p1 · · · q−1 n pnP : pi, qi ∈ P(cid:9) ∪ {∅}. The right ideals in J are called the constructible right ideals of P . A crucial condi- tion on J is given by the independence condition. Definition 2.3. We call J independent if for every X, X1, . . . , Xn ∈ J , the follow- i=1 Xi, then we must have X = Xi for some 1 ≤ i ≤ n. ing holds: Whenever X =Sn In [Li2, § 4], the Toeplitz Now assume that P is a subsemigroup of a group G. condition for P ⊆ G was introduced. For our purposes, it suffices to note that given a left Ore semigroup P , the pair P ⊆ G = P −1P is Toeplitz (see [Li2, § 8.3]). Here G = P −1P is the group of left quotients of P . In general, given a pair P ⊆ G satisfying the Toeplitz condition, we can identify C ∗ r (P ) with a full corner in a reduced crossed product by G as follows (see [Li2, § 3]): First of all, let JP ⊆G be the smallest family of subsets of G which contains J and which is closed under left translations by group elements and finite intersections. Next, for every subset X of G, let EX be the orthogonal projection in L(ℓ2(G)) onto the subspace ℓ2(X) ⊆ ℓ2(G). Alternatively, we can think of EX as the multiplication operator attached to the characteristic function of X ⊆ G. In particular, we can form EX for all X ∈ JP ⊆G. We set DP ⊆G := C ∗({EX : X ∈ JP ⊆G}) ⊆ ℓ∞(G) ⊆ L(ℓ2(G)). Here we again think of ℓ∞(G) as acting on ℓ2(G) by multiplication operators. By construction, the C-algebra DP ⊆G is a G-invariant sub-C-algebra of ℓ∞(G), where G acts on ℓ∞(G) by left translations. By [Li2, Corollary 3.10], we can identify C ∗ r (P ) with the full corner EP (DP ⊆G ⋊r G)EP of DP ⊆G ⋊r G via r (P ) ∋ Vp 7→ EP UpEP ∈ EP (DP ⊆G ⋊r G)EP . C ∗ Here, for every g ∈ G, we let Ug be the canonical unitary in the multiplier algebra of DP ⊆G ⋊r G implementing the G-action. Moreover, if P ⊆ G is Toeplitz and if P has independent constructible right ideals, [Li2, Lemma 4.2] tells us that JP ⊆G is independent (in the sense of [Li2, Defini- tion 2.5]). Since DP ⊆G is a commutative C-algebra, a natural task is to describe its spectrum. This has been done in [Li2], in the following way: We start with Definition 2.4. Let Σ be the set of all non-empty JP ⊆G-valued filters, where a JP ⊆G-valued filter is a subset F of JP ⊆G satisfying • X1 ⊆ X2 ∈ JP ⊆G, X1 ∈ F ⇒ X2 ∈ F, SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 5 • X1, X2 ∈ F ⇒ X1 ∩ X2 ∈ F, • ∅ /∈ F. By [Li2, Corollary 2.9], we can identify Spec (DP ⊆G) and Σ via ω : Spec (DP ⊆G) → Σ, χ 7→ {X ∈ JP ⊆G: χ(EX ) = 1} . For this, we used that JP ⊆G is independent. Moreover, as explained in [Li2] after Corollary 2.9, the topology of pointwise conver- gence on Spec (DP ⊆G) corresponds under ω to the topology on Σ whose basic open sets are given by U (X; X1, . . . , Xn) := {F ∈ Σ: X ∈ F, Xi /∈ F for all 1 ≤ i ≤ n} for all X, X1, . . . , Xn in JP ⊆G. 2.2. The case of ax + b-semigroups over rings of integers. Let K be a number field, i.e. a finite extension of Q. Let R be the ring of integers in K. Moreover, write K × for the multiplicative group K \ {0} and R× for the multiplicative semigroup R\{0}. We now form the ax+b-semigroup R⋊R×. We view it as a subsemigroup of the ax + b-group K ⋊ K ×. The main object of interest in this paper is the semigroup C-algebra C ∗ r (R ⋊ R×). We now set out for an explicit description of the dynamical system G y Spec (DP ⊆G) for our pair P = R ⋊ R× ⊆ K ⋊ K × = G. This leads to an explicit identification of C ∗ r (R ⋊ R×) as a full corner in a crossed product by K ⋊ K ×. Such a description has already been obtained in [C-D-L, § 5]. We present a slightly different approach In this way, we set the stage for an explicit which is more in the spirit of [Li2]. r (R ⋊ R×). This will be the topic of the next description of the primitive ideals of C ∗ section. We start with the family of constructible right ideals of R ⋊ R×. It is given by J =(cid:8)(b + a) × a×: b ∈ R, (0) 6= a ⊳ R(cid:9) ∪ {∅} , where a× = a \ {0}. This is explained in [Li1, § 2.4]. Moreover, J is independent by [Li1, Lemma 2.30]. Another useful observation is that R ⋊ R× is left Ore, and that K ⋊ K × is the corresponding group of left quotients. Therefore, our pair R ⋊ R× ⊆ K ⋊ K × is Toeplitz (see [Li2, § 8.3]). The corresponding family JP ⊆G is given by JP ⊆G =(cid:8)(b + a) × a×: b ∈ K, a a factional ideal of K(cid:9) ∪ {∅} . To describe Spec (DP ⊆G) for our pair R ⋊ R× ⊆ K ⋊ K ×, consider the finite adele space Af of K. The profinite completion R of R sits as a subring in Af . For every a, b in Af , we can form the coset b + aR ⊆ Af . Definition 2.5. We set C =(cid:8)b + aR: b, a ∈ Af(cid:9). 6 XIN LI For X = (b + a) × a× ∈ JP ⊆G, we let X(R) be b + a · R, and we set ∅(R) = ∅. Here we embed K into Af diagonally. For every prime ideal p 6= (0) of R (in the sequel, by a prime ideal we always mean a non-zero prime ideal of R), we can form the discrete valuation ring Rp and the corresponding quotient field Kp with normalized valuation vp : Kp → Z∪{∞}. Then Af =((xp)p ∈Yp Kp: vp(xp) ≥ 0 for almost all p) . Every fractional ideal a of K can be factorized in a unique way as a =Qp pvp(a) with vp(a) = 0 for almost all p. Lemma 2.6. For every X = (b + a) × a× in JP ⊆G, we have X(R) = b + a · R = {(xp)p ∈ Af : vp(−b + xp) ≥ vp(a) for all p} . Moreover, given two sets X1 and X2 in JP ⊆G, we have X1(R) ∩ X2(R) = (X1 ∩ X2)(R). Proof. The first assertion is easy to see. To prove the second claim, let Xi = (bi + ai) × a× for i = 1, 2. If X1 ∩ X2 6= ∅, then X1 ∩ X2 = (b + a) × a× with a = a1 ∩ a2 i and for some b ∈ (b1 + a1) ∩ (b2 + a2). Thus (b1 + a1 · R) ∩ (b2 + a2 · R) = (b + a1 · R) ∩ (b + a2 · R) = b + ((a1 · R) ∩ (a2 · R)) = b + (a1 ∩ a2) · R = (X1 ∩ X2)(R). If X1 ∩ X2 = ∅, then X1(R) ∩ X2(R) must be empty as well. If not, then there exists b ∈ (b1 + a1 · R) ∩ (b2 + a2 · R). Thus there is b = (bp)p ∈ Af with vp(−bi + bp) ≥ vp(ai) for all prime ideals p and i = 1, 2. But by strong approximation (see [Bour, Chapitre VII, § 2.4, Proposition 2]), there exists b ∈ K with vp(b − bp) ≥ max(vp(a1), vp(a2)) for all prime ideals p. It follows that vp(−bi + b) ≥ vp(ai) for all prime ideals p and i = 1, 2. Hence b lies in (b1 + a1) ∩ (b2 + a2). (cid:3) Lemma 2.7. We can identify Σ and C (see Definition 2.4 and 2.5) via the mutually inverse maps X(R) Σ → C, F 7→ \X∈F C → Σ, b + aR 7→(cid:8)X ∈ JP ⊆G: b + aR ⊆ X(R)(cid:9) . Proof. Let us first prove that these maps are well-defined. For every F ∈ Σ, the intersection TX∈F X(R) is non-empty. Namely, if we fix some Y ∈ F, then we have X(R) ∩ Y (R). \X∈F X(R) = \X∈F Now Y (R) is compact, and X(R) ∩ Y (R) are closed subsets of Y (R) satisfying the property that finite intersections are non-empty by the second part of Lemma 2.6 and the definition of a JP ⊆G-valued filter. Thus TX∈F X(R) ∩ Y (R) 6= ∅. Let us SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 7 now see why TX∈F X(R) is of the form b + aR. As TX∈F X(R) is non-empty, we can choose b ∈TX∈F X(R). For every X = (bX + aX ) × a× − b + X(R) = aX · R. X ∈ F, we have (1) Define np := supX∈F vp(aX ) ∈ Z ∪ {∞} for all every prime ideal p. We obviously have np ≥ 0 for almost all p. Thus there exists a = (ap)p ∈ Af with vp(ap) = np for all p, where we set vp(0) = ∞. This a satisfies TX∈F aX · R = aR. Putting all this together, we obtain (−b + X(R)) aX · R = b + aR. \X∈F X(R) = b + \X∈F (1) = b + \X∈F is well-defined. It remains to show that these maps are mutual inverses. So the first map Σ → C, F 7→TX∈F X(R) is well-defined. And using the second part of Lemma 2.6, it is easy to see that(cid:8)X ∈ JP ⊆G: b + aR ⊆ X(R)(cid:9) is a JP ⊆G-valued filter. Thus also the second map C → Σ, b + aR 7→(cid:8)X ∈ JP ⊆G: b + aR ⊆ X(R)(cid:9) Let us first prove F = (cid:8)Y ∈ JP ⊆G: TX∈F X(R) ⊆ Y (R)(cid:9). We obviously have take Y = (b + a) × a× ∈ JP ⊆G satisfying TX∈F X(R) ⊆ Y (R). Then for every prime ideal p, we must have vp(a) ≤ supX∈F vp(aX ). This implies that for every prime ideal p, there exists X ∈ F with vp(a) ≤ vp(aX ). As F is closed under finite intersections, this means that we can find X = (b + a) × a× ∈ F with vp(a) ≤ vp(a) "⊆". For the reverse inclusion, write every X ∈ F as X = (bX + aX) × a× X , and for all prime ideals p, i.e. a ⊇ a. As X and Y satisfy TX∈F X(R) ⊆ X(R) and TX∈F X(R) ⊆ Y (R), the intersection X ∩ Y cannot be empty by the second part of Lemma 2.6. As a ⊆ a, we conclude that X must be contained in Y , hence Y ∈ F. This proves "⊇". is of the form d + cR for some d, c in Af . It suffices to prove cR ⊆ aR since we Finally, let us proveT(cid:8)X ∈ JP ⊆G: b + aR ⊆ X(R)(cid:9) = b+aR. It is clear that "⊇" holds. To prove "⊆", we use our observation that T(cid:8)X ∈ JP ⊆G: b + aR ⊆ X(R)(cid:9) already know d + cR ⊇ b + aR. Let F =(cid:8)X ⊆ JP ⊆G: b + aR ⊆ X(R)(cid:9) and write every X ∈ F as X = (bX + aX ) × a× X . Then we know that vp(c) = supX∈F vp(aX ). We also know that vp(aX ) ≤ vp(a) for all prime ideals p as b + aR ⊆ X(R). Hence it follows that if vp(c) = ∞, then also vp(a) = ∞. If vp(a) is finite, then there exists X ∈ F with vp(aX ) = vp(a), and we deduce vp(c) ≥ vp(a). Thus cR ⊆ aR as desired. This proves "⊆". (cid:3) Let us denote the map Σ → C, F 7→ TX∈F X(R) by κ. def ⇔ "aR = cR further identify C with the quotient Af × Af / ∼ with (a, b) ∼ (c, d) and b − d ∈ aR = cR". Let ρ : C → Af × Af / ∼, b + aR 7→ [b, a] denote this identification. Here [b, a] is the equivalence class of (b, a) in Af × Af / ∼. Also, let π : Af × Af ։ Af × Af / ∼ be the canonical projection. It is clear that we can Lemma 2.8. The bijection ρ ◦ κ ◦ ω : Spec (DP ⊆G) → Af × Af / ∼ transports the topology of pointwise convergence on Spec (DP ⊆G) to the quotient topology of Af × Af / ∼ inherited from the natural topology of Af × Af . 8 XIN LI Proof. We have to prove that {(ρ ◦ κ)(U (X; X1, . . . , Xn)): X, X1, . . . , Xn ∈ JP ⊆G} is a basis of open sets for the quotient topology on Af × Af / ∼. We first compute κ(U (X; X1, . . . , Xn)) =(cid:8)b + aR: b + aR ⊆ X(R), b + aR * Xi(R) ∀ 1 ≤ i ≤ n(cid:9) . Write X(R) = b + a · R, Xi(R) = bi + ai · R for some b, bi in K and fractional ideals a, ai of K (1 ≤ i ≤ n). With this notation, ρ(κ(U (X; X1, . . . , Xn))) is the set of all [a, b] ∈ Af × Af / ∼ satisfying • vp(a) ≥ vp(a), vp(−b + b) ≥ vp(a) for all prime ideals p; • for all 1 ≤ i ≤ n, there exists a prime ideal p such that vp(a) < vp(ai) or vp(a) ≥ vp(ai) and vp(−bi + b) < vp(ai). From this description, we see that π−1(ρ(κ(U (X; X1, . . . , Xn)))) is clearly an open subset of Af × Af . Moreover, every open subset of Af × Af of the form π−1(U ) for some U ⊆ Af × Af / ∼ is a union of sets of the form  Yp /∈ F pnp × Yp∈ F (bp + pnp)  × Yp /∈F pnp × Yp∈F (pnp \ pmp)  for finite sets of prime ideals F and F , bp ∈ K and integers mp, np which are 0 for almost all prime ideals p. As every such subset equals π−1(ρ(κ(U (X; X1, . . . , Xn)))) for some X, X1, ..., Xn in JP ⊆G, we are done. (cid:3) Corollary 2.9. ρ ◦ κ ◦ ω : Spec (DP ⊆G) ∼= Af × Af / ∼ is a G-equivariant homeomor- phism, where G = K ⋊ K × acts on Af × Af / ∼ by left multiplication, (b, a) · [b, a] = [b + ab, aa]. Corollary 2.10. The transpose of ρ ◦ κ ◦ ω yields an identification C0(Af × Af / ∼ ) ⋊r (K ⋊ K ×) ∼= DP ⊆G ⋊r G. 3. Explicit description of primitive ideals For the sake of brevity, we write P = R ⋊ R× and G = K ⋊ K ×. The primitive ideal space of C ∗ r (P ) has been computed in [Ech-La]. Here is the result: For every a = (ap)p ∈ Af , set Z(a) = {p prime ideal: ap = 0}. Moreover, for a subset A of the set of prime ideals P, let VA be the vanishing ideal of C0(Af × Af / ∼ ) corresponding to the closed subset CA = {[b, a] ∈ Af × Af / ∼ : A ⊆ Z(a)}, i.e. VA = {f ∈ C0(Af × Af / ∼): f CA = 0}. Theorem 3.1 ([Ech-La]). The map 2P ∋ A 7→ hVAi ∈ Prim (C0(Af × Af / ∼) ⋊r (K ⋊ K ×)) is a bijection. Proof. This follows from [Ech-La, Corollary 2.7, Corollary 3.5] using the easy obser- vation that hVAi = VA ⋊r (K ⋊ K ×) = ind VA. (cid:3) SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 9 Remark 3.2. As it is shown in [Ech-La], the map Af × Af / ∼ −→ 2P , [b, a] 7→ Z(a) transports the quotient topology of the quasi-orbit space inherited from the topology of Af × Af / ∼ to the power-cofinite topology of 2P . Basic open sets in this topology 2P , the bijection of the last corollary becomes a homeomorphism. are of the form (cid:8)S ∈ 2P : S ∩ F = ∅(cid:9), for F ⊆ P finite. So with this topology on Corollary 3.3. The ideals (cid:10)V{p}(cid:11), p ∈ P, are the minimal non-zero primitive ideals of C0(Af × Af / ∼) ⋊r (K ⋊ K ×). Proof. The bijection 2P ∋ A 7→ hVAi ∈ Prim (C0(Af ×Af / ∼)⋊r (K ⋊K ×)) is clearly inclusion-preserving. (cid:3) Theorem 3.1 gives a description of the primitive ideal space of C0(Af × Af / ∼ ) ⋊r (K ⋊ K ×), hence also of DP ⊆G ⋊r G. In the sequel, we describe the minimal non-zero primitive ideals of DP ⊆G ⋊r G in an explicit way. Such a description could also be obtained for all the primitive ideals, but it is really the minimal ones we will be interested in later on. Af / ∼)⋊r(K⋊K ×) under the isomorphism C0(Af ×Af / ∼)⋊r(K⋊K ×) ∼= DP ⊆G⋊rG from Corollary 2.10. Let J{p} be the ideal of DP ⊆G ⋊r G which corresponds to the ideal(cid:10)V{p}(cid:11) of C0(Af × Proposition 3.4. J{p} =(cid:10)ER×R× − ER×p×(cid:11)DP ⊆G⋊rG. Proof. We first show "⊇". Take [b, a] ∈ C{p}, i.e. ap = 0 (a = (ap)p), and set χ = (ρ ◦ κ ◦ ω)−1[b, a]. If b + aR R, then for all x ∈ R, b + aR * x + p · R, hence χ(E(x+p)×p×) = 0. As R × p× = Sx∈R(x + p) × p×, we deduce χ(ER×R×) = χ(ER×p×) = 0. If b + aR ⊆ R, then since ap = 0, there exists a fractional ideal a such that vp(a) ≥ 1 and b + aR ⊆ b + a · R for some b ∈ K. Then b + aR ⊆ R ∩ (b + a · R) = x + (a ∩ R) · R for some x ∈ R. As vp(a) ≥ 1, we deduce a ∩ R ⊆ p and thus b + aR ⊆ x + p · R ⊆ R. This implies χ(ER×R×) = χ(E(x+p)×p× ) = 1. Since (cid:8)E(r+p)×p×: r ∈ R(cid:9) are pairwise orthogonal and ER×p× = Pr∈R E(r+p)×p×, we conclude that χ(ER×p×) = 1. Hence χ(ER×R× − ER×p× ) = 0. To prove the reverse inclusion, take [b, a] ∈ Af ×Af / ∼ and set χ = (ρ◦κ◦ω)−1[b, a]. We need to show that if χ((cid:10)ER×R× − ER×p×(cid:11)DP ⊆G⋊rG ∩ DP ⊆G) = 0, then ap = 0, where a = (ap)p. ap = 0 means that sup(cid:8)vp(a): There is (b + a) × a× ∈ JP ⊆G with χ(E(b+a)×a×) = 1(cid:9) = ∞. Assume that this supremum is finite. Then we can choose b ∈ K and a fractional ideal a of K such that χ(E(b+a)×a× ) = 1 and the supremum above agrees with vp(a). By strong approximation, we can find a ∈ K × such that vp(a) = vp(a) and a ⊆ aR. The latter condition implies χ(E(b+aR)×(aR)× ) = 1. Therefore, χ(Ad (U(b,a))(ER×R× )) = 1. As χ vanishes on (cid:10)ER×R× − ER×p×(cid:11)DP ⊆G⋊rG ∩ DP ⊆G, 10 it follows that XIN LI 0 = χ ◦ Ad (U(b,a))(ER×R× − ER×p× ) = 1 −Xx∈R χ ◦ Ad (U(b,a))(E(x+p)×p× ). Thus we must have 1 = χ ◦ Ad (U(b,a))(E(x+p)×p× ) = χ(E(b+ax+ap)×(ap)× ) for some x ∈ R. But vp(ap) = vp(a) + 1 = vp(a) + 1 > vp(a). This contradicts maximality of vp(a). (cid:3) In the sequel, we describe J{p} in a way which is more suitable for K-theoretic computations. Lemma 3.5. The smallest family J ({p}) of subsets of G = K ⋊ K × with the properties • R × (R \ p) ∈ J ({p}), • g ∈ G, X ∈ J ({p}) ⇒ g · X ∈ J ({p}), • X1, X2 ∈ J ({p}) ⇒ X1 ∩ X2 ∈ J ({p}), • X ∈ J ({p}), Y ∈ JP ⊆G ⇒ X ∩ Y ∈ J ({p}) is given by (2) {(b + a) × (a \ p · a): b ∈ K, a a fractional ideal of K} ∪ {∅} . · · · pmr r 1 1 · · · plr , b = pn·pn1 1 · · · pnr Proof. The first item is obviously fulfilled. And the set in (2) is certainly G-invariant, i.e. satisfies the second item. To prove the third property, take two fractional ideals a = pm·pm1 r with prime ideals pi. Then (a\p·a)∩(b\p·b) = ((pm\pm+1)∩(pn\pn+1))·pl1 r with li := max(mi, ni). As (pm\pm+1)∩(pn\pn+1) is either empty or pm \ pm+1 depending on whether m 6= n or m = n, we see that (a \ p · a) ∩ (b \ p · b) is either empty or pm · pl1 r = (a ∩ b) \ (p · (a ∩ b)). For the fourth property, let a and b be as above. Then (a \ p · a) ∩ b = ((pm \ pm+1) ∩ pn) · pl1 1 · · · plr r with li := max(mi, ni). Thus either (a\p·a)∩b = ∅ or (a\p·a)∩b = (pm \pm+1)·pl1 1 · · · plr 1 · · · plr r depending on whether m < n or m ≥ n. So far, we have proven that the set in (2) satisfies all the desired properties. It suffices to show that for every non-zero ideal a of R, we must have a × (a \ p · a) ∈ J ({p}). Write a = pv · a′ for some v ∈ N0 and some non-zero ideal a′ of R which is coprime to p. As every fractional ideal is of the form (aR) ∩ (cR) for some a, c in K ×, we can choose a, c ∈ K × with pv = (aR) ∩ (cR), and we can without loss of generality assume vp(a) = v. Then a(R \ p) ∩ cR = (aR ∩ cR) \ (ap ∩ cR) = pv \ pv+1. Therefore It remains to prove minimality. r \pm+1 ·pl1 r = pm·pl1 1 · · · plr r \ pm+1 · pl1 1 · · · plr 1 · · · plr a × (a \ p · a) = (a′ × (a′)×) ∩ (pv × (pv \ pv+1)) = (a′ × (a′)×) ∩ (aR × (a(R \ p)) ∩ (cR × (cR)×) ∈ J ({p}). (cid:3) For a fractional ideal a, set δa,p = Ea×a× −Ea×(p·a)× = Ea×(a\p·a). With this notation, we have J{p} = hδR,pi. SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 11 Corollary 3.6. The smallest G-invariant ideal of DP ⊆G containing δR,p is given by C ∗({EX: X ∈ J ({p})}). Corollary 3.7. We have hδR,piDP ⊆G⋊rG = C ∗({EX : X ∈ J ({p})}) ⋊r G. For our K-theoretic computations, we need Lemma 3.8. J ({p}) is independent. a× ∪ (p · ai)× ⊇ a× Proof. Assume that (b + a) × (a \ p · a) =Sn Sn that ai ⊆ a (∗). Hence, a× =Sn i=1(bi + ai) × (ai \ p · ai). Thus a \ p · a = i=1(ai \ p · ai). Moreover, we deduce that a× ⊇ a \ p · a ⊇ ai \ p · ai. This implies i . Using strong approximation as in [Li1, Lemma 2.30], we deduce i ∪ (p · a)×. Again, as in [Li1, Lemma 2.30], strong approximation implies that there exists 1 ≤ i ≤ n with a = ai. (cid:3) i=1(ai \ p · ai) ∪ (p · a)× (∗) = Sn i=1 a× All in all, we obtain Proposition 3.9. For every prime ideal p, let δR,p = ER×R× − ER×p× (with p× = p \ {0}). Moreover, set J ({p}) := {(b + a) × (a \ p · a): b ∈ K, a a fractional ideal of K} ∪ {∅} . Then the minimal non-zero primitive ideals of DP ⊆G ⋊r G are given by J{p} = hδR,pi = C ∗({EX: X ∈ J ({p})}) ⋊r G, p ∈ P. Moreover, J ({p}) is independent. 4. K-theoretic invariants Our goal in this section is to compute the K-theory of minimal non-zero primitive r (R ⋊ R×). Moreover, we also determine the torsion order of the class ideals of C ∗ of the unit in the corresponding quotients. As before, we write P = R ⋊ R× and G = K ⋊ K ×. r (R ⋊ R×) has already been computed in [C-E-L1]. Let us recall The K-theory of C ∗ the result. For every class k in the class group ClK, let ak be an ideal of R representing k. If k is the trivial class in ClK, we choose ak = R. Theorem (Theorem 8.2.1 in [C-E-L1]). The homomorphisms ιk : C ∗(ak ⋊ R∗) → C ∗ r (R ⋊ R×) given by ux 7→ VxEak×a× k = ak \ {0}) induce an isomorphism (a× k Xk (ιk)∗ : Mk∈ClK K∗(C ∗(ak ⋊ R∗)) ∼= K∗(C ∗ r (R ⋊ R×)). Here and in the sequel, the ux are the canonical unitaries in the group C-algebra C ∗(ak ⋊ R∗). 12 XIN LI As above, we write P for the set of non-zero prime ideals of R, and we set δa,p = Ea×a× − Ea×(p·a)× for every (0) 6= a ⊳ R and p ∈ P. Theorem 4.1. The minimal non-zero primitive ideals of C ∗ I{p} := hδR,piC∗ phisms κk : C ∗(ak ⋊ R∗) → I{p}, ux 7→ Vxδak,p induce an isomorphism r (R ⋊ R×) are given by r (R⋊R×), p ∈ P. Moreover, for every prime ideal p, the homomor- Xk (κk)∗ : Mk∈ClK K∗(C ∗(ak ⋊ R∗)) ∼= K∗(I{p}). r (R⋊R×) is isomorphic to a full corner of DP ⊆G⋊r G for P = R⋊R×, Proof. Since C ∗ r (R ⋊ R×) are in G = K ⋊ K ×, we know that minimal non-zero primitive ideals of C ∗ one-to-one correspondence with minimal non-zero primitive ideals of DP ⊆G⋊rG. For every prime ideal p, the ideal J{p} of DP ⊆G ⋊r G corresponds to the ideal EP J{p}EP r (R⋊R×) of C ∗ r (R ⋊ R×). of EP (DP ⊆G ⋊r G)EP , hence to the ideal I{p} = hδR,piC∗ This together with Proposition 3.9 proves the first part of our theorem. For the K-theoretic formula, just apply [C-E-L2, Corollary 3.14] with G = K ⋊ K × and I = J ({p})× = J ({p}) \ {∅}. (cid:3) Let us study the inclusion i{p} : I{p} ֒→ C ∗ r (R ⋊ R×) in K-theory. Recall that ιk : and κk : C ∗(ak⋊R∗) → I{p}, ux 7→ Vxδak,p r (R⋊R×), ux 7→ VxEak×a× C ∗(ak⋊R∗) → C ∗ are the homomorphisms from the theorems above. Moreover, we introduce another homomorphism ιk,p : C ∗(ak ⋊ R∗) → C ∗ r (R ⋊ R×), ux 7→ VxEak×(p·ak)×. It is clear that in K-theory, i{p} ◦ κk is the difference of the two homomorphisms ιk and ιk,p. To describe ιk,p in K0, we choose for every k ∈ ClK an element a(p · ak) ∈ K × such that a(p · ak) · p · ak = a[p]·k. This element a(p · ak) induces the following isomomorphism: k µa(p·ak) : C ∗((p · ak) ⋊ R∗) ∼= C ∗(a[p]·k ⋊ R∗), u(b,a) 7→ u(a(p·ak)b,a). The element a(p · ak) is determined up to a unit in R×, so that µa(p·ak) is determined up to an inner automorphism. Hence in K-theory, (µa(p·ak))∗ does not depend on the choice of a(p · ak). Lemma 4.2. (ιk,p)∗ = (ι[p]·k)∗ ◦ (µa(p·ak))∗ ◦ res (p·ak)⋊R∗ . ak⋊R∗ Proof. The image of ιk,p lies in Ck,p := C ∗((cid:8)E(r+p·ak)×(p·ak)×: r ∈ ak/p · ak(cid:9) ∪(cid:8)VxEak×(p·ak)×: x ∈ ak ⋊ R∗(cid:9)). It is clear that Ck,p can be identified with C(ak/p · ak) ⋊ (ak ⋊ R∗) ∼= C(ak ⋊ R∗/(p · ak) ⋊ R∗) ⋊ (ak ⋊ R∗) in such a way that ιk,p corresponds to the canonical embedding C ∗(ak⋊R∗) → C(ak⋊R∗/(p·ak)⋊R∗)⋊(ak⋊R∗). Now let ρk,p denote the composition ∼= C(ak ⋊ R∗/(p · ak) ⋊ R∗) ⋊ (ak ⋊ R∗) ∼= MN (p)(C ∗((p · ak) ⋊ R∗)). The last Ck,p isomorphism is the homomorphism φ from the appendix (defined before (10)), for SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 13 G = ak ⋊ R∗ and H = (p · ak) ⋊ R∗. Now consider the diagram (3) MN (p)(C ∗((p · ak) ⋊ R∗)) e⊗id C ∗((p · ak) ⋊ R∗) C ∗(ak ⋊ R∗) ιk,pCk,p ιk,p (◗◗◗◗◗◗◗◗◗◗◗◗◗ ∼= ρk,p Ck,p iCk,p ∼= µa(p·ak) C ∗(a[p]·k ⋊ R∗) t✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐ ι[p]·k C ∗ r (R ⋊ R×) where iCk,p is the canonical inclusion and e ⊗ id is the canonical embedding into the upper left corner. The lower left triangle in (3) commutes by construction. The pentagon on the right of (3) commutes in K-theory because once we compose with the inclusion C ∗ r (R ⋊ R×) ֒→ DP ⊆G ⋊r G, the homomorphisms ι[p]·k ◦ µa(p·ak) and iCk,p ◦ (ρk,p)−1 ◦ (e ⊗ id) only differ by an inner automorphism. Therefore, we obtain (ιk,p)∗ = (iCk,p)∗ ◦ (ιk,pCk,p)∗ = (ι[p]·k)∗ ◦ (µa(p·ak))∗ ◦ (e ⊗ id)−1 ∗ ◦ (ρk,p)∗ ◦ (ιk,pCk,p)∗. But ρk,p ◦ ιk,pCk,p coincides with the homomorphism ϕ from (10) in the appendix for G = ak ⋊ R∗, H = (p · ak) ⋊ R∗. Thus by Lemma A.1 in the appendix, (e ⊗ id)−1 ∗ ◦ (ρk,p)∗ ◦ (ιk,pCk,p)∗ = res (p·ak)⋊R∗ (cid:3) and we are done. ak⋊R∗ For every prime ideal p, let hp be the order of [p] in ClK and let a(php) ∈ K × satisfy a(php) · php = R. Lemma 4.3. Upon identifying K0(C ∗(R ⋊ R∗)) with the corresponding direct sum- K0(C ∗(ak ⋊ R∗)), we have that mand in Lk∈ClK and Im  (Xk (ιk)∗)−1 ◦ i{p} ◦ (Xk (κk)∗)  ∩ Z[1][R] (id − (µa(php ))∗ ◦ res php ⋊R∗ R⋊R∗ )(K0(C ∗(R ⋊ R∗))) ∩ Z[1] coincide. Here [1] is the class of the unit of C ∗(R ⋊ R∗) and [1][R] is the image of K0(C ∗(ak ⋊ R∗)) under the canonical inclusion K0(C ∗(R ⋊ R∗)) ֒→ [1] in Lk∈ClK Lk K0(C ∗(ak ⋊ R∗)). Proof. The previous lemma tells us that (4) (i{p} ◦ κ{p})∗ = (ιk)∗ − (ιk,p)∗ = (ιk)∗ − (ι[p]·k)∗ ◦ (µa(p·ak))∗ ◦ res (p·ak)⋊R∗ ak⋊R∗ . o o / / ( O O t 14 XIN LI Therefore, it suffices to consider only those k ∈ ClK which lie in the subgroup of ClK generated by [p], i.e. h[p]i =(cid:8)[R], [p], . . . , [php](cid:9). In other words, Im  (Xk (ιk)∗)−1 ◦ i{p} ◦ (Xk = Im  Xh=0 (ι[ph])∗)−1 ◦ i{p} ◦ ( ( hp−1 (κk)∗)  (κ[ph])∗)  Xh=0 hp−1 ∩ Z[1][R] ∩ Z[1][R]. Using (4), we further compute hp−1 ( (ι[ph])∗)−1 ◦ i{p} ◦ ( Xh=0 (ι[ph])∗)−1 ◦ Xh = (Xh hp−1 Xh=0 (κ[ph])∗) hp−1 (p·a[ph])⋊R∗ Then hp−1 ( (ι[ph])∗)−1◦i{p}◦( (ι[ph])∗ − (ι[ph+1])∗ ◦ (µa(p·a[ph]))∗ ◦ res a[ph]⋊R∗ ! . Let ε[ph] be the canonical embedding K0(C ∗(a[ph] ⋊R∗)) →Lk∈h[p]i K0(C ∗(ak ⋊R∗)). a[ph]⋊R∗ (cid:19) Xh=0 on Lk∈h[p]i K0(C ∗(ak ⋊ R∗)) ⊆Lk∈ClK Xh (cid:18)ε[ph] − ε[ph+1] ◦ (µa(p·a[ph]))∗ ◦ res (κ[ph])∗) =Xh (cid:18)ε[ph] − ε[ph+1] ◦ (µa(p·a[ph]))∗ ◦ res a[ph]⋊R∗ (cid:19) (xh) = z[1][R] ∈ Z[1][R]. Now let xh, 0 ≤ h ≤ hp − 1, be elements of K0(C ∗(a[ph] ⋊ R∗)) such that K0(C ∗(ak ⋊ R∗)). (p·a[ph])⋊R∗ (p·a[ph])⋊R∗ Xh=0 Since Xh (cid:18)ε[ph] − ε[ph+1] ◦ (µa(p·a[ph]))∗ ◦ res = ε[R](cid:18)x0 − ((µa(p·a ))∗ ◦ res hp−1] [p hp−1] (p·a [p [p a hp−1] + hp−2 Xh=0 ε[ph+1](cid:18)xh+1 − ((µa(p·a[ph]))∗ ◦ res ⋊R∗) (p·a[ph])⋊R∗ a[ph]⋊R∗ (cid:19) (xh) )(xhp−1)(cid:19) )(xh)(cid:19) , (p·a[ph])⋊R∗ a[ph]⋊R∗ ⋊R∗ we deduce that for every 0 ≤ h ≤ hp − 2, we must have (p·a[ph])⋊R∗ a[ph]⋊R∗ xh+1 = ((µa(p·a[ph]))∗ ◦ res )(xh). Thus (5) z[1][R] = x0 − (µa(p·a hp−1] [p ))∗ ◦ res )⋊R∗ (p·a a [p hp−1] [p hp−1] ⋊R∗ ◦ · · · ◦ (µa(p))∗ ◦ res p⋊R∗ R⋊R∗ (x0). As a(p · a[php−1]) · a(p · a[php−2]) · · · a(p · a[p])a(p)php = R, SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 15 we deduce that a(p · a[php−1]) · a(p · a[php−2]) · · · a(p · a[p])a(p) and a(php) only differ by a unit in R×. Thus (µa(p·a hp−1] [p ))∗ ◦ · · · (µa(p))∗ = (µa(php ))∗. By Lemma A.2 in the appendix, this tells us that (µa(p·a hp−1] [p ))∗ ◦ res )⋊R∗ hp−1] (p·a [p hp−1] a [p ⋊R∗ ◦ · · · ◦ (µa(p))∗ ◦ res p⋊R∗ R⋊R∗ = (µa(php ))∗ ◦ res php ⋊R∗ R⋊R∗ . Thus z[1][R] = (id − (µa(php ))∗ ◦ res php ⋊R∗ R⋊R∗ )(x0). (cid:3) Now write R∗ = µ × Γ where µ is the group of roots of unity in K, and Γ is a free abelian group. Let 1 be the unit of C ∗(R⋊Γ). Fix an element a(php) ∈ K × such that be the homomorphism C ∗(php ⋊Γ) ∼= C ∗(R⋊Γ), a(php)php = R. Moreover, let µΓ u(b,a) 7→ u(a(php )b,a). a(php ) Lemma 4.4. (id−(µΓ a(php ) )∗ ◦res php ⋊Γ R⋊Γ )(K0(C ∗(R⋊Γ)))∩ Z[1] = ((N (p)hp −1)Z)[1]. ∗ and τ php ⋊Γ ∗ Proof. Let τ R⋊Γ and τ php ⋊Γ be the canonical tracial states on C ∗(R⋊Γ) and C ∗(php⋊ Γ). We write τ R⋊Γ for the induced homomorphisms on K0. As R ⋊ Γ is amenable, it satisfies the Baum-Connes conjecture. Since R ⋊ Γ is also torsionfree, we know by [Val, Proposition 6.3.1] that τ R⋊Γ (K0(C ∗(R ⋊ Γ))) = Z. So we obtain a decomposition K0(C ∗(R ⋊ Γ)) = Z[1] ⊕ ker (τ R⋊Γ ). By Lemma A.3 in the appendix, we know that τ php ⋊Γ . Moreover, we certainly have ◦(µa(php ))∗◦res php ⋊Γ τ R⋊Γ R⋊Γ = N (p)hp · ∗ τ R⋊Γ ). Therefore, ∗ (id − (µa(php ))∗ ◦ res php ⋊Γ R⋊Γ ) preserves the direct sum decomposition K0(C ∗(R ⋊ Γ)) = Z[1] ⊕ ker (τ R⋊Γ ◦(µa(php ))∗ = τ php ⋊Γ . This shows that (µa(php ))∗ ◦ res php ⋊Γ R⋊Γ = N (p)hp · τ R⋊Γ . So we conclude that τ R⋊Γ )) ⊆ ker (τ R⋊Γ R⋊Γ (ker (τ R⋊Γ ∗ ◦ res php ⋊Γ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ). It follows that (id − (µa(php ))∗ ◦ res php ⋊Γ = Z(id − (µa(php ))∗ ◦ res php ⋊Γ R⋊Γ )(K0(C ∗(R ⋊ Γ))) ∩ Z[1] R⋊Γ )([1]) = ((N (p)hp − 1)Z)[1]. (cid:3) Let m be the number of roots of unity in K, and let 1 now denote the unit of C ∗(R ⋊ R∗). Lemma 4.5. (id − (µa(php ))∗ ◦ res php ⋊R∗ R⋊R∗ )(K0(C ∗(R ⋊ R∗))) ∩ Z[1] ⊆(cid:16) N (p)hp −1 gcd(m,N (p)hp −1) · Z(cid:17) [1]. Proof. By Lemma A.2 in the appendix, we know that res R⋊Γ R⋊R∗ ◦ (µa(php ))∗ = (µΓ a(php ))∗ ◦ res php ⋊Γ php ⋊R∗. 16 XIN LI In addition, it is clear that res php ⋊Γ php ⋊R∗ ◦ res php ⋊R∗ R⋊R∗ = res php ⋊Γ R⋊R∗ = res php ⋊Γ R⋊Γ ◦ res R⋊Γ R⋊R∗ . Therefore, the following diagram commutes: id−(µ K0(C ∗(R ⋊ R∗)) a(p hp ) )∗◦res p hp ⋊R∗ R⋊R∗ / K0(C ∗(R ⋊ R∗)) res R⋊Γ R⋊R∗ K0(C ∗(R ⋊ Γ)) id−(µΓ a(p hp ) )∗◦res p res R⋊Γ R⋊R∗ hp ⋊Γ R⋊Γ / K0(C ∗(R ⋊ Γ)) Now take x ∈ K0(C ∗(R ⋊ R∗)) with (id − (µa(php ))∗ ◦ res php ⋊R∗ Z[1C∗(R⋊R∗)]. Then R⋊R∗ )(x) = z[1C∗(R⋊R∗)] ∈ mz[1C∗(R⋊Γ)] = res R⋊Γ = res R⋊Γ R⋊R∗ (z[1C∗(R⋊R∗)]) R⋊R∗ ((id − (µa(php ))∗ ◦ res php ⋊R∗ R⋊Γ )(res R⋊Γ R⋊Γ )K0(C ∗(R ⋊ Γ)) ∩ Z[1C∗(R⋊Γ)] R⋊R∗ )(x)) R⋊R∗ (x)) = (id − (µΓ a(php ))∗ ◦ res php ⋊Γ a(php ))∗ ◦ res php ⋊Γ ∈ (id − (µΓ = ((N (p)hp − 1)Z)[1C∗(R⋊Γ)] by the previous lemma. Hence mz ∈ (N (p)hp − 1)Z which implies z ∈ N (p)hp −1 gcd(m,N (p)hp −1) · Z. (cid:3) Let ζ be a root of unity which generates µ, i.e. µ = hζi. Let p be a prime ideal with the property that for every 1 ≤ i ≤ m − 1, we have 1 − ζ i /∈ p. This implies that the order of ζ in R/p is m. Thus (R/p)∗ contains a cyclic subgroup of order m. Therefore m divides N (p)−1, hence also N (p)hp −1. This means that gcd(m, N (p)hp −1) = m. Lemma 4.6. For a prime ideal p with 1 − ζ i /∈ p for every 1 ≤ i ≤ m − 1, we have N (p)hp −1 m [1] ∈ (id − (µa(php ))∗ ◦ res php ⋊R∗ R⋊R∗ )(K0(C ∗(R ⋊ R∗))). Here 1 is the unit in C ∗(R ⋊ R∗). mPm−1 Proof. Let e be the projection 1 i=0 (u(0,ζ))i in C ∗(R ⋊ R∗). Consider the homo- morphism ϕ : C ∗(R ⋊ R∗) → MN (p)(C ∗(php ⋊ R∗)) for G = R ⋊ R∗, H = php ⋊ R∗ from (10) in the appendix. Now ϕ(u(0,ζ)) is a matrix which can be decomposed into irreducible ones. In this decomposition, we only obtain one 1-dimensional (i.e. 1 × 1) irreducible matrix, namely corresponding to the trivial coset php ⋊ R∗ of R ⋊ R∗/php ⋊ R∗, and N (p)hp −1 m-dimensional (i.e. m × m) irreducible matrices. This follows from our assumption that 1−ζ i /∈ p for all 1 ≤ i ≤ m−1. The analogues of Lemma 5.6 and Lemma 5.7 in [Li-Lu] imply that m ((µa(php ))∗ ◦ res php ⋊R∗ R⋊R∗ )[e] = [e] + N (p)hp −1 m [1]. (cid:3) / / SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 17 As an immediate consequence of the two previous lemmas, we obtain Corollary 4.7. For every prime ideal p with 1 − ζ i /∈ p for all 1 ≤ i ≤ m − 1, (id − (µa(php ))∗ ◦ res php ⋊R∗ R⋊R∗ )(K0(C ∗(R ⋊ R∗))) ∩ Z[1] = ( N (p)hp −1 m · Z)[1]. Proposition 4.8. For every prime ideal p with 1 − ζ i /∈ p for all 1 ≤ i ≤ m − 1, the r (R ⋊ R×)/I{p} is a torsion element of order N (p)hp −1 K0-class [1] of the unit of C ∗ . m Proof. Using the six term exact sequence in K-theory for the short exact sequence r (R ⋊ R×)/I{p} → 0, our claim follows once we know r (R ⋊ R×) → C ∗ i{p}−→ C ∗ 0 → I{p} that (6) in K0(C ∗ r (R ⋊ R×)). Im (i{p}) ∩ Z[1C∗ r (R⋊R×)] = ( N (p)hp −1 m · Z)[1C∗ r (R⋊R×)] As before, let [1][R] be the element of Lk∈ClK ([1][R])k =(0 if k 6= [R], [1C∗(R⋊R∗)] if k = [R]. K0(C ∗(ak ⋊ R∗)) defined by Using the identifications K∗(C ∗(ak ⋊ R∗)) ∼= K∗(I{p}), But this is precisely what we obtain by combining Lemma 4.3 with Corollary 4.7. (cid:3) Set pmax := max(cid:8)p ∈ N prime: pN (1 − ζ i) for some 1 ≤ i ≤ m − 1(cid:9). Also let n := [K : Q] and let h be the class number of K. Moreover, let Prim min be the set of minimal non-zero primitive ideals of C ∗ r (R ⋊ R×). The following result tells us how to read off splitting numbers from K-theoretic invariants, at least for all but finitely many prime numbers. Lemma 4.9. For every prime p ∈ N with p−1 formula for the splitting number gK (p) of p in K: max − 1, we have the following m > pnh gK (p) = #(cid:8)I ∈ Prim min: p(m · ord (1C∗ r (R⋊R×)/I ) + 1)(cid:9) . Proof. To prove "≤", observe that p−1 max − 1 implies that p − 1 > pmax − 1, hence p > pmax. Now assume that a prime ideal p lies above p, i.e. p ∩ Z = pZ. Then we must have 1 − ζ i /∈ p for all 1 ≤ i ≤ m − 1 since otherwise, 1 − ζ i ∈ p would imply (1 − ζ i)R ⊆ p, hence N (p)N (1 − ζ i), thus pN (1 − ζ i) for some 1 ≤ i ≤ m > pnh Xk (κk)∗ : Mk∈ClK Xk (ιk)∗ : Mk∈ClK and the fact (Pk(ιk)∗)[1][R] = [1C∗ (ιk)∗)−1 ◦ i{p} ◦ (Xk Im  (Xk (κk)∗)  K∗(C ∗(ak ⋊ R∗)) ∼= K∗(C ∗ r (R ⋊ R×)) r (R⋊R×)], (6) is obviously equivalent to ∩ Z[1][R] = ( N (p)hp −1 m · Z)[1][R]. 18 XIN LI m − 1. But this contradicts p > pmax. Thus every p ∈ P with p ∩ Z = pZ satisfies r (R⋊R×)/I{p}) + 1. This ord (1C∗ shows "≤". r (R⋊R×)/I{p}) = N (p)hp −1 . Hence p divides m · ord (1C∗ m r (R⋊R×)/I{p}) + 1 ≥ p, hence ord (1C∗ To see "≥", let p ∈ P be a prime ideal such that p divides m·ord (1C∗ r (R⋊R×)/I{p})+1. This implies m · ord (1C∗ m > pnh max − 1. Let q ∈ N be the prime number determined by p ∩ Z = qZ. Then we know by Lemma 4.5 that qnh − 1 ≥ ord (1C∗ max − 1. This implies q > pmax. Hence we deduce that 1−ζ i /∈ p for all 1 ≤ i ≤ m−1, because otherwise, we would again obtain N (p)N (1 − ζ i) and hence qN (1 − ζ i) for some 1 ≤ i ≤ m − 1. r (R⋊R×)/I{p}) = N (p)hp −1 But this cannot happen since q > pmax. Hence ord (1C∗ . Therefore, our assumption that p divides m · ord (1C∗ r (R⋊R×)/I{p}) + 1 implies that p divides N (p)hp, hence p ∩ Z = pZ (and thus p = q). This shows "≥". (cid:3) r (R⋊R×)/I{p}) ≥ p−1 r (R⋊R×)/I{p}) ≥ p−1 m > pnh m As a consequence, we obtain our main result Theorem 1.1. Theorem (Theorem 1.1). Assume that K and L are two number fields which have the same number of roots of unity. Let R and S be their rings of integers. If r (R ⋊ R×) ∼= C ∗ C ∗ r (S ⋊ S×), then K and L are arithmetically equivalent. Proof. It follows from our assumptions and the previous lemma that for all but finitely many prime numbers p ∈ N, we have gK(p) = gL(p). Hence K and L are arithmetically equivalent by [Per-Stu, Main Theorem]. (cid:3) And we also obtain Theorem 1.2. Theorem (Theorem 1.2). Assume that K and L are finite Galois extensions of r (R ⋊ R×) ∼= Q which have the same number of roots of unity. Then we have C ∗ r (S ⋊ S×) if and only if K ∼= L. C ∗ Proof. The direction "⇐" is clear: If K ∼= L, then R ∼= S and thus R ⋊ R× ∼= r (S ⋊ S×). The other direction "⇒" follows S ⋊ S×. This implies C ∗ from Theorem 1.1 and the fact that two finite Galois extensions of Q which are arithmetically equivalent must already be isomorphic (see [Per] or [Neu, Chapter VII, Corollary (13.10)]). (cid:3) r (R ⋊ R×) ∼= C ∗ 5. The number of roots of unity An obvious question is whether in the same setting as in Theorem 1.1 and Theo- rem 1.2, we can deduce from C ∗ r (S ⋊ S×) that K and L have the same number of roots of unity. If this was the case, then we could drop this extra assumption in our main result. Unfortunately, we cannot answer this question in general. However, in the following, we provide a positive answer in a special case, namely for purely imaginary number fields. r (R ⋊ R×) ∼= C ∗ SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 19 We start with the following easy observation, which is an immediate consequence of our results from the previous section. Lemma 5.1. Let K be a number field with ring of integers R. Then we have the following formula for the degree of K over Q: [K : Q] = lim sup T →∞ #(cid:8)I ∈ Prim min: ord (1C∗ r (R⋊R×)/I ) = T(cid:9) . Proof. By [Neu, Chapter I, § 8], we know that there are infinitely many prime numbers which totally split in K. This observation, combined with Proposition 4.8, gives the desired formula for the degree. (cid:3) Let K be a number field with ring of integers R. In the previous section, we only r (R ⋊ R×). Now we turn to the studied the minimal non-zero primitive ideals of C ∗ maximal primitive ideal. As observed in [Ech-La, Remark 3.9], the corresponding quotient is canonically isomorphic to the ring C-algebra A[R] from [Cu-Li1]. Let r (R ⋊ R×) ։ A[R] be the canonical projection, and let π∗ be the induced π : C ∗ homomorphism on K0. Proposition 5.2. Let K be a purely imaginary number field, i.e., K has no real embeddings. Let m be the number of roots of unity in K. Then we can compute the rank of π∗ as follows: rk (π∗) = m2 1 2 [K:Q ]−2. Proof. The proof consists of two steps. First, consider the identification A[R] ∼= C(R) ⋊e (R ⋊ R×) from [Cu-Li1]. Here R is the profinite completion of R, ⋊e stands for "semigroup crossed product by endomorphisms", and R ⋊ R× acts on C(R) by affine transformations. Using this identification, we obtain a canonical homomorphism C(R) ⋊ R ⋊ R∗ → A[R], hence a homomorphism K0(C(R) ⋊ R ⋊ R∗) → K0(A[R]). Let r be the rank of the image of this homomorphism. The first step of our argument is to show rk (π∗) = r. Here is why we have rk (π∗) = r: As before, given k ∈ ClK, let ak be an ideal of R which represents k. Then we have for every k ∈ ClK a commutative diagram (7) K0(C ∗ r (ak ⋊ R∗)) K0(C(R) ⋊ R ⋊ R∗) K0(C ∗ r (R ⋊ R×)) / K0(A[R]) All the arrows are induced by canonical homomorphisms on the level of C-algebras. Now Theorem 8.2.1 in [C-E-L1] tells us that the images of the left vertical arrows, r (R ⋊ R×)). This fact, together with commu- for all k ∈ ClK, sum up to K0(C ∗ tativity of the diagram, immediately yields the inequality rk (π∗) ≤ r. For the reverse inequality, note that the analysis in [Cu-Li2, § 6] and [Li-Lu, § 4] show that K0(C(R) ⋊ R ⋊ R∗) can be written as a stationary inductive limit with groups given by K0(C ∗(R ⋊ R∗)) and structure maps of a particular form. It follows from this description that the canonical homomorphism C ∗(R⋊ R∗) → C(R)⋊ R⋊ R∗ induces / / / 20 XIN LI a rationally surjective homomorphism K0(C ∗(R ⋊ R∗)) → K0(C(R) ⋊ R ⋊ R∗) in K-theory. All we have to do now is to consider the commutative diagram (7) for k = [R], i.e., ak = R. Since the upper horizontal arrow is rationally surjective, we immediately obtain rk (π∗) ≥ r, as desired. Secondly, the duality theorem from [Cu-Li1, § 4] allows us to identify K0(C(R) ⋊ R ⋊ R∗) with K0(C0(A∞) ⋊ K ⋊ R∗) and K0(A[R]) with K0(C0(A∞) ⋊ K ⋊ K ×) such that the diagram K0(C(R) ⋊ R ⋊ R∗) ∼= K0(A[R]) ∼= K0(C0(A∞) ⋊ K ⋊ R∗) / K0(C0(A∞) ⋊ K ⋊ K ×) commutes. Here A∞ is the infinite adele space over K. This commutative diagram shows that r coincides with the rank of the image of K0(C0(A∞) ⋊ K ⋊ R∗) → K0(C0(A∞) ⋊ K ⋊ K ×). By [Li-Lu, Corollary 4.15], we know that for every c ∈ Z× ⊆ R×, c > 1, the canonical homomorphism K0(C0(A∞) ⋊ K ⋊ (R∗ × hci)) → K0(C0(A∞) ⋊ K ⋊ K ×) is rationally injective. Hence r also concides with the rank of the image of K0(C0(A∞) ⋊ K ⋊ R∗) → K0(C0(A∞) ⋊ K ⋊ (R∗ × hci)). Using this, we now show that r = m2 1 2 [K:Q ]−2. Consider the commutative diagram K0(C0(A∞) ⋊ K ⋊ R∗) / K0(C0(A∞) ⋊ K ⋊ (R∗ × hci)) K0(C0(A∞) ⋊ R∗) K0(C0(A∞) ⋊ (R∗ × hci)) Again, all the arrows are induced by canonical homomorphisms on the level of C- algebras. Using the Pimsner-Voiculescu exact sequence, it is easy to see that the horizontal arrow at the bottom of the diagram is rationally injective. By the injectiv- ity results in [Li-Lu, § 5], we also know that the right vertical arrow in the diagram is 1 2 [K:Q ]−2. For rationally injective. This shows that r ≥ rk (K0(C0(A∞) ⋊ R∗)) = m2 the last equality, we have used the assumption that K is purely imaginary, so that rk (R∗) = 1 2 [K : Q] − 1. At the same time, we know that rk (K0(C0(A∞) ⋊ K ⋊ (R∗ × 1 2 [K:Q ]−1 by [Li-Lu, Corollary 4.15]. And finally, examining the Pimsner- hci))) = m2 Voiculescu exact sequence for C0(A∞) ⋊ K ⋊ (R∗ × hci) ∼= (C0(A∞) ⋊ K ⋊ R∗) ⋊ Z, we deduce that r ≤ m2 1 2 [K:Q ]−2. (cid:3) Corollary 5.3. Let K and L be two number fields with rings of integers R and S, r (S ⋊ S×), then K and L must have the same degree respectively. If C ∗ over Q. Moreover, if K and L are both purely imaginary, then K and L have the same number of roots of unity. r (R ⋊ R×) ∼= C ∗ Proof. The first part of our assertion is an immediate consequence of the previous r (S ⋊ S×) lemma. For the second claim, note that an isomorphism C ∗ r (R ⋊ R×) ∼= C ∗ / / / / / / O O O O SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 21 must preserve the maximal primitive ideals. Therefore, the previous proposition implies that K and L must have the same number of roots of unity. (cid:3) With the help of this corollary, we obtain the following stronger version of our main result: Theorem (Theorem 1.3). Let K and L be finite Galois extensions of Q with rings of integers R and S, respectively. Assume that either both K and L have at least one r (R ⋊ R×) ∼= real embedding, or that both K and L are purely imaginary. Then C ∗ r (S ⋊ S×) if and only if K ∼= L. C ∗ Remark 5.4. In the last theorem, we can replace the assumption that either both K and L have at least one real embedding or that both K and L are purely imaginary by the assumption that either both K and L have only two roots of unity +1 and −1, or that K and L both have more than two roots of unity. The reason is that a number field with more than two roots of unity can never have a real embedding. In this way, we obtain Theorem 1.4. However, the question remains whether C ∗ r (S ⋊ S×) excludes the possibility that K has roots of unity +1 and −1 whereas L has more than two roots of unity. r (R ⋊ R×) ∼= C ∗ Appendix A. Restriction homomorphisms In the following, we collect a few certainly well-known facts about restriction homo- morphisms in K0. Of course, we do not claim any originality here. Assume that G is a group and that H is a subgroup of G of finite index. Let N := (G : H). The restriction res H G : K0(C ∗ r (H)) is defined by sending a finitely generated projective right C ∗ r (H), i.e. we just restrict the C ∗ G in terms of homomorphisms of C-algebras. Let R ⊆ G be a full set of representatives of G/H, i.e R ∋ γ 7→ γH ∈ G/H is a bijection. From the set-theoretic bijection G = ·∪γ∈RγH ∼= ·∪γ∈RH, we obtain a unitary r (G)) → K0(C ∗ r (G)-module M to M C∗ r (H)-action. Let us now describe res H r (G)-action on M to a C ∗ (8) λγhεh)γ. ℓ2(G) ∼= Mγ∈R ℓ2(H), Xγ∈R,h∈H λγhεγh 7→ (Xh Let EH be the orthogonal projection in L(ℓ2(G)) onto ℓ2(H) ⊆ ℓ2(G). Conjugation with the unitary from (8) yields an isomorphism L(ℓ2(g)) ∼= MN (L(ℓ2(H))), T 7→ (EH λ−1 (9) Let us now represent C(G/H) ⋊r G in a canonical and faithful way on ℓ2(G) and identify C(G/H) ⋊r G with its image in L(ℓ2(G)). Then the isomorphism from r (H)) ⊆ MN (L(ℓ2(H))). Let us call this (9) identifies C(G/H) ⋊r G with MN (C ∗ identification φ, and let ϕ be the composition γ T λγ′EH)γ,γ′. (10) ϕ : C ∗ r (G) → C(G/H) ⋊r G φ −→ ∼= MN (C ∗ r (H)). 22 XIN LI let e be the minimal projection in MN (C ∗ Moreover, entry in the upper left corner, so that e ⊗ id : C ∗ embedding into the upper left corner. We then have the following r (H) → MN (C ∗ r (H)) corresponding to the r (H)) is the canonical Lemma A.1. res H G = (e ⊗ id)−1 ∗ ◦ ϕ∗. Proof. We have to show that for every finitely generated projective right C ∗ module M , we can identify r (G)- M C∗ r (H) ⊗e⊗id MN (C ∗ r (H)) and M ⊗ϕ MN (C ∗ r (H)) as right MN (C ∗ r (H))-modules. But it is easy to see that the homomorphisms M C∗ r (H) ⊗e⊗id MN (C ∗ r (H)) → M ⊗ϕ MN (C ∗ r (H)), m ⊗ x 7→ m ⊗ ex and M ⊗ϕ MN (C ∗ r (H)) → M C∗ r (H) ⊗e⊗id MN (C ∗ r (H)), m ⊗ x 7→ Xγ∈R mλγ ⊗ ϕ(λγ−1 )x are mutually inverse homomorphisms of right MN (C ∗ r (H))-modules. (cid:3) Now let H ⊆ G and H ′ ⊆ G′ be two subgroups of groups such that (G : H), (G′ : H ′) < ∞. Assume that α : G → G′ is a group isomorphism with α(H) = H ′. Such an isomorphism induces homomorphisms C ∗ r (αH ) : C ∗ r (G)) → K0(C ∗ r (H ′) as well as homomorphisms α∗ = (C ∗ r (G′) and C ∗ r (α))∗ : K0(C ∗ r (G′)), (αH )∗ = (C ∗ r (αH))∗ : K0(C ∗ r (H)) → K0(C ∗ r (H) → C ∗ r (G) → C ∗ r (α) : C ∗ r (H ′)). Lemma A.2. (αH )∗ ◦ res H G = res H ′ G′ ◦ α∗. Proof. Given a right C ∗ C ∗ m ⊗ x 7→ m ⊗ x and mC ∗ r (H ′) with (M ⊗C∗ r (G)-module M , we can obviously identify M C∗ r (α) C ∗ r (αH ) r (H ′) by the mutually inverse homomorphisms (cid:3) r (G′))C∗ r (H) ⊗C∗ r (α)−1(y) ⊗ 1 ←[ m ⊗ y. Finally, we need the following observation about traces: Let τ G and τ H denote the canonical tracial states on C ∗ r (H). We denote the induced homomor- ∗ and τ H phisms on K0 by τ G ∗ . r (G) and C ∗ Lemma A.3. τ H ∗ ◦ res H G = N · τ G ∗ . r (H)) such that MN (τ H) ◦ (e ⊗ id) = τ H. Proof. Let MN (τ H ) be the trace on MN (C ∗ An easy computation shows that MN (τ H ) ◦ ϕ = N · τ G, where ϕ is defined in (10). Our claim follows. (cid:3) SEMIGROUP C-ALGEBRAS ATTACHED TO NUMBER FIELDS 23 References [Bo-Co] J. B. Bost and A. Connes, Hecke algebras, Type III Factors and Phase Transitions with Spontaneous Symmetry Breaking in Number Theory, Selecta Math., New Series, Vol. 1, 3 (1995), 411 -- 457. [Bour] N. Bourbaki, Alg`ebre commutative. Chapitres 5 `a 7. ´El´ements de math´ematique. R´eimpres- sion inchang´ee de l'´edition originale de 1975, Springer-Verlag, Berlin Heidelberg, 2006. [Cor-Mar] G. Cornelissen and M. Marcolli, Quantum statistical mechanics, L-series and an- abelian geometry, arXiv:1009.0736v5 [C-D-L] J. Cuntz, C. Deninger and M. Laca, C-algebras of Toeplitz type associated with algebraic number fields, to appear in Math. Ann. [C-E-L1] J. Cuntz, S. Echterhoff and X. Li, K-theory for semigroup C-algebras, arXiv: 1201.4680v1 [C-E-L2] J. Cuntz, S. Echterhoff and X. Li, On the K-theory of crossed products by automorphic semigroup actions, arXiv:1205.5412v1. [Cu-Li1] J. Cuntz and X. Li, The regular C-algebra of an integral domain, in Quanta of Maths, Clay Math. Proc. 11, Amer. Math. Soc., Providence, RI, 2010, 149 -- 170. [Cu-Li2] J. Cuntz and X. Li, C-algebras associated with integral domains and crossed products by actions on adele spaces, Journal of Noncommutative Geometry 5 (2011), 1 -- 37. [Cu-Li3] J. Cuntz and X. Li, Erratum to "C-algebras associated with integral domains and crossed products by actions on adele spaces", Journal of Noncommutative Geometry 6 (2012), 819 -- 821. [dS-Per] B. de Smit and R. Perlis, Zeta functions do not determine class numbers, Bulletin (New Series) AMS 31 (1994), 213 -- 215. [Ech-La] S. Echterhoff and M. Laca, The primitive ideal space of the C-algebra of the affine semigroup of algebraic integers, arXiv:1201.5632v1. [Ha-Pa] E. Ha and F. Paugam, Bost-Connes-Marcolli systems for Shimura varieties. I. Definitions and formal analytic properties, IMRP Int. Math. Res. Rap. 2005, no. 5, 237 -- 286. [Kir] E. Kirchberg, Das nicht-kommutative Michael-Auswahlprinzip und die Klassifikation nicht- einfacher Algebren, in C-algebras (Munster, 1999), Springer, Berlin, 2000, 92 -- 141. [L-L-N] M. Laca, N. S. Larsen and S. Neshveyev, On Bost-Connes type systems for number fields, J. Number Theory 129 (2009), no. 2, 325 -- 338. [La-Rae] M. Laca and I. Raeburn, Phase transition on the Toeplitz algebra of the affine semigroup over the natural numbers, Adv. Math 225 (2010), no. 2, 643 -- 688. [La-Nesh] M. Laca and S. Neshveyev, Type III1 equilibrium states of the Toeplitz algebra of the affine semigroup over the natural numbers, J. Functional Analysis 261 (2011), no. 2, 169 -- 187. [Li1] X. Li, Semigroup C-algebras and amenability of semigroups, arXiv:1105.5539v2, to appear in J. Functional Analysis. [Li2] X. Li, Nuclearity of semigroup C-algebras and the connection to amenability, arXiv: 1203.0021v2. [Li-Lu] X. Li and W. Luck, K-theory for ring C-algebras -- the case of number fields with higher roots of unity, to appear in Journal of Analysis and Topology. [Me-Ne] R. Meyer and R. Nest, C-algebras over topological spaces: Filtrated K-theory, Canad. J. Math. 64 (2012), no. 2, 368 -- 408. [Neu] J. Neukirch, Algebraic number theory, Die Grundlehren der mathematischen Wissen- schaften, Springer-Verlag, Berlin, 1999. [Nor] M. D. Norling, Inverse semigroup C-algebras associated with left cancellative semigroups, arXiv:1202.5977v1. [Pas-Rør] C. Pasnicu and M. Rørdam, Purely infinite C*-algebras of real rank zero, J. Reine Angew. Math. 642 (2007), 51 -- 73. [Per] R. Perlis, On the equation ζK(s) = ζK ′ (s), J. Number Theory 9 (1977), 342 -- 360. [Per-Stu] R. Perlis and D. Stuart, A new characterization of arithmetic equivalence, J. Number Theory 53 (1995), 300 -- 308. [Val] A. Valette, Introduction to the Baum-Connes Conjecture, Birkhauser, 2002. From notes taken by Indira Chatterji, with an appendix by Guido Mislin. 24 XIN LI Xin Li, Department of Mathematics, Westfalische Wilhelms-Universitat Munster, Einsteinstrasse 62, 48149 Munster, Germany E-mail address: [email protected]
1609.01254	3	1609	2017-04-25T22:58:13	Gradient flow and entropy inequalities for quantum Markov semigroups with detailed balance	[ "math.OA" ]	We study a class of ergodic quantum Markov semigroups on finite-dimensional unital $C^*$-algebras. These semigroups have a unique stationary state $\sigma$, and we are concerned with those that satisfy a quantum detailed balance condition with respect to $\sigma$. We show that the evolution on the set of states that is given by such a quantum Markov semigroup is gradient flow for the relative entropy with respect to $\sigma$ in a particular Riemannian metric on the set of states. This metric is a non-commutative analog of the $2$-Wasserstein metric, and in several interesting cases we are able to show, in analogy with work of Otto on gradient flows with respect to the classical $2$-Wasserstein metric, that the relative entropy is strictly and uniformly convex with respect to the Riemannian metric introduced here. As a consequence, we obtain a number of new inequalities for the decay of relative entropy for ergodic quantum Markov semigroups with detailed balance.	math.OA	math	GRADIENT FLOW AND ENTROPY INEQUALITIES FOR QUANTUM MARKOV SEMIGROUPS WITH DETAILED BALANCE ERIC A. CARLEN AND JAN MAAS Abstract. We study a class of ergodic quantum Markov semigroups on finite- dimensional unital C ∗-algebras. These semigroups have a unique stationary state σ, and we are concerned with those that satisfy a quantum detailed balance condition with respect to σ. We show that the evolution on the set of states that is given by such a quantum Markov semigroup is gradient flow for the relative entropy with respect to σ in a particular Riemannian metric on the set of states. This metric is a non-commutative analog of the 2-Wasserstein metric, and in several interesting cases we are able to show, in analogy with work of Otto on gradient flows with respect to the classical 2-Wasserstein metric, that the relative entropy is strictly and uniformly convex with respect to the Riemannian metric introduced here. As a consequence, we obtain a number of new in- equalities for the decay of relative entropy for ergodic quantum Markov semigroups with detailed balance. 7 1 0 2 r p A 5 2 ] . A O h t a m [ 3 v 4 5 2 1 0 . 9 0 6 1 : v i X r a Keywords: quantum Markov semigroup, entropy, detailed balance, gradient flow. Subject Classification Numbers: 46L57, 81S22, 34D05, 47C90 1. introduction Let A be a finite-dimensional C∗-algebra with unit 1. We may identify A with a C∗- subalgebra of Mn(C), the C∗-algebra of n × n matrices, for some n. In finite dimension, there is no difference between weak and norm closure, and so A is also a von Neumann algebra. A Quantum Markov Semigroup (QMS) is a continuous one-parameter semigroup of linear transformations (Pt)t≥0 on A such that for each t ≥ 0, Pt is completely positive and Pt1 = 1. Associated to any QMS Pt = etL , is the dual semigroup P†t acting on S+(A), the set of faithful states of A. (When there is no ambiguity, we simply write S+.) The QMS Pt is ergodic in case 1 spans the eigenspace of Pt for the eigenvalue 1. In that case, there is a unique invariant state σ. While σ need not be faithful, a natural projection operation allows us to assume, effectively without loss of generality, that σ ∈ S+(A). Characterizations of the generators of quantum Markov semigroups on the C∗-algebra Mn(C) of all n × n matrices were given at the same time by Gorini, Kossakowski and Sudershan [30], and by Lindblad [45] in a more general setting (but still assuming norm continuity of the semigroup). Such semigroups are often called Lindblad semigroups. The notion of detailed balance in the theory of classical Markov processes has several different quantum counterparts, as discussed below. One of these is singled out here, with a full discussion of how it relates to other variants and why it is physically natural. Suffice it to say here that, as we shall see, the class of ergodic QMS that satisfy the detailed balance condition includes a wide variety of examples arising in physics. The set of faithful states S+(A) may be identified with the set of invertible density matrices σ on Cn that belong to A, as is recalled below. For ρ, σ ∈ S+(A), the relative 1 2 ERIC A. CARLEN AND JAN MAAS entropy of ρ with respect to σ is the functional D(ρkσ) = Tr[ρ(log ρ − log σ)] . (1.1) We show in Theorem 7.6 that associated to any QMS Pt = etL satisfying detailed balance, there is a Riemannian metric gL on S+ such that the flow on S+ induced by the dual semigroup P†t is gradient flow for the metric gL of the relative entropy D(·kσ) with respect to the invariant state σ ∈ S+. In several cases, we shall show that the relative entropy functional is geodesically convex for gL, and as a consequence, we shall deduce a number of functional inequalities that are useful for studying the evolution governed by Pt. In particular, we shall deduce several sharp relative entropy dissipation inequalities. Some of these are new; see e.g. Theorem 8.5 and Theorem 8.6. The Riemannian distance corresponding to gL will be seen to be a very natural analog of the 2-Wasserstein distance on the space of probability densities on Rn [72, Chapter 6]. Otto showed [54] that a large number of classical evolution equations could be viewed as gradient flow in the 2-Wasserstein metric for certain functionals, and that when the functionals were geodesically uniformly convex for this geometry, a host of useful functional inequalities were consequently valid. This is for instance the case when the functional driving the flow is classical relative entropy with respect to a Gaussian reference measure. In this case, the flow is given by the classical Ornstein-Uhlenbeck semigroup, described by the Fokker-Planck equation with linear drift. One of the inequalities that is a consequence of the uniform geodesic convexity of the relative entropy is the sharp bound on entropy dissipation for solutions of the Fokker-Planck equation. The present paper also greatly extends our previous paper [11] in which we obtained a gradient flow structure for the Fermi Ornstein-Uhlenbeck semigroup. In particular, we prove a sharp geodesic convexity result for the von Neumann entropy in this setting, thereby solving a problem that was left open in [11]. Thus, our results can be viewed as a non-commutative extension of Otto's investigation of classical gradient flows. Mielke [49, 50] investigated related variational formulations for dissipative quantum sys- tems based on Ottinger's so-called GENERIC framework [52, 53]. However, the gradient flow structure considered in Mielke's papers is in general different from the one introduced in the present paper, as the approach in [49, 50] gives rise to nonlinear evolution equations that are different from the linear Lindblad equations that we obtain here. Also, Junge and Zeng [38] have recently developed an approach to some non-commutative functional inequalities involving a non-commutative analog of the 1-Wasserstein metric. We restrict our attention to the case that A is a finite-dimensional C∗-algebra because this setting already includes many examples of physical interest, and we wish to explain our methods in a way that does not encumber them with the host of technical and topological difficulties that would follow in an infinite-dimensional setting. For example, in an infinite- dimensional setting, it would matter that A is a von Neumann algebra and not only a C∗-algebra with a unit, so that A would have a pre-dual, and then what we call S+(A) here would be the set of normal states; i.e., states that are continuous with respect to the weak-∗ topology on A, and we would require that Pt be normal for each t; i.e., continuous with respect to the weak-∗ topology on A. The innocent formula Pt = etL would require closer scrutiny, and so forth. Many of the issues involved in extending our results to a more general infinite-dimensional setting are standard but not all of them. This will be done elsewhere. We also refer to a forthcoming paper for a treatment of more general transport metrics, general entropy functionals, and their geodesic convexity properties [12]. GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 3 Here instead, we present a development of our methods that will be readily understood by people who are interested in the many physically interesting finite-dimensional exam- ples. We shall also show that our methods may be applied in the infinite-dimensional setting to particular examples without first generalizing the whole theory. We illustrate this with the family of Bose Ornstein-Uhlenbeck semigroups, which are always infinite- dimensional since all non-trivial representations of the Canonical Commutation Relations are infinite-dimensional. We use our methods to prove a new sharp relative entropy dissipa- tion inequality for this family that had recently been conjectured in [37]; see Theorem 8.5. We also prove some other sharp inequalities of this type; e.g., Theorem 8.6. While this introduction has, in the interest of brevity, used a considerable amount of terminology without explanation, the rest of the paper is elementary and quite self- contained. We do assume a basic familiarity with C∗-algebras, von Neumann algebras and completely positive maps. Many readers will have this familiarity, but for background on C∗ and von Neumann algebras, we refer to Sakai [63], and for completely positive maps to Paulsen [58]. Terminology and facts that are used here without explanation can be found in these references. We close the introduction with some preliminary definitions and by establishing some notation that will be in use throughout the paper. 1.1. Preliminary definitions and notation. A will always denote a finite-dimensional C∗-algebra with unit 1, regarded as subalgebra of some matrix algebra Mn(C). The center of A is generated by a single self-adjoint element Z ∈ A, and its spectral projections yield a decomposition of A into a finite direct sum of factors: Let {λ1, . . . , λp} for some 1 ≤ p ≤ n be the distinct eigenvalues of Z, and let Hj be the eigenspace with eigenvalue λj. Then each Hj is invariant under A, and letting Aj denote the restriction of A to Hj, Aj has a trivial center; that is, Aj is a factor. The structure of finite-dimensional factors is well-known: Hj is unitarily equivalent to Cℓj ⊗ Crj , and there is a unitary equivalence in which each A ∈ Aj takes the form Iℓj ⊗ bA where bA ∈ Mrj (C). That is, each factor may be identified in a natural way with a full matrix algebra. However, not all finite-dimensional C∗-algebras arising in mathematical physics are factors. For example, a Clifford algebra with an odd number k of generators, arising in the description of a system of k Fermi degrees of freedom, is not a factor. This example will be discussed in detail later on. Let τ denote the normalized trace on A given by τ (A) = Tr[A]/Tr[1] for all A ∈ A. Then τ is a faithful state on A. Let HA denote the Hilbert space formed by equipping A with the GNS inner product determined by τ . That is, for all A, B ∈ A, hA, BiHA = τ [A∗B] . We could use this inner product to identify HA with the space of linear functionals on A, but to keep contact with the physics literature, we do something slightly different: We use the un-normalized trace to write the general linear functional ϕ on A in the form ϕ(B) = Tr[AB] for some A ∈ A. It is easy to see that ϕ is a positive linear functional (in the sense that ϕ(B) ≥ 0 whenever B ≥ 0) if and only if A ≥ 0 in A. It follows that the set of faithful states on A may be identified with the set of strictly positive elements A of A such that Tr[A] = 1. Since A ∈ A is strictly positive in A if and only if it is strictly positive in Mn(C), we may identify the set of faithful states on A with the set of invertible density 4 ERIC A. CARLEN AND JAN MAAS matrices ρ on Mn(C) that belong to the subalgebra A. In the following, we always write S+ to denote this set, whether we think of it as a set of faithful states or as a set of density matrices. 1.1. DEFINITION (Modular operator and modular group). Let σ ∈ S+. Define a linear operator ∆σ on HA, or, what is the same thing, on A, by ∆σ(A) = σAσ−1 for all A ∈ A. ∆σ is called the modular operator. The modular generator is the self-adjoint element h ∈ A given by h = − log σ . The modular automorphism group αt on A is the group defined by αt(A) = eithAe−ith (1.2) for t ∈ C. Note that ∆σ = αi. The modular operator and the modular automorphism group are central to the charac- terization of an important class of quantum Markov semigroups. First, Davies [20] identified a class of quantum Markov semigroups that he rigorously showed to arise from coupling a finite quantum system with internal Hamiltonian h to an infinite fermion heat bath and then taking a weak coupling limit. The whole class of quantum Markov semigroups he found has the property that the semigroup commutes with the modular automorphism group αt given by (1.2) where h is the internal Hamiltonian, and then σ = e−h/Tr[e−h]. (Note that adding a scalar constant to h has no effect on (1.2).) Of course an operator on HA commutes with the modular group if and only if it commutes with the modular operator. Second, Alicki [4] showed that commutativity with respect to the modular operator is central to one natural extension of the notion of detailed balance from classical Markov chains to the quantum setting. Before explaining this fact, which is important for our work here, we first introduce some more terminology and notation. A linear operator K on A is positivity preserving in case K A ≥ 0 whenever A ≥ 0. A linear operator K on A is self-adjointness preserving in case (K A)∗ = K A∗, or, equivalently, in case K A is self-adjoint whenever A is self-adjoint. Evidently, any positivity preserving operator is self-adjointness preserving. When Pt = etL is a self- adjointness preserving semigroup, then its generator L = limt→0 t−1(Pt − I) is self- adjointness preserving as well. σ (A∗) for all A ∈ A. Moreover, for all A, B ∈ A, Let σ ∈ S+ and note that (∆σA)∗ = ∆−1 Tr[A∗∆σB] = Tr[(∆σA)∗B] and Tr[A∗∆σA] = Tr[σ1/2Aσ−1/22] , so that ∆σ is a positive operator on HA. A dagger † will be used to denote the adjoint with respect to the inner product in HA, or, what is the same, with respect to the Hilbert- Schmidt inner product. We will encounter many other inner products on A, but the GNS inner product associated to τ has a special role. We may then rewrite one of the conclusions from just above as ∆†σ = ∆σ. Because of the self-adjointness of ∆σ there is an orthonormal basis {E1, . . . , Em} of HA consisting of eigenvectors of ∆σ. Since ∆σ1 = 1, we may always assume that E1 = 1. In this case, Tr[Eγ] = τ (Eγ) = 0 for all γ > 1. Furthermore, since ∆σ is strictly positive, all GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 5 eigenvalues of ∆σ are strictly positive, hence we may write them in the form e−ωγ . Since (∆σA)∗ = ∆−1 σ A∗, it follows that for all E ∈ HA, ∆σE = e−ωE ⇐⇒ ∆σE∗ = eωE∗ . (1.3) In particular, if for ω 6= 0, e−ω is an eigenvalue of ∆σ, then so is eω, and the two eigenspaces are orthogonal and have the same dimension. It follows that there exists an orthonormal basis of HA with the properties listed in the next definition: 1.2. DEFINITION (Modular basis). Let A be a finite-dimensional C∗-algebra and let σ ∈ S+(A). Then there exists an orthonormal basis {E1, . . . , Em} of HA with the following properties: (i) {E1, . . . , Em} consists of eigenvectors of ∆σ. (ii) E1 = 1. (iiI) {E1, . . . , Em} = {E∗1 , . . . , E∗m} 2. Detailed balance There is a large literature on the detailed balance condition in a quantum setting, starting with the work of Agarwal [2], which initiated a number of investigations in the 1970's including [4, 13, 41, 67]. Already in these papers, one finds several different, and non- equivalent, definitions. A number of more recent investigations [1, 24, 26, 47, 69] have added to the variety of meanings attached to this term. We therefore carefully explain the context of the definition that we use here, and why it is the most natural for our purposes. Let Pi,j be the Markov transition matrix for a Markov chain on a finite state space S with elements {x1, . . . , xn}. Suppose that σ is a probability density on S that is invariant i=1 σiPi,j for all i. The transition matrix satisfies under this transition function: σj =Pn the detailed balance condition with respect to σ in case σiPi,j = σjPj,i for all i, j . (2.1) Let Xn be the Markov process started from the initial distribution σ, so that the process is stationary. Let Pr be measure on the path space of the process. Then (2.1) is equivalent to Pr{Xn = i, Xn+1 = j} = Pr{Xn = j, Xn+1 = i} for all i, j and all n . In other words, (Xn, Xn+1) has the same joint distribution as (Xn+1, Xn), so that (2.1) characterizes time reversal invariance. There is another characterization of (2.1) in terms of self-adjointness: The matrix P is self-adjoint on Cn equipped with the inner product if and only if (2.1) is satisfied. hf, giσ = nXk=1 σkfkgk , (2.2) There are a number of different ways one might generalize the inner product (2.2) to the quantum setting, and these give different notions of self-adjointness. Let Pt be a QMS on A. A state σ ∈ S+(A) is invariant under P†t in case P†t σ = σ for all t ≥ 0, or equivalently, L †σ = 0. These conditions are equivalent to Tr[σPt(A)] = Tr[σA] for all t > 0 and all A ∈ A, or equivalently, Tr[σL (A)] = 0 for all A ∈ A. 2.1. DEFINITION (Compatible inner product). An inner product h·,·i is compatible with σ ∈ S+(A) in case for all A ∈ A, Tr[σA] = h1, Ai. 6 ERIC A. CARLEN AND JAN MAAS If a quantum Markov semigroup Pt is self-adjoint with respect to an inner product h·,·i that is compatible with σ ∈ S+, then for all A ∈ A, Tr[σA] = h1, Ai = hPt1, Ai = h1, PtAi = Tr[σPtA] , and thus σ is invariant under P†t . 2.2. DEFINITION. Let σ ∈ S+ be a non-degenerate density matrix. For each s ∈ R, and each A, B ∈ A, define hA, Bis = Tr[(σ(1−s)/2Aσs/2)∗(σ(1−s)/2Bσs/2)] = Tr[σsA∗σ1−sB] . (2.3) Evidently each of these inner products is compatible with σ. At s = 1, this is the GNS inner product associated to the state ϕ(A) = Tr[σA]. The value s = 1/2 is also special; h·,·i1/2 is called the KMS inner product. A number of its properties are developed in [60]. The inner products in (2.3) can be written as hA, Bis = Tr[A∗∆1−s σ Bσ] . (2.4) More generally, given any function f : (0,∞) → (0,∞), define (2.5) Note that h·,·i1 is the σ-GNS inner product whether 1 is interpreted as a number, as in (2.4), or as the constant function f (t) = 1, as in (2.5). hA, Bif = Tr[A∗f (∆σ)Bσ] . Let RA denote right multiplication by A, and define Ωf way to write (2.5) is hA, Bif = Tr[A∗Ωf σB]. For all linear operators K on A, σ = Rσ ◦ f (∆σ). Then another hA, K Bif = h[Ωf σ]−1K †(Ωf σA), Bif . It follows immediately that K is self-adjoint with respect to the inner product h·,·if if and only if Ωf σ ◦ K = K † ◦ Ωf σ . 2.3. Remark. For f (t) = ts, the adjoint K ′ of K with respect to the inner product h·,·is is K ′B = σs−1K †(σ1−sBσs)σ−s . For s = 1/2, this reduces to K ′B = σ−1/2K †(σ1/2Bσ1/2)σ−1/2. Since K † is positivity preserving if and only if K is positivity preserving, it is evident that for s = 1/2, K ′ is positivity preserving if and only if K is positivity preserving. However, for other values of s, K ′ need not be positivity preserving when K is. This is one feature that sets the KMS inner product apart from the inner products h·,·is for other values of s. In [69, Definition 16], detailed balance is defined in terms of self-adjointness with respect to h·,·if , yielding a number of a priori different notions of detailed balance depending on the choice of f . The example following Proposition 18 in [69] shows that different choices of f can yield distinct classes of self-adjoint operators, and hence distinct notions of detailed balance. Specifically, it is shown in [69] that self-adjointness with respect to h·,·i1/2 is not the same as self-adjointness with respect to h·,·if where f (t) = (1 + t)/2, corresponding to the Bures metric, as discussed in [69]. The authors of [69] conclude: "The family of quantum detailed balance conditions is therefore much richer than the classical counterpart". However, it turns out that self-adjointness with respect to the GNS inner product, or indeed with respect to h·,·is for any s 6= 1/2 implies self-adjointness with respect to h·,·if GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 7 for all f . The argument leading to this conclusion, for which we have found no reference, is simple and will prove useful. The following is a simple variant on the well-known KMS symmetry condition: 2.4. LEMMA. For s ∈ R, let h·,·is be the inner product defined in (2.3). Then for all t ∈ R and all A, B ∈ A, hαit(A), Bis = hA, Bis−t = hA, αit(B)is . (2.6) In particular, αit is self-adjoint with respect to h·,·is. Proof. Using the definitions we obtain hαit(A), Bis = Tr[σs(σtAσ−t)∗σ1−sB] = Tr[σs−tA∗σ1−s+tB] = hA, Bis−t = hA, αit(B)is , which is the desired identity (cid:3) The following lemma is a result of Alicki [4] for s = 1, and the generalization to s ∈ [0, 1], s 6= 1/2, can be found in [25, Propositon 8.1]. The following short proof, a simple adaptation of Alicki's argument, is included for the reader's convenience. 2.5. LEMMA. Let σ ∈ S+ be a non-degenerate density matrix, and let s ∈ [0, 1], s 6= 1/2. Let K be any operator on A that is self-adjoint with respect to h·,·is and also preserves self-adjointness. Then K commutes with αt, for all t, real and complex. Proof. For any A, B ∈ A, hK αi(2s−1)(A), Bis = Tr[σs(K (σ2s−1Aσ1−2s))∗σ1−sB] = Tr[σs(σ2s−1Aσ1−2s)∗σ1−sK (B)] = Tr[σ1−sA∗σsK (B)] = Tr[σs(K (B∗))∗σ1−sA∗] = Tr[σsBσ1−sK (A∗)] = Tr[σ1−s(K (A))∗σsB] = hK (A), Bi1−s Since s − (2s − 1) = 1 − s, (2.6) yields hK (A), Bi1−s = hαi(2s−1)(K (A)), Bis. As B is arbitrary, αi(2s−1)K = K αi(2s−1). Since K commutes with αi(2s−1), it commutes with every polynomial in the positive self-adjoint operator αi(2s−1) = ∆2s−1 , and hence with f (∆2s−1 (cid:3) ) for every function f . In particular, K commutes with αt for all t. σ σ 2.6. Remark. Davies [20] has studied a class of quantum Markov semigroups that arise in a general model of an n-level quantum system coupled to an infinite heat bath. He studied the weak-coupling limit and gave conditions under which the weak-coupling limit produces a quantum Markov semigroup. This procedure always yields a semigroup that commutes with the evolution given by (1.2) where h is the internal Hamiltonian of the n-level system. This is true whether or not the semigroup has a particular self-adjointness property. In view of Davies' result, it is natural to focus on quantum Markov semigroups that commute with the modular operator associated to their invariant states. 2.7. Remark. The condition that the generator L of a quantum Markov semigroup Pt = etL commutes with ∆σ imposes strong restrictions on the structure of L . Consider the case A = M2(C). Let σ ∈ S+ have two distinct eigenvalues λ1, λ2 > 0. Let {η1, η2} be an orthonormal basis of C2 consisting of eigenvectors of σ: σηj = λjηj for j = 1, 2. Then ∆σ has three distinct eigenvalues 1, λ1/λ2 and λ2/λ1. The latter two eigenvalues have one-dimensional eigenspaces spanned by η1ihη2 and η2ihη1 respectively. If L commutes with ∆σ, then η1ihη2 and η2ihη1 must be eigenvectors of L with eigenvalues, say, ν and ν. Since A := η1ihη2 + η2ihη1 is self-adjoint and L is self-adjointness preserving, 8 ERIC A. CARLEN AND JAN MAAS it follows that L (A) = νη1ihη2 + νη2ihη1 is self-adjoint, which implies that ν = ν. Moreover, since both A and L (A) = νA are self-adjoint, it follows that ν is real. Thus L has at most 3 eigenvalues 1, µ and ν, and ν has multiplicity 2 (or 3 in case ν = µ). To summarize, it follows that L (η1ihη2) = νη1ihη2 and L (η2ihη1) = νη2ihη1 with ν ∈ R . (2.7) From here it is not hard to see that the value s = 1/2 is genuinely exceptional in Lemma 2.5. Suppose that L is self-adjoint with respect to the σ-KMS inner product h·,·i1/2 for σ ∈ S+, where σ has distinct eigenvalues and {η1, η2} is an orthonormal basis of C2 consisting of eigenvectors of σ. It follows from the discussion above that if (2.7) is violated, L does not commute with ∆σ. There are many such operators L on M2(C). An explicit construction is given in Appendix B. 2.8. LEMMA. Let σ ∈ S+ be a non-degenerate density matrix. Let K be any operator on A such that K αt = αtK for all t, or equivalently, ∆σK = K ∆σ. If K is self-adjoint with respect to the inner product h·,·if for some function f : (0,∞) → (0,∞), then the same holds for every function f : (0,∞) → (0,∞). Proof. Suppose that K is self-adjoint with respect to the inner product h·,·if for some function f : (0,∞) → (0,∞). Let g : (0,∞) → (0,∞) be arbitrary, and write h = g/f . Since K commutes with ∆σ, it also commutes with h(∆σ). Thus, for all A, B ∈ A, hA, K (B)ig = Tr[σA∗g(∆σ)K (B)] = Tr[σA∗f (∆σ)K h(∆σ)B] = hA, K h(∆σ)Bif = hK (A), h(∆σ)Bif = Tr[σK (A)∗f (∆σ)h(∆σ)B] = hK (A), Big , which is the desired result. (cid:3) We summarize some immediate consequences of the last lemmas in a theorem: 2.9. THEOREM. Let σ ∈ S+ be a non-degenerate density matrix, and let K be any operator on A. Then: (1) If K is self-adjoint with respect to the GNS inner product h·,·i1, and K A∗ = (K A)∗ for all A ∈ A, then K commutes with the modular automorphism group of σ and moreover, K is self-adjoint with respect to h·,·if for all f : (0,∞) → (0,∞). (2) If K commutes with the modular automorphism group of σ, and if K is self-adjoint with respect to h·,·if for some f : (0,∞) → (0,∞), then K is self-adjoint with respect to h·,·ig for all g : (0,∞) → (0,∞). This brings us to the definition of detailed balance, which is one the various definitions that may be found in the physical literature. 2.10. DEFINITION (Detailed balance). A QMS Pt on A satisfies the detailed balance condition with respect to σ ∈ S+(A) in case for each t > 0, Pt is self-adjoint in the σ-GNS inner product h·,·i1. In this case σ is invariant under P†t , and we say that the QMS Pt satisfies the σ-DBC. 2.11. Remark. Every QMS Pt = etL is self-adjointness preserving: PtA∗ = (PtA)∗. Thus, when Pt satisfies the σ-DBC, it follows from Theorem 2.9 that (2.8) for all t, t′ and all A ∈ A. This crucial observation is due to Alicki [4], and it means that Pt commutes with the time-translation governed by the Hamiltonian h corresponding αt′(Pt(A)) = Pt(αt′(A)) GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 9 to the state σ. The condition (2.8) may therefore be viewed as a quantum analog of time-translation invariance or stationarity. Theorem 2.9 asserts that in the presence of detailed balance, Pt is self-adjoint with respect to a wide variety of inner products and also that Pt commutes with the modular group. Moreover, under the condition that Pt commutes with the modular group, self- adjointness with respect to any member of this wide family of inner products implies self-adjointness with respect to all of them. This is important in what follows. The inner products defined just above include a number of inner products that arise naturally in the theory of operator algebras and mathematical physics. One that will be especially useful here is the Bogoliubov-Kubo-Mori inner product: hA, BiBKM =Z 1 0 Tr[σ1−sA∗σsB]ds =Z 1 0 Tr[σA∗∆s σB]ds = hA, Bif0 , where f0(t) =Z 1 0 tsds = t − 1 log t . The BKM inner product arises naturally in statistical mechanics and our work here as follows: For a self-adjoint operator h on a finite-dimensional Hilbert space H, and β > 0, define the Gibbs state σβ by σβ = 1 Tr[e−βh] e−βh . The free energy is the functional F(β, h) = β−1 log(Tr[e−βh]). A simple calculation using Duhamel's formula shows that for any self-adjoint A, d2 ds2F(β, h + sA)(cid:12)(cid:12)(cid:12)(cid:12)s=0 = β(cid:2)hA, AiBKM − (Tr[σβA])2(cid:3) . In particular, up to a factor of β, hA, AiBKM arises as the restriction of the Hessian of F to the space of self-adjoint matrices A satisfying Tr[σβA] = 0. It is well known that for fixed β, the function h 7→ F(β, h) is the Legendre transform of the von Neumann entropy S(ρ) on the set of density matrices. Since the gradients and Hessians of conjugate convex functions are inverse to one another, the Hessian of the free energy is the inverse of Hessian of the entropy; a fact that explains why the BKM inner product will arise naturally in the study of gradient flows of the relative entropy. 3. Generators of quantum Markov semigroups satisfying detailed balance Since a QMS Pt = etL on A that satisfies the σ-DBC for some σ ∈ S+(A) has a generator L that commutes with the modular operator ∆σ, and since ∆σ is positive with respect to the GNS inner product, ∆σ and L can be simultaneously diagonalized. In the case A = Mn(C), the diagonalization of ∆σ reduces immediately to the diagonalization of σ: Let σ = e−h be a density matrix on Cn. Let {η1, . . . , ηn} be an orthonormal basis of Cn consisting of eigenvectors of h = − log σ: hηj = λjηj. For α = (α1, α2) ∈ {(i, j) : 1 ≤ i, j ≤ n}, define numbers ωα (called the Bohr frequencies) by ωα = λα1 − λα2 , (3.1) and rank-one operators Fα given by Fα = ηα1ihηα2 where, using a standard physics notation, for η, ξ ∈ Cn, ηihξ is the rank-one operator sending ζ to hξ, ζiCn η. Evidently (3.2) F ∗α = Fα′ where α′ = (α2, α1) . ∆σFα = e−ωαFα and 10 ERIC A. CARLEN AND JAN MAAS Alicki [4] exploited such a construction to show that for σ ∈ S+ with non-degenerate spectrum (in the sense that each eigenvalue of σ is simple), the generator L of a quantum Markov semigroup on Mn(C) that satisfies the σ-DBC (so that L commutes with ∆σ) has the form described in (3.3) below; see [4, Theorem 3]. (Alicki actually considered the more general case that L is normal with respect to the σ-GNS inner product. His formula reduces to (3.3) when L is self-adjoint.) The hypothesis that σ has non-degenerate spectrum turns out to unnecessary. In the context of full matrix algebras this has been shown in [41] for the alternate canonical form given in [4, Theorem 3]. It is possible to give a simple and self-contained proof of Alicki's Theorem in a somewhat more general setting. This is done in Appendix A; the theorem proved there is: 3.1. THEOREM. Let Pt = etL be a QMS on a unital C∗-subalgebra A of Mn(C). Suppose that Pt satisfies the σ-DBC for σ ∈ S+(A) and that Pt has an extension cPt to a QMS on Mn(C). Regard the modular operator ∆σ as an operator on Mn(C) and let τ denote the normalized trace on Mn(C). Then the generator L of Pt has the form L A =Xj∈J(cid:16)e−ωj /2V ∗j [A, Vj] + eωj /2[Vj, A]V ∗j(cid:17) e−ωj/2(cid:16)V ∗j [A, Vj] + [V ∗j , A]Vj(cid:17) =Xj∈J where ωj ∈ R for all j ∈ J , and {Vj}j∈J is a set in Mn(C) with the properties: (i) τ [V ∗j Vk] = δj,k for all j, k ∈ J . (ii) τ [Vj] = 0 for all j ∈ J . (iii) {Vj}j∈J = {V ∗j }j∈J . (iv) {Vj}j∈J consists of eigenvectors of the modular operator ∆σ with (3.3) (3.4) (3.5) Conversely, given any σ ∈ S+(A), and any set {Vj}j∈J satisfying (iii) and (iv) for some {ωj}j∈J ⊆ R, the operator L given by (3.3) is the generator of a QMS Pt that satisfies the σ-DBC. ∆σVj = e−ωj Vj . 3.2. Remark. The eigenvectors of ∆σ with eigenvalues other than 1 cannot be self-adjoint on account of (1.3). However, when σ is the normalized trace, ∆σ is the identity, so that each Vj is an eigenvector of ∆σ with eigenvalue 1, thus ωj = 0. It is then possible to take each Vj to be self-adjoint, so that (3.4) reduces to L A = −Xj∈J [Vj, [Vj, A]] . (3.6) This formulation arises naturally in various applications, as we shall see in Section 6. 3.3. Remark. By Theorem 3.1, the Hilbert-Schmidt adjoint of L is given by L †ρ = Xj∈J(cid:16)e−ωj /2[Vjρ, V ∗j ] + eωj/2[V ∗j , ρVj](cid:17) e−ωj/2(cid:16)[Vjρ, V ∗j ] + [Vj, ρV ∗j ](cid:17) . = Xj∈J (3.7) 3.4. Remark. Note that the operators in {Vj}j∈J need not belong to A itself. The Fermi Ornstein-Uhlenbeck semigroup in the Clifford algebra with an odd number of generators provides an example in which they do not, as we discuss shortly. GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 11 3.5. Remark. By properties (i) and (ii), the index set J has cardinality J no greater than n2 − 1. While in general the set {Vj}j∈J is not uniquely determined, the proof in the appendix shows that J is uniquely determined. Moreover, if {Vj}j∈J and { Vj}j∈J are two such sets, there is an J × J unitary matrix Uj,k such that for all j ∈ J , Vj =Xk∈J Uj,kVk, and such that unless ωk = ωj, Uk,j = 0. Thus, in a strong sense, the sets {Vj}j∈J and {ωj}j∈J are canonically associated to L . (Indeed, since ∆σVj = e−ωj Vj, the numbers {ωj}j∈J are fixed once σ and the set {Vj}j∈J is fixed.) 4. Restriction to commutative subalgebras Let A be a unital C∗-algebra, and let σ ∈ S+(A). Let A1, A2 be eigenvectors of ∆σ: ∆σ(Aj) = λjAj, j = 1, 2. Then λ1, λ2 > 0, and since the modular operator is an automorphism, ∆σ(A1A2) = ∆σ(A1)∆σ(A2) = λ1λ2A1A2 . That is, the product of eigenvectors of ∆σ is again an eigenvector of ∆σ, and moreover, the eigenspace of ∆σ for the eigenvalue 1 is an algebra. In fact, by (1.3) it is a ∗-algebra, and it consists exactly of those elements A ∈ A that commute with σ. Clearly, σ itself always belongs to Aσ. 4.1. DEFINITION (Modular subalgebra). Let A be a unital C∗-algebra, and let σ ∈ S+(A). The σ-modular subalgebra of A, denoted Aσ, is the C∗-subalgebra of A consisting of the eigenspace of ∆σ with eigenvalue 1. Of course if σ = 1, ∆σ is the identity on A, and then Aσ = A. On the other hand, suppose that A = Mn(C) and let {η1, . . . , ηn} be an orthonormal basis of Cn consisting of eigenvectors of σ with σηj = e−λj ηj for j = 1, . . . , n. Then for each 1 ≤ i, j ≤ n, ηiihηj is an eigenvector of ∆σ with eigenvalue eλj−λi. If the numbers {λ1, . . . , λm}, which are the eigenvalues of the modular generator h, are all distinct, then the eigenspace of ∆σ for the eigenvalue 1 is exactly the span of the set {ηjihηj}n j=1 [4]. In this case Aσ is an n-dimensional commutative C∗-subalgebra of A. Let Pt = etL be an ergodic QMS on A, and let σ ∈ S+ be its unique invariant state. Suppose that Pt satisfies the σ-DBC. Since L and ∆σ commute by Theorem 2.9, the σ-modular subalgebra Aσ of A is invariant under Pt (and under P†t as well). In this c case, let P t denote the restriction of Pt to Aσ. If Aσ is commutative, then P t is an ergodic QMS on a commutative subalgebra of Mn(C), which may be identified with the transition semigroup of a classical Markov chain. This happens whenever each eigenvalue of σ is simple [4]. Therefore, consider a QMS Pt = etL on Mn(C) that satisfies the σ-DBC. Suppose that A is a unital commutative C∗-subalgebra of Mn(C) that is invariant under P†t . If Pt is ergodic, it follows using the self-adjointness of L with respect to h·,·i1, that for any ρ ∈ S+(A), σ = limt→∞ P†t ρ, and hence σ ∈ A. jections in A such thatPm In our finite-dimensional setting, there exists a finite set {E1, . . . , Em} of minimal pro- k=1 Ek = 1. Consequently, EjEk = 0 for all j 6= k, and A is the span of {E1, . . . , Em}: For all A ∈ A we have c A = mXk=1 ak Tr[Ek] Ek where ak = Tr[EkA] . (4.1) 12 ERIC A. CARLEN AND JAN MAAS Define an m × m matrix Q by Qk,ℓ = 1 Tr[Ek] Tr[EkL Eℓ] . (4.2) k=1 ρk = 1. A vector ~ρ = (ρ1, . . . , ρm) ∈ Rm is a probability vector in case ρk ≥ 0 for all k, and Pm 4.2. THEOREM. Let Pt = etL be an ergodic QMS on Mn(C) that satisfies the σ-DBC for its invariant state σ. Let A be a unital commutative C∗-subalgebra of Mn(C) that is invariant under P†t . Let {E1, . . . , Em} be a set of minimal projections in A such that Pm k=1 Ek = 1. The m × m matrix Q defined by (4.2) specifies an ergodic continuous-time Markov chain on {1, . . . , m} with jump rates Qk,ℓ from k to ℓ. The corresponding forward equation, governing the evolution of site occupation probabilities is d dt ρℓ(t) = mXk=1(cid:0)ρk(t)Qk,ℓ − ρℓ(t)Qℓ,k(cid:1) . (4.3) A time-dependent probability vector ~ρ(t) satisfies (4.3) if and only if the time-dependent state ρ(t) on A given by ρ(t) = Ek (4.4) ρk(t) Tr[Ek] mXk=1 d dt ρ(t) = L †ρ(t). Moreover, the probability vector ~σ given by σk = Tr[σEk] for satisfies k = 1, . . . , m is the unique invariant probability vector for the Markov chain, and the classical detailed balance condition is satisfied. σkQk,ℓ = σℓQℓ,k (4.5) k 6= ℓ, which makes it a transition rate matrix. Proof. We first show that the matrix Q satisfiesPm Since 0 = L 1 =Pm Then for k 6= ℓ, simple computations yield ℓ=1 L Eℓ, we havePm Tr[EkL Eℓ] = 2Xj∈J e−ωj /2Tr[EkV ∗j EℓVj]. ℓ=1 Qk,ℓ = 0 and that Qk,ℓ ≥ 0 for all ℓ=1 Qk,ℓ = 0. Let L be given in the form (3.4). Since Ek and V ∗j EℓVj are positive, Tr[EkV ∗j EℓVj] ≥ 0, showing that Qk,ℓ ≥ 0 for all k 6= ℓ. Now suppose that ak := Pℓ6=k Qk,ℓ = 0. Then from the definition, Tr[L †(Ek)Eℓ] = Tr[EkL (Eℓ)] = 0 for all k 6= ℓ, and then, sincePm ℓ=1 Qk,ℓ = 0, also Tr[L †(Ek)Ek] = 0. It would follow that L †(Ek) = 0. Since Pt is ergodic, this is impossible unless A is spanned by 1, a trivial case. Hence we may proceed assuming that ak > 0 for all k. Define the matrix Pk,ℓ by 1 ak 0 Pk,ℓ = Qk,ℓ ℓ 6= k ℓ = k . Evidently P is an m × m stochastic matrix. Define the m × m matrix M by Mk,ℓ = akδk,ℓ. Then Q = M (P − 1m) where 1m is the m × m identity matrix. The equation d ~ρ0, and it gives the dt site occupation probabilities for a continuous time Markov chain in which the jump rate ~ρ(t) = Q†~ρ(t) is solved in terms of the initial data ~ρ0 by ~ρ(t) = etQ† GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 13 for leaving site k is ak, and when such a jump occurs, the probability that site ℓ is the new site occupied is given by Pk,ℓ. The equation d dt ~ρ(t) = Q†~ρ(t) can be written as d dt ρℓ(t) = mXk=1 ρk(t)Qk,ℓ = mXk=1(cid:0)ρk(t)Qk,ℓ − ρℓ(t)Qℓ,k(cid:1) , (4.6) where the second equality follows from the first and the fact that Qℓ,ℓ = −Pk6=ℓ Qℓ,k. To show that this Markov chain satisfies the classical detailed balance condition (4.5), This gives us (4.3). observe that σEk = σk(Tr[Ek])−1Ek = Ekσ by (4.1), and therefore σkQk,ℓ = σk Tr[Ek] Tr[EkL (Eℓ)] = Tr[σEkL (Eℓ)] = Tr[σL (Ek)Eℓ] = Tr[EℓσL (Ek)] = σℓ Tr[Eℓ] Tr[EℓL (Ek)] = σℓQℓ,k , where the third equality holds since L is self-adjoint for the σ-GNS inner product. It follows immediately from (4.6) that ~σ is invariant. Finally, it is easy to check using the definition of the matrix Q in terms of the generator dt ρ(t) = L †ρ(t). Ergodicity of the Markov chain L that ~ρ(t) satisfies (4.6) if and only if d P t now follows from the ergodicity of Pt. (cid:3) c 5. Dirichlet form representation associated to a quantum Markov generator Let H = L2(X, B, µ) be the space of square integrable real-valued functions on some probability space (X, B, µ). A closed and densely defined, symmetric non-negative bilinear form E on H defines a non-negative unbounded operator −A through E(f, g) = −hf, AgiH. A special case (µ is a probability measure) of a theorem of Beurling and Deny [7, 8] states that Pt = etA is a Markov semigroup if and only if for all f ∈ H, E(bf ,bf ) ≤ E(f, f ) where bf denotes the projection of f onto the closed convex set {g ∈ H : 0 ≤ g ≤ 1 a.e.}. A powerful non-commutative extension of this theory has been developed starting with the early work of Gross [31, 32], and continuing with [3, 15, 16, 18, 21, 29]. We shall not need the whole theory at present, but the Dirichlet form representation of the generator of a QMS will be useful to us. Let Pt = etL be a QMS on A that satisfies the σ-DBC for some σ ∈ S+(A). The generator L can then be written in the canonical form specified in (3.4) of Theorem 3.1. Throughout the rest of this section we fix such a generator L , and the sets {Vj}j∈J and {ωj} that specify L in the form (3.4). Define operators ∂j on A by ∂jA = [Vj, A] so that ∂†j A = [V ∗j , A] . The operators ∂j are derivations, and we may consider them as non-commutative analogs of partial derivatives associated to L . With respect to the Hilbert-Schmidt in- ner product, we may then form non-commutative analogs of the gradient, divergence and Laplacian associated to L . We begin with the Laplacian: Given the set {Vj}j∈J , we define an operator L0 on HA by L0A = −Xj∈J ∂†j ∂jA = −Xj∈J [V ∗j , [Vj, A]] . 14 ERIC A. CARLEN AND JAN MAAS Evidently L †0 = L0, and we may write L0A =Xj∈J (V ∗j [A, Vj ] + [Vj, A]V ∗j ). Thus, by Theorem 3.1, L0 is the generator of a quantum Markov semigroup P0,t = etL0 satisfying detailed balance with respect to τ . We call this semigroup the heat semigroup associated to Pt = etL , and the operator L0 the Laplace operator associated to L . We define the Hilbert space HA,J by HA,J =Mj∈J (j) H A , (j) where each H A (j) . Thus, picking some linear ordering of J , we can write of A in H A is a copy of HA. For A ∈ HA,J and j ∈ J , let Aj denote the component We equip HA,J with the usual inner product hA, BiHA,J =Pj∈J hAj, BjiHA. Define an operator ∇ : HA → HA,J by A = (A1, . . . , AJ ) . ∇A = (∂1, . . . , ∂J A) . Thinking of elements of A as non-commutative analogs of functions on a manifold, we may think of A = (A1, . . . , AJ ) as a vector field. This point of view will be justified in the next section. We define the operator div : HA,J → HA by [Aj, V ∗j ] . div A = −Xj∈J ∂†j Aj =Xj∈J Note that div is minus the adjoint of the map ∇ : HA → HA,J , so that L0 is negative semi- definite. With these definitions, L0 = div ◦∇. We call ∇ the non-commutative gradient associated to L , and div the non-commutative divergence associated to L . 5.1. Remark. In this finite-dimensional setting, by elementary linear algebra, (cid:0)Null(div)(cid:1)⊥ = Ran(∇) . In the terminology introduced above, elements of Null(div) are divergence free vector fields. Then (5.1) says that a vector field A is a gradient if and only if it is orthogonal in HA,J to every divergence free vector field. The differential structure introduced above allows us to write the generator L of a QMS in terms of a non-commutative Dirichlet form. 5.2. LEMMA. For all s ∈ [0, 1], all j ∈ J , and all A, B ∈ A we have h∂jB, Ais = hB, esωj (e−ωj V ∗j A − AV ∗j )is . Proof. For any A, B ∈ Mn(C), h∂jB, Ais = Tr[σs(∂jB)∗σ1−sA] = Tr[σs(VjB − BVj)∗σ1−sA] σ = Tr[σsB∗V ∗j σ1−sA] − Tr[σsV ∗j B∗σ1−sA] σ(V ∗j )σsB∗σ1−sA] (V ∗j )A] − Tr[∆s = Tr[σsB∗σ1−s∆s−1 = e(s−1)ωj Tr[σsB∗σ1−sV ∗j A] − esωj Tr[V ∗j σsB∗σ1−sA] = (cid:10)B, esωj (e−ωj V ∗j A − AV ∗j )(cid:11)s , where in the fourth line we have used (3.5). (5.1) (5.2) (cid:3) GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 15 It follows from (5.2) that for all s ∈ [0, 1], and all A, B ∈ A, e(1/2−s)ωjh∂jB, ∂jAis = −(cid:10)B, e−ωj/2V ∗j [A, Vj ] + eωj /2[Vj, A]V ∗j(cid:11)s . Using the expression for L given in (3.3), we obtain Es(B, A) = −hB, L Ais , where In particular, taking s = 1/2, we see that E1/2(B, A) = −hB, L Ai1/2 , where Es(B, A) :=Xj∈J E1/2(B, A) :=Xj∈J e(1/2−s)ωjh∂jB, ∂jAis . h∂j B, ∂jAi1/2 , (5.3) which expresses L in terms of a Dirichlet form in the sense of [29, 15, 16]. As a simple consequence of the Dirichlet form representation, we state an ergodicity result. Recall that a QMS Pt = etL is ergodic in case for each t > 0, the 1-eigenspace of Pt is spanned by the identity, or, what is the same thing, the 0-eigenspace of L is spanned by the identity. We refer to [28] for an early study of ergodicity for quantum dynamical semigroups. 5.3. THEOREM. Let Pt = etL be QMS on A that satisfies the σ-DBC for σ ∈ S+(A). Let L be given in the form (3.4). Then the commutant of {Vj}j∈J equals the null space of L . In particular, Pt is ergodic if and only if the commutant of {Vj}j∈J is spanned by the identity. Proof. Suppose that A belongs to the commutant of {Vj}j∈J . By definition, this means that ∂jA = 0 for all j ∈ J , and therefore A ∈ Null(L ) by (3.3). Conversely, if L A = 0, then by (5.3), 0 = −hA, L Ai1/2 =Xj∈J h∂jA, ∂j Ai1/2 , which is the case if and only if [Vj, A] = 0 for all j. This means that A belongs to the commutant of {Vj}j∈J . (cid:3) 5.4. THEOREM. Let Pt = etL be an ergodic QMS on A that satisfies the σ-DBC for σ ∈ S+(A), and let L0 be the associated Laplacian. Then for given B ∈ HA, the equation L0X = B has a solution if and only if τ [B] = 0. Consequently, when τ [B] = 0, there is a non-trivial affine subspace of HA,J consisting of elements A for which div A = B. Proof. Since hA, L0AiHA = −h∇A,∇AiHA,J , we have Null(L0) = Null(∇). Since Pt is ergodic, it follows from Theorem 5.3 that Null(∇) is spanned by 1. Since L0 is self-adjoint on HA, the assertion now follows from the Fredholm alternative. (cid:3) The following identity will be useful going forward. 5.5. LEMMA (Chain rule identity). For all V ∈ Mn(C), ρ ∈ S+ and ω ∈ R, ρ(cid:16)V log(e−ω/2ρ) − log(eω/2ρ)V(cid:17) ds = e−ω/2V ρ − eω/2ρV . (5.4) Z 1 0 eω(s−1/2)Rρ∆s 16 ERIC A. CARLEN AND JAN MAAS Proof. Define f (s) = eω(1/2−s)ρ1−sV ρs. The right side of (5.4) equals f (1) − f (0) and f′(s) = eω(1/2−s)ρ1−s(cid:16) − ωV − log(ρ)V + V log(ρ)(cid:17)ρs = eω(1/2−s)ρ1−s(cid:16)V log(e−ω/2ρ) − log(eω/2ρ)V(cid:17)ρs . 0 f′(1 − s) ds, which yields the result. 5.6. Remark. Consider the function fω defined by Thus the left side of (5.4) equalsR 1 fω(t) :=Z 1 0 Then (5.4) can be formulated as eω(s−1/2)ts ds = eω/2 t − e−ω log t + ω . Rρfω(∆ρ)(cid:16)V log(e−ω/2ρ) − log(eω/2ρ)V(cid:17) = e−ω/2V ρ − eω/2ρV . Notice that for ω = 0, (5.4) reduces to the commutator identity Rρf0(∆ρ)([V, log ρ]) = [V, ρ] . (cid:3) (5.5) (5.6) (5.7) This identity provides a quantum analog of the classical identity for smooth, strictly positive probability densities ρ(x) on Rn: ρ(x)∇ log ρ(x) = ∇ρ(x) . (5.8) To see this, note that if A commutes with ρ, fω(∆ρ)A = A so that for each ω, the operation A 7→ Rρfω(∆ρ)A is one of the non-commutative interpretations of multiplication of A by ρ. Now applying (5.7) with V = Vj, we have Rρf0(∆ρ)(∂j log ρ) = ∂jρ for each j, which yields a quantum analog of (5.8). Lemma 5.5 and the previous remark motivate the following definition: 5.7. DEFINITION. For ρ ∈ S+, and ω ∈ R, define the operator [ρ]ω : Mn(C) → Mn(C) by [ρ]ω = Rρ ◦ fω(∆ρ) (5.9) For each ω, [ρ]ω, which is one of the non-commutative forms of multiplication by ρ, is evidently invertible, and its inverse, [ρ]−1 ω = (1/fω)(∆ρ) ◦ Rρ−1 may then be viewed as the corresponding non-commutative form of division by ρ. We remark that [ρ]−1 ω is a kernel operator that can be used to define a monotone metric on density matrices in the sense of [61]. However, the Riemannian metric on density matrices that we introduce in this paper will be different. 5.8. LEMMA. For all ω ∈ R, the maps ρ 7→ [ρ]ω and ρ 7→ [ρ]−1 Furthermore, for all A, are C∞ on S+. ω ([ρ]ωA)∗ = [ρ]−ωA∗ and consequently ([ρ]−1 ω A)∗ = [ρ]−1 −ωA∗ . (5.10) Proof. Recall the identities λ − µ log λ − log µ 0 λ1−sµs ds = Z 1 [ρ]ω = Rρfω(∆ρ) =Z 1 0 1 0 Z ∞ ds =Z 1 0 which hold for λ, µ > 0. By (5.5) and (5.9) we obtain and (t + λ)(t + µ) dt = log λ − log µ λ − µ , (5.11) eω(1/2−s)Ls ρR1−s ρ (e−ω/2Lρ)s(eω/2Rρ)1−s ds , (5.12) GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 17 and it follows from (5.11) that [ρ]−1 ω =Z ∞ 0 (t + e−ω/2Lρ)−1(t + eω/2Rρ)−1 dt . The fact that ρ 7→ [ρ]−1 ω is C∞ now follows immediately from the resolvent identity, and then the C∞-differentiability of ρ 7→ [ρ]ω is clear. Moreover, (5.10) follows from (5.12). (cid:3) It is now a simple matter to write the quantum heat flow equation ∂ρ/∂t = L †0 ρ in a form that leads directly to its interpretation as gradient flow for the relative entropy with respect to the normalized trace: For all ρ ∈ S+, L †0 ρ = div([ρ]0∇ log ρ) , where [ρ]0 is applied to each component of ∇ log ρ. As we show in the next section, it follows easily from this formula that there is a Riemannian metric on S+, which is a natural analog of the 2-Wasserstein metric, such that the quantum heat flow on S+ associated to {Vj}j∈J is a gradient flow for the relative entropy of ρ with respect to the normalized trace τ . For a comprehensive treatment of the theory of gradient flows with respect to the 2-Wasserstein metric, see [5] and [72, Chapters 23-25]. In fact, the next lemma provides the means to extend this result to the general class of quantum Markov semigroups that satisfy detailed balance with respect to some non- degenerate state σ. 5.9. LEMMA. Let Pt = etL be QMS on A that satisfies the σ-DBC for σ ∈ S+(A). and let L be given in the form (3.4). Then for all ρ ∈ S+, and all j ∈ J , ∂j(log ρ − log σ) = Vj log(e−ωj /2ρ) − log(eωj /2ρ)Vj . Proof. By (3.5) we have ∆s follows that σVj = e−sωj Vj, and thus [Vj, log σ] = −∂ss=0∆s σVj = ωjVj. It ∂j(log ρ − log σ) = [Vj, log ρ] − ωjVj = Vj log(e−ωj /2ρ) − log(eωj /2ρ)Vj , which is the desired identity. (cid:3) 5.10. THEOREM. Let Pt = etL be QMS on A that satisfies the σ-DBC for σ ∈ S+(A), and let L be given in the form (3.4). Then, for all ρ ∈ S+, Proof. Using Lemma 5.9 and (5.6) we obtain Xj∈J −L †ρ =Xj∈J ∂†j(cid:16)[ρ]ωj ∂j(log ρ − log σ)(cid:17) =Xj∈J =Xj∈J = −Xj∈J(cid:16)e−ωj /2[Vjρ, V ∗j ] + eωj /2[V ∗j , ρVj](cid:17) = −L †ρ , ∂†j(cid:16)[ρ]ωj ∂j(log ρ − log σ)(cid:17) . ∂†j(cid:16)[ρ]ωj(cid:16)Vj log(e−ωj /2ρ) − log(eωj /2ρ)Vj(cid:17)(cid:17) ∂†j(cid:16)e−ωj /2Vjρ − eωj /2ρVj(cid:17) where the final identity follows from (3.7). (cid:3) 18 ERIC A. CARLEN AND JAN MAAS 6. Examples 6.1. The infinite-temperature Fermi Ornstein-Uhlenbeck semigroup. As Segal emphasized [65], the Fermion number operator for n degrees of freedom can be represented in terms of the generators of a Clifford algebra. This permits the semigroup it generates to be realized as a QMS, a fact that was effectively exploited by Gross [31, 32] using Segal's non-commutative integration theory [64, 66]. Let {Q1, . . . , Qn} be self-adjoint operators on a finite-dimensional Hilbert space satis- fying the canonical anti-commutation relations (CAR): QjQk + QkQj = 2δjk1 . j=1. Let Γ : Cn → The Clifford algebra Cn is the 2n-dimensional algebra generated by {Qj}n Cn be the principle automorphism on Cn, i.e., the unique algebra isomorphism satisfying Γ(Qj) = −Qj for all j. The product of all of the generators Q1Q2 ··· Qn (in some order) is evidently unitary, and the CAR imply that it commutes with each Qj if n is odd, and anti- commutes with each Qj if n is even. Hence when n is odd, the center of Cn is non-trivial and Cn is not a factor. In the even case n = 2m, form the m self-adjoint unitary operators iQ2j−1Q2j, j = 1, . . . n. These all commute with one another, and we define Evidently W ∈ Cn is unitary and self-adjoint, and since W anti-commutes with each Qj, the principle automorphism is inner and is given by W = imQ1Q2 ··· Q2m . Γ(A) = W AW = W ∗AW = W AW ∗ (6.1) Let {0, 1}n be the set of fermion multi-indices, and for all α = (αj)j ∈ {0, 1}n, define Qα = Qα1 j=1 αj. Let τ be the canonical trace on Cn, determined by τ (Qα) := δ0,α. In a standard representation of Cn as an algebra of operators on (C2)⊗n due to Brauer and Weyl, τ is simply the normalized trace. See [10, 11] for references and further background. for all A ∈ Cn . 1 ··· Qαn Gross [32] defined a differential structure and a Dirichlet form on Cn as follows: For n and α =Pn b ∂j be given byb j ∈ {1, . . . , n}, let b (QjA − Γ(A)Qj) . ∂j is a skew derivation. That is, for all A, B ∈ Cn, ∂j(A) = 1 2 b Each ∂jB. Gross defined the Dirichlet form E(A, B) on Cn by E(A, B) = Pn ∂jB], and defined the operator L on Cn by −τ [A∗L B] = E(A, B). Simple computations show that for all A ∈ Cn,b ∂jA)B + Γ(A) ∂j(AB) = ( j=1 τ [( ∂jA)∗ b b b ∂†A = 1 2(cid:0)QjA + Γ(A)Qj(cid:1) and L A = 1 2 nXj=1(cid:0)QjΓ(A)Qj − A(cid:1) . (6.2) b (6.3) It is evident that L Qα = −αQα, hence −L is the fermion number operator. self-adjoint and unitary. Then, using the fact that V 2 When n is even, so that Γ(A) = W ∗AW , we define Vj = iW Qj, so that each Vj is both j = 1, we may rewrite (6.3) as L A = 1 2 nXj=1(cid:0)VjAVj − A(cid:1) = − 1 4 [Vj, [Vj, A]] , nXj=1 GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 19 which has the form (3.6). Thus, etL is a QMS satisfying the τ -DBC. Gross discussed this QMS as a fermionic analog of the classical Ornstein-Uhlenbeck semigroup; we refer to it as the infinite temperature Fermi Ornstein-Uhlenbeck semigroup. One good reason is that it is generated by the negative of the fermionic number operator. Another is that as conjectured by Gross [32] and proved in [10], it has the same optimal hypercontractivity properties as the classical Ornstein-Uhlenbeck semigroup. We shall further develop the analogy here. (The infinite temperature part of the name will be justified in the next example.) To relate the differential structure in (6.2) to the one considered here, note that when n is even, so that Γ is the inner automorphism given by (6.1), we have, with Vj defined as above,b b ∂jA = 1 2i W [Vj, A] = 1 2i W ∂jA and ∂† jA = − 1 2i ∂†j (W A) = − 1 2i ∂j(W A) (6.4) for j = 1, . . . , n. Thus, Gross's differential structure in terms of skew derivations may be substituted with the present differential structure in terms of derivations when n is even. When n is odd, this is achieved by embedding Cn is an a larger Clifford al- gebra: One can add one more generator, or, perhaps better, embed the Clifford alge- bra generated by {Q1, . . . , Qn}, in the phase space Clifford algebra with 2n generators {Q1, . . . , Qn, P1, . . . , Pn} that is discussed next. 6.2. The finite-temperature Fermi Ornstein-Uhlenbeck semigroup. Let A be the Clifford algebra Cn of dimension n = 2m for some m ∈ N. Consider a set of generators {Q1, . . . , Qm, P1, . . . , Pm} , where QjQk + QkQj = PjPk + PkPj = 2δj,k1 and QjPk + PkQj = 0 for all 1 ≤ j, k ≤ m . We think of A as the full set of phase space observables, the subalgebra generated by {Q1, . . . , Qm} as the algebra of configuration space observables, and the subalgebra gener- ated by {P1, . . . , Pm} as the algebra of momentum space observables. Form the operators Zj = 1 √2 (Qj + iPj) so that Z∗j = 1 √2 (Qj − iPj) . It is easy to check that ZjZk + ZkZj = 0 Consider the complementary orthogonal projections Nj and N⊥j defined by and ZjZ∗k + Z∗k Zj = 2δj,k1 for all 1 ≤ j, k ≤ m . 1 2 Then (6.5) implies that Nj = Z∗j Zj and N⊥j = 1 2 ZjZ∗j for all 1 ≤ j ≤ m . ZjNj = N⊥j Zj = Zj and NjZj = ZjN⊥j = 0 . Note also that ZjNk = NkZj and ZjN⊥k = N⊥k Zj for all j 6= k . (6.5) (6.6) (6.7) Moreover, {N1, . . . , Nm, N⊥1 , . . . , N⊥m} is a set of commuting orthogonal projections. For each j, QjPj commutes with both Qk and Pk for all k 6= j. Hence the operators {Q1P1, . . . , QmPm} all commute with one another. As in the previous example, let W = 20 ERIC A. CARLEN AND JAN MAAS imQm j=1 QjPj so that W is self-adjoint and unitary, and for all A ∈ A, let Γ(A) = W AW . Note that QjPjZj = iZj for each j. For any set of m real numbers {e1, . . . , em}, and any parameter β ∈ (0,∞), to be interpreted as the inverse temperature, define the free Hamiltonian h and the Gibbs state σβ by h = mXj=1 ejNj and σβ = 1 τ [e−βh] e−βh . where τ is the canonical trace as in Section 6.1. Since the Nj are commuting orthogonal projections, e−βh is the product, in any order, of the operators e−βej Nj + N⊥j . Therefore, for each 1 ≤ j ≤ m, ∆σβ (Zj) = (e−βej Nj + N⊥j )Zj(eβej Nj + N⊥j ) = eβej Zj , where we have used (6.6) and (6.7). Consequently, ∆σβ (Z∗j ) = e−βej Z∗j . Since W com- mutes with every even element of A, it follows that ∆σβ (W Zj) = eβej W Zj and ∆σβ (Z∗j W ) = e−βej Z∗j W . Define the operators Vj = W Zj, 1 ≤ j ≤ m , 2 V ∗j Vj = Nj and 1 so that 1 2 VjV ∗j = N⊥j . Then {V1, . . . , Vm, V ∗1 , . . . , V ∗m} is set of operators on A satisfying the conditions (i), (ii), (iii) and (iv) of Theorem 3.1. Therefore, the operator Lβ defined by LβA = 1 4 mXj=1heβej/2(cid:16)V ∗j [A, Vj] + [V ∗j , A]Vj(cid:17) + e−βej /2(cid:16)Vj[A, V ∗j ] + [Vj, A]V ∗j(cid:17)i (6.8) is the generator of a QMS Pt = etLβ that satisfies the σβ-DBC. It is a simple matter to diagonalize Lβ: For each 1 ≤ j ≤ m, define the four operators Kj,(0,0) = 1 , Kj,(1,0) = Zj , Kj,(0,1) = Z∗j . and Kj,(1,1) = eβej /2Nj − e−βej/2N⊥j One readily checks that this set of four operators is orthonormal in any of the inner products h·,·is based on σβ. Using the fact that for each j, Vj and V ∗j commute with Pk and Qk for all k 6= j, and using the identities VjKj,(1,1) = eβej /2Vj and Kj,(1,1)Vj = −e−βej/2Vj , we readily compute that LβZj = − cosh(βej/2)Zj Therefore, for all 0 ≤ k, ℓ ≤ 1, and LβKj,(1,1) = −2 cosh(βej/2)Kj,(1,1) . (6.9) LβKj,(k,ℓ) = −(k + ℓ) cosh(βej /2)Kj,(k,ℓ) . Let α = (α1, . . . , αm) denote a generic element of the index set {{0, 1}×{0, 1}}m, and for α = (k, ℓ) ∈ {0, 1} × {0, 1}, define α = k + ℓ. Then the functions Kα := K1,α1K2,α2 ··· Km,αm are an orthogonal (but not normalized) basis for HA consisting of eigenvectors of Lβ: LβKα = − mXj=1 αj cosh(βej/2) Kα . (6.10) GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 21 It is now easy to check that in the infinite temperature limit (i.e., β → 0), limβ→0 Lβ = L0 where L0A = − = 1 2 1 4 mXj=1(cid:0)[Vj, [V ∗j , A]] + [V ∗j , [Vj, A]](cid:1) mXj=1(cid:0)QjΓ(A)Qj + PjΓ(A)P ∗j − 2A(cid:1) , From the previous example, we recognize L0 as the negative of the number operator on A = Cn. That is, in the infinite temperature limit (β → 0), we recover the infinite temperature Fermi Ornstein-Uhlenbeck semigroup, justifying our nomenclature. b As in the infinite temperature case, there is a differential calculus that is more closely adapted to Lβ: For 1 ≤ j ≤ m, define the operators ∂jA = ∂jA = W [Vj, A] b W [V ∗j , A] and (Z∗j A − Γ(A)Z∗j ) = − (ZjA − Γ(A)Zj ) = 1 2 1 2 1 2 1 2 We readily compute that b b and that b ∂jKj,(0,0) = b b b b b ∂jKj,(0,0) = ∂jKj,(1,0) = 0 , ∂jKj,(0,1) = Kj,(0,0) and ∂jKj,1,1 = cosh(βej /2)Kj,(1,0) , (6.11) (6.12) (6.13) (6.14) ∂jKj,(0,1) = 0 , ∂jKj,(1,0) = Kj,(0,0) b and b ∂jKj,(1,1) = − cosh(βej /2)Kj,(0,1) . Again using the fact that for each j, Vj and V ∗j commute with Pk and Qk for all k 6= j, ∂j on all of A. The orthonormal basis {Kα} may be one determines the effect of viewed as consisting of analogs of multivariate Krawtchouck polynomials – the discrete ∂j and b b analogs of the Hermite polynomials. The differential operators ∂j, which are skew derivations as in the infinite temperature case, have the advantage over the closely related derivations ∂jA = [Vj, A] and ∂jA = [V ∗j , A] that they always lower the "degree" of any Kα by one, as one would expect. The operators ∂jA and ∂jA do not do this. ∂j and Using (6.11) and (6.12) one readily deduces the identities, valid for all and ∂jLβKα − Lβ ∂jKα = − cosh(βej /2) ∂jKα ∂j LβKα − Lβ ∂jKα = − cosh(βej/2) ∂ jKα . Finally, we observe that each of the vectors Kα is an eigenvector of ∆σβ . Moreover, it is easy to see that if {e1, . . . , em} is linearly independent over the integers, then ∆σβ Kα = Kα if and only if for each k, αk 6= 1. The span of the set of such Kα is the same as the span of {N1, N⊥1 , . . . , Nm, N⊥m} . (6.15) Hence in this case, the modular algebra Aσβ is the algebra generated by the commuting projections in (6.15). Let us denote this algebra, which does not depend on β, by B. While it need not be the modular algebra when {e1, . . . , em} is not linearly independent over the integers, it is easy to see (by continuity or computation) that it is always invariant under Pt. b b b b b b 22 ERIC A. CARLEN AND JAN MAAS The projections in (6.15) are not minimal in B, but the set of the 2m distinct non-zero products one can form from them is a full set of minimal projections. We may identify this c set with the discrete hypercube Qm = {0, 1}m. Set J = {1, . . . , m}, and let sj : Qm → Qm define the j-th coordinate swap defined by sj(x1, . . . , xm) = (x1, . . . ,−xj, . . . , xm). Let x denote a generic point of Qm. Define Ex = P t j (N⊥j )1−x1. The restriction N x1 For a standard representation in which the elements of A operate on C2m of Pt to B is a nearest neighbor random walk on Qm with transition rates that are readily computed using Theorem 4.2. , and τ is the normalized trace, each Ex is rank one, so that the transition rate matrix D defined in (4.2) is simply Dx,x′ = Tr[Ex, L Ex′]. Using (6.9) through (6.10), one readily computes that Dx,x′ = 0 unless x′ = sj(x) for some j, and in that case mYj=1 Dx,x′ = 2 cosh(βej ) 1 + e−βej 2 cosh(βej ) 1 + eβej xj = 1 xj = 0 , c and this gives the jump rates along the edges of Qm for the classical Markov chain corre- sponding to P t. 6.3. The Bose Ornstein-Uhlenbeck semigroup. A set {Q1, . . . , Qm, P1, . . . , Pm} of self-adjoint operators on a Hilbert space H is a representation of the Canonical Commu- tation Relations (CCR), in case for all 1 ≤ j, k ≤ m, [Qj, Qk] = 0 , [Pj, Pk] = 0 and [Qj, Pk] = iδj,k1 . All representations of the CCR are necessarily infinite-dimensional, since otherwise we would have Tr[[Qj, Pj]] = 0 which is incompatible with [Qj, Pj] = i1. The CCR algebra is the C∗-algebra generated by the unitaries {e−t1Q1, . . . , eitmQm, e−s1P1, . . . , eismPm} for all t1, . . . , tm, s1, . . . , sm, or, what is the essentially the same thing, the Weyl operators. Not only is H necessarily infinite-dimensional, but the operators {Q1, . . . , Qm, P1, . . . , Pm} are unbounded. Therefore, the CCR algebra for m Bose degrees of freedom lies outside the scope of the theory being developed in this paper. However, even without fully extending this theory to infinite dimensions, we shall be able to deduce new results for an important QMS on A, namely the Bose Ornstein-Uhlenbeck semigroup. To keep things simple in this excursion into the infinite-dimensional case, we take m = 1. Exactly as in the Fermi case, we form the operators Z = 1 √2 (Q + iP ) so that Z∗ = 1 √2 (Q − iP ) . It is easy to check that [Z, Z∗] = 1 . (6.16) In one standard representation that we may as well fix here, H = L2(R, γ(x)dx) where γ(x) = (2π)−1/2e−x2/2 and Z = ∂/∂x. Then a simple computation shows that Z∗ = x− ∂/∂x, and (6.16) is satisfied. Define the Hamiltonian h by h = Z∗Z. It is evident that for each k, the linear space of polynomials in x of degree at most k is invariant under h. Since h is self-adjoint, this means the eigenfunctions of h are orthogonal polynomials in H, and hence are the Hermite polynomials. It is well known and easy to check that the kth Hermite polynomial is an eigenfunction of h with eigenvalue k. That is, h is the Bose GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 23 number operator (for one degree of freedom). Fixing an inverse temperature β ∈ (0,∞), we define σβ as σβ =(cid:16)Tr(cid:2)e−βh(cid:3)(cid:17)−1 e−βh . Note that Tr(cid:2)e−βh(cid:3) is finite by what we have said concerning the spectrum of h. One readily finds that [Z, h] = Z, which is the differential version of the identity ∆σβ (Z) = eβZ . It follows that Z and Z∗ are eigenfunctions of the modular operator. (Note that since they are unbounded, they do not belong to the CCR algebra, and are only affiliated to its von Neumann algebra closure.) Define V1 = Z and V2 = Z∗. Then {V1, V2} is a set of operators satisfying conditions (iii) and (iv) of Theorem 3.1 (with ω1 = −β and ω2 = β), but not conditions (i) and (ii), since in the infinite-dimensional case, it is in general too much to ask that the Vj be trace-class. (However, since V1, V2 and 1 are eigenvectors of ∆σ with distinct eigenvalues there is a natural sense in which they are orthogonal so that a natural analog of (i) and (ii) is valid.) In any case, we may define LβA = 1 2heβ/2(cid:16)Z∗[A, Z] + [Z∗, A]Z(cid:17) + e−β/2(cid:16)Z[A, Z∗] + [Z, A]Z∗(cid:17)i 2{ZZ∗, A}(cid:1) , = eβ/2(cid:0)Z∗AZ − 1 2{Z∗Z, A}(cid:1) + e−β/2(cid:0)ZAZ∗ − 1 (6.17) where {A, B} denotes the anti-commutator AB + BA. The operator Lβ is in fact the generator of an ergodic QMS, as shown in [17]. These authors construct the QMS first on the infinite-dimensional analog of HA, which in this case strictly contains A, and then show that the resulting semigroup has the Feller property; i.e., it preserves A. In [17], another detailed balance condition based on self-adjointness with respect to the KMS inner product is used. However, Theorem 2.9 and what we have said above about the modular operator shows that the semigroup also satisfies the σβ-DBC as defined here. In this sense, the example falls into our framework. Simple computations show that for ∂1A = [Z, A] and ∂2A = [Z∗, A], ∂j LβA − Lβ∂jA = − sinh(β/2)∂j A (6.18) for j = 1, 2 and all A in a dense domain of analytic vectors for Lβ that is discussed in [17]. We shall use this identity on this domain later to prove a sharp entropy dissipation inequality for this semigroup, as conjectured in [37, equation (9)]. Note that the corre- sponding formula for the Fermi Ornstein-Uhlenbeck semigroup involves cosh in place of sinh. This reflects the fact that in the Fermi case, taking the infinite temperature limit (β → 0) yields a QMS with a stationary state, while for the Bose Ornstein-Uhlenbeck semigroup, this is not the case: the β → 0 limit cannot be taken in (6.3). The modular generator h has non-degenerate spectrum, and so the modular algebra Aσβ in this case is simply the set of all operators that commute with σβ, which is the same thing as the algebra generated by the spectral projections of h. In particular, Aσβ P t of Pt to is commutative and independent of β. It is easy to see that the restriction Aσβ corresponds, as in Theorem 4.2, to a birth-death process on N. c 24 ERIC A. CARLEN AND JAN MAAS 7. Riemannian metrics and gradient flow Let Pt = etL be QMS on A that satisfies the σ-DBC for σ ∈ S+(A). In this section we define a Riemannian metric on S+ that is determined by L , and for which, as we shall see, the flow given by the dual semigroup P†t , is gradient flow for the relative entropy with respect to σ. Let L be given in the standard form (3.4). Throughout this section, {Vj}j∈J and {ωj}j∈J are fixed, and we assume that Pt is ergodic. Let ρ(t), t ∈ (t0, t1), be any differentiable path in S+ regarded as a convex subset of A. For each t ∈ (t0, t1), let .ρ(t) ∈ A denote the derivative of ρ(t) in t. If ρ(t) is any differentiable path in S+ defined on (−ǫ, ǫ) for some ǫ > 0 such that ρ(0) = ρ0, then Tr[ .ρ(0)] = 0, so that by Theorem 5.4, there is an affine subspace of HA,J consisting of elements A for which .ρ(0) = div A . (7.1) We wish to rewrite (7.1) as an analog of the classical continuity equation for the time evolution of a probability density ρ(x, t) on Rn: ∂ ∂t ρ(x, t) + div[v(x, t)ρ(x, t)] = 0 . In the classical case, for ρ strictly positive, any expression of the form ∂ ∂t ρ(x, t) = div[a(x, t)] (7.2) (7.3) gives rise to (7.2) with v(x, t) = −ρ−1(x, t)a(x, t). Conversely, given (7.2) and defining a(x, t) = −ρ(x, t)v(x, t), (7.3) is satisfied. In the quantum case, there are many different ways to multiply and divide by ρ ∈ S+. Definition 5.7 gives a one-parameter family of ways to multiply A ∈ A by ρ that is In the next definition, we extend this to multiplication of vector fields relevant here. A ∈ HA,J by ρ. 7.1. DEFINITION. Let ~ω ∈ RJ . For ρ ∈ S+ we define the linear operator [ρ]~ω on HA,J by Note that [ρ]~ω is invertible with [ρ]~ω(cid:0)A1, . . . , AJ (cid:1) =(cid:0)[ρ]ω1A1, . . . , [ρ]ωJ AJ (cid:1) . ~ω (cid:0)A1, . . . , AJ (cid:1) =(cid:0)[ρ]−1 AJ (cid:1) . ω1 A1, . . . , [ρ]−1 ωJ [ρ]−1 (7.4) where we have used the fact that Rρ and ∆ρ commute. We are now ready to write (7.1) in the form of a continuity equation: Pick some ~ω ∈ RJ , and define V by V = −[ρ]−1 ~ω A. Then evidently (7.1) becomes .ρ(0) + div([ρ]~ωV) = 0 . (7.5) The vector field A in (7.1) is not unique; however according to Theorem 5.4, the set of such vector fields is an affine space, and thus, in our finite-dimensional setting a closed convex set. It follows immediately that while the vector field V in (7.5) is not unique, the set of such vector fields is a closed affine subspace of ⊕J A, and consequently there is a unique element of minimal norm in ⊕JA for any Hilbertian norm on ⊕JA. We now define the class of Hilbertian norms that is relevant here: GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 25 7.2. DEFINITION. For each ρ ∈ S+, and the given generator L , define an inner product h·,·iL ,ρ on ⊕JA by hW, ViL ,ρ =Xj∈J hWj, [ρ]ωj VjiHA . We write kVkL ,ρ for the corresponding Hilbertian norm. This norm can be viewed as a non-commutative analog of a weighted L2-norm for vector fields. 7.3. THEOREM. Let ρ(t) be a differentiable path in S+ defined on (−ǫ, ǫ) for some ǫ > 0 such that ρ(0) = ρ0. Then there is a unique vector field V ∈ ⊕JA of the form V = ∇U with U ∈ A , for which the non-commutative continuity equation .ρ(0) = − div([ρ0]~ωV) = − div([ρ0]~ω∇U ) (7.6) holds. Moreover, U can be taken to be traceless, and is then self-adjoint and uniquely determined. Furthermore, if W is any other vector field such that .ρ(0) = − div([ρ0]~ωW), then kVkL ,ρ0 < kWkL ,ρ0 . Proof. In view of the discussion above, it remains to show that the unique norm-minimizing vector field V is a gradient. To see this, let A be an arbitrary divergence-free vector field, set W := [ρ0]−1 ~ω A, and Vε := V + εW, so that .ρ(0) + div([ρ0]~ωVε) = 0 for all ε. Since kVkL ,ρ0 < kVεkL ,ρ0 for all ε, it follows that hV, WiL ,ρ0 = 0, and therefore hV, AiHA = 0. This means that V is orthogonal to the set of divergence-free vector fields, hence it is the gradient of some U ∈ A . By subtracting a multiple of the identity, we may take U to be traceless, and then U is uniquely determined, in view of Theorem 5.3 and the ergodicity of Pt. To show that U is self adjoint, define the operator Lρ by LρA = div([ρ~ω]∇A) . (7.7) A direct computation yields Then using (5.10) of Lemma 5.8, LρA =Xj∈J(cid:0)[ρ]ωj (VjA − AVj)(cid:1) V ∗j −Xj∈J (LρA)∗ =Xj∈J(cid:0)[ρ]−ωj (V ∗j A∗ − A∗V ∗j )(cid:1) Vj −Xj∈J V ∗j (cid:0)[ρ]ωj (VjA − AVj)(cid:1) . Vj(cid:0)[ρ]−ωj (V ∗j A∗ − A∗V ∗j )(cid:1) . Now use the fact that {Vj}j∈J = {V ∗j }j∈J and that for all j ∈ J , ∆σ(Vj) = e−ωj Vj and ∆σ(V ∗j ) = eωj V ∗j . It follows that (LρA)∗ = LρA∗. Using (7.7) we write (7.6) as ρ(0)U for the U found above. Since .ρ(0) is self-adjoint, it follows from what .ρ(0) = −L ρ(0)U∗. By the uniqueness of U , U is we have just shown that we also have .ρ(0) = −L self-adjoint. (cid:3) 7.4. DEFINITION. For each ρ ∈ S+, we identify the tangent space Tρ at ρ = ρ0, with the set of gradients vector fields G := {∇U : U ∈ A , U = U∗} through the one-to-one correspondence provided by (7.6). We define the Riemannian metric gL on S+ by where .ρ(0) and V are related by (7.6). k .ρ(0)k2 gL ,ρ = kVk2 L ,ρ 26 ERIC A. CARLEN AND JAN MAAS The metric we have just defined is C∞. Indeed, let A be m-dimensional and let A1, . . . , Am−1 be an orthonormal set of m − 1 self-adjoint traceless elements of HA. Then we can define a coordinate map u : S+ → Rm−1 by u(ρ) =(cid:0)Tr[A1ρ], . . . , Tr[Am−1ρ](cid:1) . Note that u(τ ) = 0. Evidently u is a one-to-one map of S+ onto an open bounded convex subset of Rm−1. We give S+, as usual, the corresponding differential structure. Conveniently, an atlas of just one chart covers the manifold. Let uk(ρ) = Tr[Akρ] by the kth coordinate function. The kth coordinate vector field is tangent to the curve t 7→ ρ + tAk for t in the open interval in which the right hand side belongs to S+. The operator div[ρ]~ω∇ is invertible on the orthogonal complement of the identity; i.e., on the span of A1, . . . , Am−1. Define the kth potential function Xk(ρ) to be the unique traceless solution X of It then follows that for the curve t 7→ ρ + tAk, .ρ(0) = div[ρ]~ω∇Xk(ρ). This means that the kth coordinate tangent vector field ∂/∂uk is given by div[ρ]~ω∇X = Ak . ∂ ∂uk = ∇Xk(ρ) , Therefore, in this coordinate system, the k, ℓ component of the metric tensor is given by [gL (ρ)]k,ℓ =Xj∈J h∇Xk(ρ), [ρ]ωj∇Xℓ(ρ)iHA , is C∞, and it follows from this that the map By Lemma 5.8, for each j, ρ 7→ [ρ]ωj ρ 7→ [div[ρ]~ω∇]−1, where the inverse is the inverse on the orthogonal complement of 1, is C∞. Thus, for each k, ℓ, [gL (ρ)]k,ℓ is a C∞ function of ρ. Now let F : S+ → R be a differentiable function. The differential of F, denoted is the unique traceless self-adjoint element in A satisfying 1 t(cid:0)F(ρ + tA) − F(ρ)(cid:1) = Tr(cid:20) δF δρ (ρ)A(cid:21) lim t→0 for all traceless self-adjoint A ∈ A. This notation is traditional in the context of gradient flows for the 2-Wasserstein metric, and it allows us to reserve the symbol D for covariant derivatives on our Riemannian manifold.) The corresponding gradient vector field, denoted gradgL F(ρ), will be interpreted using the identification of the tangent space given in Definition 7.4: it is the unique element in G satisfying δF δρ (ρ), (7.8) (7.9) for all differentiable paths ρ(t) defined on (−ǫ, ǫ) for some ǫ > 0 with ρ(0) = ρ and .ρ(0) + div([ρ]~ω∇U ) = 0 for some self-adjoint U . Combining (7.8) and (7.9), it follows that =(cid:10)gradgL F(ρ),∇U(cid:11)L ,ρ d dtF(ρ(t))(cid:12)(cid:12)(cid:12)(cid:12)t=0 (ρ), div(cid:0)[ρ]~ω∇U(cid:1)EHA −D δF δρ =(cid:10)gradgL F(ρ), [ρ]~ω∇U(cid:11)HA,J . Since this argument holds for arbitrary paths ρ(t), Theorem 7.3 implies that this identity holds for arbitrary U . Therefore, we have proved: GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 27 7.5. THEOREM. For a differentiable function F on S+, the Riemannian gradient of F with respect to the Riemannian metric gL is given by δF δρ gradgL F(ρ) = ∇ (ρ) , and the corresponding gradient flow equation (for steepest descent) is .ρ(t) = div(cid:16)[ρ(t)]~ω∇ δF δρ (ρ(t))(cid:17) . Let Pt = etL be a QMS on A that satisfies the σ-DBC for σ ∈ S+(A). Recall that the relative entropy with respect to σ is the functional D(·kσ) on S+ defined by (1.1). An easy calculation shows that for F(ρ) = D(ρσ), δF δρ = log ρ − log σ . Therefore, Theorem 5.10 and Theorem 7.5 yield: 7.6. THEOREM. Let Pt = etL be QMS on A that satisfies the σ-DBC for σ ∈ S+(A). Then ∂ ∂t ρ = L †ρ (7.10) is gradient flow for the relative entropy D(·σ) in the metric gρ,L canonically associated to L through its representation in the form (3.4). In [11], we proved the special case of Theorem 7.6 in which Pt is the infinite temperature Fermi Ornstein-Uhlenbeck semigroup, except that there we defined the metric in terms of ∂j, defined in (6.2) instead of the differential calculus associated to the skew derivations the derivations ∂j used here. The two metrics are in fact the same, and the alternate form of the metric in terms of the skew derivatives will be useful to us in the next section. b Therefore we explain the equivalence, using the notation introduced in section 6. Let ρ(t) be a smooth path in S+(Cn) defined on a neighborhood of 0 with ρ(0) = ρ. Suppose that .ρ(0) = − nXj=1 ∂j([ρ]0∂jU ) (7.11) for some self-adjoint U ∈ Cn. (We recall that since Vj is self-adjoint in this case, ∂†j = ∂j, and that ωj = 0 for each j.) Then by (6.4) and the integral representation for [ρ]0, we can rewrite (7.11) as b b nXj=1 nXj=1Z 1 0 .ρ(0) = −4 = −4 b b b b nXj=1Z 1 0 ∂j(W [ρ]0(W ∂jU )) = −4 ∂j(W ρsW ∂jU ρ1−s)ds ∂j(Γ(ρs) ∂jU ρ1−s)ds . The operation A 7→R 1 0 Γ(ρs)Aρ1−sds is precisely the non-commutative analog of "multi- plication by ρ" that was used in [11]. Thus apart form the trivial factor of 4, the realization of the tangent space and interpretation of continuity equation in [11] is the same as it is here; the two formulation of the continuity equation are equivalent. 28 ERIC A. CARLEN AND JAN MAAS The same applies to the metric. With .ρ(0) as above, k .ρ(0)k2 gL as we have defined it here is given by k .ρ(0)k2 gL = h∂j , U [ρ]0∂jUiHCn = 4 ∂jU, [ρ]0(W ∂jU )iHCn b b nXj=1 hW b ∂jU ρ1−siHCn ds . b nXj=1 nXj=1Z 1 0 h = 4 ∂jU, Γ(ρs) (7.12) The ultimate term in (7.12) is, apart from a trivial factor of 4, precisely how the metric tensor was defined in [11]. This shows two things: First, that Theorem 7.6 is an extension, and not merely an analog, of our work in [11]. Second, it makes available to us the differential calculus based on the skew derivations when studying the geometry associated to the Fermi Ornstein- Uhlenbeck semigroup. (We have explicitly discussed the infinite temperature case, but the same reasoning applies in general.) Since the skew derivations have the property of "lowering polynomial degree by one" for the eigenfunctions of L , as discussed between (6.12) and (6.13), this alternate formulation of the metric will be extremely helpful in the next section. 8. Geodesic convexity and relaxation to equilibrium In this section we develop the advantages of having written the evolution equation (7.10) as gradient flow for the relative entropy. We draw on work of Otto and Westdickenberg [56] and Daneri and Savar´e [19]. Both pairs of authors were primarily interested in infinite- dimensional problems concerning metrics on spaces of probability densities, but several of their results are new and interesting in finite dimension. The approach of Otto and Westdickenberg is thoroughly developed in the finite-dimensional setting in Section 2 of [19]. We briefly summarize what we need. Let (M, g) be any smooth, finite-dimensional Riemannian manifold. For x, y in M, the Riemannian distance dg(x, y) between x and y is given by minimizing an action integral of paths γ : [0, 1] → M running from x to y: where d2 g(x, y) = inf(cid:26)Z 1 k .γ(s)k2 0 k .γ(s)k2 g(γ(s)) ds : γ(0) = x, γ(1) = y(cid:27) , g(γ(s)) = gγ(s)( .γ(s), .γ(s)) . (If the infimum is achieved, any minimizer γ will be a geodesic.) If F is a smooth function on M, let gradgF denote its Riemannian gradient. Consider the semigroup St of trans- formations on M given by solving γ(t) = −gradgF (γ(t)); we assume for now that nice global solutions exist. The semigroup St, t ≥ 0, is gradient flow for F . For λ ∈ R, the function F is λ-convex in case whenever γ : [0, 1] → M is a distance minimizing geodesic, then for all s ∈ (0, 1), d2 ds2 F (γ(s)) ≥ λg( .γ(s), .γ(s)) . It is a standard result that whenever F is λ-convex, the gradient flow for F is λ- contracting in the sense that for all x, y ∈ M and t > 0, d dt d2 g(St(x), St(y)) ≤ −2λd2 g(St(x), St(y)) . (8.1) GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 29 Otto and Westdickenberg [56] developed an approach to geodesic convexity that takes (8.1) as its starting point. Let {γ(s)}s∈[0,1] be any smooth path in M with γ(0) = x and γ(1) = y. They use the gradient flow transformation St to define a one-parameter family of paths γt : [0, 1] → M, t ≥ 0 defined by γt(s) = Stγ(s) . Since γt is admissible for the variational problem that defines dg(St(x), St(y)), it is imme- diate that for each t ≥ 0, d2 (8.2) ds . 2 In the present smooth setting it is shown in [19, (2.8) – (2.11)] that if for all smooth curves γ : [0, 1] → M, 2 , g(γ 0(s)) (8.3) for all s ∈ (0, 1), then F is geodesically λ-convex. To see the connection between (8.3) and the contraction property, suppose that x and y are connected by a minimal geodesic γ so that d 2 d ds d ds d ds 0 (cid:13)(cid:13)(cid:13)(cid:13) γt(s)(cid:13)(cid:13)(cid:13)(cid:13) g(St(x), St(y)) ≤Z 1 g(γt(s))! ≤ −2λ(cid:13)(cid:13)(cid:13)(cid:13) dt(cid:12)(cid:12)(cid:12)(cid:12)0+ (cid:13)(cid:13)(cid:13)(cid:13) γt(s)(cid:13)(cid:13)(cid:13)(cid:13) γ(s)(cid:13)(cid:13)(cid:13)(cid:13) 0 (cid:13)(cid:13)(cid:13)(cid:13) g(x, y) =Z 1 dt(cid:12)(cid:12)(cid:12)(cid:12)0+ g(St(x), St(y)) ≤ −2λd2 d2 d ds d2 d g(γt(s)) γ0(s)(cid:13)(cid:13)(cid:13)(cid:13) 2 g(γ(s)) ds . g(x, y) , (If x and y are sufficiently close, this is the case.) Then (8.2) and (8.3) combine to yield and then, provided that St(x) and St(y) continue to be connected by a minimal geodesic for all t, the semigroup property of St yields the exponential λ-contractivity of the flow: dg(St(x), St(y)) ≤ e−λtdg(x, y) . (8.4) The local argument in [19] proves the geodesic λ-convexity of F when (8.3) is valid for all smooth paths in M, and thus leads to (8.4) without any assumptions of geodesic completeness. When (8.3) is valid for some λ > 0, and hence also (8.4) for the same λ, F has at most one fixed point x0 in M, which is necessarily a strict minimizer of F on M. We may normalize F (x0) = 0, and then under the geodesic λ-convexity of F , is it well known that for all x, d dt F (St(x)) ≤ −2λF (St(x)) , (8.5) which gives us another way to measure the rate of convergence to the fixed point under the flow St. There is also a more direct route from (8.3) to (8.5). If γ(t) is given by the gradient flow of F through γ(t) = St(x), then d dt F (γ(t)) = −kgradgF (γ(t))k2 g(γ(t)) . Define the energy function E associated to F by E(x) = kgradgF (x)k2 g(x) . (8.6) (8.7) 30 ERIC A. CARLEN AND JAN MAAS Then (8.3) applied with γt(s) = Ss+t(x) together with the semigroup property yields d dt E(St(x)) ≤ −2λE(St(x)) . Hence (8.3) not only leads to the contractivity property (8.4), but also to the exponential convergence estimates and E(St(x)) ≤ e−2λtE(x) . Moreover, combining (8.8) with (8.6) and (8.7), we obtain the inequality F (St(x)) ≤ e−2λtF (x) F (x) ≤ 1 2λ E(x) . (8.8) (8.9) In our setting, when F is a relative entropy function, (8.9) will be a generalized logarithmic Sobolev inequality. The relations between (8.4) and the bounds in (8.8) and geodesic convexity of F have all been discussed by Otto [54] as the basis of his approach to quantitative estimates on the rates of relaxation for solutions of the porous medium equation. Thus, to prove geodesic convexity of F , and hence (8.4) and (8.8), it suffices to prove (8.3). This first derivative estimate can provide a much easier route to a proof of λ- convexity of F than direct calculation of the Hessian of F followed by an estimate of its least eigenvalue. The point of view of Otto and Westdickenberg is that this approach can be especially fruitful in an infinite-dimensional setting (such as that of [54]) given all the regularity issues to go along with computing the Hessian of F . In the remainder of this section, we shall show that it is also quite fruitful in our finite-dimensional setting. Related work in the commutative setting can be found in [22, 23, 27, 44, 46, 48]. For the difficulties relating to direct computation and analysis of the Hessian even for the Fermi Ornstein-Uhlenbeck semigroup, see our previous paper [11]. 8.1. Geodesic convexity using intertwining relations. For the rest of this section, let σ ∈ S+, and fix Pt = etL , an ergodic QMS that satisfies the σ-DBC. Let L be given in the standard form (3.4), so that the data specifying L are the sets {Vj}j∈J and {ωj}j∈J . Let ∇ : HA → HA,J and div : HA,J → HA be the associated non-commutative gradient and divergence (as opposed to the associated Riemannian gradient and divergence). Let ρ : [0, 1] → S+ be a smooth path in S+, and define the one-parameter family of paths, ρt(s), (s, t) ∈ [0, 1] × [0,∞) by By what has been explained above, it we can prove that ρt(s) = P†t ρ(s) . 2 g(ρ0(s)) (8.10) d dt(cid:12)(cid:12)(cid:12)(cid:12)0+ (cid:13)(cid:13)(cid:13)(cid:13) d ds ρt(s)(cid:13)(cid:13)(cid:13)(cid:13) 2 g(ρt(s))! ≤ −2λ(cid:13)(cid:13)(cid:13)(cid:13) d ds ρ0(s)(cid:13)(cid:13)(cid:13)(cid:13) for all smooth ρ : [0, 1] → M and all s ∈ (0, 1), we will have proved the geodesic convexity of the relative entropy functional, and consequently, we shall have proved D(P†t ρσ) ≤ e−2λtD(ρσ) . We now present a simple sufficient condition for (8.10) that we shall be able to verify in a number of interesting examples. GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 31 8.1. DEFINITION. A semigroup ~Pt on HA,J intertwines with a semigroup Pt on HA in case for all t > 0, and all A ∈ HA, (8.11) (8.12) ∇PtA = ~Pt∇A . By duality, the intertwining relation ∇ ◦ Pt = ~Pt ◦ ∇ implies the identity for A ∈ HA,J . We will be particularly interested in cases in which for some λ ∈ R, P†t div(A) = div( ~Pt†A) , ~PtA = (e−λtPtA1, . . . , e−λtPtAJ ) . 8.2. Remark. A classical example is provided by the Mehler formula for the classical Ornstein-Uhlenbeck semigroup, which was first studied by Mehler in 1866. For β > 0, let γβ(x) = (β/2π)n/2e−βx2/2 be the centered Gaussian density on Rn with zero mean and variance n/β. For t > 0 and bounded continuous functions f on Rn, define Ptf by Ptf (x) =ZRn f (e−tx + (1 − e−2t)1/2y)γβ(y) dy . (8.13) Then Pt is a classical Markov semigroup; namely the Mehler or Ornstein-Uhlenbeck semi- group. The dual semigroup P †t acting on probability densities ρ on Rn is defined by A change of variables yields the dual Mehler formula: ρ(x)Ptf (x) dx P †t ρ(x)f (x) dx =ZRn ZRn ρ(cid:0)e−tx − (1 − e−2t)1/2y(cid:1)γβ(cid:0)(1 − e−2t)1/2x + e−ty(cid:1) dy . β∇ + x(cid:19) ρ(x, t) . div −x(cid:19)∇f (x, t) and ρ(x, t) = div(cid:18) 1 ∂ ∂t P †t ρ(x) =ZRn ∂ ∂t f (x, t) =(cid:18) 1 β A Taylor expansion in (8.13) and then integration by parts show that f (x, t) := Ptf (x) and ρ(x, t) := P †t ρ(x) satisfy It is immediate from (8.13) that (8.14) where ~Pt(v1, . . . vn)(x) = e−t(Ptv1(x), . . . , Ptvn(x)). We shall see below that an identity similar to (8.14) is readily proved for the Fermi and Bose Ornstein-Uhlenbeck semigroup. ∇Ptf (x) = ~Pt∇f (x) Using (8.14), which is a direct analog of (8.11) and (8.12), Ledoux [43, p. 447] gave a very simple proof of the optimal logarithmic Sobolev inequality for the classical Ornstein- Uhlenbeck semigroup. A key element in his proof is the joint convexity of (a, r) 7→ a2/r on Rn × (0,∞), for which we will need a suitable non-commutative analogue. The latter is provided by a well-known convexity inequality for matrices, which asserts that, for all ω ∈ R, the mapping (t1 + e−ω/2ρ)−1A∗(t1 + eω/2ρ)−1A dt(cid:21) ω AiHA = Tr(cid:20)Z ∞ (ρ, A) 7→ hA, [ρ]−1 (8.15) is jointly convex on S+ × A; see [35, 36]. Note that if ρ and A are scalars, the right-hand side reduces to A2/ρ. The non-commutative convexity result ultimately derives from Lieb's concavity Theorem [40]. Since P†t is completely positive, it follows from (8.15) that 0 hP†t A, [P†t ρ]−1 ω P†t AiHA ≤ hA, [ρ]−1 ω AiHA . (8.16) 32 ERIC A. CARLEN AND JAN MAAS There is a well-developed theory of monotone metrics beginning with work of Chentsov and Morozova [51] for classical Markov processes and its non-commutative extension initiated by Petz [61], and further developed in [39, 33, 59, 34, 69]. Other results from this theory will be useful in further developments. Now consider any smooth path ρ : [0, 1] → S+, and for each s ∈ (0, 1) write .ρ(s) = div A(s) where A(s) is the solution of .ρ(s) = div A(s) that minimizes hA, [ρ]−1 the definitions in Section 7, ~ω AiL ,ρ so that, by gL ,ρ( .ρ(s), .ρ(s)) =Xj∈J hAj(s), [ρ(s)]−1 ωj Aj(s)iHA . Set ρt(s) := P†t ρ(s), and suppose that the semigroup ~Pt defined by (8.12) intertwines with Pt. It follows that d ds ρt(s) = P†t div A(s) = div ~P†t A(s) . Consequently, by (8.12) and (8.16), (cid:13)(cid:13)(cid:13)(cid:13) d ds ρt(s)(cid:13)(cid:13)(cid:13)(cid:13) 2 g(ρt(s)) ≤ e−2λtXj∈J ≤ e−2λtXj∈J hP†t Aj(s), [P†t ρ(s)]−1 ωj P†t Aj(s)iHA hAj(s), [ρ(s)]−1 ωj Aj(s)iHA = e−2λt(cid:13)(cid:13)(cid:13)(cid:13) 2 , g(ρ(s)) d ds ρ(s)(cid:13)(cid:13)(cid:13)(cid:13) which clearly implies (8.10). Altogether we have proved: 8.3. THEOREM. Let σ ∈ S+, and let Pt = etL be an ergodic QMS that satisfies the σ-DBC. Let ∇ and div denote the associated non-commutative gradient and divergence. Suppose that for some λ > 0, the semigroup ~Pt defined by (8.12) intertwines with Pt. Then the relative entropy with respect to σ is geodesically λ-convex on S+ for the Rie- mannian metric (gL ,ρ)ρ. Moreover, the exponential convergence estimate holds, as well as the generalized logarithmic Sobolev inequality D(P†t ρσ) ≤ e−2λtD(ρσ) D(ρσ) ≤ 1 2λ τ(cid:2) − L †(ρ)(cid:0) log ρ − log σ(cid:1)(cid:3) . (8.17) While it is a problem of ongoing research to extend Theorem 8.3 to the infinite dimen- sional setting, the part of it concerning entropy and entropy production inequalities is relatively robust, as it relies most essentially on Lieb's convexity result, while the part of it concerning geodesic convexity is more involved and requires more work to generalize. Therefore, with regard to the infinite dimensional Bose-Ornstein-Uhlenbeck semigroup, we are not presently in a positions to make any statements about geodesic convexity of the entropy, but the situation is much better concerning entropy and entropy production. Let D[ρ] = τ(cid:2) − L †(ρ)(cid:0) log ρ − log σ(cid:1)(cid:3) be the entropy dissipation functional, which is minus the derivative of D(P†t ρσ) at t = 0. By Theorem 5.10, −L †ρ =Xj∈J ∂†j(cid:16)[ρ]ωj ∂j(log ρ − log σ)(cid:17) = div(A) . GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 33 The entropy dissipation functional D[ρ] is then given by D[ρ] =Xj∈J hAj, [ρ]−1 ωj AjiHA . Now replace ρ by ρ(t) := P†t ρ and let us assume that ρ is in the domain of L †, a trivial assumption in the finite dimensional case. Then d dt ρ(t) = P†t (L †ρ) = P†t (div(A)) , and assuming that ~Pt defined by (8.12) intertwines with Pt, we have that A(t) = e−λt(P†t A1, . . . , P†t AJ ) . Therefore, by the convexity argument used in the proof of Theorem 8.3 D[P†t ρ] = e−2λtXj∈J hP†t Aj, [P†t ρ]−1 ωj P†t AjiHA ≤ e−2λtD[ρ] . By a standard argument, the generalized log-Sobolev inequality (8.17) follows immediately for ρ in the domain of L †. In summary, if one is interested more in entropy dissipation inequalities than the geodesic convexity of the entropy Theorem 5.10 together with an intertwining relation allows one to bypass the Riemannian structure. This fact also un- derlines the utility of writing the evolution equation as gradient flow for the entropy, which is essential for the argument. In the rest of this section we explain how intertwining formulas may be proved. We con- sider two examples, already introduced, namely the Bose and Fermi Ornstein-Uhlenbeck semigroups. In the Fermi case, we are within the scope of the finite dimensional picture developed here, and we will be able to prove the geodesic convexity of the relative entropy. In the Bose case, we are in an infinite dimensional setting, and work remains to be done to rigorously prove the geodesic convexity in this case. However, by what has been explained in the preceding paragraphs, we shall obtain a rigorously valid generalized logarithmic Sobolev inequality. 8.2. Intertwining via commutation formulas. For both the Fermi and Bose Ornstein- Uhlenbeck semigroups Pt, there is a Mehler type formula for Pt from which the inter- twining can be readily checked. In the Fermi case, this can be found in formulas (4.1) and (4.2) of [10], and the formula in the Bose case is a simple adaptation of this. Fortunately however, it is not necessary to find an explicit formula for the action of the semigroup Pt to prove the intertwining identity and (8.12). In case where such identities are true, they can often be readily checked using the form of the generator L . 8.4. LEMMA. Suppose that for some numbers aj, j ∈ J , [∂j, L ] = −aj∂j (8.18) for each j ∈ J . Then defining ~Pt on HA,J by ~Pt(A1, . . . , AJ ) = (e−ta1 PtA1, . . . , e−taJ PtAJ ) , we have the intertwining relation ∂jPt = ~Pt∂j on A. Proof. Let A ∈ A and define A(t) = ∂jPtA. Then A(0) = ∂jA and d dt A(t) = ∂j L PtA = L ∂jPtA − aj∂jPtA = [L − ajI]A(t) . b b b b 34 ERIC A. CARLEN AND JAN MAAS It follows that t 7→ etaj A(t) is the unique solution of which is of course Pt∂jA. Therefore, ∂jPtA = e−taj Pt∂jA. d dt X(t) = L X(t) with X(0) = ∂jA, (cid:3) 8.5. THEOREM. Let Pt be the Bose Ornstein-Uhlenbeck semigroup with generator Lβ given in (6.17), and let σβ be its invariant state. Then for all ρ ∈ S+, D(Ptρσβ) ≤ e−2 sinh(β/2)tD(ρσβ) . Proof. Using (6.18), we may apply Lemma 8.4, and then the remarks following Theo- rem 8.3. (Note that for all t > 0, P†t is in the domain of L † so that the generalized log-Sobolev inequality is valid.) (cid:3) Results in [37, Appendix D] show that the constant 2 sinh(β/2) in Theorem 8.5 cannot be improved. We may make a similar application of Lemma 8.4, and then Theorem 8.3 itself to the Fermi Ornstein-Uhlenbeck semigroup. However, in this case, it is not the differential structure in terms of the derivations ∂j for which we have (8.18), but the skew derivations ∂j and ∂j. This was proved in (6.13) and (6.14). However, the metric can be written in ∂j and terms of ∂j just as well, bringing in the principle automorphism Γ, and indeed, this is how the metric was written in [11]. This permits us to argue as above in the Bose case, and we conclude: 8.6. THEOREM. For β ≥ 0, let Pt be the Fermi Ornstein-Uhlenbeck semigroup with generator Lβ given in (6.8), and let σβ be its invariant state. Then for all ρ ∈ S+, D(Ptρσβ) ≤ e−2λβ tD(ρσβ) where λβ = min{cosh(βej /2) : j = 1, . . . , m}. Moreover, the realtive entropy functional ρ 7→ D(ρσβ) is geodesically λβ convex in the Riemannain metric associated to Lβ. 8.3. Talagrand type inequalities. Our final results in this section extend a result from outr earlier paper [11] to the present more generla setting. The proof is step-for-step the one from our previous paper, with only minor modifications, and we shall thereofre be brief. However, it is worth recording the more general result since this subject has recently attracted the attention of other researchers [42, 62]. The connection between logarithmic Sobolev inequalities and transport inequalities of Talagrand type [68] was originally discovered and developed by Otto and Villani [55]. The fact that a Talgrand type inequlaity holds for our transport metric is further evidence that it is indeed a bona-fide transport metric. 8.7. THEOREM (Talagrand type inequality). Let L be the generator of an ergodic QMS that satisifes the σ-DBC with respect to σ ∈ S+. Suppose that the generalized logarithmic Sobolev inequality (8.17) is valid for some λ > 0. Let d(ρ1, ρ2) denote the Riemannain distance on S+ associated to L . For all ρ ∈ S+, d(ρ, σ) ≤r 2D(ρσ) λ . (8.19) Proof. Given ρ ∈ S+, define ρ(t) = Ptρ for t ∈ (0,∞). Since limt→∞ ρ(t) = σ, it follows that d(ρ, σ) ≤ arclength[ρ(·)] =Z ∞ 0 qgρ(t)( ρ(t), ρ(t)) dt . GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 35 Since the evolution described by t 7→ P†t ρ is gradient flow for the relative entropy, gρ(t)( ρ(t), ρ(t)) = − D(ρ(t)σ) so that for any 0 ≤ t1 < t2 < ∞, d dt Z t2 t1 qgρ(t)( ρ(t), ρ(t)) dt ≤ √t2 − t1pD(ρ(t1)σ) − D(ρ(t2)σ) . (8.20) Fix any ǫ > 0. Define the sequence of times {tk}, k ∈ N, D(ρ(tk)σ) = e−kǫD(ρσ) . (Since t 7→ D(ρ(t)σ) is strictly decreasing, tk is well defined.) Since D(ρ(t)σ) ≤ e−2λtD(ρσ), for each k, tk − tk−1 ≤ ǫ 2λ . Then by (8.20), with this choice of {tk}, tk−1qgρ(t)ρ(t)( ρ(t), ρ(t)) dt ≤ r ǫ Z tk 2λ Since lim ǫ→0 ∞Xk=1 e−kǫ/2pǫ(eǫ − 1)! = lim we obtain the desired bound. (e−(k−1)ǫ − e−kǫ)D(ρσ) e−kǫ/2pǫ(eǫ − 1) . 2λ = r D(ρσ) ǫ→0 ∞Xk=1 e−kǫ/2ǫ! =Z ∞ 0 e−x/2 dx = 2 , (cid:3) Appendix A. Proof of Theorem 3.1 In this appendix we present a simple and self-contained proof of Theorem 3.1. The starting point is an isometry that is crucial to the characterization of quantum Markov semigroup generators given by Gorini, Kossakowski and Sudarshan [30]: For any finite dimensional Hilbert space H, let C2(H) denote the linear operators from H to H equipped with the normalized Hilbert-Schmidt inner product hA, BiC2(H) = (dim(H))−1Tr[A∗B] so that k1kH = 1. As above, we use † for the Hermitian adjoint in C2(H). A special case deserves a special notation: let Hn denote the n×n complex matrices Mn(C) equipped with this same normalized Hilbert-Schmidt inner product. A particular orthonormal basis in Hn plays a distinguished role in what follows: For 1 ≤ i, j ≤ n, let Ei,j denote the n× n matrix whose i, j entry is 1, and whose other entries are all 0. If {e1, . . . , en} is the standard basis of Cn, then Ei,j is the rank one operator that is written as eiihej in a standard quantum mechanical notation introduced before. In this notation, one has Ei,j ⊗ Ei,j = (eiihej) ⊗ (eiihej) = ei ⊗ eiihej ⊗ ej . It follows that 1 n nXi,j=1 Ei,j ⊗ Ei,j = ΨihΨ where Ψ = 1 √n ej ⊗ ej nXj=1 is a rank one projection in Cn ⊗ Cn, and, in particular, is positive. This observation is due to Choi, and some of the simple but fundamental conclusions he drew from it are related below. More immediately, {Ei,j}1≤i,j≤n is an orthonormal basis of Hn called the matrix unit basis. (A.1) (A.2) 36 ERIC A. CARLEN AND JAN MAAS There is a natural identification of C2(Hn) with Hn ⊗ Hn that takes advantage of the multiplication on Hn: For A, B ∈ Hn, define the operator #(A ⊗ B) : Hn → Hn by #(A ⊗ B) : Mn(C) → Mn(C) , #(A ⊗ B)X = AXB , for all X ∈ Hn. It follows that the adjoint of #(A ⊗ B) as an operator on Hn is given by (cid:0)#(A ⊗ B)(cid:1)† = #(A∗ ⊗ B∗). Moreover, The left-hand side of (A.3) involves the trace for operators on Hn, whereas the right-hand side involves the trace for operators on Cn. Tr[#(A ⊗ B)] = Tr[A]Tr[B] . (A.3) A.1. LEMMA. Let {Fα} and {Gβ} be two orthonormal bases of Hn, so that {Fα ⊗ Gβ} is an orthonormal basis of Hn ⊗ Hn. Then {#(Fα ⊗ Gβ)} is orthonormal in C2(Hn). In particular, the map # is unitary from Hn ⊗ Hn into C2(Hn). Proof of Lemma A.1. We may compute the trace on Hn using any orthonormal basis, and using the matrix unit orthonormal basis {Ei,j}1≤i,j≤n, we have for all α, β and µ, ν, h#(Fα ⊗ Gβ), #(Fµ ⊗ Gν )iC2(Mn(C)) = n−2 Tr[(FαEi,jGβ)∗FµEi,jGν ] nXi,j=1 nXi,j=1 = n−2 nXi,j=1 Tr[G∗βEj,iF ∗αFµEi,jGν ] = n−2 (Gν G∗β)j,j(F ∗αFµ)i,i = n−1Tr[Gν G∗β]n−1Tr[F ∗αFµ] = δα,µδβ,ν . (cid:3) Consider any linear transformation K on Mn(C), and hence on Hn. Let {Fβ} be any orthonormal basis of Hn. Then {F ∗α} is also an orthonormal basis of Hn, and by Lemma A.1, {#(F ∗α ⊗ Fβ)} is an orthonormal basis of C2(Hn). Thus K has the expansion (A.4) cα,β#(F ∗α ⊗ Fβ) . K =Xα,β or, what is the same, K (A) =Xα,β cα,βF ∗αAFβ for all A ∈ Mn(C) , where the coefficients cα,β are given by cα,β = h#(F ∗α ⊗ Fβ), K iC2(Hn) . (A.5) (A.6) This orthonormal expansion is fundamental to the work of Gorini, Kossakowski and Sudarshan on the structure of generators of quantum Markov semigroups. A.2. DEFINITION. The n2 × n2 matrix cα,β with entries given by (A.6) is called the GKS matrix for the operator K with respect to the orthonormal basis {Fα}. When we wish to emphasize the dependence on K , we write cα,β(K ). A.3. Remark. Let A, B ∈ Mn(C) and consider the case K = #(A⊗ B). By the isometry proved in Lemma A.1, the GKS matrix of K is given by cα,β = h#(F ∗α ⊗ Fβ), #(A ⊗ B)iC2(Hn) = hF ∗α ⊗ Fβ, A ⊗ BiHn⊗Hn = n−2Tr[FαA]Tr[F ∗β B] . GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 37 In particular, the identity transformation on Mn(C) results from the choice A = B = 1, and so the GKS matrix of the identity transformation is the rank-one matrix cα,β = n−2Tr[Fα]Tr[F ∗β ] . (A.7) This formula will be useful later on. The following lemma is from [30]; for the convenience of the reader we give a short proof. A.4. LEMMA. Let K be a linear operator on Mn(C), and let {Fα} be an orthonormal basis of Mn(C). Then the GKS matrix of K with respect to {Fα} is self-adjoint if and only if (K A)∗ = K A∗ for all A ∈ Mn(C). Proof. Write K as in (A.4) and define fK (A) := (K A∗)∗. Then cα,βF ∗β AFα =Xα,β cα,βF ∗αA∗Fβ(cid:17)∗ =Xα,β cβ,αF ∗αAFβ . fK (A) =(cid:16)Xα,β By the uniqueness of the expansion (A.4), K = K if and only if cα,β = cβ,α for all α, β. (cid:3) The GKS matrix of a linear transformation K from Mn(C) to Mn(C) is closely related to the Choi matrix of K . Let {Ei,j}1≤i,j≤n be matrix unit basis of Hn. The Choi matrix of K is the element of Mn2(C) C(K ) = K (Ei,j) ⊗ Ei,j , (A.8) nXi,j=1 viewed as the n × n block matrix with entries in Mn(C) whose i, j entry is K (Ei,j). If we now identify Cn⊗Cn with Mn(C) by identifying v⊗w withPn becomes an operator on Hn. The identity provided by the following lemma was pointed out in [57], and used there to simplify part of the proof [30] of their theorem on the structure of generators of quantum Markov semigroups on Mn(C). A.5. LEMMA. Let K be a linear operator on Mn(C), and let C(K ) be defined by (A.8). i,j=1 viwjEi,j, so that C(K ) is Identify Cn ⊗ Cn with Mn(C) by identifying v ⊗ w withPn identified with an operator on Hn. Then for all F, G ∈ Hn, i,j=1 viwjEi,j, C(K ) (A.9) Proof. By direct computation, hG, C(K )FiHn = Gk,m[K (Ei,j)]k,ℓ[Ei,j]m,pFℓ,p hG, C(K )FiHn = h#(G ⊗ F ∗), K iC2(Hn) nXk,ℓ,m,p=1 nXk,ℓ,m,p=1 nXk,ℓ=1 nXk,ℓ=1 (F Ej,iG∗)ℓ,k[K (Ei,j)]k,ℓ Fℓ,p[Ej,i]p,mGk,m[K (Ei,j)]k,ℓ = = = (GEi,jF ∗)∗ℓ,k[K (Ei,j)]k,ℓ = h#(G ⊗ F ∗), K iC2(Hn) . 38 ERIC A. CARLEN AND JAN MAAS (cid:3) A fundamental theorem of Choi [14] says that a linear transformation K on Mn(C) is completely positive if and only if its Choi matrix C(K ) is positive as an operator on Hn. Indeed, in the notation of (A.2) C(K ) = nK ⊗ 1Mn(C) (ΨihΨ) , Hence, when K is completely positive, C(K ) is positive. The converse is also true: Choi used an elementary spectral decomposition [14] to show that when C(K ) is positive, then K is completely positive. A.6. Remark. The identity (A.9) shows that if {Fα} is any orthonormal basis of Hn, then the GKS matrix of K with respect to this basis is positive if and only if K is completely positive. In other words, the GKS representation (A.5) of a linear operator K on Mn(C) is well-suited to the question of whether K is completely positive or not. Going forward, it will be convenient to assume that our orthonormal bases {Fα} of Hn are indexed by α ∈ {1, . . . , n} × {1, . . . , n}, and we write α = (α1, α2). For such bases, we make the following definition: A.7. DEFINITION. Let L be an operator on Mn(C) such that L 1 = 0 and (L A)∗ = L A∗ for all A ∈ Mn. Let {Fα} be any orthonormal basis of Hn such that F(1,1) = 1. Let cα,β be the GKS matrix for L with respect to {Fα}. The (n2 − 1) × (n2 − 1) matrix with entries cα,β where α and β range over the set {(i, j) : 1 ≤ i, j ≤ n and (i, j) 6= (1, 1)} is called the reduced GKS matrix of L for the basis {Fα}. The following lemma is due to Parravinci and Zecca [57]: A.8. LEMMA. Let L be a linear operator on Mn(C), and let Pt = etL . Let {Fα} be an orthonormal basis for Hn with F(1,1) = 1. Let cα,β be the GKS matrix of L with respect to {Fα}. Then Pt is completely positive for all t ≥ 0 if and only if the reduced GKS matrix of L is positive. Proof. Suppose that Pt is completely positive for each t > 0. By (A.7), the GKS matrix of the identity transformation is cα,β(I) = δα,(1,1)δβ,(1,1) . (A.10) In particular, the reduced GKS matrix of the identity is zero. Then since cα,β(t−1(Pt − I)) = t−1cα,β(Pt) − t−1cα,β(I) , it follows that the reduced GKS matrix of t−1(Pt − I) coincides with the reduced GKS matrix of t−1Pt, and by Remark A.6 this is positive. Taking the limit t → 0, we conclude that the reduced GKS matrix of L is positive. Conversely, suppose that the reduced GKS matrix of L is positive. For small t > 0, cα,β(Pt) = cα,β(I) + tcα,β(L ) + o(t) By (A.10), this is positive for all sufficiently small t. By Remark A.6, Pt is completely positive for all sufficiently small t > 0. Then by the semigroup property, Pt is completely positive for all t > 0. (cid:3) We now temporarily put aside complete positivity, and consider a linear transformation L on Mn(C) such that L preserves self-adjointness, and such that L 1 = 0. GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 39 A.9. THEOREM. Let L be a linear operator on Mn(C) such that L 1 = 0 and (L A)∗ = L A∗ for all A ∈ Mn. Let {Fα} be any orthonormal basis of Hn such that F(1,1) = 1. Let cα,β be the GKS matrix of L for {Fα}. Then L is given by L A = −i[H, A] + cα,β (F ∗α[A, Fβ] + [F ∗α, A]Fβ ) (A.11) where H is the traceless self-adjoint matrix given by H = (c(1,1),βFβ − cβ,(1,1)F ∗β ) . (A.12) 1 2 Xα,β 6=(1,1) 2i Xβ6=(1,1) 1 Notice that only the reduced GKS matrix figures in the second term on the right in (A.11). Proof of Theorem A.9. Let cα,β be the GKS matrix for L with respect to {Fα} where F(1,1) = 1. Then by Lemma A.4, cα,β is a self-adjoint matrix. By (A.5), for all A ∈ Mn(C), (A.13) L A =Xα,β cα,βF ∗αAFβ = G∗A + AG + Xα,β 6=(1,1) cα,βF ∗αAFβ where G = c(1,1),(1,1) 2 self-adjoint matrices such that G = K + iH. Then (A.13) becomes c(1,1),βFβ. Let K = 1 1 + Xβ6=(1,1) L A = −i[H, A] + KA + AK + Xα,β 6=(1,1) cα,βF ∗αAFβ . 2 (G + G∗) and H = 1 2i (G− G∗) be the Then L 1 = 0 implies K = − 1 2 Xα,β 6=(1,1) L A = −i[H, A] + Xα,β 6=(1,1) 2 Xα,β 6=(1,1) = −i[H, A] + 1 cα,βF ∗αFβ, and thus, for all A ∈ Mn(C), AF ∗αFβ(cid:19) cα,β(cid:18)F ∗αAFβ − F ∗αFβA − 1 2 1 2 cα,β (F ∗α[A, Fβ ] + [F ∗α, A]Fβ) . Since Tr[G] = n Tr[H] = 0. 2 c(1.1),(1,1) ∈ R, and Tr[H] is the imaginary part of Tr[G], it follows that (cid:3) A.10. THEOREM. Let L be the generator of a QMS that is self-adjoint with respect to the inner product h·,·is for some s ∈ [0, 1], s 6= 1/2. Let cα,β be the GKS matrix of L with respect to a modular orthonormal basis of HA defined in Definition 1.2. Then for all α, β, eωαcα,β = cα,βeωβ , (A.14) and (A.15) where the ωα are defined in (3.1). In particular, c commutes with the diagonal matrix [δα,βeωβ], so that the eigenspaces of the latter are eigenspaces of c. cα,β = e−ωαcβ′,α′ , A.11. Remark. The conditions (A.14) and (A.15) are independent of s. Furthermore, (A.14) implies that (A.16) Therefore with an ordering of the indices α so that α ≥ β ⇐⇒ ωα ≥ ωβ, the matrix [cα,β] is block-diagonal. ωα 6= ωβ ⇒ cα,β = 0 . 40 ERIC A. CARLEN AND JAN MAAS Proof of Theorem A.10. Since L is the generator of a quantum Markov semigroup, L preserves self-adjointness and L 1 = 0. Thus, for any orthonormal basis {Fα} of Hn such that F(1,1) = 1, Theorem A.9 applies. We now fix such a basis and focus on the additional consequences of self-adjointness with respect to h·,·is. By Lemma 2.5, L commutes with the modular group, and this means that for all A, In terms of the GKS expansion for L , and making use of (3.2), L A = σ−1(L (σAσ−1))σ . Xα,β cα,βF ∗αAFβ = L A = σ−1(L (σAσ−1))σ =Xα,β cα,βeωβ−ωαF ∗αAFβ . = Xα,β cα,βσ−1F ∗ασAσ−1Fβσ Now (A.14) follows from the uniqueness of the coefficients. To prove (A.15), note that for any A, B, Tr[B∗cα,βF ∗αAFβ] =Xα,β Tr[B∗L A] =Xα,β Using Lemma A.4 we conclude that L †B =Xα,β hL A, Bis = Tr[(L (A))∗σ1−sBσs] = Tr[cα,βFβB∗F ∗αA] =Xα,β cα,βFαBF ∗β =Xα,β Tr[(cα,βFαBF ∗β )∗A] . cβ,αFαBF ∗β . Then Tr[A∗L †(σ1−sBσs)] = hA, σs−1L †(σ1−sBσs)σ−sis . The self-adjointness of L with respect to h·,·is then yields L (B) = σs−1L †(σ1−sBσs)σ−s for all B. Using the GKS expansion for a modular basis and (3.2), Xα,β cβ,ασs−1Fασ1−sBσsF ∗β σ−s =Xα,β Since F ∗γ = Fγ′ and ωγ = −ωγ′ for all γ, we can rewrite this as cα,βF ∗αBFβ =Xα,β cα,βF ∗αBFβ =Xα,β Xα,β cβ′,α′e(s−1)ωαe−sωβ F ∗αBFβ . cβ,αe(1−s)ωαesωβ FαBF ∗β . By the uniqueness of the coefficients, it follows that e−sωαcα,βesωβ = e−ωαcβ′,α′ for all α, β. However, by (A.14), c commutes with the sth power of [δα,βeωα], and thus e−sωαcα,βesωβ = cα,β. This proves (A.15). (cid:3) Proof of Theorem 3.1. By assumption Pt has an extension cPt from A to a QMS on all of Mn(C). It suffices to treat cPt, and to simplify the notation we assume the extension is done and Pt is a QMS on Mn(C). Let {Fα} be a modular basis for σ, and consider the GKS expansion By Theorem A.9, we can rewrite (A.17) as L A = −i[H, A] + L A =Xα,β 2 Xα,β 6=(1,1) 1 cα,βF ∗αAFβ . (A.17) cα,β (F ∗α[A, Fβ] + [F ∗α, A]Fβ) , (A.18) GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 41 where by (A.16), (A.12) reduces to H = = 1 2i√n Xβ6=(1,1), ωβ=0 2i√n Xβ6=(1,1), ωβ=0 1 (c(1,1),β Fβ − cβ,(1,1)F ∗β ) (c(1,1),β − cβ′,(1,1))Fβ . By (A.15) and (A.16), c(1,1),β = cβ′,(1,1), and therefore H = 0. Making use of the fact that for all γ, F ∗γ = Fγ′, we replace α with β′ and β with α′ and use (A.15) to rewrite (A.17) as L A =Xα,β cα,βFα′AF ∗β′ =Xα,β cβ′,α′FβAF ∗α =Xα,β cα,βeωαFβAF ∗α . (A.19) By Theorem A.9, we can rewrite (A.19) as 1 1 (A.20) cα,βeωα (Fβ[A, F ∗α] + [Fβ, A]F ∗α) 2 Xα,β 6=(1,1) 2i√n Xα6=(1,1), ωα=0 L A = −i[bH, A] + By (A.16), (A.12) reduces to bH = argument that led to H = 0 leads to bH = 0. Averaging (A.18) and (A.20), taking into account H = bH = 0, we obtain Now let U be an (n2 − 1) × (n2 − 1) unitary matrix that diagonalizes the reduced GKS matrix cα,β of L and which commutes with the matrix δα,βeωα, α, β 6= (1, 1) so that Uγ,α = 0 unless ωγ = ωα. We may then write cα,β[(F ∗α[A, Fβ] + [F ∗α, A]Fβ ) + eωα (Fβ[A, F ∗α] + [Fβ, A]F ∗α)] . (c(1,1),αF ∗α − cα,(1,1)Fα). The same 1 4 Xα,β6=(1,1) L A = (A.21) cα,β = 1 2 Xγ6=(1,1) U∗α,γe−ωγ /2cγUγ,β (A.22) Each cγ is non-negative since, by Lemma A.8 the reduced GKS matrix cα,β is positive, and U∗α,γeωγ /2cγUγ,β. Defining since Uγ,α = 0 unless ωγ = ωα, we also have eωαcα,β = 1 Vγ =Pβ Uγ,βFβ, we may rewrite (A.21) as 1 L A = 2 Xγ6=(1,1) cγ[e−ωγ /2(cid:0)V ∗γ [A, Vγ] + [V ∗γ , A]Vγ(cid:1) + eωγ /2(cid:0)Vγ[A, V ∗γ ] + [Vγ, A]V ∗γ(cid:1)] . (A.23) By symmetry, we may assume without loss of generality that cγ = cγ′ where as before (γ1, γ2)′ = (γ2, γ1) so that V ∗γ = Vγ′. Then the expression simplifies to cγ[e−ωγ /2V ∗γ [A, Vγ ] + eωγ /2[Vγ, A]V ∗γ ] 2Xγ which is (3.3). Simply using (A.22) directly in (A.18) leads to the alternate form L A = Xγ6=(1,1) L A = Xγ6=(1,1) cγe−ωγ /2(cid:0)V ∗γ [A, Vγ] + [V ∗γ , A]Vγ(cid:1) . Again since Uγ,α = 0 unless ωγ = ωα, (3.2) implies that for all γ and all t, σtVγσ−t = e−tωγ Vγ (A.24) 42 ERIC A. CARLEN AND JAN MAAS Letting J = {(k, ℓ) : 1 ≤ k, ℓ,≤ n and (k, ℓ) 6= (1, 1) }, we see that under the hypothe- ses of the theorem, L must have the form (3.3), and (3.5) is the differential statement of (A.23). The final step is to absorb the cj's nto the Vj's: Since cj ≥ 0 for each j, we can absorb these by by making the replacement Vj → √cjVj. This proves that the generator L of a QMS satisfying the σ-DBC has the from specified in Theorem 3.1. For the converse, if L has the specified form, one restores the cj's by normalizing the Vj's, and then writes L in its GKS form for this orthonormal basis (after including 1 and any Vj's with cj = 0). The reduced GKS matrix of L is unchanged as the argument starting from (A.13) shows. Thus, by Remark A.6, L generates a completely positive semigroup Pt, and evidently L 1 = 0, so that Pt is a QMS. The σ-DBC is then readily checked (using the fact that the Vj are eigenvectors of ∆σ). (cid:3) A.12. Remark. It is easy to check the existence of an extension of Pt from A to Mn(C) in many relevant cases; e.g., when A is a Clifford algebra with an odd number of generators. One might hope that there is a general extension using the conditional expectation. Recall that for any unital C∗-subalgebra A of Mn(C), there is the conditional expecta- tion EA which is the orthogonal projection in HMn(C) onto HA [71]. This may be written as an average over the unitaries in the commutant of A [9, 70]; the connected component of this group U that contains the identity is a Lie subgroup of SU (n), on which there ex- U∗XU dµ(U ). Evidently, EA is a completely positive map with EA1 = 1. That is, EA is a quantum Markov operator. ists a normalized Haar measure µ, and then for all X ∈ Mn(C), EAX =ZU If K is a linear transformation on A, define a linear transformation cK on Mn(C) by cK = K ◦ EA. Note that when a linear operator P on A is completely positive, so is cP, and the restriction of cP to A is simply P. Moreover, if L1 and L2 are two linear transformations of A, then cL2 cL1 = \L2L1. In this way we can "lift" any QMS Pt on A up to a one-parameter family of Markov operators cPt on Mn(C) such that for all s, t ≥ 0, cPtcPs = [Pt+s. This construction fails to yield a semigroup only because limt→0 cPt = EA (cPt − EA) = cL . The operator and not limt→0 cPt = IA. However, if Pt = etL , then lim cL is evidently a self-adjointness preserving linear transformation from Mn(C) to Mn(C), and cL 1 = 0. If Pt satisfies the σ-DBC for σ ∈ S+(A), then for all A, B ∈ Mn(C), τ [σB∗cPtA] = τ [σB∗EAPt(EAA)] = τ [(EA(Bσ)∗)Pt(EAA)] = τ [σ(EAB)∗Pt(EAA)] , where we have used the fact that since σ ∈ A, EA(σB∗) = σEAB∗. for each t ≥ 0, cPt is self-adjoint with respect to the σ-GNS inner product on Mn(C). Consequently, the same is true of cL . Therefore, the proof of Theorem 3.1 given just above shows that cL has the form speci- fied in Theorem 3.1 except that some cj's might be negative: Applicability of Lemma A.8 requires that limt→0 Pt = IMn(C). 1 t t→0 It follows that Appendix B. Note on KMS-symmetry We give a construction of a class of operators K that satisfy (K A)∗ = K A∗ for all A and that are self-adjoint with respect to the σ-KMS inner product h·,·i1/2 for some σ ∈ S+, but which do not commute with ∆σ, and consequently are not self-adjoint with respect to the σ-GNS inner product. An operator K on Mn(C) that is self-adjoint with respect to h·,·i1/2 for some σ ∈ S+ is called KMS-symmetric. GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 43 When K is completely positive with K (1) = 1, it has a Kraus representation K (A) = K∗j AKj where K∗j Kj = 1 mXj=1 mXj=1 Then bKj = ∆1/2 mXj=1 bKjσbK∗j = σ. for some set {K1, . . . , Km} ⊂ Mn(C). Evidently, K †A = Pm K †σ = σ with σ ∈ S+. The dual set of Kraus operators {bK1, . . . , bKm} is given by σ K∗j = σ1/2K∗j σ−1/2 . j=1 KjAK∗j . Suppose that Pt = Define a quantum Markov semigroup Pt by mXj=1 bK∗j bKj = 1 and It follows that the operator cK defined by mXj=1 bK∗j AbKj is completely positive with cK (1) = 1 and cK †(σ) = σ. A simple cK A = calculation shows that h cK B, Ai1/2 = hB, K Ai1/2 for all A, B ∈ Mn(C). Thus, cK is the adjoint of K with respect to the σ-KMS inner product, and cK K is a completely positive operator on Mn(C) such that cK K (1) = 1 and ( cK K )†σ = σ. ( cK K )n = et( cK K −I) . Evidently Pt is KMS-symmetric for each t > 0 since cK K is KMS-symmetric. Further- more, Pt commutes with ∆σ for each t > 0 if and only if cK K commutes with ∆σ. We To construct counterexamples, consider n = 2; the construction that follows is readily generalized. Let {u1, u2} be an orthonormal basis in C2 and let {v1, v2} be a set of two linearly independent unit vectors in C2 that are not orthogonal. Define the rank-one operators K1 and K2 by Kj = vjihuj, j = 1, 2. Evidently, K∗1 K1 + K∗2 K2 = 1, and we define K A = K∗1 AK1 + K∗2 AK2 so that K 1 = 1. Then the range of K † is spanned by (cid:8)v1ihv1 , v2ihv2(cid:9). A simple computation yields will show that the latter is not generally the case. K †(cid:0)α1v1ihv1 + α2v2ihv2(cid:1) = β1v1ihv1 + β2v2ihv2 , e−t tn n! ∞Xn=0 β2 (cid:19) =(cid:20) 1 − a (cid:18) β1 α2 (cid:19) , 1 − b (cid:21)(cid:18) α1 a (cid:19) is an eigenvector with eigenvalue 1, and hence with a = hv1, u2i2 and b = hv2, u1i2, and since {v1, v2} is not orthogonal, a + b > 0. The vector(cid:18) b (B.1) a b where σ = b a + bv1ihv1 + a a + bv2ihv2 (B.2) satisfies K †σ = σ, and hence, as we just noted, ( cK K )†σ = σ. The other eigenvalue of the matrix in (B.1) is 1 − a − b < 1, so that (B.2) gives the unique invariant state. It follows that the eigenvalues of K † are 1, 1 − a − b and 0, with Null(K †) = Span(cid:8)u1ihu2 , u2ihu1(cid:9) . Consequently, the null space of K is 2-dimensional as well, and same holds for the null space of cK K , since cK is the σ-KMS dual of K . By ergodicity, it is the only eigenspace of cK K with this property. 44 ERIC A. CARLEN AND JAN MAAS Generically, σ will have two distinct eigenvalues. (For example, take {u1, u2} to be 1 the standard basis of C2, and take v1 = 1√2(cid:18) 1 1 (cid:19) and v2 = 1√5(cid:18) 1 7(cid:20) 2 3 3 5 (cid:21).) Let {η1, η2} be an orthonormal basis of eigenvectors of σ. independent eigenvectors of cK K with the same eigenvalue. This eigenvalue can only be If Pt and ∆σ were to commute, then Remark 2.7 would imply that η1ihη2 and η2ihη1 are linearly zero by the above. But then 0 = K (η1ihη2) = K (η2ihη1). Hence we would have 2 (cid:19), so that σ = 0 = K (η2ihη1) = hv1, η2ihη1, v1iu1ihu1 + hv2, η2ihη1, v2iu2ihu2 , which would mean that hv1, η2ihη1, v1i = 0 and hv2, η2ihη1, v2i = 0 . (B.3) Suppose hv1, η2i = 0. Then since {v1, v2} is not orthogonal, and {η1, η2} is, hv2, η1i 6= 0. The second equality in (B.3) then yields hv2, η2i, but we cannot have both hv1, η2i = 0 and hv2, η2i = 0 since this would imply that η2 = 0. Under this condition, the first equality in (B.3) would yield hη1, v1i = 0. As above, this would imply hη2, v2i 6= 0, and hence hη1, v2i = 0. We cannot have both hv1, η1i = 0 and hv2, η1i = 0 since this would imply that η1 = 0. Hence K (η2ihη1) = 0 is impossible. Thus, with this choice of K , L := cK K − I is the generator of a quantum Markov semigroup with invariant state σ ∈ S+ such that L is self-adjoint with respect to the σ-KMS inner product h·,·i1/2, but such that L does not commute with ∆σ. It follows that L is not self-adjoint with respect to the GNS inner product. Acknowledgements E.C. was partially supported by NSF grant DMS 1501007, and thanks IST Austria for hospitality during a visit in June 2015. E.C. thanks the Mittag- Leffler Institute for hopitality during the final work on this paper. Both authors thank the Erwin Schrodinger Institute in Vienna for hospitality during a visit in June 2016. References [1] L. Accardi, F. Fagnola, and R. Quezada, On three new principles in non-equilibrium statistical me- chanics and Markov semigroups of weak coupling limit type, Infin. Dimens. Anal. Quantum Probab. Relat. Top., 19(2):1650009, 37 pp., 2016. [2] G. S. Agarwal, Z. Physik 258, 409-422, 1973. [3] S. Albeverio, R. Høegh-Krohn, Dirichlet Forms and Markovian semigroups on C ∗ algebras, Comm. Math. Phys. 56 173-187, 1977 [4] R. Alicki, On the detailed balance condition for non-Hamiltonian systems, Rep. Math. Phys., 10 249-258. 1976 [5] L. Ambrosio, N. Gigli, and G. Savar´e, Gradeint flows in metric spaces and in the space of probability measures, Lectures in Mathematics ETH Zurich, Birkhauser Verlag, Basel, 2005. [6] J. D. Benamou and Y. Brenier, A computational fluid mechanics solution to the Monge–Kantorovich mass transfer problem, Numer. Math., 84 375-393. 2000 [7] A. Beurling and J. Deny, Espaces de Dirichlet I: le cas elementaire, Acta Math. 99, 203-224, 1958. [8] A. Beurling and J. Deny, Dirichlet spaces, Proc. Natl. Acad. Sci. 45, 208-215, 1959. [9] E. A. Carlen Trace inequalities and quantum entropy: an introductory course. Entropy and the quan- tum, 73–140, Contemp. Math., 529, Amer. Math. Soc., Providence, RI, 2010. [10] E. A. Carlen and E. H. Lieb, Optimal hypercontractivity for Fermi fields and related noncommutative integration inequalities, Comm. Math. Phys. 155, 27-46, 1993. [11] E. A. Carlen and J. Maas, An analog of the 2-Wasserstein metric in non-commutative probability under which the fermionic Fokker-Planck equation is gradient flow for the entropy, Comm. Math. Phys. 331, 887–926, 2014. GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 45 [12] E. A. Carlen and J. Maas, On the geometry of dissipative quantum systems with detailed balance, in preparation. [13] H. J. Carmichael and D. F. Walls,: Z. Physik B23, 299-306, 1976. [14] M. D. Choi, Completely positive linear maps on complex matrices, Lin. alg. and Appl. 10 285-290 1975. [15] F. Cipriani Dirichlet forms and Markovian semigroups on standard forms of von Neumann algebras. J. Funct. Anal. 147, 259-300, 1997. [16] F. Cipriani, Dirichlet forms on noncommutative spaces, in Quantum potential theory, vol. 1954 of Lecture Notes in Math., Springer, Berlin, 161-276, 2008. [17] F. Cipriani, F. Fagnola, J. M. Lindsay, Spectral analysis and Feller property for quantum Ornstein– Uhlenbeck semigroups, Comm. Math. Phys. 210 85-105, 2000. [18] F. Cipriani and J. L. Sauvageot, Derivations as square roots of Dirichlet forms, J. Funct. Anal., 201 78-120, 2003 [19] S. Daneri and G. Savare, Eulerian calculus for the displacement convexity in the Wasserstein distance, SIAM J. Math. Anal., 40, 1104-1122, 2008. [20] E. B. Davies, Markovian Master Equation, Commun. Math. Phys. 39, 91-110, 1974. [21] E. B. Davies, J. M. Lindsay, Non-commutative symmetric Markov semigroups, Math. Z. 210 379-411, 1992. [22] M. Erbar and J. Maas, Ricci curvature of finite Markov chains via convexity of the entropy, Arch. Ration. Mech. Anal., 206, 997-1038, 2012. [23] M. Erbar, J. Maas and P. Tetali. Discrete Ricci curvature bounds for Bernoulli–Laplace and random transposition models, Ann. Fac. Sci. Toulouse Math 24, 781-800, 2015. [24] F. Fagnola and R. Reboledo, Entropy production and detailed balance for a class of quantum Markov semigroups, Open Syst. Inf. Dyn. 22, 1550013 (2015). [25] F. Fagnola and V. Umanit`a, Generators of Detailed Balance Quantum Markov Semigroups, Infin. Dimens. Anal. Quantum Probab. Relat. Top., 10, no. 3, 335-363, 2007. [26] F. Fagnola and V. Umanit`a, Generators of KMS Symmetric Markov Semigroups on B(h): Symmetry and Quantum Detailed Balance. Commun. Math. Phys. 298, 523-547, 2010. [27] M. Fathi and J. Maas. Entropic Ricci curvature bounds for discrete interacting systems, Ann. Appl. Probab., 26, 1774-1806, 2016. [28] A. Frigerio and M. Verri. Long-time asymptotic properties of dynamical semigroups on W ∗-algebras. Math. Z., 180(2):275–286, 1982. [29] S. Goldstein and J. M. Lindsay, Beurling-Deny conditions for KMS-symmetric dynamical semigroups, C. R. Acad. Sci. Paris Ser. I 317,1053-1057., 1993. [30] V. Gorini, A. Kossakowski and E. C. G. Sudarshan, Completely positive dynamical semigroups of N -level systems, J. Math. Phys. 17, 821-825,1976. [31] L. Gross, Existence and uniqueness of physical ground states, J. Funct. Anal. 10 59-109, 1972. [32] L. Gross, Hypercontractivity and logarithmic Sobolev inequalities for the Clifford-Dirichlet form, Duke Math. J. 42 383-396, 1975. [33] F. Hiai and H. Kosaki, Means for matrices and comparison of their norms, Indiana Univ. Math. J., 48, 899-936, 1999. [34] F. Hiai, H. Kosaki, D Petz, and M. B. Ruskai, Families of completely positive maps associated with monotone metrics, Lin. Alg. Appl. 439, 1749–1791, 2013. [35] F. Hiai, D Petz, Convexity of quasi-entropy type functions: Lieb's and Ando's convexity theorems revisited, J. Math. Phys. 54, 062201, 2013. [36] F. Hiai, D Petz, From quasi-entropy to various quantum information quantities, Publ. Res. Inst. Math. Sci. 48, 525-542, 2012. [37] S. Huber, R. Konig, and A. Vershynina. Geometric inequalities from phase space translations, arXiv:1606.08603, 2016. [38] M. Junge and Z. Zeng, Noncommutative martingale deviation and Poincar´e type inequalities with applications, Probab. Theory Related Fields 161, no. 3-4, pp. 449–507, 2015. [39] A. Lesniewski and M. B. Ruskai, Monotone Riemannian metrics and relative entropy on noncommu- tative probability spaces, J. Math. Phys. 40, 5702-5724, 1999. [40] E. H. Lieb Convex trace functions and the Wigner-Yanase-Dyson conjecture, Adv. Math. 11, 267-288, 1973. 46 ERIC A. CARLEN AND JAN MAAS [41] A. Kossakowski, A. Frigerio, V. Gorini, and M. Verri, Quantum detailed balance and KMS condition, Comm. Math. Phys., 57, 97-110, 1977. [42] M. Junge and Q. Zeng, Noncommutative martingale deviation and Poincarn´e type inequalities with applications, Probability Theory and Related Fields 161 449-507, 2015 [43] M. Ledoux, On an integral criterion for hypercontractivity of diffusion semigroups and extremal func- tions, J. Func. Anal. 105, 445-467, 1992. [44] M Liero and A. Mielke, Gradient structures and geodesic convexity for reaction-diffusion systems, Philos. Trans. R. Soc. Lond. Ser. A 371 20120346, 2013. [45] G. Lindblad, On the generators of quantum dynamical semigroups, Comm. Math. Phys., 48, 119-130, 1976. [46] J. Maas, Gradient flows of the entropy for finite Markov chains, J. Funct. Anal., 261, 2250-2292, 2011. [47] W. A. Majewski and R. F. Streater, Detailed balance and quantum dynamical maps, J. Phys. A: Math. Gen. 31, (1998) 7981-7995. [48] A. Mielke. Geodesic convexity of the relative entropy in reversible Markov chains, Calc. Var. Part. Diff. Equ., 48(1): 1–31, 2013. [49] A. Mielke, Dissipative quantum mechanics using GENERIC. Recent trends in dynamical systems, Springer Proc. Math. Stat., 35, 555–585. [50] A. Mielke, On thermodynamical couplings of quantum mechanics and macroscopic systems. Mathe- matical results in quantum mechanics, 331–348, World Sci. Publ., Hackensack, NJ, 2015. [51] E A. Morozova and N. N. Chentsov, Markov invariant geometry on the state manifolds (Russian), Itogi Nauki i Tekhniki 36, 69-102, 1990. [52] H. C. Ottinger. Beyond Equilibrium Thermodynamics. John Wiley, New Jersey, 2005. [53] H. C. Ottinger. The nonlinear thermodynamic quantum master equation. Phys. Rev. A, 82, 052119(11), 2010. [54] F. Otto, The geometry of dissipative evolution equations: the porous medium equation, Comm. Partial Diferential Equations, 26, 101-174, 2001. [55] F. Otto and C. Villani: Generalization of an inequality by Talagrand and links with the logarithmic Sobolev inequality, J. Funct. Anal. 173, 2000, pp. 361–400 [56] F. Otto and M. Westdickenberg, Eulerian calculus for the contraction in the Wasserstein distance, SIAM J. Math. Anal., 37, 1227-1255, 2005. [57] G. Parravinci and A. Zecca, On the generator of completely positive dynamical semigroups of N -level systems, Rep. Math. Phys., 12, 423-424, 1972 [58] V. Paulsen, Completely bounded maps and operator algebras, vol. 78 of Cambridge Studies in Advanced Mathematics, Cambridge University Press, Cambridge, 2002. [59] D. Perez-Gacia, M. M. Wolf, D. Petz, and M. B. Ruskai, Contractivity of positive and trace- preserving maps under Lp norms, Journal of mathematical physics, 47, 083506, 2006. [60] D. Petz, A dual in von Neumann algebras. Quart. J. Math. Oxford 35, 475-483, 1984. [61] D. Petz, Monotone metrics on matrix spaces, Linear Algebr. Appl. 244, 81-96, 1996. [62] C. Rouze and N. Datta, Concentration of quantum states from quantum functional and Talagrand inequalities, arXiv preprint 1704.02400. [63] S. Sakai, C ∗-algebras and W ∗-algebras, Springer-Verlag, New York-Heidelberg, 1971. Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 60. [64] I.E. Segal: A non-commutative extension of abstract integration, Annals of Math., 57, 401–457, 1953. [65] I.E. Segal.: Tensor algebras over Hilbert spaces II, Annals of Math., 63, 160-175, 1956. [66] I.E. Segal.: Algebraic integration theory, Bull. Am. Math. Soc.., 71, no. 3, pp. 419-489, 1965. [67] H. Spohn and J. L. Lebowitz, Stationary non-equilibrium states of infinite harmonic systems, Comm. Math. Phys., 54, 97-120, 1977. [68] M. Talagrand: Transportation cost for Gaussian and other product measures, Geom. Funct. Anal. 6, 1996, pp. 587–600 [69] K. Temme, M. J. Kastoryano, M. B. Ruskai, M. M. Wolf, and F. Verstraete, The χ2 divergence and mixing time of quantum Markov processes, J. Math. Phys. 51, 122201 (2010). [70] A. Uhlmann: Satze uber Dichtematrizen, Wiss. Z. Karl-Marx Univ. Leipzig 20, 633-653, 1971 [71] H. Umegaki: Conditional expectation in operator algebras I, Tohoku Math. J., 6, 1954, pp. 177-181 [72] C. Villani, Optimal transport, old and new, Springer Verlag, Berlin, 2008. GRADIENT FLOW FOR QUANTUM MARKOV SEMIGROUPS 47 Department of Mathematics, Hill Center, Rutgers University, 110 Frelinghuysen Road, Piscataway, NJ 08854-8019, USA E-mail address: [email protected] Institute of Science and Technology Austria (IST Austria), Am Campus 1, 3400 Klosterneuburg, Austria E-mail address: [email protected]
1111.1910	1	1111	2011-11-08T13:57:41	Projective representations of groups using Hilbert right C*-modules	[ "math.OA" ]	The projective representation of groups was introduced in 1904 by Issai Schur. It differs from the normal representation of groups by a twisting factor. It starts with a group T and a scalar valued function f on T^2 satisfying the conditions: f(1,1)=1, \|f(s,t)\|=1, and f(r,s)f(rs,t)=f(r,st)f(s,t) for all r,s,t in T. The projective representation of T twisted by f is a unital C-subalgebra of the C-algebra of operators on the Hilbert space l^2(T). This representation can be used in order to construct many examples of C-algebras. By replacing the scalars with an arbitrary unital C-algebra (as range of f) the field of applications is enhanced in an essential way. The projective representation of groups, which we present in this paper, has some similarities with the construction of cross products with discrete groups. It opens the way to create many K-theories. In a first section we introduce some results which are needed for this construction, which is developed in the second section. In a third section we present some examples of C*-algebras obtained by this method. Examples of a special kind (the Clifford algebras) are presented in the last section.	math.OA	math	PROJECTIVE REPRESENTATIONS OF GROUPS USING HILBERT RIGHT C-MODULES CORNELIU CONSTANTINESCU June 20, 2018 Abstract The projective representation of groups was introduced in 1904 by Issai Schur (1875-1941) in his paper [S]. It differs from the normal representation of groups (introduced by his tutor Ferdinand Georg Frobenius (1849-1917) at the suggestion of Richard Dedekind (1831-1916)) by a twisting factor, which we call Schur function in this paper and which is called sometimes multipliers or normalized factor set in the literature (other names are also used). It starts with a group T and a Schur function f for T . This is a scalar valued function on T × T satisfying the conditions f (1, 1) = 1 and f (s, t) = 1 , f (r, s)f (rs, t) = f (r, st)f (s, t) for all r, s, t ∈ T . The projective representation of T twisted by f is a unital C-subalgebra of the C-algebra L(l2(T )) of operators on the Hilbert space l2(T ). This representation can be used in order to construct many examples of C-algebras (see e.g. [C1] Chapter 7). By replacing the scalars IR or IC with an arbitrary unital (real or complex) C-algebra E the field of applications is enhanced in an essential way. In this case l2(T ) is replaced by the Hilbert right E ≈ E ⊗ l2(T ) and L(l2(T )) is replaced by LE(E ⊗ l2(T )), the E-module (cid:13)t∈T C-algebra of adjointable operators of L(E ⊗ l2(T )). The projective represen- tation of groups, which we present in this paper, has some similarities with the construction of cross products with discrete groups. It opens the way to create many K-theories. 1 In a first section we introduce some results which are needed for this con- struction, which is developed in the second section. In the third section we present examples of C-algebras obtained by this method. Examples of a spe- cial kind (the Clifford algebras) are presented in the last section. AMS Subject Classification: 22D25 (Primary) 20C25, 46L08 (Secondary) Key Words: Hilbert right C∗-modules, Projective groups representations Contents 0 Notation and Terminology 1 Preliminaries 1.1 Schur functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 7 7 1.2 E-C-algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 1.3 Some topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2 Main Part 40 2.1 The representations . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.2 Variation of the parameters . . . . . . . . . . . . . . . . . . . . . 72 2.3 The functor S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98 3 Examples 104 3.1 T := ZZ2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 3.2 T := ZZ2 × ZZ2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 2 3.3 T := ( ZZ2 )n with n ∈ IN . . . . . . . . . . . . . . . . . . . . . . . 128 3.4 T := ZZn with n ∈ IN . . . . . . . . . . . . . . . . . . . . . . . . . 133 3.5 T := ZZ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136 4 Clifford Algebras 137 4.1 The general case . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 4.2 A special case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 REFERENCES and INDEX 0 Notation and Terminology Throughout this paper we use the following notation: T is a group, 1 is its neutral element, K := l2(T ), 1K := idK := identity map of K, E is a unital C-algebra (resp. a W-algebra), 1E is its unit, E denotes the set E endowed with its canonical structure of a Hilbert right E-module ([C1] Proposition 5.6.1.5), H := E ⊗ K ≈ (cid:13)t∈T E , (resp. H := E ¯⊗K ≈ E) W (cid:13)t∈T ([C3] Proposition 2.1, (resp. [C3] Corollary 2.2)). In some examples, in which T is additive, 1 will be replaced by 0. The map LE( E) −→ E, u 7−→ h u1E 1E i = u1E is an isomorphism of C-algebras with inverse x 7−→ x · . We identify E with LE( E) using these isomorphisms. E −→ LE( E), 3 In general we use the notation of [C1]. For tensor products of C-algebras we use [W], for W-tensor products of W-algebras we use [T], for tensor products of Hilbert right C∗-modules we use [L], and for the exterior W-tensor products of selfdual Hilbert right W ∗-modules we use [C2] and [C3]. In the sequel we give a list of notation used in this paper. 1) IK denotes the field of real numbers (:= IR) or the field of complex numbers (:= IC). In general the C-algebras will be complex or real. IH denotes the field of quaternions, IN denotes the set of natural numbers (0 6∈ IN), and for every n ∈ IN ∪ {0} we put INn := { m ∈ IN m ≤ n} . ZZ denotes the group of integers and for every n ∈ IN we put ZZn := ZZ /(n ZZ ) . 2) For every set A, P(A) denotes the set of subsets of A, Pf (A) the set of finite subsets of A, and Card A denotes the cardinal number of A. If f is a function defined on A and B is a subset of A then fB denotes the restriction of f to B. 3) If A, B are sets then AB denotes the set of maps of B in A. 4) For all i, j we denote by δi,j Kronecker's symbol: δi,j :=(cid:26) 1 if 0 if i = j i 6= j . 5) If A, B are topological spaces then C(A, B) denotes the set of continuous maps of A into B. If A is locally compact space and E is a C-algebra then C(A, E) (resp. C0(A, E)) denotes the C-algebra of continuous maps A → E, which are bounded (resp. which converge to 0 at the infinity). J the character- istic function of J i.e. the function on I equal to 1 on J and equal to 0 on I \ J. For i ∈ I we put ei := (δi,j)j∈I ∈ l2(I). 6) For every set I and for every J ⊂ I we denote by eJ := eI 7) If F is an additive group and S is a set then F (S) :=(cid:8) x ∈ F S(cid:12)(cid:12) { s ∈ S xs 6= 0} is finite(cid:9) . 4 8) If E, F are vector spaces in duality then EF denotes the vector space E endowed with the locally convex topology of pointwise convergence on F , i.e. with the weak topology σ(E, F ). 9) If E is a normed vector space then E′ denotes its dual and E# denotes its unit ball: E# := { x ∈ E kxk ≤ 1} . Moreover if E is an ordered Banach space then E+ denotes the convex cone of its positive elements. If E has a unique predual (up to isomor- phisms), then we denote by E this predual and so by E E the vector space E endowed with the locally convex topology of pointwise convergence on E. 10) The expressions of the form "... C-... (resp. ... W-...)", which appear often in this paper, will be replaced by expressions of the form "... C*- ...". 11) If F is a unital C-algebra and A is a subset of F then we denote by 1F the unit of F , by P r F the set of orthogonal projections of F , by Ac := { x ∈ F y ∈ A ⇒ xy = yx} , Re F := { x ∈ F x = x∗ } , and by U n F the set of unitary elements of F . If F is a real C-algebra then ◦F denotes its complexification. 12) If F is a C-algebra then we denote for every n ∈ IN by Fn,n the C- algebra of n × n matrices with entries in F . If T is finite then FT,T has a corresponding signification. 13) Let F be a C-algebra and H, K Hilbert right F -modules. We denote by LF (H, K) the Banach subspace of L(H, K) of adjointable operators, by 1H the identity map H → H which belongs to LF (H) := LF (H, H) . For (ξ, η) ∈ H × K we put η h · ξ i : H −→ K , ζ 7−→ η h ζ ξ i and denote by KF (H) the closed vector subspace of LF (H) generated by { η h · ξ i ξ, η ∈ H }. 5 ... and by H the closed vector subspace of LF (H, K)′ generated by n ](a, ξ)(cid:12)(cid:12)(cid:12) a ∈ F , ξ ∈ Ho n ^(a, ξ, η)(cid:12)(cid:12)(cid:12) (a, ξ, η) ∈ F × H × Ko . ... 14) Let F be a W-algebra and H, K Hilbert right F -modules. We put for a ∈ F and (ξ, η) ∈ H × K, ](a, ξ) : H −→ IK , ζ 7−→ hh ζ ξ i , ai , ^(a, ξ, η) : LF (H, K) −→ IK , u 7−→ hh uξ η i , ai and denote by H the closed vector subspace of the dual H′ of H generated by H is the predual of LF (H) ([C1] Theorem 5.6.3.5 If H is selfdual then b)) and H is the predual of H ([C1] Proposition 5.6.3.3). Moreover a map defined on F is called W-continuous if it is continuous on F F . If G is a W-algebra a C-homomorphism ϕ : F → G is called a W- homomorphism if the map ϕ : F F → G G is continuous; in this case ϕ denotes the pretranspose of ϕ. 15) If F is a C-algebra and (Hi)i∈I a family of Hilbert right F -modules then we put respectively the family h ξi ξi ii∈I is summable in F) Hi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Hi :=( ξ ∈Yi∈I (cid:13)i∈I Hi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) the family h ξi ξi ii∈I is summable in F F) . Hi :=( ξ ∈Yi∈I W (cid:13)i∈I 16) ⊙ denotes the algebraic tensor product of vector spaces. 17) If F, G are W-algebras and H (resp. K) is a selfdual Hilbert right F - module (resp. G-module) then we denote by H ¯⊗K the W-tensor product of H and K, which is a selfdual Hilbert right F ¯⊗G-module ([C2] Definition 2.3). 18) ≈ denotes isomorphic. 6 If T is finite then (by [C1] Theorem 5.6.6.1 f)) LE(H) = ET,T = IKT,T ⊗ E = KE(H) . 1 Preliminaries 1.1 Schur functions DEFINITION 1.1.1 A Schur E-function for T is a map such that f (1, 1) = 1E and f : T × T −→ U n Ec f (r, s)f (rs, t) = f (r, st)f (s, t) for all r, s, t ∈ T . We denote by F(T, E) the set of Schur E-functions for T and put f : T −→ U n Ec , f : T × T −→ U n Ec , t 7−→ f (t, t−1)∗ , (s, t) 7−→ f (t−1, s−1) for every f ∈ F(T, E). Schur functions are also called normalized factor set or multiplier or two- co-cycle (for T with values in U n Ec) in the literature. We present in this subsection only some elementary properties (which will be used in the sequel) in order to fix the notation and the terminology. By the way, U n Ec can be replaced in this subsection by an arbitrary commutative multiplicative group. PROPOSITION 1.1.2 Let f ∈ F(T, E). a) For every t ∈ T , f (t, 1) = f (1, t) = 1E , f (t, t−1) = f (t−1, t) , f (t) = f (t−1) . 7 b) For all s, t ∈ T , f (s, t) f (s) = f (s−1, st)∗ , f (s, t) f (t) = f (st, t−1)∗ . a) Putting s = 1 in the equation of f we obtain so for all r, t ∈ T . Hence f (r, 1)f (r, t) = f (r, t)f (1, t) f (r, 1) = f (1, t) f (t, 1) = f (1, t) = f (1, 1) = 1E . Putting r = t and s = t−1 in the equation of f we get f (t, t−1)f (1, t) = f (t, 1)f (t−1, t) . By the above, f (t, t−1) = f (t−1, t) , f (t) = f (t−1) . b) Putting r = s−1 in the equation of f , by a), f (s, t)f (s−1, st) = f (s−1, s)f (1, t) = f (s)∗ , f (s, t) f (s) = f (s−1, st)∗ . Putting now t = s−1 in the equation of f , by a) again, f (r, s)f (rs, s−1) = f (r, 1)f (s, s−1) = f (s)∗ , f (r, s) f (s) = f (rs, s−1)∗ , f (s, t) f (t) = f (st, t−1)∗ . DEFINITION 1.1.3 We put and Λ(T, E) := { λ : T −→ U n Ec λ(1) = 1E } λ : T −→ U n Ec , δλ : T × T −→ U n Ec , t 7−→ λ(t−1) , (s, t) 7−→ λ(s)λ(t)λ(st)∗ for every λ ∈ Λ(T, E). 8 PROPOSITION 1.1.4 a) F(T, E) is a subgroup of the commutative multiplicative group (U n Ec)T×T such that f∗ is the inverse of f for every f ∈ F(T, E). b) f ∈ F(T, E) for every f ∈ F(T, E) and the map F(T, E) −→ F(T, E), f 7−→ f is an involutive group automorphism. c) Λ(T, E) is a subgroup of the commutative multiplicative group (U n Ec)T , δλ ∈ F(T, E) for every λ ∈ Λ(T, E), and the map δ : Λ(T, E) −→ F(T, E) , λ 7−→ δλ is a group homomorphism with kernel { λ ∈ Λ(T, E) λ is a group homomorphism} such that cδλ = δλ for every λ ∈ Λ(T, E). a) is obvious. b) For r, s, t ∈ T , f (r, s) f (rs, t) = f (s−1, r−1)f (t−1, s−1r−1) = = f (t−1, s−1)f (t−1s−1, r−1) = f (r, st) f (s, t) , so f ∈ F(T, E). For f, g ∈ F(T, E), cf g(s, t) = (f g)(t−1, s−1) = f (t−1, s−1)g(t−1, s−1) = f (s, t)g(s, t) = ( f g)(s, t) , ( f )∗ =cf∗ . f∗(s, t) = f (s, t)∗ = f (t−1, s−1)∗ = f∗(t−1, s−1) =cf∗(s, t) , cf g = f g , c) For r, s, t ∈ T , δλ(r, s)δλ(rs, t) = λ(r)λ(s)λ(rs)∗λ(rs)λ(t)λ(rst)∗ = λ(r)λ(s)λ(t)λ(rst)∗ , 9 δλ(r, st)δλ(s, t) = λ(r)λ(st)λ(rst)∗λ(s)λ(t)λ(st)∗ = λ(r)λ(s)λ(t)λ(rst)∗ so δλ ∈ F(T, E). For λ, µ ∈ F(T, E) and s, t ∈ T , δλ(s, t)δµ(s, t) = λ(s)λ(t)λ(st)∗µ(s)µ(t)µ(st)∗ = = (λµ)(s)(λµ)(t)(λµ)(st)∗ = δ(λµ)(s, t) , (δλ)(δµ) = δ(λµ) , δλ∗(s, t) = λ∗(s)λ∗(t)λ(st) = (δλ(s, t))∗ = (δλ)∗(s, t) , δλ∗ = (δλ)∗ , so δ is a group homomorphism. The other assertions are obvious. PROPOSITION 1.1.5 Let t ∈ T , m, n ∈ ZZ , and f ∈ F(T, E). a) f (tm, tn) = f (tn, tm). b) m ∈ IN =⇒ f (tm, tn) = m−1Qj=0 c) We define λ : ZZ −→ U n Ec , n 7−→ If tp 6= 1 for every p ∈ IN then f (tn+j, t)!(cid:18)m−1Qk=1  n−1Qj=1 −nQj=1 f (tj, t)∗ f (t−j, t) f (tk, t)∗(cid:19). if n ∈ IN if n 6∈ IN . f (tm, tn) = λ(m)λ(n)λ(m + n)∗ for all m, n ∈ ZZ . a) We may assume m ∈ IN because otherwise we can replace t by t−1. Put P (m, n) :⇐⇒ f (tm, tn) = f (tn, tm) , Q(m) :⇐⇒ P (m, n) holds for all n ∈ ZZ . f (tm, tn−m)f (tn, tm) = f (tm, tn)f (tn−m, tm) From 10 it follows for all k ∈ ZZ . P (m, n) ⇐⇒ P (m, n − m) ⇐⇒ P (m, n − km) We prove the assertion by induction. P (m, 0) follows from Proposition 1.1.2 a). By the above for all k ∈ ZZ . Thus Q(1) holds. P (1, 0) ⇐⇒ P (1, k) Assume Q(p) holds for all p ∈ INm−1. Then P (m, p) holds for all p ∈ INm−1 ∪ {0}. Let n ∈ ZZ . There is a k ∈ ZZ such that p := n − km ∈ INm−1 ∪ {0} . By the above P (m, n) holds. Thus Q(m) holds and this finishes the inductive proof. b) We prove the formula by induction with respect to m. By a), the formula holds for m = 1. Assume the formula holds for an m ∈ IN. Since f (tm, t)f (tm+1, tn) = f (tm, tn+1)f (t, tn) we get by a), = m−1Yj=0 f (tm+1, tn) = f (tm, tn+1)f (t, tn)f (tm, t)∗ = f (tk, t)∗! f (tn, t)f (tm, t)∗ = f (tn+1+j, t) m−1Yk=1 = mYj=0 f (tn+j, t) mYk=1 f (tk, t)∗! . Thus the formula holds also for m + 1. c) If m, n ∈ IN then by b), λ(m)λ(n)λ(m + n)∗ = = m−1Yk=1 f (tk, t)∗! n−1Yj=1 f (tj, t)∗ m+n−1Yj=1 f (tj, t) = 11 = m−1Yj=0 f (tn+j, t) m−1Yk=1 f (tk, t)∗! = f (tm, tn) . If m, n ∈ IN, n ≤ m − 1 then by b), If m, n ∈ IN, n ≥ m then by b), λ(m)λ(−n)λ(m − n)∗ = m−n−1Yj=1 f (tj, t) = f (t−j, t) f (tk, t)∗! = f (tm, t−n) . λ(m)λ(−n)λ(m − n)∗ = n−mYj=1 f (t−j, t)∗ = f (t−j, t) f (tk, t)∗! = f (tm, t−n) . f (tj, t)∗ = m−1Yj=1 nYj=1 f (t−n+j, t) m−1Yk=1 = m−1Yj=0 f (tk, t)∗! = m−1Yk=1 nYj=1 f (t−j, t) m−1Yk=1 = nYj=n−m+1 For all m, n ∈ IN put R(m, n) :⇐⇒ f (t−m, t−n) = λ(−m)λ(−n)λ(−m − n)∗ . By the above and by Proposition 1.1.2 a),b), λ(−1)λ(−1)λ(−2)∗ = f (t−1, t)f (t−2, t)∗ = f (t−1)∗f (t, t−2)∗ = f (t−1, t−1) , so R(1, 1) holds. Let now m, n ∈ IN and assume R(m, n) holds. Then λ(−m)λ(−n − 1)λ(−m − n − 1)∗ = f (t−j, t) m+n+1Yj=1 f (t−j, t) n+1Yj=1 f (t−j, t)∗ = = f (t−m, t−n)f (t−n−1, t)f (t−m−n−1, t)∗ = f (t−m, t−n−1) , = mYj=1 so R(m, n) ⇒ R(m, n+1). By symmetry and a), R(m, n) holds for all m, n ∈ IN. 12 COROLLARY 1.1.6 The map Λ( ZZ , E) −→ F( ZZ , E), λ 7−→ δλ is a surjective group homomorphism with kernel { λ ∈ Λ( ZZ , E) n ∈ ZZ =⇒ λ(n) = λ(1)n } . By Proposition 1.1.4 c), only the surjectivity of the above map has to be proved and this follows from Proposition 1.1.5 c). 1.2 E-C-algebras By replacing the scalars with the unital C-algebra E we restrict the cate- gory of C-algebras to the subcategory of those C-algebras which are connected in a certain way with E. The category of unital C-algebras is replaced by the category of E-C-algebras, while the general category of C-algebras is replaced by the category of adapted E-modules. DEFINITION 1.2.1 We call in this paper E-module a C-algebra F en- dowed with the bilinear maps E × F −→ F, F × E −→ F, such that for all α, β ∈ E and x, y ∈ F , (α, x) 7−→ αx , (x, α) 7−→ xα (αβ)x = α(βx) , α(xβ) = (αx)β , x(αβ) = (xα)β , α(xy) = (αx)y , (αx)∗ = x∗α∗ , (xy)α = x(yα) , (xα)∗ = α∗x∗ , α ∈ Ec =⇒ αx = xα , 1E x = x1E = x . If F, G are E-modules then a C-homomorphism ϕ : F → G is called E-linear if for all (α, x) ∈ E × F , ϕ(αx) = α(ϕx) , ϕ(xα) = (ϕx)α . 13 For all (α, x) ∈ E × F , kαxk2 = kx∗α∗αxk ≤ kxk2 kαk2 , so kxαk2 = kα∗x∗xαk ≤ kαk2 kxk2 kαxk ≤ kαkkxk , kxαk ≤ kxkkαk . DEFINITION 1.2.2 An E-C-algebra is a unital C-algebra F for which E is a canonical unital C*-subalgebra such that Ec defined with respect to E coincides with Ec defined with respect to F i.e. for every x ∈ E, if xy = yx for all y ∈ E then xy = yx for all y ∈ F . Every closed ideal of an E-C-algebra is canonically an E-module. Let F, G be E-C-algebras. A map ϕ : F −→ G is called an E-C- homomorphism if it is an E-linear C*-homomorphism . If in addition ϕ is a C-isomorphism then we say that ϕ is an E-C-isomorphism and we use in this case the notation ≈E. A C-subalgebra F0 of F is called E-C- subalgebra of F if E ⊂ F0. With the notation of the above Definition (α − ϕα)ϕx = 0 for all α ∈ E and x ∈ F . Thus ϕ is unital iff ϕα = α for every α ∈ E. The example shows that an E-C-homomorphism need not be unital. IK −→ IK × IK, x 7−→ (x, 0) If we put IT := { z ∈ IC z = 1}, E := C( IT, IC), and x : IT −→ IC , z 7−→ z and if we denote by λ the Lebesgue measure on IT then L∞(λ) is an E-C- algebra, x ∈ U n E, and x is homotopic to 1E in U n L∞(λ) but not in U n C( IT, IC). DEFINITION 1.2.3 We denote by CE (resp. by C1 algebras for which the morphisms are the E-C-homomorphisms (resp. unital E-C-homomorphisms). E) the category of E-C- the PROPOSITION 1.2.4 Let F be an E-module. 14 a) We denote by F the vector space E × F endowed with the bilinear map ((α, x), (β, y)) 7−→ (αβ, αy + xβ + xy) (E × F ) × (E × F ) −→ E × F, and with the conjugate linear map E × F −→ E × F, (α, x) 7−→ (α∗, x∗) . F is an involutive unital algebra with (1E, 0) as unit. b) The maps π : F −→ E , (α, x) 7−→ α , λ : E −→ F , α 7−→ (α, 0) , ι : F −→ F , x 7−→ (0, x) are involutive algebra homomorphisms such that π ◦ λ is the identity map of E, λ and ι are injective, and λ and π are unital. If there is a norm on F with respect to which it is a C-algebra (in which case such a norm is unique), then we call F adapted. We denote by ME the category of adapted E-modules for which the morphism are the E-linear C-homomorphisms. c) If F is adapted then F is an E-C-algebra by using canonically the injec- tion λ and for all α ∈ E and x ∈ F , kαk ≤ k(α, x)k ≤ kαk + kxk , k(α, 0)(0, x)k ≤ kαkkxk , k(0, x)k = kxk ≤ 2k(α, x)k , k(0, x)(α, 0)k ≤ kxkkαk . In particular F (identified with ι(F )) is a closed ideal of F . d) If E and F are C-subalgebras of a C-algebra G in such a way that the structure of E-module of F is inherited from G then ϕ : F −→ E × G , (α, x) 7−→ (α, α + x) is an injective involutive algebra homomorphism, ϕ( F ) is closed, F is adapted, and for all α ∈ E and x ∈ F , k(α, x)kE×F = sup{kαk ,kα + xk} . In particular every closed ideal of an E-C-algebra is adapted and CE is a full subcategory of ME. 15 e) A closed ideal G of an adapted E-module F , which is at the same time an E-submodule of F , is adapted. f ) If F is unital then it is adapted and F −→ IR+, (α, x) 7−→ sup{kαk ,kα1F + xk} is the C-norm of F . g) If y, F kαy − yαk = 0 lim for all α ∈ E+, where F denotes the canonical approximate unit of F , then F is adapted and F −→ IR+, (α, x) 7−→ sup(kαk , lim sup y, F kαy + xk) is the C-norm of F . In particular F is adapted if E is commutative. h) If F is an adapted E-module then (with the notation of b)) 0 −→ F ι−→ F π −→ λ ←− E −→ 0 is a split exact sequence in the category ME. a) and b) are easy to see. c) Since λ and ι are injective and , π(α, x) = α , (α, x) = (α, 0) + (0, x) , (α, 0)(0, x) = (0, αx) , (0, x)(α, 0) = (0, xα) we get the first and the last two inequalities as well as the identity k(0, x)k = kxk. It follows k(0, x)k ≤ k(α, x)k + k(α, 0)k = k(α, x)k + kλπ(α, x)k ≤ ≤ k(α, x)k + k(α, x)k = 2k(α, x)k . 16 d) It is easy to see that ϕ is an injective involutive algebra homomorphism. Let (α, x) ∈ ϕ( F ). There are sequences (αn)n∈IN and (xn)n∈IN in E and F , respectively, such that (αn, αn + xn) = (α, x) . lim n→∞ It follows αn ∈ E , α = lim n→∞ x − α = lim n→∞ (α, x) = ϕ(α, x − α) ∈ ϕ( F ) . Thus ϕ( F ) is closed, which proves the assertion by pulling back the norm of E × G. xn ∈ F , e) By c), F is a closed ideal of F so G is a closed ideal of F (use an approximate unit of F ). Since G is an E-submodule of F its structure of E- module is inherited from F . By d), G is adapted. f) The map F −→ E × F, (α, x) 7−→ (α, α1F + x) is an isomorphism of involutive algebras and so we can pull back the norm of E × F . g) It is easy to see that the above map is a norm. Since sup{kαk , 1 2 kxk} ≤ k(α, x)k ≤ kαk + kxk for all (α, x) ∈ E×F , F endowed with this norm is complete. For (α, x) ∈ E×F , (α, x)∗(α, x) = (α∗α, α∗x + x∗α + x∗x) , k(α, x)∗(α, x)k = sup{kαk2 , lim sup y, F kα∗αy + α∗x + x∗α + x∗xk} . For y ∈ F # + , 1 2 + x)∗(αy 2 α∗α − α∗αy (cid:13)(cid:13)(cid:13)(αy ≤(cid:13)(cid:13)(cid:13)y 1 1 2 + x) − (α∗αy + α∗x + x∗α + x∗x)(cid:13)(cid:13)(cid:13) ≤ 2(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)y 2 − x∗α(cid:13)(cid:13)(cid:13) 2 α∗x − α∗x(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)x∗αy 1 1 1 17 so lim y, F(cid:13)(cid:13)(cid:13)(αy 1 2 + x)∗(αy 1 2 + x) − (α∗αy + α∗x + x∗α + x∗x)(cid:13)(cid:13)(cid:13) = 0 . 1 2 maps F into itself and Since the map F+ → F+, y 7→ y we have by the above, kαy + xk2 = kyα∗αy + yα∗x + x∗αy + x∗xk k(α, x)k2 = sup(kαk2 , lim sup 2) = y, F (cid:13)(cid:13)(cid:13)αy 2 + x(cid:13)(cid:13)(cid:13) 2 + x)(cid:13)(cid:13)(cid:13)) = = sup(kαk2 , lim sup y, F (cid:13)(cid:13)(cid:13)(αy 2 + x)∗(αy 1 1 1 = sup{kαk2 , lim sup y, F kα∗αy + α∗x + x∗α + x∗xk} = k(α, x)∗(α, x)k . Thus the above norm is a C-norm and F is adapted. h) ι is an injective E-C-homomorphism and its image is equal to Ker π. COROLLARY 1.2.5 Let F an E-module, G a C-algebra, and ⊗σ the spatial tensor product. a) F ⊗σ G is in a natural way an E-module the multiplication being given by α(x ⊗ y) = (αx) ⊗ y , for all α ∈ E, x ∈ F , and y ∈ G. (x ⊗ y)α = (xα) ⊗ y b) If F is an E-C-algebra and G is unital then the map E −→ F ⊗σ G, α 7−→ α ⊗ 1G is an injective C-homomorphism. In particular, the E-module F ⊗σ G is an E-C-algebra. c) If F is an adapted E-module then the E-module F ⊗σ G is adapted and k(α, z)k = sup{kαk ,kα + zk} for all (α, z) ∈ E × (F ⊗σ G). 18 d) If F is an adapted E-module and G := C0(Ω) for a locally compact space Ω then C0(Ω, F ) is adapted and k(α, x)k = sup{kαk ,kαeΩ + xk} for all (α, x) ∈ E × C0(Ω, F ). a) and b) are easy to see. c) If G denotes the unitization of G then by b), F ⊗σ G is an E-C-algebra and F ⊗σ G is a closed ideal of it, so the assertion follows from Proposition 1.2.4 d),e). d) follows from c). PROPOSITION 1.2.6 a) If F, G are E-modules and ϕ : F → G is an E-linear C-homomorphism then the map ϕ : F −→ G , (α, x) 7−→ (α, ϕx) is an involutive unital algebra homomorphism, injective or surjective if ϕ is so. If F = G and if ϕ is the identity map then ϕ is also the identity map. b) Let F1, F2, F3 be E-modules and let ϕ : F1 → F2 and ψ : F2 → F3 be E-linear C-homomorphisms. Then z}{ψ ◦ ϕ = ψ ◦ ϕ. PROPOSITION 1.2.7 Let G be an E-module, F an E-submodule of G which is at the same time an ideal of G, and ϕ : G → G/F the quotient map. a) G/F has a natural structure of E-module and ϕ is E-linear. b) If G is adapted then G/F is also adapted. Moreover if ψ : G → G/F denotes the quotient map (where F is identified to {(0, x) x ∈ F}) then there is an E-C-isomorphism θ : z}{G/F → G/F such that ψ = θ ◦ ϕ. 19 a) is easy to see. z}{G/F and let x, y ∈−1 b) Let (α, z) ∈ ϕ (z). Then ψ(α, x) = ψ(α, y) and we put θ(α, z) := ψ(α, x). It is straightforward to show that θ is an isomorphism of involutive algebras. By pulling back the norm of G/F with respect to θ we see that G/F is adapted. LEMMA 1.2.8 Let {(Fi)i∈I , (ϕij )i,j∈I)} be an inductive system in the cate- gory of C-algebras, {F, (ϕi)i∈I} its inductive limit, G a C-algebra, for ev- ery i ∈ I, ψi : Fi → G a C-homomorphism such that ψj ◦ ϕji = ψi for all i, j ∈ I, i ≤ j, and ψ : F → G the resulting C-homomorphism. If Ker ψi ⊂ Ker ϕi for every i ∈ I then ψ is injective. Let i ∈ I. Since Ker ϕi ⊂ Ker ψi is obvious, we have Ker ϕi = Ker ψi. Let ρ : Fi → Fi/Ker ψi be the quotient map and ϕ′i : Fi/Ker ψi −→ F , the injective C-homomorphisms with ϕi = ϕ′i ◦ ρ , Then ψ′i : Fi/Ker ψi −→ G ψi = ψ′i ◦ ρ . ψ′i ◦ ρ = ψi = ψ ◦ ϕi = ψ ◦ ϕ′i ◦ ρ . For x ∈ Fi, since ψ′i and ϕ′i are norm-preserving, kρxk =(cid:13)(cid:13)ψ′iρx(cid:13)(cid:13) =(cid:13)(cid:13)ψϕ′iρx(cid:13)(cid:13) ≤(cid:13)(cid:13)ϕ′iρx(cid:13)(cid:13) = kρxk , kψϕixk =(cid:13)(cid:13)ψϕ′iρx(cid:13)(cid:13) =(cid:13)(cid:13)ϕ′iρx(cid:13)(cid:13) = kϕixk . Thus ψ preserves the norms on ∪i∈I ϕi(Fi). Since this set is dense in F , ψ is injective. PROPOSITION 1.2.9 Let {(Fi)i∈I , (ϕij)i,j∈I} be an inductive system in the category ME and let (F, (ϕi)i∈I ) be its inductive limit in the category of E- modules (Proposition 1.2.4 c)). a) F is adapted. 20 b) Let (G, (ψi)i∈I ) be the inductive limit in the category C1 E of the inductive system {( Fi)i∈I , ( ϕij )i,j∈I} (Proposition 1.2.6 a),b)) and let ψ : G → F be the unital C-homomorphism such that ψ ◦ ψi = ϕi for every i ∈ I . Then ψ is an E-C-isomorphism. a) Put F0 :=( (α, x) ∈ F (cid:12)(cid:12) α ∈ E, x ∈[i∈I ϕi(Fi)) , p : F0 −→ IR+ , (α, x) 7−→ inf { k(α, xi)k i ∈ I, xi ∈ Fi, ϕixi = x} . F0 is an involutive unital subalgebra of F . p is a norm and by Proposition 1.2.4 c), q(α, x) := lim (α,y)∈F0 y→x p(α, y) exists and kαk ≤ q(α, x) ≤ kαk + kxk , for every (α, x) ∈ F . kxk ≤ 2q(α, x) Let (α, x) ∈ F0. Let further i ∈ I, xi, yi ∈ Fi with ϕixi = x, ϕiyi = α∗x + x∗α + x∗x. Then (0, ϕi(α∗xi + x∗i α + x∗i xi − yi)) = ϕi((α, xi)∗(α, xi) − (α∗α, yi)) = 0 so inf i≤j kϕji(α∗xi + x∗i α + x∗i xi − yi)k = 0 . For ǫ > 0 there is a j ∈ I, i ≤ j, with kϕji(α∗xi + x∗i α + x∗i xi − yi)k < ǫ . We get p(α, x)2 ≤ k(α, ϕjixi)k2 = k(α, ϕjixi)∗(α, ϕjixi)k = = k(α∗α, α∗ϕjixi + (ϕjix∗i )α + ϕji(x∗i xi))k = = k(α∗α, ϕji(α∗xi + x∗i α + x∗i xi))k ≤ ≤ k(α∗α, ϕjiyi)k + k(0, ϕji(α∗xi + x∗i α + x∗i xi − yi))k < k(α∗α, ϕjiyi)k + ǫ . By taking the infimum on the right side it follows, since ǫ is arbitrary, p(α, x)2 ≤ p(α∗α, α∗x + x∗α + x∗x) = p((α, x)∗(α, x)) 21 and this shows that p is a C-norm. It is easy to see that q is a C-norms. By the above inequalities, F endowed with the norm q is complete, i.e. F is a C-algebra and F is adapted. b) Let i ∈ I and let (α, x) ∈ Ker ϕi. Then 0 = ϕi(α, x) = (α, ϕix) so α = 0 , ϕix = 0 , inf j∈I, j≥ikϕjixk = 0 , kψi(α, x)k = inf k ϕji(0, x)k = k(0, ϕjix)k = kϕjixk , j∈I, j≥ik ϕji(0, x)k = 0 , (α, x) ∈ Ker ψi . By Lemma 1.2.8, ψ is injective. Let (β, y) ∈ F and let ε > 0. There are i ∈ I and x ∈ Fi with kϕix − yk < ε. Then ψψi(β, x) = ϕi(β, x) = (β, ϕix) , kψψi(β, x) − (β, y)k = k ϕi(β, x) − (β, y)k = kϕix − yk < ε . Thus ψ(G) is dense in F and ψ is surjective. Hence ψ is a C-isomorphism. COROLLARY 1.2.10 We put ΦE(F ) := F for every E-module F and simi- larly ΦE(ϕ) := ϕ for every E-linear C-homomorphism ϕ. a) ΦE is a covariant functor from the category ME in the category C1 E. b) The categories C1 E and ME possess inductive limits and the functor ΦE is continuous with respect to the inductive limits. a) follows from Proposition 1.2.6. b) follows from Proposition 1.2.9. Remark. The category CE does not possess inductive limits in general. This happens for instance if ϕij = 0 for all i, j ∈ I. 22 1.3 Some topologies T is only a set in this subsection If the group T is infinite then different topologies play a certain role in the construction of the projective representations of T . It will be shown that all these topologies conduct to the same construction, but the use of them simplifies the manipulations. We introduce the following notation in order to unify the cases of C- algebras and (resp. W-algebras). DEFINITION 1.3.1 := (cid:13) e(cid:13) e⊗ := ⊗ gX :=X W := (cid:13) ) , (resp. e(cid:13) (resp. e⊗ := ¯⊗ ) , EX ) . (resp.gX := If T is a Hausdorff topology on LE(H) then for every G ⊂ LE(H), GT denotes ¯G denotes the closure of G the set G endowed with the relative topology T and in LE(H)T. Moreover TP denotes the sum with respect to T. T LEMMA 1.3.2 For x ∈ E, by the above identification of E with LE( E), is well-defined and belongs to LE(H). xe⊗1K : H −→ H , ξ 7−→ (xξt)t∈T a) The map ϕ : E −→ LE(H) , is an injective unital C-homomorphism. x 7−→ xe⊗1K 23 b) Assume E is a W-algebra. Then for every (a, ξ, η) ∈ E × H × H, the family (ξt a η∗t )t∈T is summable in EE and for every x ∈ E, ξt a η∗t+ . D ϕx , ^(a, ξ, η)E =* x , EXt∈T Thus ϕ is a W-homomorphism ([C1] Theorem 5.6.3.5 d)) with ϕ ^(a, ξ, η) = ξt a η∗t , EXt∈T where ϕ denotes the pretranspose of ϕ. c) If we consider E as a canonical unital C-subalgebra of LE(H) by using the embedding of a) then LE(H) is an E-C-algebra. a) follows from [L] page 37 (resp. [C3] Proposition 1.4). b) We have D x ¯⊗1K , ^(a, ξ, η)E = hh (x ¯⊗1K )ξ η i , ai = EXt∈T η∗t x ξt , a+ = Thus the family (ξt a η∗t )t∈T is summable in EE and h η∗t x ξt , ai =Xt∈T =Xt∈T D ϕx , ^(a, ξ, η)E =* x , h x , ξt a η∗t i . ξt a η∗t+ . EXt∈T If ϕ′ : LE(H) → E′ denotes the transpose of ϕ then ϕ′ ^(a, ξ, η) = ξt a η∗t ∈ E . By continuity ϕ′ .. LE(H)! ⊂ E and ϕ is a unital W-homomorphism. z } { EXt∈T 24 c) Let x ∈ Ec and ξ, η ∈ LE(H). By [C1] Proposition 5.6.3.17 d), (cid:10) (xe⊗1K )ξ(cid:12)(cid:12) η(cid:11) =gXt∈T xη∗t ξt = xgXt∈T =gXt∈T η∗t ((xe⊗1K )ξ)t =gXt∈T η∗t ξt = xh ξ η i . η∗t xξt = Thus for u ∈ LE(H), (cid:10) u(xe⊗1K )ξ(cid:12)(cid:12) η(cid:11) =(cid:10) (xe⊗1K )ξ(cid:12)(cid:12) u∗η(cid:11) = xh ξ u∗η i = xh uξ η i , u(xe⊗1K ) = (xe⊗1K )u , and so xe⊗1K ∈ LE(H)c. DEFINITION 1.3.3 We put for all ξ, η ∈ H (resp. and a ∈ E+) pξ,η : LE(H) −→ IR+ , X 7−→ kh Xξ η ik , (resp. pξ,η,a : LE(H) −→ IR+ , X 7−→ hh Xξ η i , ai) , pξ : LE(H) −→ IR+ , X 7−→ kXξk = kh Xξ Xξ ik1/2 , (resp. pξ,a : LE(H) −→ IR+ , X 7−→ hh Xξ Xξ i , ai1/2) , qξ : LE(H) −→ IR+ , X 7−→ pξ(X∗) , (resp. qξ,a : LE(H) −→ IR+ , X 7−→ pξ,a(X∗)) . and denote, respectively, by T1 , T2 , T3 the topologies on LE(H) generated by the set of seminorms { pξ ξ ∈ H } , { pξ,η ξ, η ∈ H } , (cid:16)resp. n pξ,η,a ξ, η ∈ H, a ∈ E+o(cid:17) , (cid:16)resp. n pξ,a ξ ∈ H, a ∈ E+o(cid:17) , (cid:16)resp. n pξ,a ξ ∈ H, a ∈ E+o ∪n qξ,a ξ ∈ H, a ∈ E+o(cid:17) . { pξ ξ ∈ H } ∪ { qξ ξ ∈ H } , Moreover k·k denotes the norm topology on LE(H). 25 Of course T2 ⊂ T3. In the C-case, T2 is the topology of pointwise conver- gence. If E is finite-dimensional then the C-case and the W-case coincide. PROPOSITION 1.3.4 Let X ∈ LE(H) and ξ, η ∈ H (resp. and a ∈ E). a) pξ,η(X) = pη,ξ(X∗) (resp. pξ,η,a(X) = pη,ξ,a(X∗)). b) pξ,η(X) ≤ pξ(X)kηk. c) If E is a W-algebra and a = xa is the polar representation of a then pξx,η,a(X) =(cid:12)(cid:12)(cid:12)D X , ^(a, ξ, η)E(cid:12)(cid:12)(cid:12) ≤ pξx,a(X)hh η η i , ai1/2 . d) If Y, Z ∈ LE(H) then pξ,η(Y XZ) = pZξ,Y ∗η(X) pξ(Y XZ) ≤ kY k pZξ(X) (resp. pξ,η,a(Y XZ) = pZξ,Y ∗η,a(X)) , (resp. pξ,a(Y XZ) ≤ kY k pZξ,a(X)) . a) From it follows h Xξ η i = h ξ X∗η i = h X∗η ξ i∗ pξ,η(X) = kh Xξ η ik = kh X∗η ξ ik = pη,ξ(X∗) , (resp. pξ,η,a(X) = hh X∗η ξ i , ai = pη,ξ,a(X∗)) . b) pξ,η(X) = kh Xξ η ik ≤ pξ(X)kηk. c) We have pξx,η,a(X) = hh X(ξx) η i , ai = hh Xξ η i x , ai = = hh Xξ η i , xai = hh Xξ η i , ai =(cid:12)(cid:12)(cid:12)D X , ^(a, ξ, η)E(cid:12)(cid:12)(cid:12) . By Schwarz' inequality ([C1] Proposition 2.3.3.9), hh X(ξx) η i , ai 2 ≤ hh X(ξx) X(ξx)i , aihh η η i , ai , 26 so pξx,η,a(X) ≤ pξx,a(X)h h η η i , ai1/2 . d) The first equation follows from pξ,η(Y XZ) = kh Y XZξ η ik = kh XZξ Y ∗η ik = pZξ,Y ∗η(X) (resp. pξ,η,a(Y XZ) = hh Y XZξ η i , ai = = hh XZξ Y ∗η i , ai = pZξ,Y ∗η,a(X)) and the second from pξ(Y XZ) = kY XZξk ≤ kY kkXZξk = kY k pZξ(X) (resp. pξ,a(Y XZ) = hh Y XZξ Y XZξ i , ai1/2 ≤ ≤ kY khh XZξ XZξ i , ai1/2 = kY k pZξ,a(X)) . LEMMA 1.3.5 Let n ∈ IN and (xi)i∈INn a family in E. Then Xi∈INn xi!∗ Xi∈INn xi! ≤ n Xi∈INn x∗i xi . We prove the relation by induction with respect to n. By [C1] Corollary 4.2.2.4 and by the hypothesis of the induction, xi!∗ Xi∈INn Xi∈INn = x∗nxn + Xi∈INn−1 ≤ x∗nxn + Xi∈INn−1 xi! =x∗n + Xi∈INn−1 x∗ixn + Xi∈INn−1 xi = (x∗nxi + x∗i xn) + Xi∈INn−1 xi ∗ Xi∈INn−1 xi ≤ x∗i xi = n Xi∈INn (x∗nxn + x∗i xi) + (n − 1) Xi∈INn−1 x∗i xi . 27 LEMMA 1.3.6 Let n ∈ IN, x ∈ En,n, and for every j ∈ INn put Then ηj := (δji1E)i∈INn ∈ (cid:13)i∈INn E . kxk ≤ √n sup j∈INn kxηjk . ∗Xj∈INn xijξj ≤ ξ∗j x∗ijxijξj = , by Lemma 1.3.5, E!# For ξ ∈ (cid:13)i∈INn h (xξ)i (xξ)i i = Xi∈INn h xξ xξ i = Xi∈INn ≤ n Xi∈INn Xj∈INn xijξj Xj∈INn (xijξj)∗(xijξj) = n Xi∈INn Xj∈INn x∗ijxij! ξj . ξ∗j Xi∈INn = n Xj∈INn (xηj)i = Xk∈INn h xηj xηj i = Xi∈INn xikηjk = xij , (xηj)∗i (xηj)i = Xi∈INn ξ∗j h xηj xηj i ξj ≤ n Xj∈INn ξ∗j ξj ≤ n sup j∈INn kxηjk2 . j∈INn kxηjk2 Xj∈INn kxk2 ≤ n sup x∗ijxij , kxηjk2 ξ∗j ξj ≤ j∈INn kxηjk2 1E , For i, j ∈ INn, so h xξ xξ i ≤ n Xj∈INn ≤ n sup COROLLARY 1.3.7 28 a) The map LE(H)T1 −→ LE(H)T1 , X 7−→ X∗ is continuous. In particular ReLE(H) is a closed set of LE(H)T1. b) T1 ⊂ T2 ⊂ T3 ⊂ norm topology. c) If E is a W-algebra then the identity map is continuous so is compact. LE(H) ... H −→ LE(H)T1 LE(H)# T1 = LE(H)# ... H d) For Y, Z ∈ LE(H) and k ∈ {1, 2}, the map LE(H)Tk −→ LE(H)Tk , X 7−→ Y XZ is continuous. e) LE(H)T3 is complete in the C-case. f ) If T is finite then T2 is the norm topology in the C-case. g) KE(H) is dense in LE(H)T3 . a) follows from Proposition 1.3.4 a). b) T1 ⊂ T2 follows from Proposition 1.3.4 b),c). T2 ⊂ T3 ⊂ norm topology is trivial. c) follows from Proposition 1.3.4 c) (and [C1] Theorem 5.6.3.5 a)). d) follows from Proposition 1.3.4 d). e) Let F be a Cauchy filter on LE(H)T3 . Put ξ 7−→ lim Y : H −→ H , X,F (Xξ) , Z : H −→ H , ξ 7−→ lim X,F (X∗ξ) , 29 where the limits are considered in the norm topology of H. For ξ, η ∈ H, h Y ξ η i = lim X,F h Xξ η i = lim X,F h ξ X∗η i = h ξ Zη i , so Y, Z ∈ LE(H) and Z = Y ∗. Thus F converges to Y in LE(H)T3 and LE(H)T3 is complete. f) follows from b) and Lemma 1.3.6. g) Let X ∈ LE(H) and ξ ∈ H. For every S ∈ Pf (T ) put PS :=Xs∈S es h · es i ∈ P r KE(H) and let FT be the upper section filter or Pf (T ). Then PSX ∈ KE(H) for every S ∈ Pf (T ) and lim S,FT PSXξ = Xξ in H (resp. in H H) ([C1] Proposition 5.6.4.1 e) (resp. [C1] Proposition 5.6.4.6 c))). Thus lim S,FT PSX = X with respect to the topology T2. Since the same holds for X∗, it follows that X belongs to the closure of KE(H) in LE(H)T3. Remark. The inclusions in b) can be strict as it is known from the case E := IK. LEMMA 1.3.8 Let G be a W-algebra and F a C-subalgebra of G. Then the following are equivalent. a) F generates G as a W-algebra. b) F # is dense in G# G . c) F is dense in G G. a =⇒ b follows from [C1] Corollary 6.3.8.7. 30 b =⇒ c is trivial. c =⇒ a follows from [C1] Corollary 4.4.4.12 a). PROPOSITION 1.3.9 Let G be a W-algebra, F a C-subalgebra of G gen- erating it as W-algebra, I a set, and F , L := (cid:13)i∈I M := G . W (cid:13)i∈I a) M is the extension of L to a selfdual Hilbert right G-module ([C2] Propo- sition 1.3 f)) and L# is dense in M # M . b) If we denote for every X ∈ LF (L) by ¯X ∈ LG(M ) its unique extension ([C3] Proposition 1.4 a)) then the map LF (L) −→ LG(M ), X 7−→ ¯X is an injective C-homomorphism and its image is dense in LG(M ) ... M . c) The map is continuous. LF (L)# T2 −→ LG(M )# T1 , X 7−→ ¯X a) By Lemma 1.3.8 a ⇒ b, F # is dense in G# and G is the extension of F to a selfdual Hilbert right G-module ([C3] Corollary 1.5 a2 ⇒ a1). By [C3] Proposition 1.8, M is the extension of L to a selfdual Hilbert right G-module. By [C3] Corollary 1.5 a1 ⇒ a2, L# is dense in M # so F # is dense in G# G M G . b) By a) and [C3] Proposition 1.4 e), the map LF (L) −→ LG(M ), X 7−→ ¯X is an injective C-homomorphism. By [C3] Proposition 1.9 b), its image is dense in LG(M ) ... M . c) Denote by N the vector subspace of ... M generated by n ^(a, ξ, η)(cid:12)(cid:12)(cid:12) (a, ξ, η) ∈ G × L × Lo . 31 By a) and [C3] Proposition 1.9 a), N is dense in = LG(M )# N . LG(M )# T1 ... M so by Corollary 1.3.7 c), For (a, ξ, η) ∈ G+ × L × L and X ∈ LF (L), by Proposition 1.3.4 c), pξ,η,a( ¯X) = (cid:10)(cid:10) ¯Xξ(cid:12)(cid:12) η(cid:11) , a(cid:11) = = hh Xξ η i , ai ≤ pξx,a(X)hh η η i , ai 1 2 , where a = xa is the polar representation of a, so the map , X 7−→ ¯X T2 −→ LG(M )# LF (L)# T1 is continuous. LEMMA 1.3.10 Let n ∈ IN, ξ ∈ (cid:13)i∈INn E, and x := [ξiδj,1]i,j∈INn ∈ En,n . Then kxk = kξk. For η ∈ (cid:13)i∈INn E and i ∈ INn, xijηj = Xj∈INn (xη)i = Xj∈INn h (xη)i (xη)i i = Xi∈INn ξiδj,1ηj = ξiη1 , h xη xη i = Xi∈INn = η∗1 Xi∈INn h ξiη1 ξiη1 i = Xi∈INn ξ∗i ξi! η1 = η∗1 h ξ ξ i η1 ≤ kξk2 η∗1η1 , η∗1ξ∗i ξiη1 = kxk ≤ kξk . On the other hand if we put ζ := (δi,11E)i∈INn then for i ∈ INn, kxηk2 ≤ kξk2 kη1k2 ≤ kξk2 kηk2 , xijζj = Xj∈INn (xζ)∗i (xζ)i = Xi∈INn (xζ)i = Xj∈INn h xζ xζ i = Xi∈INn ξiδj,11E = ξi , ξ∗i ξi = h ξ ξ i , kxk ≥ kxζk = kξk , kxk = kξk . 32 LEMMA 1.3.11 Let F, G be unital C-algebras, ϕ : F → G a surjective C-homomorphism, I a set, L := g(cid:13)i∈I F ≈ Fe⊗l2(I) , M := g(cid:13)i∈I G ≈ Ge⊗l2(I) , and for every ξ ∈ L put ξ := (ϕξi)i∈I . a) If ξ, η ∈ L and x ∈ F then ξ ∈ M , (cid:13)(cid:13)(cid:13) ξ(cid:13)(cid:13)(cid:13) ≤ kξk , g(ξx) = (ξ)ϕx , D ξ(cid:12)(cid:12)(cid:12) ηE = ϕh ξ η i . b) For every η ∈ M there is a ξ ∈ L with ξ = η, kξk = kηk. c) In the W-case the map L L −→ M M , ξ 7−→ ξ is continuous. a) For J ∈ Pf (I), Xi∈J h ϕξi ϕηi i =Xi∈J (ϕηi)∗(ϕξi) = ϕXi∈J η∗i ξi . It follows ξ ∈ M, (cid:13)(cid:13)(cid:13) ξ(cid:13)(cid:13)(cid:13) ≤ kξk , D ξ(cid:12)(cid:12)(cid:12) ηE = ϕh ξ η i. Moreover for i ∈ I, (fξx)i = ϕ(ξx)i = ϕ(ξix) = (ϕξi)(ϕx) = ξi(ϕx), fξx = ξ(ϕx) . b) Case 1 { i ∈ I ηi 6= 0} is finite For simplicity we assume { i ∈ I ηi 6= 0} = INn for some n ∈ IN. We put θ : Fn,n −→ Gn,n , [xij]i,j∈INn 7−→ [ϕxij ]i,j∈INn . 33 θ is obviously a surjective C-homomorphism. So if we put y := [ηiδj,1]i,j∈INn ∈ Gn,n , then there is an x ∈ Fn,n with θx = y, kxk = kyk ([K] Theorem 10.1.7). If we put ξ : I −→ F , i 7−→(cid:26) xi1 0 if if i ∈ INn i ∈ I \ INn and z := [xijδj1]i,j∈INn ∈ Fn,n then θz = [ϕ(xij δj1)]i,j∈INn = [yijδj1]i,j∈INn = y and by [C1] Theorem 5.6.6.1 a), kzk ≤ kxk. We get for i ∈ INn, ξi = ϕξi = ϕxi1 = yi1 = ηi . By a) and Lemma 1.3.10, kξk = kzk ≤ kxk = kyk = kηk =(cid:13)(cid:13)(cid:13)ξ(cid:13)(cid:13)(cid:13) ≤ kξk , Case 2 η arbitrary in the W-case kξk = kηk . We may assume kηk = 1. We put for every J ∈ Pf (I), ηJ : I −→ G , i 7−→(cid:26) ηi 0 if if i ∈ J i ∈ I \ J . By Case 1, for every J ∈ Pf (I) there is a ξJ ∈ L with ξJ = ηJ and kξJk = kηJk ≤ 1. Let F be an ultrafilter on Pf (I) finer than the upper section filter of Pf (I). By [C1] Proposition 5.6.3.3 a ⇒ b, ξ := lim J,F ξJ exists in L# L . For i ∈ I, ξi = ϕξi = ϕ lim J,F (ξJ )i = lim J,F ϕ(ξJ )i = ηi so ξ = η. By a), 1 = kηk =(cid:13)(cid:13)(cid:13) ξ(cid:13)(cid:13)(cid:13) ≤ kξk ≤ 1, so kξk = kηk. 34 Case 3 η arbitrary in the C-case We put for every J ∈ Pf (I) and every ζ ∈ M , ζJ : I −→ G , if if i ∈ J i ∈ I \ J . i 7−→(cid:26) ζi 0 Moreover we denote by FI the upper section filter of Pf (I), set M0 := { ζ ∈ M { i ∈ I ζi 6= 0} is finite} , and denote by M the vector subspace of KG(M ) generated by the set { ζ1 h · ζ2 i ζ1, ζ2 ∈ M0 } . Let G be the vector subspace of KF (L) generated by the set { αh · β i α, β ∈ L} . G is an involutive subalgebra of KF (L). Let (αq)q∈Q, (βq)q∈Q be finite families in L such that Let further α′, β′ ∈ M0. By Case 1, there are α, β ∈ L with α = α′ , β = β′ and we get by a), αq h · βq i = 0 . Xq∈Q αqD β′(cid:12)(cid:12) βqE(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) α′+ =Xq∈Q(cid:10) αq α′(cid:11)D β′(cid:12)(cid:12) βqE = Xq∈Q h αq α ih β βq i = h αq αiD β(cid:12)(cid:12)(cid:12) βqE = ϕXq∈Q =Xq∈Q αq h · βq i β(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = ϕXq∈Q α+ = 0 . αqD · βqE = 0 . Xq∈Q 35 It follows ([C1] Proposition 5.6.4.1 e)) Thus the linear map ψ : G −→ KG(M ) , Xq∈Q αq h · βq i 7−→Xq∈Q αqD · βqE is well-defined and it is easy to see (by a)) that ψ is an involutive algebra homomorphism. Step 1 kψk ≤ 1; we extend ψ by continuity to a map ψ : KF (L) → KG(M ) Let u :=Xq∈Q αq h · βq i ∈ G and let ζ ∈ M # (ψu)ζ =Xq∈Q 0 . By Case 1, there is an α ∈ L# with α = ζ. By a), { αqD α βqE =Xq∈Q αqϕh α βq i =Xq∈Q k(ψu)ζk = kfuαk ≤ kuαk ≤ kuk . ^z } αq h α βq i = fuα , Since M0 is dense in M ([C1] Proposition 5.6.4.1 e)), it follows kψuk ≤ kuk , kψk ≤ 1 . Step 2 M is dense in KG(M ) Let α, β ∈ M . By [C1] Proposition 5.6.4.1 e), β = lim J,FI α = lim J,FI αJ , βJ so by [C1] Proposition 5.6.5.2 a), α h · β i = lim J,FI αJ h · βJ i , which proves the assertion. Step 3 ψ is a surjective C-homomorphism 36 By Step 1, ψ is a C-homomorphism. Since its image contains M (by Case 1) it is surjective by Step 2. Step 4 The assertion Let j ∈ I. By Step 3 and [K] Theorem 10.1.7 (and [C1] Proposition 5.6.5.2 a)), there is a u ∈ KF (L) with ψu = η h · 1G ⊗ ej i , kuk = kη h · 1G ⊗ ej ik = kηk . From ψ(u((1F ⊗ ej)h · 1F ⊗ ej i)) = (η h · 1G ⊗ ej i)((1G ⊗ ej)h · 1G ⊗ ej i) = = η h · 1G ⊗ ej i , kηk = kη h · 1G ⊗ ej ik ≤ ku((1F ⊗ ej)h · 1F ⊗ ej i)k ≤ ≤ kukk(1F ⊗ ej)h · 1F ⊗ ej ik = kuk = kηk , ku((1F ⊗ ej)h · 1F ⊗ ej i)k = kηk we see that we may assume Then u = u((1F ⊗ ej)h · 1F ⊗ ej i) . u = (u(1F ⊗ ej))h · 1F ⊗ ej i . If we put ξ := u(1F ⊗ ej) ∈ L then u = ξ h · 1F ⊗ ej i, kηk = kuk = kξk, η h · 1G ⊗ ej i = ψu = ξ h · 1G ⊗ ej i) , η = η h 1G ⊗ ej 1G ⊗ ej i = ξ h 1G ⊗ ej 1G ⊗ ej i = ξ . c) Let (a, η0) ∈ G × M . By b), there is a ξ0 ∈ L with ξ0 = η0. By a), for ξ ∈ L, We put D ξ , ^(a, η0)E =DD ξ(cid:12)(cid:12)(cid:12) η0E , aE =DD ξ(cid:12)(cid:12)(cid:12) ξ0E , aE = = h ϕh ξ ξ0 i , ai = hh ξ ξ0 i , ϕai =D ξ , ^( ϕa, ξ0)E . θ : L −→ M , ξ 7−→ ξ and denote by θ′ : M′ → L′ its transpose. By the above, θ′^(a, η0) ∈ L. Since θ′ is continuous, θ′( M ) ⊂ L and this proves the assertion. 37 PROPOSITION 1.3.12 We use the notation of Lemma 1.3.11. a) If X ∈ LF (L) and ξ ∈ L with ξ = 0 then fXξ = 0; we define where ξ ∈ L with ξ = η (Lemma 1.3.11 b)). η 7−→ fXξ , X : M −→ M , b) For every X ∈ LF (L), X belongs to LG(M ) and the map LF (L) −→ LG(M ), X 7−→ X is a surjective C*-homomorphism continuous with respect to the topolo- gies Tk with k ∈ {1, 2, 3}. c) For ξ, η ∈ L, and η h · ξ i = ηD · ξE ^z } { KG(M ) =n X(cid:12)(cid:12)(cid:12) X ∈ KF (L)o . a) For i ∈ I, ϕξi = ξi = 0 so by Lemma 1.3.11 a), ^X(eiξi) = ^(Xei)ξi = ^(Xei)ϕξi = 0 . By [C1] Proposition 5.6.4.1 e) (resp. [C1] Proposition 5.6.4.6 c) and [C1] Propo- sition 5.6.3.4 c)), X(eiξi) X(eiξi) eiξi! =Xi∈I Xξ = X Xi∈I resp. Xξ = X eiξi = LXi∈I LXi∈I ^z } Xi∈I X(eiξi) =Xi∈I fXξ = { ^X(eiξi) = 0 so by Lemma 1.3.11 a) (resp. c)), 38 b) For X, Y ∈ LF (L) and ξ, η ∈ L, by Lemma 1.3.11 a), . { X(eiξi) = MXi∈I ^z } LXi∈I  resp. fXξ = ^X(eiξi) = 0 D X ξ(cid:12)(cid:12)(cid:12) ηE =D fXξ(cid:12)(cid:12)(cid:12) ηE = ϕh Xξ η i = = ϕh ξ X∗η i =D ξ(cid:12)(cid:12)(cid:12) gX∗ηE =D ξ(cid:12)(cid:12)(cid:12) fX∗ ηE , X Y ξ = XfY ξ = ^X(Y ξ) = ^(XY )ξ = gXY ξ . is a C-homomorphism. By Lemma 1.3.11 b), X ∈ LG(M ), ( X)∗ = X∗, and X Y = gXY , i.e. the map For X ∈ LF (L) and ξ, η ∈ L (resp. and a ∈ M+), by Lemma 1.3.11 a), p ξ,η( X) =(cid:13)(cid:13)(cid:13)D X ξ(cid:12)(cid:12)(cid:12) ηE(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)D fXξ(cid:12)(cid:12)(cid:12) ηE(cid:13)(cid:13)(cid:13) = kϕh Xξ η ik ≤ pξ,η(X) (resp. p ξ,η,a(X) =(cid:12)(cid:12)(cid:12)DD X ξ(cid:12)(cid:12)(cid:12) ηE , aE(cid:12)(cid:12)(cid:12) = h ϕh Xξ η i , ai = = hh Xξ η i , ϕai = pξ,η, ϕa(X)) , so by Lemma 1.3.11 b), the map is continuous with respect to the topology T1. The proof for the other topologies is similar. c) For ζ ∈ L, by Lemma 1.3.11 a), (η h · ξ i)ζ = ^z } ^z } { { η h · ξ i ζ = = η ϕh ζ ξ i = ηD ζ(cid:12)(cid:12)(cid:12) ξE =(cid:16)ηD · ξE(cid:17) ζ ^z } η h ζ ξ i = { η h · ξ i = ηD · ξE . ^z } { so by Lemma 1.3.11 b), The last assertion follows now from b). 39 2 Main Part Throughout this section we fix f ∈ F(T, E) 2.1 The representations We present here the projective representation of the groups and its main properties. DEFINITION 2.1.1 We put for every t ∈ T and ξ ∈ H, ut : E −→ H , ζ 7−→ ζ ⊗ et , Vtξ : T −→ E , s 7−→ f (t, t−1s)ξ(t−1s) . If we want to emphasize the role of f then we put V f t instead of Vt. For x ∈ E, (xe⊗1K)Vtξ : T −→ E , s 7−→ f (t, t−1s)xξ(t−1s) . PROPOSITION 2.1.2 Let s, t ∈ T , x ∈ E, ζ ∈ E, and ξ ∈ H. a) Vtξ ∈ H. c) Vt(ζ ⊗ es) = (f (t, s)ζ) ⊗ ets. b) VsVt = (f (s, t)e⊗1K )Vst. d) Vt(xe⊗1K ) = (xe⊗1K)Vt. V ∗t = ( f (t)e⊗1K )Vt−1 . f ) (xe⊗1K)Vt(ζ ⊗ es) = (f (t, s)xζ) ⊗ ets. e) Vt ∈ U n LE(H) , g) If T is infinite and F denotes the filter on T of cofinite subsets, i.e. F := { S S ∈ P(T ) , T \ S ∈ Pf (T )} , then in LE(H)T1 . lim t,F Vt = 0 40 a) For R ∈ Pf (T ), Xr∈R h (Vtξ)r (Vtξ)r i =Xr∈R(cid:10) f (t, t−1r)ξt−1r(cid:12)(cid:12) f (t, t−1r)ξt−1r(cid:11) = =Xr∈R h ξt−1r ξt−1r i =Xr∈R h ξr ξr i ≤ h ξ ξ i so Vtξ ∈ H. b) For r ∈ T , (VsVtξ)r = f (s, s−1r)(Vtξ)s−1r = f (s, s−1r)f (t, t−1s−1r)ξt−1s−1r = = f (s, t)f (st, t−1s−1r)ξt−1s−1r = f (s, t)(Vstξ)r = ((f (s, t)e⊗1K )Vstξ)r so so VsVt = (f (s, t)e⊗1K )Vst . c) For r ∈ T , (Vt(ζ ⊗ es))r = f (t, t−1r)(ζ ⊗ es)t−1r = = δs,t−1rf (t, t−1r)ζ = δr,tsf (t, s)ζ = ((f (t, s)ζ) ⊗ ets)r Vt(ζ ⊗ es) = (f (t, s)ζ) ⊗ ets . d) We have (Vt(xe⊗1K)ξ)s = f (t, t−1s)((xe⊗1K)ξ)t−1s = f (t, t−1s)xξt−1s = ((xe⊗1K )Vtξ)s so e) For η ∈ H, by Proposition 1.1.2 a),b), Vt(xe⊗1K ) = (xe⊗1K )Vt . h (Vtξ)s ηs i =gXs∈T(cid:10) f (t, t−1s)ξt−1s(cid:12)(cid:12) ηs(cid:11) = h f (t, r)ξr ηtr i =gXr∈TD ξr f (t)f (t−1, tr)ηtrE = h Vtξ η i =gXs∈T =gXr∈T 41 =gXr∈TD ξr ((( f (t)e⊗1K )Vt−1 )η)rE =D ξ (( f (t)e⊗1K )Vt−1 )ηE V ∗t Vt = ( f (t)e⊗1K)Vt−1 Vt = ( f (t)e⊗1K)(f (t−1, t)e⊗1K )Vt−1t = idH , so Vt ∈ LE(H) with V ∗t = ( f (t)e⊗1K )Vt−1. By b) and d), VtV ∗t = Vt( f (t)e⊗1K)Vt−1 = ( f (t)e⊗1K )VtVt−1 = = ( f (t)e⊗1K )(f (t, t−1)e⊗1K )Vtt−1 = idH . f) follows from c). g) Let us consider first the C-case. Let ξ, η ∈ H, t ∈ T , and ε > 0. There is an S ∈ Pf (T ) such that(cid:13)(cid:13)ηeT\S(cid:13)(cid:13) < ε. By e), (cid:10) Vtξ ηeT\S(cid:11) ≤ kVtξk(cid:13)(cid:13)ηeT\S(cid:13)(cid:13) ≤ εkξk pξ,η(Vt) = h Vtξ η i ≤ h Vtξ ηeS i + (cid:10) Vtξ ηeT\S(cid:11) < h Vtξ ηeS i + ε . so From h Vtξ ηeS i =Xs∈S η∗s f (t, t−1s)ξt−1s it follows t,F h Vtξ ηeS i = 0 , lim lim t,F pξ,η(Vt) = 0 . The W-case can be proved similarly. Remark. By e), T1 cannot be replaced by T2 in g). PROPOSITION 2.1.3 Let s, t ∈ T . a) ut ∈ LE( E, H) , b) u∗sut = δs,t1E. u∗t = h · 1E ⊗ et i. c) usu∗t = 1Ee⊗(h · et i es). 42 d) uru∗r = idH . T2Pr∈T a) For ζ ∈ E and ξ ∈ H, h utζ ξ i = h ζ ⊗ et ξ i =gXs∈T ξ∗s (ζ ⊗ et)s = ξ∗t ζ = h ζ ξt i ut ∈ LE( E, H) , u∗t ξ = ξt = h ξ 1E ⊗ et i . b) For ζ ∈ E, by a), u∗sutζ = u∗s(ζ ⊗ et) = h ζ ⊗ et 1E ⊗ es i = δs,tζ so so u∗sut = δs,t1E. c) For ζ ∈ E and r ∈ T , by a), usu∗t (ζ ⊗ er) = usδr,tζ = δr,t(ζ ⊗ es) = = ζ ⊗ h er et i es = (1Ee⊗(h · et i es))(ζ ⊗ er) , [C1] Proposition 5.6.3.4 c))) usu∗t = 1Ee⊗(h · et i es). d) For ξ ∈ H (resp. and a ∈ E+) and S ∈ Pf (T ), by c), h ξ ξ i(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xt∈T\S utu∗t − idH! = (utu∗t − idH )ξ+ , a+1/2 utu∗t − idH! =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) pξ Xt∈S resp. pξ,a Xt∈S (utu∗t − idH )ξ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xt∈S = ( Xt∈T\S hh ξ ξ i , ai)1/2! =*Xt∈S so (by a) and [C1] Proposition 5.6.4.1 e) (resp. and [C1] Proposition 5.6.4.6 c), 1/2 = and the assertion follows. 43 PROPOSITION 2.1.4 Let s, t ∈ T and x ∈ E. a) Vsut = ustf (s, t). b) u∗sVt = f (t, t−1s)u∗t−1s. c) (xe⊗1K)ut = utx. d) xu∗t = u∗t (xe⊗1K ). a) For ζ ∈ E, by Proposition 2.1.2 c), Vsutζ = Vs(ζ ⊗ et) = (f (s, t)ζ) ⊗ est = ustf (s, t)ζ so Vsut = ustf (s, t) . b) For ζ ∈ E and r ∈ T , by Proposition 2.1.2 c) and Proposition 2.1.3 a), u∗sVt(ζ ⊗ er) = u∗s((f (t, r)ζ) ⊗ etr) = δs,trf (t, r)ζ = = δt−1s,rf (t, t−1s)ζ = f (t, t−1s)u∗t−1s(ζ ⊗ er) so u∗sVt = f (t, t−1s)u∗t−1s. c) For ζ ∈ E, (xe⊗1K )utζ = (xe⊗1K)(ζ ⊗ et) = (xζ) ⊗ et = utxζ so (xe⊗1K )ut = utx. d) follows from c). DEFINITION 2.1.5 We put for all s, t ∈ T (Proposition 2.1.3 a)) ϕs,t : LE(H) −→ LE( E) ≈ E , X 7−→ u∗sXut and set Xt := ϕt,1X for every X ∈ LE(H). PROPOSITION 2.1.6 Let s, t ∈ T . 44 a) ϕs,t is linear with kϕs,tk = 1. b) For X ∈ LE(H) and x, y ∈ E, h (ϕs,tX)x y i = h X(x ⊗ et) y ⊗ es i . ϕs,t : LE(H)T1 −→ E (resp. E E) c) The map is continuous. d) ϕt,t is involutive and completely positive. e) For r ∈ T and x ∈ E, f ) If (xr)r∈T ∈ E(T ) and ϕs,t((xe⊗1K )Vr) = δs,rtf (r, t)x . X :=Xr∈T (xre⊗1K )Vr then ϕs,tX = f (st−1, t)xst−1 , Xt = xt . g) For X ∈ LE(H) and x, y ∈ E, ϕs,t((xe⊗1K )X(ye⊗1K )) = x(ϕs,tX)y , ((xe⊗1K)X(ye⊗1K ))t = xXty . a) follows from Proposition 2.1.3 a),b). b) We have h (ϕs,tX)x y i = h u∗sXutx y i = h Xutx usy i = h X(x ⊗ et) y ⊗ es i . c) The C-case 45 By b), for X ∈ LE(H), kϕs,tXk = kh (ϕs,tX)1E 1E ik = = kh X(1E ⊗ et) 1E ⊗ es ik = p1E⊗et,1E⊗es(X) . The W-case Let a ∈ E and let a = xa be its polar representation. By b), for X ∈ LE(H), h ϕs,tX , ai = hh (ϕs,tX)1E 1E i , xai = hh (ϕs,tX)x 1E i , ai = = hh X(x ⊗ et) 1E ⊗ es i , ai = px⊗et,1E⊗es,a(X) . d) For X ∈ LE(H), (ϕt,tX)∗ = (u∗t Xut)∗ = u∗t X∗ut = ϕt,t(X∗) so ϕt,t is involutive. For n ∈ IN, X ∈ ((LE(H))n,n)+, and ζ ∈ En, Xi∈INn Xj∈INn ((ϕt,tXij)ζj)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = Xi,j∈INn ζi+ = Xi,j∈INn h u∗t Xijutζj ζi i = h Xijutζj utζi i ≥ 0 ([C1] Theorem 5.6.6.1 f) and [C1] Theorem 5.6.1.11 c1 ⇒ c2) so ϕt,t is com- pletely positive ([C1] Theorem 5.6.6.1 f) and [C1] Theorem 5.6.1.11 c2 ⇒ c1). e) By Proposition 2.1.4 a),d) and Proposition 2.1.3 b), ϕs,t((xe⊗1K )Vr) = u∗s(xe⊗1K)Vrut = xu∗sVrut = xu∗surtf (r, t) = δs,rtf (r, t)x . f) By e) (and Proposition 1.1.2 a)), ϕs,tX =Xr∈T ϕs,t((xre⊗1K )Vr) =Xr∈T 46 δs,rtf (r, t)xr = f (st−1, t)xst−1 , Xt = ϕt,1X = f (t, 1)xt = xt . g) By Proposition 2.1.4 c),d), ϕs,t((xe⊗1K )X(ye⊗1K )) = u∗s(xe⊗1K )X(ye⊗1K )ut = = xu∗sXuty = x(ϕs,tX)y . DEFINITION 2.1.7 We put R(f ) :=(Xt∈T (xte⊗1K )Vt(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (xt)t∈T ∈ E(T )) , T3 S(f ) := R(f ) , Sk·k(f ) := k·k R(f ) . Moreover we put SC(f ) := S(f ) in the C-case and SW (f ) := S(f ) in the W-case. If F is a subset of E then we put S(f, F ) := { X ∈ S(f ) t ∈ T =⇒ Xt ∈ F } and use similar notation for the other S. By Proposition 2.1.2 b),d),e), R(f ) is an involutive unital E-subalgebra of LE(H) (with V1 as unit). In particular Sk·k(f ) is an E-C-subalgebra of LE(H). If T is finite then R(f ) = S(f ). By Corollary 1.3.7 e), SC(f )T3 is complete. PROPOSITION 2.1.8 For X ∈ R(f ) and s, t ∈ T , ϕs,tX = f (st−1, t)Xst−1 . T1 Let F be a filter on R(f ) converging to X in the T1-topology. By Proposition 2.1.6 c),f) (and Corollary 1.3.7 d)), ϕs,tX = lim Y,F ϕs,tY = lim Y,F f (st−1, t)Yst−1 = f (st−1, t) lim Y,F Yst−1 = = f (st−1, t) lim Y,F ϕst−1,1Y = f (st−1, t)ϕst−1,1X = f (st−1, t)Xst−1 . 47 THEOREM 2.1.9 Let X ∈ R(f ). T1 a) If (xt)t∈T is a family in E such that then Xt = xt for every t ∈ T . In particular, if T is finite then the map ET −→ S(f ), (xt ⊗ 1K)Vt X = T1Xt∈T (xte⊗1K )Vt x 7−→Xt∈T is bijective and E-linear (Proposition 2.1.2 d)). b) We have X = T3Xt∈T (Xte⊗1K )Vt ∈ S(f ) . c) (X∗)t = f (t)(Xt−1 )∗ for every t ∈ T and X∗ = T3Xt∈T ((Xt)∗e⊗1K )V ∗t ∈ T3 R(f ) . T1 T2 R(f )= d) S(f ) = R(f ). e) For ξ ∈ H and t ∈ T , (Xξ)t =gXs∈T f (s, s−1t)Xsξs−1t . f ) If T is finite and if we identify LE(H) with ET,T then X is identified with the matrix [f (st−1, t)Xst−1 ]s,t∈T , and for every r ∈ T , Vr is identified with the matrix [f (st−1, t)δs,rt]s,t∈T . 48 g) If X, Y ∈ S(f ) and t ∈ T then XY ∈ S(f ) and f (t, s)∗XtsY ∗s , XsY ∗s . (X∗Y )t =gXs∈T (XY )t =gXs∈T f (s, t)∗X∗s Yst , (X∗Y )1 =gXs∈T X∗s Ys , f (s, s−1t)XsYs−1t , (XY ∗)t =gXs∈T (XY ∗)1 =gXs∈T x 7−→ xe⊗1K h) The map E −→ S(f ), is an injective unital C-homomorphism and so S(f ) is an E-C-subalge- bra of LE(H) and ReS(f ) is closed in S(f )T1. In the W-case, SW (f ) is the W-subalgebra of LE(H) generated by R(f ) and R(f )# is dense in SW (f )# = SW (f )# , which is compact. ... H T1 i) If E is a W-algebra then SC(f ) may be identified canonically with a unital C-subalgebra of SW (f ) by using the map of Proposition 1.3.9 b). By this identification SC(f ) generates SW (f ) as W-algebra. j) If F is a closed ideal of E (resp. of E E ) then S(f, F ) is a closed ideal of S(f ) (resp. of S(f ) ). ..z}{S(f ) k) If F is a unital C-subalgebra of E such that f (s, t) ∈ F for all s, t ∈ T then S(f, F ) is a unital C-subalgebra of S(f ) and the map S(f, F ) −→ S(g), X 7−→ is an injective C*-homomorphism, where T3Xt∈T (Xte⊗1K )V g t g : T × T −→ U n F c , (s, t) 7−→ f (s, t) . This map induces a C-isomorphism Sk·k(f, F ) → Sk·k(g). l) (X, Y ) ∈ ◦z}{S(f )+=⇒ (X1, Y1) ∈ ◦E+. 49 a) By Proposition 2.1.6 c),e), Xt = ϕt,1X =gXs∈T ϕt,1((xse⊗1K)Vs) =gXs∈T δt,sf (s, 1)xs = xt . b&c&d Step 1 X = T2Pt∈T (Xte⊗1K )Vt By Proposition 2.1.3 d), Corollary 1.3.7 d), Proposition 2.1.8, and Proposi- tion 2.1.4 b),d), T2Xs∈T T2Xt∈T usu∗sXutu∗t = us(ϕs,tX)u∗t = usf (st−1, t)Xst−1 u∗t = utu∗t! = T2Xt∈T T2Xs∈T usu∗s! X T2Xt∈T = X = T2Xs∈T T2Xt∈T T2Xs∈T T2Xr∈T T2Xs∈T usu∗s(Xre⊗1K )Vr = = = T2Xs∈T T2Xr∈T usXrf (r, r−1s)u∗r−1s = usXru∗sVr = T2Xr∈T T2Xs∈T (Xte⊗1K )Vt! = usu∗s T2Xt∈T T2Xs∈T T2Xt∈T (Xte⊗1K )Vt . Step 2 b&c&d By Step 1, Corollary 1.3.7 a), and Proposition 2.1.2 d),e) (and Proposition 1.1.2 a)), (Xse⊗1K )Vs!∗ X∗ = T1Xs∈T T1Xs∈T (X∗se⊗1K )V ∗s = T1Xr∈T T1Xs∈T (( f (r)X∗r−1)e⊗1K )Vr ∈ (X∗se⊗1K )( f (s)e⊗1K)Vs−1 = = = T1 R(f ) . 50 By a), (X∗)t = f (t)(Xt−1 )∗ . By Step 1 and Proposition 2.1.2 e) (and Proposition 1.1.2 a)), X∗ = T2Xt∈T T2Xt∈T ((X∗)te⊗1K )Vt = T2Xt∈T ((Xt−1 )∗e⊗1K)V ∗t−1 = = . ((Xt−1 )∗e⊗1K )( f (t)e⊗1K )Vt = T2Xt∈T ((Xt)∗e⊗1K )V ∗t T3Xt∈T ((Xt)∗e⊗1K)V ∗t ∈ S(f ) . X∗ = Together with Step 1 this proves X = T3Xt∈T (Xte⊗1K )Vt ∈ S(f ) , T1 T2 In particular S(f ) = R(f )= R(f ). e) By b) and Corollary 1.3.7 b), in the C-case, (Xse⊗1K )Vs! ξ(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (Xξ)t = T1Xs∈T Xsf (s, s−1t)ξs−1t =Xs∈T =Xs∈T The proof is similar in the W-case. 1E ⊗ et+ =Xs∈T(cid:10) (Xse⊗1K )Vsξ(cid:12)(cid:12) 1E ⊗ et(cid:11) = f (s, s−1t)Xsξs−1t . f) For ξ ∈ H and s ∈ T , by e), (Xξ)s =Xt∈T f (t, t−1s)Xtξt−1s =Xr∈T f (sr−1, r)Xsr−1ξr . g) By b), Corollary 1.3.7 b),d), and Proposition 2.1.2 b),d), XY = T2Xs∈T T2Xt∈T (Xse⊗1K )Vs(Yte⊗1K )Vt = (Xte⊗1K )Vt! = (Xse⊗1K )Vs! T2Xt∈T T2Xt∈T T2Xs∈T (Xse⊗1K )(Yte⊗1K)VsVt = = T2Xs∈T 51 = Since by d), = T2Xt∈T T2Xs∈T (Xse⊗1K )(Yte⊗1K )(f (s, t)e⊗1K )Vst = T2Xr∈T T2Xs∈T ((f (s, s−1r)XsYs−1r)e⊗1K )Vr . T2Xr∈T ((f (s, s−1r)XsYs−1r)e⊗1K )Vr ∈ S(f ) for every s ∈ T we get XY ∈ S(f ), again by d). By Corollary 1.3.7 b) and Proposition 2.1.6 c),e), ϕt,1((f (s, s−1r)XsYs−1r)e⊗1K )Vr = f (s, s−1t)XsYs−1t . f (s, s−1t) f (s)(Xs−1)∗Ys−1t = f (s, t)∗X∗s Yst , f (s, s−1t)Xs f (s−1t)(Yt−1s)∗ = By the above, c), and Proposition 1.1.2 b), (XY )t = ϕt,1(XY ) =gXs∈TgXr∈T δt,rf (r, 1)f (s, s−1r)XsYs−1r =gXs∈T =gXs∈TgXr∈T f (s, s−1t)(X∗)sYs−1t =gXs∈T (X∗Y )t =gXs∈T f (s−1, t)∗(Xs−1)∗Ys−1t =gXs∈T =gXs∈T f (s, s−1t)Xs(Y ∗)s−1t =gXs∈T (XY ∗)t =gXs∈T f (t, t−1s)∗Xs(Yt−1s)∗ =gXs∈T =gXs∈T (X∗Y )1 =gXs∈T X∗s Ys , It follows by Proposition 1.1.2 a), f (t, s)∗XtsY ∗s . (XY ∗)1 =gXs∈T XsY ∗s . h) By c) and g), S(f ) is an involutive unital subalgebra of LE(H). Be- ing closed (resp. closed in LE(H) ... (d) and Corollary 1.3.7 c))) it is a C- subalgebra of LE(H) (resp. generated by R(f ) [C1] Theorem 5.6.3.5 b) and [C1] H 52 Corollary 4.4.4.12 a) and by [C1] Corollary 6.3.8.7 R(f )# is dense in SW (f )# , T1 which is compact by Corollary 1.3.7 c)). The assertion concerning E follows from Proposition 2.1.2 d) and Lemma 1.3.2 c). By Corollary 1.3.7 a), ReS(f ) is a closed set of S(f )T1. i) The assertion follows from h), Proposition 1.3.9 b), and Lemma 1.3.8 c) ⇒ a). j) For X ∈ S(f, F ), Y ∈ S(f ), and t ∈ T , by g), (XY )t, (Y X)t ∈ S(f, F ) so S(f, F ) is an ideal of S(f ). The closure properties follow from Proposition 2.1.6 c). k) By c) and g), S(f, F ) is a unital involutive subalgebra of S(f ) and by Proposition 2.1.6 c), S(f, F ) is a C-subalgebra of S(f ). The last assertion follows from the fact that the image of the map contains R(g). l) There are U, V ∈ S(f ) with (X, Y ) = (U, V )∗(U, V ) = (U∗,−V ∗)(U, V ) = (U∗U + V ∗V, U∗V − V ∗U ) . For t ∈ T , 0 ≤ (Ut, Vt)∗(Ut, Vt) = (U∗t ,−V ∗t )(Ut, Vt) = (U∗t Ut + V ∗t Vt, U∗t Vt − V ∗t Ut) . By g), so (U∗t Ut + V ∗t Vt) , (U∗t Vt − V ∗t Ut) X1 = (U∗U + V ∗V )1 =gXt∈T Y1 = (U∗V − V ∗U )1 =gXt∈T (X1, Y1) =gXt∈T (U∗t Ut + V ∗t Vt, U∗t Vt − V ∗t Ut) ∈ ◦E+ . Remark. It may happen that by the identification of i), SC(f ) 6= SW (f ) (Remark of Proposition 2.1.23). COROLLARY 2.1.10 53 a) If (xt)t∈T is a family in E such that (kxtk)t∈T is summable then is norm summable in LE(H) and b) The set ((xte⊗1K )Vt)t∈T (xte⊗1K)Vt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤Xt∈T (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xt∈T A :=( X ∈ S(f ) Xt∈T k(X∗)tk =Xt∈T Xt∈T k(XY )tk ≤ Xt∈T Xt∈T kxtk . kXtk < ∞) kXtk , kXtk! Xt∈T kYtk! is a dense involutive unital subalgebra of Sk·k(f ) with for all X, Y ∈ A. c) A endowed with the norm A −→ IR+, X 7−→Xt∈T kXtk is an involutive Banach algebra and Sk·k(f ) is its C-hull. a) For S ∈ Pf (T ), by Proposition 2.1.2 e), (xte⊗1K )Vt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤Xt∈S(cid:13)(cid:13)xte⊗1K(cid:13)(cid:13)kVtk =Xt∈S (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xt∈S kxtk and the assertion follows. b) By Theorem 2.1.9 c), X∗ ∈ S(f ) and k(X∗)tk = k(Xt−1 )∗k = kXt−1k 54 By Theorem 2.1.9 g), XY ∈ S(f ) and for all t ∈ T so Xt∈T k(XY )tk =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)gXs∈T k(XY )tk ≤Xt∈TXs∈T kXsk Xt∈T =Xs∈T kXt−1k =Xt∈T k(X∗)tk =Xt∈T f (s, s−1t)XsYs−1t(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤Xs∈T kXsk Xt∈T kXskkYs−1tk =Xs∈T kXtk! Xt∈T kYtk! = Xt∈T for every t ∈ T so Xt∈T kXtk . kXskkYs−1tk kYs−1tk! = kYtk! . c) is easy to see. Remark. There may exist X ∈ Sk·k(f ) for which ((Xte⊗1K )Vt)t∈T is not norm summable, as it is known from the theory of trigonometric series (see Proposition 3.5.1). In particular the inclusion A ⊂ Sk·k(f ) may be strict. COROLLARY 2.1.11 Let F be a unital C*-algebra and τ : E → F a posi- tive continuous (resp. W-continuous) unital trace. a) τ ◦ ϕ1,1 is a positive continuous (resp. W-continuous) unital trace. b) If τ is faithful then τ ◦ ϕ1,1 is faithful and V1 is finite. c) In the W-case, SW (f ) is finite iff E is finite. a) Let X, Y ∈ S(f ). By Theorem 2.1.9 g) (and Proposition 1.1.2 a)), τ ϕ1,1(XY ) = τ gXt∈T f (t, t−1)XtYt−1! = τ gXt∈T f (t, t−1)Xt−1 Yt! = 55 =gXt∈T τ (f (t, t−1)Xt−1 Yt) =gXt∈T τ (f (t, t−1)YtXt−1 ) = τ gXt∈T = τ ϕ1,1(Y X) . f (t, t−1)YtXt−1! = Thus τ ◦ ϕ1,1 is a trace which is obviously positive, continuous (resp. W- continuous), and unital (Proposition 2.1.6 c),d)). b) By Theorem 2.1.9 g), ϕ1,1 is faithful, so τ◦ϕ is also faithful. Let X ∈ S(f ) with X∗X = V1. By a), so τ ϕ1,1(XX∗) = τ ◦ ϕ1,1(X∗X) = τ ϕ1,1V1 = 1F τ ϕ1,1(V1 − XX∗) = 1F − 1F = 0 , V1 = XX∗ , and V1 is finite. c) By b), if E is finite then SW (f ) is also finite. The reverse implication follows from the fact that E ¯⊗1K is a unital W-subalgebra of SW (f ) (Theorem 2.1.9 h)). COROLLARY 2.1.12 Assume T finite and for every x′ ∈ (E′)T put a) ex′ ∈ S(f )′ and for every x′ ∈ (E′)T and the map sup ex′ : S(f ) −→ IK , X 7−→Xt∈T(cid:10) Xt , x′t(cid:11) . t∈T(cid:13)(cid:13)x′t(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)ex′(cid:13)(cid:13)(cid:13) ≤Xt∈T(cid:13)(cid:13)x′t(cid:13)(cid:13) x′ 7−→ ex′ ϕ(x′x) = (ϕx′)(x ⊗ 1K ) ϕ : (E′)T −→ S(f )′ , ϕ(xx′) = (x ⊗ 1K )(ϕx′) , is an isomorphism of involutive vector spaces such that ( [C1] Proposition 2.2.7.2) for every x ∈ E and x′ ∈ (E′)T . 56 b) If E is a W-algebra then the map (at)t∈T 7−→ (at)t∈T is an isomorphism of involutive vector spaces such that ψ : ( E)T −→ .. z}{S(f ) , ψ(xa) = (x ⊗ 1K )(ψa) , for every x ∈ E and a ∈ ( E)T . ψ(ax) = (ψa)(x ⊗ 1K ) COROLLARY 2.1.13 Assume T finite and let M be a Hilbert right S(f )- module. M endowed with the right multiplication M × E −→ M, (ξ, x) 7−→ ξ(x ⊗1K ) and with the inner-product M × M −→ E, (ξ, η) 7−→ h ξ η i1 is a Hilbert right E-module denoted by fM , LS(f )(M ) is a unital C-subalgebra of LE(fM ), and M is selfdual if fM is so. By Proposition 2.1.6 d),g) and Theorem 2.1.9 g),l), for X, Y ∈ S(f ) and x ∈ E, ϕ1,1(X(x ⊗1K )) = (ϕ1,1X)x , X ≥ 0 =⇒ ϕ1,1X ≥ 0 , ◦z}{S(f )+=⇒ (ϕ1,1X, ϕ1,1Y ) ∈ ◦E+ , (X, Y ) ∈ inf { kϕ1,1Xk X ∈ S(f )+ ,kXk = 1} > 0 and the assertion follows from Proposition 2.1.6 a),c),d) and [C1] Proposition 5.6.2.5 a),c),d). COROLLARY 2.1.14 Let n ∈ IN and let ϕ : S(f ) → En,n be an E-C- homomorphism. Then (ϕVt)i,j ∈ Ec for all t ∈ T and all i, j ∈ INn. For x ∈ E, by Proposition 2.1.2 d) and Theorem 2.1.9 h), x(ϕVt) = ϕ(xe⊗1K)(ϕVt) = ϕ((xe⊗1K)Vt) = = ϕ(Vt(xe⊗1K )) = (ϕVt)ϕ(xe⊗1K) = (ϕVt)x so (ϕVt)i,j ∈ Ec. 57 COROLLARY 2.1.15 Let S be a group and g ∈ F(S,S(f )). If we put h : (T × S) × (T × S) −→ U n S(f )c , ((t1, s1), (t2, s2)) 7−→ then h ∈ F(T × S,S(f )). (f (t1, t2)e⊗1K )g(s1, s2) The assertion follows from Theorem 2.1.9 h). COROLLARY 2.1.16 Let X ∈ S(f ) (resp. X ∈ Sk·k(f )). a) For every S ⊂ T , and (resp. k·kXs∈S (Xse⊗1K)Vs ∈ Sk·k(f )) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) S ∈ Pf (T )) < ∞ . T3Xs∈S (Xse⊗1K )Vs ∈ S(f ) γ := sup((cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xt∈S (Xte⊗1K)Vt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) b) We put for every α ∈ l∞(T ) αX : T −→ E , t 7−→ αtXt . Then αX ∈ S(f ) (resp. αX ∈ Sk·k(f )) for every α ∈ l∞(T ) and the map l∞(T ) −→ S(f ) (resp. Sk·k(f )), α 7−→ αX is norm-continuous. c) Assume E is a W-algebra and let l∞(T, E) be the C-direct product of the family (E)t∈T , which is a W-algebra ([C1] Proposition 4.4.4.21 a)). We put for every α ∈ l∞(T, E), αX : T −→ E , t 7−→ αtXt . Then αX ∈ SW (f ) for every α ∈ l∞(T, E) and the map l∞(T, E) −→ SW (f ), α 7−→ αX is continuous and W-continuous. 58 a) In the C-case the family ((Xs ⊗ 1K)Vs)s∈S is summable since SC(f )T3 is complete. By Banach-Steinhaus Theorem, γ is finite. In the W-case the summability follows now from Corollary 1.3.7 b),c) and Theorem 2.1.9 b). b) Let G be the vector subspace { α ∈ l∞(T ) α(T ) is finite} of l∞(T ). By a), the map G −→ S(f ) (resp. Sk·k(f )), α 7−→ αX is well-defined, linear, and continuous. The assertion follows by continuity. Lemma 1.3.2 b) (and Theorem 2.1.9 b)), c) Let x ∈ E, S ⊂ T , and α := xeS. For ξ, η ∈ H and a ∈ E, by a) and αX , η∗t x((eS X)ξ)t , a+ = (a, ξ, η)+ = hh αXξ η i , ai =* EXt∈T ^z } { ((eS X)ξ)taη∗t+ . h x , ((eSX)ξ)taη∗t i =* x , EXt∈T =Xt∈T Let G be the involutive subalgebra { α ∈ l∞(T, E) α(T ) is finite} of l∞(T, E) and let ¯G be its norm-closure in l∞(T, E), which is a C-subalgebra of l∞(T, E). By [C1] Proposition 4.4.4.21 a), G is dense in l∞(T, E) F , where F := l∞(T, E). Let α ∈ l∞(T, E)# and let F be a filter on G# converging to α in l∞(T, E) F ([C1] Corollary 6.3.8.7). By the above (and by Theorem 2.1.9 h)), βX = αX lim β,F and so αX ∈ SW (f ). The assertion follows. in SW (f ) .. z } { SW (f ) COROLLARY 2.1.17 Let S be a subgroup of T . Put fS := f(S × S) , KS := l2(S) , G := { X ∈ S(f ) t ∈ T \ S =⇒ Xt = 0} . a) fS ∈ F(S, E). b) G is an E-C-subalgebra of S(f ). 59 and the map c) For every X ∈ G, the family ((Xse⊗1KS )V fS T3Xs∈S (Xse⊗1KS )V fS ϕ : G −→ S(fS) , X 7−→ s s )s∈S is summable in LE(KS )T3 is an injective E-C-homomorphism. d) If X ∈ G ∩ Sk·k(f ) then ϕX ∈ Sk·k(fS) and the map G ∩ Sk·k(f ) −→ Sk·k(fS), X 7−→ ϕX is an E-C-isomorphism. e) If S is finite then the map G −→ S(fS), X 7−→Xt∈S (Xt ⊗ 1KS )V fS t is an E-C-isomorphism. a) is obvious. b) By Theorem 2.1.9 c),g), G is an involutive unital subalgebra of S(f ) and by Proposition 2.1.6 a) (resp. Proposition 2.1.6 c) and Corollary 1.3.7 c)) and Theorem 2.1.9 h), it is an E-C-subalgebra of S(f ). c) follows from Theorem 2.1.9 b) and Corollary 2.1.16 a). d) follows from c). e) is contained in d). DEFINITION 2.1.18 We denote by ST the set of finite subgroups of T and call T locally finite if ST is upward directed and S = T . [S∈ST 60 T is locally finite iff the subgroups of T generated by finite subsets of T are finite. COROLLARY 2.1.19 Assume T locally finite. We put fS := f(S × S) for every S ∈ ST and identify S(fS) with { X ∈ S(f ) t ∈ T \ S ⇒ Xt = 0} (Corollary 2.1.17 e)). a) For every X ∈ Sk·k(f ) and ε > 0 there is an S ∈ ST such that (Xt ⊗ 1K )Vt − X(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xt∈R < ε for every R ∈ ST with S ⊂ R. b) Sk·k(f ) is the norm closure of ∪s∈STS(fS) and so it is canonically iso- morphic to the inductive limit of the inductive system { S(fS) S ∈ ST } and for every S ∈ ST the inclusion map S(fS) → Sk·k(f ) is the associated canonical morphism. a) There is a Y ∈ R(f ) with kX − Y k < ε By Corollary 2.1.17 b), for R ∈ ST with S ⊂ R, 2 . Let S ∈ ST with Y ∈ S(fS). ((Xt − Yt)e⊗1K )Vt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤ kX − Y k < (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xt∈R ((Xt − Yt)e⊗1K )Vt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (Xte⊗1K)Vt − X(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xt∈R ε 2 so (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xt∈R + kY − Xk < ε 2 + ε 2 = ε . b) follows from a). Remark. The C-algebras of the form Sk·k(f ) with T locally finite can be seen as a kind of AF-E-C-algebras. PROPOSITION 2.1.20 The following are equivalent for all t ∈ T with t2 = 1 and α ∈ U n E. 61 a) 1 2 (V1 + (αe⊗1K )Vt) ∈ P r S(f ). b) α2 = f (t). By Proposition 2.1.2 b),d),e), so (Vt)∗ = ( f (t)e⊗1K )Vt , (V1 + (αe⊗1K)Vt)∗ = 1 2 (Vt)2 = ( f (t)∗e⊗1K )V1 (V1 + ((α∗ f (t))e⊗1K )Vt) , 1 2 (cid:18) 1 2 (V1 + (αe⊗1K)Vt)(cid:19)2 = 1 4 ((1E + α2 f (t)∗)e⊗1K )V1 + 1 2 (αe⊗1K)Vt . Thus a) is equivalent to α∗ f (t) = α and α2 f (t)∗ = 1E, which is equivalent to b). COROLLARY 2.1.21 Let t ∈ T such that t2 = 1 and f (t) = 1E. Then 1 2 (V1 ± Vt) ∈ P r S(f ) , (V1 + Vt)(V1 − Vt) = 0 . The assertion follows from Proposition 2.1.20. COROLLARY 2.1.22 Let α, β ∈ U n E, s, t ∈ T with s2 = t2 = 1, st = ts, (αβ∗f (st, t)∗ + βα∗f (st, s)∗) , γ := (α∗βf (s, st)∗ + β∗αf (t, st)∗) , γ′ := 1 2 1 2 and X := 1 2 ((αe⊗1K )Vs + (βe⊗1K )Vt) . a) f (s, st)f (t, st) = f (st, t)f (st, s) = f (st)∗. b) f (st, t)f (s, st) = f (st, s)f (t, st). c) X∗X = 1 2 (V1 + (γe⊗1K )Vst) , XX∗ = 1 2 (V1 + (γ′e⊗1K )Vst). d) The following are equivalent. 62 d1) X∗X ∈ P r S(f ). d2) XX∗ ∈ P r S(f ). d3) α∗βf (t, st) = β∗αf (s, st). d4) α∗βf (st, t) = β∗αf (st, s). a) and b) follow from the equation of Schur functions (Definition1.1.1) and Proposition 1.1.2 a). c) By Proposition 2.1.2 b),e) and Proposition 1.1.2 b), X∗ = 1 2 X∗X = 1 2 V1 + 1 4 1 4 (((α∗ f (s))e⊗1K )Vs + ((β∗ f (t))e⊗1K )Vt) , ((α∗β f (s)f (s, t) + β∗α f (t)f (t, s))e⊗1K )Vst = ((αβ∗ f (t)f (s, t) + βα∗ f (s)f (t, s))e⊗1K )Vst = ((α∗βf (s, st)∗ + β∗αf (t, st)∗)e⊗1K )Vst = ((αβ∗f (st, t)∗ + βα∗f (st, s)∗)e⊗1K )Vst = (V1 + (γe⊗1K)Vst) , (V1 + (γ′e⊗1K )Vst) . V1 + 1 2 1 2 = 1 2 V1 + 1 4 XX∗ = 1 2 = 1 2 V1 + 1 4 d1 ⇔ d2 is known. d1 ⇔ d3. By a), γ2 − f (st) = 1 4 (α∗βα∗βf (s, st)∗2 + β∗αβ∗αf (t, st)∗2 + 2f (s, st)∗f (t, st)∗)− −f (s, st)∗f (t, st)∗ = 1 4 (α∗βf (s, st)∗ − β∗αf (t, st)∗)2 . By Proposition 2.1.20 d1) is equivalent to γ2 = f (st) so, by the above, since α∗βf (s, st)∗ − β∗αf (t, st)∗ is normal, it is equivalent to α∗βf (s, st)∗ = β∗αf (t, st)∗ or to β∗αf (s, st) = α∗βf (t, st) . d3 ⇔ d4 follows from b). 63 PROPOSITION 2.1.23 Let X ∈ S(f ). (XtX∗t ) = (XX∗)1. X∗t Xt = (X∗X)1 , fPt∈T a) fPt∈T b) (Xt)t∈T , (X∗t )t∈T ∈ g(cid:13)t∈T E, k(Xt)t∈Tk ≤ kXk , k(X∗t )t∈Tk ≤ kXk . c) If T is finite and f is constant then there is an X ∈ S(f ) with kXk ≥ √Card T k(Xt)t∈Tk , kXk ≥ √Card T k(X∗t )t∈Tk . d) If T is infinite and locally finite and f is constant then the map S(f ) −→ g(cid:13)t∈T E, X 7−→ (Xt)t∈T is not surjective. a) follows from Theorem 2.1.9 g). b) By a), and by Proposition 2.1.6 a), (Xt)t∈T , (X∗t )t∈T ∈ g(cid:13)t∈T E k(Xt)t∈Tk2 = kϕ1,1(X∗X)k ≤ kX∗Xk = kXk2 , k(X∗t )t∈Tk2 = kϕ1,1(XX∗)k ≤ kXX∗k = kXk2 . c) Let n := Card T and for every t ∈ T put Xt := 1E, ξt := 1E. Then k(Xt)t∈Tk2 = k(X∗t )t∈Tk2 = n , k(ξt)t∈Tk2 = n . For t ∈ T , by Theorem 2.1.9 e), (Xξ)t =Xs∈T f (s, s−1t)Xsξs−1t = n1E 64 so h Xξ Xξ i = n31E , nkXk2 = kXk2 kξk2 ≥ kXξk2 = n3 , kXk2 ≥ nk(Xt)t∈Tk2 , kXk ≥ √nk(Xt)t∈Tk . d) follows from c), Theorem 2.1.9 a), and the Principle of Inverse Operator. Remark. If E is a W-algebra then it may exist a family (xt)t∈T in E such that the family ((xte⊗1K )Vt)t∈T is summable in LE(H)T2 in the W-case but not in the C-case as the following example shows. Take T := ZZ , f constant, E := l∞( ZZ ), and xt := (δt,s)s∈T ∈ E for every t ∈ T . By Proposition 2.1.23 b), ((xt ⊗ 1K )Vt)t∈T is not summable in LE(H)T2 in the C-case. In the W-case for ξ ∈ H and s, t ∈ T , Thus h ((xt ¯⊗1K )Vtξ)s ((xt ¯⊗1K )Vtξ)s i = etξs−t2 , h (xt ¯⊗1K )Vtξ (xt ¯⊗1K )Vtξ i = et kξk2 . X := T2Xt∈T (xt ¯⊗1K )Vt ∈ SW (f ) . Using the identification of Theorem 2.1.9 i), we get X ∈ SW (f ) \ SC(f ). COROLLARY 2.1.24 Let X ∈ S(f ). a) X ∈(cid:8) xe⊗1K (cid:12)(cid:12) x ∈ E(cid:9)c b) X ∈ { Vt t ∈ T }c iff iff Xt ∈ Ec for all t ∈ T . Xs−1ts = f (s, s−1ts)∗f (t, s)Xt = f (s−1, ts)f (t, s) f (s)Xt for all s, t ∈ T . c) X ∈ S(f )c iff for all s, t ∈ T Xt ∈ Ec , Xs−1ts = f (s, s−1ts)∗f (t, s)Xt = f (s−1, ts)f (t, s) f (s)Xt . In particular if f (s, t) = f (t, s) for all s, t ∈ T then X ∈ S(f )c iff Xt ∈ Ec for all t ∈ T . 65 d) ϕ1,1(S(f )c) = Ec. and X ∈ { Vt t ∈ T }c then Xt = 0. e) If the conjugacy class of t ∈ T (i.e. the set (cid:8) s−1ts(cid:12)(cid:12) s ∈ T(cid:9)) is infinite { Vt t ∈ T }c =(cid:8) xe⊗1K (cid:12)(cid:12) x ∈ E(cid:9) , S(f )c =(cid:8) xe⊗1K (cid:12)(cid:12) x ∈ Ec(cid:9) . f ) If the conjugacy class of every t ∈ T \ {1} is infinite then Thus in this case S(f ) is a kind of E-factor. g) The following are equivalent: g1) S(f ) is commutative. g2) T and E are commutative and f (s, t) = f (t, s) for all s, t ∈ T . For s, t ∈ T , x ∈ E, and Y := (xe⊗1K )Vs, by Theorem 2.1.9 g), (XY )t =gXr∈T (Y X)t =gXr∈T f (r, r−1t)XrYr−1t =gXr∈T f (r, r−1t)YrXr−1t =gXr∈T f (r, r−1t)Xrδs,r−1tx = f (ts−1, s)Xts−1x , f (r, r−1t)δr,sxXr−1t = f (s, s−1t)xXs−1t . a) follows from the above by putting s := 1 (Proposition 1.1.2 a)). b) follows from the above by putting x := 1E and t := rs (Proposition 1.1.2). c) follows from a),b), and Corollary 1.3.7 d). The last assertion follows using Proposition 1.1.5 a). d) follows from c) (and Proposition 1.1.2 a)). e) follows from b) and Proposition 2.1.23 b). f) follows from c), e), and Proposition 2.1.2 d). g1 ⇒ g2. By a), E is commutative. By Proposition 2.1.2 b), f (s, t)Vst = VsVt = VtVs = f (t, s)VtVs = f (t, s)Vts 66 and so by Theorem 2.1.9 a), st = ts and f (s, t) = f (t, s). g2 ⇒ g1 follows from c). COROLLARY 2.1.25 If IK = IR then the following are equivalent: a) S(f )c = S(f ) = Re S(f ). b) T is commutative, Ec = E = Re E, and f (s, t) = f (t, s) , f (t) = 1E , t2 = 1 for all s, t ∈ T . a ⇒ b. By Corollary 2.1.24 g1 ⇒ g2, T is commutative, E = Ec, and f (s, t) = f (t, s) for all s, t ∈ T . Since E is isomorphic with a C-subalgebra of S(f ) (Theorem 2.1.9 h)), E = Re E. By Proposition 2.1.2 e), so by Theorem 2.1.9 a), t = t−1 , f (t) = 1E, so t2 = 1. Vt = V ∗t = ( f (t)e⊗1K )Vt−1 b ⇒ a. By Corollary 2.1.24 g2 ⇒ g1, S(f )c = S(f ). For X ∈ S(f ) and t ∈ T , by Theorem 2.1.9 c), (X∗)t = f (t)(Xt−1 )∗ = (Xt)∗ = Xt so X∗ = X (Theorem 2.1.9 a)). PROPOSITION 2.1.26 Let (Ei)i∈I be a family of unital C-algebras such that E is the C-direct product of this family. For every i ∈ I, we identify Ei with the corresponding closed ideal of E (resp. of E E) and put (s, t) 7−→ f (s, t)i . fi : T × T −→ U n Ec i , a) For every i ∈ I, fi ∈ F(T, Ei). We put (by Theorem 2.1.9 b)) ϕi : S(f ) −→ S(fi) , X 7−→ ϕi is a surjective C*-homomorphism. T2Xt∈T ((Xt)ie⊗1K )V fi t . 67 b) In the C-case, if T is finite then R(f ) = Sk·k(f ) = SC(f ) is isomorphic to the C-direct product of the family (R(fi) = Sk·k(fi) = SC(fi))i∈I . c) In the C-case, if I is finite then SC(f ) (resp. Sk·k(f )) is isomorphic to d) In the W-case, SW (f ) is isomorphic to the C-direct product of the family Qi∈I SC(fi) (resp. Qi∈I Sk·k(fi)). (SW (fi))i∈I . Remark. The C-isomorphisms of b) and c) cease to be surjective in general if T and I are both infinite. Take T := ( ZZ2 )IN, I := IN, Ei := IK for every i ∈ I, and E := l∞ (i.e. E is the C-direct product of the family (Ei)i∈I ). For every n ∈ IN put tn := (δm,n)m∈IN ∈ T . Assume there is an X ∈ SC(f ) (resp. X ∈ Sk·k(f )) with ψX = (V fi ti )i∈I ), where ψ and ϕ are the maps of b) and c), respectively. Then (Xtn )i = δi,n for all i, n ∈ IN and this implies (Xt)t∈T 6∈ (cid:13)t∈T E, which contradicts Proposition 2.1.23 b). ti )i∈I (resp. ϕX = (V fi PROPOSITION 2.1.27 Let S be a finite group, K′ := l2(S), K′′ := l2(S × T ), and g ∈ F(S,S(f )) such that g(s1, s2) ∈ U n Ec (where U n Ec is identified with (U n Ec)e⊗1K ⊂ U n S(f )c) for all s1, s2 ∈ S and put h : (S × T ) × (S × T ) −→ U n Ec , ((s1, t1), (s2, t2)) 7−→ g(s1, s2)f (t1, t2) . a) h ∈ F(S × T, E); for every X ∈ S(g) put ϕX :=Xs∈S T3Xt∈T ((Xs)te⊗1K ′′)V h (s,t) ∈ S(h) . b) ϕ : S(g) −→ S(h) is an E-C-isomorphism. a) is obvious. b) For X, Y ∈ S(g) and (s, t) ∈ S × T , by Theorem 2.1.9 c),g) and Propo- sition 2.1.6 g), (ϕX∗)(s,t) = ((X∗)s)t = g(s)((Xs−1 )∗)t = 68 = g(s) f (t)((Xs−1 )t−1)∗ = h(s, t)(X(s,t)−1 )∗ = ((ϕX)∗)(s,t) , (ϕ(XY ))(s,t) = ((XY )s)t =Xr∈S =Xr∈S g(r, r−1s)gXq∈T g(r, r−1s)(XrYr−1s)t = f (q, q−1t)(Xr)q(Yr−1s)q−1t = h((r, q), (r, q)−1(s, t))X(r,q)Y(r,q)−1(s,t) = ((ϕX)(ϕY ))(s,t) , = ^X(r,q)∈S×T so ϕ is a C-homomorphism. If ϕX = 0 then X(s,t) = 0 for all (s, t) ∈ S × T , so X = 0 and ϕ is injective. Let Z ∈ S(h). For every s ∈ S put Xs := (Z(s,t) ⊗1K)V f T3Xt∈T X :=Xs∈S (Xs ⊗ 1K ′)V g t ∈ S(f ) , s ∈ S(g) . Then ϕX = Z and ϕ is surjective. PROPOSITION 2.1.28 If T is infinite and X ∈ S(f ) \ {0} then X(H #) is not precompact. Let t ∈ T with Xt 6= 0. There is an x′ ∈ E′+ (resp. x′ ∈ E+) with h X∗t Xt , x′ i > 0. We put t1 := 1 and construct a sequence (tn)n∈IN recursively in T such that for all m, n ∈ IN, m < n, Let n ∈ IN \ {1} and assume the sequence was constructed up to n − 1. Since (Proposition 2.1.23 a)) 1 2(cid:10) X∗t Xt , x′(cid:11) . (cid:12)(cid:12)(cid:12)D f (t, tm)∗f (ttmt−1 n n , tn)X∗t Xttmt−1 , x′E(cid:12)(cid:12)(cid:12) < Xs∈T(cid:10) X∗ttms−1Xttms−1 , x′(cid:11) < ∞ D X∗ttmt−1 , x′E < Xttmt−1 1 n n 4(cid:10) X∗t Xt , x′(cid:11) 69 for all m ∈ INn−1 there is a tn ∈ T with for all m ∈ INn−1. By Schwarz' inequality ([C1] Proposition 2.3.4.6 c)) for m ∈ INn−1, (cid:12)(cid:12)(cid:12)D f (t, tm)∗f (ttmt−1 ≤(cid:10) X∗t Xt , x′(cid:11)D X∗ttmt−1 n n , tn)X∗t Xttmt−1 1 n Xttmt−1 n , x′E < 2 ≤ , x′E(cid:12)(cid:12)(cid:12) 4(cid:10) X∗t Xt , x′(cid:11)2 . This finishes the recursive construction. For r, s ∈ T , by Theorem 2.1.9 e), (X(1E ⊗ er))s =gXq∈T f (q, q−1s)Xqδr,q−1s = f (sr−1, r)Xsr−1 , h X(1E ⊗ er) Xt ⊗ es i = f (sr−1, r)X∗t Xsr−1 . For m, n ∈ IN, m < n, it follows h X(1E ⊗ etm) Xt ⊗ ettm i = f (t, tm)X∗t Xt , , n n , tn)X∗t Xttmt−1 n , tn)X∗t Xttmt−1 (cid:10) h X(1E ⊗ etm ) Xt ⊗ ettm i , x′f (t, tm)∗(cid:11) =(cid:10) X∗t Xt , x′(cid:11) , h X(1E ⊗ etn ) Xt ⊗ ettm i = f (ttmt−1 (cid:10) h X(1E ⊗ etn) Xt ⊗ ettm i , x′f (t, tm)∗(cid:11) = (cid:12)(cid:12)(cid:12)D f (t, tm)∗f (ttmt−1 2(cid:10) X∗t Xt , x′(cid:11) , (cid:13)(cid:13)x′(cid:13)(cid:13)kX(1E ⊗ etm ) − X(1E ⊗ etn )kkXtk ≥ ≥(cid:12)(cid:12)(cid:10)h X(1E ⊗ etm ) − X(1E ⊗ etn ) Xt ⊗ ettm i , x′f (t, tm)∗(cid:11)(cid:12)(cid:12) ≥ ≥(cid:12)(cid:12)(cid:10)h X(1E ⊗ etm) Xt ⊗ ettm i , x′f (t, tm)∗(cid:11)(cid:12)(cid:12)− −(cid:12)(cid:12)(cid:10)h X(1E ⊗ etn) Xt ⊗ ettm i , x′f (t, tm)∗(cid:11)(cid:12)(cid:12) > >(cid:10) X∗t Xt , x′(cid:11) − 2(cid:10) X∗T XT , x′(cid:11) . 2(cid:10) X∗t Xt , x′(cid:11) = , x′E(cid:12)(cid:12)(cid:12) < 1 1 1 n Thus the sequence (X(1E ⊗ etn))n∈IN has no Cauchy subsequence and therefore X(H #) is not precompact. 70 PROPOSITION 2.1.29 Assume T finite and let Ω be a compact space, ω0 ∈ Ω, (s, t) 7−→ f (s, t)1Ω , g : T × T −→ U n C(Ω, E) , A := { X ∈ S(g) t ∈ T, t 6= 1 =⇒ Xt(ω0) = 0} , B := { Y ∈ C(Ω,S(f )) t ∈ T, t 6= 1 =⇒ Y (ω0)t = 0} . Then g ∈ F(T,C(Ω, E)) and we define for every X ∈ A and Y ∈ B, ϕX : Ω −→ S(f ) , ω 7−→Xt∈T (Xt(ω) ⊗ 1K )V f t , ψY :=Xt∈T (Y (·)t ⊗ 1K )V g t . Then A (resp. B) is a unital C-subalgebra of S(g) (resp. of C(Ω,S(f ))) ϕ : A −→ B , are C-isomorphisms, and ϕ = ψ−1. ψ : B −→ A It is easy to see that A (resp. B) is a unital C-subalgebra of S(g) (resp. of C(Ω,S(f ))) and that ϕ and ψ are well-defined. For X, X′ ∈ A, t ∈ T , and ω ∈ Ω, by Theorem 2.1.9 c),g) and Proposition 2.1.2 e), f (s, s−1t)((ϕX)(ω))s((ϕX′)(ω))s−1t = (f (s, s−1t)XsX′s−1t)(ω) = (( f (s)((Xs−1 )∗(ω))) ⊗ 1K )V f s = (Xs(ω)∗ ⊗ 1K )(V f s )∗ = (ϕX)∗(ω) (((ϕX)(ϕX′))(ω))t =Xs∈T =Xs∈T (ϕX∗)(ω) =Xs∈T =Xs∈T f (s, s−1t)Xs(ω)X′s−1t(ω) =Xs∈T s =Xs∈T (((X∗)s(ω)) ⊗ 1K )V f s−1 )∗ =Xs∈T ((Xs−1)(ω)∗ ⊗ 1K)(V f so ϕ is a C-homomorphism and we have = (XX′)t(ω) = (ϕ(XX′)(ω))t , (ψϕX)t = (ϕX)t = Xt . Moreover for Y ∈ B, (ϕψY )t(ω) = ((ψY )(ω))t = Yt(ω) which proves the assertion. 71 2.2 Variation of the parameters In this subsection we examine the changes produced by the replacement of the groups and of the Schur functions. DEFINITION 2.2.1 We put for every λ ∈ Λ(T, E) (Definition 1.1.3) Uλ : H −→ H , ξ 7−→ (λ(t)ξt)t∈T . It is easy to see that Uλ is well-defined, Uλ ∈ U n LE(H), and the map Λ(T, E) −→ U n LE(H), λ 7−→ Uλ is an injective group homomorphism with U∗λ = Uλ∗ (Proposition 1.1.4 c)). Moreover for all λ, µ ∈ Λ(T, E). kUλ − Uµk ≤ kλ − µk∞ PROPOSITION 2.2.2 Let f, g ∈ F(T, E) and λ ∈ Λ(T, E). a) The following are equivalent: a1) g = f δλ. a2) There is a (unique) E-C-isomorphism ϕ : S(f ) −→ S(g) continuous with respect to the T2-topologies such that for all t ∈ T and x ∈ E, ϕV f t = (λ(t)∗e⊗1K )V g t (we call such an isomorphism an S-isomorphism and denote it by ≈S) b) If the above equivalent assertions are fulfilled then for X ∈ S(f ) and t ∈ T , ϕX = U∗λXUλ , (ϕX)t = λ(t)∗Xt . 72 c) There is a natural bijection { S(f ) f ∈ F(T, E)} / ≈S−→ F(T, E)/{ δλ λ ∈ Λ(T, E)} . By Proposition 1.1.4 c), δλ ∈ F(T, E) for every λ ∈ Λ(T, E). a1 ⇒ a2&b. For s, t ∈ T and ζ ∈ E, by Proposition 2.1.2 c), U∗λV f t Uλ(ζ ⊗ es) = U∗λV f t ((λ(s)ζ) ⊗ es) = U∗λ((f (t, s)λ(s)ζ) ⊗ ets) = = (λ(ts)∗f (t, s)λ(s)ζ) ⊗ ets = (λ(t)∗g(t, s)ζ) ⊗ ets = (λ(t)∗e⊗1K )V g so (by Proposition 2.1.2 e)) t (ζ ⊗ es) Thus the map U∗λV f t Uλ = (λ(t)∗e⊗1K )V g t . ϕ : S(f ) −→ S(g) , X 7−→ U∗λXUλ is well-defined. It is obvious that it has the properties described in a2). The uniqueness follows from Theorem 2.1.9 b). We have ϕ((Xte⊗1K )V f t ) = (Xte⊗1K )(λ(t)∗e⊗1K)V g t = ((λ(t)∗Xt)e⊗1K )V g t so (ϕX)t = λ(t)∗Xt. a2 ⇒ a1. Put h := f δλ. By the above, for t ∈ T , so V g t = V h t and this implies g = h. t = (λ(t)∗e⊗1K )V h t t = ϕV f (λ(t)∗e⊗1K )V g c) follows from a). Remark. Not every E-C-isomorphism S(f ) → S(g) is an S isomorphism (see Remark of Proposition 3.2.3). COROLLARY 2.2.3 Let Λ0(T, E) := { λ ∈ Λ(T, E) λ is a group homomorphism } 73 and for every λ ∈ Λ0(T, E) put ϕλ : S(f ) −→ S(f ) , X 7−→ U∗λXUλ . Then the map λ 7→ ϕλ is an injective group homomorphism. By Proposition 1.1.4 c), Λ0(T, E) is the kernel of the map Λ(T, E) −→ F(T, E), λ 7−→ δλ so by Proposition 2.2.2, ϕλ is well-defined. Thus only the injectivity of the map has to be proved. For t ∈ T and ζ ∈ E, by Proposition 2.1.2 c), U∗λVtUλ(ζ ⊗ e1) = U∗λVt(ζ ⊗ e1) = U∗λ(ζ ⊗ et) = = (λ(t)∗ζ) ⊗ et = (λ(t)∗e⊗1K )Vt(ζ ⊗ e1) . So if ϕλ is the identity map then λ(t) = 1E for every t ∈ T . PROPOSITION 2.2.4 Let F be a unital C-algebra, ϕ : E → F a surjective F . We put for all ξ ∈ H, η ∈ L, and X ∈ LE(H), C-homomorphism, g := ϕ ◦ f ∈ F(T, F ), and L := g(cid:13)t∈T Xη := fXζ ∈ L , ξ := (ϕξi)i∈I ∈ L , where ζ ∈ H with ζ = η (Lemma 1.3.11 a),b) and Proposition 1.3.12 a)). Then X = T3Xt∈T ((ϕXt)e⊗1K )V g t ∈ S(g) for every X ∈ S(f ) and the map ϕ : S(f ) −→ S(g) , X 7−→ X is a surjective C-homomorphism, continuous with respect to the topologies Tk, k ∈ {1, 2, 3} such that Ker ϕ = { X ∈ S(f ) t ∈ T =⇒ Xt ∈ Ker ϕ} . 74 For s, t ∈ T and ξ ∈ H, t ξs = ( ^z { } (Xte⊗1K )V f  = ϕ(f (t, t−1s)Xtξt−1s) = g(t, t−1s)(ϕXt)ξt−1s = (((ϕXt)e⊗1K )V g t ξ)s = ^z } (Xte⊗1K)V f { t ξ)s = ϕ((Xte⊗1K )V f t ξ)s so by Lemma 1.3.11 b), By Theorem 2.1.9 b), . t ^z } { t = ((ϕXt)e⊗1K )V g (Xte⊗1K )V f T3Xt∈T (Xte⊗1K )V f X = t so by the above and by Proposition 1.3.12 b), X = T3Xt∈T ((ϕXt)e⊗1K )V g t ∈ S(g) . By Proposition 1.3.12 b), ϕ is a surjective C-homomorphism, continuous with respect to the topologies Tk (k ∈ {1, 2, 3}). The last assertion is easy to see. COROLLARY 2.2.5 Let F be a unital C-algebra, ϕ : E → F a unital C-homomorphism such that ϕ(U n Ec) ⊂ F c, g := ϕ ◦ f ∈ F(T, F ), and L := (cid:13)t∈T F . Then the map ϕ : Sk·k(f ) −→ Sk·k(g) , X 7−→ k·kXt∈T ((ϕXt) ⊗ 1L)V g t is C-homomorphism. Put G := E/Ker ϕ and denote by ϕ1 : E → G the quotient map and by ϕ2 : G → F the corresponding injective C-homomorphism. By Proposition 2.2.4, the corresponding map ϕ1 : Sk·k(f ) −→ Sk·k(ϕ1 ◦ f ) 75 is a C-homomorphism and by Theorem 2.1.9 k), the corresponding map ϕ2 : Sk·k(ϕ1 ◦ f ) −→ Sk·k(g) is also a C-homomorphism. The assertion follows from ϕ = ϕ2 ◦ ϕ1. PROPOSITION 2.2.6 Let T ′ be a group, K′ := l2(T ′), H′ := Ee⊗K′, ψ : T → T ′ a surjective group homomorphism such that ψ(t′) ∈ IN , Card −1 sup t′∈T ′ and f′ ∈ F(T ′, E) such that f′ ◦ (ψ × ψ) = f . If we put X′t′ := Xt∈ −1 ψ (t′) Xt for every X ∈ S(f ) and t′ ∈ T ′ then the family ((X′t′e⊗1K ′)V f ′ in LE(H′)T2 for every X ∈ S(f ) and the map t′ )t′∈T ′ is summable ψ : S(f ) −→ S(f′) , X 7−→ X′ := is a surjective E-C*-homomorphism. T1Xt′∈T ′ (X′t′e⊗1K ′)V f ′ t′ We may drop the hypothesis that ψ is surjective if we replace S by Sk·k. Let X ∈ S(f ). By Corollary 2.1.16 a), since ψ is surjective and Card −1 ψ(t′) ∈ IN , sup t′∈T ′ t′ )t′∈T ′ is summable in LE(H′)T2 and = ef′(t′) Xs∈ −1 ψ (t′) (Xs−1)∗ = Let X, Y ∈ S(f ). By Theorem 2.1.9 c),g), for t′ ∈ T ′, therefore X′ ∈ S(f′). it follows that the family ((X′t′e⊗1K ′)V f ′ (X′∗)t′ = ef′(t′)(Xt′−1 )∗ = ef′(t′) Xt∈ ∗ Xt −1 ψ (t′−1) 76 −1 ψ (t′) −1 ψ (t′) −1 ψ (s′) (X∗)s = (X∗)′t′ , f′(s′, s′−1t′)X′s′Y ′s′−1t′ = f (s)(Xs−1)∗ = Xs∈ = Xs∈ (X′Y ′)t′ = gXs′∈T ′ f′(s′, s′−1t′) Xs∈ Xs  Xr∈ Yr = = gXs′∈T ′ f′(s′, s′−1t′) Xs∈ XsYs−1t = = gXs′∈T ′ ψ (s′) Xt∈  Xs∈ f (s, s−1t)XsYs−1t = = gXs′∈T ′ ψ (s′) Xt∈ f (s, s−1t)XsYs−1t = Xt∈ = Xt∈ ψ (t′)gXs∈T −1 ψ (s′−1t′) −1 ψ (t′) −1 ψ (t′) −1 ψ (t′) −1 −1 −1 (XY )t = (XY )′t′ . Thus ψ is a C-homomorphism. The other assertions are easy to see. The last assertion follows from Corollary 2.1.17 d). COROLLARY 2.2.7 If we use the notation of Proposition 2.2.6 and Corol- lary 2.2.5 and define eϕ′ and eψ′ in an obvious way then eϕ′ ◦ ψ = eψ′ ◦ ϕ. For X ∈ S(f ) and t′ ∈ T ′, (eϕ′ ψX)t′ = ϕ(( ψX)t′) = ϕ Xt∈ −1 ψ (t′) Xt = Xt∈ −1 ψ (t′) ϕXt , so ϕXt , (eψ′ ϕX)t′ = Xt∈ −1 ψ (t′) −1 ψ (t′) ( ϕX)t = Xt∈ eϕ′ ◦ ψ = eψ′ ◦ ϕ . 77 PROPOSITION 2.2.8 Let F be a unital C-subalgebra of E such that f (s, t) ∈ F for all s, t ∈ T . We denote by ψ : F → E the inclusion map and put (s, t) 7−→ f (s, t) , f F : T × T −→ U n F c , H F := (cid:13)t∈T ψ : H F −→ H , F ≈ F ⊗ K , ξ 7−→ (ψξt)t∈T . Moreover we denote for all s, t ∈ T by uF s,t the corresponding operators associated with F (f F ∈ F(T, F )). Let X ∈ SC(f ) such that X( ψξ) ∈ ψ(H F ) for every ξ ∈ H F and put t , and ϕF t , V F X F : H F −→ H F , where ξ′ ∈ H F with ψξ′ = X( ψξ), and X F := (uF cal identification of F with LF ( F )) for every t ∈ T . ξ 7−→ ξ′ , t 1 )∗X F uF t ∈ F (by the canoni- a) ξ, η ∈ H F ⇒D ψξ(cid:12)(cid:12)(cid:12) ψηE = ψ h ξ η i. b) ψ is linear and continuous with(cid:13)(cid:13)(cid:13) ψ(cid:13)(cid:13)(cid:13) = 1. c) X F is linear and continuous with(cid:13)(cid:13)X F(cid:13)(cid:13) = kXk. d) For s, t ∈ T , ψϕF s,tX F = ϕs,tX , ψX F t = Xt , s,tX F = f F (st−1, t)X F ϕF st−1 . e) X F ∈ S(f F ). f ) ξ ∈ H F ⇒ X( ψξ) = (Xt ⊗ 1K )Vt( ψξ). k·kPt∈T a&b&c are easy to see. d) By a) and Proposition 2.1.6 b), ψϕF ϕF s,tX F =(cid:10) X F (1F ⊗ et)(cid:12)(cid:12) 1F ⊗ es(cid:11) , s,tX F = ψ(cid:10) X F (1F ⊗ et)(cid:12)(cid:12) 1F ⊗ es(cid:11) = 78 =D ψ(X F (1F ⊗ et))(cid:12)(cid:12)(cid:12) ψ(1F ⊗ es)E = h X(1E ⊗ et) 1E ⊗ es i = ϕs,tX . ψX F t = ψϕF 1,tX F = ϕ1,tX = Xt In particular and by Proposition 2.1.8, ψϕF s,tX F = ϕs,tX = f (st−1, t)Xst−1 = ψ(f F (st−1, t)X F st−1 ) , s,tX F = f F (st−1, t)X F ϕF st−1 . e) By c) and Proposition 2.1.3 d), for ξ ∈ H F , uF t (uF t )∗ξ = ξ , k·kXt∈T k·kXt∈T X F uF t (uF t )∗ξ , k·kXt∈T uF s ((uF s )∗X F uF t )(uF t )∗ξ . X F ξ = X F uF t (uF t )∗ξ = X F ξ = k·kXs∈T uF s (uF s )∗X F ξ = k·kXs∈T k·kXt∈T By d) and Proposition 2.1.4 b),d), X F ξ = = k·kXs∈T k·kXs∈T = k·kXt∈T k·kXr∈T k·kXs∈T s f F (st−1, t)X F uF st−1 (uF t )∗ξ = k·kXs∈T k·kXt∈T uF s X F st−1(uF s )∗V F st−1ξ = uF s X F r (uF s )∗V F r ξ = k·kXs∈T k·kXr∈T uF s (uF s )∗(X F r ⊗ 1F )V F r ξ = uF s (uF s )∗ k·kXt∈T (X F t ⊗ 1K )V F t ξ = k·kXt∈T (X F t ⊗ 1K )V F t ξ by Proposition 2.1.3 d), again. Thus X F = T2Xt∈T (X F t ⊗ 1K )V F t ∈ SC(f F ) . 79 f) For s, t ∈ T , by d), ( ψ((X F t ⊗ 1K )V F t ξt−1s) = t ξ))s = ψ((X F = f (t, t−1s)Xt( ψξ)t−1s = ((Xt ⊗ 1K )Vt ψξ)s , t ⊗ 1K )V F t ξ)s = ψ(f F (t, t−1s)X F ψ((X F t ⊗ 1K )V F t ξ) = (Xt ⊗ 1K )Vt ψξ so by b) and e), (X F X( ψξ) = ψ(X F ξ) = ψ k·kXt∈T k·kXt∈T k·kXt∈T (Xt ⊗ 1K)Vt( ψξ) . t ξ = t ⊗ 1K )V F t ⊗ 1K )V F ψ((X F t ξ) = = PROPOSITION 2.2.9 Let F be a W-algebra such that E is a unital C- subalgebra of F generating it as W-algebra, ϕ : E → F the inclusion map, and ξ := (ϕξt)t∈T ∈ L for every ξ ∈ H, where L := W (cid:13)t∈T F ≈ F ¯⊗K . a) ϕ(U n Ec) ⊂ U n F c and g := ϕ ◦ f ∈ F(T, F ). b) If ψ : LE(H) −→ LF (L) , X 7−→ ¯X is the injective C-homomorphism defined in Proposition 1.3.9 b), then ψ(SC (f )) ⊂ SW (g), ψ(SC (f )) generates SW (g) as W-algebra, and for every X ∈ SC(f ) and t ∈ T we have ( ¯X)t = ϕXt. c) The following are equivalent for every Y ∈ SW (g): c1) Y ∈ ψ(SC (f )). c2) ξ ∈ H ⇒ Y ξ ∈ H. If these conditions are fulfilled then c3) (Yt)t∈T ∈ H. c4) (Y ∗t )t∈T ∈ H. 80 c5) ξ ∈ H ⇒ Y ξ = (Yt ¯⊗1K )V g t ξ ∈ H. k·kPt∈T a) follows from the density of ϕ(E) in F F (Lemma 1.3.8 a ⇒ c). b) For x ∈ E, t ∈ T , and ξ ∈ H, (((ϕx) ¯⊗1K )V g ξ)s = g(t, t−1s)(ϕx)ξt−1s = = ϕ(f (t, t−1s)xξt−1s) = ϕ((x ⊗ 1K )Vtξs) t ((ϕx) ¯⊗1K )V g t = (x ⊗ 1K )V f t . so Let now X ∈ S(f ). By Theorem 2.1.9 b), X = T2Xt∈T (Xt ⊗ 1K)V f t so by the above and by Proposition 1.3.9 c) (and Theorem 2.1.9 d)), ¯X = (Xt ⊗ 1K )V f t = T1Xt∈T T1Xt∈T ((ϕXt) ¯⊗1K)V g t ∈ SW (g) so ψ(SC (f )) ⊂ SW (f ). By Theorem 2.1.9 a), ( ¯X)t = ϕXt for every t ∈ T . Since ϕ(E) is dense in F F (Lemma 1.3.8 a) ⇒ c)) it follows that T1 R(g) ⊂ ϕ(R(f )) so ψ(S(f )) is dense in S(g) (Lemma 1.3.8 c ⇒ a). ..z}{S(g) and therefore generates S(g) as W-algebra c1 ⇒ c2 follows from the definition of ψ. c2 ⇒ c1 follows from Proposition 2.2.8 e). c2 ⇒ c3&c4 follows from Proposition 2.1.23 b). c2 ⇒ c5 follows from Proposition 2.2.8 f). 81 LEMMA 2.2.10 Let E, F be W-algebras, G := E ¯⊗F , and . L := W (cid:13)t∈T G ≈ G ¯⊗K a) If z ∈ G# then z ¯⊗1K belongs to the closure of { w ¯⊗1K w ∈ E ⊙ F, kwk ≤ 1} in LG(L)... L . b) For every y ∈ F , the map E# E −→ G G, x 7−→ x ⊗ y is continuous. a) By [C1] Corollary 6.3.8.7, there is a filter F on { w ∈ E ⊙ F kwk ≤ 1} . By Lemma 1.3.2 b), for (a, ξ, η) ∈ G × L × L, converging to z in G# G D z ¯⊗1K , ^(a, ξ, η)E =* z , ξt a η∗t+ = lim w,F* w , ξt a η∗t+ = GXt∈T w,FD w ¯⊗1K , ^(a, ξ, η)E GXt∈T = lim which proves the assertion. b) Let (ai, bi)i∈I be a finite family in E × F . For x ∈ E, ai ⊗ bi+ =Xi∈I * x ⊗ y ,Xi∈I Since(cid:8) x ⊗ y x ∈ E#(cid:9) is a bounded set of G, the above identity proves the h x , ai ih y , bi i =* x ,Xi∈I h y , bi i ai+ . continuity. 82 PROPOSITION 2.2.11 Let F be a unital C*-algebra, S a group, and g ∈ F(S, F ). We denote by ⊗σ the spatial tensor product and put G := E ⊗σ F F ≈ Fe⊗l2(S), h : (T × S) × (T × S) −→ U n Gc , L := g(cid:13)s∈S (resp. G := E ¯⊗F ) , M := ^(cid:13) (t,s)∈T×S ((t1, s1), (t2, s2)) 7−→ f (t1, t2) ⊗ g(s1, s2) . G ≈ Ge⊗l2(T × S) , a) h ∈ F(T × S, G), LE(H) ⊗σ LF (L) ⊂ LG(M ) in the C-case, LE(H) ¯⊗LF (L) ≈ LG(M ) in the W-case . M ≈ He⊗L, b) For t ∈ T , s ∈ S, x ∈ E, y ∈ F , ((xe⊗1l2(T ))V f t ) ⊗ ((ye⊗1l2(S))V g c) In the C-case, Sk·k(f )⊗σSk·k(g) ≈ Sk·k(h) and SC(f )⊗σ SC(g) ≈ SC(h). d) In the W-case, if z ∈ G# and (t, s) ∈ T × S then (z ¯⊗1l2(T×S))V h (t,s) be- longs to the closure ofn (w ¯⊗1l2(T×S))V h s ) = ((x ⊗ y)e⊗1l2(T×S))V h (t,s)(cid:12)(cid:12)(cid:12) w ∈ (E ⊙ F )#o in LG(M ) ... e) In the W-case, SW (f ) ¯⊗SW (g) ≈ SW (h). (t,s) . M a) h ∈ F(T × S, G) is obvious. Let us treat the C-case first. For ξ, ξ′ ∈ H and η, η′ ∈ L, (cid:10) ξ′ ⊗ η′(cid:12)(cid:12) ξ ⊗ η(cid:11) =(cid:10) ξ′(cid:12)(cid:12) ξ(cid:11) ⊗(cid:10) η′(cid:12)(cid:12) η(cid:11) = Xt∈T ((ξ∗t ξ′t) ⊗ (η∗s η′s)) = X(t,s)∈T×S = X(t,s)∈T×S = X(t,s)∈T×S (ξt ⊗ ηs)∗(ξ′t ⊗ η′s) , ξ∗t ξ′t! ⊗ Xs∈S η∗s η′s! = (ξ∗t ⊗ η∗s )(ξ′t ⊗ η′s) = so the linear map H ⊙ L −→ M, ξ ⊗ η 7−→ (ξt ⊗ ηs)(t,s)∈T×S 83 E × F such that Then < ε . xi ⊗ yi − z(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I (xi ⊗ et) ⊗ (yi ⊗ es) − z ⊗ e(t,s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I < ε preserves the scalar products and it may be extended to a linear map ϕ : H ⊗ L → M preserving the scalar products. Let z ∈ G, (t, s) ∈ T × S, and ε > 0. There is a finite family (xi, yi)i∈I in so z ⊗ e(t,s) ∈ ϕ(H ⊗ L) = ϕ(H ⊗ L). It follows that ϕ is surjective and so H ⊗ L ≈ M . The proof for the inclusion LE(H) ⊗σ LF (L) ⊂ LG(M ) can be found in [L] page 37. Let us now discus the W-case. E ¯⊗ F ≈ G follows from [C2] Proposition 1.3 e), M ≈ H ¯⊗L follows from [C3] Corollary 2.2, and LE(H) ¯⊗LF (L) ≈ LG(M ) follows from [C2] Theorem 2.4 d) or [C3] Theorem 2.4. b) For t1, t2 ∈ T , s1, s2 ∈ S, ξ ∈ E, and η ∈ F , by Proposition 2.1.2 f) and [C3] Corollary 2.11, s1))((ξ ⊗ et2 ) ⊗ (η ⊗ es2)) = s1)(η ⊗ es2)) , (((xe⊗1l2(T ))V f t1 )e⊗((ye⊗1l2(S))V g = (((xe⊗1l2(T ))V f t1 )(ξ ⊗ et2 ))e⊗(((ye⊗1l2(S))V g ((((x ⊗ y)e⊗1l2(T×S)))V h = (((xe⊗1l2(T ))V f (t1,s1))((ξ ⊗ η) ⊗ e(t2,s2)) = = (h((t1, s1), (t2, s2))(x ⊗ y)(ξ ⊗ η)) ⊗ e(t1t2,s1s2) = = ((f (t1, t2)xξ) ⊗ (g(s1, s2)yη)) ⊗ et1t2 ⊗ es1s2 = t1 )(ξ ⊗ et2 ))e⊗(((ye⊗1l2(S))V g s1)(η ⊗ es2)) . We put u := ((xe⊗1l2(T ))V f s ) − ((x ⊗ y)e⊗1l2(T×S))V h By the above, u(ζ ⊗ er) = 0 for all ζ ∈ E ⊙ F and r ∈ T × S. t )e⊗((ye⊗1l2(S))V g t,s ∈ LG(M ) . 84 Let us consider the C-case first. Since E ⊙ F is dense in G , we get u(z ⊗ er) = 0 for all z ∈ G and r ∈ T × S. For ζ ∈ M , by [C1] Proposition 5.6.4.1 e), uζ = u Xr∈T×S (ζr ⊗ er)! = Xr∈T×S u(ζr ⊗ er) = 0 , which proves the assertion in this case. Let us consider now the W-case. Let z ∈ G# and r ∈ T × S and let F be a filter on (E ⊙ F )# converging to z in G G ([C1] Corollary 6.3.8.7). For η ∈ M , a ∈ G, and r ∈ T × S, D z ⊗ er , ](a, η)E = hh z ⊗ er η i , ai = h η∗r z , ai = h z , aη∗r i = = lim w,F h w , aη∗r i = lim w,FD w ⊗ er , ](a, η)E , lim w,F w ⊗ er = z ⊗ er in M M . Since u : M M → M M is continuous ([C1] Proposition 5.6.3.4 c)), we get by the above u(z⊗ er) = 0. For ζ ∈ M it follows by [C1] Proposition 5.6.4.6 c), uζ = u MXr∈T×S (ζr ⊗ er) = MXr∈T×S u(ζr ⊗ er) = 0 which proves the assertion in the W-case. c) By b), R(f ) ⊙ R(g) ⊂ R(h) so by a), Sk·k(f ) ⊙ Sk·k(g) ⊂ Sk·k(h) , Sk·k(f ) ⊗σ Sk·k(g) ⊂ Sk·k(h) , SC(f ) ⊙ SC(g) ⊂ SC(h) , SC(f ) ⊗σ SC(g) ⊂ SC (h) . Let z ∈ G#, (t, s) ∈ T × S, and ε > 0. There is a finite family (xi, yi)i∈I in E × F such that (xi ⊗ yi)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I < 1 , < ε . (xi ⊗ yi) − z(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I By b), (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xi∈I (((xi ⊗ 1l2(T ))V f t ) ⊗ ((yi ⊗ 1l2(S))V g s )) − (z ⊗ 1l2(T×S))V h < ε (t,s)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 85 and so by a), k·k T2 R(h) ⊂ R(f ) ⊙ R(g) ⊂ R(f ) ⊙ R(g) , SC(h) ⊂ SC(f ) ⊗σ SC(g) . Sk·k(h) ⊂ Sk·k(f ) ⊗σ Sk·k(g) , d) By a) and Lemma 2.2.10 a), there is a filter F on . For ξ, η ∈ M and a ∈ G, n w ¯⊗1l2(T×S)(cid:12)(cid:12) w ∈ (E ⊙ F )#o (t,s) , ^(a, ξ, η)E =D z ¯⊗1l2(T×S) , V h (t,s) , ^(a, ξ, η)E = lim w,FD (w ¯⊗1l2(T×S))V h (t,s) ^(a, ξ, η)E = (t,s) , ^(a, ξ, η)E , converging to z ¯⊗1l2(T×S) in LG(M ) ... M D (z ¯⊗1l2(T×S))V h w,FD w ¯⊗1l2(T×S)V h = lim which proves the assertion. e) By Theorem 2.1.9 h), #  ... H R(f ) # ... L  R(g) = SW (f )# ⊂ LE(H), = SW (g)# ⊂ LF (L) . By b), R(f ) ⊙ R(g) ⊂ R(h), so by Lemma 2.2.10 b), SW (f )# ⊙ R(g)# ⊂ SW (h)#, SW (f )# ⊗ SW (g)# ⊂ SW (h)# . By [C3] Proposition 2.5, SW (f ) ⊗ SW (g) ⊂ SW (h) SW (f ) ¯⊗SW (g) ≈ ... M SW (f ) ⊗ SW (g)⊂ SW (h) . For x ∈ E, y ∈ F , and (t, s) ∈ T × S, by b), ((x ⊗ y) ¯⊗1l2(T×S))V h (t,s) = ((x ¯⊗1l2(T ))V f t ) ¯⊗((y ¯⊗1l2(S))V g s ) ∈ SW (f ) ¯⊗SW (g) . Let z ∈ G#. By d), there is a filter F on n (w ¯⊗1l2(T×S))V h (t,s)(cid:12)(cid:12)(cid:12) w ∈ (E ⊙ F )#o 86 converging to (z ¯⊗1l2(T×S))V h (t,s) in LG(M ) ... M , so by the above (z ¯⊗1l2(T×S))V h (t,s) ∈ SW (f ) ¯⊗SW (g) . We get R(h) ⊂ SW (f ) ¯⊗SW (g), SW (h) ⊂ SW (f ) ¯⊗SW (g) , SW (h) = SW (f ) ¯⊗SW (g) . COROLLARY 2.2.12 Let n ∈ IN and g : T × T −→ U n (En,n)c , (s, t) 7−→ [δi,jf (s, t)]i,j∈INn . a) (S(f ))n,n ≈ S(g), b) Let us denote by ρ : S(g) → (S(f ))n,n the isomorphism of a). For X ∈ (Sk·k(f ))n,n ≈ Sk·k(g) . S(g), t ∈ T , and i, j ∈ INn, ((ρX)i,j )t = (Xt)i,j . a) Take F := IKn,n and S := {1} in Proposition 2.2.11. Then G ≈ En,n and g : T × T −→ U n Gc , (s, t) 7−→ f (s, t) ⊗ 1F . By Proposition 2.2.11 c),e), S(g) ≈ S(f ) ⊗ IKn,n ≈ (S(f ))n,n Sk·k(g) ≈ Sk·k(f ) ⊗ IKn,n ≈ (Sk·k(f ))n,n . b) By Theorem 2.1.9 b), so X = (ρX)i,j = s T3Xs∈T (Xse⊗1K)V g T3Xs∈t ((Xs)i,je⊗1K )V f s , ((ρX)i,j)t = (Xt)i,j by Theorem 2.1.9 a). 87 if n = 4m for some COROLLARY 2.2.13 Let n ∈ IN. m ∈ IN) then there is an f ∈ F( ZZn × ZZn , E) (resp. f ∈ F( ( ZZ2 )2m, E)) such that If IK = IC (resp. R(f ) = S(f ) ≈ En,n . By [C1] Proposition 7.1.4.9 b),d) (resp. [C1] Theorem 7.2.2.7 i),k)) there is a g ∈ F( ZZn × ZZn , IC) (resp. g ∈ F( ( ZZ2 )2m, IK)) such that (resp. S(g) ≈ IKn,n) . S(g) ≈ ICn,n If we put f : ( ZZn × ZZn ) × ( ZZn × ZZn ) −→ U n Ec , (resp. f : ( ZZ2 )2m × ( ZZ2 )2m −→ U n Ec , (s, t) 7−→ g(s, t) ⊗ 1E (s, t) 7−→ g(s, t) ⊗ 1E) then by Proposition 2.2.11 a),e), f ∈ F( ZZn × ZZn , E) (resp. f ∈ F( ( ZZ2 )2m, E)) and S(f ) ≈ S(g) ⊗ E ≈ IKn,n ⊗ E ≈ En,n . COROLLARY 2.2.14 Let F be a unital C-algebra, G := Ee⊗F , and h : T × T −→ U n Gc , (s, t) 7−→ f (s, t) ⊗ 1F . Then h ∈ F(T, G) and Sk·k(h) ≈ Sk·k(f ) ⊗ F , S(h) ≈ S(f )e⊗F . COROLLARY 2.2.15 If E is a W-algebra then the following are equivalent: a) E is semifinite. b) SW (f ) is semifinite. a ⇒ b. Assume first that there are a finite W-algebra F and a Hilbert space L such that E ≈ F ¯⊗L(L). Put g : T × T −→ U n F c , (s, t) 7−→ f (s, t) . 88 By Corollary 2.2.14, By Corollary 2.1.11 c), SW (g) is finite and so SW (f ) is semifinite. SW (f ) ≈ SW (g) ¯⊗L(L) . The general case follows from the fact that E is the C-direct product of W-algebras of the above form ([T] Proposition V.1.40). b ⇒ a. E is isomorphic to a W-subalgebra of SW (f ) (Theorem 2.1.9 h)) and the assertion follows from [T] Theorem V.2.15. PROPOSITION 2.2.16 Let S, T be finite groups and g ∈ F(S,S(f )) and put L := l2(S), M := l2(S × T ), and h : (S × T ) × (S × T ) −→ U n S(f )c , Then h ∈ F(S × T,S(f )) and the map ((s1, t1), (s2, t2)) 7−→ f (t1, t2)g(s1, s2) . ϕ : S(g) −→ S(h) , X 7−→ X(s,t)∈S×T ((Xs)t ⊗ 1M )V h (s,t) is an S(f )-C-isomorphism. For X, Y ∈ S(g), Z ∈ S(f ), and (s, t) ∈ S × T , by Theorem 2.1.9 c),g), (ϕ(X∗))(s,t) = ((X∗)s)t = (g(s)(Xs−1)∗)t = ((g(s)∗Xs−1)∗)t = = f (t)((g(s)∗Xs−1)t−1 )∗ = f (t)g(s)((Xs−1 )t−1)∗ = = h(s, t)((ϕX)(s−1,t−1))∗ = h(s, t)((ϕX)(s,t)−1 )∗ = ((ϕX)∗)(s,t) , h((r, u), (r, u)−1(s, t))(ϕX)(r,u)(ϕY )(r,u)−1(s,t) = g(r, r−1s)(XrYr−1s)t = ((ϕX)(ϕY ))(s,t) = X(r,u)∈S×T = X(r,u)∈S×T g(r, r−1s)f (u, u−1t)(Xr)u(Yr−1s)u−1t =Xr∈S = Xr∈S g(r, r−1s)XrYr−1s!t = ((XY )s)t = (ϕ(XY ))(s,t) , (ϕ(ZX))(s,t) = ((ZX)s)t = ((ZX)s)t = (ZXs)t = Z(Xs)t = Z(ϕX)(s,t) 89 so ϕ(X∗) = (ϕX)∗ , ϕ(XY ) = (ϕX)(ϕY ), ϕ(ZX) = Zϕ(X) and ϕ is an S(f )-C-homomorphism. If X ∈ S(g) with ϕX = 0 then for (s, t) ∈ S × T , Xs = 0 , (Xs)t = (ϕX)(s,t) = 0 , X = 0 so ϕ is injective. Let x ∈ E and (s, t) ∈ S × T . Put t ∈ S(f ) , Z := (x ⊗ 1K )V f Then for (r, u) ∈ S × T , X := (Z ⊗ 1L)V g s ∈ S(g) . (ϕX)(r,u) = (Xr)u = δr,sZu = δr,sδu,tx so and ϕ is surjective. ϕX = (x ⊗ 1M )V h (s,t) PROPOSITION 2.2.17 Let S be a finite subgroup of T and g := f(S × S). We identify S(g) with the E-C-subalgebra { Z ∈ S(f ) t ∈ T \ S ⇒ Zt = 0} of S(f ) (Corollary 2.1.17 e)). Let X ∈ S(f ) ∩ S(g)c, P+ := X∗X, and P− := XX∗ and assume P± ∈ P r S(f ). a) P± ∈ S(g)c. b) The map ϕ± : S(g) −→ P±S(f )P± , Y 7−→ P±Y P± is a unital C-homomorphism. c) For every Z ∈ ϕ+(S(g)), XZX∗ ∈ ϕ−(S(g)) and the map ψ : ϕ+(S(g)) −→ ϕ−(S(g)) , Z 7−→ XZX∗ is a C-isomorphism with inverse ϕ−(S(g)) −→ ϕ+(S(g)), Z 7−→ X∗ZX such that ϕ− = ψ ◦ ϕ+. 90 d) If p ∈ P r S(g) then (X(ϕ+p))∗(X(ϕ+p)) = ϕ+p , (X((ϕ+p))(X(ϕ+p))∗ = ϕ−p . e) If ϕ+ is injective then ϕ− is also injective, the map E −→ P±S(f )P±, x 7−→ P±(xe⊗1K )P± is an injective unital C-homomorphism, P±S(f )P± is an E-C-algebra, ϕ±(S(g)) is an E-C-subalgebra of it, and ϕ± and ψ are E-C-homo- morphisms. f ) The above results still hold for an arbitrary subgroup S of T if we replace S by Sk·k. a) follows from the hypothesis on X. b) follows from a). c) Let Y ∈ S(g) with Z = P+Y P+. By the hypotheses of the Proposition , XZX∗ = XP+Y P+X∗ = XX∗XY X∗XX∗ = = XX∗Y XX∗XX∗ = P−Y P− ∈ ϕ−(S(g)) and ψ is a C-homomorphism. The other assertions follow from X∗(XZX∗)X = P+ZP+ = P+Y P+ . d) By b) and c), (X(ϕ+p))∗(X(ϕ+p)) = (ϕ+p)X∗X(ϕ+p) = (ϕ+p)P+(ϕ+p) = ϕ+p , (X(ϕ+p))(X(ϕ+p))∗ = X(ϕ+p)(ϕ+p)∗X∗ = X(ϕ+p)X∗ = ψϕ+p = ϕ−p . e) follows from b), c), and Lemma 1.3.2. f) follows from Corollary 2.1.17 d). Remark. Even if ϕ± is injective P±S(f )P± is not an E-C-subalgebra of S(f ). 91 THEOREM 2.2.18 Let S be a finite subgroup of T , L := l2(S), g := f(S × S), ω : ZZ2 × ZZ2 → T an injective group homomorphism such that S ∩ ω( ZZ2 × ZZ2 ) = {1}, a := ω(1, 0) , c := ω(1, 1) , α1 := f (a, a) , α2 := f (b, b) , b := ω(0, 1) , β1, β2 ∈ U n Ec such that α1β2 1 + α2β2 2 = 0, γ := 1 2 X := 1 2 ((β1e⊗1K )V f (α∗1β∗1 β2 − α∗2β1β∗2 ) = α∗1β∗1β2 = −α∗2β1β∗2 , a + (β2e⊗1K)V f b ) , P+ := X∗X , P− := XX∗ . We assume f (s, c) = f (c, s) and cs = sc for every s ∈ S, and f (a, b) = −f (b, a) = 1E. Moreover we consider S(g) as an E-C-subalgebra of S(f ) (Corollary 2.1.17 e)). a) We have f (a, c) = −f (c, a) = α1 , f (b, c) = −f (c, b) = −α2 , f (c, c) = −α1α2 , γ2 = −α∗1α∗2 , V f c ∈ S(g)c . b) We have 1 2 (V f 1 , P+P− = 0 , P± = X 2 = 0, XP+ = X, P−X = X, P+X = XP− = 0, X + X∗ ∈ U n S(f ) , c ) ∈ S(g)c ∩ P r S(f ), P+ + P− = V f 1 ± (γe⊗1K )V f Y ∈ S(g) =⇒ XY X = 0 . c) The map E −→ P±S(f )P±, is a unital injective C-homomorphism; we shall consider P±S(f )P± as an E-C-algebra using this map. x 7−→ (xe⊗1K )P± d) The maps ϕ+ : S(g) −→ P+S(f )P+ , ϕ− : S(g) −→ P−S(f )P− , Y 7−→ P+Y P+ , Y 7−→ XY X∗ 92 are orthogonal injective E-C-homomorphisms and ϕ+ + ϕ− is an injec- tive E-C-homomorphism. If Y1, Y2 ∈ U n S(g) (resp. Y1, Y2 ∈ P r S(g)) then ϕ+Y1 + ϕ−Y2 ∈ U n S(f ) (resp. ϕ+Y1 + ϕ−Y2 ∈ P r S(f )). Moreover the map ψ : S(f ) −→ S(f ) , Z 7−→ (X + X∗)Z(X + X∗) is an E-C*-isomorphism such that ψ−1 = ψ , ψ(P+S(f )P+) = P−S(f )P− , ψ ◦ ϕ+ = ϕ− . If IK = IC then X +X∗ is homotopic to V f 1 in U n S(f ) and ψ is homotopic to the identity map of S(f ). Using this homotopy we find that ϕ+Y is homotopic in the above sense to ϕ−Y for every Y ∈ S(g) and ϕ+Y1 + ϕ−Y2, ϕ−Y1 + ϕ+Y2, ϕ+(Y1Y2) + P−, and ϕ+(Y2Y1 + P− are homotopic in the above sense for all Y1, Y2 ∈ S(g). e) Let s ∈ S such that sa = as. Then sb = bs , f (sc, c)f (s, c) = −α1α2 , f (sa, c)f (c, sa)∗ = −1E , f (a, s)f (s, a)∗ = f (b, s)f (s, b)∗ . f ) If sa = as for every s ∈ S then the map S × ( ZZ2 × ZZ2 ) −→ T, is an injective group homomorphism. (s, r) 7−→ s(ωr) g) If T is generated by S ∪ ω( ZZ2 × ZZ2 ) and sa = as for every s ∈ S then ϕ+ and ψ− are E-C-isomorphisms with inverse P±S(f )P± −→ S(g), Z 7−→ 2Xs∈S s , (Zse⊗1L)V g where ψ− : S(g) −→ P−S(f )P− , Y 7−→ P−Y P− . h) If sa = as and f (a, s) = f (s, a) for every s ∈ S then X ∈ S(g)c, ϕ−Y = P−Y for every Y ∈ S(g), and there is a unique S(g)-C*-homomorphism φ : S(g)2,2 → S(f ) such that φ(cid:20) 0 (α1β2 1 ) ⊗ 1L 0 (cid:21) = X . 0 93 φ is injective and φ(cid:20) V g 1 0 0 0 (cid:21) = P+ , φ(cid:20) 0 0 0 V g 1 (cid:21) = P− . i) If sa = as and f (a, s) = f (s, a) for all s ∈ S and if T is generated by 0 −γ∗ ⊗ 1L (cid:21) , S ∪ ω( ZZ2 × ZZ2 ) then φ is an S(g)-C-isomorphism and φ−1V f 0 1 =(cid:20) 1E ⊗ 1L 0 0 0 0 φ−1V f 1E ⊗ 1L (cid:21) , φ−1V f a =(cid:20) (β2γ∗) ⊗ 1L b =(cid:20) φ−1P+ =(cid:20) V g c =(cid:20) γ∗ ⊗ 1L (cid:21) , −β∗1 ⊗ 1L (cid:21) , 1 (cid:21) , 0 0 V g φ−1V f 1 0 (β1γ∗) ⊗ 1L 0 (cid:21) , 0 s =(cid:20) V g s 0 −β∗2 ⊗ 1L φ−1P− =(cid:20) 0 s (cid:21) . 0 V g φ−1V f 0 0 and for every s ∈ S j) The above results still hold for an arbitrary subgroup S of T if we replace S with Sk·k. a) By the equation of the Schur functions, f (a, a) = f (a, c)f (a, b) , f (a, b)f (c, a) = f (a, c)f (b, a) , f (a, b)f (c, b) = f (b, b) , f (b, a)f (c, b) = f (b, c)f (a, b) , f (a, b)f (c, c) = f (a, a)f (b, c) and so α1 = f (a, c) , f (c, a) = −f (a, c) = −α1 , f (c, b) = α2 , −α2 = −f (c, b) = f (b, c) , For s ∈ S, by Proposition 2.1.2 b), f (c, c) = α1f (b, c) = −α1α2 . V f c V f s = (f (c, s)e⊗1K )V f cs = (f (s, c)e⊗1K )V f c ∈ S(g)c (by Proposition 2.1.2 d)). and so V f sc = V f s V f c 94 b) By Proposition 2.1.2 b),d),e) (and Corollary 2.1.22 c)), (2V f (2V f 1 4 1 4 P+ = P− = By a), X∗ = 1 2 (((α∗1β∗1 )e⊗1K )V f a + ((α∗2β∗2)e⊗1K )V f b ) , 1 + ((α∗1β∗1 β2)e⊗1K )V f 1 + ((β1α∗2β∗2 )e⊗1K)V f c − ((α∗2β∗2 β1)e⊗1K )V f c − ((β2α∗1β∗1)e⊗1K )V f 1 ± (γ∗e⊗1K )((−α∗1α∗2)e⊗1K )V f (V f 1 2 P ∗ ± = 1 (V f 4 c ) = (V f c ) = (V f 1 2 1 2 c ) , c ) . 1 + (γe⊗1K )V f 1 − (γe⊗1K )V f c ) = P± , 1 2 1 ) = = (V f P 2 ± = 1 ± 2(γe⊗1K )V f c + (γ2e⊗1K)((−α1α2)e⊗1K )V f 1 ± (γe⊗1K)V f 1 + ((β1β2)e⊗1K )(V f so, by a) again, P± ∈ S(g)c ∩ P r S(f ). By Proposition 2.1.2 b),d), b V f 1 4 (X + X∗)2 = X 2 + XX∗ + X∗X + X∗2 = P+ + P− = V f 1 . 2α2)e⊗1K )V f c ) = P± , 1 α1 + β2 b + V f X 2 = a V f (((β2 For the last relation we remark that by the above, XY X = X(P+ + P−)Y X = XP+Y X = XY P+X = 0 . c) follows from b) and Lemma 1.3.2. a )) = 0 , d) By b) and c), the map ϕ± is an E-C-homomorphism. Let Y ∈ S(g) c so by Proposition 2.1.2 b),d) and Theorem 2.1.9 b), with ϕ±Y = 0. By b), Y = ∓Y (γe⊗1K )V f s = ∓Y (γe⊗1K )V f Xs∈S (Yse⊗1K )V f c = ∓Xs∈S ((Ysγf (s, c))e⊗1K )V f sc , which implies Ys = 0 for every s ∈ S (Theorem 2.1.9 a)). Thus ϕ± is injective. It follows that ϕ+ + ϕ− is also injective. Assume first Y1, Y2 ∈ U n S(g). By b), (ϕ+Y1 + ϕ−Y2)∗(ϕ+Y1 + ϕ−Y2) = (ϕ+Y ∗1 + ϕ−Y ∗2 )(ϕ+Y1 + ϕ−Y2) = 95 = ϕ+(Y ∗1 Y1) + ϕ−(Y ∗2 Y2) = P+ + P− = V f 1 . Similarly (ϕ+Y1 + ϕ−Y2)(ϕ+Y1 + ϕ−Y2)∗ = V f easy to see. 1 . The case Y1, Y2 ∈ P r S(g) is By b), ψ is an E-C-isomorphism with ψ−1 = ψ , ψP+ = (X + X∗)X∗X(X + X∗) = XX∗XX∗ = P− . Moreover for Y ∈ S(g), ψϕ+Y = (X + X∗)P+Y P+(X + X∗) = XY X∗ = ϕ−Y . Assume now IK = IC. By b), X + X∗ ∈ U n S(f ). Being selfadjoint its 1 in U n S(f ). spectrum is contained in {−1, +1} and so it is homotopic to V f e) We have sb = sac = asc = acs = bs. By a), f (s, c)f (sc, c) = f (s, 1)f (c, c) = −α1α2 , f (s, a)f (sa, c) = f (s, b)f (a, c) = α1f (s, b) , f (c, as)f (a, s) = f (c, a)f (b, s) = −α1f (b, s) , f (c, bs)f (b, s) = f (c, b)f (a, s) = α2f (a, s) , f (s, c)f (sc, b) = f (s, a)f (c, b) = α2f (s, a) , f (c, s)f (cs, b) = f (c, sb)f (s, b) so From f (sa, c)f (c, as)∗ = −f (s, b)f (s, a)∗f (b, s)∗f (a, s) = = −f (c, s)f (cs, b)f (c, sb)∗α2f (s, c)∗f (sc, b)∗α∗2f (c, bs) = −1E . f (s, c)f (sc, a) = f (s, b)f (c, a) , f (c, a)f (b, s) = f (c, as)f (a, s) , we get f (c, s)f (cs, a) = f (c, sa)f (s, a) f (a, s)f (s, a)∗ = f (b, s)f (s, b)∗ . 96 f) Since S and ω( ZZ2 × ZZ2 ) commute, the map is a group homomorphism. If s(ωr) = 1 for (s, r) ∈ S × ( ZZ2 × ZZ2 ) then ωr = s−1 ∈ S ∩ ω( ZZ2 × ZZ2 ), which implies s = 1 and r = (0, 0). Thus this group homomorphism is injective. g) By e) and the hypothesis of f), for every t ∈ T there are uniquely s ∈ S and d ∈ {1, a, b, c} with t = sd. Let Z ∈ P±S(f )P±. By b) and Theorem 2.1.9 b) (and Corollary 1.3.7 d)), By Proposition 2.1.2 b), ZV f V f c Z Z = ±(γe⊗1K )ZV f ((Zsf (s, c))e⊗1K)V f ((Zsbf (sb, c))e⊗1K )V f ((f (c, s)Zs)e⊗1K)V f ((f (c, sb)Zsb)e⊗1K )V f c = ±(γe⊗1K )V f sc +Xs∈S sa +Xs∈S sc +Xs∈S sa +Xs∈S ((Zsaf (sa, c))e⊗1K )V f ((Zscf (sc, c))e⊗1K )V f ((f (c, sa)Zsa)e⊗1K )V f ((f (c, sc)Zsc)e⊗1K )V f c =Xs∈S +Xs∈S c Z =Xs∈S +Xs∈S s , sb+ s sb+ and so by Theorem 2.1.9 a), Zs = ±γf (sc, c)Zsc = ±γf (c, sc)Zsc , Zsc = ±γf (s, c)Zs = ±γf (c, s)Zs , Zsa = ±γf (sb, c)Zsb = ±γf (c, sb)Zsb , Zsb = ±γf (sa, c)Zsa = ±γf (c, sa)Zsa . By e), Zsa = Zsb = 0 for every s ∈ S. We get (by a), d), and Proposition 2.1.2 b)) ϕ±(2Xs∈S s ) =Xs∈S (Zse⊗1K )V f (Zse⊗1L)V g s ±Xs∈S =Xs∈S (Zse⊗1K )V f =Xs∈S s +Xs∈S (Zse⊗1K )V f c Xs∈S s ± (γe⊗1K )V f ((γf (c, s)Zs)e⊗1K )V f (Zsce⊗1K )V f sc = Z . s = (Zse⊗1K )V f sc = 97 Thus ϕ± is an E-C-isomorphism with the mentioned inverse. h) is a long calculation using e). i) follows from h). j) follows from Corollary 2.1.17 d). Remark. An example in which the above hypotheses are fulfilled is given in Theorem 4.1.7. 2.3 The functor S Throughout this subsection we assume T finite In this subsection we present the construction in the frame of category theory. Some of the results still hold for T locally finite. DEFINITION 2.3.1 The above construction of S(f ) can be done for an arbi- trary E-module F , in which case we shall denote the result by S(F ). Moreover we shall write V F t instead of V f t in this case. If F is an E-module then S(F ) is canonically an E-module. If in addition If F is an F is adapted then S(F ) is adapted and isomorphic to S( F , F ). E-C-algebra then S(F ) is also an E-C-algebra. PROPOSITION 2.3.2 If F, G are E-modules and ϕ : F → G is an E-linear C-homomorphism then the map S(ϕ) : S(F ) −→ S(G) , X 7−→Xt∈S ((ϕXt) ⊗ 1K )V G t is an E-linear C-homomorphism, injective or surjective if ϕ is so. The assertion follows from Theorem 2.1.9 a),c),g). 98 COROLLARY 2.3.3 Let F1, F2, F3 be E-modules and let ϕ : F1 → F2, ψ : F2 → F3 be E-linear C-homomorphisms. a) S(ψ) ◦ S(ϕ) = S(ψ ◦ ϕ). b) If the sequence is exact then the sequence 0 −→ F1 ϕ −→ F2 ψ −→ F3 0 −→ S(F1) S(ϕ) −→ S(F2) S(ψ) −→ S(F3) is also exact. c) The covariant functor S : ME → ME is exact. a) is obvious. b) Let Y ∈ Ker S(ψ). For every t ∈ T , Yt ∈ Ker ψ = Im ϕ. If we identify F1 with Im ϕ then Yt ∈ F1. It follows Y ∈ ImS(ϕ), Ker S(ψ) = ImS(ϕ). c) follows from b) and Proposition 2.3.2. COROLLARY 2.3.4 Let F be an adapted E-module and put ι : F −→ F , x 7−→ (0, x) , π : F −→ E , (α, x) 7−→ α , λ : E −→ F , α 7−→ (α, 0) . 0 −→ S(F ) S(ι) −→ S( F ) S(π) −→ S(λ) ←− S(E) −→ 0 Then the sequence is split exact. PROPOSITION 2.3.5 The covariant functor S : ME → ME (resp. S : C1 E → C1 E) (Proposition 2.3.2, Corollary 2.3.3 a)) is continuous with respect to the inductive limits (Proposition 1.2.9 a),b)). 99 Let {(Fi)i∈I , (ϕij)i,j∈I} be an inductive system in the category ME (resp. C1 E) and let {F, (ϕi)i∈I} be its limit in the category ME (resp. C1 E). Then {(S(Fi))i∈I , (S(ϕij )i,j∈I)} is an inductive system in the category ME (resp. C1 E). Let {G, (ψi)i∈I} be its limit in this category and let ψ : G → S(F ) be the E-linear C-homomorphism such that ψ ◦ ψi = S(ϕi) for every i ∈ I. In the C1 E case, for α ∈ E and i ∈ I, ψ(α ⊗ 1K) = ψ ◦ ψi(α ⊗ 1K) = (S(ϕi))(α ⊗ 1K ) = α ⊗ 1K so that ψ is an E-C-homomorphism. Let i ∈ I and let X ∈ Ker S(ϕi). Then ϕiXt = 0 for every t ∈ T . Since T is finite, for every ε > 0 there is a j ∈ I, j ≥ i, with kϕjiXtk < ε Card T for every t ∈ T . Then k(S(ϕji))Xk =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)Xt∈T ((ϕjiXt) ⊗ 1K )V < ε . Fj t (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) It follows kψiXk = inf j∈I,j≥ik(S(ϕji))Xk = 0 , X ∈ Ker ψi , By Lemma 1.2.8, ψ is injective. Since ψiX = 0 , Ker S(ϕi) ⊂ Ker ψi . [i∈I ImS(ϕi) ⊂ Im ψ , Im ψ is dense in S(F ). Thus ψ is surjective and so an E-C-isomorphism. PROPOSITION 2.3.6 Let θ : F → G be a surjective morphism in the cate- gory C1 E. We use the notation of Theorem 2.2.18 and mark with an exponent if this notation is used with respect to F or to G. For every Y ∈ U n S(gG), there is a Z ∈ S(gF ) such that Z∗Z = P F + , +Y . S(θ)Z = ϕG 100 By Proposition 2.3.2 c), S(θ) is surjective and so there is a Z0 ∈ S(gF ) with kZ0k = 1 and S(θ)Z0 = Y . Put Z := P F + Z0 + X F (1 − Z∗0 Z0) 1 2 . By Theorem 2.2.18 b), Z∗Z = P F + Z∗0 Z0 + (1 − Z∗0 Z0) = P F + Z∗0 Z0 + P F 1 2 (X F )∗X F (1 − Z∗0 Z0) 1 2 = + (1 − Z∗0 Z0) = P F + . Since we get S(θ)(1 − Z∗0 Z0) = 1 − Y ∗Y = 0 S(θ)(1 − Z∗0 Z0) 1 2 = 0 , S(θ)Z = P G + Y = ϕG +Y . PROPOSITION 2.3.7 Let F be an adapted E-module and Ω a locally com- pact space. We define for X ∈ S(C0(Ω, F )) (see Corollary 1.2.5 d)) and Y ∈ C0(Ω,S(F )), ϕX : Ω −→ S(F ) , ω 7−→Xt∈T (Xt(ω) ⊗ 1K )V F t , ψY :=Xt∈T (Y (·)t ⊗ 1K )V C0(Ω,F ) t . Then ϕ : S(C0(Ω, F )) −→ C0(Ω,S(F )) , ψ : C0(Ω,S(F )) −→ S(C0(Ω, F )) are E-linear C-isomorphisms and ϕ = ψ−1. ψ induce the following E-C-isomorphisms Let ω0 ∈ Ω and assume F is an E-C-algebra. Then the above maps ϕ and S({ X ∈ C0(Ω, F ) X(ω0) ∈ E })−→ ←− { Y ∈ C0(Ω,S(F )) Y (ω0) ∈ S(E)} . Let X, X′ ∈ S(C0(Ω, F )) and Y, Y ′ ∈ C0(Ω,S(F )). By Proposition 2.1.23 b) and Corollary 2.1.10 a), ϕX ∈ C0(Ω,S(F )) , ψY ∈ S(C0(Ω, F )) 101 and it is easy to see that ϕ and ψ are E-linear. By Theorem 2.1.9 c),g), for t ∈ T and ω ∈ Ω, ((ϕX)∗(ω))t = f (t)(((ϕX)(ω)t−1 ))∗ = = f (t)Xt−1 (ω)∗ = (X∗(ω))t = ((ϕX∗)(ω))t , f (s, s−1t)((ϕX)(ω))s((ϕX′)(ω))s−1t = (((ϕX)(ϕX′))(ω))t =Xs∈T =Xs∈T f (s, s−1t)Xs(ω)X′s−1t(ω) = Xs∈T = (XX′)t(ω) = ((ϕ(XX′))(ω))t , f (s, s−1t)XsX′s−1t! (ω) = so (ϕX)∗ = ϕX∗ , (ϕX)(ϕX′) = ϕ(XX′) and ϕ is a C-homomorphism. Similarly (ψY ∗)t(ω) = (Y ∗(ω))t = f (t)(Y (ω)t−1 )∗ = f (t)((ψY )t−1 (ω))∗ = ((ψY )∗)t(ω) , f (s, s−1t)(ψY )s(ψY ′)s−1t! (ω) = ((ψY )(ψY ′))t(ω) = Xs∈T f (s, s−1t)(ψY )s(ω)(ψY ′)s−1t(ω) =Xs∈T f (s, s−1t)Y (ω)sY ′(ω)s−1t = = (Y (ω)Y ′(ω))t = (ψ(Y Y ′)t)(ω) =Xs∈T so ψY ∗ = (ψY )∗ , (ψY )(ψY ′) = ψ(Y Y ′) and ψ is a C-homomorphism. Moreover (ψϕX)t(ω) = ((ϕX)(ω))t = Xt(ω) , ((ϕψY )(ω))t = (ψY )t(ω) = (Y (ω))t , so ψϕX = X and ϕψY = Y which proves the assertion. The last assertion is easy to see. PROPOSITION 2.3.8 Let F be an adapted E-module, 0 −→ F ι−→ F π−→ E −→ 0 , 102 the associated exact sequences (Proposition 1.2.4 h)), and 0 −→ S(F ) ι0−→ π0−→ E −→ 0 z}{S(F ) j : E −→ S(E) , α 7−→ (α ⊗ 1K )V E 1 , F 1 . Then ϕ is an injective E-C-homomorphism and S(π) ◦ ϕ = j ◦ π0. (α, X) 7−→ S(ι)X + (α ⊗ 1K )V z}{S(F ) −→ S( F ) , ϕ : PROPOSITION 2.3.9 If E is commutative and F is an E-module then the map ϕ : S(E) ⊗ F −→ S(F ) , X ⊗ x 7−→Xt∈T ((Xtx) ⊗ 1K)V F t is a surjective C-homomorphism. isomorphism with inverse If in addition E = IK then ϕ is a C- ψ : S(F ) −→ S(E) ⊗ F , Y 7−→Xt∈T (V E t ⊗ Yt) . It is obvious that ϕ is surjective. For X, Y ∈ S(E) and x, y ∈ F , by Theorem 2.1.9 c),g) and Proposition 2.1.2 b),d),e), (((X∗)tx∗) ⊗ 1K )V F t = (((Xt−1 )∗x∗) ⊗ 1K)(V F t−1 )∗ = t )∗ = (ϕ(X ⊗ x))∗ , ((XsxYty) ⊗ 1K )V F s V F t = r = ϕ((X ⊗ x)(Y ⊗ y)) ((f (s, t)XsxYty) ⊗ 1K )V F ((f (s, s−1r)XsYs−1rxy) ⊗ 1K )V F r = =Xt∈T ϕ((X ⊗ x)∗) = ϕ(X∗ ⊗ x∗) =Xt∈T t =Xt∈T (( f (t)(Xt−1 )∗x∗) ⊗ 1K )V F =Xt∈T ((x∗(Xt)∗) ⊗ 1K)(V F ϕ(X ⊗ x)ϕ(Y ⊗ y) = Xs,t∈T st =Xr∈TXs∈T (((XY )rxy) ⊗ 1K )V F =Xr∈T = Xs,t∈T so ϕ is a C-homomorphism. 103 Assume now E = IK and let X ∈ S(E) and x ∈ F . Then ((Xtx) ⊗ 1K )V F V E t ⊗ (Xtx) = ψϕ(X ⊗ x) = ψXt∈T = Xt∈T t =Xt∈T t ! ⊗ x = X ⊗ x XtV E which proves the last assertion (by using the first assertion). 3 Examples We draw the reader's attention to the fact that in additive groups the neutral element is denoted by 0 and not by 1. 3.1 T := ZZ2 PROPOSITION 3.1.1 a) The map ψ : F( ZZ2 , E) −→ U n Ec , f 7−→ f (1, 1) is a group isomorphism. b) ψ({ δλ λ ∈ Λ( ZZ2 , E)}) =(cid:8) x2(cid:12)(cid:12) x ∈ U n Ec(cid:9). c) If there is an x ∈ Ec with x2 = f (1, 1) (in which case x ∈ U n Ec) then the map ϕ : S(f ) −→ E × E , X 7−→ (X0 + xX1, X0 − xX1) is an E-C-isomorphism. d) If IK = IC and if A is a connected and simply connected compact space or a totally disconnected compact space then for every x ∈ U n C(A) there is a y ∈ C(A, IC) with x = ey. e) Assume IK = IR. 104 e1) There are uniquely p, q ∈ Pr Ec with pf (1, 1) = p, p + q = 1E, e2) The map qf (1, 1) = −q . ϕ : S(f ) −→ (pE) × (pE) × ◦z}{qE , X 7−→ X , ◦z}{qE denotes the complexification of the C-algebra qE and X := (p(X0 + X1), p(X0 − X1), (qX0, qX1)) where for every X ∈ S(f ), is an E-C-isomorphism. In particular if f (1, 1) = −1E then S(f ) is isomorphic to the complexification of E. f ) Assume IK = IC, let σ(Ec) be the spectrum of Ec, and let cf11 be the function of C(σ(Ec), IC) corresponding to f11 by the Gelfand transform. Then is the spectrum of V1. n eiθ(cid:12)(cid:12)(cid:12) θ ∈ IR, e2iθ ∈ cf11(σ(Ec))o a) follows from Proposition 1.1.2 a) (and Proposition 1.1.4 a) ). b) follows from Definition 1.1.3. c) For X, Y ∈ S(f ), by Theorem 2.1.9 c),g) (and Proposition 1.1.2 a)), (X∗)0 = (X0)∗ , (X∗)1 = (x∗)2(X1)∗ , (XY )0 = X0Y0 + x2X1Y1 , (XY )1 = X0Y1 + X1Y0 , so ϕ(X∗) = ((X0)∗ + x(x∗)2(X1)∗, (X0)∗ − x(x∗)2(X1)∗) = = ((X0)∗ + x∗(X1)∗ , (X0)∗ − x∗(X1)∗) = (ϕX)∗ , (ϕX)(ϕY ) = ((X0 + xX1)(Y0 + xY1), (X0 − xX1)(Y0 − xY1)) = = (X0Y0 + xX0Y1 + xX1Y0 + x2X1Y1, X0Y0 − xX0Y1 − xX1Y0 + x2X1Y1) = 105 = ((XY )0 + x(XY )1, (XY )0 − x(XY )1 = ϕ(XY ) i.e. ϕ is an E-C-homomorphism. ϕ is obviously injective. Let (y, z) ∈ E × E. If we take X ∈ S(f ) with 1 2 (y + z), X0 := 1 2 X1 := x∗(y − z) then ϕX = (y, z), i.e. ϕ is surjective. d) is known. e1) follows by using the spectrum of Ec. e2) Put ψ : S(f ) −→ For X, Y ∈ S(f ), by Theorem 2.1.9 c),g), ◦z}{qE , X 7−→ (qX0, qX1) . ψ(X∗) = (q(X∗)0, q(X∗)1) = (q(X0)∗, qf (1, 1)∗(X1)∗) = = ((qX0)∗,−(qX1)∗) = (ψX)∗ , (ψX)(ψY ) = (qX0, qX1)(qY0, qY1) = = (q(X0Y0 − X1Y1), (q(X0Y1 + X1Y0))) = ψ(XY ) so ψ is an E-C-homomorphism. Thus by c), ϕ is an E-C-homomorphism. The bijectivity of ϕ is easy to see. f) By Proposition 2.1.2 e), V1 is unitary so its spectrum is contained in (cid:8) eiθ(cid:12)(cid:12) θ ∈ IR(cid:9). For θ ∈ IR and X ∈ S(f ), (eiθV0 − V1)X = X(eiθ − V1) = = ((eiθX0) ⊗ 1K )V0 + ((eiθX1) ⊗ 1K )V1 − (X0 ⊗ 1K )V1 − ((f11X1) ⊗ 1K )V1 = = ((eiθX0 − f11X1) ⊗ 1K)V0 + ((eiθX1 − X0) ⊗ 1K )V1 . Thus X is the inverse of eiθV0 − V1 iff X0 = eiθX1 and eiθX0 − f11X1 = 1E, i.e. (e2iθ − f11)X1 = 1E. Therefore eiθV0 − V1 is invertible iff e2iθ − cf11 does not vanish on σ(Ec). 106 COROLLARY 3.1.2 Assume IK := IR and let S be a group, F a unital C- algebra, g ∈ F(S, F ), and h : (S × ZZ2 ) × (S × ZZ2 ) −→ U n F c , ((s1, t1), (s2, t2)) 7−→ (cid:26) −g(s1, s2) g(s1, s2) if if (t1, t2) = (1, 1) (t1, t2) 6= (1, 1) . a) h ∈ F(S × ZZ2 , F ). b) S(h) ≈ ◦z}{S(g) , Sk·k(h) ≈ ◦z } { Sk·k(g). Put E := IR in the above Proposition and define f ∈ F( ZZ2 , IR) by f (1, 1) = −1 (Proposition 3.1.1 a)). By this Proposition e2), S(f ) ≈ IC. Thus by Propo- sition 2.2.11 c),e), S(h) ≈ S(g) ⊗ S(f ) ≈ ◦z}{S(g) , DEFINITION 3.1.3 We put Sk·k(h) ≈ Sk·k(g) ⊗ Sk·k(f ) ≈ ◦z } { Sk·k(g) . IT := { z ∈ IC z = 1} . EXAMPLE 3.1.4 Let E := C( IT, IC) and f ∈ F( ZZ2 , E) with f (1, 1) : IT −→ U n IC , z 7−→ z . If we put X : IT −→ IC , for every X ∈ S(f ) then the map z 7−→ X0(z2) + zX1(z2) ϕ : S(f ) −→ E , X 7−→ X is an isomorphism of C-algebras (but not an E-C-isomorphism). 107 For X, Y ∈ S(f ), by Theorem 2.1.9 c),g), (X∗)0 = (X0)∗, (X∗)1 = f (1, 1)(X1)∗ , (XY )0 = X0Y0 + f (1, 1)X1Y1, (XY )1 = X0Y1 + X1Y0 so for z ∈ IT, fX∗(z) = X∗0 (z2) + z¯z2X∗1 (z2) = X0(z2) + zX1(z2) = X∗(z) , ( X(z))( Y (z)) = (X0(z2) + zX1(z2))(Y0(z2) + zY1(z2)) = = X0(z2)Y0(z2) + zX0(z2)Y1(z2) + zX1(z2)Y0(z2) + z2X1(z2)Y1(z2) = = (XY )0(z2) + z(XY )1(z2) = gXY (z) , fX∗ = X∗ , X Y = gXY , i.e. ϕ is a C-homomorphism. If ϕX = 0 then for z ∈ IT, X0(z2) + zX1(z2) = 0 so, successively, X0(z2) − zX1(z2) = 0 , X0(z2) = X1(z2) = 0 , X0 = X1 = 0 , X = 0 and ϕ is injective. Put Let G :=(Xk∈ ZZ (ck)k∈ ZZ ∈ IC( ZZ )) ⊂ E . ckzk ∈ G ckzk(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) x :=Xk∈ ZZ and take X ∈ S(f ) with X0 :=Xk∈ ZZ X =Xk∈ ZZ c2kzk, X1 :=Xk∈ ZZ c2k+1zk . c2kz2k + zXk∈ ZZ c2k+1z2k = x Then so G ⊂ ϕ(S(f )). Since G is dense in E, ϕ(S(f )) = E and ϕ is surjective. 108 DEFINITION 3.1.5 For every x ∈ C( IT, IC) which does not take the value 0 we put w(x) := winding number of x := 1 2πiZx dz z = 1 2πi [log x(eiθ)]θ=2π θ=0 ∈ ZZ . If A is a connected compact space and γ is a cycle in A (i.e. a continuous map of IT in A), which is homologous to 0 (or more generally, if a multiple of γ is homologous to 0), then for every x ∈ C(A, U n IC) we have w(x ◦ γ) = 0. If A is a compact space and x ∈ C(A, U n IC) such that w(x ◦ γ) = 0 for every cycle γ in A then there is a y ∈ C(A, IC) with x = ey. EXAMPLE 3.1.6 Let E := C( IT, IC), f ∈ F( ZZ2 , E), and n := w(f (1, 1)). a) If n is even then there is an x ∈ U n E with winding number equal to n 2 such that the map S(f ) −→ E × E, X 7−→ (X0 + xX1, X0 − xX1) is an E-C-isomorphism. b) If n is odd then S(f ) is isomorphic to E. c) The group F( ZZ2 , E)/Λ( ZZ2 , E) is isomorphic to ZZ2 and Card ({ S(g) g ∈ F( ZZ2 , E)}/≈S ) = 2 . d) There is a complex unital C-algebra E and a family (fβ)β∈P(IN) in F( ZZ2 , E) such that for distinct β, γ ∈ P(IN), S(fβ) 6≈ S(fγ). Put α : IT −→ U n IC , z 7−→ z . Since w(f (1, 1)α−n) = 0, there is a y ∈ U n E with w(y) = 0 and f (1, 1)α−n = y2. 2 and f (1, 1) = x2 and the assertion a) If we put x := yα n 2 then w(x) = n follows from Proposition 3.1.1 c). 109 n−1 b) We put x := yα 2 . Then f (1, 1) = αx2. Take g ∈ F( ZZ2 , E) with g(1, 1) = α and λ ∈ Λ( ZZ2 , E) with (δλ)(1, 1) = x2 (Proposition 3.1.1 a),b)). Then f = gδλ. By Example 3.1.4, S(g) is isomorphic to E and by Proposition 2.2.2 a1 ⇒ a2, S(f ) is also isomorphic to E. c) follows from Proposition 3.1.1 b) and Proposition 2.2.2 a),c). d) Denote by E the C-direct product of the sequence (C( IT, ICn,n))n∈IN and for every β ∈ {0, 1}IN define fβ ∈ F( ZZ2 , E) by fβ(1, 1) : IN −→ U n Ec , n 7−→ αβ(n)1ICn,n . By a) and b), for distinct β, γ ∈ {0, 1}IN, S(fβ) 6≈ S(fγ) (Proposition 2.1.26 a)). EXAMPLE 3.1.7 Let I, J be finite disjoint sets and for all i ∈ I ∪ J and j ∈ J put Ai := Bj := IT. We define the compact spaces A and B in the following way. For A we take first the disjoint union of the spaces Ai for all i ∈ I ∪ J and identify then the points 1 ∈ Ai for all i ∈ I ∪ J. For B we take first the disjoint union of all the spaces Ai for all i ∈ I ∪ J and of the spaces Bj for all j ∈ J and identify first the points 1 ∈ Ai for all i ∈ I ∪ J and identify then also the points −1 ∈ Ai for all i ∈ I and 1 ∈ Bj for all j ∈ J. Let E := C(A, IC) and f ∈ F( ZZ2 , E) with f (1, 1) : A −→ U n IC , z 7−→(cid:26) z 1 For every X ∈ S(f ) define X ∈ C(B, IC) by X : B −→ IC , Then the map z 7−→ is an isomorphism of C-algebras. ϕ : S(f ) −→ C(B, IC) , X 7−→ X if if z ∈ Ai with i ∈ I z ∈ Ai with i ∈ J . X0(z2) + zX1(z2) X0(z) + X1(z) X0(z) − X1(z) if if if z ∈ Ai with i ∈ I z ∈ Ai with i ∈ J z ∈ Bj with j ∈ J . Let X, Y ∈ S(f ). By Theorem 2.1.9 c),g), (X∗)0 = (X0)∗ , (X∗)1 = f (1, 1)(X1)∗ , 110 (XY )0 = X0Y0 + f (1, 1)X1Y1 , (XY )1 = X0Y1 + X1Y0 . For z ∈ Ai with i ∈ I, fX∗(z) = (X∗)0(z2) + z(X∗)1(z2) = X0(z2) + z¯z2X1(z2) = = X0(z2) + zX1(z2) = ( X)∗(z) , X(z) Y (z) = (X0(z2) + zX1(z2))(Y0(z2) + zY1(z2)) = = X0(z2)Y0(z2) + zX0(z2)Y1(z2) + zX1(z2)Y0(z2) + z2X1(z2)Y1(z2) = For z ∈ Aj or z ∈ Bj with j ∈ J, = (XY )0(z2) + z(XY )1(z2) = gXY (z) . fX∗(z) = (X∗)0(z) ± (X∗)1(z) = X0(z) ± X1(z) = ( X)∗(z) , X(z) Y (z) = (X0(z) ± X1(z))(Y0(z) ± Y1(z)) = = X0(z)Y0(z) ± X0(z)Y1(z) ± X1(z)Y0(z) + X1(z)Y1(z) = Thus ϕ is a C-homomorphism. Assume X = 0. For z ∈ Ai with i ∈ I, = (XY )0(z) ± (XY )1(z) = gXY (z) . X0(z2) + zX1(z2) = 0 so, successively, X0(z2) − zX1(z2) = 0 , X0(z2) = X1(z2) = 0 , X(z) = 0 . For z ∈ Aj with j ∈ J, so Thus ϕ is injective. (cid:26) X0(z) + X1(z)= 0 X0(z) − X1(z)= 0 , X0(z) = X1(z) = 0 , X(z) = 0 . Let x ∈ C(B, IC) such that for every i ∈ I there is a family (ci,k)k∈ ZZ ∈ IC( ZZ ) with x(z) =Xk∈ ZZ ci,kzk 111 for all z ∈ Ai. Define X0, X1 ∈ E in the following way. If z ∈ Ai with i ∈ I we put X0(z) :=Xk∈ ZZ ci,2kzk , X1(z) :=Xk∈ ZZ ci,2k+1zk . If z ∈ Aj with j ∈ J then we put z′ := z ∈ Bj, X0(z) := 1 2 (x(z) + x(z′)) , X1(z) := 1 2 (x(z) − x(z′)) . It is easy to see that X0 and X1 are well defined. Then X(z) =Xk∈ ZZ ci,2kz2k + zXk∈ ZZ ci,2k+1z2k = x(z) for all z ∈ Ai with i ∈ I and X(z) = x(z) for all z ∈ Aj ∪ Bj with j ∈ J. Since the elements x of the above form are dense in C(B, IC), ϕ is surjective. EXAMPLE 3.1.8 Let E := C( IT2, IC) and f, g ∈ F( ZZ2 , E) with (cid:26) f (1, 1) : IT2 −→ U n IC , g(1, 1) : IT2 −→ U n IC , (cid:26) S(f ) −→ E, X 7−→ X0(z2 S(g) −→ E, X 7−→ X0(z1, z2 (z1, z2) 7−→ z1 (z1, z2) 7−→ z2 . 1, z2) + z1X1(z2 1, z2) 2) + z2X1(z1, z2 2) Then the maps are isomorphisms of C-algebras. Remark. S(f ) and S(g) are isomorphic but not E-C-isomorphic. EXAMPLE 3.1.9 Let E := C( IT2, IC) and f ∈ F( ZZ2 , E) with (z1, z2) 7−→ z1z2 . f (1, 1) : IT2 −→ U n IC , If we put X : IT2 −→ IC , (z1, z2) 7−→ X0(z2 1, z2 2 ) + z1z2X1(z2 1, z2 2) for every X ∈ S(f ) then the map ϕ : S(f ) −→ E , X 7−→ X 112 is an injective unital C-homomorphism with ϕ(S(f )) = G :=(cid:8) x ∈ E (z1, z2) ∈ IT2 =⇒ x(z1, z2) = x(−z1,−z2)(cid:9) . In particular S(f ) is isomorphic to E. Let X, Y ∈ S(f ). By Theorem 2.1.9 c),g), (X∗)0 = (X0)∗ , (X∗)1 = f (1, 1)(X1)∗ , (XY )0 = X0Y0 + f (1, 1)X1Y1 , (XY )1 = X0Y1 + X1Y0 so for (z1, z2) ∈ IT2, fX∗(z1, z2) = X∗0 (z2 = X0(z2 1, z2 1, z2 2) + z1z2 ¯z2 1 ¯z2 2X∗1 (z2 1, z2 2) = 2) + z1z2X1(z2 ( X(z1, z2))( Y (z1, z2)) = 1, z2 2) = X(z1, z2) , = (X0(z2 2))(Y0(z2 1, z2 = X0(z2 +z1z2X1(z2 = (XY )0(z2 2) + z1z2X1(z2 1, z2 1 , z2 1, z2 1, z2 1, z2 1, z2 1 , z2 2) + z1z2X0(z2 2) + z2 2X1(z2 1, z2 2)Y0(z2 2)Y0(z2 2) + z1z2(XY )1(z2 1z2 2)) = 1, z2 2) + z1z2Y1(z2 1, z2 1, z2 2)+ 1, z2 1 , z2 2)Y1(z2 2)Y1(z2 2) = 2 ) = gXY (z1, z2) , i.e. ϕ is a unital C-homomorphism. If X = 0 then for (z1, z2) ∈ IT2, X0(z2 1, z2 2) + z1z2X1(z2 1, z2 2) = 0 so, successively, X0(z2 1, z2 2) − z1z2X1(z2 1, z2 2 ) = 0 , X0(z2 1, z2 2) = X1(z2 1, z2 2) = 0 , and ϕ is injective. X0 = X1 = 0 , X = 0 The inclusion S(f ) ⊂ G is obvious. Let (aj,k)j,k∈ ZZ , (bj,k)j,k∈ ZZ ∈ IC( ZZ × ZZ ) and x = Xj,k∈ ZZ aj,kz2j 1 z2k 2 + Xj,k∈ ZZ bj,kz2j+1 1 z2k+1 2 ∈ G . 113 Define X0 := Xj,k∈ ZZ aj,kzj 1zk 2 , X1 := Xj,k∈ ZZ bj,kzj 1zk 2 . Then X = x. Since the elements of the above form are dense in G, ϕ(S(f )) = G. If we consider the equivalence relation ∼ on IT2 defined by (z1, z2) ∼ (w1, w2) :⇐⇒ z1 = −w1, z2 = −w2 then the quotient space IT2/∼ is homeomorphic to IT2. Thus S(f ) is isomorphic to E. EXAMPLE 3.1.10 Let E := C( IT2, IC). a) For x ∈ U n E and z ∈ IT, w(x(· , z)) and w(x(z, · )) do not depend on z, where w denotes the winding number (Definition 3.1.5). b) If x ∈ U n E and if w(x(· , 1)) = w(x(1, · )) = 0 then there is a y ∈ U n E with x = y2. c) Let f ∈ F( ZZ2 , E) and put α : IT −→ IT2 , z 7−→ (z, 1) , m := w(f (1, 1) ◦ α) , β : IT −→ IT2 , n := w(f (1, 1) ◦ β) . z 7−→ (1, z) , c1) If m + n is odd then S(f ) is isomorphic to E. c2) If m and n are even then S(f ) is isomorphic to E × E. c3) If m and n are odd then S(f ) is isomorphic to E. d) The group F( ZZ2 , E)/Λ( ZZ2 , E) is isomorphic to ZZ2 × ZZ2 and Card ({ S(f ) f ∈ F( ZZ2 , E)} / ≈S ) = 4 . 114 a) follows by continuity. b) follows from a). c) Let g ∈ F( ZZ2 , E) with g(1, 1) : IT2 −→ U n IC , (z1, z2) 7−→ zm 1 zn 2 . Then w(g(1, 1) ◦ α) = m , w(g(1, 1) ◦ β) = n . By b), there is an x ∈ U n E with f (1, 1) = x2g(1, 1). By Proposition 3.1.1 b) and Proposition 2.2.2 a1 ⇒ a2, S(f ) ≈ S(g). c1) Assume m even and put y : IT2 −→ U n IC , If h ∈ F( ZZ2 , E) with (z1, z2) 7−→ z m 2 1 z 2 n−1 2 . h(1, 1) : IT2 −→ U n IC , (z1, z2) 7−→ z2 then g(1, 1) = y2h(1, 1). By Proposition 3.1.1 b) and Proposition 2.2.2 a1 ⇒ a2, S(g) ≈ S(h) and by Example 3.1.8 a1 ⇒ a2, S(h) ≈ E. Thus S(f ) ≈ E. c2) If we put y : IT2 −→ U n IC , (z1, z2) 7−→ z m 2 n 2 1 z 2 then g(1, 1) = y2 and the assertion follows from Proposition 3.1.1 c). c3) We put y : IT2 −→ U n IC , m−1 n−1 (z1, z2) 7−→ z 1 2 2 z 2 and take h ∈ F( ZZ2 , E) with h(1, 1) : IT2 −→ U n IC , (z1, z2) 7−→ z1z2 then g(1, 1) = y2h(1, 1) so by Proposition 3.1.1 b) and Proposition 2.2.2 a1 ⇒ a2, S(g) ≈ S(h). By Example 3.1.9 S(h) ≈ E, so S(f ) ≈ E. d) follows from b), Proposition 3.1.1 b), and Proposition 2.2.2 a),c). 115 Remark. In a similar way it is possible to show that for every n ∈ IN, F( ZZ2 , ITn)/Λ( ZZ2 , ITn) is isomorphic to ( ZZ2 )n and Card ({ S(f ) f ∈ F( ZZ2 , ITn)} / ≈S) = 2n . EXAMPLE 3.1.11 Let I, J, K be finite pairwise disjoint sets and for every i ∈ I ∪ J ∪ K and k ∈ K put Ai := Bk := IT2. We define the compact spaces A and B in the following way. For A we take first the disjoint union of the spaces Ai with i ∈ I ∪ J ∪ K and then identify the points (1, 1) ∈ Ai for all i ∈ I ∪ J ∪ K. For B we take first the disjoint union of the spaces Ai with i ∈ I ∪ J ∪ K and of the spaces Bk with k ∈ K. Then we identify the points (1, 1) ∈ Ai for all i ∈ I ∪ J ∪ K and then we identify for every j ∈ J the points (z1, z2) ∈ Aj with the points (−z1,−z2) ∈ Aj and finally we identify the points (−1, 1) ∈ Ai for all i ∈ I ∪ J with the points (1, 1) ∈ Bk for all k ∈ K. Let E := C(A, IC) and f ∈ F( ZZ2 , A) such that if if if f (1, 1) : A −→ U n IC , z1 z1z2 1 (z1, z2) 7−→ We define for every X ∈ S(f ) a map X : B → IC by (z1, z2) ∈ Ai with i ∈ I (z1, z2) ∈ Ai with i ∈ J (z1, z2) ∈ Ai with i ∈ K . (z1, z2) 7→ Then the map X0(z2 X0(z2 1, z2) + z1X1(z2 1 , z2 2) + z1z2X1(z2 1 , z2) 1, z2 2) X0(z1, z2) + X1(z1, z2) X0(z1, z2) − X1(z1, z2) if if if if (z1, z2) ∈ Ai with i ∈ I (z1, z2) ∈ Ai with i ∈ J (z1, z2) ∈ Ai with i ∈ K (z1, z2) ∈ Bk with k ∈ K . S(f ) −→ C(B, IC), X 7−→ X is an isomorphism of C-algebras. The proof is similar to the proof of Example 3.1.7. EXAMPLE 3.1.12 If n ∈ IN, E := C( ITn, IC), and f ∈ F( ZZ2 ,C( ITn, IC)) then S(f ) is isomorphic either to C( ITn, IC) or to C( ITn, IC) × C( ITn, IC). 116 EXAMPLE 3.1.13 Assume E := C(A, IC), where A denotes Moebius's band (resp. Klein's bottle), i.e. the topological space obtained from [0, 2π] × [−π, π] by identifying the points (0, α) and (2π,−α) for all α ∈ [−π, π] (resp. and the points (θ,−π) and (θ, π) for all θ ∈ [0, 2π]). We put B := IT × [−π, π] (resp. B := IT2) and x : [0, 2π] × [−π, π] −→ IC , for every x ∈ E. (θ, α) 7−→(cid:26) x(2θ, α) x(2(θ − π),−α) if if θ ∈ [0, π] θ ∈ [π, 2π] a) x is well-defined and belongs to C(B, IC) for every x ∈ E. b) If f1,1(θ, α) = eiθ for all (θ, α) ∈ [0, 2π] × [−π, π] then the map ϕ : S(f ) −→ C(B, IC) , X 7−→ fX0 + eiθfX1 is a C-isomorphism. c) Let x ∈ U n E. If w(x(· , 0)) = 0 (where w denotes the winding number) then there is a y ∈ E with ey = x. d) Let x ∈ U n E and put n := w(x(· , 0)). Then there is a y ∈ E with ey = e−inθx. e) The group F( ZZ2 , A)/Λ( ZZ2 , A) is isomorphic to ZZ2 . f ) If w(f1,1(· , 0)) is even (resp. odd) then S(f ) is isomorphic to E × E (resp. to C(B, IC)). a) For α ∈ [−π, π], x(π, α) = x(2π, α) = x(0,−α) = x(π, α) so x is well-defined. Moreover x(0, α) = x(0, α) = x(2π,−α) = x(2π, α) and in the case of Klein's bottle (cid:26) x(θ,−π) = x(2θ,−π) = x(2θ, π) = x(θ, π) x(θ,−π) = x(2(θ − π), π) = x(2(θ − π),−π) = x(θ, π) if if θ ∈ [0, π] θ ∈ [π, 2π] 117 i.e. x ∈ C(B, IC). b) For X, Y ∈ S(f ) and (θ, α) ∈ [0, 2π] × [−π, π], by Theorem 2.1.9 c),g), (ϕX∗)(θ, α) = ^(X∗)0(θ, α) + eiθ^(X∗)1(θ, α) = = ^(X0)∗(θ, α) + eiθ (e−iθ(X1)∗)(θ, α) = ^z } { X0(2θ, α) + eiθ(e−2iθX1(2θ, α)) X0(2(θ − π),−α) + eiθ(e−2i(θ−π)X1(2(θ − π),−α)) =(cid:26) =( (ϕX)(ϕY ) = (fX0 + eiθfX1)(fY0 + eiθfY1) = fX0fY0 + eiθfX0fY1 + eiθfX1fY0 + e2iθfX1fY1 , X0(2(θ − π),−α) + eiθX1(2(θ − π),−α) ϕ(XY ) = ^(XY )0 + eiθ ^(XY )1 = θ ∈ [0, π] θ ∈ [π, 2π] θ ∈ [0, π] θ ∈ [π, 2π] X0(2θ, α) + eiθX1(2θ, α) = ϕX(θ, α) , if if if if = i.e. ϕ is a C-homomorphism. If ϕX = 0 then for α ∈ [−π, π], = fX0fY0 + e2iθfX1fY1 + eiθ(fX0fY1 +fX1fY0) = (ϕX)(ϕY ) , X0(2(θ − π),−α) + eiθX1(2(θ − π),−α) = 0 X0(2θ, α) + eiθX1(2θ, α) = 0 (cid:26) θ ∈ [0, π] θ ∈ [π, 2π] if if so for θ ∈ [0, π], replacing θ by θ + π and α by −α in the second relation, X0(2θ, α) − eiθX1(2θ, α) = 0 . X0(2θ, α) = X1(2θ, α) = 0 , X0 = X1 = 0 , X = 0 . It follows successively Thus ϕ is injective. Let y ∈ C(B, IC). Put ( X0 : [0, 2π] × [−π, π] −→ IC , X1 : [0, 2π] × [−π, π] −→ IC , (θ, α) 7−→ 1 2 e−i θ (θ, α) 7−→ 1 2 (y( θ 2 , α) + y( θ 2 + π,−α)) 2 (y( θ 2 , α) − y( θ 2 + π,−α)) . 118 For α ∈ [−π, π], (cid:26) X0(0, α) = 1 X0(2π,−α) = 1 (cid:26) X1(0, α) = 1 X1(2π,−α) = − 1 2 (y(0, α) + y(π,−α)) 2 (y(π,−α) + y(2π, α)) 2 (y(0, α) − y(π,−α)) 2 (y(π,−α) − y(2π, α)) so X0, X1 ∈ E. Moreover for (θ, α) ∈ [0, 2π] × [−π, π], fX0(θ, α) + eiθfX1(θ, α) = X0(2(θ − π),−α) + eiθX1(2(θ − π),−α) X0(2θ, α) + eiθX1(2θ, α) if if 2 (y(θ, α) + y(θ + π,−α) + y(θ, α) − y(θ + π,−α)) 2 (y(θ − π,−α) + y(θ, α) − y(θ − π,−α) + y(θ, α)) =(cid:26) =(cid:26) 1 1 = θ ∈ [0, π] θ ∈ [π, 2π] if if θ ∈ [0, π] θ ∈ [π, 2π] = i.e. ϕ is surjective. = y(θ, α) c) If A is Moebius's band then the assertion is obvious so assume A is Klein's bottle. The winding numbers of (cid:26) [0, 2π] −→ IC, α 7−→ x(0, α) [0, 2π] −→ IC, α 7−→ x(2π, α) are equal by homotopy, but their sum is equal to 0. Thus these winding numbers are equal to 0. The paths θ and α on A generate the homotopy group of A. Thus the winding number of x on any path of A is 0 and the assertion follows. d) The winding number of [0, 2π] −→ IC, θ 7−→ e−inθx(θ, 0) is 0 and the assertion follows from c). e) The assertion follows from d) and Proposition 3.1.1 b). f) The assertion follows from b), d), Proposition 2.2.2 a1 ⇒ a2, and Propo- sition 3.1.1 c). 119 3.2 T := ZZ2 × ZZ2 PROPOSITION 3.2.1 Let E be a unital C-algebra and let a, b, c be the three elements of ( ZZ2 × ZZ2 ) \ {(0, 0)}. Put A :=(cid:8) (α, β, γ, ε) ∈ (U n Ec)4(cid:12)(cid:12) ε2 = 1E(cid:9) and for every ∈ A and σ ∈ (U n Ec)3 denote by f and gσ the functions defined by the following tables: f a b c a βγ εγ εβ b γ c β εαγ α εα αβ gσ a b c b c a αβγ∗ αγβ∗ α2 αβγ∗ βγα∗ β2 αγβ∗ βγα∗ γ2 a) f ∈ F( ZZ2 × ZZ2 , E) for every ∈ A and the map A −→ F( ZZ2 × ZZ2 , E), 7−→ f is bijective. b) gσ ∈ { δλ λ ∈ Λ( ZZ2 × ZZ2 , E)} for every σ ∈ (U n Ec)3 and the map (U n Ec)3 −→ { δλ λ ∈ Λ( ZZ2 × ZZ2 , E)} , σ 7−→ gσ is bijective. c) The following are equivalent for all := (α, β, γ, ǫ) ∈ A and ′ := (α′, β′, γ′, ǫ′) ∈ A: c1) S(f) ≈S S(f′). c2) ε = ε′ and there are x, y, z ∈ U n Ec with y2 = αα′∗γγ′∗ , x2 = ββ′∗γγ′∗ , z2 = αα′∗ββ′∗ . c3) ε = ε′ and there are x, y ∈ U n Ec with x2 = ββ′∗γγ′∗ , y2 = αα′∗γγ′∗ . d) The following are equivalent for all := (α, β, γ, ε ∈ A) and X ∈ S(f): 120 d1) X ∈n V f t d2) t ∈ ZZ2 × ZZ2 =⇒ εXt = Xt. (cid:12)(cid:12)(cid:12) t ∈ ZZ2 × ZZ2oc . e) The following are equivalent for all := (α, β, γ, ε ∈ A) and X ∈ S(f): e1) X ∈ S(f)c. e2) t ∈ ZZ2 × ZZ2 =⇒ εXt = Xt ∈ Ec. f ) For := (α, β, γ, ε) ∈ A and X, Y ∈ S(f), (X∗)0 = X∗0 , (X∗)a = β∗γ∗X∗a , (X∗)b = εα∗γ∗X∗b , (X∗)c = α∗β∗X∗c , (XY )0 = X0Y0 + βγXaYa + εαγXbYb + αβXcYc , (XY )a = X0Ya + XaY0 + αXbYc + εαXcYb , (XY )b = X0Yb + βXaYc + XbY0 + εβXcYa , (XY )c = X0Yc + γXaYb + εγXbYa + XcY0 . g) Assume IK = IC, let σ(Ec) be the spectrum of Ec, and for every δ ∈ Ec let δ be its Gelfand transform. Then σ(Va) =n eiθ(cid:12)(cid:12)(cid:12) θ ∈ IR, e2iθ ∈ cβγ(σ(Ec))o , σ(Vb) =n eiθ(cid:12)(cid:12)(cid:12) θ ∈ IR, e2iθ ∈ cαγ(σ(Ec))o , σ(Vc) =n eiθ(cid:12)(cid:12)(cid:12) θ ∈ IR, e2iθ ∈ cαβ(σ(Ec))o . a) is a long calculation. b) is easy to verify. c1 ⇒ c2 By Proposition 2.2.2 a2 ⇒ a1 there is a λ ∈ Λ( ZZ2 × ZZ2 , E) with f = f′δλ. By b), there is a σ := (x, y, z) ∈ (U n Ec)3 with f = f′gσ. We get ε = ε′ and αα′∗ = x∗yz , ββ′∗ = xy∗z , γγ′∗ = xyz∗ . It follows xyz = αα′∗ββ′∗γγ′∗ so x2 = ββ′∗γγ′∗ , y2 = αα′∗γγ′∗ , z2 = αα′∗ββ′∗ . 121 c2 ⇒ c3 is trivial. c3 ⇒ c2 If we put z := xyγ∗γ′ then z2 = ββ′∗γγ′∗αα′∗γγ′∗γ∗2γ′2 = αα′∗ββ′∗ . c2 ⇒ c1 follows from b) and Proposition 2.2.2 a1 ⇒ a2. d) follows from Corollary 2.1.24 b). e) follows from Corollary 2.1.24 c). f) follows from Theorem 2.1.9 c),g). g) follows from f). COROLLARY 3.2.2 We use the notation of Proposition 3.2.1 and take := (α, β, γ, ε) ∈ A. a) Assume ε = 1E and there are x, y ∈ U n E with x2 = βγ , y2 = αγ. Put z := xyγ∗. a1) x, y, z ∈ U n Ec , z2 = αβ. a2) For every λ, µ ∈ {−1, 1} the map ϕλ,µ : S(f) −→ E , X 7−→ X0 + λxXa + µyXb + λµzXc is an E-C-homomorphism. a3) The map S(f) −→ E4, X 7−→ (ϕ1,1X, ϕ1,−1X, ϕ−1,1X.ϕ−1,−1X) is an E-C-isomorphism. b) Assume IK := IR , ε = 1E, and there are x, y ∈ U n E with (resp. y2 = −αγ) . x2 = −βγ , y2 = αγ , 122 Put z := xyγ∗. Then x, y, z ∈ U n Ec, z2 = −αβ (resp. z2 = αβ), and the maps S(f) −→ ( ◦E)2, X 7−→ (X0+ixXa+yXb+izXc, X0+ixXa−yXb−izXc) S(f) −→ ( ◦E)2, X 7−→ (X0+ixXa+iyXb−zXc, X0+ixXa−iyXb+zXc) are respectively E-C-isomorphisms (where ◦E denotes the complexifica- tion of E). c) Assume IK := IR, ε = −1E, and there are x, y ∈ Ec with x2 = −βγ , y2 = αγ. Put z := xyγ∗. Then x, y, z ∈ U n Ec, z2 = −αβ, and the map S(f) −→ IH ⊗ E, X 7−→ X0 + ixXa + jyXb + kzXc , where i, j, k are the canonical units of IH, is an E-C-isomorphism. d) If ε = −1E and there is an x ∈ U n Ec with x2 = αβ then for every δ ∈ U n Ec the map S(f) −→ E2,2, X 7−→(cid:20) X0 + xXc δ(Xa + xβ∗Xb) γδ∗(βXa − xXb) X0 − xXc (cid:21) is an E-C-isomorphism. The proof is a long calculation using Proposition 3.2.1 f). Remarks. d) is contained in Proposition 3.2.3 c). An example with ε = 1E but different from a) is presented in Proposition 3.3.2. PROPOSITION 3.2.3 We use the notation of Proposition 3.2.1 and take := (α, β, γ, ε) ∈ A. a) Let ϕ : S(f) → E2,2 be an E-C-isomorphism and put (cid:20) At Bt Ct Dt (cid:21) := ϕVt for every t ∈ ZZ2 × ZZ2 \ {(0, 0)}. Then ε = −1E, At, Bt, Ct, Dt ∈ Ec and At + Dt = 0 for every t ∈ ZZ2 × ZZ2 \ {(0, 0)}, and A∗a = β∗γ∗Aa , A∗b = −α∗γ∗Ab , A∗c = α∗β∗Ac , 123 B∗a = β∗γ∗Ca , a + BaCa = βγ , B∗b = −α∗γ∗Cb , A2 A2 A2 b + BbCb = −αγ , c + BcCc = αβ , b = −αγ(1E − Bb2) , A2 c = αβ(1E − Bc2) , A2 a = βγ(1E − Ba2) , A2 B∗c = α∗β∗Cc , 2AaAb + BaCb + BbCa = 0 , 2AbAc + BbCc + BcCb = 0 , 2AcAa + BcCa + BaCc = 0 , αAa = AbAc + BbCc , βAb = AaAc +BaCc , γAc = AaAb + BaCb , αBa = AbBc− AcBb , βBb = AaBc−AcBa , γBc = AaBb− AbBa , αCa = AcCb− AbCc , βCb = AcCa−AaCc , γCc = AbCa− AaCb , Aa + Ab + Ac 6= 0 , Ba + Bb + Bc 6= 3.1E . b) Let (At)t∈T , (Bt)t∈T , (Ct)t∈T , (Dt)t∈T be families in Ec satisfying the above conditions and put X′ := AaXa + AbXb + AcXc , X′′ := BaXa + BbXb + BcXc , X′′′ := CaXa + CbXb + CcXc for every X ∈ S(f). If ε = −1E then the map S(f) −→ E2,2, X 7−→(cid:20) X0 + X′ X′′′ X′′ X0 − X′ (cid:21) is an E-C-isomorphism. c) Let ε = −1E and assume there is an x ∈ Ec with x2 = βγ. Let y ∈ U n Ec (cid:21) and put z := γ∗xy. Then x, y, z ∈ U n Ec and the map ϕ : S(f) −→ E2,2 , X 7−→(cid:20) α(yXb + zXc) X0 − xXa X0 + xXa is an E-C-isomorphism such that −γy∗Xb + βz∗Xc (V0 + (x∗ ⊗ 1K )Va)) =(cid:20) 1 0 0 0 (cid:21) . ϕ( 1 2 In particular (by the symmetry of a,b,c), if ε = −1E and if there is an x ∈ Ec with x2 = βγ, or x2 = −αγ, or x2 = αβ then S(f) ≈E E2,2. 124 Remark. Take := (1E , 1E, 1E ,−1E), ′ := (1E , 1E, γ′,−1E). By c), S(f) ≈E S(f′) and by Proposition 3.2.1 c1 ⇒ c2, S(f) ≈S S(f′) implies the existence of an x ∈ U n Ec with x2 = γ′. COROLLARY 3.2.4 We use the notation of Proposition 3.2.3 and take E := IK, α = 1, and β = γ = ε = −1. Let S be a group, F a unital C-algebra, g ∈ F(S, F ), and h : ((S × ( ZZ2 )2) × (S × ( ZZ2 )2)) −→ U n F c , f(t1, t2)g(s1, s2) . ((s1, t1), (s2, t2)) 7−→ a) h ∈ F(S × ( ZZ2 )2, F ). b) S(h) ≈ S(g)2,2 , Sk·k(h) ≈ Sk·k(g)2,2. By Proposition 3.2.3 c), S(f ) ≈ IK2,2, so by Proposition 2.2.11 c),e), S(h) ≈ IK2,2 ⊗ S(g) ≈ S(g)2,2 , Sk·k(h) ≈ IK2,2 ⊗ Sk·k(g) ≈ Sk·k(g)2,2 . EXAMPLE 3.2.5 Let IK := IC and E := C( IT, IC). a) With the notation of Proposition 3.2.1, if := (α, β, γ,−1) ∈ A then S(f) ≈E E2,2. b) Card ({ S(f ) f ∈ F( ZZ2 × ZZ2 , E)} / ≈S) = 16. Put m := w(α) , n := w(β) , p := w(γ) , where w denotes the winding number. By Proposition 2.2.2 a1 ⇒ a2, we may assume α = zm, β = zn, γ = zp. a) If n + p is even then the assertion follows from Proposition 3.2.3 c). If n + p is odd then either m + p or m + n is even and the assertion follows again from Proposition 3.2.3 c). 125 b) follows from Proposition 2.2.2 a),c). Remark. Assume IK := IR and let E be the real C-algebra C( IT, IC) ([C1] Theorem 4.1.1.8 a)), ε = −1E, α : IT −→ IC , β : IT −→ IC , γ : IT −→ IC , z 7−→ z , z 7−→ −z , z 7−→ ¯z , and := (α, β, γ, ε). Then by Corollary 3.2.2 c), S(f) ≈ IH ⊗ E. EXAMPLE 3.2.6 We put E := C( IT2, IC), γ := 1E, α : IT2 −→ IC , (z1, z2) 7−→ z1 , β : IT2 −→ IC , (z1, z2) 7−→ z2 , and (with the notation of Proposition 3.2.1) := (α, β, γ,−1E ) ∈ A. a) S(f) is not commutative and not E-C-isomorphic to E2,2. b) If we put x : IT2 −→ IC , for every x ∈ E then the map (z1, z2) 7−→ x(z2 1, z2 2) S(f) −→ E2,2, X 7−→(cid:20) X0 + αβ Xc β Xa − α Xb X0 − αβ Xc (cid:21) β Xa + α Xb is a C-isomorphism. c) E2,2 ≈ S(f) 6≈E E2,2. a) By Proposition 3.2.1 d), S(f) is not commutative. Assume S(f) ≈E E2,2 and let us use the notation of Proposition 3.2.3 a). Step 1 {Aa 6= 0} ⊂ {Ab = 0} 126 Assume {Aa 6= 0} ∩ {Ab 6= 0} 6= ∅. By Proposition 3.2.3 a), 2AaAb + BaCb + BbCa = 0 , B∗a = β∗Ca , B∗b = −α∗Cb so Ba 6= 0 and Bb 6= 0 on this set. We put Aa =: Aaei Aa , Ab =: Abei Ab, Ba =: Baei Ba, Bb =: Bbei Bb , z1 =: eiθ1, z2 =: eiθ2 , with Aa, Ab, Ba, Bb ∈ IR. By Proposition 3.2.3 a), 2 Aa = θ2, 2 Ab = θ1 + π, BaCb + BbCa = −αγBaB∗b + βγBbB∗a = BaBb(ei(θ2+ Bb− Ba) − ei(θ1+ Ba− Bb)) = = BaBbei θ1 +θ2 2 (ei( θ2 −θ1 = 2BaBb sin( 2 + Ba− Bb)) = 2 + Bb− Ba) − ei( θ1 −θ2 θ2 − θ1 + Bb − Ba)ei θ1+θ2+π 2 2 . Since 2AaAb = −(BaCb + BbCa) there is a k ∈ ZZ with θ2 2 + θ1 + π 2 = θ1 + θ2 + π 2 + (2k + 1)π which is a contradiction. Step 2 {Aa 6= 0} ⊂ {Ac = 0} The assertion follows from Step 1 by symmetry. Step 3 {Aa 6= 0} = {Ab = Ac = 0} The assertion follows from Steps 1 and 2 and from Aa + Ab + Ac 6= 0. Step 4 The contradiction 127 By Step 3 and by the symmetry, the sets {Aa 6= 0}, {Ab 6= 0}, and {Ac 6= 0} are clopen and by Aa + Ab + Ac 6= 0 their union is equal to IT2. So there is exactly one of these sets equal to IT2 which implies A2 a = z2 , or A2 b = −z1 or A2 c = z1z2 and no one of these identities can hold. b) is a direct verification. c) follows from a) and b). 3.3 T := ( ZZ2 )n with n ∈ IN EXAMPLE 3.3.1 Assume f constant and put h s ti := nYi=1 (−1)s(i)t(i) for all s, t ∈ T (where ZZ2 is identified with {0, 1}) and ϕt : S(f ) −→ E , X 7−→Xs∈T h t si Xs for all t ∈ T . Then the map ϕ : S(f ) −→ E2n , X 7−→ (ϕtX)t∈T is an E-C-isomorphism. For r, s, t ∈ T , t + t = 0 , h s ti = h t si , h r + s ti = h r tih s ti , h r s + ti = h r sih r ti . For t ∈ T and X, Y ∈ S(f ), by Theorem 2.1.9 c),g), ϕt(X∗) =Xs∈T h t si (X∗)s =Xs∈T h t si (Xs)∗ = (ϕtX)∗ , 128 (ϕtX)(ϕtY ) = Xr,s∈T = Xq,r∈T h t r ih t si XrYs = Xq,r∈T h t q i XrYq−r =Xq∈T so ϕt and ϕ are E-C-homomorphisms. h t r ih t q − r i XrYq−r = h t q i (XY )q = ϕt(XY ) We have We want to prove Xt∈T h 0 ti = 2n . Xt∈T h s ti = 0 for all s ∈ T , s 6= 0, by induction with respect to Card{ i ∈ INn s(i) 6= 0}. Let i ∈ INn with s(i) 6= 0 and put r := s + ei, T0 := { t ∈ T t(i) = 0} , T1 := { t ∈ T t(i) = 1} . Then Xt∈T0 But if r = 0. By the hypothesis of the induction h r ti . h s ti = −Xt∈T1 h r ti = 2n−1 h r ti , Xt∈T1 h s ti =Xt∈T0 h r ti = Xt∈T1 Xt∈T0 h r ti =Xt∈T1 Xt∈T0 h r ti = 0 if r 6= 0 (with INn replaced by INn \ {i}, since r(i) = 0). This finishes the proof by induction. For r ∈ T and X ∈ S(f ), by the above, Xt∈T h r ti ϕtX = Xs,t∈T = Xs∈T\{r}Xt∈T h r tih t si Xs = Xs,t∈T h r + s ti Xs +Xt∈T h 0 ti Xr = 2nXr . h r + s ti Xs = Hence ϕ is bijective. 129 EXAMPLE 3.3.2 Let E := C( ITn, IC), denote by z := (z1, z2,··· , zn) the points n) for every z ∈ ITn. We identify ( ZZ2 )n with of ITn, and put z2 := (z2 1, z2 P(INn) by using the bijection 2,··· , z2 P(INn) −→ ( ZZ2 )n, I 7−→ eI and denote by I△J := (I \ J) ∪ (J \ I) the addition on P(INn) corresponding to this identification. We put λI := Qi∈I zi for every I ⊂ INn and f : P(INn) × P(INn) −→ U n Ec , (I, J) 7−→ λI∩J . Then f ∈ F(( ZZ2 )n, E) and, if we put X := XI⊂INn λI (z)XI (z2) ∈ E for every X ∈ S(f ), the map ϕ : S(f ) −→ E , X 7−→ X is an isomorphism of C-algebras. Let X, Y ∈ S(f ). By Theorem 2.1.9 c),g), fX∗ = XI⊂INn λI (XY )I (z2) = XI⊂INn λI (X∗)I (z2) = XI⊂INn λI XJ⊂INn J∩KXJ YK = XJ,K⊂INn λJ△Kλ2 gXY = XI⊂INn = XJ,K⊂INn λI λ2 I X∗I = X , f (J, J△I)2XJ YJ△I = λJ λKXJ YK = X Y so ϕ is a C-homomorphism. We put for k ∈ INn, i ∈ ZZ n, and I ⊂ INn, k :=(cid:26) 2ik + 1 if iI if k ∈ INn \ I k ∈ I 2ik , 130 iI := (iI 1, iI 2,··· , iI n) ∈ ZZ n and Let G :=( Xi∈ ZZ n n (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (ai)i∈ ZZ n ∈ IC( ZZ n)) . 2 ··· zin n ∈ G aizi1 1 zi2 2 ··· zin x := Xi∈ ZZ n aizi1 1 zi2 and for every I ⊂ INn put aiI zi1 XI := Xi∈ ZZ n 1 zi2 2 ··· zin n , X := XI⊂INn (XI ⊗ 1K )VI . Then ϕX = x and so G ⊂ ϕ(S(f )). Since G is dense in E, it follows that ϕ is surjective. We prove that ϕ is injective by induction with respect to n ∈ IN. The case n = 1 was proved in Example 3.1.4. Assume the assertion holds for n − 1. Let X ∈ Ker ϕ. Then λI (z)XI (z2) = 0 . XI⊂INn By replacing zn by −zn in the above relation, we get and so XI⊂INn−1 XI⊂INn−1 λI (z)XI (z2) − Xn∈I⊂INn λI (z)XI (z2) = Xn∈I⊂INn λI (z)XI (z2) = 0 λI (z)XI (z2) = 0 . By the induction hypothesis, we get XI = 0 for all I ⊂ INn and so X = 0. Thus ϕ is injective and a C-isomorphism. EXAMPLE 3.3.3 Let f ∈ F(( ZZ2 )3, E), put a := (0, 0, 1) , b := (0, 1, 0) , c := (0, 1, 1) , s := (1, 0, 0) , and denote by g the element of F( ZZ2 , E) defined by g(1, 1) := f (s, s) Proposi- tion 3.1.1 a). 131 a) There is a family (αi , βi , γi , εi)i∈IN7 in (U n Ec)4 such that f is given by the attached table and such that ε2 ε3 = ε1ε2 , ε5 = ε1ε4 , i = 1E for every i ∈ IN7 and ε7 = ε1ε2ε4 , ε6 = ε2ε4 , α3 = ε2ε4α1α∗2α4α6γ∗2 , α5 = α6β1γ∗2 , α7 = α4γ1γ∗2 , β2 = β1γ1γ∗2 , β5 = ε4α1α∗2α6 , γ3 = ε2α4α∗6γ1 , β3 = ε2α∗4α6β1 , β6 = ε4α1α∗2α6β1γ∗2 , γ4 = ε2ε4α2α∗4γ∗1 γ2 , β4 = ε1ε2ε4α1α∗2α4γ1γ∗2 , β7 = ε1ε2ε4α1α∗2α6 , γ5 = ε1ε4α2α∗6γ1 , γ6 = ε4α2α∗6γ2 , γ7 = ε1ε2ε4α2α∗4β1 . f a b c s a + s b + s c + s a β1γ1 ε1γ1 ε1β1 ε2γ2 ε2β2 ε3γ3 ε3β3 b γ1 ε1α1γ1 ε1α1 ε4γ4 ε5γ5 ε4γ4 ε5β5 c β1 α1 α1β1 ε6γ6 ε7γ7 ε7γ7 ε6β6 s γ2 γ4 γ6 ε2α2γ2 ε2α2 ε4α4 ε6β6 a + s b + s c + s β2 γ5 γ7 α2 α2β2 ε7α7 ε5α5 γ3 β4 β7 α4 α7 ε3α3γ3 ε3α3 β3 β5 β6 α6 α5 α3 α3β3 b) If ε1 = −1E, ε2 = ε4, γ1 = 1E, and there is an x ∈ Ec with x2 = α1β∗1 1 and (Theorem then there are P± ∈ (Ee⊗1K )c∩P r S(f ) with P++P− = V f P+S(f )P+ ≈E S(g) ≈E P−S(f )P− . 2.2.18 b)) c) If ε1 = −1E, ε2 = ε4 = γ1 = 1E, and there is an x ∈ Ec with x2 = α1β∗1 then S(f ) ≈E S(g)2,2. d) Assume ε1 = −1E, ε2 = ε4 = α1 = β1 = γ1 = 1E, γ2 = α∗2, and 2 = α4 α4 4 = α6 = 1E and put ϕ± : S(f ) −→ E2,2 , X 7−→ (cid:20) X0 + Xc ± Xs ± Xc+s Xa + Xb ± α∗2Xa+s ± α∗4Xb+s Xa − Xb ± α∗2Xa+s ∓ α∗4Xb+s X0 − Xc ± Xs ∓ Xc+s (cid:21) . Then the map S(f ) −→ E2,2 × E2,2, X 7−→ (ϕ+X, ϕ−X) is an E-C-isomorphism. 132 a) is a long calculation. b) and c) follow from a) and Theorem 2.2.18 e). d) is a long calculation using a). 3.4 T := ZZn with n ∈ IN PROPOSITION 3.4.1 Put A := U n Ec and for every α ∈ An−1 put α∗k! , fα : ZZn × ZZn −→ A , αj q−1Yk=1 (p, q) 7−→ p+q−1Yj=p where ZZn and INn are canonically identified and αn := 1E. a) For every f ∈ F( ZZn , E) and X ∈ S(f ), X ∈ S(f )c iff Xt ∈ Ec for all t ∈ T . In particular, S(f ) is commutative if E is commutative. b) fα ∈ F( ZZn , E) for every α ∈ An−1 and the map An−1 −→ F( ZZn , E), α 7−→ fα is a group isomorphism. c) The following are equivalent for all α, β ∈ An−1. c1) S(fα) ≈S S(fβ). c2) There is a γ ∈ A such that γn = (αjβ∗j ) . n−1Yj=1 c3) There is a λ ∈ Λ( ZZn , E) such that fα = fβδλ. If these equivalent conditions are fulfilled then the map S(fα) −→ S(fβ), X 7−→ U∗λXUλ 133 is an S-isomorphism and λ(1)n = (αjβ∗j ) = γn , n−1Yj=1 p ∈ ZZn =⇒ λ(p) = λ(1)p (α∗j βj) . p−1Yj=1 j 7−→ (cid:18)n−1Qk=1 1 α∗k(cid:19)n−1 if if j < n − 1 j = n − 1 . d) Let α ∈ An−1 and put β : INn−1 −→ A , Then α and β fulfill the equivalent conditions of c). e) There is a natural bijection { S(f ) f ∈ F( ZZn , E)} / ≈S −→ A/{ xn x ∈ A} . If E := C( ITm, IC) for some m ∈ IN then Card ({ S(f ) f ∈ F( ZZn , E)} / ≈S) = mn . f ) Let α ∈ An−1, β ∈ A such that βn = F :=( E if ◦E if αj, n−1Qj=1 IK = IC IK = IR , where ◦E denotes the complexification of E, and wk : S(fα) −→ F , X 7−→ βj j−1Yl=1 nXj=1 ¯αl! e 2πijk n Xj for every k ∈ INn(= ZZn ). f1) If IK = IC then the map S(fα) −→ En, X 7−→ (wkX)k∈ ZZn is an E-C-isomorphism. 134 f2) If IK = IR and n is odd then we may take β ∈ IR and the map ) 2 , X 7−→ (wnX, (wkX)k∈IN n−1 S(fα) −→ E × ( ◦E) n−1 2 is an E-C-isomorphism. f3) If IK = IR, n is even, and αj = −1 then the map n−1Qj=1 S(fα) −→ ( ◦E) n is an E-C-isomorphism. 2 , X 7−→ (wk−1X)k∈IN n 2 n−1Qj=1 2 −1, X 7−→ (wnX, w n 2 f4) If IK = IR, n is even, and αj = 1, and β = 1 then the map S(fα) −→ E × E × ( ◦E) n X, (wkX)k∈IN n 2 −1) is an E-C-isomorphism. f5) If n is even then there is a γ ∈ A such that fα( n 2 , n 2 ) = γ2. EXAMPLE 3.4.2 Let E := C( IT, IC), r ∈ ZZ n−1, z : IT → IC the canonical inclusion, and f : ZZn × ZZn −→ U n Ec , where ZZn and INn are canonically identified. Then f ∈ F( ZZn , E). Let further S rj), where ρ : ZZ → ZZn is the quotient be the subgroup of ZZn generated by ρ( n m m := Card S , ω := e 2πi n , map, h := , rj− rj! , q−1 Pj=1 (p, q) 7−→ z p+q−1 Pj=p n−1Pj=1 σ : INn −→ ZZ , p 7−→ and ϕk : S(f ) −→ E , X 7−→ 135 p h rj , p−1Xj=1 rj − m n−1Xj=1 nXp=1 (Xp ◦ zm)zσ(p)ωpk for every k ∈ INh. Then the map ϕ : S(f ) −→ Eh , X 7−→ (ϕkX)k∈INh is an E-C-isomorphism. The next example shows that the set { S(f ) f ∈ F( ZZn ,C( IT, IC))} is not re- duced by restricting the Schur functions to have the form indicated in Example 3.4.2. EXAMPLE 3.4.3 Let E := C( IT, IC) and g ∈ F( ZZn , E). Put ϕ : [0, 2π[−→ IR , θ 7−→ log (g(j, 1))(eiθ ) , n−1Yj=1 where we take a fixed (but arbitrary) branch of log. If we define r : INn−1 −→ ZZ , j 7−→( lim θ→2π ϕ(θ) − ϕ(0) 0 if if j = 1 j 6= 1 then there is a λ ∈ Λ( ZZn , E) such that g = f δλ, where f is the Schur function defined in Example 3.4.2. In particular S(f ) ≈S S(g). 3.5 T := ZZ EXAMPLE 3.5.1 Let f ∈ F( ZZ , E). a) Sk·k(f ) ≈ C( IT, E). b) If E is a W-algebra then SW (f ) ≈ E ¯⊗L∞(µ) ≈ L∞(µ, E) , where µ denotes the Lebesgue measure on IT. 136 By Corollary 1.1.6 c) and Proposition 2.2.2 a1 ⇒ a2, we may assume f constant. By Proposition 2.2.10 c),e), we may assume E := IC. Let α : IT → IC be the inclusion map. Then l2( ZZ ) −→ L2(µ), ξ 7−→Xn∈ ZZ ξnαn is an isomorphism of Hilbert spaces. If we identify these Hilbert spaces using this isomorphism then V1 becomes the multiplicator operator so η 7−→ αη L2(µ) −→ L2(µ), R(f ) −→ L∞(µ), X 7−→Xn∈ ZZ Xnαn is an injective, involutive algebra homomorphism. The assertion follows. 4 Clifford Algebras 4.1 The general case Throughout this subsection I is a totally ordered set, (Ti)i∈I is a family of groups, and (fi)i∈I ∈ Qi∈I F(Ti, E). We put ¯t := { i ∈ I ti 6= 1i } for every t ∈ Qi∈I ¯t is finite) , T :=( t ∈Yi∈I Ti(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) T is a subgroup of Qi∈I Ti (where 1i denotes the neutral element of Ti) and T ′ :=(cid:8) t ∈ T t2 = 1(cid:9) . An associated f ∈ F(T, E) will be defined in Proposition 4.1.1 b). in a bijective way the "word" ti1ti2 ··· tin, where Ti. We canonically associate to every element t ∈ T {i1, i2,··· , in} = ¯t i1 < i2 < ··· < in and 137 and use sometimes this representation instead of t ( to 1 ∈ T we associate the "empty word"). PROPOSITION 4.1.1 a) Let ti1ti2 ··· tin be a finite sequence of letters with tij ∈ Tij \{1ij} for every j ∈ INn and use transpositions of successive letters with distinct indices in order to bring these indices in an increasing order. If τ denotes the number of used transpositions then (−1)τ does not depend on the manner in which this operation was done. b) Let s, t ∈ T and let si1si2 ··· sim , ti′ 1 2 ··· ti′ ti′ n be the canonically associated words of s and t, respectively. We put for every k ∈ I, sk := sij if there is a j ∈ INm with k = ij and sk := 1k if the above condition is not fulfilled and define t in a similar way. Moreover we put (Proposition 1.1.2 a)) f (s, t) := (−1)τYk∈I fk(sk, tk) , where τ denotes the number of transpositions of successive letters with distinct indices in the finite sequence of letters si1si2 ··· simti′ 1 2 ··· ti′ ti′ n in order to bring the indices in an increasing order. Then f ∈ F(T, E). c) Let I0 be a subset of I, T0 the subgroup { t ∈ T ¯t ⊂ I0 } of T , and f0 the element of F(T0, E) defined in a similar way as f was defined in b). Then f0 = f(T0 × T0) and the map Sk·k(f0) −→ Sk·k(f ), k·kXt∈T0 t (Xte⊗1K )V f0 7−→ k·kXt∈T0 t (Xte⊗1K )V f is an injective E-C-homomorphism with image { X ∈ S(f ) (t ∈ T & Xt 6= 0) ⇒ t ∈ T0 } . 138 a) We define a new total order relation on the indices of the given word by putting for all j, k ∈ INn ij ≺ ik :⇐⇒ ((ij < ik) or (ij = ik and j < k)) . Let P be a sequence of transpositions of successive letters in order to bring the indices in an increasing form with respect to the new order and let τ′ be the number of used transpositions. Then τ − τ′ is even and so (−1)τ = (−1)τ ′ . By the theory of permutations (−1)τ ′ does not depend on P , which proves the assertion. b) By a), f is well-defined. Let r, s, t ∈ T and let ti′′ si′ ri1ri2 ··· rim , 2 ··· si′ si′ n , 1 2 ··· ti′′ ti′′ p 1 be the words canonically associated to r, s, and t, respectively. There are α, β ∈ {−1, +1} such that f (r, s)f (rs, t) = αYi∈I f (r, st)f (s, t) = βYi∈I f (ri, si)f (grisi, ti) , fi(ri,gsiti)f (si, ti) . Write the finite sequence of letters ri1ri2 ··· rimsi′ 1 2 ··· si′ si′ nti′′ 1 2 ··· ti′′ ti′′ p and use transpositions of successive letters with distinct indices in order to bring the indices in an increasing order. We can do this acting first on the letters of r and s only and then in a second step also on the letters of t. Then α = (−1)µ, where µ denotes the number of all performed transpositions. For β we may start first with the letters of s and t and then in a second step also with the letters of r. Then β = (−1)ν , where ν is the number of all effectuated transpositions. By a), α = (−1)µ = (−1)ν = β. The rest of the proof is obvious. c) follows from Corollary 2.1.17 d). COROLLARY 4.1.2 If I := IN2 then for all s, t ∈ T , f (s, t) = f1(s1, t1) f2(s2, t2) −f1(s1, t1)f2(s2, t2) if if if s2 = 12 t1 = 11 . s2 6= 12 , t1 6= 11 139 PROPOSITION 4.1.3 Let s, t ∈ T . a) f (s, t) = (−1)Card (¯s×¯t)−Card (¯s∩¯t)f (t, s). b) st = ts iff VsVt = (−1)Card (¯s×¯t)−Card (¯s∩¯t)VtVs. c) Assume ¯s ⊂ ¯t. If Card ¯s is even or if Card ¯t is odd then f (s, t) = f (t, s). If in addition st = ts then VsVt = VtVs. d) If Card I is an odd natural number and T is commutative then Vt ∈ S(f )c for every t ∈ T with ¯t = I. e) Assume t ∈ T ′. If n := Card ¯t and α :=Qi∈¯t fi(ti, ti) then f (t, t) = (−1) n(n−1) 2 α , (Vt)2 = (−1) n(n−1) 2 (αe⊗1K)V1 , n(n−1) 2 α∗ , f (t) = (−1) V ∗t = (−1) n(n−1) 2 (α∗e⊗1K)Vt . a) For i ∈ ¯s, so f (si, t) =(cid:26) (−1)Card ¯tf (t, si) (−1)Card ¯t−1f (t, si) f (s, t) = (−1)Card (¯s×¯t)−Card (¯s∩¯t)f (t, s) . i 6∈ ¯t i ∈ ¯t if if b) By Proposition 2.1.2 b), VsVt = (f (s, t)e⊗1K )Vst , Thus if st = ts then by a), VtVs = (f (t, s)e⊗1K )Vts . Conversely, if this relation holds then by a), VsVt = ((f (s, t)f (t, s)∗)e⊗1K )VtVs = (−1)Card (¯s×¯t)−Card (¯s∩¯t)VtVs . Vst = (f (s, t)∗e⊗1K )VsVt = (−1)Card (¯s×¯t)−Card (¯s∩¯t)(f (t, s)∗e⊗1K )VsVt = and we get st = ts by Theorem 2.1.9 a). = (f (t, s)∗e⊗1K )VtVs = Vts 140 c) follows from a) and b). d) follows from c) (and Proposition 2.1.2 d)). e) We have f (t, t) = (−1)(n−1)+···+2+1α = (−1) n(n−1) 2 α . By Proposition 2.1.2 b),e), (Vt)2 = (f (t, t)e⊗1K )V1 = (−1) V ∗t = f (t)Vt−1 = f (t, t)∗Vt = (−1) n(n−1) 2 n(n−1) 2 (αe⊗1K)V1 , (α∗e⊗1K )Vt . PROPOSITION 4.1.4 Let S be a finite subset of T ′ \ {1} such that st = ts and Card (¯s × ¯t) − Card (¯s ∩ ¯t) is odd for all distinct s, t ∈ S and for every t ∈ S let αt, εt ∈ U n Ec and Xt ∈ E be such that ε2 t = 1E , (Vt)2 = (αte⊗1K )V1 , Xt∈S Xt2 = 1 4 1E . X∗t = αtXt , a) P := 1 2 V1 − P = V1 +Xt∈S V1 +Xt∈S 1 2 ((εtXt) ⊗1K )Vt ∈ P r S(f ) , ((−εtXt) ⊗1K )Vt ∈ P r S(f ) . b) If s ∈ S and β ∈ Ec such that Xs = 0 and β2 = αs then P is homotopic in P r S(f ) to 1 2 (V1 + ((β∗εs)e⊗1K)Vs) . a) By Proposition 4.1.3 b),e), P ∗ = 1 2 V1 +Xt∈S ((εtX∗t α∗t )e⊗1K )Vt = 1 2 V1 +X1∈S ((εtXt)e⊗1K )Vt = P , 141 P 2 = s6=t 1 4 V1 +Xt∈S + Xs,t∈S V1 +Xt∈S V1 +Xt∈S = 1 4 = 1 4 (X 2 te⊗1K )(Vt)2 +Xt∈S ((εtXt)e⊗1K )Vt+ ((X 2 ((εsεtXsXt)e⊗1K)(VsVt + VtVs) = t αt)e⊗1K )V1 +Xt∈S (Xt2e⊗1K )V1 +Xt∈S ((εtXt)e⊗1K )Vt = ((εtXt)e⊗1K)Vt = P . b) Remark first that β ∈ U n Ec and put 1 u 7−→ ( 4 Y : [0, 1] −→ Ec + , Q : [0, 1] −→ S(f ) , u 7−→ For u ∈ [0, 1], Z : [0, 1] −→ Ec , Xt2) 1 2 , 1E − u2Xt∈S u 7−→ β∗εsY (u) , V1 + (Z(u)e⊗1K )Vs + Xt∈S\{s} 1 2 ((uεtXt)e⊗1K)Vt . αsZ(u) = β2β∗εsY (u) = βεsY (u) = Z(u)∗ , Z(u)2 + Xt∈S\{s} uXt2 = 1 4 1E so by a), Q(u) ∈ P r S(f ). Moreover Q(0) = 1 2 (V1 + ((β∗εs)e⊗1K )Vs) , Q(1) = P . COROLLARY 4.1.5 Let s, t ∈ T ′ \ {1}, s 6= t, st = ts, αs, αt, εs, εt ∈ U n Ec such that ε2 s = ε2 t = 1E , and put Ps := 1 2 (Vs)2 = (α2 se⊗1K)V1 , (V1 + ((εsα∗s)e⊗1K )Vs) , Pt := (Vt)2 = (α2 te⊗1K )V1 , (V1 + ((εtα∗t )e⊗1K)Vt) . 1 2 142 a) Ps, Pt ∈ P r S(f ); we denote by Ps∧ Pt the infimum of Ps and Pt in S(f )+ (by b) and c) it exists). b) If VsVt 6= VtVs then Ps ∧ Pt = 0. c) If VsVt = VtVs then Ps ∧ Pt = PsPt ∈ P r S(f ). a) follows from Proposition 2.1.20 b ⇒ a. b) By Proposition 4.1.3 b), VsVt = −VtVs. Let X ∈ S(f )+ with X ≤ Ps and X ≤ Pt. By [C1] Proposition 4.2.7.1 d ⇒ c, 1 2 X + 1 2 X = PsX = X = ((εsα∗s)e⊗1K )VsX = ((εsεtα∗sα∗t )e⊗1K )VsVtX = ((εsα∗s)e⊗1K )VsX , = −((εsεtα∗sα∗t )e⊗1K )VtVsX = −X so X = 0 and Ps ∧ Pt = 0. c) We have PsPt = PtPs so PsPt ∈ P r S(f ) and PsPt = Ps ∧ Pt by [C1] Corollary 4.2.7.4 a ⇒ b&d. COROLLARY 4.1.6 Let m, n ∈ IN, INm+n ⊂ I, (αi)i∈INm ∈ (U n Ec)m, and for every i ∈ INm let ti ∈ T ′ with ¯ti := INn ∪ {n + i} and titj = tjti for all i, j ∈ INm. If for every i ∈ INm, (Vti )2 = (α2 i ⊗ 1K )V1 then 1 2 V1 + 1 √m Xi∈INm (α∗i ⊗ 1K)Vti! ∈ P r S(f ) . For distinct i, j ∈ INm, Card (¯ti × ¯tj) − Card (¯ti ∩ ¯tj) = (n + 1)2 − n = n(n + 1) + 1 is odd. For every i ∈ INm put Xi := 1 2√m α∗i . Then α2 i Xi = 1 2√m αi = X∗i , Xi2 = 1 4m 1E , Xi∈INm Xi2 = 1 4 1E and the assertion follows from Proposition 4.1.4 a). 143 THEOREM 4.1.7 Let n ∈ IN such that IN2n is an ordered subset of I, S := { t ∈ T ¯t ⊂ IN2n−2 }, g := f(S × S), a, b ∈ T such that a2 = b2 = 1, i ∈ IN2n−2 =⇒ ai = bi , ¯b = IN2n−2 ∪ {2n} , ¯a = IN2n−1 , ω : ZZ2 × ZZ2 → T the (injective) group homomorphism defined by ω(1, 0) := a, ω(0, 1) := b, α1 := f (a, a), α2 := f (b, b), β1, β2 ∈ U n Ec such that α1β2 1 + α2β2 2 = 0, γ := 1 2 (α∗1β∗1 β2 − α∗2β1β∗2 ) = α∗1β∗1β2 = −α∗2β1β∗2 , 1 2 X := P+ := X∗X , P− := XX∗ . We consider S(g) as an E-C-subalgebra of S(f ) (Corollary 2.1.17 e)). ((β1e⊗1K )Va + (β2e⊗1K )Vb) , a) ab = ba, γ2 = −α∗1α∗2. We put c := ab = ω(1, 1). b) X, Vc, P± ∈ S(g)c. c) We have 1 2 P+P− = 0, P± = X 2 = 0, XP+ = X, P−X = X, P+X = XP− = 0, X + X∗ ∈ U n S(f ) . (V1 ± (γe⊗1K )Vc) ∈ P r S(f ) , P+ + P− = V1 , d) The map E −→ P±S(f )P±, x 7−→ P±(xe⊗1K )P± is an injective unital C-homomorphism. We identify E with its image with respect to this map and consider P±S(f )P± as an E-C-algebra. e) The map ϕ± : S(g) −→ P±S(f )P± , Y 7−→ P±Y P± = P±Y = Y P± is an injective unital C-homomorphism. ϕ+Y1 + ϕ−Y2 ∈ U n S(f ). If Y1, Y2 ∈ U n S(g) then f ) The map ψ : S(f ) −→ S(f ) , Z 7−→ (X + X∗)Z(X + X∗) is an E-C-isomorphism such that ψ−1 = ψ , ψ(P+S(f )P+) = P−S(f )P− , ψ◦ϕ+ = ϕ− , ψ◦ϕ− = ϕ+ . 144 If Y1, Y2 ∈ S(g) then ϕ+Y1 + ϕ−Y2 = (ϕ+Y1 + ϕ−V1)ψ(ϕ+Y2 + ϕ−V1) . g) If p ∈ P r S(g) then (X(ϕ+p)∗(X(ϕ+p)) = ϕ+p , (X(ϕ+p))(X(ϕ+p))∗ = ϕ−p . h) Let R be the subgroup {1, a, b, c} of T , h := f(R × R), d ∈ T such that α′′ := f2n(2n, 2n) . ¯d = IN2n−2 and di = ai for every i ∈ IN2n−2, and α′ := f2n−1(2n − 1, 2n − 1) , α := f (d, d) , Then α1 = αα′, α2 = αα′′, −α′α′′ = (α∗γ∗)2, c h a a αα′ α′ b −α αα′′ −α′′ c −α′ α′′ −α′α′′ b α is the table of h, P± ∈ P r S(h), and the map ϕ : S(h) −→ E2,2 , Z 7−→(cid:20) Z0 + γ∗Zc Za + α′∗γ∗Zb is an E-C-isomorphism. In particular ϕP+ =(cid:20) 1E 0 0 (cid:21) , 0 ϕP− =(cid:20) 0 (cid:21) αα′Za − αγ∗Zb Z0 − γ∗Zc 0 1E (cid:21) 0 and E2,2 is E-C-isomorphic to an E-C-subalgebra of S(f ). i) Assume I = IN2n and T2n−1 = T2n = ZZ2 . Then T ≈ S × ZZ2 × ZZ2 , ϕ± is an E-C-isomorphism with inverse P±S(f )P± −→ S(f0), Z 7−→ 2Xu∈T0 (Zu ⊗ 1K )Vu , and S(f ) ≈E S(g)2,2 145 a) is easy to see. b) follows from Proposition 4.1.3 b). c) follows from a) and Theorem 2.2.18 b),h). d) follows from Theorem 2.2.18 c). e) By b) and c), the map is well-defined. The assertion follows now from Theorem 2.2.18 d),h). f) follows from b),c), and Theorem 2.2.18 h). g) follows from b) and Proposition 2.2.17 d). h) follows from c), d), Proposition 3.2.1 a), Corollary 3.2.2 d), and Propo- sition 3.2.3 c). i) follows from Theorem 2.2.18 f). PROPOSITION 4.1.8 We use the notation and the hypotheses of Theorem 4.1.7 and assume I := IN2, T1 := ZZ2 , and T2 := ZZ2m with m ∈ IN. a) a = (1, 0), b = (0, m), c = (1, m), α = 1E, α′ = α1 = f1(1, 1), α′′ = α2 = f2(m, m), and P±S(f )P± =(cid:8) (xe⊗1K)P±(cid:12)(cid:12) x ∈ E(cid:9) . b) If m = 1 then there are α, β, γ, δ ∈ U n Ec such that f is given by the following table: f (0, 1) (0, 2) (0, 3) (0, 1) (0, 2) (0, 3) (1, 0) (1, 1) (1, 2) (1, 3) α β γ 1E α β γ γ β α∗βγ α∗γ −1E α∗γ β∗γ −1E 1E 1E δ −δ γ β α∗βγ α∗γ −δ α∗γ β∗γ −δ (1, 2) (1, 3) (1, 0) (1, 1) −β −γ −1E −α −α∗βγ −α∗γ −β −α∗γ −β∗γ −γ δ δ δ −αδ −γδ −βδ −βδ −α∗βγδ −α∗γδ −γδ −α∗γδ −β∗γδ . 146 c) We assume IK := IC and m := 1 and put for all j, k ∈ {0, 1} ϕj,k : S(f ) −→ E , Z 7−→ Z0 + (−1)jZb + ijZ(k,1) − ijZ(k,3) , If we take α := β := γ := −δ := β1 := β2 := 1E in b) then the map φ : S(f ) −→ E4 , Z 7−→ (ϕ0,0Z, ϕ0,1Z, ϕ1,0Z, ϕ1,1Z) . S(f ) −→ E2,2 × E4, Z 7−→(cid:18)(cid:20) Z0 + Z(1,2) Z(1,0) − Zb Z(1,0) + Zb Z0 − Z(1,2) (cid:21) , φZ(cid:19) is an E-C*-isomorphism. a) Use Corollary 4.1.2 and Proposition 2.1.2 b). b) Use Proposition 3.4.1 a) and Proposition 4.1.1. c) follows from b) and Proposition 3.4.1 f1. 4.2 A special case Throughout this subsection we denote by S a totally ordered set, put T := ( ZZ2 )(S), and fix a map ρ : S → U n Ec. We define for every s ∈ S, fs ∈ F( ZZ2 , E) by putting fs(1, 1) = ρ(s) ( Proposition 3.1.1 a)). Moreover we denote by fρ the Schur function f defined in Proposition 4.1.1 b) (with I replaced by S) and put Cl(ρ) := S(fρ). Remark. If S := IN2 then T = ZZ2 × ZZ2 so Cl(ρ) is a special case of the example treated in subsection 3.2. With the notation used in the left table of Proposition 3.2.1 this case appears for a := (1, 0) and b := (0.1) exactly when ε = −1E, α = −ρ(b), β = ρ(a), and γ = 1E. LEMMA 4.2.1 Pf (S) endowed with the composition law Pf (S) × Pf (S) −→ Pf (S), (A, B) 7−→ A△B := (A \ B) ∪ (B \ A) is a locally finite commutative group ( Definition 2.1.18) with ∅ as neutral ele- ment and the map Pf (S) −→ T, A 7−→ eA 147 is a group isomorphism with inverse T −→ Pf (S), x 7−→ { s ∈ S x(s) = 1} . We identify T with Pf (S) by using this isomorphism and write s instead of {s} for every s ∈ S. For A, B ∈ T , fρ(A, B) = (−1)τ Ys∈A∩B ρ(s) , where τ is defined in Proposition 4.1.1 b). PROPOSITION 4.2.2 Assume S finite and let F be an E-C-algebra. Let further (xs)s∈S be a family in F such that for all distinct s, t ∈ S and for every y ∈ E, xsxt = −xtxs , x2 s = ρ(s)1F , x∗s = ρ(s)∗xs , xsy = yxs . Then there is a unique E-C-homomorphism ϕ : Cl(ρ) → F such that ϕVs = is E-linearly independent (resp. xs for all s ∈ S. If the family (cid:18)Qs∈A generates F as an E-C-algebra) then ϕ is injective (resp. surjective). xs(cid:19)A⊂S Put ϕVA := xs1xs2 ··· xsm for every A := {s1, s2,··· , sm}, where s1 < s2 < ··· < sm, and ϕ : Cl(ρ) −→ F , X 7−→ XA⊂S XA ϕVA . It is easy to see that (ϕVs)(ϕVt) = ϕ(VsVt) and y ϕVs = (ϕVs)y for all s, t ∈ S and y ∈ E (Proposition 2.1.2 b)). Let A := {s1, s2,··· , sm} ⊂ S , B := {t1, t2,··· , tn} ⊂ S , {r1, r2,··· , rp} := A△B , where the letters are written in strictly increasing order. Then (ϕVA)(ϕVB) = xs1xs2 ··· xsmxt1xt2 ··· xtn = fρ(A, B)xr1 xr2 ··· xrp = = fρ(A, B)ϕVA△B = ϕ((fρ(A, B)e⊗1K )VA△B) = ϕ(VAVB) , 148 m(m−1) (ϕVA)∗ = x∗sm ··· x∗s2x∗s1 = (−1) 2 Yi∈INm ρ(si)∗xs1xs2 ··· xsm = (−1) 2 = (−1) = ϕ((−1) m(m−1) 2 by Proposition 4.1.3 e). (( Yi∈INm m(m−1) m(m−1) x∗s1x∗s2 ··· x∗sm = 2 Yi∈INm ρ(si)∗)e⊗1K )VA) = ϕ(V ∗A) ρ(si)∗ϕVA = YBϕVB! = XA,B∈T For X, Y ∈ Cl(ρ) (by Theorem 2.1.9 c),g)), XAϕVA! XB∈T (ϕX)(ϕY ) = XA∈T = XA,B∈T XAYBϕ(VAVB) = XA,B∈T XAYA△Cfρ(A, A△C)ϕVC = XC∈T XA∈T = XA,C∈T XAYB(ϕVA)(ϕVB) = XAYBfρ(A, B)ϕVA△B = fρ(A, A△C)XAYA△C! ϕVC = (XY )CϕVC = ϕ(XY ) , = XC∈T X∗A(ϕVA)∗ = XA∈T (ϕX)∗ = XA∈T fρ(A)∗(X∗)A fρ(A)ϕVA = XA∈T = XA∈T X∗Aϕ(VA)∗ = (X∗)AϕVA = ϕ(X∗) (Proposition 4.1.3 e)) i.e. ϕ is an E-C-homomorphism. The uniqueness and the last assertions are obvious (by Theorem 2.1.9 a)). PROPOSITION 4.2.3 Let m, n ∈ IN ∪ {0}, S := IN2n, S′ := IN2n+m, K′ := l2(P(S′)), (αi)i∈INm ∈ (U n Ec)m, ρ′ : S′ −→ U n Ec , s 7−→(cid:26) ρ(s) α2 i fρ(S) if if s ∈ S s = 2n + i with i ∈ INm , and Ai := A ∪ {2n + i} for every A ⊂ S and i ∈ INm. a) i ∈ INm =⇒ fρ′(Si) = α∗2 i , (V ρ′ Si )2 = (α2 i ⊗ 1K ′)V ρ′ ∅ . 149 b) P := 1 2 V ρ′ ∅ + 1 2√m Pi∈INm (α∗i ⊗ 1K ′)V ρ′ Si ∈ P r Cl(ρ′). c) There is a unique injective E-C-homomorphism ϕ : Cl(ρ) → PCl(ρ′)P such that ϕV ρ s = P V ρ′ s P = P V ρ′ s = V ρ′ s P for every s ∈ S. d) If m ∈ IN2 then ϕ is an E-C-isomorphism. a) By Proposition 4.1.3 e), fρ′(Si) = (−1)n(2n+1) Ys∈Si ρ′(s)∗ = (−1)n(2n−1)Ys∈S ρ(s)∗! α∗2 i fρ(S)∗ = α∗2 i , (V ρ′ Si )2 = (α2 i ⊗ 1K ′)V ρ′ ∅ . b) follows from a) and Corollary 4.1.6. c) By Proposition 4.1.3 c), for s ∈ S, V ρ′ V ρ′ s P = P V ρ′ s V ρ′ Si = V ρ′ Si V ρ′ s for every i ∈ INm so s . By b), for distinct s, t ∈ S (Proposition 4.1.3 b)), (P V ρ′ t ) = P V ρ′ s )(P V ρ′ s V ρ′ t = −P V ρ′ s = −(P V ρ′ t V ρ′ s )2 = P (ρ′(s) ⊗ 1K ′)V ρ′ = (ρ(s) ⊗ 1K ′)P , ∅ s = (ρ(s) ⊗ 1K ′)∗P V ρ′ s )∗ = P (ρ′(s)∗ ⊗ 1K ′)V ρ′ t )(P V ρ′ s ) , s s )2 = P (V ρ′ (P V ρ′ s )∗ = P (V ρ′ (P V ρ′ . By Proposition 4.2.2 there is a unique E-C-homomorphism ϕ : Cl(ρ) → PCl(ρ′)P with the given properties. Let X ∈ Cl(ρ) with ϕX = 0. Then 0 = XA⊂S A! P = (XA ⊗ 1K ′)fρ′(A, Si)V ρ′ A△Si (XA ⊗ 1K ′)V ρ′ 2√m Xi∈INmXA⊂S 1 = 1 2 XA⊂S (XA ⊗ 1K ′)V ρ′ A + and this implies XA = 0 for all A ⊂ S (Theorem 2.1.9 a)). Thus ϕ is injective. d) 150 The case m = 1 Let Y ∈ PCl(ρ′)P . Then (by Proposition 2.1.2 b)) Y = Y P = 1 2 Y + (α∗1 ⊗ 1K ′)V ρ′ S1 Y , 1 2 XA⊂S′ S1△A+ Y = XA⊂S ((α∗1fρ′(S1, A)YA) ⊗ 1K ′)V ρ′ +XA⊂S (((α∗1fρ′(S1, A1)YA1)) ⊗ 1K ′)V ρ′ S△A (cid:26) YA = α∗1fρ′(S1, (S△A)1)Y(S△A)1 YA1 = α∗1fρ′(S1, S△A)YS△A X := 2XA⊂S ϕX +XA⊂S A +XA⊂S (YA ⊗ 1K ′)V ρ′ ((α∗1fρ′(S1, A)YA) ⊗ 1K ′)V ρ′ (YA ⊗ 1K)V ρ A ∈ Cl(ρ) 1 2 S1△A = ((α∗1fρ′(S1, S△A)YS△A) ⊗ 1K ′)V ρ′ A +XA⊂S (YA1 ⊗ 1K ′)V ρ′ = Y . A1 A1 = so for every A ⊂ S. If we put then ϕX = = XA⊂S (YA ⊗ 1K ′)V ρ′ = XA⊂S Thus ϕ is surjective. The case m = 2 Let Y ∈ PCl(ρ′)P . Then ( Y = P Y = 1 √2Y = (α∗1 ⊗ 1K ′)V ρ′ 2 Y + 1 2√2 2 Y + 1 2√2 Y + (α∗2 ⊗ 1K ′)V ρ′ Y = Y P = 1 S1 S2 ((α∗1 ⊗ 1K ′)V ρ′ Y ((α∗1 ⊗ 1K ′)V ρ′ + (α∗2 ⊗ 1K ′)V ρ′ + (α∗2 ⊗ 1K ′)V ρ′ S1 S2 S2 S1 )Y ) , Y = (α∗1 ⊗ 1K ′)Y V ρ′ S1 + (α∗2 ⊗ 1K ′)Y V ρ′ S2 . 151 For every B ⊂ S put Ba := B ∪{2n + 1} , Then Bb := B ∪{2n + 2} , Bc := B ∪{2n + 1 , 2n + 2} . (S△B)c V ρ′ S1 ((YBfρ′(S1, B))⊗1K ′)V ρ′ Y = XB⊂S +XB⊂S ((YBb fρ′(S1, Bb)) ⊗ 1K ′)V ρ′ = XB⊂S ((YBfρ′(S2, B)) ⊗ 1K ′)V ρ′ (S△B)b +XB⊂S ((YBbfρ′(S2, Bb)) ⊗ 1K ′)V ρ′ = XB⊂S ((YBfρ′(B, S1))⊗1K ′)V ρ′ ((YBb fρ′(Bb, S1)) ⊗ 1K ′)V ρ′ Y V ρ′ S1 (S△B)c S△B+ + , ((YBc fρ′(S1, Bc)) ⊗ 1K ′)V(S△B)b , (S△B)c Y = V ρ′ S2 (S△B)a ((YBafρ′(S1, Ba))⊗1K ′)V ρ′ +XB⊂S +XB⊂S +XB⊂S ((YBafρ′(S2, Ba)) ⊗ 1K ′)V ρ′ S△B +XB⊂S ((YBc fρ′(S2, Bc)) ⊗ 1K ′)V ρ′ (S△B)a +XB⊂S ((YBafρ′(Ba, S1))⊗1K ′)V ρ′ +XB⊂S ((YBc fρ′(Bc, S1)) ⊗ 1K ′)V ρ′ +XB⊂S S△B +XB⊂S ((YBafρ′(Ba, S2)) ⊗ 1K ′)V ρ′ ((YBc fρ′(Bc, S2)) ⊗ 1K ′)V ρ′ Y V ρ′ S2 (S△B)a (S△B)a (S△B)c = , (S△B)b S△B+ + , +XB⊂S = XB⊂S +XB⊂S ((YBfρ′(B, S2)) ⊗ 1K ′)V ρ′ (S△B)b ((YBbfρ′(Bb, S2)) ⊗ 1K ′)V ρ′ √2Y = XB⊂S +XB⊂S +XB⊂S +XB⊂S √2Y = XB⊂S ((α∗1YBafρ′(S1, Ba) + α∗2YBbfρ′(S2, Bb)) ⊗ 1K ′)V ρ′ S△B+ (S△B)a ((α∗1YBfρ′(S1, B) + α∗2YBcfρ′(S2, Bc)) ⊗ 1K ′)V ρ′ ((α∗1YBcfρ′(S1, Bc) + α∗2YBfρ′(S2, B)) ⊗ 1K ′)V ρ′ ((α∗1YBbfρ′(S1, Bb) + α∗2YBafρ′(S2, Ba)) ⊗ 1K ′)V ρ′ (S△B)c ((α∗1YBafρ′(Ba, S1) + α∗2YBbfρ′(Bb, S2)) ⊗ 1K ′)V ρ′ (S△B)b + + , S△B+ 152 + (S△B)a ((α∗1YBfρ′(B, S1) + α∗2YBcfρ′(Bc, S2)) ⊗ 1K ′)V ρ′ ((α∗1YBcfρ′(Bc, S1) + α∗2YBfρ′(B, S2)) ⊗ 1K ′)V ρ′ ((α∗1YBbfρ′(Bb, S1) + α∗2YBafρ′(Ba, S2)) ⊗ 1K ′)V ρ′ +XB⊂S +XB⊂S +XB⊂S √2YBa = α∗1YS△Bfρ′(S△B, S1) + α∗2Y(S△B)cfρ′((S△B)c, S2) , √2YBb = α∗1Y(S△B)c fρ′((S△B)c, S1) + α∗2YS△Bfρ′(S△B, S2) , (S△B)c (S△B)b + . It follows for B ⊂ S, √2YBc = α∗1Y(S△B)bfρ′(S1, (S△B)b) + α∗2Y(S△B)a fρ′(S2, (S△B)a) = = α∗1Y(S△B)bfρ′((S△B)b, S1) + α∗2Y(S△B)afρ′((S△B)a, S2) , so by Proposition 4.1.3 a),b), YBc = 0. If we put B ∈ Cl(ρ) (YB ⊗ 1K )V ρ X := 2XB⊂S ϕX = 2XB⊂S √2 XB⊂S ((α∗2YBfρ′(B, S2)) ⊗ 1K ′)V ρ′ (YB ⊗ 1K ′)V ρ′ B! P = B + 1 S2△B , (YB ⊗ 1K ′)V ρ′ √2 XB⊂S + 1 ((α∗1YBfρ′(B, S1)) ⊗ 1K ′)V ρ′ S1△B+ then = XB⊂S and so for B ⊂ S, (ϕX)B = YB , (ϕX)Ba = 1 √2 α∗1YS△Bfρ′(S△B, S1) = YBa , (ϕX)Bb = 1 √2 α∗2YS△Bfρ′(S△B, S2) = YBb , (ϕX)Bc = 0 = YBc . Thus ϕX = Y and ϕ is surjective. Remark. If m = 3 then ϕ may be not surjective. 153 PROPOSITION 4.2.4 Let IK := IR, n ∈ IN ∪ {0}, S := IN2n, and ρ′ : IN2n+1 −→ U n Ec , s 7−→(cid:26) ρ(s) − fρ(S) if if s ∈ S s = 2n + 1 . ◦z}{ Let ([C1] Theorem 4.1.1.8 a)) by using the embedding Cl(ρ) be the complexification of Cl(ρ), considered as a real E-C-algebra E −→ ◦z}{ Cl(ρ), x 7−→ ((x ⊗ 1K )V ρ ∅ , 0) . Then there is a unique E-C-isomorphism ϕ : Cl(ρ′) → (V ρ s , 0) for every s ∈ S and ϕV ρ′ 2n+1 = (0,−( fρ(S) ⊗ 1K)V ρ S ) . ◦z}{ Cl(ρ) such that ϕV ρ′ s = We put xs :=(cid:26) (V ρ s , 0) (0,−( fρ(S) ⊗ 1K )V ρ S ) if if s ∈ S s = 2n + 1 . For s ∈ S, by Proposition 4.1.3 b), xsx2n+1 = (V ρ = (0, ( fρ(S) ⊗ 1K )V ρ By Proposition 2.1.2 b),e), s , 0)(0,−( fρ(S) ⊗ 1K )V ρ S ) = (0,−( fρ(S) ⊗ 1K )V ρ s V ρ S ) = S V ρ s ) = (0, ( fρ(S) ⊗ 1K )V ρ s )(V ρ s , 0) = −x2n+1xs . 2n+1 = (−(( fρ(S) ⊗ 1K )V ρ x2 S )2, 0) = = (−( fρ(S)2 ⊗ 1K )(fρ(S, S) ⊗ 1K )V ρ ∅ , 0) = (ρ′(2n + 1) ⊗ 1K)(V ρ ∅ , 0) , x∗2n+1 = (0, (( fρ(S) ⊗ 1K)V ρ S )∗) = = (0, ( fρ(S)∗ ⊗ 1K )( fρ(S) ⊗ 1K )V ρ S ) = (ρ′(2n + 1)∗ ⊗ 1K)x2n+1 , and the assertion follows from Proposition 4.2.2. 154 PROPOSITION 4.2.5 Let n ∈ IN ∪ {0}, S := INn, S′ l2(P(S′)), α1, α2 ∈ U n Ec, and := INn+2, K′ := ρ′ : S′ −→ U n Ec , if if if s ∈ S s = n + 1 s = n + 2 . a) There is a unique E-C-isomorphism ϕ : Cl(ρ′) → Cl(ρ)2,2 such that ρ(s) α2 1 −α2 2 s 7−→ s =(cid:20) V ρ 0 (cid:21) , ϕV ρ′ V ρ ∅ ϕV ρ′ s s (cid:21) 0 0 −V ρ n+2 = (α2 ⊗ 1K )(cid:20) 0 −V ρ V ρ ∅ 0 ∅ (cid:21) . for every s ∈ S and ϕV ρ′ n+1 = (α1 ⊗ 1K )(cid:20) 0 V ρ ∅ ϕ ϕ 1 2 1 2 b) a) Put + ((α∗1α∗2) ⊗ 1K ′)V ρ′ (V ρ′ ∅ ∅ − ((α∗1α∗2) ⊗ 1K ′)V ρ′ (V ρ′ {n+1, n+2} {n+1, n+2} ∅ 0 ) =(cid:20) V ρ ) =(cid:20) 0 0 0 V ρ 0 0 (cid:21) , ∅ (cid:21) . s xs :=(cid:20) V ρ 0 (cid:21) , s (cid:21) 0 0 −V ρ xn+2 := (α2 ⊗ 1K)(cid:20) 0 −V ρ V ρ ∅ 0 ∅ V ρ ∅ (cid:21) . for every s ∈ S and xn+1 := (α1 ⊗ 1K )(cid:20) 0 V ρ ∅ For distinct s, t ∈ S and i ∈ IN2, xsxt = −xtxs , x2 s = (ρ′(s) ⊗ 1K)(cid:20) V ρ ∅ 0 0 V ρ ∅ (cid:21) , x∗s = (ρ′(s) ⊗ 1K )∗xs , xsxn+i = −xn+ixs , x∗n+i = (ρ′(n + i) ⊗ 1K)∗xn+i , x2 n+i = (ρ′(n + i) ⊗ 1K )(cid:20) V ρ ∅ 0 0 V ρ ∅ (cid:21) , xn+1 xn+2 = −xn+2 xn+1 . 155 By Proposition 4.2.2 there is a unique E-C-homomorphism ϕ : Cl(ρ′) → Cl(ρ)2,2 satisfying the given conditions. We put for every A ⊂ S and i ∈ IN2 A := Card A , A3 := A ∪ {n + 1, n + 2} . Ai := A ∪ {n + i} , For A ⊂ S, Then for Y ∈ Cl(ρ′), It follows from the above identities that ϕ is bijective. b) By the above, ϕV ρ′ {n+1, n+2} = ϕV ρ′ ∅3 = ((α1α2) ⊗ 1K )(cid:20) V ρ ∅ (cid:21) 0 ∅ 0 −V ρ and the assertion follows. 156 V ρ ∅ 0 (cid:21) = V ρ ∅ A (cid:21)(cid:20) 0 0 (cid:21) , V ρ A (cid:21) = ∅ 0 A (cid:21)(cid:20) 0 −V ρ V ρ ∅ −V ρ 0 0 (cid:21)(cid:20) 0 −V ρ (cid:21) , V ρ A V ρ ∅ 0 ∅ A (cid:21) = ϕV ρ′ A1 ϕV ρ′ A2 (−1)AV ρ (−1)AV ρ A (−1)AV ρ (−1)AV ρ A 0 0 0 0 A 0 A 0 = (α1 ⊗ 1K )(cid:20) V ρ = (α1 ⊗ 1K )(cid:20) = (α2 ⊗ 1K )(cid:20) V ρ = (α2 ⊗ 1K )(cid:20) = ((α1α2) ⊗ 1K)(cid:20) (−1)AV ρ = ((α1α2) ⊗ 1K )(cid:20) V ρ (ϕY )11 = PA⊂S (ϕY )12 = PA⊂S (ϕY )21 = PA⊂S (ϕY )22 = PA⊂S 0 A ϕV ρ′ A3  A A (cid:21) . 0 −(−1)AV ρ 0 A ((YA + (α1α2)YA3) ⊗ 1K )V ρ ((α1YA1 − α2YA2) ⊗ 1K )V ρ (−1)A)((α1YA1 + α2YA2) ⊗ 1K)V ρ (−1)A((YA − α1α2YA3) ⊗ 1K)V ρ A . A A COROLLARY 4.2.6 Let m, n ∈ IN∪{0}, S := INn, (αi)i∈IN2m ∈ (U n Ec)2m, and ρ′ : INn+2m −→ U n Ec , s 7−→(cid:26) ρ(s) −(−1)iα2 i if if s ∈ S s = n + i . Then Cl(ρ′) ≈E Cl(ρ)2m,2m. PROPOSITION 4.2.7 Let IK := IR, n ∈ IN ∪ {0}, S := IN2n, S′ := IN2n+2, α1, α2 ∈ U n Ec, and ρ′ : S′ −→ U n Ec , ρ(s) fρ(S) s ∈ S if if . s = 2n + l with l ∈ IN2 s 7−→(cid:26) −α2 l Then there is a unique E-C-isomorphism ϕ : Cl(ρ′) → Cl(ρ) ⊗ IH such that ϕV ρ′ s = xs := if if if s ∈ S s = 2n + 1 s = 2n + 2 , if if if s ∈ S s = 2n + 1 s = 2n + 2 . V ρ s ⊗ 1 IH (((α1 fρ(S)) ⊗ 1K )V ρ (((α2 fρ(S)) ⊗ 1K )V ρ S ) ⊗ i S ) ⊗ j where i, j, k are the canonical unitaries of IH. Put V ρ s ⊗ 1 IH (((α1 fρ(S)) ⊗ 1K )V ρ (((α2 fρ(S)) ⊗ 1K )V ρ S ) ⊗ i S ) ⊗ j ∅ ⊗ 1 IH) , s = (ρ′(s) ⊗ 1K )(V ρ x2 For distinct s, t ∈ S and l ∈ IN2, by Proposition 4.1.3 b), x∗s = (ρ′(s) ⊗ 1K )∗xs , xsxt = −xtxs , xsx2n+l = −x2n+lxs, x2n+1x2n+2 = (((α1α2 fρ(S))⊗1K )V ρ )⊗k = −x2n+2x2n+1, ∅ fρ(S)2) ⊗ 1K )( fρ(S)∗ ⊗ 1K)V ρ ) ⊗ (−1 IH) = ∅ = (ρ′(2n + l) ⊗ 1K )(V ρ ∅ ⊗ 1 IH) , fρ(S)∗) ⊗ 1K)( fρ(S) ⊗ 1K )V ρ = (ρ′(2n + l) ⊗ 1K )∗x2n+l . S ) ⊗ −(i or j) = (x2n+l)∗ = (((α∗l (x2n+l)2 = (((α2 l By Proposition 4.2.2 there is a unique E-C-homomorphism ϕ : Cl(ρ′) → Cl(ρ) ⊗ IH satisfying the given conditions. 157 For X ∈ Cl(ρ′), (XA ⊗ 1K)V ρ A! ⊗ 1 IH+ ϕX = XA⊂S ((XA∪{2n+1}α1 fρ(S)fρ(A, S)) ⊗ 1K )VS△A! ⊗ i+ S△A! ⊗ j+ ((XA∪{2n+2}α2 fρ(S)fρ(A, S)) ⊗ 1K )V ρ A! ⊗ k ((XA∪{2n+1, 2n+2}α1α2 fρ(S)) ⊗ 1K )V ρ + XA⊂S + XA⊂S + XA⊂S and so ϕ is bijective. PROPOSITION 4.2.8 Let n ∈ IN ∪ {0}, S := IN2n, A′ := A ∪ {2n + 1} for every A ⊂ S, ρ′ : S′ −→ U n Ec , ∅ ± V ρ′ S′ ), and θ± : (cid:13)A⊂S P± := 1 2 (V ρ′ (θ±ξ)A := 1 √2 ξA , for every ξ ∈ (cid:13)A⊂S E and A ⊂ S. f (S) s 7−→(cid:26) ρ(s) E → (cid:13)A⊂S′ (θ±ξ)A′ := ± if if s ∈ S s = 2n + 1 , E defined by 1 √2 fρ(S△A, S)ξS△A a) fρ′(S′) = 1E , P+ + P− = V ρ′ ∅ , (V ρ′ S′ )2 = V ρ′ , ∅ V ρ′ S′ ∈ Cl(ρ′)c , P± ∈ P r Cl(ρ′)c , V ρ′ S′ P± = ±P± . b) For A ⊂ S, fρ(A, S)∗ = fρ′(S′, A)∗ = fρ′(S′, (S△A)′) . 158 c) θ± ∈ LE( (cid:13)A⊂S (θ∗ E and A ⊂ S, E) and for η ∈ (cid:13)A⊂S′ E, (cid:13)A⊂S′ (ηA ± fρ(A, S)∗η(S△A)′) = √2(P±η)A . 1 √2 ±η)A = E. d) θ∗±θ± is the identity map of (cid:13)A⊂S e) θ±θ∗± = P±. f ) For every A ⊂ S, θ±V ρ Aθ∗ ± = V ρ′ A P± = P±V ρ′ A = P±V ρ′ A P± . g) For every closed ideal F of E the map ϕ : Cl(ρ, F ) −→ P±Cl(ρ′, F )P± , X 7−→ θ±Xθ∗ ± is an E-C-isomorphism with inverse P±Cl(ρ′, F )P± −→ Cl(ρ, F ), Y 7−→ θ∗ ±Y θ± and the map ψ : Cl(ρ′, F ) −→ Cl(ρ, F ) × Cl(ρ, F ) , (θ∗+P+Y P+θ+, θ∗ −P−Y P−θ−) = (θ+Y θ+, θ∗ Y 7−→ −Y θ−) is an E-C-isomorphism. a) By Proposition 4.1.3 d),e), V ρ′ fρ′(S′) = (−1)n(2n+1) Ys∈S′ S′ ∈ Cl(ρ′)c, ρ′(s)∗ = (−1)n(2n−1) Ys∈S ρ(s)∗! ρ′(2n + 1)∗ = 1E , so b), (V ρ′ S′ )∗ = fρ′(S′)V ρ′ S′ = V ρ′ S′ , (V ρ′ S′ )2 = f (S′)∗V ρ′ ∅ = V ρ′ ∅ , P± ∈ P r Cl(ρ′)c , V ρ′ S′ P± = ±P± . b) By a), Proposition 4.1.3 c),d), Proposition 4.1.1 b), and Proposition 1.1.2 fρ(A, S)∗ = fρ′(A, S)∗ = fρ′(A, S′)∗ = 159 = fρ′(S′, A)∗ = fρ′(S′, (S△A)′) fρ′(S′) = fρ′(S′, (S△A)′) . c) For ξ ∈ (cid:13)A⊂S E, h θξ η i = XA⊂S = XA⊂S η∗A η∗A 1 √2 1 √2 1 √2 η∗A′ 1 √2 ξA ±XA⊂S ξA ±XA⊂S (ηA ± fρ(A, S)∗η(S△A)′)∗ξA η∗(S△A)′ 1 √2 fρ(S△A, S)ξS△A = fρ(A, S)ξA = = XA⊂S E, (cid:13)A⊂S′ E) and so θ ∈ LE( (cid:13)A⊂S By a) and b), (θ∗η)A = 1 √2 (ηA ± fρ(A, S)∗η(S△A)′) . (P±η)A = 1 2 ηA ± 1 2 fρ′(S′, (S△A)′)η(S△A)′ = = 1 2 (ηA ± fρ(A, S)∗η(S△A)′) = 1 √2 (θ∗ ±η)A . E and A ⊂ S, by c), d) For ξ ∈ (cid:13)A⊂S (θ∗ ±θ±ξ)A = 1 √2 ((θξ)A ± fρ(A, S)∗(θξ)(S△A)′) = = 1 2 (ξA + fρ(A, S)∗fρ(A, S)ξA) = ξA . e) For η ∈ (cid:13)A⊂S′ E and A ⊂ S, by b) and c), (θ±θ∗ ±η)A = 1 √2 (θ∗ ±η)A = (P±η)A , 160 (θ±θ∗ ±η)A′ = ± 1 √2 1 2 = ± fρ(S△A, S)(ηS△A ± fρ(S△A, S)∗ηA′) = ± 1 2 (ηA′ ± fρ′(S′, S△A)ηS△A) = ((V ρ′ ∅ 1 2 = fρ(S△A, S)(θ∗ ±η)S△A = 1 2 fρ(S△A, S)ηS△A + 1 2 ηA′ = η)A′ ± (V ρ′ S′ η)A′) = (P±η)A′ , so θ±θ∗± = P±. f) For η ∈ (cid:13)B⊂S′ Corollary 2.1.17 e)), E and B ⊂ S, by a),b),c),e) and Proposition 4.1.1 b) (and (V ρ′ A P±η)B = fρ′(A, A△B)(P±η)A△B = fρ(A, A△B)(θ±θ∗ ±η)A△B = ±η)B = (θ±V ρ Aθ∗ ±η)A△B = ±η)B , Aθ∗ (V ρ = 1 √2 1 √2 fρ(A, A△B)(θ∗ (θ±V ρ 1 √2 Aθ∗ 1 √2 ±η)B′ = ± fρ(S△B, S)(V ρ Aθ∗ fρ(S△B, S)fρ(A, S△A△B)(θ∗ ±η)S△B = = ± ±η)S△A△B = = ±fρ(S△B, S)fρ(A, S△A△B)(P±η)S△A△B = ±fρ(S△B, S)(V ρ′ A P±η)B′ = = ±fρ′(S′, S′△B′)(V ρ′ A V ρ′ = ±(V ρ′ S′ V ρ′ A P±η)S′△B′ = ±(V ρ′ S′ P±η)B′ = (V ρ′ A P±η)B′ A P±η)S△B = so by a), θ±V ρ Aθ∗ ± = V ρ′ A P± = P±V ρ′ A P± = P±V ρ′ A . g) The assertion concerning ϕ as well as the identity in the definition of ψ follow from a),d),e), and f). Thus ψ is a surjective E-C-homomorphism. For Y ∈ Ker ψ, θ∗+Y θ+ = θ∗ −Y θ− = 0 , so by a) and e), and we get i.e. ψ is injective. P+Y = P−Y = 0 Y = P+Y + P−Y = 0 161 REFERENCES [C1] Corneliu Constantinescu, C-algebras. Elsevir, 2001. [C2] Corneliu Constantinescu, W-tensor products and selfdual Hilbert right W ∗-modules. Rev. Roumaine Math. Pures Appl., 51: 5-6 (2006) 583-596. [C3] Corneliu Constantinescu, Selfdual Hilbert right W ∗-modules and their W-tensor products. Rev. Roumaine Math. Pures Appl., 55: 3 (2010) 159-196. [K] Richard V. Kadison and John R. Ringrose, Fundamentals of the theory of operator algebras. Academic Press, 1983-1986. [L] Christopher E. Lance, Hilbert C-modules. A toolkit for operator algebraist. Cambridge University Press, 1995. [S] Issai Schur U ber die Darstellung der endlichen Gruppen durch gebrochene lineare Substitutionen. J. Reine Angew. Math., 127 (1904) 20-50. [T] Masamichi Takesaki, Theory of Operator Algebra I. Springer, 2002. [W] N. E. Wegge-Olsen, K-theory and C-algebras. Oxford University Press, 1993. 162 Index Schur E-function for T , F(T, E), f , f Λ(T, E), λ, δλ (Definition 1.1.3). E-module, E-linear (Definition 1.2.1). E-C-algebra, E-W-algebra, E-C-subalgebra, (Definition 1.1.1). E-W-subalgebra, E-C-homomorphism, E-W-homomorphism, E-C*- isomorphism, ≈E (Definition 1.2.2). E (Definition 1.2.3). CE, C1 adapted, F , ME (Proposition 1.2.4). ϕ (Proposition 1.2.6). ΦE (Proposition 1.2.10). T ¯G, e(cid:13) , e⊗,fP, GT, xe⊗1K (Lemma 1.3.2). TP (Definition 1.3.1). S(f, F ), Sk·k(f, F ) (Definition 2.1.7). T1, T2, T3 (Definition 1.3.3). ut, Vt, V f (Definition 2.1.1). t ϕs,t , Xt (Definition 2.1.5). R(f ), S(f ), SC (f ), SW (f ), Sk·k(f ), Locally finite, ST (Definition 2.1.18). Uλ (Definition 2.2.1). S-isomorphism, ≈S (Proposition 2.2.2 S(F ), V F S(ϕ) (Proposition 2.3.2). IT (Definition 3.1.3). w(x), winding number of x (Definition 3.1.5). fρ, Cl(ρ) Box of subsection 4.2. Pf , △ (Definition 4.2.1). (Definition 2.3.1). a)). t 163 Corneliu Constantinescu Bodenacherstr. 53 CH 8121 Benglen e-mail: [email protected] 164
1611.04742	3	1611	2016-11-26T20:56:14	On Propagation of Fixed Points of Quantum Operations and Beyond	[ "math.OA", "math.FA" ]	We show that some abstract results on propagation of fixed points for completely positive maps on $C^*$-algebras provide a natural approach to unify recent Noether type theorems on the equivalence of symmetries with conservation laws for dynamical systems of Markov processes, of quantum operations, and of quantum stochastic maps. In addition, we obtain some new Noether type theorems, provide examples and counter-examples, and extend most of the existing results with characterisations in terms of dual infinitesimal generators of the corresponding strongly continuous one-parameter semigroups.	math.OA	math	ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND AURELIAN GHEONDEA Abstract. We show that some abstract results on propagation of fixed points for com- pletely positive maps on C ∗-algebras provide a natural approach to unify recent Noether type theorems on the equivalence of symmetries with conservation laws for dynamical sys- tems of Markov processes, of quantum operations, and of quantum stochastic maps. In addi- tion, we obtain some new Noether type theorems, provide examples and counter-examples, and extend most of the existing results with characterisations in terms of dual infinitesimal generators of the corresponding strongly continuous one-parameter semigroups. 6 1 0 2 v o N 6 2 ] . A O h t a m [ 3 v 2 4 7 4 0 . 1 1 6 1 : v i X r a 1. Introduction In view of the celebrated theorem of E. Noether [25] on the equivalence of symmetries and conservation laws, J.C. Baez and B. Fong [6] considered similar questions within the frame- work of "stochastic mechanics", in the sense of [5], for the dynamics of Markov processes. Letting {U(t)}t≥0 be a (classical) dynamical stochastic system (this is called a Markov semi- group in [6]), they show that the operator of multiplication with an observable O commutes with Ut for all t ≥ 0, an analogue for a symmetry, if and only if both its expected value hO, Utfi and the expected value of its square hO2, Utfi are constant in time for every state f (probability distribution), an analogue for a conservation law. Considering the variance hO2, fi−hO, fi2, for f an arbitrary state, the latter is equivalent with both its expected value and its variance (or standard deviation) are constant in time for every state. The appearance of the variance makes a difference when compared to the classical Noether's Theorem. It is one of our aim to show that, when viewing this from the perspective of the approach of [2], similar facts have been observed previously in closely related mathematical problems on irreversible dynamical quantum systems, e.g. as in S. Albeverio and R. Høegh-Krohn, [1], E.B. Davies [13], D. Evans [15], A. Frigerio and M. Verri [17], and E. Størmer [30], to quote a few. More precisely, letting A = {An}n∈N be a sequence of positive operators in B(H), for some Hilbert space H, such that P∞ n=1 An = I one considers the quantum operation ΦA, in the Date: September 23, 2018. 2010 Mathematics Subject Classification. 46L07, 47D07, 82C10, 81R15, 81P20, 81Q80, 70H33, 60J25, 60J35. Key words and phrases. C ∗-algebras, completely positive maps, multiplicative domains, dynamical quan- tum systems, fixed points, symmetries of dynamical systems, constants of dynamical systems, dynamical stochastic systems, dynamical Markov systems, Noether Theorem. Work supported by a grant of the Romanian National Authority for Scientific Research, CNCS UEFISCDI, project number PN-II-ID-PCE-2011-3-0119. 1 2 A. GHEONDEA Schrodinger picture and the Luders form, defined by (1.1) ΦA(T ) = ∞Xn=1 A1/2 n T A1/2 n , T ∈ B1(H), hence, ΦA is a completely positive and trace preserving linear map on B1(H). Note that in the Heisenberg picture its dual Φ♯ A is a unital normal completely positive linear map on B(H). The equivalence of assertions (i) and (ii) in the following proposition was obtained as Corollary 3.4 in [2], while the equivalence with assertions (iii) and (iv) is clear. A has the same formal expression as in (1.1) and that Φ♯ Proposition 1.1. Let ΦA be the unital quantum operation in the Schrodinger picture as in (1.1), its dual Φ♯ A. The following assertions are equivalent: A in the Heisenberg picture, and let B ∈ B(H) be a fixed point of Φ♯ (i) B commutes with all operation elements A1, A2, . . . of ΦA. (ii) B∗B and BB∗ are fixed points of Φ♯ (iii) The whole unital C ∗-algebra generated by B is fixed by Φ♯ (iv) The whole von Neumann algebra generated by B is fixed by Φ♯ A. A. A. Note that Proposition 1.1 implies that, a selfadjoint operator B ∈ B(H) commutes with all operation elements of A if and only if both B and B2 are fixed points of Φ♯ A, hence a characterisation of exactly the same type with that obtained in the Noether type theorem of [6]. There are important differences between these two results, notably the latter condition on the square of the observable is necessary even in the finite state space case for the classical Markov processes, cf. the example at page 3 in [6], while for the Luders operation it is not, cf. Theorem 3.5 in [2]. The result in [6] has been put in a setting of dynamical quantum systems by J.E. Gough, T.S. Rat¸iu, and O.G. Smolyanov in [19]. More precisely, let T = {Tt}t≥0 denote a dynami- cal system in the Schrodinger picture, that is, a norm continuous semigroup of completely positive trace-preserving linear maps on the trace-class B1(H) for some fixed Hilbert space H, for which the infinitesimal generator M takes the form, cf. [24], [18], (1.2) SL∗ (LkSL∗ L∗ kLkS) + i[S, H], S ∈ B1(H), kLk − 1 2 M(S) =Xk 1 2 k − for a collection of operators Lk ∈ B(H), k = 1, 2, . . . , and a selfadjoint operator H ∈ B(H). The constants of T are the operators A ∈ B(H) such that tr((Ttρ)A) = tr(ρA) for all density operators ρ ∈ D(H) and all t ≥ 0. Transferring to the Heisenberg picture, one considers the dual semigroup {Jt}t≥0 acting in B(H) whose set of fixed points, that is, all A ∈ B(H) such that Jt(A) = A for all t ≥ 0, coincides with the set of constants of T . The main result in [19] says that, under the technical assumption of existence of a stationary strictly positive density operator, the set of constants of the quantum dynamical system {Tt}t≥0, which coincides with the set of fixed points of {Jt}t≥0, is a von Neumann algebra and it coincides with the commutant {H, Lk, L∗ k k = 1, 2, . . .}′. In their formulation, an analogue of the second condition on the square of the observable as in [6] does not show up and, another aim of our article is to show that this happens because it is obscured by the technical assumption of existence of a stationary strictly positive density operator. ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 3 Within the same circle of ideas as in [6] and [19], K. Bartoszek and W. Bartoszek [7] recently considered a noncommutative version of dynamical stochastic system, more pre- cisely, a strongly continuous semigroup {St}t≥0 of stochastic maps with respect to some Hilbert space H, that is, trace-preserving positive linear maps on the trace-class B1(H), and an one-element measurement operator MA1/2, for some positive operator A ∈ B(H), where MA1/2(T ) = A1/2T A1/2. In this setting, they obtain several equivalent characterisations to the compatibility (commutation) of the dynamical stochastic system {St}t≥0 with the quan- tum measurement MA1/2: for example, one of these equivalent characterisations refers to A and A2 being fixed by the dual semigroup {S♯ t}t≥0 and a second one refers to the com- mutation of the infinitesimal generator s of {St}t≥0 with MA1/2. The approach used in [7] combines the probability theory methods as in [6] with operator theoretical methods. There are some important questions left unanswered in [7]: for example, to which extent the addi- tional condition of A2 to be fixed by the dual semigroup {S♯ t}t≥0 as well is really necessary? It is another aim of our article to provide an answer to this question. In this article we show that the C ∗-algebraic dilation theoretical approach as in [2] uni- fies all the results in [6], [19], and [7] under a common framework of propagation of fixed points for completely positive maps. In addition, for each of the noncommutative Noether type theorems considered in [19] and [7], we provide examples and counter-examples that clarify the necessity of the second order extra conditions imposed, obtain some new Noether type theorems, and extend most of the existing results with characterisations in terms of dual infinitesimal generators of strongly continuous one-parameter semigroups. Actually, we show that the abstract results obtained in theorems 2.2 and 2.4 short-cuts completely the probabilistic tools in the proofs of the main results in [6] and [7], while in the case of [19] they reveal what happens in case the technical assumption of existence of a stationary strictly positive density operator is removed. We briefly describe the contents of this article. Section 2 contains most of the technical results that are needed in this article. Firstly, we obtain Theorem 2.2 that shows that for a unital linear map Φ on a C ∗-algebra A that is completely positive when restricted to the unital C ∗-algebra generated by an element a ∈ A, the fixation of a, a∗a, and aa∗ propagates to the whole unital C ∗-subalgebra generated by a. Since, according to a classical theorem of Stinespring [29], positivity on commutative C ∗-algebras implies complete positivity, if a is a normal element the same conclusions as in Theorem 2.2 can be obtained for positive unital maps, as in Corollary 2.3. These results are obtained through a classical result on multi- plicative domains of M.-D. Choi [10]. Subsection 2.1 provides an equivalent characterisation of the set of fixed points for a w∗-continuous semigroup of w∗-continuous operators in terms of the null space of its w∗-infinitesimal generator, which is used in all three cases considered in sections 3, 4, and 5, in order to obtain characterisations in terms of the infinitesimal gen- erators of the dual (Markov) semigroups. We think that Theorem 2.5 is most likely known but we could not find a reference for it. Subsection 2.3 provides a semigroup version of Theorem 2.4 in [2], more precisely, letting Φ = {Φt}t≥0 be a w∗-continuous semigroup of w∗-continuous, unital, and completely positive maps on a von Neumann algebra M, by an ergodic theoretical approach we show that the set of joint fixed points MΦ is the range of a completely positive, unital, and idempotent map Ψ : M → M. This fact is the technical tool to be used, in conjunction with some classical results on injective von Neumann algebras, in clarifying the question whether the additional condition on A∗A and AA∗ in Theorem 5.6 is 4 A. GHEONDEA necessary, for infinite dimensional Hilbert spaces, by adapting the counter-example from [2] to the semigroup setting. In Section 3 we provide different proofs for the main results of [7], on compatibility of one- element quantum measurements with stochastic maps in both the discrete and continuous dynamical systems cases, more precisely, we show that these results can be obtained directly from Corollary 2.3. Example 3.1 shows, by means of the transpose map with respect to a fixed orthonormal basis, that the set of stochastic maps that are not quantum operations is quite large. In the case of a continuous dynamical stochastic system, we additionally find two more equivalent characterisations of the compatibility of one-parameter semigroups of stochastic maps {Ψt}t≥0 with an one-element quantum measurement operator MA1/2 in terms of the dual infinitesimal generator ψ♯: one by the commutation of MA1/2 with ψ♯ and the second by the fact that ψ♯ annihilates both A and A2. In Section 4 we consider the setting of dynamical systems of classical Markov processes as in [6] and show how the Noether type theorems obtained in that paper can be natu- rally recovered under our approach. In the discrete semigroup case, we point out additional equivalent characterisations through the dual semigroup while, in the case of a strongly con- tinuous semigroup, we obtain additional equivalent characterisations in terms of infinitesimal generators, dual semigroups and dual infinitesimal generators. c ⊇ CΨ In Section 5 we consider a slightly more general setting of dynamical quantum systems, when compared to that used in [19], by replacing the operator norm continuity of the one- parameter semigroup with strong continuity, and reorganise most of it in a rather different fashion and obtain new results. Firstly, we consider discrete quantum semigroups for which we obtain Noether type theorems with respect to left and right multiplication by arbitrary bounded operators. Note that, due to the fact that dynamical quantum systems consist of completely positive maps only, these results go beyond multiplication operators with normal operators, a restriction that seems difficult to overcome in the case of dynamical stochastic systems as in [7]. Then, we point out a scale of sets of constants CΨ ⊇ CΨ 2 ⊇ CΨ p ⊇ CΨ w , for Ψ a dynamical quantum system (either discrete or continuous), and discuss their relation: we show that all these sets but CΨ coincide and they make a von Neumann algebra, while the first order set of constants CΨ is the largest one and only under special conditions, as the existence of a stationary faithful state, coincides with the other sets of constants, equivalently, is a von Neumann algebra. For strongly continuous quantum semigroups, additional characterisations in terms of the infinitesimal generators and dual infinitesimal generators are obtained. In Theorem 5.12 we show that, in any infinite dimensional and separable Hilbert space, there exists norm continuous quantum semigroups for which the set of constants is not a von Neumann algebra, equivalently, it is not stable under multiplication. This result clarifies also the question why the extra condition on A2 to be a fixed point is necessary, in general, in the infinite dimensional noncommutative (quantum) case. In Appendix we provide a modern proof of Theorem 2.1 as a consequence of the Stine- spring's Dilation Theorem, following [9], that shows that the C ∗-algebraic abstract results we rely upon have a dilation theoretical character. A few words about terminology. We have used the same names "stochastic" and, re- spectively, "Markov" for both the commutative (classical) case as in Section 4, and the noncommutative (quantum) case as in Section 3, hoping that there will be no danger of confusion. This way, we left the notions of quantum stochastic and, respectively, quantum ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 5 Markov referring to the case of quantum operations in the Schrodinger picture and, respec- tively, in the Heisenberg picture, following the terminology already established in quantum physics, see [16] and [19]. We thank Marius Dadarlat for drawing our attention to the proof of M.-D. Choi's Theorem in [9] obtainable solely from the Stinespring's Dilation Theorem and for many other useful discussions on these topics, to Radu Purice for clarifying some aspects from [19], and to Carlo Beenakker for indicating [12] and [22] as sources on the significance of the transpose map in quantum information theory. 2. Preliminary Results 2.1. Propagation of Fixed Points in C ∗-Algebras. Let A and B be C ∗-algebras with unit. A linear map Φ : A → B is positive if Φ(a) ≥ 0 for all a ∈ A+, where A+ = {x∗x x ∈ A} denotes the cone of positive elements in A. Any positive map is selfadjoint, in the sense that Φ(a∗) = Φ(a)∗ for all a ∈ A, and bounded, more precisely, according to the Russo -- Dye Theorem, kΦk = kΦ(e)k, where by e we denote the unit of A. Given an arbitrary natural number n, we consider the C ∗-algebra Mn(A) of all n × n matrices with entries in A, organized as a C ∗-algebra in a canonical way, e.g. by identi- fying it with the C ∗-algebra A ⊗ Mn. This gives rise to the n-th order amplification map Φn : Mn(A) → Mn(B) defined by (2.1) Φn(A) = [Φ(ai,j)]n i,j=1 , A = [ai,j]n i,j=1 ∈ Mn(A). Φ is called n-positive if Φn is positive. Φ is called completely positive if it is n-positive for all n ∈ N. Given A a C ∗-algebra with unit, a closed linear subspace S of A is called an operator system if it is stable under the adjoint operation a 7→ a∗ and contains the unit of A. Note that any operator system is linearly generated by the cone of all its positive elements. Also, for any linear map Ψ : S → B, for B an arbitrary C ∗-algebra, the definitions of positive map, n-positive map, and completely positive map, as defined before, make perfectly sense. More generally, these definitions make sense if S is assumed to be stable under the adjoint operation only. For an arbitrary linear map Φ : A → B, the set (2.2) is called the multiplicative domain of Φ. If Φ is unital then MΦ contains the unit of A. MΦ = {a ∈ A Φ(a∗a) = Φ(a)∗Φ(a) and Φ(aa∗) = Φ(a)Φ(a∗)} We start with the following theorem, due to M.-D. Choi [10]; it is worth observing that assertion (2) is actually a property of propagation of multiplicativity which motivates the name of MΦ. The Schwarz Inequality was first obtained in a special case by R.V. Kadison in [23], that's why sometimes it is called the Kadison -- Schwarz Inequality. Following [9], we provide a short and modern proof of this theorem in Appendix, which also points out its dilation theory substance, as a consequence of the Stinespring's Dilation Theorem [29]. Theorem 2.1. Let Φ : A → B be a contractive completely positive map. Then: (1) (The Schwarz Inequality) Φ(a)∗Φ(a) ≤ Φ(a∗a) for all a ∈ A. (2) (The Multiplicativity Property) Let a ∈ A. Then: (i) Φ(a∗a) = Φ(a)∗Φ(a) if and only if Φ(ba) = Φ(b)Φ(a) for all b ∈ A. 6 A. GHEONDEA (ii) Φ(aa∗) = Φ(a)Φ(a)∗ if and only if Φ(ab) = Φ(a)Φ(b) for all b ∈ A. Consequently, (2.3) MΦ = {a ∈ A Φ(ab) = Φ(a)Φ(b), Φ(ba) = Φ(b)Φ(a), for all b ∈ A}. (3) The multiplicative domain MΦ defined at (2.2) is a C ∗-subalgebra of A and it coin- cides with the largest C ∗-subalgebra C of A such that ΦC : C → B is a ∗-homomorphism. Actually, the Schwarz Inequality is true under the more general condition that Φ is 2- positive, while the Multiplicativity Property holds for 4-positive maps: see also [26]. We are interested in fixed points of positive maps between C ∗-algebras. Given a C ∗-algebra A with unit e, let Φ : A → A be a linear map that is unital and positive. We consider the set of the fixed points of Φ (2.4) of all fixed points of Φ and it is easy to see that AΦ is an operator system. Another set of interest is the bimodule domain AΦ = {a ∈ A Φ(a) = a}, I(Φ) = {a ∈ A Φ(ab) = aΦ(b), Φ(ba) = Φ(b)a, for all b ∈ A}, (2.5) which is a C ∗-subalgebra of A containing the unit e. Clearly, (2.6) I(Φ) ⊆ AΦ ∩ MΦ. On the other hand, if Φ is completely positive and contractive, by Theorem 2.1.(2) we have (2.7) AΦ ∩ MΦ = IΦ. As shown in [2], even for the very particular case of a Luders operation Φ on B(H), where B(H) denotes the von Neumann algebra of all bounded operators on a Hilbert space H, in general we cannot expect that the set of fixed points of Φ coincides with its bimodule domain. In the following we consider a related question: given a unital positive map Φ : A → A, we want to see whether the quality of an element a ∈ A of being fixed by Φ propagates to the whole C ∗-algebra C ∗(e, a). In view of Proposition 1.1, it is not a surprise that this question is related to the concept of multiplicative domain, that is, imposing a∗a, aa∗ ∈ AΦ and a certain "locally complete positivity" condition on Φ as well. Theorem 2.2. Let A be a C ∗-algebra with unit e, let Φ : A → A be a unital linear map, and let a ∈ A be such that ΦC ∗(e,a) : C ∗(e, a) → A is completely positive. The following assertions are equivalent: (i) a, a∗a, aa∗ ∈ AΦ, that is, Φ(a) = a, Φ(a∗a) = a∗a, and Φ(aa∗) = aa∗. (ii) a ∈ AΦ ∩ MΦ, that is, Φ(a) = a, Φ(a∗a) = Φ(a)∗Φ(a), and Φ(aa∗) = Φ(a)Φ(a)∗. (iii) ΦC ∗(e,a) has the Bimodule Property, that is, Φ(ba) = Φ(b)a and Φ(ab) = aΦ(b) for (iv) C ∗(e, a) ⊆ AΦ, that is, Φ(b) = b for all b ∈ C ∗(e, a). all b ∈ C ∗(e, a). Proof. (i)⇒(ii). By assumptions it follows Φ(a∗a) = a∗a = Φ(a)∗Φ(a), Φ(aa∗) = aa∗ = Φ(a)Φ(a)∗, hence, a ∈ AΦ ∩ MΦ. Φ(xa) = Φ(x)Φ(a) = Φ(x)a and Φ(ax) = Φ(a)Φ(x) = aΦ(b) for all x ∈ C ∗(e, a). (ii)⇔(iii). By assumption and Theorem 2.1.(2), ΦC ∗(e,a) has the Bimodule Property, hence ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 7 (iii)⇒(iv). By assumption and using a straightforward induction argument, it follows that, for any n ∈ N0, we have (2.8) Φ(xan) = Φ(x)an, Φ(anx) = anΦ(x), x ∈ C ∗(e, a), and, since Φ is selfadjoint, we have Φ(a∗) = Φ(a)∗ = a∗, hence (2.9) Φ(xa∗n) = Φ(x)a∗n, Φ(a∗nx) = a∗nΦ(x), x ∈ C ∗(e, a). From (2.8) and (2.9), by a straightforward induction argument, it follows that for any mono- mial p in two noncommutive variables X and Y p(X, Y ) = X i1Y j1 · · · X imY jm, i1, . . . , jm ∈ N0, j1, . . . , jm ∈ N0, m ∈ N, it follows that Φ(p(a, a∗)) = p(a, a∗), (2.10) where p(a, a∗) ∈ A is the element obtained by formally replacing X with a and Y with a∗. Then, by linearity, it follows that (2.10) is true for any complex polynomials p in two noncommutative variables X and Y hence, since the collection of all elements of form p(a, a∗) is dense in A and ΦC ∗(e,a) is continuous, assertion (ii) follows. (iv)⇒(i). This implication is clear. As an application of Theorem 2.2 we record the special case of a normal element a, that is, a∗a = aa∗, when the condition of "locally complete positivity" follows from the condition of positivity. (cid:3) Corollary 2.3. Let Φ : A → A be a linear map which is positive and unital, and let a ∈ A be a normal element. The following assertions are equivalent: (i) Φ(a) = a and Φ(a∗a) = a∗a. (ii) Φ(b) = b for all b ∈ C ∗(e, a). Proof. Only the implication (i)⇒(ii) requires a proof. To see this, since a is normal it follows that C ∗(e, a) is a commutative C ∗-algebra hence ΦC ∗(e,a) : C ∗(e, a) → A is completely positive, see [29], and we can apply Theorem 2.2. (cid:3) One of the intrinsic deficiency of Theorem 2.2 is referring to the fact that we do not know that a has the Bimodule Property on the whole C ∗-algebra A. This deficiency is remedied for the case of quantum operations, in the Heisenberg picture, due to the overall complete positivity property. Theorem 2.4. Let A be a C ∗-algebra with unit e, let Φ : A → A be a unital completely positive linear map, and let a ∈ A. The following assertions are equivalent: (i) a, a∗a, aa∗ ∈ AΦ, that is, Φ(a) = a, Φ(a∗a) = a∗a, and Φ(aa∗) = aa∗. (ii) a ∈ AΦ ∩ MΦ, that is, Φ(ax) = aΦ(x) and Φ(xa) = Φ(x)a for all x ∈ A. (iii) C ∗(e, a) ⊆ AΦ, that is, Φ(b) = b for all b ∈ C ∗(e, a). The proof of this theorem follows the same line of argumentation as in the proof of The- orem 2.2 and we omit it. 8 A. GHEONDEA 2.2. Fixed Points of w∗-Continuous One-Parameter Semigroups. Let X be a Ba- nach space. We consider a strongly continuous one-parameter semigroup {Ψt}t≥0 of linear bounded operators on X, that is, (i) Ψt : X → X is a bounded linear operator for all t ≥ 0. (ii) ΨsΨt = Ψs+t, for all s, t ≥ 0. (iii) Ψ0 = I. (iv) R+ ∋ t 7→ Ψt(x) ∈ X is continuous for each x ∈ X. Under these assumptions, from the general theory of one-parameter semigroups, e.g. see E. Hille and R.S. Phillips [21], N. Dunford and J.T. Schwartz [14], the infinitesimal generator ψ exists as a densely defined closed operator on X, with (2.11) and (2.12) ψ(x) = lim t→0+ Ψt(x) − x t = d dt Ψt(x)t=0, x ∈ Dom(ψ), Dom(ψ) = {x ∈ X lim t→0+ Ψt(x) − x t exists in X}. In addition, e.g. see Corollary VIII.1.5 in [14], the limit (2.13) ω = lim t→∞ log kΨtk/t = inf t>0 log kΨtk/t exists with the growth bound ω < ∞ and, e.g. see Theorem VIII.1.11 in [14], for any complex number λ with Re λ > ω, the operator λI − ψ has a bounded inverse. Also, by the proof of the Hille-Yosida-Phillips Theorem, e.g. see Theorem VIII.1.13 in [14], we have (2.14) Ψt(x) = lim λ→∞ e−λt ∞Xn=0 (λ2t)n(λI − ψ)−n(x) n! , x ∈ Dom(ψ), t ≥ 0. Recall that X ♯ denotes the topological dual space of X. For every strongly continuous one-parameter semigroup {Ψt}t≥0 of bounded linear operators on X, the dual one-parameter semigroup {Ψ♯ (2.15) t}t≥0 of bounded linear operators on X ♯ exists, that is, x ∈ X, f ∈ X ♯, t ≥ 0, hΨt(x), fi = hx, Ψ♯ t(f )i, with the following properties s = Ψ♯ (i) Ψ♯ (ii) Ψ♯ (iii) Ψ♯ (iv) R+ ∋ t 7→ Ψ♯ t : X ♯ → X ♯ is a linear bounded and w∗-continuous operator for all t ≥ 0. tΨ♯ 0 = I. s+t, for all s, t ≥ 0. t(f ) ∈ X ♯ is w∗-continuous for each f ∈ X ♯. Then, e.g. see [27], {Ψ♯ t}t≥0 is a w∗-continuous semigroup of operators on X ♯ and hence, the w∗-infinitesimal generator ψ♯ exists as a w∗-closed operator on X ♯, hence a closed operator on X ♯, with (2.16) and (2.17) ψ♯(f ) = w∗- lim t→0+ Ψ♯ t(f ) − f t = w∗- d dt Ψ♯ t(f )t=0, Dom(ψ♯) = {f ∈ X ♯ w∗- lim t→0+ Ψ♯ t(f ) − f t exists in X ♯}. ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 9 The notation we use for ψ♯ looks like an abuse but actually it is not: by the R.S. Phillips's Theorem in [27], (2.18) and Dom(ψ♯) = {f ∈ X ♯ X ∋ f 7→ hx, ψ(f )i is continuous }, hψ(x), fi = hx, ψ♯(f )i, x ∈ Dom(ψ), f ∈ Dom(ψ♯), (2.19) hence, the w∗-infinitesimal generator ψ♯ of the dual w∗-continuous semigroup {Ψ♯ t}t≥0 on X ♯ is indeed the dual operator of the infinitesimal generator ψ of the strongly continuous semigroup {Ψt}t≥0 on X and, consequently, the notation for ψ♯ is fully justified. In addition, one of the major differences between the two infinitesimal generators ψ and ψ♯ is that Dom(ψ♯) may not be dense in X ♯, although it is always w∗-dense, while Dom(ψ) is always dense in X. Theorem 2.5. Let {Ψt}t≥0 be a strongly continuous semigroup of operators on a Banach space X, let {Ψ♯ t}t≥0 be the associated dual w∗-continuous semigroup of operators on X ♯, and ψ and, respectively, ψ♯, their infinitesimal generators. Considering f ∈ X ♯, the following assertions are equivalent: t(f ) = f for all real t ≥ 0. (i) Ψ♯ (ii) f ∈ Ker(ψ♯), that is, f ∈ Dom(ψ♯) and ψ♯(f ) = 0. Proof. (i)⇒(ii). This is a clear consequence of (2.16) and (2.17). (ii)⇒(i). Let λ > max{ω, 0}, where ω is defined as in (2.13). Since ψ♯ is the dual operator of ψ, as in (2.19) and (2.18), and λI − ψ is boundedly invertible, it follows that λI − ψ♯ is boundedly invertible, e.g. see Theorem 1.5 in [27]. Consequently, for any x ∈ Dom(ψ) and any g ∈ X ♯ we have ∞Xn=0 hx, e−λt (λ2t)(λI − ψ♯)−n(g) n! (λ2t)(λI − ψ)−n n! i = hx, e−λt(cid:0) ∞Xn=0 ∞Xn=0 = he−λt (λ2t)(λI − ψ)−n(x) n! (cid:1)♯(g)i , gi hence, by (2.14) it follows that (2.20) λ→∞hx, e−λt lim ∞Xn=0 (λ2t)(λI − ψ♯)−n(g) n! i = hΨt(x), gi. On the other hand, from ψ♯(f ) = 0 it follows that (λI−ψ♯)(f ) = λf hence (λI−ψ♯)−1(f ) = 1 λf . By induction we obtain (2.21) Consequently, it follows that (λI − ψ♯)−n(f ) = 1 λn f, n ≥ 0. ∞Xn=0 (λ2t)n(λI − ψ♯)−n(f ) n! = ∞Xn=0 (λt)n n! f = eλtf, 10 A. GHEONDEA hence, letting g = f in (2.20), it follows that hx, Ψ♯ t(f )i = hΨt(x), gi = lim λ→∞hx, e−λteλtfi = hx, fi, and then, since Dom(ψ) is dense in X, it follows that Ψ♯ t(f ) = f for all t ≥ 0. (cid:3) 2.3. An Ergodic Theorem in von Neumann Algebras. We first recall some definitions, in addition to those in Subsection 2.1. Let A and B be C ∗-algebras and let V ⊆ A and W ⊆ B be subspaces. For any linear map Φ : V → W and any natural number n, the n-th order amplification Φn : V ⊗ Mn → W ⊗ Mn can be defined as Φn = Φ ⊗ In, where In denotes the identity operator on Mn. Explicitly, by means of the canonical identifications Mn(V) = V ⊗ Mn and Mn(W) = W ⊗ Mn, this means (2.22) Note that, by the embeddings Mn(V) ⊆ Mn(A) and Mn(W) ⊆ Mn(B), it follows that Mn(V) and, respectively, Mn(W) have canonical norms induced by the C ∗-norms on Mn(A) and Mn(B). Consequently, we can let kΦnk denote the corresponding operator norm. Clearly, (2.23) i,j=1 ∈ Mn(V). kΦk = kΦ1k ≤ kΦ2k ≤ · · · ≤ kΦnk ≤ kΦn+1k ≤ · · · . i,j=1) = [Φ(vi,j)]n Φn([vi,j]n i,j=1, [vi,j]n The map Φ is called completely bounded if (2.24) kΦkcb = sup n≥1 kΦnk < ∞. Let CB(V,W) denote the vector space of all completely bounded maps Φ : V → W. Also, such a map Φ is called completely contractive if kΦkcb ≤ 1. A linear map Φ : V → V is called an idempotent if Φ2 = ΦΦ = Φ and, it is called a projection if it is completely contractive and idempotent. A subspace V ⊆ B(H), for some Hilbert space H, is called injective if there exists a projection Φ : B(H) → B(H) with range equal to V. A linear map Φ : A → A is called a conditional expectation if it is positive, idempotent, and it has the following bimodule property: Φ(ar) = Φ(a)r and Φ(ra) = rΦ(a), for all a ∈ A and all r ∈ Ran(Φ). By a classical result of J. Tomyama [31], a C ∗-algebra A ⊆ B(H) is injective if and only if there is a conditional expectation in B(H) with range equal to A. For a semigroup Φ = {Φt}t≥0 of unital, completely positive maps on a C ∗-algebra M, we consider MΦ the set of joint fixed points of Φ, that is, (2.25) MΦt = {a ∈ M Φt(a) = a, for all t ≥ 0}, MΦ = [t≥0 see Subsection 2.1, which is an operator system, and the joint bimodule domain (2.26) I(Φ) =\t≥0 I(Φt) = {a ∈ M Φt(ab) = aΦt(b), Φt(ba) = Φt(b)a, for all b ∈ A and all t ≥ 0}, which is clearly a C ∗-subalgebra of M and included in MΦ. In case M is a von Neu- mann algebra and each Φt is w∗-continuous, MΦ is w∗-closed and I(Φ) is a von Neumann subalgebra of M. Theorem 2.6. Let M be a von Neumann algebra and Φ = {Φt}t≥0 be a w∗-continuous semigroup of w∗-continuous, unital, completely positive maps on M. Then: ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 11 (a) There exists a completely positive, unital, and idempotent map Ψ : M → M such that the set of joint fixed points MΦ is the range of Ψ. (b) The following assertions are equivalent: (i) MΦ is stable under multiplication. (ii) MΦ is a von Neumann algebra. (iii) MΦ = I(Φ). (iv) Ψ is a conditional expectation. (c) If M = B(H) and B(H)Ψ is stable under multiplication, then B(H)Ψ is an injective von Neumann algebra. Proof. (a) For each real number t > 0, let Ψt : M → M be defined by (2.27) Φsds. Ψt = 1 t Z t 0 The integral converges with respect to the point-w∗-topology, that is, for all a ∈ M and all f ∈ M∗, we have hΨt(a), fi = 1 t Z t 0 hΦs(a), fids. It is easy to see that Ψt is w∗-continuous, unital, and completely positive and hence, by Russo -- Dye's Theorem, a completely contractive map for each t > 0. By the Alaoglu's Theo- rem, the closed unit ball of M is w∗-compact, hence by Tyhonov's Theorem the closed unit ball of CB(M) is compact with respect to the point-w∗-topology. Consequently, considering the sequence {Ψn}n∈N, there exists a subsequence {Ψkn}n∈N such that w∗- lim n→∞ Ψkn(a) = Ψ(a), a ∈ M, for some linear map Ψ : M → M. Clearly, Ψ is unital and completely positive. Let t ≥ 0 be an arbitrary real number and n ∈ N be large enough such that t ≤ n. Then Ψn − ΦtΨn = 0 t 0 0 0 1 1 = Φsds −Z n Φsds −Z t+n Φsds −Z t+n n(cid:0)Z n n(cid:0)Z n n(cid:0)Z t n(cid:0)Z t 0 kΦskds −Z t+n Φt+sds(cid:1) Φsds(cid:1) Φsds(cid:1) kΦskds(cid:1) = = 1 1 n n hence (2.28) kΨn − ΦtΨnk ≤ 2t n −−−→n→∞ 0. On the other hand, using the representation ΦtΨ − Ψ = (ΦtΨ − ΦtΨkn) + (ΦtΨkn − Ψkn) + (Ψkn − Ψ), n ∈ N, (2.29) and taking into account that, for all a ∈ M, by the defining property of the subsequence (Ψkn)n∈N, we have (ΦtΨ − ΦtΨkn)(a) = Φt(Ψ(a) − Ψkn(a)) w∗ −−−→n→∞ 0, 12 A. GHEONDEA and then of (2.28), it follows that ΦtΨ = Ψ, for all t ≥ 0. Similarly we obtain ΨΦt = Ψ for all t ≥ 0, hence (2.30) From (2.30) we get Ψkn(Ψ(a)) = 0 ΦtΨ = ΨΦt = Ψ, for all t ≥ 0. knZ kn Φs(Ψ(a))ds = Ψ(a), 1 a ∈ M, n ∈ N, and then letting n → ∞ it follows that ΨΨ = Ψ, hence Ψ is an idempotent. If a ∈ MΦ is arbitrary, then Ψkn(a) = a for all n ∈ N whence, letting n → ∞ it follows Ψ(a) = a. We have proven that MΦ ⊆ Ran(Ψ). Since, by (2.30), Ran(Ψ) ⊆ MΦ, we have MΦ = Ran(Ψ). (b) Only the equivalence of (i) and (iv) requires a proof. Assume firstly that MΦ is stable under multiplication. By the result at item (a), it follows that Ran(Ψ) = MΦ is a von Neumann algebra. Then, for arbitrary a ∈ Ran(Ψ), Ψ(a)∗Ψ(a) = a∗a = Ψ(a∗a), Ψ(a)Ψ(a)∗ = aa∗ = Ψ(aa∗), hence, by Theorem 2.1, for any b ∈ M we have Ψ(ab) = Ψ(a)Ψ(b) = aΨ(b), Ψ(ba) = Ψ(b)Ψ(a) = Ψ(b)a, consequently Ψ is a conditional expectation. stable under multiplication. Conversely, if Ψ is a conditional expectation then MΦ = Ran(Ψ) is a C ∗-algebra, hence (c) This is now a consequence of the results proven at item (a) and item (b). (cid:3) 3. Dynamical Systems of Stochastic/Markov Maps: The Noncommutative Case Let H be a Hilbert space, let B(H) be the von Neumann algebra of all bounded linear operators T : H → H and let B1(H) be the trace-class, that is, the collection of all operators T ∈ B(H) subject to the condition kTk1 = tr(T) < +∞, where T = (T ∗T )1/2 denotes the module of T and tr denotes the usual normal faithful semifinite trace on B(H). Let D(H) denote the set of states, or density operators, with respect to H, that is, the set of all positive elements ρ ∈ B1(H) with tr(ρ) = kρk1 = 1. A linear map Ψ : B1(H) → B1(H) is called stochastic if it maps states into states, equiv- alently, if it is positive, that is, Ψ(A) ≥ 0 for all A ∈ B1(H)+, and trace-preserving, that is, tr(Ψ(T )) = tr(T ) for all T ∈ B1(H). The map Ψ : B1(H) → B1(H) is called a quantum operation, if it is completely positive, see Subsection 2.1 for definition, and trace-preserving. Note that, the trace-class B1(H) is considered here as a ∗-subspace of the C ∗-algebra B(H) and, consequently, the concept of completely positive map on B1(H) makes perfectly sense. Clearly, any quantum operation is a stochastic map. We note that the definition of a quantum operation we adopt here is a bit more restrictive than usual. In quantum information theory they use the term of a quantum communication channel, or briefly a quantum channel, for what we call here a quantum operation. For a fixed Banach space X, recall that we denote its topological dual space by X ♯ and the duality map by X × X ♯ ∋ (x, f ) 7→ hx, fi, see Subsection 2.2. The topics of this article refer to the Banach space (B1(H),k · k1) and its topological dual Banach space (B(H),k · k) with the duality map B1(H) × B(H) ∋ (T, S) 7→ hT, Si = tr(T S), e.g. see Theorem 19.2 in ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 13 [11]. In particular, for a quantum operation Ψ when viewed as a trace-preserving completely positive map Ψ : B1(H) → B1(H), one usually refers to the Schrodinger picture, to which the Heisenberg picture is corresponding by duality: the dual map Ψ♯ : B(H) → B(H) is defined by hΨ(T ), Si = tr(Ψ(T )S) = tr(T Ψ♯(S)) = hT, Ψ♯(S)i, T ∈ B1(H), S ∈ B(H), k=1 AkA∗ k = I then the linear map B1(H) ∋ T 7→P∞ and it is a ultraweakly continuous (w∗-continuous) completely positive and unital linear map. Similarly, if Ψ is a stochastic linear map then its dual Ψ♯ is a ultraweakly continuous positive and unital linear map on B(H), called a Markov map. There are many quantum operations. For example, if {Ak k ∈ N} is a collection of oper- ators in B(H) such that P∞ kT Ak ∈ B1(H) is a quantum operation. The following example shows that there exist stochastic maps that are not quantum operations. The idea of using the transpose map for this kind of examples can be tracked back to W.B. Arveson [3], [4]. Stochastic maps that are not quan- tum operations, in particular the transpose map, play an important role in entanglement detectors in quantum information theory, e.g. see D. Chruscinski and A. Kossakowski [12], R. Horodecki et al. [22] and the rich bibliography cited there. Example 3.1. Let H be an arbitrary Hilbert space with dimension at least 2, for which we fix an orthonormal basis {ej}j∈J . We consider the conjugation operator J : H → H defined by Jh = h where, for arbitrary h =Pj∈J hjej, we let h =Pj∈J hjej. Then J is conjugate linear, conjugate selfadjoint, that is, it has the following property k=1 A∗ (3.1) isometric, and J 2 = I. hJh, ki = hJk, hi, h, k ∈ H, Further on, let τ : B(H) → B(H) be defined by τ (S) = JS ∗J, for all T ∈ B(H). It is easy to see that τ is isometric, that is, kτ (S)k = kSk for all S ∈ B(H), and that τ (I) = I. On the other hand, if S ∈ B(H)+ then hτ (S)h, hi = hJSJh, hi = hJh, SJhi = hSJh, Jhi ≥ 0, h ∈ H, hence τ is positive. Let us also observe that, with respect to the matrix representation of operators in B(H) associated to the orthonormal basis {ej}j∈J , τ is the transpose map: if T has the matrix representation [ti,j]i,j∈J then τ (T ) has the matrix representation [tj,i]j,i∈J . We claim now that τ leaves B1(H) invariant and the corresponding restriction map B1(H) → B1(H) is stochastic. To see this, we first observe that if T ∈ B1(H)+ we have τ (T ) ∈ B1(H)+, e.g. using that τ is the transpose map with respect to the matrix representations of opera- tors in B1(H) associated to the orthonormal basis {ej}j∈J , and the definition of the trace in terms of any orthonormal basis of H. Also, kτ (T )k1 = tr(τ (T )) = tr(T ) = kTk1. Since any operator T ∈ B1(H) is a linear combination of four positive trace-class operators, the claim follows. Finally, we show that τ is not completely positive, more precisely, it is not 2-positive. To see this, we consider the matrix units {Ei,j}i,j∈J , that is, for any i, j ∈ J , Ei,j denote the rank 1 operator on H with Ei,jej = ei and Ei,jek = 0 for all k 6= j and observe that τ (Ei,j) = Ej,i. Since dimH ≥ 2, there exist i, j ∈ J with i 6= j. Then, consider the positive finite rank operator in M2(B1(H)) defined by E =(cid:20)Ei,i Ei,j Ej,i Ej,j(cid:21) 14 A. GHEONDEA and observe that τ2(E) =(cid:20)τ (Ei,i) τ (Ej,i) τ (Ej,j)(cid:21) =(cid:20)Ei,i Ej,i Ei,j Ej,j(cid:21) τ (Ei,j) which is not positive, e.g. see [26], p. 5. Therefore, τ is a stochastic map but not a quantum operation. Remarks 3.2. (1) By means of the matrix transpose interpretation of τ as in Example 3.1, it follows easily that its dual τ ♯ : B(H) → B(H) has the same formal definition: τ (S) = JS ∗J, for all S ∈ B(H), and the same matrix transpose interpretation with respect to a fixed orthonormal basis of H. (2) The stochastic map τ described in Example 3.1 is invertible, τ −1 = τ , and antimulti- plicative, that is, τ (ST ) = τ (T )τ (S) for all S, T ∈ B1(H). The same properties are shared by its dual τ ♯. In particular, both τ and τ ♯ are ∗-antihomomorphisms. (3) In addition to the map τ described in Example 3.1, many other stochastic maps that are not quantum operations can be obtained by considering convex combinations of linear maps of type τ ◦ Ψ or Ψ ◦ τ , where Ψ are quantum operations. 3.1. Discrete Semigroups of Stochastic/Markov Maps. From the quantum measure- ments point of view, given a quantum operation Ψ, it is of interest to characterise those elements A ∈ B(H) with the property that [Ψ, MA] = 0, that is, Ψ(A∗XA) = A∗Ψ(X)A for all X ∈ B1(H), where MA : B1(H) → B1(H) denotes the one-element measurement, that is, the linear map MA(X) = A∗XA for all X ∈ B1(H) and the commutator is defined as usually [Φ, Ψ] = ΦΨ − ΨΦ. Note that, since B1(H) is a two-sided ideal of B(H), MA can be defined either as a linear map B1(H) → B1(H) or as a linear map B(H) → B(H). Actually, if we consider MA : B1(H) → B1(H) then its dual map M ♯ A : B(H) → B(H) is the one-element measurement map MA∗. A sequence {Ψn}n≥0 is called a discrete stochastic semigroup if (qs1) Ψn : B1(H) → B1(H) is a stochastic operator for all integer n ≥ 0. (qs2) Ψn+m = ΨnΨm for all integer m, n ≥ 0. (qs3) Ψ0 = I. Clearly, to any stochastic operator Ψ one associates the discrete semigroups {Ψn}n≥0 and, conversely, any discrete semigroup of stochastic operators {Ψn}n≥0 is fully determined by Ψ = Ψ1 and Ψn = Ψn, for all integer n ≥ 0. Consequently, the analysis of discrete stochastic semigroups pertains to the analysis of one stochastic operator. Remark 3.3. Let Ψ : B1(H) → B1(H) be a bounded linear map and A ∈ B(H). Then [Ψ, MA] = 0 if and only if [Ψ♯, MA∗] = 0. The one-element measurement operator MA is usually associated to a positive operator A. In this case, one rather considers the one-element measurement in the Luders form MA1/2 for some positive operator A. We show that the following theorem, obtained in [7], can be recovered as a rather direct application of Corollary 2.3. Theorem 3.4 ([7]). Let Ψ be a stochastic map on the Hilbert space H and A ∈ B(H)+. The following assertions are equivalent: (i) [Ψ, MA1/2] = 0, that is, Ψ(A1/2T A1/2) = A1/2Ψ(T )A1/2 for all T ∈ B1(H). (ii) [Ψ♯, MA1/2] = 0, that is, Ψ♯(A1/2SA1/2) = A1/2Ψ♯(S)A1/2 for all S ∈ B(H). ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 15 (iii) Ψ♯(A) = A and Ψ♯(A2) = A2. Before proceeding to the proof of this theorem, we prove two preliminary results. The first one is essentially Remark 5.4 in [7] for which we provide a coordinate free proof. Lemma 3.5. If E is a projection and C ∈ B1(H)+ such that tr(C) = tr(ECE) then C = CE = EC. Proof. Taking into account that C 1/2EC 1/2 ≤ C and that 0 ≤ tr(C − C 1/2EC 1/2) = tr(C) − tr(C 1/2EC 1/2) = tr(C) − tr(ECE) = 0, it follows that C = C 1/2EC 1/2 hence, 0 = C 1/2(I − E)C 1/2 = C 1/2(I − E)(I − E)C 1/2 = ((I − E)C 1/2)∗((I − E)C 1/2), which implies (I − E)C 1/2 = 0 hence (I − E)C = 0. From here it follows EC = C and then taking adjoints we have CE = C as well. (cid:3) The second preliminary result is a short-cut of Corollary 5.5, Corollary 5.6, and Lemma 5.7 in [7]. Lemma 3.6. Let Ψ be a stochastic map with respect to a Hilbert space H and let E be a projection such that Ψ♯(E) = E. Then (i) Ψ(ET E) = EΨ(ET E) = Ψ(ET E)E for all T ∈ B1(H). (ii) EΨ♯(ESE) = Ψ♯(ESE)E = Ψ♯(ESE) for all S ∈ B(H). (iii) Ψ♯(ESE) = EΨ♯(S)E for all S ∈ B(H). Proof. (i) It is sufficient to prove this for all T ∈ B1(H)+. With this assumption, we have tr(EΨ(ET E)E) = tr(EΨ(ET E)) = hE, Ψ(ET E)i = hΨ♯(E), ET Ei = hE, ET Ei = tr(ET E) = tr(Ψ(ET E)), and, consequently, applying Lemma 3.5 for C = Ψ(ET E), the conclusion follows. (ii) To see this, without loss of generality it is sufficient to assume that S ∈ B(H)+ is a contraction, that is, 0 ≤ S ≤ I. Then 0 ≤ ESE ≤ E hence 0 ≤ Ψ♯(ESE) ≤ Ψ♯(E) = E, which implies that the range of Ψ♯(ESE) is contained in the range of E. This implies EΨ♯(ESE) = Ψ♯(ESE) and then, by taking adjoints, we have Ψ♯(ESE)E = Ψ♯(ESE) as well. (iii) Let T ∈ B1(H) and S ∈ B(H) be arbitrary. Using assertion (ii) we have hΨ♯(ESE), Ti = hEΨ♯(ESE)E, Ti = hΨ♯(ESE), ET Ei = hESE, Ψ(ET E)i = hS, EΨ(ET E)Ei and then, using assertion (i), we have hence assertion (iii) follows. (cid:3) = hS, Ψ(ET E)i = hΨ♯(S), ET Ei = hEΨ♯(S)E, Ti, 16 A. GHEONDEA Proof of Theorem 3.4. (i)⇔(ii). This is a consequence of Remark 3.3. A1/2A1/2 = A and then Ψ♯(A2) = Ψ♯(A1/2AA1/2) = A1/2Ψ♯(A)A1/2 = A1/2AA1/2 = A2. (ii)⇒(iii). Since Ψ♯ is unital it follows that Ψ♯(A) = Ψ♯(A1/2IA1/2) = A1/2Ψ♯(I)A1/2 = (iii)⇒(ii). Letting Ψ♯ = Φ in Corollary 2.3, it follows that Ψ♯(S) = S for all S ∈ C ∗(I, A). Since Ψ♯ is w∗-continuous, by functional calculus with bounded Borel functions on σ(A), it follows that Ψ♯(S) = S for all S ∈ W ∗(A), the von Neumann algebra generated by A In particular, for any spectral projection E of A we have Ψ♯(E) = E. From in B(H). Lemma 3.6 it follows (3.2) Ψ♯(ESE) = EΨ♯(S)E, S ∈ B(H). From here, by the Spectral Theorem for A, it follows that for any function f that is continuous on σ(A) we have (3.3) Letting f (t) = g(t) = √t, t ∈ σ(A), the assertion follows. Ψ♯(f (A)Sf (A)) = f (A)Ψ♯(S)f (A), S ∈ B(H). (cid:3) Remarks 3.7. (1) Under the assumptions of Theorem 3.4, from the proof provided here and Remark 3.3, one can easily obtain the following assertions that are mutually equivalent with each of assertions (i) -- (iii), cf. [7]: (iv) [ME, Ψ] = 0 for any spectral projection E of A. (v) [ME, Ψ♯] = 0 for any spectral projection E of A. (vi) [Mf (A), Ψ] = 0 for any real function f continuous on σ(A). (vii) [Mf (A), Ψ♯] = 0 for any real function f continuous on σ(A). (2) The mutually equivalent assertions as in Theorem 3.4 can be, equivalently, written in terms of the discrete dynamical system {Ψn}n∈N0: (i) [Ψn, MA1/2] = 0 for all n ∈ N0. (ii) [Ψ♯n, MA1/2] = 0 for all n ∈ N0. (iii) Ψ♯n(A) and Ψ♯n(A2) do not depend on n ∈ N0. This way, assertions (i) and (ii) are symmetry properties while assertion (iii) is a conservation law. (3) A natural question related to Theorem 3.4 is whether the latter condition in item (iii) on A2 being fixed by Ψ♯ is really necessary for a given stochastic map Ψ. It is interesting that, for the transpose map τ as in Example 3.1 the answer is no. More precisely, let A ∈ B(H)+ be a fixed point of τ ♯. Since A is positive and taking into account that τ ♯ is antimultiplicative, see Remark 3.2.(2), it follows that τ ♯(A2) = τ ♯(A)τ ♯(A) = A2. However, the answer to this question is positive, in general, for quantum operations, see Remark 5.5.(2), and hence for stochastic maps as well. 3.2. Continuous One-Parameter Semigroups of Stochastic/Markov Maps. With notation as in the previous section, we consider a strongly continuous one-parameter semi- group Ψ = {Ψt}t≥0 of stochastic maps with respect to some Hilbert space H. Under these assumptions, we observe that {Ψt}t≥0 is uniformly bounded on B1(H). Most of the fol- lowing facts that we briefly recall refer to a particular situation of the general theory of one-parameter semigroup theory on Banach spaces, e.g. see [21] and [14], see Subsection 2.2. Given a strongly continuous semigroup Ψ = {Ψt}t≥0 of stochastic maps with respect to some Hilbert space H, the infinitesimal generator ψ exists as a densely defined closed operator on ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 17 t}t≥0 of Markov maps exists, that is, B1(H). For every strongly continuous one-parameter semigroup Ψ = {Ψt}t≥0 of stochastic maps, the dual one-parameter semigroup Ψ♯ = {Ψ♯ (3.4) hΨt(T ), Si = tr(Ψt(T )S) = tr(T Ψ♯ t(S)i, T ∈ B1(H), S ∈ B(H), t ≥ 0. Then {Ψ♯ t}t≥0 is a w∗-continuous semigroup of contractions on B(H) and hence, the w∗- infinitesimal generator ψ♯ exists as a w∗-closed operator on B(H), hence a closed operator on B(H). The w∗-infinitesimal generator ψ♯ of the dual w∗-continuous semigroup {Ψ♯ t}t≥0 of Markov maps is indeed the dual operator of the infinitesimal generator ψ of the strongly continuous semigroup {Ψt}t≥0 of stochastic maps and, consequently, the notation for ψ♯ is fully justified. t(S)) = hT, Ψ♯ Also, let us observe that, since Ψ♯ (3.5) t(I) = I, it follows that I ∈ Dom(ψ♯) and ψ♯(I) = 0. In addition, one of the major differences between the two infinitesimal generators ψ and ψ♯ is that Dom(ψ♯) may not be dense in B(H), although it is always w∗-dense, while Dom(ψ) is always dense in B1(H). The equivalence of (i) -- (iv) in the following theorem has been obtained in [7]. We add two more equivalent characterisations in terms of the dual infinitesimal generator, which actually make the proofs simpler. Theorem 3.8. Let Ψ = {Ψt}t≥0 be a strongly continuous one-parameter semigroup of sto- chastic maps on B1(H), ψ its infinitesimal generator, and let A ∈ B(H)+. With notation as before, the following assertions are equivalent: t(A) = A and Ψ♯ t(A2) = A2 for all t ≥ 0. (i) Ψ♯ (ii) [MA1/2, Ψt] = 0 for all t ≥ 0. (iii) [MA1/2, Ψ♯ t] = 0 for all t ≥ 0. (iv) [MA1/2, ψ] = 0 that is, for all T ∈ Dom(ψ) we have A1/2T A1/2 ∈ Dom(ψ) and (v) [MA1/2, ψ♯] = 0 that is, for all S ∈ Dom(ψ♯) we have A1/2SA1/2 ∈ Dom(ψ♯) and (vi) A, A2 ∈ Ker(ψ♯), that is, A, A2 ∈ Dom(ψ♯) and ψ♯(A) = ψ♯(A2) = 0. ψ♯(A1/2T A1/2) = A1/2ψ♯(T )A1/2. ψ(A1/2T A1/2) = A1/2ψ(T )A1/2. Proof. The equivalence of the assertions (i), (ii), and (iii) is a straightforward consequence of Theorem 3.4. (ii)⇒(iv). For arbitrary T ∈ Dom(ψ) and t ≥ 0, we have Ψt(A1/2T A1/2) − A1/2T A1/2 t A1/2Ψt(T )A1/2 − A1/2T A1/2 = t = A1/2 Ψt(T ) − T t A1/2 −−−→t→0+ A1/2ψ(T )A1/2, hence A1/2T A1/2 ∈ Dom(ψ) and ψ(A1/2T A1/2) = A1/2ψ(T )A1/2. and ψ(A1/2T A1/2) = A1/2ψ(T )A1/2, hence (iv)⇒(v). Let S ∈ Dom(ψ♯). Then, for any T ∈ Dom(ψ) we have A1/2T A1/2 ∈ Dom(ψ) hψ(T ), A1/2SA1/2i = hA1/2ψ(T )A1/2, Si = hψ(A1/2T A1/2), Si 18 A. GHEONDEA whence, taking into account of the continuity of the map B1(H) ∋ T 7→ A1/2T A1/2 ∈ B1(H), it follows that A1/2SA1/2 ∈ Dom(ψ♯). Consequently, hT, ψ♯(A1/2SA1/2)i = hψ(T ), A1/2SA1/2i = hψ(A1/2T A1/2), Si = hA1/2T A1/2, ψ♯(S)i = hT, A1/2ψ♯(S)A1/2i, hence, ψ♯(A1/2SA1/2) = A1/2ψ♯(S)A1/2. Then, A2 = A1/2AA1/2 ∈ Dom(ψ♯) and ψ♯(A2) = A1/2ψ♯(A)A1/2 = 0. (v)→(vi). By (3.5) we have A = A1/2IA1/2 ∈ Dom(ψ♯) and ψ♯(A) = A1/2ψ♯(I)A1/2 = 0. (vi)⇒(i). This is a consequence of Theorem 2.5. (cid:3) Remarks 3.9. (a) Under the assumptions of Theorem 3.8, each of the assertions (i) -- (vi) is equivalent with each of the following assertions, cf. [7]: (vii) d (viii) d (ix) For every spectral projection E of A we have [MA1/2, ψ] = 0, that is, for any T ∈ dthΨt(T ), Ai = d dthΨt(T ), A2i = 0 for all T ∈ B1(H). dthΨt(T ), Ani = 0 for all T ∈ B1(H) and all n ≥ 0. Dom(ψ) we have ET E ∈ Dom(ψ) and ψ(ET E) = Eψ(T )E. The equivalence of assertion (ix) is short-cut in our proof but it is an important step dur- ing the proof provided in [7]. Assertion (vii) is clearly equivalent with assertion (i), while assertion (viii) is equivalent with assertion (vii) in view of Corollary 2.3. (b) A natural question is whether the condition that A2 is a joint fixed point of Ψ, as in Theorem 3.8.(i), is a consequence of the condition that A is a joint fixed point of Ψ. The answer is negative, in general, and it will be obtained as a consequence of Theorem 5.12. 4. Dynamics for Markov Processes: The Real Commutative Case In this section we consider the setting of dynamics of Markov processes in the framework of "stochastic mechanics" in the sense of [6] and [5]. Let (X; µ) be a σ-finite measure space. A probability distribution p is an element in L1 R(X; µ) which is positive and kpk1 = 1. An observable O is an element in L∞ R (X; µ), identified with the operator of multiplication O : L1 R(X; µ) → L1 R(X; µ) (Og)(x) = O(x)g(x), g ∈ L1 R(X; µ), x ∈ X. The expected value of the observable O with respect to a probability distribution g is E(O; g) = hO, gi =ZX O(x)g(x)dµ(x), the variance of O with respect to g is while the standard deviation of O with respect to g is V (O; g) = hO2, gi − hO, gi2, σ(O; g) =phO2, gi − hO, gi2. A stochastic operator is a bounded linear operator U : L1 R(X; µ) → L1 R(X; µ) that maps probability distributions to probability distributions, equivalently, U is positive, that is, if g ∈ L1 R(X; µ) and g ≥ 0 then Ug ≥ 0, ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 19 and ZX (Ug)(x)dµ(x) =ZX g(x)dµ(x), for all g ∈ L1 R(X; µ). The latter condition can also be written as h1, Ugi = h1, gi, R(X; µ). g ∈ L1 R (X; µ) is called a Markov map if it is w∗- R (X; µ) with f ≥ 0 it follows T f ≥ 0, R(X; µ) there exists its dual operator A bounded linear operator T : L∞ R (X; µ) → L∞ Given any bounded linear operator U : L1 continuous, positive, in the sense that for any f ∈ L∞ and unital, that is, T 1 = 1. R(X; µ) → L1 (Ug)(x)f (x)dµ(x) =ZX hUg, fi =ZX R (X; µ) → L∞ U ♯ : L∞ R (X; µ), which is linear and bounded, defined by g(x)(U ♯f )(x)dµ(x) = hg, U ♯fi, f ∈ L1 In addition, U ♯ is w∗-continuous. If U : L1 its dual U ♯ : L∞ R(X; µ), g ∈ L∞ R(X; µ) → L1 R (X; µ) is a Markov operator. R (X; µ) → L∞ R (X; µ). R(X; µ) is a stochastic operator then 4.1. Discrete Stochastic/Markov Semigroups. A discrete stochastic semigroup with respect to the measure space (X; µ) is a sequence {Un}n≥0 subject to the following conditions: (ms1) Un : L1 (ms2) Un+m = UnUm for all n, m ≥ 0. (ms3) U0 = I. R(X; µ) is stochastic for all n ≥ 0. R(X; µ) → L1 Clearly, any discrete stochastic semigroup is of the form Un = U n, n ≥ 0, R (X; µ) → R R(X; µ), which is actually a Markov operator, we can equivalently discuss of discrete where U = U1 is a stochastic operator. Considering the dual operator U ♯ : L∞ L∞ Markov semigroups. The equivalence of assertions (i), (ii), (i)′, and (ii)′ in the following theorem has been obtained in [6], for which we provide a proof based on the results in Subsection 2.1, as well as complete their theorem with two more equivalent assertions in terms of duals of stochastic operators. Theorem 4.1. Let (X; µ) be a σ-finite measure space, U : L1 operator and O ∈ L∞ R (X; µ) an observable. The following assertions are equivalent: R(X; µ) → L1 R(X; µ) a stochastic (i) [O, U] = 0. (ii) For any probability distribution g on X we have hO, Ugi = hO, gi and hO2, Ugi = (i)′ [O, U n] = 0 for all n ≥ 0. (ii)′ For any probability distribution g on X, the expected values of O and O2 with respect hO2, gi. to U ng do not depend on n ≥ 0. (i)′′ [O, U ♯] = 0. (ii)′′ U ♯(O) = O and U ♯(O2) = O2. C(X; µ) → L1 C(X; µ) = L1 R(X; µ) ⊕ iL1 R(X; µ), L1 C(X; µ) → L1 C(X; µ) by we can define eU : L1 and observe that eU has the following two properties: eU (g + if ) = Ug + iUf, if g ∈ L1 f, g ∈ L1 R(X; µ), 20 A. GHEONDEA Proof. The equivalences (i)⇔(i)′, (ii)⇔(ii)′, (i)⇔(i)′′, and (ii)⇔(ii)′′ are clear. R (X; µ) we have OU ♯(f ) = U ♯(Of ). Letting f = 1 and taking into account that U ♯(1) = 1 it follows U ♯(O) = O and then letting f = O we have U ♯(O2) = OU ♯(O) = O2. (i)′′ ⇒(ii)′′. Assume that [O, U ♯] = 0 hence, for any f ∈ L∞ (ii)′′ ⇒(i)′′. The spaces L1 and, respectively, in L∞ complex stochastic operator U : L1 R(X; µ) and L∞ C(X; µ) C (X; µ). The real stochastic operator U can be naturally lifted to a R (X; µ) are naturally embedded in L1 C(X; µ). More precisely, since and ZX C(X; µ) and g ≥ 0 then eU g ≥ 0, for all g ∈ L1 C (X; µ) is unital and positive. Since L∞ (eUg)(x)dµ(x) =ZX g(x)dµ(x), C(X; µ). L∞ C (X; µ) is a commutative C (X; µ) → L∞ On the other hand, the observable O can be naturally viewed as a real valued function in Then eU ♯ : L∞ C ∗-algebra, eU ♯ is completely positive, cf. [29]. C (X; µ) and, if U ♯(O) = O and U ♯(O2) = O2, it follows that eU ♯(O) = O and eU ♯(O2) = O2. Now we can use Theorem 2.4 and conclude that eU ♯(Of ) = OeU ♯(f ) for all f ∈ L∞ hence [O,eU] = 0 and then [O, U] = 0. 4.2. Continuous Stochastic/Markov Semigroups. A continuous stochastic semigroup on (X; µ) is a strongly continuous semigroup of stochastic operators on L1 R(X; µ). The infinitesimal generator of {Ut}t≥0 is the closed and densely defined operator H in L1 R(X; µ), see Subsection 2.2. Let {Ut}t≥0 be a continuous stochastic semigroup with respect to (X; µ) and H its infinitesimal generator. Then {U ♯ t}t≥0 is a w∗-continuous semigroup of Markov maps. The w∗-infinitesimal generator of {U ♯ t}t≥0 is the w∗-closed, hence closed, and w∗- densely defined (but, in general, not densely defined) operator H ♯ in L∞ R (X; µ) which, by Phillips Theorem [27], can be described by C C(X; µ), (cid:3) H ♯f = w∗ − lim t→0+ Utf − f t , f ∈ Dom(H ♯), where Dom(H ♯) = {f ∈ L∞ R (X; µ) w∗ − lim t→0+ U ♯ t f − f t exists in L∞ R (X; µ)}. The equivalence of assertions (i) and (ii) in the next theorem has been obtained in [6], which we now obtain as a consequence of Theorem 2.4, via Theorem 4.1. We complete their theorem with four more equivalent assertions in terms of infinitesimal generators and dual. The proofs are very similar with those in Theorem 3.8 and we omit repeating the arguments, in particular, the equivalence of assertions (ii)′ and (iii)′ follows from Theorem 2.5. ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 21 Theorem 4.2. Let (X; µ) be a σ-finite measure space, {Ut}t≥0 a continuous stochastic semi- group with respect to (X; µ), H its infinitesimal generator, and O ∈ L∞ R (X; µ) an observable. The following assertions are equivalent: (i) [O, Ut] = 0 for all real t ≥ 0. (ii) For every probability distribution g on (X; µ), the expected values hO, Utgi and hO2, Utgi (iii) [O, H] = 0, in the sense that the operator of multiplication with O leaves Dom(H) are constant with respect to t ≥ 0. invariant and OHg = HOg for all g ∈ Dom(H). t (O) = O and U ♯ t ] = 0 for all real t ≥ 0. (i)′ [O, U ♯ (ii)′ U ♯ (iii)′ Both O and O2 are in the kernel of H ♯, that is, O, O2 ∈ Dom(H ♯) and H ♯(O) = t (O2) = O2 for all real t ≥ 0 H ♯(O2) = 0. 5. Constants of Dynamical Quantum Systems We now consider the setting of dynamical quantum systems as in [19]. Notation is as in Section 3. For a fixed Hilbert space H and A ∈ B(H) we have the left multiplication operator LA : B1(H) → B1(H) defined by LA(T ) = AT , for all T ∈ B1(H), and the right multiplication operator RA : B1(H) → B1(H) defined by RA(T ) = T A, for all T ∈ B1(H). Observe that, exactly with the same formal definition, we may have the left multiplication operator LA : B(H) → B(H) and, respectively, RA : B(H) → B(H). We will not use different notation for these operators, hoping that which is which will be clear from the context. For example, if LA : B1(H) → B1(H) then L♯ A : B(H) → B(H) is the operator RA : B(H) → B(H). Also, considering MA(T ) = A∗T A, the one-element quantum measurement operator, then MA = LA∗RA. We distinguish between the discrete quantum semigroups and continuous quantum semi- groups. 5.1. Discrete Quantum Semigroups. We consider a quantum operation Ψ : B1(H) → B1(H), that is, a trace preserving completely positive linear map. It gives rise naturally to the discrete quantum semigroup {Ψn}n≥0. In order to substantiate the definition of a constant of a discrete quantum semigroup {Ψn}n≥0 we first recall some natural definitions from quantum probability. Let A be a bounded observable with respect to the Hilbert space H, that is, A ∈ B(H) and A = A∗. For any state ρ ∈ D(H) one considers the expected value of A in the state ρ, (5.1) E(A; ρ) = hρ, Ai = tr(ρA), the variation of A in the state ρ, and its standard deviation, (5.2) (5.3) V (A; ρ) = hρ, A2i − hρ, Ai2 = tr(ρA2) − tr(ρA)2, σ(A; ρ) =phρ, A2i − hρ, Ai2 =ptr(ρA2) − tr(ρA)2. An operator A ∈ B(H) is called a constant of the discrete quantum semigroup {Ψn}n≥0, equivalentely, of Ψ, if for any state ρ ∈ D(H), tr(Ψn(ρ)A) does not depend on n ≥ 0, equivalently, tr(Ψ(ρ)A) = tr(ρA). Clearly, A is a constant of Ψ if and only if for any T ∈ B1(H) we have tr(Ψ(T )A) = tr(T A), equivalently, tr(T Ψ♯(A)) = tr(T A) for all T ∈ B1(H). 22 A. GHEONDEA Consequently, A ∈ B(H) is a constant of Ψ if and only if Ψ♯(A) = A, that is, A is a fixed point of Ψ♯. Formally, letting CΨ denote the set of constants of Ψ CΨ = {A ∈ B(H) for all ρ ∈ D(H), tr(Ψn(ρ)A) does not depend on integer n ≥ 0} (5.4) = {A ∈ B(H) tr(Ψ(ρ)A) = tr(ρA) for all ρ ∈ D(H)} we have, = {A ∈ B(H) Ψ♯(A) = A} = B(H)Ψ♯ , where the last equality is actually the definition of B(H)Ψ♯ as the set of all fixed points of Ψ♯, as in Subsection 2.1. We have now a first Noether Type Theorem for a discrete dynamical quantum system, in a spirit closer to [6]. Theorem 5.1. Let Ψ be a quantum operation with respect to the Hilbert space H and let A ∈ B(H). The following assertions are equivalent: (i) [LA, Ψ] = 0. (ii) A and A∗A are constants of Ψ. (iii) [RA, Ψ♯] = 0. (iv) A and A∗A are fixed points of Ψ♯. Proof. (i)⇔(iii) and (ii)⇔(iv) are clear. (iii)⇒(iv). If [RA, Ψ♯] = 0 then Ψ♯(SA) = Ψ♯(S)A for all S ∈ B(H). Letting S = I we get Ψ♯(A) = A and, since Ψ♯ is positive, hence selfadjoint, it follows that Ψ♯(A∗) = A∗. Then, letting S = A∗ we get Ψ♯(A∗A) = Ψ♯(A∗)A = A∗A. (iv)⇒(iii). Assume that Ψ♯(A) = A and Ψ♯(A∗A) = A∗A. Then Ψ♯(A∗) = A∗ and Ψ♯(A∗A) = A∗A = Ψ♯(A∗)Ψ(A). By Theorem 2.1 it follows that Ψ♯(T A) = Ψ♯(T )Ψ♯(A) = Ψ♯(T )A for all T ∈ B(H), hence [RA, Ψ♯] = 0. (cid:3) Clearly, there is a symmetric analogue of Theorem 5.1. Theorem 5.2. Let Ψ be a quantum operation with respect to the Hilbert space H and let A ∈ B(H). The following assertions are equivalent: (i) [RA, Ψ] = 0. (ii) A and AA∗ are constants of the discrete quantum semigroup {Ψn}n≥0. (iii) [LA, Ψ♯] = 0. (iv) A and AA∗ are fixed points of Ψ♯. We now consider the case of a bounded observable A ∈ B(H)+, as in (5.1) through (5.3), and reformulate Theorem 5.1 and its symmetric Theorem 5.2 to a noncommutative analogue of the Noether Type Theorem as in [6], see Theorem 4.1. Corollary 5.3. Let Ψ be a quantum operation with respect to the Hilbert space H and let A ∈ B(H) with A = A∗. The following assertions are equivalent. (i) [LA, Ψ] = 0. (i)′ [RA, Ψ] = 0. (ii) In any state ρ ∈ D(H), A and A2 have expected values with respect to Ψnρ indepen- dent of n ≥ 0. ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 23 (ii)′ In any state ρ ∈ D(H), A has expected value and standard deviation with respect to (iii) A and A2 are fixed points of Ψ♯. Ψnρ independent of n ≥ 0. In order to put the investigations in [19] in a perspective closer to our approach, we now consider a scale of sets of constants of Ψ, more precisely, let (5.5) (5.6) (5.7) (5.9) (5.10) (5.11) 2 = {A ∈ B(H) A, A∗A, AA∗ ∈ CΨ}, CΨ p = {A ∈ B(H) p(A, A∗) ∈ CΨ for all complex polynomials p CΨ in two noncommutative variables}, CΨ c = {A ∈ B(H) C ∗(I, A) ⊆ CΨ}, CΨ w = {A ∈ B(H) W ∗(A) ⊆ CΨ}, (5.8) where C ∗(I, A) denotes the C ∗-algebra generated by I and A, while W ∗(A) denotes the von Neumann algebra generated by A. Transferring these classes in the Heisenberg picture, we have 2 = {A ∈ B(H) A, A∗A, AA∗ ∈ B(H)Ψ♯ CΨ p = {A ∈ B(H) p(A, A∗) ∈ B(H)Ψ♯ CΨ } = B(H)Ψ♯ 2 , for all complex polynomials p in two noncommutative variables} = B(H)Ψ♯ p , c = {A ∈ B(H) C ∗(I, A) ⊆ B(H)Ψ♯ CΨ w = {A ∈ B(H) W ∗(A) ⊆ B(H)Ψ♯ CΨ } = B(H)Ψ♯ c , } = B(H)Ψ♯ w . (5.12) It is easy to see that CΨ is an operator system, that is, a vector space stable under taking adjoints and containing the identity I, and w∗-closed, hence closed with respect to the operator norm as well. As any other operator system, CΨ is linearly generated by the set of its positive elements but, in general, not stable under multiplication, cf. [3], [4], [8]. Theorem 5.4. Let Ψ be a quantum operation with respect to a Hilbert space H. w and this set is a von Neumann algebra. p = CΨ 2 = CΨ c = CΨ (a) We always have CΨ (b) The following assertions are equivalent: (i) CΨ is stable under multiplication. (ii) CΨ = CΨ 2 . (iii) CΨ is a C ∗-algebra. (iv) CΨ is a von Neumann algebra. c ⊇ CΨ 2 ⊇ CΨ p ⊇ CΨ Proof. (a) Clearly, CΨ w . Due to the density of the set of all operators p(A, A∗), where p is an arbitrary complex polynomial in two noncommutative variables, in C ∗(I, A), the w∗-density of C ∗(I, A) in W ∗(A), as well as the continuity and w∗-continuity of the map A 7→ tr(Ψ(ρ)A), we have the equality CΨ w . On the other hand, using the dual representations as in (5.9) and (5.11), from Theorem 2.4 we obtain CΨ representation in the Heisenberg picture as B(H)Ψ♯ selfadjoint, hence B(H)Ψ♯ 2 by Theorem 2.4 we have In order to prove that this set is a von Neumann algebra, it is preferable to use its 2 , see (5.9). Since Ψ♯ is positive it is is stable under taking the involution A 7→ A∗. If A, B ∈ B(H)Ψ♯ 2 , 2 = CΨ c . p = CΨ c = CΨ (5.13) Ψ♯(AB) = AΨ♯(B) = AB, 24 A. GHEONDEA hence B(H)Ψ♯ 2 is stable under multiplication. On the other hand, Ψ♯((A + B)∗(A + B)) = Ψ♯(A∗A + A∗B + B∗A + B∗A) = Ψ♯(A∗A) + Ψ♯(A∗B) + Ψ♯(B∗A) + Ψ♯(B∗A) hence, taking into account of (5.13), = A∗A + A∗B + B∗A + B∗A = (A + B)∗(A + B). Similarly we prove that (A + B)(A + B)∗ is a fixed point of Ψ♯. Since, clearly A + B is a fixed point of Ψ♯, it follows that B(H)Ψ♯ is stable under addition as well. On the other hand, since Ψ♯ is w∗-continuous, it follows that B(H)Ψ♯ is a von Neumann algebra. 2 2 (b) This is actually a reformulation of Lemma 2.2 in [2]. (cid:3) Remarks 5.5. (1) There are special situations when one, hence all, of the assertions (i) through (iv) in Theorem 5.4 hold(s) automatically: for example, if there exists a state ω of B(H) that is faithful, in the sense that ω(C ∗C) = 0 implies C = 0, and invariant under Ψ♯, that is, ω ◦ Ψ♯ = ω, e.g. see Theorem 2.3 in [2]. (2) However, as Theorem 4.2 in [2] shows, there exist a Luders operation ΦA with A = {A1, A2, A3, A4, A5}, with notation as in (1.1), on H = ℓ2(F2), where F2 denotes the free group on two generators, and an operator B ∈ B(H)+ with Φ♯ A(B2) 6= B2. Consequently, both conditions in Theorem 5.1 item (iv) are necessary, in general, and the set of constants CΨ, for some quantum operations Ψ, is not stable under multiplication, equivalently, it is not a von Neumann algebra, see Theorem 5.4.(b). This answers the question raised in Remark 3.7.(3) as well. A(B) = B but Φ♯ 5.2. Continuous Quantum Semigroups. A family indexed on the set of nonnegative real numbers Ψ = {Ψt}t≥0 is called a dynamical quantum system, sometimes called a dynamical quantum stochastic system, with respect to a Hilbert space H, if it is a strongly continuous semigroup of quantum operations Ψt : B1(H) → B1(H), t ≥ 0. For a dynamical quantum system Ψ, we consider its infinitesimal generator ψ, see Subsection 2.2 for the general set- ting. This definition makes a representation of the dynamical quantum system Ψ into the Schrodinger picture. Transferring a dynamical quantum system Ψ into the Heisenberg pic- ture, we get its dual, usually called dynamical quantum Markov system, Ψ♯ = {Ψ♯ t}t≥0 which is a w∗-continuous one-parameter semigroup of w∗-continuous, unital, completely positive linear maps Ψ♯ t : B(H) → B(H) to which one associates its w∗-infinitesimal generator ψ♯, as in (2.16) and (2.17). Here, an important issue is that by Phillips's Theorem [27], ψ♯ is indeed the dual of ψ. Note that our definitions are more general than those usually considered in most mathe- matical models of quantum open systems, e.g. see [19], [16] and the rich bibliography cited there, which instead of strong continuity requires the (operator) norm continuity, that is, the mapping R+ ∋ t 7→ Ψt ∈ L(B1(H)) should be continuous with respect to the operator norm of L(B1(H)). An operator A ∈ B(H) is called a constant of the dynamical quantum system Ψ = {Ψt}t≥0, if, for any density operator ρ ∈ D(H), tr(Ψt(ρ)A) does not depend on t ≥ 0, equivalently, tr(Ψt(ρ)A) = tr(ρA) for all t ≥ 0. Clearly, A is a constant of Ψ if and only if for any T ∈ B1(H) we have tr(Ψt(T )A) = tr(T A) for all t ≥ 0, equivalently, tr(T Ψ♯ t(A)) = tr(T A) for all T ∈ B1(H) and all t ≥ 0. Consequently, A ∈ B(H) is a constant of Ψ if and only if ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 25 Ψ♯ t(A) = A for all t ≥ 0, that is, A is a fixed point of Ψ♯ denote the set of constants of Ψ t for all t ≥ 0. Formally, letting CΨ (5.14) CΨ = {A ∈ B(H) for all ρ ∈ D(H), tr(Ψt(ρ)A) does not depend on real t ≥ 0} = {A ∈ B(H) tr(Ψt(ρ)A) = tr(ρA) for all ρ ∈ D(H) and all t ≥ 0} = {A ∈ B(H) Ψ♯ t(A) = A for all t ≥ 0} = B(H)Ψ♯ , where the last equality is actually the definition of B(H)Ψ♯ as the set of all joint fixed points of Ψ♯ t, t ≥ 0. In addition, as a consequence of Theorem 2.5, we have (5.15) CΨ = B(H)Ψ♯ = Ker(ψ♯) = {T ∈ B(H) T ∈ Dom(ψ♯), ψ♯(T ) = 0}. The next theorem is the continuous variant of Theorem 5.1. Theorem 5.6. Let Ψ = {Ψt}t≥0 be a dynamical quantum stochastic system with respect to the Hilbert space H, let ψ denote its infinitesimal generator, and let A ∈ B(H). The following assertions are equivalent: (i) [LA, Ψt] = 0 for all t ≥ 0. (ii) A and A∗A are constants of Ψ. (iii) [RA, Ψ♯ t] = 0 for all t ≥ 0. (iv) A and A∗A are joint fixed points of Ψ♯ (v) [LA, ψ] = 0, that is, LA leaves Dom(ψ) invariant and Aψ(T ) = ψ(AT ) for all T ∈ (vi) A, A∗A ∈ Ker(ψ♯), that is, A, A∗A ∈ Dom(ψ♯) and ψ♯(A) = ψ♯(A∗A) = 0. The equivalence of assertions (i) -- (iv) follows from Theorem 5.1, while their equivalence with assertions (v) and (vi) can be proven similarly as in the proof of Theorem 3.8 and we do not repeat the arguments. In particular, the equivalence of assertions (iv) and (vi) in the previous theorem follows from Theorem 2.5, but this is a more general fact, see (5.15). t for all t ≥ 0. Dom(ψ). There is a symmetric variant to Theorem 5.6, in which LA and RA are interchanged and, correspondingly, A∗A and AA∗ are interchanged, the continuous dynamical system variant of Theorem 5.2. We leave the reader to formulate it. In case of a bounded observable A ∈ B(H)+, with expected value, variation, and standard deviation to an arbitrary state ρ ∈ D(H) as in (5.1) through (5.3), Theorem 5.6 can be reformulated to a noncommutative analogue of the Noether Type Theorem as in [6], see Theorem 4.1. Corollary 5.7. Let Ψ = {Ψt}t≥0 be a dynamical quantum stochastic system with respect to the Hilbert space H, let ψ denote its infinitesimal generator, and let A ∈ B(H), A = A∗, be a bounded observable. The following assertions are equivalent: (i) [LA, Ψt] = 0 for all t ≥ 0. (i)′ [RA, Ψt] = 0 for all t ≥ 0. (ii) In any state ρ ∈ D(H), A and A2 have expected values with respect to Ψt independent (ii)′ In any state ρ ∈ D(H), A has expected value and standard deviation with respect to (iii) [RA, Ψ♯ (iii)′ [LA, Ψ♯ of t ≥ 0. Ψt independent of t ≥ 0. t] = 0 for all t ≥ 0. t] = 0 for all t ≥ 0. 26 A. GHEONDEA Dom(ψ). t for all t ≥ 0. (iv) A and A2 are joint fixed points of Ψ♯ (v) [LA, ψ] = 0, that is, LA leaves Dom(ψ) invariant and Aψ(T ) = ψ(AT ) for all T ∈ (v′) [RA, ψ] = 0, that is, RA leaves Dom(ψ) invariant and ψ(T )A = ψ(T A) for all T ∈ (vi) A, A2 ∈ Ker(ψ♯), that is, A, A2 ∈ Dom(ψ♯) and ψ♯(A) = ψ♯(A2) = 0. On the other hand, as in (5.5) -- (5.8), we have the joint versions of the scale of sets of Dom(ψ). constants CΨ w ⊆ CΨ in the Schrodinger picture, more precisely, • =\t≥0 CΨt CΨ • , where • = 2, p, c, w, 2 ⊆ CΨ, c ⊆ CΨ p ⊆ CΨ (5.16) and, as in (5.9) -- (5.12), the sets of joint fixed points B(H)Ψ♯ in the Heisenberg picture, (5.17) w ⊆ B(H)Ψ♯ • =\t≥0 B(H)Ψ♯ c ⊆ B(H)Ψ♯ p ⊆ B(H)Ψ♯ 2 ⊆ B(H)Ψ♯ , B(H)Ψ♯ t • , where • = 2, p, c, w. CΨ = B(H)Ψ is a w∗-closed operator system and w∗-closed, hence closed with respect to the operator norm on B(H) as well, linearly generated by the set of its positive elements but, in general, not stable under multiplication. The following theorem is a consequence of Theorem 5.4. Theorem 5.8. Let Ψ be a dynamical quantum system with respect to the Hilbert space H. w and this set is 2 = CΨ p = CΨ c = CΨ a von Neumann algebra. (a) For any dynamical quantum system Ψ we have CΨ (b) The following assertions are equivalent: (i) CΨ is stable under multiplication. (ii) CΨ = CΨ 2 . (iii) CΨ is a C ∗-algebra. (iv) CΨ is a von Neumann algebra. Remarks 5.9. (a) The main theorem in [19] states that, for a dynamical quantum (stochas- tic) system Ψ under two additional constraints, namely, that the semigroup is (operator) norm continuous and that there exists a stationary strictly positive density operator, that is, there exists ρ ∈ B1(H)+ that is strictly positive and such that Ψt(ρ) = ρ for all real t ≥ 0, then CΨ♯ = B(H)Ψ♯ is a von Neumann algebra. This theorem remains true under the general assumption that the semigroup Ψ is strongly continuous: we use Theorem 5.8 while the existence of a stationary strictly positive density operator ρ implies the existence of a normal faithful stationary state ω(T ) = tr(ρT ), T ∈ B(H), and then Theorem 2.3 in [2]. (b) In case the dynamical quantum system Ψ is (operator) norm continuous, the infin- itesimal generator ψ is bounded and, by a result of G. Lindblad [24] (and, in the finite ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 27 dimensional case, of V. Gorini, A. Kossakowski, and E.C.G. Sudarshan [18]), it takes the form (5.18) ψ(S) = ∞Xk=1 (LkSL∗ k − 1 2 SL∗ kLk − 1 2 L∗ kLkS) + i[S, H], S ∈ B1(H), for a collection of operators Lk ∈ B(H), k = 1, 2, . . . , and a selfadjoint operator H ∈ B(H). It is easy to see that it's adjoint, which is the infinitesimal generator of the dual quantum Markov semigroup {Ψ♯ t}t≥0, is ∞Xk=1 ψ♯(T ) = (L∗ (5.19) Consequently, using (5.15), it follows that the constants of Ψ are exactly the solutions T ∈ B(H) of the equation ∞Xk=1 (5.20) kLk) − i[T, H], T ∈ B(H). kLk) − i[T, H] = 0, L∗ kLkT − (L∗ kT Lk − kT Lk − L∗ kLkT − 1 2 T L∗ 1 2 T L∗ 1 2 1 2 which is an operator Riccati equation. (c) In case the dynamical quantum system Ψ is (operator) norm continuous, hence (5.18) and (5.19) hold, and Ψ has a stationary strictly positive density operator, it is proven in [19] k k = 1, 2, . . .}′, in particular, it that the set CΨ is a von Neumann algebra. w coincides with the commutant of {H, Lk, L∗ Finally, we are in a position to approach the following question: to which extent are the latter conditions on A∗A or AA∗, as in Theorem 5.6.(ii), and the latter condition on A2 as in Corollary 5.7, really necessary? Note that a positive answer to this question will answer the similar question asked for the more general case of dynamical stochastic systems as in Subsection 3.2, see Remark 3.9.(b). Example 5.10. As in [2], let F2 denote the free group on two generators g1 and g2, and let ℓ2(F2) denote the Hilbert space of all square summable functions f : F2 → C. In ℓ2(F2) a canonical orthonormal basis is made up by {δx}x∈F2, where δx(y) = 0 for all y ∈ F2, y 6= x, and δx(x) = 1. Since F2 is infinitely countable it follows that ℓ2(F2) is infinite dimensional and separable. Let Uj ∈ B(ℓ2(F2)) denote the unitary operators Ujδx = δgjx, x ∈ F2 and j = 1, 2. We consider the linear bounded operator ψ : B1(ℓ2(F2)) → B1(ℓ2(F2)) defined by ψ(S) = U1SU ∗ 1 + U2SU ∗ 2 − 2S, S ∈ B1(ℓ2(F2)), (5.21) and then let Ψt(S) = exp(tψ(S)), S ∈ B1(ℓ2(F2)), t ≥ 0. (5.22) From [24], see Remark 5.9.(b), it follows that Ψ = {Ψt}t≥0 is a (operator) norm continuous semigroup of quantum operations with respect to ℓ2(F2). Also, let L(F2) = W ∗(U1, U2) denote the group von Neumann algebra of F2. We observe, e.g. by means of (5.20), that the commutant von Neumann algebra L(F2)′ is included in the set of constants CΨ. Lemma 5.11. Let Ψ be the dynamical quantum system as in Example 5.10. Then, CΨ is stable under multiplication if and only if it coincides with L(F2)′. 28 A. GHEONDEA Proof. It is sufficient to prove that, if CΨ is stable under multiplication then it coincides with L(F2)′. To see this, assume that CΨ is stable under multiplication hence, by Theorem 5.8.(b), it is a von Neumann algebra. By Remark 5.9.(b), it follows that for any orthogonal projection E ∈ CΨ equation (5.20) holds which, in our special case, is 2 EU2 = 2E. (5.23) Consequently, for each vector h ∈ ℓ2(F2) that lies in the range of E we have 1 EU1 + U ∗ U ∗ kEU1hk2 + kEU2hk2 = hU ∗ 1 EU1h, hi + hU ∗ 2 EU2h, hi = 2hEh, hi = 2khk2, from which, after a moment of thought, we see that Ujh should lie in the range of E for j = 1, 2. We have shown that Uj leaves the range of E invariant, j = 1, 2. Since the same is true for the range of I−E, it follows that Uj commutes with all orthogonal projections in the von Neumann algebra CΨ, hence CΨ ⊆ {U1, U ∗ 2}′ = L(F2)′. The converse inclusion was observed at the end of Example 5.10. 1 , U2, U ∗ (cid:3) During the proof of the next theorem we use terminology as in Subsection 2.3. Theorem 5.12. On any infinite dimensional separable Hilbert space H, there exists a (op- erator) norm continuous semigroup of quantum operations Φ = {Φt}t≥0 with respect to H, for which: under multiplication. (a) The set of constants CΦ is not a von Neumann algebra, equivalently, it is not stable (b) There exists A ∈ B(H)+ which is a constant of Φ but A2 is not. Proof. We first show that the (operator) norm continuous dynamical quantum system Ψ as in Example 5.10 has all the required properties. To this end, it is sufficient to prove assertion (b), then assertion (a) will follow from Theorem 5.8. By a classical result of J. Hakeda and J. Tomiyama [20], a von Neumann algebra M is injective if and only if its commutant M′ is injective. By another classical result of J.T. Schwartz [28], see also J. Tomiyama [32], the von Neumann algebra L(F2) is not injective hence, its commutant L(F2)′ is not injective either. Consequently, by Lemma 5.11 and Theorem 2.6, the set of joint fixed points B(ℓ2(F2))Ψ♯ = CΨ is strictly larger than the joint bimodule set I(Ψ♯). Since B(ℓ2(F2))Ψ♯ is an operator system, hence linearly generated by its positive cone, there exits A ∈ B(ℓ2(F2))Ψ♯ \ I(Φ) with A ≥ 0. In view of Theorem 2.1, this implies A2 6∈ B(ℓ2(F2))Ψ♯. In general, if H is an infinite dimensional separable Hilbert space, then there exists a unitary operator U : ℓ2(F2) → H and let Φt = U ∗ΨtU, for all real t ≥ 0. Then Φ = {Φt}t≥0 has all the required properties. (cid:3) Theorem 5.12 answers, in the negative, also the question on whether the condition that A2 is a joint fixed point, as in Theorem 3.8.(i), is a consequence of the condition that A is a joint fixed point. Appendix: A Proof of Theorem 2.1 Let Φ : A → B be a contractive completely positive map hence, by the Russ-Dye Theorem, kΦ(e)k = kΦk ≤ 1. By the Gelfand -- Naimark Theorem, B can be faithfully represented as a C ∗-subalgebra of B(H), for some Hilbert space H, hence Ψ can be considered as a completely positive map Φ : A → B(H). Then, by Stinespring Theorem [29], there exists a ON PROPAGATION OF FIXED POINTS OF QUANTUM OPERATIONS AND BEYOND 29 triple (K; π; V ), where K is a Hilbert space, π : A → B(K) is a ∗-representation, V : H → K is a bounded linear operator, ϕ(a) = V ∗π(a)V for all a ∈ A, and such that π(A)H is a total subset of K. We briefly recall the original construction [29]. K is defined as the factorization and completion to a Hilbert space of the vector space A ⊗ H with respect to the inner product h·,·iK, defined on elementary tensors by ha ⊗ h, b ⊗ kiK = hΦ(b∗a)h, kiH, for all a, b ∈ A, h, k ∈ H, and then extended by linearity. The representation π is defined on elementary tensors by π(a)(b ⊗ h) = ab ⊗ h, for all a, b ∈ A, h ∈ H, and then extended by linearity, while the operator V is defined by V h = Φ(e) ⊗ h, h ∈ H, and, since kΦ(e)k ≤ 1 then kV k ≤ 1, equivalent to IH − V ∗V ≥ 0 and to IK − V V ∗ ≥ 0. (1) The Schwarz Inequality. For any a ∈ A we have Φ(a∗a) − Φ(a)∗Φ(a) = V ∗π(a∗a)V − V ∗π(a)∗V V ∗π(a)V = V ∗π(a)∗(IK − V V ∗)π(a)V ≥ 0, hence, Φ(a)∗Φ(a) ≤ Φ(a∗a). (2) The Multiplicativity Property. For arbitrary but fixed a ∈ A, we have kΦ(a)∗Φ(a) − Φ(a∗a)k = kV ∗π(a)∗(IK − V V ∗)π(a)V k = kV ∗π(a)∗(IK − V V ∗)1/2(IK − V V ∗)1/2π(a)V k = k(IK − V V ∗)1/2π(a)V k2 hence, Φ(a)∗Φ(a) = Φ(a∗a), if and only if (IK − V V ∗)1/2π(a)V = 0 and, taking account that, for any b ∈ A, Φ(ba) − Φ(b)Φ(a) = V ∗π(b)(IK − V V ∗)1/2(IK − V V ∗)1/2π(a)V, it follows that Φ(a)∗Φ(a) = Φ(a∗a) if and only if Φ(b)Φ(a) = Φ(ba), for all b ∈ A. In a similar fashion the latter equivalence follows. (3) MΦ is the largest C ∗-subalgebra C of A such that ΦC : C → B is a ∗-homomorphism. Clearly, MΦ is stable under taking adjoints. From (2) it follows that MΦ is stable under multiplication. Let a, b ∈ MΦ. Then Φ((a + b)∗(a + b)) = Φ(a∗a + a∗b + b∗a + b∗b) = Φ(a∗a) + Φ(a∗b) + Φ(b∗a) + Φ(b∗b) = Φ(a)∗Φ(a) + Φ(a)∗Φ(b) + Φ(b)∗Φ(a) + Φ(b)∗Φ(b) = Φ(a + b)∗Φ(a + b). Similarly we prove that Φ((a + b)(a + b)∗) = Φ(a + b)Φ(a + b)∗, hence MΦ is stable under addition as well and, consequently, a C ∗-subalgebra of A. Clearly, ΦMΦ is a ∗-homomorphism. Let C be an arbitrary C ∗-subalgebra of A such that ΦC is a ∗-homomorphism. Then, for any a ∈ C we have Φ(a∗a) = Φ(a)∗Φ(a) and Φ(aa∗) = Φ(a)Φ(a)∗, hence a ∈ MΦ. References [1] S. Albeverio, R. Høegh-Krohn, Frobenius theory for positive maps of von Neumann algebras, Comm. Math. Phys. 64(1978), 83 -- 94. [2] A. Arias, A. Gheondea, S. Gudder, Fixed points of quantum operations, J. Math. Phys. 43(2002), 5872 -- 5881. [3] W. Arveson, Subalgebras of C ∗-algebras, Acta Math. 123(1969), 141 -- 224. 30 A. GHEONDEA [4] W. Arveson, Subalgebras of C ∗-algebras. II, Acta Math. 128(1972), 271 -- 308. [5] J.C. Baez, J. Biamonte, A course on quantum techniques for stochastic mechanics, arXiv: 1209.3632 [quant-ph]. [6] J.C. Baez, B. Fong, A Noether theorem for Markov processes, J. Math. Phys. 54(2013), 013301. [7] K. Bartoszek, W. Bartoszek, A Noether theorem for stochastic operators on Schatten classes, J. Math. Anal. Appl. (to appear). [8] O. Bratelli, P. Jørgensen, A. Kishimoto, R. Werner, Pure states on Od, J. Operator Theory, 43(2000), 97 -- 143. [9] N. Brown, N. Ozawa, C ∗-Algebras and Finite-Dimensional Approximations, Amer. Math. Soc., Prov- idence, R.I., 2008. [10] M.-D. Choi, A Schwarz type inequality for positive maps on C ∗-algebras, Illinois J. Math. 18(1974), 565 -- 574. [11] J.B. Conway, A Course in Operator Theory, Amer. Math. Soc., Providence, R.I. 2000. [12] D. Chruscinski, A. Kossakowski, On the structure of entanglement witnesses and new class of positive indecomposable maps, Open Sys. Information Dyn. 14(2007), 275 -- 294. [13] E.B. Davies, Quantum stochastic processes II., Comm. Math. Phys. 19(1970), 83 -- 105. [14] N. Dunford, J.T. Schwartz, Linear Operators. Part I: General Theory, Interscience Publishers, New York 1958. [15] E.E. Evans, R. Hoegh-Krohn, Spectral properties of positive maps on C ∗-algebras, J. London Math. Soc. 17(1978), 345 -- 355. [16] F. Fagnola, R. Rebolledo, Notes on the qualitative behaviour of quantum Markov semigroups, in Open Quantum Systems III. Recent Developments, pp. 161 -- 205, Springer Verlag, Berlin 2006. [17] A. Frigerio, M. Verri, Long-time asymptotic properties of dynamical semigroups on W*-algebras, Math. Z. 180(1982), 275 -- 286. [18] V. Gorini, A. Kossakowski, E.C.G. Sudarshan, Completely positive dynamical semigroups of N-level systems, J. Math. Phys. 17(1976), 821 -- 825. [19] J.E. Gough, T.S. Rat¸iu, O.G. Smolyanov, Noether's theorem for dissipative quantum dynamical systems, J. Math. Phys. 56(2015), 022108. [20] J. Hakeda, J. Tomiyama, On some extension property of von Neumann algebras, Tohoku Math. J. 19(1967), 315 -- 323. [21] E. Hille, R.S. Phillips, Functional Analysis on Semi-groups, Amer. Math. Soc. Coll. Publ. Vol. 31, Providence, R.I., 1957. [22] R. Horodecki, P. Horodecki, M. Horodecki, K. Horodecki, Quantum entanglement, Rev. Modern Phys. 81(2009), 865 -- 942. [23] R. Kadison, A generalized Schwarz inequality and algebraic invariants for operator algebras, Ann. of Math. 56(1952), 494 -- 503. [24] G. Lindblad, On the generators of quantum dynamical semigroups, Commun. Math. Phys. 48(1976) 119 -- 130. [25] E. Noether, Invariante Variationsprobleme, Nachr. D. Konig. Gesellsch. D. Wiss. Zu Gottingen, Math-phys. Klasse. 1918, 235 -- 257. [26] V.I. Paulsen, Completely Bounded Maps and Operator Algebras, Cambridge University Press, Cam- bridge 2003. [27] R.S. Phillips, The adjoint semigroup, Pacific J. Math., 5(1955), 269 -- 283. [28] J.T. Schwartz, Non-isomorphism of a pair of factors of type III, Comm. Pure Appl. Math. 16(1963), 111 -- 120. [29] W.F. Stinespring, Positive functions on C ∗-algebras, Proc. Amer. Math. Soc. 6(1955), 211 -- 216. [30] E. Størmer, Positive linear maps on operator algebras, Acta Mathematica, 110(1963), 233 -- 278. [31] J. Tomiyama, On the projection of norm one in W ∗-algebras, Proc. Japan Acad. 33(1957), 608 -- 612. [32] J. Tomiyama, Tensor Products and Projections of Norm One in von Neumann Algebras, Seminar Lecture Notes, University of Copenhagen, Copenhagen 1970. Department of Mathematics, Bilkent University, 06800 Bilkent, Ankara, Turkey, and Institutul de Matematica al Academiei Romane, C.P. 1-764, 014700 Bucures¸ti, Romania E-mail address: [email protected] and [email protected]
1107.2094	2	1107	2012-07-02T18:58:11	Completely bounded representations of convolution algebras of locally compact quantum groups	[ "math.OA", "math.FA" ]	Given a locally compact quantum group $\mathbb G$, we study the structure of completely bounded homomorphisms $\pi:L^1(\mathbb G)\rightarrow\mathcal B(H)$, and the question of when they are similar to $\ast$-homomorphisms. By analogy with the cocommutative case (representations of the Fourier algebra $A(G)$), we are led to consider the associated map $\pi^:L^1_\sharp(\mathbb G) \rightarrow \mathcal B(H)$ given by $\pi^(\omega) = \pi(\omega^\sharp)^$. We show that the corepresentation $V_\pi$ of $L^\infty(\mathbb G)$ associated to $\pi$ is invertible if and only if both $\pi$ and $\pi^$ are completely bounded. Moreover, we show that the co-efficient operators of such representations give rise to completely bounded multipliers of the dual convolution algebra $L^1(\hat \mathbb G)$. An application of these results is that any (co)isometric corepresentation is automatically unitary. An averaging argument then shows that when $\mathbb G$ is amenable, $\pi$ is similar to a -homomorphism if and only if $\pi^$ is completely bounded. For compact Kac algebras, and for certain cases of $A(G)$, we show that any completely bounded homomorphism $\pi$ is similar to a -homomorphism, without further assumption on $\pi^$. Using free product techniques, we construct new examples of compact quantum groups $\mathbb G$ such that $L^1(\mathbb G)$ admits bounded, but not completely bounded, representations.	math.OA	math	Completely bounded representations of convolution algebras of locally compact quantum groups Michael Brannan, Matthew Daws, Ebrahim Samei May 20, 2014 Abstract Given a locally compact quantum group G, we study the structure of completely bounded homomorphisms π : L1(G) → B(H), and the question of when they are similar to ∗- homomorphisms. By analogy with the cocommutative case (representations of the Fourier algebra A(G)), we are led to consider the associated map π∗ : L1 ♯ (G) → B(H) given by π∗(ω) = π(ω♯)∗. We show that the corepresentation Vπ of L∞(G) associated to π is in- vertible if and only if both π and π∗ are completely bounded. Moreover, we show that the co-efficient operators of such representations give rise to completely bounded multipliers of the dual convolution algebra L1( G). An application of these results is that any (co)isometric corepresentation is automatically unitary. An averaging argument then shows that when G is amenable, π is similar to a ∗-homomorphism if and only if π∗ is completely bounded. For compact Kac algebras, and for certain cases of A(G), we show that any completely bounded homomorphism π is similar to a ∗-homomorphism, without further assumption on π∗. Using free product techniques, we construct new examples of compact quantum groups G such that L1(G) admits bounded, but not completely bounded, representations. Keywords: Locally compact quantum group, Fourier algebra, completely bounded homomor- phism, multipliers, corepresentation, amenability, free product. MSC classification (2010): 20G42, 22D12, 22D15, 43A30, 46L07, 46L51, 46L89, 81R50. 1 Introduction Given a locally compact group G, a uniformly bounded representation of G on a Hilbert space H is a weakly continuous homomorphism π0 from G into the invertible group of the algebra B(H) of bounded operators on H, with kπ0k := sups∈G kπ0(s)k < ∞. The study of uniformly bounded representations of locally compact groups has been central to the development of harmonic analysis and to understanding the structure of operator algebras associated to locally compact groups (see [11, 15, 32] for example). A uniformly bounded representation π0 : G → B(H) is called unitarisable if there exists an in- vertible T ∈ B(H) such that T π0(·)T −1 is a unitary representation. G is called unitarisable if every uniformly bounded representation of G is unitarisable. In [12, 16], Day and Dixmier independently showed that if G is amenable, then G unitarisable. Perhaps one of the most celebrated open prob- lems concerning uniformly bounded representations is Dixmier’s similarity problem, which asks the converse: Is every unitarisable locally compact group amenable? For recent progress on this problem, see for example [29, 31]. 1 In this paper we initiate the study of “uniformly bounded representations” of locally compact quantum groups, and generalise Dixmier’s similarity problem to this context. To motivate the objects of study in this paper, recall that for a locally compact group G, there is a bijective cor- respondence between non-degenerate bounded representations π : L1(G) → B(H), of the Banach ∗-algebra L1(G), and uniformly bounded representations π0 : G → B(H). The correspondence between π and π0 is given in the usual way by integration: (cid:0)π(ω)αβ(cid:1) =ZG ω(s)(cid:0)π0(s)αβ(cid:1) ds (ω ∈ L1(G), α, β ∈ H). ♯ (G) = π∗, where π∗ : L1 Note that kπk = kπ0k. Further, as L1(G) = L∞(G)∗ is a maximal operator space, π is automati- cally completely bounded with kπkCB(L1(G),B(H)) = kπk. Moreover, from the above correspondence, it is easy to see that a non-degenerate representation π : L1(G) → B(H) is (similar to) a ∗- representation of L1(G) if and only if π0 is (similar to) a unitary representation. Thus, if G is a locally compact quantum group with von Neumann algebra L∞(G) and convo- lution algebra L1(G), we are led to consider (completely) bounded representations π : L1(G) → B(H). A representation π : L1(G) → B(H) is called a ∗-representation if the restriction of π to the canonical dense ∗-subalgebra L1 ♯ (G) ⊆ L1(G) (with involution ω 7→ ω♯), is a ∗-representation. This is equivalent to the fact that πL1 ♯ (G) → B(H) is the representation obtained from π via the formula π∗(ω) = π(ω♯)∗. As usual, we say π is similar to a ∗-representation if T π(·)T −1 is a ∗-representation in the above sense, for some invertible T ∈ B(H). A necessary condition for a representation π : L1(G) → B(H) to be similar to a ∗-representation is that both π and π∗ extend to completely bounded maps from L1(G) to B(H) (see Proposition 3.4). We therefore begin our investigation by focusing on properties of representations π : L1(G) → B(H) where both π and π∗ are completely bounded. Completely bounded representations π : L1(G) → B(H) correspond to corepresentations Vπ ∈ L∞(G)⊗B(H), that is, operators with (∆ ⊗ ι)(Vπ) = Vπ,13Vπ,23. We show in Theorem 3.3 that when Vπ is invertible in L∞(G)⊗B(H), then π∗ is automatically completely bounded. By using the duality theory for locally compact quantum groups, and the theory of completely bounded multipliers ([13, 14, 22]) we show in Theorem 4.5 that the converse is true– if π and π∗ are completely bounded, then Vπ is invertible (with an appropriate modification when π is a degenerate homomorphism). Furthermore, invertible corepresentations share many of the nice properties of unitary corepresentations– they interact well with the unbounded antipode, and are “strongly continuous”, that is, Vπ ∈ M(C0(G) ⊗ B0(H)), see Theorem 4.9. We also show that the co- efficient operators of such a representation π naturally induce completely bounded multipliers on the dual convolution algebra L1( G) (Corollary 4.8). This generalises a classical result of de Canniere and Haagerup [15, Theorem 2.2], showing that coefficient functions of uniformly bounded representations of a group G are always completely bounded multipliers of the Fourier algebra A(G). We believe that invertible corepresentations, equivalently, π such that π∗ is also completely bounded, should be thought of as the correct quantum generalisation of a uniformly bounded group representation. An interesting corollary of these results is that a corepresentation U which is not necessarily unitary, but is at least a partial isometry, and is either injective or surjective, must automatically be unitary. This follows since if π is the associated homomorphism, then the condition on U is enough to show that π∗ is completely bounded, and then U must be invertible. An application of this is an improvement upon a construction of Kustermans: the “induced corepresentations” constructed in [25] are always unitary, without any “integrability” condition; see Section 4.4. 2 Another application of our results is to the similarity problem for amenable locally compact quantum groups. When G is amenable, Theorem 4.5 combined with some averaging techniques with invariant means shows that a representation π : L1(G) → B(H) is similar to a ∗-representation if and only if both π and π∗ are completely bounded maps (Theorem 5.1). This generalises the results of Day and Dixmier on the unitarisability of amenable groups, as well as some known results for compact quantum groups. When G is a compact Kac algebra (and for certain cases of A(G)) we are able to improve these results and show that every completely bounded representation π : L1(G) → B(H) is similar to a ∗-representation, without assumption on π∗ (see Theorems 6.2, 8.1, and 8.2). In [8], the first examples of bounded, but not completely bounded, representations of Fourier algebras were constructed. Using non-commutative Khintchine inequalities from free probability theory, we show that similar constructions hold for certain free products of compact quantum groups; such representations cannot, of course, be similar to ∗-representations. Moreover, for the representations constructed here, we show that there are co-efficient operators which fail to induce bounded multipliers of L1( G) (Corollary 7.2). This provides further evidence that completely bounded maps are indeed the “correct” maps to study. The paper is organised as follows. In Section 2, we review some basic facts about locally compact quantum groups which will be needed in this paper, fixing some notation along the way. We also make some remarks on the extended Haagerup tensor product. In Section 3 we show the correspondence between completely bounded representations π on L1(G) and corepresentations Vπ, and show that if π is similar to a ∗-representation, or Vπ is invertible, then π∗ is also completely bounded. In Section 4 we prove the converse to this result. We first study the duality theory of locally compact quantum groups, trying to find a quantum analogue of the function space L1(G)∩ A(G). We then show that coefficients of representations π, such that both π and π∗ are completely bounded, induce completely bounded multipliers. This result is then applied to prove that Vπ is invertible. In Section 5, we consider the similarity problem for locally compact quantum groups, focusing on the amenable case, and refining our results for compact Kac algebras in Section 6. In Section 7, we construct bounded representations of L1(G) which are not completely bounded, where G is a certain free product of compact quantum groups. Section 8 deals only with representations of Fourier algebras, and extends some of the results obtained in [7]. Finally Appendix A looks at weak∗-closed operators, and proves a technical result about analytic generators of σ-weakly continuous, one-parameter groups. Finally, a quick word on notation. We use h·,·i to denote the bilinear pairing between a Banach space and its dual, and (··) to denote the inner product of a Hilbert space. We use the symbol ⊗ for the von Neumann algebraic tensor product; for tensor products of Hilbert spaces, or the minimal tensor products of C∗-algebras, we write simply ⊗. We shall use standard results from the theory of operator spaces, following [17] for example. Acknowledgements: This paper was initiated at the Canadian Abstract Harmonic Analysis Symposium 2010, and the first and second authors wish to thank the third author, Yemon Choi, and the staff at the University of Saskatchewan for their hospitality. We acknowledge the financial support of PIMS, NSERC and the University of Saskatchewan. The third author was supported by an NSERC Discovery Grant. 3 2 Locally compact quantum groups For the convenience of the reader, we give a brief overview of the theory of locally compact quantum groups. Our main references are [28] and [27]. The first reference is a self-contained account of the C∗-algebraic approach to locally compact quantum groups, and the second reference discusses the von Neumann algebraic approach. For other readable introductions, see [26] and [43]. A Hopf-von Neumann algebra is a pair (M, ∆) where M is a von Neumann algebra and ∆ : M → M⊗M is a unital normal ∗-homomorphism which is coassociative: (ι⊗ ∆)◦ ∆ = (∆⊗ ι)◦ ∆. Then ∆∗ induces a completely contractive Banach algebra product on the predual M∗. We shall write the product in M∗ by juxtaposition, so hx, ωω′i = h∆(x), ω ⊗ ω′i (x ∈ M, ω, ω′ ∈ M∗). Recall the notion of a normal semi-finite faithful weight ϕ : M + → [0,∞] (see [38, Chapter VII] for example). We let + ϕ = {x ∈ M + : ϕ(x) < ∞}. nϕ = {x ∈ M : ϕ(x∗x) < ∞}, mϕ = span{x∗y : x, y ∈ nϕ}, m ϕ is equal to M +∩ mϕ. We can Then mϕ is a hereditary ∗-subalgebra of M, nϕ is a left ideal, and m+ perform the GNS construction for ϕ, which leads to a Hilbert space H, a linear map Λ : nϕ → H with dense range, and a unital normal ∗-homomorphism π : M → B(H) with π(x)Λ(y) = Λ(xy). In future, we shall tend to suppress π in our notation. Tomita-Takesaki theory gives us the modular conjugation J and the modular automorphism group (σt)t∈R. Recall that there is a (unbounded) positive non-singular operator ∇ (denoted ∆ in [38]) which induces (σt) via σt(x) = ∇itx∇−it for x ∈ M and t ∈ R. Finally, M is in standard position on H, so in particular, for each ω ∈ M∗, there are ξ, η ∈ H with ω = ωξ,η, where hx, ωξ,ηi = (xξη) for ξ, η ∈ H. A von Neumann algebraic locally compact quantum group is a Hopf-von Neumann algebra (M, ∆), together with faithful normal semifinite weights ϕ, ψ which are left and right invariant, respectively. This means that (ω ∈ M + ∗ , x ∈ m + ψ ). + ϕ , y ∈ m ϕ(cid:0)(ω ⊗ ι)∆(x)(cid:1) = ϕ(x)h1, ωi, ψ(cid:0)(ι ⊗ ω)∆(y)(cid:1) = ψ(y)h1, ωi Using these weights, we can construct an antipode S, which will in general be unbounded. Then S has a decomposition S = Rτ−i/2, where R is the unitary antipode, and (τt)t∈R is the scaling group. The unitary antipode R is a normal ∗-antiautomorphism of M, and ∆R = σ(R ⊗ R)∆, where σ : M⊗M → M⊗M is the tensor swap map. As R is normal, it drops to an isometric linear map R∗ : M∗ → M∗, which is anti-multiplicative. As usual, we make the canonical choice that ϕ = ψ ◦ R. Associated to (M, ∆) is a reduced C∗-algebraic quantum group (A, ∆). Here A is a C∗- subalgebra of M, and ∆ : A → M(A ⊗ A), the multiplier algebra of A ⊗ A. Here we identify M(A ⊗ A) with a C∗-subalgebra of M⊗M; indeed, given a C∗-algebra B ⊆ B(H), we can always identify the multiplier algebra M(B) with {x ∈ B′′ : xb, bx ∈ B (b ∈ B)}. There is a unitary W , the fundamental unitary, acting on H ⊗ H (the Hilbert space tensor product) such that ∆(x) = W ∗(1 ⊗ x)W for x ∈ M. Then W is multiplicative, in the sense that W12W13W23 = W23W12. Here, and later, we use the standard leg number notation: W12 = W ⊗ I, W13 = Σ23W12Σ23 and so forth, where here Σ : H ⊗ H → H ⊗ H is the swap map. The left-regular representation of M∗ is the map λ : ω 7→ (ω⊗ι)W . This is an injective homomorphism, the norm closure of λ(M∗) is a C∗-algebra denoted by A, and the σ-weak closure of λ(M∗) is a von Neumann algebra, which we denote by M . We define a coproduct ∆ : M → M⊗ M by 4 ∆(x) = W ∗(1 ⊗ x) W , where W = ΣW Σ. Then we can find invariant weights to turn ( M, ∆) into a locally compact quantum group – the dual quantum group to M. We then have the Pontryagin M = M canonically. There is a nice link between the scaling duality theorem which states that group of M, and the modular theory of the (left) Haar weight of M , in that τt(x) = ∇itx ∇−it for x ∈ M, t ∈ R. In this paper we will use the symbol G to indicate the abstract “object” to be thought of as a locally compact quantum group. Inspired by the classical examples of locally compact groups, we write L∞(G) for M, C0(G) for A, C b(G) for M(A) (the multiplier algebra of A), L1(G) for M∗, and L2(G) for H. We denote the locally compact quantum group dual to G by G, and use the notations L∞( G) = M , L1( G) = M∗, and so on. In order to distinguish between objects attached to G from those attached to G, we decorate elements in L∞( G), L1( G) etc. with “hats”. For example, we write x ∈ L∞( G), and ω ∈ L1( G). Since the antipode S of G is unbounded in general, there is not a natural way to turn L1(G) into a ∗-algebra. However, L1(G) always contains a dense subalgebra which has a conjugate-linear involution. We follow [24], see also [27, Section 2]. Define L1 ♯ (G) to be the collection of those ω ∈ L1(G) such that there exists ω♯ ∈ L1(G) with hx, ω♯i = hS(x)∗, ωi for each x ∈ D(S). Then L1 ♯ (G) is a dense subalgebra of L1(G), and ω 7→ ω♯ defines an involution on L1 ♯ (G). As [24, Proposition 3.1] shows, given ω ∈ L1(G), we have that ω ∈ L1 ♯ (G) if and only if λ(ω)∗ = λ(ω′) for some ω′ ∈ L1(G). Furthermore, the left regular representation λ becomes a ∗-homomorphism, when restricted to L1 When G is a locally compact group, we write V N(G) for L∞( G) (the von Neumann algebra generated by the left regular representation of λG : G → U(L2(G))), and write A(G) for L1( G), and call this the Fourier algebra of G. The Fourier algebra was first studied as a Banach algebra by Eymard [19]. The left regular representation λ : A(G) → C0(G), is a completely contrac- tive ∗-homomorphism, which allows us to identify A(G) with a dense ∗-subalgebra of C0(G), the identification being given by ♯ (G). λ(ω)(t) = hλG(t−1), ωi (ω ∈ A(G), t ∈ G). See [39] for details (note that the above identification differs from the one considered by Eymard, where t is replaced by t−1 in [19]). So as to not overburden notation, in many places (especially in Section 8) we will drop the notation λ from above, and just view A(G) as a concrete ∗-subalgebra of C0(G). 2.1 The extended Haagerup tensor product Let M and N be von Neumann algebras. The extended (or weak∗) Haagerup tensor product of M with N is the collection of x ∈ M⊗N such that x =Xi with σ-weak convergence, and such that Pi xix∗ 1/2(cid:13)(cid:13)(cid:13)Xi i(cid:13)(cid:13)(cid:13) xi ⊗ yi i and Pi y∗ i yi(cid:13)(cid:13)(cid:13) kxkeh = infn(cid:13)(cid:13)(cid:13)Xi i yi converge σ-weakly in M and N 1/2 : x =Xi xi ⊗ yio, respectively. The natural norm is then and we write M ⊗eh N for the resulting normed space. See [6] or [18] for further details. xix∗ y∗ 5 Suppose we are given (xi)i∈I ⊆ M and (yi)i∈I ⊆ N with Pi xix∗ i yi converging σ- weakly. Let M ⊆ B(H) and N ⊆ B(K). Then there are bounded maps X : H → H ⊗ ℓ2(I) and Y : K → K ⊗ ℓ2(I) given by X(ξ) =Xi x∗ i (ξ) ⊗ δi, (ξ ∈ L2(G)), where (δi)i∈I is the canonical orthonormal basis for ℓ2(I). Indeed, X ∗X = Pi xix∗ Pi y∗ i yi. Letting Σ : K ⊗ ℓ2(I) → ℓ2(I) ⊗ K be the swap map, a simple calculation shows that i and Pi y∗ i and Y ∗Y = Y (ξ) =Xi yi(ξ) ⊗ δi (X ∗ ⊗ 1)(1 ⊗ Σ)(1 ⊗ Y ) =Xi xi ⊗ yi ∈ M⊗N ⊆ B(H ⊗ K). This argument hence shows that Pi xi ⊗ yi always converges σ-weakly in M⊗N, as long as Pi∈I xix∗ i yi converge σ-weakly. Finally, note that if H = K then i and Pi∈I y∗ Xi xiyi = X ∗Y ∈ B(H), and in particular, if M = N, then Pi xiyi converges σ-weakly in M to X ∗Y . 3 Completely bounded homomorphisms and corepresen- tations In this section, we investigate the correspondence between completely bounded representations π of L1(G), and corepresentations of U ∈ L∞(G)⊗B(H) (which are not assumed unitary). We show how various naturally defined variants of π are associated to variants of U. In particular, we show that π is similar to a ∗-representation if and only if U is similar (in the sense of corepresentations) to a unitary. One of our variants of π is defined using the unbounded ∗-map on L1 ♯ (G). We show that if U is invertible, then all of the variants of π are completely bounded. A principle result, Theorem 4.7 below, is that the converse to this result is also true; and hence we have a bijection between invertible (not necessarily unitary) corepresentations and a certain class of completely bounded representations. We are interested in completely bounded homomorphisms π : L1(G) → B(H). We have the standard identifications (see [17, Chapter 7]): CB(L1(G),B(H)) ∼=(cid:0)L1(G)b⊗T (H)(cid:1)∗ ∼= L∞(G)⊗B(H). Here b⊗ denotes the operator space projective tensor product, and ⊗ the von Neumann algebraic tensor product. We use the notation that π : L1(G) → B(H) is associated with Vπ ∈ L∞(G)⊗B(H), the relationship is that We use the standard leg numbering notation, see for example [2, 41]. π(ω) = (ω ⊗ ι)Vπ (ω ∈ L1(G)). Lemma 3.1. Let π : L1(G) → B(H) be a completely bounded map, associated to Vπ ∈ L∞(G)⊗B(H). Then π is a homomorphism if and only if (∆ ⊗ ι)Vπ = Vπ,13Vπ,23. Similarly, π is an anti- homomorphism if and only if (∆ ⊗ ι)Vπ = Vπ,23Vπ,13. 6 Proof. For ω1, ω2 ∈ L1(G), while π(ω1ω2) =(cid:0)∆∗(ω1 ⊗ ω2) ⊗ ι(cid:1)Vπ = (ω1 ⊗ ω2 ⊗ ι)(cid:0)(∆ ⊗ ι)Vπ(cid:1), π(ω1)π(ω2) = (ω1 ⊗ ι)Vπ(ω2 ⊗ ι)Vπ = (ω1 ⊗ ω2 ⊗ ι)(cid:0)Vπ,13Vπ,23(cid:1), and so π is a homomorphism if and only if (∆⊗ι)Vπ = Vπ,13Vπ,23. The result for anti-homomorphisms is analogous. The condition that (∆ ⊗ ι)Vπ = Vπ,13Vπ,23 says that Vπ is a co-representation of L∞(G). Simi- Following [7], given a locally compact group G and a representation π : A(G) → B(H), we larly, we say that Vπ is a co-anti-representation when (∆ ⊗ ι)Vπ = Vπ,23Vπ,13. define π, π∗ and π by π(ω) = π(ω), π∗(ω) = π(ω)∗, π(ω) = π(ω)∗ = π(ω)∗ where ω(g) = ω(g−1) (viewing A(G) as a subalgebra of C0(G)). (ω ∈ A(G)), For ω ∈ L1(G), define ω∗ by hx, ω∗i = hx∗, ωi for x ∈ L∞(G). As ∆ is a ∗-homomorphism, it is easy to see that L1(G) → L1(G); ω 7→ ω∗ is a conjugate linear homomorphism. Recalling the definition of L1 ♯ (G) from Section 2, given π : L1(G) → B(H), we define ♯ (G)∗ → B(H); ω 7→ π(cid:0)(ω∗)♯(cid:1), π : L1 π∗ : L1 ♯ (G) → B(H); ω 7→ π(ω♯)∗, π : L1(G) → B(H); ω 7→ π(ω∗)∗. These are defined by analogy with the A(G) case, and the following results show that they are natural choices. Proposition 3.2. Let π : L1(G) → B(H) be a bounded map. Then π is completely bounded if and only if π is completely bounded, and in this case, V ∗ π = Vπ. Furthermore, π extends to a completely bounded map defined on L1(G) if and only if π∗ extends to a completely bounded map defined on L1(G), and in this case, V ∗ Proof. Let U ∈ L∞(G)⊗B(H) and define φ : L1(G) → B(H) by φ(ω) = (ω ⊗ ι)(U). Then π = Vπ∗. φ(ω∗)∗ =(cid:0)(ω∗ ⊗ ι)(U)(cid:1)∗ =(cid:0)(ω ⊗ ι)(U ∗)∗(cid:1)∗ = (ω ⊗ ι)(U ∗). It follows immediately that π is completely bounded if and only if π is, and that V ∗ π = Vπ. As π∗(ω∗) = π(ω)∗, similarly π extends to completely bounded map if and only if π∗ does, and V ∗ π = Vπ∗. We wish to show that ∗-representations of L1 ♯ (G) do give rise to completely bounded π such that also π∗ is completely bounded. To do this, we need to investigate how unitary, or more generally, invertible corepresentations interact with the antipode. Let us recall a characterisation of the antipode using ideas very close to the extended Haagerup tensor product. This is shown in [28, Corollary 5.34] in the C∗-algebra setting, but the same proofs work in the von Neumann case. Let x, y ∈ L∞(G) be such that there are families (xi), (yi) ⊆ L∞(G) with Pi xix∗ i < ∞ and P y∗ i yi < ∞, and such that x ⊗ 1 =Xi ∆(xi)(1 ⊗ yi), y ⊗ 1 =Xi (1 ⊗ xi)∆(yi). Then x ∈ D(S) and S(x) = y. Note that these sums converge σ-weakly, compare Section 2.1. For further details on characterising the antipode in this way, see [44, Proposition 5.6]. 7 Theorem 3.3. Let Vπ ∈ L∞(G)⊗B(H) be an invertible corepresentation, associated to a completely bounded homomorphism π : L1(G) → B(H). Then π is a completely bounded anti-homomorphism, and V −1 π = Vπ. Proof. Let α, β ∈ H, let (fi) be an orthonormal basis of H, and set xi = (ι ⊗ ωfi,β)(Vπ), As in Section 2.1, it is easy to see thatPi xix∗ with the sums converging σ-weakly. Now, using that (∆ ⊗ ι)(Vπ) = Vπ,13Vπ,23, i = (ι⊗ωα)(VπV ∗ yi = (ι ⊗ ωα,fi)(V −1 π ). π ) andPi y∗ (ι ⊗ ι ⊗ ωfi,β)((∆ ⊗ ι)(Vπ))(1 ⊗ yi) (ι ⊗ ι ⊗ ωfi,β)(Vπ,13Vπ,23)(1 ⊗ (ι ⊗ ωα,fi)(V −1 ∆(xi)(1 ⊗ yi) =Xi =Xi Xi π )). i yi = (ι⊗ωβ)((V −1 π )∗V −1 π ), Now, for any Hilbert space K, and S, T ∈ B(K ⊗ H), Xi (ι ⊗ ωfi,β)(S)(ι ⊗ ωα,fi)(T ) = (ι ⊗ ωα,β)(ST ). Thus Xi ∆(xi)(1 ⊗ yi) = (ι ⊗ ι ⊗ ωα,β)(Vπ,13Vπ,23V −1 π ) = V −1 π,23) = (ι ⊗ ωα,β)(Vπ) ⊗ 1. π,23V −1 π,13, and so Similarly, as ∆ is a homomorphism, (∆ ⊗ ι)(V −1 Xi (1 ⊗ xi)∆(yi) =Xi (1 ⊗ (ι ⊗ ωfi,β)(Vπ))(ι ⊗ ι ⊗ ωα,fi)(V −1 π,23V −1 π,13) = (ι ⊗ ι ⊗ ωα,β)(Vπ,23V −1 π,23V −1 π,13) = (ι ⊗ ωα,β)(V −1 π ) ⊗ 1. Thus (ι ⊗ ωα,β)(Vπ) ∈ D(S) and S(cid:0)(ι ⊗ ωα,β)(Vπ)(cid:1) = (ι ⊗ ωα,β)(V −1 Let ω∗ ∈ L1 ♯ (G), so we have that π ). hx, (ω∗)♯i = hS(x)∗, ω∗i = hS(x), ωi (x ∈ D(S)). It follows that (π(ω)αβ) = h(ι ⊗ ωα,β(Vπ), (ω∗)♯i = hS((ι ⊗ ωα,β(Vπ)), ωi = h(ι ⊗ ωα,β)(V −1 π ), ωi = hV −1 π , ω ⊗ ωα,βi. As such ω are dense in L1(G), it follows that π is completely bounded, with Vπ = V −1 claimed. π , as We now show that any ∗-homomorphism on L1 ♯ (G) gives rise to a completely bounded homo- morphism π such that also π∗ is completely bounded. We use a result of Kustermans that any ∗-representation has a “generator”– that is, Vπ exists and is unitary. The interaction between a unitary corepresentation and the antipode (which, in some sense, we generalised in the previous proposition) is of course well-known, see [50, Theorem 1.6] for example. 8 ♯ (G) → B(H) be a homomorphism which is similar to a ∗-homomorphism. Proposition 3.4. Let π : L1 Then π and π∗ extend by continuity to completely bounded homomorphisms L1(G) → B(H), and π and π extend by continuity to completely bounded anti-homomorphisms L1(G) → B(H). Proof. We first show this in the case when π is a ∗-homomorphism. Then π(ω♯) = π(ω)∗ for each ω ∈ L1 ♯ (G), and so π∗ = π. As π is a ∗-homomorphism, it is necessarily contractive (see [37, Chapter I, Proposition 5.2] for example) and so extends by continuity to a homomorphism defined on all of L1(G). To show that π extends to a completely bounded homomorphism, we use a non-trivial result of Kustermans. Suppose for the moment that π : L1 ♯ (G) → B(H) is non-degenerate. By [24, Corollary 4.3], there exists a unitary U ∈ L∞(G)⊗B(H) (actually, U ∈ M(C0(G) ⊗ B0(H)) ⊆ L∞(G)⊗B(H)) such that π(ω) = (ω ⊗ ι)U for each ω ∈ L1 ♯ (G). It follows immediately that π does indeed extend to a completely bounded homomorphism from L1(G), and that thus actually U = Vπ. As U is unitary, so invertible, the previous propositions shows that π∗ is also completely bounded. If π is degenerate, then as π is a ∗-homomorphism, we can orthogonally decompose H as H1 ⊕ H2, where π restricts to H1, and π(ω)ξ = 0 for each ξ ∈ H2, ω ∈ L1 ♯ (G). Let π1 be the restriction of π to H1; so π1 : L1(G) → B(H1) is a completely bounded homomorphism. It follows immediately that the same must be true of π and of π∗. If π is only similar to a ∗-representation, then there exists an invertible T ∈ B(H) such that ♯ (G) → B(H), ω 7→ T −1π(ω)T is a ∗-homomorphism. Then θ extends to a completely bounded θ : L1 homomorphism, and hence so also does π = T θ(·)T −1. The same result holds for π∗, as π∗(ω) = π(ω♯)∗ = (T −1)∗θ∗(ω)T ∗ (ω ∈ L1 ♯ (G)). Having established that π and π∗ extend to completely bounded maps L1(G) → B(H), it follows from Proposition 3.2 that also π and π extend to completely bounded maps L1(G) → B(H) which are easily seen to be anti-homomorphisms. 4 Invertible corepresentations In the previous section, we showed that a necessary condition for π : L1(G) → B(H) to be similar to a ∗-representation is that both π and π∗ are completely bounded. Furthermore, if π is associated to a corepresentation Vπ, and Vπ is invertible, then π∗ is completely bounded. In this section, we use the duality theory of locally compact quantum groups to show the converse: if π and π∗ are completely bounded, then Vπ is invertible (together with an appropriate interpretation of this when π is a degenerate homomorphism). Identification of “function” spaces 4.1 In [7], an important component of the argument was to use the space L1(G) ∩ A(G), identified as a function space on G. For a quantum group, we would like to be able to talk about the space L1(G) ∩ L1( G); to make sense of this, we shall embed dense subspaces of L1(G) and L1( G) into L2(G), making use of the fact that we always identify L2(G) with L2( G). This section is slightly technical– the key result is Proposition 4.3 below, which is used in the following section. We first introduce the non-standard, but instructive notation L1(G) ∩ L2(G) =(cid:8)ω ∈ L1(G) : ∃ ξ ∈ L2(G),hx∗, ωi = (ξΛ(x)) (x ∈ nϕ)(cid:9). 9 This space is denoted by I in [27, Section 1.1] and [28, Notation 8.4]. Then ϕ is the weight with GNS construction (L2(G), ι, Λ) where λ(L1(G)∩ L2(G)) forms a σ-strong∗ core for Λ, and we have that Then [28, Result 8.6] shows that L1(G) ∩ L2(G) is a left ideal in L1(G) and hx∗, ωi =(cid:0)Λ(λ(ω))(cid:12)(cid:12)Λ(x)(cid:1) Λ(cid:0)λ(ωω1)(cid:1) = λ(ω)Λ(cid:0)λ(ω1)(cid:1) (x ∈ nϕ, ω ∈ L1(G) ∩ L2(G)). (ω ∈ L1(G), ω1 ∈ L1(G) ∩ L2(G)). This formula is of course immediate from the fact that Λ is a GNS map, but this reasoning is circular(!) if one is following [28, Section 8]. Further, [28, Result 8.6] can easily be adapted (or just perform the obvious calculation) to show that L1(G) ∩ L2(G) is a left L∞(G) module, and Λ(cid:0)λ(xω1)(cid:1) = xΛ(cid:0)λ(ω1)(cid:1) (x ∈ L∞(G), ω1 ∈ L1(G) ∩ L2(G)). Continuing to follow [28, Section 8], we find that there is a norm continuous, one-parameter group (ρt)t∈R of isometries on L1(G), such that ρt is an algebra homomorphism for each t, and with σt(λ(ω)) = λ(ρt(ω)) for ω ∈ L1(G) (where (σt)t is the modular automorphism group associated to the dual left Haar weight ϕ). As observed before [27, Proposition 2.8], each ρt maps L1 ♯ (G) into itself, and ρt(ω♯) = ρt(ω)♯ for ω ∈ L1 ♯ (G). Finally, for ω ∈ L1(G)∩L2(G), also ρt(ω) ∈ L1(G)∩L2(G) and Λ(λ(ρt(ω))) = ∇it Λ(λ(ω)). ♯ (G), and ρi(ω)♯ ∈ Lemma 4.1. The collection of ω ∈ L1(G) ∩ L2(G) with ω ∈ D(ρi), ρi(ω) ∈ L1 L1(G)∩ L2(G) is dense in L1(G). Furthermore, the resulting collection of vectors Λ(λ(ω)) is dense in L2(G). Proof. We use a “smearing argument”, compare (for example) [26, Proposition 5.21]. Let ω ∈ L1 ♯ (G) ∩ (L1(G) ∩ L2(G)). Such ω are dense in L1(G) by [27, Lemma 2.5]. For n ∈ N and z ∈ C define ω(n, z) = n √π ZR exp(−n2(t + z)2)ρt(ω) dt. Then ω(n, 0) is analytic for ρ and ρz(ω(n, 0)) = ω(n, z). As ρ is norm continuous, it follows that ω(n, 0) → ω in norm as n → ∞. For x ∈ D(S), using that ρt(ω)♯ = ρt(ω♯), hS(x)∗, ω(n, z)i = = n √π ZR √π ZR n exp(−n2(t + z)2)hS(x)∗, ρt(ω)i dt exp(−n2(t + z)2)hx, ρt(ω♯)i dt = hx, (ω♯)(n, z)i. ♯ (G) and ω(n, z)♯ = (ω♯)(n, z). Hence ω(n, z) ∈ L1 For x ∈ nϕ, hx∗, (ω♯)(n, z)i = = where n n √π ZR exp(−n2(t + z)2)hx∗, ρt(ω♯)i dt √π ZR exp(−n2(t + z)2)(cid:0)Λ(λ(ρt(ω♯)))(cid:12)(cid:12)Λ(x)(cid:1) dt =(cid:0)ξ(cid:12)(cid:12)Λ(x)(cid:1), √π ZR exp(−n2(t + z)2) ∇it Λ(λ(ω♯)) dt. n ξ = 10 In particular, (ω♯)(n, z) ∈ L1(G) ∩ L2(G). It follows that {ω(n, 0) : n ∈ N, ω ∈ L1 that ρz(ω(n, 0)) ∈ L1 a similar calculation to that above yields that ♯ (G) ∩ L2(G)} is dense in L1(G), analytic for ρ, satisfies ♯ (G) for all z, and satisfies that ρz(ω(n, 0))♯ ∈ L1(G) ∩ L2(G). Furthermore, Λ(λ(ω(n, 0))) = n √π ZR exp(−n2t2) ∇it Λ(λ(ω)) dt. This converges to Λ(λ(ω)) in norm, as n → ∞. In particular, as ω varies, the collection of vectors Λ(λ(ω(n, 0))) is dense in L2(G). For the following, let T be the Tomita map, which is the closure of Λ(n ϕ∩ n ϕ) → L2(G), Λ(x) 7→ Λ(x∗). For further details, see [38, Chapter 1] (where this map is denoted by S, a notation which we avoid, as it clashes with the antipode). Proposition 4.2. Let X = {ω ∈ L1(G) ∩ L2(G) : Λ(λ(ω)) ∈ D( T ∗)}. Then X is dense in L1(G), and Λ(λ(X)) is dense in L2(G). For x ∈ D(S)∗ and ω ∈ X, we have that xω ∈ X, and T ∗ Λ(λ(xω)) = S(x∗) T ∗ Λ(λ(ω)). ∗ Proof. Let ω ∈ L1(G) ∩ L2(G) be given by Lemma 4.1. As λ(ρt(ω)) = σt(λ(ω)), an analytic continuation argument shows that λ(ω) ∈ D(σi) with σi(λ(ω)) = λ(ρi(ω)). As also ρi(ω) ∈ L1 ♯ (G) and ρi(ω)♯ ∈ L1(G) ∩ L2(G), we have that Λ(cid:0)λ(ρi(ω)♯)(cid:1) = Λ(cid:0)λ(ρi(ω))∗(cid:1) = T Λ(cid:0)λ(ρi(ω))(cid:1) = T Λ(cid:0)σi(λ(ω))(cid:1) = T ∇−1 Λ(cid:0)λ(ω)(cid:1) = T ∗ Λ(cid:0)λ(ω)(cid:1). As such ω are dense in L1(G), it follows that X is dense. As the collection of vectors Λ(λ(ω)) is dense in L2(G), it also follows that Λ(λ(X)) is dense. Now, D(S)∗ = D(τi/2). For x ∈ L∞(G) and t ∈ R, τt is implemented as τt(x) = ∇itx ∇−it. Thus, that x ∈ D(τi/2) means that ∇−1/2x ∇1/2 extends to a bounded operator, namely τi/2(x). Thus also J ∇−1/2x ∇1/2 J ⊆ S(x∗). As T ∗ = J ∇−1/2 = ∇1/2 J, we conclude that T ∗x T ∗ ⊆ S(x∗). Thus, for x ∈ S(x)∗ and ω ∈ X, S(x∗) T ∗ Λ(λ(ω)) = T ∗x T ∗ T ∗ Λ(λ(ω)) = T ∗xΛ(λ(ω)) = T ∗ Λ(λ(xω)), as required. Here we used that T ∗ = ( T ∗)−1, see [38, Lemma 1.5, Chapter 1]. Let us make one final definition: L1 ♯ (G) ∩ L2 ♯ (G) = {ω ∈ L1(G) ∩ L2(G) : ω ∈ L1 ♯ (G) and ω♯ ∈ L1(G) ∩ L2(G)}. Then [27, Proposition 2.6] shows that L1 is dense in L2(G). ♯ (G) ∩ L2 ♯ (G) is dense in L1(G), and Λ(λ(L1 ♯ (G) ∩ L2 ♯ (G))) Proposition 4.3. Let ω1 ∈ L1 ♯ (G), and let ω2 ∈ X be as in the previous proposition. Let x ∈ L∞(G), set ξ = Λ(λ(ω1)) and η = T ∗ Λ(λ(ω2)), and set ω = ωxξ,η. Then ω ∈ L1( G) ∩ L2(G) and Λ(λ(ω)) = Λ(λ((xω1)ω2)). When we take x = 1, as ω1, ω2 vary, such ω are dense in L1( G). ♯ (G) ∩ L2 11 Equivalently, T π α,β is determined by the dual pairing T π α,β = (ι ⊗ ωα,β)Vπ ∈ L∞(G). α,β, ωi =(cid:0)π(ω)α(cid:12)(cid:12)β(cid:1) hT π (ω ∈ L1(G)). Proof. Let x ∈ n ϕ. As λ(xω1) ∈ n ϕ and n ϕ is a left ideal, both λ(xω1)∗x ∈ n ϕ and x∗λ(xω1) ∈ n ϕ, ϕ. Thus Λ(x∗λ(xω1)) ∈ D( T ) and T Λ(x∗λ(xω1)) = Λ(λ(xω1)∗x). Thus so that x∗λ(xω1) ∈ n ϕ ∩ n ∗ hx∗, ωi = (x∗xξη) =(cid:0)x∗xΛ(λ(ω1))(cid:12)(cid:12) T ∗ Λ(λ(ω2))(cid:1) =(cid:0)x∗ Λ(λ(xω1))(cid:12)(cid:12) T ∗ Λ(λ(ω2))(cid:1) =(cid:0)Λ(λ(ω2))(cid:12)(cid:12) T Λ(x∗λ(xω1))(cid:1) =(cid:0)Λ(λ(ω2))(cid:12)(cid:12)Λ(λ(xω1)∗ x)(cid:1) = ϕ(cid:0)x∗λ(xω1)λ(ω2)(cid:1) =(cid:0)Λ(λ((xω1)ω2))(cid:12)(cid:12)Λ(x)(cid:1). As x ∈ n ϕ was arbitrary, this shows that ω ∈ L1( G) ∩ L2(G) with Λ(λ(ω)) = Λ(λ((xω1)ω2)) as claimed. As above, the collection of allowed ξ is dense in L2(G), and by Proposition 4.2, the collection of allowed η is also dense in L2(G). Hence certainly the collection of thus constructed ω will be dense in L1( G). 4.2 Coefficients of representations Let π : L1(G) → B(H) be a completely bounded (anti-)homomorphism with associated co(-anti- )representation Vπ. Given α, β ∈ H, a co-efficient of π is the operator (1) Note that the latter definition of T π completely bounded – a situation we will address in Section 7. α,β makes sense even if π is bounded, but not necessarily In this subsection and the next one, we study the analytic structure of co-efficients of completely bounded representations π : L1(G) → B(H). We show that if π∗ is also completely bounded, then co-efficients of π give rise to completely bounded multipliers of L1( G), a tool which will allow us to prove the main theorems of this section (Theorems 4.7 and 4.9) below. Proposition 4.4. Suppose that both π and π∗ extend to completely bounded maps L1(G) → B(H). For every α, β ∈ H, we have that T π∗ Proof. Let ω ∈ L1 β,α, ω♯i =(cid:0)π(ω♯)β(cid:12)(cid:12)α(cid:1) =(cid:0)π(ω♯)∗α(cid:12)(cid:12)β(cid:1) = (π∗(ω)αβ) = hT π∗ hT π α,β ∈ D(S) with S(T π∗ β,α)∗ = T π∗ Comparing this with the definition of ♯, we might hope that T π α,β. This is indeed true, but a little care is needed, see Proposition A.1 in the appendix. As S◦∗◦S = ∗, we also see that T π∗ β,α ∈ D(S) with S(T π α,β ∈ D(S) with S(T π∗ β,α, as claimed. α,β)∗ = T π β,α. ♯ (G), so that α,β)∗ = T π α,β, ωi. Given a co-efficient T π α,β, arguing as in Section 2.1 and as in the proof of Theorem 3.3, we see that ∆(T π α,β) = (ι ⊗ ι ⊗ ωα,β)(Vπ,13Vπ,23) =Xi (ι ⊗ ωfi,β)(Vπ) ⊗ (ι ⊗ ωα,fi)(Vπ) =Xi fi,β ⊗ T π T π α,fi, the sum converging σ-weakly. This observation, combined with the previous proposition, shows that the hypotheses of the next theorem are not so outlandish; compare with the proof of Theo- rem 4.7 below. 12 Theorem 4.5. Let G be a locally compact quantum group, let x ∈ L∞(G), and suppose that σ∆(x) = Pi bi ⊗ ai ∈ L∞(G) ⊗eh L∞(G). Suppose furthermore that each bi ∈ D(S)∗, and Pi S(b∗ i ) < ∞. Then, for ξ, η ∈ L2(G), i )∗S(b∗ xλ(ωξ,η) = λ(cid:16)Xi ωaiξ,S(b∗ i )η(cid:17), the sum converging absolutely in L1( G). Proof. Let ω1, ω2 ∈ L1(G) be given by Proposition 4.3. Let ξ = Λ(λ(ω1)), η = T ∗ Λ(λ(ω2)); recall that such choices are both dense in L2(G). Then ωξ,η ∈ L1( G) ∩ L2(G) and Λ(λ(ωξ,η)) = Λ(λ(ω1ω2)). For each i, let ξi = ai(ξ) and ηi = S(b∗ i )(η). Then Xi Similarly, i ai, ωξ,ξi ≤(cid:13)(cid:13)(cid:13)Xi S(b∗ i )∗S(b∗ ha∗ kξik2 =Xi kηik2 ≤(cid:13)(cid:13)(cid:13)Xi Xi a∗ i ai(cid:13)(cid:13)(cid:13)kξk2. i )(cid:13)(cid:13)(cid:13)kηk2. It follows that Pi ωξi,ηi converges absolutely in L1( G). By Proposition 4.2, ηi = T ∗ Λ(λ(biω2)), and also ξi = ai Λ(λ(ω1)). By Proposition 4.3, Thus, for y ∈ nϕ, Λ(λ(ωξi,ηi)) = Λ(cid:0)λ(cid:0)(aiω1)(biω2)(cid:1)(cid:1). Xi (cid:0)Λ(λ(ωξi,ηi))(cid:12)(cid:12)Λ(y)(cid:1) =Xi (cid:0)Λ(cid:0)λ(cid:0)(aiω1)(biω2)(cid:1)(cid:1)(cid:12)(cid:12)Λ(y)(cid:1) hy∗, (aiω1)(biω2)i =Xi =Xi = h∆(y∗x), ω1 ⊗ ω2i = hy∗, x(ω1ω2)i =(cid:0) Λ(λ(x(ω1ω2)))(cid:12)(cid:12)Λ(y)(cid:1). h∆(y∗)(ai ⊗ bi), ω1 ⊗ ω2i Thus, Λ(cid:0)xλ(ωξ,η)(cid:1) = xΛ(λ(ω1ω2)) = Λ(λ(x(ω1ω2))) =Xi Λ(λ(ωξi,ηi)), which completes the proof, as Λ injects, and such ξ, η are dense. i )∗S(b∗ We remark that we could weaken the hypothesis that σ∆(x) ∈ L∞(G) ⊗eh L∞(G) to just requiring that ∆(x) = Pi ai ⊗ bi, the sum converging σ-weakly in L∞(G)⊗L∞(G), and with Pi a∗ i ai < ∞ and Pi S(b∗ We can interpret this result in terms of left multipliers or centralisers on L1( G). We shall follow [14] (see also [22], but be aware of differing language). A left multiplier of L1( G) is a linear map L : L1( G) → L1( G) with L(ω1 ω2) = L(ω1)ω2 for each ω1, ω2 ∈ L1( G). If additionally L is completely bounded, then an equivalent condition is that the adjoint L∗ : L∞( G) → L∞( G) satisfies ∆L∗ = (L∗ ⊗ ι) ∆. i ) < ∞. 13 Given data as in the above theorem, we may define a σ-weakly continuous linear map L∗ : L∞( G) → B(L2(G)) by L∗(x) =Xi S(b∗ i )∗xai (x ∈ L∞( G)). (2) Then clearly the preadjoint L : B(L2(G))∗ → L1( G) satisfies λ(Lωξ,η) = xλ(ωξ,η) for all ξ, η ∈ L2(G). Thus L factors through the quotient map B(L2(G))∗ → L1( G) and induces a map on L1( G), still denoted by L. Hence we can regard L∗ as a map on L∞( G). We record the following for use later. Proposition 4.6. With notation as above, L∗ ∈ CB(L∞( G)), kL∗kcb ≤(cid:13)(cid:13)(cid:13)Xi S(b∗ i )∗S(b∗ 1/2(cid:13)(cid:13)(cid:13)Xi i )(cid:13)(cid:13)(cid:13) a∗ i ai(cid:13)(cid:13)(cid:13) 1/2 , and (L∗ ⊗ ι)( W ) = (1 ⊗ x) W . Proof. The norm estimate is a standard calculation from the definition of L∗ given by (2). The last assertion follows from [14, Proposition 2.4]; indeed, this is just a calculation using the definition of λ and that xλ(ωξ,η) = λ(Lωξ,η). We remark that it follows from [14] that x ∈ C b(G) (compare with Theorem 4.9 below) and that x ∈ D(S). 4.3 When we get an invertible corepresentation We will work in a little generality, and deal with possibly degenerate homomorphisms. Given a homomorphism π : A → B(H) of a Banach algebra A on a Hilbert space H, the essential space of π is He, the closed linear span of {π(a)ξ : a ∈ A, ξ ∈ H}. The following shows that if π and π∗ are completely bounded, then Vπ is an invertible operator (suitably interpreted in the degenerate case), and thus we have a converse to Theorem 3.3. Theorem 4.7. Let π : L1(G) → B(H) be a completely bounded representation such that π∗ extends to a completely bounded representation. Letting He be the essential space of π, we have that VπVπ = VπVπ is a (not necessarily orthogonal) projection of L2(G) ⊗ H onto L2(G) ⊗ He. Furthermore, both Vπ and Vπ have ranges equal to L2(G) ⊗ He. Hence Vπ and Vπ restrict to operators on L2(G) ⊗ He, and are mutual inverses in B(L2(G) ⊗ He). Proof. Fix α, β ∈ H. By Proposition 3.2, we know that Vπ = V ∗ the discussion before Theorem 4.5, fi,α)∗ =Xi β,α)∗ =Xi β,fi)∗ ⊗ (T π π , and thus T π T π fi,β ⊗ T π α,β = (T π β,α)∗. As in α,β) = σ∆(T π σ∆(T π α,fi. (T π As π∗ is completely bounded, by Proposition 4.4, (T π fi,β)∗ = T π β,fi ∈ D(S) with S(cid:0)(T π fi,β)∗(cid:1) = S(cid:0)T π β,fi(cid:1) = (T π∗ fi,β)∗ = T π β,fi, 14 where the last equality uses Proposition 3.2. Moreover, Xi S(cid:0)(T π fi,β)∗(cid:1)∗ S(cid:0)(T π fi,β(T π∗ T π∗ fi,β)∗ =Xi (ι ⊗ ωfi,β)(Vπ∗)(cid:0)(ι ⊗ ωβ,fi)(Vπ∗)(cid:1)∗ π∗) = (ι ⊗ ωβ)(Vπ∗V ∗ π∗), (ι ⊗ ωfi,β)(Vπ∗)(ι ⊗ ωfi,β)(V ∗ fi,β)∗(cid:1) =Xi =Xi and similarly Pi(T π tion 4.6 to conclude that if we define α,fi)∗T π α,fi = (ι ⊗ ωα)(VπV ∗ π ). We can hence apply Theorem 4.5 and Proposi- L∗ : L∞( G) → B(L2(G)); x 7→Xi T π∗ fi,β xT π α,fi, (3) then L∗ ∈ CB(L∞( G)), and (L∗ ⊗ ι)( W ) = (1 ⊗ T π (ι ⊗ L∗)(W ∗)W = T π α,β ⊗ 1. That is, α,β ⊗ 1 =Xi (1 ⊗ T π∗ T π fi,β)W ∗(1 ⊗ T π α,fi)W =Xi (1 ⊗ T π∗ fi,β)∆(T π α,fi). α,β) W . As W = σW ∗σ, this shows that So, let ω1, ω2 ∈ L1(G), and consider Xi fi,β)∆(T π α,fi), ω1 ⊗ ω2i =Xi α,fi, ω1(cid:0)ω2T π∗ hT π fi,β(cid:1)π(ω1)α(cid:12)(cid:12)fi(cid:1) =Xi hT π∗ π(ω1)α,fi, ω2i h(1 ⊗ T π∗ =Xi (cid:0)π(cid:0)ω2T π∗ = h(ι ⊗ ωπ(ω1)α,β)(Vπ∗Vπ), ω2i. fi,βT π fi,β(cid:1)i =Xi (cid:0)π(cid:0)ω1(cid:0)ω2T π∗ fi,β(cid:1)(cid:1)α(cid:12)(cid:12)fi(cid:1) However, we know that this is equal to hT π α,β ⊗ 1, ω1 ⊗ ω2i. So we conclude that (ι ⊗ ωπ(ω)α,β)(Vπ∗Vπ) = hT π α,β, ωi1 =(cid:0)π(ω)α(cid:12)(cid:12)β(cid:1)1 (ω ∈ L1(G), α, β ∈ H). Letting He be the essential space of π, we conclude that Vπ∗Vπ = 1 on L2(G) ⊗ He ⊆ L2(G) ⊗ H. ♯ (G) is dense in L1(G), we see that the essential space for π∗ also equals He. Letting (ei) be an orthonormal basis for L2(G), a simple calculation shows that As L1 Vπ∗(ξ ⊗ γ) =Xi ei ⊗ π∗(ωξ,ei)(γ) ∈ L2(G) ⊗ He (ξ ∈ L2(G), γ ∈ H). So Vπ∗ has range contained in L2(G) ⊗ He. It follows that Vπ∗Vπ is actually a (not necessarily orthogonal) projection from L2(G) ⊗ H onto L2(G) ⊗ He. Furthermore, the range of Vπ∗ must actually be L2(G) ⊗ He. Now set φ = π∗, so φ is also a completely bounded homomorphism L1(G) → B(H). Thus the same argument now applied to φ shows that Vφ∗V φ = VπV ∗ π∗ = VπVπ is a (not necessarily orthogonal) projection from L2(G) ⊗ H onto L2(G) ⊗ He where He is the essential space of φ, which agrees with the essential space for φ∗ = π. We also conclude that the range of Vφ∗ = Vπ is L2(G) ⊗ He. Following [27, Section 4], the opposite quantum group to G is Gop, where L∞(Gop) = L∞(G) and ∆op = σ∆. That is, we reverse the multiplication in L1(G) to get L1(Gop). Then Rop = R and 15 t = τ−t for each t. Thus Sop = Ropτ op τ op and ω ∈ L1 ♯ (Gop), −i/2 = Rτi/2 = S −1 = ∗ ◦ S ◦ ∗. For x ∈ D(Sop) = D(S)∗, It follows that ω∗ ∈ L1 hx, ω♯,opi = hSop(x)∗, ωi = hS(x∗), ωi = hS(x∗)∗, ω∗i. ♯ (G), as for each y ∈ D(S), setting x = y∗ ∈ D(S)∗, we see that hS(y)∗, ω∗i = hS(x∗), ωi = hx, ω♯,opi = hy, (ω♯,op)∗i. So (ω∗)♯ = (ω♯,op)∗, and reversing this argument shows that L1 ♯ (G) = L1 ♯ (Gop)∗. Thus, if we set φ = π : L1(Gop) → B(H) then φ is a completely bounded representation, and φ = π, so that the essential space of φ is He. For ω ∈ L1 ♯ (G), and φ∗(ω) = φ(ω♯,op)∗ = π(((ω∗)♯)∗)∗ = π((ω∗)♯) = π(ω). Hence, as π is completely bounded, also φ∗ is completely bounded; and so φ∗ = π and similarly φ = π∗. Applying the previous argument to φ, we conclude that Vφ∗V φ = VπVπ is a (not necessarily orthogonal) projection onto L2(G) ⊗ He. We also see that the range of Vφ∗ = Vπ is L2(G) ⊗ He. Finally, we also have that VφV φ = VπVπ∗ is a (not necessarily orthogonal) projection onto L2(G) ⊗ He. It follows that ♯ (Gop), we have that ω∗ ∈ L1 ker(cid:0)VπVπ(cid:1) = Im(V ∗ π V ∗ π )⊥ = Im(Vπ∗Vπ)⊥ =(cid:0)L2(G) ⊗ He(cid:1)⊥ , and analogously, also ker(VπVπ) = Im(VπVπ∗)⊥ = (L2(G)⊗ He)⊥. So we conclude that ker(VπVπ) = ker(VπVπ). So VπVπ and VπVπ are projections with the same range and kernel, and hence are equal. The above theorem says that if both π and π∗ are completely bounded and non-degenerate, then Vπ is an invertible corepresentation of L∞(G), and, informally, (S ⊗ ι)(Vπ) = V −1 π . This is well-known in, for example, the theory of algebraic compact quantum groups (compare with [41, Proposition 3.1.7(iii)]). It is interesting that our arguments seem to require a lot of structure– for example, by using the duality theory for locally compact quantum groups. From Proposition 4.6 and the proof of Theorem 4.7, we obtain the following corollary, show- α,β considered in Theorem 4.7 naturally induce completely bounded ing that the co-efficients T π multipliers on L1( G). Corollary 4.8. Let π : L1(G) → B(H) be a completely bounded representation such that π∗ extends to a completely bounded representation. Then for any α, β ∈ H, the co-efficient T π α,β represents a completely bounded left multiplier L : L1( G) → L1( G), determined by λ(Lω) = T π α,β λ(ω) (ω ∈ L1( G)). Moreover, kLkcb ≤ kπkcbkπ∗kcbkαkkβk. Proof. The fact that L is a completely bounded left multiplier was already observed in the proof of Theorem 4.7. Moreover, from Proposition 4.6 and the proof of Theorem 4.7, we deduce that kLkcb = kL∗kcb ≤ k(ι ⊗ ωβ)(Vπ∗V ∗ π∗)k1/2k(ι ⊗ ωα)(V ∗ π Vπ)k1/2 ≤ kπkcbkπ∗kcbkαkkβk. 16 When G is a locally compact group, the above result is classical and is due to de Canniere and Haagerup [15, Theorem 2.2]. (See also [7, Theorem 3] for when G is the dual of a locally compact group.) α,β ∈ C b(G), s 7→(cid:0)π0(s)αβ(cid:1). Let G be a locally compact group, π : L1(G) → B(H) a bounded representation, and π0 : G → B(H) the uniformly bounded representation associated to π. The correspondence π ↔ π0 given in Section 1 implies that each co-efficient operator T π α,β ∈ L∞(G) is actually (a.e.) equal to the continuous function ϕπ Let B0(H) be the compact operators on H, so that M(B0(H)) = B(H). We hence get the strict topology on B(H), where a bounded net (Tα) is strictly-null if and only if the nets (Tαx) and (xTα) are norm-null, for each x ∈ B0(H). One consequence of the above correspondence is that the corepresentation Vπ ∈ L∞(G)⊗B(H) is actually a member of C b str(G, B(H)), the space of bounded continuous maps G → B(H), where B(H) is given the strict topology. Indeed, Vπ corresponds to the map G → B(H); s 7→ π0(s). str(G,B(H)) is the multiplier algebra M(C0(G) ⊗ B0(H)). This follows, as M(C0(G) ⊗ B0(H)) is isomorphic to str(G,B(H)). Indeed, given F ∈ C b C b str(G,B(H)) and f ∈ C0(G,B0(H)) = C0(G)⊗B0(H)), clearly the pointwise products F f, f F are members of C0(G,B0(H)), and so F is a multiplier, and a partition of unity argument shows that every multiplier arises in this way. The following gives a summary of the results of this section. The argument about multiplier When G is a locally compact quantum group, the natural replacement for C b algebras should be compared with [50, Section 4], which in turn is inspired by [2, Page 441]. Theorem 4.9. Let G be a locally compact quantum group, and let π : L1(G) → B(H) be a non-degenerate, completely bounded representation. Then the following are equivalent: 1. π extends to a completely bounded map L1(G) → B(H); 2. Vπ is invertible. In this case, V −1 π = Vπ, and Vπ ∈ M(C0(G) ⊗ B0(H)). π Proof. The equivalences have been established by Theorem 3.3 and Theorem 4.7. Suppose now that V −1 exists. Recall that the coproduct on L∞(G) is implemented by the fundamental unitary by ∆(x) = W ∗(1 ⊗ x)W for x ∈ L∞(G). Furthermore, W ∈ M(C0(G) ⊗ B0(L2(G))). Then π,23 ∈ M(cid:0)C0(G) ⊗ B0(L2(G)) ⊗ B0(H)(cid:1). Vπ,13 = (∆ ⊗ ι)(Vπ)V −1 π,23 = W ∗ 12Vπ,23W12V −1 Thus Vπ ∈ M(C0(G) ⊗ B0(H)) as claimed. 4.4 Application to (co)isometric corepresentations A corepresentation U ∈ L∞(G)⊗B(H) is a coisometric corepresentation if U is a coisometry, namely UU ∗ = 1. As ∆ is a ∗-homomorphism, we see that U ∗ is a co-anti-representation. Proposition 4.10. Let U ∈ L∞(G)⊗B(H) be a coisometric corepresentation, associated to π : L1(G) → B(H). Then π is a completely bounded representation, with Vπ = U ∗. Proof. This follows almost exactly as the proof of Theorem 3.3. Let α, β ∈ H, let (fi) be an orthonormal basis of H, and set xi = (ι ⊗ ωfi,β)(U), yi = (ι ⊗ ωα,fi)(U ∗). 17 Then as before, we see that Xi ∆(xi)(1 ⊗ yi) = (ι ⊗ ι ⊗ ωα,β)(U13U23U ∗ 23) = (ι ⊗ ωα,β)(U) ⊗ 1, and similarly Xi (1 ⊗ xi)∆(yi) = (ι ⊗ ι ⊗ ωα,β)(U23U ∗ 23U ∗ 13) = (ι ⊗ ωα,β)(U ∗) ⊗ 1. Hence (ι⊗ ωα,β)(U) ∈ D(S) with S((ι⊗ ωα,β)(U)) = (ι⊗ ωα,β)(U ∗). As before, it now follows that π is completely bounded with Vπ = U ∗. The following is now immediate from Theorem 4.7. Corollary 4.11. Let U ∈ L∞(G)⊗B(H) be a coisometric corepresentation. Then U is unitary. Corollary 4.12. Let U ∈ L∞(G)⊗B(H) be an isometric corepresentation. Then U is unitary. Proof. As in the proof of Theorem 4.7, following [27, Section 4], the opposite quantum group to G is Gop, where L∞(Gop) = L∞(G) and ∆op = σ∆. It follows that corepresentations of G are co-anti-representations of Gop, and vice versa. The proof follows from the observation that U ∗ is thus a coisometric corepresentation of Gop. An application of this result is to the theory of induced corepresentations in the sense of Kustermans, [25]. In this paper, a theory is developed allowing one to “induce” a corepresentation from a “smaller” quantum group to a larger one. However, in general the resulting corepresentation is only a coisometry, and not unitary (see [25, Notation 5.3]), and under a further “integrability” condition, and with an elaborate argument, it is shown that this corepresentation is unitary, [25, Proposition 7.5]. Our result shows that a further condition is not necessarily, and the induced corepresentation is always unitary. 5 The similarly problem We recall from [4, Definition 3.2] (for example) that G is amenable if there is a state m on L∞(G) such that hm, (ι ⊗ ω)∆(x)i = hm, xih1, ωi for x ∈ L∞(G), ω ∈ L1(G). Using the unitary antipode, we see that we could have equivalently used the obvious “left” variant of this definition instead. We call m a right invariant state. Based on the results of the previous section, we are now a position to state the “similarity problem” for locally compact quantum groups: Suppose G has the property that every completely bounded homomorphism π : L1(G) → B(H), with π∗ also extending to a completely bounded ho- momorphism, is similar to a ∗-homomorphism. Then is G necessarily amenable? By Theorem 4.9, this is equivalent to asking if having all invertible corepresentations being similar to unitary corep- resentations forces G to be amenable? When G is commutative, this reduces to the standard conjecture that a unitarisable locally compact group G is amenable. When G is cocommutative, [7] proves this conjecture in the affirmative– there is of course the stronger conjecture that actually, no condition on π∗ is needed, and we provide more evidence for this in Section 8 below. In the following, we shall show how to use an invariant state to “average” an invertible corep- resentation to a unitary corepresentation. This is well-known in the compact quantum group case (see, for example, [51, Theorem 5.2]). The non-compact case is similar, but we provide the details, as we are unaware of a good reference. 18 Theorem 5.1. Let G be an amenable locally compact quantum group, and let π : L1(G) → B(H) be a non-degenerate homomorphism. The following are equivalent: 1. π is similar to a ∗-homomorphism; 2. both π and π extend to completely bounded (anti-)homomorphisms L1(G) → B(H). Proof. That (1) implies (2) is Proposition 3.4. Now suppose that (2) holds, and that π is non-degenerate, so by Theorem 4.7, Vπ is invertible. Let m ∈ L∞(G)∗ be a right invariant state. Set T = (m ⊗ ι)(V ∗ π Vπ) ∈ B(H). This means (by definition) that hT, ωi = hm, (ι ⊗ ω)(V ∗ As Vπ is invertible, so is the positive operator V ∗ π Vπ ≥ ǫ1. It follows that for any ω ∈ B(H)+ V ∗ ∗ , hT, ωi = hm, (ι ⊗ ω)(V ∗ π Vπ)i ≥ hm, (ι ⊗ ω)(ǫ1)i = ǫhω, 1i. π Vπ)i for ω ∈ B(H)∗. π Vπ, and so, for some ǫ > 0, we have that Hence T ≥ ǫ1, and so we may define V = (1 ⊗ T 1/2)Vπ(1 ⊗ T −1/2) ∈ L∞(G)⊗B(H). Then, as ∆ is a unital ∗-homomorphism, (∆ ⊗ ι)(V ) = (1 ⊗ 1 ⊗ T 1/2)Vπ,13Vπ,23(1 ⊗ 1 ⊗ T −1/2) = (1 ⊗ 1 ⊗ T 1/2)Vπ,13(1 ⊗ 1 ⊗ T −1/2)(1 ⊗ 1 ⊗ T 1/2)Vπ,23(1 ⊗ 1 ⊗ T −1/2) = V13V23. So V is a corepresentation, say inducing a homomorphism φ : L1(G) → B(H). For ω ∈ L1(G), we have that φ(ω) = (ω ⊗ ι)(V ) = T 1/2π(ω)T −1/2, and so φ is similar to π. We next show that V is unitary. For ω1 ∈ L1(G) and ω2 ∈ B(H)∗, we may define ψ ∈ B(H)∗ by hR, ψi = hV ∗ π (1 ⊗ R)Vπ, ω1 ⊗ ω2i (R ∈ B(H)). Then we see that (ι ⊗ ψ)(V ∗ π Vπ) = (ι ⊗ ω1 ⊗ ω2)(cid:0)V ∗ π,23V ∗ π,13Vπ,13Vπ,23(cid:1) = (ι ⊗ ω1 ⊗ ω2)(cid:0)(∆ ⊗ ι)(V ∗ π Vπ)(cid:1). Thus hT, ψi = hm, (ι ⊗ ψ)(V ∗ π Vπ)(cid:1)i π Vπ)i = hm, (ι ⊗ ω1 ⊗ ω2)(cid:0)(∆ ⊗ ι)(V ∗ = hm, (ι ⊗ ω1)∆((ι ⊗ ω2)(V ∗ π Vπ))i = hm, (ι ⊗ ω2)(V ∗ π Vπ)ih1, ω1i = h1 ⊗ T, ω1 ⊗ ω2i. As ω1, ω2 were arbitrary, we conclude that V ∗ V ∗V = (1 ⊗ T −1/2)V ∗ as required. π (1 ⊗ T )Vπ = 1 ⊗ T . Thus π (1 ⊗ T 1/2)(1 ⊗ T 1/2)Vπ(1 ⊗ T −1/2) = (1 ⊗ T −1/2)(1 ⊗ T )(1 ⊗ T −1/2) = 1, Arguing as in Theorem 3.3, it is now easy to see that φ is a ∗-homomorphism, because V is unitary. 19 We can also deal with the case of degenerate representations of L1(G). The proof of [7, Propo- sition 6] actually shows the following: Lemma 5.2. Let A be a ∗-algebra, and let π : A → B(H) be a homomorphism. Let πe : A → B(He) be the subrepresentation given by restricting to the essential space He. Suppose there is a (not necessarily orthogonal) projection Q from H to He such that π(a)Q = π(a) for all a ∈ A. Then π is similar to ∗-representation if and only if πe is similar to a ∗-representation. Theorem 5.3. Let G be an amenable locally compact quantum group, and let π : L1(G) → B(H) be a homomorphism. The following are equivalent: 1. π is similar to a ∗-representation; 2. both π and π extend to completely bounded (anti-)representations L1(G) → B(H). Proof. That (1) implies (2) is again Proposition 3.4. If (2) holds, then Theorem 4.7 shows that Vπ restricts to an invertible operator on L2(G) ⊗ He; that is, Vπe is invertible. So Theorem 5.1 shows that πe is similar to a ∗-representation. By Lemma 5.2 it suffices to construct a projection Q : H → He with π(ω)Q = π(ω) for ω ∈ L1(G). By Theorem 4.7, P = VπVπ is a projection of L2(G) ⊗ H onto L2(G) ⊗ He. The proof actually shows that ker P = ker VπVπ = (L2(G)⊗ He)⊥ = L2(G)⊗ H ⊥ e = {0}, and so He ∩ H ⊥ e = {0}. Similarly, L2(G) ⊗ He + L2(G) ⊗ H ⊥ e = H. It follows that there is a projection Q of H onto He with ker Q = H ⊥ e . Now, α ∈ H ⊥ e . Thus L2(G)⊗He∩L2(G)⊗ H ⊥ e = L2(G) ⊗ H, and so He + H ⊥ if and only if e 0 = (απ(ω)β) = (απ(ω∗)∗β) = (π(ω∗)αβ) (ω ∈ L1(G), β ∈ H). It follows that for ω ∈ L1(G), α ∈ ker Q, we have that π(ω)α = 0 = π(ω)Qα. Clearly π(ω)Q = π(ω) on He, and thus it follows that π(ω)Q = π(ω) on H, as required. 6 Special cases In this section, we take an approach which is closer in spirit to [7]. A cost is that we shall have to assume that G is coamenable. Recall that G is coamenable if L1( G) has a bounded approximate identity. This is equivalent to a large number of other conditions, see [4, Theorem 3.1]. In particular, the proof of this reference shows that L1( G) has a contractive approximate identity It is conjectured that the coamenability of G is equivalent to G being amenable, in this case. but except when G is discrete (see [42]) it is only known that G coamenable implies that G is amenable. A benefit to the approach of this section is that in special cases (as for G being a SIN group in [7]) we can show that if π : L1(G) → B(H) is completely bounded, then π is similar to a ∗-homomorphism, without assumption about π. Theorem 6.1. Let G be a locally compact quantum group such that G is coamenable. Let π : L1(G) → B(H) be a completely bounded homomorphism, such that π extends a completely bounded map L1(G) → B(H). Then π is similar to a ∗-homomorphism. Proof. Let α, β ∈ H and ξ, η ∈ L2(G). From Corollary 4.8, T π application of Theorem 4.5 to T π λ(cid:0)ωaiξ,S(b∗ i )η(cid:1), λ(ωξ,η) ∈ λ(L1( G)). Moreover, an λ(ωξ,η) =Xi β,α β,α shows that T π β,α 20 where, if (fi) is an orthonormal basis for H, In particular, Xi kωaiξ,S(b∗ i )ηk ≤(cid:16)Xi ≤(cid:13)(cid:13)(cid:13)Xi ai = T π β,fi, S(b∗ fi,α)∗(cid:1) = T π α,fi. i ) = S(cid:0)(T π kaiξk2(cid:17)1/2(cid:16)Xi 1/2(cid:13)(cid:13)(cid:13)Xi i ai(cid:13)(cid:13)(cid:13) S(b∗ a∗ kS(b∗ i )ηk2(cid:17)1/2 i )(cid:13)(cid:13)(cid:13) 1/2 i )∗S(b∗ kξkkηk ≤ kξkkηkkαkkβkkVπkkVπk = kξkkηkkαkkβkkπkcbkπkcb. Let (ωk) be a contractive approximate identity for L1( G). exists ω′ k ∈ L1( G) with It follows that for each k, there (T π α,β)∗λ(ωk) = T π β,α λ(ωk) = λ(ω′ k) with kω′ kk ≤ kαkkβkkπkcbkπkcb. As λ(L1( G)) is σ-weakly dense in L∞(G), it follows that for ω ∈ L1(G), hT π α,β, ωi = h(T π k), ω∗i = lim The final equality follows by a simple calculation, see [13, Lemma 8.10]. Thus α,β)∗λ(ωk), ω∗i = lim α,β)∗, ω∗i = lim k hλ(ω′ k h(T π k hλ(ω)∗, ω′ ki. (π(ω)αβ) = hT π α,β, ωi ≤ kαkkβkkπkcbkπkcbkλ(ω)k. As λ(L1(G)) is dense in C0( G), by continuity, there is a homomorphism φ : C0( G) → B(H) with kφk ≤ kπkcbkπkcb, φ ◦ λ = π. As G is coamenable, it follows that G is amenable (see [4, Theorem 3.2]) and that hence C0( G) is a nuclear C∗-algebra (see [4, Theorem 3.3]). The similarly problem has an affirmative answer for nuclear C ∗-algebras (see [9, Theorem 4.1]) so φ is similar to a ∗-homomorphism. As λ is a ∗-homomorphism on L1 ♯ (G)) is also similar to a ∗-homomorphism. The above reference to [9] needed that φ (that is, π) is non-degenerate. But as C0( G) has a bounded approximate identity, we can simply follow [7, Proposition 6] to deal with the case when π is degenerate. ♯ (G), it follows that π (restricted to L1 The above proof is interesting because in special cases, it allows us remove the hypothesis that π is completely bounded (or even bounded on all of L1(G)). This idea was used for SIN groups in [7], the key point being that if G is SIN, then the Plancheral weight on V N(G) can be approximated by tracial states in A(G). It seems unlikely that many genuinely “quantum” groups will have this property, and so we shall restrict attention to the compact case, where we have many non-trivial examples. Recall that a locally compact quantum group G is compact if C0(G) is unital. Here, the existence of Haar weights follows from simple axioms, see [49]. We shall be interested in the case when the Haar state is tracial– by [49, Theorem 2.5] this is equivalent to G being a compact Kac algebra. See [3, 47, 48] for some genuinely “quantum” examples of compact Kac algebras. 21 Theorem 6.2. Let G be a compact Kac algebra, and let π : L1(G) → B(H) be completely bounded homomorphism. Then π is similar to a ∗-homomorphism. Proof. As the dual G is discrete, L1( G) is unital. Indeed, if Λ : L∞(G) → L2(G) is the GNS map, then it is not hard to show that the state ωΛ(1) is the unit of L1( G). Following the proof of λ(ωΛ(1)) = λ(ω) for some ω whose norm is controlled. As Theorem 6.1, we wish to show that T π β,α S = R in this case, we really have to show that =Xi ϕ(cid:16)R(cid:0)(T π fi,α)∗(cid:1)∗ R(cid:0)(T π fi,α)∗(cid:1)(cid:17) =Xi ϕ(cid:16)(T π fi,α)∗T π fi,α(cid:17) < ∞. 2 fi,α)∗(cid:1)Λ(1)(cid:13)(cid:13)(cid:13) Xi (cid:13)(cid:13)(cid:13)R(cid:0)(T π ϕ(cid:16)T π Xi fi,α(T π where ϕ is the (normal, tracial) Haar state on L∞(G), which satisfies ϕ ◦ R = ϕ. However, as ϕ is a trace, this sum is fi,α)∗(cid:17) =Xi ϕ(cid:16)(T π α,fi)∗T π α,fi(cid:17) = ϕ(cid:0)(ι ⊗ ωα)(V ∗ π Vπ)(cid:1) ≤ kπk2 cbkαk2. Then we can just continue as in the proof of Theorem 6.1. The above proof could be adapted to a general Kac algebra provided that we can find a bounded approximate identity (ωξk,ηk) for L1( G) such that, for each k, we have that ωηk is tracial (by scaling ηk and ξk, we can suppose that kηkk = 1 for each k, but we also need at least that supk kξkk < ∞). In particular, we can only apply this to G if G is a SIN group, by using the proof of [40, Proposition 3.2] (and so we have reproved [7, Theorem 20]). However, this argument will not extend to other classes of groups. For a general compact quantum group, we can consider the Hopf ∗-algebra A ⊆ C0(G) formed from the matrix coefficients of irreducible representations. On this algebra, we have reasonably explicit formulae for the antipode S and the Haar weight ϕ (see [41, Section 3.2] for example). A little calculation shows that if, in reasonable generality, we have something like Xi ϕ(cid:0)S(b∗ i )∗S(b∗ i )(cid:1) =Xi ϕ(bib∗ i ), then already ϕ is a trace. So to extend the above result to general compact quantum groups (if it is true at all!) probably needs a new idea. 7 The importance of complete boundedness Up to this point, we have restricted our attention to completely bounded representations π : L1(G) → B(H), and derived some interesting structure results for such representations. A natural question which arises is: what can be said about bounded representations π which are not com- pletely bounded, if such representations even exist? In [8] Choi and the 3rd named author answered the existence part of this question by constructing a bounded, but not completely bounded, repre- sentation π : A(G) → B(H), for any group G which contains F2 as a closed subgroup. In particular, for G = F2, every completely bounded representation of A(G) is similar to a ∗-representation ([7, Theorem 20], see also Theorem 6.2 above) but there are bounded representations of A(G) not similar to ∗-representations. In this section, we will consider free products of compact quantum groups, and Khintchine in- equalities for reduced free products, to construct examples of non-commutative/non-cocommutative compact quantum groups G such that L1(G) admits bounded representations π : L1(G) → B(H) which are not similar to ∗-representations. To be precise, we will show the following. 22 Theorem 7.1. Let G be a non-trivial compact quantum group, and let eG = ∗i∈NGi be the compact quantum group free product of countably many copies of G (or let eG = T ∗ G). Then there exists a bounded representation π : L1(eG) → B(H) which is not completely bounded. In particular, π is not similar to a ∗-representation. For completely bounded representations π : L1(G) → B(H) (with π∗ also completely bounded), a key result in the preceding sections was Corollary 4.8, which showed that each co-efficient operator T π α,β ∈ L∞(G) (α, β ∈ H) induces a completely bounded multiplier of the dual convolution algebra L1( G). The following corollary shows that for non-completely bounded representations, there is no hope of generalising Corollary 4.8. Corollary 7.2. Let eG and π : L1(eG) → B(H) be as in the statement and proof of Theorem 7.1. λ(ω) (ω ∈ L1(beG)) Then there exist vectors α, β ∈ H with the property that the map λ(ω) 7→ T π does not induce a bounded left multiplier of L1(beG). Proof of Corollary 7.2. We use an argument similar to the proof of [15, Theorem 2.3]. Suppose, to get a contradiction, that each co-efficient T π α,β induces a multiplier of L1(beG), as in Corollary 4.8. α,β Then for each α, β ∈ H, as L1(beG) is unital, there exists a unique ω π α,β = α,β). Define a sesquilinear map Q : H × H → L1(beG) by setting Q(α, β) = ω π λ(ω π α,β. We claim that Q is separately continuous. To see this, fix β ∈ H and suppose limn→∞ kα − αnk = 0 and limn→∞ Q(αn, β) = ω ∈ L1(beG). Then for any ω ∈ L1(eG), n→∞(cid:0)π(ω)αn(cid:12)(cid:12)β(cid:1) n→∞hλ(Q(αn, β)), ωi = lim n→∞hT π =(cid:0)π(ω)α(cid:12)(cid:12)β(cid:1) = hT π α,β ∈ L1(beG) such that T π Therefore λ(ω) = λ(Q(α, β)), giving ω = Q(α, β), and the closed graph theorem implies that Q is continuous in the first variable. The same argument applies to the second variable, proving the claim. α,β, ωi = hλ(Q(α, β)), ωi. hλ(ω), ωi = lim αn,β, ωi = lim Since any separately continuous sesquilinear map is bounded (see for example [35, Theorem beG) ≤ Dkαkkβk for all α, β ∈ H. As a 2.17]), there is a constant D > 0 such that kω π consequence, we have α,βkL1( (cid:0)π(ω)αβ(cid:1) = hλ(ω π ≤ kω π α,β), ωi = hλ(ω∗)∗, ω π α,βi α,βkkλ(ω∗)∗k ≤ Dkαkkβkkλ(ω∗)k (ω ∈ L1(eG)), showing that π extends to the bounded anti-representation σ : C0(beG) → B(H) determined by σ ◦ λ = π. Since eG is compact, C0(beG) is nuclear, and it follows that σ is automatically completely bounded. But then kπkcb = kπkcb = kσ ◦ λkcb ≤ kσkcb < ∞, contradicting the fact that π is not completely bounded. 7.1 Reduced free products Before proceeding with the proof of Theorem 7.1, let us recall the construction of the reduced free product of a collection of C∗-algebras equipped with distinguished states. For further details see 23 [1], [46] or the introduction to [33]. Let I be an index set, and for each i ∈ I let Ai be a C∗-algebra (assumed unital), and let ϕi be a state on Ai with faithful GNS construction (πi, Hi, ξi). Let DI be the collection of all “words” in I, that is, tuples (i1,· · · , in) with ij 6= ij+1 for each j; we allow the empty word ∅. Set H 0 i1 ⊗· · ·⊗ H 0 as (ij) varies through DI; we interpret the space for ∅ ∈ DI as being a copy of C. Thus H is some sort of generalised Fock space construction. Let Ω ∈ H be the vector 1 ∈ C. in : (ij) ∈ DI, i1 6= i(cid:9)(cid:17) i ⊆ Hi, and let H be the Hilbert space direct sum of H 0 For each i ∈ I there is an obvious isomorphism i1 ⊗ · · · ⊗ H 0 i = ξ⊥ in i ) ⊗(cid:16)M(cid:8)H 0 H ∼= (C ⊕ H 0 ∼= Hi ⊗(cid:16)M(cid:8)H 0 i1 ⊗ · · · ⊗ H 0 in : (ij) ∈ DI, i1 6= i(cid:9)(cid:17). Let Ui be the unitary implementing this isomorphism, and define πi : Ai → B(H); a 7→ U ∗ i (πi(a)⊗ 1)Ui, a faithful ∗-representation of Ai on H. A description of Ui is given in [1], but let us just note that if ai ∈ Ai with ϕi(ai) = 0, then πi(ai)Ω = aiξi ∈ H 0 i . Let A be the algebraic free product of the (Ai), amalgamated over units. Then π = ∗i∈I πi is a faithful ∗-representation of A on B(H). By definition, the reduced free product ∗i∈AAi is the closure of π(A). The vector state induced by Ω is the free product of the states, denoted ∗i∈Iϕi. Notice that (π, H, Ω) is the GNS construction for ∗i∈Iϕi. Denote by A the full free product, which is the completion of A under the norm given by considering all ∗-representations of Ai on a common Hilbert space. If each Ai were a von Neumann algebra, with normal states ϕi, then the (reduced) von Neumann algebraic free product is simply the weak operator closure of ∗i∈IAi in B(H). In [48] Wang studied free products of compact quantum groups (see also [41, Section 6.3]). If for each i we have a compact quantum group Gi, then we form G = ∗i∈I Gi as follows. Let C0(G) be the reduced free product of (C0(Gi), ϕi), where ϕi is the Haar state. For each i let ιi : C0(Gi) → A be the inclusion, and let ∆i be the coproduct on C0(Gi). Set ρi = (ιi ⊗ ιi)∆i : C0(Gi) → A ⊗ A. By the universal property of A, there is a ∗-homomorphism ∆ : A → A ⊗ A such that ∆ιi = ρi for each i. It is easy to see that ∆ is coassociative, and by using the corepresentation theory of compact quantum groups, one can verify the density conditions to show that (A, ∆) is a compact quantum group. Then [48, Theorem 3.8] shows that the Haar state on (A, ∆) is just the free product of the Haar states on each C0(Gi). It follows that the reduced version of A is just C0(G), and thus ∆ drops to a coproduct on C0(G). We require the following non-commutative Khintchine inequality describing the operator space structure of the linear span of a family of centred freely independent operators. See [21, Proposi- tion 7.4] or the introduction to [33]. In fact, we shall only use the k = 1 case, which is attributed to Voiculescu [45]. Theorem 7.3. Let I be an index set, and for each i ∈ I, let Ai be a von Neumann algebra with faithful normal state ϕi. Let (A, ϕ) = (∗i∈IAi,∗i∈Iϕi) be their von Neumann algebraic free product. Then for any xi ∈ Ai with ϕi(xi) = 0, and any finitely supported family {ai}i∈I ∈ Mk(C) (k ∈ N), 24 we have a∗ i aiϕi(x∗ maxn max ≤(cid:13)(cid:13)(cid:13)Xi∈I ≤ 3 maxn max i∈I kai ⊗ xikMk(C)⊗A,(cid:13)(cid:13)(cid:13)Xi∈I ai ⊗ xi(cid:13)(cid:13)(cid:13)Mk(C)⊗A i∈I kai ⊗ xikMk(C)⊗A,(cid:13)(cid:13)(cid:13)Xi∈I Mk(C) 1/2 i xi)(cid:13)(cid:13)(cid:13) i xi)(cid:13)(cid:13)(cid:13) aia∗ ,(cid:13)(cid:13)(cid:13)Xi∈I ,(cid:13)(cid:13)(cid:13)Xi∈I 1/2 Mk(C) a∗ i aiϕi(x∗ i ϕi(xix∗ 1/2 i )(cid:13)(cid:13)(cid:13) Mk(C)o aia∗ i ϕi(xix∗ i )(cid:13)(cid:13)(cid:13) 1/2 Mk(C)o. Now let G be a non-trivial compact quantum group, set Gi = G for each i ∈ N, and let eG = ∗i∈NGi be the free product. (We will show how to modify the following arguments to address the case of eG = T ∗ G in Remark 7.7.) Fix once and for all a non-trivial irreducible unitary representation U ∈ L∞(G)⊗Md(C) and for each i ∈ N, let U i ∈ L∞(eG)⊗Md(C) be the irreducible unitary representation of eG induced by the inclusion L∞(Gi) ֒→ L∞(eG). Note that the U i’s are i = {(ι⊗ρ)U i : ρ ∈ Md(C)∗} be the co-efficient pairwise inequivalent representations. Finally, let L∞ space of U i and let X be the weak∗-closed linear span of all the L∞ i ’s. Lemma 7.4. There is a constant C > 0 such that kxΩkL2( eG) ≤ kxkL∞( eG) ≤ CkxΩkL2( eG) (x ∈ X). Proof. The lower bound is immediate. To get the upper bound, fix x ∈ X and write x as an L2-convergent series x =Pi∈N xi ∈ X where each xi is a member of L∞ i . As U is non-trivial, the Haar state annihilates all coefficients of U, and so we can apply the k = 1 version of Theorem 7.3 to see that i xi)(cid:17)1/2 ,(cid:16)Xi∈N i ,kxΩkL2( eG),(cid:16)Xi∈N kx∗ L2(Gi)(cid:17)1/2o. i ξik2 kxkL∞( eG) ≤ 3 maxn max = 3 maxn max i kxikL∞( eG),(cid:16)Xi∈N i kxikL∞ i )(cid:17)1/2o ϕi(xix∗ ϕi(x∗ Since the spaces L∞ C1, C2 > 0 such that i are all finite dimensional and isometrically isomorphic, there exist constants kxikL∞ i ≤ C1kxiξikL2(Gi) and kx∗ i ξikL2(Gi) ≤ C2kxiξikL2(Gi) (i ∈ N). The lemma now follows by taking C = 3 max{C1, C2}. Lemma 7.5. The pre-annihilator ⊥X is a closed two-sided ideal in L1(eG). Proof. By linearity, it is enough to show that if ω ∈ ⊥X and x = (ι ⊗ ρ)U i for some i ∈ N and ρ ∈ Md(C)∗, then hx, ωω′i = 0 = hx, ω′ωi for ω′ ∈ L1(eG). However, hx, ωω′i = h∆(x), ω ⊗ ω′i = hU i 13U i 23, ω ⊗ ω′ ⊗ ρi = hU i, ω ⊗ (ω′ ⊗ ι)(U i)ρi = 0, as ρ′ = (ω′ ⊗ ι)(U i)ρ ∈ Md(C)∗ and (ι ⊗ ρ′)(U i) ∈ X. Similarly, hx, ω′ωi = 0. Consider now the space ℓ2(N, Md), which is a Banach algebra under pointwise operations. 25 Proposition 7.6. The map φ : L1(eG) → ℓ2(N, Md); ω 7→(cid:0)(ω ⊗ ι)(U i)(cid:1)i∈N is well-defined, bounded, and is an algebra homomorphism. Furthermore, φ drops to give a iso- morphism between L1(eG)/⊥X and ℓ2(N, Md). Proof. The dual space of ℓ2(N, Md) is ℓ2(N, M ∗ ρi)(U i) ∈ L∞ ϕi(xi) = 0 for each i. As the xiξi are pairwise orthogonal in the Fock space H, it follows that d ). Let ρ = (ρi) ∈ ℓ2(N, M ∗ i=1 xi ∈ X. Observe that xΩ = Pn i . For n ∈ N, let x = Pn d ), and let xi = (ι ⊗ i=1 xiΩ = Pn i=1 xiξi as kxΩk2 = nXi=1 kxiξik2 ≤ nXi=1 kxik2 ≤ nXi=1 kρik2 ≤ kρk2. Letting C be the constant from Lemma 7.4, we see that nXi=1 (cid:12)(cid:12)(cid:12) hρi, (ω ⊗ ι)(U i)i(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12) hxi, ωi(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)hx, ωi(cid:12)(cid:12) nXi=1 ≤ kωkkxk ≤ CkωkkxΩk ≤ Ckωkkρk. As n and ρ were arbitrary, it follows that φ is well-defined, and kφk ≤ C. For each i, as U i is a corepresentation, the map L1(eG) → Md(C); ω 7→ (ω ⊗ ι)(U i) is a homomorphism. It follows that φ is also a homomorphism. Notice that φ(ω) = 0 if and only if (ω ⊗ ι)(U i) = 0 for each i, if and only if ω ∈ ⊥L∞ for each i, if and only if ω ∈ ⊥X. So φ drops to give an injective map L1(eG)/⊥X → ℓ2(N, Md). d ) → (⊥X)⊥ = X is the map (ρi) 7→Pi(ι⊗ρi)(U i). As U is irreducible, d → L2(G), ρ 7→ the map M ∗ Λ((ι ⊗ ρ)(U)) is injective. Thus there is a constant C ′ such that kρk ≤ C ′k(ι ⊗ ρ)(U)ξik for all ρ ∈ M ∗ d → L∞(G), ρ 7→ (ι ⊗ ρ)(U) is injective, and so also the map M ∗ d , i ∈ N. Then, for (ρi) ∈ ℓ2(N, M ∗ d ), Notice that φ∗ : ℓ2(N, M ∗ i k(ρi)k2 ≤ C ′2Xi ≤ C ′2(cid:13)(cid:13)(cid:13)Xi k(ι ⊗ ρi)(U i)ξik2 = C ′2(cid:13)(cid:13)(cid:13)Xi (ι ⊗ ρi)(U i)(cid:13)(cid:13)(cid:13) = C ′2(cid:13)(cid:13)φ∗((ρi))(cid:13)(cid:13)2 2 . 2 (ι ⊗ ρi)(U i)Ω(cid:13)(cid:13)(cid:13) Thus φ∗ is bounded below, which implies that φ is a surjection, and so the proof is complete. 7.2 Proof of the main theorem The final part of the construction is to find a bounded-below homomorphism θ : ℓ2(N, Md) → B(H) for a suitable Hilbert space H. Given ξ, η ∈ H, denote by θξ,η the rank-one operator H → H; γ 7→ (γη)ξ. Recall from [8, Section 3.1] that if we denote by (δi) the canonical orthonormal basis of ℓ2(N0) or ℓ2(N), and define θ0 : ℓ2(N) → B(ℓ2(N0)) by θ0(δi) = θδi,δi+δ0, then θ0 extends by linearity and continuity to a bounded below algebra homomorphism. Set H = Cd⊗ℓ2(N0), and regard ℓ2(N, Md) as Md⊗ℓ2(N). Now define θ = (ι⊗θ0) : ℓ2(N, Md) → Md ⊗ B(ℓ2(N0)) ∼= B(H), an algebra homomorphism. As Md is finite-dimensional, clearly θ is bounded, and bounded below. We can now prove the main result of this section. 26 Proof of Theorem 7.1. Form X as above, and the algebra isomorphism φ : L1(eG)/⊥X → ℓ2(N, Md). Let q : L1(eG) → L1(eG)/⊥X be the quotient map. Let π = θ ◦ φ ◦ q, so that π is a bounded repre- sentation of L1(eG). corepresentation Vπ ∈ L∞(eG)⊗B(H). Set Towards a contradiction, suppose that π is completely bounded, and hence associated to a where the sum converges σ-weakly (think of U as being a block diagonal matrix, with diagonal U i ⊗ θδi,δi ∈ L∞(eG)⊗Md⊗B(ℓ2(N0)) ∼= L∞(eG)⊗B(H), entries U i, which are all unitary). For ω ∈ L1(eG), ξ ∈ Cd and k ∈ N0, we have that π(ω)(ξ ⊗ δk) = δk,0(ω ⊗ ι)(U i)(ξ) ⊗ δi + δk,i(ω ⊗ ι)(U k)(ξ) ⊗ δk. U =Xi∈N Consequently, (ω ⊗ ι)(Vπ − U)(ξ ⊗ δk) = π(ω)(ξ ⊗ δk) − (ω ⊗ ι)(U k)(ξ) ⊗ δk = δk,0(ω ⊗ ι)(U i)(ξ) ⊗ δi. If we now set xi = (ι ⊗ ωδ0,δi)(Vπ − U) ∈ L∞(eG)⊗Md, then Pi x∗ i xi converges σ-weakly to (ι ⊗ ωδ0,δ0)((Vπ − U)∗(Vπ − U)). However, for ω ∈ L1(eG) and ξ ∈ Cd, also (ω ⊗ ι)(xi)(ξ) ⊗ δi =Xi Xi (ω ⊗ ι)(Vπ − U)(ξ ⊗ δ0) =Xi (ω ⊗ ι)(U i)(ξ) ⊗ δi. It follows that xi = U i for all i, and hence each xi is unitary, which contradictsPi x∗ i xi converging. We conclude that π is not completely bounded, and hence cannot be similar to a ∗-representation. Let us make a remark about π∗. As each U i is a unitary corepresentation, we have that (ω♯ ⊗ ι)(U i) = (ω ⊗ ι)(U i)∗ for each ω ∈ L1 operation, it follows that φ ◦ q : L1(eG) → ℓ2(N, Md) is a ∗-homomorphism. Thus ♯ (eG). Thus if we give ℓ2(N, Md) the pointwise ∗ π∗(ω) = π(ω♯)∗ = θ(cid:0)φ(q(ω))∗(cid:1)∗ (ω ∈ L1 ♯ (eG)). It follows that π∗ extends to a bounded homomorphism on L1(eG). It is easy to compute explicitly what θ(a∗)∗ is, for each a ∈ ℓ2(N, Md), and then adapting the argument in the previous proof will show that π∗ is also not completely bounded. Remark 7.7. We now briefly address the case where eG = T ∗ G in Theorem 7.1. Let z denote the canonical Haar unitary generator of L∞(T), and let U ∈ L∞(G) ⊗ Md(C) be our fixed non- trivial irreducible unitary representation of G. For each i ∈ N, consider the tensor product unitary representation of eG given by V i = zi ⊠ U ⊠ z−i ∈ L∞(eG) ⊗ Md(C) (where zi is viewed as a It follows from [48] that the representations {V i}i∈N are one-dimensional representation of T). pairwise inequivalent and irreducible. Our claim is that the above proof for eG = ∗i∈NGi goes through unchanged with the family {U i}i∈N replaced by {V i}i∈N. To see this, observe that the only facts that we used about eG and {U i}i∈N are: (1) each representation U i is d-dimensional, (2) the elements of the co-efficient spaces L∞ i associated to 27 U i are centred with respect to the Haar state, and (3) the von Neumann algebras Ai generated i are pairwise freely independent in (L∞(eG), ϕ). For the family {V i}i∈N, conditions (1) by L∞ and (2) are automatically satisfied. To see that condition (3) is also satisfied, note that if A0 ⊆ L∞(G) ֒→ L∞(eG) denotes the von Neumann algebra generated by the co-efficient space of U, then Ai = ziA0z−i for each i ∈ N. Using this fact, condition (3) is easily seen to be a simple consequence of the definition of free independence (see [30, Definition 5.3], for example) and the fact that z is ∗-free from A0. 8 For the Fourier algebra In this section, we collect some further special cases for the Fourier algebra. These add further weight to the conjecture that every completely bounded representation π : A(G) → B(H) is similar to a ∗-representation. The main results of this section are as follows. Theorem 8.1. Let G be an amenable locally compact group, and let π : A(G) → B(H) be a completely contractive representation. Then π is a ∗-representation. Theorem 8.2. Let G be an amenable locally compact group which contains an open SIN subgroup K ⊆ G. Then every completely bounded representation π : A(G) → B(H) is similar to a ∗- representation. The proofs of the above results rely on the fact that when G is an amenable locally compact group, A(G) is always a 1-operator amenable completely contractive Banach algebra, see [34]. Definition 8.3. Let A be completely contractive Banach algebra, and denote by m : Ab⊗A → A the multiplication map. A net {dα} in Ab⊗A is called a bounded approximate diagonal for A if {dα} is bounded and for every a ∈ A, (b ⊗ c) · a = b ⊗ ca (a, b, c ∈ A). and am(dα) → 0, a · (b ⊗ c) = ab ⊗ c, a · dα − dα · a → 0 where Ab⊗A has the usual A-bimodule structure, An element M ∈ (Ab⊗A)∗∗ is called a virtual diagonal provided that (a ∈ A). Ruan proved in [34] that the a has a bounded approximate diagonal in Ab⊗A if and only if it has a virtual diagonal in (Ab⊗A)∗∗ (and, in turn, these are equivalent to A being operator amenable). Now let L > 0. A is called L-operator amenable if it has a virtual diagonal M ∈ (Ab⊗A)∗∗ such that kMk ≤ L. The operator amenability constant of A is N := inf{L : A is L-operator amenable}. By Alaoglu’s theorem, the operator amenability constant is attained, that is, A has a N-virtual diagonal in (Ab⊗A)∗∗. Since any virtual diagonal has norm at least 1, the operator amenability a · M = M · a, am∗∗(M) = a constant is at least 1. Proof of Theorem 8.1: Since G is amenable, A(G) has a contractive approximate identity. Hence it follows from [7, Remark 7] (see also Lemma 5.2) that π is a ∗-representation if and only if πe is a ∗-representation, where πe = πHe is the essential part of π. So, without loss of generality, we can assume that π is non-degenerate. 28 Put A = π(A(G)). Since A(G) is 1-operator amenable and π is a complete contraction, it follows that A is also 1-operator amenable. It follows from [5, Theorem 7.4.18(ii)] that A is a (commutative) C∗-algebra. (Note that the proof of [5, Theorem 7.4.18(ii)] is for unital operator algebras but it can be easily modified to apply to our case as well). We now follow the proof of [7, Corollary 9]. Indeed, the adjoint operation on a commutative C∗-algebra (in particular, on A) is a complete isometry. Recall that π∗ is defined by π∗(u) = π(u)∗ (u ∈ A(G)). Now, the map u 7→ u is an anti-linear complete isometry, and thus kπ∗kcb = kπkcb. Hence by [7, Theorem 8], there is an invertible operator T ∈ B(H) and a ∗-homomorphism σ : C0(G) → B(H) such that Moreover, T can be chosen so that π(u) = T −1σ(u)T (u ∈ A(G)). 1 ≤ kTkkT −1k ≤ kπk2 cbkπ∗k2 cb ≤ 1. So kTkkT −1k = 1, and so, by replacing T with T /kTk, we can assume that kTk = kT −1k = 1 which means that T is unitary. Hence π(·) = T −1σ(·)T = T ∗σ(·)T We now turn to the proof of Theorem 8.2. For a locally compact group G and an open is a ∗-homomorphism, as claimed. subgroup K of G, let G = Sx∈IxK denote the decomposition of G to distinct left cosets of K (i.e. xK ∩ yK = ∅ if x 6= y). For every element u ∈ A(G) and x ∈ I, write ux = uχxK, where χxK is the characteristic function of the coset xK. Since K is open, each χxK is a norm-one idempotent in the Fourier-Stieltjes algebra B(G) = C ∗(G)∗ [19, Proposition 2.31]. Since A(G) is a closed ideal in B(G), we conclude that ux ∈ A(G). We let A(xK) = A(G)χxK. Since K is open, the canonical embedding of the Fourier algebra A(K) into A(G) (that is, extending functions by zero outside of K) is completely isometric, allowing us to identify A(K) with its image A(eK) unambiguously. In what follows, we shall consider the translation operators Lx : A(K) → A(xK) defined by (Lxu)(y) = u(x−1y), (u ∈ A(K)). Since left translation on A(G) is completely isometric, Lx : A(K) → A(xK) is always a completely isometric algebra isomorphism. Finally, for any homomorphism π : A(G) → B(H), let πx (x ∈ I) denote the restriction of π to the ideal A(xK). Proof of Theorem 8.2: Since G is amenable, A(G) has a contractive approximate identity, and we may again assume that π is non-degenerate. 29 As above, we have that G = Sx∈IxK. For every x ∈ I, the mapping πx ◦ Lx : A(K) → B(H) defines a completely bounded homomorphism of A(K) on H, and so by [7, Theorem 20], πx ◦ Lx extends continuously to a bounded representation σx : C0(K) → B(H) with kσxk ≤ kπk2 cb. (5) (6) Let {fα}α∈∆ ⊂ A(K) be a contractive approximate identity for C0(K). For every u ∈ A(G), we have lim α σx(fα)π(u) = lim α σx(fα(Lx−1ux)) = σx(Lx−1ux) = π(ux). This, together with the fact that π is non-degenerate, implies that Ex, the limit of the net {σx(fα)} in the strong operator topology of B(H), exists. Moreover, for every x ∈ I, Ex is an idempotent and for every x 6= y, ExEy = 0. (4) Now let A = π(A(G)) ⊆ B(H) . Since A is the closure of the range of a completely bounded homomorphism of the operator amenable Banach algebra A(G), it is itself operator amenable. Hence, by combining [5, Proposition 7.4.12(ii)] and [20, Corollary 4.8], it follows that A′′ = A s.o.t , s.o.t where A is commutative and for every x ∈ I, refers to the closure of A in the strong operator topology of B(H). In particular, A′′ Ex ∈ A′′. Therefore, again by combining [5, Proposition 7.4.12(ii)] and [20, Lemma 4.4], there is an invertible operator T ∈ B(H) such that T −1ExT is an orthogonal projection for all x ∈ I. Let Px = T −1ExT and ρx(·) = T −1σx(·)T (x ∈ I). For every x, y ∈ I with x 6= y, it follows from (4) that PxPy = 0. This implies that, for every u, v ∈ C0(K), ρx(u)∗ρy(v) = ρx(u)∗P ∗ x Pyρy(v) = 0, and ρx(u)ρy(v)∗ = ρx(u)PxP ∗ y ρy(v)∗ = 0. Now let u ∈ A(G). For every x, y ∈ I with x 6= y, it follows from (5) that (cid:0)T π(ux)T −1(cid:1)∗(cid:0)T π(uy)T −1(cid:1) = ρx(Lx−1ux)∗ρy(Ly−1uy) = 0. Similarly (6) implies that Hence {(T −1π(ux)T )∗(T −1π(ux)T )}x∈I is a family of pairwise orthogonal positive operators on H, and therefore (cid:0)T π(ux)T −1(cid:1)(cid:0)T π(uy)T −1(cid:1)∗ kT π(u)T −1k2 = k(T π(u)T −1)∗(T π(u)T −1)k =(cid:13)(cid:13)(cid:13)Xx∈I = 0. (T π(ux)T −1)∗(T π(ux)T −1)(cid:13)(cid:13)(cid:13) = sup x∈I k(T π(ux)T −1)∗(T π(ux)T −1)k ≤ kTk2kT −1k2 sup x∈I kπ(ux)k2 ≤ kTk2kT −1k2kπk4 cbkuk2 ∞. Thus T π(·)T −1 extends continuously to a bounded homomorphism of C0(G) into B(H), and hence, so does π. The result now follows. 30 One class of amenable locally compact groups having open SIN subgroups are the amenable, totally disconnected groups. In particular, if we assume further that they are non-compact, then they give us examples of non-SIN groups satisfying the assumption of Theorem 8.2. These groups, for instance, include Fell groups (see [36, Section 2] for more details). A Duality for closed operators In this appendix, we will prove the following result. Proposition A.1. Let G be a locally compact quantum group, and let x, y ∈ L∞(G) be such that hx, ω♯i = hy, ωi for all ω ∈ L1 ♯ (G). Then x ∈ D(S) with S(x)∗ = y. Firstly, note that as S = Rτ−i/2, we have that ω ∈ L1 ♯ (G) if and only if there is ω′ ∈ L1(G) with hx, ω′i = hτ−i/2(x), ω∗i (x ∈ D(τ−i/2)). Set D = L1 Indeed, if this holds, then for x ∈ D(S), we have that hx, ω′ ◦ Ri = hτ−i/2(R(x)), ω∗i = hS(x)∗, ωi, so ω♯ = ω′ ◦ R. The converse follows similarly. It follows that the proposition above is equivalent to the following. Proposition A.2. Let G be a locally compact quantum group, and let x, y ∈ L∞(G) be such that hx, ω′i = hy, ωi for all ω ∈ D. Then x ∈ D(τ−i/2) with y = τ−i/2(x). ♯ (G)} and for ω ∈ D let ω′ = (ω∗)♯. ♯ (G)∗ = {ω∗ : ω ∈ L1 Perhaps the “standard” proof of the proposition would be to “smear” x and y by the one- parameter group (τt); compare [26, Proposition 4.22] or the proof of [28, Proposition 5.26], for example. Instead, we wish to indicate how to prove the result by using closed operators. Given a topological vector space E and a subspace D(T ) ⊆ E, a linear map T : D(T ) → E is closed and densely defined if D(T ) is dense in E, and the graph G(T ) = {(x, T (x)) : x ∈ D(T )} is closed in E × E. Define also G ′(T ) = {(T (x), x) : x ∈ E}, which is closed if and only if G(T ) is. Let E be a Banach space– we shall consider the case when E has the norm topology, and when E ∗ has the weak∗-topology. Let T : D(T ) → E be a linear map. Set D(T ∗) = {µ ∈ E ∗ : ∃λ ∈ E ∗, hλ, xi = hµ, T (x)i (x ∈ D(T ))}, and with a slight abuse of notation, set G(T ∗) to be the collection of such (µ, λ). Notice that: • G ′(−T ∗) = G(T )⊥, and so G(T ∗) is weak∗-closed; • D(T ) is dense if and only if (λ, 0) ∈ G(T )⊥ =⇒ λ = 0, which is equivalent to G(T ∗) being the graph of an operator (that is, the choice of λ being unique); • G(T ) = ⊥(G(T )⊥) = ⊥G ′(−T ∗) and so D(T ∗) is weak∗-dense, because G(T ) is the graph of an operator. It follows that T ∗ is a closed, densely defined operator, for the weak∗-topology. We can similarly reverse the argument, and start with T ∗, and define a closed, densely defined operator T by setting D(T ) = {x ∈ E : ∃y ∈ E, hµ, yi = hT ∗(µ), xi (µ ∈ D(T ∗))} and T (x) = y. Furthermore, these two constructions are mutual inverses, so if we start with T ∗, form T , and then form the adjoint of T , we recover T ∗. We are not aware of an entirely satisfactory reference for this construction, but see [23, Section 5.5, Chapter III]. 31 Proof of Proposition A.2. We apply the preceding discussion to τ−i/2 acting on L∞(G). As τ−i/2 is the analytic generator of the σ-weakly continuous group (τt), it follows from [10] that τ−i/2 is σ-weakly closed and densely defined. Let τ∗,−i/2 be the pre-adjoint, which from the above is hence a (norm) closed, densely defined operator on L1(G). It follows that D = D(τ∗,−i/2) and that ω′ = τ∗,−i/2(ω) for ω ∈ D. The hypothesis of the proposition now simply states that x ∈ D((τ∗,−i/2)∗) and y = (τ∗,−i/2)∗(x). The claim now follows from the fact that (τ∗,−i/2)∗ = τi/2. We remark that, for example, this argument gives a very easy proof that L1 ♯ (G) is dense in L1(G), without the need for a “smearing” argument (compare with the discussion before Proposi- tion 3.1 in [24]). References [1] D. Avitzour, “Free products of C ∗-algebras”, Trans. Amer. Math. Soc. 271 (1982) 423–435. [2] S. Baaj, G. Skandalis, “Unitaires multiplicatifs et dualit´e pour les produits crois´es de C ∗-alg`ebres”, Ann. Sci. ´Ecole Norm. Sup. 26 (1993) 425–488. [3] T. Banica, R. Speicher, “Liberation of orthogonal Lie groups”, Adv. Math. 222 (2009) 1461–1501. [4] E. B´edos, L. Tuset, “Amenability and co-amenability for locally compact quantum groups”, Internat. J. Math. 14 (2003) 865–884. [5] D. Blecher, C. Le Merdy, “Operator algebras and their modules: an operator space approach”, London Mathematical Society Monographs, New Series, 30 (Oxford University Press, Oxford, 2004). [6] D. P. Blecher, R. R. Smith, “The dual of the Haagerup tensor product”, J. London Math. Soc. (2) 45 (1992) 126–144. [7] M. Brannan, E. Samei, “The similarity problem for Fourier algebras and corepresentations of group von Neumann algebras”, J. Funct. Anal. 259 (2010) 2073–2097. [8] Y. Choi, E. Samei, “Quotients of the Fourier algebra, and representations that are not completely bounded”, preprint, see arXiv:1104.2953 [math.FA] [9] E. Christensen, “On nonselfadjoint representations of C ∗-algebras”, Amer. J. Math. 103 (1981) 817– 833. [10] I. Cioranescu, L. Zsid´o, “Analytic generators for one-parameter groups”, Tohoku Math. J. 28 (1976) 327–362. [11] M. Cowling, U. Haagerup, “Completely bounded multipliers of the Fourier algebra of a simple Lie group of real rank one”, Invent. Math. 96 (1989) 507–549. [12] M. Day, “Means for the bounded functions and ergodicity of the bounded representations of semi- groups”, Trans. Amer. Math. Soc. 69 (1950) 276–291. [13] M. Daws, “Multipliers, Self-Induced and Dual Banach Algebras”, Dissertationes Math. 470 (2010) 62pp. [14] M. Daws, “Multipliers of locally compact quantum groups via Hilbert C∗-modules”, J. London Math. Soc. 84 (2011) 385–407. 32 [15] J. de Canniere, U. Haagerup, “Multipliers of the Fourier algebras of some simple Lie Groups and their discrete subgroups”, Amer. Journ. Math. 107 (1985) 455–500. [16] J. Dixmier, “Les moyennes invariantes dans les semi-groupes et leurs applications”, Acta Sci. Math. (Szeged) 12 (1950) 213–227. [17] E. G. Effros, Z.-J. Ruan, “Operator spaces”,London Mathematical Society Monographs, New Series, 23 (Oxford University Press, New York, 2000). [18] E. G. Effros, Z.-J. Ruan, “Operator space tensor products and Hopf convolution algebras”, J. Oper- ator Theory 50 (2003) 131–156. [19] P. Eymard, “L’alg`ebre de Fourier d’un groupe localement compact”, Bull. Soc. Math. France 92 (1964) 181–236. [20] J. A. Gifford, “Operator algebras with a reduction property”, J. Austral. Math. Soc. 80 (2006) 297–315. [21] M. Junge, “Embedding of the operator space OH and the logarithmic ‘little Grothendieck inequality’” Invent. Math. 161 (2005) 225–286. [22] M. Junge, M. Neufang, Z.-J. Ruan, “A representation theorem for locally compact quantum groups”, Internat. J. Math. 20 (2009), 377–400. [23] T. Kato “Perturbation theory for linear operators.” Second edition. Grundlehren der Mathematischen Wissenschaften, Band 132. (Springer-Verlag, Berlin-New York, 1976). [24] J. Kustermans, “Locally compact quantum groups in the universal setting”, Internat. J. Math. 12 (2001) 289–338. [25] J. Kustermans, “Induced corepresentations of locally compact quantum groups”, J. Funct. Anal. 194 (2002) 410–459. [26] J. Kustermans, “Locally compact quantum groups” in Quantum independent increment processes. I, Lecture Notes in Math. 1865, pp. 99–180 (Springer, Berlin, 2005). [27] J. Kustermans, S. Vaes, “Locally compact quantum groups in the von Neumann algebraic setting”, Math. Scand. 92 (2003) 68–92. [28] J. Kustermans, S. Vaes, “Locally compact quantum groups”, Ann. Sci. ´Ecole Norm. Sup. 33 (2000) 837–934. [29] N. Monod, N. Ozawa, “The Dixmier problem, lamplighters and Burnside groups”, J. Funct. Anal. 258 (2010) 255–259. [30] A. Nica and R. Speicher, R., “Lectures on the combinatorics of free probability” LMS Lecture Notes Series 335 (Cambridge University Press, 2006). [31] G. Pisier, “Are unitarizable groups amenable?” in “Infinite groups: geometric, combinatorial and dynamical aspects”, 323–362, Progr. Math. 248 (Birkhauser, Basel, 2005). [32] T. Pytlik, R. Szwarc, “An analytic family of uniformly bounded representations of free groups”, Acta Math. 157 (1986) 287–309. 33 [33] ´E. Ricard, Q. Xu, “Khintchine type inequalities for reduced free products and applications” J. Reine Angew. Math. 599 (2006) 27–59. [34] Z.-J.Ruan, “The operator amenability of A(G)”, Amer. J. Math. 117 (1995), 1449–1474. [35] W. Rudin, “Functional Analysis”, (McGraw-Hill Publishing Co., New York, 1973). [36] V. Runde, N. Spronk, “Operator amenability of Fourier-Stieltjes algebras. II”, Bull. Lond. Math. Soc. 39 (2007) 194–202. [37] M. Takesaki, “Theory of operator algebras. I.”, Encyclopaedia of Mathematical Sciences, 124. Oper- ator Algebras and Non-commutative Geometry, 5. (Springer-Verlag, Berlin, 2002). [38] M. Takesaki, “Theory of operator algebras. II.” Encyclopaedia of Mathematical Sciences, 125. Oper- ator Algebras and Non-commutative Geometry, 6. (Springer-Verlag, Berlin, 2003). [39] M. Takesaki, N. Tatsuuma, “Duality and subgroups. II”, J. Functional Analysis 11 (1972) 184–190. [40] K. F. Taylor, “The type structure of the regular representation of a locally compact group”, Math. Ann. 222 (1976) 211–224. [41] T. Timmermann, “An invitation to quantum groups and duality”, (European Mathematical Society, Zurich, 2008). [42] R. Tomatsu, “Amenable discrete quantum groups”, J. Math. Soc. Japan 58 (2006) 949–964. [43] S. Vaes, “Locally compact quantum groups”, PhD. thesis, Katholieke Universiteit Leuven, 2001. Available from http://wis.kuleuven.be/analyse/stefaan/ [44] S. Vaes, A. Van Daele, “Hopf C ∗-algebras”, Proc. London Math. Soc. 82 (2001) 337–384. [45] D. Voiculescu, “A strengthened asymptotic freeness result for random matrices with applications to free entropy” Internat. Math. Res. Notices 1 (1998) 41–63. [46] D. Voiculescu, “Symmetries of some reduced free product C∗-algebras” in Operator algebras and their connections with topology and ergodic theory (Bu¸steni, 1983) 556–588. Lecture Notes in Math. 1132 (Springer, Berlin, 1985). [47] S. Wang, “Quantum symmetry groups of finite spaces”, Comm. Math. Phys. 195 (1998) 195–211. [48] S. Wang, “Free products of compact quantum groups”, Comm. Math. Phys. 167 (1995) 671–692. [49] S. L. Woronowicz, “Compact quantum groups” in Sym´etries quantiques (Les Houches, 1995) pp. 845–884 (North-Holland, Amsterdam, 1998). [50] S. L. Woronowicz, “From multiplicative unitaries to quantum groups”, Internat. J. Math. 7 (1996) 127–149. [51] S. L. Woronowicz, “Compact matrix pseudogroups”, Comm. Math. Phys. 111 (1987) 613–665. 34 Michael Brannan Department of Mathematics and Statistics, Queen’s University Jeffery Hall, University Avenue, Kingston, Ontario, Canada K7L 3N6. Email: [email protected] Matthew Daws School of Mathematics, University of Leeds, LEEDS LS2 9JT United Kingdom Email: [email protected] Ebrahim Samei Department of Mathematics and Statistics McLean Hall University of Saskatchewan 106 Wiggins Road, Saskatoon, Saskatchewan, Canada S7N 5E6 Email: [email protected] 35

Subsets and Splits

No saved queries yet

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